Semi-supervised bearing fault diagnosis with adversarially trained phase-consistent network

Abstract

In this study, we propose an adversarially-trained phase-consistent network (APCNet), which is a semi-supervised signal classification approach. Generating ground-truth labels of large-sized datasets are costly and impractical in many real-world situations. Semi-supervised learning methods, which train models with partially-labeled datasets, provide practical solutions to those applications. Hence, we aim to present a bearing fault diagnosis approach that is robust on datasets with extremely small labeled to unlabeled data ratio. APCNet suggests three novelties for this objective: specialized encoder architecture for vibration signal data, phase-consistency regularization, and adversarially trained latent distribution alignment of labeled and unlabeled distribution. We conduct experiments on two public bearing datasets to evaluate the performance of APCNet as compared to other baseline methods. We interpret the model capabilities with numerous experiments of different data label ratio and latent distribution analysis. The results show that APCNet performs well on datasets with small labeled to unlabeled data ratio. Also, we show that APCNet achieves our objectives of capturing important features of vibration signals and modeling the true data distribution effectively.