A novel fuzzy dynamic Bayesian network for dynamic risk assessment and uncertainty propagation quantification in uncertainty environment

Risk assessment (RA) plays a vital role in safety engineering. The conventional RA approaches have limited capabilities in handling time dependence and data uncertainty. Although dynamic Bayesian network (DBN) is robust in inference under uncertainty due to its flexible structure and capability of modeling the interdependencies of variables, it still has some defects in quantifying the uncertainty (probability range) propagation over time, and dealing with inaccurate or insufficient data (data uncertainty). This study is aimed to propose a novel fuzzy dynamic Bayesian network (FDBN) methodology to improve the ability of dynamic risk assessment (DRA) methods to quantify and propagate uncertainty arise from inaccurate or insufficient data. The methodology incorporates the fuzzy set theory (FST) with DBN to conduct DRA under data uncertainty while quantifying the uncertainty propagation over time. The proposed methodology represents the causality of variables in the time dimension and adopts expert elicitation and FST to determine the probability of causality. Triangular fuzzy numbers are used throughout the entire dynamic modeling process of DBN to completely retain the uncertainty information. A comparison between the proposed novel FDBN and crisp value based DBN verifies the credibility, rationality and robustness of the proposed methodology. A multi-variable risk assessment of the cotton warehouse is presented here to illustrate the potential of the proposed methodology in dealing with dynamic risk with uncertainty.

Automated summary

Risk assessment is crucial in safety engineering, and conventional methods may have limitations in handling time dependence and data uncertainty. This study introduces a novel fuzzy dynamic Bayesian network (FDBN) methodology to improve dynamic risk assessment by incorporating fuzzy set theory and expert elicitation to quantify uncertainty propagation over time under data uncertainty. The comparison with traditional methods shows the credibility and robustness of the proposed methodology.