Data Generation Method for Domain Adaptation in Fault Diagnosis Using Motor Current Signals

Recently, the demand for data-driven fault diagnosis using motor current signals has increased, however, collecting labeled abnormal data in actual environments is difficult. We proposed a method to generate current signal data by combining simulation data, generated using the multiple coupled circuits model, with normal real data. The data are combined using frequency and phase matching and data assimilation. Our proposed method can reproduce more diverse features than conventional methods and enables more domain-invariant features to be learned by sim-to-real domain adaptation. We evaluated the proposed and conventional data generation methods on the task of classifying the number of broken rotor bars in an induction machine using five unsupervised domain adaptation algorithms. Experimental results demonstrate that the proposed method outperformed all conventional methods with respect to classification accuracy. This study was the first to adopt sim-to-real domain adaptation for current-signal-based fault diagnosis.