A bearing fault data augmentation method based on hybrid-diversity loss diffusion model and parameter transfer

Abstract

The fault diagnosis of mechanical equipment can prevent potential mechanical failures, avoid property damage and personal injury, and ensure the stable and safe operation of mechanical equipment. Data driven is an important aspect of intelligent fault diagnosis. When data is scarce, it can seriously affect the accuracy of fault diagnosis and make it difficult to ensure the smooth and safe operation of machinery. Faced with this challenge, a classifier-free guidance diffusion model combining hybrid loss and diversity loss (CFGDMHD) is proposed for data augmentation of fault samples. This new data augmentation method generates samples with the same data distribution as real samples from random noise through diffusion process. CFGDMHD can generate multi-class samples simultaneously without the need for additional classifier guidance in the joint training of unconditional diffusion models and conditional diffusion models. This work proposes diversity loss to improve the diversity of generated samples. We conducted experiments using a bearing dataset. The results indicate that the sample quality and diversity generated by this method are excellent, which can help improve the accuracy of fault diagnosis and ensure the safe operation of mechanical systems.