Dynamic safety analysis using advanced approaches

Process systems are prone to accidents as they deal with hazardous material at high temperature and/or pressure. Process plants are also characterized as complex systems where a dense cluster of pipes and equipment may cause a chain of accidents. Therefore, implementation and maintenance of safety measures through risk assessment is crucial to maintain risk below the acceptance criteria. Risk assessment methodologies such as quantitative risk analysis (QRA) and probabilistic safety analysis (PSA) comprise different steps among which accident scenario analysis is a common task. Accident scenario analysis includes accident sequence modeling and associated consequence assessment. Among many techniques available to conduct accident scenario analysis, bow-tie (BT) and Bayesian network (BN) are the most popular. Both techniques are graphical methods illustrating an accident scenario completely and taking advantage of robust probabilistic reasoning engines. BT technique addresses causes and consequences of an accident scenario in a transparent manner that is readily tractable and communicable with stakeholders. However, it suffers limitations of being static and unable to model conditional dependencies. These limitations significantly reduce BT's efficacy to do dynamic risk analysis. In the present study, Bayesian updating and real-time monitoring of operational parameters in the form of physical reliability models are used to overcome these limitations. Physical reliability models provided the analyst with a deeper insight into the behavior of risk while Bayes' rule helps to capture variations over time and to learn from experiences. Bayesian network is an alternative technique to conventional methods such as fault tree and bow-tie, with ample potential in risk assessment and safety analysis. Mapping fault tree and bow-tie into Bayesian network, it is shown that how conditional dependencies, multi-state variables, common cause failures can be considered and most importantly, probability updating can be conducted. Advanced aspects of Bayesian networks such as object-oriented Bayesian networks (OOBN) and discrete-time Bayesian networks (DTBN) are examined in this study. The former decomposes a large network to sub-models with desired level of abstraction, facilitating the modeling and capturing of dependencies. The latter explicitly takes time into account to model sequential failures by means of dynamic gates. To improve the performance of DTBN, an innovative algorithm is introduced to reduce the size of probability tables. Further, two new relationships are developed for dynamic gates cold spare and sequential enforcing gates to make them compatible with most distribution functions. Applying Bayesian networks in the field of domino effects, both propagation pattern and probability of domino effect at different stages are calculated. In this study, the efficacy of BN in safety analysis and accident scenario modeling of a variety of applications such as loss of well control, risk-based design of safety systems and domino effect is examined.