张坤鹏, 郝鹏, 段于辉, 等. 基于深度学习的多级曲线加筋壁板布局优化设计[J]. 中国舰船研究, 2021, 16(4): 86–95. doi: 10.19693/j.issn.1673-3185.02188
引用本文: 张坤鹏, 郝鹏, 段于辉, 等. 基于深度学习的多级曲线加筋壁板布局优化设计[J]. 中国舰船研究, 2021, 16(4): 86–95. doi: 10.19693/j.issn.1673-3185.02188
ZHANG K P, HAO P, DUAN Y H, et al. Layout optimization design of hierarchical curvilinearly stiffened panels based on deep learning[J]. Chinese Journal of Ship Research, 2021, 16(4): 86–95. doi: 10.19693/j.issn.1673-3185.02188
Citation: ZHANG K P, HAO P, DUAN Y H, et al. Layout optimization design of hierarchical curvilinearly stiffened panels based on deep learning[J]. Chinese Journal of Ship Research, 2021, 16(4): 86–95. doi: 10.19693/j.issn.1673-3185.02188

基于深度学习的多级曲线加筋壁板布局优化设计

Layout optimization design of hierarchical curvilinearly stiffened panels based on deep learning

  • 摘要:
      目的   在航空、航天、船舶等领域的结构设计中,因结构的功能性需求,存在大量开口薄壁结构,导致承载性能明显降低。曲线加筋方式虽然在改善开口结构的承载性能方面潜力极大,但设计变量的激增对结构优化提出了挑战。利用数据驱动的深度学习方法,针对含曲筋布局的开口多级加筋壁板开展优化设计。
      方法   针对开口多级曲线加筋壁板,提出结构参数特征图像化表征方法,建立面向结构响应特征学习的深度学习网络模型,实现数据驱动下的结构优化设计,并与经典的由结构数值参数构造的代理模型进行比较。
      结果   结果显示,所提基于图像识别的结构特征学习模型的预测精度改善了约1倍;基于学习模型开展结构的优化设计,多级正交加筋结构的承载性能提升了10.78%,多级曲线加筋结构的承载性能提升了18.19%。
      结论   研究表明,对于设计变量数众多且数量动态变化的开口多级加筋壁板,基于深度学习的结构优化方法更有效;相较传统直筋构型,曲筋方式能更有效地对开口结构承载性能进行补强。

     

    Abstract:
      Objectives   Due to the functional requirements of structures, a large number of thin-walled structures with cutouts are adopted in the structural design of aviation, aerospace, shipbuilding and other fields, leading to a significant reduction in the bearing capacity of such structures. Although the curved stiffening method has great potential in improving the load-bearing performance of open structures, the sharp increase in design variables presents a challenge for structural optimization.The data-driven deep learning method is used to optimize the design of hierarchical stiffened thin-walled structures with cutouts reinforced by curvilinear stiffeners.
      Methods   For structures with cutouts, the hierarchical curvilinearly stiffened method is designed, and the image representation method of structural parameters is proposed. The deep learning network model for structural response feature-learning is established to realize structural optimization design under data-driven conditions.
      Results   The results show that compared with the classical surrogate models constructed by structural numerical parameters, the prediction accuracy of the proposed structural response feature-learning model based on image recognition is improved roughly twofold. In the optimization design of structures based on the learning model, the bearing capacity of hierarchical orthogonal stiffened structures increased by 10.78%, and the bearing capacity of hierarchical curvilinearly stiffened structures increased by 18.19%.
      Conclusions   The results show that this deep learning-based structural optimization method is more effective for hierarchical stiffened structures with large numbers of design variables and dynamic changes in the number of design variables. Compared with traditional straightly stiffened panels, the curvilinearly stiffened panel is more effective in strengthening the bearing capacity of thin-walled structures with cutouts.

     

/

返回文章
返回