DGNSCA: A Domain Generalized Network Leveraging Similarity and Class-Conditional Contrast for Robust Fault Diagnosis of Rotating Machinery Under Unseen Operating Conditions

Timely and accurate detection of bearing faults is critical to the reliability of rotating machinery, yet models trained under fixed conditions often fail on unseen domains owing to distribution shifts. To overcome this limitation, we develop DGNSCA—a Domain Generalization Network that leverages Similarity guidance, Class-conditional feature alignment, and Adversarial training to learn domain-agnostic yet class-discriminative features. The architecture is built on a 1D convolutional neural network and incorporates three key modules: global distribution alignment to extract domain-invariant features, class-conditional contrastive learning to enhance intra-class consistency across domains, and adversarial training to improve generalization. DGNSCA further introduces a domain-similarity-guided decision module that adaptively aggregates per-source predictions via similarity weights, enabling robust inference without access to target-domain data. Comprehensive evaluations on two widely recognized benchmark datasets show that DGNSCA consistently surpasses current leading methods regarding accuracy, precision, and F1-score. Ablation analyses further validate the distinct contribution of each module within the framework. These results indicate that DGNSCA provides a reliable and scalable solution for intelligent fault diagnosis under realistic, distribution-shifted operating conditions.