Controladores neurais adaptativos.

Abstract

The principal objective of this Thesis is to demonstrate that for a class of non-linear systems with partially known models, it is possible to employ a multi-layer artificial neural network (MLANN), for an adaptative control strategy, without the need for off-line training of the neural network. Initially, some experimental results, as well as simulation results, related with the position and speed control of a d.c. motor are presented. The experimental results were obtained both with the conventional controllers like the PID controller and reference model adaptive controller and the non-conventional controller like a neural network controller. Based on the simulations studies and comparisons of the performance of the conventional and neural controllers, the configuration of a neural network controller is proposed. Using the principal neural controller strategies proposed in the literature, and more specifically the direct and indirect controllers and the ones based on non-linear functions, experimental studies were carried out for a d.c. motor drive system with these controllers. These investigations revealed that if the jacobian of a plant is known, it is possible to transform the direct neural controller into an adaptive neural controller. Based on the plant jacobian, the concept of a passive state for the plant under control is introduced. The use of this concept enables the on-line training o f the multi-layer artificial neural network in real time with a fair degree of reliability and the necessity for a prolonged off-line training of the MLANN is thus avoided. This also affords a possibility for on-line fine-tuning of the neural controller. Some architectural aspects and other characteristics of the MLANN, as an adaptive controller, have been verified. A method to determine the optimum number of neurons in the hidden layer is presented. The adaptation factor for a direct adaptive neural controller is defined and an experimental procedure to determine its value, is presented. With a view to experimentally demonstrate the possibility of the generalization of the direct neural adaptive controller, concepts of fuzzy control are combined with those of the neural control to implement a position controller for a rigid inverted pendulum arm mounted on the d.c. motor shaft.