Both reliable and unreliable predictions matter: Domain adaptation for bearing fault diagnosis without source data

Abstract

Rolling bearing fault diagnosis is crucial for maintaining the reliability and safety of industrial systems. Recently, it has attracted increasing attention to transferring a diagnosis model from the source domain to the target domain without source data in real-world diagnosis scenarios due to confidentiality and efficiency concerns. However, existing approaches are sub-optimal as they simply exploit confidently pseudo-labeled target samples, and simultaneously overlook the intrinsic structural characteristics of the feature space. Besides, the reliability of fault pseudo-labels is always estimated with entropy, whose accuracy could be improved through more sophisticated strategies. To address these issues, we propose to explore the correlation between features and pseudo-labels in the target domain to maintain the balance between feature discriminability and feature diversity. In addition, we develop a voting-based strategy associated with data augmentation for more accurate reliability estimation of fault pseudo-labels. The proposed method is able to utilize both the reliable samples and unreliable samples for diagnosis model transfer via self-supervised training and distribution structure discovering respectively. Extensive experiments on two bearing fault benchmarks demonstrate the effectiveness and superiority of our proposed method. The source code is publicly available at: https://github.com/BdLab405/SDALR.