Data-based die spotting in sheet metal forming

在钣金成形中，30%的工具开发成本是由于工具加工。原因是制造误差、力流中所有元件的弹性变形以及成形过程中板材厚度的变化。由于在仿真模型中考虑所有的影响和不确定性是不现实的，它们的影响只能在真实的模具上进行校正。磨合包括对有效表面进行点样和机械加工，以根据均匀的点样图案和规定的材料流生产出优质零件。模具定位是一项人工、耗时、经验性强的体力劳动。必须处理异构数据和抽象信息，并且必须考虑与后续过程的交互。熟练工人的短缺迫使科学界为自动化工具熟悉化奠定基础。目前还没有发现点样图像与烧蚀量之间的相关方法，也没有描述结合策略的数学形式化。无法确定用于芯片定位的自动化解决方案。最后，没有记录将经过训练的工具状态转移到后续几代工具。在不同的人工智能方法的组合中，可以看到自动化芯片定位的有前途的方法。以下研究问题（FF）需要回答和假设（H）测试：FF 1：哪种类型的神经网络与网格拓扑结构是适合自动生成的主动表面的成形工具，考虑到工具-机器的相互作用？H1.1：具有处理空间数据能力的NN可以解决上述设计问题。H 1.2：表示学习或符号AI算法代表其他解决方案。H 1.3：使用模拟进行预训练，提高了ML模型的鲁棒性，尽管数据量很小。FF 2：什么类型和拓扑结构的神经网络是适合学习的设计功能方面的机器参数，如力和速度的历史？H 2.1：设计功能根据成形问题和机床的描述识别参数。必须为两者定义形式描述。同样，在模拟数据上预训练NN可以提供一种解决方案。FF 3：如何实现芯片定位的自动化？ H 3.1：基于相机的系统捕获并分析有效表面的点样图像和成形部件的2D图像，以确定所需的烧蚀量。H 3.2：可以根据以前的数据和模拟计算优化的机器和工艺参数。H 3.3：学习函数g（工具熟悉）的经验可以用于改进函数f（工具设计）。

30 % of the development costs of a tool in sheet metal forming are due to the tool machining. The causes are manufacturing inaccuracies, elastic deformations of all elements in the force flow and the change in sheet thickness during forming. Since it is not realistic to consider all influences and uncertainties in simulation models, their effects can only be corrected on the real die. The run-in includes the spotting and mechanical machining of the active surfaces to produce a good part based on a uniform spotting pattern and a defined material flow. The die spotting is manual, time-consuming and experience-based work with strong physical stress. Heterogeneous data and abstract information must be processed and interactions with subsequent processes must be taken into account. The shortage of skilled workers is forcing the scientific community to create foundations for automating tool familiarization. No method of correlation between spotting image and the amount of ablation has been found. No mathematical formalization of an incorporation strategy has been described. No automated solutions for die spotting could be identified. Finally, no transfer of the trained tool state to subsequent generations of tools has been documented. Promising methods for automatable die spotting are seen in the combination of different AI approaches. The following research questions (FF) need to be answered and hypotheses (H) tested: FF 1: Which type of NN with which mesh topology is suitable for automated generation of the active surfaces of the forming tools considering the tool-machine interaction? H 1.1: NNs with the ability to process spatial data can solve the above design problem. H 1.2: Representation learning or symbolic AI algorithms represent other solutions. H 1.3: Pre-training with simulations increases the robustness of ML models despite small amounts of data. FF 2: What type and topology of NN is suitable for learning the design function in terms of machine parameters such as force and velocity histories? H 2.1: The design function identifies parameters based on descriptions of the forming problem and the machine. A formal description must be defined for both. Again, pre-training NNs on simulation data could provide a solution. FF 3: How can the task of die spotting be automated? H 3.1: A camera-based system captures and analyzes spotting images of the active surfaces and 2D images of the formed part to determine the amount of ablation required. H 3.2: Optimized machine and process parameters can be calculated based on previous data and simulations. H 3.3: Experiences from learning function g (tool familiarization) can be used to improve function f (tool design).