Академический Документы
Профессиональный Документы
Культура Документы
3, JUNE 2003
489
I. INTRODUCTION
OMPUTED torque or inverse dynamics technique is a special application of feedback linearization of nonlinear systems. A number of works related to computed torque control of
robotic manipulators have been published [1]. The computed
torque controller is utilized to linearize the nonlinear equation
of robot motion by cancellation of some, or all, nonlinear terms
[1]. However, the objection to the real-time use of such control
scheme is the lack of knowledge of uncertainties, which include
parametric variations and external disturbances of the system.
The adaptive control technique is essential for providing a stable
and robust performance for a wide range of applications (e.g.,
robot control, process control, etc.), and most of the applications
are inherently nonlinear with uncertainties [2][4]. Therefore,
several computed torque controllers have tried to circumvent the
problem of uncertainties using adaptive techniques [5][7].
Recently, much research has been done on applications of
wavelet neural networks, which combine the capability of
artificial neural networks in learning from processes [8][11]
and the capability of wavelet decomposition [12][15] for identification and control of dynamic systems [16][20]. In Zhang
Manuscript received May 28, 2000. This work was supported by the National
Science Council of Taiwan, R.O.C., under Grant NSC 90-2213-E-155-014. This
paper was recommended by Associate Editor P. E. Borne.
The author is with the Department of Electrical Engineering, Yuan Ze University, Chung-Li 32026, Taiwan, R.O.C. (e-mail: rjwai@saturn.yzu.edu.tw).
Digital Object Identifier 10.1109/TSMCB.2003.811125
490
IEEE TRANSACTIONS ON SYSTEMS MAN AND CYBERNETICSPART B: CYBERNETICS, VOL. 33, NO. 3, JUNE 2003
where
(3)
,
and
denote uncertainties, and
where
called the lumped uncertainty and is defined as
is
(4)
491
(8)
, and
and
are
where the control gain
positive constants. Substituting (8) into (3) and using (7), it can
be obtained that
(9)
(5)
A linear scale detects the position signal of the slider
the robust control law is defined as follows:
. Now,
(6)
is a computed torque controller and
is a comwhere
pensated controller. The computed torque controller with the
WNN uncertainty observer is the main tracking controller, and
the compensated controller is designed to compensate the minimum approximation error of the WNN uncertainty observer.
(10)
492
IEEE TRANSACTIONS ON SYSTEMS MAN AND CYBERNETICSPART B: CYBERNETICS, VOL. 33, NO. 3, JUNE 2003
WNN are
and its derivative; the output of the WNN is the
estimated lumped uncertainty . The signal propagation and
the basic function in each layer are introduced in the following
paragraph.
For every node in the input layer, the net input and the net
output are represented as
C. Compensated Controller
To develop the compensated controller, first, the minimum
approximation error is defined as follows:
(16)
where
(11)
and
. Moreover, a family of wavelets
where
is constructed by translations and dilations performed on a
single fixed function called the mother wavelet. In the mother
that is derived
wavelet layer each node performs a wavelet
from its mother wavelet. The first derivative of a Gaussian
is adopted as a mother
function
wavelet in this study. It may be regarded as a differentiable
version of the Haar mother wavelet, just as the sigmoid is a
differentiable version of a step function, and it has the universal
approximation property [19]. For the th node
is an optimal weight vector achieves the minimum approximais the absolute value. In this study, it
tion error, in which
is assumed that the amplitude of is less than a positive conif the system existent uncertainties are
stant (i.e.,
bounded. Although the existent uncertainty in practical application is difficult to measure, a more conservative value of bound
is selected according to the powerful approximated ability of
the WNN uncertainty observer. From (3), (6), and (10), an error
equation is obtained as follows:
(12)
and
are, respectively, the translation and dilation
where
vectors in the th term of the th input to the node of layer 2,
and is the total number of the wavelets with respect to the input
nodes. In addition, each node in the wavelet layer is denoted
by , which multiplies the input signals and outputs the result
of the product, i.e., the product of monodimensional wavelets
with respect to the input node [19]. For the th-rule node
(17)
is a stable matrix, and
where
Then, a Lyapunov function is defined as
(18)
(13)
represents the th input to the node of layer 3;
,
where
which are the weights between the mother wavelet layer and the
is the number
wavelet layer, are assumed to be unity;
of wavelets with complete connection if each input node has
the same wavelet nodes. Furthermore, the single node in the
output layer is labeled as , which computes the overall output
as the summation of all input signals
(14)
is the output action strength
where the connecting weight
represents
of the th output associated with the th wavelet;
. On the other
the th input to the node of layer 4, and
can be rewritten as follows [23]:
representation,
(15)
, in which
is initialwhere
ized to be zero and adjusted during on-line operation;
, in which
is determined by the selected
.
