Self-­Organising Maps for Processing Tomographic Data

Amal El­Gehani and Trevor York

School of Electrical & Electronic Engineering, University of Manchester

ABSTRACT

The aim of this work is to explore the use of a specific type of artificial neural network, the Self­ Organising Map (SOM), for characterising tomographic data. This is motivated by a desire to extract useful information from tomographic data without recourse to the often deleterious step of image reconstruction. A significant advantage of the SOM is that it is unsupervised and therefore does not require known solutions during training and this eliminates the need for knowledge of material distribution in a vessel for the purposes of calibration. The Kohonen and Counter­Propagation Artificial Neural Network toolboxes for MATLAB have been applied to a measurement data set that has been acquired using an electrical capacitance tomography system. The data set comprises 780 distributions of plastic pellets in air in stratified and bubble configurations. Experiments have been conducted to explore the classification of flow regime, volume of material and orientation of material. Results, presented as 2D maps, suggest an accuracy greater than 85% might be realised.

Keywords Electrical tomography; artificial neural network; self­organising map