Prediction and compensation of subsequent deformation in robotbased incremental sheet forming by application of machine learning

渐进板料成形是一种柔性的、与工件无关的小批量板料成形工艺。由于该工艺的几何精度仍然很低，工业应用还很少，甚至还没有全部实现。其原因主要是缺少精确模拟成形过程的可能性，阻碍了对回弹和后续变形的补偿方法的使用。虽然有多种基于有限元的模拟方法，但由于过程的增量性质而导致的模拟误差的汇总，阻碍了它们的应用。在研究项目中，将通过使用机器学习来追求数据驱动的方法，与有限元模拟相反，机器学习不需要对成形过程进行详细的建模。建立多层人工神经网络(ANN)，根据常用的工艺参数、零件几何形状和刀具轨迹对成形实验的几何精度进行预测，并建立工艺数据库以训练神经网络。最好在实验系列中获得较宽光谱的工艺数据，其中将形成一个系统变化的部分，并用改变的工艺参数进行测量。通过这种方法生成大范围的训练数据，使得神经网络的泛化成为可能，使其适用于任何零件。为了考虑零件几何形状对所产生的几何精度的影响，零件几何形状将被转换为具有固定参数的格式，从而可用于机器学习。这是通过开发几种表示方法来实现的，这些方法的性能用质量标准进行评估。它评价几何表示和零件几何参数的逼近和相关性的质量，利用建立的工艺数据库和开发的性能最好的几何表示，训练多层神经网络。同时，利用参考成形实验中采集的测试数据集对其预测性能进行了验证。然后利用训练好的神经网络，根据预测的几何精度对刀具轨迹进行修正，从而提高零件的几何精度。为了实现这一点，需要扩展现有的刀具路径规划方法，因为ANN的数据驱动性质可能导致极少的预测误差，否则会导致错误的刀具路径。

Incremental sheet forming (ISF) is flexible, workpiece-independent process for manufacturing sheet metal parts in small lot sizes. An industrial application has yet rare or not all taken place due to the still low geometric accuracy of the process. The reason for this is mainly the missing possibility for a precise simulation of the forming process, hindering the use of compensation approaches for springback and subsequent deformation. While there are multiple FEM-based simulation approaches, their application is prevented by summing up simulation errors, caused by the incremental nature of the process. During the research project, a data driven approach will be pursued by the usage of machine learning, which, in contrast to FEM-simulations, does not need a detailed modelling of the forming process. A multi-layer artificial neural network (ANN) will be build up, predicting the resulting geometric accuracy of a forming experiment based on common process parameters, part geometry and the course of the tool path.To be able to train the ANN, a process database will be built up. A preferably wide spectrum of process data will be acquired in an experimental series, in which a systematically varied part will be formed and measured with alternated process parameters. By the this way generated wide range of the training data, a generalisation of the ANN is enabled making it applicable to any part.To take the influence of the part geometry on the resulting geometric accuracy into account, the part geometry will be transformed into a format with a fixed number of parameters which is therefore usable for machine learning. This is achieved by the development of several representation approaches whose performances are evaluated with a quality criterion. This assesses the quality of the approximation and correlation of the parameters of the geometry representation and the part geometry.Utilizing the built up process database and the developed geometry representation with the highest performance, a multi-layer ANN will be trained. Meanwhile its prediction performance is validated with a test dataset gathered in reference forming experiments. Afterwards the trained ANN is used to improve the geometric accuracy of a part by modifying the tool path based on the predicted geometric accuracy. To execute this, existing tool path planning approaches need to be extended as the data driven nature of the ANN can lead to rare prediction errors otherwise resulting in false tool paths.