System Health Diagnosis and Prognosis Using Dynamic Bayesian Networks
Bartram, Gregory Walsh
:
2013-08-12
Abstract
This dissertation develops a methodology that provides information to make optimal decisions with respect to the mission and maintenance of a system. A great amount of information is needed about a system to make such decisions, including its current condition and predictions of its future state. The problem is broken into four subproblems. In the first, a dynamic Bayesian network based system modeling approach is developed for use when the available information is heterogeneous, i.e. available in various formats from various sources (e.g., laboratory data, operational data, expert opinion, mathematical models, and reliability data). The resulting system model accounts for uncertainty and is amenable to the system health management tasks of diagnosis, prognosis, and decision-making. In the second subproblem, a diagnosis approach for systems is developed for systems in the presence of heterogeneous information. The approach accounts for and quantifies uncertainty in the probability of damage, isolation, and quantification. Next, a prognosis is considered when the available information is heterogeneous. The prognosis methodology accounts for uncertainty in diagnosis and is subjected to validation. Finally, a methodology for decision-making problem is developed for problems where the available information is heterogeneous and when assigning multiple systems to multiple missions. The methodology accounts for uncertainty in diagnosis and prognosis. The system modeling, diagnosis, prognosis, and decision-making problems are illustrated using a hydraulic actuator with multiple possible faults, while the diagnosis problem is demonstrated using a cantilever beam with possible damage at the support or a mid-span crack. The proposed methodology is general and can be applied to many different systems.