基于深度神经网络代理模型的盾构隧道密封垫断面优化

SHIELD TUNNEL GASKET SECTION OPTIMIZATION BASED ON DEEP NEURAL NETWORK SURROGATE MODEL

合理的弹性橡胶密封垫断面形状是保障盾构隧道管片接缝防水设计性能的关键。密封垫断面优化设计时,需要反复进行材料大变形、接触分析等复杂的非线性计算,极大限制了优化效率。为此,以闭合压力与有效接触压力占比为双控目标,提出了一种结合深度神经网络代理模型的结构优化算法。在遗传算法框架下,深度神经网络代理模型可以实现由断面形状到接触应力场的快速映射。同时,迁移学习的引入实现了不同类型断面形状代理模型的知识复用,仅利用小样本即可建立高精度的接触应力预测模型,从而有效提高了闭合压力约束条件下的密封垫结构断面优化效率。

The key to ensuring the waterproof design performance of a shield tunnel segment joint is a reasonable cross-sectional shape of an elastic rubber gasket.The frequent nonlinear calculations linked to large material deformation and contact analysis in the gasket section optimization design process limit the optimization efficiency.A structural optimization approach compiled with a deep neural network surrogate model is proposed,with the closure pressure and the ratio of effective contact pressure as the dual control objectives.The depth neural network surrogate model is utilized in the framework of the genetic algorithm to swiftly transform the section form into the contact stress field;Simultaneously,the use of transfer learning allows for the reuse of knowledge from surrogate models with variable sectional shapes,enabling for the creation of high-precision contact stress prediction models with few samples.As a result,the optimization efficiency of a sealing gasket structural section under its closure pressure constraint is significantly increased.