This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY 1

Optimal Tracking Control of Motion Systems

Anusha Mannava, S. N. Balakrishnan, Member, IEEE, Lie Tang, and Robert G. Landers, Senior Member, IEEE

Abstract—Tracking control of motion systems typically requires friction compensation was presented in [6]. The drawback to
accurate nonlinear friction models, especially at low speeds, and this methodology is that modeling nonlinear friction is difficult
integral action. However, building accurate nonlinear friction and the model changes substantially over time. Biaxial contour
models is time consuming, friction characteristics dramatically
change over time, and special care must be taken to avoid windup control subject to friction has been studied in [7]. The friction
in a controller employing integral action. In this paper a new is observed with a state observer. Based on the results of state
approach is proposed for the optimal tracking control of motion observation, friction is estimated and compensated. However,
systems with significant disturbances, parameter variations, and experimental results indicate that the observer suffers from a
unmodeled dynamics. The ‘desired’ control signal that will keep long settling time and high frequency oscillations.
the nominal system on the desired trajectory is calculated based
on the known system dynamics and is utilized in a performance Many studies have investigated cross-coupled controllers to
index to design an optimal controller. However, in the presence of regulate the contour error (i.e., the distance from the actual posi-
disturbances, parameter variations, and unmodeled dynamics, the tion to the contour the system is attempting to track). Although
desired control signal must be adjusted. This is accomplished by several control techniques that handle cross-coupling effects
using neural network based observers to identify these quantities, were proposed (e.g., [8]–[11]) and show good results, the effect
and update the control signal on-line. This formulation allows
for excellent motion tracking without the need for the addition of of introducing such controllers on the overall system dynamics
an integral state. The system stability is analyzed and Lyapunov was not examined [12]. Subsequently, adaptive robust control
based weight update rules are applied to the neural networks [13]–[15] was introduced for systems with uncertainties and can
to guarantee the boundedness of the tracking error, disturbance effectively handle bounded disturbances while preserving tran-
estimation error, and neural network weight errors. Experiments sient performance. Model-based position synchronization [16],
are conducted on the linear axes of a mini CNC machine for the
contour control of two orthogonal axes, and the results demon- [17] and sliding mode based robust controllers [18]–[20] are
strate the excellent performance of the proposed methodology. some of the other approaches taken towards minimizing con-
tour errors. Other contour error control methods include iter-
Index Terms—Lyapunov stability analysis, motion systems,
neural network observer, optimal control, uncertainty estimation. ative learning control [21], hierarchical control [22], and Ly-
punov-based control [23]. To further improve controller results,
other studies considered adaptive feedrate schemes based on
I. INTRODUCTION methods such as fuzzy logic [24], [25] and predictive control
[26]. Recently, contour control has been extended to five axis
machines [27].

I N many applications, precise motion control is critical to

ensure the quality of the application. For example, axis er-
rors in machining operations directly result in part geometric
In this paper an optimal tracking controller for uncertain sys-
tems is proposed and applied to contour tracking applications.
The desired control signal history (i.e., the amount of control
errors. When a motion system contains multiple axes, an im- required to keep the states of the nominal system on the de-
portant quality metric is contour error [1], which is defined as sired trajectories) is computed a priori using the known system
the shortest distance between the desired contour and the actual dynamics. A performance index is defined in terms of devia-
position. The desired contour error accuracy can be obtained tion from the desired states, as well as deviation from the de-
by reducing the individual axial errors, provided the design has sired control signal. An optimal controller is constructed that
high tracking bandwidth and robustness [2]. One such technique can provide tracking without the utilization of an integrator.
is zero phase tracking error control (ZPTEC) [3]. This method Therefore, saturation due to integral windup does not occur and
and its extensions (e.g., [4], [5]) demonstrate good tracking per- anti-windup schemes are not required. Controllers are designed
formance and zero phase error between the reference and output. for each motion axis separately and the contour error is mini-
A methodology using feedback, feedforward, and model-based mized by reducing the axial errors. Since motion systems are
susceptible to disturbances such as nonlinear friction, cogging
Manuscript received March 17, 2011; revised July 09, 2011; accepted torques, etc., and parameter variations and unmodeled dynamics
September 09, 2011. Manuscript received in final form September 10, 2011.
exist, neural network based observers are used to estimate these
Recommended by Associate Editor Y. Bestaoui. This work was supported in
part by NASA under Grant ARMD NRA NNH07ZEA001N-IRAC1 and by effects. This information is used to adjust the desired control
the National Science Foundation. The views of the authors do not necessarily signal on-line to compensate for disturbances, parameter varia-
represent the views of the NSF or the NASA.
tions, and unmodeled dynamics. A proof is provided that guar-
The authors are with the Department of Mechanical and Aerospace
Engineering, Missouri University of Science and Technology, Rolla, MO antees the boundedness of the tracking error, disturbance esti-
65401-0050 USA (e-mail: [email protected]; [email protected]; [email protected]; mation error, and neural network weights.
[email protected]).
The advantages of the proposed method are two-fold. First,
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. traditional control methodologies [28] for multi-axis motion
Digital Object Identifier 10.1109/TCST.2011.2168608 control usually rely on integrators to avoid the constant bias in

