Advances in Flight Control Systems

ADVANCES IN
FLIGHT CONTROL SYSTEMS

Edited by Agneta Balint
Advances in Flight Control Systems
Edited by Agneta Balint
Published by InTech
Janeza Trdine 9, 51000 Rijeka, Croatia
Copyright 2011 InTech

All chapters are Open Access articles distributed under the Creative Commons
Non Commercial Share Alike Attribution 3.0 license, which permits to copy,
distribute, transmit, and adapt the work in any medium, so long as the original
work is properly cited. After this work has been published by InTech, authors
have the right to republish it, in whole or part, in any publication of which they
are the author, and to make other personal use of the work. Any republication,
referencing or personal use of the work must explicitly identify the original source.
Statements and opinions expressed in the chapters are these of the individual contributors
and not necessarily those of the editors or publisher. No responsibility is accepted
for the accuracy of information contained in the published articles. The publisher
assumes no responsibility for any damage or injury to persons or property arising out
of the use of any materials, instructions, methods or ideas contained in the book.
Publishing Process Manager Ivana Lorkovic

Technical Editor Teodora Smiljanic
Cover Designer Martina Sirotic
Image Copyright Oshchepkov Dmitry, 2010. Used under license from Shutterstock.com
First published March, 2011

Printed in India
A free online edition of this book is available at www.intechopen.com

Additional hard copies can be obtained from [email protected]
Advances in Flight Control Systems, Edited by Agneta Balint

p. cm.
ISBN 978-953-307-218-0
free online editions of InTech
Books and Journals can be found at
www.intechopen.com
Contents
Preface IX
Chapter 1 Adaptive Fight Control Actuators

and Mechanisms for Missiles, Munitions
and Uninhabited Aerial Vehicles (UAVs) 1
Ronald Barrett
Chapter 2 Adaptive Backstepping Flight Control

for Modern Fighter Aircraft 23
L. Sonneveldt, Q.P. Chu and J.A. Mulder
Chapter 3 Hybrid Adaptive Flight Control

with Model Inversion Adaptation 53
Nhan Nguyen
Chapter 4 Application of Evolutionary Computing

in Control Allocation 77
Hammad Ahmad, Trevor Young, Daniel Toal and Edin Omerdic
Chapter 5 Fault Tolerant Flight Control,

a Physical Model Approach 93
Thomas Lombaerts, Ping Chu,
Jan Albert (Bob) Mulder and Olaf Stroosma
Chapter 6 Design of Intelligent Fault-Tolerant Flight

Control System for Unmanned Aerial Vehicles 117
Yuta Kobayashi and Masaki Takahashi
Chapter 7 Active Fault Diagnosis and Major Actuator

Failure Accommodation: Application to a UAV 137
Franois Bateman, Hassan Noura and Mustapha Ouladsine
Chapter 8 Fault-Tolerance of a Transport Aircraft

with Adaptive Control
and Optimal Command Allocation 159
Federico Corraro, Gianfranco Morani and Adolfo Sollazzo
VI Contents
Chapter 9 Acceleration-based 3D Flight Control for UAVs:

Strategy and Longitudinal Design 173
Iain K. Peddle and Thomas Jones
Chapter 10 Autonomous Flight Control System

for Longitudinal Motion of a Helicopter 197
Atsushi Fujimori
Chapter 11 Autonomous Flight Control

for RC Helicopter Using a Wireless Camera 217
Yue Bao, Syuhei Saito and Yutaro Koya
Chapter 12 Hierarchical Control Design of a UAV Helicopter 239

Ali Karimoddini, Guowei Cai, Ben M. Chen, Hai Lin and Tong H. Lee
Chapter 13 Comparison of Flight Control System

Design Methods in Landing 261
S.H. Sadati, M.Sabzeh Parvar and M.B. Menhaj
Chapter 14 Oscillation Susceptibility of an Unmanned Aircraft

whose Automatic Flight Control System Fails 275
Balint Maria-Agneta and Balint Stefan
Preface
Nonlinear problems in flight control have stimulated cooperation among engineers

and scientists from a range of disciplines. Developments in computer technology al-
lowed for numerical solutions of nonlinear control problems, while industrial recogni-
tion and applications of nonlinear mathematical models in solving technological prob-
lems is increasing.
The aim of the book Advances in Flight Control is to bring together reputable research-
ers from dierent countries in order to provide a comprehensive coverage of advanced
and modern topics in flight control not yet reflected by other books. This product com-
prises 14 contributions submitted by 38 authors from 11 dierent countries and ar-
eas. It covers most of the currents main streams of flight control researches, ranging
from adaptive flight control mechanism, fault tolerant flight control, acceleration based
flight control, helicopter flight control, comparison of flight control systems and funda-
mentals. According to these themes the 14 contributions are grouped in six categories,
corresponding to six parts of the book.
The first part of the book is a collection of four chapters (1, 2, 3, 4) dedicated to the adap-
tive flight control.
In Chapter 1 the purpose is to introduce the technical community to some of the adap-
tive flight control mechanisms and structures which have either lead directly to or ac-
tually flown in various classes of missiles, munitions and unhabited aircraft. Although
many programs are not open for publication, glimpses of a select flow have made it to
the public arena at various level. This chapter is centered on airing several supersys-
tem-level advances to flight-proven missiles, munitions and UAVs. Toward that end,
basic models were used to lay out proof-of-concept flight hardware which was then
fabricated, bench and/or ground tested and incorporated in flight vehicles. In the early
years the adaptive aircraft were often simply flown, just to prove the concept worked.
More recently, aircraft using adaptive flight control mechanisms have been flown o
against conventional benchmark aircraft so as to demonstrate systematic superiority,
thereby proving that flight control systems employing adaptive aerostructures result
in some combination of lower power consumption, higher bandwidth, reduction in
total aircraft empty weight, greater flight-speed, shock resistance, lower part count,
lower cost etc. On normal occasions, adaptive aerostructures have ever been shown to
be enabling that is, the aircraft class would not be able to fly without them.
In Chapter 2 an integrated, though cascaded Lyapunov-based adaptive backstepping

approach is taken and used to design a flight path controller for nonlinear high-fidelity
X Preface
F-16 model. Adaptive backstepping allows assuming that the aerodynamic force and
moment models may not be known exactly, and even that they may change in flight
due to causes as structural damage and control actuator failness. To simplify the math-
ematical approximation partition the flight envelope into multiple connecting operat-
ing regions, called hyperboxes, is proposed. In each hyperbox a locally valid linear-
in-parameters nonlinear model is defined. The coecients of these local models can be
estimated using the update laws of the adaptive backstepping control laws. The num-
ber and size of the hyperboxes should be based on a priori information on the physical
properties of the vehicle on hand, and may be defined in terms of state variables as
Mach number, angle of attack and engine thrust. To interpolate between the local mod-
els to ensure smooth model transitions B-spline neural networks are used. Numerical
simulations of various maneuvers with aerodynamic uncertainties in the model and
actuator failures are presented. The maneuvers are performed at several flight condi-
tions to demonstrate that the control laws are valid for the entire flight envelope.
In Chapter 3 a hybrid adaptive flight control method as another possibility to reduce

the eect of high-gain control, is investigated. The hybrid adaptive control blends both
direct and indirect adaptive control in a model inversion flight control architecture. The
blending of both direct and indirect adaptive control is sometimes known as composite
adaptation. The indirect adaptive control is used to update the model inversion con-
troller by two parameter estimation techniques: 1) an indirect adaptive law based on
the Lyapunov theory and 2) a recursive least-squares indirect adaptive law. The model
inversion controller generates a command signal using estimates of the unknown plant
dynamics to reduce the model inversion error. This directly leads to a reduced tracking
error. Because the direct adaptive control only needs to adapt to a residual uncertainity,
its adaptive gain can be reduced in order to improve stability robustness. Simulations
of the hybrid adaptive control for a damaged generic transport aircraft and a pilot-in-
loop flight simulator study show that the proposed method is quite eective in provid-
ing improved command tracking performance for a flight control system.
In Chapter 4 the purpose is to design a compensator using an evolutionary computing

technique (i.e. generic algorithms) to compensate the interaction between control al-
location and actuator dynamics.
The interaction of first order actuator dynamics and control allocation and the struc-
ture of the compensator is established for first order and second order actuator dynam-
ics. The tuning of the compensator parameters using generic algorithm is described.
Simulation and results for tuned compensator are shown for a range of first and second
order actuator dynamics. At the end conclusions are established.
The second part of this book consists on four chapters (5, 6, 7, 8) dedicated to the fault
tolerant flight control.
In Chapter 5 a physical modular approach, where focus is placed on the use of math-
ematical representation based on flight dynamics., is used. All variables and quanti-
ties which appear in the model have a physical meaning and thus are interpretable
in this approach, and on avoids so called black and grey box models, where the con-
tent has no clear physical meaning. Besides the fact that this is a more transparent
approach, allowing the designers and engineers to interpret data of each step, it is
assumed that these physical models will facilitate certification for eventual future real
Preface XI
life applications, since monitoring of data is more meaningfull. Adaptive nonlinear dy-
namic inversion is selected as the preferred adaptive control method in this modular or
indirect approach. The advantage of dynamic inversion is the absence of any need for
gain scheduling and an input-output decoupling of all control channels. Adaptation
of the controller is achieved by providing up-to-date aerodynamic model information
which is collected in a separate identification module.
In Chapter 6 an intelligent flight control system is presented that can discriminate be-
tween faults and natural disturbances in order to evaluate and deal with the situation.
In the control system, an evaluator of flight conditions is designed on the basis of the
dynamics of a controlled object. Moreover, to deal with the situation adaptively, a new
flight-path-planning generator is introduced on the basis of the evaluation. In the study
each system is designed by neural network. The learning based systematical design
method is developed that uses evaluation functions for the subsystems. A six-degree-
of-freedom nonlinear simulation is carried out.
In Chapter 7 a nonlinear UAV model, which allows simulating asymmetrical control

surface failures is presented. In fault-free mode, a nominal control law based on an
Eigenstructure assignment strategy is designed. As the control surface positions are
not measured, a diagnosis system is performed with a bank of observers able to esti-
mate the unknown inputs. However, as the two ailerons oer redundant eects, isolat-
ing a fault on these actuators requires an active diagnosis method. In the last part, a
pre-computed F.T.C. strategy, dedicated to accommodate for a ruddervator failure, is
depicted.
In Chapter 8 a scheme of a fault-tolerant flight control system is proposed. It is com-

posed by the core control laws, based on the DAMF technique, to achieve both robust-
ness and reconfiguration capabilities and the CA system, based on the active set meth-
od, to properly allocate the control eort on the healthy actuators. Numerical results of
a case study with a detailed model of a large transport aircraft are reported to show the
eectiveness of the proposed fault-tolerant control scheme.
The third part of this book consists on one chapter (9) dedicated to the acceleration
based flight control.
In Chapter 9 the manoeuvre autopilot solution presented moves away from the more
mainstream methods, recently reported in the literature and instead returns to the con-
cept of acceleration control which has been commonly used in missile applications, and
to a limited extent in aircraft applications, for a number of decades. However, whereas
acceleration control has traditionally been used within the framework of linearized
flight control (the aircraft or missile dynamics are linearized, typically about a straight
and level flight condition), the algorithms and mathematics presented in this chapter
extend the fundamental acceleration controller for operate equally eectively over the
entire 3D flight envelope. The result of this extension is that the aircraft then reduces to
a point mass with a steerable acceleration vector from a 3D guidance perspective. This
abstraction which is now valid over the entire flight envelope is the key to significantly
reducing the complexity involved in solving the manoeuvre flight control problem.
The fourth part of this book consists on three chapters (10, 11, 12) dedicated to helicop-
ter flight control.
XII Preface
In chapter 10 a flight control design for the longitudinal motion of helicopter is pre-
sented. The flight mission considered here is that a helicopter hovers at a start position,
moves to goal position with keeping a specified cruise velocity and hovers again at the
goal. The characteristics of the linearized equation of the helicopter are changed dur-
ing this flight mission because the trim values of the equation are widely varied. Gain
scheduling is one of candidates to stabilize the vehicle for the entire flight region. The
flight control system is constructed as a double loop control system, which consists of
an inner-loop controller and outer-loop controller. The former is needed for stabilizing
the controlled plant, while the latter is used for tracking the reference which is given
to accomplish the flight mission. To design the inner-loop controller, the longitudinal
motion of a helicopter is modelled by a linear interpolative polytopic model whose
varying parameter is the flight velocity. The outer-loop controller is designed by taking
into consideration the steady-state of the controlled variable.
In chapter 11 an autonomous flight of a RC helicopter with small-wireless camera and

a simple artificial marker which is set on the ground is analyzed. It is thought that a
more wide-ranging flight is possible if the natural feature points are detected from the
image obtained by the camera on RC.
In chapter 12 an analytical approach to design and analyze the whole system including
the inner-loop and outer-loop controllers for a small-scale UAV helicopter is presented.
Here, in the proposed hierarchical structure, the inner-loop is responsible for the in-
ternal stabilization of the UAV in the hovering state and control of the linear velocities
and heading angle angular velocity whereas the outer-loop is used to drive the system,
which is already stabilized by the inner-loop, to follow a desired path while keeping
the system close to the hovering state. This strategy is an intuitive way of controlling
such a complex system. Another reason that compels to employ such a control system
is that the UAV model cannot be fully linearized. In practice, the heading angle of
the UAV cannot be restricted to a small range of variation, it depends on the mission
and could be in any direction. This imposes nonlinearity, which can be modelled by a
transformation. To handle this semi-linearized model, the linear and nonlinear parts
are separated, and then the linear part is controlled by the inner-loop and the nonlinear
part by the outer-loop.
The fifth part of this book consists on one chapter (13) dedicated to the comparison of
flight control systems.
In chapter 13 an overview concerning conventional, fuzzy logic-based, neural net-

based adaptive control techniques is provided to the reader. Practical control schemes
applicable in the area of control system design are introduced. The control laws are
demonstrated on a three-degree-of-freedom simulation with linearized aerodynamic
and engine models. The chapter is focused on aircraft landing manoeuvres. This part
of flight needs to be strongly assisted by human pilot.
The sixths part of this book consists on one chapter (14) dedicated to the oscillation
susceptibility when the control system fails.
In chapter 14 the oscillation susceptibility of an unmanned aircraft whose automatic

flight control system fails, is analysed. The analysis is focused on the research of os-
cillatory movement around the centre of mass in a longitudinal flight with constant
Preface XIII
forward velocity (mainly in the final approach and landing phase). The analysis is
made in a mathematical model defined by a system of three nonlinear ordinary dif-
ferential equations, which govern the aircraft movement around its centre of mass, in
such a flight. This model is deduced from a set of nine nonlinear dierential equations,
governing the movement of the aircraft around its centre of mass. It is shown that in
a longitudinal flight with constant forward velocity, if the elevator deflection outruns
some limits, oscillatory movement appears. This is proved by means of coincidence
degree theory and Mawhins continuation theorem. Numerical results concern an Aero
Data Model in Research Environment and consist in: identification of the range of the
elevator deflection for which steady state exists; computation of the manifold of steady
states; identification of stable and unstable steady states; simulation of successful and
unsuccessful manoeuvres; simulation of oscillatory movements.
I would like to express my sincere gratitude to all chapter authors for the time and ef-
fort they generously gave to the book. I would like to thank the publisher, InTech Edu-
cation and Publishing, for cooperation in publication. Special thanks go to Ms. Ivana
Lorkovic, process manager, who initiated this project.
February 26, 2011.
Agneta M. Balint
West University of Timisoara
Faculty of Physics
Timisoara, 300223,
Romania
1
Adaptive Fight Control Actuators and

Mechanisms for Missiles, Munitions and
Uninhabited Aerial Vehicles (UAVs)
Ronald Barrett
The University of Kansas, Lawrence, Kansas
USA
1. Introduction
The purpose of this chapter is to introduce the technical community to some of the adaptive
flight control mechanisims and structures which have either lead directly to or actually
flown in various classes of missiles, munitions and uninhabited aircraft. Although many
programs are not open for publication, glimpses of a select few have made it to the public
arena at various levels.
This chapter is centered on airing several supersystem-level advances to flight-proven
missiles, munitions and UAVs. Toward that end, basic models were typically used to lay out
proof-of-concept flight hardware which was then fabricated, bench and/or ground tested,
and incorporated in flight vehicles. In the early years, the adaptive aircraft were often
simply flown, just to prove the concept worked. More recently, aircraft using adaptive
flight control mechanisms have been flown off against conventional benchmark aircraft so
as to demonstrate systemic superiority, thereby proving that flight control systems
employing adaptive aerostructures result in some combination of lower power
consumption, higher bandwidth, reduction in total aircraft empty weight, greater flight
speed, shock resistance, lower part count, lower cost etc. On several occasions, adaptive
aerostructures have even been shown to be "enabling;" that is, the aircraft class would not
be able to fly without them.
Although adaptive materials have been known for more than 120 years, the Aerospace
industry has only more recently become aware of their basic characteristics. Starting in the
mid 1980's Ed Crawley's group at MIT laid the foundations of what would become an
active and vibrant branch of aerospace technology. With simple experiments on bending
and extension-twist coupled plates, this group demonstrated that airloads on
aerodynamic surfaces could be actively manipulated by using conventionally attached
piezoelectric actuators.1-3 Although the structures resulting from these projects were not
incorporated in flightworthy aircraft, their significance cannot be overstated as they
introduced the technical community to the possibility of aircraft flight control with
adaptive aerostructures and by doing so effectively started the entire field. Table 1
summarizes a collection of programs which have lead directly to a series of flight proven
adaptive subscale aerospace systems.
2 Advances in Flight Control Systems
Project p=piezoelectric, s=shape memory alloy Agency/

c=component testing only f=flight tested Sponsor
v=entire vehicle configuration tested with adaptive device
1 Bending-Twist Coupled Aeroelastic PZT Plate (1985-87) c,p MIT
2 Adaptive Flap (1987-89) c,p MIT
3 Twist-Active Subsonic DAP Missile Wing (1989-90) c,p ARO
4 Twist-Active DAP Rotor (1990-91) c,p ARO
5 Aeroservoelastic Twist-Active Wing (1990-92) c,p Purdue
6 Twist-Active Supersonic DAP Wing (1991-92) c,p KU
7 Constrained Spar Torque-Plate Missile Fin (1991-92) c,p KU
8 Free-Spar DAP Torque-Plate Fin (1992-93) c,p KU
9 Pitch-Active DAP Torque-Plate Rotor (1992-93) c,p KU
10 Subsonic Twist-Active DAP Wing (1993-94) c,p WL/MNAV
11 Subsonic Twist-Active SMA Wing (1993-94) c,s WL/MNAV
12 Subsonic Camber-Active DAP Wing (1993-94) c,p WL/MNAV
13 Subsonic Camber-Active SMA Wing (1993-94) c,s WL/MNAV
14 Supersonic Twist-Active DAP Wing (1993-94) c,p WL/MNAV
15 Supersonic Twist-Active SMA Wing (1993-94) c,s WL/MNAV
16 Supersonic Camber-Active DAP Wing (1993-94) c,p WL/MNAV
17 Supersonic Camber-Active SMA Wing (1993-94) c,s WL/MNAV
18 UAV with Flexspar Stabilator (Mothra 1993-94) v,f,p AAL
19 Flexspar TOW-2B Wing (1993-94) v,p NSF
20 Solid State Adaptive Rotor (SSAR) (1994-95) c,p NSF
21 Aeroservoelastic Flexspar Fin (1994-95) c,p AAL
22 UAV with Solid State Adaptive Servopaddle Rotor (95-96) v,f,p NSF
23 MAV with Flexspar Stabilator (1994-97) v,f,p DoD CDTO
24 Barrel-Launched Adaptive Munition (1995-97) v,p AFOSR
25 Smart Compressed Reversed Adaptive Munition (1995-97) v,p WL/MNAV
26 Rotationally Active Linear Actuator (RALA 1995-97) c,p WL/Boeing
27 Pitch-Active Torque-Plate Wing (1997-98) c,p AAL
28 Range-Extended Adaptive Munition (1998-99) v,p DARPA
29 Hypersonic Interceptor Test Technology (1998-2000) v,p SMDC/Schafer
30 Coleopter MAV with Flexspar Stabilators (1998-2001) v,f,p DARPA
31 UAVs with Pitch-Active SMA Wings (2000-01) v,f,s AAL
32 Light Fighter Lethality MicroFlex Actuator (2000-01) v,p TACOMARDEC
33 Pitch-Active Curvilinear Fin Actuator (2001-02) c,s AMCOM
34 SC Range-Ex Adaptive Munition (SCREAM) (2000-03) v,p TACOM ARDEC
35 Thunder Multilaminate RALA Fin (2000-03) c,p AFRL/MNAV
36 Centerline Precompression RALA Fin (2000-03) c,p AFRL/MNAV
37 Center Pivot Flexspar Fin (2002-03) c,p ARL
38 PBP StAB (2003-) v,f,p TU Delft/TNO
39 Convertible UAV with PBP Grid Fin (2003-) v,f,p TU Delft/KU
40 Convertible UAV with PBP Turning Vane Flaps(2003-) v,f,p TU Delft/KU
41 Extended-Range Adaptive Gravity Weapons (2003-) v,f,p AFRL/Boeing
42 PBP Supersonic NAV (2004-) v,f,p Lutronix
43 PBP Morphing Wing UAV (2005-) v,f,p TU Delft/KU
Table 1. Summary of Adaptive Aerostructures Projects with Direct Connections to

Flightworthy Adaptive Uninhabited Aerial Vehicles
Adaptive Fight Control Actuators and Mechanisms
for Missiles, Munitions and Uninhabited Aerial Vehicles (UAVs) 3
Fig. 1.1 Historical Overview of Adaptive Aerostructures Projects with Direct Connection to
UAV Flight
2. Missile, munition and supersonic Uninhabited Aerial Vehicle (UAV) flight

control
In the late 1980's, up through 1990, a series of flight control mechanisms were being
explored which could drive flaps with piezoelectric bender elements.4 Although these flight
control devices were tested on a small scale, the goal of the investigators was to eventually
transition these devices to full-scale, inhabited aircraft.
In 1990, a paradigm shift took was triggered which lead one branch of adaptive
aerostructures technology down a large scale/inhabited aircraft path, the other to
uninhabited aircraft. The driving philosophy behind the split was simple:
Aircraft will benefit the most from this line of technology when adaptive materials can be
integrated into aircraft primary structure.
Because there is no current or planned FAA/MilHandbook5 database for materials certification,
these materials -by law- will not be allowed in primary structure of FAR 23/25, 27/29
certified aircraft and will not allowed in the primary structure of most inhabited military aircraft
because structural designers have no A- or B-basis mechanical properties to lay out designs
with.
Because this philosophy split occured so long ago, the uninhabited branch of technology has
had ample time to mature and be demonstrated in flight, thankfully, without the myriad of
restrictions which begjile their inhabited counterparts. Among the uninhabited aircraft
which were and still are being robustly pursued in this obscure corner of the adaptive
aerostructures world are missiles, munitions and UAVs of many sizes. Because funding for
military research was comparatively easier to obtain, some of the first applications
development efforts were centered on missile and fin research. These eventually matured
into system-level designs, bench test and eventually flight tests.
2.1 Early fin and wing designs

In 1990, an important advance in piezoelectric actuation was made.5-8 The invention of
Directionally Attached Piezoelectric (DAP) actuators allowed otherwise isotropic
piezoelectric elements to behave as if they were highly orthotropic. This was achieved by
attaching piezoelectric elements along a longitudinal axis, leaving the sides free to expand
or contract and thereby effectively nulling the lateral stiffness of the element while
maintaining the longitudinal stiffness. This property was especially important as it allowed
direct active manipulation of structures in twist, rather than just bending. Eventually, DAP
actuators were developed with Orthotropy Ratios, OR in excess of 100 which makes them to
this day one of the preferred actuator classes for twist activation, especially when cost,
robustness and ease of manufacturing are considered. This high OR allowed both coupled
and uncoupled structures to be directly manupulated in twist for the first time. If one
considers an active laminate with an isotropic active material bonded to an uncoupled
substrate, one can see clearly from laminated plate theory that the shear and twist activation
terms are zero for CAP elements and non-zero for DAP elements:
Conventionally Attached Piezoelectric Actuators:
A11 A12 0 B11 B12 0 (A11 + A12 )

A12 A11 0 B12 B11 0 (A11 + A12 )
N 0 0 2A 66 0 0 2B 66 0 0
= =
0 (B11 + B12 )
M CAP (1)
B11 B12 0 D11 D12 0
B12 B11 0 D12 D11 0 0 (B + B )
11 12
0 0 2B 66 0 0 2D 66 CAP 0 0 CAP
Directionally Attached Piezoelectric Actuators:
A11 A12 2A16 B11 B12 2B16 (A11 + A12 )

A12 A 22 2A 26 B12 B 22 2B 26 (A 22 + A12 )
N A16 A 26 2A 66 B16 B 26 2B 66 0 (A16 + A 26 )
= =
0 (B11 + B12 )
M DAP (2)
B11 B12 2B16 D11 D12 2D16
B12 B 22 2B 26 D12 D 22 2D 26 0 (B 22 + B12 )

B16 B 26 2B 66 D16 D 26 2D 66 DAP 0 (B16 + B 26 )
DAP
In the late 1980's it was not readily apparent which among the many approaches to adaptive
flight control would actually work in a real flight environment. Several important papers
relating aeroelastic tailoring, geometric sweep and aeroelastic coupling were authored,
comparing CAP and DAP elements.9-11 These papers showed that with aeroelastic coupling,
small deflections could be aeroelastically magnified to control authority levels which were
consistent with aircraft flight control.
Although DAP elements were first integrated into a subscale missile wing in 1989, a new
design incorporating DAP elements on a torque plate was conceived and reduced to practice
in 1990.6-8 This approach allowed large rotations of an aerodynamic shell while the loads
were taken up by a high strength internal structure. Figure 2 shows the Constrained Spar
DAP Torque-Plate Missile Fin. Because the fin was intended for use in a TOW missile, it was
capable of 5 deflections with a corner frequency better than 30Hz and a maximum power
consumption under 50mW.12-14 During the this program, an important design philosophy
was evolved that is still seen as critical advice for adaptive aerostructures designers to this
day:
i. Minimize the amount of work done by adaptive materials on passive structure
ii. Employ aerodynamic and mass balancing principles so that the adaptive structures
resist only transient external airloads and inertial loads.
iii. Use coefficient of thermal expansion mismatch to precompress piezoelectric actuator
elements.
The US Air Force was generous in its support of this area after the concept was initially
proven on the bench and the wind tunnel. The first study examining piezoelectric flight
control for Air Force missiles was centered on low aspect ratio fin and wing manipulation.
In 1993, Wright Laboratory commissioned a study examining DAP and CAP-activated
surfaces. This was one of the first times that finite element methods were used to capture the
behavior of subsonic and supersonic active lifting surfaces. figure 3 shows a FEM model of a
supersonic double-circular arc camber-active DAP fin and NACA 0012 subsonic twist-active
DAP fin.15
Fig. 2. Constrained Spar DAP Torque-Plate Missile Fin13
Fig. 3. Finitel Element Models of Active Supersonic & Subsonic Fins15

Although the technology worked, at the time it did not work well enough as the levels of
manipulation of low aspect ratio aerodynamic surfaces being actuated in twist and camber
was simply far too low. Still, for higher aspect ratios, the design was shown not only to
work, but work well. However, because most missiles employ low aspect ratio aerodynamic
surfaces, a new approach was needed.
2.2 Flexspar: The first flight enabling adaptive system

In an effort to dramatically increase control surface deflections of low aspect ratio fins and in
keeping with principles i. and ii above, a new flight control device was invented and reduced
to practice. This flight control system would be shown to be one of the most useful and widely
used in aircraft employing adaptive materials for flight control. Figure 4 shows the general
layout of the device which was designed around a symmetric, balanced airfoil shell.
Fig. 4. Flexspar Fin Anatomy

Fig. 5. Tip-Joint Flexspar Fin Geometric Parameters

If one examinies the basic construction of a piezoelectric bender element, then a simple
expression can be laid out which relates laminate curvature, k to cure parameters, material
characteristics and active free strain levels of the actuator, L. As has been shown over the
past 20 years, it is important to use thermally induced precompression in all flightworthy
adaptive aerostructures as shown in Equation 3. Equation 4 is the solution of the bending
curvature considering a symmetric, balanced laminate composed of two sheets of CAP
actuator material bonded on either side of an uncoupled substrate. By using the geometry of
Figure 5, the curvatures commanded can be translated into control surface deflections, d. Of
course these estimations assume a frictionless, balanced system operating without geometric
binding (which typically set in on real systems for rotation angles in excess of 15.
A11 + A12
0
0
D11 + D12 lam {}
= A11 + A12
0
0 T
D11 + D12 a 0 a
+{ }
(3)
A11 + A12
0
0 T
D11 + D12
s
{ }
+
0
0 s B11 + B12
B11 + B12
0 0
a
{}
=
( )
Ea ts ta + 2t b ta + t 2a
t + 2t 2 t (4)
Es t 3s
+ Ea
(s b) a
+ (t s + 2t b )t2a + 2 t3a
12 2 3

1 (1 cos(l ))+ (l
otot sin (l o ))
1
o
= 2sin 1 (5)
2l l
These expressions have been regularly used for more than a decade to predict Flexspar
actuator deflection levels with experimental and predicted results typically within 5% of
each other.16-18
Fig. 6. Typical Flexspar Deflection s and Correlation Levels17

The Flexspar actuator configuration is still, to this day, one of the more well used actuation
schemes for flight control. It comes in two major variants: the Tip-Joint Flexspar
arrangement as shown in Fig. 5. This configuration is particularly well suited to low
subsonic flight control using symmetric, balanced aerodynamic surfaces. A high moment
configuration called a Shell-Joint Flexspar actuator is used for high subsonic and faster
control surfaces. In the Fall of 1994 invention disclosures were submitted to Auburn
University where the Flexspar was invented. Because the University failed to either file for
patents or revert the rights to the inventors the Flexspar design can now be used royalty-free
by one and all. The first missile system to incorporate the Flexspar design was the TOW-2B
which used the Flexspar to manipulate wing deflections. Figure 7 shows the TOW-2B missle
mounted in the wind tunnel during testing. Because the Flexspar wings were both
aerodynamically and inertially balanced and they employed symmetric airfoils, the wing
pitch deflections were not affected by airspeed.
Fig. 7. Flexspar-Equipped TOW-2B Missile in Wind Tunnel
2.2 Cruise missile and gravity weapon applications

2.2.1 Smart compressed reversed adaptive munition
In 1995 the first of the gravity weapons programs was commissioned by the US Air Force
employing adaptive flight control mechanisms. The program goal was to compress gravity
weapons into bays the size of the weapon warheads. The driving factor in weapon
compression came from the limited size of the F-22 internal weapon bays which were sized
for AIM-120 air-to-air missiles, but not the existing slate of minimally compressed gravity
weapons. Because conventionally guided gravity weapons of the time could not fit within
the bay, a new approach was undertaken. Although a Flexspar configuration would have
worked well, an antagonistic piezoelectric actuator was selected to drive a fin set as the
design flight speed ranged from mid subsonic through low supersonic. Because of large
shifts in position of center of pressure, the transonic flight regime is often the most
challenging to flight control actuator designers as large rotations at high bandwidth against
high moments are typically prescribed.
Because the designers were allowed to rearrange the weapon configuration itself, a new
configuration was developed which took the most advantage of the 1940's-era Mk83
warhead design. This configuration called for a reversal in warhead direction such that the
base of the warhead would fly first. This would allow for a stable bluff-body relase
(important for weapon egress) and full strakes along the length of the weapon to maintain
suitable levels of CNa and provide a housing to accommodate the antagonistic piezoelectric
actuators. The entire weapon design took advantage of other artifacts including more than
80 in3 of volume in large fuse well. Extensive bench and wind tunnel testing showed that
full 10 fin deflections could be resist all airloads without degradation through the
transonic flight regime at frequencies in excess of 40Hz. Power consumption studies
demonstrated that the actuators could be accommodated over the entire flight duration for
less than 2cc of zinc-air batteries.19,20 Figure 8 shows the weapon configuration and during
wind tunnel testing.
Fig. 8. Smart Compressed Reversed Adaptive Munition (SCRAM)19,20

Space constraints prevent the full chronicling of the program, but suffice it to say that this
effort demonstrated that adaptive aerostructures could be used to increase weapon loadout
by an order of magnitude.
2.2.2 Weapon integration and design technology

In 1997, following the success of the SCRAM program an effort was undertaken to provide
vernier control for a new family of small penetrator weapons. The canard actuator used a
modified form of a shell-joint Flexspar actuator called a Rotationally Active Linear Actuator
(RALA). Although this unclassified program is many years old, details are not yet approved
for public release.
The actuator set designed for the GBU-39 was intended to enhance terminal guidance and
went through extensive bench and wind tunnel testing, showing full deflection capability
through the transonic and low supersonic flight speeds.22,23
Fig. 9. GBU-39 Small Diameter Bomb with Adaptive Canards21
2.2.3 Miniature cruise missile airframe technology demonstrator

Elements of the SCRAM and WIDT programs are currently alive and well in this
USAF/Boeing project. Started in 2003, this effort is centered on demonstrating various
advanced technologies including an adaptive on an extended range weapon system. As
with the GBU-39 WIDT program, technical details have not yet been approved for release.
Fig. 10. Boeing/USAF Miniature Cruise Missile with Adaptive Wings24
2.3 Hard-launched munitions and supersonic Nano Aerial Vehicles (NAVs)

2.3.1 Barrel-Launched Adaptive Munition (BLAM) program
In 1995 the Barrel-Launched Adaptive Munition (BLAM) program was initated to enhance
aerial gunnery by increasing the hit probability and the probability of a kill given a hit in
close-in aerial gunnery. To do this, a proof-of concept nontactical round was developed.
Figure 11 shows the general arrangement of the test article.25-28
The most significant challenges that all hard-launched adaptive munition designs must
overcome is clearly associated with launch loads. With respect to launch loads, all flight,
storage and handling loads are trivial. In addition to launch loads, the round must also be
able to deal with certain environmental factors that are also challenging for aerospace
systems. The short summary below illustrates some of these challenges.
Fig. 11. Barrel-Launched Adaptive Munition (BLAM) Configuration

2.3.1.1 Setback Accelerations
Setback accelerations strongly influence structural layout and material choices and are the
driving condition behind length limitations of actuators for hard-launched actuators.
Although munitions designers use exacting profiles which are specific to a gun, round,
muzzle velocity and charge type combination to predict peak setback accelerations, some
basic boundaries can be gleaned from fundamental physics and empirical trends for first-
order design. If one assumes a constant acceleration along the length of a barrel (a traveling-
charge profile), a round starting from 0 and exiting at a finite muzzle velocity, then a lower
bound below which it is not possible to go:
V2
amin = muzzle (6)
2L barrel
Because there is no upper bound which can be obtained by simple physics, generalized
trends from interior ballistic profiles can be obtained. By examining the acceleration profiles
of instrumented weapons like the Hypervelocity Weapon System, a rough upper bound can
be gleaned for initial design purposes.29
2
1.45Vmuzzle
a peak (7)
L barrel
For larger caliber rounds which are currently fielded, setback accelerations on the order of
5,000 30,000 gs are typical. The Navy's ERGM projectile is typical of the current families of
guided 5 (127mm) cannon shells and is designed for 12,000gs of setback acceleration while
the LCCM projectiles withstand 15,000gs.30
2.3.1.2 Setforward, Balloting and Ringing
Although secondary to setback accelerations, setforward accelerations have extremely
detrimental effects on hard-launch round components and subsystems. Setforward
accelerations are induced as the supersonic round exits the barrel into comparatively still
air. This typically causes a large decelration force on most rounds with a pulse of
approximately one order of magnitude lower than the setback acceleration. Setforward
accelerations are the principal loads which induce buckling and end crush-out failure modes
of many families of adaptive actuators. Reference 30 lists the design setforward accelerations
for the ERGM round to be approximately 2,500gs.
2.3.1.3 Rotational Accelerations

Most of the gun-launched munitions which are in use today are spin stabilized via the
rifling in the barrel. Such rifling induces acceleration rates of several hundred thousand
rad/s2. As is the case with acceleration rates, Froude scaling principles hold when arriving
at estimates for smaller (or larger) rounds, which indicates that lower caliber rounds will
encounter acceleration levels as the reciprocal of the scale factor.
2.3.1.4 Thermal Environment
From Ref. 25 - 28, it can be seen that minimum operational and storage temperatures have a
strong influence on the design of the actuator elements as they rely upon CTE mismatch to
precompress actuator elements. Precompression levels at depressed temperatures must be
carefully matched to thickness ratios and launch accelerations to ensure actuator survival of
setback accelerations. Actuator material selection must be made with strong consideration of
the operational and storage temperatures. Ref. 30 lists temperature environments which are
typical of military munitions as: -40C (-40F) to +63C (145F) storage -9C (+15F) to
+63C (145F) in a tactical/operational environment. References 30 - 36 lay out many other
daunting environmental considerations which must be taken into account when laying out
an adaptive munition.
2.3.1.5 Current Progress
The forefront of modern guided round research has progressed far beyond the BLAM which
is now more than a decade old. These rounds range in size from just a few milimeters in
caliber and up and employ several families of adaptive actuators, guiding rounds with
control authorities of just over 1g to many tens of g's. Not surprising, these projects are
proprietary and/or restricted by ITARs and EARs. Several scattered efforts have
intermittently surfaced, but most projects are still out of the public eye.36
2.3.2 Supersonic Nano Aerial Vehicles (NAVs)

The latest international incarnations of hard-launched aircraft comes in the form of supersonic
Nano Aerial Vehicles (NAVs). Because conventional, subsonic NAVs are highly sensitive to
the many adverse factors which become more severe with reduced scale, it is only logical that
many of those problems can be skirted if the NAV is launched supersonically and flown for
only a few seconds. The missions for these NAVs is nonlethal and primarily centered on
reconnaissance. Figure 12 shows a CAD model of a supersonic NAV.
Fig. 12. Supersonic Nano-Aerial Vehicle (NAV) Design37

The flight control of these aircraft employ some of the latest adaptive actuators. These
advanced "Post-Buckled Precompressed" (PBP) actuators have been shown to generate
significantly higher deflections than conventional actuators and are ideal for small aircraft
like NAVs.37
3. Uninhabited Aerial Vehicle (UAV) & Micro Aerial Vehicle (MAV) flight
control
Subsonic Uninhabited Aerial Vehicle (UAV) and Micro Aerial Vehicle (MAV) flight control
with adaptive aerostructures draws its lineage back to some early experiments done on
flight control devices which produced large changes in commanded lift coefficient.
Although flight control mechanisms in rotary- and fixed-wing subsonic UAVs differ
sharply, they share some common roots and even took advantage of some of the same
families of actuators.
3.1 Fixed-wing UAV beginnings

As part of a National Science Foundation program investigating flight control with adaptive
materials, the first fixed-wing aircraft using adaptive materials for all flight control was
designed, built and flown in September of 1994. Using a Tip-Joint Flexspar configuration
akin to the configuration shown in Figures 4 and 5, the aircraft executed basic maneuvers
expected of micro-light aircraft using vertical and horizontal stabilator flight control.16
Fig. 13. Mothra, The First UAV with Flexspar Stabilators for Flight Control
3.2 Foundations of rotary-wing UAV flight control

The first serious attempts at achieving high control authority deflections of rotor systems
was made in 1992. These early efforts employed the same class of torque-plates that drove
missile fins, but in the roots of rotor blade systems.38 Although the first stages of the Solid
State Adaptive Rotor (SSAR) was not selected for funding by the US Army, the founding
experiments that went into the effort were instrumental in proving its feasibility. In 1994 the
National Science Foundation picked up the project and supported it all the way through
flight test. Figure 14 shows the earliest incarnation of the SSAR on a hover stand.
The first rotary-wing aircraft to fly using adaptive aerostructures for all flight control took to
the air in December of 1996. Space constraints prevent its being chronicled completely, but it
employed a pair of piezoelectric DAP servopaddles mounted on a teetering rotor system.
The DAP servopaddles were driven by a brush contact assembly which allowed the rotor
system to respond to basic cyclic commands at speeds in excess of 2.7/rev. Flight tests were
conducted against a benchmark aircraft, ultimately demonstrating maneuver authority
nearly identical to the baseline aircraft. The big difference was that the aircraft shed 40% of
its flight control system weight, leading to an 8% reduction in total gross weight, a 26% drop
in parasite drag and a cut in part count from 94 components down to 5. Figure 15 shows the
SSAR aircraft Gamara in on the bench and during flight test.40
Fig. 14. Solid State Adaptive Rotor with Root Torque-Plate Actuator38,39
Fig. 15. Gamara, The First Rotary-Wing UAV to Fly with Adaptive Materials for All Flight
Control,39
3.3 The DoD's first MAV -- kolibri

In 1994 the DoD CounterDrug Technology Office commissioned a program that would
eventually lead to the US Military's first Micro Aerial Vehicle (MAV). In 1995, managers at
Lutronix Corp. read about the success of the SSAR program in the technical literature and
decided to fold the adaptive technology used in these various programs into their aircraft.
The mission specification for the MAV called for a 24hr loiter with acoustic signature levels
under 65db at 10ft. Accordingly, an electric-tethered configuration was chosen. Because of
the highly constrained rotor diameter and the limited adaptive materials manufacturing
techniques of the day, it was decided that instead of a DAP-torque-plate rotor configuration,
a Flexspar stabilator configuration would be used. These Flexspar stabilators would be
placed in the rotor wash at the bottom of a graphite-truss frame fuselage counterrotating
coleopter.
Fig. 16. DoD's First MAV -- Kolibri Stabilator, Aircraft & in Flight
Because the Kolibri was so severely weight-critical, any opportunity to shed weight was
taken. Accordingly, the flight control system was a prime target for weight reduction.
Because the aircraft body times to double amplitude were on the order of several tens of
miliseconds, extremely fast actuators were a necessity. The conventional servoactuators on
the open market were simply not fast enough to catch the aircraft and their weights were
prohibitive. Flexspar actuators on the other hand were extremely lightweight with a mass of
only 380mg each and exhibited a corner frequency of 47 Hz -- almost double the bandwidth
required to maintain flight. So for the first time, adaptive flight control mechanisms were
not only enhancing technologies, but they actually enabled an entire class of aircraft to take to
the air. Not surprisingly, the flight control system also included adaptive materials in the
Tokin DO-16 piezoelectric gyros which were used to sense pitch, roll and yaw accelerations.
3.4 The first free-flight rotary-wing MAV

Following the success of the Kolibri, the DoDCDTO handed the program off to DARPA,
thereby kicking off DARPA's much touted MAV program of the late '90's. Although the
Kolibri satisfied the 24 hour hover endurance requirement with a tether, there was a strong
desire to shed the tether. As a result a decision was made to go with an internal combustion
engine. Although the boost in power was tremendous, the noise and structural vibrations
were also boosted by an order of magnitude. As with the Kolibri, Flexspar stabilators and
piezoelectric gyros allowed smooth flight in turbulent atmospheric conditions up through
18kt gusting winds. Figure 17 shows the aircraft overview and in flight. Ultimately, the
aircraft was the only one of three finalist MAVs which successfully flew at DARPA's 3-day
Fly-Off at Quantico Marine Corps Base, Virginia in September of 2000. Fly-offs were also
conducted in several other locations including MacDill AFB, Florida, again, with the
LuMAV appearing as the only rotary-wing/VTOL aircraft in the air. The aircraft
performance specifications included a 15 minute endurance with all-weather capability
including rain rates in excess of 12" (30cm)/hr, dust and snow capability through 18kt gust
fields, flight in 100F, 100% humidity environments and 15g wall-strikes. The aircraft was
designed to carry a single submicrovideo camera an a GPS navigation suite.44
Fig. 17. Lutronix MAV Configuration & Flying at MacDill AFB, Florida
3.5 The XQ-138 convertible UAV

As the LuMAV project came to a resoundingly successful conclusion, a follow-on design
was sought. Although the LuMAV was clearly quite capable and flew circles around
competitors, it was not selected for follow-on funding by DARPA. Instead, DARPA
managers recommended approaching Boeing, which in turn recommended a new corporate
partner on the Future Combat System (FCS) program, Singapore Technologies Engineering.
A new aircraft configuration was independently conceived and reduced to practice in the
summer of 2001 which employed the best of the rotary-wing and fixed-wing worlds.
Impressed with the new aircraft performance and promise, ST Engineering purchased the
rights to the aircraft and paid for its production. Initially, the XQ-138, a convertible
coleopter, used conventional flight control actuators in its grid-fin empennage and turning
vane flaps. Following component development efforts, these actuators were replaced by
piezoelectric mechanisms. Figure XX18 shows the overall configuration of the XQ-138.45
Fig. 18. The 11" Rotor Diameter XQ-138a Overall Configuration

More than 300 flight tests were conducted in all types of atmospheric conditions including
gusts through 26 kts, rain at 9"/hr, 100F (38C) heat at 100% humidity, winter flights in
snow at 22F (-6C), dust, sand and finally flight in smoke plumes from exploded tanks.
Figure 19 shows a sequence of photos of the aircraft flying off an FCS prototype on Redstone
Arsenal, Alabama in April of 2002. These tests were followed by live-fire Battle-Damage
Assessment (BDA) tests on the Hellfire Range of Eglin AFB in May of 2002. Although all
variants of the aircraft used piezoelectric gyros at the core of its GNC package, the
conversion of the aircraft to piezoelectric flight controls lent marked improvements in all
aspects, eventually leading to a total empty weight savings in excess of 10% which allowed
the range to be expanded by 30nmi to 100nmi and more than an hour and a half of
endurance. Variants of the aircraft survive today as Singapore Technologies Engineering's
FanTail UAV line of aircraft.
Fig. 19. The Piezoelectric FCS-Equipped XQ-138 Convertible Coleopter UAV
3.6 Low and zero net passive stiffness structures

In 2004 an important innovation was made which dramatically improved the performance
of adaptive aerostructures. It was discovered how to simultaneously improve both
deflection and force with minimal weight volume and cost penalties.46 This discovery was
shown to dramatically improve flight control actuator performance and has been integrated
into a number of flight control systems.47-53 Several variants of Low Net Passive Stiffness
(LNPS), Zero Net Passive Stiffness (ZNPS) as Post-Buckled Precompressed (PBP) actuator
elements have been built into aircraft which are currently undergoing development.
4. Nomenclature
Symbol Description Units
A,B,D in-plane, coupled, bending laminate stiffnesses N/m, N, N-m
B actuator width mm (in)
Dp actuator power density per unit mass, volume, cost W/g, W/cc, W/$
E stiffness GPa (msi)
F applied end force N (lbf)
M Mach number ~
M applied moment vector N-m/m (in-lb/in)
N applied force vector N/m (lb/in)
OR Orthotropy Ratio = EL/ET ~
t thickness mm (in)
y out of plane displacement dimension mm (in)
z through thickness dimension mm (in)
angle of attack deg
PBP beam angle deg
o PBP end rotation angle deg
laminate in-plain strain strain
laminate curvature rad/m (rad/in)
piezoelectric free element strain strain
stress GPa (msi)
Subscripts
a actuator
b bond
c cost
ex external
l laminate
L longitudinal
m mass
s substrate
t thermally induced
T transverse
v volume
Acronyms
AAL The Adaptive Aerostructures Laboratory
AFOSR US Air Force Office of Scientific Research,
AFRL Air Force Research Lab
AMCOM US Army Aviation and Missile Command
ARO US Army Research Office
DAP Directionally Attached Piezoelectric
DARPA Defense Advanced Research Projects Agency

DoD CDTO Department of Defense CounterDrug Technology Office
FCS Future Combat System
LAV Light Armored Vehicle
MAV micro aerial vehicle
NAV nano aerial vehicle
NSF National Science Foundation
PZT lead zirconate titanate
SMDC Space and Missile Defense Command
TACOM-ARDEC US Army Tank-Automotive and Armaments
Command/Armament Research, Development and Engineering Center
TNO Toegepast Natuurwetenschappelijk Onderzoek
TU Delft The Technical University of Delft, Netherlands
UAV uninhabited aerial vehicle
WL Wright Laboratory
5. References
[1] Crawley, E., Lazarus, K. and Warkentin, D., "Embedded Actuation and Processing in
Intelligent Materials," presented at the 2nd international Workshop on Composite
Materials and Structures for Rotorcraft, Troy, NY, Sept., 1989.
[2] Lazarus, K., and Crawley, E., "Multivariable Active Lifting Surface Control using Strain
Actuation: Analytical and Experimental Results," paper presented at the Third
International Conference on Adaptive Structures, sponsored by the ASME, 9 - 11
November, 1992, San Diego.
[3] Lazarus, K. B., Crawley, E. F., and Bohlmann, J. D., "Static Aeroelastic Control Using
Strain Actuated Adaptive Structures," Proceedings of the First Joint U.S./Japan
Conference on Adaptive Structures, Maui, Hawaii, October, 1990.
[4] Spangler, R. L., and Hall, S. R., "Piezoelectric Actuators for Helicopter Rotor Control,"
Paper presented at the 31st Structures, Structural Dynamics and Materials Conference,
Long Beach, California, April, 1990.
[5] Barrett, R., "Intelligent Rotor Blade Actuation through Directionally Attached
Piezoelectric Crystals, 46th American Helicopter Society National Conference and
Forum, Washington, D.C., May, 1990.
[6] Barrett, R., Intelligent Rotor Blade and Structures Development using Piezoelectric
Crystals, MS Thesis, the University of Maryland, College Park, Maryland 1990.
[7] Barrett, R., Method and Apparatus for Sensing and Actuating in a Desired Direction,
US Pat. 5,440,193, Aug. 1995.
[8] Barrett, R., "Actuation Strain Decoupling Through Enhanced Directional Attachment in
Plates and Aerodynamic Surfaces," proceedings of the First European Conference
on Smart Structures and Materials, Glasgow, Scotland, 12 - 14 May 1992, IOP
Publishing, Bristol, UK 1992, pp. 383 - 386.
[9] Ehlers, S. M., and Weisshaar, T. A., "Static Aeroelastic Behavior of an Adaptive
Laminated Piezoelectric Composite Wing," AIAA-90-1078-CP, April, 1990, pp.
1611-1623.
[10] Ehlers, S. M., and Weisshaar, T., "Adaptive Wing Flexural Axis Control," paper
presented at the Third International Conference on Adaptive Structures, sponsored by
the ASME, 9 - 11 November, 1992, San Diego.
[11] Ehlers, S. M., and Weisshaar, T. A., "Effect of Material Properties on Static Aeroelastic
Control," paper presented at the 33rd Structures, Structural Dynamics and
Materials Conference, Dallas, Texas, 15 April, 1992.
[12] Barrett, R., "Active Plate and Missile Wing Development Using EDAP Elements," Journal
of Smart Materials and Structures, Institute of Physics Publishing, Ltd., Techno
House, Bristol, UK, Vol. 1, No. 3, pp. 214226, ISSN 096.
[13] Barrett, R., "Active Plate and Missile Wing Development Using DAP Elements," AIAA
Journal,March, 1994.
[14] Barrett, R., "Active Composite Torque-Plate Fins for Subsonic Missiles," paper presented
at the Dynamic Response of Composite Structures Conference, New Orleans,
Louisiana, August 30 - September 1, 1993.
[15] Barrett, R., "Advanced Low-Cost Smart Missile Fin Technology Evaluation," Contractor
Report to the United States Air Force Armament Directorate, Eglin Air Force Base,
Florida, Contract No. F0 8630-93-C-0039, BAT, November 1993.
[16] Barrett, R., Brozoski, F., and Gross, R. S., "Design and Testing of a Subsonic All-Moving
Adaptive Flight Control Surface," AIAA Journal, published by the AIAA, Reston,
VA,Volume 35, No. 7, July 1997, pp. 1217 - 1219.
[17] Barrett, R. and Brozoski, F., "Missile Flight Control using Active Flexspar Actuators,"
Journal of Smart Materials and Structures, Institute of Physics Publishing, Ltd.,
Techno House, Bristol, UK, Vol. 5, No. 2, March 1996, pp. 121-128.
[18] Barrett, R., "Active Aeroelastic Tailoring of an Adaptive Flexspar Stabilator," Journal of
Smart Materials and Structures, Vol. 5, No. 6 December 1996, Techno House, Bristol,
UK, 1996, pp. 723 730.
[19] SCRAM Report Barrett, R., Design, Construction and Testing of a Proof-of-Concept
Smart Compressed Reversed Adaptive Munition, Final Report to the USAF
Armament Directorate, Wright Laboratory, Eglin AFB, FL contract no. AF-FO8630-
95-K-0079, September, 1996.
[20] Barrett, R., and Stutts, J., Development of a Piezoceramic Flight Control Surface
Actuator for Highly Compressed Munitions, proceedings of the 39th Structures,
Structural Dynamics and Materials Conference 20 - 23 April 1998, Long Beach, CA,
AIAA, Washington, D.C. 1998, paper no. AIAA-98-2034.
[21] Anon., "GBU-39 SDB - Small Diameter Bomb,"
http://www.globalsecurity.org/military/systems/munitions/sdb.htm, January
2006.
[22] Barrett, R., Construction and Test Report for the Rotationally Active Linear Actuator
(RALA) Adaptive Canard, Final Report for McDonnell Douglas, St. Louis, MO
contract no. FO 8630-95-C-0009, August, 1997.
[23] Knowles, G., R. Barrett and M. Valentino, Self-Contained High Authority Control of
Miniature Flight Control Systems for Area Dominance, SPIE 11th International
Symposium on Smart Structures and Materials, San Diego, CA, Mar. 2004.
[24] Avila, C.A., "Precision Engagement," http://www.aviationnow.com/

conferences/html/ad03/avila_session_4b.pdf January 2006.
[25] Barrett, R., "Invention and Evaluation of the Barrel Launched Adaptive Munition
(BLAM)," final report for USAF contract no. F-49620-93-C-0063, USAF Wright
Laboratory Flight Vehicles Branch, WL/MNAV August, 1995.
[26] Barrett, R., and Stutts, J. "Barrel-Launched Adaptive Munition BLAM Experimental
Round Research," final report for USAF contract no. F-49620-C-0063, USAF Wright
Laboratory Flight Vehicles Branch, WL/MNAV February, 1997.
[27] Winchenbach, G., "Cone Aerodynamics Test", Aeroballistic Research Facility Ballistic
Spark Range Technical Report, USAF Wright Laboratory Flight Vehicles Branch,
WL/MNAV June 1996.
[28] Barrett, R. and Stutts, J., "Modeling, Design and Testing of a Barrel-Launched Adaptive
Munition," proceedings of the 4" Annual SPIE Symposium on Smart Structures and
Materials, San Diego, CA, 3-6 March 1997.
[29] Anon., "Actual Results and Analysis," Hypervelocity Weapon System (HVWS) Field
Experiment 1 Final Report, Volume 1 Technical Report, published by Technology
Gateways, Inc., Niceville, Florida, 1993.
[30] Hertlein, Robert and Mark Miner, Extended Range Guided Munition (ERGM) Safe and
Arm Device and Height-of-Burst Sensor, paper presented at the NDIA Fuse
Conference, 9 April 2993, New Orleans, LA.
(http://www.dtic.mil/ndia/2003fuze/hertlien.pdf
[31] Pike, J. Trajectory Correctable Munition (TCM), published by Global Security.org,
Alexandria, Virginia, October, 2002.
(http://www.globalsecurity.org/military/systems/munitions/tcm.htm).
[32] Anon., M732A2 Proximity Fuse and M782 Multi-Option Fuse for Artillery (MOFA)
Data Sheets, published by Alliant Techsystems, Inc. Edina, Minnesota, 2003.
(http://www.atk.com/productsPrecision/descriptions/products/fuses/
artilleryfuzes.htm)
[33] Lee, Gary, Range-Extended Adaptive Munition (REAM) Final Report from Lutronix
Corporation to the Defense Advanced Research Projects Agency (DARPA), Del
Mar, California, April 1999.
[34] Lee, Gary, 40/50 Caliber Range-Extended Adaptive Munition (REAM) Final Report
from Lutronix Corporation to the US Army TACOM-ARDEC Del Mar, California,
May 2000.
[35] Lee, Gary, Shipborne Countermeasure Range-Extended Adaptive Munition
(SCREAM), Final Report from Lutronix Corporation to the US Army TACOM-
ARDEC, Del Mar, California, May 2003.
[36] Rabinovitch, O., and J. R., Vinson, On the Design of Piezoelectric Smart Fins for Flight
Vehicles, Journal of Smart Materials and Structures, IOP Publishing, Ltd., Techno
House, Bristol, UK, Vol. 12, No. 5, pp. 686 695, Oct., 1993.
[37] Barrett, R. and P. Tiso, PBP Adaptive Actuator Device and Embodiments,
International Patent Application number PCT/NL2005/000054, via TU Delft, 18
February 2005.
[38] Barrett, R., High Bandwidth Electric Rotor Blade Actuator Study, Phase I SBIR
proposal Submitted to the U.S. Army Aviation Systems Command, St. Louis, MO
27 June 1992.
[39] Barrett, R. and Stutts, J., Design and Testing of a 1/12th Scale Solid State Adaptive
Rotor, Journal of Smart Materials and Structures, Vol. 6, No. 4 August 1997, Techno
House, Bristol, UK, 1997, pp. 491 - 497.
[40] Barrett, R., Frye, P., and Schliesman, M., Design, Construction and Characterization of
a Flightworthy Piezoelectric Solid State Adaptive Rotor, Journal of Smart Materials
and Structures Vol. 7, No. 3, June 1998, pp. 422431.
[41] Lee, G., Design and Testing of the Kolibri Vertical Take-Off and Landing Micro Aerial
Vehicle, final report for the Department of Defense CounterDrug Technology
Office, November 1997.
[42] Barrett, R., and Howard, N., Adaptive Aerostructures for Subscale Aircraft, refereed
proceedings of the 20th Southeastern Conference on Theoretical and Applied
Mechanics, Pine Mountain, GA, 17 April 2000.
[43] Barrett, R., Burger, C. and Melin J. P., Recent Advances in Uninhabited Aerial Vehicle
(UAV) Flight Control with Adaptive Aerostructures, 4th European Demonstrators
Conference, 10 15 Dec. 2001, Edinburgh, Scotland.
[44] Barrett, R. and Lee, G., Design Criteria, Aircraft Design, Fabrication and Testing of
Sub-Canopy and Urban Micro-Aerial Vehicles, AIAA/AHS International Powered
Lift Conference, Alexandria, Virginia, Nov. 2000.
[45] Barrett, R., Convertible Vertical Take-Off and Landing Miniature Aerial Vehicle, US
patent 6,502,787, 22 Feb. 2002.
[46] Barrett, R. and P. Tiso, PBP Adaptive Actuator Device and Embodiments,
International Patent Application number PCT/NL2005/000054, via TU Delft, 18
February 2005.
[47] Barrett, R., McMurtry,R., Vos, R., Tiso, P., and De Breuker, R., "Post-Buckled
Precomprecompressed Piezoelectric Flight Control Acutator Design, Development
and Demonstration," Journal of Smart Materials and Structures, Vol. 15, No. 5,
October 2006, pp. 1323 - 1331.
[48] Vos, R., De Breuker, R., Barrett, R., and Tiso, P., Morphing Wing Flight Control via
Postbuckled Precopressed Piezoelectric Actuators, Journal of Aircraft, Vol. 44, No. 4,
pp. 1060 - 1068, July-August 2007.
[49] Vos, R., Barrett, R., De Breuker, R. and Tiso, P., "Post-buckled Precompressed Elements:
A New Class of Control Actuators for Morphing Wing UAVs," Journal of Smart
Materials and Structures, Vol. 16, No. 3, June 2007, pp. 919 - 926.
[50] Barrett, R., Post-Buckled Precompressed (PBP) Subsonic Micro Flight Control
Actuators," Journal of Smart Materials and Structures, vol. 17, no. 5, 10pp., October
2008.
[51] Vos, R., and Barrett, R., "Dynamic Elastic Axis Shifting: An Important Enhancement of
Piezoelectric Postbuckled Precompressed Actuators," The Journal of the American
Institute of Aeronautics and Astronautics, Vol. 48, No. 3 March 2010.
[52] Barrett, R., "Hypermaneuverability and Visual Cloaking; New Adaptive Aerostructures
Technologies for Uninhabited Aerial Vehicles (UAVs)," The Aeronautical Journal,
Royal Aeronautical Society, London, UK, Vol. 114, No. 1156, June 2010.
[53] Vos, R., and Barrett, R., "Post-Buckled Precompressed Techniques In Adaptive
Aerostructures: An Overview," MD-08-1306 Journal of Mechanical Design, Vol. 132,
Issue 3, March 2010.
2
Adaptive Backstepping Flight Control for

Modern Fighter Aircraft
L. Sonneveldt, Q.P. Chu and J.A. Mulder
Delft University of Technology
The Netherlands
1. Introduction
Inertial trajectory control is essential for UAVs which must follow predetermined paths
through three-dimensional space (Healy and Liebard, 1993, Kaminer et al., 1998, Boyle et al.,
1999, Singh et al., 2003, Tsach et al., 2003, Ren and Beard, 2004, Wegener et al., 2004, Ren and
Atkins, 2005, No et al., 2005, Clough, 2005, Papadales et al., 2005, Narasimhan et al., 2006,
Kaminer et al., 2007). Other applications of trajectory control include formation flight, aerial
refueling, and autonomous landing maneuvers (Pachter et al., 1994, Proud et al., 1999,
Fujimori et al. 2000, Singh et al., 2000, Pachter et al., 2001, Wang et al., 2008).
Two different approaches can be distinguished in the design of these trajectory control
systems. The most popular approach is to separate the guidance and control laws: a given
reference trajectory is converted by the guidance laws to velocity and attitude commands for
the autopilot, which in turn generates the actuator signals (Ren and Beard, 2004, Pachter et
al., 1994, Pachter et al., 2001). Usually, the assumption is made that the autopilot response to
heading and airspeed commands is first order in nature to simplify the design.
The other design approach is to integrate the guidance and control laws into one system, in
order to achieve better stability guarantees and improved performance. Kaminer et al. (1998)
use an integrated guidance and control approach to trajectory tracking in which the
trimmed flight conditions along the reference trajectory are the command input to the
tracking controllers. Singh (2003) uses a combination of sliding-mode control and adaptive
control.
In this chapter an integrated, though cascaded Lyapunov-based adaptive backstepping
(Krsti et al., 1992, Singh and Steinberg 1996) approach is taken and used to design a flight-
path controller for a nonlinear high-fidelity F-16 model. Adaptive backstepping allows
assuming that the aerodynamic force and moment models may not be known exactly, and
even that they may change in flight due to causes as structural damage and control actuator
failures. There is much literature available on adaptive backstepping control system design
for aircraft and missiles (see, for example, (Singh and Steinberg, 1996, Hrkegrd, 2003,
Farrell et al., Kim et al., 2004, Shin and Kim, 2004, Farrell et al., 2005, Sonneveldt, et al., 2006,
Sonneveldt, et al. 2007)). Most of these designs consider control of the aerodynamic angles ,
, and . Due to the higher relative degree, however, the design of trajectory controllers as
discussed here is much more complicated, as the required analytical calculation of the
derivatives of the intermediate control variables leads to a rapid explosion of terms. This
phenomenon is the main motivation for the authors of (Singh et al., 2003) to select a sliding-
mode design for the outer feedback loops. Another disadvantage of (adaptive) backstepping
flight control system design is that the contribution of the control-surface deflections to the
aerodynamic forces cannot be taken into account. For these reasons, the constrained
adaptive backstepping approach of (Farrell et al., 2005, Sonneveldt et al., 2007, Yip 1997) is
used here. This method makes use of command filters to calculate the derivatives of the
intermediate controls, which greatly simplifies the design. Additionally, these filters can be
used to enforce magnitude and rate limits on the state and input variables.
To simplify the mathematical approximation of the unknown aerodynamic force and
moment characteristics, we propose to partition the flight envelope into multiple connecting
operating regions called hyperboxes. In each hyperbox a locally valid linear-in-the-
parameters nonlinear model is defined. The coefficients of these local models can be
estimated using the update laws of the adaptive backstepping control laws. The number and
size of the hyperboxes should be based on a priori information on the physical properties of
the vehicle on hand, and may be defined in terms of state variables as Mach number, angle
of attack and engine thrust. In this study we use B-spline neural networks (Cheng et al.,
1999, Ward et al., 2003) to interpolate between the local models to ensure smooth model
transitions. Numerical simulations of various maneuvers with aerodynamic uncertainties in
the model and actuator failures are presented. The maneuvers are performed at several
flight conditions to demonstrate that the control laws are valid for the entire flight envelope.
The chapter is outlined as follows. First, the nonlinear dynamics of the aircraft model are
introduced in Sec. II. In Sec. III the adaptive control system design is presented decomposed
in four cascaded feedback-loop designs. The aerodynamic model identification process
including the B-spline neural networks is discussed in Sec. IV. Section V validates the
performance of the control laws using numerical simulations performed in
MATLAB/Simulink. A summary of the results and the conclusions are given in Sec. VI.
Finally, an appendix on the concept of constrained adaptive backstepping is included.
2. Aircraft model description

The aircraft model used in this study is that of an F-16 fighter aircraft with geometry and
aerodynamic data as reported in (Nguyen et al., 1979). The aerodynamic data in tabular
form have been obtained from wind-tunnel tests and are valid up to Mach 0.6 for the wide
range of -20 deg 90 deg and -30 deg 30 deg. The control inputs of the model are
the elevator, ailerons, rudder, and leading-edge flaps, as well as the throttle setting. The
leading-edge flaps are not used in the control design. The control-surface actuators are
modeled as first-order low-pass filters with rate and magnitude limits as given in
(Sonneveldt et al., 2007). Before giving the equations of motion for the F-16 model, some
reference frames to describe the aircraft motion are needed. The reference frames used in
this paper are the Earth-fixed reference frame FE , used as the inertial frame; the vehicle-
carried local Earth reference frame FO , with its origin fixed in the center of gravity of the
aircraft, which is assumed to have the same orientation as FE ; the wind-axes reference
frame FW , obtained from FO by three successive rotations of , , and ; the stability-axes
reference frame FS , obtained from FW by a rotation of ; and the body-fixed reference
frame FB , obtained from FS by a rotation of , as is also indicated in Fig. 1. The body-fixed
reference frame FB can also be obtained directly from FO by three successive rotations of yaw
angle , pitch angle , and roll angle . More details and transformation matrices are given
in, for example, (Lewis and Stevens, 1992, Cook, 1997).
Adaptive Backstepping Flight Control for Modern Fighter Aircraft 25
Fig. 1. Aircraft reference frames

Assuming that the aircraft has a rigid body, which is symmetric around the XZ body-fixed
plane, the relevant nonlinear coupled equations of motion can be described by (Lewis and
Stevens, 1992):
V cos cos

X 0 = V sin cos (1)
V sin
1
( D + T cos cos ) g sin
m
1

X1 = L sin + Y cos + T ( sin sin cos sin cos )
(2)
mV cos

1 g

mV L cos Y sin + T ( cos sin sin + sin cos )
V cos
cos sin
0 0 sin + cos sin tan cos tan
cos cos
cos sin cos
X 2 = cos tan 1 sin tan X 3 + 0 X1 (3)
cos cos
sin 0 cos
0 cos cos sin

(c r + c p) q + c L + c N + h q
1 2 3 ( 4 e )

( )
X 3 = c 5 pr c 6 p r + c7 M he r

2 2
( ) (4)

(
( c 8 p c 2 r ) q + c 4 L + c 9 N + he q )
T T T T
where X 0 = x y z , X 1 = V , X 2 = , X 3 = p q r , and the
definition of the inertia terms ci ( i = 1," ,9 ) is given in, for example, (Sonneveldt et al., 2007).
The engine angular momentum he is assumed to be constant. These 12 differential equations

are sufficient to describe the complete motion of the aircraft; other states such as the attitude
angles , , and are functions of X 3 , and their dynamics can be expressed as

