Research On Subgrade Settlement Prediction Method Based On Multi-source Data In Seasonal Frozen Area

With the improvement of railway design speed,it is necessary to control the subgrade settlement more strictly to ensure the safety and stability of traffic.In the complex geological conditions and climate environment,temperature,water content and soil stress are the main factors affecting the settlement of the subgrade in the seasonal frozen area.However,in the current research based on the measured data to predict the settlement of the subgrade,there is a lack of consideration of temperature,water content,soil stress and other factors.Therefore,it is of great practical significance to study a method of subgrade settlement prediction based on multisource data to make it more stable and accurate to predict the settlement of high-speed railway subgrade.Based on the national key research and development program "Intelligent Sensing Theory and Method of Road Infrastructure(2018YFB1600200)",this paper studies the multivariable subgrade settlement prediction method based on temperature,water content and soil stress.According to the relevant monitoring data of the subgrade of Harbin-Dalian high-speed railway,The BP neural network and support vector regression(SVR)models optimized by the thought evolution method were constructed respectively,and the applicability and limitations of each model were analyzed.Aiming at the limitation of single model prediction,the multi-model fusion algorithm of Extreme Learning Machine(ELM)was proposed,which combined the advantages of two models and made up for the instability of single model prediction.The main work contents of this paper are as follows:(1)Analysis of season frozen area embankment settlement rule,according to the season frozen zone of subgrade frost heave and thawing settlement phenomenon,determine the main influence factors influencing season frozen area roadbed settlement,and analyzes each factor’s influence on the subgrade frost heaving and,for the temperature,water content and soil stress of the three variables as the input parameters to predict the theoretical basis for subgrade settlement.(2)Choosing the appropriate multi-source data of subgrade settlement prediction method,the BP neural network and support vector machine algorithm mechanism,and introduces the principle of the evolution of thinking method to optimize the BP neural network,analyses the applicability of the BP neural network and support vector machine(SVM)regression and the differences of the two,which provides the theoretical foundation for constructing the combination of the two models.(3)Preprocess the subgrade monitoring data of Harbin-Dalian high-speed railway,select the appropriate interpolation method and the number of sample sets,select the best model parameters,transfer function and training function through the simulation experiment of sample data,and construct the prediction model of BP neural network optimized by the thought evolution method;The parameters of SVR model were optimized by network search method and K-re-cross validation.The optimal kernel function was selected through the simulation experiment of sample data,and the SVM regression model was constructed.The two models are applied to the subgrade settlement prediction of engineering examples,and the effectiveness of the thought evolution method-BP model and SVR model is proved through the application of practical engineering.(4)Aiming at the shortcomings of unstable accuracy and strong randomness in single model prediction,the ELM multi-model fusion algorithm was proposed,and the variable weight combination prediction of the thought evolution method-BP neural network and SVR model was carried out by the extreme learning machine.In order to probe into a single model prediction accuracy with low impact on the fusion algorithm,established the contrast experiment,the experimental results show that the evolution of thinking in law-BP neural network model,support vector regression model prediction accuracy is higher or lower under different conditions,ELM multimodel fusion algorithm can get smaller error of predicted results,It is proved that the ELM multi-model algorithm has higher stability and accuracy than the single model.