Self-supervised feature learning for motor fault diagnosis under various torque conditions

Abstract

Even though research on deep learning-based motor fault diagnosis has been actively conducted, however, several challenges still make it difficult to be effectively applied in industry. First, it is hard to acquire labeled data due to resource shortages; specifically, annotation is expensive and time-consuming, thereby making labeled data scarce. Second, even if acquired data were available, they could have large distributional differences, since motors used in practical settings generally operate under various torque conditions. This paper thus proposes a novel self-supervised feature learning framework for motor fault diagnosis, namely notch filter augmentation-based multi-channel self-supervised learning (NFA-MSSL). The proposed NFA-MSSL method includes a preprocessing step that extracts the instantaneous amplitude of the stator current signal to maximize the use of unlabeled stator current signals, thereby increasing the model performance with limited labeled data. Furthermore, NFA-MSSL also includes a notch filter augmentation (NFA) method for stator current signals that can extract torque-invariant, fault-related features by transferring the frequency domain knowledge to the time domain. The proposed method is validated by applying it to an open-source induction motor dataset and a permanent magnet synchronous motor dataset. The results show that the proposed method properly extract fault-related features from the stator current signal, thereby outperforming conventional self-supervised and supervised algorithms under various torque conditions and insufficient labeled data.