Beschreibung
Manufacturing of tool molds represents a single part production characterized by varying designs and various different process steps. The associated milling processes require a precise and complex process planning, which subsequently has to be optimized by running-in tests and adaptions to meet the quality specifications. Moreover, high costs of the raw material and the milling tools require a particularly careful and therefore time-consuming choice of process parameters, mainly based on human experience. Often, subsequent rework becomes necessary. This results in additional efforts during the process. For that purpose, machine learning can be used to find correlations between the process parameters in the process planning and the resulting shape error prior to the first cut. Hereby, the choice of the machine learning algorithm and its hyperparameters largely defines the prediction quality. As a disadvantage, finding the optimum of these hyperparameters to model a process with machine learning can be a tedious, timeconsuming and error-prone procedure that also highly relies on the experience of the respective user. Automated machine learning (AutoML) offers a method to automatically search for a well-performing set of hyperparameters for a specific machine learning application. This study shows the performance improvements achieved by AutoML to predict shape errors that can occur during milling. For this purpose, a series of experimental investigations was conducted to collect representative data in a varying pocket milling process of cold working steel 1.2842. The design of experiment is supposed to ensure a variety of process parameters. As a novel addition, the machine learning model is incorporating the time-variant behavior such as tool wear. Additionally, the study is making a more realistic approach as it is considering error influences from CAD until the machined part in contrast to other studies. We show that we can achieve substantial improvements in terms of prediction RMSE by using the AutoML tool autosklearn; depending on the data between a factor of five and three orders of magnitude compared to plain default settings. This study demonstrates the high potential of using automated machine learning regarding the reduction of efforts in process planning due to improved prediction of shape errors and the ease of using state-of-the-art machine learning.
