A wind turbine bearing fault diagnosis method based on fused depth features in time–frequency​ domain

Abstract

Diagnosis of bearing faults has significant meaning to the maintenance of wind turbines in real industry. Well-performed bearing fault diagnosis generally requires effective features extracted from vibration signals. However, conventional methods have shortages at obtaining comprehensive information of vibration signals. Aiming at this problem, this paper proposes a feature extraction method for wind turbine bearing fault diagnosis. This method contains three useful steps to realize accuracy improvement. Firstly, the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is applied to extract time-domain feature, meanwhile fast Fourier transform (FFT) is performed twice for deep frequency-domain features extraction. Secondly, the recursive feature elimination (RFE) combined with the chi-square test is utilized to select optimal feature subset from the obtained time–frequency​ features. Thirdly, several classifiers are modeled, fed with these optimal features, and used for fault diagnosis. The industrial data from Case Western Reserve University (CWRU) and Jiangxi wind farm are taken in the experimental analysis. Numerical results show that the proposed method is effective and applicable in real wind turbine rolling bearing fault diagnosis.