Direct Flow Process Estimations From Tomographic Data Using Artificial Neural Systems

Junita Mohamad-Saleh, Brian S. Hoyle, Frank J. W. Podd, D. Mark Spink

Institute of Integrated Information Systems, School of Electronic and Electrical Engineering, University of Leeds, Leeds LS2 9JT, UK. E-mail: B.S.Hoyle@leeds.ac.uk.

ABSTRACT

The paper deals with the goal of component fraction estimation in multi-component flows, and component height and interface orientation estimations in two-component flows of various flow regimes. These are critical measurements in many process systems. Electrical Capacitance Tomography (ECT) has been an attractive sensing technique for this task due to its low-cost, non- intrusion and fast response. However, typical systems, which include practicable real-time reconstruction algorithms have shown to give inaccurate results. Such systems also depend upon an intermediate image that must be interpreted to yield useful process information. The existing approaches to direct component fraction measurements have a performance that is typically flow- regime dependent, and they fail to discriminate component fractions in three-component flows. Although accurate estimations of component height and interface orientation can be obtained using iterative methods, the approach is normally computation-intensive that fast, real-time processing cannot be achieved. In this investigation, an artificial neural network approach has been used to directly estimate the component fractions in two-component and three-component flows, and component heights and orientations in two-component flows from ECT measurements. A two- dimensional finite-element electric field model of a 12-electrode ECT sensor is used to simulate ECT measurements for corresponding component fractions, component heights and interface orientations. The Principal Components Analysis method is used to reduce the raw measurement data to a mutually independent set. Multi-Layer Perceptron (MLP) neural systems are trained with sets of such reduced ECT data with the corresponding component fractions, component heights and interface orientations. The trained MLPs are tested with test patterns consisting of unlearned simulated ECT data as well as real static ECT data of gas-water flows. The best-performed MLPs give means of absolute errors of less than 1% for the estimation of component fractions and heights, and less than 10° for the estimations of interface orientations based on the simulated measurements. The mean absolute error based on the real ECT data is less than 3% for the water fraction and height estimations and, less than 20° for the gas-water interface orientations.

Keywords Electrical capacitance tomography, direct model estimation, neural network interpretation, principal components analysis.