Active learning for the prediction of shape errors in milling

Abstract

In machining processes, various influences, such as workpiece and tool geometry, process parameters or tool wear, can lead to decreasing part quality. Thus, predicting these influences to enable better process planning is essential. However, no general, analytical model is available to facilitate this. Data-driven approaches, on the other hand, are costly due to the required data labeling efforts. To address this, the authors present a data-driven active machine learning approach to predict shape errors based on process data enhanced by a material removal simulation. Using two representative pocket milling datasets, it is shown that this approach can yield better (up to 5 % decrease of RMSE) and more consistent (up to 85 % decrease in standard deviation of RMSE) model accuracy for a given budget of tactile probing points compared to a passive learning strategy.