Data-driven Detection And Diagnosis Of Incipient Faults In Traction Systems Of High-speed Trains

Benefitted from their inherent characteristics such as the comfort,high efficiency,and convenience,high-speed trains have been one of the most popular transportation means among multiple cities in China.Traction systems are regarded as the core part of high-speed trains whose reliability is critically important to ensure the whole trains to operate in safety.As the serving time of high-speed trains increases,it is inevitable that the aging components will induce incipient faults such that the performance together with the reliability and safety of trains will be negatively affected.Early detection and diagnosis of incipient faults in high-speed trains can buy more time for operators to take remedial actions and thereby will improve the safety of trains.Since high-speed trains are typically nonlinear and nonGaussian systems,it is difficult to establish mathematical expressions by means of the first principles especially when externally time-varying disturbances and noises attend.Fortunately,there are huge amount of real-time data acquired from running trains,which contains sufficient information such as the operating conditions.Therefore,based on the off-line and online data from trains,this dissertation focuses on data-driven detection and diagnosis of faults(FDD)techniques and on emphasis of practical applications to high-speed trains.This dissertation begins with a thorough survey of FDD methods for high-speed trains.From the theoretical aspect,these FDD methods can be generally categorized into three classes: signal analysisbased,model-based and data-driven methods.In addition,some recent research together with the pros and cons of these FDD methods are then followed.It is well known that,the core of the design of data-driven FDD methods includes data modeling and construction of test statistics.Therefore,this dissertation will develop multiple FDD strategies from two above aspects.From the data analysis and modeling aspect,this dissertation will propose improved FDD methods based on local information,fault information,as well as systemic and noise information hidden in original data collected from traction systems.In addition,this dissertation will also design two new test statistics based on the expected FDD performance.One is to design a test statistic of great sensitiveness to incipient faults;and another is to develop a test statistic which is not only sensitive to incipient faults but also robust to unknown noises.Salient novelties and contributions of the developed FDD methods in this dissertation cover the following items:(i)The first focus is,based on the switched characteristics of traction systems,on discovering the dissimilarity of data in each mode and then proposes the multi-mode principal component analysis(PCA)-based and multi-mode kernel PCA-based FDD methods.(ii)The second aim,by sufficiently exploiting the distribution information of faults in both detection and diagnosis phases,is to develop an incipient FDD method for high-speed trains.By using a variationally weighted matrix related to data,the proposed scheme belongs to nonlinear projection methods but has the high-computation ability like linear ones.(iii)The third attempt is to design a deep PCA method to achieve improved fault detectability for traction systems.This developed scheme is dedicated to exploring multiple data processing layers to extract more accurate signal features.Furthermore,the sufficient and necessary conditions of the proposed scheme,together with the feasibility for traction systems,are given via theoretical derivations.(iv)Based on discussions on the commonly used test statistics for FD methods,the fourth work is to develop a PCA and Kullback-Leibler divergence(KLD)-based scheme to overcome the drawbacks of the standard approach.The associated investigations also cover construction of a symmetric evaluation function in single and multivariate cases,the derivation of an uniform form of these evaluation functions,transforming non-Gaussian signals of traction systems into Gaussian ones,etc.(v)The final attention is paid to the design of a new robust method based on PCA and Hellinger distance to achieve FDD tasks for traction systems in high-speed trains.This study is dedicated to developing the Hellinger distance-based test statistics which is of high robustness to unknown uncertainties and of considerable fault detectability.In addition,the satisfactory advantages of the proposed scheme are strictly proven by theoretical derivations and also illustrated by experimental analysis.All proposed methods in this dissertation have been tested on traction systems,and experimental results have illustrated their effectiveness.It is of great confidence to affirm that,this dissertation expands the application scope of data-driven FDD techniques on the one hand,and provides researchers and practitioners with informative guidance on the other hand.