Diffusion-based hybrid virtual-physical fault diagnosis method for rolling bearings lacking available fault samples

Lacking available fault data remains to be a major obstacle for applying intelligent bearing fault diagnosis in engineering practice. To address this problem, this paper proposes a diffusion-based hybrid virtual-physical fault diagnosis method. The diffusion model is first pre-trained using experimental healthy data to learn the baseline characteristics of actual data. Then, guided by simulation data from a phenomenological signal model, a conditional diffusion process is then conducted to generate experiment-like synthetic data. During this process, a hybrid virtual-physical time-frequency loss (HVP-TF loss) is specifically designed to ensure that the conditional diffusion process preserves the pre-learned baseline features of experimental data while replicating fault characteristics from the simulated data. Finally, statistical features are extracted from both the generated synthetic data and experimental healthy data to train lightweight classification algorithms, thereby establishing the hybrid virtual-physical fault diagnosis model. The developed model demonstrates superior diagnostic performance on fault experimental data from our self-built test rig.