Improving preventive maintenance strategies for enhanced pavement friction restoration performance via explainable machine learning

Maintaining sufficient pavement surface friction is crucial for roadway safety, particularly under wet conditions. Preventive Maintenance (PM) treatments have proven effective in restoring pavement friction levels and reducing the likelihood of crashes. However, capturing the friction deterioration behavior after different PM treatments remains challenging due to complex interactions among traffic, climate, and pavement conditions. This study analyzed pavement friction number (FN) using data from the Long-Term Pavement Performance (LTPP) Special Pavement Study 3 (SPS-3) to evaluate the effectiveness of 4 PM treatments at improving and maintaining high levels of friction. Four machine learning models, namely Artificial Neural Network (ANN), Support Vector Regression (SVR), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost), were developed to capture the complex relationships between 10 input features and FN. The SHapley Additive exPlanations (SHAP) was employed to interpret model predictions and quantify the contribution of each input and interaction variables. Among all models, XGBoost achieved the best performance, with an R 2 of 0.76 and RMSE of 5.46 on the test dataset. Results indicate that slurry seals are the most effective treatment for improving pavement friction over a five-year period, followed by chip seal. The XGBoost model, combined with a cluster-based data splitting strategy, provided the most robust and accurate FN predictions. Sensitivity analysis shows that slurry seal maintains FN above the investigatory level (FN = 40) when Annual Average Daily Truck Traffic (AADTT) is below approximately 2000 trucks/day, but its effectiveness declines sharply beyond this threshold. Additionally, pre-treatment surface condition strongly influences pavement friction performance, with higher International Roughness Index (IRI) values corresponding to degraded texture and lower friction. Finally, the SHAP framework offered interpretable insights into data-driven pavement management decisions. These findings provide transportation agencies with actionable insights to guide treatment selection and optimize maintenance planning for safer pavement management. • Effectiveness and cost-effectiveness of Preventive Maintenance were evaluated for pavement friction restoration. • Machine Learning models were trained and tested for predicting long-term pavement friction performances. • SHAP framework offers a clear, interpretable understanding of how key factors influence skid resistance.