LGSTA-GNN: A Local-Global Spatiotemporal Attention Graph Neural Network for Bridge Structural Damage Detection

Abstract

Accurate detection of structural damage is essential for ensuring the safety and reliability of bridges. However, traditional vibration-based approaches often struggle to capture rich feature representations and adequately model spatial dependencies among sensors. This study proposes a novel bridge damage detection framework, LGSTA-GNN, which integrates local–global spatiotemporal learning with graph neural networks. The framework first extracts multi-scale temporal–frequency features using a multi-scale feature extraction module. A local graph feature extraction module then models intrinsic spatial relationships through graph convolutions, while a global graph attention module adaptively captures inter-sensor dependencies by emphasizing structurally informative nodes. A benchmark dataset generated from a scaled bridge model under progressive damage states is used to evaluate the proposed method. Extensive experiments demonstrate that LGSTA-GNN outperforms multiple graph neural network variants and conventional deep learning techniques, achieving superior accuracy, precision, recall, and F1-score. The confusion matrix and t-SNE visualization further verify its enhanced discriminative capability and robustness. Ablation studies confirm the contribution of each module, highlighting the effectiveness of global attention in identifying subtle structural deterioration. Overall, LGSTA-GNN provides an effective and interpretable solution for intelligent bridge damage detection, with strong potential for practical structural health monitoring and real-time safety assessment.