wavelets and
(20)
To satisfy
controller
(21)
sgn
(22)
493
(23)
,
is negative semidefinite [i.e.,
Since
], which implies that and
are bounded. Let
, and integrate function
function
with respect to time
(24)
is bounded, and
Because
bounded, then
is nonincreasing and
(25)
Differentiate
Since all the variables on the right side of (17) are bounded, it
is uniformly continuous
implies is also bounded. Then
[3]. By using Barbalats lemma [2], [3], it can be shown that
. Therefore,
as
. As a result,
the robust control system is asymptotically stable. Moreover, the
tracking error of the system will converge to zero according
. On the other hand, the guaranteed convergence
to
of tracking error to be zero does not imply convergence of the
estimated value of the lumped uncertainty to its real value. The
persistent excitation condition [3] should be satisfied for the estimated value to converge to its theoretic value.
D. Convergence Analyzes of WNN
According to the assumption of the bound of the minimum
approximation error, the stability proof of the robust control
system can be guaranteed; however, the convergence condition
of the WNN must be satisfied. If the parameters of the WNN
are bounded, the convergence property of the WNN can be
guaranteed. From (15), the output of the WNN is bounded if
the weights between the output layer and the wavelet layer are
for as
bounded. Define the constrain set
) or (
(27)
is a two-norm of vector. According to the projecwhere
tion algorithm [23], the adaptation law (21) can be modified
as follows:
if
or
and
if
and
(28)
(31)
494
IEEE TRANSACTIONS ON SYSTEMS MAN AND CYBERNETICSPART B: CYBERNETICS, VOL. 33, NO. 3, JUNE 2003
Fig. 3. Simulated responses of robust control system. (a)(c) Case 1. (d)(f) Beginning at Case 1 and changing to Case 2 at 5 s. (g)(i) Beginning at Case 1 and
changing to Case 3 at 5 s.
,
(32)
By use of Barbalats lemma [2], [3], as shown in Section III-C,
as
. Thus, the stability property also can
be guaranteed. The effectiveness of the proposed robust control system for the control of motor-toggle servomechanism will
495
Fig. 3. (Continued). Simulated responses of robust control system. (a)(c) Case 1, (d)(f) Beginning at Case 1 and changing to Case 2 at 5 s; (g)(i) Beginning
at Case 1 and changing to Case 3 at 5 s.
kg
kg
kg
m
m
kg
m
m
m
(33)
in which
is the lead of the screw,
is the mass of slider ,
and and are the friction coefficient and gravity acceleration,
respectively. In addition, the gains of the robust control system
are given as
(34)
in this study is selected as an
For simplicity, the matrix
identity matrix. All the gains in the robust control system are
chosen to achieve the best transient control performance in
both simulation and experimentation considering the limitation
of the control effort and the requirement of stability. Moreover,
the initialization of the network parameters described in [19]
is adopted to initialize the parameters of the wavelets in this
study. The WNN has two, 14, seven, and one neurons at the
input, mother wavelet, wavelet, and output layer, respectively.
Three simulation cases due to a periodic sinusoidal command
are addressed as follows:
496
Fig. 4.
IEEE TRANSACTIONS ON SYSTEMS MAN AND CYBERNETICSPART B: CYBERNETICS, VOL. 33, NO. 3, JUNE 2003
Experimental results of robust control system: (a)(c) Nominal case; (d)(f) Parameter variation case.
trol characteristics of the robust control system under the occurrence of parameter variation can be clearly observed.
V. CONCLUSIONS
This study successfully demonstrates the application of a robust control system to control the position of a slider of the
motor-toggle servomechanism. First, the mathematical model
of the motor-toggle servomechanism was introduced. Then, the
theoretical bases of the proposed computed torque controller,
WNN uncertainty observer, and compensated controller were
described in detail. Moreover, simulation and experimentation
were carried out using a periodic sinusoidal reference trajectory
to test the effectiveness of the proposed robust control system.
The major contributions of this study are i) the successful development of a robust control methodology, in which an WNN
is utilized to adapt the lumped uncertainty on line and a compensated controller is designed to compensate the minimum approximation error of the WNN uncertainty observer, and ii) the
successful application of the robust control system to control the
slider position of the motor-mechanism coupling system considering the existence of uncertainties.
ACKNOWLEDGMENT
The author would like to express his gratitude to the referees
and Associate Editor for their kind comments and suggestions.