1 sin tan cos tan

= 0 cos sin X 3 (5)
sin cos
0
cos cos
The thrust model of (Nguyen et al., 1979) is implemented, which calculates the thrust as a
function of altitude, Mach number, and throttle setting t . This model is given in tabular
form. The aerodynamic forces L, Y, and D (expressed in the wind reference frame FW ) and
moments L , M , and N (expressed in body fixed frame FB ) are summations of the various
aerodynamic contributions stored in lookup tables. As an example, the pitch moment M is
given by

( )
M = qSc C m ( , , e ) + C ZT xcgr xcg + C mLEF 1 LEF
25
+

(6)

+
qc
2VT
(
)
C mq ( ) + C mq ( ) 1 LEF
LEF
25

+ C m ( ) + C mds ( , e )

Other aerodynamic forces and moments are given in similar form; for a detailed discussion,
see (Nguyen et al., 1979).
3. Adaptive control design

In this section we aim to develop an adaptive guidance and control system that
( )
T
asymptotically tracks a smooth prescribed inertial trajectory Y ref = x ref y ref z ref with
position states X 0 = ( x y z ) . Furthermore, the sideslip angle has to be kept at zero to
T
enable coordinated turning. It is assumed that the reference trajectory

( )
T
Y ref
= x ref
y ref
z ref
satisfies
x ref = V ref cos ref (7)
y ref = V ref sin ref (8)
with V ref , ref , zref , and their derivatives continuous and bounded. It also assumed that the
components of the total aerodynamic forces L, Y, and D and moments L , M and N are
uncertain, and so these will have to be estimated. The available controls are the control-
surface deflections ( e a r ) and the engine thrust T. The Lyapunov-based control
T
design based on (Farrell et al., 2005, Sonneveldt et al., 2007) is done in four feedback loops,
starting at the outer loop.
3.1 Inertial position control

We start the outer-loop feedback control design by transforming the tracking control
problem into a regulation problem:
z01 cos sin 0

Z0 = z02 = sin cos (
0 X 0 Y ref ) (9)
z03 0 0 1
where we introduce a vehicle carried vertical reference frame with origin in the center of
gravity and X-axis aligned with the horizontal component of the velocity vector (Ren and
Beard, 2004, Proud et al., 1999). Differentiating Eq. (9) now gives
(
V + z02 V ref cos ref

)
(
Z 0 = z01 + V ref sin ref )

(10)
z ref V sin

We want to control the position errors Z0 through the flight-path angles and , and the
total airspeed V. However, from Eq. (10) it is clear that it is not yet possible to do something
about z02 in this design step. Now we select the virtual controls
( )
V des,0 = V ref cos ref c01 z01 (11)
c 03 z03 z ref
des,0 = arcsin , < < (12)
V 2 2
where c01 > 0 and c03 > 0 are the control gains. The actual implementable virtual control
signals V des and des , as well as their derivatives, V des and des , are obtained by filtering the
virtual signals with a second-order low-pass filter. In this way, tedious calculation of the
virtual control derivatives is avoided (Swaroop et al., 1997). An additional advantage is that
the filters can be used to enforce magnitude or rate limits on the states (Farrell et al., 2003,
2007). As an example, the state-space representation of such a filter for V des,0 is given by
q2
q 1 ( t )
= V2 (13)
q2 ( t ) 2 VV SR 2

M (
S V des,0 )
q1 q2

V V
V des q1
des = (14)
V q2
where SM ( ) and SR ( ) represent the magnitude and rate limit functions as given in (Farrell
et al., 2007). These functions enforce the state V to stay within the defined limits. Note that if
the signal V des,0 is bounded, then V des and V des are also bounded and continuous signals.
When the magnitude and rate limits are not in effect, the transfer function from V des,0 to
V des is given by
V des V2
= 2 (15)
V des,0
s + 2 VV + V2
and the error V des,0 V des can be made arbitrarily small by selecting the bandwidth of the
filter to be sufficiently large (Swaroop et al., 1997).
3.2 Flight-path angle and airspeed control

In this loop the objective is to steer V and to their desired values, as determined in the
previous section. Furthermore, the heading angle has to track the reference signal ref , and
we also have to guarantee that z02 is regulated to zero. The available (virtual) controls in this
step are the aerodynamic angles and , as well as the thrust T . The lift, drag, and side
forces are assumed to be unknown and will be estimated. Note that the aerodynamic forces
T
also depend on the control-surface deflections U = e a r . These forces are quite
small, because the surfaces are primarily moment generators. However, because the current
control-surface deflections are available from the command filters used in the inner loop, we
can still take them into account in the control design. The relevant equations of motion are
given by
X 1 = A1F1 ( X ,U ) + B1G1 ( X ,U , X 2 ) + H 1 ( X ) (16)
where

0 0 V g sin V cos cos 0 0

1 cos T 1 1
A1 = 0 0 , H1 = cos sin cos , B1 = 0 0
mV cos mV cos mV cos

0 sin 0 T g 0 0 1
mV cos sin sin V cos
are known (matrix and vector) functions, and

L ( X ,U )
T

( )
F1 = Y ( X ,U ) , G1 = L ( X ,U ) T sin sin

D ( X ,U )

( )
L ( X ,U ) T sin cos

are functions containing the uncertain aerodynamic forces. Note that the intermediate
control variables and do not appear affine in the X 1 subsystem, which complicates the
design somewhat. Because the control objective in this step is to track the smooth reference
( )
signal X 1des = V des des des with X1 = (V ) , the tracking errors are defined as
T T
z11

Z1 = z12 = X1 X1des (17)
z13
To regulate Z1 and z02 to zero, the following equation needs to be satisfied (Kanayama et al.,
1990):
c 11 z11

B1G 1 ( X ,U , X 2 ) = V ref ( c 02 z02 + c 12 sin z12 ) A1F1 H 1 + X 1des (18)
c 13 z13
where F1 is the estimate of F1 and

T

G1 ( X ,U , X 2 ) = ( 0 )
L ( X ,U ) + L ( X ,U ) + T sin sin
(19)

( 0 )
L ( X ,U ) + L ( X ,U ) + T sin cos

with the estimate of the lift force decomposed as L ( X ,U ) = L 0 ( X ,U ) + L ( X ,U )

The estimate of the aerodynamic forces F1 is defined as
F1 = TF1 ( X ,U )

F1 (20)
where TF1 is a known (chosen) regressor function and is a vector with unknown constant
F1
parameters. It is assumed that there exists a vector F1 such that
F1 = TF1 ( X ,U ) F1 (21)
=
This means the estimation error can be defined as . We now need to determine
F1 F1 F1
the desired values and . The right-hand side of Eq. (18) is entirely known, and so the
des des
left-hand side can be determined and the desired values can be extracted. This is done by
introducing the coordinate transformation
(
x L 0 ( X ,U ) + L ( X ,U ) + T sin cos ) (22)
(
y L 0 ( X ,U ) + L ( X ,U ) + T sin sin ) (23)
which can be seen as a transformation from the two-dimensional polar coordinates
L 0 ( X ,U ) + L ( X ,U ) + T sin
T
and to Cartesian coordinates x and y. The desired signals T des ,0 y0 x0 are given by
T des ,0 c11 z11

B1 y 0 = V ref ( c02 z02 + c12 sin z12 ) A1F1 H 1 + X 1des (24)
x c13 z13
0
Thus, the virtual control signals are equal to
L ( X ,U ) des,0 = x02 + y 02 L 0 ( X ,U ) T sin (25)

and
arctan ( y 0 / x0 ) if x0 > 0

arctan ( y 0 / x0 ) + if x0 < 0 and y0 0

des,0 = arctan ( y 0 / x0 ) if x0 < 0 and y0 <0 (26)
/2 if x0 = 0 and y0 >0

/ 2 if x0 = 0 and y0 <0
Filtering the virtual signals to account for magnitude, rate, and bandwidth limits will give
the implementable virtual controls des , des and their derivatives. The sideslip-angle
T
command was already defined as ref = 0 , and thus X 2des = des des 0 and its derivative
are completely defined. However, care must be taken because the desired virtual control
des,0 is undefined when both x0 and y 0 are equal to zero, making the system momentarily
uncontrollable. This sign change of L 0 ( X ,U ) + L ( X ,U ) + T sin can only occur at very low
or negative angles of attack. This situation was not encountered during the maneuvers
simulated in this study. To solve the problem altogether, the designer could measure the
rate of change for x0 and y 0 and devise a rule base set to change sign when these terms
approach zero. Furthermore, problems will also occur at high angles of attack when the
control effectiveness term L will become smaller and eventually change sign. Possible
solutions include limiting the angle-of-attack commands using the command filters or
proper trajectory planning to avoid high-angle-of-attack maneuvers. Also note that so far in
the control design process, we have not taken care of the update laws for the uncertain
aerodynamic forces; they will be dealt with when the static control design is finalized.
3.3 Aerodynamic angle control

Now the reference signal X 2des = [ des des des ]T and its derivative have been found and
we can move on to the next feedback loop. The available virtual controls in this step are the
angular rates X 3 . The relevant equations of motion for this part of the design are given by
X 2 = A2 F1 ( X ,U ) + B2 ( X ) X 3 + H 2 ( X ) (27)
where
cos sin
tan + tan sin tan cos 0 0
cos cos
1 1
A2 = 0 0 , B2 = cos tan 1 sin tan
mV cos
sin 0 cos
0 1 0

g
T ( sin tan sin + sin tan cos sin tan cos ) tan cos cos
V
1 sin g
H2 = T + cos cos
mV cos V

g
T cos cos + cos sin
V
are known (matrix and vector) functions. The tracking errors are defined as
Z2 = X 2 X 2des (28)
To stabilize the Z2 subsystem, a virtual feedback control X 3des,0 is defined as
B2 X 3des,0 = C 2 Z2 A2 F1 H 2 + X 2des , C 2 = C 2T > 0 (29)
The implementable virtual control (i.e., the reference signal for the inner loop) X 3des and its
derivative are again obtained by filtering the virtual control signal X 3des,0 with a second-order
command-limiting filter.
3.4 Angular rate control

In the fourth step, an inner-loop feedback loop for the control of the body-axis angular rates
T
X 3 = p q r is constructed. The control inputs for the inner loop are the control-surface
T
deflections U = e a r . The dynamics of the angular rates can be written as
( )
X 3 = A3 F3 ( X ,U ) + B3 ( X )U + H 3 ( X ) (30)
where
c 3 0 c4 ( c 1r + c 2 p ) q

A3 = 0 c7 (
0 , H 3 = c 5 pr c 6 p 2 r 2 )
c 4
c 9
( c8 p c 2 r ) q
0

are known (matrix and vector) functions, and
L0 L L a L r
e
F3 = M0 , B3 = M e M a M r

N 0 N e N a N r
are unknown (matrix and vector) functions that have to be approximated. Note that for a
more convenient presentation, the aerodynamic moments have been decomposed: for
example,
M ( X ,U ) = M0 ( X ,U ) + M e e + M a a + M r r (31)
where the higher-order control-surface dependencies are still contained in M0 ( X ,U ) . The

control objective in this feedback loop is to track the reference signal X 3des = [ p ref q ref r ref ]T
with the angular rates X 3 . Defining the tracking errors
Z3 = X 3 X 3des (32)
and taking the derivatives results in
( )
Z 3 = A3 F3 ( X ,U ) + B3 ( X )U + H 3 ( X ) X 3des (33)
To stabilize the system of Eq. (33), we define the desired control U 0 as
A3 B 3U 0 = C 3Z3 A3F3 H 3 + X 3des , C 3 = C 3T > 0 (34)
where F3 and B3 are the estimates of the unknown nonlinear aerodynamic moment functions
F3 and B3 , respectively. The F-16 model is not over-actuated (i.e., the B3 matrix is square). If
this is not the case, some form of control allocation would be required (Enns, 1998, Durham,
1993). The estimates are defined as
F3 = TF3 ( X ,U )
, B = T ( X )
F3 3i B3
, for i = 1," , 3
B3 (35)
i i
where TF3 and TB3 are the known regressor functions, and are vectors with
i
F3 B3i
unknown constant parameters, and B 3i represents the ith column of B3 . It is assumed that
there exist vectors F3 and B3 such that
i
F3 = TF3 ( X ,U ) F3 , B3i = TB3 ( X ) B3 (36)

i i
=
This means that the estimation errors can be defined as and
= .
F3 F3 F3 B3 B3 B3 i i i
The actual control U is found by applying a filter similar to Eq. (13) to U 0 .
3.5 Update laws and stability properties

We have now finished the static part of our control design. In this section the stability
properties of the control law are discussed and dynamic update laws for the unknown
parameters are derived. Define the control Lyapunov function
1 T 2 2 cos z12
V= Z0 Z0 + z11 +
2
+ z13
2
+ Z2T Z2 + Z3T Z3 +
2 c 02 (37)
1
( ( ) ( )) ( )
3
T 1
+ trace F1 F1
+ trace
F1
T 1
F3 F3

F3 + trace
T 1
B3 B3

B3
2 i =1
i i i
with the update gains matrices F1 = TF1 > 0, F3 = TF3 > 0 , and B3 = TB3 > 0 . Taking the
i i
derivative of V along the trajectories of the closed-loop system gives

V = c 01 z01
2
( ) ( )
+ z02 z01 + V V des,0 z01 + z02 z01 + V ref sin z12 c03 z03
2
(
V sin sin des,0 z03 + )
c 11 z11
2

c
c 02

V ref sin z12 z02 + 12 sin 2 z12 c 13 z13

2
( F1 1 (
+ B G ( X ) G ( X ) +
+ Z1T A1TF1 1 2 1 2 ))
( ( ))
+ Z1T B1 G 1 ( X 2 ) G 1 X 2des,0 Z2TC 2 Z2 + Z2T A2 TF1 (
+ ZT B X X des,0 ZTC Z +
F1 2 2 3 3 3 3 3 ) (38)
3
U + ZT A B U U 0 trace
T 1 T 1
+ Z3T A3 TF3
+ T
F3 B3i B3i i 3 3 3 ( )
F1 F1
trace
F1

F3 F3
+
F3

i =1
3
trace T 1
B B B
i =1 3i 3i 3i
To cancel the terms in Eq. (38), depending on the estimation errors, we select the update
laws

( T T
)
F1 = F1 F1 A1 a Z1 + A2 Z2 , F3 = F3 F3 A3 Z3 ,
T
(
B3 = proj B3 B3 B3 A3 Z3U i
i i
T
i i
) (39)
= A T
with A1 aTF1 F1 1 F1 F1 1 1 2 1(
+ B G ( X ) G ( X )
2 )
The update laws for B3 include a projection operator (Ioannou and Sun, 1995) to ensure that
certain elements of the matrix do not change sign and full rank is maintained always. For
most elements, the sign is known based on physical principles. Substituting the update laws
in Eq. (38) leads to
c
V = c01 z01
2
c03 z03
2
c11 z11
2
V ref 12 sin 2 z12 c13 z13
c02
2
(
Z2T C 2 Z2 Z3T C 3Z3 + V V des,0 z01 + )
(40)
( ) ( ( )) ( )
V sin sin des,0 z03 + Z1T B1 G1 ( X 2 ) G 1 X 2des,0 + Z2T B2 X 3 X 3des,0 + Z3T A3 B 3 U U 0 ( )
where the first line is already negative semi-definite, which we need to prove stability in the
sense of Lyapunov. Because our Lyapunov function V equation (37) is not radially
unbounded, we can only guarantee local asymptotic stability (Kanayama et al., 1990). This is
sufficient for our operating area if we properly initialize the control law to
ensure z12 / 2 . However, we also have indefinite error terms due to the tracking errors
and due to the command filters used in the design. As mentioned before, when no rate or
magnitude limits are in effect, the difference between the input and output of the filters can
be made small by selecting the bandwidth of the filters to be sufficiently larger than the
bandwidth of the input signal. Also, when no limits are in effect and the small bounded
difference between the input and output of the command filters is neglected, the feedback
controller designed in the previous sections will converge the tracking errors to zero (for
proof, see (Farrell et al., 2005, Sonneveldt et al., 2007, Yip, 1997)).
Naturally, when control or state limits are in effect, the system will in general not track the
reference signal asymptotically. A problem with adaptive control is that this can lead to
corruption of the parameter-estimation process, because the tracking errors that are driving
this process are no longer caused by the function approximation errors alone (Farrell et al.,
2003). To solve this problem we will use a modified definition of the tracking errors in the
update laws in which the effect of the magnitude and rate limits has been removed, as
suggested in (Farrell et al., 2005, Sonneveldt et al., 2006). Define the modified tracking errors
Z 1 = Z1 1 , Z 2 = Z2 2 , Z 3 = Z3 3 (41)
with the linear filters
( (
1 = C 1 1 + B1 G 1 ( X ,U , X 2 ) G 1 X ,U , X 2des,0 ))
(
2 = C 2 2 + B2 X 3 X des,0
3 ) (42)
(
3 = C 3 3 + A3 B 3 U U 0 )
The modified errors will still converge to zero when the constraints are in effect, which
means the robustified update laws look like

( T )
F1 = F1 F1 A1 a Z1 + A2 Z2 , F3 = F3 F3 A3 Z3 ,
T T = proj AT Z U
B3 B3 B3 B3
i
3 3 i i
( i i
) (43)
To better illustrate the structure of the control system, a scheme of the adaptive inner-loop
controller is shown in Fig. 2.
4. Model identification
To simplify the approximation of the unknown aerodynamic force and moment functions,
thereby reducing computational load, the flight envelope is partitioned into multiple
connecting operating regions called hyperboxes or clusters. This can be done manually
using a priori knowledge of the nonlinearity of the system, automatically using nonlinear
optimization algorithms that cluster the data into hyperplanar or hyperellipsoidal clusters
(Babuka, 1998) or a combination of both. In each hyperbox a locally valid linear-in-the-
parameters nonlinear model is defined, which can be estimated using the update laws of the
Lyapunov-based control laws. The aerodynamic model can be partitioned using different
state variables, the choice of which depends on the expected nonlinearities of the system. In
this study we use B-spline neural networks (Cheng et al., 1999, Ward et al., 2003) (i.e., radial
basis function neural networks with B-spline basis functions) to interpolate between the
local nonlinear models, ensuring smooth transitions. In the previous section we defined
parameter update laws equation (43) for the unknown aerodynamic functions, which were
written as
F1 = TF1 ( X ,U )
, F = T ( X ,U )
F1 3 F3
, B = T ( X )
F3 3i B3

B3 (44)
i i
Now we will further define these unknown vectors and known regressor vectors. The total
force approximations are defined as
qc
L = qS C L0 ( , ) + C L ( , e ) + C Lq ( ) + C L ( , ) e
2V e

pb rb
Y = qS C Y0 ( , , e ) + C Yp ( , ) + C Yr ( , ) + C Y ( , ) a + C Y ( , ) r (45)
2V 2V a r

(
D = qS C ( , , ) + C ( , )

D0 e

D e e )
Fig. 2. Inner-loop control system

and the moment approximations are defined as
pb rb
L = qS C L0 ( , , e ) + C Lp ( , ) + CLr ( , ) + CL ( , ) e + C L ( , ) a + C L ( , ) r
2V 2 V e a r

qc
M = qS CM0 ( , ) + CMq ( )
+ CM ( , ) e (46)
2V e

pb rb
N = qS C N0 ( , , e ) + C Np ( , ) + C Nr ( , ) + CN ( , ) e + C N ( , ) a + C N ( , ) r
2V 2V e a r

Note that these approximations do not account for asymmetric failures that will introduce
coupling of the longitudinal and lateral motions of the aircraft. If a failure occurs that
introduces a parameter dependency that is not included in the approximation, stability can
no longer be guaranteed. It is possible to include extra cross-coupling terms, but this is
beyond the scope of this paper. The total nonlinear function approximations are divided
into simpler linear-in-the parameter nonlinear coefficient approximations: for example,
C L0 ( , ) = CTL ( , )CL (47)

0 0
where the unknown parameter vector CL contains the network weights (i.e., the unknown
0
parameters), and CL is a regressor vector containing the B-spline basis functions
0
(Sonneveldt et al., 2007). All other coefficient estimates are defined in similar fashion. In this
case a two-dimensional network is used with input nodes for and . Different scheduling
parameters can be selected for each unknown coefficient. In this study we used third-order
B-splines spaced 2.5 deg and one or more of the selected scheduling variables , and e .
Following the notation of Eq. (47), we can write the estimates of the aerodynamic forces and
moments as
L = TL ( , , e )
, Y = T ( , , )
L Y e Y
= T ( , , )
, D
D e

D
(48)
, N = T ( , , )
L = TL ( , , e )
, M
L
= TM ( , , e ) M N e

N
which is a notation equivalent to the one used in Eq. (44). Therefore, the update laws
equation (43) can indeed be used to adapt the B-spline network weights. In practice
nonparametric uncertainties such as 1) un-modeled structural vibrations 2) measurement
noise, 3) computational round-off errors and sampling delays, and 4) time variations of the
unknown parameters, can result in parameter drift. One approach to avoiding parameter
drift taken here is to stop the adaptation process when the training error is very small (i.e. a
dead zones (Babuka, 1998, Karason and Annaswamy, 1994)).
5. Simulation results
This section presents the simulation results from the application of the flight-path controller
developed in the previous sections to the high-fidelity, six-degree-of-freedom F-16 model of
Sec. 2. Both the control law and the aircraft model are written as C S-functions in
MATLAB/Simulink. The simulations are performed at three different starting flight
conditions with the trim conditions: 1) h= 5000 m, V= 200 m/s, and ==2.774 deg; 2) h=0 m,
V =250 m/s, and ==2.406 deg; and 3) h= 2500 m, V= 150 m/s, and ==0.447 deg; where h
is the altitude of the aircraft, and all other trim states are equal to zero.
Furthermore, two maneuvers are considered: 1) a climbing helical path and 2) a
reconnaissance and surveillance maneuver. The latter maneuver involves turns in both
directions and some altitude changes. The simulations of both maneuvers last 300 s. The
reference trajectories are generated with second-order linear filters to ensure smooth
trajectories. To evaluate the effectiveness of the online model identification, all maneuvers
will also be performed with a 30% deviation in all aerodynamic stability and control
derivatives used by the controller (i.e., it is assumed that the onboard model is very
inaccurate). Finally, the same maneuvers are also simulated with a lockup at 10 deg of the
left aileron.
5.1 Control parameter tuning

We start with the selection of the gains of the static control law and the bandwidths of the
command filters. Lyapunov stability theory only requires the control gains to be larger than
zero, but it is natural to select the largest gains of the inner loop. Larger gains will, of course,
result in smaller tracking errors, but at the cost of more control effort. It is possible to derive
certain performance bounds that can serve as guidelines for tuning (see, for example, Krsti,
et al., 1993, Sonneveldt et al., 2007). However, getting the desired closed-loop response is
still an extensive trial-and-error procedure. The control gains were selected as c01 = 0.1 ,
c02 = 10 5 , c03 = 0.5 , c11 = 0.01 , c12 = 2.5 , c13 = 0.5 , C 2 = diag ( 1,1,1 ) , C 3 = diag ( 2, 2, 2 ) .
The bandwidths of the command filters for the actual control variables e , a , and r are
chosen to be equal to the bandwidths of the actuators, which are given in (Sonneveldt et al.,
2007). The outer-loop filters have the smallest bandwidths. The selection of the other
bandwidths is again trial and error. A higher bandwidth in a certain feedback loop will
result in more aggressive commands to the next feedback loop. All damping ratios are equal
to 1.0. It is possible to add magnitude and rate limits to each of the filters. In this study
magnitude limits on the aerodynamic roll angle and the flight-path angle are used to
avoid singularities in the control laws. Rate and magnitude limits equal to those of the
actuators are enforced on the actual control variables.
The selected command-filter parameters can be found back in Table 1. As soon as the
controller gains and command-filter parameters have been defined, the update law gains
can be selected. Again, the theory only requires that the gains should be larger than zero.
Larger update gains means higher learning rates and thus more rapid changes in the B-
spline network weights.
5.2 Manoeuvre 1: upward spiral

In this section the results of the numerical simulations of the first test maneuver, the
climbing helical path, are discussed. For each of the three flight conditions, five cases are
considered: nominal, the aerodynamic stability and control derivatives used in the control law
perturbed with +30% and with -30% with respect to the real values of the model, a lockup of
the left aileron at +10 deg, and a lockup at -10 deg. No actuator sensor information is used. In
Fig. 3 the results are plotted of the simulation without uncertainty, starting at flight condition
1. The maneuver involves a climbing spiral to the left with an increase in airspeed. It can be
seen that the control law manages to track the reference signal very well and closed-loop
tracking is achieved. The sideslip angle does not become any larger than 0.02 deg. The
aerodynamic roll angle does reach the limit set by the command filter, but this has no
consequences for the performance. The use of dead zones ensures that the parameter update
laws are indeed not updating during this maneuver without any uncertainties. The responses
at the two other flight conditions are virtually the same, although less thrust is needed due to
the lower altitude of flight condition 2 and the lower airspeed of flight condition 3. The other
control surfaces are also more efficient. This is illustrated in Tables 24, in which the mean
absolute values (MAVs) of the outer-loop tracking errors, control-surface deflections, and
thrust can be found. Plots of the parameter-estimation errors are not included. However, the
errors converge to constant values, but not to zero, as is common with Lyapunov-based
update laws (Sonneveldt et al., 2007, Page and Steinberg, 1999).
Table 1. Command-filter parameters
Table 2. Manoeuvre 1 at flight condition 1: mean absolute value of tracking errors and
control inputs
control inputs
control inputs
The response of the closed-loop system during the same maneuver starting at flight
condition 1, but with +30% uncertainties in the aerodynamic coefficients, is shown in Fig. 4.
It can be observed that the tracking errors of the outer loop are now much larger, but in the
end, the steady-state tracking error converges to zero. The sideslip angle still remains within
0.02 deg. Some small oscillations are visible in Fig. 4j, but these stay well within the rate and
magnitude limits of the actuators. In Tables 24 the MAVs of the tracking errors and control
inputs are shown for all flight conditions. As was already seen in the plots, the average
tracking errors increase, but the magnitude of the control inputs stays approximately the
same. The same simulations have been performed for a -30% perturbation in the stability
and control derivatives used by the control law, and the results are also shown in the tables.
It appears that underestimated initial values of the unknown parameters lead to larger
tracking errors than overestimates for this maneuver. Finally, the maneuver is performed
with the left aileron locked at 10 deg [i.e., adamaged = 0.5 ( a 10 / 180 ) ]. Figure 5 shows the
response at flight condition 3 with the aileron locked at -10 deg.
Except for some small oscillations in the response of roll rate p, there is no real change in
performance visible; this is confirmed by the numbers of Table 4. However, from Tables 2
and 3 we observe that aileron and rudder deflections become larger with both locked aileron
failure cases, whereas tracking performance hardly declines.
5.3 Manoeuvre 2: reconnaissance

The second maneuver, called reconnaissance and surveillance, involves turns in both
directions and altitude changes, but airspeed is kept constant. Plots of the simulation at
flight condition 3 with -30% uncertainties are shown in Fig. 6. Tracking performance is again
excellent and the steady-state tracking errors converge to zero. There are some small
oscillations in the rudder deflection, but these are within the limits of the actuator. We
compare the MAVs of the tracking errors and control inputs with those for the nominal case
in Table 5 and observe that the average tracking errors have not increased much for this
case. The degradation of performance for the uncertainty cases is somewhat worse at the
other two flight conditions, as can be seen in Tables 6 and 7. The sideslip angle always
remains within 0.05 deg for all flight conditions and uncertainties. Corresponding with the
results of maneuver 1, overestimation of the unknown parameters again leads to smaller

tracking errors. Simulations of maneuver 2 with the locked aileron are also performed.
Figure 7 shows the results for flight condition 1 with a locked aileron at _10 deg. Some very
small oscillations are again visible in the roll rate, aileron, and rudder responses, but
tracking performance is good and steady-state convergence is achieved.
Table 6 confirms that the results of the simulations with actuator failure hardly differ from
the nominal case. There is only a small increase in the use of the lateral control surfaces. The
same holds at the other flight conditions, as can be seen in Tables 5 and 7.
control inputs
control inputs
control inputs
6. Conclusions
In this paper a nonlinear adaptive flight-path control system is designed for a high-fidelity
F-16 model. The controller is based on a backstepping approach with four feedback loops
that are designed using a single control Lyapunov function to guarantee stability. The
uncertain aerodynamic forces and moments of the aircraft are approximated online with B-
spline neural networks for which the weights are adapted by Lyapunov-based update laws.
Numerical simulations of two test maneuvers were performed at several flight conditions to
verify the performance of the control law. Actuator failures and uncertainties in the stability
and control derivatives are introduced to evaluate the parameter-estimation process. The
results show that trajectory control can still be accomplished with these uncertainties and
failures, and good tracking performance is maintained. Compared with other Lyapunov-
based trajectory control designs, the present approach is much simpler to apply and the
online estimation process is more robust to saturation effects. Future studies will focus on
the actual trajectory generation and the extension to formation-flying control.
Appendix constraint adaptive backstepping

Backstepping [21] is a systematic, Lyapunov-based method for nonlinear control design,
which can be applied to nonlinear systems that can be transformed into lower-triangular
form, such as the system of Eq. (A.1):
x 1 = f ( x1 ) + g ( x1 ) x2 (A.1)
The name backstepping refers to the recursive nature of the control law design procedure.
Using the backstepping procedure, a control law is recursively constructed, along with a
control Lyapunov function (CLF) to guarantee global stability. For the system Eq. (A.1), the
aim of the design procedure is to bring the state vector x1 to the origin. The first step is to
consider x2 as the virtual control of the scalar x1 subsystem and to find a desired virtual control
law 1 ( x1 ) that stabilizes this subsystem by using the control Lyapunov function V1 ( x1 ) :
1 2
V1 ( x1 ) = x1 (A.2)
2
The time derivative of this CLF is negative definite
V ( x )
V1 ( x1 ) = 1 1 f ( x1 ) + g ( x1 ) 1 ( x1 ) < 0, x1 0 (A.3)
x1
If only the virtual control law
x 2 = 1 ( x1 ) (A.4)
could be satisfied. The key property of backstepping is that we can now step back through
the system. If the error between x2 and its desired value is defined as
z = x 2 1 ( x1 ) (A.5)
the system Eq. (6) can be rewritten in terms of this error state
x 1 = f ( x1 ) + g ( x1 ) 1 ( x1 ) + z
1 ( x1 ) (A.6)
z = u
x1
( f ( x ) + g ( x ) ( x ) + z )
1 1 1 1
The control Lyapunov function Eq. (A.2) can now be expanded with a term penalizing the
error state z
1 2
V2 ( x1 , z ) = V1 ( x1 ) + z (A.7)
2
In simplified notation the time derivative of V2 is equal to
V
V2 = 1 f + g ( 1 + z ) + z u 1 f + g ( 1 + z )
x1 x 1
(A.8)
V V
= 1 f + g 1 + z 1 g + u 1 f + g ( 1 + z )
x1
1x x 1
which can be rendered negative definite with the control law
1 V
u = cz + f + g ( 1 + z ) 1 g , c > 0 (A.9)
x1 x1
This design procedure can also be used for a system with a chain of integrators. The only
difference is that there will be more virtual states to backstep through. Starting with the
state farthest from the actual control, each step of the backstepping technique can be
broken up into three parts: 1. Introduce a virtual control 1 and an error state z , and rewrite
the current state equation in terms of these; 2. choose a CLF for the system, treating it as a
final stage; 3) choose an equation for the virtual control that makes the CLF stabilizable.
The CLF is augmented at subsequent steps to reflect the presence of new virtual states, but
the same three stages are followed at each step. Hence, backstepping is a recursive design
procedure.
For systems with parametric uncertainties there exists a method called adaptive
backstepping (Kannelakopoulos et al., 1991), which achieves boundedness of the closed-
loop states and convergence of the tracking error to zero.
Consider the parametric strict-feedback system
x 1 = x2 + 1T ( x1 )
# # #
(A.10)
x n 1 = xn + nT 1 ( x1 ," , xn 1 )
x n = ( x ) u + nT ( x )
where ( x ) 0 for all x \ n , is a vector of unknown constant parameters, and i are

known (smooth) function vectors. The control objective is to asymptotically track a given
reference y r ( t ) with x1 . All derivatives of y r ( t ) are assumed to be known.
The adaptive backstepping design procedure is similar to the normal backstepping
procedure, only this time a control law (static part) and a parameter update law (dynamic
part) are designed along with a control Lyapunov function to guarantee global stability. The
control law makes use of a parameter estimate , which is constantly adapted by the
dynamic parameter update law. Furthermore, the control Lyapunov function now contains
an extra term that penalizes the parameter estimation error = .
Theorem A.1 (Adaptive Backstepping Method): To stabilize the system Eq. (A.9) an error
variable is introduced for each state
zi = xi y r( ) i 1
i 1
(A.11)
along with a virtual control law
i 1
( )
i 1 i1
i xi , , yr( i 1) = ci zi zi 1 iT + x k + 1 + ( ki 11) y r( ) + i 1 i + i 1 i zk
k
(A.12)

k = 1 x k y r k = 2

for i = 1, 2," , n , where the tuning function i and the regressor vectors i are defined as
( )
i xi , , y r( i 1) = i 1 + i zi (A.13)
and
( ) i 1
i 1
i xi , , y r( i 2 ) = i k (A.14)
k =1 x k
( )
where xi = ( x1 , x2 ," , xi ) , y ri( )= y r , y r ," , y r( ) . ci > 0 are design constants. With these new
i i
variables the control and adaptation laws can be defined as
u=
1
( x )
n ( r )
x , , y ( n 1) + y ( n )
r

(A.15)
and

( )
= n x , , yr( n 1) = Wz (A.16)
where = T > 0 is the adaptation gain matrix and W the regressor matrix
( )
W z , = (1 ," , i ) (A.17)
The control law Eq. (A.15) together with the update law Eq. (A.17) renders the derivative of
the Lyapunov function
1 n 2 1 T 1
V= zi + 2
2 i =1
(A.18)
negative definite and thus this adaptive controller guarantees global boundedness of x ( t )
and asymptotically tracking of a given reference y r ( t ) with x1 .
Proof of this theorem can be found in Sec. 4.3 of (Krsti et al., 1992).
The standard adaptive backstepping procedure as has been discussed so far has a number of
drawbacks.
1. The analytic calculation of the virtual control derivatives is tedious, especially for large
systems;
2. The procedure can only handle systems that can be transformed into a lower-triangular
form;
3. Constraints on the inputs and states are not taken into account.
The third drawback can be a major problem when designing for flight control, because the
actuators of an aircraft have rate, bandwidth, and magnitude constraints. When the control
signal demanded by the backstepping controller cannot be generated by the actuators, that
is, the actuators saturate, stability can no longer be guaranteed. The problem becomes worse
Fig. 3. Manoeuvre 1: climbing helical path performed at flight condition 1 without any
uncertainty or actuator failures
Fig. 4. Manoeuvre 1: climbing helical path performed at flight condition 2 with +30%
uncertainties in the aerodynamic coefficients
Fig. 5. Manoeuvre 1: climbing helical path performed at flight condition 3 with left aileron
locked at -10 deg
Fig. 6. Manoeuvre 2: reconnaissance and surveillance performance at flight condition 3 with

-30% uncertainties in the aerodynamic coefficients
Fig. 7. Manoeuvre 2: reconnaissance and surveillance performance at flight condition 1 with

left aileron locked at +10 deg.
when online parameter update laws are used, because these tend to be aggressive while
seeking the desired tracking performance. Because the desired control signal is not achieved
during saturation, the tracking error will increase. Because this tracking error is not just the
result from the parameter estimation error, the update law may unlearn during these
saturation periods.
In (Farrell et al., 2003, 2005) a method is proposed that fits within the recursive adaptive
backstepping design procedure and deals with the constraints on both the control variables
and the intermediate states used as virtual controls.An additional advantage of the method
is that it also eliminates the two other drawbacks of the adaptive backstepping method, that
is, the time consuming analytic computation of virtual control signal derivatives and the
restriction to nonlinear systems of a lower-triangular form.
The proposed method extends the adaptive backstepping framework in two ways.
1. Command filters are used to eliminate the analytic computation of the time derivatives of
the virtual controls. The command filters are designed as linear, stable, low-pass filters with
unity gain from its input to its output. The inputs of these filters are the desired (virtual)
control signals and the outputs are the actual (virtual) control signal and its time derivative.
Using command filters to calculate the virtual control derivatives, it is still possible to prove
stability in the sense of Lyapunov in the absence of constraints on the control input and state
variables.
2. A stable parameter estimation process is ensured even when constraints on the control
variables and states are in effect. During these periods the tracking error may increase
because the desired control signal cannot be implemented due to these constraints imposed
on the system. In this case the desired response is too aggressive for the system to be feasible
and the primary goal is to maintain stability of the online function approximation. The
command filters keep the control signal and the state variables within their mechanical
constraints and operating limits, respectively. The effect these constraints have on the
tracking errors can be estimated and this effect can be implemented in modified tracking
error definitions. These modified tracking errors are only the result of parameter estimation
errors as the effect of the constraints on the control input and state variables has been
removed. These modified tracking errors can thus be used by the parameter update laws to
ensure a stable estimation process.
The command filtered adaptive backstepping approach is summarized in the following
theorem.
Theorem A.2 (Constrained Adaptive Backstepping Method): For the parameter strict-
feedback system Eq. (15) the tracking errors are again defined as
zi = xi y r( ) i 1
i 1
(A.19)
for i = 1, 2," , n . The nominal or desired virtual control laws can be defined as
i0 = ci z zi 1 iT + + i 1 i + 1 , i = 1, 2," , n 1 (A20)
where
zi 1 = zi i , i = 1, 2," , n (A.21)
are the modified tracking errors and where

i = ci i + ( i i0 ) , i = 1, 2," , n 1 (A.22)
are the filtered versions of the effect of the state constraints on the tracking errors zi . The
nominal virtual control signals i0 are filtered to produce the magnitude, rate, and
bandwidth limited virtual control signals i and its derivatives i that satisfy the limits
imposed on the state variables. This command filter can for instance be chosen as (Farrell et
al., 2005)
q2
q 1 i q1
= n2 , = (A.23)
q2 2n SR 2 ( )
SM i0
q1 q 2
i q2
n
where SM () and SR () represent the magnitude and rate limit functions, respectively. These
saturation functions are defined similarly as
M if x M

SM ( x ) = x if x < M

M if x M
The effect of implementing the achievable virtual control signals instead of the desired ones
is estimated by the i filters. With these filters the modified tracking errors zi can be defined.
It can be seen from Eq. (A.21) that when the limitations on the states are not in effect the
modified tracking error converges to the tracking error. The nominal control law is defined
in a similar way as
u0 =
1
(x)
(
c n zn zn 1 nT + n 1 + y r( )
n
) (A.24)
which is again filtered to generate the magnitude, rate, and bandwidth limited control signal
u. The effect of implementing the limited control law instead of the desired one can again be
estimated with
n = c n n + ( u u0 ) (A.25)
Finally, the update law that now uses the modified tracking errors is defined as
n

= i zi (A.26)
i =1
The resulting control law will render the derivative of the control Lyapunov function
1 n 2 1 T 1
V= zi + 2
2 i =1
(A.27)
negative definite, which means that the closed-loop system is asymptotically stable.
7. References
Clough, B. T. (2005), Unmanned Aerial Vehicles: Autonomous Control Challenges, a
Researchers Perspective, Journal of Aerospace Computing, Information, and
Communication, Vol. 2, No. 8, pp. 327347, doi: 10.2514/1.5588.
Wegener, S., Sullivan, D., Frank, J., and Enomoto, F. (2004), UAV Autonomous Operations
for Airborne Science Missions, AIAA 3rd Unmanned Unlimited Technical
Conference, Workshop and Exhibit, AIAA Paper 2004-6416.
Papadales, B., and Downing, M. (2005), UAV Science Missions: A Business Perspective,
Infotech@Aerospace, AIAA Paper 2005-6922.
Tsach, S., Chemla, J., and Penn, D. (2003), UAV Systems Development in IAI-Past, Present
and Future, 2nd AIAA Unmanned Unlimited Systems, Technologies, and
Operations-Aerospace Land, and Sea Conference, AIAA Paper 20036535.
Kaminer, I., Pascoal, A., Hallberg, E., and Silvestre, C. (1998), Trajectory Tracking for
Autonomous Vehicles: An Integrated Approach to Guidance and Control, Journal
of Guidance, Control, and Dynamics, Vol. 21, No. 1, pp. 2938, doi:10.2514/2.4229.
Boyle, D. P., and Chamitof, G. E. (1999), Autonomous Maneuver Tracking for Self-Piloted
Vehicles, Journal of Guidance, Control, and Dynamics, Vol. 22, No. 1, pp. 5867,
doi: 10.2514/2.4371.
Singh, S. N., Steinberg, M. L., and Page, A. B. (2003), Nonlinear Adaptive and Sliding Mode
Flight Path Control of F/A-18 Model, IEEE Transactions on Aerospace and
Electronic Systems, Vol. 39, No. 4, pp. 12501262, doi: 10.1109/TAES.2003.1261125.
Ren, W., and Beard, R. W. (2004), Trajectory Tracking for Unmanned Air Vehicles with
Velocity and Heading Rate Constraints, IEEE Transactions on Control Systems
Technology, Vol. 12, No. 5, pp. 706716, doi:10.1109/TCST.2004.826956.
Ren, W., and Atkins, E. (2005), Nonlinear Trajectory Tracking for Fixed Wing UAVs via
Backstepping and Parameter Adaptation, AIAA Guidance, Navigation, and
Control Conference and Exhibit, AIAA Paper 2005-6196.
No, T. S., Min, B. M., Stone, R. H., and K. C. Wong, J. E. (2005), Control and Simulation of
Arbitrary Flight Trajectory-Tracking, Control Engineering Practice, Vol. 13, No. 5,
pp. 601612, doi:10.1016/j.conengprac.2004.05.002.
Kaminer, I., Yakimenko, O., Dobrokhodov, V., Pascoal, A., Hovakimyan, N., Cao, C., Young,
A., and Patel, V. (2007), Coordinated Path Following for Time-Critical Missions of
Multiple UAVs via L1 Adaptive Output Feedback Controllers, AIAA Guidance,
Navigation, and Control Conference and Exhibit, AIAA Paper 2007-6409.
Pachter, M., DAzzo, J. J., and J. L. Dargan (1994), Automatic Formation Flight Control,
Journal of Guidance, Control, and Dynamics, Vol. 17, No. 6, pp. 13801383.
Proud, A. W., Pachter, M., and DAzzo, J. J. (1999), Close Formation Flight Control,AIAA
Guidance, Navigation, and Control Conference, AIAA Paper 1999-4207.
Fujimori, A., Kurozumi, M., Nikiforuk, P. N., and Gupta, M. M. (2000), Flight Control
Design of an Automatic Landing Flight Experiment Vehicle, Journal of Guidance,
Control, and Dynamics, Vol. 23, No. 2, pp. 373376, doi:10.2514/2.4536.
Singh, S. N., Chandler, P., Schumacher, C., Banda, S., and Pachter, M. (2000), Adaptive
Feedback Linearizing Nonlinear Close Formation Control of UAVs, American
Control Conference, Inst. of Electrical and Electronics Engineers, Piscataway, NJ,
pp. 854858.
Pachter, M., DAzzo, J. J., and Proud, A. W. (2001), Tight Formation Control, Journal of
Guidance, Control, and Dynamics, Vol. 24, No. 2, pp. 246254, doi:10.2514/2.4735.
Wang, J., Patel, V., Cao, C., Hovakimyan, N., and Lavretsky, E. (2008), Novel L1 Adaptive
Control Methodology for Aerial Refueling with Guaranteed Transient
Performance, Journal of Guidance, Control, and Dynamics, Vol. 31, No. 1, pp. 182
193, doi:10.2514/1.31199.
Healy, A., and Liebard, D. (1993), Multivariable Sliding Mode Control for Autonomous
Diving and Steering of Unmanned Underwater Vehicles, IEEE Journal of Oceanic
Engineering, Vol. 18, No. 3, pp. 327339, doi:10.1109/JOE.1993.236372.
Narasimhan, M., Dong, H., Mittal, R., and Singh, S. N. (2006), Optimal Yaw Regulation and
Trajectory Control of Biorobotic AUV Using Mechanical Fins Based on CFD
Parametrization, Journal of Fluids Engineering, Vol. 128, No. 4, pp. 687698,
doi:10.1115/1.2201634.
Kannelakopoulos, I., Kokotovi, P. V., and Morse, A. S. (1991), Systematic Design of
Adaptive Controllers for Feedback Linearizable Systems, IEEE Transactions on
Automatic Control, Vol. 36, No. 11, pp. 12411253, doi:10.1109/9.100933.
Krsti, M., Kanellakopoulos, I., and Kokotovi, P. V. (1992), Adaptive Nonlinear Control
Without Overparametrization, Systems and Control Letters, Vol. 19, pp. 177185,
doi:10.1016/0167-6911(92)90111-5.
Singh, S. N., and Steinberg, M. (1996), Adaptive Control of Feedback Linearizable
Nonlinear Systems With Application to Flight Control, AIAA Guidance,
Navigation, and Control Conference, AIAA Paper 1996-3771.
Hrkegrd, O. (2003), Backstepping and Control Allocation with Applications to Flight
Control, Ph.D. Thesis, Linkping Univ., Linkping, Sweden.
Farrell, J., Polycarpou, M., and Sharma, M. (2003), Adaptive Backstepping with Magnitude,
Rate, and Bandwidth Constraints: Aircraft Longitude Control, American Control
Conference, American Control Conference Council, Evanston, IL, pp. 38983903.
Kim, S. H., Kim, Y. S., and Song, C. (2004), A Robust Adaptive Nonlinear Control
Approach to Missile Autopilot Design, Control Engineering Practice, Vol. 12, No.
2, pp. 149154, doi:10.1016/S0967-0661(03)00016-9.
Shin, D. H., and Kim, Y. (2004), Reconfigurable Flight Control System Design Using
Adaptive Neural Networks, IEEE Transactions on Control Systems Technology,
Vol. 12, No. 1, pp. 87100, doi:10.1109/TCST.2003.821957.
Farrell, J., Sharma, M., and Polycarpou, M. (2005), Backstepping Based Flight Control with
Adaptive Function Approximation, Journal of Guidance, Control, and Dynamics,
Vol. 28, No. 6, pp. 10891102, doi:10.2514/1.13030.
Sonneveldt, L., Chu, Q. P., and Mulder, J. A. (2006), Constrained Adaptive Backstepping
Flight Control: Application to a Nonlinear F-16/MATV Model, AIAA Guidance,
Navigation, and Control Conference and Exhibit, AIAA Paper 2006-6413.
Sonneveldt, L., Chu, Q. P., and Mulder, J. A. (2007), Nonlinear Flight Control Design Using
Constrained Adaptive Backstepping, Journal of Guidance, Control, and Dynamics,
Vol. 30, No. 2, pp. 322336, doi:10.2514/1.25834.
Yip, P.-C. P. (1997), Robust and Adaptive Nonlinear Control Using Dynamic Surface
Controller with Applications to Intelligent Vehicle Highway Systems, Ph.D.
Thesis, Univ. of California at Berkeley, Berkeley, CA.
Cheng, K. W. E., Wang, H., and Sutanto, D. (1999), Adaptive B-Spline Network Control for
Three-Phase PWM AC-DC Voltage Source Converter, IEEE 1999 International
Conference on Power Electronics and Drive Systems, Inst. of Electrical and
Electronics Engineers, Piscataway, NJ, pp. 467472.
Ward, D. G., Sharma, M., Richards, N. D., and Mears, M. (2003), Intelligent Control of Un-
Manned Air Vehicles: Program Summary and Representative Results, 2nd AIAA
Unmanned Unlimited Systems, Technologies and Operations Aerospace, Land and
Sea, AIAA Paper 2003-6641.
Nguyen, L. T., Ogburn, M. E., Gilbert, W. P., Kibler, K. S., Brown, P. W., and Deal, P. L.
(1979), Simulator Study of Stall Post-Stall Characteristics of a Fighter Airplane
with Relaxed Longitudinal Static Stability, NASA Langley Research Center,
Hampton, VA.
Lewis, B. L., and Stevens, F. L. (1992), Aircraft Control and Simulation, Wiley, New York,
pp. 154, 110115.
Cook, M. V. (1997), Flight Dynamics Principles, Butterworth-Heinemann, London, pp. 11
29.
Swaroop, D., Gerdes, J. C., Yip, P. P., and Hedrick, J. K. (1997), Dynamic Surface Control of
Nonlinear Systems, Proceedings of the American Control Conference.
Kanayama, Y. J., Kimura, Y., Miyazaki, F., and Noguchi, T. (1990), A Stable Tracking
Control Method for an Autonomous Mobile Robot, IEEE International Conference
on Robotics and Automation, Inst. of Electrical and Electronics Engineers,
Piscataway, NJ, pp. 384389.
Enns, D. F. (1998), Control Allocation Approaches, AIAA Guidance, Navigation, and
Control Conference and Exhibit, AIAA Paper 1998-4109.
Durham, W. C. (1993), Constrained Control Allocation, Journal of Guidance, Control, and
Dynamics, Vol. 16, No. 4, pp. 717725, doi:10.2514/3.21072.
Ioannou, P. A., and Sun, J. (1995), Stable and Robust Adaptive Control, PrenticeHall,
Englewood Cliffs, NJ, pp. 555575.
Babuka, R. (1998), Fuzzy Modeling for Control, Kluwer Academic, Norwell, MA, pp. 4952.
Karason, S. P., and Annaswamy, A. M. (1994), Adaptive Control in the Presence of Input
Constraints, IEEE Transactions on Automatic Control, Vol. 39, No. 11, pp. 2325
2330, doi:10.1109/9.333787.
Krsti, M., Kokotovi, P. V., and Kanellakopoulos, I. (1993), Transient Performance
Improvement with a New Class of Adaptive Controllers, Systems and Control
Letters, Vol. 21, No. 6, pp. 451461, doi:10.1016/0167-6911(93)90050-G.
Sonneveldt, L., van Oort, E. R., Chu, Q. P., and Mulder, J. A. (2007), Comparison of Inverse
Optimal and Tuning Functions Designs for Adaptive Missile Control,
AIAA Guidance, Navigation, and Control Conference and Exhibit, AIAA Paper
2007-6675.
Page, A. B., and Steinberg, M. L. (1999), Effects of Control Allocation Algorithms on a
Nonlinear Adaptive Design, AIAA Guidance, Navigation, and Control
Conference and Exhibit, AIAA, Reston, VA, pp. 16641674; also AIAA Paper 1999-
4282.
3
Hybrid Adaptive Flight Control with

Model Inversion Adaptation
Nhan Nguyen
NASA Ames Research Center
United States of America
1. Introduction
Adaptive ight control is a potentially promising technology that can improve aircraft stability
and maneuverability. In recent years, adaptive control has been receiving a signicant amount
of attention. In aerospace applications, adaptive control has been demonstrated in many ight
vehicles. For example, NASA has conducted a ight test of a neural net intelligent ight
control system on board a modied F-15 test aircraft (Bosworth & Williams-Hayes, 2007).
The U.S. Air Force and Boeing have developed a direct adaptive controller for the Joint
Direct Attack Munitions (JDAM) (Sharma et al., 2006). The ability to accommodate system
uncertainties and to improve fault tolerance of a ight control system is a major selling
point of adaptive control since traditional gain-scheduling control methods are viewed as
being less capable of handling off-nominal ight conditions outside a normal ight envelope.
Nonetheless, gain-scheduling control methods are robust to disturbances and unmodeled
dynamics when an aircraft is operated as intended.
In spite of recent advances in adaptive control research and the potential benets of
adaptive control systems for enhancing ight safety in adverse conditions, there are several
challenges related to the implementation of adaptive control technologies in ight vehicles
to accommodate system uncertainties. These challenges include but are not limited to: 1)
robustness in the presence of unmodeled dynamics and exogenous disturbances (Rohrs et al.,
1985); 2) quantication of performance and stability metrics of adaptive control as related to
adaptive gain and input signals; 3) adaptation in the presence of actuator rate and position
limits; 4) cross-coupling between longitudinal and lateral-directional axes due to failures,
damage, and different rates of adaptation in each axis; and 5) on-line reconguration and
control reallocation using non-traditional control effectors such as engines with different rate
limits.
The lack of a formal certication process for adaptive control systems poses a major hurdle
to the implementation of adaptive control in future aerospace systems (Jacklin et al., 2005;
Nguyen & Jacklin, 2010). This hurdle can be traced to the lack of well-dened performance
and stability metrics for adaptive control that can be used for the verication and validation
of adaptive control systems. Recent studies by a number of authors have attempted to address
metric evaluation for adaptive control systems (Annaswamy et al., 2008; Nguyen et al., 2007;
Stepanyan et al., 2009; Yang et al., 2009). Thus, the development of veriable metrics for
adaptive control will be important in order to mature adaptive control technologies in the
future.
Over the past several years, various model-reference adaptive control (MRAC) methods have
been investigated (Cao & Hovakimyan, 2008; Eberhart & Ward, 1999; Hovakimyan et al., 2001;
Johnson et al., 2000; Kim & Calise, 1997; Lavretsky, 2009; Nguyen et al., 2008; Rysdyk & Calise,
1998; Steinberg, 1999). The majority of MRAC methods may be classied as direct, indirect,
or a combination thereof. Indirect adaptive control methods are based on identication
of unknown plant parameters and certainty-equivalence control schemes derived from the
parameter estimates which are assumed to be their true values (Ioannu & Sun, 1996).
Parameter identication techniques such as recursive least-squares and neural networks have
been used in many indirect adaptive control methods (Eberhart & Ward, 1999). In contrast,
direct adaptive control methods adjust control parameters to account for system uncertainties
directly without identifying unknown plant parameters explicitly. MRAC methods based on
neural networks have been a topic of great research interest (Johnson et al., 2000; Kim & Calise,
1997; Rysdyk & Calise, 1998). Feedforward neural networks are capable of approximating a
generic class of nonlinear functions on a compact domain within arbitrary tolerance (Cybenko,
1989), thus making them suitable for adaptive control applications. In particular, Rysdyk
and Calise described a neural net direct adaptive control method for improving tracking
performance based on a model inversion control architecture (Rysdyk & Calise, 1998). This
method is the basis for the intelligent ight control system that has been developed for the
F-15 test aircraft by NASA. Johnson et al. introduced a pseudo-control hedging approach for
dealing with control input characteristics such as actuator saturation, rate limit, and linear
input dynamics (Johnson et al., 2000). Hovakimyan et al. developed an output feedback
adaptive control to address issues with parametric uncertainties and unmodeled dynamics
(Hovakimyan et al., 2001). Cao and Hovakimyan developed an L1 adaptive control method
to address high-gain control (Cao & Hovakimyan, 2008). Nguyen developed an optimal
control modication scheme for adaptive control to improve stability robustness under fast
adaptation (Nguyen et al., 2008).
While adaptive control has been used with success in many applications, the possibility of
high-gain control due to fast adaptation can be an issue. In certain applications, fast adaptation
is needed in order to improve the tracking performance rapidly when a system is subject to
large uncertainties such as structural damage to an aircraft that could cause large changes
in aerodynamic characteristics. In these situations, large adaptive gains can be used for
adaptation in order to reduce the tracking error quickly. However, there typically exists a
balance between stability and fast adaptation. It is well known that high-gain control or fast
adaptation can result in high frequency oscillations which can excite unmodeled dynamics
that could adversely affect stability of an MRAC law (Ioannu & Sun, 1996). Recognizing
this, some recent adaptive control methods have begun to address fast adaptation. One such
method is the L1 adaptive control (Cao & Hovakimyan, 2008) which uses a low-pass lter
to effectively lter out any high frequency oscillation that may occur due to fast adaptation.
Another approach is the optimal control modication that can enable fast adaptation while
maintaining stability robustness (Nguyen et al., 2008).
This study investigates a hybrid adaptive ight control method as another possibility to
reduce the effect of high-gain control (Nguyen et al., 2006). The hybrid adaptive control blends
both direct and indirect adaptive control in a model inversion ight control architecture.
The blending of both direct and indirect adaptive control is sometimes known as composite
adaptation (Ioannu & Sun, 1996). The indirect adaptive control is used to update the model
Hybrid Adaptive Flight Control with Model Inversion Adaptation 55
inversion controller by two parameter estimation techniques: 1) an indirect adaptive law

based on the Lyapunov theory, and 2) a recursive least-squares indirect adaptive law. The
model inversion controller generates a command signal using estimates of the unknown
plant dynamics to reduce the model inversion error. This directly leads to a reduced tracking
error. Any residual tracking error can then be further reduced by a direct adaptive control
which generates an augmented reference command signal based on the residual tracking
error. Because the direct adaptive control only needs to adapt to a residual uncertainty, its
adaptive gain can be reduced in order to improve stability robustness. Simulations of the
hybrid adaptive control for a damaged generic transport aircraft and a pilot-in-the-loop ight
simulator study show that the proposed method is quite effective in providing improved
command tracking performance for a ight control system.
2. Hybrid adaptive ight control

Consider a rate-command-attitude-hold (RCAH) inner loop ight control design. The
objective of the study is to design an adaptive law that allows an aircraft rate response to
accurately follow a rate command. Assuming that the airspeed is regulated by the engine
thrust, then the rate equation for an aircraft can be written as
= + (1)

where = p q r is the inner loop angular rate vector, is the uncertainty in the plant
model which can include nonlinear effects, and is the nominal plant model where
= F1 + F2 + G (2)
with F1 , F2 , G R 33 as nominal state transition and control sensitivity matrices which

are assumed to be known, = is the outer loop attitude vector which has

slower dynamics than the inner loop rate dynamics, and = a e r is the actuator
command vector to ight control surfaces.
Fig. 1. Hybrid Adaptive Flight Control Architecture

Figure 1 illustrates the proposed hybrid adaptive ight control. The control architecture
comprises: 1) a reference model that translates a rate command into a desired acceleration
command, 2) a proportional-integral (PI) feedback control for rate stabilization and tracking,
3) a model inversion controller that computes the actuator command using the desired
acceleration command, 4) a neural net direct adaptive control augmentation, and 5) an
indirect adaptive control that adjusts the model inversion controller to match the actual plant
dynamics. The tracking error between the reference trajectory and the aircraft state is rst
reduced by the model inversion indirect adaptation. The neural net direct adaptation then
further reduces the tracking error by estimating an augmented acceleration command to
compensate for the residual tracking error. Without the model inversion indirect adaptation,
the possibility of a high-gain control can exist with only the direct adaptation in use since
a large adaptive gain needs to be used in order to reduce the tracking error rapidly. A
high-gain control may be undesirable since it can lead to high frequency oscillations in the
adaptive signal that can potentially excite unmodeled dynamics such as structural modes. The
proposed hybrid adaptive control can improve the performance of a ight control system by
incorporating a model inversion indirect adaptation in conjunction with a direct adaptation.
The inner loop rate feedback control is designed to improve aircraft rate response
characteristics such as the short period mode and the dutch roll mode. A second-order
reference model is specied to provide desired handling qualities with good damping and
natural frequency characteristics as follows:

s2 + 2 p p s + 2p m = c p lat (3)

s2 + 2 q q s + q2 m = cq lon (4)

s2 + 2 r r s + r2 rm = cr rud (5)
where m , m , and m are reference bank, pitch, and heading angles; lat , lon , and rud are the
lateral stick input, longitudinal stick input, and rudder pedal input; p , q , and r are the
natural frequencies for desired handling qualities in the roll, pitch, and yaw axes; p , q , and
r are the desired damping ratios; and c p , cq , and cr are stick gains.
Let pm = m , q m = m , and rm = m be the reference roll, pitch, and yaw rates. Then the
reference model can be represented as
t
m = K p m Ki m d + cc (6)
0

where m = pm q m rm , K p = diag 2 p p , 2 q q , 2 r r , Ki = diag 2p , q2 , r2 , c =

diag c p , cq , cr , and c = lat lon rud .
A model inversion controller is computed to obtain an estimated control surface deection
command to achieve a desired acceleration d as

= G 1 d F1 F2 (7)
where F1 , F2 , and G are the unknown plant matrices to be estimated by an indirect adaptive
law which updates the model inversion controller; and moreover G is ensured to be invertible
by verifying its matrix conditioning number.
In order for the controller to track the reference acceleration m , the desired acceleration d is
computed as
t
d = m + K p e + Ki e d u ad (8)
0
where e = m is dened as a rate tracking error, and u ad is a direct adaptive signal
designed to reduce the tracking error to small bound away from zero in order to provide
stability robustness.
Because the true plant dynamics are unknown, the model inversion controller incurs a
modeling error equal to
t
d = m K p e Ki e d + u ad (9)
0
but from Eq. (7) the model inversion controller is also equal to

d = F1 F1 F2 F2 G G (10)
where = is the unknown plant model error.

Comparing these two equations, the tracking error equation is formed as
e = Ae + Bu ad + B F1 + B F2 + B G B (11)

where e = 0
t
e d e is the tracking error, F1 = F1 F1 , F2 = F2 F2 , G = G G , and

0 I 0
A= , B= (12)
Ki K p I
The direct adaptive signal u ad is computed from a single-layer sigma-pi neural network
u ad = W (13)

where W R m3 is a neural network weight matrix, and = C1 C2 C3 C4 C5 R m 1
is a basis function with Ci , i = 1, . . . , 5, as inputs into the neural network consisting of control
commands, sensor feedback, and bias terms; dened as follows

C1 = V 2 (14)

C2 = V 2 1 2 2 (15)

C3 = V 2 (16)

C4 = p q r (17)

C5 = 1 T (18)
where T in C5 is an engine throttle parameter.
These basis functions are designed to model the unknown nonlinearity that exists in the
unknown plant model. For example, the aerodynamic force in the x- axis for an aircraft can be
expressed as
1
Fx = T Tmax + V 2 S CL0 + CL + CL + CL + CL
2
1
V 2 S CD0 + CD + CD + CD + CD (19)
2
where the engine thrust is replaced by T Tmax and Tmax is the maximum engine thrust.
Thus, C1 , C2 , and C3 are designed to model the product terms of , , , and in the
aerodynamic and propulsive forces. Similarly, C4 models the cross-coupling terms of the
aircraft rates in the moment equations, and C5 models the effects the gravity and propulsive
force. Alternatively, the basis function can also be formed from a subset of Ci , i = 1, 2, . . . , 5.
The update law for the neural net weights W is due to Rysdyk and Calise (Rysdyk & Calise,
1998) and is given by

W = e PB + e PB W (20)
where = > 0 R mm is an adaptive gain matrix, > 0 R is an e-modication

parameter (Narendra & Annaswamy, 1987), . is a Frobenius norm, and P = P > 0 R 66
solves the Lyapunov equation
PA + A P = Q (21)
for some positive-denite matrix Q = Q > 0 R 66 .
The goal is to compute F1 , F2 , and G by a model inversion indirect adaptive law. The indirect
adaptive law updates the estimates of F1 , F2 , and G so that the model inversion controller
can accommodate as much as possible the effects of the unknown plant dynamics. Two
approaches are considered: 1) an indirect adaptive law based on the Lyapunovs direct
method, and 2) a recursive least-squares indirect adaptive law for parameter estimation of
the unknown plant model. Both of these approaches are described as follows:
2.1 Lyapunov-Based indirect adaptive law

The hybrid adaptive control with model inversion adaptation can be implemented by the
following indirect adaptive law

= e PB + e PB (22)

where = W W W R 3 p is a weight matrix, =
R p1 is an input matrix of state and control vectors, = diag ( , , ) > 0 R p p is an
adaptive gain matrix, and = diag ( I, I, I ) > 0 R p p is an e-modication parameter
matrix.
Then the estimates of F1 , F2 , and G can be computed as
F1 = F1 + W (23)
F2 = F2 + W (24)

G = G + W (25)
The basis functions , , and are designed to model the nonlinearity in the plant model
error. For example, if the plant model error is given by
= A1 + A2 + A3 (26)

then W = A1 A2 A3 and = I I I .
The tracking error then becomes
e = Ae + BW + B B (27)
The indirect adaptive law (22) can be shown to provide a stable estimation of the unknown
plant matrices F1 , F2 , and G as follows:
Proof: The matrix A is Hurwitz. Let W = W + W and = + where the asterisk
symbol denotes the ideal weight matrices that cancel out the unknown plant model error
and the tilde symbol denotes the weight deviations. The ideal weight matrices are unknown
but they may be assumed constant and are bounded to stay within a -neighborhood of the
plant model error , assuming that the input or the command c L is bounded. Then

= sup W + (28)
,,
Choose the following Lyapunov candidate function

V = e Pe + tr W 1 W + 1 (29)
where tr (.) denotes the trace operation.

The time derivative of the Lyapunov candidate function is computed as

V = e Pe + e P e + 2tr W 1 W
+ 1
(30)
which upon substitution yields

V = e PA + A P e + 2e PB W +

+ 2tr W e PB + e PB W e PB + e PB (31)
Utilizing the trace operation tr ( XY ) = YX, where X is a column vector and Y is a row vector,
then
2tr W e PB = 2e PBW (32)

2tr e PB = 2e PB (33)
Completing the square yields
2 2
W W

2tr W e PB W + W = 2 e PB 2 + W 2

e PB W W 2
2
(34)
2

2tr e PB + 2 e PB min ( ) 2 +

2
2
max ( ) e
PB

( ) max ( ) 2 (35)
2 min
where . is a Frobenius norm, and min and max are the maximum and minimum
eigenvalues, respectively.
Then, substituting back into V gives

V e Qe + 2e PB e PB W W 2
2
2

e PB min ( ) max ( ) 2 (36)
Since B = 1, it can be established that

V min ( Q) e2 + P e 2 W W 2
2
2
min ( ) + max ( ) 2 (37)
which can also be expressed as

V e min ( Q) e P 2 + W 2 + max ( ) 2
2 2
+ P W + min ( ) P (38)
Let S be a compact set dened as

2 2
S = e, W, : min ( Q) e + P W + min ( ) P r (39)
where
r = P 2 + W 2 + max ( ) 2 (40)
Then V 0 outside the compact set S . Also there exist functions 1 , 2 KR where

1 e , W , = min ( P ) e2 + min 1 W + min 1
2 2
(41)

2 e , W , = max ( P ) e2 + max 1 W + max 1
2 2
(42)
such that
1 e , W , V 2 e , W , (43)
Then, according to Theorem 3.4.3 of (Ioannu & Sun, 1996), the solution is uniformly ultimately
bounded. Therefore, the hybrid adaptive control results in stable and bounded tracking error;
i.e., e, W, L .
It should be noted that the bounds on e, W , and depends on . To improve the
tracking performance, the magnitudes of must be kept small. This is predicated upon how
well the neural network can approximate the nonlinear uncertainty in the plant dynamics.
Increasing the adaptive gains and improves the tracking performance but at the same
time degrades stability robustness. On the other hand, the values of and must also be kept
sufciently large to ensure stability robustness, but large values of and can degrade the
tracking performance. Thus, there exists a trade-off between performance and robustness in
selecting the adaptive gains and and the e-modication parameters and .
To ensure that the indirect adaptive law will result in a convergence of the estimates F1 , F2 ,
and G to their steady state values, the input signals must be sufciently rich to excite all
frequencies of interest in the plant dynamics. This condition is known as a persistent excitation
(PE) (Ioannu & Sun, 1996).
2.2 Recursive Least-squares indirect adaptive law

The tracking error equation (11) can be expressed as

e = Ae + Bu ad + B (44)
Suppose the plant model error can be written as
= + = (45)
where is the estimation error of . Then, the estimated plant model error is
= = F1 F2 G (46)
where is the estimated acceleration.

