| In recent years,rock tunnel has developed rapidly as an important part of China’s infrastructure construction.Tunnel construction has also gradually passed the stage of long,large and deep tunnels.Due to the highly uncertain geological conditions in the surrounding area and the limited number of experienced experts,the construction process faces great challenges,especially in assessing the quality of the surrounding rock and the safety of the excavation.In addition,geological surveys in the initial stage are limited and can hardly reflect the technical and hydrological characteristics of the whole tunnel.Therefore,there is an urgent need to use the continuously circulating information about the tunnel wall to accurately assess the quality of the surrounding rock.Currently,mainly contact and non-contact methods are used for in-situ assessment.Although these methods improve the acquisition of geological parameters,there are also the following problems:(1)although the contact method is simple and direct,its process is dangerous,time-consuming,and subjective;(2)non-contact methods,especially digital photography,are easily influenced by the external environment.In addition,most non-contact methods lack the recording and analysis of environmental and photographic parameters during the recording process.The number of features that can be extracted by current non-contact methods is relatively limited,and there is also a lack of efficient feature extraction methods,which makes it difficult to integrate expert judgments.Although the amount of information contained in the excavated area is huge,ideal methods for extracting geological information are still lacking.Moreover,combining the extracted key features to build an accurate predictive model for the quality of the surrounding rock is inevitably challenging.This undoubtedly leads to great difficulties in the safety assessment of tunneling based on the modeling of on-site geological conditions.Against this background,this paper is based on two projects addressing the two key scientific issues of rock structure characterization and rock refinement classification: "Science and Technology Project of Yunnan Provincial Ministry of Transport"(No.25 of 2018)and "Innovation Team in Key Areas of the Innovative Talent Promotion Plan of the Ministry of Science and Technology"(No.2016RA4059).Using field measurements,data statistics,intelligent algorithms and numerical simulations,quantitative extraction algorithms for rock face features are proposed.Then,based on the fusion of heterogeneous data from multiple sources,a machine learning model for rock mass evaluation is constructed.Finally,a safety evaluation of tunnel construction in a complex geological environment is conducted.(1)Taking the Mengzi-Pingbian highway tunnel project in Yunnan Province as an example,the on-site geological and environmental data are collected by digital photography and measurement techniques.On this basis,the correlation between the image quality and the tunnel construction environment and equipment parameters are analyzed.Then,based on the surrounding rock classification methods widely used at home and abroad,the main classification parameters and indicators are established based on image processing extractions and field measurements.On this basis,the optimal field acquisition time,internal parameters for photography,equipment arrangement parameters,etc.are proposed.The study shows that the optimized parameters can help to accurately and efficiently collect the basic data required for the classification of the surrounding rock.(2)Based on the technology of image photography,the photographic image database of the tunnel excavation surface is created.For the main features of the surrounding rock such as weak interlayer,joint fissure,groundwater and obvious structure,deep learning algorithms such as Deeplab V3+,Fra Seg Net and InceptionResnet-V2 are used to extract the respective features.At the same time,hyperparameter and moduli optimization method is applied to achieve more accurate feature information classification and refined semantic segmentation.Different postprocessing algorithms for graphics are proposed for the most important features of different rock masses.In this way,the effective conversion of the image with rock mass features from pixels to coordinates and then to information is realized.Finally,the geometric parameters of the main features of rock masses are obtained,including the structural type,weak interlayer,groundwater area and distribution,and the parameters of the common trace(inclination angle,length,distance,density,thickness,etc.).(3)Using the structure from motion(Sf M)algorithm to obtain three-dimensional point cloud data of the tunnel wall,a three-dimensional method is proposed to characterize the occurrence of discontinuous structural planes.In this way,the calculation of local curvature,automatic classification of normal vector,calculation of occurrence and visualization of discontinuity features are realized.Then,a hybrid machine learning model(RS-SMOTE-GBT)is proposed for the classification of 3D traces.Thus,the 3D discontinuity trace can be accurately extracted from the 3D discontinuity.The results show that the unbalanced ratio of data structure can be optimized with smote technology.At the same time,the weak performance of 3D trace classification caused by the imbalance of data is mitigated.Finally,the generalization ability and robustness of the proposed model are systematically analyzed.(4)Based on image processing extraction and field measurements,a 13-dimensional multidimensional heterogeneous database containing rock geometry,tunnel environment,physical and mechanical parameters is constructed.Then,a hybrid TPE-GBRT prediction model and its comparison algorithm are built using metamachines and ensemble machine learning models.The tree-structured Parzen estimator is used to optimize the hyperparameters of the model.Using this algorithm,the parameter combination for the optimal prediction of the hybrid machine learning model is determined.Using the optimized model,the rock mass rating(RMR)index for the surrounding rock classification is accurately predicted.At the same time,the ranking of feature importance affecting the classification is appropriately determined.Finally,the validity and applicability of the model prediction are analyzed.(5)Based on the information about the occurrence of structural planes obtained by machine vision,the geological environment of the rock is constructed based on the discrete fracture network(DFN).Then,a three-dimensional tunnel model 3DEC is constructed based on the excavation surface information.The continuous excavation process of the tunnel is simulated by using the settlement displacement of the vault crown as an evaluation measure and combining it with the strength reduction method.During the simulation,the stress-strain and shear slip during the tunnel excavation are monitored to determine the stability characteristics and safety condition during the continuous excavation of the tunnel.Then,based on the response surface design method,the relationship between the numerical modeling parameters and the safety factor of the tunnel is calculated.Finally,the influence of the parameters and their coupling states on the safety factor is shown. |
PDF Full Text Request