基于极限梯度提升和探地雷达时频特征的水泥路面脱空识别

Cement Pavement Void Identification Based on XGBoost and GPR Time-frequency Features

摘要:

针对探地雷达（GPR）数据解译依赖于人工经验，存在费时费力和主观偏差的问题，提出了基于极限梯度提升（XGBoost）和GPR时频特征的水泥路面脱空识别方法。采用正演模拟、室内试验和现场试验获得了脱空病害数据源，建立含有标签的脱空GPR数据集；通过重采样方法统一GPR数据采样频率，并对预处理后的GPR数据进行时频域特征提取，建立了包含18个时域和12个频域特征的数据集。以时频域特征为输入，是否存在脱空病害为输出，采用XGBoost算法构建脱空识别模型，并与随机森林（RF）和人工神经网络（ANN）算法进行对比。结果表明，模型的识别准确率排序为XGBoost（98.10%）>ANN（95.10%）>RF（93.17%），XGBoost模型识别精度最高，并能在实际路面上准确定位脱空区域。

Abstract:

Ground penetrating radar （GPR） is an effective method of void detection， but GPR data interpretation depends on human experience， being time consuming and laborious， or even existing subjective bias. To address above issues， a cement pavement void identification method based on XGBoost and GPR time-frequency features was proposed. To automatically identify cement pavement void area， the finite difference time domain method， lab and field tests were carried out， and GPR void dataset with label was created. Then， the resampling method was used to obtain the same sample frequency. Thirty time and frequency domain features， including 18 time-domain features and 12 frequency-domain features， were extracted from the post-processed GPR data. Taking the time-frequency domain feature as an input， and void label as an output， XGBoost was used to build a void identification model. The random forest （RF） and artificial neural network （ANN） were also trained to compare with XGBoost. The comparison results indicate that the accuracy ranking is XGBoost（98.10%）>ANN（95.10%）>RF（93.17%）. The accuracy of the XGBoost method is the highest and verified by field tests.