Contrastive Regularization-Based Domain Generalization for Rolling Bearing Fault Diagnosis under Unseen Rotational-speed Conditions

Abstract

To address the performance degradation of rolling bearing fault diagnosis under varying operating conditions, where data from unseen conditions are unavailable during training, a contrastive regularization-based domain generalization method is proposed. First, omni-dimensional dynamic convolution is employed to adaptively characterize vibration-feature variations across multiple source domains and extract discriminative fault representations. Second, class-specific feature perturbation is introduced to enrich intra-class variations, and a class-conditional domain-discrepancy-aware interpolation mechanism is developed to quantify the feature shifts of different fault classes across source domains and adaptively regulate the Beta interpolation distribution, thereby expanding the class-conditional feature space according to actual cross-condition variations. Subsequently, self-supervised contrastive regularization is imposed at both the feature and classifier-output levels to enhance the consistency of same-class representations and classification responses, while the classification objective preserves inter-class discriminability. Finally, experiments on rolling bearing domain generalization tasks demonstrate that the proposed method achieves an average diagnostic accuracy of 97.53%, indicating its effectiveness for fault diagnosis under unseen rotational-speed conditions.

Conflict of Interest Statement

The authors declare no conflicts of interest.