Development of a deep learning-based image processing technique for local phase fraction of multiphase transient flow

Multiphase flow measurement techniques are essential to capture flow behavior and understand the underlying phenomena. These methods rely on specific instrumentation to monitor flows and record signal responses, either in situ , instantaneously, or as time-averaged data. However, intrusive techniques can disturb the current flow, particularly when measuring local volume fraction of the phases, which may affect and alter the natural flow behavior. To address these challenges, combining complementary techniques is often necessary for accurate characterization of complex flows. Therefore, the development of a suitable technique that enhances both the accuracy and the feasibility of measurements in complex flows is fundamental to achieving reliable flow characterization. This research aims to employ Convolutional Neural Networks (CNNs) in image-based analysis to predict phase fraction distributions. A framework was developed to estimate local volume fractions of the phases by applying image processing to high-speed recordings of air-oil-water flows. Experimental investigations were conducted in a rock-flow cell with a transparent pipe, enabling direct visualization and detailed recording of flow dynamics. Grayscale and RGB image-processing approaches were employed to identify and segment the flow structures, enabling the determination of the local volume fraction of the phases within the pipe. The use of CNNs for two-phase flow pattern identification is well-established in the literature. Therefore, this work emphasizes a measurement technique phase distribution captured by two-dimensional image data are mapped into three-dimensional volumetric results through the implementation of a virtual probe. This is achieved by computing pixel-wise chord lengths based on the circular geometry of the pipe to account for curvature effects. The results indicated that the CNN-based approach achieved high accuracy despite the inherent limitations of imaging techniques. Furthermore, RGB processing proved particularly effective as the number of phases increased, improving performance in three- and four-phase flow and showing consistent and reliable results.