Research And Implementation Of Bayesian Network Structure Learning Method For System Reliability Modeling

As the complexity and integration of electro-mechanical systems increase, these systems are becoming much more intelligent than ever. However, this makes the systems more frequently to get into failures. For complex systems, the failures are not only from the system components themselves, but more probably from the uncertainty relations between them. These associations constitute the causal logic relations between component failures in the system. This has made system reliability assessment more and more difficult, even challenging. Traditional methods of system failure modeling and analysis such as Fault Trees, have been unable to solve this problem.This makes the Bayesian network based structural learning methods gradually applied to the field of system reliability evaluation, and widely recognized. And now, Bayesian networks are widely used in artificial intelligence, machine learning and image processing et al. Therefore, this paper tries to propose some reliability modeling approaches through Bayesian networks. In the framework of Bayesian network theory, these methods can infer the corresponding BNs from the system reliability data. This paper mainly includes the following parts:(1). Describe the history of system reliability and its role, and introduce the basic principles of reliability theory and the related methods.(2). Introduce and analyze in detail the basic concepts, principles and applications of Bayesian networks, and then reveal the relationships between system reliability and Bayesian networks.(3). Introduce the Bayesian network learning methods, and focus on the structure learning approaches. Based on the thoughts of traditional genetic algorithm, we propose the CH-DGA algorithm to model the system reliability with Bayesian networks. It combines the CH score and the dual chromosome encoding of structures so that we can build the system reliability Bayesian network model. This method uses a pair of chromosomes to encode a particular BN structure, and special search operators to ensure that the whole space of BN structures is searched. Simulations show the effectiveness and accuracy of the proposed approach.(4). A hybrid evaluation rule(M-B rule) is proposed based on the MDL criterion and BIC evaluation, which aims to avoid the phenomenon of over-fittings. This rule can effectively combine the advantages of those two criteria, and make a more strict evaluation of the structure. In addition, an improved hill climbing method(IHC method) is introduced to guide the updating process of Bayesian network structures. The IHC method can prevent the abnormal interrupt in structure learning process so as to reduce the occurrence of the under-fitting phenomena. We call this method M-B method. Through the simulation and the reliability estimation applications, we found that the learning results of the M-B technique are in line with our expectations.