Application of Machine Learning for Classification of Nuclear Reactor Operational Status Using Magnetic Field Sensors

Abstract

The nuclear fuel cycle forms the basis for producing special nuclear materials used in nuclear weapons via a series of interdependent industrial operations. These industrial operations each produce characteristic emanations that can be gathered to ascertain signatures of facility operations. Machine learning and deep learning techniques were applied to time series magnetic field sensor data collected at the High Flux Isotope Reactor (HFIR) to assess the feasibility of determining the ON/OFF operational state of the reactor. When data collected by the sensor near the cooling fans, position 9, are transformed to the frequency domain, it was found that both machine and deep learning methods were able to classify the operational state of the reactor with a balanced accuracy of over 90%. This result suggests that the utilized methods show promise for application as techniques to verify declared activities involving nuclear reactors. Additional effort is recommended to develop models and architectures that will more fully capitalize on the data’s temporal nature by incorporating the magnetic field’s time dependence to improve the model’s robustness and classification performance.

Keywords

• machine learning
• deep learning
• random forests
• nuclear reactors
• frequency domain
• classification