Shah et al. Res. Trends Int. J. Technol. Innov., July - September 2026, 1 (3) : 66-73
1Department of Computer Science, Birla Institute of Technology and Science, Pilani, India; 2Department of Information Technology, International Institute of Information Technology Bangalore, India
Article History
Accepted : 07 Jul 2026
Published : 10 Jul 2026
Publication Issue
Volume 1, Issue 3
July - September 2026
Page Number66–73
Manual inspection routines for structural health monitoring rely on hand-crafted threshold rules that struggle to generalise to novel or compound defect patterns. This paper represents a bridge sensor network as a spatial graph and trains a graph convolutional network to detect four defect classes: crack propagation, corrosion-induced deterioration, delamination and load-path anomaly. Evaluated on a labelled dataset of 9,300 sensor-network inspection records, the GCN model achieved an F1-score of 0.91 for crack-propagation detection, exceeding a widely used rule-based threshold analyser F1-score of 0.79 on the same test set.
Keywords - structural health monitoring, graph convolutional network, defect detection, bridge engineering, sensor networks
Undetected structural defects such as progressive crack propagation have contributed to costly bridge closures and, in severe cases, structural failures, and rule-based threshold analysers, while fast, frequently miss novel or compound variants of known defect patterns that a learned graph representation may generalise to better.
Bridge sensor-network readings were converted to a spatial graph representation with nodes encoding per-sensor strain, vibration and tilt features, and a three-layer graph convolutional network was trained on 9,300 inspection records labelled by structural engineers for the presence of crack propagation, corrosion-induced deterioration, delamination and load-path anomaly, benchmarked against a rule-based threshold analyser on an identical held-out test split.
The GCN model achieved an F1-score of 0.91 for crack-propagation detection and 0.87 averaged across all four defect classes, exceeding the rule-based analyser F1-scores of 0.79 and 0.74 respectively on the same test structures, with the largest relative improvement observed for defect patterns involving multiple interacting sensor readings that rule-based analysis under-detected.
Graph-based deep learning on structural sensor-network representations detects a meaningfully broader range of defect instances than rule-based threshold analysis alone, suggesting value as a complementary inspection tool rather than a replacement. Future work will explore combining both approaches in an ensemble inspection pipeline.
[1] Kipf T. N. and Welling M., Semi-supervised classification with graph convolutional networks, ICLR, 2017. [2] Farrar C. R. and Worden K., An introduction to structural health monitoring, Philosophical Transactions of the Royal Society A, 2007.
© 2026 The Author(s). Published by IJEIA Editorial Office. This is an open access article under the Creative Commons Attribution 4.0 International License (CC BY 4.0).
Parth Shah, Divyansh Rastogi (2026). Graph-Based Deep Learning for Automated Structural Defect Detection in Bridge Health Monitoring Sensor Networks. IJEIA, 1(3), 66-73.