Data-Driven Modelling of Metal Bending Processes

复杂弯曲件的质量在很大程度上取决于用于此目的的多阶段弯曲和矫直操作的设计、协调和实施。一个特殊的挑战是控制跨阶段或数量依赖性效应。目前，这通常只能通过使用专业知识来实现，并且由于相互作用的复杂性而并不总是成功的。特别地，影响例如半成品性质、基础摩擦学系统或热和机械工具和机器行为的时间和空间波动的扰动变量（随机性质）是有问题的。在复杂弯曲线材零件中显著提高单个特征的精度和可重复性的出发点是将多级机电矫直机与类似的弯曲单元组合在一起，形成完整的机器系统（机电弯曲机（MSA））。然而，为了实现面向目标的适用性，面向目标的，跨阶段和数量依赖的建模是必需的，这样的制造过程，机器系统和相应的工具的设计。因此，计划中的研究项目的目的是研究数据驱动的混合模型，并将人工智能方法集成在一起，以及它们在这种多阶段和数量依赖的弯曲过程中的适用性。除其他事项外，这应该能够跨阶段和件数进行充分的缺陷或数据跟踪。因此，研究的重点是数据收集和数据库，组件属性和质量特性之间的关系和相互作用的领域。以这种方式收集的数据的有针对性的组合，使用适当的混合人工智能模型，允许整合专家知识和物理条件，应该能够通过可解释性组件进行交互式数据分析。这为第一个资助期的关键成果提供了基础--初始参考数据集，其中包含基于数据的工艺流程表示，即多阶段矫直和冲压弯曲，以及结合领域知识和学习分析结果对机器学习方法的初步调整。

The quality of complex bended parts depends significantly on the design, coordination and implementation of the multi-stage bending and straightening operations used for this purpose. A particular challenge is the control of cross-stage or quantity-dependent effects. At present, this is often only possible by using expert knowledge and is not always successful due to the complexity of the interactions. In particular, temporally and spatially fluctuating disturbance variables (of a stochastic nature), which affect e.g. the semi-finished product properties, the underlying tribological system or the thermal and mechanical tool and machine behavior, are problematic. A starting point for dramatic improvements in, among other things, the accuracy and repeatability of individual features in complex bent wire parts is the combination of a multi-stage mechatronic straightener with similar bending units to form a complete machine system (mechatronic bending machine (MSA)). However, in order to achieve target-oriented applicability, target-oriented, cross-stage and quantity-dependent modeling is required for the design of such a manufacturing process, machine system and corresponding tools. The aim of the planned research project is therefore the investigation of data-driven, hybrid models with the integration of AI methods and their applicability in such multi-stage and quantity-dependent bending processes. Among other things, this should enable sufficient defect or data tracking across stages and piece numbers. Accordingly, research focuses on the areas of data collection and data bases, relations and interactions between component properties and quality characteristics. A targeted combination of the data collected in this way, using appropriate hybrid AI models that allow the integration of expert knowledge and physical conditions, should enable interactive data analysis through components of explainability. This provides the basis for a key result of the first funding period - initial reference data sets with a data-based representation of the process sequence, i.e. multi-stage straightening and punch bending, with an initial adaptation of machine learning methods in combination with domain knowledge and the learned analysis results.