Machine learning-based smoothness prediction for asphalt concrete overlay pavements considering milling effects

초록

The smoothness of asphalt concrete (AC) overlay pavements has often been predicted using models originally developed for new AC pavements. However, these two pavement types exhibit distinct performance behaviors, and the presence or absence of milling strongly influences performance. The regression model proposed by the American Association of State Highway and Transportation Officials (AASHTO) has limitations in representing the unique characteristics of AC overlays. To address this gap, this study developed machine learning models for predicting AC overlay smoothness using three algorithms: random forest (RF), extreme gradient boosting (XGB), and TabNet. The number of predictive features was logically reduced by incorporating pavement distress indicators that reflect combined design effects. Two modeling approaches were adopted: an integrated model including all overlays and separate models distinguishing between milling and non-milling. The RF model from the integrated approach yielded the most accurate and stable predictions (R2 = 0.854). The initial international roughness index (IRI) was the most influential feature, followed by major distress variables. Contrary to expectations, non-milled overlays exhibited lower IRI values, suggesting better smoothness. The separate RF models for milled and non-milled sections achieved predictive performance comparable to the integrated RF model. In non-milled sections, initial IRI primarily governed smoothness, whereas early deterioration was more pronounced in milled sections, particularly those with lower initial IRI. These findings indicate that the separate models better capture the distinct behaviors of milled and non-milled overlays.

키워드