For reliability analysis of complex structures with time-consuming implicit performance functions, the computational cost required by direct sampling methods is usually unaffordable for engineering applications. However, the adaptive metamodel embedded in sampling methods can significantly improve reliability analysis efficiency. Therefore, a new metamodel-based directional importance sampling method (Meta-DIS-AK) is proposed for reliability analysis in this article. The main novelty of Meta-DIS-AK is constructing the quasi-optimal DIS density (DIS-D) by the Kriging model and giving a simple rejection sampling algorithm for accurately extracting DIS-D samples. By the constructed quasi-optimal DIS-D, the failure probability is transformed into a two-step estimation of augmented failure probability and a correction item. The updating of the first step is designed to ensure that the constructed DIS-D is well approaching to theoretical optimal DIS-D and obtain augmented failure probability estimation, and the second step updating can guarantee the accuracy of correction item estimation. Meta-DIS-AK not only overcomes the difficulties of constructing DIS-D and extracting importance direction vector samples in the existing DIS but also inherits the advantages of DIS in dealing with high-dimensional and small failure probability problems. Compared with the existing efficient sampling methods combined with AK, the results of the examples fully verify the accuracy and efficiency of the proposed Meta-DIS-AK.
The global reliability sensitivity index (GRSI) can measure the effect of model input variable on the failure probability of the structure and provide guidance for the reliability-based design optimization. In this paper, to efficiently estimate the GRSI, an equivalent form of the GRSI is derived by elaborative combination of Bayes’ theorem and the law of total expectation in the successive intervals without overlapping, and it not only makes the total computational cost independent of the dimensionality of model inputs, but also avoids approximating the probability density function approximation used in the existing Bayes’ theorem based global reliability sensitivity analysis. For further improving the efficiency of estimating the GRSI by the equivalent form, two algorithms are presented by nesting the adaptive Kriging (AK) into Monte Carlo simulation (MCS) and importance sampling (IS), respectively, which are abbreviated as AK-MCS and AK-IS. Results of one numerical example and four engineering applications show that the number of model evaluations by the AK-IS is less than $\text{2\%} $ of that by direct IS, and the model evaluation number by AK-MCS is less than $\text{4\%} $ of that by direct MCS under the convergent condition. The results illustrate that the proposed methods for estimating the GRSI are practical for engineering applications.