Potential of systematically generated training datasets on the accuracy and generalization of AI-based approaches for the automated identification of machine control signals

The automated identification of data sources in machine tools is steadily gaining importance due to the growing use of data-driven methods. However, due to small size and volatile information content of training data sets, even promising AI approaches are increasingly reaching their limits. With systematically generated unique reference trajectories, large-scale data sets with high information content can be created. In this paper, the effects of these data sets on accuracy and generalization of AI-based approaches are investigated in comparison to the training with process data. In particular, the subset optimizing the identification of real-world process data is addressed.