Sampling-based system reliability-based design optimization using composite active learning Kriging. (15th October 2020)

• Record Type: Journal Article
• Authors: Zhang, Jinhao; Xiao, Mi; Li, Peigen; Gao, Liang
• Is Part Of: Computers & structures. Volume 239(2020)
• Publication Date: 2020-10-15
• Subjects: System reliability-based design optimization -- Stochastic sensitivity analysis -- Kriging metamodel -- Composite active learning -- Uncertainty quantification; Structural engineering -- Data processing -- Periodicals; Electronic data processing -- Structures, Theory of -- Periodicals; 624.171
• DOI: 10.1016/j.compstruc.2020.106321
• Languages: English
• ISSNs: 0045-7949

Abstract

Highlights: A new method is proposed for system reliability-based design optimization. A composite active learning strategy is proposed for locally refining Kriging. Multiple Kriging uncertainty is considered in the termination of Kriging update. Results of four examples verify the accuracy and efficiency of proposed method. Abstract: This paper proposes a sampling-based system reliability-based design optimization (SRBDO) method with local approximation of constraints. To enhance the optimization efficiency of SRBDO problems with time-consuming constraints, Kriging metamodels are employed to replace the true constraint functions. A new composite active learning strategy based on the possibility of correctly predicting the state of the cut-set system is developed to locally approximate constraints. Furthermore, to ensure the accuracy of the system reliability analysis at the final SRBDO solution, the Kriging update in the developed strategy is terminated by quantifying the influence of the Kriging uncertainty on the prediction of the system failure probability and the confidence that the solution satisfies the prescribed system failure probability. This approach can avoid the unnecessary burden of Kriging construction during system reliability analysis at intermediate solutions far from the final solution. Based on the updated Kriging metamodel, the system failure probability of constraints is estimated by Monte Carlo simulation, and its partial derivative is calculated by stochastic sensitivity analysis. The performance of the proposed method is tested and verified by using four examples. Compared with the existing methods, the proposed method has high computational accuracy and efficiency for solving SRBDO problems.