Significance Support Vector Machine for High-Speed Train Bearing Fault Diagnosis

Abstract

Wheelset bearing is the most critical component in High-Speed Train (HST) and crucial for HST safe and efficient operation. As a wide applying method for bearing health monitoring, Support Vector Machine (SVM) may easily lead to overfitting due to giving the same attention to all samples especially when training dataset contains outliers. For measuring the sample importance corresponding with normal or outlier in feature space of SVM training process, sample significance is represented by utilizing sample spatial aggregation information of feature space in which the normal samples having a good aggregation and outliers caused by noise or uncertainty tending to deviate from the sample aggregation group. Then by introducing sample significance, an improved Significance SVM (SSVM) is proposed for enhancing the robustness of SVM with noisy samples or uncertainty. SSVM overcomes the drawback of SVM and gives outlier samples less attention during model training by assigning significant coefficients to samples in the model training process. Combing SSVM, multi-domain features and feature selection, a noise robustness fault diagnosis method is successfully applied on HST bearing vibration dataset. The experiment results verify the effectiveness and the stability of the proposed method under different noise levels.