Research On Health Assessment And Prediction Method Of Switch Equipment For Speed-Up Turnouts

The modern railway industry is developing towards intelligence、networking、comfort、convenience and the requirements for various railway equipment are becoming higher and higher.Among them,switch equipment,as key equipment for switching switches and changing the direction of train operation,its health status affects the safety and efficiency of train operation.At the railway site,the judgment of switch equipment’s health status mainly relies on on-site staff’s judgment on data such as switch power and current.This method is highly subjective and inaccurate,and the focus of attention on switches is almost always on the post-fault diagnosis.Insufficient attention to pre-failure health status makes it difficult to form an effective equipment maintenance plan.To address the above issues,this article takes the ZYJ7 switch machine as the research object and proposes the assessment and prediction of the health status of switch equipment before failure.Using the Holder coefficient method to screen the time-domain features of the power curve of the switch machine,Self Organizing Map(SOM)neural network is used to extract the Health Index(HI)of the switch equipment,and the Fisher algorithm is used to divide the health stages of the switch machine.Finally,the e Xtreme Gradient Boosting(XGBoost)method is used to predict the state of the switch machine.The main research work of this thesis is as follows:(1)Feature parameter selection.A feature parameter screening method based on the Holder coefficient method is proposed to address the issue of minor differences in features and multiple feature parameters in different degraded states of switch machines.Firstly,preprocess the power curve data of the switch machine,divide the power curve into five stages,extract the time-domain features of each stage,and then use the Holder coefficient method to calculate the correlation between each feature parameter and the actual health status of the switch equipment.Finally,select the feature parameter set with the highest impact value.Compared with traditional feature parameter selection methods,the Holder coefficient method is more suitable for high-dimensional data.It can mine and filter feature parameters that are closely related to the health status of switch machine,constructing a feature parameter space for switch machines.(2)Health status assessment.In response to the issue of inaccurate manual evaluation of the status of switch equipment,by analyzing the daily work data of switch machine,SOM algorithm is used to calculate the health factors that can characterize the health status of switch equipment,and the health status of the equipment is digitized.Verify the effectiveness of the SOM neural network algorithm using actual power curve data from a certain railway bureau site.The experimental results show that compared with other algorithms,the SOM equipment condition evaluation model completed by data training shows significant improvement in trend and monotonicity,reaching 0.848 and 0.684,respectively,which can effectively track the health status of switch equipment.(3)Health stage division.In response to the issue of not accurately corresponding to each health stage of switch equipment on site,measures cannot be taken in a timely manner,using the calculated health factor curve as the object,the Fisher optimal segmentation algorithm is used to partition the health stages of switch equipment.Obtain the optimal health order of 3,complete the scientific division of the health status of switch equipment,and provide health factor intervals and segmentation thresholds for each health stage,providing data support for accurately corresponding health status of switch equipment.(4)Prediction of the development trend of health status.In response to the current problem of low prediction accuracy and limited lifespan prediction for switch equipment,this study uses the classic machine learning algorithm Extreme Gradient Boosting(XGBoost)to analyze the state prediction of switch equipment.Firstly,model parameters are determined based on the health factor data characteristics of switch equipment,and a prediction model structure is constructed.Secondly,health factors are used as input to predict subsequent health factor data.Compared with the random forest(RF)algorithm,Gradient Boost Decision Tree(GBDT)algorithm,and Long Short Term Memory(LSTM)algorithm,the XGBoost algorithm has better performance in predicting the health factors of switch equipment.The average relative error of prediction results is 0.193,the root mean square error is 0.0413,and the absolute coefficient is 0.703.It can achieve early detection of equipment degradation information in the early stages of minor degradation and scientifically assist equipment management personnel in making maintenance decisions for switch machine.