Parameter estimation from tomographic data using self-organising maps

Abstract

The paper reports on the potential of using a type of artificial neural network, the self-organising map, for processing tomographic data from pipe separators to estimate interface levels. This is motivated by a desire to estimate process parameters without recourse to image reconstruction. Results show direct quantitative estimation of volume fraction of two-component flow mixtures containing oil and water from electrical capacitance tomography measurements. Parameter extraction from the trained feature map is realised using Gaussian mixture modelling. Parametric information of a mixture is determined by using the probability estimation of sample map and comparing the result with the model's topology. The SOM Toolbox in MATLAB was used for training and developing the models. After preparing the training data the SOM mixture model can be trained in less than 20 seconds. 75% of the two-component mixture test samples are classified within 5% of the sample's true composition.