Evaluation Of Highway Subgrade Distresses Based On Ground Penetrating Radar And Convolutional Neural Network

Highway subgrade disease detection is an important part of the evaluation of road conditions.Accurate inspection results have a decisive influence on highway maintenance and maintenance decisions.Research shows that the pavement ground penetrating radar(GPR),combining properties of road materials and technology of GPR,shows more efficient and non-destructive than other non-destructive testing technologies(NDT).This paper uses ground penetrating radar technology to research and evaluate the diseases.Working principle of GPR and the main factors influencing the detecting veracity are introduced in this paper.And also,the signals of different road materials and pavement diseases(crack,void,soil loose)are summarized.The algorithms used to deal with signals and recognize different diseases in recent years are introduced emphatically.The data analysis of ground penetrating radar(GPR)technology in detecting subgrade distresses relies on manual identification,but the manually processed GPR image used for classifying distresses is inefficient and inaccurate at present.In this case,an application about convolutional neural networks was putted forward to classify subgrade distresses automatically.The cascade CNN is consisting of two convolutional neural networks,which were utilized to recognize distresses using low resolution images and high resolution images separately.The processes of developing convolutional neural networks mainly included training,validating,and testing.After developing the cascade CNN,training and testing results were used to verify stability of the cascade CNN.At last,the cascade CNN was compared with other methods to show its superiority.The results show that the accuracies of the cascade CNN in the training and validating are 97.46% and 95.80%.Stability analysis shows that the cascade CNN has great stability on both the emission frequencies and highway structures.In addition,comparing with the Sobel edge detection and K-value clustering analysist,the CNN-based method shows higher recognition accuracy.