1063-6536/$26.00 © 2011 IEEE

 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

the tracking error, which may cause integral “wind up”, leading where , is an -dimensioned Lagrange multiplier and
to actuator saturation and loss of transient performance [29], is a positive definite symmetric matrix determined
[30]. Thus, anti-windup schemes have to be used. Although from the nonlinear Riccati equation
anti-windup schemes may not be difficult to apply, they may
deteriorate controller performance, which is critical for high (5)
precision motion control tasks. This deterioration (e.g., increase and
in response time) is due to the fact that the anti-windup scheme (6)
changes the closed-loop dynamics when saturation occurs. The
methodology proposed in this paper requires no integral oper- is defined as the tracking error. Note that the structure of the
ation and thus no anti-windup scheme, helping to improve the performance index in (2) results in a control law that does not
controller performance and reduce the controller complexity. require integral action to track a desired trajectory.
Second, by integrating the uncertainty estimation neural net-
work as a part of an observer formulation, the uncertainties B. Controller Design for Uncertainties
are learned very fast without high frequency oscillations [31], Since disturbances, parameter variations, and unmodeled
enabling excellent transient performance. Moreover, the un- dynamics affect the motion system, the true system model is
certainty estimation provides an estimate of the nonlinearities, written as
such as nonlinear friction and cogging, which are difficult to
model. This knowledge of system behavior can be used in the (7)
analysis of the motion system dynamics and in identifying
disturbance bounds. where is the state-dependent uncertainty. Since we
consider only matched uncertainties (that is uncertainties ap-
II. CONTROL DESIGN PROCEDURE pear only in equations containing the control signals), the coef-
ficient matrix, contains zeros and constants. Entries
The proposed optimal tracking controller design procedure is
with constants correspond to the equations containing the con-
now given.
trol components and, therefore, contain the uncertainties.
A. Controller Design for Nominal System Assumption 1: The uncertainty is bounded by

Consider the nominal system dynamics (8)

(1) This is to ensure the boundedness of the estimation and motion

tracking errors. Since the true dynamics are different from (1),
where is the state vector and is the control vector. the steady state control defined in (4) does not hold and should
The state coefficient matrix is and the control coefficient be obtained from
matrix is . The system is required to follow the desired
signal . In order to achieve this, the performance index (9)
to be minimized is defined as
However, since the uncertainty is unknown to the designer,
it must be estimated. A neural network-based observer is pro-
posed next to explicitly estimate . Neural Networks are ad-
vantageous since they allow the estimation of the uncertainty
(2) reliably, which in turn can be used in the control law.

where and are symmetric weighting matrices that C. Neural Network Based Estimation
penalize deviations between the states and the desired states and
The output of the neural network is given by
between the control and the desired control, respectively. The
term is the desired control signal, computed from the (10)
known system dynamics in (1) as
where is a matrix of estimated weights and is the
(3) number of neurons in the network and is the basis
function vector, which is defined as
The desired control signal is the amount of control required to
keep the nominal system on the desired trajectory. The desired
trajectory can be any non-zero, time-varying quantity from
(11)
which the corresponding time-varying desired control can be
computed. If , the problem reduces to a regulation where kron denotes the Kronecker product and
problem and the desired control for a stable system will be
zero. The total control signal, obtained by minimizing the cost (12)
function in (2) is given by
Trigonometric functions are chosen to construct the basis func-
(4) tion vector since these functions form an orthogonal basis and
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

MANNAVA et al.: OPTIMAL TRACKING CONTROL OF MOTION SYSTEMS 3

Fig. 1. Proposed controller block diagram.

