Data-driven sensor recalibration and fault diagnosis in nuclear power plants
Open Access
- Graduate Program:
- Nuclear Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- March 01, 2019
- Committee Members:
- Asok Ray, Dissertation Advisor
- Justin Kyle Watson, Committee Chair
- Kostadin Ivanov, Committee Member
- Kenan Unlu, Committee Member
- Shashi Phoha, Outside Member
- Justin Kyle Watson, Dissertation Advisor
- Kostadin Ivanov, Dissertation Advisor
- James A Turso, Special Member
- Xin Jin, Special Member
- Keywords:
- sensor recalibration
- fault diagnosis
- nuclear power plants
- online monitoring
- autoregressive support vector regression
- resistance temperature detector
- Abstract:
- This dissertation explores techniques for online monitoring of nuclear power plants, especially pressurized water reactor (PWR) plants, which must have the capabilities to examine and diagnose the health of instrumentation and component, recalibrate faulty sensor measurements, and send maintenance request to the control room. Such techniques will enhance the functionality and reliability of the control and monitoring system and reduce the instrumentation maintenance labor requirement and cost. Two data-driven methods are introduced for sensor recalibration. The first method is recursive adaptive filtering that estimates the plant state parameters from a set of redundant sensor measurements. It corrects the redundant measurements based on the principle of best linear least-squares estimation and also detects and isolates anomalous measurements by adjusting their weights, in real time, based on a sequential log likelihood ratio test of sensor data. The second method is autoregressive support vector regression that is a virtual sensing technique; it predicts unknown measurements without the sensor redundancy. A support vector machine is built by “learning” from historical time series measurements and uses measurements from other sensors from previous time instants to estimate the current unknown. The feasibility of both approaches is validated with simulation and experimental data for PWR applications. From these perspectives, an online monitoring scheme is proposed to expand the monitoring capabilities for prognosis of sensor and component degradation. A symbolic dynamics modeling method is used to extract statistical features of time series data at the fast time scale and detect sensor and component degradation when the measurements have not shown observable anomalies at a slow time scale. The extracted features have been shown to produce distinguishable patterns between normal and faulty temperature sensor measurements. This dissertation contains detailed descriptions of the proposed algorithms, theoretical evaluations, pertinent results, and an outlook of how the research will be applied in real plants.