The model inversion adaptation using the recursive least-squares indirect adaptive law is
given by
1
= 2 R (47)
m
1
R = 2 R R (48)
m
where R = R > 0 R p p is a positive denite covariance matrix and m2 is a normalization
factor
m2 = 1 + R (49)
The recursive least-squares indirect adaptive law can be derived as follows:
The estimation error can be minimized by considering the following cost function
t 2
1
J () = d (50)
2m2 0
To minimize the cost function, the gradient of the cost function with respect to the weight
matrix is computed and set to zero, thus resulting in
t
1
J = d = 0 (51)
m2 0
Equation (51) is then written as

t t
1 1
d = d (52)
m2 0 m2 0
Let t
1
R 1 = d > 0 (53)
m2 0
Differentiating Eq. (53) yields
dR1 1
= 2 (54)
dt m
It is noted that
dR1
R 1 R = I R + R1 R = 0 (55)
dt
Solving for R yields Eq. (48).
Also, differentiating Eq. (52) yields
1 1
R1 + = 2 (56)
m2 m
Solving for yields the recursive least-squares indirect adaptive law (47) .
The recursive least-squares indirect adaptive law can be shown to provide a stable estimation
of the unknown plant matrices F1 , F2 , and G as follows:
Proof: The steady state ideal weight matrix is assumed to be bounded by a
-neighborhood where

= sup (57)
,,
The ideal weight matrix W is assumed to be bounded inside a neighborhood where

= sup W + sup W + (58)
,, ,,
Choose the following Lyapunov candidate function

L = e Pe + tr W 1 W + R1 (59)
The only difference between L and V is in the last term. Then, the time rate of change of the
Lyapunov candidate function is computed as

L = e Qe + 2e PB W +

2tr W e PB + e PB W

2

dR1
+ tr + (60)
m2 dt
Further simplication yields

2

L e Qe + 2e PB + 2e PB + e PB W 2 W
1 2
2 + 2 (61)
m m
L is then bounded by

L min ( Q) e2 + P e 2 + 2 + W 2
2 1

2
2

P e W 2 + 2 (62)
m m
which can also be expressed as

L e min ( Q) e P 2 + 2 + W 2
2 1

+ P W 2 2 (63)
m
L < 0 if

> 2 (64)
and
2

min ( Q) e + P W > P 2 + 2 + W 2

> P 2 + 4 + W 2 (65)
Let C be a compact set dened as

2

C = e, W, : min ( Q) e + P W r or 2 (66)
where
r = P 2 + 4 + W 2 (67)
Then L 0 outside the compact set C , and so according to Theorem 3.4.3 of (Ioannu &
Sun, 1996), the solution is uniformly ultimately bounded. Therefore, the hybrid adaptive
control results in stable and bounded tracking error; i.e., e, W, L . Thus, the recursive
least-squares indirect adaptive law is stable.
The parameter convergence of the recursive least-squares depends on the persistent excitation
condition on the input signals (Ioannu & Sun, 1996). The update law for the covariance matrix
R has a very similar form to the Kalman lter with Eq. (48) as the differential Riccati equation
for a zero-order plant dynamics. The recursive least-squares indirect adaptive law can also
be implemented in a discrete time form with various modications such as with an adaptive
directional forgetting factor (Bobal et al., 2005) according to
1
i +1 = i + Ri+1 i i+1
i i (68)
m2i+1
1
Ri+1 = Ri i+11 + i+1 Ri i iT Ri (69)
where and are dened as
i+1 = m2i+1 1 (70)
i + 1 = i+1 i1 (1 i +1 ) (71)
The directional forgetting factor is calculated as

i +1 (1 + i +1 ) i +1
1
i +1 = 1 + ( 1 + ) ln ( 1 + i + 1 ) + 1 (72)
1 + i +1 + i +1 1 + i +1
where is a constant, and and are parameters with the following update laws
2
1
i +1 =
i +1 i +1 i i (73)
i +1 = i +1 (1 + i ) (74)
2

k+1 = i+1 k + (1 + i+1 ) i+1
i i (75)
3. Flight control simulations

3.1 Generic transport model
To evaluate the hybrid adaptive ight control method, a simulation was conducted using
a NASA generic transport model (GTM) which represents a notional twin-engine transport
aircraft as shown in Fig. 2 (Jordan et al., 2004). An aerodynamic model of the damaged aircraft
is created using a vortex lattice method to estimate aerodynamic coefcients and derivatives.
A damage scenario is modeled corresponding to a 28% loss of the left wing. The damage
causes an estimated C.G. shift mostly along the pitch axis with y = 0.0388c and an estimated
mass loss of 1.2%. The principal moment of inertia about the roll axis is reduced by 12%, while
changes in the inertia values in the other two axes are not as signicant. Since the damaged
aircraft is asymmetric, the inertia tensor has all six non-zero elements. This means that all the
three roll, pitch, and yaw axes are coupled together throughout the ight envelope.
Fig. 2. Left Wing Damaged Generic Transport Model

A level ight condition of Mach 0.6 at 4572 m is selected. Upon damage, the aircraft is
re-trimmed with T = 0.0705W, = 5.9o , = 3.2o , a = 27.3o , e = 0.5o , r = 1.3o . The
remaining right aileron is the only roll control effector available. In practice, some aircraft can
control a roll motion with spoilers, which are not modeled in this study. The reference model is

specied by p = 2.3 rad/sec, q = 1.7 rad/sec, r = 1.3 rad/sec, and p = q = r = 1/ 2.
slotine
The state space model of the damaged aircraft is given by

p 1.3568 0.2651 0.5220 p
q = 0.0655 0.8947 0.0147 q
r 0.0836 0.0042 0.5135 r

0 10.9985 8.9435
+ 0.0007 2.7041 0.0064
0 0.1841 2.8822

3.2190 0.0451 1.3869 a
+ 0.3391 3.4656 0.0245 e (76)
0.0124 0.0007 2.2972 r

1 0 0.1024 p
= 0.0059 0.9723 0.0004 q
0.0031 0.0002 0.9855 r

0 0 0
+ 0.0028 0.4799 0.0235
0.0507 0.0133 0.1751

0 0 0 a
+ 0.0240 0.0700 0.0011 e (77)
0.0019 0.0001 0.0588 r
4 4
No Adaptation Direct
2 Reference Model 2 Reference Model
q, deg/sec
q, deg/sec
0 0
2 2
4 4
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
4 4
Hybrid Indirect Hybrid RLS
2 Reference Model 2 Reference Model
q, deg/sec
q, deg/sec
0 0
2 2
4 4
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
Fig. 3. Pitch Rate

The pilot pitch rate command is simulated with a series of ramp input longitudinal stick
command doublets, corresponding to the reference pitch angle between 3.1o and 3.1o .
The tracking performance of the baseline ight control with no adaptation versus the three
30 30
20 20
p, deg/sec
p, rad/sec
10 10
0 0
10 10
20 20
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
30 30
20 20
p, deg/sec
p, deg/sec
10 10
0 0
10 10
20 20
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
Fig. 4. Roll Rate
adaptive control methods is compared in Figs. 3 to 6. With no adaptation, there is a signicant

overshoot in the ability for the baseline ight control system to follow the reference pitch rate
as shown in Fig. 3. The performance progressively improves rst with the direct adaptive
control alone, then with the hybrid Lyapunov-based indirect adaptive control, and nally
with the hybrid recursive least-squares (RLS) indirect adaptive control. The Lyapunov-based
indirect adaptive control performs better than the direct adaptive control alone as expected,
since the presence of the Lyapunov-based indirect adaptive law further enhances the ability
for the ight control system to adapt to damage.
0.6
0.4 0.4
r, deg/sec
r, deg/sec
0.2 0.2
0 0
0.2 0.2
0.4 0.4
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
0.6
0.4 0.4
r, deg/sec
r, deg/sec
0.2 0.2
0 0
0.2 0.2
0.4 0.4
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
Fig. 5. Yaw Rate

The most drastic improvement is provided by the hybrid RLS indirect adaptive control which
results in a very good tracking performance in all three control axes. In the pitch axis, the
hybrid RLS indirect adaptive control tracks the reference pitch rate very accurately. In the roll
and yaw axes, the roll and yaw rate responses are maintained close to zero. In contrast, both
30 30
||e||, deg/sec
|| ||, deg/sec
20 20
10 10
e
0 0
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
30 30
||e||, deg/sec
|| ||, deg/sec
20 20
10 10
e
0 0
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
Fig. 6. Tracking Error Norm
the direct adaptive control and the hybrid Lyapunov-based indirect adaptive control improve
the roll and yaw rate responses, but the response amplitudes are still signicant and therefore
can be objectionable particularly in the roll rate.
Figure 6 is the plot of the tracking error norm for all the three angular rates to demonstrate
the effectiveness of the hybrid adaptive control method. The hybrid Lyapunov-based indirect
adaptive control reduces the tracking error by roughly half of that with the direct adaptive
control alone and by a factor of three when there is no adaptation. Moreover, the hybrid
RLS indirect adaptive control drastically reduces the tracking error by more than an order
of magnitude over those with the direct adaptive control and with the baseline ight control.
20 20
0 0
, deg
, deg
20 20
40 40
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
20 20
0 0
, deg
, deg
20 20

40 40
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
Fig. 7. Bank Angle

The attitude responses of the damaged aircraft are shown in Fig. 7 to 9. When there is
no adaptation, the damaged aircraft exhibits a rather severe roll behavior with the bank
angle ranging from almost 40o to 20o . The direct adaptive control improves the situation
10 10
8 8
6 6
, deg
, deg
4 4
2 2
0 0
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
10 10
8 8
, deg
6
, deg
4 4
2 2
0 0
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
Fig. 8. Angle of Attack
signicantly and cuts down the bank angle to a range between about 30o and 10o . With the
hybrid RLS indirect adaptive control, the bank angle is essentially maintained at its trim value.
The angle of attack as shown in Fig. 8 is in a reasonable range. The angle of attack when there
is no adaptation goes through a large swing from 1o to 9o , but the hybrid RLS indirect adaptive
control reduces the angle of attack to a range between 3o and 8o .
Figure 9 shows the plot of the sideslip angle. In general, ying with sideslip angle is not a
recommended practice since a large sideslip angle can cause an increase in drag and more
importantly a decrease in the yaw damping. With no adaptation, the largest negative sideslip
angle is about 3o . This is still within a reasonable limit, but the swing from 3o to 1o can
cause objectionable handling qualities. With the hybrid RLS indirect adaptive control, the
sideslip angle is retained virtually at zero.
2 2
0 0
, deg
, deg
2 2
4 4
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
2 2
0 0
, deg
, deg
2 2
4 4
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
Fig. 9. Sideslip Angle

The control surface deections are plotted in Figs. 10 to 12. Because of the wing damage, the
damaged aircraft has to be trimmed with a rather large aileron deection. This causes the
roll control authority to severely decrease. Any pitch maneuver can potentially run into a
control saturation in the roll axis due to the pitch-roll coupling that exists in a wing damage
scenario. With the maximum aileron deection at 35o , it can be seen clearly that a roll control
saturation is present in all cases, being the worst when there is no adaptation and the best
with the hybrid RLS indirect adaptive control. The range of aileron deection when there
is no adaptation is quite large. As the aileron deection hits the maximum position limit, it
tends to over-compensate in the down swing because of the large pitch rate error produced by
the control saturation. Both the direct adaptive control alone and the hybrid Lyapunov-based
indirect adaptive control alleviate the situation somewhat but the control saturation is still
present. The hybrid RLS indirect adaptive control is apparently very effective in dealing with
the control saturation problem. As can be seen, it results in only a small amount of control
saturation, and the aileron deection does not vary widely. The hybrid RLS indirect adaptive
control essentially enables the aileron to operate almost at its full authority, whereas with the
other control methods, only partial control authority is possible.
40 40
30 30
a, deg
a, deg
20 20
10 10
0 0
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
40 40
30 30
a, deg
a, deg
20 20
10 10
0 0
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
Fig. 10. Aileron Deection

Figure 11 is the plot of the elevator deection that shows similar elevator deections to be
within a range of few degrees for all the four different controllers. This implies that the roll
control contributes mostly to the response of the damaged aircraft.
The rudder deection is shown in Fig. 12. With no adaptation, the rudder deection is quite
active, going from 5o to 0o . While this appears small, it should be compared relative to the
rudder position limit, which is usually reduced as the airspeed and altitude increase. The
absolute rudder position limit is 10o but in practice the actual rudder position limit may
be less. Therefore, it is usually desired to keep the rudder deection as small as possible. The
direct adaptive control results in a maximum negative rudder deection of 4o and the hybrid
Lyapunov-based indirect adaptive control further reduces it to 2o . The hybrid RLS indirect
adaptive control produces the smallest rudder deection and keeps it to less than 0.5o from
the trim value.
2 2
1 1
e, deg
e, deg
0 0
1 1
2 2
3 3
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
2 2
1 1
e, deg
e, deg
0 0
1 1
2 2
3 3
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
Fig. 11. Elevator Deection
2 2
0 0
r, deg
r, deg
2 2
4 4
6 6
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
2 2
0 0
r, deg
r, deg
2 2
4 4
6 6
0 10 20 30 40 0 10 20 30 40
t, sec t, sec
Fig. 12. Rudder Deection
3.2 Piloted ight simulator

The Crew-Vehicle System Research Facility (CVSRF) at NASA Ames Research Center houses
two motion-based ight simulators, the Advanced Concept Flight Simulator (ACFS) and the
Boeing 747-400 Flight Simulator for use in human factor and ight simulation research. The
ACFS has a highly customizable ight simulation environment that can be used to simulate a
wide variety of transport-type aircraft. The ACFS employs advanced y-by-wire digital ight
control systems with modern features that can be found in todays modern aircraft. The ight
deck includes head-up displays, a customizable ight management system, and modern ight
instruments and electronics. Pilot inputs are provided by a side stick for controlling aircraft in
pitch and roll axes.
Recently, a piloted study has been conducted in the ACFS to evaluate a number of adaptive
control methods (Campbell et al., 2010). A high-delity ight dynamic model was developed
to simulate a medium-range generic transport aircraft. The model includes aerodynamic

models of various aerodynamic surfaces including aps, slats, and other control surfaces. The
aerodynamic database is based on Reynolds number corrected wind tunnel data obtained
from wind tunnel testing of a sub-scale generic transport model. The ground model with
landing gears as well as ground effect aerodynamic model are also included.
A number of failure and damage emulations were implemented including asymmetric
damage to the left horizontal tail and elevator, ight control faults emulated by scaling the
control sensitivity matrix (B-matrix failures), and combined failures. Eight different NASA
test pilots were requested to participate in the study. For each failure emulation, each pilot
was asked to provide Cooper-Harper Ratings (CHR) for a series of ight tasks, which included
large amplitude attitude capture tasks and cross-wind approach and landing tasks.
Fig. 13. Advanced Concept Flight Simulator at NASA Ames

Seven adaptive control methods were selected for the piloted study that include
e-modication (Narendra & Annaswamy, 1987), hybrid adaptive control (Nguyen et al.,
2006), optimal control modication (Nguyen et al., 2008), metric-driven adaptive control
using bounded linear stability method (Nguyen et al., 2007), L1 adaptive control (Cao &
Hovakimyan, 2008), adaptive loop recovery (Calise et al., 2009), and composite adaptive
control (Lavretsky, 2009). This is by no means an exhaustive list of new advanced adaptive
control methods that have been developed in the past few years, but this list provides an
Fig. 14. Pilot Evaluation of Adaptive Flight Control
initial set of adaptive control methods that could be implemented under an existing NASA
partnership with the industry and academia sponsored by the NASA Integrated Resilient
Aircraft Control (IRAC) project.
The study generally conrms that adaptive control can clearly provide signicant benets
to improve aircraft ight control performance in adverse ight conditions. The study also
provides an insight of the role of pilot interactions with adaptive ight control systems. It
was observed that many favorable pilot ratings were associated with those adaptive control
methods that provide a measure of predictability, which is an important attribute of a ight
control system design. Predictability can be viewed as a measure of how linear the aircraft
response is to a pilot input. Being a nonlinear control method, some adaptive control methods
can adversely affect linear behaviors of a ight control system more than others. Thus, while
these adaptive control methods may appear to work well in a non-piloted simulation, they
may present potential issues with pilot interactions in a realistic piloted ight environment.
Thus, understanding pilot interaction issues is an important consideration in future research
of adaptive ight control.
With respect to pilot handling qualities, among the seven adaptive ight controllers evaluated
in the study, the optimal control modication, the adaptive loop recovery, and the composite
adaptive control appeared to perform well over all ight conditions (Campbell et al., 2010).
The hybrid adaptive control also performs reasonably well in most cases. For example, with
the B-matrix failure emulation, the average CHR was 5 for 8 capture tasks with the baseline
dynamic inversion ight controller. The average CHR number was improved to 3 with the
hybrid adaptive control. In only one type of failure emulations that involved cross-coupling
effects in aircraft dynamics, the performance of the hybrid adaptive ight controller fell below
that for the e-modication which is used as the benchmark for comparison.
Future NASA research in advancing adaptive ight control will include ight testing of some
of the new promising adaptive control methods. Previously, NASA conducted ight testing
of the Intelligent Flight Control (IFC) on a NASA F-15 aircraft up until 2008 (Bosworth &
Fig. 15. Cooper-Harper Rating Improvement of Various Adaptive Control Methods
Williams-Hayes, 2007). In January of 2011, NASA has successfully completed a ight test
program on a NASA F-18 aircraft to evaluate a new adaptive ight controller based on the
Optimal Control Modication (Nguyen et al., 2008). Initial ight test results indicated that the
adaptive controller was effective in improving aircrafts performance in simulated in-ight
failures. Flight testing can reveal new observations and potential issues with adaptive control
in various stages of the design implementation that could not be observed in ight simulation
environments. Flight testing therefore is a critical part of validating any new technology such
as adaptive control that will allow such a technology to transition into production systems in
the future.
4. Conclusions
This study presents a hybrid adaptive ight control method that blends both direct and
indirect adaptive control within a model inversion ight control architecture. Two indirect
adaptive laws are presented: 1) a Lyapunov-based indirect adaptive law, and 2) a recursive
least-squares indirect adaptive law. The indirect adaptive laws perform on-line parameter
estimation and update the model inversion ight controller to reduce the tracking error. A
direct adaptive control is incorporated within the feedback loop to correct for any residual
tracking error.
A simulation study is conducted with a NASA wing-damaged transport aircraft model.
The results of the simulation demonstrate that in general the hybrid adaptive control offers
a potentially promising technique for ight control by allowing both direct and indirect
adaptive control to operate cooperatively to enhance the performance of a ight control
system. In particular, the hybrid adaptive control with the recursive least-squares indirect
adaptive law is shown to be highly effective in controlling a damaged aircraft. Simulation
results show that the hybrid adaptive control with the recursive least-squares indirect
adaptive law is able to regulate the roll motion due to a pitch-roll coupling to maintain a
nearly wing-level ight during a pitch maneuver.
The issue of roll control saturation is encountered due to a signicant reduction in the roll
control authority as a result of the wing damage. The direct adaptive control and the hybrid
adaptive control with the Lyapunov-based indirect adaptive law restore a partial roll control
authority from the control saturation. On the other hand, the hybrid adaptive control with the
recursive least-squares indirect adaptive law restores the roll control authority almost fully.
Thus, the hybrid adaptive control with the recursive least-squares indirect adaptive law can
demonstrate its effectiveness in dealing with a control saturation.
A recent piloted study of various adaptive control methods in the Advanced Concept Flight
Simulator at NASA Ames Research Center conrmed the effectiveness of adaptive control in
improving ight safety. The hybrid adaptive control was among the methods evaluated in the
study. In general, it has been shown to provide an improved ight control performance under
various types of failure emulations conducted in the piloted study.
In summary, the hybrid adaptive ight control is a potentially effective adaptive control
strategy that could improve the performance of a ight control system when an aircraft
operating in adverse events such as with damage and or failures.
5. References
Annaswamy, A.; Jang, J. & Lavretsky, E. (2008). Stability Margins for Adaptive Controllers in
the Presence of Time-Delay, AIAA Guidance, Navigation, and Control Conference,
Honolulu, Hawaii, August 2008, AIAA 2008-6659.
Bobal, V.; Buhm, J.; Fessl, J. & Machacek,
J. (2005). Digital Self-Tuning Controllers: Algorithms,
Implementation, and Applications, Springer-Verlag, ISBN 1852339802, London,
Bosworth, J. & Williams-Hayes, P. (2007). Flight Test Results from the NF-15B IFCS Project with
Adaptation to a Simulated Stabilator Failure, AIAA Infotech@Aerospace Conference,
Rohnert Park, California, May 2007, AIAA-2007-2818.
Calise, A.; Yucelen, T.; Muse, J. & Yang, B. (2009). A Loop Recovery Method for Adaptive
Control, AIAA Guidance, Navigation, and Control Conference, Chicago, Illinois,
August 2009, AIAA-2009-5967.
Campbell, S.; Kaneshige, J.; Nguyen, N. & Krishnakumar, K. (2010). An Adaptive Control
Simulation Study using Metrics and Pilot Handling Qualities Evaluations, AIAA
Guidance, Navigation, and Control Conference, Toronto, Canada, August 2010,
AIAA-2010-8013.
Cao, C. & Hovakimyan, N. (2008). Design and Analysis of a Novel L1 Adaptive Control
Architecture with Guaranteed Transient Performance. IEEE Transactions on Automatic
Control, Vol. 53, No. 2, March 2008, pp. 586-591.
Cybenko, G. (1989). Approximation by Superpositions of a Sigmoidal Function. Mathematics
of Control Signals Systems, Vol. 2, No. 4, 1989, pp. 303-314.
Eberhart, R. L. & Ward, D. G. (1999). Indirect Adaptive Flight Control System Interactions.
International Journal of Robust and Nonlinear Control, Vol. 9, No. 14, December 1999,
pp. 1013-1031.
Hovakimyan, N.; Kim, N.; Calise, A. J.; Prasad, J.V. R. & Corban, E. J. (2001). Adaptive
Output Feedback for High-Bandwidth Control of an Unmanned Helicopter, AIAA
Guidance, Navigation and Control Conference, Montreal, Canada, August 2001,
AIAA-2001-4181.
Ioannu, P.A. & Sun, J. (1996). Robust Adaptive Control, Prentice-Hall, ISBN 0134391004.
Jacklin, S. A.; Schumann, J. M.; Gupta, P. P.; Richard, R.; Guenther, K. & Soares, F.
(2005). Development of Advanced Verication and Validation Procedures and
Tools for the Certication of Learning Systems in Aerospace Applications, AIAA
Infotech@Aerospace Conference, Arlington, VA, September 2005, AIAA-2005-6912.
Johnson, E. N.; Calise, A. J.; El-Shirbiny, H. A. & Rysdyk, R. T. (2000). Feedback Linearization
with Neural Network Augmentation Applied to X-33 Attitude Control, AIAA
Guidance, Navigation, and Control Conference, Denver, Colorado, August 2000,
AIAA-2000-4157.
Jordan, T. L.: Langford, W. M.; Belcastro, Christine M.; Foster, J. M.; Shah, G. H.; Howland, G.
& Kidd, R. (2004). Development of a Dynamically Scaled Generic Transport Model
Testbed for Flight Research Experiments, AUVSI Unmanned Unlimited, Arlington,
VA, 2004
Kim, B. S. & Calise, A. J. (1997). Nonlinear Flight Control Using Neural Networks. AIAA
Journal of Guidance, Control, and Dynamics, Vol. 20, No. 1, 1997, pp. 26-33.
Lavretsky, E. (2009). Combined / Composite Model Reference Adaptive Control, AIAA
Guidance, Navigation, and Control Conference, Chicago, Illinois, August 2009,
AIAA-2009-6065
Narendra, K. S. & Annaswamy, A. M. (1987). A New Adaptive Law for Robust Adaptation
Without Persistent Excitation. IEEE Transactions on Automatic Control, Vol. 32, No. 2,
February 1987, pp. 134-145.
Nguyen, N.; Krishnakumar, K.; Kaneshige, J. & Nespeca, P. (2006). Dynamics and
Adaptive Control for Stability Recovery of Damaged Asymmetric Aircraft, AIAA
Guidance, Navigation, and Control Conference,Keystone, Colorado, August 2006,
AIAA-2006-6049.
Nguyen, N.; Bakhtiari-Nejad, M. & Huang, Y. (2007). Hybrid Adaptive Flight Control
with Bounded Linear Stability Analysis, AIAA Guidance, Navigation, and Control
Conference, Hilton Head, South Carolina, August 2007, AIAA 2007-6422.
Nguyen, N.; Krishnakumar, K. & Boskovic, J. (2008). An Optimal Control Modication
to Model-Reference Adaptive Control for Fast Adaptation, AIAA Guidance,
Navigation, and Control Conference, Honolulu, Hawaii, August 2008, AIAA
2008-7283.
Nguyen, N. & Jacklin, S. (2010). Neural Net Adaptive Flight Control Stability, Verication
and Validation Challenges, and Future Research, In: Applications of Neural Networks
in High Assurance Systems, Schumann, J. & Liu, Y., (Ed.), pp. 77-107, Springer-Verlag,
ISBN 978-3-642-10689-7, Berlin.
Rohrs, C.E.; Valavani, L.; Athans, M. & Stein, G. (1985). Robustness of Continuous-Time
Adaptive Control Algorithms in the Presence of Unmodeled Dynamics. IEEE
Transactions on Automatic Control, Vol. 30, No. 9, September 1985, pp. 881-889.
Rysdyk, R. T. & Calise, A. J. (1998). Fault Tolerant Flight Control via Adaptive Neural
Network Augmentation, AIAA Guidance, Navigation, and Control Conference,
Boston, Massachusetts, August 1998, AIAA-1998-4483.
Sharma, M.; Lavretsky, E. & and Wise, K. (2006). Application and Flight Testing of an Adaptive
Autopilot On Precision Guided Munitions, AIAA Guidance, Navigation, and Control
Conference, Keystone, Colorado, August 2006, AIAA-2006-6568.
Steinberg, M. L. (1999). A Comparison of Intelligent, Adaptive, and Nonlinear Flight Control
Laws, AIAA Guidance, Navigation, and Control Conference, Portland, Oregon,
August 1999, AIAA-1999-4044.
Stepanyan, V.; Krishnakumar, K.; Nguyen, N. & Van Eykeren, L. (2009). Stability and
Performance Metrics for Adaptive Flight Control, AIAA Guidance, Navigation, and
Control Conference, Chicago, Illinois, August 2009, AIAA-2009-5965.
Yang, B.-J.; Yucelen, T.; Calise, A. J. & Shin, J.-Y. (2009). LMI-based Analysis of Adaptive
Controller, American Control Conference, June 2009.
4
Application of Evolutionary Computing in

Control Allocation
Hammad Ahmad, Trevor Young, Daniel Toal and Edin Omerdic
University of Limerick,
Republic of Ireland
1. Introduction
This issue of interaction of control allocation and actuator dynamics and has been dealt with
by very few researchers. What was not considered in most control allocation algorithms is
the fact that the control surfaces are manipulated by either hydraulics or electric actuators,
and constitute a dynamic system which cannot produce infinite accelerations. In other
words, if a control was initially at rest, and later commanded to move at its maximum rate
in some direction for a specified amount of time, it would gradually build up speed until it
reached the commanded rate. The final position of the control would therefore not be the
same as that calculated using the commanded rate and the time during which it was
instructed to move (Bolling 1997). In this chapter, a method, which post-processes the
output of a control allocation algorithm, is developed to compensate for actuator dynamics.
The method developed is solved for a diagonal matrix of gain corresponding to individual
actuators. This matrix is then multiplied with the commanded change in control effector
settings as computed by the control allocator and actuators dynamics interactions. The basic
premise of this method is to post process the output of the control allocation algorithm to
overdrive the actuators so that at the end of a sampling interval the actual actuator positions
are equivalent to the desired actuator positions (Oppenheimer and Doman 2004). The
overdriving of the actuators is done by multiplying the change in commanded signal with
the identified gain matrix which is called the compensator. This identification is done by
using a soft computing technique (i.e. genetic algorithms). The simulation setup including
control allocator block, compensator and actuator rig makes a non-linear set up. During the
identification of the compensator using this setup by soft computing technique such as
genetic algorithms, the likelihood of the solution being a global minimum is high as
compared to other optimisation techniques. This is why genetic algorithms have been used
in this analysis rather than other techniques such as linear programming. The main
contribution is to design a compensator using an evolutionary computing technique (i.e.
genetic algorithms) to compensate the interaction between control allocation and actuator
dynamics. It should be mentioned that in this method the model of the actuator does not
need to be known. The simulation setup consists of excitation signals, the control allocation
block, the compensator and the actuators rig.
When designing control allocation typically the actuator dynamics are ignored because the
bandwidth of the actuators is larger than the frequencies of the rigid body modes of the
aircraft. Fig. 1 shows a control allocator with actuator dynamics neglected. If there is a case
in which actuator frequencies are comparable with the bandwidth of the rigid body modes
then the actuator dynamics cannot be neglected, as shown in Fig. 2.
Find
min u ud
v u 2 u~ cmd
subject to 11 1
Bc u = v
uuu
Fig. 1. Control allocation with actuator dynamics neglected
Fig. 2. Control allocation with actuator dynamics

In this case the output of the control allocator should match the output of the actuator
dynamics. In reality the output of the control allocation is attenuated due to the presence of
non-negligible actuator dynamics. The loss of the gain from the CA output signal is
compensated by the scheme shown in Fig. 3. In the second order dynamics of the actuator
the rate could be estimated using a Kalman filter if the rate sensing is not available. The
Kalman filter is an efficient recursive filter that estimates the state of a dynamic system from
a series of noisy measurements. The matrix of gains as shown in Fig. 3 is tuned offline using
GA. The structure of the compensator is taken from (Oppenheimer and Doman 2004).
1.1 Control allocation for aircraft: graphical illustration

Control allocation is merely a mapping (i.e. linear or non-linear) from total virtual demands
in terms of body angular accelerations to the control position setting subject to rate and
position constraints. An illustration of control allocation is given in Fig. 4.
Section 2 describes the interaction of first order actuator dynamics and control allocation
and the structure of the compensator is established in this section for first order actuator
dynamics. Similarly, in section 2.2 the structure of the compensator is established for second
Application of Evolutionary Computing in Control Allocation 79
u
ucmd u
Find
min u ud ~
v u 2 ucmd ucmd
subject to u
Bc u = v
uuu
Fig. 3. Structure of compensator with actuator dynamics with diagonal gain matrix M of
dimension (11X11)
Fig. 4. Control allocation for aircraft: graphical illustration

order actuator dynamics. In section 3 tuning of the compensator parameters using genetic
algorithm is described. In section 4 simulation and results for a tuned compensator are
shown for a range of first and second order actuator dynamics. Finally, in section 5 some
conclusions are established.
2. First-order actuator dynamics interaction

In this section, the effects of first-order actuator on the system are shown in Fig. 2
(Oppenheimer and Doman 2004). Let the dynamics of a single actuator be represented by a
continuous time first order transfer function of the form
(1)
The discrete time solution to the first-order actuator dynamic equation for one sample
period is given by
T (2)
e
where is the sampling time. This result does not depend on the type of hold because is
specified in terms of its continuous time history, over a sample interval (Franklin et
al. 1998). The most common hold element is zero-order hold (ZOH) with no delay, i.e.
, (3)
Performing substitution
(4)
in Eq. (2) yields
T (5)
e
Defining
T
e
(6)

Eq. (5) can be written as a difference equation of standard form
1 u k (7)
The signal , is held constant over each sampling period. The command to actuator is
given by
(8)
The command increment change in actuator position over one sample as shown in Fig. 5 is
defined by
(9)
Fig. 5. Command increment change in actuator position with gain matrix M equal to
Identity matrix I of dimension (11X11)
where is the actuator command coming from the control allocator. Since the effector
commands are held constant for one sample period then appear to be a step
command from the measured position . Substituting Eq. (8) in Eq. (7) gives
1 (10)
If 1, the incremental commanded signal from the control allocation algorithm,
is attenuated by the actuator dynamics, thus 1 . The objective is to find the
gain M that changes the output of the control allocation algorithm such that 1
(Oppenheimer and Doman 2004). Hence
1 M (11)
The gain M is tuned by using the genetic algorithm in section 3.2. If there is a bank of first
order actuator dynamics, then the gain M is chosen to be a diagonal matrix of dimensions
(11x11), as shown in Fig. 3 and Fig. 5.
2.1 Example showing effect of first-order actuator dynamics

Let us consider an example with
4.2 4.2 5.0 5.0 0 0 0 0 0 2.6 0.7

10 0.9 0.9 2.9 2.9 9.4 9.4 6.9 6.9 80.5 7.7 5.8
0.2 0.2 0.1 0.1 0 0 0 0 0 0 0
Position in (deg) and rate in (deg/s) constraints are defined as follows:
20 20 12 12 23 23 23 23 12 25 25
(12)
20 20 15 15 17 17 17 17 3 25 25
45 45 45 45 37 37 37 37 0.5 50 50
(13)
45 45 45 45 37 37 37 37 0.5 50 50
3X1
Fig. 6. Block diagram with desired demand produced by the control allocator and compared
with the actual demand when there is no actuator dynamics included1
First the time response of control allocation without actuator dynamics is shown in Fig. 6
and Fig 7. It can be seen that if the actuators are fast enough to cater for the rigid body
1 cf. (from Latin confer) means compare

modes, there is no need to consider the actuator dynamics and hence one to one mapping
between the control allocator and control surfaces is sufficient. This would not be the case
with the non aerodynamic actuators, so actuator dynamics cannot be ignored. It can be seen
from the results shown in Fig. 8 and Fig. 9 that how the actuator dynamics affects the
outcome of the control allocator. It can also be seen that how the control allocator command
is attenuated. The first-order actuator dynamics used for this example are given as
0.6128
(14)
0.6128
Fig. 7. Comparison of results with no dynamics of actuator involved
Fig. 8. Block diagram with desired demand produced by the control allocator and compared
with the actual demand when there is actuator dynamics included
Fig. 9. Time responses of desired and actual responses of virtual demand with actuator
dynamics included
In the following section the second order actuator dynamics are parameterised for the
design of the compensator.
2.2 Second-order model dynamics interaction

In this section, the effects of second order actuator on the system is shown in Fig. 2
(Oppenheimer and Doman 2004). Let the dynamics of a second actuator be represented by a
continuous time second order function of the form
(15)
2
The state space representation of this transfer function is given
0 1 0
2
(16)
1 0
(17)
0 1
The discrete time solution to the second-order actuator dynamic Eq. (16) to Eq. (17) for one
sample period is given by
T (18)
e
where is the sampling time. This result does not depend on the type of hold because
is specified in terms of its continuous time history, over a sample interval (Franklin
et al. 1998). A zero-order hold (ZOH) with no delay is given by
, (19)
Performing substitution
(20)
In Eq. (18) yields
T (21)
e
Defining,
T
, ,
e
, ,
(22)
(23)
The first state variable equation can be written as
1 , , , (24)
Parameterizing Eq. (24) will give
1 (25)
where , , , and , .
The objective is to find M to modify the , as shown in Fig. 3 such that 1
.
M (26)
Solving for M gives (Oppenheimer and Doman 2004)
1
M (27)

These parameters , and are tuned using genetic algorithm optimisation. Here it is
assumed that the positions and rate of change of actuators are available. If there is a bank of
second order actuator dynamics then M is chosen to be a diagonal matrix of dimension
(11X11).
In the following a stochastic evolutionary algorithm technique was discussed and applied to
tune the parameters for the compensator design in section 3.
3. Tuning of compensator to mitigate interaction using GAs

The idea is to combine the design objective in the form of a cost function that is to be
optimised using an optimizer such as a Genetic Algorithm. Where the cost function includes
the time domain objective; the tracking error is transformed into the integrated square of
error between the commanded signal and actual output u of the actuators. In addition there
is another design objective, exception handling (e.g. division by zero) and this is also
included in the cost function. The schematic is shown in Fig. 10.
Fig. 10. Cost function error generated by the simulation

Numerically the cost function for tracking error is given by
(28)
where the is the simulation run time.

Numerically the cost function for exception handling is given by
penalty exception generated

(29)
0 no exception
In Eq. (29) the penalty is assigned as a large number like 1010 so that the individual
generating this exception would most likely not be selected in the next generation because of
having very low fitness value. Numerically the combined cost function is given as
(30)
This cost function is then minimised to tune the parameters for the compensator. In the next
section GA based optimisation details are given.
3.1 GA based optimisation

Genetic Algorithms are a part of Evolutionary Computing which is a rapidly growing area
of Artificial Intelligence. GA take up the process of evaluating the relative fitness of the
individuals of a large population called genes, to select for a new generation, and mimic mutations
and crossing over (mixing genes from two parent genes to form offspring genes), the so called
evolution phenomenon (Lindenberg 2002). Unlike biological evolution, in GA the gene
controls some other processes like compensator parameters in this work, and is evaluated
by comparing properties of the process instead of simply computing some function on the
gene space (Lindenberg 2002). Genes used in GA are encoded as bit strings, and their fitness
(Lindenberg 2002) is a relation described by a real valued fitness function f on the set of bit
strings ,, such that gene is fitter than gene if .
Optimisation using GA begins with a set of solutions (represented by chromosomes) called
the population. Solutions from one population (based on some selection criteria) are taken
and used to form a new population. The flow chart of illustration of GA is shown in Fig. 11.
After selection of an encoding method (binary encoding for this case) and fitness function
(cost function value for an underlying gene), the algorithm proceeds in the following steps
(Lindenberg 2002).
Fig. 11. Flow chart explanation of Genetic Algorithms

Select an initial population indexed by P at random from some subset of
Repeat the following
evaluate the fitness of all the and use them to assign selection probabilities to
select a new population, dropping genes of low fitness and duplicating fit ones,
keeping index set (the size of population)
apply the genetic operators of mutation and crossing over (after forming couples)
This is repeated until some condition (for instance maximum number of generations or
improvement of the best solution) is satisfied. The main advantage (Leigh 2004) of GA over
other optimizers is their parallelism, GA is travelling in a search space using more
individuals so they are less likely to get stuck in a local minima. The most important
attributes of GA are mutation and cross over. A good cross over rate is expected to take
better parts of parent genes to the next generation. Mutation on the other hand changes the
individuals and if it is kept to a safe low level it helps the population to avoid falling in local
minima. This makes GA different from other optimisers, and particularly suitable for non-
convex optimisation problems like the compensator parameter optimisation in this research.
The main disadvantage linked with GA is the higher computation time and required
resources, but this can be avoided if there is a possibility to stop the GA anytime in the
routine. Also with the ever increasing processing power of computers over time this
constraint diminishes.
3.2 Optimizing routine using GA

Numerically the optimizing problem is given as Find by minimizing .
min (31)
where is a diagonal gain matrix of dimension (11X11). The GA optimising routine is

formulated by using the MATLAB Genetic Algorithm Direct Search Toolbox. A flow chart
representation of the optimisation routine is shown in Fig. 12.
Fig. 12. Flow chart for tuning compensator parameters using GA

The complete process shown in Fig. 11 can be summarised as:

The GA main function calls the evaluation function, giving searched parameters to
calculate compensator parameters
The evaluation function calculates compensator parameters and calls the simulation
model giving the parameters for the compensator
The simulation model runs the simulation for the given compensator parameter (i.e.
individual of population) and returns the value of error between and actual
The evaluation function calculates the cost function value for given errors and returns
to the main GA function
This is repeated for the total number of genes in one generation (population), and then
one generation completes, and so the remaining generations are iteratively completed
The above process is repeated until the cost function attains convergence, or the
maximum number of generations is reached.
In the next section the simulation results are given to show how the compensator mitigates
the interaction between the control allocation and actuator dynamics.
4. Simulation results
During simulation, a mixture of actuator dynamics was used. In the case of redundant
control surfaces diagonal gain matrices were tuned by the GA. The control surfaces were
approximated by the transfer functions as shown in Table 1.
Control surfaces Number of control surfaces Transfer functions

270
Ailerons 4
22.19 270
0.6128
Elevators 4
0.6128
0.0087
Stabilizer 1
0.0087
270
Rudders 2
22.19 270
Table 1. Aerosurfaces actuator dynamics (Esteban and Balas 2003)

The virtual control signal, , consists of chirps of amplitude 0.1, 0.15, 0.1 ( / ) in roll,
pitch and yaw angular accelerations respectively. The frequencies of chirps ranged from
0.1 1 in 20 seconds. In the processing of the GA routine exception handling is carried
out to avoid breaking the GA optimisation process. For example if there is an individual (i.e.
gains in diagonal matrix) in the population that gives division by zero that would break the
simulation. This is dealt with in an exception handling block, which will give a penalty to
that individual without breaking the simulation. In the next generation that individual
would not be selected.
Simulations are done with compensation (Fig. 13 and Fig. 14) and without compensator (Fig.
15 and Fig 16). As can be seen clearly from the results with no compensation there is serious
attenuation and mismatch, but as soon as the compensation is turned on, is
achieved because sufficient control authority exists.
Deviations in the case of no compensation case means that the desired control surface
positions coming out of the control allocator are different from the actual position of control
surfaces. This interaction between the control allocator and the actuator dynamics results in
serious consequence if the bandwidths of the actuators are not high or, in other words, the
actuators are slow.
Fig. 13. Implementation scheme for compensator when the compensator is switched on
Fig. 14. Desired angular accelerations ( ) and actual angular acceleration ( ) in rad/s2
when compensation is on
Fig. 15. Implementation scheme for compensator when the compensator is switched off
Fig. 16. Desired angular accelerations ( ) and actual angular acceleration (B ) in rad/s2
when compensation is off
5. Conclusions
This chapter details the application of genetic algorithms for the design and tuning of a
compensator to alleviate the effects of control allocation and actuator dynamics interaction.
The effects of non-negligible actuator dynamics have been investigated first. It was observed
that, for the Boeing 747-200, the actuator dynamics cannot be ignored if the excitations are in
the range of 0.1 to 1 Hz, which normally depends on the pilot dynamics. Another
observation suggests that the bandwidths of the actuators are smaller than the rigid body
modes of the aircraft and should not be neglected. The benefit of using a soft-computing
methodology for tuning the compensator gains is to avoid the optimisation converging to a
local minima and it is seen that the likelihood of the genetic algorithms converging to local
minima solution is less as compared to other techniques. In this methodology the model of
the actuator is not needed to be known because this methodology was designed to be used
on the actuator rig. In the case of the second order actuator, the rates should be either
measured or observed. GAs are used offline and the band limited chirps signal is used as the
excitation signal in the simulation. However, in the real system a band limited pseudo-
random binary signal (PRBS) for this type of identification process could be used as an
excitation signal rather than chirp because the later gives cyclic loading on the actuator,
which could be problematic.
6. References
Bolling, J.G., (1997) Implementation of Constrained Control Allocation Techniques Using an
Aerodynamic Model of an F-15 Aircraft, MSc. thesis, Virginia Polytechnic Institute
and State University, Virginia, USA.
Esteban, A. M., Balas, G.J. (2003) A B747-100/200 aircraft fault tolerant and fault diagnostic
benchmark, AEM-UoM 2003-1, Aerospace Engineering and Mechanics
Department, University of Minnesota, USA.
Franklin, G.F., Powell, J.D. and Workman, M. (1998) Digital control of dynamic systems, 3rd
ed., Addison-Wesley Longman, Inc., California
Lindenberg, F.M. (2002) Adaptive Compute Systems lecture notes, Technical University
Hamburg-Harburg Germany.
Oppenheimer, M.W. and Doman, D.B. (2004) 'Methods for compensating for control
allocator and actuator interactions', Journal of Guidance, Control, and Dynamics,
27(5), pp. 922-927.
5
Fault Tolerant Flight Control,

a Physical Model Approach
Thomas Lombaerts, Ping Chu,
Jan Albert (Bob) Mulder and Olaf Stroosma
Delft University of Technology
the Netherlands
1. Introduction
Safety is of paramount importance in all transportation systems, but especially in civil
aviation. Therefore, in civil aviation, a lot of developments focus on the improvement of safety
levels and reducing the risks that critical failures occur. When one analyses recent aircraft
accident statistics (Civil Aviation Safety Data 1993-2007 (2008); Smaili et al. (2006)), there are two
major categories of accidents which can be attributed to a single primary cause, as illustrated
in gure 1. The largest category is "collision with ground" (controlled ight into terrain, CFIT)
where a fully functional aircraft hits terrain due to the loss of situational awareness by the
pilot, which counts for as much as 26% of the accidents. This percentage is decreasing over the
years thanks to the continuously evolving amount and manner of cockpit display information.
The second major category is "loss of control in ight", which can be attributed to mistakes
made by the pilot or a technical malfunctioning. This category counts for 16% of all aircraft
accident cases and is not decreasing.
Fig. 1. Accident statistics, source: Civil Aviation Safety Data 1993-2007 (2008)
Analysing a major part of the accidents in the latter category has led to a common conclusion:
from a ight dynamics point of view, with the technology and computing power available
at this moment, it might have been possible to recover the aircraft in many accident
situations in this category, on the condition that non-conventional control strategies would
have been available. These non-conventional control strategies involve the so-called concept
of active fault tolerant ight control (FTFC), where the control system is capable to detect
the change in the aircraft behaviour and to adapt itself so that it can handle the perturbed
aircraft dynamics. Earlier research projects in FTFC involve the Self-Repairing Flight Control
System (SRFCS) program (Corvin et al. (1991)), the MD-11 Propulsion Controlled Aircraft
(PCA) (KrishnaKumar & Gundy-Burlet (n.d.)), the Self-Designing Controller for the F-16
VISTA (Ward & Barron (1995)), Recongurable Systems for Tailless Fighter Aircraft in the
X-36 RESTORE program (Brinker & Wise (1999); Calise et al. (2001)), the NASA Intelligent
Flight Control System (IFCS) F-15 program (Intelligent Flight Control: Advanced Concept
Program (1999)) and Damage Tolerant Flight Control Systems for Unmanned Aircraft by
Athena/Honeywell (Gavrilets (2008)). There are many alternative control approaches to
achieve FTFC. In all these control approaches, there remain some problems and limitations,
varying from the limitation to a restricted number of failure cases to the limitation of the
type and extent of damage which can be compensated for due to xed model structures for
identication. Another frequently encountered issue are convergence problems. Besides, black
box structures like for neural networks reduce the transparency of the approach. Moreover,
for many approaches it is not clear what will happen when the reference model behaviour is
not achievable in post-failure conditions.
The research approach as elaborated in this chapter uses a physical modular approach, where
focus is placed on the use of mathematical representations based on ight dynamics. All
quantities and variables which appear in the model have a physical meaning and thus are
interpretable in this approach, and one avoids so-called black and grey box models where
the content has no clear physical meaning. Besides the fact that this is a more transparent
approach, allowing the designers and engineers to interpret data in each step, it is assumed
that these physical models will facilitate certication for eventual future real life applications,
since monitoring of data is more meaningful.
Adaptive nonlinear dynamic inversion has been selected as the preferred adaptive control
method in this modular or indirect approach. The advantages of dynamic inversion are
the absence of any need for gain scheduling, and an effective input-output decoupling
of all control channels. Adaptation of the controller is achieved by providing up-to-date
aerodynamic model information which is collected in a separate identication module.
The structure of this chapter is as follows. Section 2 provides information on the high delity
RECOVER simulation model which has been used in this research project. A global overview
of the fault tolerant control architecture is given in section 3, and further explanations of some
of the individual modules are added in sections 4 and 5. Simulation results are discussed
in section 4.2 for the aerodynamic model identication, section 5.3 for the autopilot and in
section 5.5 for the manual control approach. Finally, section 6 presents some conclusions and
recommendations for future research.
2. The RECOVER benchmark simulation model

The presented work is part of a research project by the Group for Aeronautical Research and
Technology in Europe (GARTEUR). This group has established ight mechanics action group
FM-AG(16) with the specic goal to investigate the possibilities of fault tolerant control in
aeronautics and to compare the results of different reconguring control strategies applied to
a reference benchmark ight trajectory. That benchmark scenario is inspired by the so-called
Fault Tolerant Flight Control, a Physical Model Approach 95
Bijlmermeer disaster of EL AL ight 1862, where a Boeing 747-200 Cargo aircraft of Israels
national airline EL AL lost two engines immediately after take-off from Amsterdam airport
Schiphol in the Netherlands and crashed into an apartment building in the neighbourhood
while trying to return to the airport. A detailed simulation model of this damaged aircraft
is available from the Dutch Aerospace Laboratory NLR. This RECOVER (REcongurable
COntrol for Vehicle Emergency Relief) benchmark model is discussed in detail in ref. Smaili
et al. (2008; 2006) and has been used (also in earlier versions) by a number of investigators
and organizations (Maciejowski & Jones (2003); Marcos & Balas (2003); Szaszi et al. (2002)).
More information about the reference benchmark scenario can be found in ref. Lombaerts
et al. (2005; 2006). Other control strategies and results applied to the same benchmark model
as part of the framework of FM-AG(16) can be found in ref. Alwi (2008); Cieslak et al. (2008);
Hallouzi & Verhaegen (2008); Joosten et al. (2007; 2008). Related FDI work can be found in ref.
Varga (2007); Varga & Hecker (2004).
The simulation benchmark for evaluating fault tolerant ight controllers as discussed in ref.
Smaili et al. (2006) contains six benchmark fault scenarios, enumerated in g. 2(a). These
failure cases have varying criticality. Fig. 2(b) shows the failure modes and structural damage
conguration of the Flight 1862 accident aircraft, which is the most important fault scenario
in the simulation benchmark.
(a) GARTEUR FM-AG(16) RECOVER (b) Failure modes and structural damage conguration
benchmark fault scenarios, source: of the Flight 1862 accident aircraft, suffering right wing
Smaili et al. (2008) engine separation, partial loss of hydraulics and change
in aerodynamics, source: Smaili et al. (2008)
Fig. 2. GARTEUR FM-AG(16) RECOVER benchmark fault scenarios and conguration
The rudder runaway, the vertical tail separation and the EL AL engine separation have been
used as scenarios for this chapter. In the case of a rudder runaway (also called hardover),
the rudder moves quickly to an extreme position. More precisely, the rudder deects to the
left, inducing a yawing tendency of the aircraft to the left. The rudder deection limit in this
scenario depends on the ight speed, since aerodynamic blowdown is taken into account in
the RECOVER simulation model. As a result the maximum rudder deection is slightly below
15 for an airspeed around 270 knots, and even close to 25 (the physical maximum deection
limit imposed by the rudder hardware structure) for an airspeed of 165 knots. The vertical tail
separation leads to the loss of all rudder control surfaces as well as the loss of all damping in
the roll and yaw axes. Mind that loss of hydraulics is not considered in this situation. The El
Al engine separation scenario is an accurate simulation of ight 1862, validated by black box
data of the accident, where the loss of hydraulics is taken into account.
3. Global overview of the physical modular approach

Globally, the overall architecture of this modular approach consists of three major assemblies,
namely the controlled system, the Fault Detection and Identication (FDI) assembly and the
Fault Tolerant Flight Control (FTFC) assembly, as shown in g. 3. The controlled system
comprises the aircraft model and the actuator hardware. Possible failures in this controlled
system are structural failures and actuator hardware failures in the latter. Sensor failures have
not been considered in this research, since it has been assumed that effects of these failures
can be minor thanks to sensor redundancy and sensor loss detection. However, the latter
mechanism is recommended for future research.
Fig. 3. Overview of the modular physical approach for fault tolerant ight control
The Fault Detection and Identication (FDI) architecture consists of several components. The
core of this assembly is the two step method (TSM) module, described in section 4.1. This
module consists of a separate aircraft state estimation step followed by an aerodynamic model
identication step, where the latter is a joint structure selection and parameter estimation
(SSPE) procedure. The state estimation step is a nonlinear problem solved by an Iterated
Extended Kalman Filter. The preferred SSPE algorithm is Adaptive Recursive Orthogonal
Least Squares. In case a structural failure occurs (in the aircraft structure or in one of the
control surfaces), re-identication is triggered when the average square innovation exceeds
a predened threshold. For successful identication of the control derivatives of every
individual control surface, control effectiveness evaluation is needed after failure. This can
be done by inserting multivariate orthogonal input signals in the actuators. Although this
must be done carefully such that the damaged aircraft cannot be destabilized, it is necessary
in order to obtain sufcient control surface efciency information for the control allocation
module, to be discussed later. A valid approach might be to introduce these evaluation signals
only when strictly needed, i.e. when successful reconguration is not possible due to a lack of
information about this control efciency. The two step method is ideally suited to deal with
structural failures, but for the detection of actuator failures a separate actuator monitoring
algorithm is needed, such as an Actuator Health Monitoring System (AHMS).
Four other functions can be grouped to form the Fault Tolerant Flight Control (FTFC)
assembly. The core for this group are indirect adaptive control and control allocation. Indirect
adaptive control can be achieved by adaptive nonlinear dynamic inversion (ANDI) described
in section 5.1. In this setup, the control structure consists of three inversion loops, as
elaborated in section 5.2. This control structure relies on the principle of time scale separation.
The advantages of ANDI are that it removes the need for gain scheduling over different
operating conditions and it effectively decouples input-output relationships. Moreover, the
NDI control algorithm automatically involves some form of control allocation, due to the
structure of the control law. This structuring allows a clear separation how different failure
types are dealt with. Structural failures independent of the control surfaces are detected by
the TSM and this damage information is supplied to the ANDI algorithm by means of the
aerodynamic derivatives. On the other hand control surface related failures, aerodynamics or
actuator related, are identied by the TSM or the actuator monitoring algorithm respectively.
This information is sent to the adaptive control allocation block. The preferred control
allocation approach is found to be the control distributor concept (CDC), (Oppenheimer &
Doman (2006)), combined with the weighted pseudo inverse (WPI). This method assumes
the presence of a large amount of similar control surfaces. This assumption holds for the
type of aircraft considered in this research project, such as the Boeing 747. This approach
ts in the modular setup of the global procedure, where the CDC principle takes into
account aerodynamic changes and the WPI provides the adaptivity with reference to actuator
failures. Furthermore, a reference model denes the reference signal that the closed loop
conguration has to track. However, this reference model needs to be adaptive such that
its signals are limited based upon the achievable performance of the damaged closed loop
conguration. This reference signal adaptation can be achieved by Pseudo Control Hedging.
This modulation is based upon the difference between the demanded input signal and the
achieved input signal by the actuators. This reference signal adaptation is primarily driven
by saturation effects. Besides, this hedging operation takes into account the updated model
information from the FDI assembly. In this way, one makes sure that no unreachable reference
signals are given to the closed loop conguration. This PCH operation can be considered as
a rst degree of safe ight envelope enforcement, based upon input saturation effects and
updated model information. Experiments have shown that especially the throttle channel is
prone to saturation effects.
During this research, it has been found that safe ight envelope enforcement is a crucial aspect
in this control setup, and it is part of recommended future research. This protection algorithm,
based upon the achievable performance which can be estimated based upon actuator status,
aerodynamic derivatives and control derivatives, should contribute on two levels. On one
hand, it has to assist the PCH algorithm by limiting the reference model output appropriately.
Moreover, the output of the control allocation block, and thus the input to the actuators,
should be limited based upon this reachable ight envelope information. A tabular overview
of the preferred method for every component in the global overview of the fault tolerant ight
control setup can be found in table 1.
In the following sections, further explanations will be given about some of the individual
modules.
4. Fault detection and identication

4.1 Two step method
The identication method considered in this study is the so-called two step method (Chu
(2007); Chu et al. (1994); Laban (1994); Mulder (1986); Mulder et al. (1999)). There are
many other identication algorithms mentioned in the literature like maximum likelihood
component preferred method

aircraft state estimation Iterated Extended Kalman Filter
aerodynamic model Adaptive Recursive
identication Orthogonal Least Squares
identication triggering average square innovation & threshold
control effectiveness evaluation multivariate orthogonal input functions
control allocation Control Distributor Concept
& Weighted Pseudo Inverse
indirect adaptive control triple layer Adaptive
Nonlinear Dynamic Inversion
actuator monitoring Actuator Health Monitoring System
reference signal adaptation Pseudo Control Hedging
Table 1. Preferred methods for FTFC modules
identication (MLI) and other one step identication routines, but not all of them are
applicable on line. One of the few procedures which can be implemented in real time is the
so-called ltering method developed at German Aerospace Centre DLR, see ref. Jategaonkar
(2006). This is a joint state and parameter estimation algorithm, but very complex. Another
algorithm can be found in ref. Morelli (2000), which is frequency based. The advantage of
the two step method is that it is easier to implement on-line. The aim is to update an a
priori aerodynamic model (obtained by means of windtunnel tests and CFD calculations)
by means of on-line ight data. The rst step is called the Aircraft State Estimation phase,
where the second one is the Aerodynamic Model Identication step. In the Aircraft State
Estimation procedure, an Iterated Extended Kalman Filter is used to determine the aircraft
states, by making use of the redundant but perturbed information from all sensors (air data,
inertial, magnetic and GPS measurements). By means of this state information, it is possible
to construct the combined aerodynamic and thrust forces and moments acting on the aircraft.
Mass and inertia are considered as known constants in these calculations. In the absence of
a structural failure, real time mass and inertia can be calculated by integrating fuel ow and
subtracting it from the total take off values. Future research is aimed at taking into account
changing masses and inertia in the presence of structural failures. Finally the aerodynamic
derivatives can be deduced in the second step by means of structure selection and parameter
estimation.
This second step involves the aerodynamic properties of the aircraft. Thanks to the separation
of the kinematics, these aerodynamics become linear in the parameters. Model structure
determination is the rst concern, and a regression method can be applied in order to estimate
the so-called aerodynamic model parameters. Simultaneous recursive structure selection and
parameter estimation is performed in this phase, by means of so-called Adaptive Recursive
Orthogonal Least Squares (AROLS), (Lombaerts et al. (2010c)). This method determines the
model structure by means of an orthogonal QR-decomposition, which is extended recursively
to make the method suitable for online applications. Special monitoring criteria verify
continuously when the structure needs to be extended. For a detailed discussion of these
criteria see ref. Lombaerts et al. (2010c).
The model structure selection procedure considers a set of candidate regressors based upon a
Taylor series expansion with respect to the aircraft states and control inputs (control surface
deections and engine settings) relevant for the aerodynamic forces and moments. The
candidate regressors are shown in table 2.
dependent candidate regressors

variable
q c
1, , 2 (X only), V , e i , ih , spi , f i , Ti
pb rb , m , qc , , , 2 , 2 , 3 , 2 3 m = 3, . . . , 8}
X; Z; M { , 2V , 2V V e
pb rb , 2 pb , 2 , rb , n n = 2, . . . , 5}
{ 2V , 2V 2V 2V
pb rb q c
1, , 2V , 2V , ai , r i , spi , {, V , Ti }
qc
Y; L; N { , V , e , , , , , m = 2, . . . , 8}
m 2 2 3 2 3
pb pb
{ 2V , 2V
rb
, 2 2V , 2 , 2V rb
, n n = 2, . . . , 5}
Table 2. States and control inputs which inuence forces and moments
p r a r
CY < 0 CYp 0 CYr > 0 CYa 0 CYr >0 nominal
Y CY 0 CYp 0 CYr > 0 CYa 0 CYr >0 engine separation
CY 0 CYp 0 CYr 0 CYa 0 CYr =0 tail loss
Cl < 0 Cl p < 0 Cl r > 0 Cl a <0 Cl r >0 nominal
L Cl < 0 Cl p 0 Cl r 0 Cl a <0 Cl r >0 engine separation
Cl 0 Cl p < 0 Cl r > 0 Cl a <0 Cl r =0 tail loss
Cn > 0 Cn p < 0 Cn r < 0 Cn a >0 Cn r <0 nominal
N Cn 0 Cn p < 0 Cn r < 0 Cn a >0 Cn r <0 engine separation
Cn 0 Cn p < 0 Cn r 0 Cn a >0 Cn r =0 tail loss
Table 3. Predicted lateral aerodynamic derivative changes for asymmetric failure scenarios
studied
In table 2, {. . .} indicate regressor candidates for AROLS related to aerodynamic cross

coupling effects in case of aircraft asymmetries as well as higher order aerodynamic effects,
for example due to structural failures. The identied parameters contain valuable information
about the physical state of the aircraft. For example, the parameters Cm and Cn represent
measures for static stability. In fact, Cn is also known as Weathercock stability. The derivatives
Cme , Cla and Cnr represent elevator-, aileron- and rudder-effectiveness respectively.
A structural failure will result in changing aerodynamic derivatives. Based upon ight
dynamics theory (Mulder et al. (January 2006)), it is possible to predict the changes in the
aerodynamic derivatives in the engine separation and tail loss scenarios, as illustrated in table
3. Each parameter is an aerodynamic derivative, e.g. CY represents the change in lateral force
Y caused by a change in sideslip angle . In the engine separation scenario for example, wing
sweep combined with the loss of the leading edge results in a larger negative Y-force caused
by a positive sideslip angle .
Various validation tests by means of a posteriori batch process identication, least squares
residual analysis and reconstruction of velocity and angular rate components have shown
that this method works well and can in fact be used for identication purposes of damaged
aircraft. Identication of component as well as structural failures is possible in this setting, see
ref Lombaerts et al. (2007) for a more elaborate discussion of validation methods and results.
4.2 Applications
Two application examples are considered here. First, a static directional stability analysis
is performed for the tail loss scenario. Thereafter, the model structure of the dimensionless
aerodynamic force coefcient CZ is elaborated and its parameter values are estimated at the
same moment in the engine separation scenario.
4.2.1 Tail loss

A natural consequence of the tail loss scenario is the huge reduction in directional static
stability. This can be seen in the behaviour of the aerodynamic derivative Cn , as shown in g.
4. A positive value for Cn indicates static directional stability. From g. 4, it is clear that, while
the nominal aircraft is stable, the damaged aircraft is observed to be directionally statically
unstable.
0.16
nominal
loss of vertical tail
0.14
0.12
0.1
Cn
0.08
0.06
0.04
0.02
0.02
0 10 20 30 40 50 60
time, [s]
Fig. 4. Comparison of Weathercock stability derivative for undamaged and damaged aircraft
with loss of vertical tail
4.2.2 Engine separation

In this example, the dimensionless aerodynamic force coefcient CZ is considered in the
engine separation scenario. Applying the AROLS routine to this stretch of simulation data
leads to the results as shown in g. 5. Fig. 5(a) shows the results prior to failure, while g.
5(b) focuses on the post-failure situation. As can be seen in this gure, the t is accurate in
both cases. The optimal residual, which is updated over the entire time history at every time
step, is minimal. The real residual, which is not updated a posteriori, gives an indication of the
reduction in the residual over the time interval. Fig. 5(a) reveals the initial large residual due to
initialization. Fig. 5(b) shows the initial large residual caused by the failure. This is minimized
in very short term, but starts increasing thereafter due to slow dynamics which become
dominant. Between 55s and 60s, the residual is again minimized, thanks to the incorporation
of these slow dynamics in the aerodynamic model.
The parameter estimation results, together with their standard deviations, can be found in
table 4. These results highlight the difference between the situations before and after failure.
Because there is no anomaly before the failure, the important independent variables for the
vertical force are the conventional ones, namely a constant, the angle of attack and pitch
qc
rate V . However, after the failure a violent roll-dive manoeuvre follows. The inuence of the
damage on the change in aerodynamics is represented by additional contributions by sideslip
pb
and roll rate 2V . These are a consequence of the asymmetric damage, as a result of which
0.2 0.2
measured measured
0.3 reconstructed 0.1 reconstructed
est
est
CZ & CZ
CZ & CZ
0.4 0
0.5 0.1
0.6 0.2
0 5 10 15 20 25 30 35 40 45 50 50 55 60 65 70 75
0.04 0.02
optimal residual
optimal residual
0.02 0
0 0.02
0.02 0.04
0 5 10 15 20 25 30 35 40 45 50 50 55 60 65 70 75
0.2 0.6
actual residual
actual residual
0.1 0.4
0 0.2
0.1 0
0.2 0.2
0 5 10 15 20 25 30 35 40 45 50 50 55 60 65 70 75
time [s] time [s]
(a) AROLS structure selection and (b) AROLS structure selection and
parameter estimation result for CZ prior parameter estimation result for CZ after
to failure failure
Fig. 5. AROLS structure selection and parameter estimation results for CZ before and after
the failure
the decoupling of longitudinal and lateral regressors does not hold anymore in this scenario.
The standard deviations give an indication of the accuracy of the results.
prior failure post failure

estimated () estimated ()
1 0.1583 (1%) 0.4131 (2%)
4.4016 ( 0%) 4.9047 (1%)
qc
V 1.6124 (22%) 105.1966 (1%)
4.2319 (1%)
pb
2V 9.4708 ( 0%)
rb
2V
all others
Table 4. SSPE results for CZ for engine separation scenario, before and after failure
A more elaborate discussion of this example can be found in ref. Lombaerts et al. (2010c).
5. Fault Tolerant Control method

Fault Tolerant Flight Control is achieved by making use of a triple layered adaptive
nonlinear dynamic inversion algorithm. First, the basics of the control approach are explained.
Thereafter, each of the three individual control loops is described. Finally, some application
examples on a selection of failure scenarios will show the failure handing capabilities of this
fault tolerant control setup.
5.1 Control approach: Adaptive nonlinear dynamic inversion

A major advantage of dynamic inversion is its ability to naturally handle changes of operating
condition, which removes the need for gain scheduling, e.g. as is the case for classical control
methods. This is especially advantageous for control of space re-entry vehicles, due to their
extreme and wide operating conditions which include hypersonic speed during re-entry and
subsonic regions during the terminal glide approach phase to the runway. Another advantage
is its natural property of decoupling the control axes, i.e. no coupling effects remain between
steering channels and the different degrees of freedom. NDI control has been implemented in
the Lockheed F-35 Lightning II, (Balas (2003); Walker & Allen (2002)).
Nonlinear dynamic inversion considers original nonlinear systems of the afne form:
x = a (x) + b (x) u (1)
and provides a solution for the physical control input u by introducing an outerloop virtual
control input :
u = b 1 (x) [ a (x)] (2)
which results in a closed-loop system with a decoupled linear input-output relation:
x = (3)
A linear outer loop control law is sufcient to enforce exponentially stable tracking dynamics,
where the control gains can be determined by the required closed loop characteristics.
Dynamic inversion is a popular control method for ight control and aircraft guidance, (Balas
et al. (1992); Campa et al. (2005); da Costa et al. (2003); Ramakrishna et al. (2001); Reiner et al.
(1996); van Soest et al. (2006)) as well as reconguring control, (Ganguli et al. (2005; 2006);
Oppenheimer & Doman (2006); Ostroff & Bacon (2002)).
The main assumption in NDI is that the plant dynamics are assumed to be perfectly known
and therefore can be cancelled exactly. However, in practice this assumption is not realistic,
not only with respect to system uncertainties but especially to unanticipated failures for the
purpose of fault tolerant ight control. In order to deal with this issue, one can make use of
robust control methods as outer loop control to minimize or suppress undesired behaviour
due to plant uncertainties which cause imperfect plant dynamic cancellation. Other control
methods such as neural networks also have been proposed in the literature, in order to
augment the control signal as a compensation for the non-inverted dynamics, as explained
previously. However, another solution to the weakness of classical NDI, namely its sensitivity
to modelling errors, is the use of a real time identication algorithm, which supplies updated
model information to the dynamic inversion controller. These augmented structures are called
adaptive nonlinear dynamic inversion (ANDI). The latter procedure is the method of choice
for this research which led to the results presented in this paper.
5.2 Autopilot control loops

Three consecutive inversion loops have been implemented, namely a body angular rate loop,
an aerodynamic angle loop and a navigation loop. The developed control laws are described
below.
5.2.1 Body angular rate loop

Consider the dynamic equation of an aircraft:
= I1 M a I1 I (4)
T T
where = p q r are the rotational rates and M a = L a M a Na the aerodynamic
moments acting on the aircraft. The inertia matrix I stands for:

Ixx Ixy Ixz
I = Iyx Iyy Iyz (5)
Izx Izy Izz
where the moments of inertia Ixy , Iyx , Iyz and Izy are assumed to be zero.
Fig. 6. NDI rate control inner loop

Rewriting the aircraft dynamics into the form of eq. (2):
1
a b Cla 0 b Clr
e 0 1 2 1
= cCme 0 V S
r 2
b Cna 0 b Cnr

p p p bClstates
I q + q I q cCmstates (6)

r r r bCnstates
T
where the virtual inputs p q r are the time derivatives of the rotational rates of the
aircraft. All control and state derivatives in this expression are estimated via the real time
identication procedure.
5.2.2 Aerodynamic angle loop

The middle loop quantities are roll angle , angle of attack and sideslip angle :
1
1 sin tan cos tan
p v b u2b 2
q 0
= V 2 w2b V wb
r w2b 2
2b 2
u
0
V vb V vb

0

1 ( A + g cos cos )
2 w2 z (7)
V b

1 2 Ay + g cos sin

V 2 vb
5.2.3 Navigation loop

The procedure used in this step is inspired by the method used in ref. Holzapfel (2004),
although the application for this study implies some important deviations compared to the
conventional method, since an adaptive model needs to be taken into account. Main crux in
this deviating approach is that this inversion loop consists of two separate steps. First, the
kinematics based virtual inputs are transformed towards the roll angle and the symmetric
aerodynamic forces through a physically interpretable nonlinear mapping. Consequently,
the aforementioned force components are translated into commanded angle of attack and
dimensionless thrust values via a classical NDI-setup as used before, which involves a local
gradient determination step.
The kinematics based inversion is as follows:

FA X = m V + g sin (8)

V 2 FAY 2
FA Z = cos m2 g+ + (V )2 (9)
cos cos

cos
= arctan (10)
+ g cos
V
where FA Z corresponds to | Frequired | and is in fact required .

