A new multiple stochastic Kriging model for active learning surrogate-assisted reliability analysis

The Kriging model-assisted reliability analysis method is widely recognized as an effective way to evaluate structural failure probability. However, accurately estimating failure probability is challenging due to the inherent limitations of the Kriging model in accounting for response noise during the modeling process. This limitation undermines the accuracy of emulation in reliability analysis, significantly reducing the confidence of the reliability evaluation. To overcome this challenge, this paper proposes an active learning Lasso-based multiple stochastic Kriging model-Monte Carlo simulation method. First, a Voronoi-based adaptive proximity-guided sampling strategy is presented to sample important MCS points near the limit state surface by continuously identifying sensitive Voronoi cells. These identified MCS points are then used to select the stochastic Kriging model components, thereby ensuring that the selection process prioritizes the most informative regions. Second, a Lasso-based model selection strategy is proposed to account for the model-form uncertainty in the multiple stochastic Kriging modeling process, which optimizes and selects the best ensemble of multiple stochastic Kriging model components for the framework of the surrogate ensemble-assisted reliability analysis method. The effectiveness of the proposed method is demonstrated through numerical and engineering case studies. Results show that the proposed method provides more accurate failure probability estimation with fewer calls to limit state functions compared to existing methods, improving predictive accuracy and computational efficiency in structural reliability analysis.