Predictive modeling of pavement final service life using machine learning techniques
12th International Conference on Bearing Capacity of Roads, Railways, and Airfields, BCRRA 2026, Ljubljana, Slovenya, 22 - 24 Haziran 2026, cilt.99, ss.1300-1306, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Cilt numarası: 99
- Doi Numarası: 10.1016/j.trpro.2026.07.182
- Basıldığı Şehir: Ljubljana
- Basıldığı Ülke: Slovenya
- Sayfa Sayıları: ss.1300-1306
- Anahtar Kelimeler: machine learning, pavement condition index (PCI), pavement management systems (PMS), pavement service life, predictive modeling, random forest
- Süleyman Demirel Üniversitesi Adresli: Evet
Özet
Pavement Management Systems (PMS) require reliable estimates of the service life achieved after rehabilitation to prioritize projects and allocate budgets effectively. However, post-rehabilitation service life is difficult to model with conventional statistical methods because it is driven by interacting structural, condition, and traffic-related factors and often exhibits nonlinear behavior. This study addresses the practical problem of predicting the Final Service Life (Final SL) of rehabilitated pavement sections using routinely collected PMS indicators. A dataset comprising 141 sections was used, including condition indices (PCI, RCI, RI), rutting measures, overlay thickness, and average daily traffic (ADT). Three predictive approaches (Linear Regression, Random Forest, and XGBoost) were compared under an 80/20 train–test split using MSE, MAE, and R². Among the evaluated models, Random Forest provided the most accurate and stable predictions (MSE = 13.46, R² = 0.60), improving upon Linear Regression (MSE = 17.09, R² = 0.48) and XGBoost (MSE = 15.12, R² = 0.55). Importantly, the analysis identified the dominant drivers of post-rehabilitation service life—PCI change, post-treatment PCI, rutting severity, overlay thickness, and traffic load—offering interpretable insights for maintenance decision-making. Overall, the results indicate that ML-based prediction can support more proactive and evidence-based rehabilitation planning within PMS. Future work will investigate robust validation and hyperparameter/ensemble strategies to further enhance generalization and reliability.