NEURAL-NETWORK FOR PATTERN ASSOCIATION IN ELECTRICAL CAPACITANCE TOMOGRAPHY

The paper describes the basic principles of an artificial neuron, the multilayer perceptron network and the back-propagation training algori thm, applied to electrical capacitance tomography systems for real-tim e, noninvasive imaging and measurement of multicomponent flows such as gas/oil and water/oil. Particular attention is given to the problems of distortion of the field ('soft field' error) and limitations on spa tial resolution (imposed by the number of electrodes) in conventional image reconstruction algorithms for current systems. In addressing the se issues, for the first time, an artificial neural network is employe d to replace conventional image reconstruction algorithms. The system consists of a simulation program for a single layer multiple output ne twork, using a variant of the back-propagation training algorithm with the principle of pattern association. The input vector consists of pr eprocessed capacitance measurements, and the output of the network dir ectly corresponds to the spatial image. Two similar networks are train ed for gas/oil flow (small difference in permittivity) and water/oil f low (large difference in permittivity) with results compared. The resu lts illustrate the feasibility of such networks for relatively accurat e image reconstruction, without the difficulties associated with conve ntional algorithms.