A Deep Learning Tool for the Assessment of Pavement Smoothness and Aggregate Segregation during Construction

Pavement construction monitoring and quality assurance (QA) practices are mostly based on costly, discrete, and destructive methods. Most quality assurance programs are based on pavement construction procedures encompassing in-situ coring for layer thickness determination, density measurements, laboratory testing to measure volumetric properties, and smoothness measurements in case of the availability of a profiler. The main objective of this study was to develop a machine learning-based classifier for predicting pavement roughness and aggregate segregation based on digital image analysis, image recognition, and deep learning machine models. The developed Convolution Neural Networks (CNN) models were trained, tested, and validated using 600-pavement surface images extracted from the Louisiana Department of Transportation and Development (LaDOTD) Pavement Management System (PMS) and 129 pavement images collected from three construction sites a few days after paving. These images were randomly divided into 70%, 15%, and 15% for the training, testing, and validation phases, respectively. The roughness model achieved 93.8% and 92.6% accuracy in the training and validation stages; respectively, and predicted the International Roughness Index (IRI) values with a coefficient of determination R² of 0.98 and a Root-Mean Square Error (RMSE) of 3.5%. In addition, the developed image-processing model for the detection of aggregate segregation achieved adequate accuracy. Furthermore, the developed segregation detection procedure adequately described the relationship between mix density and segregation.路面施工监测与质量保证（QA）作业大多依赖成本高昂、离散且具有破坏性的检测手段。当前多数质量保证流程基于路面施工程序构建，涵盖用于测定层厚的现场取芯、密度测量、用于检测体积特性的实验室试验，以及在配备路面断面检测仪（profiler）时开展的平整度检测。本研究的核心目标是基于数字图像分析、图像识别与深度学习模型，开发一款用于预测路面平整度与集料离析的机器学习分类器。所开发的卷积神经网络（CNN）模型，采用从路易斯安那州交通与发展部（LaDOTD）路面管理系统（PMS）中提取的600幅路面表面图像，以及从3个施工路段摊铺数日后采集的129幅路面图像，完成训练、测试与验证流程。上述图像被随机划分为训练、测试与验证三个阶段的数据集，占比分别为70%、15%与15%。该平整度模型在训练与验证阶段的准确率分别达到93.8%与92.6%；其预测的国际平整度指数（IRI）值的决定系数R²为0.98，均方根误差（RMSE）为3.5%。此外，所开发的用于集料离析检测的图像处理模型同样达到了良好的检测精度。进一步而言，所构建的离析检测流程能够充分表征混合料密度与离析之间的关联关系。