To address the dual challenges of scarce historical time-series data and limited representational capacity of standalone models in pavement performance prediction, this study proposes an Engineering-heuristic-constrained Perturbation Data Augmentation Framework and a hybrid Bidirectional Long Short-Term Memory–Extreme Gradient Boosting (BiLSTM–XGBoost) model. The augmentation framework generates high-quality virtual samples by applying controlled perturbations aligned with engineering variability—to both covariates (e.g., traffic volume and layer thickness) and Pavement Condition Index (PCI) sequences—while enforcing the physical constraint of monotonic year-on-year deterioration. This expands 10 typical road sections into 1200 training samples. A two-stage prediction architecture is then developed: BiLSTM first extracts high-order temporal features from historical PCI sequences; these features are then fused with covariates and engineering features as input to XGBoost for final regression. Evaluated on an independent test set, the hybrid model outperforms the standalone models and the ANN model, achieving an R2 of 0.771, with RMSE, MAE, and MAPE as low as 2.043, 1.706, and 1.859%, respectively. This work provides an accurate and practical tool for pavement performance prediction under data scarcity, supporting informed decision-making in pavement management systems.
Zhou et al. (Sat,) studied this question.