Fault Feature Transfer and Recognition of Bearings Under Multi-condition

Abstract

Although increasingly being applied, the deep learning fault diagnosis method is hard to deliver high-precision accuracy due to sample size limit under practical working conditions. To solve this problem, a new bearing fault diagnosis method is proposed. Firstly, converted the original vibration signals of 10 types of states into two-dimensional image data and proportionally divided into three different datasets. Then, the structure of ResNet34 network is improved to extract the weight parameters pre-trained on ImageNet dataset, and the feature parameters are transferred to the Case Western Reserve University (CWRU) dataset by transfer learning to process the small sample fault diagnosis. Finally, the fault diagnosis model is obtained by repeatedly adjusting the hyperparameter training. Compared with the experimental results of several other methods, the final accuracy of this improved method can reach 99.21%. The test results under different working conditions also demonstrate that this transfer learning method has higher identification accuracy than the existing methods and can meet the requirements of fault diagnosis in actual industrial production.