24

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 32, NO. 1, FEBRUARY 2002

Dynamic Triangular Neural Controller for Stepper Motor Trajectory Tracking

Edgar N. Sanchez, Senior Member, IEEE, Alexander G. Loukianov, and Ramon A. Felix

AbstractIn this paper, we present a novel neural controller for this kind of motor. This controller is developed as follows. Modifying published results for nonlinear identification using dynamic neural networks (NNs), an NN identifier of triangular form is implemented. Then, based on this mode, a control law using sliding modes is derived. This neural identifier and the proposed control law allow trajectory tracking for stepping motors. Applicability of the approach is tested via simulations. Index TermsElectric motors, neural identifier, sliding modes, triangular forms.

I. INTRODUCTION INCE the publication of the seminal paper (see [1]), there has been increasing interest in applying neural networks (NNs) to identification and control of nonlinear systems. Most of these application are based on feed-forward ones [2], [3]. Lately, the use of dynamic NNs, which allows a more efficient modeling of dynamic systems, has been increasing [4], [5]. A very efficient algorithm for nonlinear identification, which ensures error exponential convergence using dynamic NNs, has been proposed in [6]. On the other hand, there has recently been a strong interest in applying nonlinear control methodology to electric motors [7]. One kind of motor which has been used to test nonlinear technique is the stepper motor. Permanent magnet stepper motors have begun to displace direct current motors in servo systems and direct drives. Today, these motors are widely used by industry for a variety of applications such as automotive control, printers, etc. Basically, these kinds of controllers have been designed using exact feedback linearization technique without including any robustification term [8][11]. Therefore, these controllers are functions of all plant parameters and external disturbance such as the load torque. In practice, some of these parameters are subjected to variations, and the load torque is unknown. Moreover, these control schemes do not take into account practical limitation on the magnitude of the stator voltage. On the other hand, a fruitful and relatively simple approach, especially when dealing with nonlinear plants subjected to disturbances, is based on the use of variable structure control

Manuscript received June 24, 2000; revised February 6, 2002. This work was supported by CONACYT, Mexico, under Project 32059A. This paper was recommended by Associate Editor J. Lee. The authors are with the Advanced Studies and Research Center of the National Polytechnic Institute (CINVESTAV), Unidad Guadalajara, Guadalajara, Jalisco, CP 45091, Mexico (e-mail: sanchez@gdl.cinvestav.mx). Publisher Item Identifier S 1094-6977(02)04677-1.

(VSC) with sliding mode [12]. First and foremost, this approach enables high accuracy and robustness in the presence of disturbances and plant parameter variations. Second, the control variables of the basic sliding mode control law rapidly switch between extreme limits, which is ideal for the direct operation of the switched mode power converters of electrical motors. It is known, however, that the sliding mode controller only achieves the robustness property of the closed-loop system with respect to matched uncertainties [13], while, for example, the load torque and variation of the moment of inertia constitute unmatched disturbances. In standard control schemes, to reject load torque effects, this torque is usually estimated by an observer, which depends on the plant parameters. To the best of our knowledge, the use of NNs in electric motors identification and control is mainly limited to application of the excellent approximation capabilities of feed-forward NNs for modeling, as in [14]. In this paper, we present a novel approach using a combination of NNs, sliding mode, and feedback linearization techniques. It is assumed that all of the state variables, including angle position, velocity, and stator currents, can be measured by an encoder and current sensors; by using the encoder, it is possible to measure the angle position as well as to calculate the velocity. Modifying published results for nonlinear identification using dynamic NNs [6], a neural identifier of the so-called triangular form [15], is proposed. Such an identifier enables us to use feedback linearization technique to design a nonlinear sliding surface with linearized sliding mode dynamics. Note that only one surface is designed for two components of the vector control input. This leads to the so-called singular sliding modes [12]. In this case, all of the control resources are fully used to achieve the control goal, that is, trajectory tracking for the stepper motor. Hence, by proper combination of a new neural identifier and discontinuous controller, the effect of the matched and unmatched nonlinearities and plant parameter uncertainties and disturbances can be suppressed, and total invariance is obtained when the motion of the system is in sliding mode. The paper is organized as follows. Section II presents the mathematical model of the stepper motor. In Section III, dynamic NNs of high order are described. The application of this kind of NNs to identification, as well as the proposed identifier, are discussed in Section IV. The control law (based on this identifier), feedback linearization, and sliding modes techniques, as applied to stepper motor trajectory tracking, are developed in Section V. The corresponding simulation results are included in Section VI. Finally, relevant conclusions are presented in Section VII.