have the capability to represent nonlinear functions accurately. in this paper and in other studies that utilize neural networks.
Also, they allow the constraint on the basis function vector mag- By using this observer structure, transient performance can be
nitude to be formulated as tuned via the observer gains.
Theorem 1: A stable weight update rule with full state mea-
(13) surement given by
where is the basis function norm upper bound. From the
(19)
universal function approximation property of neural networks
[32], it can be stated that there exists an ideal neural network stabilizes the estimation error , the tracking error , and the
with an optimum weight vector and basis function vector adaptive weights of the online network. Adaptation rates
that approximates to an accuracy of are used along with a modification factor of
to enforce the bound on the neural network weights and pro-
(14)
vide robustness. The proof of Theorem 1 is provided in the
Different architectures for neural networks exist in the literature Appendix.
[33]. In this paper, a linear-in-the-parameter network is chosen.
The reasons are three fold: to keep the architecture simple and D. Desired Control Signal Adjustment
mathematically tractable, and not computationally intensive. With the approximated uncertainty, the desired control can
Assumption 2: The ideal weight is bounded by now be computed from

(15) (20)
Assumption 3: The neural network approximation error is where . The system uncertainty can be taken
bounded by into account in the control design by redefining the desired con-
trol signal based on the uncertainty estimation. The total control
(16)
signal is the combination of the redefined desired control and
This error can be reduced by increasing the number of neurons feedback control signals, similar to (4). Essentially, the feed-
in the neural network design. The state estimator structure is back component of the control is used only to shape the tran-
sient dynamics. A block diagram representation of the proposed
(17) controller is given in Fig. 1. The estimated state vector and es-
timated weights are both initialized to zero.
The estimation error is then expressed as where the
estimation error dynamics are E. Design Parameter Selection
(18) The stability properties of the controller formulation are ana-
lyzed using the Lyapunov function
where is the error between the actual and
ideal neural network weights. Although the observer estimates (21)
system states are assumed to be measurable, the utility of the
modified observer structure is the estimation of the process un- where “tr” is the trace operator and and are adaptation
certainties. In fact, the modified observer structure is the major rates chosen as positive definite matrices. For ease of design,
difference between how process uncertainties are estimated these rates are chosen such that only the diagonal elements are
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

TABLE I
SELECTION CRITERIA FOR OBSERVER DESIGN PARAMETERS

non-zero. Through Lyapunov stability analysis (see Appendix),

the bounds on the errors in tracking, estimation, and neural net-
work weights can be deduced, respectively, as

(22)

(23)

(24)

where

(25)

(26)

The matrix is positive definite since adaption rates and

observer gain are chosen as positive definite. The matrix
is negative definite since the adaptation rate is chosen
as positive definite while the closed-loop matrix for the nom-
inal controller, , is negative definite. If the
error norms are greater than the bounds then the Lyapunov func-
tion becomes negative definite and the error decreases. Once
the error norm is within the bound, the errors continue to re-
main in that region. It is possible to design the observer and Fig. 2. Mini CNC machine used in experimental studies.
the controller such that the error bounds are small. Table I pro-
vides selection criteria for the observer parameters. The rate of
convergence of neural network weights depends on the adap- mini CNC machine shown in Fig. 2. The machine consists of
tive rate . High adaptation rate enables fast estimation of three orthogonal axes and a spindle. The control system block
strong nonlinearities and time-varying system parameters; how- diagram is built using MATLAB Simulink and downloaded to
ever, high adaptation rate is detrimental to the reduction of es- a xPC target computer with a 800 MHz Pentium III processor
timation error as shown in (22). The larger the adaptation rate, and 384 MB RAM via Ethernet. A schematic of the experi-
the greater the oscillations will be during estimation. In order to mental system setup is shown in Fig. 3. The computer running
obtain smoother estimation signals, the observer gain must Target xPC is equipped with a National Instrument’s PCI-6711
be large too, while satisfying the constraint on . The tracking analog output board and a National Instrument’s PCI-6602
error in the weight update rule given in (19) is useful only in counter/timer board. The PCI-6711 board, with a range of 10
demonstrating the stability of the system and the learning rate V and 12 bits of resolution, is used to output control signals
accompanying it must be small to avoid a large bound on the to the axis amplifiers, while the PCI-6602 counter/timer board
tracking error. The bound in (24) shows that for larger uncer- measures the encoder feedback. The axis motors are equipped
tainties the tracking performance is poorer, unless the nominal with incremental optical encoders providing a linear position
closed-loop linear system , which is a part of sensing resolution of 0.108 m in quadrature mode. All exper-
, has large eigenvalues. The weighting matrices and can iments are conducted with a sampling frequency of 1000 Hz.
be chosen accordingly.
IV. AXIS MODELING
III. EXPERIMENTAL SETUP Ignoring the electrical dynamics, which are faster than the
In order to test the effectiveness of the proposed optimal mechanical dynamics, each axis is modeled in the continuous
tracking controller, experimental studies are conducted on the domain as
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

MANNAVA et al.: OPTIMAL TRACKING CONTROL OF MOTION SYSTEMS 5

Fig. 3. Experimental setup schematic.

