Pattern Of Life Analysis Of A Facility Under Nuclear Safeguards

摘要

In a single-purpose facility, activities associated with modes of operation are reflected by any and all "signs of life" that are represented by facility-related data streams. When appropriately analyzed together, the "signs of life" collectively reveal "patterns of life" of the facility and, as such, may confirm the nature of declared activities or even reveal the presence of undeclared activities otherwise hidden when only individual data streams are scrutinized. While current practices for safeguards verification are largely based on manual and independent interpretation of individual data streams, the value in developing methods that can integrate and utilize the wealth of information contained within large and heterogeneous datasets is widely recognized. A multitude of disparate data streams were collected from a nuclear training facility at Los Alamos National Laboratory to develop a general approach for fusing heterogeneous data streams and validating classes of declared activities. Implicit and/or explicit detection of facility misuse or material diversion would deter would-be proliferators through the threat of early detection. Large data streams have been integrated to identify and classify activities of interest at a LANL facility that is typically used for training of safeguards professionals, such as IAEA inspectors. First, features were extracted from the individual data streams, then cross correlation analysis and machine learning were used to down-select the feature library to a set that carries the most relevant information, and finally supervised learning methods were used to classify modes of facility operations. The fusion of these disparate data streams yields more accurate characterization of facility operations than any data stream individually. The results of the preliminary analysis show that this approach can distinguish operation modes with a rather high degree of confidence. The approach presented here is readily generalizable and applicable to other types of facilities with different sensor types and other data sources.