Local Linear Generalized Autoencoder-Based Incipient Fault Detection for Electrical Drive Systems of High-Speed Trains

Features of incipient faults are tiny in high-speed trains’ electrical drive systems. Noises and disturbances in the external environment and sensors can mask incipient faults. Therefore, fault detection (FD) of incipient faults is a challenge. This paper proposes a new FD scheme using a novel manifold learning method named local linear generalized autoencoder (LLGAE). The prominent characteristics of the LLGAE-based FD method are three-fold: 1) it can realize FD for electric drive systems even without the physical model or expertise; 2) it still has good results for non-Gaussian electrical drives; 3) it entirely takes into account the locally linear structure of samples. Mathematical derivations have proved the proposed method. Through an experimental platform of high-speed trains, the proposed method is validated.