(27)
Fig. 4. -axis velocity model and experimental results comparison.
(28)
(29) TABLE II
LINEAR AXIS SYSTEM PARAMETERS
(30)

where is the position (mm), is the time constant is

the axis speed (mm/s), is the gain ((mm/s)/V), V is the com-
mand voltage (V), is the measurement (mm), and is the un-
known nonlinear friction (mm/s), which experimental studies
have found to be well modeled as Coulomb friction. The non- both increased to 2.5 (mm/s)/V and these values were used to
linear friction is more complicated than this model, and sub- design the desired controller. The desired control signal for each
stantially changes over time; however, it represents the major axis is
portion of the unknown disturbances and will be used to com-
pare to the results of the disturbance estimator. The system pa- (32)
rameters for each axis are given in Table II. To determine the
model parameters for the axes, a pyramid command voltage where provides the uncertainty estimate, which is
signal is sent to the servo motor and the corresponding axis mainly dominated by the nonlinear friction in Case I. However,
speed is recorded. The model parameters are then determined in Case II, the uncertainty is dominated by the nonlinear friction
using a Particle Swarm Optimization algorithm [34], [35]. To and the difference between the estimated and true gain. The
demonstrate the model performance, the simulated -axis ve- total control signal for each axis is
locity using the developed model is compared with the mea-
sured velocity for a pyramid command voltage, as shown in (33)
Fig. 4. The results demonstrate that -axis model fits the ex-
perimental data very well. The average error percentage is The feedback portion of the controllers for both axes are
designed using the linear quadratic regulator (LQR) with
% % (31) weighting matrices and ,
adaptation rates , and ,
The state-space model for each axis is given by (7), where and an observer gain . The axes track
the diamond profile given in Fig. 5 for two cycles with
, and mm/s and a sampling interval of 1 ms. A constant velocity
. The desired state vector is , interpolator is used to generate the desired position signals.
where is the desired position (mm) and is the desired ve- In order to avoid unlearning and relearning the uncertainty
locity (mm/s). each time the axes change direction, two neural networks are
used: one for each direction of motion. Note that the neural
V. EXPERIMENTAL RESULTS network weight update rule in Theorem 1 and the feedback law
The proposed method is tested on the mini CNC machine for were developed for the full state measurement. Since only the
diamond contours using the - and -axes for two cases. The position of each axis is measured (30), a pseudo-measurement
number of neurons used in both case studies is 9, which is de- of velocity is created by taking the forward difference of the
termined using trial and error by observing the control perfor- measurement .
mance. In Case I the model parameters in Table II were used to The system performance is now analyzed in terms of axial
design the desired controller. In Case II the model gains were and contour errors. The contour error for a linear contour is
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

Fig. 5. Diamond contour schematic.

Fig. 6. Case I axial and contour error signals.

(34)

