Exploring a Bayesian approach for structural modelling of common cause failures
Common Cause Failures (CCFs) are a class of dependent failures that occur to complex technological systems, such as nuclear power plants, where redundant components serve as multiple layers of defence. For the purposes of quantitative assessment of CCFs, parametric models are used. A common feature of all parametric models is the difficulty in parameter estimation due to limited available observational data. The Unified Partial Method (UPM) for CCF modelling is a systematic methodology that takes into consideration physical and operational system defences. This research explores the application of the Influence Diagram (ID) formalism in order to extend UPM, through an example of Emergency Diesel Generators from nuclear power plants. The proposed model incorporates intermediate stages in the modelling process, namely root causes and coupling factors, to allow for a representation of the CCF mechanisms. Moreover, it captures interactions existing amongst the system's defences, in their contribution to risk. With an underlying Bayesian approach to risk, the model quantifies operational experience, accounts for the epistemic uncertainty, and allows for a coherent combination of expert opinion with observations. This thesis proposes a model structure, which integrates with the ICDE generic database for CCFs. Finally, the ID formalism allows for the propagation of uncertainty within the model structure, and provides a tool for decision-making. The construction of the ID model has been entirely based on expert judgment: the model network has been constructed with the help of experts, whilst a suggested model quantification methodology has been explored. This thesis documents the building process, and explores the behaviour of the resulting model. Findings within this research suggest the feasibility of the proposed methodology for development of a CCF model with a structural and exploratory character.