Intelligent Identification Method Of Internal Defects Of Engineering Structures Based On Cross-domain Translation Of Ground Penetrating Radar Data

The detection and identification of internal defects in underground engineering structures is very important.Structural defects such as cracks,cavities,and leakages inside the underground engineering structure directly affect the service life of the structure,cause high maintenance costs,and even cause safety accidents,threatening people’s lives.Ground penetrating radar(GPR)is a non-destructive detection technology widely used in the detection of underground engineering structures.Identifying internal defects based on collected radar data is a current research hotspot and an urgent need.In recent years,the intelligent identification technology based on deep learning has developed rapidly.However,in actual engineering,the amount of GPR data is limited,the labeling is difficult,and the shape and distribution of internal defects are complex.As a result,the intelligent identification model trained based on simulated GPR data is difficult to generalize to actual GPR data,which poses a huge challenge to the intelligent identification of defects driven by data.Therefore,focusing on the urgent needs of identifying the location,category,and shape of engineering internal defects based on actual GPR data,researching unsupervised translation of GPR data and intelligent identification methods for internal defects in engineering structures will be helpful for accurate diagnosis of underground engineering structural defects,effective evaluation of engineering health status,it is of great significance to ensure the safe operation and longevity of the engineering.The main research contents and contributions of this paper are as follows:(1)Aiming at the inherent problem that the deep learning model trained on simulated GPR data is difficult to generalize to GPR data,a "actual-’resemble-simulated’" GPR image translation method based on unsupervised deep learning is proposed.The actual radar image is converted into a " resemble-simulated GPR image ",which retains the internal defect position and waveform information while incorporating the style of the simulated GPR image.Geometric consistency constraints are introduced into the network model to avoid discontinuous and distorted defect waveforms during image translation.Experimental verification is carried out on simulated GPR images and a variety of actual GPR images,which proves the universality of the translation algorithm for GPR images.(2)Aiming at the requirement of accurate identification of defect type and location,a GPR image identification method based on target detection network is studied.Taking the translated GPR image as the input of the network,there is no need to mark the actual GPR image,and the target detection model based on the simulation GPR image training is used to identify the translated " resemble-simulated GPR image ",the type identification and location of internal defects in actual GPR images are realized indirectly.Using GPRs with different center frequencies and manufacturers to collect actual GPR images in various scenarios for experimental verification,it proves the effectiveness of the translation algorithm for intelligent identification of internal defects in actual GPR images.(3)Aiming at the problem of defect boundary and contour identification,the "image-data"conversion method of the resemble-simulated GPR image after translation is studied,which provides support for mining internal physical law information based on GPR data,and solves the "stuck neck" problem that is difficult to describe the outline of internal defects based on actual ground-penetrating radar data.On this basis,a defect contour intelligent identification network integrated with the back-projection imaging branch was created,and the backprojection imaging data was used to impose prior constraints on the network to guide the network to better learn defect contour information and achieve defect contour characterization.Simulation experiments,model tests and field tests are carried out comprehensively to verify the effectiveness of the proposed method for the contouring of internal defects in GPR data.