A Simulated-to-Real Transfer Fault Diagnosis Method Based on Prototype Clustering Subdomain Adversarial Adaptation Network for HST Bogie Bearing

In practical high-speed train (HST) fault diagnosis, obtaining labeled data is challenging. Simulated data can provide labeled information for diagnosing unlabeled data. However, the disparity between simulated and real domains poses challenges for transfer learning. Furthermore, the neglect of fine-grained domain matching will impede the alignment of subdomains. To address these issues, a simulated-to-real transfer fault diagnosis method is proposed. A simulated and real data fusion (SRF) method is developed to enhance the similarity between simulated and real data by integrating machine-specific characteristics with the inherent features of fault bearings. Based on the fused data, a prototype clustering subdomain adversarial adaptation network (PCSAN) is developed to realize transfer diagnosis. PCSAN utilizes prototype clustering to obtain representative subdomain prototypes and introduces multikernel maximum mean discrepancy (MK-MMD) to achieve adaptation between similar prototypes in the simulated and real domains. Finally, a joint objective function is proposed to accelerate model convergence and achieve alignment of conditional and marginal distributions. The proposed method is validated using a public bearing dataset and a self-build HST bogie bearing dataset. Compared to other advanced methods, this approach demonstrates superiority in transfer diagnosis from simulated domains to real domains.