Research On Ground Penetrating Radar Echo Data Processing Based On Deep Learning

Ground penetrating radar(GPR)is a nondestructive detection method using electromagnetic radiation to locate shallow geological subsurface features and underground utilities buried in the ground.Compared with other conventional underground detection methods,it has the advantages of fast detection speed,continuous detection process,high resolution,easy operation,low detection cost and wide detection range.Therefore,it has been one of the hotspots in international academic circles.Quantitative analysis of attributes of underground environment by GPR signal imaging processing technology has always been the focus of scientific researchers in this field.However,due to the complexity of detection environment,the diversity of geometric shape and dielectric properties of underground obstacles,and the interference of various noises,the existing technology can not accurately quantitatively analyze the underground environment.In order to improve the GPR's capability of quantitative analysis and semi-quantitative analysis for targets,an in-depth research on deep learning and GPR technology has been carried out in this paper.The main works are listed as below.(1)The research on basic theory and data preprocessing of ground penetrating radar(GPR).This paper introduces the working principle and mode of ground penetrating radar(GPR),introduces the forward numerical simulation software GprMax,and uses GprMax to simulate the detection model of GPR,and analyses the direct wave removal algorithm based on the echo data experiment.(2)The research on GPR imaging technology based on time domain back projection(BP)imaging technology.According to the problem of high sidelobe and clutter energy,this paper improves the simple addition of the original algorithm into weighted addition,and proposes to calculate the weight coefficient using cosine distance.A backward projection algorithm CBP based on cosine theorem is proposed.At the same time,the idea of 0-1 coding and step processing is combined with the traditional BP algorithm,and the backward projection algorithm EBP based on coding is proposed.The theoretical analysis and simulation data experiments verify the superiority of the two algorithms.At the same time,by proposing two new algorithms to improve the original algorithm,the analysis shows that the ground penetrating radar echo data has strong local correlation,which lays a foundation for subsequent research.(3)The research on recognition algorithms of underground obstacles based on convolutional neural network(CNN).In this paper,the CNN structure design for underground target recognition is carried out.We have designed three sets of experiments to verify the ability of the CNN network to identify the shape,properties and comprehensive properties of underground targets.The feasibility and superiority of this idea are verified from both accuracy and real-time.(5)The research on imaging algorithms of underground obstacles location based on Faster R-CNN.Combined with the target detection algorithm of deep learning,this paper proposes an underwater target location recognition algorithm based on Faster R-CNN.Through the training and structure optimization of Faster R-CNN,the automatic localization and recognition of underground targets based on ground penetrating radar simulation data is realized.The feasibility of the underground target location method based on Faster R-CNN is verified by experimental results.