Dendrite-Network-Assisted Multi-Source Domain Generalization for Fault Diagnosis in Rotating Machinery

Abstract

Intelligent fault diagnosis is crucial for the safe, efficient operation of rotating machinery. Domain adaptation models can mitigate distribution shifts from diverse operating conditions, but their over-reliance on annotated target domain data limits their use under unknown and dynamically changing conditions. To tackle this matter, DDDG, a dendrite-network-assisted multi-source domain generalization framework for fault diagnosis in rotating machinery, is proposed. First, a network integrating a dendrite module is designed to learn logical feature interactions and enhance nonlinear expressiveness. Second, a complementary global-specific representation learning mechanism is developed. Class-aware entropic optimal transport performs soft distribution-level alignment to reduce excessive inter-source discrepancies, while domain-specificity-preserving contrastive learning maintains a relative local similarity hierarchy within each fault category, thereby preventing complete representation collapse across operating conditions. Third, source-specific classifiers are adaptively combined using normalized relative gating coefficients generated by an independent domain descriptor, enabling sample-dependent soft coordination among the available source experts. Extensive experiments were conducted on public bearing datasets, and the results demonstrate that DDDG achieves high accuracy under unseen operating conditions, exhibiting strong generalizability.