Deep Learning Based Ground Penetrating Radar Image Permittivity Inversion Research

Ground penetrating radar(GPR)is an important geophysical tool,which has the detection characteristics of fast,nondestructive,continuous and portable.It has been widely used in civil engineering,geotechnical engineering and nondestructive detection.In the field of civil engineering,GPR is mainly used in advanced detection,tunnel structure detection,highway defect detection and building detection.In the field of geotechnical engineering,GPR is mainly used for rock detection,soil composition and water content analysis.In the field of nondestructive detection,GPR is mainly used in internal tunnel defects detection and underground pipe detection.GPR has a broad application prospect in practical engineering.However,the GPR images were mainly interpreted by manual analysis,which has strong subjectivity.The permittivity images have the characteristics of intuitive and clear,so that the efficiency and accuracy of engineering detection can be improved by converting GPR images into permittivity images for analysis.Therefore,an accurate and effective GPR inversion method has became an urgent need of practical detection engineering.The main study works and achievements of this paper are summarized as follows:(1)In order to solve the problem of transforming GPR B-scan into permittivity images,this paper proposed an deep learning based inversion network named PINet according to the characteristics of GPR images.Due to the serious attenuation of GPR signal along the depth direction,a dimension compressor was designed to compress the time dimension of B-scan and extract effective features.Aiming at the problem of poor spatial correspondence between GPR B-scan image and permittivity image,a global feature encoder was designed to extract the global features of B-scan.Finally,a permittivity decoder was designed to reconstruct the permittivty image according to the compressed feature maps.(2)Aiming at the problem of complex applications and detection structures,a general model data set was established to verify the inversion performance of PINet and the applicability of PINet to different frequency GPR data.Firstly,permittivity models with different shapes,sizes and target permittivity were designed.GPR B-scans were generated based on the permittivity models.Then,the PINet was trained,verified and tested on the general model data set.The permittivitity images were successfully reconstructed,which proved the effectiveness of the PINet.The accuracy and superiority of PINet network were proved by comparing with the existing deep learning based inversion network.Finally,the engineering field test was implemented,and the collected GPR data was inverted by the PINet.The inversion results proved the excellent inversion performance of the PINet and the applicability of PINet to real data.(3)Aiming at the application of GPR inversion network to tunnel lining defects detection,this paper established a tunnel lining defect data set and verified the performance of the PIINet.Firstly,permittivity models with cracks,delamination,voids and non-compactness defects were generated according to the engineering practice.The GPR B-scans were generated based on the permittivity models.Then,the PINet was trained,verified and tested on the tunnel lining defect data set.The tunnel lining permittivity models were successfully reconstructed,which proved the accuracy of the PINet.Finally,the modelling test was implemented,and the real data were inverted by the PINet.The inversion results proved the excellent inversion performance of the PINet and the applicability of PINet to practical engineering detection.