Developing Signatures-based Safeguards For Enrichment Facilities

The International Atomic Energy Agency’s (IAEA) current approach to safeguards at enrichment facilities relies on attended non-destructive assay (NDA), weight measurements, and destructive analysis in addition to unattended methods such as the Online Enrichment Monitoring System (OLEM). These measurement systems are designed to detect specific anomalies at enrichment facilities. This work hypothesizes that combining all these sensors together through machine learning will result in a more robust safeguards system. For example, information from the OLEM system could help detect excess production of enriched material rather than simply monitoring enrichment. The concept of a signature matrix-based approach is applied wherein a wide range of existing process measurements are represented in several correlation matrices. An attention-based convolutional autoencoder is then used to detect and locate anomalies within the signature matrices. This work will compare this new machine learning based approach to a traditional safeguards approach. Discussion on the potential to reduce the frequency of attended measurements in light of this new approach is also provided.