BFD-DDPM: Denoising Diffusion Probabilistic Models for Bearing Fault Diagnosis

Rotating machinery fault diagnosis is of paramount importance in industrial production. In this field, the proposed generation model breaks through the restriction of machine learning on the completeness of sample set in the process of model training. While Generative Adversarial Networks (GANs) have shown potential in this domain, their susceptibility to training collapse due to stringent parameter selection remains a concern. In addressing these challenges, we propose BFD-DDPM, a novel bearing fault diagnosis grounded in the principles of unsupervised learning via the diffusion model. Utilizing the Denoising Diffusion Probabilistic Models (DDPM) - an enhanced generative model, normal samples of mechanical vibration signals are studied in an unsupervised manner to capture their prevalent distribution in a latent space. Subsequently, equipment fault diagnosis is achieved by locating the nearest generated data of test samples within this space via denoising diffusion, utilizing the Manhattan distance metric. Compared to unsupervised learning-based fault diagnosis methods, this paper demonstrates the advantages of DDPM in terms of accuracy and the robustness of parameter selection.