A comparison of DBN model performance in SIPPRA health monitoring based on different data stream discretization methods

摘要

? 2023The energy and industry sectors depend upon the reliability of complex engineering systems (CESes), such as nuclear power plants or manufacturing plants; it is important, therefore, to monitor system health and make informed decisions on maintenance and risk management practices. One proposed approach is to use causal-based models such as Dynamic Bayesian Networks (DBN), which contain the structural logic of and provide graphical representations of the causal relationships within engineering systems. A current challenge in CES modeling is fully understanding how different data stream discretizations used in developing underlying conditional probability tables (CPTs) impact the DBN's system health estimates. This paper demonstrates the impact that different time discretization strategies have on the performance of DBN models built for CES health assessments. Using simulated nuclear data of a sodium fast reactor (SFR) experiencing a transient overpower (TOP), different strategies for discretizing CES data streams are used to construct the CPTs for a health-based DBN model. This study finds that these strategies generate different models with varying levels of performance for determining different assessments of overall system health. By understanding how these design factors impact the model's health assessments, future risk models can be developed to provide a more meaningful assessment of a system's health, resulting in more informed decisions.