A Novel Unbalanced Fault Diagnosis Method With Diffusion Model in Rotating Machinery

In terms of the fault diagnosis of rotating machinery, overcoming the fault diagnosis under data unbalance is always the key to improving diagnostic accuracy. Sample expansion is an effective way to solve the data unbalance. The generation adversarial network can effectively expand the sample by designing a generator and discriminator to generate simulation data in an adversarial way. However, the training process is complicated and the mode breakdown is easy to occur. Therefore, a data unbalanced fault diagnosis method for rotating machinery based on the diffusion model is proposed in this paper. The method mimics the physical diffusion process by constantly adding noise to the data samples in the forward process and continuously de-noising the data samples in the reverse process, to stably expand the samples and avoid the problems of instability and mode collapse in the training process. By comparing different sample expansion methods, it is proved that this method can effectively improve the diagnosis accuracy of data unbalance. At the same time, to comprehensively evaluate the quality of the extended sample, this paper puts forward a set of evaluation frameworks, and verifies the extended sample through the three indicators of MMD, PSNR and SSIM, further demonstrating the superior performance of the fault diagnosis model proposed in this paper.