Six sigma robust optimization method based on a pseudo single-loop strategy and RFR-DBN with insufficient samples. (December 2021)

• Record Type: Journal Article
• Authors: Yu, Huijie; Yang, Jiaqi; Ding, Xiaohong; Wang, Haihua; Wang, Shenlong
• Is Part Of: Computers & structures. Volume 257(2021)
• Publication Date: 2021-12
• Subjects: Design for six sigma -- Uncertainty -- Robust optimization -- Pseudo single-loop optimization strategy -- Deep belief networks; Structural engineering -- Data processing -- Periodicals; Electronic data processing -- Structures, Theory of -- Periodicals; 624.171
• DOI: 10.1016/j.compstruc.2021.106653
• Languages: English
• ISSNs: 0045-7949

Abstract

Highlights: A highly efficient six sigma robust optimization method is proposed considering insufficient samples. Pesudo single-loop strategy is applied to plan an assessment region by setting a reasonable threshold. Insufficient samples is utilized based on the ensemble of random forest regression and deep belief networks. The results of the case demonstrates the validity of the proposed method. Abstract: Design for six sigma has become increasingly important in complex optimization work considering uncertainty. In this paper, we present a six sigma robust optimization method based on a pseudo single-loop optimization strategy and an ensemble of random forest regression and deep belief networks (RFR-DBN). To verify its validity, we take a lightweight passenger car seat with insufficient samples as an example. We utilize intractable insufficient samples in a complex optimization problem to learn the key features for various responses and extract them separately for surrogate models from the RFR-DBN. In addition, by employing multi-island genetic algorithm and Monte Carlo simulation based on descriptive sampling, we perform quality improvement and quality assessment to find the optimal solution. Through the pseudo single-loop optimization strategy, we avoid extensive calculations in the optimization process. We demonstrate from the analytical results that the proposed method is a solution to the efficiency of optimization and insufficient samples.