Research on a prediction model for asphalt pavement deflection in full-scale test track integrating multi-source features and machine learning

Abstract

To ensure asphalt pavement durability under environmental and traffic loads, this study investigates permanent deformation caused by temperature changes and repeated axle loads. Based on long-term full-scale track monitoring, a feature fusion model integrating cumulative load passes and temperature was developed. We evaluated XGBoost, ANN, Stacking, and LSTM models for deformation prediction. Results show the Stacking model achieved the highest accuracy, with a test R 2 of 0.902 and an RMSE 18.95% lower than linear regression. XGBoost demonstrated better stability, while LSTM and ANN exhibited limitations with sequential patterns. SHAP analysis identified pavement structure, axle load, and ambient temperature as the dominant factors. This study investigates pavement deflection as an indicator of structural performance under temperature changes and repeated axle loads, with implications for permanent deformation assessment.