Time spectrum repaint domain-wise diffusion model for bearing fault diagnosis under time-varying conditions in...

Fault diagnosis of axle box bearings is crucial to the safety of high-speed trains. However, the diagnostic accuracy remains suboptimal due to the complex and variable operational conditions during train service, coupled with the scarcity of fault samples resulting from conservative maintenance strategies. To overcome these limitations and improve diagnostic capabilities, this paper proposes the Time Spectrum Repaint Domain-wise Diffusion (TSR-DD) method, which improves the generalization capability of transfer learning via short-term segment extraction of time-varying operational condition data to construct equivalent steady-state condition datasets. The core innovation of this method lies in its diffusion model trained on stable data that is specifically developed to enable a short-term completion process for variable-speed time-frequency images. Through this process, the proposed diffusion model masks and extends short-term variable-speed monitoring data to generate equivalent approximate data. Building upon these generated datasets, a robust speed-generalization diagnosis method is introduced, which integrates domain generalization, thereby enabling accurate diagnosis of arbitrary speed data trained with limited and continuous monitoring data. The experimental validation using real axle bearing data from high-speed trains confirms the effectiveness and superiority of the TSR-DD method, demonstrating its high accuracy and stability in diagnosing high-speed train bearing faults.