Decoupling Global and Local Alignments: Dual-Stream Collaborative Regularization for Single-Source Cross-Condition Fault Diagnosis

Single-source cross-condition fault diagnosis faces significant challenges due to drastic distribution shifts caused by fluctuating operating conditions. This issue becomes more severe when a single model must simultaneously generalize across global structural shifts and fine-grained class-wise patterns. Existing methods typically rely on a single-stream architecture, which suffers from an inherent optimization conflict between domain invariance and class discriminability, often combined with unreliable pseudolabels. To address these limitations, this study proposes a Dual-Stream Collaborative Regularization Adaptation Network (DS-CRAN). The proposed framework explicitly decouples global marginal alignment and class-conditional alignment into two parallel streams, thereby reducing gradient conflicts. Furthermore, a synergistic regularization mechanism is introduced to enhance representation quality. A mask-based consistency constraint enforces robustness against local signal disturbances, while a class-conditioned reconstruction task prevents feature collapse and preserves semantic information despite label noise. Extensive experiments on bearing and gear datasets, particularly under large speed and load variations, demonstrate that DS-CRAN consistently outperforms state-of-the-art methods. The results validate the effectiveness of the decoupling strategy and collaborative regularization in achieving resilient diagnosis.