Adaptation learning control scheme for a high-performance permanent-magnet stepper motor using online random training of neural networks

• Published in: IEEE transactions on industry applications Vol. 37; no. 2; pp. 495 - 502
• Main Authors: Rubaai, A., Kotaru, R.
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.03.2001; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; Computer simulation; Feedforward neural networks; Mathematical models; Motors; Neural networks; Neurofeedback; Noise measurement; Nonlinear dynamics; Signal processing; State estimation; State feedback; Steppers; Studies; Torque; Training; Trajectories; Trajectory
• ISSN: 0093-9994; 1939-9367
• DOI: 10.1109/28.913714

This paper addresses the problem of controlling the speed of a permanent-magnet stepper motor assumed to operate in a high-performance drives environment. An artificial neural network (ANN) control scheme which uses continual online random training (with no offline training) to simultaneously identify and adaptively control the speed of the stepper motor is proposed. The control scheme utilizes two three-layer feedforward ANNs: (1) a tracker identification neural network which captures the nonlinear dynamics of the motor over any arbitrary time interval in its range of operation; and (2) a controller neural network to provide the necessary control actions to achieve trajectory tracking of the motor speed. The inputs to the controller neural network are not constructed from the actual motor/load dynamics, but as a feedback signal, from the estimated state variables of the motor supplied by the neural identifier and the reference trajectory to be tracked by the actual speed. A full nonlinear model (with no simplifying assumptions) is used to model the motor dynamics, and to the best of the authors' knowledge this represents the first such attempt for this device. This paper also makes use of a very realistic and practical scheme to estimate and adaptively learn the noise content in the speed-load torque characteristic of the motor. Simulations reveal that the neural controller adapts and generalizes its learning rate to a wide variety of loads, in addition to providing the necessary abstraction when measurements are contaminated with noise.