Speaker
Description
Wind-induced vibrations pose significant challenges to the stability and serviceability of long-span cable-supported bridges. This study presents an innovative data-driven framework based on Ensemble Bayesian Neural Networks (BNNs) to predict vertical and torsional dynamic responses of the Hardanger Bridge in Norway under severe windstorms. The approach addresses critical issues in Structural Health Monitoring (SHM), including multicollinearity among wind-related features and the nonlinear interactions between environmental conditions and bridge responses. The proposed method employs Variance Inflation Factor (VIF)-based feature selection and predictor normalisation to enhance model stability. An ensemble of BNNs is trained through rigorous hyperparameter optimisation and k-fold cross-validation to improve generalisation and capture both aleatoric and epistemic uncertainties. The model is validated using field measurements from multiple storms, where 20 neural models collectively forecast wind-induced RMS acceleration responses with high fidelity. Compared to conventional machine learning approaches, such as Regularised Support Vector Machines, the Ensemble BNN framework exhibits superior performance in test scenarios, achieving R² values exceeding 92% for vertical and 94% for torsional predictions. Results confirm its robustness and scalability in realistic SHM conditions. This study not only advances the predictive capabilities for bridge dynamics but also provides a reliable foundation for intelligent decision-making in bridge safety management.
