Flutter Prediction From Flight Flutter Test Data

JOURNAL OF AIRCRAFT
Vol. 38, No. 2, March– April 2001
Flutter Prediction from Flight Flutter Test Data
G. Dimitriadis¤ and J. E. Cooper†

University of Manchester, Manchester, England M13 9PL, United Kingdom
The most common approach to ight utter testing is to track estimated modal damping ratios of an aircraft
over a number of ight conditions. These damping trends are then extrapolated to predict whether it is safe to
move to the next test point and also to determine the utter speed. In the quest for more reliable and ef cient
ight utter testing procedures, a number of alternative data analysis methods have been proposed. Five of these
approaches are compared on two simulated aeroelastic models. The comparison is based on both the accuracy of
prediction and the ef ciency of each method. It is found that, for simple aeroelastic systems, the Nissim and Gilyard
method (Nissim, E., and Gilyard, G. B., “Method for Experimental Determination of Flutter Speed by Parameter
Identi cation,” AIAA Paper 89-1324, 1989) yields the best utter predictions and is also the least computationally
Downloaded by CALIFORNIA INST OF TECHNOLOGY on May 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/2.2770
expensive approach. However, for larger systems, simpler approaches such as the damping t and envelope function
methods are found to be most reliable.
Nomenclature Introduction
Ai
Ai
= polynomial coef cients in characteristic
equation, utter margin method (FM)
= aerodynamic matrix coef cients, Nissim and
D URING the design stage of a new aircraft, the aeroelastic be-
havior is estimated by producing detailed mathematical and
aeroelastic wind-tunnel models. The critical phenomenon is utter,
Gilyard method (NG) a violent unstable oscillation, which must be avoided throughout
a.m/ = moving average (MA) coef cients, the ight envelope. Airworthiness regulations require that stability
autoregressivemoving average method (ARMA) throughoutthe required ight regime is demonstratedby tests of the
b.m/ = autoregressive (AR) coef cients, ARMA actual ying aircraft, commonly termed ight utter tests. Flight
Cs = structural damping matrix, NG utter tests consist of ying an aircraft at a range of subcritical air
CN = damping matrix, NG speeds while applying some form of excitation to the structure. The
env.t/ = decay envelope or envelope function, envelope response of the structure is measured at a number of measurement
function method stations, and the data is curve tted to determine the stability at
F = utter margin, FM the current ight speed and predict whether it is safe to proceed to
FN = forcing vector, NG the next test point. Of interest is the speed at which utter is pre-
F § .l/, G.1/, = Jury stability criteria, ARMA dicted. This process is repeated at numerous ight conditions until
G.¡1/ the envelope is cleared. In practice,the most common response data
g.t / = forcing function, NG analysis procedure is to estimate the damping present in the aircraft
J = model order, ARMA response and to track its variation with air speed.
Ks = structural stiffness matrix, NG Even in the days of high-speed computers and sophisticated data
KN = stiffness matrix, NG measurement and analysis tools, utter testing remains as much an
M N = mass matrix, NG art as a science. Subcritical damping data cannot always be safely
m = number of degrees of freedom, NG extrapolatedto obtain an accurate prediction for the utter velocity.
nf = number of points used in identi cation, NG Nonlinearities in the control system or in the aerodynamics and
q = dynamic pressure, FM structure of an aircraft can critically affect the aeroelastic behavior.
q = generalized coordinates, NG Finally, the aeroelasticstability can change from positive to negative
S = shape function, envelope function method with an increase in air speed of only a few knots, and the whole
tN = time centroid of decay envelope, envelope procedure is very dangerous and time consuming.
function method There are various methods to predict the utter speed using ight
u.i / = discrete values of input, ARMA utter test data. The purpose of this paper is to examineand compare
Vf = utter velocity a number of the approaches that are used in industry and appear in
V1 = freestream velocity, NG the literature. This comparison has not been attempted before to the
y.i / = discrete values of output, ARMA authors’ knowledge. Both the ef ciency and quality of the predic-
y.t / = impulse response, envelope function method tions obtainedby the methodsare comparedusing simulatedsystems
y H .t/ = Hilbert transform of the impulse response, to provide a comprehensive overview of contemporary utter pre-
envelope function method diction procedures. A more accurate and ef cient utter prediction
¯ = reference length, NG would decrease the time taken for ight utter testing as well as
¯i = real part of ith eigenvalue, FM increase safety.
³ = damping ratio, ARMA
¸i = i th eigenvalue, FM
Flutter Prediction Methods
½ = density of air, NG
! = natural frequency, ARMA Damping Ratio Variation with Air Speed
!i = imaginary part of ith eigenvalue, FM Traditionally, the most widely used indicators of the stability of
an aeroelastic system are the modal dampings and their variation
Received 20 May 1999; revision received 15 September 2000; accepted with freestream velocity. At utter, the damping in at least one of
for publication 15 November 2000. Copyright ° c 2001 by the American the modes is zero, thus causing self-excited oscillations. Beyond
Institute of Aeronautics and Astronautics, Inc. All rights reserved. this speed, the system becomes unstable. In the typical ight utter
¤ Lecturer, Aerospace Engineering, Oxford Road. test, the damping ratios for all of the signi cant modes are evalu-
† Senior Lecturer, Aerospace Engineering, Oxford Road. Member AIAA.
ated at a number of subcriticalair speeds using system identi cation
355
356 DIMITRIADIS AND COOPER
methods. The damping ratio trends are then curve tted by a poly- Some experimentalevaluationof the method suggeststhat using a
nomial, or by hand, and extrapolated to estimate the utter velocity quadratic t of the utter margin can yield results that are very sensi-
and to examine the stability at the next proposed ight test. tive to errors or uncertaintyin the experimental data.6 An alternative
There are numerous system identi cation methods, both in is to use linear extrapolation, again in a least-squares sense.
the time and the frequency domain, that allow the calculation of
the modal parameters of a vibrating system.1 Here, a version of the Envelope Function
rational fraction polynomial method2 is used. Obviously, the accu- The envelope function7 was originally proposed as a tool to pro-
racy of predictingthe utter speed dependson how well the damping vide an assessmentof overall stability to complement standard anal-
ratios are estimated. ysis. However, it has since been used in practice to provide ight
utter clearance of the ALTOS high-altitude research aircraft.
Flutter Margin Method The basis of the method is that the impulse response of any stable
The utter margin method (FM) was rst presented in Ref. 3. The damped system is decaying, with the shape of the decay in the time
basis of the approach is the quest for a more fundamental stability domain being described by the decay envelope. As the damping
criterion than just tracking the damping present in the system. In its in a given aeroelastic system decreases, the decay envelope grows
original form, the approach only covers binary utter; however, in wider, eventually becoming a rectangle as the damping becomes
Ref. 4 an extension of the technique to trinary utter is presented. zero. When the position of the centroid of the decay envelope and
The FM employs the Routh stability criterion.5 For a two-degree- the way that it shifts on the time axis as the damping decreases are
of-freedomsystem, for which the characteristicequation is a quartic evaluated, it is possible to assess the stability of the system.
of the form For an aeroelastic system with impulse response y.t /, the decay
envelope, or envelope function, is given by

