Dictionary Learning for Bearing Fault Diagnosis

Monitoring the vibration signal is an effective way for automatic diagnosis of mechanical faults in rotary electric machines. This is done in order to to determine the state of health of electrical machines, especially those used in Hybrid Electric Vehicle (HEV) and Electric Vehicle (EV) applications, whose state of health needs to be monitored continuously. Feature extraction is the key step in vibration diagnosis. Usually, the time domain vibration signal is transformed into another domain like Fourier or wavelet for feature extraction. In this paper a dictionary learning method is proposed for finding an optimal transformation for feature extraction. A label consistent sparse representation is applied to vibration data collected from an electric machine to diagnose three classes of bearing faults with a higher accuracy, compared to other extraction methods found in the literature. The trained transform projects the vibration signal from different classes of faults to different basis function which improves the classification task.