An investigation into the dynamic implementation of a 16-electrode FDM Electrical Impedance Tomography System

Goldswain, Gareth

An investigation into the dynamic implementation of a 16-electrode FDM Electrical Impedance Tomography System

Thesis / Dissertation

2005

Abstract

This thesis documents one phase of an ongoing project - a project which was initiated in 1999, and whose research goal is to investigate, and provide insight into, the possibility of using impedance tomography to measure the mass flow rate of an air-gravel-seawater mixture. The ultimate goal of the research is to develop an instrument which can be placed in-line with an industrial scale airlift, and used to monitor its flow characteristics. Since the project's conception, a number of researchers have been involved with various stages of its development; the research of G. Teague produced the most noteworthy results and since their publication (in his Ph.D. thesis), the drive of the project has been to pursue his recommendations. Teague implemented a dual-plane, 16-electrode, Frequency Division Multiplexed (FDM) Electrical Impedance Tomography measurement system using square-wave excitation, as well as Neural Network (NN) based image reconstruction algorithms. Research succeeding Teague's investigated the use of sine-wave excitation on an single plane, 8-electrode system, and using the reconstruction software developed previously, demonstrated superior outcomes. This thesis is specifically involved with the upgrading of the abovementioned 8-electrode system, to a 16-electrode system capable of measuring real flows. Only static environments had been investigated prior to this research thus, the project's dynamic implementation issues needed to be addressed - a crucial step in the system's industrialisation. Besides the design and fabrication of a new, open-ended, 16-electrode electrode ring, hard-ware issues which needed to be addressed included: the development of a new data acquisition system and the upgrading of measurement hardware. There exists (now) measurement hard-ware capable of capturing EIT data from a 16-electrode rig using sine-wave excitation. The electrode ring contains three sets of electrodes - one plane of 'line' electrodes and two planes of guard electrodes (positioned on either side of the line electrodes). This variety of electrodes was realised to enable an investigation into the effects of electrode shape and configuration on NN based reconstruction algorithms. The reason for investigating the use of line electrodes is that previous research had shown significantly improved results in 2-dimensional Finite Element Method (FEM) simulations of periphery voltages in EIT. Regarding the system's reconstruction software and signal processing: a new reconstruction method has been proposed and is investigated. Kernel Ridge Regression (KRR) is an 'intelligent' generalisation technique which has been demonstrated to outperform classical NN type approaches in similar problems. The experiments of this research benchmark its performance against the best performing Multi-Layer Perceptron (MLP) NN found in Teague's research. Static training data was attained in much the same way as in forerunning phases of the research. Three data sets were captured; measurements obtained using line electrodes, point electrodes (realised using one plane of guards) and line + guard electrodes were recorded, normalised and split into training and test data. On all datasets KRR outperformed MLP NNs in predicting volume fractions of a static 2-phase mixture. Data captured using line electrodes displayed improved results to data captured using point electrodes, when using MLPs to predict volume fractions. The results of KRR were unchanged. The use of guard electrodes in conjunction with line electrodes showed a further improvement in MLP and KRR generalisation ability. Dynamic experiments were performed using two techniques: firstly, a laboratory scale airlift was used to emulate flow regimes similar to that expected in the final application. Results of testing MLPs and KRR, trained using the static data of previous experiments, on real data acquired from the airlift were poor. Thus, simulated boundary measurements with added Gaussian noise were used as training data, the results of which demonstrated a strong correlation between mean predicted volume fractions and average recorded flow rate. The second dynamic experiment involved obtaining real flow data, captured from a high velocity pumping loop, together with reconstructed conductivity 'images' of each data frame. The data was supplied, and reconstructed by, Dr. A. Wilkinson's world-class research group at the University of Cape Town. The volume fractions of each frame were used as target values in database population and were estimated by finding what percentage of pixels in the image fell below a certain conductivity threshold. These target values, together with the raw voltage measurements used to reconstruct each frame's image, were used as training and testing instances. Testing KRR and MLPs on the abovementioned data produced good results. Specifically, KRR's mean absolute prediction error was less than 1 % of the vessel's cross-sectional area. The findings of the research lead to the following conclusions and recommendations: 1. KRR consistently outperforms MLPs in EIT reconstruction. 2. Data collected using line electrodes, especially used concurrently with guard electrodes, is superior when using KRR or MLPs to predict volume fractions of 2-phase mixtures. 3. Boundary measurements simulated using the Finite Element Method (with added Gaussian noise), produce a more representative training database than the empirical training methods used previously. 4. KRR, using simulated training data, performs well in predicting volume fractions of a rising-air-bubbles flow. 5. Kernel Ridge Regression, when trained using conventional images of a flow, can be used to predict volume fractions of similar flows directly. 6. Simulated measurements should become the method of choice for training database population for the prediction of real flows. 7. The real-time implementation issues of kernel methods should be investigated. 8. The system should be tested on high velocity flows.

Keywords

Electrical Engineering

Reference:

Collections

Masters

Full item page