Research On Deformation Prediction Method Of Tunnel Surrounding Rock Based On Multi-Model Fusion

Since the 21 st century,Chinese highway development has reached a new stage of rapid development.The construction of highways has gradually extended to mountainous areas,which has also promoted the development of highway tunnels.Tunnel construction will lead to changes in the stress-strain balance of surrounding rock,including vault settlement and peripheral convergence.If the deformation of surrounding rock isn’t monitored and predicted in time,major safety accidents such as large deformation of surrounding rock,collapse or rock burst will occur.Therefore,timely monitoring and accurate prediction of vault settlement and surrounding convergence are of great significance for reducing accident risks and improving tunnel construction quality.The current prediction of the surrounding rock deformation mainly has the following problems: lack of comprehensive analysis of the overall monitoring section deformation of the tunnel,single prediction model,low accuracy of multi-step prediction,etc.In response to the above problems,this paper proposes a multimodel fusion-based surrounding rock deformation prediction model.This method fuses the time-series data clustering model,sequence decomposition model,traditional time-series data prediction model and deep learning to build a multi-model fusion surrounding rock deformation prediction model,so as to obtain a prediction model applicable to the deformation of surrounding rock in multiple sets of cross-sections of the whole tunnel.The main points of this paper are as follows:(1)Analyze the knowledge of surrounding rock deformation mechanism and data resource requirements.Firstly,this paper introduces the mechanism of vault settlement and peripheral convergence deformation during tunnel construction in detail;Secondly,the deformation law of vault settlement and peripheral convergence under different surrounding rock grades is given;Finally,the overall framework of vault settlement and peripheral convergence prediction is proposed.It includes data pre-processing,surrounding rock deformation sequences clustering,sequences decomposition and reconstruction,prediction model construction and verification.(2)Construct a time series clustering model of vault settlement and peripheral convergence.By analyzing multiple sets of deformation monitoring sections under different surrounding rock grades in the tunnel,a clustering model based on soft-DTW and K-Mediods is constructed to cluster the vault settlement and peripheral convergence time series of different surrounding rock grades.Thereby multiple groups of clusters with different deformation trends are obtained.And compared with K-Mediods clustering methods based on different similarity measures,the effectiveness of soft-DTW-K-Mediods clustering is verified.(3)Construct a decomposition and reconstruction model of surrounding rock deformation data based on singular spectrum analysis.The clustering centers of vault settlement and peripheral convergence in each category are extracted.The center sequence is decomposed and reconstructed into three sub-sequences of trend sequence,periodic sequence and noise sequence by singular spectrum analysis for subsequent modeling and analysis.(4)Construct a prediction model of surrounding rock deformation based on multi-model fusion.Firstly,it is judged whether the subseries obtained by decomposition in(3)is smooth.If the series is smooth,the ARMA model is used for prediction,and the non-smooth series is predicted using LSTM.To improve the prediction accuracy of the LSTM model,the paper combines two loss functions: MSE and MAE.The LSTM prediction model based on MAE＿MSE combined loss is constructed and optimized.Then,the prediction result is calculated by integrating the three groups of sub-sequence results.Finally,the validity of the model is verified with other deformed data in each category.The accuracy of the proposed model is verified through the establishment of multiple sets of comparative models and engineering applications.