Enhancing Nuclear Material Storage
Container Surveillance with Automation
and Machine Learning Toolkits

Nuclear material storage containers at U.S. national laboratories must meet performance requirements set by the DOE to ensure the safety of workers, public, and the environment. As part of these requirements, container inspections must be performed routinely and are often in high-dose environments where maintaining as-low-as-reasonably-achievable dose requirements limit the number and frequency of surveillance. Additionally, these inspections are human-intensive tasks that require a greater-thanfamiliarity knowledge of each container’s status by subject matter experts. Observing and documenting all anomalies of concern, such as dents and corrosion, both during and between inspections challenges the current surveillance paradigm. Improvements to the frequency and accuracy of inspections can be achieved with an autonomous system capable of detecting anomalous features of nuclear storage containers in real-time. Our implementations use a supervised machine learning approach with regionbased convolutional neural network architectures (Mask R-CNN and Detectron2). Using both real and simulated container images, model performance is comparable to that achieved with the COCO dataset. Detection of all anomalous features, especially those not trained into supervised models, presents a larger challenge. Our approach leverages the use of two separate techniques: unsupervised machine learning models and sensor fusion for online automated training. We evaluate the improvements feasible from unsupervised anomaly detection to identify previously unseen defects and allow datasets to be entirely constructed from images of undamaged nuclear storage containers. A second improvement leverages supervised learning instance segmentation models on 2D images augmented with additional 3D scanning hardware to generate “ground truth” true-positive regions autonomously in a bench-top implementation. These techniques are combined to build a more general toolset that can better detect nuclear material storage container anomalies and improve container inspections overall.