Image identification for two-phase flow patterns based on CNN algorithms

Abstract

Flow patterns are essential and useful to model the interfacial structures and heat transfer in gas-liquid two-phase flow. However, the current two-phase flow patterns classification methods mostly depend on direct visual observation. This study adopted a new flow pattern classification method based on convolutional neural network (CNN) algorithms to achieve an automatic and objective identification of two-phase flow patterns. A database of 696 test conditions, including 105642 condensing flow pattern images of methane and tetrafluoromethane in a horizontal circular tube, is collected as the input of the data-driven algorithms. After 80% of image data is fed to train and fit the parameters in the algorithms, the trained models with acceptable universality are obtained to identify five flow patterns: annular flow, bubbly flow, churn flow, slug flow and stratified flow. Compared with the manual classification, the proposed method can accurately predict two-phase flow patterns with a prediction accuracy of more than 90.63% and 91.45% for the test dataset and the entire database, respectively. The average accuracy for predicting all data points in the database is more than 97.56%. The results showed that using images as input, CNN algorithms can provide objective prediction with satisfactory accuracy and universality for two-phase flow pattern identification.