Predictive risk modelling with bayesian network for maritime applications

• Main Author: Zhang, Guizhen
• Other Authors: Chan Siew Hwa
• Format: Theses and Dissertations
• Language: English
• Published: 2017
• Subjects: DRNTU::Engineering::Maritime studies
• Online Access: http://hdl.handle.net/10356/71563
• Institution: Nanyang Technological University

Maritime accidents have so far still occurred frequently, threatening the safety for seafarers at sea, the economic performance of shipping companies and the sustainability of the environment. One way to improve maritime safety and reduce the related pollution is to carry out the Formal Safety Assessment, a rational and systematic process to assess the risks associated with shipping activities and evaluate the cost effectiveness of potential risk control options (RCOs). The increasingly popular Bayesian Network (BN) method has been recommended for risk assessment (Step 3 of FSA) to the International Maritime Organization (IMO). However, issues of “data availability”, “data quality” and “dependence on experts” all challenge the use of BN for maritime applications. Therefore, this study aims to enhance the BN methodology to address the limitations of the data constraint for maritime accidents prediction. The modelling of both the Low Probability, High Consequence (LP-HC) accidents and High Probability, Low Consequence (HP-LC) accidents are addressed in this study, with consideration of their inherent differences in nature. The extension of traditional BN with interval probabilities was proposed in the present study for the modelling of LP-HC accidents, for which probability elicitation from experts is the main source both for model structure construction and parameterization. The elicited probabilities often carry different levels of uncertainty due to incompleteness in human knowledge, since experts are not only asked about their own expertise but also about the probability of others’ failures during the elicitation process. Interval probability numbers can represent the imprecision in the experts’ judgement. The extended BN with interval probabilities was applied to the predictive modelling for ship collision causation probability, which enabled the quantification of epistemic uncertainty explicitly in the risk assessment results. An excel-based elicitation tool was developed, utilizing linguistic terms, which allowed rapid elicitation of conditional interval probabilities. Inferences were made directly with the interval probability parameters. Different levels of discrepancies were revealed from the computed results using inputs from different experts, which in turn verified the existence of uncertainty in risk modelling. A discussion was also provided on how the uncertainty in risk assessment propagated to the decision-making process and influenced the ranking of potential RCOs. Probabilistic modelling of the seafarers’ occupational accidents and injuries was then studied as a stereotype of the HP-LC accidents. Compared with the LP-HC accidents, modelling of the HP-LC accidents does not need to rely on the experts’ opinions, since more data could be obtained due to their higher occurrence frequency. Therefore, an extensive empirical survey was carried out using a carefully designed questionnaire, to collect first hand data about the seafarers’ injury experience as well as their behavioral safety practices regarding the potential risk factors. Unlike the historical injury accident databases, the survey captured information for both the accident and ordinary cases, which enabled the probabilistic prediction for the injury occurrences. Meanwhile, the survey data supplemented the experts’ opinion by allowing the verification of potential risk factors with empirical data. The predictive BN model was built consisting of the major risk factors identified from the survey. Insightful results regarding the frequency, circumstances, and causes of injuries aboard merchant ships were discovered. Finally, the BN model was verified with several validation tests.