Abstract: The automatic identification method for asphalt pavement structural damage based on ground penetrating radar (GPR) mapping and deep neural network has the problem of limited data volume and unbalanced distribution of types, and the accuracy and stability of identification still need to be improved. The pavement structural damage identification technology based on heterogeneous GPR mapping was proposed. GPR was used to collect structural cracks and interlayer discontinuous diseases of asphalt pavement, and the measured profile maps was obtained. Based on the time-domain finite difference method, the echo features of cracks and interlayer discontinuities in the homogeneous model were numerically simulated, and combined with the measured maps to form dataset 1#. The "asphalt-aggregate" two-phase medium model was constructed based on the CT scan images of core samples, and the echo features of cracks and interlayer discontinuities in the two-phase medium model were simulated, which was combined the measured maps to form dataset 2#. YOLO v5 deep neural network was trained by dataset 1# and 2#, respectively. The research results show that the mAP@0.5 tested in YOLO v5 model using datasets 1# and 2# are 93.79% and 96.33%, which demonstrates that heterogeneous mapping features can enrich network training samples and improve the identification accuracy of deep learning model.

Key words: highway engineering; structural crack; interlayer discontinuity; ground-penetrating radar; deep neural network

摘要： 基于雷达图谱与深度神经网络的沥青路面结构损伤自动辨识方法存在数据量少且种类不均衡的问题，识别准确性与稳定性仍有待提高。提出基于非均质雷达图谱的路面结构损伤识别技术。采用探地雷达采集沥青路面结构裂缝与层间不连续病害，获取实测剖面图；基于时域有限差分法，模拟裂缝与层间不连续在匀质模型中的回波特征，与实测图谱组成数据集1#；基于芯样CT扫描图构建“沥青-集料”二相介质模型，模拟裂缝与层间不连续在二相介质模型中的回波特征，与实测图谱组成数据集2#；采用数据集1# 和2#，分别训练YOLO v5深度神经网络。研究结果表明：数据集1# 和2# 在YOLO v5模型测试集上的mAP@0.5为93.79%与96.33%，证明非均质图谱特征可丰富网络训练样本，并提高深度学习模型识别的准确性。

关键词: 道路工程；结构裂缝；层间不连续；探地雷达；深度神经网络