 |
International Society of Science and Applied Technologies |
| Towards Real-World CNC Thermal Error Compensation: A Bayesian-Optimized CNN-LSTM Approach and Diagnostic Analysis | ||||
| Author |
Nguyen Khanh Hoa Ngo
|
|||
| Co-Author(s) |
Ming Chang Chih; Meng-Hua Li
|
|||
| Abstract | Thermally induced deformation remains a major
source of inaccuracy in high-precision CNC machining. This paper proposes a practical thermal error prediction framework that combines XGBoost-based temperature-sensitive point selection, a hybrid CNN–LSTM predictor, and Bayesian optimization for automated hyperparameter tuning. The framework is evaluated on nine independent long-duration runs collected from a three-axis CNC machining center under a repeated 11-phase spindle-speed program, with active spindle speeds ranging from 1600 to 7200 rpm and one spindle-stop phase. Run-wise validation and a strict-causal protocol are adopted to avoid optimistic leakage and assess realistic deployability. Results show strong generalization for the Y-axis, achieving a test R! of 0.843, while the Z-axis remained positively predictable under the independent test run. In contrast, the X-axis exhibited limited generalization under temperature and spindlespeed signals, suggesting the need for additional machine-state features for reliable worktable-axis compensation. The study therefore contributes a practical Bayesian-optimized CNN–LSTM framework for CNC thermal error prediction, together with axis-wise diagnostic insight into the limits of thermal compensation based solely on temperature and spindle-speed signals.
|
|||
| Keywords | Thermal error compensation, CNC machine tools, XGBoost, CNN–LSTM, Bayesian optimization | |||
| Article #: RQD2026-142 | ||||
Proceedings of 31st ISSAT International Conference on Reliability & Quality in Design |