Feature set evaluation and fusion for motor fault diagnosis

This paper proposes a novel approach to the feature fusion in motor fault diagnosis with the main aim of improving the performance and reliability of clustering and identification of the fault patterns. In addition, the significance of individual feature sets in specific fault scenarios, which is normally gained by engineers through experience, is investigated by using flexible Non-Gaussian modeling of the historical data. Furthermore the comparison is made by applying individual and fusion of feature sets to the probabilistic distributions of trained models using a Maximum a Posteriori (MAP) approach. To carry out the task, current waveforms are collected non-invasively from three-phase DC motors. Waveforms are then compressed into time, frequency and wavelet feature sets to form the input to the clustering algorithm. The result demonstrates the suitability of specific feature sets in different motor modes and the efficiency of fusion which is carried out with a Winner Takes All (WTA) approach.