Georgia Institute of Technology
Generative Transformers Enabled with Diffusion for Electrical Imaging of Structural Damage
Abstract
dc:description.abstractElectrical Impedance Tomography (EIT) is a non-invasive imaging modality that re- constructs the internal conductivity distribution of a domain from boundary voltage measurements, proposing a promising framework for structural health and damage detection. However, the EIT inverse problem is severely ill-posed and classical regularization approaches are limited in their ability to produce spatially sharp, noise free reconstructions in complex structural geometries. This thesis presents the Transformer-Enabled Diffusion (TED) model, a novel deep learning architecture developed to address these limitations in the context of two-dimensional EIT-based structural damage detection. The proposed TED architecture integrates four computational components: a one- dimensional Convolutional Neural Network (CNN) encoder that extracts local features from differential boundary voltage measurements, a multi-head self-attention Transformer that captures global patterns through encoded representation, a parallel feedforward bypass network that stabilizes training and supplements the Transformer path, and a diffusion mod- ule that iteratively sharpens the predicted conductivity field. An empirical hyperparameter study across 15 iterations was conducted, with reconstruction quality assessed by visual inspection of predicted conductivity fields applied to a damaged carbon fiber truss. The study identifies a configuration capable of producing damage localizations in both single and multiple damage scenarios.
Degree
thesis:*- Name thesis:degree_name
- Civil Engineering, MS
- Department thesis:degree_department
- Civil and Environmental Engineering
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Jones, Brandon Marcellus
- Advisor dc:contributor.advisor
-
- Smyl, Danny
- Committee members dc:contributor.committeemember
-
- Tien, Iris
- Sherman, Ryan
Rights
- Language dc:language.iso
- English
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/81642