Fast Imaging Of Surface Magnetic Resonance Sounding Based On Residual Convolutional Neural Network

Magnetic resonance sounding(MRS),as an advanced geophysical method for direct and quantitative detection of groundwater,has played an important role in water exploration in arid areas,water resources investigation in pastoral areas,and advanced prediction of water bodies in tunnels and mines.However,the scarcity of fresh water sources and the frequent occurrence of geological disasters caused by groundwater are still the two most prominent water problems facing China.Therefore,the development of new high-precision MRS groundwater rapid imaging methods have always been an urgent need in this field.This paper proposes a deep learning method using Residual Convolutional Neural Network(RCNN)in view of the shortcomings of the existing MRS inversion methods at domestic and abroad that need to manually adjust the regularization parameters and the inversion takes a long time,which only needs to train one network with strong generalization ability can be applied to most groundwater models,which has the advantages of rapidity,high efficiency and wide application range.This research is of great significance for promoting the progress of magnetic resonance inversion methods,and also has reference for promoting the engineering application of MRS technology.First,the network training research of MRS inversion method based on RCNN is carried out.Factors that affect the training effect of RCNN include network structure,dataset size,sample noise level,and the choice of training parameters.By comparing the loss function,root mean square error,training time and adaptation effect of new sample data after network training under different conditions,the network structure is finally determined as Res Net50,and the number of data sets is 10~5,of which 80%are training sets,20%is the validation set,and the network generalization ability is the strongest when the maximum 200 n V noise level is added to the sample.Then,the generalization ability analysis of the RCNN inversion method is carried out.The influence of MRS data on RCNN inversion results under different noise levels,different resistivities and different coil sizes was tested and analyzed,and it was concluded that the RCNN inversion method is suitable for noise levels less than 300n V,resistivity greater than 10?m,and measuring coil side lengths greater than 50m.At the same time,compared with the existing classical QT inversion methods,the average time-consuming of the RCNN inversion method is less than 1 s,while the QT inversion method is more than 30 s.The RCNN inversion method has the advantage of being fast and efficient.Finally,through an example of ground magnetic resonance measurement near Hongyan Lake in Tongliao City,using the RCNN inversion method,the credible vertical distribution results of underground aquifers are effectively obtained,which proves the effectiveness of the method.The comparative analysis with the results of drilling and pumping experiments at the same location further verifies the accuracy of the ground magnetic resonance imaging method based on RCNN proposed in this paper.The innovative work of this paper is as follows:1.For the first time,a deep neural network is used to solve the MRS imaging problem,and an RCNN-based MRS data inversion method is proposed,which has higher computational efficiency and is suitable for complex groundwater models,and gets rid of the limitation of traditional methods that require manual adjustment of regularization parameters.2.The water content-relaxation time feature matrix of the groundwater model is used as the data set of RCNN,so that RCNN can comprehensively learn the mapping relationship between MRS data and water content and relaxation time,so that it can face the unlearned subsurface model(such as unknown resistivity)and measurement methods(such as different coil sizes)can show good generalization ability.