Explainable Machine Learning for Bubble Leakage Detection at Tube Array Surfaces in Pool

Early detection of bubble generation from tube arrays in systems such as fast reactor steam generators, Pressurized Water Reactor (PWR) cores, and Liquefied Natural Gas (LNG) regasification units is critical for safety. While various methods have been proposed, they face challenges such as high spatial resolution requirements, rapid response times, and varying strengths and weaknesses, suggesting the need for a combined approach. This study integrates ultrasonic testing (UT) with Machine Learning (ML) to identify the presence, location, and direction of bubbles within a complex tube array that cause signal attenuation. A Convolutional Neural Network (CNN) successfully achieved 100% identification accuracy. Furthermore, a method was developed that uses an autoencoder as a feature extractor, combined with a One-Class Support Vector Machine (SVM) and k-means. This approach achieved high accuracy and a correct decision basis. It also demonstrated strong generalization, successfully detecting anomalies without requiring labels for anomalous data, enabling robust bubble identification.