Integrating self-supervised learning with domain adaptation based on Multi-view Structural Consistency and joint learning for bearing fault diagnosis

Abstract Rotary machinery is an indispensable part of modern industry; it greatly reduces labor and improves production efficiency. Therefore, accurate diagnosis and early warning of faults help minimize repair costs, increase efficiency, and prevent problems. Deep learning-based fault diagnostic methods often assume training and testing data are in the same domain and require large amounts of labeled data without taking advantage of the amount of unlabeled data. To solve this problem, we propose an end-to-end learning method called Self-Supervised Domain Adaptation Network with Multi-view Structural Consistency (SSDA-MSC). The proposed method integrates knowledge extracted from the labeled source domain with representative features learned from unlabeled data to enhance transferability to the target domain, thus maximizing the informativeness of the labeled samples while effectively leveraging the vast potential of unlabeled datasets. The proposed method is evaluated on two bearing datasets: Paderborn University (PU) and Hanoi University of Science and Technology (HUST). The experiments show that the model achieves an average classification performance of 87.76% on the PU dataset, 93.36% on the HUST dataset for the test set of the target domain with only 50 training samples per class of the source domain, outperforming previous methods and showing that this method has potential for industrial applications, particularly in contexts where unlabeled data are abundant but labeled samples are limited. The source code of the model will be available.