Abstract
Excessive vibration of stay cables in strong winds has been a concern for bridge operators, which impairs the durability of both the cables and the bridge structure. This paper develops a data-driven approach to predict the amplitude of the cable vibration using an ensemble learning model. The model aims to predict cable vibrations in both in-plane and out-of-plane directions, with the wind speed, wind direction, turbulence intensity, and deck acceleration as input variables. Especially, the deck acceleration is included considering the deck-cable interaction and vehicle effects, which significantly improved the accuracy of the prediction. Furthermore, the model is interpreted with local interpretable model-agnostic explanations (LIME) and partial dependence plot (PDP) methods. The former demonstrates the relative importance of input variables on a global scale, and the latter indicates the correlation between individual input variables with the prediction target. The investigation is validated using the data harnessed from structural health monitoring (SHM) of a 1088-m cable-stayed bridge during three typhoon events. The adopted Gradient boosting regression tree (GBRT) model demonstrated better performance than other state-of-the-art machine learning models. The developed approach can provide guidance on preventive maintenance of stay cables to avoid damage due to excessive vibration.