Intelligent metallic loose part monitoring in three-dimensional structures using convolutional neural networks and the position-invariant loss function

Loose parts in the primary systems of nuclear power plants pose a significant safety risk as they can collide with equipment inner surfaces, particularly in high-velocity areas like the steam generator. Such collisions, especially near welded joints in the steam generator tubes, can cause cracks and lead to coolant leakage. To mitigate these risks, nuclear plants use loose parts monitoring systems that trigger alarms based on rule-based algorithms. However, current methods rely on the analyst's proficiency, and timely analysis is challenging. To address these issues, this paper proposes a convolutional neural network (CNN) model to estimate the mass of any detected loose parts and their impact location. The CNN model uses a position-invariant loss function derived from Cartesian coordinates and the Wigner-Ville distribution. The proposed method is validated using impact signals acquired from a quarter-scale testbed specifically designed to simulate a nuclear power plant's primary system steam generator. This study experimentally confirms that the proposed method can estimate the impact location and mass of internal loose parts in three-dimensional structures more accurately and rapidly than existing methods.