Identificação de sistemas não lineares utilizando aprendizagem profunda e modelagem orientada a blocos.

Abstract

In this masters dissertation, connections between the areas of systems identification and deep learning are explored. In this context, a method for nonlinear systems identification based on deep neural networks and nonlinear block-oriented models is proposed. The proposed approach is formulated from the separation of linear and non-linear parts of the Hammerstein and Wiener models. In this sense, a neural network structure that reflects this separation strategy is used and the identification procedure is performed by training the deep neural network with input and output data from the systems of interest. To test the proposed method, six different Hammerstein systems and six Wiener systems of different complexities were simulated in order to generate input and output data for the application of identification techniques. As a reference, another identification method was adopted which, although also based on the separation of linear and nonlinear parts of block-oriented models, employs orthonormal and radial basis functions for the identification task. The comparison of the results obtained reveals that the proposed method, in general, provides more accurate models.