Data-driven dynamic response forecasting and anomaly detection in long-span bridges

Long-span bridges, due to their high flexibility and low damping capacity, are particularly vulnerable to hazardous vibrations, such as vortex-induced vibrations (VIVs). Effective monitoring and early warning systems require advanced predictive models capable of accurately forecasting dynamic responses and detecting anomalies. This study proposes a Bayesian-optimized bidirectional long short-term memory sequence-to-sequence regression model for predicting vibrational responses and detecting abnormal structural behavior in long-span bridges. The proposed framework is designed with two configurations: one optimized for accurate prediction under normal conditions and another tailored to capture abrupt changes for anomaly detection. The model is trained and validated using structural health monitoring and weigh-in-motion data from an operational bridge, incorporating key environmental and operational conditions. By leveraging Bayesian optimization for hyperparameter tuning, the framework improves predictive accuracy and generalization. The model’s performance is evaluated based on the prediction accuracy, the impact of time lag selection, and sensitivity of key input variables. Additionally, anomaly detection is achieved by analyzing residuals between predicted and measured responses, enabling early identification of hazardous events such as VIVs. The validation results demonstrate high predictive accuracy, with RMSE values below 0.05 m/s² and R2 exceeding 95%, alongside robust anomaly detection capabilities. These findings highlight the effectiveness of the proposed framework in enhancing the safety, reliability, and early warning capabilities of long-span bridges.