Convolutional neural network for flow boiling patterns classification

Abstract

Identifying flow patterns is crucial for understanding two-phase flow behaviors, which are relevant in areas such as liquid-gas mixtures, refrigeration, and convective boiling. Visual image processing allows for the automation of interpreting these two-phase flow patterns. This article aims to enhance the accuracy of classifying two-phase flow patterns during the convective boiling of isobutane in a 1 mm diameter horizontal tube. To achieve this, two-phase liquid-gas flow patterns were classified using a convolutional neural network (CNN) based on ResNet50 architecture. CNN results were compared with the kNN approach using the same dataset. A discussion is also presented. Using images from a high-speed camera, five unique flow patterns were detected: isolated bubble, plug, slug, churn, and wavy-annular flow. The CNN method showed encouraging results with an accuracy of 93%.