XGBoost-Based Digital Twin Model for Predicting Trajectory Errors in a Hexapod Coordinated Machining System Using Positioning Accuracy and Vibration Data

Dynamic errors in robotic machining can degrade part quality, particularly in flexible platforms that are susceptible to both geometric and inertial disturbances. This work introduces a data-driven digital twin for pointwise prediction of circular trajectory errors in a hexapod-based machining cell, using a compact sensing configuration that combines ballbar measurements with tri-axial vibration signals. Deviations measured by ballbar, acceleration data, and CMM-measured profiles are synchronized in the angular domain via a unified pipeline for denoising, resampling, and phase alignment. Sliding-window vibration statistics and the ballbar path error are used as inputs to XGBoost, multilayer perceptron, and random forest regressors. Model performance is evaluated under a deployment-relevant leave-one-run-out protocol and a conventional random 70:30 point split. XGBoost achieves micrometer-level accuracy on unseen runs, with RMSE around 5 µm, R² exceeding 0.80, and near-complete coverage within a ±20 µm tolerance band. Compared to baseline models, it also provides improved suppression of extreme residuals. Feature importance and ablation studies show that the ballbar path error captures the dominant geometric component, while compact hybrid feature sets-combining this anchor with selected vibration descriptors-retain most of the predictive accuracy and enable practical offline batch-level compensation.