Research On Inversion Of Two-dimensional Borehole-to-Surface Resistivity Method Based On Deep Learning

The borehole-to-surface electrical method is a method of electrical resistivity in which the power is supplied in drilled borehole and the measurement is conducted on the ground.Compared with the conventional ground DC resistivity method,the borehole-to-surface model can better reflect the resistivity distribution of abnormal objects in deep underground.Linear inversion method has a complete theory and has significant effects,but it also has the problem of exacerbating the non-uniqueness of the solution and causing the inversion to fall into a local minimum problem.In view of low identifiability for the results,huge data as well as slow pace of linear inversion method,this paper introduces the deep learning theory in the field of computer vision into the electrical nonlinear inversion,and carries out the study of inversion of 2D borehole-to-surface electrical method based on convolutional neural networks with deep learning and application of measured data.In this paper,numerical simulation is carried out based on the finite difference method,and the borehole-to-surface electrical method in the DC resistivity method is applied to simulate the pole-pole array.The model library is established with samples of single abnormal object,double abnormal object,single-tier abnormal object and double-tier abnormal object in this paper.The convolutional neural networks with deep learning for inversion of borehole-to-surface electrical method are established on the basis of structure of encoder-decoder-skip.The activation function of Re LU and application of dropout are able to avoid overfitting and improve computation speed.After setting of parameters of the neural network,the training for it will be carried out by using the sample library.When the weight optimization is achieved,the algorithm cases will be carried out with multiple abnormal objects and analysis of Loss functions,Average and Intersection over Union(Io U)will be presented in this network.It proves that the convolutional neural network based on deep learning can perform an inversion of more accurate location and physical properties of abnormal objects in two-dimensional model.On the validation set,the cross-entropy loss function gradually reaches convergence after 100 trainings,and the Average validation accuracy can reach 0.950-0.960 with intersection ratio Io U?0.5.The neural network inversion experiment is carried out based on the measured data of borehole-to-surface electrical method in the north-west of Abaga banner,Xilin Gol League,Inner Mongolia.The inversion method based on deep learning can quickly obtain the inversion results.From the experiments,it can prove that the results are roughly corresponding to the ones of traditional inversion methods.The deep learning is focused on the classifications;however,the results are not as continuous as the traditional inversion method.Moreover,due to the large interval between the electrodes observed by which the measured data is collected,there are some errors in the resistivity distributions obtained by interpolating the measured data which produces the mapping of apparent resistivity data as the input of neural network.