Application of convolutional neural networks for the classification of two-phase flow patterns

Characterization of the two-phase flow is a critical task due to the great number of applications in which it is involved, such as power generation or efficient use of energy, processes that are typically present in the nuclear, cryogenic and petrochemical industries, among others. Identification of two-phase flow patterns is a complicated task due to the diversity of factors on which these flow patterns depend. The performance of the process of training, validation, and classification tests of images of two-phase flow patterns using convolutional neural networks is presented. In order to carry out this procedure, series of frames were extracted from experimental videos. Afterwards, an image catalog with samples of representative flow patterns with the types of slug, annular, semi-annular, and dispersed annular flow was defined, in addition a category of superior annular was included, described by presenting the annular flow pattern in the upper section while the appearance of disturbances at the bottom of the image, this last category represents a transition process between the annular flow and another flow pattern. The pattern identification effectiveness obtained with the proposed scheme was greater than 90%, for the five flow patterns used, demonstrating that convolutional neural networks can perform two-phase flow regimes identification.