Wavelet packet measurements and neural networks applied to stator short-circuit diagnosis

Abstract

Detecting stator failure is crucial for maintaining reliability in manufacturing processes. The diagnosis in the early stages is challenging, and the industrial environment imposes even more significant difficulties on this task. The voltage unbalances in the power supply are one of the most significant obstacles to correctly identifying stator faults since they cause effects similar to failures. Also, different mechanical torque levels may confuse the diagnosis. This combination of adverse conditions is often neglected in motor health monitoring studies. Therefore, this work develops a new approach for induction motor short-circuit classification. Here, the predictive power, a predictability measure based on relative entropy, is used to extract relevant features from wavelet components. Experiments show that multi-layer perceptron learned better the patterns extracted from the predictive power than root mean square, mainly for incipient faults. The results demonstrated that the predictive power is a reliable stator fault indicator, considering up to 1% of short-circuited turns and a wide range of voltage unbalances and load levels.