Rolling Bearing Fault Severity Recognition via Data Mining Integrated With Convolutional Neural Network

Rolling bearing vibration signals exhibit typically complex modulation characteristics, and usually present nonstationary features. The defect of a rolling bearing is mainly manifested by impulses carried by resonance vibration, and the resonance frequency is independent of the operation conditions. Recent studies have correlated the characteristics of impulses with the fault severity of rolling bearings. However, the impulses are extracted manually, and the fault severity is evaluated by manually analyzing the target impulses or the matched atoms. This paper takes advantage of impulses, and proposes a novel intelligent rolling bearing fault severity recognition method. The method includes two modules, i.e., impulse mining and fault recognition. Recently, matrix profile (MP) has emerged as a promising method of mining the motifs in a time-domain signal. In the first module, MP is firstly introduced to the field of fault diagnosis, and it is conducted to mine the impulses from vibration signals. In the second module, convolutional neural network (CNN) combined with softmax regression is applied to automatically extract the discriminatory features from the mined impulses and accomplish the fault severity recognition. The proposed method is evaluated by the lab experimental bearing signals, and further validated by the signals collected from a wind turbine simulator with faulty high-speed shaft roller bearing. The results demonstrate that the proposed method can better recognize the fault severities of rolling bearings than the comparison methods.