A Physics-Conditioned Denoising Diffusion Probabilistic Model for Vibration-Based Wheel Flat Diagnosis Under Variable-Speed Conditions

Wheel flats deteriorate normal wheel-rail contact conditions and induce periodic impacts, thereby affecting the operational stability and reliability of railway vehicles. In practical engineering applications, the number of fault samples available for diagnosis is usually limited. Moreover, under variable-speed conditions, fault-related time-frequency characteristics in vehicle vibration responses vary significantly with operating speed, which further increases the difficulty of constructing labeled training datasets. To address these issues, a physics-conditioned denoising diffusion probabilistic model (DDPM)-based data augmentation method is proposed in this study. Impact occurrence timing and peak amplitude are extracted from vibration responses as response-derived physical conditions to characterize the temporal distribution of wheel-rail impacts and the variation in impact intensity. Subsequently, the physical conditions are transformed into two-dimensional conditional feature maps and introduced into the diffusion generation process, enabling the generation model to learn the evolution patterns of fault-related time-frequency characteristics caused by operating-condition variations. The proposed method is validated using a vehicle-track dynamic simulation dataset and a scaled vehicle-track test rig dataset. Compared with the original datasets, the proposed method improves the diagnosis accuracy from 87.8% to 94.4% on the simulation dataset and from 69.2% to 81.2% on the scaled vehicle-track test rig dataset. These results indicate that the proposed physics-conditioned DDPM can generate samples that better preserve impact-related time-frequency characteristics and improve wheel flat diagnosis performance under limited fault samples.