Hazard Risk Assessment Of Tunnel During Construction Based On Machine Learning

The total number of mountain tunnels in China will increase by 2000 km/year,with an average of 6 km/d,of which a significant number will be deep buried and ultra-long tunnels with complex geological conditions,such as tectonic zones,active faults,and weak zones.During the construction of deep buried tunnels,water inrush,collapses and large deformation incidents occur frequently,becoming bottlenecks that hinder the construction of deep buried tunnels.There is an urgent need to assess and predict tunnel construction risks.However,existing research lags engineering applications.Therefore,in view of the key scientific problems such as the untimely and low accuracy of tunnel engineering disaster risk prediction,combined with the actual needs of China’s deep-buried ultra-long tunnel construction,this study carried out the research on the prediction and application of tunnel engineering disaster risk based on machine learning.The thesis establishes a database of tunnel squeezing deformation,collapse,and water inrush disaster by widely collecting cases at home and abroad.Using machine learning methods as the main research tool,the risk prediction and evaluation methods of large deformation,collapse and water inrush disaster of tunnel have been studied.These risk prediction methods are also applied to actual projects to provide new ideas and ways for tunnel engineering disaster risk prediction and evaluation.This thesis mainly obtains the following results and conclusions:(1)In order to predict the squeezing level quickly and accurately,this thesis presents a new prediction method based on evidence theory and support vector machine(SVM).The strength stress ratio(SSR),tunnel burial depth(H),tunnel equivalent diameter(D),support stiffness(K)and rock quality index([BQ])are selected as the evaluation indexes for the squeezing level prediction.A framework for the identification of tunnel squeezing is then established based on three typical tunnel squeezing classification schemes.Based on the statistics and analysis,the tunnel database of squeezing cases is constructed,and the Gaussian distribution membership model of each evaluation index is established to objectively determine the basic probability assignment(BPA)of each index.Finally,the generated BPAs are fused by the Dempster-Shafer(D-S)synthetic formula to achieve the prediction of tunnel squeezing levels.In terms of the maximum deformation prediction,the prediction method based on SVM model is established,and three optimization algorithms of grid search(GS),particle swarm optimization algorithm(PSO)and genetic algorithm(GA)are used to optimize the hyperparameters of SVM,and the prediction performance of the three optimization algorithms is compared.Overall,the GS-SVM model performs significantly better than the GA-SVM and PSO-SVM models,which is more suitable for tunnel extrusion deformation prediction.(2)A risk assessment method of tunnel collapse is proposed based on case analysis,advanced geological prediction,and D-S evidence theory.Firstly,11 risk factors are selected as the risk evaluation indexes,combined with the dynamic response parameters of advanced geological prediction,the tunnel collapse risk assessment index system is formed.Based on the number of risk index intervals,to avoid the influence of subjective factors,the BPA of each index is determined by the European distance method,and the index weight is estimated according to the similarity and support between the indexes.Finally,the D-S evidence fusion theory and the maximum probability principle are applied to determine the risk level of tunnel collapse.To improve the accuracy of the evaluation results,the tunnel comprehensive advanced geological forecast system is proposed,and the quantitative classification of unfavorable geological conditions is established.The proposed method can improve the accuracy of the evaluation results.This method is applied to the collapse risk evaluation of Yuxi tunnel through F14 fault,and according to the evaluation results,the corresponding risk mitigation measures are put forward for dynamic risk regulation.Moreover,based on Monte Carlo(MC)simulation technology to verify the effectiveness of the proposed method.(3)The Bootstrap algorithm in statistics is introduced to quantify the impact of uncertainty in traditional point prediction models.Combining multiple intelligent algorithms with Bootstrap,the interval(probability)prediction model based on Bootstrap-SVM-BPNN algorithm is established to provide accurate point prediction and reliable interval prediction for the water inflow of tunnel during construction.The effectiveness of the point prediction is evaluated in terms of the coefficient of determination(R~2)and root mean square error(RMSE),and the quality of the interval prediction is quantitatively assessed in terms of prediction interval coverage probability(PICP),mean prediction interval width(MPIW),and coverage width-based criterion(CWC).The prediction performance of the proposed method is tested based on the collected tunnel water inflow case data.In point prediction,through the comparative analysis of R~2 and RMSE of six models,the prediction performance of Bootstrap-SVM method is significantly better than the single prediction models such as multiple linear regression(MLR),generalized recurrent neural network(GRNN),extreme learning machine(ELM),backpropagation neural network(BPNN)and SVM.In the interval prediction,the prediction interval of the tunnel water inflow under different confidence levels can be clearly and reliably constructed,which can better cover the actual value,and show the good prediction performance of this model.Based on the SMOGN method,the prediction performance of the model can be effectively improved.(4)The collapse risk evaluation method and the water inflow prediction method proposed in this thesis were successfully applied to the Gaofeng Tunnel.The prediction results are in accordance with the reality,showing the accuracy of this method.For the risk levels corresponding risk adjustment measures are given,which can be used as a reference for similar projects.The research results of this thesis not only provide a more feasible way to enhance the applicability and practicality of traditional data mining and machine learning methods in the field of tunnel engineering,but also provide a new vision for the further development of tunnel disaster risk prediction and evaluation methods,which is expected to provide some theoretical support and practical reference for the construction safety,disaster prevention and mitigation of deep buried ultra-long tunnel projects under complex geological conditions.