An interpretable cross-domain bearing fault diagnosis method based on domain-aware and informed-guided domain adaptation
TL;DR
Validation on multiple public bearing datasets demonstrates that IGDATL achieves higher accuracy and superior domain generalization performance, providing an innovative and practical solution for cross-domain intelligent diagnosis of complex industrial systems.
Abstract
Accurate cross-domain fault diagnosis of bearings is critical to the collaborative and intelligent operation of mechanical systems. However, structural differences between devices and variations in operating parameter settings cause the collected sensor data to exhibit notable nonlinear feature distribution shifts. This leads to a marked diminishment in the feature adaptability of existing fault diagnosis methods and a drastic increase in diagnostic errors. To address this issue, this paper proposes an informed-guided domain adaptation transfer learning framework (IGDATL) for cross-domain bearing fault diagnosis. Firstly, a domain-aware squeeze-and-excitation Net is constructed. This network utilizes Domain-Specific Batch Normalization (DSBN) to adaptively standardize different domain features through perceiving domain discrepancies, which significantly enhances the comparability and consistency of feature spaces across domains. Meanwhile, a Squeeze-and-Excitation (SE) attention mechanism is integrated into the feature extraction network to amplify the coupling responses between channels that are relevant to fault information. Secondly, an informed-guided progressive alignment module is designed that embeds the dynamic alignment evolution regularity based on prior constraints into the domain adaptation process. It resolves the ambiguity in the coordination between domain and subdomain adaptation and improves the effectiveness and reliability of cross-domain feature migration. Finally, the informed-guided progressive alignment module is coordinated with an adversarial module to achieve comprehensive alignment across multiple dimensions, encompassing category distinguishability, domain distribution consistency, and intra-class structural tightness. Validation on multiple public bearing datasets demonstrates that IGDATL achieves higher accuracy and superior domain generalization performance, providing an innovative and practical solution for cross-domain intelligent diagnosis of complex industrial systems.