In Case I, the LQR controller is designed based on the nom-

inal model parameters given in Table II. In this case

, and .
The axial and contour error signals are given in Fig. 6. The
plots on the left side of Fig. 6 are the complete error signals.
The maximum absolute contour error is 22.1 m. The spikes in
the error signals correspond to changes in the direction of an
axis and are due to the fact that the diamond contour is discon-
tinuous at these points and a constant velocity interpolator is
utilized. It can be observed that the - and -axes have max-
Fig. 7. Case I uncertainty and axis control signals.
imum tracking errors of 31.2 and 30.9 m, respectively, during
the transient phases. The transient error could be reduced if con-
stant acceleration or jerk interpolators are used. The plots on the the axis. This variation is most likely due to the position-depen-
right-hand side of Fig. 6 are the detailed steady-state error sig- dent nature of the nonlinear friction, which appears to be more
nals. The mean values of the absolute values of the -, -axis, pronounced in the -axis than in the y axis. The axis control
and contour errors, respectively, are 0.125, 0.109, and 0.110 m. signals are also given in Fig. 7. The control signals temporarily
These values are within two resolutions of the sensor signals, saturate when either axis changes direction; however, this does
demonstrating the proposed controller is able to provide excel- not affect the system stability or performance during steady op-
lent tracking performance even though an integral state is not eration. The experimental results also demonstrate that the xPC
utilized in the controller. target computer is capable of executing the entire control system
The axis uncertainty estimations, computed using the neural at a sample time of 1 ms without overloading the CPU, as a CPU
network based observers, are shown in Fig. 7. These estimations overload error occurs when the CPU time required by the con-
account for disturbances, parameter variations, and unmodeled trol application is greater than the sample time.
dynamics. The mean uncertainty estimations for the -axis in The uncertainty and axis tracking error transient responses
the positive and negative directions of motion are 0.701 and at the first corner are shown in Fig. 8. It can be observed that
0.314 mm/s, respectively, which are within 0.07 mm/s of the the tracking error and uncertainty estimation stabilize in ap-
nonlinear friction values given in Table II. This indicates the proximately 0.05 s and no high frequency oscillations are ob-
uncertainty in the -axis is mainly due to the nonlinear friction. served, indicating that the neural network-based estimator per-
The mean uncertainty estimations for the -axis in the positive forms very well during the transient phase.
and negative direction of motion are 0.470 and 1.05 mm/s, In order to test the robustness of the proposed control
respectively, which are within 0.25 mm/s of the nonlinear fric- methodology, in Case II the LQR controller is designed
tion values given in Table II. This indicates the uncertainty in based on the nominal time constants in Table II and gains
the -axis is due to the nonlinear friction as well as other effects of 2.5 (mm/s)/V, increases of 32% and 56%, re-
such as model parameter variations. The uncertainty estimation
in the -axis also varies slightly during the steady operation of spectively. In this case
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

MANNAVA et al.: OPTIMAL TRACKING CONTROL OF MOTION SYSTEMS 7

Fig. 10. Case II uncertainty and axis control signals.

Fig. 8. Case I uncertainty and axis tracking error transient responses at first
corner.

Fig. 11. Case II uncertainty and axis tracking error transient response at first
corner.

Fig. 9. Case II axial and contour error signals. now is mainly due to the nonlinear friction and the change is
model gain, which is given by

(35)

where is the variation in the model gain. The mean un-

and . The axial and contour error signals certainty estimations for the -axis in the positive and nega-
are given in Fig. 9. The plots on the left side of Fig. 9 are the tive directions of motion are 2.19 and 1.69 mm/s, respec-
complete error signals. The maximum absolute values of the -, tively, compared to values of 2.17 and 1.74 mm/s computing
-axis, and contour errors, respectively, are 31.4, 33.5, and 23.8 using (35). The mean uncertainty estimations for the -axis in
m. These errors are within 8% of the corresponding errors the positive and negative directions of motion are 2.34 and
in Case I. Again the spikes in the error signals correspond to 3.16 mm/s, respectively, compared to values of 2.53 and 3.41
a change in the direction of an axis and are due to the fact mm/s computing using (35). These results indicate the neural
that the diamond contour is discontinuous at these points. The network based observers are estimating the uncertainty in the
plots on the right side of Fig. 9 are the zoomed-in error signals. - and -axes very well. The axis control signals are also given
The mean values of the absolute values of the -, -axis, and in Fig. 10. Again, the control signals temporarily saturate when
contour errors, respectively, are 0.145, 0.168, and 0.126 m. either axis changes direction; however, this does not affect the
Similar to Case I, these values are within two resolutions of the system stability or performance during steady operation.
sensor signals, again indicating excellent tracking during steady The uncertainty and axis tracking error transient responses at
operation. Comparing the axial and contour errors generated in the first corner for Case II are shown in Fig. 11. In comparison
both cases, it can be seen that the controller performance is not with the transient responses shown in Fig. 8, the settling time for
affected by the model uncertainties considered in this study. both tracking error and uncertainty estimation remain approxi-
The axis uncertainty estimations, computed using the neural mately 0.05 s and no high frequency oscillations are observed.
network based observers, are shown in Fig. 10. The uncertainty However, it is noted that the uncertainty estimations in Case II
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

Fig. 12. Case II position and velocity estimation.

