Reviewing large volumes of surveillance imagery data for safeguards-relevant events requires IAEA nuclear safeguards inspectors to expend significant time and effort while maintaining intense focus. Thus, a solution was sought to increase the efficiency and reduce the time burden upon inspectors performing surveillance data reviews. To that end, the authors have worked closely with a group of inspectors to identify surveillance review use cases and collect their associated requirements. This paper presents the work done in the development of a novel deep learning (DL) system to integrate the results of trained DL models into the safeguards surveillance data review workflow. DL convolutional neural network models were trained to identify movements of spent fuel casks in CANDU nuclear power plants from the spent fuel ponds to the dry storage areas, as these data reviews pose a particularly heavy burden for the inspectors. To date, the DL results observed in tests to find safeguards-relevant activities have demonstrated detection accuracy that is far superior to the customarily used motion detection algorithm. Using remotely transmitted camera data, DL activity detection is performed automatically on dedicated servers, whose results are integrated into the Next Generation Surveillance Review (NGSR) software. Events identified via DL are flagged in NGSR, and can save inspectors time by creating a report template prefilled with a list of predefined safeguards-relevant activities. NGSR functions as a decision support tool, as these DL-based events can be subsequently analyzed and interpreted by inspectors. By quickly and accurately identifying safeguards-relevant objects and activities in large quantities of surveillance imagery data, the DL system shows promise as a pathway to a more efficient workflow for surveillance review.
Year
2024
Abstract