Guiding preventive maintenance timing decisions using interpretable prediction of pavement performance

Pavement preventive maintenance timing decisions relies on accurate pavement performance prediction and benefit-cost analysis. However, current research gaps persist in incorporating interpretable performance forecasting for fine-grained spatial dimension while neglecting systematic evaluation of lifecycle cost-effectiveness. This study proposes an interpretable tree-based model to achieve 100-meter pile prediction of the Pavement Condition Index (PCI) for guiding preventive maintenance timing decisions. Firstly, we integrate multi-source data, including pavement materials, climate, traffic loads, and historical performance, to train machine learning models. Secondly, hyperparameter optimization enhances sensitivity to multivariate factors, achieving a prediction accuracy of 0.8173. Simultaneously, SHapley Additive exPlanations (SHAP) was employed to quantitatively analyze feature importance, thereby improving the transparency of the models. Finally, by evaluating benefit-cost ratios under varying PCI thresholds, we identify that initiating preventive maintenance at a PCI of 87 minimizes intervention frequency while maximizing cost efficiency. This strategy ensures that the PCI value remains above the minimum acceptable threshold throughout the analysis period, thereby significantly supporting long-term maintenance expenditures. The proposed framework provides an interpretable, data-driven basis for pavement preventive maintenance, contributing to sustainable infrastructure management.