| To remedy the natural defects of the traditional linear inversion method of Electrical Resistivity Tomography(ERT)which tends to fall into local minima,overreliance on the initial geological model and non-uniqueness of the solution.In this paper,a study on the nonlinear inversion method of high-density resistivity method is carried out based on the intelligent algorithm represented by neural network.The engineering practicality of the GASA-BP algorithm proposed in this paper is verified through five aspects: theoretical analysis,numerical simulation,algorithm optimization,joint inversion,and engineering application.The main abstract of the full paper is as follows.(1)Theoretical aspects.In this paper,starting from the exploration theory of ERT,the formula of stable point source current field in the ground and the formula of apparent resistivity under the condition of inhomogeneous medium are derived;on the basis of discussing different device types and arrangements of ERT,the traditional least squares inversion formula of ERT is derived;on the basis of analyzing different transfer functions of BP neural network and the performance of training algorithm,the sample division method,modeling method and inversion process of BP neural network for ERT inversion are proposed.(2)Numerical simulation.Based on the COMSOL finite element analysis platform,the ideal exploration depth of the ERT is explored and the inversion model size is determined.The finite-difference numerical simulation method of the two-dimensional geoelectric field is introduced in detail,and the apparent resistivity data of the geoelectric model is calculated by the finite-difference method,which is the key to constructing the training samples and is the basis of the full-text study.(3)Aspects of nonlinear inversion algorithm research.This paper combines the principle of BP neural network inversion and proposes a nonlinear inversion method for ERT based on BP neural network.It is found that the BP neural network algorithm is more advantageous than the traditional least squares method to identify the morphology and size of the heterodyne in the geoelectric model.To overcome the defects of BP neural network with random initial parameters and easy to fall into local minima,this paper proposes a joint GASA-BP inversion algorithm by combining it with a genetic algorithm with global search capability and a simulated annealing algorithm with strong local retrieval capability.The joint inversion algorithm improves the defect that BP neural network is easy to fall into local minima and overcomes the "premature" phenomenon of genetic algorithm.Through model simulation,the GASA-BP algorithm proposed in this paper has high convergence and high inversion accuracy.(4)Joint inversion.In this paper,the GASA-BP algorithm combining intelligent algorithm and BP neural network has high stability and inversion accuracy,but the practical engineering effect of the algorithm is also affected by the training samples of the network.To overcome this problem,the joint inversion algorithm of GASA-BP algorithm and traditional least squares method is proposed in this paper.The joint inversion algorithm takes the least-squares inversion model based on the measured apparent resistivity data as the geological a priori information,performs a derivative transformation of the least-squares inversion model by the relevant random field theory,and uses the resulting model for the GASA-BP algorithm training.It is found that the joint inversion method proposed in this paper has better inversion capability for engineering real-world data,which is an effective initiative to enhance the engineering utility of the neural network-based ERT inversion method.(5)Engineering application.In this paper,ERT field data acquisition experiments were conducted,and more data were obtained.By means of pre-processing,the degree of error of the measured data is corrected,and the data are used for algorithm research and performance verification.The effectiveness and feasibility of the GASA-BP algorithm proposed in this paper are proved from the practical point of view. |
