Speaker
Description
Wind-induced vibrations pose significant risks to the safety and serviceability of long-span bridges. However, the rarity of extreme wind events and sensor failures during such events often result in sparse monitoring datasets, impeding the reliability of predictive models. This study introduces SMOGN-LSTM, a novel predictive framework that integrates statistical data augmentation and deep regression learning to address small-sample challenges in structural health monitoring (SHM) of bridges. The framework combines the Synthetic Minority Over-sampling technique with Gaussian Noise (SMOGN) to augment sparse wind and vibration datasets, and a Long Short-Term Memory (LSTM) network to learn complex nonlinear relationships and forecast bridge responses. Applied to the Hardanger Bridge in Norway, the framework demonstrates strong performance in predicting both vertical and torsional accelerations under critical windstorms. Comparative ablation studies against alternative data augmentation and regression methods confirm the superiority of the proposed framework in enhancing prediction accuracy and data sparsity mitigation. The results underscore the framework’s potential to improve safety assessments and operational decisions in SHM, particularly for scenarios where real-time data is limited or incomplete. SMOGN-LSTM thus offers a robust and practical tool for ad-vancing predictive analytics in bridge structures under challenging data conditions.