REFERENCES
[1] R. Ortega and M. W. Spong, Adaptive motion control of rigid robots:
A tutorial, in Proc. IEEE Conf. Decision Contr., 1988, pp. 15751584.
[2] K. J. Astrom and B. Wittenmark, Adaptive Control. New York: Addison-Wesley, 1995.
[3] J.-J. E. Slotine and W. Li, Applied Nonlinear Control. Englewood
Cliffs, NJ: Prentice-Hall, 1991.
[4] R. Johansson, Adaptive control of robot manipulator motion, IEEE
Trans. Robot. Automat., vol. 6, pp. 483490, Aug. 1990.
[5] C. Y. Su and T. P. Leung, A sliding mode controller with bound estimation for robot manipulators, IEEE Trans. Robot. Automat., vol. 9, pp.
208214, Apr. 1993.
[6] J. Imura, T. Sugie, and T. Yoshikawa, Adaptive robust control of robot
manipulators-theory and experiment, IEEE Trans. Robot. Automat.,
vol. 10, pp. 705710, Oct. 1994.
[7] M. Teshnehlab and K. Watanabe, Self tuning of computed torque gains
by using neural networks with flexible structures, Proc. Inst. Elect.
Eng.Control Theory Appl., vol. 141, no. 4, pp. 235242, 1994.
[8] M. H. R. Fazlur Rahman, R. Devanathan, and Z. Kuanyi, Neural network approach for linearizing control of nonlinear process plants, IEEE
Trans. Ind. Electron., vol. 47, pp. 470477, Apr. 2000.
[9] J. R. Noriega and H. Wang, A direct adaptive neural-network control
for unknown nonlinear systems and its application, IEEE Trans. Neural
Networks, vol. 9, pp. 2734, Jan. 1998.
497
[10] M. Zhihong, H. R. Wu, and M. Palaniswami, An adaptive tracking controller using neural networks for a class of nonlinear systems, IEEE
Trans. Neural Networks, vol. 9, pp. 947955, Sept. 1998.
[11] O. Omidvar and D. L. Elliott, Neural Systems for Control. New York:
Academic, 1997.
[12] Y. C. Pati and P. S. Krishnaprasad, Analysis and synthesis of feedforward neural networks using discrete affine wavelet transformations,
IEEE Trans. Neural Networks, vol. 4, pp. 7385, Jan. 1993.
[13] B. Delyon, A. Juditsky, and A. Benveniste, Accuracy analysis for
wavelet approximations, IEEE Trans. Neural Networks, vol. 6, pp.
332348, Apr. 1995.
[14] C. F. Chen and C. H. Hsiao, Wavelet approach to optimising dynamic
systems, Proc. Inst. Elect. Eng.Contr. Theory Appl., vol. 146, no. 2,
pp. 213219, 1999.
[15] T. Lindblad and J. M. Kinser, Inherent features of wavelets and pulse
coupled networks, IEEE Trans. Neural Networks, vol. 10, pp. 607614,
May 1999.
[16] Q. Zhang and A. Benveniste, Wavelet networks, IEEE Trans. Neural
Networks, vol. 3, pp. 889898, Nov. 1992.
[17] J. Zhang, G. G. Walter, Y. Miao, and W. N. W. Lee, Wavelet neural
networks for function learning, IEEE Trans. Signal Processing, vol. 43,
pp. 14851496, June 1995.
[18] Q. Zhang, Using wavelet network in nonparametric estimation, IEEE
Trans. Neural Networks, vol. 8, pp. 227236, Mar. 1997.
[19] Y. Qussar, I. Rivals, L. Personnaz, and G. Dreyfus, Training wavelet
networks for nonlinear dynamic input-output modeling, Neurocomputing, vol. 20, pp. 173188, 1998.
[20] L. M. Reyneri, Unification of neural and wavelet networks and fuzzy
systems, IEEE Trans. Neural Networks, vol. 10, pp. 801814, July
1999.
[21] F. J. Lin, R. F. Fung, and R. J. Wai, Comparison of sliding mode and
fuzzy neural network control for motor-toggle servomechanism, IEEE
Trans. Mechatron., vol. 3, pp. 302318, Dec. 1998.
[22] R. J. Wai, F. J. Lin, and K. K. Shyu, A toggle mechanism servo system
with variable structure model-following control, Int. J. Syst. Sci., vol.
30, no. 11, pp. 12131226, 1999.
[23] L. X. Wang, A Course in Fuzzy Systems and Control. Englewood
Cliffs, NJ: Prentice-Hall, 1997.