Image Based Bubbly Flow Feature Identification Using Deep Learning

Abstract Gas-liquid two-phase flow has high potential in heat transfer and mixing capabilities, and therefore it is utilized in various technologies such as nuclear reactor and chemical plants. There are several flow regimes since the gas-liquid interface transforms constantly. For the sake of safety and optimization in operating plants, it is crucial to understand the behavior of the gas-liquid interface. We have focused on extracting the bubble features in the bubbly flow by filming the bubbly flow with a high-speed camera and training convolutional neural network (CNN) for feature extraction. The assumption made was bubbles in the bubbly flow being ellipsoids. Since void fraction and interfacial area concentration are one of the geometric parameters in the two-phase flow models, like two-fluid model, it becomes possible to evaluate the flow field of the two-phase flow quickly and quantitively by calculating these parameters from the extracted features. We have compared two-phase flow parameters with the conventional object detection method using bounding boxes, and the new ellipse fitting method to identify the best region proposal shape. As a result, the conventional method showed higher accuracy in extracting bubble features under our flow conditions.