Prediction Models For Railway Track Geometry Degradation Using Machine Learning Methods

The dramatically increasing demand for rail transport brings challenges to track maintenance on railways.To improve maintenance efficiency,numerous scholars have studied track geometry prediction models based on machine learning methods.However,the applicability and generality of these models are still unclear,which affects their wide application in the railway industry.Therefore,this thesis compares and analyses several popular machine learning-based track geometry prediction models,including Support Vector Machine(SVM),Grey Model(GM)and Deep Neural Network(DNN),to provide a reference for railway engineers to choose a reasonable prediction model.On this basis,the machine learning models with good prediction performance are further improved to enhance their prediction accuracy and lay the foundation for subsequent practical applications.The following research work and related findings are carried out in this thesis:(1)Analysis and evaluation of track geometry dataVarious statistical methods are used to analyze and evaluate the track longitudinal level of a railway line.The results show that there is little difference between the peak high and low unevenness at the same wavelength for the left and right tracks;there are different degradation cycles for the line and that the track geometry degradation increases approximately linearly over the different cycles;there is variability in the degradation of the different track sections.(2)Comparison analysis among machine learning-based prediction models for railway track geometrySeveral popular machine learning-based track geometry prediction models are compared and analyzed,including DNN,GM and SVM.The results show that the DNN model has high prediction accuracy for both the track longitudinal level(continuous functions)and standard deviation of the track longitudinal level(discrete functions);the GM(1,1)model has moderate prediction accuracy for both the track longitudinal level and standard deviation of the track longitudinal level;the SVM model has high prediction accuracy for the standard deviation of the track longitudinal level but has poor prediction performance for the track longitudinal level.(3)Prediction models for railway track geometry based on an improved DNNTo overcome the shortcomings of traditional DNN models,this thesis uses the DNN models optimized by a random search algorithm to predict the different wavelengths track longitudinal level.The results show that the short-term prediction accuracy of the optimised DNN model for the track longitudinal level is higher than the confidence interval of 95%,which is higher than the traditional DNN model;the long-term prediction for the track longitudinal level shows that the optimised DNN model can correctly predict the growth pattern and the location characteristics of track geometry,and the maintenance effect can be considered by the model.This thesis presents a comparative analysis of track geometry prediction models based on SVM,GM and DNN and an optimised DNN prediction model with high prediction accuracy is developed,which helps to promote the development of track geometry prediction models and has certain theoretical significance and application value for the guidance of track maintenance.