Comparison of Two Machine Learning Models for Predicting Volumetric Errors From On-The-Fly R-Test Type Device Data and Virtual End Point Constraints

On-the-fly virtual end-point constraints consists in moving all five axes of the machine tool while nominally maintaining the coincidence of a sensing head centre point with a master ball centre attached to the workpiece table. The sensing head detects the deviations from the nominal coincidence as a 3D volumetric error vector. More than one ball can be so measured, and a fixed length ball bar is also measured for detecting isotropic scaling effects. Initial processing of data using the SAMBA (scale and master ball artefact) method eliminates setup errors and provides estimates of inter- and intra-axis errors as well as volumetric error vectors. Two ML models are trained and compared, Neural Network (NN) and eXtreme Gradient Boosting (XGBoost), to find the most suitable model and the required amount of training data to predict volumetric errors of a five-axis machine tool with wCBXfZY(S)t topology based on axis commands. The results show that NN marginally outperforms XGBoost and a kinematic model with ratios of prediction error over volumetric error norms of 0.12, 0.13 and 0.14, respectively.