Bearing Fault Diagnosis Based on Diffusion Model and One-Class Support Vector Machine

Abstract

Fault diagnosis is of great importance for reducing economic losses and ensuring the safety of equipment. As an important part of industrial machine, it is necessary to perform fault diagnosis on bearings. However, in reality, normal data are often more abundant than fault data, making it challenging to recognize faults. To address this issue, the diffusion model with u-net has been introduced, for its excellent feature extraction ability. In this paper, the feature knowledge extracted by diffusion model is sent into one-class support vector machine (OCSVM) for anomaly detection instead of using raw data. Firstly, train the diffusion model with normal data, and then take out the trained encoder in u-net. Next, input test data which includes both normal data and abnormal data into the trained encoder to extract feature knowledge. Finally, train OCSVM with normal data and then send the extracted feature knowledge into trained OCSVM for detection. Compared to directly using raw data for fault diagnosis, the proposed method achieves superior accuracy.