1094-6977/02$17.00 2002 IEEE

SANCHEZ et al.: DYNAMIC TRIANGULAR NEURAL CONTROLLER

25

II. MOTOR MODEL The permanent magnet stepper motor model used in this paper is directly taken from [9]; and is formulated as follows:

Furthermore,

is a vector defined as . . . . . .

. . .

. . . a

with the input to the NNs, and smooth sigmoid function formulated by (1) where and

currents in phase and phase of the motor, respectively; and voltages applied on the windings of the phase and phase , respectively; and resistance and self inductance in each of the phase windings, respectively; angular displacement; angular velocity of the motor shaft; number of rotor teeth; motor torque constant; rotor load inertia; viscous friction; unknown external load torque. , , The following state variables are defined: , and . Henceforth, (1) can be rewritten as

For the sigmoid, is a positive constant and is a small positive . real number. Hence, Defining the vector

. . .

. . .

(3) can be rewritten as

or (4) . where Equation (4) describes the behavior of the th neuron. B. Modified Neural Model (2) Equation (2) clearly fulfills the conditions of triangular form systems. III. DYNAMIC NEURAL NETWORKS OF HIGH ORDER A. Neural Model Now, we will consider the NN structure presented in [6]. For the sake of completeness, a brief description of this NN is included here. In [6], high-order recurrent NNs are defined as (3) where where th neuron state; number of high-order connections; collection of no ordered subsets of , ; adjustable weights of the NN; no negative integers. (5) or (6) On the basis of the described recursive high-order neural network (RHONN), we propose the following modifications. First, they consider only first-order connections. Second, we consider that some of the components of the vector pass through any function nonnecessary to a sigmoid one. Third, we consider that some of the neural weights could be fixed. Hence

26

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 32, NO. 1, FEBRUARY 2002

Next, we propose to modelize the stepper motor by the following NN:

(7) It is obvious that this neural model also fulfills the conditions , , of triangular systems. In order to avoid zero crossing, , and are fixed and selected as , , , and . Additionally, for (6)

with and , and for any . In order to identify the stepper motor model, it is assumed that this motor is represented by the proposed RHONN (7). Then, two possible models can be built: 1) the parallel model 2) the series-parallel model where is the th component of the RHONN and is the stepper motor state. We indistinctly use either the former or the latter. A. On-Line Weight Update Law This law is taken from [6] and adapted to fit (6). The identi. Replacing with (9) fication error is defined as and neglecting , then

Derivating with respect to time

and selecting the weight update law as and finally . The scheme, to be proposed, is based on the following fact. Given a desired trajectory, expressed on state variables as , a stepper motor with state , and an NN as (7), then it is possible to establish the inequality

(10) , , with and a positive constant expressing the updating gain. Then, the identification error fulfills ,

where stands for the Euclidean norm, and . Hence, it is possible to divide the tracking problem into two parts: , which can be achieved by an 1) minimization of on-line identification algorithm, based on dynamic NNs; , for which a tracking algorithm 2) minimization of is developed by means of the sliding modes technique, on the basis of the neural identifier. IV. IDENTIFICATION To derive an identification scheme, it is useful to represent (6) as in the following lemma [6]. Lemma 1: The RHONN model (6) can be written as

Therefore, this updating law ensures exponential convergence of the identification error. In order to counteract the presence of unmodeled dynamics and the lack of persistence of excitation, in [6], the following update law is proposed:

(11) where is given as

(8) , is the RHONN initial state. where It is easy to see that is a filtered vector , and using this lemma, the RHONN time response can be formulated as (9) where It is clear that . with integer , and and positive constants. This law guarantees the exponential convergence of the error, and boundness of the weights and their derivatives. It is known in the adaptive control literature as the modification, and its robustness properties are well known [16].

