Application of Artificial Intelligence for Automated Defect Detection on Nuclear Power Plant Fuel Asssembly Components

Abstract

For manufacturing nuclear fuel, Framatome's various plants and workshops carry out rolling operations of chemical and thermal treatment on zirconium tubes. After the finishing stages of these tubes, a series of checks are carried out, including a visual inspection of the final external surface, to ensure the quality and integrity of these cladding tubes, essential for nuclear safety. Today, this visual inspection is automated but requires a significant amount of time for an expert to review the rejected components by the current analysis system. To minimize the number of components rejected by the existing system while respecting the same level of defect detections, a study of the application of Artificial Intelligence was conducted. Two convolutional networks of a similar architecture have been developed: one for making a decision to sanction the tube and one to classify the type of rejection. In this article we present the architecture of the chosen neural networks, as well as the performance obtained for a large set of evaluation data from the rejections of produced tubes. Finally, we put these results into perspective by exposing the potential productivity gain using this system in production assisted by Artificial Intelligence.