Incipient fault diagnosis for high-speed train traction systems via improved LSTM

Abstract

The incipient fault diagnosis of high-speed train traction systemsis a fundamental but essential task to guarantee the safety and reliability of high-speed trains.Traditional intelligent data-driven methods require comprehensive expertiseand cannot diagnose incipient faults in the early period.This paper focuses on the fault diagnosis approach for incipient time-varying faults and unknown faultsusing sensor signals.Considering that the sequential signal has non-stationary characteristicsand the incipient faults are slow time-varying,an improved LSTM (long short-term memory) network that isan unsupervised recurrent neural network, is developed.Combined with the GRU (gated recurrent unit) decoder,an LSTM-based encoder-decoder model is proposed to extract the feature vectors,which has the robustness to non-stationary characteristics.Then, the t-SNE (t-distributed stochastic neighbor embedding) approach is used to reduce the dimension of extracted feature vectors,and the DBSCAN (density-based spatial clustering of applications with noise)is utilized for clustering and recognizing faults to realize the unsupervised fault diagnosis.The proposed approach is verified using the data from a semi-physical platformof the CRH$_2$ traction system, in which the known and unknown faults are illustrated.The results show that the proposed unsupervised approachcan achieve a fault diagnosis rate of over 95%.