Doctoral Dissertations

Date of Award

8-1991

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Electrical Engineering

Major Professor

J. S. Lawler

Committee Members

J. Milton Bailey, James C. Hung, Robert E. Uhrig

Abstract

This dissertation addresses the on-line neural net control of a brushless DC machine (BDCM) in the presence of uncontrolled external disturbances. Its major contribution is its treatment of disturbance rejection in neural net control systems. First, a design model of the BDCM is derived that yields the machine parameters in terms of its geometry. This model establishes the nonlinear plant to be controlled and provides the design equations for an axial gap permanent magnet synchronous machine with a trapezoidal emf. The engineering assumptions needed to establish a linearized plant model for PI control are discussed. Next, a framework is established for the neural net control of the plant using back propagation. Back propagation was selected because of its ability to map multiple continuous inputs to continuous outputs. It was hoped the slow learning normally associated with back propagation could be overcome by judicious selection of the input variables, the neural net structure (the bi-net structure), and methods reported in the literature to increase learning speed. It was found, however, that none of these techniques brought the learning speed capability required of on-line control. Finally, a probabilistic neural net (PNN) control was formulated for the nonlinear plant. The PNN addressed all the difficulties encountered with the back propagation network, but posed new problems. In particular, the PNN was able to learn nearly instantaneously compared to back propagation, but it had a problem with training set selection. It was important to minimize the similarity between input patterns in the training set to avoid biasing the output toward the most common control. Allowing a nonorthogonal training set resulted in the inability to recognize aberrant behavior caused by unknown external disturbances. Several parameters were introduced to PNN control that affect its performance, and these performance sensitivities were determined through extensive simulation. It is concluded that PNN control is a viable means of on-line control. When compared to a proportional integral controller with current feedback, it has superior noise rejection properties and comparable robustness and disturbance rejection properties. Compared to back propagation neural nets, it is easy to train, is capable of complete on-line learning (no off-line training set), and accommodates unknown external disturbances very well.

Download
Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS