Neural-network learning of two-phase flow regimes from EIT-spectrograms

Abstract This study is concerned with the classification of two-phase flow regimes. The data considered are provided by electrical measurements at the boundary of the probed medium. In this context, our aim is to measure and characterize flow regimes, and their transitions, in real time. To this end, we investigate a neural-network approach based on spectrogram representations of time-dependent measurements in order to improve the accuracy compared to time series analyses. In the present study, dynamical tests in an experimental two-phase flow loop are considered, making use of a sensor developed at the Atomic Energy Commission coupled with a simultaneous voltage excitation method. A total of 80 configurations are studied, with different air/water mass flow rates leading to five different flow regime classes, namely slug, plug, wavy, stratified and annular. For each configuration, the available electrical measurements are processed into a normalized impedance matrix, whose important features are extracted using singular value decomposition. The time evolution of the maximum singular value is then passed to a neural network, which is trained from labeled data while making use of the stratified k -fold cross-validation method. Two different network architectures are considered: the first one processes raw time-series, leading to a regime classification accuracy of 89.4%. The second one uses spectrogram images, resulting in a more accurate classification reaching 97.4%. A comparison with other methods for identifying two-phase flows is also proposed, which supports the main contribution of this study in providing evidence of the advantages of using spectrograms of electrical impedance tomography measurements.