Periodic Impact Feature Enhanced Diffusion Model for Rolling Bearing Fault Diagnosis Under Small Sample Conditions

Rolling bearing fault diagnosis fundamentally depends on the extraction of periodic impact features. However, accurate extraction of periodic impact features requires numerous high-quality fault samples, which are difficult to obtain in real industrial environments, significantly limiting diagnostic accuracy. To overcome this challenge under small sample conditions, a periodic impact feature enhanced diffusion model (PIFE-DM) is proposed for the fault diagnosis of rolling bearings. The proposed model integrates a multi-scale residual dilated module and a densely connected module to enhance its sensitivity to the periodic impact features of the fault signals. Additionally, a triple loss function is proposed, which includes the time-domain coherence loss, the impact feature loss, and the dynamic response loss, and effectively preserves the periodic impact features of the fault signals. Experimental validations on simulated signals, CWRU benchmark datasets, and bearing test rig datasets indicate that the PIFE-DM model accurately extracts the periodic impact features from small fault samples, generates high-quality fault signals, and enable effective diagnosis of bearing faults under small sample conditions. Specifically, the PIFE-DM model achieves diagnostic accuracies of 97.00% and 96.33% on the real datasets, the improvements of 4.7% and 6.33% over the original small sample datasets. These results outperform the existing data augmentation methods and highlight the potential of the PIFE-DM model in diagnosing faults under small sample conditions.