Condition Monitoring And Fault Prediction Of Railway Turnout

Railway turnout is one of the indispensable important equipment in the railway signal system.It is the key equipment to change the direction of train operation and ensure the stable operation of the train.It is also a signal device that is prone to failure.Once a failure occurs,it often causes problems such as train delays and even derailment.With the increasingly busy railway transportation business,in order to ensure the operation efficiency,the reliability of turnout system becomes more and more prominent.At present,the main maintenance mode of turnout equipment in China are“planned repair”and“on-condition repair”.The maintenance work consumes huge energy of personnel,and many problems such as missed inspections are often arised in the maintenance process.Fault prediction and health management technology realizes the equipment condition perception and health condition monitoring.Through the analysis of monitoring data,it can predict the time point of fault occurrence,which greatly improves the maintenance efficiency.It is the upgrading and development of the previous"planned repair"and also the development direction of equipment management in the future.Under the guidance of PHM maintenance management concept,this thesis aims to realize the automatic monitoring of the operation condition of turnout,and the prediction of the remaining useful life of turnout on some mechanical predictable faults,so as to further the development of"on-condition repair".In order to achieve the above goals,this thesis proposes a turnout condition monitoring method based on kernel principal component analysis(KPCA).Based on the feature extraction of on-site historical current and power data,Hotelling-T~2and square prediction error index are used to measure from the inside and outside of the monitoring data main component space to realize the monitoring of turnout equipment.For predictable mechanical fault,feature fusion and genetic programming methods are used to predict the remaining useful life of the turnout system.Finally,the proposed scheme is compared with other mainstream algorithms to verify its superiority.The main work of the thesis can be summarized as follows:(1)In this thesis,the whole turnout conversion process is divided into four parts by the way of signal segmentation.As for the method,the idea of signal segmentation based on gradient is introduced into the previous time-based signal segmentation method,which makes the segmentation result more accurate.Considering the diversity of turnout monit-oring signals,this thesis not only uses power data but also adds current signals as the data support in the condition monitoring process.According to the above,a KPCA-based turnout condition monitoring method was proposed.(2)This thesis proposes a turnout remaining useful life prediction scheme based on feature fusion and genetic programming.Considering that a single monitoring signal feature only contains part of the turnout degradation information,we rank the features based on their importance after extracting characteristics,and use the variant correlation-based feature selection method to select the optimal degradation feature set.Then auto associative kernel regression method is used to fuse it,and finally the remaining useful life prediction is achieved by genetic programming algorithm.(3)In this thesis,a condition monitoring and fault prediction system for turnouts is designed and tested.C#is used to complete data management,condition monitoring and other interface display work.Finally,we test the proposed condition monitoring scheme by using the historical monitoring data of two groups of turnouts in Changsha electric power section of Guangzhou Railway Group.The true positive rates are greater than 85%(positive if abnormal conditions are found),and the rate of missed alarms is less than 10%.For the verification of the fault prediction scheme,this thesis uses Turkish scholars'open data set for turnout monitoring and Changsha electric power section field data set to show that the prediction scheme has good applicability.The results show that the mean absolute errors on the two data sets are 4.1118 and 14.9299 respectively,which prove that the method has good prediction performance and meets the requirements of“on-condition repair”.Figure 54,Table 11,References 74...