¸4 C A3 ¸3 C A 2 ¸2 C A1 ¸ C A0 D 0 (1) p
env.t / D y.t /2 C y H .t /2 (5)
the Routh stability criterion reduces to
where y H .t / is the Hilberttransformof the impulse response,de ned
[A 2 . A1 = A 3 / ¡ .A 1 = A3 /2 ¡ A0 ] > 0 (2) as
y H .t / D F ¡1 fIm[Y .!/]g ¡ j Re[Y .!/] (6)

Hence, for a two-degree-of-freedom system, there is a quantity
described by Eq. (2) that has to be positive for the system to be
where Y .!/ is the Fourier transform of y.t / and Im and Re are
stable and becomes zero when instability is reached. This quantity imaginary and real part, respectively.The time centroid of the decay
is termed the utter margin and is given by (in slightly modi ed
envelope is given by
form)
µ³ ´ ³ ´¶2 "³ ´ R tmax
!22 ¡ !21 ¯22 ¡ ¯12 !22 C !12 env.t /t dt
FD C C 4¯1 ¯2 tN D R0 tmax (7)
2 2 2 0
env.t / dt
³ ´2 # "³ ´³ ´ The upper limit of integration, tmax , serves to de ne the rectangle
¯2 C ¯1 ¯2 ¡ ¯1 !22 ¡ !12 within which the integration takes place. For a single-degree-of-
C2 ¡
2 ¯2 C ¯1 2 freedom system, when the damping is zero, the time centroid lies
at t D tmax =2. For a multi-degree-of-freedom system, it is suggested
³ ´2 # 2 that t ¼ tmax =2 is an adequate approximation for the position of the
¯2 C ¯ 1 time centroid.7
C (3) Because tN tends to increase as the damping drops, its inverse is
2
usually employed as the signi cant shape parameter, that is,
which is obtained from Eq. (3), after substituting for the two sets of
complex conjugate eigenvalues, ¸1 ; : : : ; ¸4 , such that S D 1=tN (8)
¸1;2 D ¯1 § i !1 ; ¸3;4 D ¯2 § i !2 in which case the value of S at the utter condition is S D 2= tmax for
a single-degree-of-freedom system. For multi-degree-of-freedom
The application of the method is quite straightforward. For a systems S ¼ 2=tmax .
two-degree-of-freedom system, the response to a known input at The envelopefunction utter testingprocedureis to evaluate S at a
a particular (subcritical) air speed is recorded and the eigenvalues number of subcritical air speeds. The variation of S with air speed is
of the system are calculated.1 These are then used to compute the then curve tted using a polynomial, as with the damping method,
utter margin. Where some further derivation3 is used, it can be and extrapolated to the point where S D 2= tmax , thus yielding the
shown that the utter margin is a quadratic function of the dynamic utter velocity.
pressure, that is, The impulse responseof an aeroelastic system may be obtainedin
two ways: 1) direct measurement of the aircraft response to impul-
F D B2 q 2 C B1 q C B0 (4) sive excitation, for example, stick jerk, and 2) application of other
excitation functions, for example, white noise, chirp. The Fourier
where B0 , B1 , and B2 are coef cients to be evaluated. Hence, if transforms of the response and excitation are then divided to yield
the utter margin is known at three different air speeds, it can be the frequency response function (FRF). The impulse response is
tted by a second-order polynomial and subsequently extrapolated. obtained as the inverse Fourier transform of the FRF.
Flutter occurs when F D 0. In practice, to counteract the effects of In the present implementation, the second approach was chosen.
experimental uncertainty, the utter margin is estimated at a wider The disadvantage of the rst technique is that, in practice, it is im-
range of subcriticalair speedsand then tted in a least-squaressense. possible to apply a perfect impulse to a system as complex as an
That the utter margin is derived for a two-degree-of-freedom aircraft; hence, the measured response may not be close enough to
system does not imply that the method cannot be used with larger the impulse response. The second approach also allows the smooth-
systems. In fact, the utter mechanism is often dependent on two ing of some of the measurement noise by careful choice of the
modes only. If it is known beforehand which two modes will cause Fourier transform window. Additionally,the impulse responses can
utter, the FM can be applied successfully; otherwise, all possible be exponentially weighted8 to improve the noise-to-signal ratio at
pairs of modes must be examined. the tail end of the response.
DIMITRIADIS AND COOPER 357
Nissim and Gilyard Method are usually very small compared to structural inertia terms. In other
The two methods already described deal with attempting to eval- words, matrices C and K in Eq. (12) can be broken down into struc-
uate the stability of a given aeroelastic system and, hence, the utter tural and aerodynamic part, that is,
velocity. The Nissim and Gilyard method (NG),9 which is an exten- C D Cs C 12 ½ V1 A1 ; K D Ks C 12 ½V12 A0 (16)
sion to a technique developed by Skingle et al.,10 adopts a different
approach by attempting to identify the whole system, including its with subscript s denoting structural terms. If the identi cation pro-
aerodynamic variation with freestream air speed by estimating its cess described earlier is repeated at two distinct velocities, then the
equations of motion. Then, the identi ed system can be solved for structuraland aerodynamicmatrices can be evaluatedseparatelyand
different velocities to obtain the utter speed. the behaviorof the system at any air speed can be predictedeither by
The equations of motion for a forced aeroelastic system are integrating the equations of motion or by calculating the system’s
MN qR C CP
N q C Kq
N D Fg.t
N / (9) eigenvalues. Finally, the utter speed can be obtained by means of
a suitable iterative calculation, for example, evaluating the system
If Eq. (9) is transposed to the frequency domain and premultiplied damping at increasing speeds until utter is reached.
by MN ¡1 , the equations of motion become It should be noted that the NG requires modal responses to work.
f¡I!2 C C| ! C Kgq.!/ D Fg.!/ (10) Consequently, if only physical coordinates z are available, then the
modal matrix U should be evaluated and used to obtain the modal
where coordinates from
N ¡1 C;
CDM N N ¡1 K;
KDM N N ¡1 FN
FDM zDU q (17)
Autoregressive Moving Average-Based Method

