Adversarial Sim2Real Transfer via Cross-Domain Coupling Networks: A Single-Source Generalization Approach for Bearing Fault Diagnosis

In recent years, domain generalization-based fault diagnosis methods have achieved impressive results. Most existing methods extract common diagnostic knowledge from multiple source domains and generalize it to unknown domains. However, in real industrial settings, a challenging scenario often arises where only a single domain is available for training, yet diagnosis is needed across many unknown domains. To address this issue, this study proposes a simulation-enhanced multicategory coupled generative adversarial network (MCCGAN) integrated with an adversarial strategy, designed for deployment under unknown operation conditions. Initially, a physical nonlinear vibration model of the bearing is developed to generate ample simulated fault samples, capturing the physical characteristics of faulty bearings across diverse operation conditions. Subsequently, to explore the common transformation relationship between simulated and real data, the MCCGAN architecture is meticulously designed. Leveraging the coupling and personality layers within MCCGAN, the model effectively captures both the shared and unique features between simulated and real data. Finally, an adversarial-based domain generalization network, augmented with conditional maximum mean discrepancy (MMD), is employed to further enhance model generalization capabilities and mitigate localized misclassifications. Extensive comparative experiments conducted on three datasets demonstrate that the proposed diagnostic model achieves high accuracies of 99.73%, 99.67%, and 99.82% in single-domain generalization diagnostic tasks, strongly validating the effectiveness and superiority of the proposed method.