Just-in-time Learning-aided Nonlinear Fault Detection for Traction Systems of High-speed Trains

Traction systems in high-speed trains exhibit significant dynamic characteristics, which mainly arise from operation-point changes. Most existing fault detection methods provide static data models for global structures, especially for traditional multivariate statistical analysis methods constrained by constant operating points. The symptoms of incipient faults are slight and easily hidden. Despite the moderate effect of incipient faults, they will compromise the overall performance and remaining life of traction systems in the long run. Therefore, a just-in-time slow feature analysis method is proposed in this study. The salient advantages of the proposed method are: 1) It can be applied to dynamic non-linear systems; 2) It can detect incipient faults subject to environments containing noise and unknown disturbances; 3) It mitigates false alarms caused by parameter mutation during mode-switching. A series of experiments are carried out on a traction system platform to verify the effectiveness and superiority of the proposed method.