and multiplication by | ! denotes differentiation. Divide both sides
of Eq. (10) by g.!/ and rearrange The equations of motion of a dynamic system can be expressed
as a sum of the regressive response terms equal to the value of
C| !Hq .!/ C KHq .!/ ¡ F D I! 2 Hq .!/ (11) the regressive forcing term. This is the basis of the representation
of dynamical systems by autoregressive moving-average (ARMA)
where Hq .!/ D q.!/=g.!/. For an m-degree-of-freedom system, models, where the AR part denotesthe terms containingthe response
Eq. (11) can be rearranged and expanded as y and the MA part denotes the white noise excitation terms u.
0 1 0 1
Hq1 .!1 / ::: Hqm .!1 / | !1 Hq1 .!1 / ::: | !1 Hqm .!1 / ¡1 0 1 !12 Hq1 .!1 / ::: !21 Hqm .!1 /
B C T B C
B Hq1 .!2 / : : : Hqm .!2 / | !2 Hq1 .!2 / ::: | !2 Hqm .!2 / ¡1C K B !2 H .! /
B 2 q1 2 ::: !22 Hqm .!2 /
C
C
B :: :: :: :: :: :: C @
:: C C TA D B C (12)
B B :: :: :: C
@ : : : : : : : A @ : : :
¡ ¢ ¡ ¢ ¡ ¢ ¡ ¢ FT ¡ ¢ ¡ ¢A
Hq1 !n f ::: Hqm !n f | !n f Hq1 !n f ::: | !n f Hqm !n f ¡1 2 H
!n f q1 !n f ::: !n2 f Hqm !n f
Equation (12) is of the form The general form of an ARMA is12
TX D B (13) X
2J X
2J ¡ 1
b.m/y.i C m/ D a.m/u.i C m/ (18)
where X is to be evaluated in a least-squares sense. Strictly speak- mD0 m D0
ing, n f only needs to be equal to m for a successful identi cation;
The order of the model, J , can be equal to the number of modes of
however, to counteract the effect of noise in the responses, usu-
the system to be modeled but, in the presenceof experimentalnoise,
ally n f > m and the equation is solved in a least-squares sense. For
it is usually taken to be greater than the number of modes.
increased accuracy, multiple forcing vectors can be applied to the
To identify a given dynamic system, the ARMA equation is ap-
system. This is particularly relevant to the case where the responses
plied to a set of single input/single output sampled data. The un-
include a high level of noise.
known coef cients, a, b, and J are evaluated using a parameter
To obtain an identi ed model of the system at all air speeds, a
estimation algorithm, the simplest of which is the least-squarespro-
second identi cation needs to take place at a different freestream
cedure. This evaluation can be simpli ed by dividing the ARMA
velocity. According to Ref. 11, the unsteady aerodynamic matrix
equation throughout by b.2J /, so that the leading AR coef cient is
Q.| !/ for an aeroelastic system is given by
" always unity.
³ ´ ³ ´2 If Eq. (18) is evaluated at instances i D 1; : : : ; k, then the coef -
1 b b
Q.| !/ D ½ V12 AN 0 C AN 1 | ! C AN 2 .| !/2 cients can be obtained using
2 V1 V1 0 1
y.2J / ::: y.1/ u.2 J / : : : u.1/
³ ´X # B y.2 J C 1/ ::: y.2/ u.2J C 1/ : : : u.2/C
V1
nL
AN .n C 2/
B C
C (14) B :: :: :: :: C
b | ! C .V1 =b/bn
@ : : : : A
nD1
y.2J C k ¡ 1/ ::: y.k/ u.2J C k ¡ 1/ : : : u.k/
where n L depends on the desired accuracy but is usually no more
than four, b is a reference length, and AN 0 ; : : : ; AN n L are matrix coef - 8 9
cients. For a quasi-steadyaerodynamic representation,the lag terms >
> b.2J ¡ 1/>
>
>
> :: >
> 8 9
> >
are ignored leading to >
>
> : >
>
> > y.2J C 1/>
>
> >
< = < y.2J C 2/> =
Q.| !/ D 12 ½V12 A0 C 12 ½ V1 A1 | ! (15) b.0/
£ D¡ :: (19)
>
> a.2J ¡ 1/>
> >
> : >
>
> > > >
where Ai D b AN i for i D 1; 2; A0 are aerodynamic stiffness terms;
i >
>
>
:: >
>
>
:
y.2J C k/
;
and A1 are aerodynamic damping terms. Terms A2 , representing >
> : >
>
: ;
aerodynamic inertia terms, have also been neglected because they a.0/
This matrix equation can be solved using any number of schemes, Consequently, the utter condition is de ned as the rst air speed
such as those described in Refs. 13 and 14. The eigenvalues of the at which any of the criteria G.1/, G.¡1/, and F § .l/ becomes zero.
system can be obtained by forming the characteristic polynomial Hence, the system can be identi ed at a range of velocities, and the
G.¹/ D ¹2 J C b.2J ¡ 1/¹2 J ¡ 1 C ¢ ¢ ¢ C b.1/¹ C b.0/ D 0 (20) values of the criteria plottedagainstair speed and then curve tted by
a polynomial to yield the points were they intersectthe velocity axis
and, subsequently,evaluating its roots and, hence, the utter velocities. This particular implementation of
£¡ p ¢ ¤ the Jury stability criterion is disadvantageous compared to the FM
¹1;2 D exp ¡³ ! § j ! 1 ¡ ³ 2 1t (21) in that the criteria do not vary in a predictablemanner with velocity.
Instead, if polynomial tting is attempted, the order will not be
The ARMA method, as described up to now is a system iden-
known, as with the damping t and envelope methods.
ti cation technique. Matsuzaki et al.12 suggest a procedure for us-
ing an ARMA representation of an aeroelastic system as a means Preliminary Validation of the Methods
of predicting the utter velocity. The basis of the approach is the The methods were applied to two simulated aeroelastic models,
Jury determinant method for evaluating the stability of a discrete- one of them modeling a simple, three-degree-of-freedom, rectan-
time system, which is very similar to the Routh– Hurwitz criterion gular wing with control surface, referred to as the Hancock model
for continuous-time systems (used in conjunction with the FM). (see Ref. 15) and the other modeling a four-enginedcivil transport,7
The Jury stability criterion applies to the characteristic polynomial, referred to as the Sim-2 model. The Hancock model was chosen to
Eq. (20), such that the system is stable if demonstrate the methods on the most simplistic aeroelastic model
possible. The Sim-2 model was chosen because it is a multimodal,
G.1/ > 0; G.¡1/ > 0 (22)
industrially validated model and is characteristic of a number of