have larger magnitudes during the steady phase due to the incor- APPENDIX
rect model gain used, indicating the neural network based esti- BOUNDS ON TRACKING, DISTURBANCE ESTIMATION, AND
mator performs very well during the transient phase even when NEURAL NETWORK WEIGHT ERRORS
the model parameters are much different than the nominal ones. This section shows the derivation that leads to Theorem 1
The measured axis position, estimated axis position and esti- presented in the section neural network-based estimation. The
mated axis velocities are plotted in Fig. 12. The average abso- idea here is to first formulate a Lyapunov function in terms of
lute velocity estimation errors for - and -axis are the errors in the approximate system states, tracking errors and
(mm/s) and (mm/s), respectively. It can be seen that the error between the ideal weight and the actual weight of the
the observer works very well in estimating the axis position and neural networks. Second, a weight update law will be selected
velocity since no high frequency oscillations are observed and to make the rate of change of the Lyapunov function less than or
the estimation errors are small. equal to zero. The bound on the inequality will lead to bounds
on the errors. Let

VI. SUMMARY AND CONCLUSION (36)

This study presented a new optimal tracking controller for be a candidate Lyapunov function. Taking the derivative of (36)
motion systems containing significant disturbances, model pa- with respect to time
rameter variations, and unmodeled dynamics. The performance
index was modified with the introduction of the desired con-
trol signal and a neural network observer was used to modify
the control signal on-line to account for disturbances, model pa- (37)
rameter variations, and unmodeled dynamics. Lyapunov-based
weight update rules were used to guarantee the boundedness of The tracking error dynamics are given by
the tracking error, disturbance estimation error, and neural net-
work weight errors. The optimal tracking controller was imple- (38)
mented on a mini CNC and the results demonstrated excellent
tracking results, without the use of integral control, even when By combining (37) and (38), the Lyapunov function derivative
there was significant variation in the model gains. The exper- can be expressed as
imental results showed the neural network estimator was able
to accurately estimate the disturbances acting on the motion
system and the model parameter variations that were present.
The experimental results also demonstrate that the neural net-
work based estimator performs very well during the transient
phase and no high frequency oscillations are observed even
when the model parameters are much different from the nominal
ones. The neural network based estimator enhances the under-
standing of the system dynamics and could be used to analyze
the nonlinearities further. (39)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

MANNAVA et al.: OPTIMAL TRACKING CONTROL OF MOTION SYSTEMS 9

Computing and using (25) and (26), respectively, ap- where . After completing the square
plying the trace property , and adding and sub- on the term with
tracting the term , (39) can be rewritten as

(40)

In order to drive the error in the approximate system, , the (46)

tracking error, , and the error, between the ideal neural net-
work weights for perfect approximation of the model uncertain- Since
ties and the actual neural network weights to be bounded by
parameters that a designer can choose, the weight update rule
given in (19) is chosen. Furthermore, is a negative definite
matrix. Substituting the weight update rule in (19) into (40)
(47)

Equation (46) becomes

(41)

Since , (41) can be rewritten as

(48)

The negative definiteness of the Lyapunov function is proved if

and either of the two following conditions is satisfied:

a)

(42)

Equation (42) can be reduced to

b)
(43)

Applying norms to every term on the right-hand side of (43)

Applying these conditions, the bounds on the tracking, estima-

tion, and neural network weight errors, respectively, are given
by (22)–(24). Thus, the ultimate upper bounds of these errors
have been derived. The selection of the design parameters to
(44) reduce these bounds has been discussed previously. The system
performance is given by the contour error in (34). Since the con-
where and denote maximum and minimum tour error is a direct function of servomechanism tracking errors,
eigenvalues, respectively. By applying Assumptions 1–3, (44) the contour error is bounded by
can be expressed as
(49)

Since the system performance, which is given by the contour

error, is described by the performance of subsystems that do not
interact, the stability of the system is guaranteed by the stability
(45) of the subsystems.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

REFERENCES [25] K. Su and M. Cheng, “Contouring accuracy improvement using cross-