The subsequent aerodynamic forces based inversion is as follows:

FA X FA X F FA X
F FAY
Tc
A Xcomm
0

TC
= AY Tc
FAY FAY0 (11)
FA Z FA Z FA Zcomm FA Z
Tc 0
where the symbol denotes the left inverse.

More information about how these control laws have been derived in detail can be found in
ref. Lombaerts (2010); Lombaerts et al. (2010b). Summarizing, the global setup of the third
dynamic inversion loop can be found in gure 7.
Fig. 7. Third NDI autopilot loop, featuring VTAS , and control. LGD stands for local
gradient determination. TSM represents the two step method model identication block
elaborated in section 4.
Linear controllers act on each separate NDI loop, as indicated by "LC" in g. 6 and 7. These
linear controllers involve proportional and proportional-integral control, and gains have been
selected to ensure favourable ying qualities by means of damping ratio and natural
frequency n while complying with the time scale separation principle. Optimization of
these gain values has been achieved by means of multi-objective parameter synthesis (MOPS)
optimization, see ref. Lombaerts (2010); Looye (2007); Looye & Joos (2006).
5.3 Evaluations of autopilot experiments on RECOVER

Implementation of these control laws results in the controller performance illustrated in the
following simulation results. Besides the unfailed scenario the following failures have been
investigated: vertical tail loss and engine separation. In each scenario, the failure occurs
exactly at t = 50s.
5.3.1 Nominal unfailed

A complex manoeuvre is performed to evaluate the performance of the fault tolerant
controller. Three joint commands are given, more precisely a course change of = 160
at t = 70s, an altitude change towards h = 5000m at t = 80s and a velocity change of
V = 50m/s around t = 110s. This scenario allows to evaluate the performance as well as
eventual inadvertent couplings between the different channels. The tracking of the reference
signals is satisfactory and couplings are minimal to even non-existent, as can be seen in g.
8. In the control surface deection time histories in gures 9 and 10(a), it can be seen that the
elevators compensate rst for the vertical lift component loss induced by the turn, and then
initiate the altitude change. Ailerons and rudders cooperate in order to achieve a coordinated
turn. It can be seen that the spoilers assist the ailerons to set the turn up. The specic forces in
g. 10(b) indicate an acceptable behaviour. The responses of the unfailed aircraft will serve as
a comparison basis to evaluate the performance of the controller in failed situations.
tracking quantities
200 States
0.1 1
pbody
[]
0 0 0
phi
0.1 1
0 200 400 600 0 200 400 600
200
0 100 200 300 400 500 0.1 0.2
qbody
theta
10 0 0.1
0.1 0
0 200 400 600 0 200 400 600
[]
0
0.05 5
rbody
0 0
10 0.05 psi 5
0 100 200 300 400 500 0 200 400 600 0 200 400 600
200
Vtas [m/s]
200 10000
VTAS
150 5000
150 100 0
he
0 200 400 600 0 200 400 600

4
100 x 10
0 100 200 300 400 500 0.2 5
alpha
0.1 0
10000 0 5
xe
0 200 400 600 0 200 400 600

h [m]
3 4
5000 x 10 x 10
5 4
0 2
beta
5 0
ye
0 0 200 400 600 0 200 400 600

0 100 200 300 400 500
time [s]
(a) tracking quantities (b) states
Fig. 8. Tracking quantities and states for the unfailed scenario
15 5
inner elevator right inner aileron right
10 inner elevator left inner aileron left
outer elevator right outer aileron right
5 outer elevator left 0 outer aileron left
5 5
0 100 200 300 400 500 0 100 200 300 400 500
0.5 1
stabilizer angle outer flaps
0 upper rudder inner flaps
0.5
lower rudder
0.5
0
1
0.5
1.5
2 1
0 100 200 300 400 500 0 100 200 300 400 500
(a) deections of elevators, stabilizer and (b) deections of ailerons and aps
rudders
Fig. 9. Deections of elevators, stabilizer, rudders, ailerons and aps for the unfailed scenario
Specific forces in body axes

2 4
spoiler #1
Axb [m/s2]
1.5 spoiler #2 2
spoiler #3
spoiler #4 0
1
spoiler #5
2
spoiler #6 0 100 200 300 400 500
0.5
0.1
0
Ayb [m/s2]
0 100 200 300 400 500
0
5
spoiler #7 0.1
4 spoiler #8 0 100 200 300 400 500
spoiler #9 5
3
spoiler #10
Azb [m/s2]
2 spoiler #11
spoiler #12 10
1
0 15
0 100 200 300 400 500 0 100 200 300 400 500
(a) deections of spoilers (b) specic forces
Fig. 10. Deections of spoilers and specic forces for the unfailed scenario
5.3.2 Vertical tail loss

For this failure scenario, the reference signals have been slightly adapted. No speed change
is requested here, because this makes it more demanding for the control surfaces. When the
speed increase is maintained, the increasing aerodynamic damping reduces the disturbing
effect of the failure. The altitude capture has been changed accordingly to 1500m, in order to
prevent throttle saturation. Aerodynamic damping is the very reason why, in practice, control
laws actually work with calibrated airspeed CAS, in this way the control actions are related to
the dynamic pressure and control surface efciency is maintained.
Figure 12(b) reveals the compensating behaviour of the ailerons which takes into account the
loss of the vertical tail and the cessation of rudder functioning after failure, as can be seen
in g. 12(a). It can be observed that the ailerons effectively take over the function of yaw
damper, assisted by the spoilers in g. 13(a), since the rudders are lost. Figure 13(b) shows
that the oscillation in the lateral specic force Ay is only damped in the longer term, which
can be explained by the fact that ailerons and spoilers have no direct inuence on the yawing
moment, but on the rolling moment.
tracking quantities
200 States
0.1 0.5
pbody
[]
0 0 0
phi
0.1 0.5
0 200 400 600 0 200 400 600
200
0 100 200 300 400 500 0.1 0.4
qbody
theta
10 0 0.2
0.1 0
0 200 400 600 0 200 400 600
[]
0
0.05 5
rbody
0 0
psi
10 0.05 5
0 100 200 300 400 500 0 200 400 600 0 200 400 600
140
Vtas [m/s]
140 2000
VTAS
130 1000
130 120 0
he
0 200 400 600 0 200 400 600

4
120 x 10
0 100 200 300 400 500 0.2 5
alpha
0.1 0
2000 0 5
xe
0 200 400 600 0 200 400 600

h [m]
4
1000 x 10
0.2 4
0 2
beta
0.2 0
ye
0 0 200 400 600 0 200 400 600

0 100 200 300 400 500
time [s]
Fig. 11. Tracking quantities and states for the tail loss scenario
5.3.3 Engine separation

The engine separation scenario is a very sensitive situation to combine commands in
heading, altitude and speed simultaneously. Crucial in this context is to avoid engine throttle
saturation. Therefore, in this experiment only a heading change has been considered, as shown
15 20
10 inner elevator left 10 inner aileron left
0 10
5 20
0 100 200 300 400 500 0 100 200 300 400 500
0.5 1
0 upper rudder inner flaps
0.5
lower rudder
0.5
0
1
0.5
1.5
2 1
0 100 200 300 400 500 0 100 200 300 400 500
rudders
Fig. 12. Deections of elevators, stabilizer, rudders, ailerons and aps for the tail loss scenario
15 4
spoiler #1
Axb [m/s2]
spoiler #2 2
10 spoiler #3
spoiler #4 0
spoiler #5
5 2
spoiler #6 0 100 200 300 400 500
1
0
Ayb [m/s2]
0 100 200 300 400 500 0
1
15
spoiler #7 2
spoiler #8 0 100 200 300 400 500
10 spoiler #9 5
spoiler #10
Azb [m/s2]
spoiler #11
5 spoiler #12 10
0 15
0 100 200 300 400 500 0 100 200 300 400 500
Fig. 13. Deections of spoilers and specic forces for the tail loss scenario
in g. 14(a). Moreover, a limited maximum roll angle has been imposed, due to the restricted
safe ight envelope as explained in section 3. It has been found that altitude and speed
changes are also feasible separately, but these are not discussed in this section.
The time histories of the states in g. 14(b) reveal that the aircraft in post failure conditions ies
with a small nonzero roll angle and sideslip angle, due to the asymmetric damage, despite a
zero commanded sideslip angle. The control surface deections in gures 15 and 16(a) conrm
the cessation of functioning of the control surfaces which are powered by the hydraulic circuits
connected to engines number 3 and 4, as illustrated in g. 2(b). The remaining operative
surfaces are successful in keeping the aircraft in equilibrium and under control, although
with restricted authority. The nonzero lateral specic force in g. 16(b) is a consequence of
the sideslipping ight.
Two additional interesting quantities to investigate are the throttle setting and the average
square innovation, which triggers the re-identication routine as explained in ref. Lombaerts
et al. (2009; 2010a). Figure 17(a) conrms that the throttle setting does not saturate, however
the remaining control margins in order to remain inside the safe ight envelope are severely
restricted. This is due to the asymmetric thrust which needs to be compensated by the
control surfaces. The spike at t = 50s is caused by the feedforward path in the controller,
which is needed to compensate for the instantaneous speed loss of the two dead engines.
Figure 17(b) depicts the values for the average square innovation for each force and moment
channel separately. At t = 50s, it can be seen that the threshold for X is exceeded, and a
tracking quantities
200 States
0.05 0.2
pbody
[]
100 0 0
phi
0.05 0.2
0 200 400 600 0 200 400 600
0
0 100 200 300 400 500 0.1 0.2
qbody
theta
5 0 0.1
0.1 0
0 200 400 600 0 200 400 600
[]
0
0.02 5
rbody
0 0
psi
5 0.02 5
0 100 200 300 400 500 0 200 400 600 0 200 400 600
140
Vtas [m/s]
140 620
VTAS
130 600
130 120 580
he
0 200 400 600 0 200 400 600
4
120 x 10
0 100 200 300 400 500 0.2 5
alpha
0.1 0
620 0 5
xe
0 200 400 600 0 200 400 600
h [m]
4
600 x 10
0.05 4
0 2
beta
0.05 0
ye
580 0 200 400 600 0 200 400 600
0 100 200 300 400 500
time [s]
Fig. 14. Tracking quantities and states for the engine separation scenario
15 20
10 inner elevator left 10 inner aileron left
0 10
5 20
0 100 200 300 400 500 0 100 200 300 400 500
10 1
upper rudder 0.5 inner flaps
5 lower rudder
0
0
0.5
5 1
0 100 200 300 400 500 0 100 200 300 400 500
rudders
Fig. 15. Deections of elevators, stabilizer, rudders, ailerons and aps for the engine
separation scenario

20 4
spoiler #1
Axb [m/s2]
15 spoiler #2 2
spoiler #3
spoiler #4 0
10
spoiler #5
2
spoiler #6 0 100 200 300 400 500
5
1
0
Ayb [m/s2]
0 100 200 300 400 500 0.5
0
0.4
spoiler #7 0.5
spoiler #8 0 100 200 300 400 500
0.3
spoiler #9 5
spoiler #10
Azb [m/s2]
0.2
spoiler #11
spoiler #12 10
0.1
0 15
0 100 200 300 400 500 0 100 200 300 400 500
Fig. 16. Deections of spoilers and specic forces for the engine separation scenario
re-identication procedure is triggered for CX . It has become necessary to include the sideslip
angle , which has become signicant due to the sideslipping ight, as an additional regressor
in the identication procedure. This leads to a successful new identication procedure which
is performed extremely quickly as can be seen in this gure. This result conrms the benecial
contribution from the identication routine in this fault tolerant ight control setup.
throttle behaviour average square innovation as trigger for reidentification
1.6 12
X

1.4 Z
10

m

average square innovation

1.2 Y
8 l

1 n
Tc []
6
0.8
4
0.6
2
0.4
0.2 0
0 100 200 300 400 500 0 100 200 300 400 500
time [s] time [s]
(a) throttle behaviour (b) average square innovation as trigger for

re-identication
Fig. 17. Spoilers and specic forces for the engine separation scenario
5.4 Manual control loops

A manual variant of this fault tolerant controller has been developed as well. This variant
consists of the body angular rate inner loop as described in section 5.2.1, augmented by the
sideslip coordination axis only from the aerodynamic angle middle loop as explained in
section 5.2.2. Throttle control is by the conventional autothrottle. As a result, the pilot steers
roll rate p by means of the control wheel, pitch rate q with the control columns, and nally
the pedals can be used for creating a nonzero sideslipping ight, although this is rarely used.
Since dynamic inversion is used in all control loops, these steering channels are effectively
decoupled.
5.5 Simulator evaluation of manual controller

This manual control setup has been applied in the SIMONA (SImulation, MOtion and
NAvigation) Research Simulator (SRS), see g.18(a). It is a pilot-in-the-loop ight simulator
developed, built and operated by Delft University of Technology. It provides researchers with
a exible powerful tool that can be adapted to various uses, see ref. Stroosma et al. (2003).
The simulators exible software architecture and high-delity cueing environment allows
the integration of a variety of aircraft simulation models, such as the aforementioned Boeing
747 benchmark simulation model from ref. Smaili et al. (2006). Its inputs and outputs were
standardized to t the SRS software environment and the SIMULINKTM simulation model
as well as NDI-controller were converted to C code using Real-Time Workshop. Finally the
models were integrated with the pilot controls, aircraft instruments (Figure 18(b)) and other
cueing devices of the SRS (i.e. outside visual and motion systems). On the ight deck of
the SRS the evaluation pilot was presented with ight instruments representative of a large
transport aircraft, a control column with large transport aircraft feel system dynamics, a
central pedestal with dual engine controls and a wide collimated view on a virtual outside
world. The simulators motion system was tuned to give the pilot realistic inertial motion
cues in nominal and failure conditions. The test pilots were four Boeing 747 captains (one
retired) and one other wide body captain on Airbus A330 and Boeing 767. All were familiar
with the research simulator practices used for this investigation.
(a) outside view (b) cockpit view
Fig. 18. The SIMONA (SImulation, MOtion and NAvigation) Research Simulator (SRS) at
Delft University of Technology, photo by Joost Ellerbroek
The adaptive NDI control system has been validated on two failure scenarios, namely the
engine separation failure and the rudder runaway scenarios. Fig. 19 shows the evaluation
trajectory during the piloted simulation runs in SIMONA. The trajectory consists of four
main phases, namely altitude capture, bank angle capture, localizer intercept and glideslope
intercept. For every phase, required and adequate performance specications have been
dened for the relevant longitudinal as well as lateral quantities. The scheme presented in
g. 20 assists the pilot in rating the handling qualities (Cooper & Harper (1969)) of the aircraft
while taking into account the performance of the aircraft with respect to the aforementioned
requirements. Fig. 21 shows the time histories of a selection of the most important aircraft
states. These conrm the evaluation trajectory as shown in g. 19. Moreover, altitude and roll
angle plots show altitude and roll angle captures which have been executed by the test pilot
in order to evaluate the post-failure handling qualities of the aircraft.
The handling qualities results for the algorithm show that, especially for the El Al Flight
1862 scenario, conventional ight control was restored to acceptable levels while physical
and mental workload were reduced signicantly. This is illustrated in Figure 22 where an
example is given of lateral handling quality pilot ratings for the localizer capture task.
It can be seen that, for this task, both the baseline and fault-tolerant y-by-wire (FBW)
aircraft were rated Level 1 (Rating 1-3). After separation of the right-wing engines (Figure
22), lateral handling qualities degraded to Level 2 for the conventional aircraft with the
classical control system. The recongured aircraft (FBW) shows about Level 1 handling
qualities after incurring signicant damage due to the loss of the right-wing engines. This
was substantiated by measured pilot control activities, representative of workload, indicating
no pilot compensation after reconguration. For the rudder runaway failure, however, Level
2 handling qualities remained after reconguration despite the fact that no sustained pilot
compensation was required. The difference was most probably caused by the fact that this
initial setup is a rate control and hold loop instead of a rate control attitude hold type. As a
consequence, angular rate disturbances are corrected for automatically by the controller but
subsequent disturbances from the equilibrium attitude had to be compensated for by the pilot
himself. The use of a rate control attitude hold setup will solve this issue.
Figure 23 illustrates the physical workload analysis results by depicting the average pilot
forces. In the graph, a distinction is made between roll, pitch and yaw channel, as illustrated
Fig. 19. Trajectory of the piloted simulation runs in SIMONA.
by the three graphs separated vertically. In each control channel, six cases have been studied,
namely unfailed, engine separation and rudder runaway, each time with classical and fault
tolerant control. In each case, the workload gure of each of the ve pilots is represented
individually by means of bar plots, after which the mean and standard deviations are
superimposed on these bar plots for every case, in order to facilitate mutual comparisons.
First of all, the unfailed conditions conrm that this is a good comparison basis between classic
and FTFC, since both have the same ratings. Comparing classic control with FTFC for failed
congurations shows that overall values for average manual control forces over all pilots
decrease for FTFC in the failure scenarios. In addition, in the failure scenarios the standard
deviations also reduce from classic control towards FTFC. At rst sight this seems not the
case for the pedal forces. Closer inspection of the experimental data, however, reveals that
this is caused by the deviating performance of pilot no 2 (probably due to misconception of
the control principle within the fault tolerant controller). Finally, searching for overlap of the
errorbars between classic and FTFC shows that this overlap does not occur. This observation
makes the trends signicant, despite the limited number of experiment subjects.
As a global conclusion, which is supported by the graphs above, it can be stated that this fault
tolerant ight controller improves the handling qualties and reduces physical pilot workload
considerably in failure conditions.
6. Conclusions and future work

Summarizing, it can be stated that, following numerous experiments, fault tolerant ight
control using a physical modular approach is successful in recovering damaged aircraft.
The designed methods are capable to accommodate the damage scenarios which have been
investigated in this project. It has been found that the engine separation scenario, based upon
Fig. 20. Cooper Harper Handling Qualities Rating Scale, source: Cooper & Harper (1969)
Selection of aircraft states rudder runaway scenario Selection of aircraft states rudder runaway scenario
0.4 1000
altitude [m]
pitch [rad]
0.2 500
0 0
0 200 400 600 800 1000 1200 1400 0 200 400 600 800 1000 1200 1400
angle of attack [rad]
0.2 5
heading [rad]
0.1 0
0 5
0 200 400 600 800 1000 1200 1400 0 200 400 600 800 1000 1200 1400
flight path angle [rad] angle of sideslip [rad]
true airspeed [m/s]
0.5 150
0 100
0.5 50
0 200 400 600 800 1000 1200 1400 0 200 400 600 800 1000 1200 1400
time [s] time [s]
0.2 1
roll angle [rad]
classic classic
0 FTFC 0 FTFC
0.2 1
0 200 400 600 800 1000 1200 1400 0 200 400 600 800 1000 1200 1400
time [s] time [s]
Fig. 21. Comparison of a selection of aircraft states for the rudder runaway scenario
(a) classical control (b) fault tolerant control
Fig. 22. Localizer capture task handling qualities ratings for classical control and fault
tolerant control
Average exerted pilot force during complete simulation run
6
roll force [Nm]
0 classic FTFC classic FTFC classic FTFC

no failure no failure engine engine rudder rudder
separation separation runaway runaway
40
pitch force [Nm]
30
20
10

300
pilot 1
yaw force [N]
pilot 2
200 pilot 3
pilot 4
pilot 5
100 mean

Fig. 23. Total average manual control forces during the simulation runs
El Al ight 1862, is survivable with adaptive control techniques. Experiments have also shown
that the two step method is successful for real time identication of damaged aircraft models,
including a real time static stability analysis. Autopilot control based upon adaptive nonlinear
dynamic inversion shows good failure handling capabilities.
An important aspect which has not been considered in this research is sensor loss detection.
Despite the presence of redundant sensors, recent aircraft accidents (Lombaerts (2010)) have
shown that sensor loss detection cannot be avoided and current monitoring techniques are not
always sufcient. More elaborate ight envelope protection algorithms, taking into account
a.o. minimum control airspeed limits, are another important topic for future research. Finally,
an important next step in the development of fault tolerant ight control technologies is to
validate them in real ight on board of manned as well as unmanned research aircraft. This is
one of the major challenges for the future.
7. References
Alwi, H. (2008). Fault Tolerant Sliding Mode Control Schemes with Aerospace Applications, PhD
thesis, University of Leicester.
Balas, G. (2003). Flight control law design: An industry perspective, European Journal of Control,
special issue 9(23): 207226.
Balas, G., Garrard, W. & Reiner, J. (1992). Robust dynamic inversion control laws for aircraft
control, Proceedings of the AIAA Guidance, Navigation and Control Conference, AIAA,
Washington, DC, pp. 192205.
Brinker, J. & Wise, K. (1999). Nonlinear simulation analysis of a tailless advanced ghter
aircraft recongurable ight control law, AIAA Guidance, Navigation and Control
Conference and Exhibit, number AIAA-99-4040, Portland, OR.
Calise, A. J., Lee, S. & Sharma, M. (2001). Development of a recongurable ight control law
for the x-36 tailless ghter aircraft, AIAA Journal of Guidance, Control and Dynamics
24(5): 896902.
Campa, G., Seanor, B., Gu, Y. & Napolitano, M. R. (2005). Nldi guidance control laws for close
formation ight, American Control Conference, Portland, OR, USA.
Chu, Q. (2007). Lecture Notes AE4-394, Modern Flight Test Technologies and System Identi cation,
Delft University of Technology, Faculty of Aerospace Engineering.
Chu, Q., Mulder, J. & Sridhar, J. (1994). Decomposition of aircraft state and parameter
estimation problems, Proceedings of fhe 10th IFAC Sympium on System Identi ation, Vol.
Vol. 3, pp. pp. 6166.
Cieslak, J., Henry, D., Zolghadri, A. & Goupil, P. (2008). Development of an active
fault-tolerant ight control strategy, AIAA Journal of Guidance, Control and Dynamics
31: 135147.
Civil Aviation Safety Data 1993-2007 (2008). Technical report, Civil Aviation Authority of the
Netherlands (CAA-NL).
Cooper, G. & Harper, R. J. (1969). The use of pilot rating in the evaluation of aircraft handling
qualities, Technical Report TN D-5153, NASA.
Corvin, J. H., Havern, W. J., Hoy, S. E., Norat, K. F., Urnes, J. M. & Wells, E. A. (1991).
Self-repairing ight control systems, volume i: Flight test evaluation on an f-15
aircraft, Final Report WL-TR-91-3025.
da Costa, R., Chu, Q. P. & Mulder, J. A. (2003). Re-entry ight controller design using nonlinear
dynamic inversion, Journal of Spacecraft and Rockets 40(1): 6471.
Ganguli, S., Papageorgiou, G., Glavaski, S. & M., E. (2005). Piloted simulation of fault
detection, isolation and reconguration algorithms for a civil transport aircraft, AIAA
Guidance, Navigation and Control Conference and Exhibit, number AIAA-2005-5936, San
Francisco, CA.
Ganguli, S., Papageorgiou, G., Glavaski, S. & M., E. (2006). Aircraft fault detection,
isolation and reconguration in the presence of measurement errors, AIAA Guidance,
Navigation and Control Conference and Exhibit, number AIAA-2006-6551, Keystone, Co.
Gavrilets, V. (2008). Damage tolerant ight control systems for unmanned aircraft, ICAS/ATIO
Conference.
Hallouzi, R. & Verhaegen, M. (2008). Fault-tolerant subspace predictive control applied to a
boeing 747 model, AIAA Journal of Guidance, Control and Dynamics 31: 873883.
Holzapfel, F. (2004). Nichtlineare adaptive Regelung eines unbemannten Fluggertes, PhD thesis,
Lehrstuhl fr Flugmechanik und Flugregelung, Technische Universitt Mnchen.
Intelligent Flight Control: Advanced Concept Program (1999). Final Report BOEING-STL 99P0040,
The Boeing Company.
Jategaonkar, R. (2006). Flight Vehicle System Identi cation: A Time Domain Methodology, Vol. 216
of Progress in Astronautics and Aeronautics Series, rst edn, AIAA.
Joosten, D., van den Boom, T. & Lombaerts, T. (2007). Effective control allocation in
fault-tolerant ight control with mpc and feedback linearization, Proceedings of the
European Conference on Systems and control, Kos, Greece, pp. 35523559.
Joosten, D., van den Boom, T. & Lombaerts, T. (2008). Fault-tolerant control using dynamic
inversion and model-predictive control applied to an aerospace benchmark, the
Proceedings of the 17th IFAC world congress, Vol. 17, pp. 1203012035.
KrishnaKumar, K. & Gundy-Burlet, K. (n.d.). Intelligent control approaches for aircraft
applications, Technical report, NeuroEngineering Laboratory, NASA Ames Research
Center.
Laban, M. (1994). Online aircraft aerodynamic model identi cation, PhD thesis, Delft University
of Technology.
Lombaerts, T. (2010). Fault Tolerant Flight Control. A Physical Model Approach, PhD thesis, Delft
University of Technology.
Lombaerts, T., Breeman, J., Joosten, D., van den Boom, T., Chu, Q., Mulder, J. & Verhaegen, M.
(2005). Specications modelling document for Garteur AG16 fault tolerant control,
Technical report, Delft University of Technology.
Lombaerts, T., Chu, Q., Mulder, J. & Joosten, D. (2007). Real time damaged aircraft model
identication for reconguring control, Proceedings of the AIAA AFM Conference and
Exhibit, number AIAA-2007-6717, Hilton Head, SC.
Lombaerts, T., Chu, Q.-P., Mulder, J. A. & Joosten, D. (2009). Flight control reconguration
based on a modular approach, Proceedings of the 7th IFAC SAFEPROCESS Symposium
on Fault Detection, Supervision and Safety of Technical Processes, pp. 259264.
Lombaerts, T., Chu, Q.-P., Mulder, J. A. & Joosten, D. (2010a). Flight control reconguration
based on a modular approach, Control Engineering Practice . under review.
Lombaerts, T., Joosten, D., J.H.Breemand, Smaili, H., Chu, Q., van den Boom, T., Mulder,
J. & Verhaegen, M. (2006). Assessment criteria as specications for reconguring
control, proceedings of the AIAA Guidance, Navigation, and Control Conference and
Exhibit, number AIAA-2006-6331, Keystone, CO.
Lombaerts, T., Looye, G., Chu, Q. P. & Mulder, J. A. (2010b). Pseudo control hedging and its
application for safe ight envelope protection, AIAA Guidance, Navigation and Control
Conference and Exhibit.
Lombaerts, T., Van Oort, E., Chu, Q. P., Mulder, J. A. & Joosten, D. (2010c). Online aerodynamic
model structure selection and parameter estimation for fault-tolerant control, Journal
of Guidance, Control and Dynamics . to be published.
Looye, G. (2007). An Integrated Approach to Aircraft Modelling and Flight Control Law Design,
PhD thesis, Delft University of Technology.
Looye, G. & Joos, H.-D. (2006). Design of autoland controller functions with multiobjective
optimization, AIAA Journal of Guidance, Control and Dynamics 29(2): 475484.
Maciejowski, J. & Jones, C. N. (2003). Mpc fault-tolerant ight control case study: Flight
1862, Proceedings of the 5th IFAC Symposium on Fault Detection, Supervision and Safety of
Technical Processes SAFEPROCESS, Washington DC, USA, pp. 121126.
Marcos, A. & Balas, G. (2003). A boeing 747U100/200 aircraft fault tolerant and diagnostic
benchmark, Technical Report AEMUUoMU2003U1, Department of Aerospace and
Engineering Mechanics, University of Minnesota.
Morelli, E. (2000). Real-time parameter estimation in the frequency domain, Journal of
Guidance, Control and Dynamics 23(5): 812818.
Mulder, J. (1986). Design and evaluation of dynamic ight test manoeuvers, PhD thesis, TU Delft,
Faculty of Aerospace Engineering.
Mulder, J., Chu, Q., Sridhar, J., Breeman, J. & Laban, M. (1999). Non-linear aircraft ight path
reconstruction review and new advances, Progress in Aerospace Sciences 35(7): 673726.
Mulder, J., van Staveren, W., van der Vaart, J. & de Weerdt, E. (January 2006). Lecture Notes
AE3-302, Flight Dynamics, Delft University of Technology, Faculty of Aerospace
Engineering, Delft, The Netherlands.
Oppenheimer, M. & Doman, D. (2006). Efcient reconguration and recovery from damage
for air vehicles, Proceedings of the AIAA Guidance, Navigation and Control Conference
and Exhibit, number AIAA-2006-6552.
Ostroff, A. J. & Bacon, B. J. (2002). Enhanced ndi strategies for recongurable ight control,
Proceedings of the American Control Conference, Anchorage, AK.
Ramakrishna, V., Hunt, L. & Meyer, G. (2001). Parameter variations, relative degree, and stable
inversion, Automatica 37: 871880.
Reiner, J., Balas, G. J. & Garrard, W. L. (1996). Flight control design using robust dynamic
inversion and time-scale separation, Automatica 32(11): 14931504.
Smaili, H., Breeman, J. & Lombaerts, T. (2008). A simulation benchmark for aircraft
survivability assessment, Proceedings of the International Congress of Aeronautical
Sciences, number 2008-9.3.2.
Smaili, M., Breeman, J., Lombaerts, T. & Joosten, D. (2006). A simulation benchmark for
integrated fault tolerant ight control evaluation, Proceedings of the AIAA Modelling
and Simulation Technologies Conference and Exhibit, number AIAA-2006-6471.
Stroosma, O., van Paassen, M. & Mulder, M. (2003). Using the simona research simulator
for human-machine interaction research, AIAA Modeling and Simulation Technologies
Conference.
Szaszi, I., Ganguli, S., Marcos, A., Balas, G. J. & Bokor, J. (2002). Application of fdi to a
nonlinear boeing 747 aircraft, 10th Mediterranean Conference on Control and Automation,
Lisbon, Portugal.
van Soest, W., Chu, Q. P. & Mulder, J. A. (2006). Combined feedback linearization and model
predictive control for re-entry ight, AIAA Journal of Guidance, Control and Dynamics
29(2): 427434.
Varga, A. (2007). Design of least order residual generators for fault detection and isolation with
application to monitoring actuator/surface faults for a boeing 747-100/200 aircraft,
Technical report, German Aerospace Center (DLR).
Varga, A. & Hecker, S. (2004). Methods for threshold selection for robust residual evaluation,
Technical report, German Aerospace Center (DLR).
Walker, G. & Allen, D. (2002). X-35b stovl ight control law design and ying qualities,
Proceedings of the Biennial International Powered Lift Conference and Exhibit, number
AIAA-2002-6018.
Ward, D. & Barron, R. (1995). A self-designing receding horizon optimal ight controller,
Proceedings of the American Control Conference, Seattle, Washington.
6
Design of Intelligent Fault-Tolerant

Flight Control System
for Unmanned Aerial Vehicles
Yuta Kobayashi and Masaki Takahashi
Keio University
Japan
1. Introduction
Recently, unmanned aerial vehicles (UAVs) have gained worldwide attention. Because the
safety of people on board does not need to be considered, small UAVs can easily be made
for low-cost. Therefore, a UAV can be used to observe disasters, to surveil for a long time,
and so on. However, it also has several disadvantages such as unreliability and worse
performance in unexpected situations. Because small UAVs must be easily made for low-
cost, adding a redundant on-board actuator or sensor in order to deal with unexpected
situations is unsuitable. Thus, several researchers have proposed a flight control system
using a software redundancy approach.
For fault detection, methods using multiple-model adaptive estimation (MMAE) (Guillaume
Ducard & Hans P. Geering, 2008), and system parameters (Mohammad Azam et al, 2005) have
been proposed. However, because these methods design a model or parameters for only each
assumed fault in designing, unexpected faults cannot be detected. On the other hand, another
method discriminates between faults and natural disturbances like gusts of wind. (Jovan D.
Boskobic et al, 2005) However, this is not easy because the expected disturbances are assumed
in designing. Currently, the demand for a UAV flight control system is to discriminate
between faults and natural disturbances fundamentally with a simple algorithm.
In this research, an intelligent flight control system was developed that can discriminate
between faults and natural disturbances in order to evaluate and deal with the situation. In
the proposed control system, an evaluator of flight conditions was designed on the basis of
the dynamics of a controlled object. Moreover, to deal with the situation adaptively, a new
flight-path-planning generator was introduced on the basis of the evaluation. In this study,
each subsystem was designed by a neural network. Moreover, the learning-based
systematical design method was developed that uses evaluation functions for the
subsystems. To verify the effectiveness of the proposed flight control system, a six-degree-
of-freedom nonlinear simulation was carried out.
2. Aircraft motion
The UAV treated in this research is a double-delta-wing UAV shown in Fig. 1. The coordinate
system is defined in Table 1. The motion equation of an aircraft is derived from Newtonian
dynamics. Six-degree-of-freedom nonlinear equation of motion is shown in Eq. (1).
y Q, M
er1 r
er 2
V el 2
el1
CG
U
P, L
x
Vc
R, N
z
Fig. 1. Body axis
(
X = m U + QW VR + g sin )
Y = m (V + UR PW g cos sin )
Z = m ( W + PV UQ g cos cos )
(1)
L = PI RJ + QR ( I I ) PQJ
x xz z y xz
M = QI + PR ( I I ) + ( P R ) J
y x z
2 2
xz
N = RI PJ + PQ ( I I ) + QRJ
z xz y x xz
In Eq. (1), X, Y, and Z indicate each axiss external force term except for gravitational force
(including aerodynamic force, thrust force). In addition, ,, and indicate Euler angle of
each axis. The proper nonlinear model shown in Eq. (1) is used in the numerical simulation.
In contrast, the linearized model based on Eq. (1) is used to design the controller. Many of
parameters of motion equation are decided on the basis of the wind-tunnel experiment. The
parameters that cannot be acquired in the experiment are estimated by the method using
nonlinear function. (Kato et al, 1982)
Each elevon steerage angle of the double-delta-wing UAV is expressed in Eq. (2) by using
elevator steerage angle e and aileron steerage angle a .
el1 = el2 = 1 2 ( e + a )
(2)
er1 = er2 = 1 2 ( e a )
3. Fault-tolerant system
The block diagram of the proposed intelligent fault-tolerant flight control system is shown
in Fig. 2. It is composed of fault detection, fault identification, and fault accommodation
(FDIA). In this section, the brief summary of each system is represented.
Design of Intelligent Fault-Tolerant Flight Control System for Unmanned Aerial Vehicles 119
Velocity & External force &

Distance
Angular velocity Moment
Forward U X x
Starboard V Y y
Down W Z z
Roll P L
Pitch Q M
Yaw R N
Table 1. Coordinate system and symbol
Detector Guidance command

Actuator command
NN
Actuator
Identifier Estimate Value Observed
NN
Value
Observed
Value Flight
Actuator
Model
Estimator
NN
Flight Path
Sensor Navigation Generator
Control
GPS Kalman NN
MDM/MDP
Distributor
IMU Filter
Guidance
MDM/MDP
Flight Condition , Failure Position

Fig. 2. Block diagram of proposed flight control system
3.1 Fault detection

Fault detection is to distinguish faults from natural phenomena like gusts of wind. To
achieve this, this research focused on how each influence on the dynamics of an aircraft, and
then an estimator and a detector were designed. The estimator can estimate the ideal state of
an UAV. The detector evaluates the flight conditions of an UAV by using the error
information between the observed and estimated values.
3.2 Fault identification

Fault identification is to locate a broken actuator. To achieve this, an identifier was designed.
Generally, to identify a fault, a method is used that sets a threshold value of error
information between an actuator command and a steerage value. However, this method
depends on designers thought, and inevitably the design work gets into trial and error. By
contrast, this research focused on the nonlinear mapping ability of a neural network to
flexibly respond to changes.
3.3 Fault accommodation

Fault accommodation is to stabilize the flight conditions of an UAV when a fault emerges.
To achieve this, a distributor and a flight path generator were designed. The distributor
switches the distribution matrix that sends a control command to actuators on the basis of
the location of a broken actuator. This countermeasure results in the maximum application
of the remaining actuators. The flight path generator generates a new flight path which
automatically takes account of both flight stability and following capability of mission
trajectory on the basis of the evaluation result.
4. Specific design of each component

4.1 Guidance and control
In this research, a coupled motion between longitudinal and lateral-directional is controlled.
This is because the roll angular velocity is controlled by limiting the derivative value of the
bank angle command. Therefore, the motion of UAV can be separated into longitudinal and
lateral-directional motions. In the guidance and control system, longitudinal guidance,
lateral-directional guidance, longitudinal control, and lateral-directional control were
designed separately. The guidance and control laws were designed by multiple delay model
and multiple design point (MDM/MDP) method. The block diagram of longitudinal
guidance, lateral-directional guidance, longitudinal control, and lateral-directional control
are shown in Figs. 3 to 6.
4.2 Estimator
The estimator achieves nonlinear dynamics of the UAV approximately by using nonlinear
mapping ability of feedforward-type neural network. It estimates next state vectors of the
UAV from previous state vectors and actuator steerage commands.
The structure is three-layer neural network shown in Fig. 7. Input layer has 15 neurons,
hidden layer has 18, and output layer has 9. In Fig. 4, the index of obs means the observed
value, the index of cmd means the actuator steerage command, and the index of est
Pitch Pitch
Height rate rate*
command Delay
1/s Model Longitudinal dh/dt
Feedback/
Motion
Feedforward Height
Ground speed Point Mass
Gain t
command Delay t* Approximation
Ground
1/s Model speed
Fig. 3. Block diagram of longitudinal guidance system

Bank Bank
angle angle*
y Delay
Lateral- y
command Model
1/s Feedback/ directionl
Feedforward Motion dy/dt
Gain Sideslip angle = 0
Point Mass
Approximation
Fig. 4. Block diagram of lateral-directional guidance system
Pitch rate
command Longitudinal Pitch rate
Feedback/ e *
1/s Delay e Motion
Feedforward Model Short-period
Short-
Gain Mode
Approximation Normal
acceleration
Fig. 5. Block diagram of longitudinal control system
Bank angle
command
a Delay a * Roll rate
1/s Model
Bank angle
Feedback/ Lateral -
Sideslip angle Feedforward directional Sideslip angle
command Gain Motion
1/s
r Delay r * Yaw rate
Model
Fig. 6. Block diagram of lateral-directional control system
u obs
v obs
w obs
obs u est
ob
obss v est
obs w est
p obs est
q obs est
robs est
el 1 cm d p est
el 2 cm d q est
er 1 cm d rest
er 2 cm d
r cm d
t cmd
cm d
Fig. 7. Structure of estimator neural network

means the estimated value. The transfer functions of each layer are shown in Eqs. (3) to (5),
where each tansig and purelin means tangent-sigmoid function, linear function shown
in Eqs. (6) and (7) respectively. In Eqs. (3) to (5), neti, netj, and netk mean the input of input
layer, hidden layer, and output layer respectively.
In addition, back propagation (BP) is applied for the learning of neural network. The flight
data acquired with the six-degree-of-freedom nonlinear simulation is used as a teach signal.
f i = tansig ( neti ) (3)
(
f j = tansig net j ) (4)
f k = purelin ( netk ) (5)
2
f (x) = 1 (6)
1 + exp ( 2 x )
f (x) = x (7)
4.3 Detector
The detector discriminates the influence of fault on the UAV from that of natural
disturbance such as gusts of wind by focusing on the impact on the dynamics of the UAV. It
uses the error between observed value and estimated value as the information about the
dynamics of the UAV. Moreover, the error between actuator steerage command and the real
actuator steerage angle is used for the evaluation of the flight condition. The derivative of
bank angle is also used.
Because input-output characteristic is unknown, the structure of the detector is three-layer
neural network shown in Fig. 8. Input layer has 10 neurons, hidden layer has 20, and output
layer has 1. The transfer functions of each layer are shown in Eqs. (8) to (10).
pest pobs
qest qobs
rest robs
uest uobs
vest vobs Flight
west wobs Condition
cmd obs
qcmd qobs
cmd
obs
Fig. 8. Structure of detector neural network

f i = purelin ( neti ) (8)
(
Because there is no explicit teach signal, a genetic algorithm (GA) was applied for the
learning of neural network. In GA, 50 individuals that encode the connection weight of a
neural network were prepared. Both fitness proportionate and elite selection strategies were
used. Moreover, with repeating random crossover and mutation, the individual that had the
highest fitness was acquired. 3 cases about gust in different directions and 9 cases about left
elevon-1 fault in different angles are used as the simulation case. Both gusts of wind and
fault are occurred in horizontal flight. The fitness function is shown in Eq. (11), where td is
the detection time, t failure is the initiation time of fault, and ad is the constant value of
detector for evaluation. In the evaluation, to detect the fault more quickly has higher score.
In addition, in the gusts of wind cases, when the detector did false detection, the value of
fitness function becomes zero.
0 t < t
d ( failure )
J= (11)
( (
exp ad td t failure

)) ( td t failure )
4.4 Identifier
The identifier locates where the broken actuator is by using the information of both actuator
steerage command and actuator steerage angle. Neural network shown in Fig. 9 is located in
each actuator and the location of broken actuator is identified by the outputs of each neural
network.
Because input-output characteristic is unknown, the structure of the identifier is three-layer
neural network. Input layer has 3 neurons, hidden layer has 18, and output layer has 1. The
cmd
obs Identifier Value
cmd obs
Fig. 9. Structure of identifier neural network

transfer functions of each layer are shown in Eqs. (12) to (14). GA is applied for the learning
of neural network. As the simulation case, 9 cases about left elevon-1 fault in different
angles happened in horizontal flight are used. Equation (15) is the fitness function, where ti
is the identification time, t failure is the initiation time of fault, and ai is the constant value of
identifier for evaluation. In the evaluation, to identify the location of broken actuator more
quickly has higher score.
f i = purelin ( neti ) (12)
(
i (
0 t < t
failure )
J= (15)
( (
exp ai ti t failure

)) (t t
i failure )
4.5 Distributor
The distributor switches the distribution matrix by using the outputs of the detector and the
identifier. When the distribution matrix was changed, the elevator, aileron, and rudder
commands from the control system are divided into 5 actuator commands (left elevon-1, left
elevon-2, right elevon-1, right elevon-2, and rudder) to separate the broken actuator. The
switching algorithm is to change the command for the broken actuator to zero and to realize
the maximum use of the remaining actuators. The structure of the distributor is shown in
Fig. 10.
Fig. 10. Structure of distributor

4.6 Flight path generator

The flight path generator is located in parallel with the guidance system. It generates a new
flight path which considers both flight stability and following capability of mission
trajectory under the condition where the elevon fault is occurred.
In general, there are two turning methods. One is to use a bank angle and the other a
sideslip angle. To assure robustness against the rudder fault, the turning method using the
bank angle was adopted in the guidance system. On the other hand, to assure robustness
against the elevon fault, the turning method using the sideslip was adopted in the flight
path generator.
Generally, because drag increases when the sideslip angle is allowed to changes in the
turning flight, too much energy is used. However, the emergency situation such as an
elevon fault is an exception because keeping the flight stable is more important than saving
energy. Therefore, the flight path generator has been designed that enables the sideslip
angle to change.
The flight path generator calculates the desired sideslip angle command by using Eq. (16),
where refstandard is the standard sideslip angle command which achieves the turning flight in
mission trajectory. To generate a new flight path by changing the radius adaptively, the
flight path generator calculates K depending on the fault level.
ref = K ref standard (16)
Because input-output characteristic is unknown, the structure of the flight path generator is
three-layer neural network shown in Fig.11. Input layer has 6 neurons, hidden layer has 1,
and output layer has 1. The input signals of the flight path generator are the signals from
both the detector and the identifier which are integrated in a given time. The transfer
functions of each layer are shown in Eqs. (17) to (19), where shifti is the width of parallel
shift. Equation (17) is the symmetric double sigmoid function. (Akihiko Shimura & Kazuo
Yoshida, 2001)
Flight Condition
Identifier Value el1
Identifier Value el 2
K
Identifier Value er1
Identifier Value er 2
Identifier Value r
Fig. 11. Structure of flight path generator
1
fi =
2
{tansig ( neti - shifti ) + tansig ( neti + shifti ) } (17)
(
f j = exp net j ) (18)
GA is applied for the learning of neural network. As the simulation case, 7 cases about
conducting the turning flight after left elevon-1 fault in different angles happened in
horizontal flight are used. The termination conditions of each simulation case are as
follows.
(A) 120 < [deg] < 300 x[m] < 500
(B) height < 0.18
(C) [deg] < 4.9 29 < [deg]
(D) [deg] < 9.9 9.9 < [deg]
Equation (20) is the fitness function, where aref 1 and aref 2 are the constant value for the
following capability of mission trajectory. Yref is the y-direction target value of mission
trajectory. Yref is the y-direction target value generated by the flight path generator. time ,
time failure , and timestable are respectively the simulation time, the initiation time of fault , and
the time when the error value between the real height and that of mission trajectory is
controlled within the constant value. In the evaluation, both the following capability of
mission trajectory and the flight stability are evaluated.
In addition to Eq. (20), the termination conditions are also evaluated. When the simulation
was stopped because of the termination condition except for (A), the value of fitness
function becomes zero because the stability is lost.
(
J = aref 1 exp aref 2 Yref Yref )
(20)
timestable
+ exp
(
time -failure time
)

5. Numerical simulation
5.1 Simulation condition
The effectiveness of the proposed intelligent fault-tolerant flight control system was verified
with the six-degree-of-freedom nonlinear simulation. The airframe model, external
environment model, and guidance/control law were considered as a mathematical model in
the simulation. In the airframe model, the actuator characteristic was expressed using the
second order time delay model with restrictions of position and velocity. In addition, the
characteristic of sensor was assumed to be ideal that there were no errors in both static and
dynamic conditions. As the external environment model, only wind was used. The constant
wind model was constructed by using the MIL-F-9490D method applied to the ALFLEX
simulation. (NAL/NASDA ALFLEX Group, 1994) It considered the difference of the
airframe size by multiplying the scale-dependent constant value. The actuator-fixed fault
such as Lock-in-place, Hard-over and Float was adopted as the actuator fault model. (Jovan
D. Boskobic et al, 2005)
As the mission trajectory, turning above devastated district in constant height to observe
was applied. It is shown in Fig. 12. In this mission, the UAV flaw at the height of 30m in the
velocity of 20m/s and the constant wind was from +x direction. The gusts of wind was
expressed by changing the scale-dependent constant value of constant wind in 3 seconds.
(Kohichiroh Yoshida et al, 1994) In addition, not only the learned fault, left elevon-1 fault,
but also non-learned fault, rudder fault, was considered. The conditions of fault and gust are
represented in Table 2.
In the simulation, the proposed intelligent fault-tolerant flight control system and the flight
control system designed by MDM/MDP method were compared.
Fig. 12. Mission trajectory
Condition Time Direction

Gust1 15s From y minus
Failure 30s
Gust2 90s From x plus
Table 2. Conditions of disturbance
5.2 Simulation results

In this section, the simulation results under the condition shown in section 5.1 are
represented.
First, Figs. 13 to 19 respectively show the results under the conditions where the left elevon-
1 was fixed at 9 degree for the flight trajectory, the time history of bank angle, sideslip angle,
and actuator steerage. In addition, Table 3 shows the effective area.
Second, the results for detection, identification, and accommodation are shown. The output
of the detector and the identifier are respectively shown in Figs. 20 and 21. Figure 22 shows
the relationship between the fixed angle of broken elevon and the y-direction target value
generated by the flight path generator. Moreover, the coherence functions between the
observed value and the estimated value for velocity u and angular velocity q are compared
under the conditions of a fault and gust of wind in Figs. 23 and 24.
Finally, Figs. 25 and 26 show the results under the condition where the rudder was fixed at -
8 degree for the flight trajectory and the time history of actuator steerage.
5.3 Evaluation
From the results in Figs. 13 to 19, we confirmed how each method deals with the fault in
which the elevon is fixed at the angle.
The conventional system generates a bank angle command and achieves a turning flight by
using an elevon. On the other hand, the proposed flight control system stabilizes the
airframe by using redundant elevon in horizontal flight as soon as the fault happens. After
that, it generates a sideslip angle command and achieves a turning flight by using a rudder.
Proposed System
Normal System
40
-z [m]
20
0
2000
1000 2000
0 1000
-1000 0
-1000
x [m] y [m]
Fig. 13. Flight trajectory (left elevon-1 fault)
20
obs
cmd
bank angle [deg]
10
Straight Straight
Turning
-10
0 50 100 150 200 250 300 350
Gust Failure Gust time [s]
Fig. 14. Time history of bank angle (normal system)
20
obs
cmd
bank angle [deg]
Straight Straight
10 Straight
Turning Turning
-10
0 50 100 150 200 250 300 350
Fig. 15. Time history of bank angle (proposed system)
10 obs
sideslip angle [deg]
cmd
5
-5
Straight Straight
-10 Turning
0 50 100 150 200 250 300 350
Fig. 16. Time history of sideslip angle (normal system)
10 obs
sideslip angle [deg]
cmd
5
-5
Straight Straight Straight
-10 Turning Turning
0 50 100 150 200 250 300 350
Fig. 17. Time history of sideslip angle (proposed system)
10
0
-10
0 50 100 150 200 250 300 350
10
0
-10
0 50 100 150 200 250 300 350
10
0
-10
0 50 100 150 200 250 obs
300 350350
10 cmd
0
-10
0 50 100 150 200 250 300 350
10
0
-10
0 50 100 150 200 250 300 350
15
10
5
0
0 50 100 150 200 250 300 350
Gust Failure Gust time[s]
Fig. 18. Time history of actuator steerage (normal system)
10
0
-10
0 50 100 150 200 250 300 350
10
0
-10
0 50 100 150 200 250 300 350
10
0
-10
0 50 100 150 200 250 obs
300 350
350
cmd
10
0
-10
0 50 100 150 200 250 300 350
10
0
-10
0 50 100 150 200 250 300 350
15
10
5
0
0 50 100 150 200 250 300 350
Gust Failure Gust time[s]
Fig. 19. Time history of actuator steerage (proposed system)

Flight Condition
2
0
Failure
0 10 20 30 40 50 60
Gust time[s]
Fig. 20. Output of detector neural network
2
el1
-2
0 10 20 30 40 50 60
2 time[s]
el 2
-2
0 10 20 30 40 50 60
time[s]
2
er1
-2
0 10 20 30 40 50 60
time[s]
2
er 2
-2
0 10 20 30 40 50 60
time[s]
2
r
-2
0 10 20 30 40 50 60
Gust time[s] Failure
Fig. 21. Output of identifier neural network
Proposed System
Target Value of Y direction [m]
1400 Normal System
1200
1000
800
-10 -5 0 5 10
degree of the locked angle[deg]
Fig. 22. Target value generated by flight path generator
1 Fault
Coherence Function
Gust
0.5
0
0 10 20 30 40 50
Frequency [Hz]
Fig. 23. Coherence function, u
1
Coherence Function
Fault
Gust
0.5
0
0 10 20 30 40 50
Frequency [Hz]
Fig. 24. Coherence function, q
40
-z [m]
20
0
2000
1000 2000
0 1000
0
-1000 -1000
x [m] y [m]
Fig. 25. Flight trajectory (rudder fault)

10
0
-10
0 50 100 150 200 250 300 350
10
0
-10
0 50 100 150 200 250 300 350
10
0
-10
0 50 100 150 200 250 obs
300 350
350
10 cmd
0
-10
0 50 100 150 200 250 300 350
10
0
-10
0 50 100 150 200 250 300 350
15
10
5
0
0 50 100 150 200 250 300 350
time[s]
Gust Failure Gust
Fig. 26. Time history of actuator steerage (rudder fault)
Fault position Elevon Rudder

Range of movement [deg] -1313 -88
Effective area[deg] -139 -88
Table 3. Effective area of proposed system

The results in Figs. 13, 15, 17, and 19, confirm that the vibration motion is generated in the
horizontal flight after turning flight by using the proposed method. This vibration frequency
is about 0.067 Hz. This is because the resonation with the vibration occurs at the
longitudinal short cycle mode and the lateral-directional dutchroll mode when the turning
flight is changed to the horizontal flight in order to deal with the fault. However, this
vibration fits into the stable area of both an attack angle and a sideslip angle that is
established when designed and shown in section 4.6 as the termination conditions.
Therefore, the vibration is considered to be an allowable range.
From the results in Figs. 25 and 26, we confirmed that the proposed flight control system
generates a bank angle command and achieves a turning flight by using an elevon when a
rudder fault happens.
These results confirm that the proposed system can detect, identify and accommodate both
learned and non-learned faults.
From the simulation results, we confirmed that the proposed flight control system can
stabilize the airframe in fault situations shown in Table 3.
Figure 20 shows the output of a detector which means the evaluation value of a flight
condition. We confirmed that the detector can distinguish the fault from the gust of wind.
The flight control system can distinguish between the fault and the gusts from various
directions because a number of directional gusts are considered in the learning of neural
network. Figures 23 and 24 show that the gust has a wider range of frequency where the
coherence function takes the value of approximately 1 than the fault. If the disturbance is
estimated, the motion of the system is the same as the model assumed when the control
system is designed. On the other hand, the motion of the system with the fault is different
from the assumed model. Therefore, the proposed model-based detector can accurately
detect faults.
Figure 21 shows the output of an identifier which means the evaluation value of the fault
position. It was confirmed that the proposed identifier can identify the fault position
because only the broken actuator indicates the abnormal value.
Figure 22 shows the performance of a flight path generator. The horizontal axis indicates a
fixed angle of a broken elevon and the vertical axis indicates a new target value of y
direction that is calculated by the flight path generator. The results confirm that the higher
the level of a fault, the gentler the turning based on a new target value generated by the
flight path generator. In this research, the actuator error between the stable and the broken
conditions means the fault level. Moreover, the error from a mission trajectory is considered
in the evaluation function. Therefore, the proposed flight control system can generate a
suitable target value of turning in accordance with the situation.
The proposed flight control system focuses on the change in dynamics caused by a fault. It is
designed by considering the elevon fault that enormously influences the airframe because
an elevon plays the roles of both an aileron and an elevator. The simulation results confirm
the proposed system can perform well in both learned and non-learned fault situations.
6. Conclusion
This research aimed at proposing an intelligent fault-tolerant flight control system for an
unmanned aerial vehicle (UAV). In particular, the flight control system was developed that
has estimator, detector, identifier, distributor, and flight path generator. The proposed
system distinguishes a fault from a disturbance like a gust of wind and automatically
generates a new flight path suited to the fault level. To verify the effectiveness of the
proposed method, a six-degree-of-freedom nonlinear simulation was carried out. In the
simulation, we assumed that the fault in left elevon-1, which was learned in designing each
neural network, or the fault in the rudder, which was not learned, would be generated in a
horizontal flight. The simulation results confirm that the proposed flight control system can
detect, identify and accommodate the fault and keep a flight stable. Moreover, the proposed
system can distinguish a fault from a gust and keep a flight stable automatically. It is
expected that the proposed design method can be used in broader flight areas by expanding
the learning area.
7. References
Akihiko Shimura and Kazuo Yoshida, Non-Linear Neuro Control for Active Steering for
Various Road Condition, The Japan Society of Mechanical and Engineers, Vol. 67, No.
654(2001), pp. 407-413.
Brian L. Steavens and Frank L. Lewis, Aircraft Control and Simulation 2nd Edition, JOHN
WILEY & SONS, INC. (2003)
Guillaume Ducard and Hans P. Geering, Efficient Nonlinear Actuator Fault Detection and
Isolation System for Unmanned Aerial Vehicles, AIAA, Journal of Guidance, Control,
and Dynamics, Vol. 31, No.1 (2008), pp. 225-237.
Jovan D. Boskovic, Sarah E. Bergstrom ,and Raman K. Mehra, Robust Integrated
Flight Control Design Under Failures, Damage, and State-Depenndent
Disturbances, AIAA, Journal of Guidance, Control, and Dynamics, Vol. 28, No.5 (2005),
pp. 902-916.
Kanichiro Kato, Akio Oya, and Kenzi Karasawa, Introduction of Aircraft Dynamics,
University of Tokyo Press, (1982).
Kohichiroh Yoshida, kazumichi Mototsuna and Yasushi Kumakura, Elementary knowledge
of marine technology, Seizandou,(1994)
Masaki Takahashi, Teruma Narukawa and Kazuo Yoshida, Robustness and Fault-Tolerance
of Cubic Neural Network Intelligent Control Method : Comparison with Sliding
Mode Control, The Japan Society of Mechanical and Engineers, Vol. 69, No. 682(2003),
pp. 1579-1586.
Mohammad Azam, Krishana Pattipati, Jeffrey Allanach, Scott Poll, and Ann Patterson-Hine,
In-flight Fault Detection and Isolation in Aircraft Flight Control Systems, Aerospace
Conference, 2005 IEEE, (2005), pp. 3555- 3565.
NAL/NASDA ALFLEX Group, Flight simulation model for Automatic Landing Flight
Experiment (Part I : Free Flight and Ground Run Basic Model), Technical Report of
National Aerospace Laboratory, Vol. 1252 (1994).
Taro Tsukamoto, Masaaki Yanagihara, and Takanobu Suito, Feasibility Study of
Lateral/Directional Control of Winged Re-entry Vehicle with Split Elevons,
Technical Report of National Aerospace Laboratory, Vol. 1379 (1999).
Toshinari Shiotsuka, Kazusige Ohta, Kazuo Yoshida and Akio Nagamatsu, Identification
and Control of Four-Wheel-Steering Car by Neural Network, The Japan Society of
Mechanical and Engineers, Vol. 59, No. 559(1993), pp. 708-713.
Tsuyoshi Hatake, Junichiro Kawaguchi, and Tatsushi Izumi, Control in Aerospace,
CORONA PUBLISHING CO., LTD. (1999).
7
Active Fault Diagnosis and Major Actuator

Failure Accommodation: Application to a UAV
Franois Bateman1 , Hassan Noura2 and Mustapha Ouladsine3
1 French Air Force Academy, Salon de Provence
2 United Arab Emirates University, Al-Ain
3 Paul Cezanne University, Marseille
1,3 France
2 United Arab Emirates
1. Introduction
Interest in Unmanned Aerial Vehicles (UAVs) is growing worldwide. Nevertheless there are
numerous issues that must be overcome as a precondition to their routine and safe integration
in military and civilian airspaces. Chief among these are absence of certication standards and
regulations addressing UAV systems, poor reliability record of UAV systems and operations.
Standards and regulations for airworthiness certication and ight operations in the military
and civilian airspaces are being studied (Brigaud, 2006). In this respect, the USAR standard
suggests a mishap rate of one catastrophic mishap per one million hours (Brigaud, 2006). To
reach such performances, upcoming technologies have the promise of signicantly improving
the reliability of UAVs.
In this connection, a detailed study (OSD, 2003) shows that most of the breakdowns are due
to system failures such as propulsion, data link and Flight Control Systems (FCS). These
latter include all systems contributing to the aircraft stability and control such as avionics,
air data system, servo-actuators, control surfaces/servos, on-board software, navigation, and
other related subsystems. As regards FCS, it is recommended in (OSD, 2003) to incorporate
emerging technologies such as Self-Repairing Flight Control Systems (SRFCS) which have the
capability to diagnose and to repair malfunctions.
In this respect, Fault-tolerant control (FTC) are control systems that have the ability to
accommodate failures automatically in order to maintain system stability and a sufcient
level of performance. FTC are classied into passive and active methods. The analytical
fault-tolerant control operation can be achieved passively by the use of a control law designed
to guarantee an acceptable degree of performance in fault-free case and to be insensitive to
some faults. However, the passive methods are unsuitable to deal with a signicant number
of faults. In particular, for an aircraft, it may be tricky to design an a priori controller able
to accommodate the whole of the faults affecting the control surfaces. By contrast, an active
FTC consists of adjusting the controllers on-line according to the fault magnitude and type, in
order to maintain the closed-loop performance of the system. To do so, a fault detection and
isolation (FDI) module which provides information about the fault is required (Noura et al.,
2009). Active FTC mechanisms may be implemented either via pre-computed control laws or
via on-line automatic redesign.
In this respect, FDI and FTC applied to aeronautical systems have received considerable
attention in the literature. However, regarding the control surface failures, some problematics
tackled in this chapter are underlined:
severe failures are considered and the control surfaces may abruptly lock in any position
in their deection range,
each control surface being driven indenpendently, an actuator failure produces
aerodynamical couplings between the longitudinal and the lateral axis,
the UAV is equipped with an autopilot which masks the failure effects,
the aircraft studied is a small UAV and the control surface positions are not measured,
which makes the fault detection difcult,
the control surfaces have redundant effects, which complicates the fault isolation,
the FTC must take into account the bounds existing on the control surface deections and
the ight envelope.
In this chapter, a nonlinear UAV model which allows to simulate assymetrical control surface
failures is presented (Bateman et al., 2009). In fault-free mode, a nominal control law based
on an Eigenstructure Assignment (EA) strategy is designed. As the control surface positions
are not measured, a diagnosis system is performed with a bank of observers able to estimate
the unkown inputs. However, as the two ailerons offert redundant effects, isolating a fault on
these actuators requires an active diagnosis method (Bateman et al., 2008a). In the last part,
a precomputed FTC strategy dedicated to accommodate for a ruddervator failure is depicted
(Bateman et al., 2008b).
All the models can be simulated with MATLAB-SIMULINK. Files and tutorial can be
downloaded at http://www.lsis.org/bateman/UAV.zip.
2. Aircraft model
The aircraft studied in this paper and shown in Fig.1 is an inverted V-tail UAV. It is assumed
that its controls are fully indenpendent: x is the throttle, ar , al , f r , f l , er , el control the
right and left ailerons, the right and left aps, the right and left inverted V tail control surfaces
respectively. These latter controls are named ruddervators because they combine the tasks
of the elevators and rudder. In the fault-free mode, the ailerons and the ruddervators are
known as the primary control surfaces, they produce the roll, the pitch and the yaw. As far as
the aps are concerned, in the fault-free mode, they are only used to produce a lift increment
during takeoff and a drag increment during landing. They are known as the secondary control
surfaces.
It is assumed that each one of the primary control surfaces may lock at any arbitrary position
on its deection range. To compensate for the fault, the FTC exploits the redundancies
provided by the remaining control surfaces. In this pespective, the UAV model has to consider
the aerodynamic effects produced by each control surface.
The following dynamic model of the aircraft is presented in the case of a rigid-body aircraft,
the weight m is constant and the centre of gravity c.g. is xed position. Let R E = (O, xE , y E , zE )
be a right-hand inertial frame such that zE is the vertical direction downwards the earth, =
( x, y, z) denotes the position of c.g. in R E . Let Rb = (c.g., xb , y b , zb ) be a right-hand body xed
frame for the UAV, at t = 0 R E and Rb coincide. The linear velocities = (u, v, w) and the
angular velocities = ( p, q, r ) are expressed in the body frame Rb where p, q, r are roll, pitch
and yaw respectively. The orientation of the rigid body in R E is located with the bank angle
Active Fault Diagnosis and Major Actuator Failure Accommodation: Application to a UAV 139
Fig. 1. The RQ7A Shadow 200 UAV
, the pitch angle and the heading angle . The transformation from Rb to R E is given by a
transformation matrix TbE :

cos cos sin sin cos cos sin cos sin cos + sin sin
TbE = cos sin sin sin sin + cos cos cos sin sin sin cos (1)
sin sin cos cos cos
Forces FxRb , FyRb , FzRb acting on the aircraft are expressed in Rb , they originate in gravity F grav ,
propulsion F prop , and aerodynamic effects F aero. According to Newtons second law:
R
u Fx b p u
v = 1 R
Fy b q v (2)
m
w FzRb r w
where denotes the cross product. Let Rw = (c.g., xw , y w , zw ) be the wind reference frame
where xw is aligned with the true airspeed V. The orientation of the body reference frame
in the wind reference frame is located with the angle of attack and the sideslip . The
transformation from Rb to Rw is given by a transformation matrix Tbw :

cos cos sin sin cos
Tbw = cos sin cos sin sin (3)
sin 0 cos
Furthermore, the aerodynamic state variables (V, , ) and their time derivatives can be
formulated using Tbw from (Rauw, 1993).

V = u 2 + v2 + w 2
w
= arctan (4)
u
v
= arctan
u 2 + w2
For the sake of clarity, the forces are written in the reference frame where their expressions are
the simplest. They are transformed into the desired frame by means of the matrices TbE and
Tbw or their inverse.

