Uncertainty quantification methods for robust online monitoring and recalibration interval extension

Safe, efficient, and economic operation of nuclear facilities (nuclear power plants, fuel fabrication and storage, used fuel processing, etc.) relies on accurate, timely, and reliable measurement of process variables. During operation, components of nuclear facilities, including sensors, may degrade due to age, environmental exposure, and even maintenance interventions. These factors (which could lead to failure of the sensing element) result in anomalies, such as signal drift and response time changes in the measured signal, and challenge the ability to reliably distinguish between signal changes due to plant or subsystem performance deviations and those due to sensor or instrumentation issues. Online monitoring (OLM) is a non-invasive approach to assess measurement accuracy and component condition and provides an alternative to periodic sensor recalibration (the dominant current approach to addressing these instrumentation problems), which is costly, radiation-intensive, and time-consuming. No U.S. plant has implemented OLM for the purpose of extending sensor calibration intervals except for short-term demonstration and research purposes. This is partly due to several technical gaps that must be addressed in order to provide the technical and regulatory bases for these technologies for existing and future plants. Addressing these gaps will require the ability to better quantify sources of OLM uncertainty and associated bounds. This paper presents research supported by the Advanced Sensors and Instrumentation pathway of the U.S. Department of Energy's Nuclear Energy Enabling Technology program on a data-driven uncertainty quantification (UQ) method for OLM. A multi-output Gaussian process technique for UQ, which incorporates both spatial and temporal correlations from measured data in computing the uncertainty bounds, is proposed. The resulting UQ method can track changes in the uncertainty bounds as operating conditions change. Assessment of the proposed UQ methodology using data from an instrumented flow loop indicates the feasibility of generating error bounds on measurement data that are time-dependent. Further, results indicate that the approach may provide insights into the development of fault detection techniques for OLM sensor fault identification and localization.