ISEANet: An interpretable subdomain enhanced adaptive network for unsupervised cross-domain fault diagnosis of rolling bearing

Unsupervised Domain Adaptation (UDA) has gained widespread application in bearing fault diagnosis across various operational conditions, attributed to its commendable transfer diagnosis efficacy. However, global Domain Adaptation (DA) under the influence of noise interference often overlooks subdomain distribution, leading to local distinctions among multiple categories. To address these challenges, this paper proposes an Interpretable Subdomain Enhanced Adaptive Network (ISEANet), which enhances subdomain representation from key facets: initial sample processing, intermediate feature mapping, and subdomain discrepancy calculation. Firstly, the Sparse Subsegment-guided Noise Reduction (SSNR) layer is formulated to enhance the physical knowledge. Subsequently, Lightweight Multi-Feature Extraction Module (LMFEMod) is designed to comprehensively capture domain discriminable features from local and global perspectives to enhance the coordinated adaptation and interpretability between physical knowledge and feature mapping. Moreover, a novel subdomain metric method, Improved Local Maximum Mean Discrepancy (ILMMD), is proposed. ILMMD introduces a priori probability distributions between different labels, replacing the original hard labels. This modification aims to increase the distance between clustering centers and bridge subdomain gaps during Subdomain Adaptation (SA), and further enhances the reliability of subdomain discrepancy calculations. Comparative tests with other prevalent methods on public and Lanzhou University of Technology (LUT) bearing dataset for the transfer task are conducted, and the results show that ISEANet exhibits excellent cross-domain diagnostic performance.