T
F grav R E = 0 0 g
T
k(z)
F prop Rb = x 0 0 (5)
V
T
F aero Rw = qS CD Cy CL
The model of the engine propeller is given in (Boifer, 1998), is the air density, k is a constant
1
characteristic of the propeller engine, q = V 2 and S denote the aerodynamic pressure and
2
a reference surface. The aerodynamic force coefcients are expressed as linear combination
of the state elements and control inputs. The values of these aerodynamic coefcients can be
found in the attached MATLAB les.
S 2
CD = CD0 + C + CDar | ar | + CDal | al | + CD f r | f r | + CD f l | f l | + CDer | er | + CDel | el |
b2 L
Cy = Cy + Cyar ar + Cyal al + Cy f r f r + Cy f l f l + Cyer er + Cyel el (6)
CL = CL0 + CL + CLar ar + CLal al + CL f r f r + CL f l f l + CLer er + CLel el
The relationships between the angular velocities, their derivatives and the moments M R b
x ,
Rb Rb
My , Mz applied to the aircraft originate from the general moment equation. J is the inertia
matrix. R
p Mx b p p
q = J1 R
My b q J q

(7)
r R r r
Mz b
The moments are expressed in Rb , they are due to aerodynamic effects and are modeled as
follows:

M Rx b MyRb MzRb = qS bCl cCm bCn (8)
where c and b are the mean aerodynamic chord and the wing span. The aerodynamic moment
coefcients are expressed as a linear combination of state elements and control inputs as
bp br
Cl = Cl + Cl p + Clr + Clar ar + Clal al + Cler er + Clel el + Cl f r f r + Cl f l f l (9)
2V 2V
cq
Cm = Cm0 + Cm + Cmq + Cmar ar + Cmal al + Cmer er + Cmel el + Cm f r f r + Cm f l f l
2V
bp br
Cn = Cn + Cnp + Cnr + Cnar ar + Cnal al + Cner er + Cnel el + Cn f r f r + Cn f l f l
2V 2V
Equations (6) and (9) make obvious the aerodynamic forces and moments produced by each
control surface. This is useful to model the fault effects and the redundancies provided by the
healthy control surfaces.
The FDI/FTC problem is rst an attitude control problem, thus the heading angle , the x and
y coordinates are not studied in the sequel.
With regard to the kinematic relations, the bank angle and the pitch angle time derivatives are
(Boifer, 1998):
= p + q sin tan + r cos tan
= q cos r sin
(10)
The relationship between the time derivative of the position of the aircrafts centre of gravity
, the transformation matrix TbE and the linear velocities allows to write:
z = u sin + v sin cos + w cos cos (11)
Let X = ( V p q r z) T the state vector and U = (x ar al f r f l

er el ) T the control vector. All the state vector is measured and Y is the measure vector.
From above, the model of the UAV can be written as a nonlinear model afne in the control
X = f (X) + g(X)U (12)

Y = CX
Practically, the nonlinear aircraft model has been implemented with MATLAB in a sfunction.
In the fault-free mode, for a given operating point {Xe0 , Ue0 }, where Ue0 denotes the trim
positions of the controls, the linearized model of the aircraft can be written as
x = Ax + Bu (13)
y = Cx (14)
2.1 Fault model

In the fault-free mode, the control surface deections are constrained: asymmetrical aileron
deections produce the roll control, pitch is achieved through deecting both ruddervators
in the same direction and yaw is achieved through deecting both ruddervators in opposite
direction. As for the aps, their deection is symmetrical.
The faults considered are stuck control surfaces. For t t f , the faulty control vector
Uf (t) = U f , where t f is the fault-time and U f are the stuck control surface positions. For
the simulations, a fault is modeled as a rate limiter response to a step. The slew rate is chosen
equal to the maximum speed of the actuators. Let Uh be the remaining surfaces, then the state
equation (12) in faulty mode becomes:
X = f (X) + g f (X)Uf + gh (X)Uh (15)
3. The nominal controller

A linear state feedback controller with reference tracking is designed. It is based on an
EA method which allows to set the aircraft handling qualities (Magni et al., 1997). This
method allows to set the modes of the closed-loop (CL) aircraft with respect to the standards
(MIL-HDBK-1797, 1997) and to decouple some state and control variables from some modes.
The design is based on the fault-free linearized model (13) which modal analysis shows that
the spiral mode isopen-loop (OL) unstable.
Let Y = V z the tracked vector and Yref the reference vector. The autopilot is depicted
in gure 2 and the nominal control law is:
u = L + Kx (16)
where = Yref Y augments the state vector with the state variables which have to be tracked
to zero and T
= V z (17)
Ue0
?
- L -+
+
Yre f - X - Y
- - UAV C

6 6
+
?
K m Xe0
x
Fig. 2. The UAV control law
For a straight and level ight stage, Table 1 illustrates the EA strategy. A 0 means that the
mode and the state variable or the control are decoupled. On the contrary, an means that
they are coupled.
In the fault-free mode, the ruddervators produce the pitch and the yaw, thus the longitudinal
and the lateral axis are coupled. To take this into account, the autopilot is designed by
considering the complete linearized model of the aircraft (13). For example, a coupling has
been set between the spiral mode and the angle of attack which are a lateral mode and a
longitudinal state variable respectively. This is illustrated by the highlighted cells in Table 1.
This approach signicantly differs from the classical method which consists in designing two
autopilots, one for the longitudinal axis, another one for the lateral axis. However, from the
FTC point of view, a control surface failure upsets the equilibrium of forces and moments
and produce signicant couplings between longitudinal and lateral axis. Because of this, the
method adopted to design the nominal autopilot could be used to design the fault-tolerant
controllers. From (13), (16) and (17)

x A + BK BL x 0
= + Yre f (18)
C 0 I
with C R 49 and C(i, j) = 1 for {i, j} = {1, 1}, {2, 5}, {3, 3}, {4, 9} else C(i, j) = 0.
Matrices K R 79 and L R 74 are computed in order to set the state space matrixs
eigenvalues and eigenvectors in (18). These latter dene the aircrafts modes. The state space
matrix in (18) also writes:

A 0 B
+ K L = F+G K L (19)
C 0 0
Let i the i th eigenvalue (or closed-loop pole) corresponding to the eigenvector

v i R 13 ,
next i
F+G K L v = i

vi (20)
let

i
wi = K L v R7 (21)
then i
v

F i I G = 0 (22)
wi
mode short period phugoid throttle roll dutchroll spiral, V , z

OL poles 4.2 9.6i 0.03 0.5i 0.0002 21 1.7 5.6i 0.073
CL poles 10 10i 2 2i 1 100 5 5i 1 .25i 1.5 1 .5i
eigenvector
v 1,2
v 3,4
v5
v6
v 7,8

v 9,10
v 11

v 12,13
0 0 0
0 0 0
V 0 0 0 0
0 0 0
0 0 0
p 0 0 0 0
q 0 0 0
r 0 0 0 0
z 0

0
V
z
x
ar 0 0 0 0
al 0 0 0 0
f r 0 0 0 0 0 0 0 0
f l 0 0 0 0 0 0 0 0
er 0
el 0
Table 1. Eigenstructure assignment strategy, airspeed 25ms1 , height 200m

F i I G is a 13 20 matrix and the size of its null space is equal to the number of control
inputs, here seven. In order to set the eigenvector structure (Table 1) while reducing the
solution space, some constraints are added. To hide the i th mode to the jth state variable
or control input, a mask is introduced such that
j

1 20
maski = 0 0 1 0 0 (23)
In order to nd a unique solution vector, for each eigenvalue i , six masks are dened and the
system to solve writes
F i I G
mask i1
vi

= 0 (24)
... wi
mask i6

Let P = v 13 and Q =
v 1 ...

w1 . . .

w 7 , then according to (21), matrices K and L write:

K L = QP 1 (25)
Fig. 4 shows the nominal autopilot functionning in the [0s, 16s] fault-free time interval.
4. Fault diagnosis
The class of faults addressed here are stuck control surfaces. However, the proposed diagnosis
system can also deal with actuator the loss of efciency.
To process for the faults, the diagnosis system could be realized by measuring the actuator
positions. This approach which requires potentiometers, wiring and acquisition board is
complex to implement and induces an increase of weight. Without these measurements, the
control inputs appear as unknown inputs which have to be estimated. This can be achieved
by the use of observers able to estimate the unknown inputs of a system.
In this connection, the problem of unknown, constant or slowy varying input estimation using
banks of Kalman lters is discussed in (Kobayashi & Simon, 2003), (Ducard & Geering, 2008).
The unknown inputs are declared as state variables, under the condition that the system is
observable, the problem consists in estimating an augmented state vector.
However, to catch for the actuator fault transients, the observer has to estimate time varying
inputs. Such an estimation is possible if these inputs are observable. The input observability
problem was addressed by Patton in (Hou & Patton, 1998) who gave some necessary and
sufcient conditions to prove input observability for linear time invariant systems with
unknown initial conditions.
As far as the observer is concerned, Xiong (Xiong & Saif, 2003) proposed an Unknown Input
Decoupled Functionnal Observer (UIDFO) which has no boundedness conditions as for time
varying inputs and does not require differentiation of the measured outputs.
The fact remains that control surfaces offer redundancies that make aileron failures not
isolable. In these conditions, an active diagnosis strategy has to be considered.
Input observability of the UAV is studied in subsection 4.1, the UIDFO is briey described in
subsection 4.2 and a diagnosis system based on a bank of UIDFO is detailed in subsection 4.3.
The active diagnosis strategy is presented in subsection 4.4.
4.1 Input observability

Given, the linearized model (13).
Denition 1. The input u (t) is said to be observable if y(t) = 0 for t 0 implies u (t) = 0 for t > 0
(Hou & Patton, 1998).
Let C, sys and AC refer to the system matrix and the observability pencil respectively,
they write:
A B I 0 A I
sys = and AC = (26)
CD 00 C 0
Kroeneckers theory of singular pencils shows that any pencil M N with dimension m n
can be brought into the canonical quasidiagonal form:

Mc Nc
I Nf
P ( M N) Q =

(27)
M I
Mr Nr
P and Q are nonsingular constant matrices with dimensions m m and n n.
the nite eigenvalues are in the square and regular pencil I Jf and Jf is in Jordan
canonical form. It is built with Jordan blocks Jf i with dimension i i.
the innite eigenvalues are in the square and regular pencil J I and J is compound
with Jordan blocks J i of size i i.
M r Nr is a singular
pencil
T and
has a block diagonal structure, each block takes the form
I 0
M r i Nr i = T with dimension ri+1 ri . 0 stands for zero vector and 0 T
0 I
denotes its transpose. The ri are called Kroeneckers row indices.
M c Nc is a singular
pencil
and has a block diagonal structure, each block takes the form
: M cj Ncj = I 0 0 I . The c j c j+1 are called Kroeneckers column indices.
This decomposition is applied to sys and AC . The following theorem contains a numerically
effective test for input observability:
Theorem 1. System (13) is input observable if and only if the block M csys Ncsys vanishes in
equation (27) and dim(Jf sys )=dim(Jf AC ) (Hou et Patton)(Hou & Patton, 1998).
Nevertheless, due to the numerical unreliability of the computation, this form is not suitable
and the staircase form, computed with the GUPTRI algorithm, is used to exhibit the
Kroenecker form (Demmel & Kgstrom, 1993).
All the state vector is measured and theorem 1 is applied to assess the ruddervator
observability. For this outpout vector, the matrix system and the observability pencil have
the following structures:
AC has no nite eigenvalue, nine singular pencils with Kronecker raw indice equal to
one,
sys has no nite eigenvalue, seven singular pencils with Kronecker raw indice equal to
one and two 2 2 jordan blocks containing innite eigenvalues.
According to theorem 1 the ruddervator positions are observable. With the same measured
outputs, theorem 1 is applied to assess the aileron observability. The matrix system and the
observability pencil have the following structure:
AC is unchanged
sys has no nite eigenvalue, one singular pencil with Kronecker column indice equal to
one, eight singular pencils with Kronecker raw indice equal to one and one 2 2 jordan
block containing innite eigenvalues.
Due to the presence of the singular pencil with Kronecker column equal to one, the aileron
positions are not both observable. Thus, if one of the ailerons breaks down, the faulty control
is not isolable. This is depicted in Fig. 3 where a down lock position of the left aileron has the
same effects as a top lock position of the right aileron. In this case, an active diagnosis must
be used to discriminate the faulty control surface.

y
m -
p>0 p>0

z ?

al al
el el
er er
ar ar
Fig. 3. Aerodynamic effects induced by right and left aileron failures

4.2 The unknown input decoupled functional observer

In this part, results established in (Xiong & Saif, 2003) are recalled. The following dynamic
system driven by both known and unknown inputs is considered
x = Ax + Bu + Gd (28)
y = Cx
where x R n is the state vector, u R m is the known input vector, d R is the unknown
input vector and y R o is the output vector. A, B, G and C are matrices with appropriate
dimensions, C and G are assumed to be full rank.
The UIDFO detailed in (Xiong & Saif, 2003) provides an estimation d of the unknown input
d and an estimation z of linear combination of state Tx. Theoretically, no boundedness
conditions are required for the unknown inputs and their derivatives.
z = Fz + Hy + TBu + TGd (29)

d = (Wy Ez) with R +
Matrices F, H, T, W and E are all design parameters. in order to satisfy the following
conditions

FT TA + HC = 0 F is stable,

E = (TG) T P with P solution of: PF + F T P = Q
and Q, a semi-positive denite matrix, (30)

ET = G T T T PT = WC

rank(TG) = rank(G) =
These matrices exist if and only if

(i) rank(CG) = rank(G),
(ii) all unstable transmission zeros of system (A, G, C) are unobservable modes of (A, C).
To prove their existence and to calculate them, system (28) is transformed into "Special
Coordinate Basis" (SCB) (X.Liu et al., 2005).
4.3 Diagnosis system performed with a bank of UIDFO

The right aileron, the left aileron, the right ruddervator and the left ruddervator positions
are unknown. In order to estimate each one of these unknown input positions, a bank of
four UIDFOs is implemented. However, it has been proven above that the ailerons are not
both input observable. It means that, if one of these two controls is faulty, one of the two
estimations is wrong whereas the other is right, but in any case it is impossible to know which
one is which.
Let ar , al , er and el the positions estimated by each UIDFO, ar , al , er and el the known
control inputs computed by the nominal controller and ar , al , er and el the unknown
actual control surface positions. In the fault-free mode, the actual control surface positions are
equal to those computed by the controller and to their estimations. When an actuator breaks
down, the faulty actual control surface position is equal to its estimation but differs from that
computed by the controller. Residuals are obtained by processing the differences between the
positions computed by the controller with the estimated positions. The fault detection and
isolation process consists in monitoring residuals by comparing them with thresholds.
The rst UIDFO estimates the unknown right aileron actual position ar by processing the
measurement vector y and the known input u 1 = (al , er , el ) T . Let bi the column of the
control matrix B associated with the i control input, then B1 = (bal , ber , bel ) T and G1 =
(bar ),
z1 = F 1 z1 + H1 y + T1 B1 (al , er , el ) T + T1 G1 ar (31)
ar = 1 (W 1 y E 1 z1 )
The other three ones UIDFO equations write
z2 = F 2 z2 + H2 y + T2 B2 (ar , er , el ) T + T2 G2 al (32)
al = 2 (W 2 y E 2 z2 )
with B2 = (bar , ber , bel ) T and G2 = (bal ),
z3 = F 3 z3 + H3 y + T3 B3 (ar , al , el ) T + T3 G3 er (33)
er = 3 (W 3 y E 3 z3 )
with B3 = (bar , bal , bel ) T and G3 = (ber ),
z4 = F 4 z4 + H4 y + T4 B4 (ar , al , er ) T + T4 G4 el (34)
el = 4 (W 4 y E 4 z4 )
with B4 = (bar , bal , ber ) T and G4 = (bel ).
For all the UIDFOs, condition (i) is assessed and condition (ii ) is checked by computing the
staircase forms of the system matrices (A, Gj , C, 0) with j = {1, . . . , 4} and the observability
pencil (A, C) with the GUPTRI algorithm.
Error signals are generated by comparison between the control positions i and the estimated
positions i where i {ar , al , er , el }. In order to avoid false alarm that may arise from the
transient behavior, these signals are integrated on a duration to produce residuals ri such
that t+

ri = i ( ) i ( )d (35)
t
Let i the corresponding threshold and i a logical state such that i = 1 if ri > i else
i = 0. Then, to detect and to partially isolate the faulty control surface, an incidence matrix
is dened as follows:
Faulty control ar el al er
right aileron 1 0 1 0
left aileron 1 0 1 0
right ruddervator 0 0 0 1
left ruddervator 0 1 0 0
Table 2. The incidence matrix
This matrix reveals that fault on right aileron and fault on left aileron are not isolable.
In order to illustrate the above-mentioned concepts, three failure scenarios are studied: a
non critical ruddervator loss of efciency 50%, a catastrophic ruddervator locking and a non
critical aileron locking. For the three cases, the fault occurs at faulty time t f = 16s whereas
the UAV is turning and changing its airspeed (see Fig.4, Fig.7, Fig.10). These two manoeuvres
involve both ailerons and ruddervators.
4.3.1 Ruddervator loss of efciency

For the right ruddervator, a 50% loss of efciency is simulated. The nominal controller is
robust enough to accommodate for the fault as it is depicted in Fig.4. The actual control
surface positions and their estimations are shown in Fig. 5. As far as the right ruddervator is
concerned, after time t f , its control signal differs from its estimated position and this difference
renders the fault obvious. The residuals are depicted in Fig. 6, with respect to (35) and to the
incidence matrix Tab. 2, the right ruddervator is declared to be faulty.
4.3.2 Ruddervator locking

At time t f , the right ruddervator is stuck at position 0 . As it is illustrated in Fig.7, the nominal
controller cannot accommodate for the fault and the UAV is lost. The actual control surface
positions and their estimations are shown in Fig.8, As regards the right ruddervator, its control
signal differs from its estimated position and the residual analysis Fig.9 allows to declare this
control surface to be faulty. However, the control and measurement vectors diverge quickly,
thus the signal acquisition of the estimated positions has to be sampled fast.
4.3.3 Aileron locking

In the event of an aileron failure, the nominal controller is robust enough to accommodate for
the fault. However, the incidence matrix shows that faults on right and left ailerons cannot be
isolated. This is due to the redundancies offered by these control surfaces that are not input
observable. This aspect is depicted in Fig. 10, 11, 12 where the left aileron locks at position
+5 at time t f = 16s. Fig.10 shows that this fault is non critical (it is naturally accommodated
by the right aileron). However, as it is shown in Fig. 11, both estimations of aileron positions
are consistent and the corresponding residuals exceed the thresholds. As a consequence, it is
not possible to isolate the faulty aileron.
4.4 Active diagnosis

To overcome this problem, an active fault diagnosis strategy is proposed. It consists in exciting
one of the aileron (here the right aileron) with a small-amplitude sinusoidal signal. If the
left aileron is stuck, the measures contain a sinusoidal component which is detected with a
selective lter. If the right aileron is stuck, the sinusoidal excitation cannot affect the state
vector and the measures do not contain the sinusoidal components. This point is depicted in
Fig. 13, a fault is simulated on the left aileron next to the right aileron. In the rst case, the
left aileron is stuck at 5 , after the fault has been detected, the right aileron is excited with
a 1 sin (20t) signal. This component distinctly appears in the estimation of the left aileron
position and allows to declare the left aileron faulty. In the second case, the right aileron
is stuck at +5 , after the fault has been detected, the right aileron is excited with the same
sinusoidal signal. As this control surface is stuck, the sinusoidal signal does not appear in
the estimation of the right aileron position and this control surface is declared faulty. Note
that this method allows only to detect stuck ailerons. To deal with a loss of efciency, three
control surfaces are excited: the right aileron, the right and the left aps. The excitation signals
are such that they little affect the state and the measurement vectors. This is achieved by
choosing the amplitudes of these excitations in the nullspace of the (bar , b f r , b f l ) matrix or if
the nullspace does not exist, the excitation vector can be chosen as the right singular vector
corresponding with the smallest singular value of the (bar , b f r , b f l ) matrix. If the right aileron
is faulty, the excitation is not fullled and the measures contain a sinusoidal component. On
the contrary, if the left aileron is faulty, the right aileron and the aps fulll the exctitation
Bank angle ()
40
20
20
0 5 10 15 20 25 30
Airspeed (m/s)
26.5
26
25.5
25
24.5
0 5 10 15 20 25 30
height (m)
200.1
200.05
200
199.95
199.9
0 5 10 15 20 25 30
time (s)
Fig. 4. Right ruddervator loss of efciency: the tracked state variables

Right aileron positions () Left aileron positions ()
3 2
2 1
1 0
0 1
1 2
2 3
0 10 20 30 0 10 20 30
Right ruddervator positions () Left ruddervator positions ()

5
actual actual
control 0 control
estimation estimation
0 2
5 6
10 10
0 10 20 30 0 10 20 30
time (s) time (s)
Fig. 5. Right ruddervator failure: the estimation process

Right aileron residual Left aileron residual
0.3 0.3
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
5 10 15 20 25 30 5 10 15 20 25 30
Right ruddervator residual Left ruddervator residual

0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
5 10 15 20 25 30 5 10 15 20 25 30
time (s) time (s)
Fig. 6. Right ruddervator failure: the fault detection and isolation process
Bank angle ()
50
50
100
150
0 2 4 6 8 10 12 14 16 18
Airspeed (m/s)
26.5
26
25.5
25
24.5
0 2 4 6 8 10 12 14 16 18
height (m)
201
200
199
198
197
0 2 4 6 8 10 12 14 16 18
time (s)
Fig. 7. Right ruddervator stuck: the tracked state variables

4 1
3 0
2 1
1 2
0 3
1 4
10 12 14 16 10 12 14 16

5 0
0 5
actual
control
5
estimation 10 actual
control
estimation
10 15
10 12 14 16 10 12 14 16
time (s) time (s)
Fig. 8. Right ruddervator failure: the estimation process

0.3 0.5
0.25 0.4
0.2
0.3
0.15
0.2
0.1
0.05 0.1
0 0
10 12 14 16 10 12 14 16

1.5 1.5
1 1
0.5 0.5
0 0
10 12 14 16 10 12 14 16
time (s) time (s)
Fig. 9. Right ruddervator failure: the fault detection and isolation process
Bank angle ()
40
20
20
0 2 4 6 8 10 12 14 16 18 20
Airspeed (m/s)
26.5
26
25.5
25
24.5
0 2 4 6 8 10 12 14 16 18 20
height (m)
200.1
200.05
200
199.95
199.9
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 10. Left aileron stuck: the tracked state variables

15 10
10 5
5 0
0 5
5 10
10 15
0 5 10 15 20 0 5 10 15 20

5
0 actual
actual
0 control 2
control
estimation estimation
4
5 6
10 10
0 5 10 15 20 0 5 10 15 20
time (s) time (s)
Fig. 11. Left aileron failure: the estimation process

2 2
1.5 1.5
1 1
0.5 0.5
0 0
14 14.5 15 15.5 16 16.5 14 14.5 15 15.5 16 16.5
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
14 14.5 15 15.5 16 16.5 14 14.5 15 15.5 16 16.5
time (s) time (s)
Fig. 12. Left aileron failure: the fault detection and isolation process
Left aileron positions () Right aileron positions ()

6 6
actual
4 control 4
2 estimation
2
0
0
2
2 actual
4
control
6 4 estimation
14 15 16 17 18 14 15 16 17 18
Left aileron failure, residual Right aileron failure, residual

0.02 0.014
0.012
0.015
0.01
0.008
0.01
0.006
0.004
0.005
0.002
0 0
14 15 16 17 18 14 15 16 17 18
time (s) time (s)
Fig. 13. Left or right aileron stuck: the active diagnosis method
signals, as these latter have no effect on the state vector, the measures do not contain sinusoidal
components.
5. Fault-tolerant control
The faults considered are asymmetric stuck control surfaces. When one or several control
surfaces are stuck, the balance of forces and moments is broken, the UAV moves away from
the fault-free mode operating point and there is a risk of losing the aircraft. This risk is all the
more so critical that it affects the ruddervators, these latter producing the pitch and the yaw
moments.
So a fault may be accommodated only if an operating point exists and the design of the FTC
follows this scheme.
1. It is assumed that the faulty control surface and the fault magnitude are known. This
information is provided by the fault diagnosis system described above.
2. The deection constraints of the remaining control surfaces are released e.g symmetrical
deections for aps, asymmetrical deections for ailerons.
3. For the considered faulty actuator and its fault position, a new operating point is
computed.
4. For this new operating point a linear state feedback controller is designed with an EA
strategy. This controller aims to maintain the aircraft handling qualities at their fault-free
values.
5. The accommodation is achieved by implementing simultaneously the new operating point
and the fault-tolerant controller.
5.1 Operating point computation

The operating point exists if the healthy controls offer sufcient redundancies and its value
depends on:
the considered ight stage,
the faulty control surface,

the fault magnitude.
In the following {Xe , Ue } denote the operating point in faulty mode, Uhe the trim positions
of the remaining controls and Uf the faulty controls. According to (15) and when k control
surfaces are stuck, computing an operating point is equivalent to solve the algebraic equation:
0 = f (Xe ) + gh (Xe )Uhe + g f (Xe )Uf (36)
To take into account the ight stage envelope and the remaining control surface deections,
the operating point computation is achieved with an optimization algorithm. This latter aims
at minimizing the cost function:
J = qV (V Ve0 )2 + q ( e0 )2 + q ( e0 )2 (37)
under the following equality and inequality constraints:

a control surface is stuck
Uf = U f (38)
the ight envelope and the healthy control deections are bounded:

Xmin X Xmax
(39)
Uhmin Uh Uhmax
according to the desired ight stage, some equality constraints are added
ight level
= = V = = = p = r = q = z = 0
(40)
= p=q =r =0
climb or descent with a ight path equal to

= = V = = = p = r = q = 0
(41)
= p = q = r = 0 , =
turn
= = V = = = p = r = q = z = 0
(42)
p = q = 0 , = e , r = re
This strategy aims at keeping the operating point in faulty mode the closest to its fault-free
value. As the linearized model i.e. the state space and the control matrices strongly depends
of the operating point, the open-loop poles (and consequently the open-loop handling
qualities) are little modied.
The computation of an operating point for a faulty ruddervator is described in the sequel.
The right ruddervator is stuck on its whole deection range [20 , +20 ] and the remaining
controls are trimmed in order to maintain the UAV ight level with an airspeed close to 25m/s
and an height equal to 200m. The results of the computation are illustrated in Fig 14. They
show that an operating point exists in the [13 , +3 ] interval. However, for some fault
positions, the actuator positions are close to their saturation positions. This will drastically
limit the aircraft performance. For example, a fault in the 1 position can be compensated
with a throttle trimmed at 90%. It is obvious that this value will limit the turning performance.
Indeed, during the turn, due to the bank angle, the lift force decreases and to keep a constant
height, increasing the throttle control is necessary. As the throttle range is limited, the bank
angle variations will be reduced. This is all the more critical that the aircraft has a lateral
unstable mode. Note that, from now on, there are couplings between the longitudinal and the
lateral axes. Indeed, to obtain these faulty operating points, the longitudinal and the lateral
state variables are coupled e.g. the sideslip angle must differ from zero to achieve a ight level
stage.
For each fault position in the [13 , +3 ] interval, the operating point and the related
linearized model are computed. The root locus is depicted in Fig. 15 and shows that the
open-loop poles are little scattered.
To complete this work, similar studies should be conducted for the left ruddervator, the right
and left ailerons.
1 20
X (%)
()
0.8 0
e
ar
e
0.6 20
0.4
10 5 0 40
10 5 0
er stuck on [13,3] er stuck on [13,3]
20 40
trims in faulty mode
()
()
0 20 trims in faultfree mode

e
e
al
fr
20 0
40 20
10 5 0 10 5 0
stuck on [13,3] stuck on [13,3]
er er
40 10
()
()
20 0
e
e
el
fl
0 10
20 20
10 5 0 10 5 0
stuck on [13,3] stuck on [13,3]
er er
Fig. 14. Right ruddervator stuck, the remaining control trim positions
0.89 0.8 0.7 0.54 0.38 0.18
8
0.95
6
4
0.988
0
20 17.5 15 12.5 10 7.5 5 2.5
2 spiral mode
dutchroll mode
0.988
4
shortperiod mode
6 spiral mode
0.95 phugoid mode
8
propulsion mode
0.89 0.8 0.7 0.54 0.38 0.18
10
20 15 10 5 0
Fig. 15. Right ruddervator stuck, the poless map

5.2 Fault-tolerant controller design

Fault-tolerant control (FTC) strategy has received considerable attention from the control
research community and aeronautical engineering in the last two decades (Steinberg, 2005).
An exhaustive and recent bibliographical review for FTC is presented in (Zhang & Jiang,
2008). Even though different methods use different design strategies, the design goal for
recongurable control is in fact the same. That is, the design objective of recongurable control
is to design a new controller such that post-fault closed-loop system has, in certain sense the
same or similar closed-loop performance to that of the pre-fault system (Zhang & Jiang, 2006).
In this work, the FTC objective consists in keeping the faulty UAV handling qualities identical
to those existing in fault-free mode. Moreover, the tracked outputs (, , V, z) should have
the same dynamics that in fault-free mode. As the computation of the faulty operating point
induced couplings between the longitudinal and the lateral axes, and as each healthy actuator
is driven separately, FTC controllers identical to the nominal controller are kept, i.e. linear state
feedback fault-tolerant controllers which design is based on an EA strategy. This is illustrated
in Tab. 3 where the proposed EA strategy aims at accommodating a right ruddervator failure.
Note that with respect to Tab. 1, the closed-loop poles are unchanged, but the eigenvectors
are modied, particularly to decouple the modes from the faulty control. The design of the
mode short period phugoid throttle roll dutchroll spiral, V , z
CL poles 10 10i 2 2i 1 100 5 5i 1 .25i 1.5 1 .5i
eigenvector
v 1,2
v 3,4
v5
v6

v 7,8
v 9,10
v 11

v 12,13
0
0 0
V 0 0 0 0
0
0 0 0
p 0 0 0 0
q 0 0
r 0 0 0 0
z 0 0

V
z
x 0
ar 0 0 0
al 0 0 0
f r 0 0 0 0 0
f l 0 0 0 0 0
er 0 0 0 0 0 0 0 0
el 0
Table 3. Fault-tolerant controller, EA strategy for a ruddervator failure

FTC is similar to those presented in section 3. Similar studies could be conducted for the other
control surfaces.
5.3 Fault-tolerant controller implementation

A fault is described by the type of control surface and its fault magnitude. This information is
provided by the FDI system studied above. For a given fault, a new operating point and a FTC
controller must be theoretically computed. As far as the operating points are concerned, they
are precomputed, tabulated and selected with respect to the fault. In the same way, a FTC
should be designed for each operating point and its corresponding linearized model. This
method has been adopted to compensate for right ruddervator failures. Practically, it enables
to accommodate for them in the [5 , 0 ] interval with a 1 step study. Consequently, six
fault-tolerant controllers should be designed.
In order to reduce this number and for the six faulty linearized models, a single fault-tolerant
controller is kept, the one which minimizes the scattering of the poles. For a right ruddervator
failure, this fault-tolerant controller is the one designed for a 2 fault position.
Outside this interval, the faults are too severe to be accommodated.
5.4 Results of simulations

Results of simulations are depicted in Fig. 16. The right ruddervator is stuck in the 0 position
at time t f = 16s. This case is similar to the one studied in the paragraph 4.3.2. After time t f ,
the fault is detected, isolated and its magnitude is estimated, then the fault-tolerant controller
efciently compensates for the fault and the aircraft can continue to y and to maneuver.
However, for the reasons explained in subsection 5.1, the bank angle cannot exceed 10 and
the nonlinear effects due to the throttle saturation (see. Fig. 16) affects the dynamics of the
airspeed.
The same fault-tolerant controller is tested for various stuck positions simulated in the
[5 , 0 ] interval and occuring at time t f = 16s. As it is shown in Fig. 18, all these faults
are accommodated with this unique fault-tolerant controller.
Bank angle ()
10
10
0 20 40 60 80 100 120
Airspeed (m/s)
27
26
25
24
0 20 40 60 80 100 120
height (m)
200.5
200
199.5
0 20 40 60 80 100 120
time (s)
Sideslip angle ()
2
0
2
0 20 40 60 80 100 120
Fig. 16. Right ruddervator stuck, the tracked state variables
6. Conclusion
A UAV model has been designed to deal with asymmetrical control surfaces failures that upset
the equilibrium of moments and produce couplings between the longitudinal and the lateral
axes. The nominal controller aims at setting the UAV handling qualities and it is based on an
eigenstructure assignment strategy. Control surface positions are not measured and, in order
to diagnose faults on these actuators, input observability has been studied. It has proven that
faults on the ailerons are not isolable. Next, a bank of Unknown Input Decoupled Functional
Throttle (%)
100
50
0
0 20 40 60 80 100 120
Right and left ailerons ()
right aileron
20
10
left aileron
0
10
0 20 40 60 80 100 120
Right and left flaps ()
right flap
20
10
left flap
0
10
0 20 40 60 80 100 120
Right and left ruddervators () right ruddervator
5
left ruddervator
0
5
10
0 20 40 60 80 100 120
time (s)
Fig. 17. Right ruddervator stuck, the controls

Bank angle ()
20
0
0
reference
20 1
0 5 10 15 20 25 30 35 40 45 50
2
Airspeed (m/s)
27 3
4
26 5
25
24
0 10 20 30 40 50 60 70 80
Height (m)
200.1
200
199.9
0 20 40 60 80 100 120
Sideslip angle ()
2
0
2
12 14 16 18 20 22 24 26
time (s)
Fig. 18. Right ruddervator stuck, the ruddervator actual position, estimation, control position
and the residual
Observers has been implemented in order to detect, isolate and estimate faults. To process
with the faulty ailerons, an isolation method based on a signal processing method has been
presented. Future works should also take into account the redundancies provided by the right
and left aps. The fault accommodation consists in computing a new operating and a related
fault-tolerant controller. For this latter, the objectives of the settings are identical to those
pursued in the fault-free mode. However, the results of simulations show the importance
of actuator saturations, especially in faulty mode, where to compensate for the fault, the
remaining actuator strokes may be signicantly reduced and may affect the control stability.
Our present works deal with FTC designs which aim at setting the handling qualities while
sizing the stability domain with respect to the ight envelope by considering the actuator
saturations.
The rst author would like to acknowledge Pr. T. Hermas for proofreading the initial
manuscript.
7. References
Bateman, F., Noura, H. & Ouladsine, M. (2008a). An active fault tolerant procedure for an uav
equipped with redundant control surfaces, 16th Mediterranean Conference on Control
and Automation, Ajaccio, France.
Bateman, F., Noura, H. & Ouladsine, M. (2008b). A fault tolerant control strategy for an
unmanned aerial vehicle based on a sequential quadratic programming algorithm,
Conference on Decision and Control, Cancun, Mexico.
Bateman, F., Noura, H. & Ouladsine, M. (2009). Fault tolerant control strategy for an
unmanned aerial vehicle, 7th IFAC SafeProcess, Barcelona, Spain.
Boifer, J. (1998). The dynamics of ight, Wiley.
Brigaud, R. (2006). Working towards a usar stanag, Euro UAV, Paris, France.
Demmel, J. & Kgstrom, B. (1993). The generalized schur decomposition of an arbitrary
pencil a - zb: robust software with error bounds and applications, ACM Transaction
on Mathematical Software 19(2): 175201.
Ducard, G. & Geering, H. (2008). Efcient nonlinear actuator fault detection and isolation for
unmanned aerial vehicles, Journal of Guidance, Control and Dynamlics 31(1): 225237.
Hou, M. & Patton, R. (1998). Input observability and input reconstruction, Automatica
34(6): 789794.
Kobayashi, T. & Simon, D. (2003). Application of a bank of kalman lters for aircraft engine
fault diagnostics, Technical Report NASA Report 212526, NASA.
Magni, J. F., Bennami, S. & Terlouw, J. (1997). Robust Flight Control, a design challenge, Springer.
MIL-HDBK-1797 (1997). U.s. military handbook mil-hdbk-1797, Technical report, U.S
Department Of Defense.
Noura, H., Theilliol, D., Ponsart, J. & Chamsedinne, A. (2009). Fault-tolerant Flight Control
Systems, Springer.
OSD (2003). Unmanned aerial vehicle reliability study, Technical report, Ofce of the Secretary
of Defense.
Rauw, M. (1993). A Simulink environment for ight dynamics and control analysis, PhD thesis,
Delft University of Technology, Faculty of Aerospace Engineering.
Steinberg, M. (2005). Historical review of research in recongurable ight control, Journal of
Aerospace Engineering 219(4): 263275.
Xiong, Y. & Saif, M. (2003). Unknown disturbance inputs estimation based on state functional
observer design, Automatica 39: 13901398.
X.Liu, Chen, B. & Lin, Z. (2005). Linear system toolkit in matlab : structural decomposition
and their applications, Journal of Control, theory and application 3: 287294.
Zhang, Y. & Jiang, J. (2006). Issues on integration of fault diagnosis and recongurable control
in active fault-tolerant control systems, IFAC Safe Process, Beijing, China.
Zhang, Y. & Jiang, J. (2008). Bibliographical Review on Recongurable Fault Tolernat Control,
Annual Reviews in Control 32(2): pp. 229252.
8
Fault-Tolerance of a Transport Aircraft

with Adaptive Control and
Optimal Command Allocation
Federico Corraro, Gianfranco Morani and Adolfo Sollazzo
Italian Aerospace Research Centre (CIRA)
Italy
1. Introduction
How to achieve high performance and reliability against various unforeseen events,
uncertainties and other changes in plant dynamics has been a very challenging issue for
control system design in recent years. Reconfigurable flight controls aim to guarantee
greater survivability in all the cases in which the systems to be controlled may be poorly
modelled or the parameters of the systems may be subjected to large variations with respect
to the operating environment. A suitable approach to the problem of flight control
reconfiguration consists in redesigning its own structure and/or re-computing control gains
in the case of unexpected events or large model and environmental uncertainties. A number
of different approaches have been proposed and developed in the past years (Patton, 1997).
In this chapter a Direct Adaptive Model Following (DAMF) algorithm has been used for
reconfiguration purposes. It is possible to find in literature a great amount of proposed
techniques to implement (Bodson & Groszkiewicz, 1997; Calise et al., 2001; Boskovic &
Mehra, 2002; Kim et al., 2003; Tandale & Valasek, 2003). The Lyapunov theory described in
(Kim et al., 2003; Tandale & Valasek, 2003) has very attractive features both in terms of
effectiveness and implementation and it has been used to develop the fault-tolerant scheme
described in this chapter.
Another important matter in flight control reconfiguration is the Control Allocation (CA)
problem. It concerns the possibility to exploit actuators redundancy with respect to the
variables to be controlled in order to redistribute the control effort among the available
control effectors. In this way the control commands needed to attain the desired moments
can be computed even in presence of actuator failures, while also dealing with position and
rate limits of the control effectors. A great amount of techniques for control allocation are
available in literature (Virnig & Bodden, 2000; Enns, 1998; Buffington & Chandler, 1998;
Durham & Bordignon, 1995; Burken et al., 2001). The technique used in this chapter is the
one introduced by Harkegard (Harkegard, 2002) based on active set methods, which is very
effective for real-time applications and converge in a finite number of steps.
Therefore in this chapter a scheme of a fault-tolerant flight control system is proposed. It is
composed by the core control laws, based on the DAMF technique, to achieve both
robustness and reconfiguration capabilities, and the CA system, based on the active set
method, to properly allocate the control effort on the healthy actuators. Numerical results of
a case study with a detailed model of a large transport aircraft are reported to show the
effectiveness of the proposed fault-tolerant control scheme.
The chapter is organized as follows. Sec. 2 explains the proposed flight control system
architecture and all its features. Sec. 3 and 4 report detailed descriptions of each element
composing the Flight Control System (FCS). Sec. 5 contains the most meaningful results of
the numerical evaluation. Finally, Sec. 6 reports the main outcomes and the conclusions
about the current results of the research project.
2. The fault-tolerant flight controller scheme

The logical scheme of the full fault-tolerant flight controller is reported in Fig. 1. The picture
shows four main elements, the autopilot (A/P), the actuators health monitor, the Stability
and Controllability Augmentation System (SCAS), based on the DAMF, and the Control
Allocation module. The last two elements represent the Fault-Tolerant Control System
(FTCS) that is the object of this chapter.
Fig. 1. The first layer scheme of the full Flight Control System
Although adaptive control exhibits great reconfiguration capabilities, in case of in-flight
faults, abrupt and dramatic changes in control effectiveness and/or plant dynamics may
occur, such that the adaptive controller may not be able to recover the vehicle. Therefore, an
adaptive controller could take advantage of a control allocation module to ensure the
generation of the demanded moments by the optimal control system, both in healthy and
faulty conditions of the actuation system.
The remaining two elements are not the focus of this work and they are developed with
classic techniques. In details, the A/P is designed by means of the classic sequential loop
closures, implementing the typical guidance modes for the aircraft (see Table. 1).
Longitudinal Lateral
Altitude Hold/Select Heading Hold/Select
GlideSlope Intercept Localizer Intercept
Approach Lon Approach Lat
Table 1. List of Autopilot modes
Fault-Tolerance of a Transport Aircraft with Adaptive Control and Optimal Command Allocation 161
Also the health monitoring of actuators is a very trivial system based on the comparison
between the input and the output of each actuators. In the numerical validation it is
supposed to use a monitoring system with the capability to detect an actuator fault within
10 seconds, and to pass the binary information healthy/faulty to the control allocation
system. In the following two sections the elements of the FTCS are briefly recalled.
3. Adaptive control system

The core module of the whole flight control system is the SCAS that is in charge of
guaranteeing vehicle attitude control and stability. As already said, the proposed algorithm
for this module has been designed using a Direct Adaptive Model-Following method
(Boskovic & Mehra, 2002; Kim et al., 2003; Tandale & Valasek, 2003), having the advantage
of strong robustness against model parameter uncertainty, and a good capability of reacting
to system parameters variation. Moreover, the model following strategy lets the designer to
define in a clear and simple way the reference dynamics for the system, thus making this
control strategy very attractive among other available robust control techniques. In the
following some recalls about the DAMF are given.
3.1 Theoretical recalls

DAMF is a Model Reference Control Strategies and it earns its robustness properties by
means of a direct adaptation of control loop gains to asymptotically reach two objectives.
The first is a null error between the output of the reference model and the one of real plant.
The second objective is to minimize the control effort. As already said, the proposed
adaptation algorithm is based on the Lyapunov theory. A mathematical description of the
method, fully reported in (Kim et al., 2003), follows. Let us consider the linear model of a
plant:
x = Ax + Bu + d
(1)
y = Cx
with x n the state vector, y l the output vector, u m the control vector, A nxn,
B nxm, C lxn and the term d represents the trim data. The reference system dynamics is
written in term of desired input-output behaviour:
y m = Am y m + Bm r (2)
where ym is the desired output for the plant, r is the given demand, Am and Bm represent the
reference linear system. The control laws structure is defined as follows:
u = C 0 (G0 x + + r + K 0 ym ) (3)
where G0, C0 and v are proper terms generated by the adaptation rules, instead K0 is a feed-
forward gain matrix off-line computed. It is now possible to calculate the error function as
follows:
e = y ym (4)
and to evaluate the error dynamics, in terms of the plant parameters and the reference
system dynamics:
e = (CA + CBC0G0 )x + CBC0 r + CBC0 + CBC0K 0 ym + Cd Am ym Bmr (5)
Assuming a desired error system dynamics, expressed as:
e = Ae e + (6)
where Ae is a stable and properly chosen matrix, and represents a bounded forcing
function, it is possible to write the following identities:
CA + CBC0*G0* = AeC
CBC0* = Bm
(7)
CBC0* * = Cd
CBC0* K0 = Am Ae
Equations 7 allow to write the expressions of the optimal terms G0*, C0*, v* and K0 to obtain a
perfect model inversion that guarantees the asymptotical stability of the closed loop system
and the asymptotical null error.
G0* = Bm1 (AeC CA)

C0* = (CB) 1 Bm
(8)
* = Bm 1Cd
K0 = Bm 1 (Am Ae )
In order to evaluate the left hand terms (the gains of the controller), Equations 8 require
matrix Bm to be invertible and CB matrix to be pseudo-invertible. While the former is a
design parameter, the latter, called high frequency gain, is a structural characteristic of the
plant. Anyway, modern aircrafts have typically a sufficient redundancy order for the control
surfaces, thus ensuring not to lose rank order even in the case of single and often double
actuators failure. Concerning the C matrix, no sensor failure cases are addressed in this
chapter, anyway the device redundancy or several techniques, available in literature (f.i.
Kalman filtering), may ensure a full state feedback, even though each signal may lose
accuracy in case of sensor failure.
It should be anyway noted that the control parameters of Equation 8 do not take into
account the system parameters variation. However, the system parameters uncertainties can
be modelled by a proper variation of the matrices in Equation 1. Finally, a set of adaptation
rules is necessary to react to the system parameters variation and uncertainty, Lyapunov
theory furnishes a very efficient solution. First of all, let us define the differences between
the actual adaptive parameters and the optimal ones:
G = G0 G0*
= C0* 1 C0 1 (9)
= 0 0*
After some manipulations (Kim et al., 2003), here left out for the sake of brevity, it is now
possible to write the real expression of the error dynamics taking into account a parameters
variation:
e = Ae e + Bm G x + Bm u + Bm (10)
It is allowed to impose the Lyapunov stability condition for the error system. So, let us
consider the Lyapunov candidate function:
GT G T T
V = eT Pe + tr + tr + (11)
1 2 3
where P is the matrix solution of Lyapunov equation:
ATe P + PAe = Q (12)
with Q is a positive definite weighting matrix. By calculating the time derivative of the
Lyapunov candidate function and by casting it to get null, the following conditions can be
found, that represent the adaptation rules for the control laws parameters.
G 0 = 1Bm
T
PexT
C = C BT PeuT C
0 2 0 m 0 (13)
T
0 = 3 Bm Pe
Moreover by taking into account the Equations 10, 11 and 13 it is possible to demonstrate
the non-positiveness of Lyapunov candidate function derivative:
V = eT Pe 0 (14)
which assures the asymptotical stability for the error dynamic system.
3.2 Implementation of the SCAS module

The SCAS module is made of two nested sub-modules, taking advantage of the dynamics
separation principle, being the angular rate dynamics sufficiently faster than those of the
attitude ones. This two modules architecture also leads to a relevant reduction of the overall
complexity (in terms of states number) of the adaptive algorithm. The detailed structure of
each Multi Input Multi Output (MIMO) controller is reported in Fig. 2. The design of both
inner and outer loops consists in tuning some parameters. First of all, the matrices Am and
Bm, representing the dynamics of the Reference Model, must be selected with the limitation
that the former must trivially be chosen with negative eigenvalues and the latter must be
chosen invertible. These two matrices actually define the control system performance
requirements. For both attitude and rates regulators, a couple of very simple reference
models made of two diagonal systems (1st order and decoupled systems) have been chosen.
The desired error dynamics are chosen through the matrix Ae by which, it is also possible to
modify the system capability to reject noise and disturbances.
The matrix Q, used in the Equation 12 for the calculation of P, has the meaning of a
weighting matrix. By fine tuning this matrix, it is possible to give more or less relevance to
the tracking requirement of one or more output variables with respect to the others. Finally,
the three parameters 1, 2 and 3(evaluated by means of a trial and error procedure) are used
to regulate the adaptive capability. As a reminder, in Table 2 all the design parameters are
reported.
Fig. 2. SCAS sub-module logical architecture
Am, Bm Ae Q 1, 2, 3
1 0 0 1 0 0 4.5 0 0 0 .8 0 0
Loops
0.06
Inner

0 1 0 , 0 1 0 0 4.5 0 0 1 .2 0 0.1
0 0 4.5 0.1
0 0 1 0 0 1 0 0 1
Loops
Outer
0.01
0.5 0 0.5 0 1.5 0 1 0
0 , 0 0.5 0.01
0.5 0 0.5
0 0.75
0.2
Table 2. SCAS module parameters
4. The control allocation module

As mentioned in the introduction a control allocation algorithm can be very useful for
control reconfiguration purposes due to its ability of managing actuator redundancy, so to
redistribute control effort after a failure event. Moreover, it may be a great support for an
optimal control strategy, such as the DAMF that works well in the case of limited faults (i.e.
the plant dynamics do not change dramatically) and in any case it does not take into account
the limited range and limited rate of the control variables.
4.1 Theoretical background

In this section the control allocation problem is briefly introduced. Given a control vector
up m as computed by the control system (see Equation 3), a desired moment vector
vdes l can be defined as:
vdes = CBup (15)
where the matrix product CB is the high frequency gain of the healthy system (no faults). In
the event of one or more faults, system defined in Equation 1 becomes:
x = Ax + B fault u + Bulock + d
(16)
y = Cx
where Bfault is the control matrix of the failed plant which can be expressed as:
B fault = B (17)
where is a diagonal matrix with the elements i = 0 for i-th actuator failed and i = 1 for a
healthy actuator. This matrix accounts for the fact that failed actuators cannot be used
anymore to change the systems dynamics. The term Bulock accounts for a residual moment
due to an actuator locked in a fixed position, so we set ulock(j) = 0 if j-th actuator is healthy
and ulock(j) = uj if j-th actuator is locked at uj. After this setting, the residual moment to be
attained by the failed system is defined as
v = vdes CBulock (18)
which is the moment to be attained with the failed high frequency gain matrix Bfault.
Therefore the goal of control allocation is to find a control m such that CBfault =v. The
new control vector shall also satisfy the constraints on maximum and minimum values,
which can be computed at each instant depending on the actual position and rate limits, that
is, umin umax.
Generally speaking a solution to the above problem may not exist or it may be not unique
depending on the rank of matrix CBfault. If there are more solutions, the exceeding control
authority can be exploited to choose the solution which is the nearest to a reference control
vector up m (for example the one computed by the control system). A common approach
to solve the control allocation problem is based on the following weighted least square
formulation (Harkegard, 2002):
uw = min
umin u umax
(
Wu u up ) 2
(
+ Wv CB fault u v ) 2
(19)
where 2 is the L2-norm, up is a reference control vector, Wu and Wv are non-singular

weighting matrices. The first term of the above minimization problem allows to choose,
among all the feasible control vectors which minimize the L2-norm of the error CBfaultu-v,
the one minimizing the norm of (u-up). The weighting factor defines the relative degree of
importance between the moments error CBfaultu-v and the control error (u-up). Obviously
should be chosen large enough to ensure the minimization of the error in attaining the
desired moments.
In this chapter we will address the problem of control allocation with the use of a technique
based on the active set method and described in (Harkegard, 2002). Active set methods are
very common in constrained quadratic programming. They only consider active constraints
(equality constraints) and disregard inactive constraints. Therefore active set algorithms
move on the surface defined by the set of the active constraints (named Working Set) to get
an improved solution. The working set is continuously updated during the execution of the
algorithm. In fact, whenever a new constraint is violated, it is added to the current working
set. On the other hand, if a feasible solution has been found by the algorithm, but a
Lagrange multiplier related to the current working set is negative, the corresponding active
constraint is dropped in order to get an improved solution. An exhaustive description of the
control allocation algorithm used in this chapter can be found in (Harkegard, 2002). Some
recalls are given in the section below.
4.2 Control allocation algorithm

For each iteration step of the algorithm the following optimization problem is solved:
min A(p + uk ) b ;
2
p
Wv B Wv v (20)
pi = 0 , i Wk , A = , b =
Wuup
Wu
where uk is the starting solution at the iteration step k, the set Wk is the current working set,
that is, the set containing the active constraints (i.e. saturated controls) which are expressed
through the equality pi=0, while the remaining inequality constraints are disregarded.
Solution to the least square problem of Equation 21 consists of finding the optimal
perturbation p which can be obtained by using a simple pseudo inversion method
(Harkegard, 2002). Once the constraints on actuator limits have been set
Cu U ;
+ I + umax (21)
C = , U =

I umin
(Imxm is the identity matrix), if the solution uk+p is feasible, the Lagrange multipliers i
associated with the active constraints are computed. If they are non-negative, the optimum
solution is obtained otherwise the i-th constraint is dropped from the active set because a
better solution can be found according to the meaning of Lagrange multipliers (Luenberger,
1989). In the case that uk+p is not feasible, the maximum step is calculated such that uk+p
is still feasible and a new constraint is added to the working set. This iterative procedure is
then repeated until a suitable solution and a working set with negative related Lagrange
multiplier are found.
4.3 Remarks
As above described control allocation algorithm has the aim of redistribute the control effort
among the healthy surfaces to achieve the moments needed to keep the system along
reference trajectory. In view of these considerations we argue that control allocation can be
very useful, when used in conjunction with a direct adaptive control in those critical failure
scenarios which can be hardly handled by the only use of the adaptive controller.
Nevertheless, in order to be effective for reconfiguration purposes, control allocation needs a
Fault Detection (FD) system, which gives information about the health of the surfaces
actuators. This aspect could make unfeasible the use of a CA scheme. Anyway, in the
following sections it will be shown that, in order to obtain a satisfactory performance of the
CA module, only limited failure information are needed. In fact, also a very simple
monitoring algorithm, based on the actuator model and on the surface actual position, can
be sufficient to establish whether an actuator is failed or not. The results show that the use of
a CA scheme allows significant improvements of the control system performances also in
the event of very critical failures and it only needs limited information about actuators
health. These features make the proposed control architecture very appealing for
reconfiguration purposes.
5. Numerical validation
The FCS has been applied in a case study with a large transport aircraft. The works has been
performed within the GARTEUR Action Group 16, project focused on Fault-Tolerant
Control. In that project a benchmark environment (Smaili et al., 2006) has been developed
modelling a bunch of surface actuators faulty conditions. A brief summary of all these
conditions is given in Table 3, while a detailed explanation of the benchmark can be found
in (Smaili et al., 2006).
Several manoeuvres are considered in the benchmark to be accomplished in the various
faulty conditions. The test results are here shown both in terms of time histories of the state
variables and with a visual representation of the trajectories performed by the airplane.
Stuck Ailerons:
Both inboard and outboard ailerons are stuck.
Stuck Elevators:
Both inboard and outboard elevators are stuck.
Stabilizer Runaway:
The stabilizer goes at the maximum speed toward
the maximum deflection.
Rudder Runaway:
The upper and lower rudders go at the maximum
speed toward the maximum deflection.
Loss of Vertical Tail:

The vertical tail separates from the aircraft.
Table 3. Failures considered in the test campaign

Only the most meaningful conditions are here reported and discussed. To better
demonstrate the improvement of fault-tolerance achieved by adopting the adaptive control
in conjunction with the Control Allocation, comparison is made between three versions of
the FCS, the first is a baseline SCAS developed with classic control techniques. The two
remaining FCS are based on the adaptive SCAS with and without the CA respectively. As
above said, only limited FD information are supposed to be provided, that is, the
information about whether an actuator is failed or not but the current position of the failed
actuator will be considered as unknown. The CA parameters have been set to:
= 106
Wu = I 3 3 (22)
3 3
Wv = I
5.1 Straight flight with stabilizer failure

In this condition, while in straight and levelled flight, the aircraft experiences a stabilizer
runaway to maximum defection that generates a pitching down moment. The initial flight
condition data are summarized in Table 4.
Altitude True Airspeed Heading Mass Flaps

[m] [m/s] [deg] [kg] [deg]
600 92.6 180 263,000 20
Table 4. Flight condition data
(a) Trajectories
(b) Time plots

Fig. 3. Straight flight with stabilizer runaway with classic technique (dotted line), DAMF
(solid line) and DAMF+CA (dashed line)
Fig. 3 (a) shows the great improvement achieved thanks to the adoption of the control
allocation. Note that the classic technique, for this failure condition, shows adequate
robustness. This is caused by its structure. In fact, the longitudinal control channel (PI for
pitch-angle above proportional pitch-rate SAS) affects only the elevators, while the stabilizer is
supposed to be operated by the pilot separately. In this way, the stabilizer runway results to be
a strong, but manageable disturbance. Instead, the DAMF tries to recover the attitude
lavishing stronger control effort on the faulty stabilizer, the most effective surface, with bad
results. The awareness of the fault on the stabilizer gives the chance to the CA technique to
compensate by moving the control effort from this surface to the elevators, thus achieving the
same results of the classical technique. As it is also evident in the time plots of Fig. 3 (b) when
the failure is detected and isolated (here it is supposed to be done in 10 sec after the failure
occurs), the aircraft recovers a more adequate attitude to carry out properly the manoeuvre.
5.2 Right turn and localizer intercept with rudder runaway

This manoeuvre consists in the interception of the localizer beam, parallel to initial flight
path, but opposite in versus. So, in the early stage of the manoeuvre, a right turn is
performed, and then the capture and the tracking of the localizer beam are carried out. The
fault, instead, consists in a runaway of both upper and lower rudder surfaces, so giving a
strong yawing moment opposite to the desired turn. The initial flight condition data are
summarized in Table 4.
In this failure case, a classical technique is totally inadequate to face such a failure, so
leading the aircraft to crash into the ground. Instead, the DAMF shows to be robust enough
to deal with this failure condition and it makes the aircraft to accomplish the manoeuvre,
even though with reduced performance. The control allocation technique, instead, shows a
sensible improvement of the robustness (see Fig. 4), if compared to the DAMF technique.
The awareness of the fault (detected 10 sec after it actually occurs) allows the control laws to
fully exploit all the efficient effectors, thus accomplishing the manoeuvre smoothly. It is
worth noting that in this case the DAMF without CA is robust enough to accomplish the
manoeuvre, even though with degraded performances.
5.3 Right turn and localizer intercept with loss of vertical tail
The manoeuvre, here considered, is the same described in the previous subsection, but the
failure scenario consists in the loss of the vertical tail (Smaili et al., 2006). The initial flight
condition data are summarized in Table 4. This is both a structural and actuation failure, in
fact, the loss of the rudders strongly affects the lateral-directional aerodynamics and stability,
compromising the possibility to damp the rotations about the roll and yaw axes. In this case
(see Fig. 5), the classical technique is not able to reach lateral stability. Instead, no significant
differences are evidenced between the two versions of the adaptive FCS (with and without
CA). In fact, the information about the efficiency of the differential thrust is already available
to the DAMF, due to the linear model of the bare Aircraft. Thus, as the tracking errors increase,
the core control laws raise the control effort for both the rudders (failed) and the differential
thrust. The latter is efficient enough to ensure the manoeuvrability.
6. Conclusions
In this chapter a fault-tolerant FCS architecture has been proposed. It exploits the main
features of two different techniques, the adaptive control and the control allocation. The
contemporaneous usage of these two techniques, the former for the robustness, and the
latter for the explicit actuators failure treatment, has shown significant improvements in
terms of fault-tolerance if compared to a simple classical controller and to the only adaptive
(a) Trajectories
(b) Time plots

Fig. 4. Right turn and localizer intercept with rudder runaway with classic technique (dotted
line), DAMF (solid line) and DAMF+CA (dashed line)
(a) Trajectories
(b) Time plots

Fig. 5. Loss of vertical tail failure scenario, while performing a right turn & localizer intercept
runaway with classic technique (dotted line), DAMF (solid line) and DAMF+CA (dashed line)
controller. The ability of the DAMF to on-line re-compute the control gains guarantees both
robustness and performance, as shown in the proposed test cases. However, the
contemporary usage of a control allocation scheme allowed improving significantly the
fault-tolerance capabilities, at the only expense of requiring some limited information about
the vehicle actuators health. Therefore the proposed fault-tolerant scheme appears to be
very promising to deal with drastic off-nominal conditions as the ones induced by severe
actuators failure and damages thus improving the overall adaptive capabilities of a
reconfigurable flight control system.
7. References
Bodson, M. & Groszkiewicz, J. E. (1997). Multivariable Adaptive Algorithms for
Reconfigurable Flight Control, IEEE Transaction on Control Systems Technology, Vol.
5, No. 2, pp. 217-229.
Boskovic, J. D. & Mehra, R. K. (2002), Multiple-Model Adaptive Flight Control Scheme for
Accommodation of Actuator Failures, Journal of Guidance, Control, and Dynamics,
Vol. 25, No. 4, pp. 712-724.
Buffington, J. & Chandler, P. (1998), Integration of on-line system identification and
optimization-based control allocation, Proceedings of AIAA Guidance, Navigation, and
Control Conference and Exhibit, Boston, MA.
Burken, J. J., Lu, P., Wu, Z. & Bahm, C. (2001), Two Reconfigurable Flight-Control Design
Methods: Robust Servomechanism and Control Allocation, Journal of Guidance,
Control, and Dynamics, Vol. 24, No. 3, pp. 482-493.
Calise, A. J., Hovakimyan, N. & Idan, M. (2001). Adaptive output feedback control of
nonlinear systems using neural networks, Automatica, Vol. 37, No. 8, pp. 12011211.
Durham, W. C. & Bordignon, K. A. (1995), Closed-Form Solutions to Constrained Control
Allocation Problem, Journal of Guidance, Control and Dynamics, Vol. 18, No. 5, pp.
1000-1007.
Enns, D. (1998), Control Allocation Approaches, Proceedings of the AIAA Guidance, Navigation
and Control Conference, Boston, MA.
Harkegard, O. (2002), Efficient Active Set Algorithms for Solving Constrained Least squares
Problems in Aircraft Control Allocation, Proceedings. of the 41st IEEE Conference on
Decision and Control, Vol. 2, pp. 1295-1300.
Kim, K. S., Lee, K. J. & Kim, Y. (2003), Reconfigurable Flight Control System Design Using Direct
Adaptive Method, Journal of Guidance, Control, and Dynamics, Vol. 26, No. 4, pp. 543-550.
Luenberger, D. G. (1989), Linear and Nonlinear Programming, 2nd ed., Addison-Welsey, 1989,
Chapter 11.
Patton, R. J. (1997). Fault-Tolerant Control Systems: The 1997 Situation, Proceedings of the
IFAC Symposium on Fault Detection, Supervision and Safety for Technical Processes, Vol.
2, pp. 10331055.
Smaili, M. H., Breeman, J., Lombaerts, T. J. & Joosten, D. A. (2006), A Simulation Benchmark
for Integrated Fault Tolerant Flight Control Evaluation, Proceedings of AIAA
Modeling and Simulation Technologies Conference and Exhibit, Keystone, CO.
Tandale, M. & Valasek, J. (2003), Structured Adaptive Model Inversion Control to
Simultaneously Handle Actuator failure and Actuator Saturation, Proceedings. of the
AIAA Guidance, Navigation and Control Conference, Austin, TX.
Virnig, J. & Bodden, D. (2000), Multivariable Control Allocation and Control Law
Conditioning when Control Effector Limit, Proceedings of the AIAA Guidance,
Navigation and Control Conference, Denver, CO.
9
Acceleration-based 3D Flight Control for UAVs:

Strategy and Longitudinal Design
Iain K. Peddle and Thomas Jones
Stellenbosch University
South Africa
1. Introduction
The design of autopilots for conventional flight of UAVs is a mature field of research. Most
of the published design strategies involve linearization about a trim flight condition and the
use of basic steady state kinematic relationships to simplify control law design (Blakelock,
1991);(Bryson, 1994). To ensure stability this class of controllers typically imposes significant
limitations on the aircrafts allowable attitude, velocity and altitude deviations. Although
acceptable for many applications, these limitations do not allow the full potential of most
UAVs to be harnessed. For more demanding UAV applications, it is thus desirable to
develop control laws capable of guiding aircraft though the full 3D flight envelope. Such an
autopilot will be referred to as a manoeuvre autopilot in this chapter.
A number of manoeuvre autopilot design methods exist. Gain scheduling (Leith &.
Leithead, 1999) is commonly employed to extend aircraft velocity and altitude flight
envelopes (Blakelock, 1991), but does not tend to provide an elegant or effective solution for
full 3D manoeuvre control. Dynamic inversion has recently become a popular design
strategy for manoeuvre flight control of UAVs and manned aircraft (Bugajski & Enns, 1992);
(Lane & Stengel, 1998);(Reiner et al., 1996);(Snell et al., 1992) but suffers from two major
drawbacks. The first is controller robustness, a concern explicitly addressed in (Buffington et
al., 1993) and (Reiner et al., 1996), and arises due to the open loop nature of the inversion
and the inherent uncertainty of aircraft dynamics. The second drawback arises from the
slightly Non Minimum Phase (NMP) nature of most aircraft dynamics, which after direct
application of dynamic inversion control, results in not only an impractical controller with
large counterintuitive control signals (Hauser et al., 1992) (Reiner et al., 1996), but also in
undesired internal dynamics whose stability must be investigated explicitly (Slotine & Li,
1991). Although techniques to address the latter drawback have been developed (Al-
Hiddabi & McClamroch, 2002);(Hauser et al., 1992), dynamic inversion is not expected to
provide a very practical solution to the 3D flight control problem and should ideally only be
used in the presence of relatively certain minimum phase dynamics.
Receding Horizon Predictive Control (RHPC) has also been applied to the manoeuvre flight
control problem (Bhattacharya et al., 2002);(Miller & Pachter, 1997);(Pachter et al., 1998), and
similarly to missile control (Kim et al., 1997). Although this strategy is conceptually very
promising the associated computational burden often makes it a practically infeasible
solution for UAVs, particularly for lower cost UAVs with limited processing power.
The manoeuvre autopilot solution presented in this chapter moves away from the more
mainstream methods described above and instead returns to the concept of acceleration
control which has been commonly used in missile applications, and to a limited extent in
aircraft applications, for a number of decades (see (Blakelock, 1991) for a review of the major
results). However, whereas acceleration control has traditionally been used within the
framework of linearised flight control (the aircraft or missile dynamics are linearised,
typically about a straight and level flight condition), the algorithms and mathematics
presented in this chapter extend the fundamental acceleration controller to operate equally
effectively over the entire 3D flight envelope. The result of this extension is that the aircraft
then reduces to a point mass with a steerable acceleration vector from a 3D guidance
perspective. This abstraction which is now valid over the entire flight envelope is the key to
significantly reducing the complexity involved in solving the manoeuvre flight control
problem.
The chapter thus begins by presenting the fundamental ideas behind the design of gross
attitude independent specific acceleration controllers. It then highlights how these inner
loop controllers simplify the design of a manoeuvre autopilot and motivates that they lead
to an elegant, effective and robust solution to the problem. Next, the chapter presents the
detailed design and associated analysis of the acceleration controllers for the case where the
aircraft is constrained to the vertical plane. A number of interesting and useful novel results
regarding aircraft dynamics arise from the aforementioned analysis. The 2D flight envelope
illustrates the feasibility of the control strategy and provides a foundation for development
to the full 3D case.
2. Autopilot design strategy for 3D manoeuvre flight

For most UAV autopilot design purposes, an aircraft is well modelled as a six degree of
freedom rigid body with specific and gravitational forces and their corresponding moments
acting on it. The specific forces typically include aerodynamic and propulsion forces and
arise due to the form and motion of the aircraft itself. On the other hand the gravitational
force is universally applied to all bodies in proportion to their mass, assuming an
equipotential gravitational field. The sum of the specific and gravitational forces determines
the aircrafts total acceleration. It is desirable to be able to control the aircrafts acceleration
as this would leave only simple outer control loops to regulate further kinematic states.
Of the total force vector, only the specific force component is controllable (via the
aerodynamic and propulsion actuators), with the gravitational force component acting as a
well modelled bias on the system. Thus, with a predictable gravitational force component,
control of the total force vector can be achieved through control of the specific force vector.
Modelling the specific force vector as a function of the aircraft states and control inputs is an
involved process that introduces almost all of the uncertainty into the total aircraft model.
Thus, to ensure robust control of the specific force vector a pure feedback control solution is
desirable. Regulation techniques such as dynamic inversion are thus avoided due to the
open loop nature of the inversion and the uncertainty associated with the specific force
model.
Considering the specific force vector in more detail, the following important observation is
made from an autopilot design simplification point of view. Unlike the gravitational force
vector which remains inertially aligned, the components that make up the specific force
vector tend to remain aircraft aligned. This alignment occurs because the specific forces arise
Acceleration-based 3D Flight Control for UAVs: Strategy and Longitudinal Design 175
as a result of the form and motion of the aircraft itself. For example, the aircrafts thrust
vector acts along the same aircraft fixed action line at all times while the lift vector tends to
remain close to perpendicular to the wing depending on the specific angle of attack. The
observation is thus that the coordinates of the specific force vector in a body fixed axis
system are independent of the gross attitude of the aircraft. This observation is important
because it suggests that if gross attitude independent measurements of the specific force
vectors body axes coordinates were available, then a feedback based control system could
be designed to regulate the specific force vector independently of the aircrafts gross
attitude. Of course, appropriately mounted accelerometers provide just this measurement,
normalized to the aircrafts mass, thus practically enabling the control strategy through
specific acceleration instead.
With gross attitude independent specific acceleration controllers in place, the remainder of a
full 3D flight autopilot design is greatly simplified. From a guidance perspective the aircraft
reduces to a point mass with a fully steerable acceleration vector. Due to the acceleration
interface, the guidance dynamics will be purely kinematic and the only uncertainty present
will be that associated with gravitational acceleration. The highly certain nature of the
guidance dynamics thus allows among others, techniques such as dynamic inversion and
RHPC to be effectively implemented at a guidance level. In addition to the associated
autopilot simplifications, acceleration based control also provides for a robust autopilot
solution. All aircraft specific uncertainty remains encapsulated behind a wall of high
bandwidth specific acceleration controllers. Furthermore, high bandwidth specific
acceleration controllers would be capable of providing fast disturbance rejection at an
acceleration level, allowing action to be taken before the disturbances manifest themselves
into position, velocity and attitude errors.
With the novel control strategy and its associated benefits conceptually introduced the
remainder of this chapter focuses on the detailed development of the inner loop specific
acceleration controllers for the case where the aircrafts motion is constrained to the 2D
vertical plane. No attention will be given to outer guidance level controllers in the
knowledge that control at this level is simplified enormously by the inner loop controllers.
The detailed design of the remaining specific acceleration controllers to complete the set of
inner loop controllers for full 3D flight are presented in (Peddle, 2008).
3. Modelling
To take advantage of the potential of regulating the specific acceleration independently of
the aircrafts gross attitude requires writing the equations of motion in a form that provides
an appropriate mathematical hold on the problem. Conceptually, the motion of the aircraft
needs to be split into the motion of a reference frame relative to inertial space (to capture the
gross attitude and position of the aircraft) and the superimposed rotational motion of the
aircraft relative to the reference frame. With this mathematical split, it is expected that the
specific acceleration coordinates in the reference and body frames will remain independent
of the attitude of the reference frame. An obvious and appropriate choice for the reference
frame is the commonly used wind axis system (axial unit vector coincides with the velocity
vector). Making use of this axis system, the equations of motion are presented in the desired
form below. The dynamics are split into the point mass kinematics (motion of the wind axis
system through space),
= (C + g cos ) V
(1)
W W W

