Contextually Aware Nuclear Evaluation System

Abstract

The ideal radiation detector system for homeland security would provide an unambiguous signal indicating the presence of a nuclear weapon, special nuclear materials (SNM) or a radiological weapon. In many cases, differentiation between signals from benign sources, such as naturally occurring radioactive material (NORM), medical isotopes and background fluctuations, and those from contraband agents can be achieved through spectroscopic classification and identification. The effectiveness of this differentiation depends largely on the spectroscopic resolution of the system. Currently, radiation portal monitors (RPMs) and hand held radioisotope identifiers (RIIDs) are used to detect and evaluate radioactive materials. When these systems detect radiation, a human expert is often required to intervene to resolve the situation. The expert typically considers the spectroscopic signal while incorporating contextual information (e.g., the size and weight of the container) and past experience in similar situations to make a determination to pass the shipment or initiate a detailed search. This requires a combination of a deep understanding of a specific problem and a hierarchical series of rules. We are developing a contextually aware nuclear evaluation system (CANES). CANES is based on machine learning (ML) algorithms that are trained through repeated exposure to sample incidents and are subsequently able to return an assessment (classification) of a new incident. We present here the overall system design including data inputs, feature vector extraction, and system operation. We will show how the system is optimized to ingest domain expertise directly from nuclear analysts to provide both decision support assistance and complete analysis.