Multimodal Data Analytics for Nuclear Facility Monitoring

Recent progress in the development of multisensor networks has opened opportunities for indirect physical sensing for proliferation detection applications, where numerous data modalities can be monitored in real-time—including acceleration, magnetic field, pressure, temperature, humidity, ambient light, and passive IR, among others. In networks of multisensors, the space of possible measurements is large: a network composed of only five nodes generating seven data products per second corresponds to over three million data products per day. This work demonstrates the viability of indirect physical sensing for characterization of nuclear facility operations and explores reduced sets of sensing modalities required for specific detection priorities. Data were collected in four distinct experimental campaigns at the 88-Inch Cyclotron at Lawrence Berkeley National Laboratory. Using supervised learning, the cyclotron operational status was consistently classified to above 95% accuracy with a multilayer feed-forward neural network and the full suite of sensing modalities as inputs. To provide guidance in feature engineering for multisensor devices in nonproliferation applications, permutation importance was used to evaluate the sensing modalities according to their contribution to each model’s classification success. The lowest-contributing modality was eliminated from the model inputs, models were retrained on the resulting subset of the input features, and classification performance was re-evaluated. This process was recursively implemented to yield the classification accuracy as a function of sensor loadout. The result is a feature engineering framework that provides a quantitative basis for optimized multisensor design for specific deployment scenarios.