Dynamic Prediction and Early Warning of Bridge Structural Deformation Performance Based on LSTM-SSA-BDLM

Abstract:

Understanding the anticipated behavior of bridge structures is crucial for early identification of potential structural issues or failure modes. This paper proposes a novel predictive framework for bridge structural performance based on Long Short-Term Memory Networks （LSTM）. The block maxima （BM） method is employed to extract hourly deflection extremes from monitoring data， which serve as key indicators for assessing bridge safety. The method involves sliding window LSTM predictions for periodic deflection extreme sequences， integrating Singular Spectrum Analysis （SSA） and error-updating Bayesian Dynamic Linear Models （BDLM） to effectively extract long-term trends and cyclic deflection changes caused by environmental factors. This process effectively reduces the impact of noise while preserving the critical information of vehicle load effects. Applications in three real engineering cases demonstrate that proposed method significantly improves prediction accuracy compared to sliding window LSTM and BDLM approaches. Furthermore， the paper proposes a dynamic warning threshold setting method based on extreme value theory， which effectively avoids the limitations of static warning indicators and utilizes the confidence intervals of predictions for proactive warning.