and for l D 1; 3; : : : ; 2J ¡ 1, commercial aircraft. Nevertheless, in both cases the utter mech-
anism is essentially binary. Figures 1 and 2 show the variation of
F § .l/ D
X l¤ § Yl¤
>0 (23) natural frequencies and damping ratios for both models. In the case
of the Sim-2 model, only some representativemodes are plotted.The
where X 2¤ J ¡ 1 and Y2¤J ¡ 1 are .2J ¡ 1/ £ .2J ¡1/ matrices, given by utter speeds of the two models were 44.07 m/s for the Hancock
0 1 model and 398 kn for the Sim-2 model (5000 ft).
b.2 J / b.2 J ¡ 1/ ::: : : : b.3/ b.2/ With both models, and for all of the methods, the excitationwave-
B 0 b.2 J / b.2J ¡ 1/ : : : b.4/ b.3/ C forms used were frequencysweeps,sweepingthroughall of the natu-
B C
B C ral frequenciesof the models so as to exciteall modes. The excitation
X 2¤ J ¡ 1 DB 0 0 b.2J / : : : b.5/ b.4/ C
forces were applied through control surfaces and were symmetric.
B :: :: :: C
@ : : : A Additionally, all responses were contaminated by 5% rms simu-
lated noise to approximate the effects of experimental uncertainty.
0 ::: ::: ::: ::: b.2J /
Each method was tested in a variety of velocity ranges, as shown in
(24) Tables 1 and 2 for the Hancock and Sim-2 models, respectively. In
0 1 each of the ranges, a number of tests were performed at increments
b.2J ¡ 2/ b.2J ¡ 3/ ::: ::: b.1/ b.0/ of 7% of the true utter speed. Results for the NG are not presented
Bb.2J ¡ 3/ b.2J ¡ 4/ : : : : : : b.0/ 0 C for the Sim-2 model, for reasons that will be explained later.
B C Noted that the validationof the methods presented here is speci c
B C
Y2¤J ¡ 1 D Bb.2J ¡ 4/ b.2J ¡ 5/ : : : b.0/ 0 0 C to the particular models and implementations chosen.
B :: :: :: C
@ : : : A
Damping Fit Method
b.0/ 0 ::: ::: ::: 0 Preliminary tests of the method showed that curve ts by a range
(25) of polynomial orders (between 3 and 8) can give the optimal utter
Fig. 1 Variation of natural frequencies and dampings with air speed, Hancock model.
Fig. 2 Variation of natural frequencies and dampings with air speed, Sim-2 model.
Table 1 Modulus of errors in utter estimates, Hancock model
Velocity range
(% of utter Damping
speed) t, % FM, % Envelope, % NG, % ARMAX, %
23– 43 14.4 6.8 13.0 2.5 46.2
23– 57 5.1 3.9 2.1 4.7 17.5
23– 70 9.1 2.5 0.6 1.3 21.0
23– 84 0.1 1.1 4.8 0.1 15.4
23– 98 0.5 0.3 0.1 0.1 0.3
45– 66 17.5 1.3 6.9 0.8 23.6
45– 79 0.9 1.7 1.9 0.7 8.7
45– 93 0.1 0.1 0.3 0.2 1.6
68– 88 0.2 0.8 0.4 0.7 6.1
Table 2 Modulus of errors in utter estimates, Sim-2 model
Velocity range
(% of utter damping
speed) t, % FM, % Envelope, % ARMAX, %
23 – 43 28.4 —— 5.5 53.3
23 – 57 17.0 18.8 35.1 32.9 Fig. 3 Flutter prediction using damping t; 20% rms noise.
23 – 70 9.2 15.5 16.2 6.5
23 – 84 4.6 3.8 2.5 8.7 Because, in practice, the utter speed is not known in advance, it is
23 – 98 0.3 0.8 0.7 1.0 impossible to predict which order of t will give the best estimate.
45 – 66 2.8 20.8 9.4 12.1
45 – 79 2.3 5.5 6.0 8.7
The sensitivity of the method to noise is governed by the sen-
45 – 93 0.5 —— 1.3 2.2 sitivity of the rational fraction polynomial (RFP) procedure. The
68 – 88 2.9 —— 0.5 2.0 deteriorating quality of the latter with increasing levels of noise af-
fects the utter prediction negatively, as is seen in Fig. 3, where a
utter prediction is shown for results with 20% rms noise. In Fig. 3,
the circles are estimated damping ratio data, the solid curve is the
predictions. Additionally, changing the polynomial order by one best-case polynomial t of the estimated damping data, the crosses
can give wildly inaccurate results. In general, orders between three are theoretical damping data, and the dashed curve the best-case
and six are the most suitable choices; however, utter predictionsin polynomial t of the theoretical damping ratio data (obtained from
Tables 1 and 2 were out by up to 17.5% of the true utter speed for the the equations of motion). A comparison between the latter and the
Hancock model and 28.4% for the Sim-2 model. Furthermore, if too best-case t of the estimated damping ratios shows that the utter
high an order is used (typically above 10, depending on how many speed has been overestimatedby quite a margin (4.4%) and also that
points are tted), the matrix used for the polynomial t becomes the estimated subcritical behavior of the damping curve is wrong.
rank de cient, and the t itself fails completely.
Damping trends tend to be relatively smooth and no more than
quartic in shape even when hard utter is encountered.In actual fact, FM
no case was encounteredwhere the use of polynomial orders higher As mentioned earlier, successful application of the FM to a
than six was needed. However, this is by no means a general result. multi-degree-of-freedom (DOF) aeroelastic system depends on the
knowledge of the utter mechanism, that is, which two modes will and F also becomeszero. This phenomenonis shown in Fig. 4 for the
combine to cause utter. When the technique is applied to these Hancock model, whose utter mechanism includes the wing torsion
two modes, a representative system is analyzed. Note that utter and control surface torsion modes, the damping of the control sur-
margins calculated from any combination of modes containing the face torsion mode becoming zero at utter. Figure 4 shows utter
mode that becomes unstable at utter will become zero at the utter margins calculated for all three combinationsof modes. In the wing
velocity. In Ref. 16 it is shown that Eq. (3) can be rewritten as bending– wing torsion case, the utter margin never becomes zero.
" ³ ´2 #(³ ´2 "³ ´ In the other two cases, which contain the control surface torsion
¯2 ¡ ¯1 !22 ¡ !21 2 !22 C !12 mode, the utter margin becomes zero at utter. This phenomenon
F D 1¡ C.¯1 C ¯2 /
¯2 C ¯1 2 2 is because any system formed of two modes from another system,
one of which is unstable, will also be unstable. However, Fig. 4
³ ´2 #) shows that the variation of the utter margin with air speed for the
¯2 C ¯1 wing bending-controlsurface torsion case is not quadratic. In other
C (26)
2 words, a number of spurious utter margins, or parameters indi-
cating the stability of a multi-DOF system, can be formed using
It is obvious that, if either ¯1 or ¯ 2 become zero, then the FM but these parameters will not be the utter margin, as de-
" ³ ´2 # ned by Zimmerman and Weissenburger.3 Only if the actual utter
¯2 ¡ ¯1 mechanism is used will the utter margin be obtained.
1¡ D0 (27)
¯2 C ¯1 The applicationof the FM to the Sim-2 model was not as straight-
forward as in the Hancock case. The Sim-2 model contains 23
modes, of which mode 12 becomes unstable at utter. The ut-

