Thermal error prediction for a five-axis machine-tool feed system using error decoupling and data–mechanism integrated modeling

Abstract

Thermal-error prediction models for machine-tool feed systems can benefit from combining mechanistic knowledge with data-driven learning, but the respective roles of the two components are often unclear. This study presents an error-decoupling based data–mechanism integrated modeling method. A first-order Taylor term represents the dominant, smooth temperature rise error (TRE), whereas the remaining time-varying nonlinear residual error component (NREC) is modeled using BiLSTM ResXGBoost residual learning. The two predicted components are reconstructed additively to obtain the total thermal-error prediction. In the reported Y-axis experiment on a five-axis machining center, the reconstructed model achieved an RMSE of 0.066 μm under the tested operating condition. The results demonstrate the feasibility of explicitly allocating the smooth thermal trend and nonlinear residual component to the mechanism-informed and data-driven models, respectively.