IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 33, NO. 3, MAY/JUNE 1997 697 Identification and Control of Induction Motor Stator Currents Using Fast On-Line Random Training of a Neural Network Bruce Burton, Associate Member, IEEE, Farrukh Kamran, Member, IEEE, Ronald G. Harley, Fellow, IEEE, Thomas G. Habetler, Senior Member, IEEE, Martin A. Brooke, Member, IEEE, and Ravi Poddar Abstract— Artificial neural networks (ANN’s), which have no off-line pretraining, can be trained continually on-line to identify an inverter-fed induction motor and control its stator currents. Due to the small time constants of the motor circuits, the time to complete one training cycle has to be extremely small. This paper proposes and evaluates a new form of the random weight change (RWC) algorithm, which is based on the method of random search for the error surface gradient. Simulation results show that the new form of the RWC, termed continually on- line trained RWC (COT-RWC), yields performance very much the same as conventional backpropagation with on-line training. Unlike backpropagation, however, the COT-RWC method can be implemented in mixed digital/analog hardware and still have a sufficiently small training cycle time. The paper also proposes a VLSI implementation which completes one training cycle in as little as 8 s. Such a fast ANN can identify and control the motor currents within a few milliseconds and, thus, provide self-tuning of the drive while the ANN has no prior information whatsoever of the connected inverter and motor. Index Terms—Induction motor, motor current regulator, neu- ral network, on-line training. I. INTRODUCTION T HE induction motor is a nonlinear system, the parameters of which vary with time and operating conditions. For high-performance applications, such as vector control and direct self control, it is necessary for the controller design to be based on observers and estimation techniques [1], which depend on a simplified model of the motor. Artificial neural networks (ANN’s) provide an alternative method of observing the input/output relationships of the motor. A previously presented scheme [2] proposed that an ANN be trained off- line, i.e., with data obtained a priori to mimic existing stator current controllers. Once sufficiently well-trained, the ANN could replace the original current controller with the advantage of increased speed of execution and fault tolerance. With this approach, no further training of the network is possible, after Paper IPCSD 96–53, approved by the Industrial Drives Committee of the IEEE Industry Applications Society for presentation at the 1995 Industry Applications Society Annual Meeting, Lake Buena Vista, FL, October 8–12. Manuscript released for publication October 7, 1996. B. Burton and R. G. Harley are with the Department of Electrical Engi- neering, University of Natal, Durban, 4001 South Africa. F. Kamran, T. G. Habetler, M. Brooke, and R. Poddar are with the School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA 30332 USA (e-mail: thabetler@ee.gatech.edu). Publisher Item Identifier S 0093-9994(97)02940-X. the drive is commissioned. Therefore, the performance of such an off-line trained ANN approach depends upon the amount and quality of training data used, which in turn depends on system complexity and the range of operating conditions involved and is also sensitive to parameter variations. Since the induction machine is a deterministic system for which the equations are well known, training of an ANN from a random initial condition is not necessarily required. Reference [3] proposes a current regulator for induction ma- chines which maps the electrical equivalent circuit equations onto a feedforward neural network and does not require training. However, like the off-line trained ANN scheme [2], the approach in [3] is also prone to degraded perfor- mance because of parameter variations. In order to account for unknown parameter variations, an observer-based scheme like [1] may be used, but unmodeled nonlinearities, such as magnetic saturation, can only be accounted for using an adaptive nonlinear-model-based controller, or by using an ANN which is trained while the drive controller (including the ANN) is operating on-line. Such an ANN scheme was proposed [4] which used continual on-line training (with no off-line training) to identify and adaptively control the currents and, therefore, the torque and, thus, the speed of an induction machine. Simulated results showed that this scheme could produce high dynamic performance similar to that achieved with conventional vector control. With an on- line trained scheme, custom tailoring of the ANN architecture and weights to match the structure of the motor equations, although possible, is not necessary. The scheme in [4] incorporates three ANN’s which are on-line trained using backpropagation, with two different rates of execution; a relatively slow rate for the two ANN’s which accomplishes the rotor speed identification and control functions and a much faster rate for the ANN performing the stator current regulation. The stator current loop control scheme, illustrated in Fig. 1, must run at a high enough rate to cope with the comparatively small electrical time constants of the machine. In order to achieve on-line training at a reasonable sampling frequency, e.g., 10 kHz, one epoch (one ANN weights update cycle based on the training error) needs to be completed in approximately 50 s. Due to the lack of a suitable ANN application specific integrated circuit (ASIC), the adaptive on-line trained current controller ANN of [4] was implemented in software [5], [6] on a transputer, but 0093–9994/97$10.00  1997 IEEE