ter mechanism consists of mode 12 and mode 4. The utter margin
variation with speed for the actual utter mechanism is shown in
Fig. 5. Notice that the variation is not exactly quadratic because
such a variation can only be obtained for systems with no structural
damping,4 whereas the Sim-2 model has nonzero damping at zero
speed, which is equivalent to structural damping. Unfortunately,
mode 4 is highly, damped with respect to most of the other modes
and, hence, does not feature prominently in any of the FRFs, even
at speeds which are low compared to the utter speed. Because, for
this work, the eigenvalues were obtained using a frequency-domain
method, the eigenvalue of mode 4 could not always be obtained.
The FM results of Table 2 have gaps were the eigenvalues of the
modes were not obtained.As a consequence,an alternativeapproach
was used for the simulated utter tests described later. The FM was
applied to the combination of modes 12 and 3, the latter being a
prominent mode in the FRFs. Equation (3) was applied to these two
modes yielding a spurious utter margin that did not vary quadrat-
ically and had to be curve tted by higher-order polynomials. Nev-
ertheless, because the spurious utter margin became zero at utter,
it served as a stability parameter. The combination of modes 12
Fig. 4 Flutter margin variation for the three possible utter mecha- and 3 was chosen after all of the other possible combinations were
nisms, Hancock model. tested.
Fig. 5 True utter margin variation with air speed, Sim-2 model.
As with the damping t approach, the FM’s sensitivity to noise The advantageof applyingthe method to a large systemwith many
depends on that of the RFP. It is at these conditions that the au- measurement positions, such as the Sim-2 model, is that responses
thors of Ref. 6 suggest a linear t of the utter margin gives better from all of these positionscan be used to calculateshape parameters.
estimates for the utter velocity. However, Fig. 6 shows that this lat- Even if some of the responses overshoot the decay envelope utter
ter approach would provide worse results than the normal quadratic criterion because of measurement error, on average the criterion is
t on the Hancock model. satis ed. This statement can be veri ed by Fig. 8, which shows a
plot of shape parameters from each measuring position in the case
Envelope Function Method where the impulse responses were corrupted with 10% rms noise.
The main considerations in the successful application of the en- Hence, the envelope function utter prediction procedure can be
velope function method are the number and velocities of subcritical modi ed to include curve tting of the shape parameters of all of
tests, the order of the polynomial t of the decay envelope centroid, the measured impulse responses (17 for the Sim-2 model).
and the values of tmax , and noise. When testing a multi-DOF system, The presence of noise in the system responses causes a dete-
a number of responses can be measured, all of which will yield an rioration in the quality of the estimated impulse responses. This
envelopefunction and a shape parameter at each air speed, as seen in deterioration,in turn, causes signi cant scatter in the results for the
Fig. 7 for the Hancock model, where the shape parameter variation position of the decay envelope centroid. However, Fig. 8 demon-
with air speed for each DOF is plotted. Hence, the shape parame- strates that, for the Sim-2 model, the existenceof noise has a positive
ters from every measured response can be curve tted to yield an effect. Noise will appear in the steady-state responses thus displac-
estimate for the utter velocity. ing the time centroid to the right and making the transient response
Fig. 6 Flutter margin variation with velocity, 10% rms noise, with Fig. 8 Shape parameter variation with velocity, Sim-2 model, 10%
linear t, Hancock model. rms noise.
Fig. 7 Shape parameter centroid variation with velocity, Hancock model.

Table 3 Flutter speed estimates using Table 4 Flutter estimates by the NG method
envelope method, Hancock model using different numbers of test speeds
Best polynomial Best utter speed Number of

tmax order estimates % jErrorj air speeds Flutter speed estimate, m/s
23 – 43 Range, % 2 44:0761 § 0:0434
10 2 54.7619 24.2 3 44:0931 § 0:0600
8 2 49.7932 13.0 4 44:1506 § 0:0720
6 2 34.7932 21.1 5 44:1208 § 0:0656
23 – 57 Range, % 6 44:0532 § 0:0715
10 3 43.5895 1.1 7 44:1218 § 0:0564
8 3 43.1423 2.1 8 44:1468 § 0:0526
6 3 44.1552 0.2
23 – 70 Range, %
10 3 44.2338 0.4
of motion. However, during the course of the present research, the
8 3 44.3199 0.6 possibility that utter predictions might improve if more than two
6 4 45.1514 2.4 velocities are used was examined. The basic premise of this idea
23 – 84 Range, %
is that if there are measurement errors in the responses used in
10 5 42.9772 2.5 the identi cation process then the identi ed equations of motion
8 5 46.1875 4.8 will also contain errors, resulting in less accurate utter predictions.
6 4 45.9741 4.3 Hence, if tests are carried out at more than two velocities and the
23 – 98 Range, % aerodynamic and structural matrices obtained by a least-squares
10 6 44.0838 0.0 procedure, the utter estimates might improve.
8 4 44.0438 0.1 To test this idea, the Hancock model was used in the 10– 30 m/s
6 3 43.9444 0.3 velocity range (or 0.23V f – 068V f ). The rst test was carried out
45 – 66 Range, % with two velocities, one at 10 m/s and one at 30 m/s. The second
10 2 41.2831 6.3 test was carried out with three velocities, one at 10, one at 20, and
8 2 41.0387 6.9 one at 30 m/s. In all, seven tests were carried out with up to eight
6 2 42.6185 3.3 velocities, all within the same range. The responses in all of the
45 – 79 Range, % tests were contaminated with 10% rms white noise, chosen from a
10 3 41.0289 6.9 normal distribution with zero mean and unity variance. Each test
8 3 43.2193 1.9 case was repeated 100 times to provide a large population of utter
6 3 44.3230 0.6 speed estimates. The con dence bounds of the mean utter estimate
45 – 93 Range, % for each test case were calculated using the Student’s T test17 with
10 4 44.1043 0.1 a con dence level of 1%. The results are given in Table 4. It can
8 4 44.1887 0.3 be seen that the best result, both in terms of mean utter speed and
6 4 44.3050 0.5 con dence bounds, occurs in the case where only two speeds are
68 – 88 Range, % used. The conclusion drawn from these results is that there is no
10 2 44.4217 0.8 advantage in using more than two test speeds to identify a simple
8 2 44.2729 0.4 system such as the Hancock model with the NG method. A higher
6 2 44.6352 1.3 number of air speeds would only be required if the aerodynamic lag
terms had not been neglected in the derivation of the method.
As the noise level increases, the quality of identi cation by the
less signi cant. Hence, apart from shape parameters obtained from NG decreases, as can be seen in the left-hand side of Fig. 9. The
measurementstationsdominated by the rigid-bodymodes, the shape Hancock model contains no structural damping so that the damping
parameter variations in Fig. 8 are relatively at at subcritical speeds. ratios should be zero at zero velocity, but the quality of identi ca-
As utter is approached, the transient response takes longer to de- tion in the presence of 10% rms noise is so low that the predicted
cay and becomes more signi cant, resulting in a drop in the shape dampings at zero velocity are nonzero. Additionally, parts of the
parameters. Hence, utter is approached when the shape parameter identi ed frequency variation are inaccurate. The adverse effects of
trends are no longer at. Thus, subcritical variations in the shape experimental errors can be overcome by using more than one forc-
parameter can be ignored, making the detection of utter a simpler ing vector during identi cation. With four forcing vectors and with
process. Of course, at high noise levels, this advantage is lost be- 20% rms noise, for the same test case, the quality of identi cation
cause, the higher the noise amplitude, the more sudden the drop in was much improved, as seen in the right-hand side of Fig. 9
the shape parameter near utter. The NG method was not applied successfullyto the Sim-2 model.
As with the damping t method, the appropriate order of the The large number of modes in the Sim-2 model, coupled with the
polynomial t of the decay envelope centroid is far from obvious. fact that the equations of motion are stiff, that is, some of the system
The problem is that a single polynomial order can not be chosen eigenvalues are orders of magnitude larger than the low-frequency
to be adequate for a variety of test cases. Finally, the value of tmax eigenvalues, caused matrix T in Eq. (13) to be badly scaled and,
does not appear to in uence greatly the utter predictions. Table 3 hence, nearly singular. As a consequence,the resulting equations of
shows utter predictions obtained by the envelope function method motion were highly inaccurate.
for the Hancock model. For each velocity range, three values of tmax
are applied. None of these values appears to yield consistentlymore ARMA-Based Method
accurate predictions. Table 3 also shows the polynomial orders that The MA part of an ARMA model implies white noise excitation;
yield the best utter predictions for each test case. however, because all of the other methods were evaluated using
frequency sweep excitation signals, an autoregressive moving av-
NG erage with exogenous inputs (ARMAX) model was used instead
The NG modeling of the equations of motion of a system is very of ARMA to ensure a fair comparison of the methods. Exogenous
accurate for low-order models. The factors that affect the accuracy inputs refers to inputs other than white noise.
of the NG are the presence of noise and the velocities at which the The rst dif culty in the implementation of the method is the
subcritical tests take place. A further consideration is the number evaluation of the ARMAX coef cients. In the absence of noise,
of air speeds at which tests are carried out. The NG only needs two this evaluation can be accomplished quite successfully even using a
tests at two different velocities to provide the complete equations simple least-squaresprocedurewith J equal to the number of modes.
Fig. 9 Comparison of performance of NG with 1 and 4 forcing vectors.
Fig. 11 F stability criteria using ARMA method, 5% noise, third-

