Robust and efficient dual-graph neural networks for structural damage detection and localization

Effective and timely detection of structural damage is crucial for maintaining the integrity, safety, and longevity of civil infrastructure systems. Traditional structural health monitoring techniques, especially vibration-based methods, often encounter limitations such as sensitivity to measurement noise and reliance on manual feature extraction. Moreover, advanced deep learning methods and single-graph models often struggle to accurately localize and assess damages of varying scales and positions within complex, high-dimensional structures. To address these limitations, this study introduces a novel dual-graph convolutional network approach integrating both spatial (sensor topology) and feature-based adjacency matrices. A one-dimensional convolutional neural network preprocessing module is also incorporated to enhance computational efficiency through effective feature extraction and data compression. Comprehensive evaluations were conducted on two established civil engineering benchmark datasets, the Leibniz University Test Structure for Monitoring (LUMO) and Qatar University Grandstand Simulator (QUGS). The proposed method achieved 98.8 % accuracy on LUMO and 99.9 % on QUGS. Robustness tests showed accuracies above 96.5 % (LUMO) and 97 % (QUGS), even at low signal-to-noise ratio (0 dB). CNN-based compression significantly reduced training time. For instance, training time dropped from 295.7 to 12.1 h (LUMO) and from 1712.85 to 148.24 h (QUGS), maintaining accuracy above 98 %. The proposed dual-graph approach thus offers substantial improvements in structural damage detection and localization for complex civil structures. • Dual-graph approach for structural damage detection in civil infrastructure is proposed. • Novel GCN fuses sensor topology and feature correlations to improve accuracy. • Validated on LUMO and QUGS datasets with high accuracy, noise robustness, and fast convergence. • Outperforms state-of-the-art methods on QUGS and LUMO datasets. • 1D-CNN module reduces training time while preserving high detection accuracy.