Answers are grounded in this paper and cite their sources.
Bridge deflection is an important structural response that provides valuable information on structural stiffness, service condition, and operational safety. The random temperature variations and complex mechanical systems inherent in long-span bridges pose significant challenges to achieving accurate and efficient deflection prediction in long-term structural health monitoring. A GRU-FA-ELM model is proposed in this study to address the trade-off between accuracy and efficiency in bridge deflection prediction. In the proposed model, a gated recurrent unit (GRU) extracts temporal dependency features from multivariate monitoring data and generates the final hidden state vector as a compact representation of the input time window. A feature attention (FA) module is then applied to this vector to refine the feature contribution distribution and highlight informative hidden features related to deflection prediction. An extreme learning machine (ELM) serves as a fast regression module that maps the enhanced features to the predicted bridge deflection. Optuna is also adopted to jointly optimize the key hyperparameters. The two-stage training strategy integrating supervised regression training and ELM regressor training was developed to reduce the training complexity and improve the computational efficiency. Field monitoring data from a cable-stayed bridge were used to validate the proposed GRU-FA-ELM model. The results indicate that the proposed model achieves higher prediction accuracy when compared with GRU, BiGRU, long short-term memory (LSTM), bidirectional LSTM, and GRU-ELM models on both small-sample and large-sample datasets. Furthermore, its computation time is reduced by approximately 22–51% compared with the recurrent baseline models, while being only marginally longer than that of the GRU-ELM model. The proposed model strikes an effective balance between accuracy and efficiency for online bridge deflection prediction.
No takes yet. Share an insight, caveat, or question.
Xi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: