A Class-Enhanced Multi-Source Domain Adaptation Method for Rotary Machinery Fault Diagnosis

In practical industrial scenarios, rotating machinery fault diagnosis is frequently challenged by variable operating conditions and unavailable fault data, resulting in the non-identical distribution between test data and training data, and the negligence of few fault data. Therefore, a class-enhanced multi-source domain adaptation method is devised, which reduces distributional discrepancies among different operating domains through adversarial training and incorporates class-sensitive learning to boost focus on few classes, thereby obtaining more robust and diverse invariant features. Firstly, adversarial learning framework between the multi-source fault identification module and the domain energy discrimination module is built to promote cross-domain feature alignment. Wherein, the multi-source fault recognition module consists of a multi-channel complex convolutional network with strong complex feature extraction capability, and meanwhile combines the class-enhanced loss in decision space penalizes few class errors more heavily, thereby improving class boundary discrimination. Second, a quality metric-based fusion strategy is proposed that dynamically assigns source-domain weights based on their contribution to the target domain, enabling dynamic fusion for better generalization. Experiment results performed on the Paderborn University bearing dataset show that the proposed method achieves average accuracies of 95.16%, 93.08%, and 91.02% for the migration tasks at three unbalanced ratios, superior to other advanced relevant methods.