TY  - EJOU
AU  - Wang, Mengyao 
AU  - Zhu, Ailian 
AU  - Zhu, Longji 
AU  - Lan, Chen 
AU  - Li, Yang 

TI  - DLPC-GNN A Dual-Layer Progressive Physics-Constrained Graph Neural Network for Asphalt Pavement Distress Prediction and Maintenance Strategy Classification
T2  - Computer Modeling in Engineering \& Sciences

PY  - 2026
VL  - 148
IS  - 2
SN  - 1526-1506

AB  - Accurate prediction of asphalt pavement distress is essential for proactive maintenance and life-cycle infrastructure management. However, existing data-driven methods often struggle to jointly represent multi-source inspection data, distress evolution mechanisms, and spatial propagation relationships among pavement sections. To address these limitations, this study proposes a Dual-Layer Progressive Physics-Constrained Graph Neural Network (DLPC-GNN) for asphalt pavement distress prediction and maintenance strategy classification. The proposed model represents pavement deterioration using a dual-layer graph structure. At the microscopic level, cracks, surface deterioration, and structural moisture-induced damage are modeled as physically associated distress nodes. At the macroscopic level, pavement-section supernodes are connected according to spatial distance, traffic-flow correlation, and structural continuity. To improve physical consistency and interpretability, seepage, fatigue crack growth, hydrodynamic pressure, and bounded deflection evolution mechanisms are incorporated into physics-informed edge initialization and auxiliary physical constraints. The model is validated using multi-source inspection data from the Huizhou section of the Huizhou–Shenzhen Expressway. Experimental results show that DLPC-GNN outperforms conventional machine learning models, deep learning models, and purely data-driven graph neural networks in both continuous distress-state prediction and maintenance strategy classification. For the continuous distress-state regression task, DLPC-GNN achieves an MAE of 2.78 and an RMSE of 4.62. For the maintenance strategy classification task, it achieves a Recall of 88.8%. Compared with the purely data-driven GNN, DLPC-GNN reduces MAE and RMSE by 15.1% and 21.4%, respectively, and improves Recall by 12.2%. Ablation and robustness analyses further demonstrate that the dual-layer graph structure, physics-informed edge initialization, and physical constraints improve prediction stability under small-sample, noisy, and feature-missing conditions. These results indicate that DLPC-GNN provides a physics-guided and partially interpretable graph-learning framework for maintenance-oriented asphalt pavement distress assessment.
KW  - Asphalt pavement; graph neural network; physics-constrained learning; distress evolution; hydro-mechanical coupling; maintenance strategy classification

DO  - 10.32604/cmes.2026.085279