Machine-learning prediction of pavement fatigue life ratio from structural inputs and TSDD measurements using physics-guided data generation

Fatigue damage is a critical mode of structural deterioration in flexible pavements, often initiating beneath the surface and progressing over time. While recent developments in traffic speed deflection devices (TSDDs) have enabled efficient network-level structural response measurement, current applications of TSDD data typically rely on simplified indices or threshold-based approaches that do not fully account for the mechanistic complexity of fatigue behavior. This study introduces a machine-learning (ML) prediction framework developed from physics-guided data generation to estimate the fatigue life ratio (FLR)—the ratio of remaining to initial fatigue life—using TSDD slope measurements combined with asphalt concrete (AC) thickness, damaged modulus, and binder type. A comprehensive analytical database was generated using viscoelastic simulations across actual pavement structures, five representative binders, and multiple damage levels. Separate Random Forest and XGBoost (extreme gradient boosting) models were developed to estimate AC modulus and FLR, respectively. Binder-specific FLR models were trained using 5-fold cross validation and integrated into a consolidated model capable of routing inputs based on binder type. The developed ML framework was evaluated using field data from a TSD (traffic speed deflectometer) over a temperature loop in Denmark, providing an initial field-based check of the AC modulus prediction component and illustrating how FLR estimation could be implemented using field TSDD measurements. The results demonstrate that FLR predictions informed by full slope profiles, damaged AC modulus, AC thickness, and binder classification outperform traditional single-index methods, providing a scalable and mechanistically consistent approach.