Extrapolative digital greybox models for describing and predicting the macroscopic system behavior of TiAlN-coated cutting tools

具有硬涂层的刀具构成了当今使用的大多数刀具。它们保护基材不受磨粒磨损，提高耐化学腐蚀性并降低摩擦系数，从而在车削和铣削操作过程中延长刀具寿命。对于刀具寿命的估计，存在各种经验或物理模型。由于这些涂层的性能取决于大量的制造参数，因此没有一致的模型来预测硬涂层的磨损行为。拟议研究项目的目的是在考虑涂层的特殊机械、化学结构和摩擦学特性的情况下，获得涂层刀具与工件材料和工艺参数相互作用时的摩擦学因果关系的知识，并将这些知识映射到模型中。因此，应使用灰箱模型。这些模型依赖于人工神经网络(黑盒)，该网络使用基于加工参数、测量和机械加工测试的元数据的训练数据。磨损率预测和刀具剩余寿命的物理和经验模型被附加到这些模型(白盒)中。这使得该模型能够在训练数据范围有限的情况下预测物理上有意义的解决方案。在机械加工测试中，通过在数控车床的工作空间中集成力、温度、结构声和表面测量技术以及显微镜来实现部分自动化，为同步建模创建了一个不断增长的数据库。为了最大限度地减少实验工作量，从而减少材料和能源成本，在每种情况下，使用主动学习根据模型的当前训练状态来确定下一个实验参数。记录的测量和元数据有多种格式可用，例如图像数据、连续测量或离散测量点。他们被处理并与他们的元数据一起存储在数字实验室书籍中。随后，对数据进行均化和约简，以获得用于训练神经网络的平衡数据集。对于最终的验证，使用模型未知的切割参数和涂层进行盲法测试，以确认灰箱模型的内插和外推能力。

Tools with a hard coating make up the majority of cutting tools used today. They protect the substrate from abrasive wear, increase chemical resistance and reduce the coefficient of friction, resulting in increased tool life during turning and milling operations. Various empirical or physically based models exist for the estimation of tool life. Since the properties of these coatings depend on a large number of manufacturing parameters, no consistent models exist to predict the wear behavior of hard coatings. The objective of the proposed research project is to gain knowledge about the tribological cause-and-effect relationships of the wear behavior of coated cutting tools in interaction with the workpiece material and the process parameters under consideration of the particular mechanical, chemical-structural and tribological properties of the coating and to map this knowledge into a model. Therefor greybox models shall be used. These models are rely on artificial neural networks (blackbox) that use training data based on process parameters, measurements and metadata from machining tests. Physical and empirical models for wear rate prediciton and remaining tool life are additionally integrated into these models (whitebox). This allows the model, with the limited scope of training data due to the high experimental effort, to predict physically meaningful solutions. In machining tests, which are partially automated by integrating force, temperature, structure-borne sound and surface measurement techniques as well as microscopy in the workspace of a CNC lathe, a continuously growing database is created for the simultaneous model building. In order to minimize the experimental effort and thus the material and energy costs, the next experimental parameters are determined in each case on the basis of the current training state of the model using Active Learning. The recorded measurement and metadata are available in a wide variety of formats, such as image data, continuous measurement or discrete measurement points. They are processed and stored with their metadata in an digital lab book. Subsequently, the data is homogenized and reduced to obtain a balanced data set for training the neural networks. For final validation, blind tests, with cutting parameters and coating unknown to the model, are performed to confirm the interpolation and extrapolation capabilities of the greybox model.