A joint collaborative adaptation network for fault diagnosis of rolling bearing under class imbalance and variable operating conditions

Abstract

With the dynamic evolution of electromechanical equipment processing tasks, rolling bearing fault diagnosis is often hindered by variable operating conditions and imbalanced fault data, which compromise the recognition of minority fault types and cause significant domain shifts. Multi-source domain adaptation, by integrating data from multiple sources, can alleviate domain shifts and partially mitigate the class imbalance issues, but a dedicated class-aware mechanism is still needed to further enhance performance on minority fault classes. To jointly tackle these challenges, a joint collaborative adaptation network (JCAN) is developed within a multi-source domain adaptation framework that integrates transfer learning, information fusion, and class-aware techniques. Specifically, JCAN extracts domain-invariant features through adversarial training, and enhances sensitivity to underrepresented fault classes by class-aware technique. The adversarial framework comprises a complex convolutional feature extractor and a domain energy discriminator to facilitate cross-domain feature adaptation. Class attention mechanism and class-overlap optimization loss dynamically adjust the focus on imbalanced classes. Moreover, joint domain alignment mechanism minimizes distributional divergence between different domains to ensure consistent feature representation. Further, JCAN integrates multi-source domain information for collaborative decision, where a soft selection-based decision fusion strategy evaluates the source domain contributions, soft attenuating low-contribution sources during information fusion. Experiments on the Paderborn University (PU) and Mechanical Comprehensive Diagnosis Simulation Platform (MCDSP) bearing datasets validate the effectiveness of the proposed JCAN in fault diagnosis tasks under class imbalance and variable operating conditions, as well as the outperformance compared to advanced methods.