[1] A. Poo, J. G. Bollinger, and W. Younkin, “Dynamic error in type con- coupled control and position error compensator,” Int. J. Mach. Tools
touring systems,” IEEE Trans. Ind. Appl., vol. IA-8, no. 4, pp. 477–484, Manuf., vol. 48, no. 12–13, pp. 1444–1453, 2008.
Jul. 1972. [26] L. Tang and R. G. Landers, “Predictive contour control with adap-
[2] K. Erkorkmaz and Y. Altintas, “High speed CNC system design part tive feedrate,” IEEE/ASME Trans. Mechatron., pp. 1–11, 2011, DOI:
III: High speed tracking and contouring control of feed drives,” Int. J. 10.1109/TMECH.2011.2119324.
Mach. Tools Manuf., vol. 41, no. 11, pp. 1637–1658, 2001. [27] Y. Altintas and B. Sencer, “High speed contouring control strategy for
[3] M. Tomizuka, “Zero phase error tracking algorithm for digital control,” five-axis machine tools,” CIRP Annals-Manuf. Technol., vol. 59, no. 1,
ASME J. Dyn. Syst., Meas., Control, vol. 109, pp. 65–68, 1987. pp. 417–420, 2010.
[4] M. Tomizuka, M. S. Chen, S. Renn, and T.-C. Tsao, “Tool positioning [28] S. Chen and K. Wu, “Contouring control of smooth paths for multiaxis
for noncircular cutting with lathe,” ASME J. Dyn. Syst., Meas., Control, motion systems based on equivalent errors,” IEEE Trans. Control Syst.
vol. 109, no. 2, pp. 176–179, 1987. Technol., vol. 15, no. 6, pp. 1151–1158, Nov. 2007.
[5] T.-C. Tsao and M. Tomizuka, “Adaptive zero phase error tracking al- [29] N. J. Krikelis and S. K. Barkas, “Design of tracking systems subject to
gorithm for digital control,” ASME J. Dyn. Syst., Meas., Control, vol. actuator saturation and integrator windup,” Int. J. Control, vol. 39, no.
109, no. 4, pp. 349–354, 1987. 4, pp. 667–682, 1987.
[6] H. Zhang and R. G. Landers, “Precision motion control methodology [30] N. Kapoor, A. R. Teel, and P. Daoutidis, “An anti-windup design for
for complex contours,” ASME J. Manuf. Sci. Eng., vol. 129, no. 6, pp. linear systems with input saturation,” Automatica, vol. 34, no. 5, pp.
1060–1068, 2007. 559–574, 1998.
[7] C. Chen and K. Lin, “Observer-based contouring controller design of [31] K. Rajagopal, A. Mannava, S. N. Balakrishnan, K. Kalmanje, and N.
a biaxial stage system subject to friction,” IEEE Trans. Control Syst. Nguyen, “Neuroadaptive model following controller design for non-
Technol., vol. 16, no. 2, pp. 322–329, Mar. 2008. affine non-square aircraft system,” presented at the AIAA Guid., Nav.,
[8] Y. Koren, “Cross-coupled biaxial computer controls for manufacturing Control Conf., Chicago, IL, 2009.
systems,” ASME J. Dyn. Syst., Meas., Control, vol. 102, no. 4, pp. [32] F. M. Ham and I. Kostanic, Principles of Neurocomputing for Science
265–272, 1980. and Engineering. New York: McGraw Hill, 2001.
[9] Y. Koren and C. C. Lo, “Variable-gain cross-coupling controller for [33] S. Haykin, Neural Networks: A Comprehensive Foundation. Engle-
contouring,” CIRP Annals, vol. 40/1, pp. 371–374, 1991. wood Cliffs, NJ: Prentice-Hall, 1999.
[10] P. K. Kulkarni and K. Srinivasan, “Cross coupled compensators for [34] J. Kennedy and R. C. Eberhart, “Particle swarm optimization,” in Proc.
contouring control of multi-axial machine tools,” in Proc. North Amer. IEEE Int. Conf. Neural Netw., 1995, pp. 1942–1948.
Manuf. Res. Conf., 1985, pp. 558–566. [35] L. Tang, J. Ruan, R. G. Landers, and F. W. Liou, “Variable powder flow
[11] P. K. Kulkarni and K. Srinivasan, “Optimal contouring control of mul- rate control in laser metal deposition processes,” ASME J. Manuf. Sci.
tiaxial feed drive servomechanism,” ASME J. Eng. for Ind., vol. 111, Eng., vol. 130, no. 4, p. 041016, 2008.
no. 2, pp. 140–148, 1989.
[12] T. C. Chiu and B. Yao, “Adaptive robust contour tracking of machine
tool feed drive systems—A task coordinate frame approach,” presented
at the Amer. Control Conf., Albuquerque, NM, 1997. Anusha Mannava received the B.S. degree from Jawaharlal Nehru Techno-
[13] B. Yao and M. Tomizuka, “Adaptive robust control of MIMO nonlinear logical University, Hyderabad, India, in 2008 and the M.S. degree in electrical
systems with guaranteed transient performance,” presented at the IEEE engineering from the Missouri University of Science and Technology, Rolla, in
Conf. Decision Control, New Orleans, LA, 1995, “,” in . 2010.
[14] B. Yao and M. Tomizuka, “Adaptive robust control of a class of multi- Her research interests include optimal control and neural networks.
