Computational and Interpretational Considerations for Multivariate Analytics in Nuclear Nonproliferation Multisensor Arrays

• Activity Number: 62 - Data Fusion: An Exploration of Practical Aspects
• Type: Topic Contributed
• Date/Time: Sunday, July 28, 2019 : 4:00 PM to 5:50 PM
• Sponsor: Section on Physical and Engineering Sciences
• Author(s): Marylesa Howard and Aaron Luttman and Bethany Goldblum and Christopher Stewart and Zoe Gastelum and Boian Alexandrov and Margaret Hoeller and Daniel J. Champion*
• Companies: Nevada National Security Site and Nevada National Security Site and University of California Berkeley and University of California Berkeley and Sandia National Laboratories and Los Alamos National Laboratory and Nevada National Security Site and Nevada National Security Site
• Keywords: Non-negative Matrix Factorization; Probabilistic Principal Component Analysis; Multi-modal Sensing; Nuclear Nonproliferation

Abstract

Nuclear nonproliferation R&D involves the development of technologies for detecting nuclear materials and detonations. This has focused on the design of physical systems, like radiation detectors, but there is a shift towards development of statistical approaches to extracting actionable information from diverse data. Rather than enhancing the quality of the measurements, instead develop analysis methods to enhance our ability to detect more subtle signatures indicative of illicit activities in the data we already have. Here we present experiments that provide data for the design and development of mathematical and statistical approaches to detecting patterns of life associated with nuclear experimentation. The sensor systems are compact and mobile, with multiple sensing modalities but low-quality data. In this work we present and compare the uses of probabilistic principal component analysis (PPCA) and variants of non-negative matrix factorization (NNMF) to show that, while both approaches successfully classify patterns of life around our experiments, NNMF provides a greater level of human interpretability of the classification, whereas PPCA is more computationally efficient.