CNN-OBA

Pool scrubbing has proven to be an effective method for preventing the release of radioactive aerosols into the environment during severe accidents. In this process, particles are washed out of a water reservoir through bubble rise and interaction processes. The hydrodynamics of bubbles is therefore of great importance for understanding particle retention. Characterizing the size and shape distribution of bubbles based on optical measurements (e.g., using high-speed cameras) is crucial for describing two-phase flow, as this allows for the collection of essential geometric information for the study of mass and heat transfer. However, in regions with dense bubble swarms, bubbles often overlap, which complicates image analysis.

Furthermore, most analysis approaches simplistically assume a uniform bubble size in swarm regions, whereas in reality, bubble sizes here follow a log-normal distribution. This research project therefore focuses on the development of an innovative method for analyzing overlapping bubbles using efficient Convolutional Neural Networks (CNN). Neural networks are capable of identifying overlapping, blurred, and non-spherical bubbles within an image segment and can thus contribute in the next step to characterizing the flow much more accurately, thereby significantly improving the modeling of bubble behavior.

The goal of this work is to use a machine learning model (CNN) to improve the reliability of image processing analysis for highly overlapping bubbles with high gas content (>15%). The results of this work are intended to increase the accuracy of bubble parameter predictions. The entire workflow is divided into three phases: bubble detection and segmentation, bubble reconstruction, and application and optimization. Corresponding models are developed and trained to improve the accuracy of bubble image detection, reduce the number of outliers, shorten data processing time, and significantly reduce the number of misidentified bubbles compared to standard detection methods.