Predicting the Power Level of a Nuclear Reactor by Combining Multiple Modalities

Detection and quantification of nuclear reactor operations, specifically power level, is of interest for the nonproliferation community for validating the declared operating wattage of a facility. Under the Multi-Informatics for Nuclear Operations Scenarios venture, a collocated research nuclear reactor and reprocessing facility at Oak Ridge National Laboratory have been instrumented with various sensors recording data through approximately one year of operations. Prolonged monitoring has been found prone to outages of sensors at different times, presenting a challenge for combining information from multiple modalities. In this paper, we focus on fusing information from seismic, acoustic, radiation, and electro-magnetic modalities to predict the power level of the reactor. Features of the different modalities will be combined through a classifier that is designed to be robust against data gaps, so that the probability vector of reactor power level can be estimated regardless of gaps in training or testing data sets.