Density-Based Dataset Balancing for Incremental Retraining in CNC Axis and Spindle Current Prediction

Accurate prediction of the axis and main spindle current is essential for reliable model-based process monitoring in Computer Numerical Control (CNC) machining. However, modern manufacturing is characterized by a high level of product variety, short life cycles, and frequently changing process conditions. This results in unbalanced and constantly changing data distributions, which present a challenge to model training. Traditional data collection strategies generate large and redundant datasets that require substantial computational and storage resources. This work introduces a density-based approach to dataset balancing for incremental retraining that enables efficient and robust model training under evolving operating conditions. The proposed approach operates at the data level via a density-controlled replay memory, rather than modifying model parameters. A reduced, interpretable feature representation is derived to represent the underlying data structure, providing a basis for systematic discretization. Based on this representation, the adaptive memory mechanism prioritizes samples according to their spatial density, selectively removing redundant information, and ensuring adequate coverage of regions critical for model performance. The proposed method has been evaluated across multiple machines, machining parameters, and geometries. The results show that the balancing strategy yields significantly better results than static training while reducing the required data volume by over 40% compared to incremental data expansion. Consequently, it facilitates the development of data-efficient, scalable and robust training strategies for intelligent machining systems, while also supporting improved adaptability in the monitoring of flexible production environments.