Presentation Information

[Structure-10]Machine Learning Approaches for Predicting Asphalt Pavement Structural Integrity Using PMS Data

*Kazutaka Suzuki1, Kiyoyuki Kaito2, Kotaro Sasai2 (1. NEXCO-CENTRAL (Japan), 2. Osaka University Graduate School of Engineering (Japan))

Keywords:

asphalt Pavement Structural Integrity,road surface condition measurement data,machine learning,preventive maintenance,FWD

In Japan, asphalt pavement repairs on expressways tend to become more frequent over time due to fatigue cracking that develops from the bottom of the upper base course, as well as damage that deepens when water penetrates through surface cracks.To prevent such deterioration, it is essential to implement preventive maintenance before the damage progresses to deeper layers, which requires an accurate assessment of the structural integrity of the pavement. For example, NEXCO, which manages Japan’s expressways, evaluates pavement structural integrity using a damage index (Di) obtained through Falling Weight Deflectometer (FWD) measurements. However, FWD measurements require traffic restrictions, and the measurement process is costly and labor-intensive, making it difficult to perform evaluations at high density and high frequency.We have been developing a predictive model for the damage index (Di) using machine learning based on data from the Pavement Management System (PMS). In this study, we examined explanatory variables suitable for prediction and optimized hyperparameters for an existing neural network (NN) model to further improve its predictive accuracy. Furthermore, SHAP analysis was employed to calculate the contribution of explanatory variables such as pavement structure, road surface condition measurement data (rutting, IRI, cracking ratio), and heavy vehicle traffic volume to the prediction results of the damage index (Di), and the results were examined. Finally, we developed a model applying LightGBM, which is considered to be faster in training and inference and to achieve higher predictive accuracy compared to NN, and conducted a comparison of prediction accuracy with the NN model.As a result of applying these models, although further validation is required, the findings indicate the potential to quickly assess the structural integrity of asphalt pavement over a 250 km section of expressway.