V = AW g sin W (2)
PN = V cos W (3)
PD = V sin W (4)
and the rigid body rotational dynamics (attitude of the body axis system relative to the wind
axis system),
Q = M I yy (5)
= Q + (C W + g cos W ) V (6)
with, W the flight path angle, V the velocity magnitude, PN and PD the north and down
positions, g the gravitational acceleration, Q the pitch rate, M the pitching moment, I yy the
pitch moment of inertia, the angle of attack and AW and C W the axial and normal specific
acceleration coordinates in wind axes respectively. Note that the point mass kinematics
describe the aircrafts position, velocity magnitude and gross attitude over time, while the
rigid body rotational dynamics describes the attitude of the body axis system with respect to
the wind axis system (through the angle of attack) as well as how the torques on the aircraft
affect this relative attitude. It must be highlighted that the particular form of the equations of
motion presented above is in fact readily available in the literature (Etkin, 1972), albeit not
appropriately rearranged. However, presenting this particular form within the context of the
proposed manoeuvre autopilot architecture and with the appropriate rearrangements will be
seen to provide a novel perspective on the form that explicitly highlights the manoeuvre
autopilot design concepts. Expanding now the specific acceleration terms with a commonly
used pre-stall flight aircraft specific force and moment model yields,
AW = (T cos D ) m (7)
C W = (T sin + L ) m (8)
with,
T = T T + TC T (9)
L = qSC L (10)
D = qSC D (11)
M = qSC m (12)
where m is the aircrafts mass, T the thrust time constant, S the area of the wing, C L , C D
and C m the lift, drag and pitching moment coefficients respectively and,
q = V 2 2 (13)
the dynamic pressure. Expansion of the aerodynamic coefficients for pre-stall flight (Etkin &
Reid, 1995) yields,
( )
C L = C L0 + C L + C LQ c 2V Q + C L E
E
(14)
C D = C D0 + C L2 Ae (15)
( )
C m = C m0 + C m + C mQ c 2V Q + C m E
E
(16)
where A is the aspect ratio, e the Oswald efficiency factor and standard non-dimensional
stability derivative notation is used. Note that it is assumed in this chapter that the non-
dimensional stability derivatives above are independent of both the point mass and rigid
body rotational dynamics states. Although in reality the derivatives do change somewhat
with the system states, for many UAVs operating under pre-stall flight conditions this
change is small. Furthermore, with the intention being to design a feedback based control
system to regulate specific acceleration, the adverse effect of the modelling errors will be
greatly reduced thus further justifying the assumption.
g
m I yy W
T AW Point mass kinematics V

Rigid Body Rotational Dynamics
E CW PN
[ Q]
T T
W V PN PD
PD
V
g cos W
g cos W
= f ( PD )
Fig. 1. Split between the rigid body rotational dynamics and the point mass kinematics.
Figure 1 provides a graphical overview of the particular form of the dynamics presented
here. The dash-dotted vertical line in the figure highlights a natural split in the aircraft
dynamics into the aircraft dependent rigid body rotational dynamics on the left and the
aircraft independent point mass kinematics on the right. It is seen that all of the aircraft
specific uncertainty resides within the rigid body rotational dynamics, with gravitational
acceleration being the only inherent uncertainty in the point mass kinematics. Of course left
unchecked, the aircraft specific uncertainty in the rigid body rotational dynamics would
leak into the point mass kinematics via the axial and normal specific acceleration, thus
motivating the design of feedback based specific acceleration controllers.
Continuing to analyze Figure 1, the point mass kinematics are seen to link back into the
rigid body rotational dynamics via the velocity magnitude, air density (altitude) and flight
path angle. If it can be shown that the aforementioned couplings do not strongly influence
the rigid body rotational dynamics, then the rigid body rotational dynamics would become
completely independent of the point mass kinematics and thus the gross attitude of the
aircraft. This in turn would provide the mathematical platform for the design of gross
attitude independent specific acceleration controllers.
Investigating the feedback couplings, the velocity magnitude and air density couple into the
rigid body rotational dynamics primarily through the dynamic pressure, which is seen in
equations (10) through (12) to scale the magnitude of the aerodynamic forces and moments.
However, the dynamics of most aircraft are such that the angle of attack and pitch rate
dynamics operate on a timescale much faster than that of the velocity magnitude and air
density dynamics. Thus, assuming that a timescale separation either exists or can be
enforced through feedback control, the dynamic coupling is reduced to a static dependence
where the velocity magnitude and air density are treated as parameters in the rigid body
rotational dynamics.
The flight path angle is seen to couple only into the angle of attack dynamics, via
gravitational acceleration. The flight path angle coupling term in equation (6) represents the
tendency of the wind axis system to rotate under the influence of the component of
gravitational acceleration normal to the velocity vector. The rotation has the effect of
changing the relative attitude of the body and wind axis system as modelled by the angle of
attack dynamics. However, the normal specific acceleration (shown in parenthesis next to
the gravity term in equation (6)) will typically be commanded by an outer loop guidance
controller to cancel the gravity term and then further to steer the aircraft as desired in
inertial space. Thus, the effect of the flight path coupling on the angle of attack dynamics is
expected to be small. However, to fully negate this coupling, it will be assumed that a
dynamic inversion control law can be designed to reject it, the details of which will be
discussed in a following section. Note however, that dynamic inversion will only be used to
reject the arguably weak flight path angle coupling, with the remainder of the control
solution to be purely feedback based.
With the above timescale separation and dynamic inversion assumptions in place, the rigid
body rotational dynamics become completely independent of the point mass kinematics and
thus provide the mathematical platform for the design of gross attitude independent specific
acceleration controllers. With all aircraft specific uncertainty encapsulated within the inner
loop specific acceleration controllers and disturbance rejection occurring at an acceleration
level, the design is argued to provide a robust solution to the manoeuvre flight control
problem. The remainder of this article focuses of the design and simulation of the axial and
normal specific acceleration controllers, as well as the associated conditions for their
implementation.
4. Decoupling the axial and normal dynamics

Control of the axial and normal specific acceleration is dramatically simplified if the rigid
body rotational dynamics can be decoupled into axial and normal dynamics. This
decoupling would allow the axial and normal specific acceleration controllers to be
designed independently. To this end, consider equations (7) and (8), and notice that for
small angles of attack and typical lift to drag ratios, the equations can be well approximated
as follows,
AW (T D ) m (17)
CW L m (18)
With these simplifying assumptions, the thrust no longer couples into the normal dynamics
whose states and controls include the angle of attack, pitch rate and elevator deflection. On
the other hand, the normal dynamics states still drive into the axial dynamics through the
drag coupling of equation (17). However, through proper use of the bandwidth-limited
thrust actuator the drag coupling can be rejected up to some particular frequency. Assuming
that effective low frequency disturbance rejection can be achieved up to the open loop
bandwidth of the thrust actuator, then only drag disturbance frequencies beyond this
remain of concern from a coupling point of view.
Considering now the point mass kinematics, it is clear from equation (2) that the axial
specific acceleration drives solely into the velocity magnitude dynamics. Thus
uncompensated high frequency drag disturbances will result in velocity magnitude
disturbances which in turn will couple back into the rest of the rigid body rotational
dynamics both kinematically and through the dynamic pressure. However, the natural
integration process of the velocity magnitude dynamics will filter the high frequency part of
the drag coupling. Thus, given acceptable deviations in the velocity magnitude, the thrust
actuator need only reject enough of the low frequency portion of the drag disturbance for its
total effect on the velocity magnitude to be acceptable. By acceptable it is meant that the
velocity magnitude perturbations are small enough to result in a negligible coupling back
into the rigid body rotational dynamics.
To obtain a mathematical hold on the above arguments, consider the closed loop transfer
function from the normalized drag input to the axial specific acceleration output,
AW (s)
SD ( s) (19)
D( s) m
Through proper control system design, the gain of the sensitivity transfer function above
can be kept below a certain threshold within the controller bandwidth. The bandwidth of
the axial specific acceleration controller will however typically be limited to that of the
thrust actuator for saturation reasons. For frequencies above the controller bandwidth, the
sensitivity transfer function will display some form of transient and then settle to unity gain.
Considering the velocity magnitude dynamics of equation (2), the total transfer function of
the normalized drag input to velocity magnitude is then,
V (s ) S (s )
= D (20)
D(s ) m s
Note that the integrator introduced by the natural velocity dynamics will result in
diminishing high frequency gains. Equation (20) can be used to determine whether drag
perturbations will result in acceptable velocity magnitude perturbations. Conversely, given

the expected drag perturbation spectrum and the acceptable level of velocity magnitude
perturbations, the specifications of the sensitivity transfer function can be determined. To
ease the process of determining acceptable levels of velocity magnitude perturbations and
expected levels of drag perturbations, it is convenient to write these both in terms of normal
specific acceleration. The return disturbance in normal specific acceleration due to a normal
specific acceleration perturbation can then be used to specify acceptable coupling levels.
Relating first normalized drag to normal specific acceleration, use of equation (18) provides
the following result,
Dm 1
= (21)
CW RLD
where use has been made of the fact that lift is related to drag through the lift to drag ratio
RLD . Then, equation (18) can be used to capture the dominant relationship between velocity
perturbations and the resulting normal specific acceleration perturbations. Partially
differentiating equation (18) with respect to the velocity magnitude yields the desired result,
C W qSC L VSC L C
= =2 W (22)
V V m m V
Combining equations (20) to (22) yields the return disturbance sensitivity function,
C W V (s ) D m
SCW ( s)
V D(s ) m C W
(23)
C 1
= 2 W SD (s )
VRLD s
Given an acceptable return disturbance level, the specifications of the sensitivity function of
equation (19) can be determined for a particular flight condition. With the velocity
magnitude and lift to drag ratio forming part of the denominator of equation (23), the
resulting constraints on the sensitivity function are mild for low operating values of normal
specific acceleration. Only during very high acceleration manoeuvres, does the sensitivity
specification become more difficult to practically realize. This is illustrated through an
example calculation at the end of section 5.
Given the above arguments, the drag coupling into the axial specific acceleration dynamics
can be ignored if the associated sensitivity function constraint is adhered to when designing
the axial specific acceleration controller. With the coupling of the drag term ignored, the
axial dynamics become independent of the normal dynamics allowing the controllers to be
designed separately. Furthermore, note that the axial dynamics also become independent of
the velocity magnitude and the air density. Thus, unlike the normal specific acceleration
controller, there is no need for the axial specific acceleration controller to operate on a
timescale much faster than these variables. This greatly improves the practical viability of
designing an axial specific acceleration control system since most thrust actuators are
significantly bandwidth limited. The axial and normal dynamics can thus be decoupled as
follows,
Axial Dynamics:
T = 1 T T + 1 T TC (24)
AW = 1 m T + D m (25)
Normal Dynamics:
L LQ L E g cos W L
1 0
mV

mV mV V mV
=
+ + (26)
MQ Q M E E M0
Q M
I
I yy I yy I
yy yy
L LQ L E L0
CW = + E + (27)
m m Q m m
where dimensional stability and control derivative notation has been used to remove clutter.
Finally, notice that the normal dynamics are simply the classical short period mode
approximation (Etkin & Reid, 1995) but have been shown here to be valid for all point mass
kinematics states (i.e. all gross attitudes) with the flight path angle coupling term acting as a
disturbance input. Intuitively this makes sense since the physical phenomena that manifest
themselves into what is classically referred to as the short period mode are not dependent
on the gross attitude of the aircraft. Whether an aircraft is flying straight and level, inverted
or climbing steeply, its short period motion remains unchanged.
5. Axial specific acceleration controller

In this section a controller capable of regulating the axial specific acceleration is designed.
Attention will also be given to the closed loop sensitivity function constraint of equation (23)
for a specific return disturbance level. With reference to the axial dynamics of equations (24)
and (25) define the following Proportional-Integrator (PI) control law with enough degrees
of freedom to allow for arbitrary closed loop pole placement,
Tc = K A AW K EEA (28)
E A = AW AWR (29)
where AWR is the reference axial specific acceleration command. The integrator in the
controller is essential for robustness towards uncertain steady state drag and thrust actuator
offsets. It is straightforward to show that given the desired closed loop characteristic
equation,
c ( s ) = s 2 + 1s + 0 (30)
the feedback gains that will fix the closed loop poles are,
K A = m ( T 1 1 ) (31)
K E = m T 0 (32)
These simple, closed form solution gains will ensure an invariant closed loop axial specific
acceleration dynamic response as desired. The controller design freedom is reduced to that
of selecting appropriate closed loop poles bearing in mind factors such as actuator
saturation and the sensitivity function constraint of the previous section. Investigation of the
closed loop sensitivity function for this particular control law yields the following result,
s Ts + 1 0
(33)
SD ( s) =
T 0 1 s 2 + 1s + 0
For actuator saturation reasons the closed loop axial dynamics bandwidth is typically
limited to being close to that of the open loop thrust actuator and thus for reasonable closed
loop damping ratios the second order term in parenthesis above can be well approximated
by a first order model to simplify the sensitivity function as follows,
s 1
SD ( s) (34)
T 0 A s + 1
Here, A = 1 ( T 0 ) is the approximating time constant calculated to match the high

frequency sensitivity function asymptotes. Substituting equation (34) into equation (23)
yields the return disturbance transfer function,
CW 1
SCW ( s) = 2 (35)
VRLD T 0 As + 1
Given the maximum allowable gain of the return disturbance transfer function , a lower
bound constraint on the natural frequency ( n ) of the closed loop axial control system is
calculated by satisfying the inequality
n 2C
W T (36)
T VRLD max
where, T = 1 T is the open loop bandwidth of the thrust actuator and the subscript max
denotes the maximum value of the term in parenthesis. The following example illustrates
the practical feasibility of adhering to the sensitivity function constraint. Consider a UAV
that is to fly with a minimum velocity of 20 m/s, with a maximum normal specific
acceleration of 4 g and a minimum lift to drag ratio of 10. Then, for more than 20 dB of
return disturbance rejection, the natural frequency of the closed loop system should have
the following relationship to the open loop thrust bandwidth,
n T 2 T (37)
For this specific example, thrust actuators with a bandwidth of below 4 rad/s (time constant
of greater than 0.25 s) will require that the closed loop natural frequency is greater than that
of the thrust actuator. Despite the fairly extreme nature of this example (low velocity
magnitude, high acceleration and low lift to drag ratio), thrust time constants on the order
of 0.25 s are still practically feasible for UAVs. The deduction is thus that the axial specific
acceleration controller will be practically applicable to most UAVs.
6. Normal specific acceleration controller

This section presents the design and associated analysis of a closed form normal specific
acceleration controller that yields an invariant dynamic response for all point mass
kinematics states. The design is based on the linear normal dynamics of equations (26) and
(27) with the velocity magnitude and air density considered parameters and the flight path
angle coupling rejected using dynamic inversion. To consider the velocity magnitude and
air density as parameters requires a timescale separation to exist between these two
quantities and the normal dynamics. Therefore, it is important to investigate any upper
limits on the allowable bandwidth of the normal dynamics as this will in turn clamp the
upper bandwidth of the velocity magnitude and the air density (altitude) dynamics.
Furthermore, it is important to investigate the eligibility of the normal dynamics for
effective dynamic inversion of the flight path angle coupling term. Thus, before continuing
with the normal specific acceleration controller design, the natural normal dynamics are
analyzed in detail.
6.1 Natural normal specific acceleration dynamics

Consider the dynamics from the elevator control input to the normal specific acceleration
output. The direct feed-through term in the normal specific acceleration output implies that
the associated transfer function has as many zeros as it does poles. The approximate
characteristic equation for the poles is easily be shown to be,
L M L MQ M
p( s) = s 2 + Q s + (38)
mV I yy mV I yy I yy

where use has been made of the commonly used simplifying assumption (Etkin & Reid,
1995),
LQ mV 1 (39)
Considering equation (38) it is important to note that the normal dynamics poles are not
influenced by the lift due to pitch rate or elevator deflection. The importance of this will be
made clear later on in this section. The zeros from the elevator input to the normal specific
acceleration output can be shown, after some manipulation, to be well approximated by the
roots of the characteristic equation,
s 2 LQ ( lT lD ) I yy s L ( lT lN ) I yy = 0 (40)
with the following characteristic lengths defined,
lN M L (41)
lT M E L E (42)
lD MQ LQ (43)
where, lN is the length to the neutral point, lT is the effective length to the tail-plane and lD
is the effective damping arm length. Note that only the simplifying assumption of equation
(39) has been used in obtaining the novel characteristic equation for the zeros above.
Completing the square to find the roots of equation (40) gives,
2 2
s LQ ( lT lD ) 2 I yy = LQ ( lT lD ) 2 I yy + L ( lT lN ) I yy (44)

For most aircraft the effective length to the tail-plane and effective damping arm lengths are
very similar. This is because most of the damping arises from the tail-plane which is also
typically home to the elevator control surface. Thus the moment arm lengths for pitch rate
and elevator deflection induced forces are very similar. As a result, the first term on the
right hand side of equation (44) is most often negligibly small and to a good approximation,
the zeros from elevator to normal specific acceleration are,
z1,2 LQ ( lT lD ) 2 I yy L ( lT lN ) I yy (45)
Analysis of equation (45) reveals that the only significant effect of the lift due to pitch rate
derivative on the zeros is that of producing an offset along the real axis. As previously
argued, the effective tail-plane and damping arm lengths are typically very similar and as a
result, even this effect is usually small. Thus, it can be seen that to a good approximation,
the lift due to pitch rate plays no role in determining the elevator to normal specific
acceleration dynamics.
On the other hand, the effective length to the tail-plane and the length to the neutral point
typically differ significantly. With this difference scaled by the lift due to angle of attack
(which is usually far greater than the lift due to pitch rate) it can be seen that the second
term in equation (45) will dominate the first in determining the zero positions. Thus al-
though the lift due to elevator deflection played no role in determining the system poles, it
plays a large role in deter-mining the zeros. Knowing the position of the zeros is important
from a controller design point of view because not only do they affect the dynamic response
of the system but they also impose controller independent limitations on the systems
practically achievable dynamic response. These limitations are mathematically described by
Bodes sensitivity and complementary sensitivity integrals as discussed in (Freudenberg &
Looze, 1985); (Goodwin et al., 2001).
Expanding on the above point, it is noted that for most aircraft the effective length to the
tail-plane is far greater than the length to the neutral point and so the zeros are real and of
opposite sign. The result, as intuitively expected, is that the dynamics from the elevator to
normal specific acceleration are NMP since a Right Half Plane (RHP) zero exists. A RHP
zero places severe, controller independent restrictions on the practically attainable upper
bandwidth of the closed loop normal specific acceleration dynamics. Furthermore,
designing a dynamic inversion control law in a system with NMP dynamics, particularly
when the NMP nature of the system is weak, tends to lead to an impractical solution with
internal dynamics that may or may not be stable (Hauser et al., 1992); (Hough, 2007).
Since this NMP dynamics case is by far the most common for aircraft, the limits imposed by it
shall be investigated further in the following subsection. The goal of the investigation is to seek
a set of conditions under which the effects of a RHP zero become negligible, equivalently
allowing the NMP nature of a system to be ignored. With these conditions identified and
satisfied, the design of the normal specific acceleration controller can continue based on a set of
simplified dynamics that do not capture the NMP nature of the system.
6.2 Analysis of the NMP dynamics case

Ignoring the typically negligible real axis offset term in equation (45), the transfer function
from the elevator to normal specific acceleration is of the form,
n2 z0 s
G( s ) = k (46)
s + 2n s + n2 z0
2
where z0 is the RHP zero position. Note that the effect of the left half plane zero has been
neglected because it is most often largely negated through pole-zero cancellation by un-
modelled dynamics such as those introduced through servo lag. The transfer function of
equation (46) can be written as follows,
G( s ) = Gn (s ) Gn ( s ) s z0 (47)
where,
n2
Gn (s ) = k (48)
s + 2ns + n2
2
is a nominal second order system with no zeros. Equation (48) makes it clear that as the
position of the zero tends towards infinity, so the total system transfer function converges
towards Gn (s ) . The purpose of the analysis to follow is to investigate more precisely, the
conditions under which G(s ) can be well approximated by Gn (s ) . To this end, a time
response analysis method is employed. Consider the Laplace transform of the systems step
response,
Y (s ) = Gn (s ) s Gn (s ) z0 (49)
Equation (49) makes it clear that the total step response is the nominal system step response
less the impulse response of the system scaled by the inverse of the RHP zero frequency.
Since the nominal response gradient is always zero at the time of the step, the system must
exhibit undershoot. The level of undershoot will depend of the damping, speed of response
of the system and the zero frequency. If the level of undershoot is small relative to unity
then it is equivalent to saying that the second term of equation (47) has a negligible effect.
Thus, by investigating the undershoot further, conditions can be developed under which the
total system response is well approximated by the nominal system response.
A closed form solution for the exact level of undershoot experienced in response to a step
command for a system of the form presented in equation (46) is provided below,
y min = k 1 ( sin sin ) e

( ) tan
(50)

where,
= cos 1 ( ) (51)
= tan 1 ( 1 2 ( + r ) ) (52)
r = n z0 (53)
Derivation of this novel result involves inverse Laplace transforming equation (49) and
finding the time response minima through calculus. The above equations make it clear that
the undershoot is only a function of the ratio between the systems natural frequency and
the zero frequency ( r ) and the systems damping ratio ( ). Figure 2 below provides a plot
of the maximum percentage undershoot as a function of r 1 for various damping ratios.
100
= 0.7
90 = 0.5
= 0.2
Maximum percentage undershoot
80
70
60
50
40
30
20
10
0
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
RHP zero frequency normalised to the natural frequency
Fig. 2. Maximum undershoot of a 2nd order system as a function of normalized RHP zero
frequency for various damping ratios.
It is clear from Figure 2 that for low percentage undershoots, the damping ratio has little
influence. Thus, the primary factor determining the level of undershoot is the ratio of the
systems natural frequency to that of the zero frequency. Furthermore, it is clear that for less
than 5% maximum undershoot, the natural frequency should be at least three times lower
than that of the zero. With only 5% undershoot the response of the total system will be well
approximated by the response of the nominal system with no zero. Thus by making use of
the maximum undershoot as a measure of the NMP nature of a system the following novel
frequency domain design rule is developed,
n < z0 3 (54)
for the NMP nature of the system to be considered negligible. This rule implies that the
system poles must lie within a circle of radius z0 3 in the s-plane. Thus, an upper bound is
placed on the natural frequency of the system if its NMP nature is to be ignored.
6.3 Frequency bounds on the normal specific acceleration controller

Given the results of the previous two subsections, the upper bound on the natural frequency
of the normal specific acceleration controller becomes,
n < L ( lT lN ) I yy 3 (55)
where the typically negligible offset in the zero positions in equation (45) has been ignored.
Adhering to this upper bound will allow the NMP nature of the system to be ignored and
will thus ensure both practically feasible dynamic inversion of the flight path angle coupling
and no large sensitivity function peaks (Goodwin et al., 2001) in the closed loop system.
Note that given the physical meaning of the characteristic lengths defined in equations (41)
through (43), the approximate zero positions and thus upper frequency bound can easily be
determined by hand for a specific aircraft.
It is important to note that the upper bound applies to both the open loop and closed loop
normal specific acceleration dynamics. If the open loop poles violate the condition of
equation (55) then moving them through control application to within the acceptable
frequency region will require taking into account the effect of the system zeros. Thus, for an
aircraft to be eligible for the normal specific acceleration controller of the next subsection, its
open loop normal dynamics poles must at least satisfy the bound of equation (55). If they do
not then an aircraft specific normal specific acceleration controller would have to be
designed. However, most aircraft tend to satisfy this bound in the open loop because open
loop poles outside the frequency bound of equation (55) would yield an aircraft with poor
natural flying qualities i.e. the aircraft would be too statically stable and display significant
undershoot and lag when performing elevator based manoeuvres. Interestingly, the
frequency bound can thus also be utilized as a design rule for determining the most forward
centre of mass position of an aircraft for good handling qualities.
In term of lower bounds, the normal dynamics must be timescale separated from the
velocity magnitude and air density (altitude) dynamics. Of these two signals, the velocity
magnitude typically has the highest bandwidth and is thus considered the limiting factor.
Given the desired velocity magnitude bandwidth (where it is assumed here that the given
bandwidth is achievable with the available axial actuator), then as a practical design rule the
normal dynamics bandwidth should be at least five times greater than this for sufficient
timescale separation. Note that unlike in the upper bound case, only the closed loop poles
need satisfy the lower bound constraint. However, if the open loop poles are particularly
slow, then it will require a large amount of control effort to meet the lower bound constraint
in the closed loop. This may result in actuator saturation and thus a practically infeasible
controller. However, for typical aircraft parameters the open loop poles tend to already
satisfy the timescale separation lower bound.
With the timescale separation lower bound and the NMP zero upper bound, the natural
frequency of the normal specific acceleration controller is constrained to lying within a
circular band in the s-plane as shown in Figure 3 (poles would obviously not be selected in
the RHP for stability reasons). The width of the circular band in Figure 3 is an indication of
the eligibility of a particular airframe for the application of the normal specific acceleration
controller to be designed in the following subsection.
For most aircraft this band is acceptably wide and the control system to be presented can be
directly applied. For less conventional aircraft, the band can become very narrow and the
two constraint boundaries may even cross. In this case, the generic control system to be
presented cannot be directly applied. One solution to this problem is to design an aircraft
specific normal specific acceleration controller. However, this solution is typically not
desirable since the closeness of the bounds suggests that the desired performance of the
particular airframe will not easily be achieved practically. Instead, redesign of the airframe
and/or reconsideration of the outer loop performance bandwidths will constitute a more
practical solution.
Im(s)
NMP upper bound s-plane
Re(s)
Feasible pole Timescale separation

placement region lower bound
Fig. 3. NMP upper bound and timescale separation lower bound outlining feasible pole
placement region.
6.4 Normal specific acceleration controller design

Assuming that the frequency bounds of the previous section are met, the design of a
practically feasible normal specific acceleration controller can proceed based on the
following reduced normal dynamics,
L g cos W L
1 0 0
mV
+ M V mV
= MQ Q E E + M0
(56)
Q M I

I I yy yy I
yy yy
L L
CW = 0 + 0 E + 0 (57)
m Q
m
The simplifications in the dynamics above arise from the analysis of subsection 6.1 where it
was shown that to a good approximation, the lift due to pitch rate and elevator deflection
only play a role in determining the zeros from elevator to normal specific acceleration.
Under the assumption that the upper bound of equation (55) is satisfied, the zeros
effectively move to infinity and correspondingly these two terms become zero. Thus, the
simplified normal dynamics above will yield identical approximated poles to those of
equation (38), but will display no finite zeros from elevator to normal specific acceleration.
To dynamically invert the effect of the flight path angle coupling on the normal specific
acceleration dynamics requires differentiating the output of interest until the control input
appears in the same equation. The control can then be used to directly cancel the
undesirable terms. Differentiating the normal specific acceleration output of equation (57)
once with respect to time yields,
L L L g cos W
C W = C W + Q + (58)
mV m mV
where the angle of attack dynamics of equation (6) have been used in the result above.
Differentiating the normal specific acceleration a second time gives,

= MQ L C + M + L MQ C
C W W W
I yy mV I yy mVI yy
(59)
L M L M L M L g MQ
sin
+ E
E + 0 0 + cos W +

mI yy mV I yy
W W

mI yy mI yy
where use has been made of equations (56) to (58) in obtaining the result above. The elevator
control input could now be used to cancel the effect of the flight path angle coupling terms
on the normal specific acceleration dynamics. However, the output feedback control law to
be implemented will make use of pitch rate feedback. Upon analysis of equation (6), it is
clear that pitch rate feedback will reintroduce flight path angle coupling terms into the
normal specific acceleration dynamics. Thus, the feedback control law is first defined and
substituted into the dynamics, and then the dynamic inversion is carried out. A PI control
law with enough degrees of freedom to place the closed loop poles arbitrarily and allow for
dynamic inversion (through EDI ) is defined below,
E = KQQ KCC W K EEC + E DI

(60)
E C = C W C WR (61)
with C WR the reference normal specific acceleration command. The integral action of the
control law is introduced to ensure that the normal specific acceleration is robustly tracked
with zero steady state error. Offset disturbance terms such as those due to static lift and
pitching moment can thus be ignored in the design to follow. It is best to remove the effect
of terms such as these with integral control since they are not typically known to a high
degree of accuracy and thus cannot practically be inverted along with the flight path angle
coupling. Upon substitution of the control law above into the normal specific acceleration
dynamics of equation (59), the closed loop normal dynamics become,

= L M E K E + MQ L M E K C
C W E C Q W
mI yy I yy mV I yy
(62)
M L M L M E L M E
+ + Q + KC KQ C W
I yy mVI yy mI yy mVI yy
E C = C W C WR (63)
when,
g M I
E = Q KQ cos W + yy sin (64)

M E W
W
V M E
DI

and the static offset terms are ignored. Note that the dynamic inversion part of the control
law is still a function of the yet to be determined pitch rate feedback gain. Given the desired
closed loop characteristic equation for the normal dynamics,
c ( s ) = s 3 + 2 s 2 + 1s + 0 (65)
the closed form solution feedback gains can be calculated by matching characteristic
equation coefficients to yield,
I yy M L
KQ = 2 + Q (66)
M E I yy mV
mI yy M L L
KC = 1 + 2 (67)
L M E I yy mV mV
mI yy
KE = 0 (68)
L M E
Substituting the pitch rate feedback gain into equation (64) gives,
g I yy L C + g cos W
E = 2 cos W W sin W (69)
DI
M
V E mV V
where use has been made of equation (1) to remove the flight path angle derivative. The
controller design freedom is reduced to that of placing the three poles that govern the closed
loop normal dynamics. The control system will work to keep these poles fixed for all point
mass kinematics states and in so doing yield a dynamically invariant normal specific
acceleration response at all times.
7. Simulation
To verify the controller designs of the previous subsections, they are applied to an off-the-
shelf scale model aerobatic aircraft, the 0.90 size CAP232, used for research purposes at
Stellenbosch University. In the simulations and analysis to follow, the aircraft is operated
about a nominal velocity magnitude of 30 m/s and a nominal sea level air density of 1.225
kg/m3. The modelling parameters for the aircraft are listed in the table below and were
obtained from (Hough, 2007).
m = 5.0 kg A = 5.97 C L0 = 0.0 C m0 = 0.0

I yy = 0.36 kgm 2
T = 0.25 s C L = 5.1309 C m = 0.2954
c = 0.30 m e = 0.85 C LQ = 7.7330 C mQ = 10.281
S = 0.50 m 2 C D0 = 0.02 CL = 0.7126 Cm = 1.5852

E E
Table 1. Model parameters for the Stellenbosch University aerobatic UAV

Given that the scenario described in the example at the end of section 5 applies to the aerobatic
UAV in question, the closed loop natural frequency of the axial specific acceleration controller
should be greater than or equal to the bandwidth of the thrust actuator (4 rad/s) for a return
disturbance of -20 dB. Selecting the closed loop poles at {-43i}, provides a small buffer for
uncertainty in the actuator lag, without overstressing the thrust actuator. Figure 4 provides a
Bode plot of the actual and approximated return disturbance transfer functions for this design,
i.e. equation (23), with the actual and approximated sensitivity functions of equations (33) and
(34) substituted respectively. Also plotted are the actual and approximated sensitivity
functions themselves as well as the term in parenthesis in equation (23), i.e. the normalized
drag to normalized velocity perturbation transfer function. Figure 4 clearly illustrates the
greater than 20 dB of return disturbance rejection obtained over the entire frequency band due
to the appropriate selection of the closed loop poles. The figure also shows how the return
disturbance rejection is contributed towards by the controller at low frequencies and the
natural velocity magnitude dynamics at high frequencies. The plot thus verifies the
mathematics of the decoupling analysis done in section 4.
Open loop analysis of the aircrafts normal dynamics reveals the actual and approximated
poles (shown as crosses) in Figure 5 and the actual and approximated elevator to normal
specific acceleration zeros of {54.7, -46.7} and {54.5, -46.6} respectively. The closeness of the
poles in Figure 5 and the similarity of the numerical values above verify equations (38) and
(40). The approximate zero positions are used in equation (55) to determine the upper NMP
frequency bound shown in Figure 5. The lower timescale separation bound arises as a result
of a desired velocity magnitude bandwidth of 1 rad/s (a feasible user selected value). Notice
both the large feasible pole placement region and the fact that the open loop poles naturally
satisfy the NMP frequency constraint, implying good open loop handling qualities.
The controller of subsection 6.4 is then applied to the system with desired closed loop
complex poles selected to have a constant damping ratio of 0.7 as shown in Figure 5. The
desired closed loop real pole is selected equal to the real value of the complex poles. The
corresponding actual closed loop poles are illustrated in Figure 5. Importantly, the locus of
actual closed loop poles is seen to remain similar to that of the desired poles while the upper
NMP frequency bound is adhered to. Outside the bound the actual poles are seen to diverge
quickly from the desired values.
20 Actual sensitivity function

Approximated sensitivity function
Normalised drag to velocity transfer
10
Actual return disturbance
Approximated return disturbance
0
Magnitude (dB)
-10
-20
-30
-40
-50
-1 0 1 2
10 10 Frequency (rad/sec) 10 10
Fig. 4. NMP upper bound and timescale separation lower bound outlining feasible pole
placement region.
20
Desired CL Poles
15
Actual CL Poles
10 Frequency Bounds
Imaginary Axis [rad/s]
-5
-10
-15
-20
-20 -10 0 10 20
Real Axis [rad/s]
Fig. 5. Actual and approximated open loop (CL) poles, actual and desired closed loop poles
and upper and lower frequency bounds.
The controller of subsection 6.4 is then applied to the system with desired closed loop
complex poles selected to have a constant damping ratio of 0.7 as shown in Figure 5. The
desired closed loop real pole is selected equal to the real value of the complex poles. The
corresponding actual closed loop poles are illustrated in Figure 5. Importantly, the locus of
actual closed loop poles is seen to remain similar to that of the desired poles while the upper
NMP frequency bound is adhered to. Outside the bound the actual poles are seen to diverge
quickly from the desired values.
Figure 6 shows the corresponding feedback gains plotted as a function of the RHP zero
position normalized to the desired natural frequency ( r 1 ). The feedback gains are
normalized such that their maximum value shown is unity. Again, it is clear from the plot
that the feedback gains start to grow very quickly, and consequently start to become
impractical, when the RHP zero is less than 3 times the desired natural frequency. The
results of Figures 5 and 6 verify the design and analysis of section 6.
Given the analysis above, the desired normal specific acceleration closed loop poles are
selected at {-108i, -10}. The desired closed loop natural frequency is selected close to that of
the open loop system in an attempt to avoid excessive control effort. With the axial and
normal specific acceleration controllers designed, a simulation based on the full, nonlinear
dynamics of section 3 was set up to test the controllers. Figure 7 provides the simulation
results.
The top two plots on the left hand side of the figure show the commanded (solid black line),
actual (solid blue line) and expected/desired (dashed red line) axial and normal specific
acceleration signals during the simulation. The normal specific acceleration was switched
between -1 and -2 g s (negative sign implies pull up acceleration) during the simulation
while the axial specific acceleration was set to ensure the velocity magnitude remained
within acceptable bounds at all times.
1 KQ
KC
KE
Normalised feedback gains
0.5
-0.5
-1
2 4 6 8 10 12
RHP zero frequency normalised to the natural frequency
Fig. 6. Normalized controller feedback gains as a function of the RHP zero position
normalized to the desired natural frequency.
Importantly, note how the axial and normal specific acceleration remain regulated as
expected regardless of the velocity magnitude and flight path angle, the latter of which
varies dramatically over the course of the simulation. As desired, the specific acceleration
controllers are seen to regulate their respective states independently of the aircrafts velocity
magnitude and gross attitude. The angle of attack, pitch rate, elevator deflection and
normalized thrust command are shown on the right hand side of the figure. The angle of
attack remains within pre-stall bounds and the control signals are seen to be practically
feasible.
Successful practical results of the controllers operating on the aerobatic research aircraft and
other research aircraft at Stellenbosch University have recently been obtained. These results
will be made available in future publications.
1 8
AW - [g's]
- [deg]
0.5
4
0
-0.5 0
0 2 4 6 8 10 0 2 4 6 8 10
100
Q - [deg/s]
CW - [g's]
-1 50
-2 0
-50
0 2 4 6 8 10 0 2 4 6 8 10
40 0
V - [m/s]
E - [deg]
-2
30
-4
20
0 2 4 6 8 10 0 2 4 6 8 10
300 1
TC/mg - [g's]
W - [deg]
200
100 0.5
0
-100 0
0 2 4 6 8 10 0 2 4 6 8 10
Time - [s] Time - [s]
Fig. 7. Simulation results illustrating gross attitude independent regulation of the axial and
normal specific acceleration.
8. Conclusion and future work

An acceleration based control strategy for the design of a manoeuvre autopilot capable of
guiding an aircraft through the full 3D flight envelope was presented. The core of the
strategy involved the design of dynamically invariant, gross attitude independent specific
acceleration controllers. Adoption of the control strategy was argued to provide a practically
feasible, robust and effective solution to the 3D manoeuvre flight control problem and the
non-iterative nature of the controllers provides for a computationally efficient solution.
The analysis and design of the specific acceleration controllers for the case where the
aircrafts flight was constrained to the 2D vertical plane was presented in detail. The aircraft
dynamics were shown to split into aircraft specific rigid body rotational dynamics and
aircraft independent point mass kinematics. With a timescale separation and a dynamic
inversion condition in place the rigid body rotational dynamics were shown to be
dynamically independent of the point mass kinematics, and so provided a mathematical
foundation for the design of the gross attitude independent specific acceleration controllers.
Under further mild conditions and a sensitivity function constraint the rigid body rotational
dynamics were shown to be linear and decouple into axial and normal dynamics. The
normal dynamics were seen to correspond to the classical Short Period mode approximation
dynamics and illustrated the gross attitude independent nature of this mode of motion.
Feedback based, closed form pole placement control solutions were derived to regulate both
the axial and normal specific accelerations with invariant dynamic responses. Before
commencing with the design of the normal specific acceleration controller, the elevator to
normal specific acceleration dynamics were investigated in detail. Analysis of these
dynamics yielded a novel approximating equation for the location of the zeros and revealed
the typically NMP nature of this system. Based on a time domain analysis a novel upper
frequency bound condition was developed to allow the NMP nature of the system to be
ignored, thus allowing practically feasible dynamic inversion of the flight path angle
coupling.
Analysis and simulation results using example data verified the functionality of the specific
acceleration controllers and validated the assumptions upon which their designs were
based. Future research will involve extending the detailed control system design to the full
3D flight envelope case based on the autopilot design strategy presented in section 2.
Intelligent selection of the closed loop poles will also be the subject of further research.
Possibilities include placing the closed loop poles for maximum robustness to parameter
uncertainty as well as scheduling the closed loop poles with flight condition to avoid
violation of the NMP frequency bound constraint.
9. References
Al-Hiddabi, S.A. & McClamroch, N.H. (2002). Aggressive longitudinal aircraft trajectory
tracking using nonlinear control. Journal of Guidance, Control and Dynamics, Vol. 25,
No. 1, January 2002, pp. 2632, ISSN 07315090.
Bhattacharya, R.; Balas, G.J.; Kaya, M.A. & Packard, A (2002). Nonlinear receding horizon
control of an F-16 aircraft. Journal of Guidance, Control and Dynamics, Vol. 25, No. 5,
September 2002, pp. 924931, ISSN 07315090.
Blakelock, J.H. (1991). Automatic Control of Aircraft and Missiles (2nd Ed.), John Wiley and
Sons, ISBN 0471506516, New York.
Buffington, J.; Adams, R. & Banda, S. (1993). Robust, nonlinear, high angle-of-attack control
design for a super-manoeuvrable vehicle, Proceedings of the AIAA Guidance,
Navigation and Control Conference, pp. 690700, Monterey, August 1993.
Bugajski, D.J. & Enns, D.F. (1992). Nonlinear control law with application to high angle-of-
attack flight. Journal of Guidance, Control and Dynamics, Vol. 15, No. 3, May 1992, pp.
761767, ISSN 07315090.
Etkin, B. (1972). Dynamics of Atmospheric Flight, John Wiley and Sons, ISBN-10 0471246204,
New York.
Etkin, B. & Reid, L.D. (1995). Dynamics of Flight Stability and Control (3rd Ed.), John Wiley and
Sons, ISBN 0471034185, New York.
Freudenberg, J.S. & Looze, D.P. (1985). Right-half plane poles and zeros and design tradeoffs
in feedback systems. IEEE Transactions on Automatic Control, Vol. 30, No. 6, June
1985, pp. 555-565, ISSN 00189286.
Goodwin, G.C.; Graebe, S.F. & Salgado, M.E. (2001). Control System Design, Prentice Hall,
ISBN-10 0139586539, Upper Saddle River.
Hauser, J.; Sastry, S. & Meyer, G. (1992). Nonlinear control design for slightly non-minimum
phase systems: Application to V/STOL aircraft. Automatica, Vol. 28, No. 4, April
1992, pp. 665679, ISSN 00051098.
Hough, W.J. (2007). Autonomous Aerobatic Flight of a Fixed Wing Unmanned Aerial Vehicle,
Masters thesis, Stellenbosch University.
Kim, M.J.; Kwon, W.H.; Kim, Y.H. & Song, C. (1997). Autopilot design for bank-to-turn
missiles using receding horizon predictive control scheme. Journal of Guidance,
Control and Dynamics, Vol. 20, No. 6, November 1997, pp. 12481254, ISSN
07315090.
Lane, S.H. & Stengel, R.F. (1988). Flight control design using non-linear inverse dynamics.
Automatica, Vol. 24, No. 4, April 1988, pp. 471483, ISSN 00051098.
Leith, D.J. & Leithead, W.E. (2000). Survey of gain scheduling analysis & design,
International Journal of Control, Vol. 73, No. 11, July 2000, pp. 1001-1025, ISSN
00207179.
Miller, R.B. & Pachter, M. (1997). Manoeuvring flight control with actuator constraints.
Journal of Guidance, Control and Dynamics, Vol. 20, No. 4, July 1997, pp. 729734,
ISSN 07315090.
Pachter, M.; Chandler P.R. & Smith, L. (1998). Manoeuvring flight control. Journal of
Guidance, Control and Dynamics, Vol. 21, No. 3, May 1998, pp. 368374, ISSN
07315090.
Peddle, I.K. (2008). Acceleration Based Manoeuvre Flight Control System for Unmanned Aerial
Vehicles, PhD thesis, Stellenbosch University.
Reiner, J.; Balas, G.J. & Garrard, W.L. (1996). Flight control design using robust dynamic
inversion and time-scale separation. Automatica, Vol. 32, No. 11, November 1996,
pp. 14931504, ISSN 00051098.
Slotine, J.E. & Li, W. (1991). Applied Nonlinear Control, Prentice Hall, ISBN-10 0130408905,
Upper Saddle River.
Snell, S.A.; Enns, D.F. & Garrard, W.L. (1992). Nonlinear inversion flight control for a super-
manoeuvrable aircraft. Journal of Guidance, Control and Dynamics, Vol. 15, No. 4, July
1992, pp. 976984, ISSN 07315090.
10
Autonomous Flight Control System

for Longitudinal Motion of a Helicopter
Atsushi Fujimori
University of Yamanashi
Japan
1. Introduction
Recently, unmanned aerial vehicles (UAVs) have been developed for the purposes of scientic
observations, detecting disasters, surveillance of trafc and army objectives (Wilson, 2007;
Langelaan & Rock, 2005; Cho et al., 2005). This paper presents an autonomous ight control
design to give insights for developing helicopter-type UAVs.
A Helicopter is generally an unstable aircraft. Once it is stalled, it is not easy to recover
its attitude. A control system is therefore needed to keep the vehicle stable during ight
(Bramwell, 1976; Padeld, 1996; Johnson & Kannan, 2005). This paper presents a ight
control design for the longitudinal motion of helicopter to establish autopilot techniques of
helicopters. The ight mission considered in this paper is that a helicopter hovers at a start
position, moves to a goal position with keeping a specied cruise velocity and hovers again at
the goal. The characteristics of the linearized equation of the helicopter is changed during this
ight mission because the trim values of the equation are widely varied. Gain scheduling (GS)
is one of candidates to stabilize the vehicle for the entire ight region. In this paper, a ight
control system is designed as follows. The ight control system is constructed as a double
loop control system (Fujimori et al., 1999; Fujimori et al., 2002) which consists of an inner-loop
controller and an outer-loop controller. The former is needed for stabilizing the controlled
plant, while the latter is used for tracking the reference which is given to accomplish the
ight mission. To design the inner-loop controller, the longitudinal motion of a helicopter is
rst modeled by a linear interpolative polytopic model whose varying parameter is the ight
velocity. A GS state feedback law is then designed by linear matrix inequality (LMI) (Boyd
et al., 1994; Fujimori et al., 2007) so as to stabilize the polytopic model for the entire ight
region. On the other hand, the outer-loop controller is designed by taking into consideration
the steady-state of the controlled variable.
The rest of this paper is organized as follows. Section 2 shows equations of the longitudinal
motion of helicopter. Section 3 gives a ight mission and shows a double loop control system
adopted in this paper. The details of the controller designs are presented in Section 4. Section
5 shows computer simulation in Matlab/Simulink to evaluate the proposed ight control
system. Concluding remarks are given in Section 6.
control plane
a1
c
x tip path plane

d
V c
horizontal plane
u w
x
V

z
Fig. 1. Helicopter in forward ight
2. Equation of longitudinal motion of helicopter

Figure 1 shows a helicopter considered in this paper. The angular velocity of the main rotor
is . The main rotor produces the thrust T which is needed for not only lifting the vehicle
against the gravity but also moving transitionally and rotationally. It depends on the tilting
angle of the control plane.
( x, y, z) represent the body-xed-axes whose origin is located at the center of gravity of the
vehicle. The forward velocity is V whose x- and z-axes elements are u and w, respectively.
The longitudinal motion of helicopter consists of the transitional motion with respect to x-
and z-axes and the rotational motion around y-axis; that is, the pitch angle denoted as and
its derivative q(= ). It is represented as the following equations (Bramwell, 1976; Padeld,
1996):
m(u + qw) = X mg sin (1)

m(w qu) = Z + mg cos (2)
Iyy q = M (3)
where m and Iyy are respectively the mass and the moment of inertia of the vehicle. g is the
gravity acceleration. The external forces X, Z and the moment M are given by
X = T sin(c a1 ) D cos (4)

Z = T cos(c a1 ) D sin (5)
M = Th R sin(c a1 ) (6)
where is dened as tan. c is the cyclic pitch angle which is one of the control inputs for
the longitudinal motion of the helicopter. a1 is the angle between the control plane and the tip
path plane. h R is the distance of the hub from the center of gravity. D is the drag of the vehicle
and is given by
1
D = V 2 SCD (7)
2
Autonomous Flight Control System for Longitudinal Motion of a Helicopter 199
where is the atmospheric density, S the representative area and CD the drag coefcient. The
thrust T can be calculated by integrating the lift over the whole blade. This results in the
following expression for the thrust coefcient:
T Nc 2
CT = = C {( + 2 )0 c i } (8)
(R)2 R2 4R l 3
where
V V vi
= cos c , c = sin c , i = (9)
R R R

c = c (10)
R is the radius of the rotor blades, c the chord length, N the number of the blades and Cl the
lift slope of the blades. vi is the induced velocity through the rotor. 0 is the collective pitch
angle which is another control input. According to Van Hoydonck (2003), the dimensionless
induced velocity i through the rotor is approximated by
i = CT CTGl (11)
where CTGl is the thrust coefcient which is given by Glauerts hypotheses. is a time constant
of i .
Summarizing the above equations, dene the state and the input vectors as

x p = [ u w q i ] T 5 , u p = [ 0 c ] T 2 . (12)
The equation of the longitudinal motion of the helicopter is then written as
x p = f p ( x p , u p ). (13)
Equation (13) is referred as the nonlinear plant Pnl hereafter.

Letting xe and he be horizontal and the vertical positions of the helicopter from the start, they
are given by
xe = u cos + w sin (14)

he = u sin w cos . (15)

Dening p as p = [ xe he ] T , they are compactly given as
p = g p ( x p ). (16)
In this paper, numerical values of the Eurocopter Deutschland Bo105 (Padeld, 1996) were
used in simulation. They are listed in Table 1. Since this paper considers hovering and forward
ight, the trim condition is given in level ight. Letting x p and u p be the state and the input
in trim, respectively, f ( x p , u p ) = 0 holds. Figure 2 shows variations of x p and u p with respect
to the ight velocity V. It is seen that all trim values are changed in the range of V [0, 60]
[m/s].
u [m/s] w [m/s]
60 0
2
40
4
6
20
8
0 10
0 20 40 60 0 20 40 60
[deg]
i
0 0.06
2
0.04
4
6
0.02
8
10 0
0 20 40 60 0 20 40 60
V [m/s] V [m/s]
(a) States in trim x p
0 [deg] c [deg]
10 5
4
9
3
8
2
7
1
6 0
0 20 40 60 0 20 40 60
V [m/s] V [m/s]
(b) Inputs in trim u p
Fig. 2. Trim values with respect to forward velocity V

parameter value unit

Cl 6.113 [1/rad]
c 0.27 [m]
m 2200 [kg]
Ibl 231.7 [kgm2 ]
Iyy 4973.0 [kgm2 ]
R 4.91 [m]
44.4 [rad/s]
N 4 [-]
CD S 1.5 [m2 ]
hR 1.48 [m]
Table 1. Parameters of Bo105 (Padeld, 1996)
Vr
Vc1
Vc2
0
tc1 tc2 tc3 tc4 tc5 t
Fig. 3. Flight velocity prole Vr
3. Construction of ight control system

Let the start position be the origin of the coordinates ( xe , he ). A ight mission considered in
this paper is to navigate the helicopter from the start (0,0) to the goal, denoted as ( xr , hr ),
with keeping its attitude stable. To design a control system, the followings are assumed to be
satised:
(i) The motion in y-axis direction is not taken into account.
(ii) x p is measurable.
(iii) The trim values x p and u p are known in advance.
To realize the ight mission, this paper constructs a tracking control system whose controlled
variable is the ight velocity. The ight region is divided into six phases with respect to the
ight velocity as shown in Fig. 3. They are referred as follows.
Kin zp
r + zr + + +v up
Kp Kout [E -F] Pnl xp p
- - gp
Fig. 4. Flight control system
0 t < tc1 : initial hovering phase

tc1 t < tc2 : acceleration phase
tc2 t < tc3 : cruise phase
tc3 t < tc4 : deceleration phase
tc4 t < tc5 : low speed phase
tc5 t : approach phase
From the initial hovering phase to the low speed phase, the reference of the ight velocity
is given by Vr shown in Fig. 3. In this paper, the total time of the ight is not cared. But the
integrated value of Vr for t [0, tc5 ] must be less than xr not to overtake the goal before the
approach phase. In the approach phase, the reference is generated to meet the position of the
helicopter p = [ xe he ] T with the goal r = [ xr hr ] T .
Taking into consideration the above, a double loop control system (Fujimori et al., 1999;
Fujimori et al., 2002) is used as a ight control system in this paper. It is shown in Fig. 4. Pnl
represents the nonlinear helicopter dynamics given by Eq. (13), Kin is the inner-loop controller,
Kout is the outer-loop controller and K p is a gain. The controlled variable from the initial

hovering phase to the low speed phase is given by z p = [u w] T and its reference is given

by zr = [ur wr ] T . In the approach phase, another loop is added outside of (zr z p )-loop,

where p = [ xe he ] T is the controlled variable and r = [ xr hr ] T is its reference.
Kin consists of
Kin = [ E F ] (17)
where E is a feedforward gain for tracking the reference, while F is a feedback gain
for stabilizing the plant. Since the trim values are widely varied as shown in Fig. 2, the
characteristics of the linearized plant is also varied. Then, F is designed by a GS technique
in terms of LMI formulation (Boyd et al., 1994).
The reference zr from the initial hovering phase to the low speed phase is generated by the
ight velocity prole shown in Fig. 3, and zr in the approach phase is derived from the
positional error r p . The switch of the reference is done at t = tc5 .
4. Design of control system

4.1 Linear interpolative polytopic model
The objective of ight control in this paper is that the controlled variable is regulated to the
specied trim condition. Linearized models along with the trim is therefore used for controller
design. Letting x p (V ), u p (V ) be respectively the state and the input in trim where the ight
velocity is V, the perturbed state and the input are dened as

x p (t) = x p (t) x p (V ), u p (t) = u p (t) u p (V ) (18)
The linearized equation of Eq. (13) is then given as
x p (t) = A p (V )x p (t) + B p (V )u p (t) (19)
where
f p ( x p , u p ) f p ( x p , u p )
A p (V ) = , B p (V ) = . (20)
x Tp u Tp
Although matrices A p and B p are functions with respect to V, it is hard to get their explicit
representations because of complicated dependence of V as described in Section 2. Then, A p
and B p are approximated by interpolating multiple linearized models in the trim condition.
For the range of the ight velocity V [0, Vu ], r points {V1 , , Vr }, called the operating
points, are chosen as
0 V1 < < Vr Vu . (21)
The linearized model for V = Vi is a local LTI model representing the plant near the i-th
operating point. Linearly interpolating them, a global model over the entire range of the ight
velocity is constructed as

x (t) = A (V )x (t) + B (V )u (t)
p p p p p
(22)
z p (t) = C p x p (t) + z p (V )
where
r r
A p (V ) = i (V ) A pi , B p (V ) = i (V ) Bpi ,
i =1 i =1

10000
Cp = . (23)
01000
i (V ) satises the following relations.
0 i (V ) 1 (i = 1, , r ) (24)
r
i (V ) = 1 (25)
i =1
Equation (22) with Eq. (23) is called the linear interpolative polytopic model in this paper.
4.2 Design of Kin

Under assumption (ii), consider a state feedback law
u p (t) = F (V )x p (t) + E(V )v(t) (26)

where v is a feedforward input for tracking zr and is given by v = zr z p when designing

Kin . The closed-loop system combining Eq. (26) with Eq. (22) is given by

x (t) = A (V )x (t) + B (V ) E(V )v(t)
p F p p
(27)
z p (t) = C p (V )x p (t) + D p (V ) E(V )v(t) + z p (V )

A F (V ) = A p (V ) B p (V ) F (V ).
The steady-state controlled variable is given by
z p () = C p A 1
F B p Ev + z p . (28)
v is then given so as to meet z p () with the reference zr ; that is, z p () zr . E is designed as
E = (C p A 1 1
F Bp ) . (29)
Next, F (V ) is designed so that the closed-loop system is stable over the entire ight range and
H2 cost is globally suppressed (Fujimori et al., 2007). The controlled plant is newly given by

x (t) = A (V )x (t) + B (V )u (t) + B (V )w (t)
p p p p p 1 1
(30)
z1 (t) = C1 (V )x p (t) + D1 (V )u p (t)
where z1 and w1 are respectively the input and the output variable for evaluating H2 cost.
B1 (V ), C1 (V ) and D1 (V ) are matrices corresponding to z1 and w1 . Substituting Eq. (26)
without v into Eq. (30), the closed-loop system is

x (t) = A (V )x (t) + B (V )w (t)
p F p 1 1
(31)
z1 (t) = C1F (V )x p (t)

C1F (V ) = C1 (V ) D1 (V ) F (V ).
In this paper, F (V ) is designed so as to minimize the integration of H2 cost over V [0, Vu ].

That is, the objective is to nd F (V ) such that (Boyd et al., 1994).
Vu
inf trW (V )dV subject to
F,P,W 0

P(V ) P(V ) B1 (V )
> 0, (32a)
() W (V )

He( P(V ) A F (V )) + V dV
dP
()
<0 (32b)
C1F (V ) Iq

He( A) is dened as He( A) = A + A T where () means the transpose of the element located
at the diagonal position. P(V ) > 0 is the parameter dependent Lyapunov variable. To derive
nite number of inequalities from Eq. (32), the following procedures are performed. First

dene new variables X (V ) = P1 (V ) and M (V ) = F (V ) X (V ). A p (V ) and B p (V ) are given
by the polytopic forms Eq. (23). B1 (V ), C1 (V ), D1 (V ), X (V ), M (V ) and W (V ) are also given

by similar polytopic forms. The range of V = aV is given as aV [ aV , aV ]. Furthermore,
dP/dV are approximated as
dP dX 1
= X 1 X (33)
dV dV
dX X Xi Xi
i +1 = (34)
dV Vi+1 Vi Vi
Pre- and post-multiplying diag{ X, I } to Eq. (32) and using the polytopic forms, the following
LMIs are derived as a sufcient condition.
r 1
inf
Mi ,Xi ,Wi i =1
tr(Wi Vi ) subject to

Xi B1i
>0 (i = 1, , r ), (35a)
() Wi

He( A pi Xi B pi Mi ) aV X
Vi ()
i
< 0, (35b)
C1i Xi D1i Mi Iq

He( A pj X j B pj M j ) aV X
Vi ()
i
< 0, (35c)
C1j X j D1j M j Iq

He( A pi X j B pi M j
()
+ A X B M ) 2a Xi
pj i pj i V Vi
<0 (35d)
C X D M
1i j 1i j
2Iq
+C1j Xi D1j Mi
(i = 1, , r 1, j = i + 1), aV = aV , aV
If Mi , Xi and Wi (i = 1, , r ) satisfying the above LMIs, F (V ) is given by

1
r r
F (V ) = i ( V ) Mi i ( V ) Xi . (36)
i =1 i =1
4.3 Design of Kout

Since v in Eq. (26) is a feedforward input from the ight velocity reference, the tracking error
will be occurred by model uncertainties and/or disturbances. Let us evaluate this in the LTI
representation. Let Tz p v (s) be a transfer function from v to z p . z p converges to a constant zr if
there are no model uncertainties in Tz p v (s) because Tz p v (0) = I. If Tz p v (0) is varied as Tz p v (0) =
I + due to model uncertainties, we have the following steady-state error:

e0 = zr z p () = zr (37)
Model Vi [m/s] GS-SF

Ppoly1 { 0, 50 } Fgs1
Ppoly2 { 0, 25, 50 } Fgs2
Ppoly3 { 0, 10, 15, 40, 50 } Fgs3
Table 2. Operating points of polytopic models
Model Vd [m/s] Fixed-SF

Plti1 0 Ff ix1
Plti2 25 Ff ix2
Plti3 50 Ff ix3
Table 3. Design points of LTI models

To reduce the error, a feedback from z p ; that is, an outer-loop is added as shown in Fig. 4. The
transfer function from zr to z p is given by
Tz p zr (s) = ( I + Tz p v (s)Kout (s))1 Tz p v (s)( I + Kout (s)) (38)

The steady-state error is then
1
e1 = zr z p () = zr lim sTz p zr (s) zr
s 0 s (39)
= ( I + Tz p v (0)Kout (0))1 zr .
This means that the steady-state error e1 with the outer-loop is reduced by ( I +
Tz p v (0)Kout (0))1 . Summarizing the above, the design requirements of Kout are given as
follows:
(i) Kout must stabilize Tz p v .
(ii) The amplitude of ( I + Tz p v ( j )Kout ( j ))1 should be small in the low frequency region.
5. Simulation
To evaluate the proposed ight control system, a ight simulator was built on
MATLAB/Simulink. For design and discussion hereafter, the notations about plant models are
given as follows: Pl pv (V ) is a linear parameter varying (LPV) model obtained by linearizing
Pnl . Ppoly (V ) is the linear interpolative polytopic model given by Eq. (22) with Eq. (23). Plti (Vd )
is an LTI model where the ight velocity is xed at Vd .
Two cases of ight control system with respect to the state feedback gain F were compared in
simulation. One is that F was designed by GS where the plant model was Ppoly (V ). Another
is that F was designed by LQR where the plant model was Plti (Vd ). The former is referred to
as GS-SF, while the latter is referred to as Fixed-SF. The parameter values of the ight velocity
prole in Fig. 3 were given as follows:
( xr , hr ) = (3000, 0) [m], Vc1 = 50, Vc2 = 15 [m/s],
tc1 = 5, tc2 = 30, tc3 = 60, tc4 = 80, tc5 = 100 [s].
gap metric between Plti and Plpv

1
0.8 P
lti1
P
lti2
0.6 Plti3
0.4
0.2
0
0 10 20 30 40 50
V [m/s]
(a) LTI models
gap metric between Ppoly and Plpv
P
poly1
0.5 P
poly2
Ppoly3
0.4
0.3
0.2
0.1
0
0 10 20 30 40 50
V [m/s]
(b) polytopic models
Fig. 5. -gap metric

H cost
2
2.4
F
fix3
Fgs1
2.3
Fgs2
Fgs3
2.2
2.1
1.9
1.8
1.7
0 10 20 30 40 50
V [m/s]
Fig. 6. H2 cost
5.1 Evaluation of design models

According to Section 4.1, three linear interpolative polytopic models were obtained. Table 2
shows the operating points chosen for the models. While, the design points Vd of three LTI
models are shown in Table 3. The -gap metric is one of criteria measuring the model error
in the frequency domain. It had been introduced in robust control theories associated with
the stability margin (Vinnicombe, 2001). The -gap metric between two LTI models, P1 (s) and
P2 (s), is dened as

( P1 , P2 ) = ( I + P2 P2 )1/2 ( P1 P2 )( I + P1 P1 )1/2 (40)
The range is [0, 1]. A large means that the model error is large. The -gap metric is used
for evaluating the model Ppoly (V ) and Plti (Vd ). Figure 5 shows -gap metric between Plti (Vd )
and Pl pv (V ) and between Ppoly (V ) and Pl pv (V ). ( Plti (V ), Pl pv (Vd )) was rapidly increased
when V was shifted from Vd . On the other hand, the maximum of ( Ppoly (V ), Pl pv (V ))
was reduced according to the number of the operating points. It was seen that Ppoly (V )
appropriately approximated Pl pv (V ) over the entire range of the ight velocity.
5.2 Design of F and H2 cost

The design parameters for designing F in Eq. (30) were given as follows.

