Prediction Of Longitudinal Tunneling-induced Settlement Using Long Shortterm Memory Network

The ground disturbance caused by earth pressure balanced shield tunneling in cities causes ground settlement and affect the safety of buildings and traffic.The degree of disturbance is influenced by shield operational parameters,geological parameters,tunnel geometric parameters and abnormal tunneling conditions,etc.Based on finite element method and formation loss rate,it is difficult to consider the timely and accurate prediction of longitudinal ground settlement under the coupling effect of multiple factors.To solve the above problems,this study used traditional machine learning(Back-propagation Neural Network and Random Forest)and deep learning algorithms(Long Short-Term Memory networks and Bi-directional Long Short-Term Memory)to establish tunneling-induced longitudinal ground settlement prediction models,to provide effective abnormal ground settlement warning and control strategies for tunneling.The main results of this study are as follows:(1)A scripted longitudinal ground settlement database was established.Considering the factors that affect the soil response of tunneling,select the tunnel depth,the groundwater table,modified count of standard penetration test,modified uniaxial compressive strength,modified count of dynamic penetration test,type of soil layer at the tunnel face,torque,tunneling speed,thrust,chamber pressure,cutter rotational speed,shield machine operating status and the distance between the cutter and the monitoring point are 13 input parameters,and the real-time ground settlement is the output parameter.The script written in Python language can quickly read and process the data of the geological survey report,the operational parameter record,the tunneling record and the daily report of the ground settlement detection,and organize all the input and output parameters into a database with the ring as the storage unit.The results indicate that the scripting of establishing database can greatly improve the efficiency of data processing,and provide a fast and accurate method for the construction and expansion of the algorithm model training database.(2)The longitudinal settlement models based on machine learning algorithms were established and the prediction performance of 4 machine learning algorithms were compared.This study established longitudinal settlement prediction models based on BP neural network,random forest algorithm,long short-term memory neural network and bidirectional long-shortterm memory neural network,used particle swarms optimization and k-fold cross-validation method to determine the optimal hyperparameters of machine learning algorithms,and set the cyclic learning rate to improve the predictive performance and training efficiency of deep learning algorithms.To compare the settlement prediction performance of different algorithms,this study used performance evaluation indicators,such as mean absolute error(MAE),Pearson correlation coefficient(R)and the predicted curve of the longitudinal settlement.The results indicate that the prediction accuracy of BP neural network is low,and it is difficult to predict the development law of longitudinal settlement.The prediction accuracy of bidirectional longshort-term memory neural network is the best,and the predicted settlement development process fits well with the actual,followed by the random forest and long-short-term memory neural network.(3)An intelligent selection model of operational parameters with the goal of settlement control was proposed.This study used bidirectional long-short-term memory neural network to train the data with small settlement stability value(less than 10 mm)in the database to get the tunneling parameter prediction model.Before tunneling,the operational parameters can be predicted through the geological parameters,tunnel geometric parameters and shield operational parameters.Besides,the longitudinal settlement prediction model was used to test the settlement control effect,and the grid search method was used to optimize the operational parameters at the points with poor control effect of settlement to meet the control demand.The results indicate that the tunneling-induced settlement based on the predicted shield operational parameters is less than the measured ground settlement,and the intelligent selection model of shield operational parameters can realize effective control of settlement through the prediction and optimization of shield operational parameters.(4)The intelligent software of predicting and controlling tunneling-induced ground settlement was developed.Based on Android,integrated tunneling-induced longitudinal ground settlement prediction model and shield operational parameters intelligent selection model,and supporting the shield monitoring system,a mobile device application was developed which can predict the maximum ground settlement,longitudinal settlement and shield operational parameters in real time.The results indicate that the intelligent software has simple interface and operation and it can predict ground settlement and recommend tunneling parameters in real time during the tunneling process of the shield machine.