SANCHEZ et al.: DYNAMIC TRIANGULAR NEURAL CONTROLLER

27

V. CONTROL LAW The goal of the proposed scheme is to force the stepper motor position to track a specified trajectory. In order to achieve this tracking, we developed a control law based on the sliding modes technique. Given the trajectory , the output tracking error and a new variable are defined by and (12)

with

positive constants. Applying (17) and (18) to (16) (19)

, where and stands for the absolute value. Since is always positive, the discontinuous control law (17) and (18), under the following condition:

Then, their first derivatives with respect to time are derived as (13) and guarantees convergence of the state to the sliding surface in a finite time. Once this is achieved, the sliding mode motion is described by the linear second-order system

where as

. Introducing , then a new variable

is formulated

(14) that yields

with desired eigenvalues and . Hence, and , which ensures tracking of the desired trajectory . A crucial property of the sliding mode control (17) and (18), when applied to (7), is that it yields the invariant , where . The subspace on this invariant subspace is referred to as the dynamic of (19), zero dynamics. To derive these dynamics, from must be the calculated equivalent control substituted in the fourth equation of (7), with values given by (18) (20) where

The time derivative of

is given by

(15) Substituting , in (15), and introducing a function

and

Then, vector driven to zero, which means

defined by (12) and (14), is , , and

Thus then where Finally, the NN identifier (7), expressed in new coordinates , and , is rewritten as , (21) . Since and is bounded, the solution of the zero dynamics (19) converges exponentially to a compact set [17]. Hence, the solution of the closed-loop system (7), (17), (18), (14), and (12) is ultimately bounded, and the tracking error (12) exponentially tends to zero. Remark 1: Since the proposed control scheme is based on the complete order continuous time plant model (1) or (2), there is no chattering in the closed-loop system. For discrete time applications, it is required to use the discrete time sliding mode control or unit control [12]. The complete scheme is implemented as follows. The varying weights of the NN identifier (7) are adapted on-line by means of the learning algorithm (11); then, the values of the adapted and fixed weights of (7) are used to calculate the control law (17) and

(16) To ensure tracking of law is proposed: , the following discontinuous control (17) (18)

28

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 32, NO. 1, FEBRUARY 2002

Fig. 1.

Control structure.

Fig. 2.

External load torque.

(18). Fig. 1 shows these different blocks of the complete scheme , with , for this application. where VI. SIMULATION RESULTS In order to illustrate its applicability, via simulations, the proposed scheme is applied to a stepper motor with parameters selected as the following. 1) Stepper Motor: V s/rad N m s /rad N m s/rad 2) Neural Network Identifier:

Fig. 3. Resistance increase.

3) Control Law:

The integration method used is the DormandPrince, with a s. Identification starts at . In order to step of ensure convergence of the identifier, a time interval of 1 s is allocated; then the control law is incepted. To test the robustness of the scheme, the following disturbances, incepted at the same time as the control law, are applied: a square wave load torque N m, and an increase with a with amplitude of constant rate of 0.5 /s. The time evolution of these disturbances are shown in Figs. 2 and 3. The simulation results are presented as follows. Fig. 4 shows position tracking, Fig. 5 shows angular velocity tracking, Fig. 6 shows position tracking error, and Fig. 7 shows the respective control actions. As can be seen, tracking is achieved in spite of external disturbances and parameter variations. To complete the simulation results, the following is included: the identification errors for the currents (see Fig. 8) and the weights time evolution for the third state of the neural identifier (see Fig. 9). From these results, it is noteworthy to state the following. Remark 2: The simulated external load torque (see Fig. 2) is a square wave. This kind of external disturbance looks similar to the effects of dry friction and backlash, among others.