6.039 10.977

154.03 49.188

0.001I5 052

B1 = 3.954
7.187 , C1 = , D1 =
025 I2
0 0

0.38495 0.12395
FixedSF 1
60 10
40 0
w [m/s]
u [m/s]
20 10
u w
0 20
ur w
r
20 30
0 50 100 150 0 50 100 150
25 6
20 4
[deg]
[deg]
15 2
0
10 0
5 2
0 50 100 150 0 50 100 150
t [s] t [s]
(a) Controlled variables and inputs
FixedSF 1
3000
x [m]
2000
e
1000
0
0 20 40 60 80 100 120 140
t [s]
100
0
h [m]
100
e
200
300
0 20 40 60 80 100 120 140
t [s]
(b) Positions
Fig. 7. Time responses using Ff ix1

FixedSF 2
60 10
u
40 u 0
r
w [m/s]
u [m/s]
20 10
0 20 w
w
r
20 30
0 50 100 150 0 50 100 150
25 6
20 4
[deg]
[deg]
15 2
0
10 0
5 2
0 50 100 150 0 50 100 150
t [s] t [s]
FixedSF 2
3000
x [m]
2000
e
1000
0
0 20 40 60 80 100 120 140
t [s]
100
0
h [m]
100
e
200
300
0 20 40 60 80 100 120 140
t [s]
(b) Positions

FixedSF 3
60 10
u
40 u 0
r
w [m/s]
u [m/s]
20 10
0 20 w
w
r
20 30
0 50 100 150 0 50 100 150
25 6
20 4
[deg]
[deg]
15 2
0
10 0
5 2
0 50 100 150 0 50 100 150
t [s] t [s]
FixedSF 3
3000
2000
x [m]
e
1000
0
0 20 40 60 80 100 120 140
t [s]
100
0
h [m]
100
e
200
300
0 20 40 60 80 100 120 140
t [s]
(b) Positions

GSSF 1
60 10
u
40 u 0
r
w [m/s]
u [m/s]
20 10
0 20 w
w
r
20 30
0 50 100 150 0 50 100 150
25 6
20 4
[deg]
[deg]
15 2
0
10 0
5 2
0 50 100 150 0 50 100 150
t [s] t [s]
GSSF 1
3000
x [m]
2000
e
1000
0
0 20 40 60 80 100 120 140
t [s]
100
0
h [m]
100
e
200
300
0 20 40 60 80 100 120 140
t [s]
(b) Positions
Fig. 10. Time responses using Fgs1

GSSF 2
60 10
u
40 u 0
r
w [m/s]
u [m/s]
20 10
w
0 20
w
r
20 30
0 50 100 150 0 50 100 150
25 6
20 4
[deg]
[deg]
15 2
0
10 0
5 2
0 50 100 150 0 50 100 150
t [s] t [s]
GSSF 2
3000
x [m]
2000
e
1000
0
0 20 40 60 80 100 120 140
t [s]
100
0
h [m]
100
e
200
300
0 20 40 60 80 100 120 140
t [s]
(b) Positions

GSSF 3
60 10
u
40 u 0
r
w [m/s]
u [m/s]
20 10
0 20 w
wr
20 30
0 50 100 150 0 50 100 150
25 6
20 4
[deg]
[deg]
15 2
0
10 0
5 2
0 50 100 150 0 50 100 150
t [s] t [s]
GSSF 3
3000
x [m]
2000
e
1000
0
0 20 40 60 80 100 120 140
t [s]
100
0
h [m]
100
e
200
300
0 20 40 60 80 100 120 140
t [s]
(b) Positions

They were used for both of GS-SF and Fixed-SF. Three GS-SF gains denoted as Fgsi (i =
1, 2, 3) were designed according to Section 4.2, while three Fixed-SF gains denoted as Ff ixi
(i = 1, 2, 3) were designed by LQR technique in which the weights of the quadratic index
were given by C1T C1 and D1T D1 .
Figure 6 shows the H2 cost of the closed-loop system which the designed F is combined with
Eq. (30). The H2 cost by Ff ix3 was minimized at V = 40 [m/s] which was near the design
point Vd = 50 [m/s], but was increased in the low ight velocity region. The H2 cost by Ff ix1
and Ff ix2 showed the similar result. On the other hand, the H2 cost by Fgs2 and Fgs3 was
kept small over the entire ight region. The H2 cost by Fgs1 was small in the middle ight
velocity region but was increased in the low and the high ight velocity regions.
5.3 Tracking evaluation

The ight mission given in Fig. 3 was performed in Simulink. Figures 7 - 12 show the time
histories of the closed-loop system with the three GS-SF and three Fixed-SF gains. In the case
of Ff ix1 shown in Fig. 7, the controlled variables u and w tracked their references until the
acceleration phase (5 t < 30 [s]) but they were diverged in the cruise phase (30 t < 60
[s]). In the deceleration phase (60 t < 80 [s]), the closed-loop system was stabilized again
but it was de-stabilized in the approach phase (t 100 [s]). Although the closed-loop system
remained stable for the entire ight region in the case of Ff ix2 shown in Fig. 8, oscillatory
responses were observed in the cruise and approach phases. The responses using Ff ix3
shown in Fig. 9 were better than those using Ff ix2 .
On the other hand, the three GS-SF gains provided stable responses as shown in Figs. 10 -
12, In particular, The responses by Fgs3 showed improved tracking and settling properties
compared to other cases.
Summarizing the simulation in MATLAB/Simulink, polytopic model Ppoly3 made the -gap
metric smaller than other models for the entire ight region. Fgs3 designed by using Ppoly3
showed better control performance.
6. Concluding remarks
This paper has presented an autonomous ight control design for the longitudinal motion of
helicopter to give insights for developing autopilot techniques of helicopter-type UAVs. The
characteristics of the equation of helicopter was changed during a specied ight mission
because the trim values of the equation were widely varied. In this paper, gain scheduling
state feedback (GS-SF) was included in the double loop ight control system to keep the
vehicle stable for the entire ight region. The effectiveness of the proposed ight control
system was evaluated by computer simulation in MATLAB/Simulink. The model error of
the polytopic model was smaller than that of LTI models which were obtained at specied
ight velocity. Flight control systems with GS-SF showed better control performances than
those with xed-gain state feedback. The double loop ight control structure was useful for
accomplishing ight mission considered in this paper.
7. References
[1] Boyd, S.; Ghaoui, L. E.; Feron, E. & Balakrishnan, V. (1994). Linear Matrix Inequalities in
System and Control Theory, SIAM, Vol. 15, Philadelphia.
[2] Bramwell, A. R. S. (1976). Helicopter Dynamics, Edward Arnold, London, 1976.
[3] Cho, S.-J.; Jang, D.-S. & Tahk, M.-L. (2005). Application of TCAS-II for Unmanned Aerial
Vehicles, Proc. CD-ROM of JSASS 19th International Sessions in 43rd Aircraft Symposium,
Nagoya, 2005
[4] Fujimori, A.; Kurozumi, M.; Nikiforuk, P. N. & Gupta, M. M. (1999). A Flight Control
Design of ALFLEX Using Double Loop Control System, AIAA Paper, 99-4057-CP,
Guidance, Navigation and Control Conference, 1999, pp. 583-592.
[5] Fujimori, A.; Nikiforuk, P. N. & Gupta, M. M. (2001). A Flight Control Design of
a Reentry Vehicle Using Double Loop Control System with Fuzzy Gain-Scheduling,
IMechE Journal of Aerospace Engineering, Vol. 215, No. G1, 2001, pp. 1-12.
[6] Fujimori, A.; Miura, K. & Matsushita, H. (2007). Active Flutter Suppression of a
High-Aspect-Ratio Aeroelastic Using Gain Scheduling, Transactions of The Japan Society
for Aeronautical and Space Sciences, Vol. 55, No. 636, 2007, pp. 34-42.
[7] Johnson, E. N. & Kannan, S. K. (2005). Adaptive Trajectory Control for Autonomous
Helicopters, Journal of Guidance, Control and Dynamics, Vol. 28, No. 3, 2005, pp. 524-538.
[8] Langelaan, J. & Rock, S. (2005). Navigation of Small UAVs Operating in Forests, Proc.
CD-ROM of AIAA Guidance, Navigation, and Control Conference, San Francisco, 2005.
[9] Padeld, G. D. (1996). Helicopter Dynamics: The Theory and Application of Flying Qualities
and Simulation Modeling, AIAA, Reston, 1996.
[10] Van Hoydonck, W. R. M. (2003). Report of the Helicopter Performance, Stability and Control
Practical AE4-213, Faculty of Aerospace Engineering, Delft University of Technology,
2003.
[11] Vinnicombe, G. (2001). Uncertainty and Feedback (H loop-shaping and the -gap metric),
Imperial College Press, Berlin.
[12] Wilson, J. R. (2007). UAV Worldwide Roundup 2007, Aerospace America, May, 2007, pp.
30-38.
11
Autonomous Flight Control for

RC Helicopter Using a Wireless Camera
Yue Bao1, Syuhei Saito2 and Yutaro Koya1
1Tokyo City University,
2Canon Inc.
Japan
1. Introduction
In recent years, there are a lot of researches on the subject of autonomous flight control of a
micro radio control helicopter. Some of them are about flight control of unmanned
helicopter (Sugeno et al., 1996) (Nakamura et al., 2001). The approach using the fuzzy
control system which consists of IF-Then control rules is satisfying the requirements for the
flight control performance of an unmanned helicopter like hovering, takeoff, rotating, and
landing. It is necessary to presume three dimensional position and posture of micro RC
helicopter for the autonomous flight control.
A position and posture presumption method for the autonomous flight control of the RC
helicopter using GPS (Global Positioning System), IMU (Inertial Measurement Unit), Laser
Range Finder (Amida et al., 1998), and the image processing, etc. had been proposed.
However, the method using GPS cannot be used at the place which cannot receive the
electric waves from satellites. Therefore, it is a problem that it cannot be used for the flight
control in a room. Although the method which uses various sensors, such as IMU and Laser
Range Finder, can be used indoors, you have to arrange many expensive sensors or
receivers in the room beforehand. So, these methods are not efficient. On the other hand, the
method using an image inputted by a camera can be used in not only outdoors but also
indoors, and is low price. However, this method needs to install many artificial markers in
the surroundings, and it is a problem that the speed of the image inputting and image
processing cannot catch up the speed of movement or vibration of a RC helicopter. A
method presuming the three dimensional position and posture of a RC helicopter by the
stereo measurement with two or more cameras installed in the ground was also proposed.
In this case, the moving range of the RC helicopter is limited in the place where two or more
cameras are installed. Moreover, there is a problem for which a high resolution camera must
be used to cover a whole moving range. (Ohtake et al., 2009)
Authors are studying an autonomous flight of a RC helicopter with a small-wireless camera
and a simple artificial marker which is set on the ground. This method doesnt need to set
the expensive sensors, receivers, and cameras in the flight environment. And, we thought
that a more wide-ranging flight is possible if the natural feature points are detected from the
image obtained by the camera on the RC helicopter. This chapter contains the following
contents.
a. Input method of image from a small, wireless camera which is set on a RC helicopter.
b. Extraction method of feature points from an image of flight environment taken with a
camera on RC helicopter.
c. Calculation method of three dimensional position and posture of RC helicopter by
image processing.
d. Experiment of autonomous flight of a RC helicopter using fuzzy logic control.
2. Composition of system
The overview of a micro RC helicopter with coaxial counter-rotating blades used in our
experiment is shown in Fig.1. Since this RC helicopter can negate a running torque of a body
by a running torque between an up propeller and a down propeller, it has the feature that it
can fly without being shakier than the usual RC helicopter. The composition of our
experiment system for automatic guidance of RC helicopter is shown in Fig.2. A small
wireless camera is attached on the RC helicopter as shown in Fig.3, and the image of the
ground is acquired with this camera, and this image is sent to the receiver on ground, and
then sent to the computer through a video capture. The position and posture of the RC
Fig. 1. Micro RC helicopter with the coaxial contra-rotating rotors
Fig. 2. Composition of automatic guidance system

Autonomous Flight Control for RC Helicopter Using a Wireless Camera 219
Fig. 3. RC helicopter equipped with a micro wireless camera
Fig. 4. An artificial marker
Fig. 5. Coordinate axes and attitude angles

helicopter are computed with image processing by the computer set on the ground, and, this
image processing used the position and shape of an artificial marker in the camera image
like Fig.4. The three dimensional position of the RC helicopter ( x(t ), y(t ), z(t )) and the
changing speed of the position ( x (t ), y (t ), z(t )) and the attitude angles (t ) and changing
speed of attitude angles (t ) can be obtained by this calculation. Fig.5 shows the relation
between these coordinate axes and attitude angles.
This micro RC helicopter is controlled by four control signals, such as Aileron, Elevator,
Rudder, and Throttle, and, the control rule of fuzzy logic control is decided by using
measurement data mentioned above. The control signals are sent to micro RC helicopter
through the digital-analog converter.
3. Image processing
3.1 Image input
The micro wireless camera attached on the RC helicopter takes an image by interlaces
scanning. If the camera takes an image during RC helicopter flying, since the vibration of the
RC helicopter is quicker than the frame rate of the camera, the image taken by the camera
will be a blurred image resulting from an interlace like Fig.6. We devised a method skipping
the odd number line (or, even number line) of input image to acquire an clear input image
while the RC helicopter is flying.
Fig. 6. A blurring image acquired by wireless camera
3.2 Feature point extraction

Feature point detection is defined in terms of local neighborhood operations applied to an
image such as an edge and corner. Harris operator (Harris and Stephens, 1988) (Schmid et
al., 1998) and SUSAN operator (Smith and Brady, 1997) are well known as common feature
detectors. The methods (Neumann and You, 1999) (Bao and Komiya, 2008) to estimate
position and attitude by using natural feature points or marker in the input image are
proposed too. Our prototype experiment system used the Harris operator which can extract
the same feature points in higher rate more than other feature point extraction methods
(Schmid et al., 1998). First, we obtain a grayscale image from camera. Let us consider taking
an image patch over an area (u , v ) from an image I and shifting it by (x, y). The Harris
matrix M can be found by taking the second derivative of the sum of squared differences
between these two patches around (x , y ) = (0, 0). M is given by:
2I I I
G 2 G
x x y
M = (1)
I I 2I
G
x y G
2
y
Let the standard deviation of G in the equation be with Gaussian function for performing
smoothing with Gaussian filter. The strength of a corner is decided by second derivative.
Here, the eigenvalue of M is (1 , 2 ) , and the value of eigenvalue can be got from the
following inference.
If 1 0 and 2 0 , then there are no features at this pixel (x , y).
If either 1 or 2 is large positive value, then an edge is found.
If 1 and 2 are both large positive values, then a corner is found.
Because the exact calculation of eigenvalue by the method of Harris will increase
computational amount, the following functions R were proposed instead of those
calculation methods.
R = det( M ) k(tr( M ))2 (2)
R = 12 k( 1 + 2 )2 (3)
The det expresses a determinant and tr expresses the sum of the diagonal element of a
matrix, and k is a value decided experientially.
The kanade-Tomasi corner detector (Shi and Tomasi, 1994) uses min (12) as measure of
feature point. For example, Fig.7 shows a feature point detection using Harris operator for
photographed image. The Harris operator detects the corner point mainly from the image as
a feature point.
The position of feature point is estimate able by related position information of an artificial
marker to feature point from camera image after coordinate transformation. The flight
control area of RC helicopter can be expanded(see Fig.8) by using the information of natural
feature points around an artificial marker. Harris operator is suitable for detecting natural
points. Our system saves the areas including the natural feature points as templates when
the artificial marker is detected. In the range that can take the image of the artificial marker,
the system uses the position information of the artificial marker. If the system can't take the
image of an artificial marker, the position of the helicopter is estimated by template
matching between the area of natural feature points and the template area.
Fig. 7. Feature point extraction by a Harris operator
Fig. 8. The expansion of flight area

3.3 Detection of a marker
Fig. 9. The flow chart of marker detection

The image of the marker photographed with the wireless camera on micro RC helicopter is
shown in Fig.4. The yaw angle can be calculated by the center of the marker and a position
of a cut of triangle of the marker, and the angle of the pitch and the angle of roll can be
acquired by coordinate transformation between a camera coordinate system and a world
coordinate system. The flow chart about the image processing for the marker extraction and
the calculation of position and posture are shown in Fig.9. First, a binarization processing is
performed at the input image from the camera installed on the micro RC helicopter. Then,
the marker in the image is searched, and the outline of the marker is extracted. If the outline
cannot be extracted, an image is acquired from the camera again and a marker is searched again.
A marker center is searched after the outline of a marker is extracted. The method of search
is shown in Fig.10. The maximum values and the minimum values about the x-coordinate
and y-coordinate are searched from of the extracted outline, and their middle values are
given as the coordinates of the marker center. A length of major axis and a length of minor
axis of the marker are calculated by the distance between the center coordinates of the
marker and the pixel on the outline of the marker. The calculation method of the major axis
and a minor axis is shown in Fig.11. When the center coordinate is defined as P( xC , yC ) , and
the coordinate of the pixel which is on the outline is defined as I(x,y), the distance PI from
the center to the pixel of the outline is calculated by equation (4).
PI = ( x xc )2 + ( y yc )2 (4)
For obtaining the maximum value G1 of PI and the minimum value G2 of PI, all pixels on
the outline are calculated. And the segment of G1 is defined as the major axis PO , and the
segment of G2 is defined as a minor axis PQ . The position and posture of the micro RC
helicopter are calculated by the method shown in Section 4.
ymax
ymax-ymin
2 P (xc,yc)
ymin
X
xmin xmax-xmin xmax
2
Fig. 10. The marker center
O
I(x,y)
PI
G1
G2 P (xc,yc)
X
Fig. 11. The calculation method of the major axis and the minor axis of the marker
Since the vibration of the RC helicopter is quicker than the frame rate of the camera, when
the camera which attached on the micro RC helicopter takes the image, the taken image has
become a blurred image resulting from the interlace like Fig.6. Therefore, an exact result
cannot be obtained by a usual image processing. The image processing that we devised
scans with skipping the odd (or even) number lines only about y axial direction at pixel
scanning. The method of scanning is shown in Fig.12. If the pixels processed are on odd
lines, the even number lines are skipped only about y axial direction. About x axial
direction, the system scans 1 pixel at a time like the usual pixel scanning. By this method, a
stable profile tracking can be performed without being blurred by interlace.
edge pixel
attention pixel
odd-numbered line
even-numbered line
Fig. 12. The method of outline tracking
4. Calculation of position and posture

4.1 Calculation of the position coordinates of RC helicopter
In order to measure the position of the RC helicopter, it is necessary to calculate the focal
distance of the wireless camera on the RC helicopter first. If the size of marker and the
location of z axial direction of RC helicopter are got, the focal distance f of the wireless
camera can be calculated easily. The three-dimensional coordinates of a moving RC
helicopter are calculated by using the focal distance, a center of gravity and the value of
radius of the marker. The image formation surface of the camera and the circular marker
surface are parallel if RC helicopter is hovering above the circular marker like Fig.13.
Lets consider that a marker is in the center of the photographed image. The radius of
circular marker in the camera image is defined as D1 , and the center coordinates of the
circle marker are defined as ( xC 1 , yC 1 ) . The radius of the actual marker is defined as d1, and
the center coordinates of the actual marker in a world coordinate system are defined as
( x1 , y 1 , z1 ) . Then, the focal distance f of a camera can be calculated from the following two
equations from parallel relation.
z1 : d1 = f : D1 (5)
z1 D1
f = (6)
d1
When the RC-helicopter is in moving, the radius of the marker in the image after moving is
defined as D2 , the center coordinates of the marker after moving are defined as ( xC 2 , yC 2 ) ,
and the center coordinates of actual marker is defined as ( x2 , y 2 , z2 ) . Then, the following
equation is acquired.
z2 : d1 = f : D2 (7)
Here, since the focal distance f and the radius d1 of actual marker are not changing, the
following relation is obtained from equation (5) and equation (7).
D1 : D2 = z2 : z1 (8)
z2 can be acquired by the following equation. Moreover, x2 and y 2 can be calculated by

the following equations from parallel relation. Therefore, the coordinate of the helicopter
after moving is computable by using equation (9), equation (10), and equation (11), using the
focal distance of the camera.
D1 z1
z2 = (9)
D2
X2 z2
x2 = (10)
f
Y2 z2
y2 = (11)
f
Camera coordinate
D2 D1
X
Camera coordinate (Xc1,Yc1)

D1 f (Xc2,Yc2)
X Y
Y
f (Xc1,Yc1)
z2
z1 (x1,y1) (x2,y2)
y (x1,y1) y
z
z
d1 d1
x x
World coordinate World coordinate
Fig. 13. The location calculation method of RC helicopter
4.2 Calculation of the attitude angle of RC helicopter

The relation between an angle of RC helicopter and an image in the camera coordinate
system is shown in Fig.14. When RC helicopter is hovering above a circular marker, the
circular marker image in the camera coordinate system is a right circle like an actual marker.
If RC helicopter leans, the marker in a camera coordinate system becomes an ellipse. To
calculate the attitude angle, first, the triangular cut part of the circular marker is extracted as
a direction feature point. Then the deformation of the marker image is corrected for
calculating a yaw angle using the relation between the center of the circular marker and the
location of the direction feature point of the circular marker. The pitch angle and the roll
angle are calculated performing coordinate transformation from the camera coordinate
system to the world coordinate by using the deformation rate of the marker in the image
from the wireless camera.
Camera
coordinate
RC
helicopter
Marker
Ground
Fig. 14. Relation between attitude angle of RC helicopter and image in wireless camera.
Calculation of a yaw angle
The value of yaw angle can be calculated using the relation of positions between the center
of circular marker image and the direction feature point of the circular marker image.
However, when the marker image is deforming into the ellipse, an exact value of the yaw
angle cannot be got directly. The yaw angle has to be calculated after correcting the
deformation of the circular marker. Since the length of the major axis of the ellipse does not
change before and after the deformation of marker, the angle between x axis and the
major axis can be correctly calculated even if the shape of the marker is not corrected.
As shown in Fig.15, the center of a marker is defined as point P , the major axis of a marker
is defined as PO , and the intersection point of the perpendicular and x axis which were
taken down from Point O to the x axis is defined as C . The following equation is got if
OPC is defined as '.
OC
' = arctan (12)
PC
Here, when the major axis exists in the 1st quadrant like Fig.15(a), is equal to the value of
', and when the major axis exists in the 2nd quadrant, is calculated by subtracting '
from 180 degrees like Fig.15(b). If the x -coordinate of Point O is defined as xO, the value of
is calculated by the following equation.
( xo 0)
= (13)
180 ( xo < 0)
Fig. 15. An angle between the major axis and coordinate axes
Next, the angle between the major axis and the direction of direction feature point is
calculated. When taking a photograph from slant, a circular marker transforms and becomes
an ellipse-like image, so the location of the cut part has shifted compared with the original
location in the circular image. The marker is corrected to a right circle from an ellipse, and
the angle is calculated after acquiring the location of original direction feature point. First,
the value for deforming an ellipse into a right circle on the basis of the major axis of an
ellipse is calculated. The major axis of an ellipse is defined as PO like Fig.16, and a minor
axis is defined as PQ. The ratio R of the major axis to a minor axis is calculated by the
following equation.
PO G1
R= = (14)
PQ G2
If this ratio multiplies along the direction of a minor axis, an ellipse can be transformed to a
circle. The direction feature point of the marker in the ellipse is defined as a, and the point of
intersection formed by taking down a perpendicular from Point a to the major axis PO is
defined as S . If the location of the feature point on the circle is defined as A, point A is on
the point of intersection between the extended line of the segment aS and a right circle.
Because aS is a line segment parallel to a minor axis, the length of a line segment aS is
calculated by the following equations.
AS = aS R (15)
When the line segment between Point A and the center of the marker is defined as PA , the
angle which the line segment PA and the major axis PO make is calculated by the
following equations.
AS
= arctan (16)
PS
Finally, a yaw angle is calculable by adding to .
O
S
G1
a
A Q

G2
X
P
Fig. 16. An angle between the direction feature point and the major axis
Calculation of pitch angle and roll angle
By using the deformation rate of the marker in an image, a pitch angle and a roll angle can
be calculated by performing coordinate transformation from a camera coordinate system to
a world coordinate system. In order to get the pitch angle and rolling angle, we used a weak
perspective projection for the coordinate transformation (Bao et al., 2003).
Fig.17 shows the principle of the weak perspective projection. The image of a plane figure
which photographed the plane figure in a three-dimensional space by using a camera is
defined as I, and the original configuration of the plane figure is defined as T. The relation
between I and T is obtained using the weak perspective projection transformation by the
following two steps projection.
a. T' is acquired by a parallel projection of T to P paralleled to camera image surface C.
b. I is acquired by a central projection of T ' to C .
The attitude angle ' is acquired using relation between I and T. The angle ' shown in
Fig.18 expresses the angle between original marker and the marker in the camera coordinate
system. In that case, the major axis G1 of the marker image and a minor axis G2 of the
marker image can show like Fig.19.
p
Two dimension image T
3-dimensional space
o
Two dimension image T
p m
Camera imaging surface C O

M
P
Photography image I
z
y
Fig. 17. The conceptual diagram of weak central projection
O
X
P
G1

G2
Fig. 18. The schematic diagram of the attitude angle '

Q G2
P
L
Camera G1

S
Ground U
Fig. 19. Calculation of an attitude angle

Fig. 19 shows the calculation method of . PQ is transformed into LP if along optical axis of
a camera an inverse parallel projection is performed to the minor axis PQ. Since the original
configuration of a marker is a right circle, LP becomes equal to the length G1 of the major
axis in a camera coordinate system. is calculated by the following equation.
G
' = arcsin 2 (17)
G1
To get the segment TU, SU is projected orthogonally on the flat surface parallel to PQ . PQ
and TU are in parallel relationship and LP and SU are also in parallel relationship.
Therefore, the relation between and can be shown by equation (18), and the inclination
of the camera can be calculated by the equation (19).
'= (18)
G2
= arcsin (19)
G1
5. Control of RC helicopter
Control of RC helicopter is performed based on the position and posture of the marker
acquired by Section 4. When RC helicopter is during autonomous hovering flight, the position
data of RC helicopter are obtained by tracking the marker from definite height. The fuzzy rule
of the Throttle control input signal during the autonomous flying is defined as follows.
If z(t ) is PB and z(t ) is PB, Then Throttle is NB
If z(t ) is PB and z(t ) is ZO, Then Throttle is NS
If z(t ) is PB and z(t ) is NB, Then Throttle is ZO
If z(t ) is ZO and z(t ) is PB, Then Throttle is NS
If z(t ) is ZO and z(t ) is ZO, Then Throttle is ZO
If z(t ) is ZO and z(t ) is NB, Then Throttle is PS
If z(t ) is NB and z(t ) is PB, Then Throttle is ZO
If z(t ) is NB and z(t ) is ZO, Then Throttle is PS
If z(t ) is NB and z(t ) is NB, Then Throttle is PB
The fuzzy rule design of Aileron, Elevator, and Rudder used the same method as Throttle.
Each control input u(t) is acquired from a membership function and a fuzzy rule. The
adaptation value i and control input u(t) of a fuzzy rule are calculated from the following
equations.
n
i = Aki ( x k ) (20)
k =1
ici
r
u(t ) = i =r 1 (21)
i =1i
Here, i is the number of a fuzzy rule, n is the number of input variables, r is the quantity of a
fuzzy rule, Aki is the membership function, x k is the adaptation variable of a membership
function, and ci is establishment of an output value (Tanaka, 1994) (Wang et al., 1997).
6. Experiments
In order to check whether parameter of a position and a posture can be calculated correctly,
we compared actual measurement results with the calculation results by several
experiments. The experiments were performed indoors. In the first experiment, a wireless
camera shown in Fig.20 is set in a known three-dimensional position, and a marker is put on
the ground like Fig.21.
The marker is photographed by this wireless camera. A personal computer calculated the
position and the posture of this wireless camera and compared the calculated parameters
with the actual parameters.
Table 1 shows the specification of the wireless camera and Table 2 shows the specification of
the personal computer. A marker of 19cm radius is used in experiments because it is
considered that the marker of this size can be got easily when this type of wireless camera
which has the resolution of 640x480 pixels photographs it at a height between 1m and 2m.
Table 3 shows experimental results of z axis coordinates. Table 4 shows experimental results
of moving distance. Table 5 shows experimental results of yaw angle ( +). Table 6 shows
experimental results of angle. According to the experimental results, although there are
some errors in these computed results, these values are close to actual measurement.
Fig. 20. The wireless camera
Fig. 21. The first experiment
Maker RF SYSTEM lab.

Part number Micro Scope RC-12
Image sensor 270,000pixel , 1/4 inch , color CMOS
Lens 0.8mm Pin lens
Scan mode Interlace
Effective distance 30m
Time of charging battery About 45 minutes
Size 151835(mm)
Weight 14.7g
Table 1. The specification of the wireless camera
Maker Hewlett Packard

Model name Compaq nx 9030
OS Windows XP
CPU Intel Pentium M 1.60GHz
Memory 768MB
Table 2. The specification of PC
Actual distance (mm) 800 1000 1200 1400

Calculated value 785 980 1225 1372
Table 3. The experimental results of z axis coordinates
Actual moving distance (mm) 50 -50 100 -100

Computed value of x axis coordinates 31 -33 78 -75
Computed value of y axis coordinates 29 -33 101 -89
Table 4. The experimental results of moving distance
Actual degree (degree) 45 135 225 315

Table 5. The experimental results of yaw angle ( angle angle)
Actual degree (degree) 0 10 20 40

Table 6. The experimental results of angle
In next experiment, we attached the wireless camera on RC helicopter, and checked if
parameters of a position and a posture would be calculated during the flight. Table 7 shows
the specification of RC helicopter used for the experiment. A ground image like Fig.22 is
photographed with the wireless camera attached at RC helicopter during the flight. The
marker is detected by the procedures of Fig.9 using image processing program. A
binarization was performed to the inputted image from the wireless camera and the outline
on the marker was extracted like Fig. 23. The direction feature point was detected from the
image of the ground photographed by the wireless camera like Fig.24.
Fig. 25 shows the measurement results on the display of a personal computer used for the
calculation. The measurement values in Fig.25 were x-coordinate=319, y-coordinate=189,
z-coordinate = 837, angle =10.350105, angle = -2.065881, and angle '=37.685916. Since
our proposal image input method which can improve blurring was used, the position and
the posture were acquirable during flight. However, since the absolute position and posture
of the RC helicopter were not measureable by other instrument during the flight. We
confirmed that by the visual observation the position and the posture were acquirable
almost correctly.
Length 360mm(Body) , 62mm(Frame)

Width 90mm
Height 160mm
Gross load 195g
Diameter of a main rotor 350mm
Gear ratio 9.857:1
Motor XRB Coreless Motor
Table 7. The specification of RC helicopter
Fig. 22. An image photographed by the wireless camera
Fig. 23. The result of marker detection
Fig. 24. The result of feature point extraction

Fig. 25. The measurement results during flight

At the last, the autonomous flight control experiment of the RC helicopter was performed by
detecting the marker ,calculating the position and the posture,and fuzzy control. Fig. 26
shows a series of scenes of a hovering flight of the RC helicopter. The results of image
processing can be checked on the display of the personal computer. From the experimental
results, the marker was detected and the direction feature point was extracted correctly
during the autonomous flight. However, when the spatial relation of the marker and the RC
helicopter was unsuitable, the detection of position and posture became unstable, then the
autonomous flight miscarried. We will improve the performance of the autonomous flight
control for RC helicopter using stabilized feature point detection and stabilized position
estimation.
7. Conclusion
This Chapter described an autonomous flight control for micro RC helicopter to fly indoors.
It is based on three-dimensional measuring by a micro wireless camera attached on the
micro RC helicopter and a circular marker put on the ground. First, a method of measuring
the self position and posture of the micro RC helicopter simply was proposed. By this
method, if the wireless camera attached on the RC helicopter takes an image of the circular
marker, a major axis and a minor axis of the circular marker image is acquirable. Because
this circular marker has a cut part, the direction of the circular marker image can be
Time 1 Time 2
Time 3 Time 4
Fig. 26. The experiment of autonomous flight
acquired by extracting the cut part as a direction feature point of the circular marker.
Therefore, the relation between the circular marker image and the actual circular marker can
be acquired by a coordinate transform using the above data. In this way, the three-
dimensional self position and posture of the micro RC helicopter can be acquired with
image processing and weak perspective projection. Then, we designed a flight control
system which can perform fuzzy control based on the three-dimensional position and
posture of the micro RC helicopter. The micro RC helicopter is controlled by tracking the
circle marker with a direction feature point during the flight.
In order to confirm the effectiveness of our proposal method, in the experiment, the position
and the posture were calculated using an image photographed with a wireless camera fixed
in a known three-dimensional position. By the experiment results, the calculated values near
the actually measuring values were confirmed. An autonomous flight control experiment
was performed to confirm that if our proposal image input method is effective when using a
micro wireless camera attached on the micro RC Helicopter. By results of the autonomous
flight control experiment of the RC helicopter, the marker was detected at real-time during
the flight, and it was confirmed that the autonomous flight of the micro RC helicopter is
possible. However, when the spatial relation of the marker and the RC helicopter was
unsuitable, the detection of position and posture became unstable and then the autonomous
flight miscarried. We will improve the performance of autonomous flight control of the RC
helicopter to more stable. We will improve the system so that the performance of the
autonomous flight control of the RC Helicopter may become stability more.
8. Reference
Amida, O.; Kanade, T. & Miller, J.R. (1998). Vision-Based Autonomous Helicopter Research
at Carnegie Mellon Robotics Institute 1991-1997, American Helicopter Society
International Conf. Heli, Japan.
Harris, C. & Stephens, M. (1988). A Combined Corner and Edge Detecter, Proc. 4th Alvey
Vision Conf., pp.147-151.
Nakamura, S.; Kataoka, K. & Sugeno, M. (2001). A Study on Autonomous Landing of an
Unmanned Helicopter Using Active Vision and GPS, J.RSJ Vol.18, No.2, pp.252-260.
Neumann, U. & You, S. (1999). Natural Feature Tracking for Augmented-reality, IEEE
Transactions on Multimedia, Vo.1, No.1, pp.53-64.
Ohtake, H.; Iimura, K. & Tanaka, K. (2009). Fuzzy Control of Micro RC Helicopter with
Coaxial Counter-rotating Blades, journal of Japan Society for Fuzzy Theory and
Intelligent Informatics, Vol.21, No.1, pp.100-106.
Schmid, C.; Mohr, R. & Bauckhage, C. (1998). Comparing and Evaluating Interest Points,
Proc. 6th Int. Conf. on Computer Vision, pp.230-235.
Shi, J. & Tomasi, C. (1994). Good Features to Track, Proc. IEEE Conf. Comput. Vision Patt.
Recogn., pp.593-600.
Smith, S. M.; & Brady, J. M. (1997). SUSAN - A New Approach to Low Level Image
Processing, Int. J. Comput. Vis., vol.23, no.1, pp.45-78.
Sugeno, M. et al. (1996). Inteligent Control of an Unmanned Helicopter based on Fuzzy
Logic., Proc. of American Helicopter Society 51st Annual Forum., Texas.
Tanaka, K. (1994). Advanced Fuzzy Control, Kyoritsu Shuppan Co.,LTD, Japan.
Wang, G.; Fujiwara, N. & Bao, Y. (1997). Automatic Guidance of Vehicle Using Fuzzy
Control. (1st Report). Identification of General Fuzzy Steering Model and
Automatic Guidance of Cars., Systems, Control and Information, Vol.10, No.9, pp.470-
479.
Bao, Y.; Takayuki, N. & Akasaka, H. (2003). Weak Perspective Projection Invariant Pattern
Recognition without Gravity Center Calculation, journal of IIEEJ, Vol.32, No.5,
pp.659--666
Bao, Y. & Komiya, M. (2008). An improvement Moravec Operator for rotated image, Proc. of
the ADVANTY 2008 SYMPOSIUM , pp.133-138.
12
Hierarchical Control Design of a UAV Helicopter

Ali Karimoddini, Guowei Cai, Ben M. Chen, Hai Lin, Tong H. Lee
Graduate School for Integrative Sciences and Engineering (NGS) and Department of
Electrical and Computer Engineering, National University of Singapore
Singapore
1. Introduction
Nowadays, control design of Unmanned Aerial Vehicles (UAVs) has emerged as an attractive
research area, due to the wide range of UAV applications in various military and civilian
areas such as terrain and utility inspections, coordinated surveillance, search and rescue
missions, disaster monitoring, rapid emergency response, aerial mapping, trafc monitoring,
and reconnaissance missions (see, e.g., (Metni et al., 2007), (Kuroki et al., 2010 ), (Campbell
& Campbell, 2010 )). They can also be used as complex test-bed dynamic systems for
implementation and verication of the control schemes for different research purposes (Kim
& Sukkarieh, 2007), (Saripalli et al., 2003), (Bortoff, 1999). Several research groups are involved
in the modeling and control of UAVs (Bortoff, 1999), (Gavrilets et al., 2000), (Cai et al., 2006).
The control methods such as the neural network approach (Enns & Si, 2003), the differential
geometry method (Isidori et al., 2003), feedback control with decoupling approach (Peng et
al., 2009), and the model predictive approach (Shim et al., 2003) have been applied for the
ight control of the UAV helicopters. In this chapter, however, we have used an analytical
approach to design and analyze the whole system including the inner-loop and the outer-loop
controllers for a small-scale UAV helicopter. Here, in the proposed hierarchical structure, the
inner-loop is responsible for the internal stabilization of the UAV in the hovering state and
control of the linear velocities and heading angle angular velocity whereas the outer-loop is
used to drive the system, which is already stabilized by the inner-loop, to follow a desired path
while keeping the system close to the hovering state. This hierarchical strategy is an intuitive
way of controlling such a complex system. However, there is another reason that compels us
to employ such a control structure. Indeed, the UAV model cannot be fully linearized, since,
in practice, we cannot expect the heading angle of the UAV to be restricted to a small range
of variation as depending on the mission, the heading of the UAV could be in any direction.
This will impose some kind of nonlinearity on the system, which can be modeled by a simple
transformation. To handle this semi-linearized model of the UAV, we can separate the linear
and nonlinear parts, and then control the linear part in the inner-loop and the nonlinear part
in the outer-loop.
In this hierarchy, for the inner-loop, we have used an H controller to both stabilize the system
and suboptimally achieve the desired performance of the UAV attitude control. Assuming
that the inner-loop has already been stabilized by an H controller, a proportional feedback
controller combined with a nonlinear compensator block have been used in the outer-loop to
bring the UAV into the desired position with desired heading angle.
Although designing a proportional feedback controller for SISO systems is straightforward,

the situation for MIMO systems is different. This is because, in MIMO systems, it is not easy
to use the popular tools, such as the Nyquist stability theorem or the root-locus approach,
that are well-established for the SISO systems. The current approaches employed for MIMO
systems are rather complicated and are mostly extensions of the existing results for SISO
systems (Wang et al., 2007). In this chapter, we propose a design method of a decentralized
P-controller for MIMO systems that, although conservative, can be effectively used in practical
problems, particularly for the case that the system is close to a decoupled system. The
approach is an extension of the Nyquist theorem to MIMO systems, and its application to
the NUS UAV system provides a successful ight controlled system.
The test-bed is Helion (Fig. 1), the rst developed UAV helicopter in our NUS UAV research
group (Peng et al., 2007). In (Cai et al., 2008a), a systematic procedure for the construction of
a small-scale UAV helicopter is described, and in (Cai et al., 2005), the hardware parts of the
NUS UAV, including both the avionic system and the ground station, are illustrated in detail.
Fig. 1. The NUS UAV helicopter.

The remaining parts of this chapter are organized as follows. In Section II, the model and
the structure of the NUS UAV is described. The UAV model consists of two decoupled
subsystems. In Section III, a hierarchical controller, including an inner-loop and an outer-loop
controller, is designed for both subsystems. Actual ight tests are presented in Section IV, and
the chapter is concluded in Section V. For the convenience of the reader, a nomenclature part
is provided at the end of this chapter.
2. Modeling and structure of the UAV helicopter

A typical UAV helicopter consists of several parts: physical parts such as engine and fuselage;
ground station to monitor the ight situation and collect realtime ight data, and the avionic
system to implement the control strategy to have an autonomous ight control. Among these
elements, the avionic part is in the center of our interest in this chapter and we will focus on
the control structure which is embedded in the airborne computer system. Here, the avionic
Hierarchical Control Design of a UAV Helicopter 241
system consists of an airborne computer system which can be extended modularly by some
extension boards such as A\D card, DC-DC convertor card, and serial communication board.
In addition, the avionic system has been equipped with some analog and digital sensors to
collect the information of the current state of the UAV. The major sensor used in the avionic
system, is the NAV-IMU sensor. The IMU sensor provides three axis velocities, acceleration
and angular rates in the body frame, as well as longitude, latitude, relative height and heading,
pitch and roll angles. Moreover, the avionic system has a fuel level sensor as well as a magnetic
RPM sensor to measure the speed of the rotor. Furthermore, it comprises ve servo actuators
that could manipulate the helicopter to move forward and backward, up and down, to turn
left and right, to regulate the nose angle and nally to control the spinning speed of the
rotor. All of these servos are controlled by a servo board as a local controller. In addition,
the servo board gives the ability of driving the servo system into either the manual mode or
the automatic mode. In the manual mode, a pilot can drive the helicopter by a radio controller
which is useful in the emergency situations; however, in the automatic mode the helicopter is
under the control of the computer system and all control signals are generated by the avionic
system and the computer board, autonomously.
Using some basic physical principles, we can obtain a general nonlinear UAV model. These
principles will result in several equations that represent the effects of different factors such
as gravity, the main rotor, and tail rotor forces and moments. The model equations will be
obtained in two coordinate systems: the body frame and the ground frame. The body frame
is centered at the center of gravity of the UAV, and the ground frame is an NED (North - East
- Down) coordinate system (Stevens & Lewis, 1992) with a xed origin at the starting point
of the UAV ight. Clearly, the UAV dynamic equations should be derived in the body frame,
while the position of the UAV is considered in the ground frame.
Neglecting the gyroscopic effect of the engine-driven train, the equations of the helicopter
motion in the body frame are obtained as follows:
b =
V b V b + Bbg + m1F (1)

b = J b J
1
b + J 1 M (2)
where in these equations, denotes the cross product of the vectors, and the concatenation
of two matrices or vectors represents the normal matrix multiplication. Moreover, F and M
are the resultant force and moment in the body frame, including those generated from the
main rotor, the tail rotor and the fuselage. Other symbols denition can be found in the
nomenclature part provided at the end of this chapter.
The Euler angles that show the orientation of the body frame relative to the ground frame are
as follows:

1 tan sin tan cos
= 0 cos sin b (3)
sin cos
0 cos cos

where is a vector that contains the Euler angles to describe the attitude of the
helicopter with respect to the NED frame.
The relation between the UAV position in the ground frame and the UAV velocity in the body
frame is:
P g = Bb V
b (4)
where Bb is the transformation matrix from the ground frame to the body frame, which has
the following form:

cos cos cos sin sin
Bb = cos sin + sin sin cos cos cos + sin sin sin sin cos (5)
sin sin + cos sin cos sin cos + cos sin cos cos cos
The details of this UAV model are described in (Peng et al., 2007). From the above model
description, it can be seen that the UAV model is nonlinear. Furthermore, the main problem
encountered in the modeling of our UAV is that the process of buying a radio-control
helicopter from the market and upgrading it to an autonomous ying vehicle leaves us with
many unknown parameters and aerodynamic data. Therefore, for practical reasons, we need
to linearize the UAV model and then, identify the resulting linearized model through the
recorded in-ight data. The in-ight data can be collected through the manual mode and by
injecting perturbed input signals to the ying helicopter. We have obtained a linearized model
at the hovering state as it has been presented in (Cai et al., 2006). By hovering, we mean that
0 = 0,
V 0 = 0, 0 = 0, 0 = 0 . The obtained linearized model is as follows:
x = Ax + Bu + Ew (6)
where the state variable x includes 11 variables as x = [ Vzb (m/s) zb (rad/s) wz f (rad/s)
Vxb (m/s) Vyb (m/s) xb (rad/s) yb (rad/s) (rad) (rad) as (rad) bs (rad) ] . These parameters
are shown in Fig. 2. wz f is the yaw rate feedback, which is related to pedal by a rst-order
differential equation (Cai et al., 2008b).
Furthermore, the control input u includes commands to the servos embedded for the control
of the helicopter blades as u = [ col (rad) pedal (rad) roll (rad) pitch (rad) ] .
Matrices
A, B, and E areobtained as follows:

A1 038 B1 032 E1 032
A= ,B= ,E=
083 A2 082 B2 081 E2
where
0.6821 0.1070 0 15.6491 0
A1 = 0.1446 5.5561 36.6740 , B1 = 1.6349 58.4053 ,
0 2.7492 11.1120 0 0

0 0 0.1778 0
0 0 0 0.3104

0 0 0.3326 0.2051
0.5995
0 0 0.0802 0.2940
E1 = 1.3832 , B2 = , E2 = ,
0 0 0 0
0
0 0 0 0
0.0496 2.6224 0 0
2.4928 0.1740 0 0

0.1778 0 0 0 0 9.7807 9.7808 0
0 0.3104 0 0 9.7807 0 0 9.7807

0.3326 0.5353 0 0 0 0 75.7640 343.86

0.1903 0.2940 0 0 0 0 172.620 59.958
A2 = .
0 0 1 0 0 0 0 0

0 0 0 1 0 0 0 0
0 0 0 1 0 0 8.1222 4.6535
0 0 1 0 0 0 0.0921 8.1222
Remark 1. In the linearized model described by (6), the saturation level for the servos is | max | = 0.5.
We need to provide a control law such that the resulting control signals always remain within the linear
unsaturated range.
Fig. 2. Helicopter states in the body frame.
Although (6) describes the relation between the control input and the state variable x, it still
does not describe the whole dynamics of the system, and particularly,
P and are not reected
in the model. Thus, considering (2) and (4), a more complete model containing P and is as
follows:
x = Ax + Bu + Ew

= zb (7)

Pg = Bb Vb
Remark 2. Matrix Bb includes some time-dependent terms. Therefore, matrix Bb can not be considered
as a constant term and it is not simple to integrate both sides of (7) in order to obtain the position in
the ground frame. This is due to the fact that the body frame is a moving coordinate system. Hence, to
obtain the displacement, it is necessary to rst obtain the velocities in a xed coordinate system such as
the ground frame. Then, the displacement can be calculated by integrating of the velocity vector in the
xed coordinate system.
The presence of nonlinear terms of Bb , in the third equation of (7), makes it difcult to design a
controller for the system; however, we can further simplify the model. Indeed, matrix Bb in (5),
which introduces some nonlinear terms to the model, can be linearized at the hovering state.
In practice, the heading angle of the helicopter can take any arbitrary value; however, the roll
and pitch angles are usually kept close to the hovering condition. Therefore, linearizing matrix
Bb at the hovering state will result in:

cos sin 0
R 021
Bb = sin cos 0 = (8)
012 1
0 0 1
The physical interpretation is that by keeping and close to zero, the Euler rotation in a
three-dimensional space will be converted into a simple rotation in a two-dimensional space
with respect to . In this case, the rotation matrix is:

cos sin
R= (9)
sin cos
In the following section, as we design the outer-loop controller, it will be shown that this new
formulation of Bb helps us to keep the system decoupled, even after using the outer-loop
controller.
The semi-linearized UAV model, presented in (7) can be controlled in separate parts: the linear
part in the inner-loop and the nonlinear part in the outer-loop, as described in the following
section.
3. Controller design
We use a hierarchical approach to design a controller for the UAV (Fig. 3). In this framework,
the system is stabilized in the inner-loop, and then it is driven to track a desired trajectory in
the outer-loop. Besides this rational strategy, there is another reason that compels us to select
this particular architecture. As mentioned previously, we need to derive the model equations
in two coordinate systems: the velocities and accelerations should be obtained in the body
coordinate system as a moving frame, whereas the displacement must be derived by the
integration of the velocities in a xed frame. The velocity transformation from the body frame
to the ground frame, modeled by matrix Bb , imposes some kind of nonlinearity as described
in (7). This nonlinearity can be handled in the outer-loop. Using this control strategy, we have
separated the nonlinear term from the linear part and put it in the outer-loop.
Pre f - Outer-loop ui - Inner-loop - Helicopter

- Control Law - Control Law
Fig. 3. Schematic diagram of the ight control system.
In this control architecture, the references for the inner-loop controller, u i , are the linear
velocities (Vxb , Vyb , and Vzb ) and the yaw rate, zb , which should all be provided by the
outer-loop. The outer-loop, however, is responsible for the control of the position and heading
angle of the UAV and will guide the UAV to follow a desired trajectory. Therefore, the
references for the outer-loop are the position ( Xr , Yr , Zr ) and the yaw angle r . In other words,
the UAV will follow the generated path by the position and the yaw angle control in the
outer-loop, and the linear velocity and the angular rate control in the inner-loop.
looking at matrices A, B, and E in (6), we can see that, the system has been already decoupled
into two independent parts. Therefore, (6) can be rewritten into two separate subsystems as
follows:
x1 = A1 x1 + B1 u1 + E1 w (10)
x2 = A2 x2 + B2 u2 + E2 w (11)
where x1 = [ Vzb (m/s) zb (rad/s) wz f (rad/s) ] , u1 = [ col pedal ] , x2 = [ Vxb (m/s) Vyb (m/s)
xb (rad/s) yb (rad/s) (rad) (rad) as (rad) bs (rad) ] , and u2 = [ roll (rad) pitch (rad) ] .
Considering (7), (8), (10), and (11), the above-mentioned hierarchical control strategy can be
implemented for the decoupled model of our UAV, as shown in Fig. 4 and Fig. 5. In these
gures, subscripts g, b, and r stand for ground frame, body frame, and reference, respectively.
Moreover, matrices C1 and C2 are:

100 10000000
C1 = , C2 = (12)
010 01000000
Fig. 4. Control schematic for Subsystem 1.
Fig. 5. Control schematic for Subsystem 2.

Due to the special structure of the linearized form of Bb , Subsystem 1 is a fully linearized
model. However, in the outer-loop of Subsystem 2, the term R1 appears as a nonlinear
element, and thus Subsystem 2 is more complicated than Subsystem 1. In the following parts,
we will describe the control design for both subsystems.
3.1 Controller for subsystem 1

Subsystem 1 is a fully linearized model so that linear design tools can be applied to the system.
We will use the H control design technique for the inner-loop and a P-controller for the
outer-loop.
3.1.1 Inner-loop controller

Using an H controller for the inner-loop, both robust stability and proper performance of the
system can be achieved simultaneously. To design the H controller, using notation analogous
with (Chen, 2000), we dene the measurement output simply as the state feedback in the form
of y1 = C11 x with C11 = I. Since our primary task is to design a control law to internally
stabilize the system, and to achieve a good response of the state variables that are directly
related and linked to the outer-loop, while considering the constraints on the inputs and some
state variables, we dene the controlled output h1 in the form of h1 = C12 x + D12 u, where

023
3.1623 0 44.7214 0
0
C12 =
0 3.1623 0 , D12 =
0 28.2843 (13)
032
0 0 1.7321
The nonzero entries of C12 and D12 are used for tuning the controller, and here, are determined
experimentally to achieve the desired performance. Meanwhile, the H design guarantees
internal stability and robustness of the system. Indeed, H control design minimizes the effect
of the wind gust disturbance, i.e., minimizes the H norm of the closed-loop transfer function
from the disturbance w to the controlled output h1 , denoted by T1 . The H norm of the transfer
function T1 is dened as follows:
T1 = sup max [ T1 ( j )] (14)

0 <
where max [.] denotes the maximum singular value of the matrix.
It should be highlighted that the H norm is the worst case gain in T1 (s). Therefore,
minimization of the H norm of T1 is equivalent to the minimization of the disturbance
effect from the disturbance w to the controlled output h1 in the worst case situation. Having
the matrix C12 and D12 , one can nd the which is the optimal H performance for the

closed-loop system from the disturbance input w to the controlled output h1 over all the
possible controllers that internally stabilize the system. As practically, is not achievable,
we will try to reach which is slightly larger than .
With this choice of the control parameters, D11 and D12 are full rank and the quadruples
( A1 , B1 , C12 , D12 ) and ( A1 , E1 , C11 , D11 ) are left invertible and are free of invariant zeros.
Therefore, we have a so-called regular problem, for which we can use well-established H
control theory (Chen, 2000). The resulting closed loop system suboptimality minimizes the
H norm of the transfer function from the disturbance w to the controlled output h1 . To design
this controller we consider the control law in the following form:
u1 = F1 x1 + G1 r1 (15)
where r1 = (Vzr , zr ) is the reference signal generated by the outer-loop controller, G1 =

(C1 ( A1 + B1 F1 )1 B1 )1 is the feedforward gain, and F1 is the H control gain that can be
achieved as follows:

F1 = ( D12 D12 )1 ( D12

C12 + B1 P1 )) (16)
where matrix P1 is the positive semi-denite solution of the following H algebraic Riccati
equation:
A1 P1 + P1 A1 + C12

C12 + P1 E1 E1 P1 /2

( P1 B1 + C12 D12 )( D12 D12 )1 ( D12

C12 + B1 P1 ) = 0 (17)
For this system and these control parameters values, the value of is 1.4516. Hence, we
select = 1.4616. Therefore, matrices F1 and G1 are obtained as follows:

0.0935 0.0005 0.0027 0.1371 0.0066
F1 =
0.0008 0.0364 0.0481
, G1 = 0.0020 0.2748 (18)
To evaluate the controller performance and its effect on the disturbance attenuation, we
simulated the closed loop system with an initial state of x1 (0) = [1.5 0 0] , and also we injected
wind gust disturbance for 20 sec (Fig. 6). The injected disturbance has a maximum amplitude
of 3 m/s along the z axis (the other directions do not affect the dynamics of Subsystem 1). The
controlled system reaches the steady hovering state after 3.5 sec, and the disturbance effect is
reduced to less than 0.25%. The control inputs are within the unsaturated region.
Wind gust (m/s) 4
0
0 5 10 15 20 25 30 35 40
2
V (m/s)
0
b
z
2
0 5 10 15 20 25 30 35 40
0.5
(rad/s)
0
b
z
0.5
0 5 10 15 20 25 30 35 40
0.2
col
Inputs
0
ped
0.2
0 5 10 15 20 25 30 35 40
Time(s)
Fig. 6. Simulation of the inner-loop of Subsystem 1.
3.1.2 Outer-loop controller

In the outer-loop of Subsystem 1, we use a P-controller K P1 (Fig. 7). We can redraw this system
as shown in Fig. 8, in which Gin1 = S1 C1 (SI ( A1 + B1 F1 ))1 B1 G1 . It can be shown that Gin1 is
a 22 multi-variable system. Unfortunately, in general, designing a P-controller for an MIMO
system is difcult. However, if we consider K p1 in diagonal form as K p 1 = k p 1 I22 , we can
apply the generalized Nyquist theorem (Postlethwaite & MacFarlane, 1979) to design k p1 such
that it stabilizes the system as described in the following part.
+ ?g- - +g - X1 = A1 X1 + B1 u1
-+ - kp - G1 - C1 - 1/s

Zr

VZr

6 Zg

r Zr g
F
Fig. 7. Control structure of Subsystem 1.
+ ?-
- +j - k p1 - Gin1
Zr VZ Zg
r
r Z g
r
Fig. 8. Redrawing the control structure of Subsystem 1.
3.1.3 Stability analysis

The characteristic loci of Gin1 are shown in Fig. 9, where the dash-dot lines correspond to the
innite values. In Subsystem 1, Fig. 8, the inner-loop has already been stabilized, using an H
controller. Therefore, due to the presence of the integral term, Gin1 has two poles at the origin
and the remaining poles are in the LHP plane. Hence, Gin1 has no pole in the Nyquist contour.
It follows from the form of the characteristic loci of Gin1 in Fig. 9, that k p1 (0 , ) will keep
the entire system stable. However, in practice, we are subjected to the selection of small values
of k p1 to avoid saturation of the actuators. k p1 = 1.5 is a typical value.
Fig. 9. Characteristic loci of Gin1 .

3.1.4 Tuning the controller
With the above outer-loop controller, the stability of the whole system has been achieved;
however, the controller in the form of K p1 = k p1 I22 with only one control parameter is
not an appropriate choice. We need to have more degrees of freedom to tune the controller
and achieve better performance. By considering the proportional feedback gain K p1 =
diag{K p11 , K p12 } as a diagonal matrix, we have more degrees of freedom and can control each
of the output channels in a decentralized manner, while keeping the system decoupled.
Uncertainty analysis usually is used to investigate the effect of the plant uncertainty. Here,
we borrow this idea to analyze the effect of deviation of the diagonal entries of the matrix
K p1 = k p1 I22 in the controller part. Alternatively, one can dene K p1 as follows:

K p11 0 K p 0
K p1 = = k p1 I22 + 11 (19)
0 K P12 0 K P
12

The objective is to design = diag K p , K p such that it does not affect the stability of
11 12
the system. In fact, is the tuning range (Fig. 10).
Following from Fig. 10, one can extract the internal model of the system as:

y = Gin1 k p1 ( I + Gin1 k p1 )1 v + ( I + Gin1 k p1 )1 Gin1 z
(20)
x = ( I + Gin1 k p1 )1 v ( I + Gin1 k p1 )1 Gin1 z
To simplify the notation, (20) can be rewritten as:

y = G11 v + G12 z
(21)
x = G21 v + G22 z
-
+ ?

v - + e - k p1 I - +n - Gin1 y

6
-
x z
Fig. 10. Tuning the controller using uncertainty analysis.
v - G11 k -
-+ y
6

-
G21
B
B
B
- G12 B
B
B
z BN
- G22 -+ k -x
6
Fig. 11. Redrawing Subsystem 1 for uncertainty analysis.

Therefore, Fig. 10 can be redrawn as it is shown in Fig. 11. In the new diagram, since the
nominal system with = 0 is stable, all Gij are stable. G11 , G12 , and G21 are outside the
uncertain loop and cannot be affected by block ; however, the loop includes G22 and may
affect the internal stability of the system due to perturbations of the elements of . Since G22
and are stable, according to the generalized Nyquist theorem, the characteristic loci of the
loop transfer function should not encircle the point (-1+j0), or equivalently, |i ( G22 )| < 1.
To satisfy this condition, since | i ( G22 )| ( G22 ) sup(( G22 )) = G22 , it is

sufcient that G22 < 1. Using norm properties, we have:
G22 G22 (22)
Therefore, the sufcient condition for the stability of the system is:
G22 < 1 (23)
For these values of the controller and plant and for a frequency range of (0, 10000), we obtain
G22 = 0.6986. Therefore, the perturbation of K p1 should be such that 1.4315.
Recall that has a diagonal structure, and hence, all diagonal entries of K p1 should have less
than a 1.4315-unit deviation from their nominal value. In fact, using this approach, we rst
obtained a nominal controller that provides the stability of the system, and then, we attempted
to tune the controller to improve the performance, while keeping the system stable. After
tuning the controller, the value of K p1 = diag{0.5, 0.7} was selected as an appropriate value
that satises the above mentioned condition and gives a satisfactory performance. The method
is conservative as is structured and real, but applying to the UAV plant it has provided
sufcient degree of freedom for tuning the controller and improving the performance.
To simulate the resulting system, let the outer-loop reference be ( Zr , r ) = (2, 0.5) and the
current position and heading angle be ( Zg , g ) = (0, 0). The system will reach its target after
approximately 8 sec as shown in Fig. 12.
2 0.5
1.8 0.45
1.6 0.4
1.4 0.35
1.2 0.3
(rad)
Zg(m)
1 0.25
0.8 0.2
0.6 0.15
0.4 0.1
0.2 0.05
0 0
0 10 20 30 40 0 10 20 30 40
Time(s) Time(s)
Fig. 12. Simulation of the outer-loop of Subsystem 1.
3.2 Designing the controller for subsystem 2

3.2.1 Inner-loop controller
For Subsystem 2, described by (11), we use an H controller for the inner-loop
controller design, similar to Subsystem 1. Analogous with Subsystem 1, we dene h2
as h2 = C22 x2 + D22 u2 , where

028
0.3162 0 0 0 0000

0 0.3162 0 0 0000

0 0 0.3162 0 0 0 0 0

C22 = 0 0 0 0.3162 0 0 0 0

0 0 0 0 1000

0 0 0 0 0100

0 0 0 0 0010
0 0 0 0 0001

5.4772 0
D22 = 0 5.4772
082
= 0.0731, and choosing = 0.0831, we have:
With these parameters, we obtain
0.0017 0.1683 0.0486 0.0081 1.9336 0.1974 0.3227 2.1444
F2 =
0.0815 0.0461 0.0087 0.0535 0.3908 1.0690 1.1712 0.4659
Moreover, G2 = (C2 ( A2 + B2 F2 )1 B2 )1 , is the feedforward gain for Subsystem 2 and can

be calculated as:
0.0029 0.2335
G2 = 0.0978 0.0632
The simulation of the system is shown in Fig. 13. In this gure, the initial state of the system
is x2 (0) = [1.5 0 0 0 0.17 0 0 0] . The injected disturbance has a maximum amplitude of 10
m/s along the x and y axes (the z direction does not affect the dynamics of Subsystem 2). The
controlled system reaches the steady hovering state after 3.5 sec, and the disturbance effect is
reduced to less than 0.25%. In addition, the control inputs are within the unsaturated region.
10
Wind gust
Xaxis
Yaxis
(m/s)
10
0 5 10 15 20 25 30 35 40
5
Vx
Velocity
b
(m/s)
0 Vy
b
5
0 5 10 15 20 25 30 35 40
Ang. Velocity
1

x
b
0 y
(rad/s)
1
0 5 10 15 20 25 30 35 40
0.5

angle
(rad)
0.5
0 5 10 15 20 25 30 35 40
0.5
rol
Inputs
0
pitch
0.5
0 5 10 15 20 25 30 35 40
Time(s)
Fig. 13. Simulation of the inner-loop of Subsystem 2.
3.2.2 Outer-loop controller

Although the outer-loop of Subsystem 2 is similar to the outer-loop of Subsystem 1, the main
difference lies in the presence of the nonlinear term, R1 , in the outer-loop of Subsystem 2, as
shown in Fig. 4. In this structure, it can be seen that the error signal is the difference between
the actual position and the reference position, which both are in the ground frame. Therefore,
the resulting control signal, which is the reference for the inner-loop, will be obtained in the
ground frame; however, the inner-loop is in the body frame. Hence, it is reasonable that
we transform the control signal to the body frame before delivering it to the inner-loop as
the reference to be tracked. To implement this idea, we can use the transformation term,
R, to obtain a control signal in the body frame. The new structure is shown in Fig. 14, in
which Gin2 = C2 (SI ( A2 + B2 F2 ))1 B2 G2 is a 2 2 multi-variable system. In Fig. 15, it is
shown that the inner-loop block Gin2 is very close to a decoupled system with equal diagonal
elements. Indeed, Subsystem 2 corresponds to the dynamics of the helicopter for the x y
plane movement. In practice, we expect the dynamics of the UAV in the x and y directions
to be similar and decoupled, since the pilot can easily drive the UAV in either of directions
independently. Using this concept, we can take the block Gin2 out so that the two rotation
matrices R and R1 will cancel each other.
Fig. 14. Control diagram of Subsystem 2.
G11 G12
0 0
Magnitude (dB)
20 20
40 40
60 60
80 80
100 100
1 0 1 2 1 0 1 2
10 10 10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)
G21 G22
0 0
Magnitude (dB)
20 20
40 40
60 60
80 80
100 100
2 0 2 1 0 1 2
10 10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)
Fig. 15. Bode plot of entries of Gin2

The remaining job is simple, and we can repeat the procedure of designing the outer-loop
controller for Subsystem 1 and design a P-controller in the form of K p2 = diag{K p21 , K P22 }
that stabilizes Gin2 (Fig. 16). As an appropriate choice of control parameters, we can select
K p2 = diag{0.3 , 0.3}. Rationally, K p21 and K P22 should be the same, since we expect a similar
behavior of the UAV system in the x and y directions.
Fig. 16. Redrawing the control diagram of Subsystem 2.

For an outer-loop reference at ( xr , yr ) = (2, 2) and the current UAV position at ( x g , y g ) =
(0, 0), the simulation results are shown in Fig. 17, in which the UAV reaches to the desired
position after approximately 10 sec, smoothly and without overshooting.
2.5 2.5
2 2
1.5 1.5
X (m)
Yg(m)
g
1 1
0.5 0.5
0 0
0 10 20 30 40 0 10 20 30 40
Time(s) Time(s)
Fig. 17. Simulation of the outer-loop of Subsystem 2.
4. Experimental results
Before using the designed controller in an actual ight test, we rst evaluate it through a
hardware-in-the-loop simulation platform (Cai et al., 2009). In this platform, the nonlinear
dynamics of the UAV has been replaced with its nonlinear model, and all software and
hardware components that are involved in a real ight test remain active during the
simulation. Using the hardware-in-the-loop simulation environment, the behavior of the

system is very close to the real experiments.
Then, we conducted actual ight tests to observe the in-ight behavior of the helicopter. First,
we used the UAV in the hovering state for 80 sec. Figure 18 shows the state variables in the
hovering experiment at ( x, y, z, ) = (16, 34, 10, 1.5). To evaluate the hovering control
performance, the position of the UAV is depicted in a 2D x y plane in Fig. 19. As can be seen,
the position of the UAV has at most a 1-meter deviation from the desired hovering position,
which is quite satisfactory. The control inputs are also shown in Fig. 20. All of the control
inputs are within the unsaturated region.
Next, we used the UAV to follow a circle with a diameter of 20 meters as a given
trajectory. This trajectory determines the reference ( x (t), y(t), z(t), (t)) for the system. With
this trajectory, the UAV should complete the circle within 63 sec, while keeping a xed
altitude. Then, it will hover for 7 sec. In Fig. 21, it is shown that the UAV is able to
follow this trajectory successfully. The UAV path tracking in the x y plane is shown
in Fig. 22. Moreover, to have a better insight of the system behavior, all of the states of
the UAV and the control inputs are represented in Fig. 23 and Fig. 24, respectively. These
results show that the UAV is able to track the desired trajectory in situations close to the
hovering state. The small deviations in the hovering mode or path tracking mode could
be due to environmental effects such as wind disturbances or the GPS signal inaccuracy
as with the installed sensors the measurable steady accuracy of the heading angle is 2.5o
and the positioning accuracy of the GPS is 3m(1). Videos of the hovering experiment
and circle path tracking are available at http://uav.ece.nus.edu.sg/video/hover.mpg and
http://uav.ece.nus.edu.sg/video/circle.mpg , respectively.
All state variables

20
Position (m)
X
g
0
Yg
20 Zg
40
0 10 20 30 40 50 60 70
Velocity (m/s)
2
Vx
b
Vy
0 b
Vz
b
2
0 10 20 30 40 50 60 70
Angular velocity Angular position
1

0
(rad)
2
0 10 20 30 40 50 60 70
1
x
b
0
y
(rad/s)
b
1
z
b
2
0 10 20 30 40 50 60 70
Time (s)
Fig. 18. State variables of the UAV for the hovering.

xy plane
30
31
32
33
y (m)
34
35
36
37
38
17.2 17 16.8 16.6 16.4 16.2 16 15.8 15.6 15.4 15.2
x (m)
Fig. 19. UAV position in x y plane at hovering
5. Conclusion
In this chapter, we presented a systematic approach for the ight control design of a
small-scale UAV helicopter in a hierarchical manner. In this structure, the lower level aims
at stabilization of the system, and the upper level focuses on the reference tracking. For
the disturbance attenuation and stabilization of the UAV, we used an H controller in the
inner-loop of the system. Due to the presence of some nonlinear terms in the outer-loop of
the system, we rst compensated for the nonlinearity by an inverse rotation; then, we used a
decentralized P-controller to enable the UAV to follow a desired trajectory. We also proposed
a new method of designing a P-controller for MIMO systems that was successfully applied
to the UAV system. The simulations and actual ight tests show the efcacy of the control
structure. In the future, we will use this structure to accomplish more complex missions
such as formation control (Karimoddini et al., 2010). Such missions will require an embedded
decision-making unit to support the tasks and to switch between the controllers. This concept
will guide us in designing a hybrid supervisory controller in the path planner level of the UAV
to comprehensively analyze the reactions between the continuous dynamics of the system and
discrete switching between the controllers (Karimoddini et al., 2009).
6. Acknowledgments
The authors gratefully acknowledge the technical support of Mr. Dong Xiangxu and Mr. Lin
Feng during the implementations and ight tests.
0.1
rol
0.05
0
0 10 20 30 40 50 60 70 80
0.1
pitch
0.05
0
0 10 20 30 40 50 60 70 80
0.1
col
0.2
0.3
0 10 20 30 40 50 60 70 80
0.01
ped
0.01
0 10 20 30 40 50 60 70 80
time (s)
Fig. 20. Control signals at hovering
20
Actual
30
Xg (m)
Reference
40
50
0 10 20 30 40 50 60 70
10
Yg (m)
20
30
0 10 20 30 40 50 60 70
5
Zg (m)
10
15
0 10 20 30 40 50 60 70
5
(rad)
5
0 10 20 30 40 50 60 70
Time (s)
Fig. 21. Tracking a desired path.

xy plane
5
Actual
Reference
0
10
15
y (m)
20
25
30
35
40
60 55 50 45 40 35 30 25 20 15 10
x (m)
Fig. 22. Circle path tracking in x y plane.
All state variables

Position (m)
0 X
g
20 Yg
Z
40 g
60
10 20 30 40 50 60 70
Velocity (m/s)
2
V
x
b
V
0 y
b
Vz
b
2
10 20 30 40 50 60 70
Angular position
4

(rad)
2

0
2
10 20 30 40 50 60 70
Angular velocity
2

x
(rad/s)
0
y
b

z
b
2
10 20 30 40 50 60 70
Time (s)
Fig. 23. States of the UAV in the circle path tracking behavior.
0.1
rol 0
0.1
0 10 20 30 40 50 60 70
0.1
pitch
0.1
0 10 20 30 40 50 60 70
0
col
0.2
0.4
0 10 20 30 40 50 60 70
0
ped
0.02
0.04
0 10 20 30 40 50 60 70
time (s)
Fig. 24. Control inputs in the circle path tracking behavior.
Nomenclature
A = system matrix
as = longitudinal blade angle
B = control input matrix
Bb = transformation matrix from the ground to the body frame
bs = lateral blade angle
E = disturbance matrix
F = resultant force in the body frame
g = acceleration due to gravity, g = [0, 0, g]
J = inertia matrix

M = resultant moment in the body frame
P = position of the UAV, P = ( X, Y, Z )
u = control input vector
V = linear velocity vector of the UAV, V = (Vx , Vy , Vz )
= angular velocity vector of the UAV, = ( x , y , z )
w = wind gust disturbance in the body frame, w = (w x , wy , wz )
= pitch angle
= roll angle
= yaw angle
roll = roll channel input
pitch = pitch channel input
pedal = pedal channel input
col = collective channel input
g = the value of in the ground frame
b = the value of in the body frame
r = the value of as control reference
7. References
Bortoff, S. (1999). The university of toronto rc helicopter: a test bed for nonlinear control,
Proceedings of the IEEE International Conference on Control Applications, IEEE, Vol. 1,
pp. 333338.
Cai, G., Peng, K., Chen, B. M. & Lee, T. H. (2005). Design and assembling of a uav helicopter
system, International Conference on Control and Automation, Vol. 2, pp. 697702.
Cai, G., Chen, B. M., Peng, K., Dong, M. & Lee, T. H. (2006). Modeling and control system
design for a uav helicopter, 14th Mediterranean Conference on Control and Automation,
pp. 16.
Cai, G., Feng, L., Chen, B. M. & Lee, T. H. (2008a). Systematic design methodology and
construction of uav helicopters, Mechatronics, 18(10): 545558.
Cai, G., Chen, B. M., Peng, K., Dong, M. & Lee, T. H.(2008b). Comprehensive modeling
and control of the yaw channel of a UAV helicopter, IEEE Transactions on Industrial
Electronics, 55(9): 34263434.
Cai, G., Chen, B. M., Lee, T. H. & Dong, M. (2009). Design and implementation
of a hardware-in-the-loop simulation system for small-scale UAV helicopters,
Mechatronics, 19(7): 10571066.
Campbell, M. E. & Wheeler, M. (2010). Vision-Based Geolocation Tracking System for
Uninhabited Aerial Vehicles, AIAA Journal of Guidance, Control, and Dynamics, 33(2):
521532.
Chen, B. M. (2000). Robust and H Control, Springer, London.
Dong, M., Chen, B. M., Cai, G. & Peng, K. (2007). Development of a real-time onboard and
ground station software system for a UAV helicopter, AIAA Journal of Aerospace
Computing, Information, and Communication, 4(8): 933955.
Enns, R. & Si, J. (2003). Helicopter trimming and tracking control using direct neural dynamic
programming, IEEE Transactions on Neural Networks, IEEE, Vol. 14, pp. 929-939, 2003.
Gavrilets, V., Shterenberg, A., Dahleh, M. & Feron, E. (2000). Avionics system for a small
unmanned helicopter performing aggressive maneuvers, Proceedings of the 19th
Digital Avionics Systems Conferences, Vol. 1, pp. 1E2/11E2/7.
Isidori, A., Marconi, L. & Serrani, A.(2003) Robust nonlinear motion control of a helicopter,
IEEE Transactions on Automatic Control, IEEE, Vol. 48, pp. 413426, 2003.
Karimoddini, A., Lin, H., Chen, B.M. & Lee, T.H. (2009) Developments in hybrid modeling
and control of unmanned aerial vehicles, Proceedings of the 7th IEEE International
Conference on Control and Automation, Christchurch, New Zealand, pp. 228233.
Karimoddini, A., Lin, H., Chen, B.M. & Lee, T.H. (2010) Hybrid formation control of the
Unmanned Aerial Vehicles, Mechatronics, doi:10.1016/ j.mechatronics.2010.09.007.
Kim, J. & Sukkarieh, S. (2007). Real-time implementation of airborne inertial-SLAM, Journal of
Robotics and Autonomous Systems, 55(1): 6271.
Kuroki, Y., Young, G.S. & Haupt, S.E. (2010). UAV navigation by an expert system for
contaminant mapping with a genetic algorithm, Journal of Expert Systems with
Applications, 37(6 ): 46874697.
Metni, N. & Hamel, T. (2007). A UAV for bridge inspection: Visual servoing control law with
orientation limits, Journal of Automation in Construction, 17(1): 310.
Peng, K., Dong, M., Chen, B. M., Cai, G., Lum, K. Y. & Lee, T. H. (2007). Design and
implementation of a fully autonomous ight control system for a uav helicopter,
Chinese Control Conference, pp. 662667.
Peng, K. , Cai, G., Chen, B. M., Dong, M., Lum, K. Y. & Lee, T. H. (2009). Design
and implementation of an autonomous ight control law for a UAV helicopter,
Automatica, 45(10), pp. 23332338.
Postlethwaite, I. & MacFarlane, A. G. I. (1979). A Complex Variable Approach to the Analysis of
Linear Multivariable Feedback Systems, in Lecture Notes in Control and Information
Sciences, Vol. 12/1979, Springer, Berlin, pp. 5876.
Saripalli, S., Montgomery, J. & Sukhatme, G. (2003). Visually guided landing of an unmanned
aerial vehicle, IEEE Transactions on Robotics and Automation, 19(3): 371380.
Shaferman, V. & Shima, T. (2008). Unmanned Aerial Vehicles Cooperative Tracking of Moving
Ground Target in Urban Environments, AIAA Journal of Guidance, Control, and
Dynamics, 31(5): 13601371.
Shim, D. H., Kim, H. J. & Sastry, S. (2003). Decentralized nonlinear model predictive control of
multiple ying robots, Proceedings of the 42nd IEEE Conference on Decision and Control,
IEEE, Hawaii, pp. 36213626.
Stevens, B. L. & Lewis, F. L. (1992). Aircraft control and simulation, Wiley, New York.
Wang, Q.G., Lin, C., Ye, Z., Wen, G., He, Y. & Hang, C. C. (2007). A quasi-lmi approach
to computing stabilizing parameter ranges of multi-loop pid controllers, Journal of
Process Control, 17(1): 5972.
13
Comparison of Flight Control System Design

