Control of dynamic systems using feedforward neural networks

John Gustav Kuschewski, Purdue University

Abstract

We investigate dynamic system control using feedforward neural networks (FNNs) that use generalized weight adaptation algorithms. We first analyze a method for control of linear, time-invariant, single-input single-output dynamic systems using an adaptive linear element (Adaline). The main feature of the control structure is a single feedback loop. We then present an algorithm, based on root locus concepts, for determining the proper range of the Adaline's learning rate. A case study, that includes computer simulations and laboratory experiments, is carried out to evaluate the performance of this method when applied to the position control of a permanent magnet armature-controlled dc servomechanism, the main element of which is a permanent magnet armature-controlled dc servomotor. We then present methods for identification and control of dynamic systems using Adaline, two-layer, and three-layer FNNs equipped with generalized weight adaptation algorithms. The main feature of the control structure is the coordination of feedforward and feedback loops. The FNNs considered contain odd nonlinear operators in their neurons and in their weight adaptation algorithms. We carry out two case studies to evaluate the performance of the proposed methods. The first case study involves a nonlinear, time-invariant, dynamic system consisting of an inverted pendulum controlled by an armature-controlled dc motor through a gear train. We use computer simulations to evaluate the proposed methods of on-line FNN based identification of the system's forward and inverse dynamics. Specifically, our interest is in the effect the type of nonlinear functions in the neurons and in the weight adaptation algorithms have on identification performance. We then evaluate the proposed methods of FNN based control via on-line identification of the system's inverse dynamics combined with the coordination of feedforward control method. The second case study involves FNN based position control of a dc servomechanism. In this case study, we evaluate the performance of the FNN based control method.

Degree

Ph.D.

Advisors

Zak, Purdue University.

Subject Area

Electrical engineering|Artificial intelligence

COinS