| dc.description.abstract | Monitoring pavement conditions using crowdsourced vehicular data can significantly contribute to real-time and cost-effective pavement maintenance decision-making. This paper presents the development of various machine learning models for predicting crucial pavement characteristics essential for ensuring roadway safety, including roadway longitudinal grade, cross slope, international roughness index, surface rutting, and pavement skid resistance. The study collected vehicular sensing data and paired it with field pavement characteristics obtained using innovative instruments on selected asphalt and concrete sections in Oklahoma. Seven machine learning models were trained using the AutoGluon platform, yielding highly accurate predictions for safety-related roadway characteristics. The weighted ensemble L2, random forest, and category boosting (CatBoost) models exhibited the highest accuracy, with R-squared values exceeding 0.9, while the k-nearest neighbor algorithm and LightGBM models showed lower competitiveness. The inference latency of the models varied, with CatBoost demonstrating the lowest latency and weighted ensemble L2 achieving the highest accuracy at the expense of slightly higher inference latency. The choice of model depends on the specific application, whether it be pavement network management or real-time roadway condition monitoring. The findings from this research empower transportation agencies to efficiently screen the pavement network for further inspection or maintenance, thus enhancing transportation safety by providing instant alerts to drivers about potential high-risk pavement sections, resulting in safer and more reliable transportation infrastructure. | |
