Machine learning models for predicting VIV amplitude of streamlined steel box girders

Abstract

The vortex-induced vibrations (VIVs) can cause severe consequences for long-span bridges, including structural fatigue, human-body comfort and vehicle safety, and become significant issues in terms of the wind-resistance. Currently, estimations of VIV amplitudes are typically based on wind tunnel tests, field monitoring, computational fluid dynamics and theoretical models. In this paper, a machine learning (ML) approach is developed for rapidly estimating the VIV amplitude of a streamlined steel box girder during preliminary structural design. The VIV amplitude forecasting models are constructed using four well-established and widely used ML algorithms, e.g., support vector machine (SVM), random forest (RF), gradient boosting decision tree (GBDT), and artificial neural network (ANN). These four models take geometric and dynamic features as inputs, and output the VIV amplitude. The hyper-parameters are optimized to attain acceptable accuracy. To address the low interpretability of complex ML models, this paper introduces the Shapley additive explanations (SHAP) method to explain the VIV prediction pattern. The results indicate that the constructed ML models with optimal hyper-parameters show excellent performance for predicting the VIV amplitude in a validated and accurate manner. The SHAP explanation supports these results and offers valuable additional insights.