Fig. 4. Position tracking.

Fig. 5. Velocity tracking.

SANCHEZ et al.: DYNAMIC TRIANGULAR NEURAL CONTROLLER

29

Remark 3: As shown in Fig. 8, for the motor currents, the neural identifier is able to estimate the respective states, in the sense that it adequately reproduces their dynamics. Similar results are obtained for the two other states as well. VII. CONCLUSIONS In this paper, we have discussed a novel scheme for trajectory tracking of stepper motors. We have modified a published algorithm in order to implement an on-line identifier developed on the basis of a new structure of NNs: the so-called triangular form. Based on this identifier, we derive a new control law, using sliding modes, which ensures trajectory tracking. Simulations results are very encouraging, especially those related to robustness of the proposed scheme in presence of parameter variation and external disturbances. Research is being conducted to implement this new control law in real time. ACKNOWLEDGMENT The authors would like to thank the useful comments of anonymous reviewers, which helped to improve the paper. REFERENCES
[1] K. S. Narendra and K. Parthasarathy, Identification and control of dynamical systems using neural networks, IEEE Trans. Neural Networks, vol. 1, pp. 427, Jan. 1990. [2] M. M. Gupta and D. H. Rao, Eds., Neuro-Control Systems, Theory and Applications. Piscataway, NJ: IEEE Press, 1994. [3] K. Hunt, G. Irwin, and K. Warwick, Eds., Neural Networks Engineering in Dynamic Control Systems. New York: Springer-Verlag, 1995. [4] K. Suykens, L. Vandewalle, and R. de Moor, Artificial Neural Networks for Modeling and Control of Nonlinear Systems. Norwell, MA: Kluwer, 1996. [5] A. S. Poznyak, E. N. Sanchez, J. P. Perez, and W. Yu, Nonlinear adaptive trajectory tracking using dynamic neural networks, IEEE Trans. Neural Networks, vol. 10, pp. 14021411, Nov. 1999. [6] E. B. Kosmatopoulos et al., Dynamical neural networks that ensure exponential identification error convergence, Neural Netw., vol. 7, no. 2, pp. 299314, 1997. [7] R. Ortega, A. Loria, P. J. Nicklasson, and H. Sira-Ramirez, PassivityBased Control of Euler-Lagrange Systems, Mechanical, Electrical and Electromechanical Applications. New York: Springer-Verlag, 1998. [8] M. Zibri et al., Static sliding mode control of a PM stepper motor, in Proc. Eur. Control Conf., Karlsruhe, Germany, Aug. 31Sept. 3 1999. [9] M. Zribi and J. N. Chiasson, Position control of a PM stepper motor by exact linearization, IEEE Trans. Automat. Contr., vol. 36, pp. 620625, May 1991. [10] M. Bodson et al., Feedback control of a permanent magnet stepper motor, IEEE Trans. Contr. Syst. Technol., vol. 1, pp. 514, 1993. [11] J. N. Chiasson and R. T. Novotnak, Nonlinear speed observer for the PM stepper motor, IEEE Trans. Automat. Contr., vol. 38, pp. 15841588, Oct. 1993. [12] V. I. Utkin, Sliding Modes in Control and Optimization. New York: Springer-Verlag, 1992. [13] B. Drajenovic, The invariance conditions in variable structure systems, Automatica, vol. 5, pp. 287295, 1969. [14] S. I. Moon, A. Keyhani, and S. Pillutla, Nonlinear neural network modeling of an induction motor, IEEE Trans. Contr. Syst. Technol., vol. 7, pp. 203211, 1999. [15] S. Celikovsky and E. Aranda-Bricaire, Constructive nonsmooth stabilization of triangular systems, Syst. Control Lett., vol. 36, pp. 2137, 1999. [16] P. A. Ioannou and J. Sun, Robust Adaptive Control. Englewood Cliffs, NJ: Prentice-Hall, 1996. [17] K. H. Khalil, Nonlinear Systems. Englewood Cliffs, NJ: Prentice-Hall, 1996.

