One of the activities of interest for nuclear non-proliferation is monitoring nuclear facilities. This is a challenging task because in realistic scenarios, the events of interest occur under variable environmental or operating conditions and a limited set of labeled data, covering only a subset of events, is available. Predictions about new or future events needs to be made based on the subset of data available. Traditional machine learning algorithms such as supervised (e.g., classification) and unsupervised learning work entirely with a labeled and unlabeled dataset, respectively, and couldn’t be used to make predictions about unseen events directly. In this work, we use concepts from transductive learning (a form of semi-supervised learning), clustering, and other distance-based methods to make relative quantitative predictions about events that may not be covered by available labeled data. We apply our approach to prediction of reactor power level where data about intermittent power level may not be covered by the labeled dataset. We term this problem as an intermittent data evaluation and evaluate our approach using sensors positioned near a collocated research nuclear reactor and reprocessing facility at Oak Ridge National Laboratory for the Multi-Informatics for Nuclear Operations Scenarios (MINOS) venture.
Year
2022
Abstract