order model.
Fig. 10 F and G stability criteria using ARMA method, no noise,
Hancock model.
applicationof the method to the Hancock wing with control surface
model with 5% rms noise in the response. These included standard
Then, the response of the ARMAX model to a given signal is almost least squares, instrumental matrix with delayed observations, dou-
identical to that of the actual system. In Fig. 10, all of the stability ble least squares14 and the recursive ltering method.13 Figure 11
criteria are plotted for a range of velocities up to the utter velocity shows the estimated F stability criteria for a best-case application
for the Hancock noise-free case. Criterion F5¡ goes negative at an using double least squares and J equal to the number of modes.
air speed very close to the actual utter velocity. However, it can Despite the noise, F5¡ is correctly identi ed as the rst criterion
be seen that the drop in the value of F5¡ is very abrupt, giving to go negative; however, the curves are obviously very rugged and
the impression of hard utter, even though Fig. 1 shows that the would not admit a successful polynomial curve t. The utter ve-
Hancock model does not undergo hard utter. If a curve t of the locity was identi ed accurately, but only after tests were performed
criterion was attempted at low subcritical speeds, it would fail to at a high subcritical speed. For J D 5 (number of modes C 2), the
predict utter. References 12 and 18 suggest a linear curve t very resulting criteria are worse. The authors of Refs. 12 and 18 suggest a
close to the utter velocity; however, such a procedure would carry maximum likelihood approach for order determination and param-
considerable risks in the case of an experimental test, be it in the eter estimation; however, they do not mention any applications of
wind tunnel or in the air. that approach to the case were there are measurement errors in the
When the responses contain noise, the evaluation of an accurate responses in either of these references.
ARMAX representationof an aeroelastic system becomes very dif- The best results for the Hancock model were obtained when us-
cult. A number of parameter estimation schemes were used in the ing J equal to the number of modes with double least squares. The
response results were decimated,that is, the time step was increased Hancock model. This is because, due to the high number of modes,
because when the time step is very small the curve t is more sen- there are many more F § .l/ stability criteria. Figure 13 shows the
sitive to noise corruption.14 Nevertheless, the ARMAX results in F C .l/ criteria for a sample application[criteria G.1/ and G.¡1/ do
Table 1 are very disappointing, with only test cases very close to not go negative and the F ¡ .l/ criteria go negative after the F C .l/
utter yielding acceptable predictions. ones]. The critical criterion is F C .45/ because it assumes negative
For the Sim-2 model, the best ARMAX ts were obtained using values rst. Hence, the utter velocities in Table 2 were obtained by
normal least squares and J equal to the number of modes. It was polynomial curve ts of F C .45/.
foundthat small amounts of noise improved the ts. Figure 12 shows
a typical t of the response of transducernumber 7 (located at the tip
of the tailplane, midchord). The agreement between the ARMAX Sensitivity of the Methods to Noise
model and the Sim-2 results is very good. The performance of the An indication of the sensitivity of the methods to measurement
ARMAX method is generallybetter for the Sim-2 model than for the noise was obtained by the application of the techniques to data
Fig. 12 Sample ARMAX t of the Sim-2 model response to chirp input.
Fig. 13 F+ (l) stability criteria variation with air speed for Sim-2 model.
Table 5 Variation of error in utter estimates with increasing noise level, Hancock model
Damping t Envelope NG FM ARMAX based

