Research On Prediction Model And Optimization Method Of Ground Settlement Of Shield Tunnel Based On Statistical Machine Learning

The shield method is the most widely used construction method in the construction of subway tunnels.The prediction of ground deformation caused by shield tunnel construction is one of the most important topics to ensure the safety of subway construction.At present,China's rail transit is still in the stage of large-scale high-speed construction.On the one hand,the intensive rail transit construction has put forward higher requirements for the prediction and control of strata deformation,on the other hand,a large number of engineering practice experiences have created conditions for the accurate prediction of strata deformation.Due to the multi-factor and nonlinear characteristics of ground deformation response caused by shield tunneling,it is difficult for traditional theories and models to reflect this complex relationship.Therefore,it has become an important research direction to use machine learning model to realize intelligent prediction of ground subsidence.However,the existing machine learning sedimentation prediction models have the problems of poor generalization ability and weak explanability.Taking Chengdu Metro Line 6 as the background,this paper introduces the tree model in machine learning model to predict the surface subsidence caused by shield tunnel construction,and puts forward the corresponding optimization method on the basis of comparative analysis with various algorithms.The main research contents and results are as follows:(1)Study on the mechanism and influencing factors of ground deformation caused by shield tunneling.Based on the process of shield tunneling,the mechanism of ground deformation caused by shield tunneling is comprehensively analyzed and summarized,with emphasis on the horizontal deformation law of ground and the diachronic variation law of ground subsidence.Based on the construction process,the corresponding numerical model is established to analyze the influence of various factors on the surface settlement in the process of tunnel excavation,which provides a theoretical basis for the subsequent research.(2)Extraction of modeling elements of prediction model based on surface subsidence analysis.The monitoring data of ground subsidence and parameters of shield tunneling in three sections were collected and sorted out.The distribution law and development trend of ground subsidence caused by shield tunnel construction are analyzed,and the factors that affect the ground subsidence value are statistically analyzed,and the corresponding database is established.The results show that the surface subsidence groove curve in the interval basically conforms to the Gaussian distribution,the width coefficient of the subsidence groove curve is negatively correlated with the tunnel buried diameter ratio(H/D),and the formation loss is significantly correlated with the formation type.The diachrony development curve distribution of subsidence is mainly related to the formation type and grouting.The subsidence development curve is obviously different under different strata conditions,but basically conforms to the four stages of subsidence development.Through the analysis of the factors affecting the surface subsidence,it is found that the stratum type is the main factor affecting the subsidence magnitude,the special surrounding environment will affect the distribution of subsidence value,and the tunneling parameters mainly reflect the tunneling state of shield.(3)Comparison of subsidence prediction models.Five machine learning algorithms,including linear regression model(LR),BP neural grid model(BPNN),support vector regression model(SVR),two tree models CART model and XGBOOST model,which are widely used in underground engineering,are considered to build corresponding prediction models respectively.The differences in learning ability,generalization ability and interpretability of each model were compared,and the optimal model for predicting surface subsidence was determined.Taking the prediction error of the test set as the index to evaluate the generalization ability of the model,the performance ranking of the model is XGBOOST > CART > SVR > BPNN > LR.Especially in a few measuring points with large settlement value,the prediction effect of the two tree models is far better than that of the other three models,and the tree model also has better interpretability.(4)A multi-stage dynamic prediction model for subsidence prediction is proposed.Considering the time effect of ground subsidence caused by shield tunneling and the existing priori information,a multi-stage prediction model of ground subsidence is proposed to solve the problem that the engineering significance of the existing prediction model is not clear.The dynamic prediction of subsidence is realized and the model data set is optimized.(5)The regularization optimization of prediction model is realized based on Bayesian method.Starting from the regularization term of XGBoost algorithm,the Bayesian optimization method is introduced to realize the automatic optimization of the regularization parameters of the model,which further improves the generalization ability of the model itself.The optimized model has better prediction effect for surface subsidence.