Multirepresentation Dynamic Adaptive Network for Cross-Domain Rolling Bearing Fault Diagnosis in Complex Scenarios

Recently, cross-domain fault diagnosis employing transfer learning methods has made remarkable outcomes. However, many existing methods struggle to perform satisfactorily when confronted with class imbalance, different fault severities, and daunting cross-machine tasks. To deal with these challenges, the multirepresentation dynamic adaptive network (MDAN) is proposed in this work. The main contributions of the MDAN include three aspects. First, a novel cyclic frequency-domain attention mechanism (CFDAM) is constructed to enhance the rolling bearings' key state information expression. Second, a multirepresentation operation is proposed to generate different feature subspaces, enabling the extraction of more comprehensive features across various scales. Third, the domain adaptation module incorporates the dynamic factor with a normalized weighting (NW) strategy to adaptively evaluate the relative importance of marginal and conditional distribution alignment. Specifically, through cyclostationary analysis, the states of rolling bearings are first characterized by spectral coherence (SCoh) maps. Then, the transferable features extracted by the CFDAM-based feature extraction network are aligned dynamically in respective subspaces based on multikernel maximum mean discrepancy (MK-MMD). Ultimately, the feature vectors from different subspaces are concatenated for label prediction. The effectiveness of the proposed MDAN was validated by the Case Western Reserve University (CWRU) rolling bearing dataset and two self-conducted rolling bearing datasets (obtained from rotor rolling bearing testbed and centrifugal pump testbed). The superiority of MDAN is also demonstrated by the comparison with several state-of-the-art methods.