An active learning strategy of reliability-based design and optimization by parallel adaptive sequential importance candidate region method
Mengchuang Zhang,Zhizhuo Zhang,Shasha Xia,Xiaochuan Li,Qin Yao,Zhiping Yin
DOI: https://doi.org/10.1007/s00158-023-03724-4
IF: 4.279
2024-02-02
Structural and Multidisciplinary Optimization
Abstract:Reliability-based design and optimization (RBDO) is increasingly critical for high-quality products. However, existing methods, such as active learning Kriging (ALK) and sequential optimization and reliability assessment (SORA), encounter challenges in efficiently and accurately handling complex nonlinear probabilistic constraints. This paper introduces a strategy for ALK called parallel adaptive candidate region (PACR), which leverages clustering methods (K-means or DBSCAN) inspired by machine learning. The novelty is that this method incorporates advanced techniques for RBDO, including quasi-sequence importance sampling (SIS) with a Markov chain based on quasi-samples obtained by ALK model. Another novelty is that a new learning function for reliability constraints, known as the influence factor function with a limit state constraints g ( x ) = 0 (IFG), has been used to simplify the calculation loop. The efficiency of the methods is validated through various mathematical tests. Then, a specific high-fidelity case study for a planetary roller screw mechanism (PRSM) is conducted to discover an optimal high-precision design with a reasonable load distribution under uncertain constraints. With our method, the standard deviation of the load sharing coefficient decreases by 35.90%, from 0.6221 to 0.3988, thereby affirming the effectiveness of our proposed approach in relatively complex nonlinear optimization problems with uncertain constraints.
mechanics,computer science, interdisciplinary applications,engineering, multidisciplinary