Research On Methods And Application Of Probabilistic Safety Assessment Based On Bayesian Networks

With the development of science and technology, the performance of modern engineering and technological systems, such as nuclear power plants, aviation and spaceflight systems, and weapon systems and so on, is improved and the structures of them are becoming more and more complicated. The safety accidents of these systems usually result in casualties, equipment destruction, environment pollution, losing wealth and other serious outcomes, so much as political influences; hence, the safety assessment of these systems is of great importance. The current probabilistic safety assessment methods meet some difficulties while dealing with complex dynamic systems which have human and software interactions, process variables and characteristics such as multiple states, non-coherence, failure correlations, etc. New modeling and assessment methods are desired. Bayesian networks developed in the last century, which is famed for its mathematical foundations and powerful modeling and analysis abilities including making use of all kinds of information, are focused by experts of many domains. According to the complicated characteristics of modern engineering and technological systems, several issues of the probabilistic safety assessment methods based on Bayesian networks are studied in this thesis, with respect to:(1) The static systems. The static Bayesian networks of binary-state systems are constructed based on event trees and fault trees. The computation method of outcome probabilities is proposed, and the definitions and computation method of importance of generic systems are also studied. The explicit, mixed and implicit models of common cause problems are proposed. In addition, the algorithms of solving minimal path sets (minimal cut sets) of monotone binary-state systems and maximum lower vectors (minimal upper vectors) of monotone multi-state systems are proposed, and the correctness of the latter is also proved. The probabilistic safety assessment of gas leak problems is performed based on the models mentioned above and the results are compared with those of current methods.(2) The sequential systems. The dynamic Bayesian networks and the discrete-time Bayesian networks are constructed based on event trees and dynamic fault trees. The definitions and computation methods of outcome probabilities and importance of sequential systems are studied. The bank transaction system is analyzed based on the models mentioned above.(3) The control systems. The probabilistic convergence of simple loop is studied, according to the deficiency exists in other literatures, the sufficient and necessary conditions are proposed based on the construction method. The Bayesian networks with loops are constructed for probabilistic safety assessment on simple feedback control systems. The process control systems are modeled using extended dynamic Bayesian networks. The feedback control system of chemical liquid level and the process control system of holdup tank are analyzed based on the models mentioned above.(4) The systems with human and software interactions. The probabilistic safety assessment model based on object-oriented Bayesian networks is constructed using two-phase method according to the interaction processes between software, human and systems. A new parameter learning method of object-oriented Bayesian networks based on evidence theory is studied, and besides, a new method for fusing information of multiple sources based on evidence theory is also proposed in order to determine the prior distribution. The safety gate system is analyzed based on the models mentioned above.In the end of the thesis, the contents are summarized and future work is prospected.