variable nonlinear systems,” presented at the IFAC World Congr., San
Francisco, CA, 1996.
[15] B. Yao, “Adaptive robust control of nonlinear systems with applica-
tion to control of mechanical systems,” Ph.D. dissertation, Dept. Mech.
Eng., Univ. California Berkeley, Berkeley, 1996. S. N. Balakrishnan (M’10) received the Ph.D. de-
[16] L. F. Yang and W. H. Wang, “Synchronization of twin-gyro preces- gree in aerospace engineering from the University of
sion under cross-coupled adaptive feedforward control,” AIAA J. Guid., Texas, Austin.
Control, Dyn., vol. 19, no. 3, pp. 534–539, 1996. He is currently a Curators’ Professor of Aerospace
[17] D. Sun, “Position synchronization of multiple motion axes with adap- Engineering with the Department of Mechanical and
tive coupling control,” Automatica, vol. 39, no. 6, pp. 997–1005, 2003. Aerospace Engineering, Missouri University of Sci-
[18] Z. Jamaludin, H. V. Brussel, and J. Swevers, “Classical cascade and ence and Technology, Rolla. His research interests in-
sliding mode control tracking performances for a XY feed table of a clude neural networks, optimal control, and control
high-speed machine tool,” Int. J. Precision Technol., vol. 1, no. 1, pp. of large-scale and impulse systems. His papers using
65–74, 2007. the development of techniques in these areas include
[19] Y. Altintas, K. Erkorkmaz, and W.-H. Zhu, “Sliding mode controller applications to missiles, spacecraft, aircraft, robotics,
design for high speed feed drives,” Annals CIRP, vol. 49/1, pp. temperature, animal population control, and manufacturing.
265–270, 2000.
[20] F.-J. Lin, P.-H. Shieh, and P.-H. Shen, “Robust recurrent-neural-net-
work sliding-mode control for the X-Y table of a CNC machine,” IEE
Proc. Control Theory Appl., vol. 153, no. 1, pp. 111–123, 2006. Lie Tang received the B.E. and M.E. degrees
[21] K. Barton and A. Alleyne, “A cross-coupled iterative learning con- in electrical engineering from Hohai University,
trol design for precision motion control,” IEEE Trans. Control Syst. Nanjing, China, in 2001 and 2005, respectively, and
Technol., vol. 16, no. 6, pp. 1218–1231, Nov. 2008. the Ph.D. degree in mechanical engineering from
[22] H. Moghadam, R. G. Landers, and S. N. Balakrishnan, “Hierarchical Missouri University of Science and Technology,
position-contour control of linear axes,” presented at the Int. Symp. Rolla, in 2009.
Flexible Autom., Tokyo, Japan, 2010. He is currently a Post Doctoral Researcher with
[23] N. Uchiyama, S. Takagi, S. Sano, and K. Yamazaki, “Robust con- the Department of Mechanical and Aerospace En-
touring control for multi-axis feed drive systems,” in Proc. IEEE Conf. gineering, Missouri University of Science and Tech-
Emerg. Technol. Factory Autom., 2006, pp. 840–845. nology (formerly University of Missouri—Rolla).
[24] M. Cheng, K. Su, and S. Wang, “Contour error reduction for free-form His research interests include modeling, simulation,
contour following tasks or biaxial motion control systems,” Robot. and control of mechatronics systems, manufacturing processes, and hydrogen
Comput.-Integr. Manuf., vol. 25, no. 2, pp. 323–333, 2009. fuel cells.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

MANNAVA et al.: OPTIMAL TRACKING CONTROL OF MOTION SYSTEMS 11

Robert G. Landers (SM’11) received the Ph.D. de- processes and alternative energy systems and has over 100 technical refereed
gree in mechanical engineering from the University publications.
of Michigan, Ann Arbor, in 1997. Dr. Landers was a recipient of the Society of Manufacturing Engineers’ M.
He is currently an Associate Professor of me- Eugene Merchant Outstanding Young Manufacturing Engineer Award in 2004.
chanical engineering and the Associate Chair for He is a member of ASEE and ASME and a senior member of SME. He is cur-
Graduate Affairs in the Department of Mechan- rently an Associate Editor for the IEEE TRANSACTIONS ON CONTROL SYSTEM
ical and Aerospace Engineering at the Missouri TECHNOLOGY, the ASME Journal of Dynamic Systems, Measurement, and Con-
University of Science and Technology (formerly trol, and the ASME Journal of Manufacturing Science and Engineering.
University of Missouri Rolla), Rolla. His research
and teaching interests include the areas of modeling,
analysis, monitoring, and control of manufacturing