Research On Inversion Of Ground Penetrating Radar Data And Automatic Detection Of Road Diseases Based On CNN

Ground penetrating radar is widely used in engineering and environmental detection due to the advantages of efficient,accurate and non-destructive detection,such as road structure and quality detection,environmental pollutant investigation and detection.The traditional ground penetrating radar velocity inversion is a nonlinear and ill-conditioned optimization problem.The method to solve this ill-conditioned inversion problem is usually to continuously optimize the target between the forward data of the inversion model and the real data through an iterative algorithm.Therefore,there are two main problems: one is the low efficiency of the inversion,and the other is the accuracy of the forward algorithm.In this paper,a method based on end-to-end self-encoding structure convolutional neural network for inversion imaging of common-offset ground penetrating radar data velocity is studied.This deep learningbased ground penetrating radar inversion method does not require forward calculation,but transforms the nonlinear ill-conditioned optimization problem into a mapping problem from the ground penetrating radar data space to the physical model space of the underground medium.First,a sufficient number of ground penetrating radar data with markers are simulated through forward modeling;Then,the neural network model is trained to learn the intrinsic mapping relationship between the ground penetrating radar data,which contains information such as amplitude,waveform,and two-way travel time,and the physical property data of the underground medium.Using the trained neural network model,the ground penetrating radar data simulated by different media models are inverted and tested,including homogeneous layered medium,undulating interface layered medium,and stochastic equivalent medium with undulating interface established by small-scale non-uniformity ellipsoid autocorrelation function,the results demonstrate the feasibility of the method.Finally,two measured ground penetrating radar data are inverted and predicted and imaged.One is the measured road ground penetrating radar data,which inverts the road structure and speed information;the other is the ground penetrating radar data of environmental pollutant investigation,the range and boundary of pollutants can be delineated through the inversion results.After the above experiments,the results show that once the training of the neural network model is completed,this convolutional neural networkbased common-offset ground penetrating radar data inversion method can directly and quickly and accurately invert the velocity structure and various information of the underground medium,such as dielectric properties inside the layer.With the rapid development of my country's highway industry and the wide application of road ground penetrating radar,a large amount of measurement data has been generated,and the identification of road diseases is usually manual,inefficient and time-consuming.Therefore,this paper proposes a method based on Center Net to achieve automatic identify road disease from ground penetrating radar images.Firstly,the models of three common types of road diseases were constructed by referring to the literature,and the corresponding ground penetrating radar data images of the diseases were simulated by forward modeling.In order to prevent the problem of over-fitting and enhance the generalization ability of the model,the data set is augmented by data augmentation.Then,a strategy of transfer learning is introduced to train fine-tuning on a pretrained neural network with a dataset consisting of simulated GPR images.Finally,the trained target detection network model is automatically detected and recognized on the test set and the measured road GPR images.The results show that the method can achieve better detection and recognition results in the case of less sample data sets,which proves the effectiveness of the method,which can quickly census diseases from road ground penetrating radar detection images.