Recognition of hidden distress in asphalt pavement based on convolutional neural network

Abstract

ABSTRACT Ground penetrating radar (GPR) is widely used in nondestructive tests of asphalt pavements, but the characteristics of its detection data need to be extracted manually, and it is very difficult to determine the hidden distress efficiently and accurately inside the asphalt pavement. Therefore, to recognize hidden distress inside asphalt pavement intelligently and improve the efficiency and accuracy of GPR for asphalt pavement detection, this research used the convolutional neural network (CNN) algorithm combined with GPR detection technology. Significantly, the image of the dielectric properties distribution of the road was obtained by processing the GPR detection data, and the GPR detection image data set was established. Moreover, this research dealt with the imbalance of the data set by weighting the loss value. Next, this research designed a CNN model with a simple structure, called visual GPR (VGPR), and used the GPR image data set to train, validate, and test the VGPR. Finally, the model was fine-tuned according to the results, and the optimal hyperparameters were selected. The results show that VGPR converges fast in the training process, and no obvious overfitting phenomenon exists. During the test, the weighted average F1 score in the comprehensive metrics reached 99.626%, indicating that VGPR has excellent generalization performance. When VGPR uses a graphics processing unit for computing, the image recognition efficiency can meet the engineering requirements. A comparison of VGPR with other CNN models reveals that when using the same GPR detection image data set, VGPR has the best recognition effect. In summary, the method detailed herein can quickly and accurately identify the hidden distress of the asphalt pavement and provide a guarantee for the maintenance and operation of the expressway.