An approach for rapid prediction of textile draping results for variable composite component geometries using deep neural networks
An approach for rapid prediction of textile draping results for variable composite component geometries using deep neural networks
复制标题
一种使用深度神经网络快速预测可变复合部件几何形状的纺织品悬垂结果的方法
DOI：
10.1063/1.5112512
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
L. Kärger
中科院分区：
文献类型：
--
作者：
C. Zimmerling;Daniel Trippe;B. Fengler;L. Kärger
连续纤维增强塑料（CoFRP）在低密度下具有优异的机械性能，因此在过去几十年中在对重量敏感的行业中引起了越来越多的关注。与金属相比，CoFRP的制造由多个步骤组成，通常包括织物的成型过程（覆盖）。然而，在纺织品成形过程中管理固有的复杂、各向异性和非线性材料行为并避免成形缺陷是批量生产中的巨大挑战。为了在制造之前评估可成形性，可以应用虚拟过程模拟。为了获得最佳的零件质量，组件设计和应用的工艺参数必须相互补充，这反过来又需要大量的优化迭代，并迅速超过合理的计算时间。在获得最佳工艺参数方面已经做出了相当大的努力，但是很少考虑几何形状适应性以实现可制造性。使用卷积神经网络（CNN）的深度学习技术能够从提供的样本中学习复杂的系统动态。在这里介绍的工作中，CNN用于快速预测可变组件几何形状的纺织品成形结果。生成高度变化的几何形状和相应的覆盖示例的大型数据库，并在其上训练CNN。该论文表明，CNN能够再现潜在的形成动力学，并且它们可以很好地推广到未知的测试几何形状。与传统的元模型方法相比，所提出的方法不仅估计标量零件质量属性，而且预测完整的剪切应变场，这有利于工程解释。该方法被证明在不同的几何形状，从简单的形状到复杂的几何形状。由于计算成本低廉，CNN在组件设计期间为实时几何迭代提供即时反馈。连续纤维增强塑料（CoFRP）在低密度下具有优异的机械性能，因此在过去几十年中在对重量敏感的行业中引起了越来越多的关注。与金属相比，CoFRP的制造由多个步骤组成，通常包括织物的成型过程（覆盖）。然而，在纺织品成形过程中管理固有的复杂、各向异性和非线性材料行为并避免成形缺陷是批量生产中的巨大挑战。为了在制造之前评估可成形性，可以应用虚拟过程模拟。为了获得最佳的零件质量，组件设计和应用的工艺参数必须相互补充，这反过来又需要大量的优化迭代，并迅速超过合理的计算时间。在获得最佳工艺参数方面已经做出了相当大的努力，但是很少考虑几何形状适应性以实现可制造性。深度学习技术使用…
Continuous fibre reinforced plastics (CoFRPs) offer remarkable mechanical properties at low density and have thus drawn increasing attention in weight-sensitive industries over the last decades. Contrasting metals, manufacturing of CoFRPs consists of multiple steps, often comprising a forming process of a textile (draping). However, managing the inherently complex, anisotropic and non-linear material behaviour during textile forming and avoiding forming defects is a great challenge in serial production. To assess formability prior to manufacture, virtual process simulations can be applied. For optimum part quality, component design and applied process parameters must complement each other, which in turn requires a high number of optimisation iterations and quickly exceeds reasonable computation times. Considerable effort has been made with respect to obtaining optimum process parameters, however considering geometry adaptions to achieve manufacturability is rarely addressed. Deep Learning techniques using convolutional neural networks (CNN) are capable of learning complex system dynamics from supplied samples. In the work presented here, CNNs are used to rapidly predict textile forming results of variable component geometries. A large database of highly variant geometries and corresponding draping examples is generated, on which the CNNs are trained. The paper shows, that CNNs are capable of reproducing the underlying forming dynamics and that they generalise well to unknown test geometries. Contrasting traditional meta-model approaches, the presented method estimates not just a scalar part quality attribute, but predicts the complete shear strain field, which facilitates engineering interpretation. The method is demonstrated on different geometries ranging from simple shapes to complex geometries. Being computational inexpensive, CNNs give immediate feedback for real-time geometry iterations during component design. Thus, CNNs are considered a promising and time-efficient tool to reflect manufacturability during part and process design.Continuous fibre reinforced plastics (CoFRPs) offer remarkable mechanical properties at low density and have thus drawn increasing attention in weight-sensitive industries over the last decades. Contrasting metals, manufacturing of CoFRPs consists of multiple steps, often comprising a forming process of a textile (draping). However, managing the inherently complex, anisotropic and non-linear material behaviour during textile forming and avoiding forming defects is a great challenge in serial production. To assess formability prior to manufacture, virtual process simulations can be applied. For optimum part quality, component design and applied process parameters must complement each other, which in turn requires a high number of optimisation iterations and quickly exceeds reasonable computation times. Considerable effort has been made with respect to obtaining optimum process parameters, however considering geometry adaptions to achieve manufacturability is rarely addressed. Deep Learning techniques using...
影响因子：
6.3
作者：
Kaerger, Luise;Galkin, Siegfried;Wolf, Klaus
通讯作者：
Wolf, Klaus
影响因子：
6.3
作者：
Kaerger, Luise;Bernath, Alexander;Henning, Frank
通讯作者：
Henning, Frank