Modeling Pavement Performance and Distresses: A Machine-Learning Approach

This paper presents the development of three robust machine learning models aimed at predicting pavement performance in terms of three different indicators. International Roughness Index (IRI), fatigue and longitudinal cracking were investigated. The study started by leveraging a comprehensive dataset extracted from the Long-Term Pavement Performance (LTPP) database incorporating a range of significant variables such as age, the duration since last maintenance, temperature, precipitation levels, pavement material properties, Equivalent Single Axle Loads (ESALs), and the current state of fatigue and longitudinal cracks. Regions with dry and non-freeze climates were selected. To develop the three predictive models of IRI, fatigue cracking, and longitudinal cracking, each model was subjected to a group of six machine learning algorithms. XGBoost, Random Forest, Decision Trees, Bayesian Regression, K Nearest Neighbors (KNN), and Ridge Regression were implemented for each model to identify the most promising algorithm. Then, each algorithm performance was tested and evaluated using two metrics, R² and Mean Absolute Error (MAE). The results of the analysis indicated that the XGBoost algorithm outperformed its counterparts across all output variables, offering the lowest MAE of 0.17 and R² of 0.729 for modeling IRI. It also developed an R² of 0.672 and 0.692 and MAE of 4.95 and 2.96% for fatigue and longitudinal cracking, respectively. The three developed predictive models using the machine learning technique can provide stakeholders in the field of transportation management with reliable tools to predict pavement conditions and hence can help in properly allocating maintenance funds and save costs. The findings of this study contribute to paving the way for more integration of machine learning in the maintenance and construction of durable road infrastructure.