Robust subdomain distribution alignment with class-wise weight selection mechanism for mechanical fault transfer diagnosis

Abstract Lots of fault transfer diagnosis methods have been presented to tackle the issue of class-wise distribution shift under the unsupervised scenario. Pseudo labels serve as the anchors during the fine-grained distribution alignment between the source and target domains, directly determining the overall diagnostic performance. However, most of them neglect the confidence of pseudo label, particularly the outlier disturbances induced by abnormal samples, thereby causing negative transfer in some cases. To bridge this gap, a new robust subdomain distribution alignment (RSDA) method is developed. In RSDA, a feature separation basic model is first constructed to extract the intrinsic fault features from the domain-specific noise background, so as to guarantee a noise-free feature alignment space. Then, an entropy-guided robust class-wise weight selection mechanism based on Shannon entropy and Jensen–Shannon divergence is designed to reduce the mismatch degree of the source-domain decision boundary with respect to target-domain samples. Besides, the progressive training strategy is proposed to learn the domain-invariant feature representations. Ultimately, the proposed RSDA is evaluated on several cross-machine transfer tasks built by three public bearing datasets, and experimental results show its superiority and effectiveness compared with the state-of-the-art distribution alignment methods.