Quantifying two-phase fluid quality through an optical supervised machine learning technique

Abstract

Accurately quantifying two-phase fluid quality is essential for enhancing the efficiency and reliability of thermofluidic energy systems. This research presents a novel, non-intrusive approach to empirically estimate steam quality, using shadowgraphy imaging, combined with an innovative supervised machine learning segmentation technique. This contribution is original in adapting supervised pixel-level classification, traditionally applied in biomedical imaging, to two-phase flows, representing a cross-disciplinary methodology. The technique uniquely enables thresholding of complex steam flow interactions, including a ‘discontinuous’ annular flow. A dedicated experimental test facility enabled repeatable steam quality measurements, verified by a condensate trap, yielding a standard deviation of only 0.02. A Photron SA4 camera, operated with lower specification settings, captured 2D images through a 20 mm internal sight glass at 30 kg/h and 1.2 to 1.4 BarA. Two image segmentation approaches were evaluated for the varying compositions: an individual-trained method, where separate classifiers were developed for each quality (5 images out of 200), and a multi-condition trained method, where a generalised classifier was developed using training samples from multiple quality conditions (2 images each at 3 qualities). Using supervised machine learning tools within FIJI, steam quality estimations for both methodologies showed strong agreement with the condensate trap measurements, achieving a good accuracy within +0.02 to -0.04 over the quality range of x = 0.77 to x = 0.9. These findings demonstrate that this unique method can be integrated into real-time flow measurements, combining optical diagnostics with automated segmentation to enable accurate monitoring of flow regimes under industrial constraints.