Fault Diagnosis of Rolling Bearing Based on a Priority Elimination Method

Abstract

Aiming at the fault diagnosis accuracy of rolling bearings is not high enough, and unknown faults cannot be correctly identified. A priority elimination (PE) method is proposed in this paper. First, the priority diagnosis sequence of faults was determined by comparing the ratios of the inter-class distance to the intra-class distance for all faults. Then, the model training and fault diagnosis were carried out in order of the priority sequence, and the samples of the fault that had been identified were eliminated from the data set until all faults were diagnosed. For the diagnosis model, the stacked sparse auto-encoder network (SSAE) was selected to extract the features of the vibration signal. The extreme gradient boosting algorithm (XGBoost) was chosen to identify the fault type. Finally, the method was tested and verified by experimental data and compared with classical algorithms. Research results indicate the following: (1) with the addition of PE based on SSAE-XGBoost, the fault diagnosis accuracy can be improved from 96.3% to 99.27%, which is higher than other methods; (2) for the test set with the samples of unknown faults, the diagnosis accuracy of SSAE-XGBoost with PE can reach 92.34%, which is nearly 6% higher than that without PE and is also obviously higher than other classical fault diagnosis methods with or without PE. The PE method can not only improve the diagnosis accuracy of faults but also identify unknown faults, which provides a new method and way for fault diagnosis.