Neural network reconstruction for tomography of a gravel-air-seawater mixture

This paper presents the first practical implementation of electrical impedance tomography for the imaging of a gravel-air-seawater mixture, wherein the image reconstruction is performed using a neural network. Although research into the use of neural networks for image reconstruction has been done, previous work has predominantly made use of simulated capacitance or impedance readings to train the networks. In contrast, this paper adopts a new practical approach whereby the networks are trained and tested using real data from an existing impedance tomography system. The practical aspects of generating this database are discussed and the network training details are given. Results indicate that the trained networks are able to discriminate among the seawater, gravel and air phases and reconstruct images of bubble configurations not included in the training database. In terms of the intended application, accurate predictions of volume fraction are required. Consequently, this paper examines an alternative solution to this problem, whereby the networks are trained to predict the volume fractions directly rather than first performing an image reconstruction as is standard. Results indicate that these volume fraction predictors outperform the image reconstruction networks in terms of the accuracies of the volume fraction predictions for each of the three phases.