Classificação de falhas em maquinas eletricas usando redes neurais, modelos wavelet e medidas de informação

Abstract

This work presents a methodology for diagnosis and classification of faults in three-phase induction motors connected directly to the power grid. The proposed method is based on the analysis of the stator current signals, with and without the presence of faults in the bearings, stator and rotor. These faults cause the presence of specific frequency components that are related to the machine rotational speed. The signals were analyzed using wavelet-packet decomposition, which allows a multiresolution evaluation of the signals. Using this decomposition, we estimated some predictability measures, such as relative entropy, predictive power and normalized error variance, obtained with the predictability component analysis. With this measures, we verified which were the most predictable components. In this work, normalized error variance and the predictive power were used as inputs to three topologies of artificial neural networks used as classifiers: multilayer perceptron, radial basis function and Kohonen self-organizing maps. We tested six different input vectors to the artificial neural networks, in which we vary the predictability measures and the number of elements of the vectors. The studies were performed considering samples of signals from different motors, with various kinds of faults, working under several load conditions and with voltage unbalance. The signals were firstly classified in two patterns: with and without the presence of faults. After, we detected the kind of fault was present in the signal: bearing, stator or rotor fault. Last, the samples were classified inside the subgroup in which they were.