Research On Structural Reliability Analysis Method Based On Surrogate Model Under Random Uncertainty

Aleatory uncertainty is an inherent property of structures and cannot be eliminated in engineering.It widely exists in mechanical structures and seriously affect the safety and reliability of structures.In engineering,the performance function of a component or system is often an implicit function,which makes it difficult to apply with traditional reliability analysis methods.The surrogate model methods replace the original numerical simulation with a simple model.Thus,it has the advantages of small computational burden and easy application.Therefore,reliability analysis by using accurate surrogate model methods can significantly reduce computational burden and improve computational efficiency.The main works and innovations of the thesis are as follows:(1)When using the weighted response surface for reliability analysis,there are problems of reduced computational efficiency and matrix singularity.In order to address these problems,two weighting methods in response surface are analyzed;and the reasons of decreasing in computational efficiency and the singularity of the matrix are studied.The study found that if there are samples with a response value of 0 or very small,matrix singularity will occur when using the iterative weighting response surface.When the sample size is large,the computational efficiency of using iterative weighting response surface is decreasing.Because the samples near the limit state function are more important than others in weighting response surface,this paper proposes a new experimental design method based for response surface.The closer the experimental point is to the true limit state,the greater the role it plays in construction of the response surface.In view of the fact that the response surface model gradually approaches the true limit state surface with the iteration process,a new experimental design method is proposed to combine the most probable point(MPP)for selecting the new training point of the next iteration.The research results show that the samples selected by the proposed method are closer to the limit state surface to improve the computational accuracy.(2)A moving least square method based on sample screening is proposed.Compared with the general moving least square methods,the selected samples are more evenly distributed near the MPP based on the response value of the MPP.The results show the approximation accuracy of the moving least square method at the MPP,the accuracy and efficiency are improved compared with traditional moving least square methods.(3)A convergence criterion of Kriging model is proposed.Considering the stricter convergence conditions of traditional Kriging model,a relaxation factor criterion is developed,which allows for lower prediction accuracy at sample points that have less impact on the Kriging model.Thus,the computational efficiency is improved for krigingbased surrogate models with certain accuracy requirement.