Data And Model Selection To Detect Sparse Events From Multiple Sensor Modalities

Monitoring nuclear facilities is important for nuclear nonproliferation. However, the activities or events of interest are likely to be sparse and occur under variable conditions. In this work, we focus on predicting the power level of a nuclear reactor, where only a few observations (imbalance of ~1:1000) are available for the intermediate power levels (10-90%) using data from multiple sensor modalities (seismic, acoustic, thermal, electromagnetic, and effluent). These sensors are 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. While combining data from multiple modalities offers opportunities for detecting signals that may not be fully captured by any individual modalities, it also poses a few challenges: 1) Not all of the modalities may provide useful information. 2) A number of features could be computed for each modality and some of these features may be less informative or lack robustness across different reactor startups. 3) Often these reactor startups occur in different environmental conditions which may lead to operation signatures that vary across different startups. 4) Depending upon the physical phenomenon that each sensor is designed to capture, different machine learning models may be most effective. In this paper, we present a systematic approach to select data, features, and models to improve prediction of the reactor power level.