Methods in Landing
S.H. Sadati, M.Sabzeh Parvar and M.B. Menhaj
Aerospace Engineering Department,
Amir Kabir University, Hafez Ave., Tehran,
Iran
1. Introduction
The development and application of most present-day systems and control theory were
spurred on by the need to resolve aerospace problems. This is roughly the problem of
analyzing and designing flight control systems for tactical missiles or aircraft. The control
laws used in current tactical missile or aircraft are mainly based on classical control design
techniques. These control laws were developed in the 1950s and have evolved into fairly
standard design procedures [1].
Current autopilot design processes contain time-and resource-consuming trial-and-error
approaches. Especially late changes in the flight control laws contribute to high cost and
delay of first delivery. The automatic landing mode development is a good example of a
process with trial-and-error design phases, because of the many parameters the system has
to be robust against. These parameters originate from different runway, terrain, ILS and
weather characteristics, and from aircraft uncertainties like configuration and landing
weight. Additional uncertainties arise from uncertain aerodynamic parameters, actuator
model uncertainties, etc. The autonomous aircraft landing is an issue that implies three main
aspects: the performance of equipment, the process models and the ethics. Generally, the
landing is not a standard flight task as it could be thinking. We consider it a nonstandard
flight stage because it has a very high sensitivity versus environment perturbation and to
the psychological factors.
In the last three decades, optimality-based designs have been considered to be the most
effective way for a guided missile engaging the target [2-4]. However, it is also known from
the optimal control theory that a straightforward solution to the optimal trajectory shaping
problem leads to a two point boundary-value problem [2], which is too complex for real-
time onboard implementation. Based on the reasons given above, advanced control theory
must be applied to a control system to improve its performance. One of the best ways to
solve this problem is to approach the artificial intelligence modeling technology based on
fuzzy logic and neural network [5].
Intelligent control is a control technology that replaces the human mind in making
decisions, planning control strategies, and learning new functions whenever the
environment does not allow or does not the presence of a human operator. Artificial neural
networks and fuzzy logic are two potential tools for use in applications in intelligent control
engineering. Artificial neural networks offer the advantage of performance improvement
through learning by means of parallel and distributed processing. Many neural control
schemes with back propagation training algorithms, which have been proposed to solve the
problems of identification and control of complex nonlinear systems, exploit the nonlinear
mapping abilities of neural networks [6,7]. Recently, adaptive neural network algorithms
have also been used to solve highly nonlinear flight control problems. A fuzzy logic-based
design that can resolve the weaknesses of conventional approaches has been cited above.
The use of fuzzy logic control is motivated by the need to deal with highly nonlinear flight
control and performance robustness problems. It is well known that fuzzy logic is much
closer to human decision making than traditional logical systems. Fuzzy control based on
fuzzy logic provides a new design paradigm such that a controller can be designed for
complex, ill-defined processes without knowledge of quantitative data regarding the input-
output relations, which are otherwise required by conventional approaches [8-11]. An
overview of neural and fuzzy control designs for dynamic systems was presented by Dash
et al.[12].
2. Aircraft model
Aircraft landing process enhanced several phases that define the so-called standard landing
trajectory. The landing operation concerning two controlled maneuvers: first for guiding the
aircraft in the horizontal plane, in order to align it onto the axe of the runway and
the second, for aircraft guiding in the vertical plane in order to do the approaching of
runway surface. Basically, the automatic landing systems provide the information for
instrument navigation along the standard trajectory. In this paper only control about the
longitudinal axis is considered. With this restriction, the equations of motion describing the
aircraft take the form [13]:
g
V = (Tbal D) g sin (1)
W
g L
= ( cos ) (2)
V W
qSC
q = (C m ) (3)
Iy
h = V sin (4)
x = V cos (5)
= q (6)
The state variables are V = airspeed, = flight path angle, = pitch angle, and q = pitch rate,
h is the altitude, T is the engine thrust force. The aerodynamic and propulsive forces are
assumed to have the following form:
Cm = C m0 + Cmq q + Cm + C m e + Cm f (7)
e f
Comparison of Flight Control System Design Methods in Landing 263
2 mg cos gl
CLbal = (8)
SVl 2
C Lb C L0
bal = (9)
C L
Vl 2S
Tbal = C D(CL ) + mg sin gl (10)
bal 2
bal = (Cm0 + Cm bal + C m f ) / Cm (11)

f e
C D ( ) = B0 + B1 + B2 2 (12)
C 0 + C 1 , *
C L ( ) = ,
2 * (13)
C
0 + C 1 + C 2 , max
The model parameter data given in Table 1 refer to a Boeing B727 aircraft powered by three
JT8D-17 turbofan engines [13].
3. Conventional design methods

Generally, the controller is in a form of proportional-integral-derivative (PID) parameters,
and the control gains are determined by using classical control theory, such as the root locus
method, Bode method or Nyquist stability criterion [14-16]. Modern control theory has been
used extensively to design the flight control system, such as in the linear quadratic
techniques [17,18], generalized singular linear quadratic technique [19], H design
technique and synthesis technique [20].
A. Optimal Control
The optimization of dynamic processes frequently requires the solution of an optimal
control problem. That is, given a process model in terms of differential equations for the
state variables, find the vector of control functions that minimizes some performance index,
subject to boundary conditions and possibly additional equality or inequality constraints on
the state and controls. Application of the necessary condition for optimal control of systems
defined by ordinary differential equations results in a two-point boundary value problem. A
suitable Method to solve such problems is the well-known Multiple Shooting Method [21].
The method reduces the boundary value problem to the solution of a set of algebraic
equations and initial value problems with partially unknown initial conditions that are solve
by Newton iteration. Based upon this method, this boundary value problem is solved
numerically by the Multiple Shooting code BNDSCO [22]. The optimal trajectory problem
consists of minimizing [2]
J (u) = [( e ebal )2 + ( H * H )2 ]dt (14)
with boundary conditions are,

V (0) = V0 ,
(0) = 0 ,
q(0) = q0
(15)
h(0) = h0 , h(t f ) = h f
X (0) = X0
(0) = 0, , (t f ) = f
and the final time t f is free

The variational Hamiltonian is formed by adjoining the right-hand sides of the system of
state differential Eqs. (1-6) with the costate variables
H = ( e eb )2 + ( H * H )2 +
g g L
V ( (Tb D) g sin ) + ( ( cos )) (16)
W V W
qSC
+q ( C m ) + h (V sin ) + x (V cos ) + (q )
Iy
The costate variables are defined by differential equation
dv H
= , v = V , , q , h , x , (17)
dt v
g g L q CSC m
V = V [ ( Dv )] + 2 ( cos ) q v h sin x cos (18)
W V W Iy
g sin
= V g cos hV cos + xV sin (19)
V
qCSC mq
q = q (20)
Iy
h = 2( H * H ) (21)
x = 0 (22)
= 0 (23)
Solving these equations with the appropriate boundary conditions is a two-point boundary-
value problem (TPBVP). Generally, at the initial point, all six states are specified and the
costates are free; whereas, at the final time, the H , is fix and the remaining four states are
free. Thus, at the final time all costates are free. The free final time t f can be considered as an
additional variable of the problem and, by the standard transformation
t = tf
d / d = t f d / dt (24)
The equations are transformed into a system with the independent variable ranging in
the interval 0 1 . Further, the additional trivial differential equation
d (25)
tf = 0
d
is added to the system. At the final time, the transversality condition requires that
H (t f ) = 0
The optimal control function is obtained by means of the minimum principle. With respect
to the elevator, we find the relation
qCSC m
e = eb q e (26)
2Iy
Result of optimal trajectory for the selected B-727 is presented in Fig (3) through (8). The
costate variable q , which is responsible for the determination of the extermal controller, is
shown in Fig.8.
B. Pole Placement Method
An alternative and very powerful method for designing feedback gains for autostabilization
systems is the pole placement method. The method is based on the manipulation of the
equations of motion in state space form and makes full use of the appropriate computational
tools in the analytical process. Practical application of the method to airplanes is limited
since it assumes that all state variables are available for use in an augmentation system,
which is not usually the case. The open loop state equation may then be written
X = AX + BU & Y = CX (27)
0.04436 29.988 0 32.1629 0 0

-0.001458 0.53075 1 0.007357 0 0

A= 0 0.4914 0.431 0 0 & B = -0.5849 (28)

0 0 1 0 0 0
0 210 0 210 0 0

To track the target successfully, the closed-loop tracking system must be fast enough. To
improve the flying qualities of this airplane more short-period damping is needed. A short
period mode damping ratio of 0.7 was chosen to have gives a good margin of stability.
Therefore, the pole placement method is employed to give the augmented aircraft the
following closed loop characteristic polynomial
aug (s ) = ( s + 0.00019)(s 2 + 11.2 s + 64)(s 2 + 0.12s + 0.0684) . The feedback gain matrix required
to give the augmented aircraft the characteristic polynomial was determined
K = [ 0.1411 85.6292 17.6339 14.6896 0.0004 ]
4. Intelligent control techniques

Intelligent control achieves automation via the emulation of biological intelligence. It either
seeks to replace a human who performs a control task or it borrows ideas from how
biological systems solve problems and applies them to the solution of control problems. In
this section we will provide an overview of several techniques used for intelligent control
and discuss challenging industrial application domains where these methods may provide
particularly useful solutions. The objective here is not to provide a comprehensive
treatment. We only seek to present the basic ideas to give a favor of the approaches.
A. Neural Net-based Control Design
Artificial neural networks are circuits, computer algorithms, or mathematical
representations loosely inspired by the massively connected set of neurons that form
biological neural networks. The success of neural networks is mainly attributed to their
unique features:
1. Parallel structures with distributed storage and processing of massive amounts of
information.
2. Learning ability made possible by adjusting the network interconnection weights and
biases based on certain learning algorithms.
The neural flight control architecture is based upon the augmented model inversion
controller, developed by Rysdyk and Calise [23]. The NN adaptation rule results from
nonlinear stability analysis, which ensures that the error signals and network weights are
bounded. One of the common methods for controlling nonlinear dynamical systems is based
on approximate feedback linearization [24]. The form that is employed in each control
channel depends on the relative degree of the controlled variable. To simplify our
discussion, we assume that the system has full relative degree, where each controlled
variable (element of the state vector x) has a relative degree of two
x = f ( x , x , ) (29)
In the case of aircraft, typically x , R , where the element of x correspond to the pitch attitude
angle. A variant of this form arises in which angular rate is controlled. Here, the equation of
motion for that degree of freedom is expressed in first order form [25]. A pseudo-control is
defined such that the dynamic relation between it and the system state is linear
v = f ( x , x , ) (30)
Ideally, the actual controls are obtained by inverting Eq. (30). The total pseudo control
signal is constructed of three components
= rm + pd ad (31)
where rm the pseudo-control component is generated by the reference model, pd is the

output of the linear compensator, and ad is generated by the adaptive element introduced
to compensate for the model inversion error. The linear compensator is designed so that the
error dynamics are stabilized. This is most often achieved using standard proportional-
derivative (PD) controllers, although additional integral action can be incorporated to
improve steady state performance. For the second order system, PD compensation is
expressed by
pd = [ K P K D ] e (32)
The compensator gain matrices K P , K D R are chosen so that the tracking error dynamics
given by
e = A e + B[ vad ] (33)
0 I 0
A= ,B= (34)
K P K D I
are stable, i.e., the eigenvalues of A are prescribed. It is evident from Eq. (33) that the role of
the adaptive component vad is to cancel .The adaptive signal is chosen to be the output of
a single hidden layer [26].
ad = W T (V T x ) (35)
where V and W are the input and output weighting matrices, respectively, and is a
sigmoid activation function. Although ideal weighting matrices are unknown and usually
cannot be computed, they can be adapted in real time using the following NN weights
training rules [27]:
W
( )
= V T x eT PB + e W
W
(36)
V = V xeT PBW T + e V (37)
where W , V and W , V are the positive definite learning rate matrices, and is the e-
modification parameter. Here, P is a positive definite solution of the Lyapunov equation
AT P + PA + Q = 0 for any positive definite Q .
B. Fuzzy Logic-Based Control Design
The existing applications of fuzzy control range from micro-controller based systems in
home applications to advanced flight control systems. The main advantages of using fuzzy
are as follows:
1. It is implemented based on human operators expertise which does not lend itself to
being easily expressed in conventional proportional integral-derivative parameters of
differential equations, but rather in action rules.
2. For an ill-conditioned or complex plant model, fuzzy control offers ways to implement
simple but robust solutions that cover a wide range of system parameters and, to some
extent, can cope with major disturbances.
The aircraft landing procedures admit a linguistic describing. This is practiced, for example,
in case of guiding for landing in non-visibility conditions or in piloting learning. This
approach permits to build a model for landing control based on the reasoning rules using
the fuzzy logic. The process requires the control of the following parameters: the current
altitude to runway surface (H), the aircraft's vertical speed and aircraft flight speed. The goal
of the control is formulated as follow: the aircraft should touch the runway (H becomes 0) at
the conventional point of landing with admitted vertical touch speed and the recommended
landing speed. The input of FLC normally includes the error between the state variable and
its set point, ( e = xd x ) and the first derivative of the error, e . A typical form of the
linguistic rules is represented as
Rule i Th: If e is A i and e is Bi then u * is C i
Where A i ,B i , and C i are the fuzzy sets for the error, the error rate, and the controller
output at rule i, respectively, and u * is the controller output.
The resulting rule base of FLC is shown in Table 2. The abbreviations representing the fuzzy
sets N, Z, P, S, and B in linguistic form stand for negative, zero, positive, small, and big,
respectively, for example negative big (NB). Five fuzzy sets in triangular membership
functions are used for FLC input variables, e and e , and FLC output, u * . For the fuzzy
inference or rule firing, Mamdani-type min-max composition is employed. In the
defuzzification stage, by adopting the method of center of gravity, the deterministic control
u is obtained. The membership functions have been designed for input and output diagram
using the trapezoidal shapes, as shown in Figures 1, 2. The fuzzy control system design with
a simple longitudinal aircraft model given by Eqs. (1-6).
Advantages over Conventional Designs
1. Fuzzy guidance and control provides a new design paradigm such that a control
mechanism based on expertise can be designed for complex, ill-defined flight dynamics
without knowledge of quantitative data regarding the input-output relations, which are
required by conventional approaches. A fuzzy logic control scheme can produce a higher
degree of automation and offers ways to implement simple but robust solutions that cover
a wide range of aerodynamic parameters and can cope with major external disturbances.
2. Artificial Neural networks constitute a promising new generation of information
processing systems that demonstrate the ability to learn, recall, and generalize from
training patterns or data. This specific feature offers the advantage of performance
improvement for ill-defined flight dynamics through learning by means of parallel and
distributed processing. Rapid adaptation to environment change makes them
appropriate for guidance and control systems because they can cope with aerodynamic
changes during flight.
5. Simulation results and discussion

Simulations are performed at sea level, airspeed of 210 ft/s, corresponding to the flare
maneuver configuration of the Boeing 727. The simulation results are presented in Figs 3 to
8. Figure 3, which depicts the flight speed variation, demonstrates that the engines can
regulate slight speed until that is compromised for attitude rate control. Time histories of the
controls are shown in Fig. 4. The time response of pitch angle is shown in fig.5. A
comparison between the commanded altitude profile and the actual aircraft response is
presented in Fig. 6, it shows that the difference between the actual and desired trajectory
(the fuzzy logic, neural net-based adaptive and optimal controls) is kept less than about 6ft.
This figure, so shows that the sink rate (the rate of descent) is reduced to less than 1.0 ft/sec,
which is small enough to achieve a smooth landing. The fuzzy logic, neural net-based
adaptive and optimal control approaches do the flare maneuver well, while the Pole
Placement Method has substantially large error. Neural network adaptation signal vad for
compensate inversion error is presented in Fig. 7. Summarizing the results presented so far,
the nonlinear controller performance for this maneuver has been found very good.
NB NS Z PS PB
1
0.8
Degree of membership
0.6
0.4
0.2
-5 0 5 10 15
Height
Fig. 1. Altitude Membership Functions
NB NS Z PS PB
1
0.8
Degree of membership
0.6
0.4
0.2
-20 -15 -10 -5 0 5

Force
Fig. 2. Force Membership Functions

211
NN
optimal
fuzzy
poleplace
210
209
208
Speed (ft/s)
207
206
205
204
0 1 2 3 4 5 6 7 8
Time
Fig. 3. Time response of the airspeed
0
NN
optimal
fuzzy
poleplace
-2
-4
de (deg)
-6
-8
-10
-12
0 1 2 3 4 5 6 7 8
Time
Fig. 4. Time response of the elevator

12
NN
optimal
fuzzy
11.5 poleplace
11
10.5
pitch angle (deg)
10
9.5
8.5
8
0 1 2 3 4 5 6 7 8
Time
Fig. 5. Time response of the pitch angle
35
NN
NN
optimal
30
fuzzy
fuzzy
poleplace
25 poleplace
command
20
15
Altitude (m)
10
-5
-10
0 1 2 3 4 5 6 7 8
Time
Fig. 6. Desired and actual flare trajectories

0.04
neural network(NN)
0.02
0
ouput neural network
-0.02
-0.04
-0.06
-0.08
0 1 2 3 4 5 6 7 8
Time
Fig. 7. NN adaptation signal ad
0.01
optimal
-0.01
-0.02
Lambda q
-0.03
-0.04
-0.05
-0.06
-0.07
0 1 2 3 4 5 6 7 8
Time
Fig. 8. Time response of the q

6. Conclusions
It has been the general focus of this paper to summarize the basic knowledge about
intelligent control structures for the development of control systems. For completeness,
conventional, adaptive neural net-based, fuzzy logic-based, control techniques have been
briefly summarized. Our particular goal was to demonstrate the potential intelligent control
systems for high precision maneuvers required by aircraft landing. The proposed model
reveals the functional aspect for realistic simulation data. The method does not require the
existing controller to be designed based on a linear model.
max = 17.2 * = 12deg
g = 32.2ft 2 I y = 3 106
C 0 = 0.7125 B0 = 0.1552
B1 = 0.12369rad 1 S = 0.156 10 4 ft 2
B 2 = 2.4203rad2 W = 150000Ib
C 2 = 9.0277rad 2 C 1 = 6.0877rad 1
Table 1. model parameter data B-727
Fuzzy set, e
Fuzzy set, e NB NS Z PS PB
NB PB PB PS PS NS
NS PB PS PS NS NB
Z PB PS Z NS NB
PS PB PS NS NS NB
PB PS NS NS NB NB
Table 2. Rule base for FLC
7. References
[1] Chicago, IL, U.S.A. Locke, A. S., Guidance, D. Van Nostrand Co., Princeton, NJ, U.S.A (1955).
[2] Bryson, A. E., Jr. and Y. C. Ho, Applied Optimal Control., Blaisdell, Waltham, MA, U.S.A
(1969).
[3] Lin, C. F., Modern Navigation, Guidance, and Control Processing, Prentice-Hall, Englewood
Cliffs, NJ, U.S.A (1991).
[4] Zarchan, P., Tactical and Strategic Missile Guidance, 2nd Ed., AIAA, Inc., Washington, D.C.,
U.S.A (1994).
[5] Bezdek, J., "Fuzzy Models: What are they and Why," IEEE Trans. Fuzzy Syst., Vol.1 No.1,
pp, 1-6 (1993).
[6] Miller, W. T., R. S. Sutton, and P. J. Werbos., Neural Networks for Control., MIT Press,
Cambridge, MA, U.S.A. Mishra (1991).
[7] Narendra, K. S. and K. Parthasarthy., Identification and control of dynamical systems using
neural networks. IEEE Trans. Neural Networks, 1(1), 4-27 (1990).
[8] Mamdani, E. H. and S. Assilian., An experiment in linguistic synthesis with a fuzzy logic
controller. Int. J. Man Machine Studies, 7(1), 1-13 (1975).
[9] Lee, C. C., Fuzzy logic in control systems: fuzzy logic controller part I. IEEE Trans. Syst.
Man and Cyb., 20(2), 404-418 (1990).
[10] Lee, C. C., Fuzzy logic in control systems: fuzzy logic controller part II. IEEE Trans. Syst.
Man and Cyb., 20(2), 419-435 (1990).
[11] Driankov, D., H. Hellendoorn, and M. Reinfrank., "An Introduction to Fuzzy Control".
Springer, Berlin, Germany. Driankov (1993).
[12] Dash, P. K., S. K. Panda, T. H. Lee and J. X. Xu., Fuzzy and neural controllers for dynamic
systems: an overview. Proc. Int. Conf. Power Electronics, Drives and Energy Systems,
Singapore (1997).
[13] BULIRSCH,R., F. Montone, and H. Pesch, "Abort Landing in the Presence of Windshear
as a minimax optimal Control Problem, Part 1: Necessary Conditions", J. Opt.
Theory Appl., Vol. 70,pp. 1-23 (1991).
[14] Price, C. F. and R. S. Warren, "Performance Evaluation o f Homing Guidance Laws for
Tactical Missiles," TASC Tech (1973).
[15] Nesline, F. W., B. H. Wells, and P. Zarchan, "Combined optimal/classical approach to
robust missile autopilot design," AIAA J. Guid. Contr., Vol.4,No.3, pp.316-322 (1981).
[16] Nesline, F.W. and M.L. Nesline, How Autopilot Requirements Constrain the
Aerodynamic Design of Homing Missiles, Proc. Amer. Contr. Conf., San Diego, CA,
USA, pp.176-730 (1984).
[17] Stallard, D. V., "An Approach to Autopilot Design for Homing Interceptor Missiles,"
AIAA Paper 91-2612, AIAA, Washington, D.C., U.S.A, pp. 99-113 (1991).
[18] Lin, C.F., J. Cloutier, and J. Evers, Missile Autopilot Design Using a Generalized
Hamiltonian Formulation, Proc. IEEE 1st Conf. Aero. Contr. Syst., Westlake Village,
CA, USA, pp. 715-723 (1993).
[19] Lin, C. F. and S. P. Lee, "Robust missile autopilot design using a generalized singular
optimal control technique," J. Guid., Contr., Dyna., Vol. 8, No. 4, pp. 498-507 (1985).
[20] Lin, C. F. Advanced Control System Design. Prentice- Hall, Englewood Cliffs, NJ, U.S.A (1994).
[21] Stoer J. and R. Burlisch, Introduction to Numerical Analysis, Springer Verlag, New York, (1980).
[22] Oberle, H.J, "BNDSCO-A Program for the Numerical Solution of Optimal Control
Problems," Internal Report No.515-89/22, Institute for Flight Systems Dynamics,
DLR, Oberpfaffenhofen, Germany (1989).
[23] Rysdyk, R., B. Leonhardt, and A.J. Calise, Development of an Intelligent Flight
Propulsion Control System: Nonlinear Adaptive Control, AIAA- 2000-3943, Proc.
Guid. Navig. Contr. Conf., Denver, CO, USA (2000).
[24] Isodori, A., Nonlinear Control Systems, Springer Verlag, Berlin (1989).
[25] Calise, A. J., S. Lee, and M. Sharma, "Development of a reconfigurable flight control law
for the X-36 tailless fighter aircraft," Proc. AIAA Guid. Navig. Contr. Conf., Denver,
CO, USA., AIAA-2000-3940 (2000).
[26] Hornik, K., M. Stinchcombe, M. and H. White, Multilayer Feedforward Networks are
Universal Approximators, Neural Networks, Vol. 2, pp. 359-366 (1989).
[27] Johnson, E. and A.J. Calise, Neural Network Adaptive Control of Systems with Input
Saturation, Proc. Amer. Contr. Conf., pp. 35273532. (2001).
14
Oscillation Susceptibility of an
Unmanned Aircraft whose Automatic
Flight Control System Fails
Balint Maria-Agneta and Balint Stefan
West University of Timisoara
Romania
1. Introduction
Interest in oscillation susceptibility of an aircraft was generated by crashes of high
performance fighter airplanes such as the YF-22A and B-2, due to oscillations that were not
predicted during the aircraft development. Flying qualities and oscillation prediction, based
on linear analysis, cannot predict the presence or the absence of oscillations, because of the
large variety of nonlinear interactions that have been identified as factors contributing to
oscillations. Pilot induced oscillations have been analyzed extensively in many papers by
numerical means.
Interest in oscillation susceptibility analysis of an unmanned aircraft, whose flight control
system fails, was generated by the need to elaborate an alternative automatic flight control
system for the Automatic Landing Flight Experiment (ALFLEX) reentry vehicle for the case
when the existing automatic flight control system of the vehicle fails.
The purpose of this chapter is the analysis of the oscillation susceptibility of an unmaned
aircraft whose automatic flight control system fails. The analysis is focused on the research
of oscillatory movement around the center of mass in a longitudinal flight with constant
forward velocity (mainly in the final approach and landing phase). The analysis is made in a
mathematical model defined by a system of three nonlinear ordinary differential equations,
which govern the aircraft movement around its center of mass, in such a flight. This model
is deduced in the second paragraph, starting with the set of 9 nonlinear ordinary differential
equations governing the movement of the aircraft around its center of mass.In the third
paragraph it is shown that in a longitudinal flight with constant forward velocity, if the
elevator deflection outruns some limits, oscillatory movement appears. This is proved by
means of coincidence degree theory and Mawhin's continuation theorem. As far as we
know, this result was proved and published very recently by the authors of this chapter
(research supported by CNCSIS-UEFISCSU, project number PNII IDEI 354 No. 7/2007)
and never been included in a book concerning the topic of flight control.The fourth
paragraph of this chapter presents mainly numerical results. These results concern an Aero
Data Model in Research Environment (ADMIRE) and consists in: the identification of the range
of the elevator deflection for which steady state exists; the computation of the manifold of
steady states; the identification of stable and unstable steady states; the simulation of
successful and unsuccessful maneuvers; simulation of oscillatory movements.
2. The mathematical model

Frequently, we describe the evolution of real phenomena by systems of ordinary differential
equations. These systems express physical laws, geometrical connections, and often they are
obtained by neglecting some influences and quantities, which are assumed insignificant
with respect to the others. If the obtained simplified system correctly describes the real
phenomenon, then it has to be topologically equivalent to the system in which the small
influences and quantities (which have been neglected) are also included. Furthermore, the
simplified system has to be structurally stable. Therefore, when a simplified model of a real
phenomenon is build up, it is desirable to verify the structural stability of the system. This
task is not easy at all. What happens in general is that the results obtained in simplified
model are tested against experimental results and in case of agreement the simplified model
is considered to be authentic. This philosophy is also adapted in the description of the
motion around the center of gravity of a rigid aircraft. According to Etkin & Reid, 1996;
Cook, 1997, the system of differential equations, which describes the motion around the
center of gravity of a rigid aircraft, with respect to an xyz body-axis system, where xz is the
plane of symmetry, is:
D D D
V cos cos cos sin sin cos = r sin q sin cos
V

g X
sin +
V m V
D
V D g Y
sin + cos = p sin cos r cos cos + sin cos +
V V m V
D D D
V sin cos sin sin + cos cos = p sin + q cos cos +
V

g Z
cos cos +
V m V

( )
D D
I x p I xz r = I y I z q r + I xz p q + L (1)

I qD
y ( )
= ( I z I x ) p r I xz p 2 r 2 + M
D
( )
D
I z r I xz p = I x I y p q I xz q r + N

D
D
= q cos r sin
D
= q sin + r cos
cos

The state parameters of this system are: forward velocity V, angle of attack , sideslip angle
, roll rate p, pitch rate q, yaw rate r, Euler roll angle , Euler pitch angle and Euler yaw
angle . The constants Ix , Iy and Iz -moments of inertia about the x, y and z-axis, respectively;
Oscillation Susceptibility of an Unmanned Aircraft whose Automatic Flight Control System Fails 277
Ixz - product of inertia, g -gravitational acceleration; and m - mass of the vehicle. The aero
dynamical forces X, Y, Z and moments L, M, N are functions of the state parameters and the
control parameters: a - aileron deflection; e - elevator deflection; and r - rudder deflection
(the body flap, speed break, c, ca are available as additional controls but, for simplicity,
they are set to 0 in the analysis to follow). When the automatic flight control system is in
function, then the control parameters are functions of the state parameters, describing how
the flight control system works. When the automatic flight control system fails, then the
control parameters are constant. This last situation will be analyzed in this chapter.
A flight with constant forward velocity V is defined as a flight for which V = const (i.e.
D
V = 0 ).
In a flight with constant forward velocity V the following equalities hold:
D D g X
cos sin sin cos = r sin q sin cos sin +
V m V
D g Y
cos = p sin cos r cos cos + sin cos + (2)
V m V
D D g Z
sin sin + cos cos = p sin + q cos cos + cos cos +
V m V
D
Replacing V by 0 in the system (1), the equalities(2) are obtained.
If in a flight with constant forward velocity V one has ( 2 n + 1 ) , then the following
2
equalities hold:
g X
( 1 ) r V sin + m V 0
n

g Y
sin cos + 0 (3)
V m V
g Z
( 1 ) p + V cos cos + m V 0
n+1

D
Replacing = 0 and (2 n + 1) in (2), the equalities (3) are obtained.
2
If in a flight with constant forward velocity V one has ( 2 n + 1 ) , then the following equality
2
holds:
Y
g sin sin cos cos cos sin + sin cos cos cos + sin +
m (4)
X Z
+ cos cos + sin cos 0
m m
D D
Equation (4) is the solvability (compatibility) condition, with respect to , , of the system
(2) when ( 2 n + 1 ) .
2
D D
If ( 2 n + 1 ) and equality (4) holds, then the system (2) can be solved with respect to , ,
2
obtaining in this way the explicit system of differential equations, which describes the motion around
the center of mass of the aircraft in a flight, with constant forward velocityV:
D g
= q p cos tan r sin tan + cos cos cos +
V cos
1 Z X
+ sin sin + cos sin
cos m V m V
D 1 g 1 Y
= p sin r cos + sin cos +
cos V cos m V

( )
D D
I p I r = I I q r + I p q + L
x xz y z xz

( )
D
I y q = ( I z I x ) p r I xz p 2 r 2 + M (5)
D
xz ( x y)
I r I pD = I I p q I q r + N
z xz
D
D
= q cos r sin

D q sin + r cos
=
cos

D D
System (5) is obtained solving system (2) withD respectD to , and replacing in system (1)
the equations (1)1, (1)2 , (1)3 with the obtained and .
A longitudinal flight is defined as a flight for which the following equalities hold:
p r 0 and a = r = 0 . (6)
A longitudinal flight is possible if and only if Y = L = N = 0 for = p = r = = = 0 and

a = r = 0 .
This result is obtained from (1) taking into account the definition of a longitudinal flight.
The explicit system of differential equations which describes the motion of the aircraft in a
longitudinal flight is:
D X Z
V = g sin( ) + cos + sin
m m
D g X Z
= q + V cos ( ) m V sin + m V cos
D (7)
q = M
Iy

D
= q
This result is obtained from (1) taking into account the definition of a longitudinal flight.
In system (7) X, Z, M depend only on , q , and e . These dependences are obtained
replacing in the general expression of the aerodynamic forces and moments:
= p = r = = = 0 and a = r = 0 .
The explicit system of differential equations which describes the motion around the center of gravity
of the aircraft in a longitudinal flight with constant forward velocity V is:
D g X Z
= q + cos ( ) sin + cos
V m V m V
D M
q = (8)
Iy
D
= q

D
This system is obtained from (7) taking into account V = 0 .
A longitudinal flight with constant forward velocity is possible if the following equalities hold:
Y = L = N = 0 for = p = r = = = 0 and a = r = 0 (9)
X Z
g sin ( ) + cos + sin = 0 (10)
m m
D
This result is obtained from system (7), taking into account the fact that V is equal to zero.
In (8) X , Z , M depend on , q , , e and V . Taking into account (10), the system (8) can be
written as:
D g g Z 1
= q + cos ( ) sin ( ) tan +
V V m V cos
D M
q = (11)
Iy
D
= q

The system (11) describes the motion around the center of gravity of an aircraft in a
longitudinal flight with constant forward velocity V and defines the general nonlinear
model.
In system (11) the functions Z = Z( , q , ; e ,V ) and M = M( , q , ; e ,V ) are considered
known. When the automated flight control system fails, e and V are parameters.
Frequently, in a research environment for the description of the movement around the
center of the gravity of some types of aircrafts in a flight with constant forward velocity V,
the explicit system of differential equations (12) is employed by Balint et al., 2009a,b,c;
2010a,b; Kaslik & Balint, 2007; Goto & Matsumoto, 2000.
The model defined by equations (12) is called Aero Data Model In a Research Environment
(ADMIRE).
System (12) can be obtained from (5) substituting the general aero dynamical forces and moments
(see for example section 4), assuming that and are small and making the approximations (13).
Due to the approximations (13), the ADMIRE model defined by (12) is also called the
simplified ADMIRE model.
The simplified system which governs the longitudinal flight with constant forward velocity V of the
ADMIRE aircraft is (14).
System (14) is obtained from (12) for = p = r = = 0 , a = r = c = ca = 0 and defines the
simplified nonlinear model of the motion around the center of gravity of the aircraft in a
longitudinal flight with constant forward velocity V.
In system (14) g ,V , z , z e , m , mq , m D , c 2 , a2 , a , m e are considered constants (see Section 4).

D g 2
= q p + cos cos + z + y + y p ( , ) p + yr ( ) r +
V
+ y a a + z e e + y r r
D
= p r + g sin cos z + y y ( , ) p y ( ) r +
p r
V
+ y a a z e e + y r r

D
p = i1 q r + l ( ) + lp p + lr ( ) r + l a a + l r r

D 2
q = i2 p r + m + mq q mD p + y p p + y + y r ( ) r +

g c (12)
+ m D cos cos 2 a2 sin + y a a + m c c +
V a

+ m e e + y r r
D
r = i3 p q + n + np ( , ) p + nr ( , ) r + n a + n ( ) ca + n r
a ca r
D
= p + ( q sin + r cos ) tan
D
= q cos r sin
D
= q sin + r cos
cos
in (5)1 cos 1 ; p cos tan p ; r sin tan 0 ; cos 1 ; sin 0

in (5)2 cos 1
(
I I + I I
xz x z y 0; ) ( ) 2
I y I z I z I xz
i1 ;
I xz
0;
2 2 (13)
I x I z I xz I x I z I xz Iy
2
( )
I xz + I x I y I x
i3 ;
(
I xz I y I z I xz) 0;
I I I2 2
I x I z I xz
x z xz
D g
= q + cos + z + z e e
V
D g c2
q = m + mq q + m D cos a2 sin + m e e (14)
V a
D
= q

3. Theoretical proof of the existence of oscillatory movements

Interest in oscillation susceptibility of an aircraft is generated by crashes of high
performance fighter airplanes such as the YF-22A and B-2, due to oscillations that were not
predicted during the aircraft development (Mehra & Prasanth, 1998).Flying qualities and
oscillation prediction are based on linear analysis and their quasi-linear extensions. These
analyses cannot, in general, predict the presence or the absence of oscillations, because of the
large variety of nonlinear interactions that have been identified as factors contributing to
oscillations. The effects of some of these factors have been reported by Mehra et al., 1977;
Shamma & Athans, 1991; Kish et al., 1997; Klyde et al., 1997. The oscillation susceptibility
analysis in a nonlinear model involves the computation of nonlinear phenomena including
bifurcations, which lead sometimes to large changes in the stability of the aircraft.
Interest in oscillation susceptibility analysis of an unmanned aircraft, whose flight control
system fails, was generated by the elaboration of an alternative automatic flight control for
the case when the existing automatic flight control system of the aircraft fails Goto &
Matsumoto, 2000. Numerical results concerning this problem for the ALFLEX unmanned
reentry vehicle, considered by Goto & Matsumoto, 2000, for the final approach and landing
phase are reported by Kaslik et al., 2002; 2004 a; 2004 b; 2005 a; 2005 b; 2005 c; Caruntu et al.,
2005.
A theoretical proof of the existence of oscillatory solutions of the ALFLEX unmanned
reentry vehicle in a longitudinal flight with constant forward velocity and decoupled flight
control system is reported Kaslik & Balint, 2010. Numerical results related to oscillation
susceptibility analysis along the path of the longitudinal flight equilibriums of a simplified
ADMIRE-model, when the automated flight control system fails are reported by Balint et al.,
2009 a; 2009 b; 2009 c; 2010 a.
In this section we give a theoretical proof of the existence of oscillatory solutions in
longitudinal flight in the simplified ADMIRE model of an unmanned aircraft, when the
flight control system fails. This result was established by Balint et al., 2010b.
The simplified system of differential equations which governs the motion around the center
of mass in a longitudinal flight with constant forward velocity of a rigid aircraft, when the
automatic flight control system fails, is given by (see Balint et al., 2010b):
D g
= z + q + cos + z e e
V
D g c2
q = m + mq q + m cos a2 sin + m e e (15)
V a
D
= q

In this system, the state parameters are: angle of attack , pitch rate q and Euler pitch
angle . The control parameter is the elevator angle e . V is the forward velocity of the
aircraft, considered constant and g is the gravitational acceleration.
The aero dynamical data appearing in (15) are given in section 4.
The following proposition (Balint et al., 2010b) addresses the existence of equilibrium states
for the system (15).
Proposition 1. ( , q ) T is an equilibrium state of the system (15) corresponding to e if and only
if is a solution of the equation:
A 2 + B e + C e2 + D = 0 (16)
q is equal to zero and is a solution of the equation:
V
cos = z + z e e (17)
g
where A, B, C, D are given by:
c2 2
( )
2
A = m m z + a2 2 z 2
a2
c2 2
( )(
B = 2 m m z m e m z e + 2 ) a2
a2 2 z z e
(18)
c2 2
( )
2
C = m e m z e + a2 2 z e 2
a2
2 2
g c2
D= 2
a2 2
V a2
Proof. By computation.
Proposition 2. Equation (16) has real solutions if and only if e satisfies:
4 AD
e 2
(19)
B 4 A C
Proposition 3. If z < 0 and is a real solution of Eq.(16), then Eq.(17) has a solution if and only
if for the following inequality holds:
1 g 1 g
z e e + z e e (20)
z V z V
Remark. For e = 0 the solutions of Eq.(16) are:
g 2 c2 2
a2 2
= V 2 a2 (21)
c2 2
( )
2
a2 2 z 2 + m m z
a2
and both verify (20) for z < 0 .

In the following assume that z < 0 and consider e , e defined as follows:
e = inf { e e < 0, a real solution of eq.(16) for which (20) holds}
e = sup { e e > 0, a real solution of eq.(16) for which (20) holds}
Let I be the closed interval I = [ e , e ] .

Proposition 4.
a. If e I , then for the system (15) there exists a countable infinity of equilibriums corresponding to
e , namely for any n Z
B 1 2 2
( 2
)
1,2 = 2 A e 2 A B e 4 A C e + D ; q = 0 ;

V
g
(
1,2 = 2 n arccos z n z e e

)
{ }
b. If e I = e , e , then the equilibriums corresponding to e are saddle-node bifurcation points.
c. If e I , then for the system (15) there are no equilibriums corresponding to e .
Proposition 4 translates into the following necessary and sufficient condition for the
existence of equilibrium states for (15): e I = e , e . At = e , e saddle-node

bifurcation occurs.
It can be easily verified that the following proposition is valid.
Proposition 5.
a. If ( (t ), q(t ), (t )) T is a solution of the system (15), then (t ) is a solution of the third order
differential equation:

( )
V
g
z + mq + z mq m +
c
m sin + 2 a2 cos =
a (22)
g

g c2
(
= m z m D cos + z a2 sin + m z e z m e e
V V a
)
b. If (t ) is a solution of (22), then
1 g c
(t ) = mq m cos 2 a2 sin m e e
m V a
q(t ) = (t )
(t ) = (t )
is a solution of the system (15).
A solution (t ) of Eq.(22) is called monotonic oscillatory solution if the derivative (t ) is a

strictly positive or a strictly negative periodic function.
Proposition 6. a. If there exists an increasing oscillatory solution (t ) of Eq.(22)and T > 0 is the
period of (t ) , then there exists n N * such that (t + T ) = (t ) + 2 n and there exists a 2n -
periodic solution x(s ) of the equation:

( ) g c
x " = 2( x ') 2 + z + mq x ' e x + z mq m + m sin s + 2 a2 cos s e 2 x
V a
(23)
g g c
( )

m z m D cos s + z 2 a2 sin s + m z e z m e e e 3 x
V V a
satisfying
2 n
e x( s ) ds = T (24)
0
b. If for n N * there exists a 2 n - periodic solution x(s ) of Eq.(23), then there exists an increasing
oscillatory solution (t ) of Eq.(22) satisfying (t + T ) = (t ) + 2 n , where T is given by (24).
Proof. See Balint et al., 2010b.
In order to prove that Eq.(22) has an increasing oscillatory solution, it is sufficient to prove
that there exists a 2n - periodic solution of the Eq.(23).
Denoting x1 = x and x2 = ( z + mq ) e x x ' e 2 x , eq.(23) is replaced by the system:
x1 ' = ( z + mq ) e x1 x2 e 2 x1

( ) V
g
( g
V
) c
x2 ' = z m e + m z e e + m z m cos s + z 2 a2 sin s e x1
a
(25)
g c2
z mq m + m sin s + a2 cos s
V a
Proposition 7. a. If there exists a decreasing oscillatory solution (t ) of (22) and T > 0 is the period
of (t ) , then there exists n N * such that (t + T ) = (t ) 2 n and there exists a 2 n - periodic
solution x(s ) of the equation:

( )
g
V
c
x " = 2( x ')2 z + mq x ' e x + z mq m + m sin s + 2 a2 cos s e 2 x +
a
(26)
g
( ) V
g c2 3x
+ m z m cos s + z a2 sin s + (m e z m e ) e e
V a
satisfying
2 n
T= e x( s ) ds (27)
0
b. If for n N * , there exists a 2 n - periodic solution x(s ) of(26), then there exists a decreasing
oscillatory solution (t ) of (22), satisfying (t + T ) = (t ) 2 n with T given by (27).
It follows that in order to prove that Eq.(22) has a decreasing oscillatory solution, it is
sufficient to prove that there exists a 2n - periodic solution of the Eq.(26).
Denoting by x1 = x and x2 = ( z + mq ) e x + x ' e 2 x Eq.(26) is replaced by the system:
x1 ' = ( z + mq ) e x1 + x2 e 2 x1

( ) V
g
( g
V
c
)
x2 ' = z m e + m z e e + m z m cos s + z 2 a2 sin s e x1
a
(28)
g c2
+ z mq m + m sin s + a2 cos s
V a
Hence, in order to show the existence of a decreasing oscillatory solution of Eq.(22) it is

sufficient to show the existence of a 2n - periodic solution of (28).
For a continuous 2 n periodic function we define:
f M
= max f (s ) f L
= min f (s )
s0, 2 n s0, 2 n
Proposition 8. If f is a smooth 2n periodic function, then
2 n
1
f M
f L
+ f (s ) ds .
2 0

Let X , Y be two infinite dimensional Banach spaces. A linear operator L : DomL X Y is
called a Fredholm operator if Ker L has finite dimension and Im L is closed and has finite
codimension. The index of a Fredholm operator L is the integer
i(L ) = dim KerL co dim Im L .
In the following, consider L : DomL X Y a Fredholm operator of index zero, which is
not injective. Let P : X X and Q : Y Y be continuous projectors, such that
KerQ = Im L , Im P = KerL , X = KerL KerP and Y = Im L Im Q . The operator
LP = L DomL KerP : DomL KerP Im L is an isomorphism. Consider the operator
K PQ : Y X defined by K PQ = K P 1 ( I Q ) . Let X be an open bounded set and
N : Y be a continuous nonlinear operator. We say that N is L compact if K PQ N is
compact, QN is continuous and QN ( ) is a bounded set in Y . Since ImQ is isomorphic to
KerL , there exists an isomorphism I : Im Q KerL .
Mawhin, 1972; Gaines & Mawhin, 1977 established the Mawhins continuation theorem : Let
X be an open bounded set, let L be a Fredholm operator of index zero and let N be L compact
on . Assume:
a. Lx Nx for any ( 0, 1 ) and x DomL .
b. QNx 0 for any x KerL .
c. Brouwer degree deg B ( IQN , KerL , 0) 0 .

Then Lx = Nx has at least one solution in DomL .
Consider the Eq.(22) and denote by:
= ( z + mq )
g c
= z mq m + m sin + 2 a2 cos
V a
(29)
( )V
g c2
= m z e z m e e + z a2 sin
a
g
(
= m z m
V
)
With these notations Eq.(22) can be written as:
+ + = cos (30)
As concerns the quantities , , , we make the following assumptions:
> 0, > 0 and 2 > 4 (31)
Remark that the above inequalities can be assured as follows:
z + mq < 0 ( > 0)
g 2 c 2
z mq m > m + 22 a2 2 ( > 0) (32)
V a
c 2
4g
( )
2
( z + mq ) 2 > 4 ( z mq m ) + m + 22 a2 2 2
> 4
V a
Remark also that since z < 0 for the following inequality holds:
( m z e
)
z m e e
g
V
c
( g
) c
z 2 a2 m z e z m e e + z 2 a2
a V a
(33)
In terms of , , , the systems (25) and (28) can be written in the forms:
x1 ' = e x1 x2 e 2 x1
x
(34)
x2 ' = (s ) cos s e 1 (s )
and
x1 ' = e + x2 e
x1 2 x1
x
(35)
x2 ' = (s ) cos s e 1 + (s )
respectively.
Theorem 1. If the inequalities (32) hold and
( m z e
z m e e + ) g
V
c
z 2 a2 >
a
(36)
then for any n N * the system (2.20) has at least one 2n -periodic solution.
Theorem 2. If inequalities (32) hold and
( m z e
)
z m e e
g
V
c
z 2 a2 <
a
(37)
then for any n N * the system (2.21) has at least one 2n -periodic solution.
The conclusion of this section can be summarized as:
Theorem 3. If inequalities (32) and (36) hold, then for any n N * equation (22) has at least one
solution (t ) , such that its derivative (t ) is a positive 2n -periodic function (i.e. (t ) is an
increasing oscillatory solution).
If inequalities (32) and (37) hold, then for any n N * equation (22) has at least one solution (t ) ,
such that its derivative (t ) is a negative periodic function (i.e. (t ) is a decreasing oscillatory
solution).
4. Numerical examples
To describe the flight of ADMIRE (Aero Data Model in a Research Environment) aircraft
with constant forward velocity V , the system of differential equations (12) is employed:
where:
e
z = a C N z e = a C N y = a C y yr ( ) = a C yr ( ) y r = a C y r
D

y a = a C y a y p ( , ) = a C yp ( , ) m = a2 C m c1 C N + c 2 a CT + C m

a C N

( )
D D
e
e
m e = a2 C m c1 C N + C m a C Ne c
m c = a2 C m mq = a2 C mq + C m

m D = a2 C m

l ( ) = a1 C l ( ) lp = a1 C lp lr ( ) = a1 C lr ( ) l r = a1 C l r l a = a1 C l a
(
n = a3 C n + c 3 C y ) (
np ( , ) = a3 C np ( , ) + c 3 C yp ( , ) ) n ca ( ) = a3 c 3 C n ca ( )
(
nr ( , ) = a3 C nr ( , ) + c 3 C yr ( ) ) (
n r = a3 C n r + c 3 C y r ) (
n a = a3 C n a + c 3 C y a )
y = a2 c 2 a C y
yr = a2 c 2 a C yr ( ) yp = a2 c 2 a C yp ( , )
y r = a2 c 2 a C yr y a = a2 c 2 a C y a CT = 0.157[rad 1 ] C lp = 0.28[rad 1 ]
C N = 3.295[ rad 1 ] C lr ( ) = (0.344 + 0.02)[rad 1 ] C Ne = 1.074[rad 1 ]

C n = 0.0907[ rad 1 ]
Cm = 0.267[ rad 1 ] C n r = 0.0846[rad 1 ] e
Cm = 0.426[rad 1 ]
C n a = 0.051[rad 1 ] c
Cm = 0.2[rad 1 ] C n ca ( ) = 0.49 + 0.0145[ rad ]
D

Cm = 0.44[rad 1 ] C mq = 1.45[rad 1 ] C 1 ( ) = 0.896 2 0.47 0.04 [rad ]
C y = 0.804[ rad 1 ] C y r = 0.185[rad 1 ] C y a = 0.122[rad 1 ] C yr ( ) = 2.725 2 [rad ]
C yp ( , ) = ( 6.796 + 0.315 ) 2 + (0.237 0.498) 10 3 [rad ]
( )
C nr ( , ) = 1.572 2 0.368 1.07 2 0.005[rad ]
C l = 0.024[ rad ]; C l = 0.192[rad ]; C np ( , ) = ( 2.865 + 0.3 ) 2 [rad ]

r a
H = 500[m] M = 0.25 as = 338[ m s 1 ] = 1.16[ kg m3 ]

g = 9.81[ m s 2 ] V = M as = 84.5[m s 1 ] g / V = 0.116[s 1 ]
S = 45[ m2 ] c = 5.2[m] b = 10[m] m = 9100[ kg ] xG = 1.3[m] ze = 0.15
2 2 2
I x = 21000[ kg m ] I y = 81000[ kg m ] I z = 101000[ kg m ]
a = 0.485[ s 1 ] a1 = 88.743[s 1 ] a2 = 11.964[ s 1 ] a3 = 18.45[s 1 ]
c1 = 0.25 c 2 = 0.029 c 3 = 0.13 i1 = 0.952 i2 = 0.987 i3 = 0.594
All the other derivatives are equal to zero.
The system which governs the longitudinal flight with constant forward velocity V of the
ADMIRE aircraft, when the automatic flight control fails, is:
D g
= q + cos + z + z e e
V
D g c2
q = m + mq q + m D cos a2 sin + m e e (38)
V a
D
= q

When the automatic flight control system is in function, then e in (38) is given by:
e = k + kq q + k p (39)
with k = 0.401; k q = 1.284 and k p = 1 8 .

System (38) is obtained from the system (12) for = p = r = = 0 a = r = c = ca = 0 .
The equilibriums of (38) are the solutions of the nonlinear system of equations:
g
q + V cos + z + z e e = 0

g c2
m + mq q + m cos a2 sin + m e e = 0 (40)
V a
q = 0

System (40) defines the equilibriums manifold of the longitudinal flight with constant
forward velocity V of the ADMIRE aircraft.
It is easy to see that (40) implies:
A 2 + B e + C e 2+ D = 0 (41)
where A,B,C,D are given by:
c2 2
( )

2
A = m m z + a2 2 z 2
a2
c2 2
(
B = 2 m m z m e m z e + 2 )(
) a2
a2 2 z z e
c2 2
( )

2
C = m e m z e + a2 2 z e 2
a2
g 2 c2 2
D= a2 2
V 2 a2
Solving Eq.(41) two solutions 1 = 1(e) and 2 = 2(e) are obtained. Replacing in (17) 1 =
1(e) and 2 = 2(e) the corresponding 1 =1(e)+2k and 2 =2(e) +2k are obtained ( k Z ) .
Hence a part of the equilibrium manifold MV ( k = 0) is the union of the following two
pieces:
P1 = { ( 1 ( e ) ,0,1 ( e ) )# e I} ; P2 = { ( 2 ( e ) ,0, 2 ( e ) )# e I} .
The interval I where e varies follows from the condition that the angles 1 ( e ) and 2 ( e )
have to be real.
Using the numerical values of the parameters for the ADMIRE model aircraft and the
software MatCAD Professional it was found that:
e = -0.04678233231992 [rad] and e = 0.04678233231992[rad].
The computed 1 ( e ) , 1 ( e ) , 2 ( e ) , 2 ( e ) are represented on Fig.1, 2.
( ) ( ) ( ) ( ) and ( ) > ( ) for ( , ) .

Fig.1 shows that 1 e = 2 e , 1 e = 2 e 1 e 2 e e e e
Fig.2 shows that ( ) = ( ) , ( ) = ( ) and ( ) < ( ) for ( , ) .

1 e 2 e 1 e 2 e 1 e 2 e e e e
The eigenvalues of the matrix A( e ) are: 1= - 22.6334; 2= - 1.5765; 3= 1.0703 x 10-8 0 .

For e > e the equilibriums of P1 are exponentially stable and those of P2 are unstable. These
facts were deduced computing the eigenvalues of A( e ) .
More precisely, it was obtained that the eigenvalues of A( e ) are negative at the
equilibriums of P1 and two of the eigenvalues are negative and the third is positive at the
equilibriums of P2. Consequently, e is a turning point. Maneuvers on P1 are successful and
on P2 are not successful, Fig.3, 4.
Moreover, numerical tests show that when e ', e " e , e , the maneuver e ' e " ( )
transfers the ADMIRE aircraft from the state in which it is at the moment of the maneuver in
the asymptotically stable equilibrium 1 ( e ") ,0, 1 ( e ") . ( )
Fig. 1. The 1(e) and 2(e) coordinates of the equilibriums on the manifold MV.
Fig. 2. The 1(e)+2k and 2(e)+2k coordinates of the equilibriums on the manifold MV.
Fig. 3. A successful maneuver on P1 :

11 = 0.078669740237840 [rad]; q11= 0 [rad/s] ; 11= 0.428832005303479 [rad]
12 = 0.065516737567037 [rad]; q12= 0 [rad/s] ; 12= - 0.698066723826469 [rad]
Fig. 4. An unsuccessful maneuver on P2 :

21 = 0.064883075974905 [rad]; q21= 0 [rad/s] ; 21= 0.767462467841413 [rad]
12 = 0.065516737567037 [rad]; q12= 0 [rad/s] ; 12= 0.698066723826469 [rad] instead of
21 = 0.064883075974905 [rad]; q21= 0 [rad/s] ; 21= 0.767462467841413 [rad]
22 = 0.046845089090947 [rad]; q22= 0 [rad/s] ; 22= 1.036697186364400 [rad].
Fig. 5. Oscillation when e = - 0.05 [rad] and the starting point is :

1 = 0.086974288419088 [rad]; q1 = 0 [rad/sec]; 1= 0.159329728679884[rad].
Fig. 6. Oscillation when e = 0.048 [rad] and the starting point is :

1 = 0.086974288419088 [rad]; q1 = 0 [rad/sec]; 1= 0.159329728679884[rad].
The behavior of the ADMIRE aircraft changes when the maneuver e ' e " is so that
(
e ' e , e ) (
and e " e , e ) . Computation shows that after such a maneuver and q
oscillate with the same period and tends to + or . (Figs.5, 6)
The oscillation presented in Figs. 5,6 is a non catastrophic bifurcation, because if e is reset,
then equilibrium is recovered, as it is illustrated in Fig.7.
Fig. 7. Resetting e = 0.048[ rad ] < eo after 3000 [s] of oscillations to e = eo , equilibrium is
recovered.
7. Conclusion
For an unmanned aircraft whose automatic flight control system during a longitudinal flight
with constant forward velocity fails, the following statements hold:
1. If the elevator deflection is in the range given by formula (19), then the movement
around the center of mass is stationary or tends to a stationary state.
2. If the elevator deflection exceeds the value given by formula (36), then the movement
around the center of mass becomes oscillatory decreasing and when the elevator
deflection is less than the value given by formula (37), then the movement around the
center of mass becomes oscillatory increasing.
3. This oscillatory movement is not catastrophic, because if the elevator deflection is reset
in the range given by (19), then the movement around the center of mass becomes
stationary.
4. Numerical investigation of the oscillation susceptibility (when the automatic flight
control system fails) in the general non linear model of the longitudinal flight with
constant forward velocity reveals similar behaviour as that which has been proved
theoretically and numerically in the framework of the simplified model. As far as we
know, in the general non linear model of the longitudinal flight with constant forward
velocity the existence of the oscillatory solution never has been proved theoretically.
5. A task for a new research could be the proof of the existence of the oscillatory solutions
in the general model.
8. References
Balint, St.; Balint, A.M. & Ionita, A. (2009a). Oscillation susceptibility along the path of the
longitudinal flight equilibriums in ADMIRE model. J. Aerospace Eng. 22, 4 (October
2009) 423-433 ISSN 0893-1321.
Balint, St.; Balint, A.M. & Ionita, A. (2009b). Oscillation susceptibility analysis of the
ADMIRE aircraft along the path of longitudinal flight equilibrium. Differential
Equations and Nonlinear Mechanics 2009. Article ID 842656 (June 2009) 1-26 . ISSN:
1687-4099
Balint, St.; Balint, A.M. & Kaslik, E. (2010b) Existence of oscillatory solutions along the path
of longitudinal flight equilibriums of an unmanned aircraft, when the automatic
flight control system fails J. Math. Analysis and Applic., 363, 2(March 2010) 366-382.
ISSN 0022-247X.
Balint, St.; Kaslik E.; Balint A.M. & Ionita A. (2009c). Numerical analysis of the oscillation
susceptibility along the path of the longitudinal flight equilibria of a reentry
vehicle. Nonlinear Analysis:Theory, Methods and applic., 71, 12 (Dec.2009) e35-e54.
ISSN: 0362-546X
Balint, St.; Kaslik, E. & Balint, A.M. (2010a). Numerical analysis of the oscillation
susceptibility along the path of the longitudinal flight equilibria of a reentry
vehicle. Nonlinear Analysis: Real World Applic., 11, 3 (June 2010)1953-1962.ISSN:
1468-1218
Caruntu, B.; Balint, St. & Balint, A.M. (2005). Improved estimation of an asymptotically
stable equilibrium-state of the ALFLEX reentry vehicle Proceedings of the 5th
International Conference on Nonlinear Problems in Aviation and Aerospace Science,
pp.129-136, ISBN: 1-904868-48-7, Timisoara, June 2-4, 2004, Cambridge Scientific
Publishers, Cambridge.
Cook, M. (1997). Flight dynamics principles, John Wiley & Sons, ISBN-10: 047023590X, ISBN-
13: 978-0470235904 , New York.
Etkin, B. & Reid, L. (1996). Dynamics of flight: Stability and Control, John Wiley & Sons, ISBN:
0-471-03418-5, New York.
Gaines, R.& Mawhin J. (1977). Coincidence Degree and Nonlinear Differential Equations,
Springer, ISBN: 3-540-08067-8, Berlin New York.
Goto, N. & Matsumoto, K.(2000). Bifurcation analysis for the control of a reentry vehicle.
Proceedings of the 3rd International Conference on Nonlinear Problems in Aviation and
Aerospace Science,pp.167-175 ISBN 0 9526643 2 1 Daytona Beach, May 10-12, 2000
Cambridge Scientific Publishers, Cambridge
Kaslik, E. & Balint, St.. (2007). Structural stability of simplified dynamical system governing
motion of ALFLEX reentry vehicle, J. Aerospace Engineering, 20, 4 (Oct. 2007) 215-
219. ISSN 0893-1321.
Kaslik, E.; Balint, A.M.; Chilarescu, C. & Balint, St. (2002). The control of rolling maneuver
Nonlinear Studies, 9,4, (Dec.2002) 331-360. ISSN: 1359-8678 (print) 2153-4373
(online)
Kaslik, E.; Balint, A.M.; Grigis, A. & Balint, St. (2004a). The controllability of the path
capture and steady descent flight of ALFLEX. Nonlinear Studies, 11,4, (Dec.2004)
674-690. ISSN: 1359-8678 (print) 2153-4373 (online)
Kaslik, E.; Balint, A.M.; Birauas, S. & Balint, St. (2004 b). On the controllability of the roll rate
of the ALFLEX reentry vehicle, Nonlinear Studies 11, 4 (Dec.2004) 543-564. ISSN:
1359-8678 (print) 2153-4373 (online)
Kaslik, E.; Balint, A.M.; Grigis, A. & Balint, St. (2005a). On the set of equilibrium states
defined by a simplified model of the ALFLEX reentry vehicle Proceedings of the 5th
International Conference on Nonlinear problems in aviation and aerospace science, pp.359-
372. ISBN: 1-904868-48-7, Timisoara, June 2-4 , 2004, Cambridge Scientific
Publishers, Cambridge
Kaslik, E.; Balint, A.M.; Grigis, A. & Balint, St. (2005b) Considerations concerning
the controllability of a hyperbolic equilibrium state. Proceedings of the 5th
International Conference on Nonlinear problems in aviation and aerospace science, pp.383-
389 ISBN: 1-904868-48-7, Timisoara, June 2-4 , 2004, Cambridge Scientific
Publishers, Cambridge
Kaslik, E.; Balint, A.M.; Grigis, A. & Balint, St. (2005c) Control procedures using domains
of attraction, Nonlinear Analysis, 63, 5-7, (Nov.-Dec.2005) e2397-e2407 ISSN: 0362-
546X
Kaslik, E.& Balint, St. (2010). Existence of oscillatory solutions in longitudinal
flight dynamics. Int. J. Nonlinear Mechanics, 45(2) (March 2010) 159-168. ISSN: 0020-
7462
Kish, B.A. ; Mosle, W.B. & Remaly, A.S. (1997) A limited flight test investigation of pilot-
induced oscillation due to rate limiting. Proceedings of the AIAA Guidance,
Navigation, and Control Conference AIAA-97-3703 New Orleans (August 1997)
1332-1341
Klyde, D.H.; McRuer, D.T.; & Myers, T.T. (1997). Pilot-induced oscillation analysis and
prediction with actuator rate limiting. J. Guidance, Control and Dynamics, 20,1,
(Jan.1997) 81-89 ISSN 0731-5090
Mawhin, J. (1972). Equivalence theorems for nonlinear operator equations and coincidence
degree theory, Journal of Differential Equations 12, 3, 610-636 (Nov. 1972). ISSN: 0022-
0396
Mehra, R.K. & Prasanth, R.K. (1998). Bifurcation and limit cycle analysis of nonlinear pilot
induced oscillations, (1998) AIAA Paper98-4249 AIIA Atmosphere Flight Mechanics
Conf., 10-12, August 1998, Boston, MA.
Mehra, R.K. ; Kessel, W.C. & Carroll, J.V. (1977). Global stability and control analysis of
aircraft at high angles of attack. ONR-CR215-248, vol.1-4 (June 1977) 81-153.
Shamma,I. & Athans, M. (1991). Guaranted properties of gain scheduled control for linear
parameter varying plants Automatica, 27,3, 559-564 (May 1991), ISSN: 0005-1098.

Advances in Flight Control Systems

Uploaded by

Copyright:

Available Formats

Advances in Flight Control Systems

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Advances in Flight Control Systems

Uploaded by

Copyright:

Available Formats

What are some of the topics covered in the document?

What are some of the topics covered in the document?

What are some of the references cited in the document?

What are some of the references cited in the document?

ADVANCES IN

FLIGHT CONTROL SYSTEMS

Copyright 2011 InTech

Publishing Process Manager Ivana Lorkovic

First published March, 2011

A free online edition of this book is available at www.intechopen.com

Advances in Flight Control Systems, Edited by Agneta Balint

Chapter 1 Adaptive Fight Control Actuators

Chapter 2 Adaptive Backstepping Flight Control

Chapter 3 Hybrid Adaptive Flight Control

Chapter 4 Application of Evolutionary Computing

Chapter 5 Fault Tolerant Flight Control,

Chapter 6 Design of Intelligent Fault-Tolerant Flight

Chapter 7 Active Fault Diagnosis and Major Actuator

Chapter 8 Fault-Tolerance of a Transport Aircraft

Chapter 9 Acceleration-based 3D Flight Control for UAVs:

Chapter 10 Autonomous Flight Control System

Chapter 11 Autonomous Flight Control

Chapter 12 Hierarchical Control Design of a UAV Helicopter 239

Chapter 13 Comparison of Flight Control System

Chapter 14 Oscillation Susceptibility of an Unmanned Aircraft

Nonlinear problems in flight control have stimulated cooperation among engineers

In Chapter 2 an integrated, though cascaded Lyapunov-based adaptive backstepping

In Chapter 3 a hybrid adaptive flight control method as another possibility to reduce

In Chapter 4 the purpose is to design a compensator using an evolutionary computing

In Chapter 7 a nonlinear UAV model, which allows simulating asymmetrical control

In Chapter 8 a scheme of a fault-tolerant flight control system is proposed. It is com-

In chapter 11 an autonomous flight of a RC helicopter with small-wireless camera and

In chapter 13 an overview concerning conventional, fuzzy logic-based, neural net-

In chapter 14 the oscillation susceptibility of an unmanned aircraft whose automatic

February 26, 2011.

Adaptive Fight Control Actuators and

Project p=piezoelectric, s=shape memory alloy Agency/

Table 1. Summary of Adaptive Aerostructures Projects with Direct Connections to

2. Missile, munition and supersonic Uninhabited Aerial Vehicle (UAV) flight

2.1 Early fin and wing designs

Conventionally Attached Piezoelectric Actuators:

A11 A12 0 B11 B12 0 (A11 + A12 )

A11 A12 2A16 B11 B12 2B16 (A11 + A12 )

Fig. 2. Constrained Spar DAP Torque-Plate Missile Fin13

Fig. 3. Finitel Element Models of Active Supersonic & Subsonic Fins15

2.2 Flexspar: The first flight enabling adaptive system

Fig. 4. Flexspar Fin Anatomy

Fig. 5. Tip-Joint Flexspar Fin Geometric Parameters

Fig. 6. Typical Flexspar Deflection s and Correlation Levels17

Fig. 7. Flexspar-Equipped TOW-2B Missile in Wind Tunnel

2.2 Cruise missile and gravity weapon applications

Fig. 8. Smart Compressed Reversed Adaptive Munition (SCRAM)19,20

2.2.2 Weapon integration and design technology

Fig. 9. GBU-39 Small Diameter Bomb with Adaptive Canards21

2.2.3 Miniature cruise missile airframe technology demonstrator

Fig. 10. Boeing/USAF Miniature Cruise Missile with Adaptive Wings24

2.3 Hard-launched munitions and supersonic Nano Aerial Vehicles (NAVs)

Fig. 11. Barrel-Launched Adaptive Munition (BLAM) Configuration

2.3.1.3 Rotational Accelerations

2.3.2 Supersonic Nano Aerial Vehicles (NAVs)

x ref = V ref cos ref (7)

y ref = V ref sin ref (8)

X 1 = A1F1 ( X ,U ) + B1G1 ( X ,U , X 2 ) + H 1 ( X ) (16)

B2 X 3des,0 = C 2 Z2 A2 F1 H 2 + X 2des , C 2 = C 2T > 0 (29)

A3 B 3U 0 = C 3Z3 A3F3 H 3 + X 3des , C 3 = C 3T > 0 (34)

Since B = 1, it can be established that