Noise
% rms Error, % STD Error, % STD Error, % STD Error, % STD Error, % STD
5 3.1 1.1 ¡3.0 5.7 ¡0.1 0.2 2.4 0.7 7.4 19.9
10 ¡1.3 1.3 ¡9.6 8.1 ¡0.5 1.0 7.9 1.9 10.1 16.2
15 5.8 7.7 ¡15.8 14.2 ¡0.7 1.4 13.3 2.6 12.0 15.6
20 50.0 112.6 ¡26.8 13.8 1.8 4.9 29.9 8.1 20.8 19.9
from the Hancock model corrupted with simulated noise of increas- estimates. At this early stage in the testing procedure, all methods
ing amplitude. The noise was white noise with a square frequency indicated stability for the next test point.
spectrum, incorporating frequencies up to twice the highest natural 4) The ight speed kept being increased by the same increment,
frequency of the system. The noise amplitude was measured as a and new estimates for the stability parameters were obtained and
percentage of the rms amplitude of the clean response. Tests were added to the curve- tting procedure.With each new estimate, higher
carried out with all of the methods at noise levels of 5, 10, 15, ordersof polynomialcurve ts were possible.The utter margin was
and 20% at an air speed range between 20 and 80% of the utter only tted by second-orderpolynomials,but the other three stability
speed. Each test was repeated 30 times, and the mean utter predic- parameters were tted by polynomials of all possible orders, with
tions as well as their standard deviations (STDs) were calculated. the new estimate for the utter speed being calculated as the mean
The mean errors in the utter predictions and STDs are given in of all of the acceptable roots. After all of the new estimates were
Table 5. obtained, stability checks were performed for the next test speed.
The damping t approach appears to perform well up to a noise 5) The simulated ight test was ended at the test speed at which all
amplitude of 15% (even though, at that amplitude, the spread of methods predicted that the next test speed would be within 20% of
the results is quite high), but fails completely at 20% rms noise the latest utter speed estimates. For each of the methods, the nal
amplitude. The envelope function has a tendency to underestimate utter speed estimate was taken to be the rst estimate at which the
the utter speed with increasing noise level. The NG approach (with stability check failed.
four excitation vectors) is very robust, yielding very small errors For the NG, the procedure was slightly different. At each test
and STDs even at noise levels of 20%. The FM yields unacceptable point, only the current and initial responses were used to provide
predictions at noise amplitudes higher than 10%. The performance a utter estimate. Additionally, the NG was only applied to the
of the ARMAX-based method is characterizedby a very high spread Hancock model utter tests.
of predictions,as witnessedby the high valuesof STDs. If an error of All results were obtained for responses contaminated by 5% rms
up to 10% in the utter estimates is considered acceptable, then the white noise. The noise was simulated by taking the inverse Fourier
maximum allowable noise levels for each method are damping t, transform of a frequencysignal with constantamplitude and random
15%; envelope method, 10%; NG, 20%; FM, 10%; and ARMAX- phase. The simulated utter test procedure was repeated 30 times,
based, 5%. and the mean utter speed prediction and 5% con dence interval
Three observations must be made. First, the ight conditions at was calculated for each method.
which the tests are performed affect the sensitivity of the meth-
ods to noise. If test were performed very close to the actual utter
condition, then the methods would have been far less sensitive to Comparison Between Methods
noise. Second, the sensitivity of the methods to noise also depend
From the discussion already presented, it should be clear that all
on the particular implementation. For example, the way in which
of the utter predictionmethods investigatedhere can yield accurate
the damping ratios and envelope function values are curve tted and
predictions under certain circumstances. The crucial consideration
extrapolatedcontributesto noise sensitivity.Third, in real ight ut-
that divides the techniques is how wide is this set of circumstances
ter tests, the measurement noise is not necessarily white, and there
for each one of them. A further consideration is the ease of use
may also be systematic errors included in the responses. Hence, the
and the speed of the calculations involved. During a ight utter
conclusions drawn in this section are only indicative.
test, it is often very important to know as soon as possible whether
proceeding to the next test point is safe.
Simulated Flutter Test The methods can be separated into two categories:
Simulated utter tests were performed for both models using all 1) The rst category consists of methods that identify the equa-
of the utter prediction methods. The procedure for all approaches tions of motion. This category contains only the NG method, which
(apart from the NG technique) was as follows: requires the DOF or modal responses. The identi cation requires
1) The response of each model was rst obtained at a low sub- only two tests at two different velocities. The equations of motion
critical speed, equal to 22.7% of the actual utter speed. In the case can then be solved to yield the utter speed.
of the Sim-2 model, the response also depends on the ight altitude, 2) The second category consists of methods that curve t a stabil-
which was xed at 5000 ft. The responses were analyzed to provide ity criterion.This categorycontainsthe rest of the methods(damping
estimates for the damping ratios, utter margin, envelope function t, FM, envelope, and ARMA). They all calculate a parameter that
shape parameter, and ARMA-based stability criteria. characterizes the stability of a system at each test velocity. The pa-
2) The ight speed was increased by an increment equal to 7% of rameter variation with air speed is curve tted by a polynomial and
the actual utter speed,and estimates of the four stability parameters then extrapolated to the condition for instability, to yield the utter
mentioned in step 1 were obtained. velocity.
3) The ight speed was increased again by the same increment The results of Tables 1 and 2 show that, as expected, the predic-
of the actual utter speed and estimates of the four stability param- tions from all of the methods tend to improve when the tests are
eters were obtained. Each of the parameters was tted with respect performed near the actual utter condition. For example, the pre-
to ight speed by a second-degree polynomial, whose roots were dictions for the air speed range 23– 98% (of the utter speed) are
subsequently obtained. Any roots that were complex, lower to the never more than 0.5% out. However, this is only a general trend. In
current ight speed, or occurred at points of positive curvature were Table 1, for example, as the test speed range is widened, the damping
ignored. The remaining roots provided the rst estimates for the t prediction error falls from 14.4 to 5.1%, increases to 9.1%, and
utter speed, a maximum of one estimate for each stability param- then falls again to 0.1%. Because, with the inclusion of noise, all of
eter. Subsequently, checks were performed to ensure that the next the processes are stochastic, there are always going to be glitches
ight speed would not be within 20% of any of the utter speed in the general trends. Nevertheless, the patterns that emerge show
Table 6 Flutter predictions for the Hancock Table 8 Comparative computational costs
model, simulated utter test
Method Time, s FLOPs
Mean utter speed Mean
Method estimate, m/s error, % ARMA 103.9080 558,279,907
Damp 119.0050 580,848,971
Damping t 43:73 § 1:054 ¡0.8 FM 93.4040 420,815,792
Envelope 46:00 § 1:093 4.4 Envelope 58.6300 111,513,021
FM 45:34 § 0:251 2.9 NG 19.2780 16,367,706
ARMAX-based 24:50 § 2:176 ¡44.4
NG 44:13 § 0:142 0.1
damping t method does not suffer from any of these dif culties.
Any aeroelastic system, no matter how complex, will yield damping
Table 7 Flutter predictions for the SIM-2 data and will have zero damping at the utter speed.
model, simulated utter test A nal note should be made on the ARMA-based method. Its per-
formance in both the stand-alonetests and the simulated utter tests
Mean utter speed Mean was much poorer than the performance of any of the other methods.
Method estimate, kn error, %
The Jury stability criteria,which are employed to pinpointthe utter
Damping Fit 394:15 § 5:88 ¡1.0 velocity,have a very indifferent low subcritical variation,that is, at
Envelope 386:82 § 2:47 ¡2.8 and noisy, as seen in Fig. 13. Only near utter does the variation
