An Innovative Crack Detection Algorithm to Support Automated Inspection of Nuclear Reactor Cores

Throughout the nuclear sector, Condition Monitoring and Inspection (CM&I) is used to evaluate, assess and quantify the health and condition of key components and critical infrastructure. Of particular interest is the inspection of nuclear reactors which, depending on reactor design, can be achieved via Remote Visual Inspections (RVI) whereby a camera is deployed within a nuclear reactor to gather images and videos for visualization and analysis. The current manual analysis process is labour-intensive and extremely time-consuming. It is also subjective, prone to human error and difficult to repeat. More recent practices on the detection and analysis of such defects, focus on automating the process, but assign a label on the whole image or a patch of it (image classification). In this paper, we establish a framework for automatically detecting defects from inspection footage to a pixel level, by implementing and modifying a well-known deep neural network. After creating a dataset of images acquired from software that processes and stitches the raw video footage, we train and asses our model. We evaluate the trained model and report the testing accuracy on unseen data. The results are promising and represent an attempt for an automatic pixel-level crack and defect detection system to a field where this family of methods has scarcely been explored: nuclear inspection. This approach provides a low-level localization of cracks, removes bias that is often present with a classification or subjectivity associated with manual approaches and allows repeatability of results.