Study Of Surrogate Model Approximation And Surrogate-enhanced Structural Reliability Analysis

In recent decades, along with the significant technical progress of new materials,advanced manufacturing technology and simulation analysis based on scientific computing,engineering structural analysis and design techniques are also showing a diversifieddevelopment trend. Several advanced structure design ideologies, which includemulti-objective structure optimization, multidisciplinary design optimization(MDO),structural reliability analysis and design, and robust design, has to be reflected in the actualengineering structural design applications. These new design principles also fulfill currentescalating structural design requirements, which means shorten the design cycle, reducedevelopment costs, improve design quality, and more friendly to the environment andcustomers. On the other hand, even the computer simulation capacity and computationalefficiency has made great progress, widely usage of new materials and new structures designcould still make traditional structure analysis theory of too difficult or complex in assistingstructure designs and fulfilling the structure design requirements. Therefore, it is necessary tostudy new design technologies which adapt to contemporary structural analysis capabilities.As an efficient solution to these problems in structure design, surrogate modeling techniqueshad become a popular research topic. The research focus of this dissertation lays on severalcritical issues related to approximation of surrogate models and its applicability in structuralreliability analysis. The content of dissertation can be summarized as follows:(1)Improved Artificial Neural Networks models and application in characterizing impactdamage of composite structuresAs the structure design application of composite is still advanced than its correspondinganalytical model development, the artificial neural network model (ANN) was studied andextended to composite structure designs based on abundant physical test results on composite materials. It is necessary to improve the traditional modeling process of ANN because of thefeature of training sample data from physical structure test. A cross-validation scheme isutilized to determine the optimal network structural and initial parameters. The complexnonlinear function approximation problem has verified the applicability of proposed ANNmodeling method. The internal damages induced in honeycomb sandwich structure are thencharacterized and predicted with proposed ANN model whose training sample data comesfrom standard low-velocity impact tests and non-destructive damage testing. The results haveproved the hypothesis that surrogate ANN models can approximate responses of compositestructures efficiently, which enabled a possible option in assisting composite structure designs(2) A Two-Phase Hybrid model for structural reliability analysisA Two-Phase Hybrid Method (TPHM) aiming at structural reliability analysis isproposed in current research based on complete performance discussion on basic reliabilitymethods. The TPHM is expected to highlight the advancements of each basic reliabilitymethod and obtain the structure failure probabilityPf efficiently. In the first phase, the firstorder reliability methods (FORM) are utilized to promote the sampling efficiency of surrogatemodels, while FORM can help to locate newly added sample points as close as possible to thelimit state surface. In the second phase, the Monte Carlo method and surrogate models isadopted to estimate the structure failure probabilityPf accurately. The chosen4test examplesdemonstrated the success of sequential adaptive sampling schemes and verified theapplicability of TPHM.(3)Doubly weighted moving least square method and its application in structuralreliability analysisThe object of this research is extending the moving least square method (MLS) tostructural reliability analysis, so the basic hypothesis and principle is firstly explained in detail.However, it is necessary to improve the traditional MLS in order to further embody thefeature of structural reliability analysis. Besides the original weight system in MLS, anadditional weight system to sample points is devised, which enable to assign more weightfactors to sample points that locate more closer to limit state surface or most probable failurepoint. The highlight of sample points close to limit state surface or most probable failure pointcan promote the approximate of themselves, which would do great good to failure probabilityestimation.3numerical test problems had verified the applicability of MLS and doubly weighted MLS in structural reliability analysis, furthermore, the second additional weightsystem enhanced the performance of failure probability estimations.(4)Utilization of simulation sensitivity information in boosting structural reliabilityanalysisThe possibility of combining both the property of computer simulation tools andstructural reliability analysis is discussed in current study. The sensitivity information, whichhad been able to calculated inexpensively by commercial finite element programs, is studiedto boost the capability of structural reliability methods and two possible methodsproposed:1.the simulation sensitivity information outputed by finite element code are treat asthe partial derivatives of performance function to random variables in first order reliabilityanalysis, and three practical engineering structure analysis verified that the cheap sensitivityinformation do great good to convergence of FORM, which is exactly the reason why theaccuracy of failure probability is unacceptable for many problem;2. A sensitivity enhancekriging (SEK) is reformulated based on discussion of kriging theory, and the numericalexample help conclude that the sensitivity information can promote the approximation ofsurrogate models remarkably. An analogical Two-Phase Hybrid Method (TPHM) based onsensitivity information and SEK is proposed and two reliability problems are tested. The testproblem have demonstrated the utilization of sensitivity information in estimating failureprobabilities.