مدل ریسک مبتنی بر شبکه بیزی بهبود یافته و کاربرد آن در ارزیابی ریسک سانحه Improved Bayesian Network-Based Risk Model and Its Application in Disaster Risk Assessment

1 Introduction Risk is the consequence of interactions between risk factors and risk-bearing objects (Grandell 1991) in a multidimensional and multilayered system. Risk assessment, the core of risk science, is a comprehensive evaluation and estimation of the occurrence of risks and losses (Zhang 2013), and constitutes an important research area in the field of management and decision making. Qualitative risk assessment is mainly based on expert knowledge, whereas quantitative risk assessment uses mathematical methods (Bu¨hlmann 1996). Considering risk cognition and risk information, both risk and its assessment are uncertain, or rather fuzzy and random (Pate´-Cornell 1996). There are several culprits that create this uncertainty: (1) the randomness of attributes of risk such as time, frequency, and intensity; (2) the incompleteness and ambiguity of environmental information; and (3) dependency on subjective knowledge. Therefore, the expression and handling of fuzzy and random information is one of the key issues in modern risk assessment modeling (Yu 2017). There are many risk assessment methods, including qualitative and quantitative ones. Classic methods, such as the analytic hierarchy process (Saaty 1980), fuzzy comprehensive assessment (Yang and Yang 1998), and grey system theory (Deng 1990), are used widely. The analytic hierarchy process (AHP) combines qualitative judgment with quantitative analysis to handle subjective preference in a quantitative way (Al-Harbi 2001). Fuzzy comprehensive assessment (FCA) can process ambiguous information with quantitative mathematical expressions (Ruan et al. 2005). Grey evaluation applies grey relational analysis (GRA) to assessment modeling with incomplete data (Zheng and Hu 2009). However, these classical methods are mainly based on subjective experience and expert knowledge. Neither weight calculation in AHP nor affiliation determination in FCA takes advantage of objective data, indicating a relatively strong subjectivity and low credibility in assessment. Furthermore, these methods have defects in describing nonlinear interactions between risk factors. In the past two decades, some emerging research techniques have been utilized in risk assessment. Prominent among them have been the neural network (NN) approaches discussed by Hagan and Beale (2002), cloud models (Li and Liu 2004), event tree analysis (ETA) applied by Ericson (2005) to system safety studies, and Petri net (PN) techniques that Girault and Valk (2003) introduced into systems engineering. NN in particular has capability in parallel computing, self-learning, and fault tolerance, which can achieve nonlinear modeling for complex systems. Cloud modeling combines fuzzy theory with probability theory for the expression of uncertain information. ETA and PN can describe the interinfluence of factors visually and achieve the rigorous inference of multisource information.