A Condition-Disentangled Representation Optimization Framework for Domain Generalization Under Distribution Shift: Application to Fault Diagnosis

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Domain generalization under distribution shift remains a fundamental challenge in representation learning, where predictive models trained on multiple source domains are expected to generalize to previously unseen operating conditions. For rotating machinery fault diagnosis, variations in operating conditions often introduce domain-specific characteristics that interfere with fault-related representations, resulting in significant performance degradation across unseen domains. To address this issue, this paper formulates cross-condition fault diagnosis as a condition-disentangled representation optimization problem and proposes a Condition-Disentangled Representation Optimization Framework. The proposed framework jointly optimizes fault-discriminative and condition-related representations through a time-frequency collaborative architecture. Specifically, an asymmetric orthogonal constraint is introduced to encourage feature disentanglement between fault and operating-condition representations. A class-domain prototype regularization strategy is developed to improve intra-class compactness and inter-domain consistency, while a bidirectional prototype alignment mechanism further enhances cross-domain semantic correspondence. These objectives are integrated into a unified representation optimization framework and optimized jointly using a weighted objective function. Numerical experiments and comparative analyses on multiple benchmark datasets demonstrate that the proposed framework achieves the highest average accuracy on both datasets and the best performance among the compared methods on most cross-condition transfer tasks, while remaining competitive on the remaining task. The results indicate that optimizing condition-disentangled representations effectively improves robustness, feature separability, and generalization capability under distribution shift, providing an effective optimization framework for cross-condition intelligent fault diagnosis.