Abstract
Civil infrastructure is the backbone of society, providing pathways for goods and services distribution. However, in the United States, the overall condition of infrastructure in 2021 has been rated as a C- by the American Society of Civil Engineers (ASCE), indicating that infrastructure will be unable to support future capacity needs. To address this, proactive and cost-effective maintenance strategies are critical, requiring comprehensive data to assess structural health accurately. This research aims to develop a digital twin framework that informs stakeholders about the structural condition of assets. A digital twin combines virtual models and simulations with real-world measurements to facilitate informed decision-making, ensuring the twin reflects the current state of the physical system.
This dissertation seeks to leverage graph neural networks (GNNs) to enable predictive capabilities within the digital twin framework. The primary goal is to create a robust, asset-level digital twin capable of accurately predicting subsurface damage based on real-world measurements, contributing to more efficient infrastructure management and maintenance. Within the proposed framework, digital image correlation (DIC) has been selected as the primary measurement strategy, offering a non-invasive full-field surface deformation approach that aligns with the spatial simulations derived from the GNN model.
First, a heterogeneous GNN framework was formulated to connect historical 3D finite element measurements with current surface structural behavior data. Training data was generated through finite element simulations incorporating randomly positioned damage of varying depths while maintaining elastic behavior. MeshGraphNet, a mesh-based GNN architecture developed by Google DeepMind, was adapted for binary node-level damage classification. Three key analyses were conducted to optimize performance: sensitivity analysis to determine optimal training data requirements, grouped feature removal to assess feature contributions, and permutation importance to identify classification drivers. The model achieved a 69.1% F1 score for damage detection. Relative to the current iterative optimization approach, the model demonstrated a 15% improvement in performance.
Next, real-world noise incorporation was investigated using an ideal experimental configuration from a DIC setup to obtain three-dimensional structural behavior from small scale tensile testing. The previous dataset was expanded to include varying load magnitudes at damage onset while maintaining elastic conditions, addressing class imbalance in damaged instances. The training set augmentation incorporated noise-free, noisy, and varying noise threshold versions of each instance, enabling the model to distinguish signal from noise. Under noisy conditions, the model achieved a 43.79% F1 score for damage detection, with performance improving significantly for instances with greater damage where signal-to-noise ratios increased.
Finally, the framework's real-world applicability was validated through extensive experimental testing. Seven DIC testing conditions were evaluated across four coupon geometries to assess model robustness. To ensure accurate validation, a novel laser scanning algorithm was developed for precise coordinate alignment between camera and FEA domains. Additionally, code previously designed to calculate the discrepancy between FEA and DIC points, was modified and integrated with pyANSYS to handle mesh interpolation and address unmatched points between experimental and computational domains. The proposed alignment was able to achieve a rate up to 80% across all tests, demonstrating the effectiveness of the approach. The pre-trained GNN achieved performance metrics on experimental data comparable to those obtained on simulated noisy data from 5%-75% F1 score of the damaged class. The performance of the models was affected by the damage percentage relative to the graph, coordinate alignment and the noise present in the experiments. The results confirmed that the noise-robust training successfully translated to real-world conditions with varying signal-to-noise ratios.
This research, presented in the form of three journal manuscripts, establishes graph neural networks as effective tools for bridging computational and experimental domains in structural health monitoring. The developed framework demonstrates that robust damage detection can be achieved even with limited sensor data and measurement uncertainties. By enabling reliable integration of partial experimental observations with finite element models, this work provides a foundation for implementing digital twins in structural systems, advancing the field toward automated, real-time damage assessment capabilities.
