A Parallel & Hybrid Electrical Impedance Tomography System for Three Phase Flow Measurement

Abstract

Electrical Impedance Tomography (EIT), recognized for its non-intrusive, cost-effective, and radiation free imaging capabilities, has gained significant attention within the biomedical and industrial fields. Traditionally, 2D EIT has been extensively used in applications such as breast cancer detection and lung imaging. It excels the conventional imaging methods like X-rays, CT scans, and MRIs by providing real-time data monitoring [1-2]. However, the limitations of 2D EIT become apparent in complex, three-dimensional scenarios where a more detailed and accurate imaging representation is required. [3].
This thesis introduces a post-processing framework for three-dimensional EIT that integrates the Gauss-Newton (GN) algorithm with a Generative Adversarial Network (GAN) to enhance conductivity image quality. This work fills a critical research gap by being one of the first to extend deep learning postprocessing to the 3D EIT. Unlike most prior studies that restrict post-processing to 2D, this extends the concept into the 3D domain, enabling high-resolution volumetric enhancement. In addition, this thesis addresses key limitations found in direct deep learning approaches for EIT reconstruction. Prior studies focus on end-to-end models that map boundary voltage measurements directly to conductivity distributions. However, these models typically require large, high-resolution datasets and struggle to generalize across different electrode setups, excitation patterns, and anatomical complexities. By contrast, the hybrid approach introduced overcomes these shortcomings, allowing the system to remain adaptive to different hardware setups while still benefiting from learning-based enhancement. This makes the proposed method more flexible, interpretable, and deployable in practical scenarios, particularly where direct mapping struggles due to limited data or domain shift.
Key contributions include the development of a 3D GAN model designed for post-processing GN-based conductivity volumes, achieving improved structural clarity, boundary definition, and noise resilience. The system is implemented within a cylindrical domain using a dual-ring electrode configuration with current excitation applied between the rings—an approach that captures vertical current paths and interplane interactions representative of real-world three-phase systems. The architecture operates efficiently on GPU hardware, taking advantage of parallelism. The model demonstrates consistent robustness under varying noise conditions, significantly outperforming traditional GN reconstructions. By preserving the physical consistency of traditional solvers and leveraging deep learning for refinement, it provides a scalable, explainable, and hardware-efficient framework for advancing 3D imaging in multi-phase flow applications.
This thesis presents a parallel and hybrid 3D EIT reconstruction system designed for simulated three-phase flow monitoring. “Parallel” refers to the use of GPU-accelerated computation on embedded platforms, while “hybrid” reflects the integration of a traditional Gauss-Newton solver with a Generative Adversarial Network for post-processing enhancement.

Date of Award	Jun 2025
Original language	American English
Supervisor	Mahmoud Meribout (Supervisor)