Research On Methods And Application Of Reliability And Reliability Sensitivity Based On Surrogate Models

The reliability methods based on surrogate models are important components in the methodology of reliability and reliability sensitivity analysis,and they are widely used in the academic field of reliability research.The application of surrogate models balanced the computational efficiency and calculating accuracy in a certain extent.In this dissertation,on the assumption of that the uncertainties could be described by the stochastic models,the applications of surrogate models to reliability and reliability sensitivity analysis are studied and kinds of analysis methods are proposed.The artificial neural network(ANN),Kriging metamodel and polynomial response surface are the specific realization of the surrogate models.At the same time,a general-purpose reliability analysis and design software system is developed.Furthermore,the research on reliability engineering applications is also carried out,the key structure of a missile and the typical mechanism of a glide bomb are selected as the specific engineering examples.The main contents and the primary highlights could be briefly described as follows:1.With the introduction of Markov Chain Monte Carlo(MCMC)method,the traditional experimental design of ANN is improved.Then,3 kinds of ANN-based numerical simulation methods for reliability analysis are proposed with the combination of importance sampling,truncated importance sampling and multi-center importance sampling respectively.The proposed methods take full advantages of the adaptability of Markov Chain and the outstanding nonlinear fitting capability of ANN,and reduce the dependency on the design of experiment.The well-trained ANN could be used to accurately predicate the responses of the numerical simulation samples and thus guarantees the accuracy of the results.For the reason of that the responses are efficiently predicated by ANN,the calculating efficiency is greatly increased as well.Results of numerical and engineering examples show that the proposed method could be used to deal with kinds of reliability problems,and the failure probability could be accurately obtained with high calculating efficiency.2.In order to deal with the implicit reliability problems which are widely existed in practical engineering,a series of adaptive numerical simulation methods are proposed based on active learning Kriging metamodel.The proposed methods inherit the advantages of active learning mechanism,Markov chain and Kriging metamdel,and greatly increase the efficiency on the premise of that the calculating accuracy is guaranteed.The introduction of Markov chain improves the adaptability of the methods,and ensures that all the experimental samples which are used to construct the Kriging metamodel are located in the interested regions.The active learning mechanism fully utilizes the predicted variance of Kriging model,and improves the fitting ability,thus increases the predicating accuracy to the numerical simulation samples.The introduction of Kriging model reduces the response calculating times to the actual performance function,thus greatly increases the calculating efficiency.The application results to several examples prove the applicability and efficiency of the proposed methods.3.On the basis of polynomial response surface,a series of semi-analytical methods are proposed for response cumulative distribution function(CDF)sensitivity and importance measure analysis.Under the assumption that the input-output response function could be approximated by the polynomial response surface,the probability moments and the partial derivatives of probability moments with respect to distribution parameters are analytically derived,and the detailed CDF sensitivity solving formulas are proposed based on the fourth moment method as well.At the same time,the formulas to estimate conditional expectation and conditional variance at the instances of independent and correlated input variables are derived respectively,and the detailed estimating processes are given,which achieves the estimation to variance-based importance measure.With the combination of polynomial response surface,the proposed methods estimate the CDF sensitivity and the variance-based importance measure of input variables semi-analytically,and enables to efficiently estimate the reliability sensitivity indices.The obtained ranking order of the input variables to the responses could be used to provide reference information to improve the structural or mechanical design.4.A series of numerical simulation methods are proposed based on Kriging metamodel,focusing on CDF sensitivity and importance measure analysis.With the consideration of that the fitting capacity of Kriging metamodel is excellent,the proposed methods show excellent applicability,and could be used to efficiently solve the reliability sensitivity problems with the calculating accuracy maintained.The proposed methods improve the numerical simulating strategy,reduce the numerical simulation samples,and decrease the computational demand in a further step.The numerical simulation solving strategy guarantees the accuracy of the reliability sensitivity results which are obtained on the basis of Kriging metamodel.Results of several validation examples confirm the applicability of the proposed methods.5.A general-purpose reliability analysis and design software system was developed,which makes the reliability and reliability sensitivity analysis methods could be realized in a software engineering approach.The developed software integrates kinds of probability distribution models and a variety of analysis methods for reliability and reliability sensitivity analysis,and could be used to solve various kinds of reliability problems.The software provides necessary interfaces to commercial CAE tools,which enables to communicate with other commercial CAE tools,and makes the software could be used to solve the implicit reliability problems,thus enhances the applicability of the software.The developed software provides a familiar Windows-based user interface(UI)with a simple operate way,and illustrates the reliability analysis results with necessary figures and tables.At the same time,the software enables to generate simple reliability analysis reports.Applications to numerous numerical and engineering examples demonstrate the powerful solving ability and the application prospects of the developed software.6.The engineering applications of the proposed methods are carried out with a certain type of missile rudder structure and a glider-bomb wing unfolding mechanism.The parameterization for the FEM model of rudder structure and the virtual prototype model of wing unfolding mechanism is accomplished respectively,and the parameterized CAE models are obtained.The Kriging metamodel which could be used to accurately predicate the mechanical properties of the rudder structure and the movement characteristic responses of the wing unfolding mechanism is constructed respectively.The failure probability,CDF sensitivity and variance-based importance measures are obtained for the rudder structure and the wing unfolding mechanism with the numerical simulation approach.Under the premise of considering inputs' uncertainties,the reliability assessments are obtained and some specific approaches which may improve the structural or mechanical properties are proposed.