A denoising diffusion probabilistic model-based fault sample generation approach for imbalanced intelligent fault diagnosis

Abstract In practical industrial applications, the data imbalance problem caused by limited fault sample size can seriously affect the fault diagnosis accuracy, which will decrease the equipment reliability and safety. One existing solution is to augment the limited fault samples using generative models, with the most extensively utilized being generative adversarial networks (GANs). However, GAN is prone to mode collapse during training. Hence, this paper proposes a denoising diffusion probabilistic model (DDPM) based sample generation approach to augment fault samples for imbalanced intelligent fault diagnosis (IIFD). DDPM adds noise to the image in the diffusion process and eliminates the noise in the reverse process, so as to expand the fault samples ultimately, avoiding the mode collapse problem of GAN. In addition, an evaluation indicator framework is proposed to assess the generated sample quality. Three experiment cases are implemented to verify the proposed approach. Experimental results demonstrate that the sample quality generated by DDPM is much superior to that of other approaches. This approach can overcome the problem of unsatisfactory accuracy in IIFD effectively.