| Abstract: |
For civil infrastructure managers worldwide, the structural integrity of steel and concrete bridges is a critical concern. Fatigue life and deterioration rates have become crucial in determining maintenance schedules, especially with monitors for aging bridge networks and hundreds of thousands of bridges across the United States being rapidly overloaded. Although traditional empirical methods and physics-based models are cornerstones in the degradation mechanism modeling, they struggle with effectively capturing the complex nonlinear nature of deterioration for real-world bridge structures. For the last twenty years, machine learning (ML) algorithms have provided revolutionary data-driven high-fidelity predictive capabilities that are complementary to and often surpass traditional approaches. This review paper is a meta-analysis of the studies available in existing literature on predicting fatigue life and deterioration of bridge infrastructure using ML approaches such as artificial neural networks (ANN), support vector machine(s) (SVM), random forests (RF), gradient boosting methods(GBM, XGBoost, Light GBM)), convolutional neural networks (CNN) and physics-informed neural networks(PINN). This review (n = 30 studies from 2005–2024) systematically evaluates methodology, data-set characteristics, predictive accuracy and practical deployment barriers. They demonstrate that, when used alone on their own, single-algorithm approaches are outperformed consistently by both ensemble learning and deep learning models achieving MAE reductions as high as 35% compared to regression baselines. Important shortcomings include limited bridge typology generalization, minimal real-world sensor deployment and no standard benchmark datasets. The paper ends with directions for future research towards hybrid physics-ML frameworks, federated learning for distributed monitoring networks, and explainable AI (XAI) to garner engineer trust in decision support. |
