Modeling system based on fuzzy dynamic Bayesian network for fault diagnosis and reliability prediction

Abstract

This paper presents a new method to model a complex system with uncertain and dynamic information for fault diagnosis and reliability prediction. Some system is very complicated, and the relationships between components are also complex. In addition, the exact failure mode of each component may not always be known in many cases. What's mo re, the performance of the system is dynamic and changing with time. Therefore, the conventional Bayesian Network (B N) is not suitable. It is almost impossible to model and analyze its reliability by using conventional methods. In view of this, the article presents a system reliability modeling and assessing method using Fuzzy Dynamic Bayesian Network (FDBN) through fusing various test information. To determine the prior and posterior likelihood, the FDBN-based system fault provides the required quantities. The quantitative analysis of a FDBN can proceed along two lines, the forward (or predictive) analysis and backward (or diagnostic) analysis. Hence this method not only gives accurate reliability prediction, but also fault diagnose. The main steps for modeling are as follows: Firstly, modeling the BN. The topology of BN is founded based on failure analysis and the node information basically includes the failure state and failure rate. Secondly, the fuzzy set theory is presented to BN. Considering both the fuzziness of fault state and fault rate and the uncertainty of fault logical relationship between components, the fault states of system and components are described by fuzzy numbers, fault rates are denoted by fuzzy subsets, and the relationship between components is described as conditional probability table of BN. Therefore, the Bayesian networks are capable of handing the fuzzy information. Thirdly, Static BN can be extended by introducing time dependence of the related to dynamic Bayesian networks (DBN) model, to capture the dynamic variables at different times between the dynamic behaviors of the static network. This paper introduces dynamic process to Bayesian network to model a dynamic system. The sufficient statistics of posterior time slices are estimated using Bayesian probability statistical method, and then the time-variant transition probabilities are learned with both current sufficient statistics and estimated sufficient statistics. At last, this method is applied in the reliability analysis of visual scene matching system of airplane. The results will show that the proposed method is able to make full use of fuzzy information and dynamic process in reliability analysis and can also improve the efficiency of fault diagnosis and reliability prediction.