Research on the prediction of serviceability of mountain bridges based on the Graph-Guided and Deep Attention Network model

Abstract

China’s infrastructure has transitioned from rapid construction to long-term operation, making accurate prediction of bridge service behavior critical for structural safety. The 2024 Zhashui bridge collapse underscores the urgent need for precise predictive models to mitigate catastrophic consequences. Conventional prediction methods suffer from limited accuracy, while deep learning models often lack transparent treatment of dependencies among structural health monitoring variables, especially for mountain-span bridges under geological and environmental uncertainties. We propose Graph-Guided and Deep Attention Network (CG-AttenDNN), a hybrid deep learning framework integrating candidate causal-graph information and attention mechanisms. First, the PCMCI+ algorithm is used to identify statistically supported lagged and contemporaneous dependency relationships from structural health monitoring (SHM) data, from which a graph representation is constructed. Graph Convolutional Networks (GCN) then model spatial dependencies based on the graph links, while an attention-enhanced deep neural network (DNN) captures temporal dynamics. Finally, Shapley Additive Explanations (SHAP) interpretability analysis deciphers feature contributions. Validated on Lianhua Bridge’s GIS-FIM platform, our model achieves R2 scores of 0.8558 (pier settlement) and 0.9180 (deflection), outperforming baseline models by 8.65% and 3.17% respectively. SHAP analysis confirms that lateral vibration acceleration and pier-top inclination dominate predictions. By embedding data-driven graph structure into deep learning, CG-AttenDNN provides accurate and interpretable predictions for mountain-span bridges, enabling proactive maintenance and risk management while recognizing that observational SHM data alone cannot prove physical causation.