Deep Learnig Inversion Method Of Electrical Resistivity Survey Based On Prior Constraints

Electrical Resistivity Survey(ERS)is a common engineering geophysical detection tool.Owing to its wide application scope,high exploration efficiency,low economic cost and flexible operation,ERS has been applied in geological exploration,disease detection and geological forward-prospecting for geological hazards in the construction and running of highways,railway,tunnels,dams,hydraulic power plants and urban underground space engineering.The inverse problem of ERS is to reconstruct the subsurface distribution of electrical resistivity with the measured apparent resistivity of underground space.It is the primarily method to reveal the location and geometry of subsurface anomaly of electrical resistivity.However,linear inversion,commonly used at present,suffers from the dependence of initial model,local optimization,and insufficient accuracy.To solve the above problems,focusing on the rapid development of artificial intelligence algorithms and starting from the non-linear nature of resistivity inversion,an intelligent inversion scheme based on the deep learning method was presented in this paper.Theoretical analysis,numerical simulation and field experiments were used to carry out the research.Starting from the feasibility of deep learning method to solve the inverse problem of resistivity data,ERSInvNet network was designed for deep learning of resistivity.On this basis,resistivity deep learning method based on constraints of prior information is studied,and three constraint items were made including input constraint based on prior tier feature map,loss function data item constraint based on depth weighting function information and regularization constraint based on prior smoothing penalty.Finally,the developed inversion scheme was introduced into the field of surface detection and in-tunnel geological forward-prospecting.Numerical,physical and field experiments utilizing proposed deep learning inversion were conducted respectively.The primarily work and results of my study is shown as follow:(1)The feasibility and realizing of deep learning solution of resistivity inversion problem.In this thesis,the applicability of inversing resistivity data using deep learning method was analyzed.Then,the spatial corresponding relation between the anomalous resistivity area in real model and measured apparent resistivity profile was discussed.The deep learning inversion scheme based on deep convolution network was then proposed with the deep learning network of ERSInvNet for resistivity inversion.These works provided basis for building an end-to-end mapping relation between the input(the apparent resistivity)and output(resistivity model)dataset for deep learning inversion.(2)Deep learning inversion method of resistivity data based on prior information.In order to solve the problem of fuzzy output,low inversion accuracy and false anomalies of ERSInvNet network,a solution idea of applying effective prior information to constraint in deep learning inversion was proposed.The prior input structure of tier feature map was then constructed,which overcame the fuzzy output of ERSInvNet network.The prior constraint structure of depth weighting function was proposed to improve the inversion accuracy of resistivity abnormal region and inversion results of deep anomalous body.The prior constraint of smoothing penalty was constructed to solve the serious problem of false anomaly in inversion results,and the deep learning inversion scheme for resistivity inversion based on prior information was then established.Compared with the traditional linear least square inversion method,ERSInvNet inversion imaging provided the location and geometry of resistivity abnormal areas with more accuracy and closer inversion results of model resistivity to its real value.Effective prior information significantly improves the precision and accuracy of inversion.Additionally,a domain transfer method based on migration learning is proposed to build a foundation for inversion interpretation of field measured data(3)The realization and application of deep learning inversion for Surface ERT.Based on the prior information based resistivity deep learning inversion method,surface ERT deep learning inversion imaging was implemented and its field test was conducted.And then,the deep learning inversion method was introduced into the cross-hole ERT survey after adapting modification.The input data reorganization strategy with global eigenvector prior information was proposed.The priori constraint dataset of middle-distance weighting was made.The deep learning network for cross-hole ERT inversion,ERSInvNet-C,was then designed.Based on these works,the deep learning inversion imaging in cross-hole ERT was realized and tested in physical model experiments(4)Deep learning inversion of in-tunnel resistivity forward-prospecting and its application.The in-tunnel focused sounding observation system was optimized in this thesis.On this basis,the inversion dataset ERSDSet-T of deep learning inversion of resistivity was built.Then,the ERSInvNet-T network for in-tunnel resistivity inversion was designed based on fully connected reconstruction of 3D feature and convolution output resistivity model.Finally,the efficiency of the proposed deep learning inversion method was tested by numerical and field experiments.