| Abstract: |
Bridge infrastructure globally is at an increasing risk of fatigue-related deterioration, and shortening of service lives due to increased traffic loads, increased environmental aggressiveness, and the ageing of many existing structures. Conventional empirical and semi-empirical deteriorations models such as linear damage accumulation (Miner's rule) and mechanistic finite-element simulations are resource-intensive, dependent on idealized material assumptions, and lack scalability across heterogeneous bridge inventories. To this end, in this paper a rigorous empirical study is developed to explore the forecasting performance of six popular machine learning (ML) algorithms—Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), Artificial Neural Network (ANN/MLP), Long Short-Term Memory( LSTM) and Gaussian Process Regression(GPR)—to predict remaining fatigue life & annual deterioration rate for steel girder and reinforced/prestressed concrete bridges. A total of 1,465 bridge records were used in this study by compiling datasets from Federal Highway Administration National Bridge Inventory (FHWA-NBI), Indian Bridge Management System (BMS), and IABSE technical reports. Through correlation analysis and importance ranking by SHAP, we identified 18 engineered features that represent structural geometry, material properties, traffic loading, environmental exposure, and inspection history. XGBoost achieved the best R² (0.941 on the test set) as well as lowest RMSE (3.09 years) was the top algorithm in predicting fatigue life, beating all competing algorithms. For the temporal deterioration rate estimation task, LSTM exhibited better performance compared to wDNN (R² = 0.926). The results confirm the potential of using ensemble tree-based and recurrent deep-learning architectures for data-driven bridge lifecycle assessment with meaningful implications for national bridge management and preventative maintenance scheduling. |
