Research On Synchronous Grouting Parameter Inversion And Surface Settlement Prediction Method For Shield Construction Based On Machine Learning

As one of the main underground tunnel construction methods,the shield method is widely used in urban underground construction.The degree of ground deformation caused by shield tunneling is important for evaluating engineering safety.To predict the ground deformation in shield construction,many scholars have used numerical analysis methods and machine learning methods that have emerged in recent years to conduct research.Although the calculation results of the numerical analysis method have high accuracy,the parameters obtained in the project are different from the actual values,which often requires a lot of debugging and calibration of the established FEM model,which is inefficient.The accuracy of the machine learning method depends to a large extent on the training data set,and there are problems that the results lack theoretical support and the model generalization ability is poor.This paper starts from an actual project in Beijing,combines numerical simulation with machine learning algorithms,establishes a numerical calculation surrogate model to predict the results of existing numerical models under different input parameters,and intelligent optimization algorithm is introduced to realize the back analysis of soil parameters and grouting parameters.Finally,a surrogate model prediction method for the surface settlement in shield construction is proposed based on the above.The main contents and results of this paper are as follows:(1)Starting from the actual project,combined with the on-site monitoring data and records,discuss and summarize the law of the influence scope and degree of shield construction on the surface subsidence.A numerical model is established for analysis,and the influence of grouting parameters and cover to depth ratio on ground settlement in shield construction is explored.(2)The performance comparison of different machine learning algorithms to predict the results of existing numerical models and the establishment of surrogate models.Use numerical models to generate data sets for machine learning algorithm training,and select three types of machine learning: Backpropagation Neural Network,Random Forest,and Support Vector Machine Regression,which are widely used in the field of geotechnical engineering.The algorithm uses random search method to set the hyperparameters of each algorithm,compares the prediction accuracy of different algorithms on the results of existing numerical models,and selects the optimal algorithm to establish a numerical calculation surrogate model.The test results show that although the BP-ANN has the smallest discreteness of prediction error,it has low computational efficiency.Although SVR has the highest computational efficiency,the average error dispersion of the prediction results is the largest.RF can have both computational efficiency and accuracy and is more suitable for predicting numerical simulation results than BP-ANN and SVR.Finally,RF is used to establish a numerical calculation surrogate model.(3)Compare the performance of three different optimization algorithms,namely particle swarm(PSO),differential evolution(DE),and artificial fish swarm(AFSA),and combine the best optimization algorithm with the surrogate model(AFSA-RF)to realize the back analysis of soil parameters and grouting parameters,the inversion values are input into the numerical model for calculation and compared with the monitoring data to verify its validity.The results show that the inversion value obtained by AFSA is the closest to the standard value,and has the best inversion capability.Combining AFSA with the surrogate model,the soil layer parameters and grouting parameters were obtained based on the field monitoring data.The obtained inversion value can reduce the average error between the numerical calculation result and the monitoring data by about 35.8%,and improve the calculation accuracy.(4)A surface subsidence prediction method based on surrogate model is designed.The data set is established by building numerical models and the surrogate model of surface subsidence prediction is established by using RF.Input the monitoring data into AFSA-RF to obtain the inversion value of soil parameters,and input the inversion value according to the depthweighted average,together with the cover to depth ratio and construction parameters,into the surrogate model,to realize the prediction of surface settlement.Using the monitoring data collected during the shield advancement process,the data set is continuously expanded.The results show this method can effectively predict the surface subsidence caused by shield construction,and the expansion of the data set can improve the prediction accuracy.