Maximização da correntropia por simulação com poda aplicado a detecção de estruturas e estimação de parâmetros de modelos NARX

Abstract

In the last decades, due to the growing complexity of dynamic systems and the growing demand for better performance, the area of systems identification has emphasized the use of non-linear models to represent dynamic systems. In this context, Non-linear autoregressive with exogenous inputs models (NARX) are heavily used due to to their simplicity, flexibility and capacity of better representation. However, such models rely heavily on structure selection and the most traditional algorithms have limitations when the data is contaminated by non-gaussian distribution noises. Noting this, in this thesis, the objective is to present a new identification method called simulated correntropy maximization with pruning which uses concepts of learning based on information theory. In this work basic concepts about systems identification and correntropy, methods based on orthogonal least squares and simulated error reduction, and the new proposed methodology. The proposed method is applied and compared to the traditional methods in some study cases. The first experiment is composed by three SISO numeric dynamic systems in the presence of bimodal noise. The second study case is a set taken from a benchmark system called Silver Box. The third is a real dynamic system. The obtained results validate the performance of the proposed method when compared to other algorithms of structure detection and parameter estimation, showing that the proposed method presents a better and more robust performance in the presence of non-gaussian distribution noise.