ADAPTIVE CONTROL OF ELECTRIC DRIVES USING NEURAL NETWORKS CM. Lim ‘Y-Z. Ouyang, V.K. Ouyang and T.H. Lee Electronic & Computer Engineering Department ‘The Chinese High School ‘Ngee Ann Polytechnic Bukit Timah Road ‘Singapore Singapore Abstract A simple adaptive control scheme using neural networks is applied to DC drives through simulation and real-time experimental studies. The primary objective of the paper is to demonstrate the effectiveness of this control scheme for different operating conditions. The adaptive neural control scheme consists of two parts. One isa neural estimator which is used to estimate the dynamics of the drive system under control. The other is a neural controller which is used to generate a control signal so that a pre-defined system output is regulated (0 a desired value. Simulation and experimental results are presented to demonstrate the effectiveness of this neural adaptive control scheme. 1. Ivtropuction Recently, the application of neural networks (NNs) to tackle problems for which conventional control schemes prove inefficient has received much attention [1]. These problems emerged largely due tothe fact that the real world ‘cannot be precisely modelled with mathematical equations. On the other hand, an ordinary person can tackle these problems without much effort. Its therefore not surprising that researchers working on intelligent control schemes have taken to modelling them after the way the human brain works. NNs are based on the way the human brain is structured, Basically, a NN consists of layers of simple processing elements known as neurons, connected 10 one another. The performance of the NN is determined by the sirengths of the connections between the nodes, also known as weights. Although the structure of the artificial NN is similar to that of natural ones, the precise operations can vary greatly. Nevertheless, the basic morphological features and the general approach to computation are ‘common ia both natural and artificial Ns. This can result in programs that can efficiently solve a wide variety of problems for which conventional approaches are inefficient ‘but natural intelligence does very well. In this paper, we apply an adaptive neural control scheme to two drive systems in order to demonstrate its efficiency over a wide range of operating conditions. Digital simulation and real-time experimental studies will be used to evaluate the performance of this control scheme. 7803-3773-5/97/810.00 © 1997 IEEE 2. ADAPTIVE NEURAL CONTROL SCHEME ~~ | : poet FT , | I Figure I: Closed-loop System Figure 1 shows the block diagram of the overall closed-loop system which is equipped with an adaptive neural ccontrolter. Essentially, the adaptive neural control scheme ‘consists of a single-layered neural estimator which is used to estimate the dynamics of the drive under contol Specifically, the drive dynamics is given by elk) = Yo wi =D kD @ where xk = yk: x(k= 1) = uk i; ton n+Ltontm and y isthe drive output, isthe contol signal and y isthe estimated drive output. “The integer constants n> and im > 0 ate predefined by the designer. The Widrow-Hoff rule (2}is used to update the weights, w, so that the error, = ok) - yh), is minimised, i. 5 tetk-+ 1208) wk+1 - ea x(k) @ where WCE) = (oe Wy OD Wag Drones Wye OT CK) = [Dy oka seem OTand & €(0,2) and € is chosen to be close to zero, ‘The control scheme makes use of a neural controller to ‘generate the control signal such thatthe drive output tracks the desired reference with a zero steady-state error. This neural controller is of the form. | @ 2B) = Diy E+ Dina MA Dros pe rene OO a 9 se (DMEM OT w where yqris the reference signal. In order to enhance the damping characteristics of the overall system, the above drive output is modified to yu (k= WE) Ky SO) a) where Ky is a constant [3]. The above modification can be readily made by replacing y(.) by yu.) in the above expressions. Furthermore, the increase in the computational burden due to this modification is not significant, Most importantly, the system damping can be enhanced as a derivative term possesses. anticipatory characteristics and takes early corrective actions. 3, Drive SySTEMS AND RESULTS 3.1 Simulation The above adaptive neural controller has been tested using simulation studies by applying it fo control the speed of a DC motor. In the simulations, the transfer function of the DC motor is of the form o K Vin Tes 6) ‘where «is the motor speed, Vin is the applied voltage, K is the DC gain and T'is the motor mechanical time constant, Figure 2. shows the results obtained when the DC motor ‘mechanical time constant was varied from 20 ms to 120 ms after the 75th sampling period. For comparison purposes, the results obtained when a PID controller was applied are included in Fig. 2. Simulation results show that the adaptive neural controller is superior to the PID controtler in the presence ofa large change in the system parameter. Tan Figure 2a: Simulation results with neural control scheme Figure 2b: Simulation results with neural control scheme Figure 24: Simulation results with PID controller 32 Experimental In order (0 further demonstrate the effectiveness of the adaptive neural control scheme, experimental studies have also been performed. Specifically, an experimental one- link manipulator, consisting of a long, light metal rod with 4 payload attached to one end and driven by a DC servo- ‘motor, has been chosen to test the adaptive neural control scheme in realtime. Figure 3 shows the overall setup of the experimental system. Figure 4 shows the experimental results obtained for different payloads and reference signals, Another one-link manipulator was chosen for experimental studies. Figure 5 shows the experimental results obtained for different initial conditions “The effect of setting Ky to zero on the performance of the neural controller was examined and the results obtained are as shown in Fig. 5.Micro-Computer PA = Power Amplifier DS Drive System Figure 3: Experimental system Discusstons Both the simulation and real-time experimental results show that the neural control scheme is able to provide good performance to the systems under contro. It is of interest to take note of the experimental results shown in Figures Se and Sf where Ky is set to zero . 5. CONCLUSIONS ‘An adaptive neural control scheme has been proposed and ‘applied to different drive systems. Simulation and ‘experimental results have shown that this control scheme is effective over a wide range of operating conditions. 6. ACKNOWLEDGEMENTS. ‘The work reported in this paper is part of the Gifted Programme of The Chinese High School. Y.Z. Ouyang, Y.K. Ouyang and TH. Lee are deeply indebted to Ngee ‘Ann Polytechnic for providing them with the necessary facilities and equipment to complete this project, and to the Gifted Programme Department of The Chinese High Schoo! for its support throughout the duration of this project. 17. REFERENCES UU] D. Psaltis, A. Sideris and A. A. Yamamura, “A ‘Matilayered Neural Network Controller", /EBE Control Systems Magazine, pp. 17-20, 1988 (21 V. Etxebarria, “Adaptive control of discrete systems using neural networks", JEE Proc. D, Control Theory Appl., Vol. 141, No. 4 , 1994, pp. 209-215 [3] C.M. Lim, “Correspondence: Adaptive Control of Discrete Systems using Neural Networks”, IEE Proc.-Control Theory Appl, Vol. 142, No. 5, September 1995, pp. 526-528 Angular Position (V) x 6 Reference e 3 6 sawpling tine x 108 fingulan Position (W) x 25 sanpling tine x 108 Angular Position (V) x a 1 Output Reference a a 6 sampling tine x 108 Figure 4: Experimental results (a) =normal payload (b),(¢) =heavy payloadfmgular Festtion (x 5 2 1 Output t a Reference i ee ‘ 2 (a ¢ ‘ Angular Position W) x 5 2 Reference: 1 N a 4 6 (eo) sampling tine x 168 Reference 4 6 (eo) sausling tine x 188 Centro! Signal (W) x 2 i 4 Control Signal (W) x 9 2 a Cantrol Signal (W) x 3 t 1 Figure 5: Experimental results (@)-@) Ky > 05 ©) Kw=0 t (a) . (fe) (b) 4 om * 4 t sampling tine x 186 4 L Sunpling tine x 10