FM 432:82 § 5:38 8.8 become signi cant. Hence, any polynomial curve t that uses low
ARMAX 164:08 § 9:87 ¡58.8 subcritical values of the stability criteria will yield large errors. In
essence, the present research into the ARMA-based method sheds
doubts on the suitability of the Jury stability criteria for the analysis
that, for the Hancock model, the highest errors are obtained with of ight utter test data.
the ARMAX-based method and the lowest with the NG method, In terms of computational effort, the category 1 method requires
whereas, for the Sim-2 model, the highest errors are again obtained more data, especially if multiple excitation vectors are used, but
from the ARMAX-based method whereas the envelope function fewer test speeds than category 2 methods. As Table 8 shows, the
method yields the lowest errors. category1 method is less expensivethan category2 methods, mainly
Simulated ight utter predictions from the methods are given because of the smaller number of tests. The computationaltime and
in Tables 6 (Hancock model) and 7 (Sim-2 model). Two striking number of oating point operations (FLOPs) data were obtained
features of the results in Tables 6 and 7 are the poor performances for a speed range of 10– 43 m/s for the Hancock model. Category 2
of the ARMAX-based method for both models and the FM on the method tests took place every 3 m/s in that range (a total of 12
Sim-2 model. The reason for the failure of the ARMAX method is tests). The category 1 method was applied with four excitation vec-
the inability to obtain accurate ARMAX coef cients in the presence tors. Of course, these results should only be seen as indicators of
of noise, as detailed in the section on preliminary validation of comparativecomputationalcosts, not as absoluteassessments of the
the ARMAX method. The FM overestimated the utter speed of computational ef ciency of each method.
the Sim-2 model by 8.8% because the fake utter margin variation
dropped very abruptly near utter, that is, did not supply adequately Conclusions
early warning of impending instability. Five different utter prediction methods were implemented, and
As far as the accuracy of prediction, the equation-of-motion- their performance was evaluated on two simulated systems. The
identi cation method is best, as indicated by the Hancock model criteria used in this evaluation were the quality of predictions and
results. The NG provides consistently high-qualitypredictions even method complexity. The NG was found consistently to provide the
when the subcritical tests are carried out at relatively low velocities most reliable predictionsfor the simple Hancock model; however, it
and under high levels of noise. The accuracy of the method derives failed completely for the larger Sim-2 model. The FM can yield ac-
from a number of responsesto different inputs at the same test speed ceptable results when applied to an aeroelastic system whose utter
being employed in the calculations. mechanism is known, but it requires ight tests to be carried out at
On the other hand, it was found that the category 1 method could relatively high subcritical air speeds in the presence of high levels
not be successfully applied to the Sim-2 model. Only category 2 of noise corruption. The damping t and envelope function method
methods could provide results for that model, the damping t ap- were foundto give the best results for the Sim-2 model. The ARMA-
proach being the most successful,followed by the envelopemethod. based method was found to suffer in the presence of measurement
However, category 2 methods require polynomial curve tting and, error, mainly because of the inaccuracy of the ARMA- tting proce-
except in the case of the FM, the order of the polynomial is un- dure. Overall, the results presented show that the simplest methods
known. One way of determining which order of polynomial is best appear to be the more robust while giving good accuracy. The more
is to obtain a preliminary value for the utter velocity. This can be complex methodsneed further developmentto become more univer-
provided either by using a mathematical model of the system un- sally effective. Nevertheless, it was found that there was no general
der considerationor from earlier wind-tunnel and ight utter tests. application of the techniques investigated here that would work for
Additionally,with the exception of the envelope method, category 2 every aeroelastic model. The method implementations had to be
methods are complementary, that is, they all require a knowledge speci cally adapted for each model.
of the eigenvalues of the system. Once these have been evaluated
for a particular test case, utter predictions can be obtained very References
quickly and inexpensively.Hence, the damping t, FM, and ARMA 1 Cooper, J. E., “Parameter Estimation Methods for FlightFlutter Testing,”
methods could conceivably be applied simultaneously at each test CP-566, AGARD, May 1995.
case. 2 Géradin, M., and Rixen, D., Mechanical Vibrations, Wiley, 1997.
3 Zimmerman, N. H., and Weissenburger, J. T., “Prediction of Flutter Onset
Note that the present research indicatesthat, for large systems, the
damping t method is the most effective. This method also happens Speed Based on Flight Testing at Subcritical Speeds,” Journal of Aircraft,
to be the most widely used approach in practical ight utter tests. Vol. 1, No. 4, 1964, pp. 190– 202.
4 Price, S. J., and Lee, B. H. K., “Evaluation and Extension of the Flutter-
In a purely academic sense, category 1 methods are preferable to
Margin Method for Flight Flutter Prediction,” Journal of Aircraft, Vol. 30,
the damping t technique because they provide a fuller picture of No. 3, 1993, pp. 395– 402.
the dynamics of a given system. Nevertheless, in more realistic 5 Routh, E. J., Advanced Part of a Treatise on the Dynamics of a System
applications, category 1 methods are confronted by a number of of Rigid Bodies, 5th ed., Vol. 2, MacMillan, 1930.
practical problems such as the choice of the number of modes, the 6 Dickinson, M., “CF-18 Flight Flutter Test (FFT) Techinques,” CP-566,
need for modal responses, and associated numerical problems. The AGARD, May 1995.
7 Cooper, J. E., Emmett, P. R., Wright, J. R., and Scho eld, M. J., “Enve- Aircraft, Vol. 18, No. 10, 1981, pp. 862– 868.
lope Function: A Tool for Analyzing Flutter Data,” Journal of Aircraft, Vol. 13 James, P. N., Souter, P., and Dixon, D. C., “A Comparison of Param-
30, No. 5, 1993, pp. 785– 790. eter Estimation Algorithms for Discrete Systems,” Chemical Engineering
8 Cooper, J. E., “Modal Parameter Identi cation Using Exponential Science, Vol. 29, 1974, pp. 539– 547.
Weighting,” Proceedings of the First International Conference on the In- 14 Cooper, J. E., “Comparison of Some Time-Domain-System Identi ca-
tegration of Dynamics, Condition Monitoring and Control, Manchester. A. tion Techniques Using Approximate Data Correlations,” Journal of Modal
A. Balkema Publishers, Rotterdam. 1999. Analysis, Vol. 4, No. 2, 1989, pp. 51 – 57.
9 Nissim, E., and Gilyard, G. B., “Method for Experimental Determination 15 Dimitriadis, G., and Cooper, J. E., “A Method for Identi cation of Non-
of Flutter Speed by Parameter Identi cation,” AIAA Paper 89-1324, 1989. Linear Multi-Degree-of-Freedom Systems,” Proceedings of the Institute of
10 Gaukroger, D. R., Skingle, C. W., and Heron, K. H., “An Application Mechanical Engineers, Pt. G, Vol. 212, 298, 1998, pp. 287– 298.
16
of System Identi cation to Flutter Testing,” Journal of Sound and Vibration, Bennett, R. M., “Application of Zimmerman Flutter-Margin Criterion
Vol. 72, No. 2, 1980, pp. 141– 150. to a Wind-Tunnel Model,” NASA, TM 84545, 1982.
11 17
Eversman, W., and Tewari, A., “Consistent Rational-Function Approx- Stroud, K. A., Engineering Mathematics, 2nd ed., MacMillan, New
imation for Unsteady Aerodynamics,” Journal of Aircraft, Vol. 28, No. 9, York, 1970.
1991, pp. 545– 552. 18 Matsuzaki, Y., and Torii, H., “Response Characteristics of a Two-
12 Matsuzaki, Y., and Ando, Y., “Estimation of Flutter Boundary from Dimensional Wing Subjected to Turbulence Near the Flutter Boundary,”
Random Responses Due to Turbulence at Subcritical Speeds,” Journal of Journal of Sound and Vibration, Vol. 136, No. 2, 1990, pp. 187– 199.

Flutter Prediction From Flight Flutter Test Data

Uploaded by

Copyright:

Available Formats

Flutter Prediction From Flight Flutter Test Data

Uploaded by

Document Information

Original Title

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

Flutter Prediction From Flight Flutter Test Data

Uploaded by

Copyright:

Available Formats

JOURNAL OF AIRCRAFT

Vol. 38, No. 2, March– April 2001