Data-driven and deep learning-based detection and diagnosis of incipient faults with application to electrical traction systems

Abstract

Incipient faults in electrical drive systems will evolve into faults or failures as time goes on. Successful detection and diagnosis of incipient faults can not only improve the safety and reliability but also provide optimal maintenance instructions for electrical drive systems. In this paper, an integration strategy of data-driven and deep learning-based method is proposed to deal with incipient faults. The salient advantages of the proposed method can be summarized as: (1) The moving average technique is firstly introduced into the canonical correlation analysis (CCA) framework, which makes the new residual signals more sensitive to incipient faults than the traditional CCA-based method; (2) Based on the defined residual signals, the new test statistics cooperating closely with Kullback–Leibler divergence (KLD) are proposed from the probability viewpoint, which can greatly improve the fault detectability; (3) It is of high computational efficiency because the estimation of probability density functions of residual signals is skilly avoided; (4) Based on the new developed test statistics, the fault matrices are defined and regarded as the input of convolutional neural network (CNN) whose feature extraction ability is highly improved compared with the traditional method, which helps to accurately diagnose of incipient faults; (5) The proposed method can be implemented without any priori knowledge on system information. Theoretical analysis and three sets of experiments on a practical electrical drive system demonstrate the effectiveness of the proposed method.