For many facilities, including nuclear reactors, operators log text information around events in the facility processes (e.g., starting and securing pumps, open and closing relays, voltage levels, etc.). Associated with each event in the log, there is often a plethora of sensor and control data collected from plant control and monitoring instrumentation (i.e., Supervisory Control And Data Acquisition (SCADA) systems), such as coolant temperature at various process points, radiation levels, etc. Finally, there may also be additional signals from indirect sensors, such as current measurements from power lines. There is currently a challenge for detecting and understanding these controls and sensor data behind important events for nuclear facilities. To address this, we have built interpretable machine learning tools that classify events around sequences of various sensor data and compute which features contributed the most to that event. This is performed using recent approaches in the literature, such as SHAP (SHapley Additive exPlanations) and Local Surrogate (LIME) combined with clustering text data and building machine learning models using time series from SCADA and internal measurement systems, and external indirect sensors. We particularly focus on methods that not only explain important sensor signals but also important time periods for events. We have focused on data collected by the Multi-Informatics for Nuclear Operations Scenarios (MINOS) project, which seeks the combination of multiple, disparate data modalities to characterize the operations at a nuclear facility, specifically instrumenting and studying the High-Flux Isotope Reactor (HFIR) located at Oak Ridge National Laboratory in Oak Ridge, TN. The tools developed by this work on interpretable methods that takes advantage of sensor control systems will improve awareness of operating conditions, including sensing, and potentially allow for anomaly detection, cyber-physical security, and predictive maintenance and monitoring.
Year
2022
Abstract