Study On Active Learning Reliability Method And Its Application To Structural Engineering

Despite siginificant progress made in structural reliability analysis,the reliability assesments for pratical engineering structure still remains a challenging task.Active learning reliability methods(ALRMs)represented by AK-MCS,which combines the adapative metamodel and simution method judiciously,have received extensive attention and development in recent years.However,AK-MCS requires a huge require a large computational effort for problmes with small failure probiblities.Advanced ALRMs has been proposed to overcome this shortcoming,however at the expense of sacrificing the generality and simplicity of AK-MCS.Besides,the current researchs about ALRMs mostly focuses on Kriging with few studies on other widely used meatamodels.Therefore,this dissertation aims at developing and perfecting ALRMs to improve the accuracy and efficiency of structural reliability analysis,as well as to apply the proposed method and provide referenc to practical engineering structures.The marjor research contents are presented as follow:(1)Since AK-MCS is difficult to apply to the problems with small failure probabilities,this paper presents a novel ALRMs called AK-SDMCS based on the idea of spherical decomposition.Firstly,the parameter space is decomposed into a series of mutually nonoverlapping subregions.With the fomula for calculating failure probabilities and coefficient of variation(Co V)derived,as well as the sampling method and Co V control strategy developed,this paper propsed an improved Monte Carlo Method named SDMCS(Spherical Decomposition Monte Carlo Simulation).The algorithm AK-SDMCS is put forward by combining SDMCS with adaptive Kriging.The numerical examples illustrate that AK-SDMCS significantly reduces the candidate sample size under the preservation of AK-MCS’s generality,and effectively reduced the calls to the real performance function compared with other improved methods.Finally,the reason why AK-SDMCS is only applicable to low-to-moderate problems is discussed.(2)For the current researchs about ALRMs mostly focuses on Kriging,and in view of the different predictive incentives between SVR(Support Vector Regression)and Kriging,this paper proposes a novel ALRM ASVR-MCS based on adaptvie SVR to explore the adaptvie SVR’s potential to solve structural realiability problems.Additionally,this paper introduces two learning functions applicable to adaptive SVR and proposes a novel learning function based on weighted penalty from the perspective of optimization.Furthermore,by testing and comparing the performance of ALRMs based on different types of metamodels(Kriging,SVM,SVR),it concludes that ASVR-MCS provides decent results in sloving various categories of reliability analysis problems.Besides,for the superior mapping capacities and flexible regression mechanisms,ASVR-SVR is expert in dealing with performance function with discontinuous derivatives.(3)This paper establishs the general framework of ALRMs applied in pratical enginnnering,designs the corresponding program flowchart and complete the programming work on MATLB.Taking MATLB as the main control program,the application of ALRMs to engineering is realized by using the finite element software’s API interface or other way to call the engineering numerical model.Besides,two examples are given to discuss or verify the feasibility of applying ALRMs to practical structural engineering.