Deep learning-based prediction of wind-induced lateral displacement response of suspension bridge decks for structural health monitoring

Wang, Zhi-wei, Lu, Xiao-fan, Zhang, Wen-ming, Fragkoulis, Vasileios C ORCID: 0000-0001-9925-9167, Zhang, Yu-feng and Beer, Michael ORCID: 0000-0002-0611-0345 (2024) Deep learning-based prediction of wind-induced lateral displacement response of suspension bridge decks for structural health monitoring. Journal of Wind Engineering and Industrial Aerodynamics, 247. p. 105679.

Text JWEIA paper_author version copy 2.pdf - Author Accepted Manuscript Available under License Creative Commons Attribution. Download (5MB) | Preview

Abstract

Monitoring the wind-induced lateral displacement (WLD) of the bridge deck is crucial for structural health monitoring (SHM) of suspension bridges. An accurate WLD prediction model can aid the bridge SHM systems in abnormal data detection and reconstruction, structural response estimation under specific wind events, and structural condition assessment. However, WLD prediction faces challenges due to stochastic wind action and complex aerodynamic effects acting on the bridge deck. To address this, a deep learning-based framework was proposed for predicting the WLD response of the suspension bridge deck. This framework decomposed the WLD response into two components, namely the quasi-static and the dynamic one. Two separate deep-learning tasks were employed to predict these components using the lateral wind speed as input. In Task 1, a recurrent neural network (RNN) based on the gated recurrent unit (GRU) was built, whereas a fully convolutional neural network (CNN) based on U-Net was built in Task 2. Novel loss functions tailored to each task were established to facilitate accurate predictions. Measured data from the SHM system of the Jiangyin Yangtze River Bridge, China, was used as a case study to verify the proposed predictive framework's feasibility and high accuracy. The extreme value-weighted loss function in Task 1 enhanced the prediction accuracy for the extreme quasi-static WLD, while the time-frequency cross-domain loss functions in Task 2 effectively integrated the prediction accuracies in both time and frequency domains for the dynamic component of WLD. However, trade-offs were identified between the prediction errors of extreme and non-extreme values, as well as between the time- and frequency-domain prediction accuracies.

Item Type:	Article
Uncontrolled Keywords:	7 Affordable and Clean Energy
Divisions:	Faculty of Science and Engineering > School of Engineering
Depositing User:	Symplectic Admin
Date Deposited:	04 Mar 2024 09:03
Last Modified:	15 Mar 2024 05:26
DOI:	10.1016/j.jweia.2024.105679
Related URLs:	Publisher
URI:	https://livrepository.liverpool.ac.uk/id/eprint/3179064