Machine learning models for predicting volumetric errors based on scale and master balls artefact probing data

The volumetric accuracy of machine tools is important to both machine tool manufacturers and users. Predicting volumetric errors (VEs) is a pre-requisite for their compensation yielding increased dimensional quality of machined parts. However, predicting VEs in five-axis machine tools is challenging due to the complexity of error sources and their associated physics-based model. Machine learning (ML) is used to predict VEs under no load and stable thermal conditions. Data is acquired using a scale and master ball artefact (SAMBA) and on-machine touch probing. A general process for determining the minimum number of balls required to generate data to satisfactorily train an ML model is proposed. The VEs prediction is verified using synthetic data for inter-axis and some intra-axis geometric errors, and then validated using only experimental data. Different datasets based on decreasing number of balls are tested to train either a Neural Networks (NN) or an eXtreme Gradient Boosting (XGBoost) algorithm to compare their performances. The results show that, both NN and XGBoost are effective to predict VEs of a five-axis machine tool with wCBXfZY(S)t topology regardless of the geometric error parameter values. By using only experimental data of twenty balls to train the models, XGBoost outperforms NN in all four error metrics and processing time. A time efficient scheme was tested whereby only two master balls plus one scale bar dataset and an additional master ball (when only the spindle rotates) were used for training NN.