Data-driven prediction and interpretation of fatigue damage in a road-rail suspension bridge considering multiple loads

In long-span suspension bridges, fatigue is a significant concern for steel members as they are continuously under multiple effects such as wind, temperature, and traffic loads. Therefore, it is important to quantify the fatigue damage at critical details for effective maintenance countermeasures. This paper proposes an ensemble machine-learning-based method with physical interpretation, which predicts fatigue damage with monitored parameters of temperature, wind, train, and roadway loads. Interpretation of the machine learning model is achieved regarding the different contributions of the monitored parameters to the fatigue damage. Datasets from the structural health monitoring (SHM) system of a long-span rail-road suspension bridge are utilized to verify the proposed method. They are divided into 80% of training datasets and 20% of testing datasets. The ensemble machine learning model is established with Gradient Boosting Regression Tree (GBRT), and the fatigue damage for top chords on the upstream and downstream sides is predicted. The results are compared with other state-of-the-art machine learning algorithms, namely, artificial neural network (ANN), support vector machine (SVM), decision tree (DT), and random forest (RF). It is found that GBRT demonstrates the highest accuracy for fatigue damage prediction with R-2 of 0.9016 and 0.9180, respectively. Furthermore, interpretation of the results is carried out with feature importance estimates (FIE) and individual conditional expectation (ICE). FIE results indicate that the train load has the most significant influence on fatigue damage of the suspension bridge, and ICE results can provide insights on better maintenance plans to avoid severe fatigue damage.

Automated summary

This paper proposes a machine-learning-based method to predict fatigue damage in long-span suspension bridges using monitored parameters of temperature, wind, train, and roadway loads. The results show that the method achieves high accuracy and provides valuable information for maintenance work.