Bearing Fault Detection and Classification Based on Temporal Convolutions and LSTM Network in Induction Machine

One of the critical components in most electromechanical systems are the bearing system. Therefore, a proper condition monitoring method that can classify the type and the severity of electrical machine faults in different load levels is crucial to avoid unwanted downtime and loss of operation. Non-invasive condition monitoring methods based on electrical signatures of machine in an electromechanical system, are considered as simple and cost-effective approaches for the fault detection process. In this paper, a deep learning approach based on a combination of temporal convolutions and Long Short Term Memory (LSTM) network is used for fault diagnosis. The two architectures are both shown to be effective for time-series classification and sequence modeling. Temporal convolutions are shown to be competent in feature extraction for time-series classification; however, they are rarely studied in bearing fault detection and classification in an electromechanical system. The presented method does not need any preprocessing or predetermined signal transformation, and uses the raw time-series sensor data. In this regard, three different faults, as inner race, outer race, and balls are considered for validity of the proposed method. The results show that healthy cases can be separated from faulty cases in different load levels with high accuracy (95.8%).