Loading-environment coupled prediction of pavement performance based on RIOHTrack data and PSO-XGBoost modeling

Abstract

This study develops an interpretable pavement performance prediction framework to quantify the coupled effects of loading and environmental factors on rutting and deflection. A high-quality dataset was established through preprocessing. Empirical Mode Decomposition was then used to separate trend, periodic and noise components. Key variables were selected using feature importance and distance correlation. PSO-XGBoost, Random Forest (RF) and Backpropagation Neural Network (BPNN) models were compared using data before and after denoising. SHAP was applied to interpret the best-performing model. The results demonstrated that denoising improved the prediction accuracy of rutting and deflection by 6.98% and 20.08%, respectively. PSO-XGBoost outperformed RF by 0.55% and 5.58% and BPNN by 8.37% and 16.53% in predicting rutting and deflection, respectively. SHAP showed that rutting was mainly driven by cumulative axle counts and aggravated by high temperature and humidity. Deflection was more sensitive to temperature, while higher deflection values were more likely to induce rutting. This framework enables refined analysis of pavement performance evolution.