Study On Ballastless Track Subgrade Settlement Prediction

Along with the national high-speed railway large-scale construction and therunning speed of the railway big increases in, ballastless track subgradesettlement has become the influence the running speed of the high speed railway,operation life and safety of main factors.Embankment settlement by climate,geology, as well as influence of various factors, the settlement is very complex.at the same time settlement standard is very strict. The current prediction andevaluation method is mainly to legal parameters. Parameters on the assumptionthat the legal is to settlement regularity, and on the basis of known to fixedsample data according to the parameters of the model are fitting model.Therefore, sample study ability and the model generalization ability is bad, evencan appear fitting failure to lead the spread data.In this paper the ballastless track subgrade settlement principle and routinealgorithm was firstly analyzed and then introduced the neural network algorithm.Neural network algorithm as artificial intelligence method of the most matureand most widely algorithm, it has good nonlinear mapping ability, can be verygood to approach the function of any order. Meanwhile for neural networklearning speed is slow, easily into local minimum value disadvantagesimprovement; And then the BP neural network model and ballastless tracksubgrade settlement to together, establish the previous four groups settlement forfour dimensional attribute value after a group of sedimentation value for theresult of the input/output several model, determines the corresponding nodenumber of each layer, and then through the measured data validation analysis,greatly improving the prediction model of learning ability, but generalizationability is still very weak.Support vector machine is based on VC of statistical learning theory. Astructural risk minimization algorithm, in a small sample data to establish the very good nonlinear mapping model, in predictive control aspects are of veryhigh stability and robustness. At last of this paper using SVM (support vectormachine) establish the same input/output model, through the analysis of the samedata to greatly validate and improve the model later prediction ability.Finallythrough the two groups of measured data all prediction model is compared, andthe results show that neural network is very good solution to the conventionalalgorithm learning ability difference characteristics. Support vector machine inimproving the ability of the sample, meanwhile, very good solution to the neuralnetwork relying too much on experience of the defect, greatly improving themodel generalization ability.