EIT-based two-phase flow classification using a neural-network on time-series data

Anes SAOUDI^1*, Cédric BELLIS², Mathieu DARNAJOU³, Guillaume RICCIARDI¹

¹CEA, DES, IRESNE, Nuclear technology department Center of Cadarache, 13108, Saint-Paul-lez-Durance, France
²Aix Marseille Univ, CNRS, Centrale Med, LMA UMR 7031 Marseille, France
³FLUIIDD, La Ciotat, France

^*Email: anes.saoudi@cea.fr

ABSTRACT

In Pressurized Water Reactors PWRs, the two phases exhibit different electrical conductivities, which can be leveraged using Electrical Impedance Tomography to reconstruct the conductivity distribution of the flowing medium and characterize flow regimes. This article presents an analysis of data from EIT modules developed at the Atomic Energy Commission, utilizing simultaneous voltage excitations with Frequency Division Multiplexing, in a method known as ONe Excitation for Simultaneous High-speed Operation Tomography. Given the need for high-rate measurements to identify two-phase flow regimes and their transitions, dynamic experimental tests were conducted in the horizontal two-phase flow loop at the University of South-Eastern Norway. A total of 80 tests, varying in Air/Water mass flow rates, led to the identification of five flow regimes: Slug, Plug, Wavy, Stratified, and Annular. Each regime's measurements were mapped into a Normalized Impedance Matrix, with key features extracted using Singular Value Decomposition. The maximum singular values were then used as raw time-series inputs for a Neural Network. To ensure robust training and classification, the Stratified k-fold cross-validation method was employed, enabling accurate identification of different flow regimes.

Keywords: Neural-Network, Two-phase flow, Industrial process, Tomography.

Industrial Application: Nuclear and Petroleum industries.