Machine learning-based technology for asphalt concrete pavement performance decision-making in hot and humid climates

Abstract

Highway state agencies incur significant budget savings through optimal allocation of pavement Maintenance, Rehabilitation, and Reconstruction (MR&R) activities. These activities require robust prediction models that can handle large-scale, real-world data and can forecast pavement performance in the long run. Unfortunately, the traditional performance prediction models have been questionable in terms of efficiency and accuracy, are based on a limited number of explanatory variables, and are designated to predict short-term (up to five years) pavement conditions. Therefore, the goal of this study was to propose a machine learning-based technology that can predict the field performance by up to 11 years of Asphalt Concrete (AC) overlays placed on asphalt pavements in Southern states in the US based on key project conditions. The proposed technology resulted from assessing the prediction accuracy of machine learning algorithms, including Decision-Tree (DT), eXtreme Gradient Boosting (XGBoost), Artificial Neural Network (ANN), and ensemble-learning method, in forecasting the Pavement Condition Index (PCI) as the pavement performance indicator. For each algorithm, six models were developed sequentially based on historical pavement condition data collected from the Louisiana Department of Transportation and Development (LaDOTD) Pavement Management System (PMS) database. The six models learned from 892 log miles of randomly placed AC overlay sections in Louisiana. The output of these models was the future PCI of AC overlays at a biannual rate from one to 11 years. The findings showed that XGBoost and ensemble learning showed similar performance during model training and were further evaluated using the testing dataset. During model testing, the ensemble learning method yielded higher prediction accuracy than other algorithms, with R2 values decreasing from 0.77 at age 1 to 0.67 at age 11, Root Mean Square Error (RMSE) values increasing from ±1.65 to ±4.74, and Mean Absolute Error (MAE) increasing from 1.24 to 4.59. The resulting models can be used by state agencies to replace traditional prediction techniques.