Robust bubble feature extraction in gas-liquid two-phase flow using object detection technique

Abstract

ABSTRACT Identification of flow regime and feature extraction of bubble characteristics in gas-liquid two-phase flow systems have been significant issues in various engineering disciplines, including nuclear thermal-hydraulics field. In nuclear safety analysis codes, flow regime transition criteria and bubble size-dependent correlations are often utilized to obtain solutions for the governing equations. However, even today, they still rely on flow regime maps and empirical criteria that are based on visual observations and fully-developed flow conditions. In the present study, the AI-based object detection approach was adapted by utilizing convolutional neural networks to extract bubbly flow features. The model network is capable of extracting real-time bubbly flow features in an internal multiphase flow system. From the coordinate information of each individual detected bubbles, features such as void fraction, bubble number density, and average bubble sizes can be acquired instantly. The present approach has the potential to replace the traditional hand-coded image analysis method and open up a path for a fully automated and versatile two-phase flow feature extraction.