Uso de redes neurais e conhecimento a priori na identificação de sistemas dinâmicos não lineares.

• Ano de defesa: 2001
• Autor(a) principal: Gleison Fransoares Vasconcelos Amaral
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: Engenharia Elétrica; Engenharia elétrica
• Link de acesso: http://hdl.handle.net/1843/BUOS-8D8EZL

This work investigates the use of artificial neural networks like multilayer perceptron MLP in identifying nonlinear dynamic systems. The main issues investigated were the possibility and feasibility of using a priori information to both constrain the network topology and the weights of the same during training. In order to test hypotheses, work with two home studies, which have become "test bench" in the identification of nonlinear systems. The systems used were the Duffing-Ueda oscillator, the main objective was to train a network in order to reproduce the bifurcation of the wishbone. To this end, we used information on symmetry of fixed points of the autonomous system to restrict the network topology. For the case of electric heater, the basic objective was to train a network in such a way as to ensure that the same behavior had previously known steady-state system. The main results have shown that it was possible to train a network that reproduces well the bifurcation diagram of the Duffing-Ueda system, within the range of parameters investigated, including presenting the bifurcation of the wishbone. How far does the authors' knowledge, MLP networks with such characteristics have not been reported in the literature. The use of symmetry information of fixed points Seems to have been important in achieving this result. Was obtained from a network that represents both dynamic and steady state of an electric heater to do this using constraints on the weights of the network during its training. A comparison of pruning techniques investigated showed that in general, they can reduce the network size without significant performance loss, but could not improve performance in steady state using such tools. Thus, it is believed that the procedures developed and tested in this work, have an important potential use of neural networks in problems of identification of nonlinear systems.