Deep-Learning Framework for Dynamic Deflection Prediction of Suspension Bridges under Temperature, Traffic, and Wind Loads

A data-driven predictive model for dynamic deflection response can aid the safe operation and maintenance of a suspension bridge. In particular, it lays the groundwork for bridge condition assessment and damage detection, dynamic reliability evaluation, abnormal structural health monitoring (SHM) data reconstruction, and digital twin system construction. This study proposed a deep learning framework for predicting the deflection response of a suspension bridge based on the SHM system and the weigh-in-motion (WIM) system. In Task 1, an attention-long short-term memory model was developed to estimate the quasi-static deflection component using air temperature, wind speed, and gross vehicle weight as input features. In Task 2, an analytical model was built to transform the WIM data into a time series of traffic flow state coefficient at equal time intervals. These, along with the fluctuating wind speeds, were input into an attention-enhanced time-frequency generative adversarial network (AE-TFGAN) to predict the dynamic deflection component. The AE-TFGAN incorporated an attention-gated U-Net generator and a time-frequency dual discriminator, supported by a novel loss function that enhances prediction accuracy in both time and frequency domains. A case study on the Yangsigang Yangtze River Bridge in China validated the proposed framework, demonstrating that the integration of attention mechanisms and time-frequency features significantly improves prediction performance, outperforming conventional baseline models.