Redes neurais com estados de eco aplicadas em controle dependente dos estados

Abstract

In around 1764, it emerged a new branch of science – the theory of control - when James Watt was given a model Newcomen engine to repair. He realised that it was hopelessly inefficient and began to work to improve the design. He did a velocit controller to solve the problem. The evolution of these systems is shown in State-dependent Riccati equation (SDRE) techniques. Although it is a very advanced technique in relation to the capacity of performing control, some problems have to be faced for its use, as the necessity of computational resources of high level and cost, which may impede the use of SDRE in some systems. The solution for these problems is pointed out in this study by the use of Echo State Neural Networks (ESNs). These neural networks have inputs and outputs and the inputs are processed through the use of algorithms in order to reach the desired results, and for that the neural network has to be under a task of training. After that we use the behavioral SDRE data for the training followed by the neural network efficiency test for the system control and for the computational cost. The results are compared to the ones obtained with the ESN control. This test was realized for a micro eletromechanical system and the control of the active suspension of a half-car. The results were positive as the ESN could perform the control in a short time in relation to the SDRE. There is also a fixed structure which makes possible some adjusts for different kinds of control.