A Self-Supervised Bearing Failure Detection Method with STFT and Point-wise CNN Feature Extraction

Abstract

As the most important parts, bearings are widely used in many types of mechanical and electronic equipment. Bearing failure often leads to serious consequences. Therefore, the fault detection of bearings is the key to ensuring the safe and reliable operation of equipment. At present, most feature extraction methods are based on signal processing algorithms. However, for some types of bearings, these feature extraction methods may lose some important information. The bearing vibration data is unbalanced data, which may lead to diagnostic errors. A new self-supervised bearing failure detection method with STFT point-wise CNN feature extraction and auto-encoder is proposed in this paper. Firstly, a point-wise CNN extracts the features from the signal after STFT. Then the auto-encoder compresses and reconstructs the features. Finally, the distance between the auto-encoder input and output is used to judge the signal is abnormal or not. The CWRU bearing vibration signal data set is used as test case. We also proposed the distance ratio as an evaluation index in the experiment. The proposed bearing fault detection method in this paper has a lower data requirement and better effect based on the results.