Identificação caixa-cinza de sistemas não lineares utilizando representações narmax racionais e polinomiais.

Abstract

This thesis investigates the use of prior information in the identification of non-linear dynamic systems. This enables obtaining more robust models with leas parameters, greater generalization capacity and better performance in describing the behaviour observed in the real data. A study of different representations is presented and usend in the identification of non-linear systems. The performance acting of NARMAX rational and polynomial models and other representations is investigated. These representations are compared in the identification of chaotic systems and also in the case of systems with onde input and one output. In the latter, it is verified if the models identified from dynamic data are able to fit the system static curve previously known. The main contribution of this thesis are algorithms developed to use prior information in the identification of NARMAX rational and polynomial models. The first refers to the choice of the model analysis and validation. A detailed study of these representations is presented. With them it is possible to understand which role is played by each term in the model. The relationship between term clusters and the static function of the model allows one to choose a viable model structure that will be able, at least in principle, to represent the ststic characteristic desired. The developed techniques are applied to several dynamic systems, real and simulated.