Model-Based Diagnostic Frameworks for Fault Detection and System Monitoring in Nuclear Engineering Systems

Abstract

The high operations and maintenance (O&M) cost for nuclear plants is one of the most significant challenges facing the industry today. The research in this thesis is motivated by the ongoing effort to utilize automation and improved operator support technologies to reduce O&M costs in nuclear power plants. A diagnostic framework is first developed for the problem of monitoring equipment health and sensor calibration status in nuclear engineering systems. This is achieved by utilizing real-time data from sensors that are already in place for system monitoring to perform automated diagnostics of equipment degradation. Given the long-time scale over which component degradation typically proceeds, some of the sensors may also inevitably degrade and become unreliable. The need to simultaneously consider equipment and instrument faults is both a technical necessity and a desired capability. The automation of these monitoring tasks contributes to the reduction of the overall O&M cost by reducing the required human resources and by providing better maintenance scheduling. Early detection of slow degradation over the course of plant operation requires sufficient detection sensitivity from the diagnostic framework. The problem is more complicated in the presence of various sources of uncertainty and possible changes of operating conditions due to plant drifts. To resolve these difficulties and provide the desired capability, the proposed framework is a hybrid integration of quantitative model-based diagnosis, statistical change detection and probabilistic reasoning. Physics-based models are developed to describe the fault-free behavior of system components. Quantitative residuals are generated from the analytical redundancy in each model and serve as fault symptoms for model-based diagnosis. Statistical change detection methods are used to detect changes in the residuals in the presence of uncertainty. Measurement and modelling uncertainty are robustly treated by methods of statistical change detection and probabilistic reasoning. A system level diagnosis framework is proposed to deal with the lack of local sensors to each component. The overall framework has been implemented and demonstrated with a high-pressure feedwater system whose available sensor set is insufficient for the construction of standalone models for most major components. Results from the demonstration showed that the system level approach can be used to construct models and perform diagnostics for systems with limited instrumentation. Both component faults and sensor faults can be detected, and the effects of uncertainty can be mitigated by the proposed probabilistic reasoning framework. Areas for future work were identified and include the investigation of a dynamic Bayesian network to treat the effects of uncertainty in the diagnosis as well as the investigation of using high fidelity simulation codes to construct simulation-based surrogate models of the basic plant components.