Research Of Geoelectric Model Identification For DC Resistivity Method Based On Convolutional Neural Network

DC resistivity method is an important electrical detection technology,which is widely used in environmental monitoring,mineral prospecting and other fields.At present,there are two kinds of data processing methods for DC resistivity method: linear method and nonlinear method.In the linear method,there are always problems of information loss and low accuracy of the identification result.In the nonlinear method,it is easy to get the result into local minimum when using neural network to identify the abnormal body model.In order to solve the problems in the nonlinear method,the geoelectric model identification network for DC resistivity method based on convolutional neural network is established through the Tensor Flow framework of deep learning.Through the data test,the result accuracy of identification network reaches more than 91%.The main research contents are as follows:1)Based on the numerical calculation of finite difference method,the sample database of abnormal body models with different buried depths and different resistivity is established.Based on the framework structure of Tensor Flow,different convolutional neural networks are established for the single abnormal body model,double target abnormal body model and ten typical abnormal body models.2)For different abnormal body models,identification networks are established respectively.After analyzing and comparing the networks,Relu activation function,entropy loss function and dropout technology are selected to improve network accuracy.After training the network with three abnormal body model samples,the accuracy results are above 96% and the loss is reduced to less than 0.12.3)In order to improve the accuracy of identification,noise is added to the sample data of high-resistance abnormal body,which increases the anti-interference of the sample.The test sample is tested by the trained recognition network,and the recognition accuracy is above 93%.4)The parameters of batch size,the size of the convolution kernel,dropout rate of discarded and optimization algorithm is analyzed,which will influence the recognition precision of convolutional neural network.The ten typical abnormal body is as an example to recognize the network.The test results show that training network with the parameters of batch size 80,the convolution kernels size 3×3,dropout rate of discarded 0.6 and Adam optimization algorithm has the best training effect and the highest output precision.5)The accuracy of the identification network is tested by using double target abnormal body model data,ten typical abnormal body model data and outdoor experimental data,and the results show that the test accuracy are all reached more than 91%.It is proved that the identification network can identify the geoelectric model of DC resistivity method effectively.