Mining Knowledge from Unlabeled Data for Fault Diagnosis: A Multi-Task Self-Supervised Approach

Deep learning-assisted fault diagnosis has achieved significant success in recent years due to its capability of automatic feature learning and intelligent decision-making. Nonetheless, supervised methods are limited by their demands for annotations and fail to explore the growing unlabeled data generated by monitoring devices. An urgent need arises for an efficient approach to utilizing massive unlabeled data to facilitate fault diagnosis. However, existing unsupervised approaches usually rely on a single task for representation learning and lack the synergistic consideration of the cooperation of multiple tasks. To this end, a multi-task self-supervised approach is proposed to comprehensively mine diagnostic knowledge from unlabeled data. Three self-supervised tasks, namely Contrastive similarity matching, Pseudo Label predicting, and Intra-sample temporal relation reasoning (CPLI), are designed to learn representations of vibration signals at inter-instance, instance, and inner-instance levels, respectively. They are meticulously designed and combined to work corporately. The first task focuses on estimating similarities among pairs of augmented samples, while the second task helps this process by guiding the model to identify augmentation methods. As an important complement, the third task delves into the temporal relations among pieces of a time series. Three case studies demonstrate the superiority of the CPLI over state-of-the-art methods in terms of domain adaptability and diagnostic accuracy. These findings highlight its potential for leveraging unlabeled monitoring data to benefit fault diagnosis.