Performance of neural network in image reconstruction and interpretation for electrical capacitance tomography

An Artificial Neural Network (ANN) has successfully performed the task of image reconstruction for Electrical Capacitance Tomography on simulated measurements, and has additionally provided image interpretation to identify the flow. This paper analyses the performance of such tomographic processing and demonstrates its flexibility and tolerance under wide noise and parameter variation. It also illustrates the feasibility of training a single network to reconstruct images in water-continuous as well as oil-continuous flow. The front-end network is a Single-Layer-Multi-Output-Network (SLMON) with 54 neurones in the input layer (corresponding to 12-electrode capacitance measurement, excluding the adjacent measurements), fully connected to a grid of 100 neurones in the output layer (corresponding to the spatial image), with no intra-layer connections. The back-end network is a Multi-Layer Perceptron (MLP) with one hidden layer trained with backpropagation algorithm to identify different flow regimes, namely: stratified, bubble, core and annular flow