Graph Neural Networks for Structural Damage Detection on the Hell Bridge Test Arena Proceedings Paper

Avila, CC, Marasco, G, Azizinamini, A. (2026). Graph Neural Networks for Structural Damage Detection on the Hell Bridge Test Arena . Procedia Structural Integrity, 85 985-992. 10.1016/j.prostr.2026.08.080

cited authors

  • Avila, CC; Marasco, G; Azizinamini, A

abstract

  • Structural Health Monitoring (SHM) is critical for the lifecycle management of aging steel bridges. The Hell Bridge Test Arena (HBTA) dataset provides a unique full-scale benchmark with controlled damage scenarios, making it an ideal platform for validating data-driven algorithms. Previous studies using the HBTA dataset have primarily relied on manual feature extraction and classical machine learning techniques. These approaches typically treat sensor data as independent vectors, neglecting the physical connectivity of the truss structure and the spatial dependencies inherent to vibration propagation. This study proposes a Geometric Deep Learning framework that explicitly incorporates these spatial relationships. A Graph Neural Network (GNN) is developed for supervised damage identification by representing the sensor network as a graph aligned with the bridge topology. In addition, a Graph Neural Network Autoencoder is trained exclusively on undamaged data to evaluate its capability for damage localization. The performance of the proposed models is benchmarked against classical approaches, including Support Vector Machine (SVM) and Multi-Layer Perceptron (MLP). Results demonstrate that the GNN significantly improves damage identification accuracy by leveraging structural topology, highlighting the benefits of physics-informed geometric deep learning for real-world SHM applications.

publication date

  • January 1, 2026

published in

Digital Object Identifier (DOI)

start page

  • 985

end page

  • 992

volume

  • 85