Bubble visualization in rod bundle flow channel using deep learning-based bubble detection

Uesawa, Shinichiro  ; Ono, Ayako ; Yoshida, Hiroyuki  

Bubble visualization using a high-speed video-camera has been used as measurement techniques of bubble diameters, interfacial area concentrations, and void fractions in dispersed bubbly flow. However, the bubble detection was difficult under the condition of the high void fraction because the overlapping bubbles for the sight direction of the camera increased with the increase in the void fraction. In this study, we developed the deep learning-based bubble detector with Shifted window Transformer (Swin Transformer) to overcome the issue. To verify the performance, we used the synthetic bubble images obtained by Generative Adversarial Networks (GAN) and obtained average precisions (APs) for the number of the training dataset. The result showed that the AP was large enough for 50 datasets, and bubble detection was possible even with a small number of the training data. Additionally, we confirmed that the detector can detect and segment individual bubbles in overlapping bubbles obtained in the visualization experiments of pipe and bundle flows. By using the detection results, we estimated the interfacial area concentrations and void fractions. In comparison with commonly used relations, the results were in good agreement with the relations. Thus, the detector can measure not only bubble diameters but also interfacial area concentrations and void fractions.