Physics-Informed Dataset Optimization for CNC Axis-Current Prediction

Accurate prediction of axis and spindle current signals is essential for model-based CNC
process monitoring. Although representative datasets are essential, high product diversity, short life cycles, and varying process conditions lead to incrementally collected datasets with imbalanced or shifted distributions. Therefore, this paper proposes Physics-informed Dataset Optimization (PiDO), a data-centric PiML approach that enriches training datasets with physics-consistent samples before model retraining. PiDO uses locally parameterized physical equations (PEs) for axis- and spindle-current dynamics, derived from an energy-balance-based current model, to generate samples in a reduced discretized feature space. Physical consistency is constrained by local PE parameterization, feature-space bounds, NC-code, machinekinematic information, and statistically observed ranges of unresolved variables. The approach is evaluated on 11 POM-C impeller geometries on two three-axis CNC milling machines (DMC 60H and CMX 600V), resulting in 22 validation datasets, using real-data-only training as baseline. The evaluation is based on RMSE, R$^2$, significance tests, and dataset-balance measures. ... mehrCompared with a global PE, the proposed local PE parameterization reduced RMSE by 43.0–74.8% and increased R$^2$ from 0.072–0.448 to 0.821–0.947. In the downstream prediction task, PiDO improved dataset balance and achieved significant positive RMSE effects in several validation scenarios, particularly in sparse data regimes, without significant deterioration compared to the real-data baseline.