Baksh, Md. Al-Amin (2013) Predictive accident modeling through Bayesian network. Masters thesis, Memorial University of Newfoundland.
- Accepted Version
Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission.
Accident modeling methodologies in literature such as System Hazard Identification, Prediction and Prevention (SHIPP)  consider accident precursors to assess the likelihood of accident occurrence and to design preventive, controlling and mitigating measures for improving the industrial process safety. The SHIPP methodology considers five engineering safety barriers represented using fault and event tree to model the cause consequence relationship between the failure of safety barriers and potential adverse events [1, 2]. In this method, to evaluate the probabilities of end events' occurrence, a restrictive assumption is used that the severity of the adverse events increases only through sequential failures of the five safety barriers considered. -- First, it is strengthen by appending two important non-mechanical safety barriers viz. human and management & organizational factors. We propose to improve the shortcoming of the SHIPP methodology in the following ways. First, we relax the restrictive sequential event assumption in SHIPP methodology by allowing non-sequential failure of safety barriers to cause adverse event of any order. Secondly, in the prediction of posterior probabilities of adverse events for real time industrial data, we include an important mechanical safety barrier viz. 'Damage Control and Emergency Management Barrier (DCEMB)' and as a result we include an adverse event of highest order viz. 'Catastrophe'. Further, posterior probabilities of adverse events are calculated using Bayesian network approach. The posterior probabilities are used to update the safety barrier failure probabilities through a backward analysis and in turn update the estimates of the likelihood continually. The utility of this approach is tested and demonstrated with the data from a liquefied natural gas (LNG) process facility. The method allows for continual updating of occurrence probability for adverse events and failure probabilities of safety barriers for successive monthly data from industry.
|Item Type:||Thesis (Masters)|
|Additional Information:||Includes bibliographical references (leaves 68-84).|
|Department(s):||Science, Faculty of > Computer Science|
|Library of Congress Subject Heading:||Industrial safety--Mathematical models; Bayesian statistical decision theory; Probabilities.|
Actions (login required)