Managing the Noise: Harnessing the Power of Graph Neural Networks to Detect Subsurface Damage under Noisy Experimental Settings

Abstract Aging infrastructure with limited budgets has driven stakeholders to explore new methods for monitoring structural health. Structural health monitoring represents a strategy to address this challenge using physical measurements to characterize performance. This approach has often leveraged various sensing modalities to describe operational response to loading environment, with most strategies relying on continuous measurement. In recent years, the complexity and economic constraints associated with continuous monitoring have motivated alternatives. The digital twin framework offers a model-first approach, where real-world data informs virtual models (the “twin”), enables dynamic updates based on measurements and further offers enhanced assessment and predictive capabilities. This work formulates a graph neural network-powered prototype digital twin of a structural system that can identify unseen damage through updating intact finite element models of small-scale representations of structural elements based on full-field measurements. Within this study, MeshGraphNet is trained on noisy strain and displacement data using noise statistics characterized from a physical digital image correlation experimental setup, which differs from most existing methods that rely on idealized or synthetically generated noise. Three models were trained on noise-free data, data with noise, and data with improved signal-to-noise ratios. Results demonstrate that model performance correlates positively with damage severity, achieving an F1 score of 43.79% for the damaged class across the data set. Performance improved substantially when evaluated on specific damage severity ranges, reaching 78.75% for instances with 45%–50% damage, as shear strain exceeds the noise floor at higher damage severities. When compared to robust principal component analysis and denoising autoencoders, the proposed mixed data training strategy outperforms these preprocessing methods. Validation on a physical dogbone specimen demonstrates consistent performance across simulated and experimental data, indicating potential for real-world implementation. Future work aims to extend the framework to more extensive experimental data and assess the scalability of the trained model for larger and more complex structural systems.