Fig. 6. Position tracking error.

Fig. 7. Control actions.

Fig. 8.

Identification errors.

Fig. 9.

x weights time evolution.

30

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 32, NO. 1, FEBRUARY 2002

Edgar N. Sanchez (M85SM95) was born in Sardinata, Colombia, in 1949. He received the B.S.E.E. degree in power systems from Universidad Industrial de Santander (UIS), Bucaramanga, Colombia, in 1971, the M.S.E.E. degree in automatic control from the Advanced Studies and Research Center of the National Polytechnic Institute (CINVESTAV), Mexico City, Mexico, in 1974, and the Docteur Ingenieur degree in automatic control from Institut Nationale Polytechnique de Grenoble, Grenoble, France, in 1980. In 1971, 1972, 1975, and 1976, he was with different electrical engineering consulting companies in Bogota, Colombia. In 1974, he was a Professor in the Department of Electrical Engineering, UIS. From January 1981 to November 1990, he was a Researcher at the Electrical Research Institute, Cuernavaca, Mexico. He was a Professor of the graduate program in electrical engineering at the Universidad Autonoma de Nuevo Leon (UANL), Monterrey, Mexico, from December 1990 to December 1996. Since January 1997, he has been with CINVESTAV, Guadalajara Campus, Mexico, as a Professor of the electrical engineering graduate programs. His research interest centers on neural networks and fuzzy logic as applied to automatic control systems. He has published more than 50 technical papers in international journal and conferences, and has served as a reviewer for various international journals and conferences. Dr. Sanchez was granted a U.S. National Research Council Award as a Research Associate at NASA Langley Research Center, Hampton, VA (January 1985March 1987). He is a member of the Mexican National Research System, the Mexican Academy of Science, and the Mexican National Academy of Engineering. He has also been a member of international conferences (IPCs).

Alexander G. Loukianov was born in Moscow, Russia, in 1946. He received the Dipl. Eng. degree from the Polytechnic Institute, Moscow, in 1975, and the Ph.D. degree in automatic control from the Institute of Control Sciences of the Russian Academy of Sciences, Moscow, in 1985. He was with the Institute of Control Sciences from 1978 to 1995, and was Head of the Discontinuous Control Systems Laboratory from 1994 to 1995. From 1995 to 1997, he held a visiting position at the University of East London, London, U.K., and since April 1997, he has been with the Advanced Studies and Research Center of the National Polytechnic Institute (CINVESTAV), Guadalajara Campus, Mexico, as a Professor of electrical engineering graduate programs. His research interests include nonlinear systems control and variable structure systems with sliding mode as applied to electric drives and power systems control, robotics, SPICE, and automotive control. From 1992 to 1995, he was in charge of an industrial project between his institute and the largest Russian car plant. He has also been in charge of several international projects supported by INTAS and INCO-COPERNICUS, Brussels, Belgium. He has published more than 50 technical papers in international journals and conferences, and has served as reviewer for various international journals and conferences.

Ramon A. Felix was born in Guamuchil, Sinaloa, Mexico, in 1976. He received the B.Sc. degree from Instituto Tecnologico de Mar (ITMAR), Mazatlan Campus, Mazatlan, Sinaloa, in 1998, and the M.Sc. degree in electrical engineering from the Advanced Studies and Research Center of the National Polytechnic Institute (CINVESTAV), Guadalajara Campus, Mexico, in 2000. He is currently pursuing the Ph.D. degree at CINVESTAV. His research interests include neural control and VSC and their applications to electric machines and electric power systems.