基于ResNet失稳模态图像识别和先验知识的加筋圆柱壳精细化优化设计

Refined optimal design of ring-stiffened cylindrical shells based on ResNet buckling mode image recognition and domain knowledge

  • 摘要:
    目的 为了自动甄别加筋圆柱壳失稳模态的类别,利用先验知识提升协同分解优化方法求解加筋圆柱壳精细化优化设计的效率。
    方法 训练ResNet神经网络对加筋圆柱壳失稳模态图像进行识别,提炼基于设计变量与约束特征量耦合关系等的先验知识,提出特定分组策略和资源分配策略。
    结果 数值实验结果表明,训练的ResNet神经网络对加筋圆柱壳失稳模态类别识别的准确率达98%;算例采用先验知识策略的最终优化效果较无先验知识策略结果平均减重7.06%。
    结论 所提优化策略效果显著,对加筋圆柱壳优化设计具有重要参考价值。

     

    Abstract:
    Objectives In the fields of underwater vehicles and aerospace, ring-stiffened cylindrical shells are widely used. Achieving a refined design for such structures is crucial to reducing structural weight while meeting strength and stability requirements. This study focuses on solving the optimization challenges of ring-stiffened cylindrical shells, with each rib as a design variable, to overcome the difficulties in automatically distinguishing whether the buckling mode is global or local. Additionally, it aims to improve the effectiveness and efficiency of the collaborative decomposition optimization algorithm with the assistance of domain knowledge.
    Methods To achieve these objectives, ResNet is trained for the buckling mode image recognition of ring-stiffened cylindrical shells. Using ABAQUS for parametric modeling, numerous finite element simulations are conducted under different design variable combinations to generate buckling mode image datasets under various conditions. These images are preprocessed and divided into training, validation, and test sets for training the ResNet101 model. Meanwhile, based on domain knowledge regarding the coupling relationship between design variables and constraint quantities in ring-stiffened cylindrical shell design, specific grouping and resource allocation strategies are proposed.
    Results The experimental results show that the trained ResNet has an outstanding performance in identifying the buckling modes of ring-stiffened cylindrical shells, with an accuracy rate of 98%. In this case, the volume of the optimized solution using the domain knowledge-driven algorithm is significantly reduced compared to the initial solution. Specifically, the volume is reduced by 38.4% compared to the initial solution, and further reduced by an average of 7.06% using this algorithm compared to the solution without domain knowledge.
    Conclusions In conclusion, combining the neural network-based image recognition technology with the collaborative decomposition optimization algorithm, supported by domain knowledge, effectively solves the optimization problem of ring-stiffened cylindrical shells with different ribs. The high recognition accuracy of the neural network ensures the accurate stability constraint calculations, and the innovative strategies based on domain knowledge enhance the algorithm’s optimization performance and provide a valuable reference for designing ring-stiffened cylindrical shells in related fields.

     

/

返回文章
返回