Data-driven Health Assessment And Prediction For Railway Vehicle Door System

Railway Vehicle Door System(RVDS)is one of the most important subsystems in passenger compartments of railway vehicles.As the only path for passenger boarding,its performance and health state directly affect the vehicles' operational reliability and the passengers' safety.In the vehicles' operation,door system is often with high failure rate due to its complex mechatronic structure,frequent open and close movements,and crowded passenger flow environment.It is of great significance to improve the safety,reliability,reduce the failure rate and maintenance cost of the RVDS by accurate health assessment and fault prediction.With the development of sensoring,data storage and communication technologies,abundant data are being accumulated by the metro companies,which contain rich information on the running conditions of RVDS.Therefore,data-driven methods have become an effective means to study health assessment and fault prediction of RVDS.In this thesis,data-driven health assessment and prediction have been carried out for a typical sliding plug door system.Firstly,system structure,mechanism and data acquisition of RVDS are studied.Then,the common faults are summarized,which are grouped into two types,sudden faults and gradual faults.The main causes of these common faults are also analyzed.Secondly,a health assessment method based on feature fusion has been proposed for degradation assessment of the driving screw with poor lubrication.Multiple features in time or frequency domains are extracted from the online measurements of the rotation angle/position,speed and current of the drive motor in RVDS.These high-dimensional multi-domain features are filtered by Relief algorithm.Then,Principal Component Analysis(PCA)is used to eliminate the redundant information among the degradation features.A Health Index(HI)is finally constructed by Kernel smoothing.The feasibility and accuracy of the proposed method are verified on a test bench of RVDS.Thirdly,a similarity evaluation based Remaining Useful Life(RUL)prediction method has been developed.The time series of Health Index is taken as the input variable,and the uncertainty of the initial health state is transformed into the time delay problem of HI sequence fragments in the similarity evaluation,and then a density weighting method for model aggregation based on kernel density estimation is proposed.The weighted RUL of similar historical time series is taken as RUL of the current sample.The validity of the developed method is verified by bearing data set,and its prediction performance is higher than several existing classical methods.Finally,a health assessment method via JS divergence has been proposed based on the on-line operational RVDS with high uncertainty data.The three parameters of driving motor are reconstructed.The motor's angle is regarded as the independent variable and the motor's speed is regarded as the response variable,and the statistic distribution parameters of the speed variable with a fixed motor angle are defined as random variables.Based on these random variables,the entropy based JS divergence is introduced to measure the distance between the current statistic distribution and the referenced one.Then the distance is transformed into a Health Index which can represent the health state of RVDS.The case study on Guangzhou Line 5 can show that the developed method is feasible and effective.