The Research On Inversion Methods Of Ground Penetrating Radar

Ground Penetrating Radar(GPR)is a non-destructive detection tool that is widely used at present.It is usually used to detect underground unknown structural information.It measures unknown underground structures by transmitting high-frequency electromagnetic wave signals and receiving reflected electromagnetic wave signals.The measurement process does not need to destroy the underground structure,and has the advantages of high precision,convenience and speed.However,at present,the analysis of GPR signals mainly relies on manual calculations.The signal map returned by GPR is not intuitive,and users need a lot of complicated professional calculations to get the desired information.Traditional manual calculation methods have certain limitations in computational efficiency,accuracy and stability,and inversion technology is designed to solve this difficulty,aiming at converting the received electromagnetic wave echo signal graph into underground structure model,so as to improve the efficiency and accuracy of analysis.Dielectric constant is a physical quantity that describes the response of a material in electromagnetic field.It can usually be used to represent the condition of underground media.Different materials have different dielectric constants,and the dielectric constant of the same material is also affected by factors such as its density and humidity.The required data can be obtained directly through the permittivity model of underground structure,which greatly simplifies the user’s analysis of ground penetrating radar signals.This paper aims to solve the inversion problem by means of deep learning,and convert two-dimensional ground-penetrating radar B-scan signals into intuitive subsurface structure information through neural networks.The main research contents are as follows:(1)Aiming at the problem that ground penetrating radar is difficult to obtain a large amount of data,a forward algorithm based on FDTD is proposed.The algorithm replaces the derivative calculation with the second-order central difference,and solves Maxwell’s equations in one-dimensional,two-dimensional and three-dimensional spaces respectively.At the same time,we introduced an additional source in the simulation area to simulate the GPR signal source,and used PMC boundary conditions to deal with the electromagnetic waves at the boundary of the simulation area to ensure the accuracy and stability of the calculation results.(2)Aiming at the problem of low inversion accuracy of ground penetrating radar signal images,an inversion network GINet based on deep learning is proposed to invert the dielectric constant of ground penetrating radar signal images,and the ground penetrating radar two-dimensional B-The scan signal is transformed into a subsurface permittivity model.The algorithm first performs bicubic interpolation on the ground-penetrating radar signal image,and then performs Fourier transform on the interpolated image.The original image and the transformed image are spliced and then inverted by GINet.The image of the inversion result is again subjected to bicubic interpolation to adjust the image resolution.In order to verify the performance of the network,this paper uses simulated data to test the network,and the verification data is generated by the simulation of the time domain finite difference algorithm.At the same time,this paper also compares our network with some existing deep learning inversion networks,quantitatively evaluates the performance of our network.