Research On Constructing Reliability Analysis Model Based On Bayesian Networks And Multi-source Information

Reliability is an important index to mark the technical level of a product. Improving product reliability level is the key to enhance the core competitiveness of the manufacturing industry. Reliability analysis is an important basis for reliability design and a fundamental measure to improve the reliability level of the product. Traditional reliability analysis methods are ineffective for analyzing some widespread problems in complex system, such as polymorphism, uncertainty logical relationship and correlation failure. Bayesian networks have solid theoretical foundation and are suitable for the expression and analysis of uncertainty, so they are widely used in the field of system reliability analysis. However, it often is difficult to construct models by using Bayesian networks due to the lack of reliability data. In order to solve this problem, methods for analyzing and processing different types of multi-source information, such as, the results of traditional reliability analysis methods, expert opinions, field data, to construct and perfect Bayesian network model for system reliability analysis are proposed in the paper. The research achievements of this paper can expand the basic theory of reliability and promote the application of Bayesian network in system reliability analysis. The application of these research achievements in practice can help technical personnel for improving the ability of reliability modeling and analysis for complex system and play an important role in promoting the development of modern equipment manufacturing industry and ensuring the safe and reliable operation of major equipment.The main achievements of this Ph D work are listed as follows:1. Based on the analysis and summary of the basic theories of the Bayesian networks representation, reasoning and modeling, the methods for using Bayesian networks in reliability analysis were discussed. It was analyzed that Bayesian networks had an advantage of dealing with some uncertain problems, such as polymorphism, uncertainty logical relationship and correlation failure. It was proved that Bayesian network was a powerful tool to solve the problems of complex systems reliability analysis.2. Based on the analysis of the advantage and disadvantage of FMEA, FTA and Bayesian networks, two modeling methods, modeling by information extraction and modeling by FFB(FMEA-FTA-BN), were put forward. The modeling approaches were shown. Aim at the shortcomings of modeling by expert experience, such as large subjective deviation and low efficiency, the method using structure matrix to represent the causality in FMEA, FTA and Bayesian networks was proposed. Then, the integration algorithm and rationality test method were presented. As an example, a Bayesian network of the wind turbine gear box was constructed by using the method of FFB. The result of this example proved that the proposed method could integrate FMEA and FTA information quickly and effectively to establish the initial Bayesian network model.3. Based on the analysis of the problems existing in course of obtaining multi-state multi-parent node conditional probability, aim at the fuzziness and subjective deviation of expert opinions, a method dealing with expert opinions by fuzzy group decision making was put forward. According to fuzzy mathematics theory, the linguistic variables from different experts were transformed into triangular fuzzy numbers. By steps viz. equalization, defuzzification and normalization, the precision probabilities were obtained. An example was used to prove that the proposed method was effective. Aim at large workload and difficult judgment for obtaining the conditional probabilities of multi-parent node by use of expert opinions, a conditional probability calculation method of multi-state multi-parent node was put forward. The standardized form of fuzzy number was used to represent a conditional probability under the single cause events independent influence and the calculation models of the conditional probabilities under different causality conjunction were given. An example was used to prove that the proposed method could reduce the workload of collecting expert opinions and had a stronger applicability.4. Based on the characteristics analysis of field data, an update learning method for perfecting Bayesian networks was studied. Aim at the difficulty for determining update occasion, a parameter update method relied on performance monitor was presented. Using collected field data, the prediction results of the existing Bayesian networks were evaluated by scoring rule. If the deviation of the prediction was large, the corresponding parameter would be updated by taking the sample size of field data as the equivalent sample size of expert opinion. Then using Bayes net toolbox(BNT) for Matlab as simulation tool and choosing a classical Bayesian network-Asia net as simulation model, a simulation experiment was carried out. The experimental results shown that the proposed method could select appropriate model update occasion, combine expert opinion with field data effectively, make the parameters approach the ideal value gradually and it was more suitable for updating Bayesian networks by using the field data collected step by step.5. According to the analysis of the function requirements, software for system reliability analysis based on Bayesian networks was designed by Visual Basic+Matlab+Access hybrid programming, which could realize such functions as reliability information collection, Bayesian networks modeling and system reliability analysis. Some pivotal technologies to realize the functions of the system including database design, translating FMEA and FTA into structure matrix and the VB+Matlab hybrid programming were presented. The software developed based on the above theories can give facilities for technical personnel and it is helpful for popularization and application of these theories.