Intelligent identification of induction motor conditions at several mechanical loads

This paper presents an assessment of three classification models, all based on computational intelligence techniques, for the automatic identification of three possible conditions found in induction motors: healthy, with a half broken rotor bar or with one broken bar. Motors with full-load, half load and no load were considered. Based on evidence previously reported, the power spectral densities of the absolute value of a motor currents and vibrations are suitable as signatures of possible damages. However, a simple statistical analysis over such raw signals is not enough to accurately identify such conditions. In order to obtain good identification performance, we looked for accurate characterizations of vibration signals, as well as for suitable classifiers, in order to improve performance rates previously reported. We found that a feature extraction method, based on simple statistics of Discrete Wavelet Transforms (DWT), was able to characterize well the conditions with different loads. Classifiers analyzed were based on three strategies: one-nearest neighbors (1-NN), Support vector machines (SVM) and multi-layer perceptrons (MLP). Our 3-fold validated experiments reported up to 100% accuracy for motors with full load and no load; a 99.87% of accuracy was obtained in motors with half load, using 1-NN.