Data-driven robust design for industrial cylindrical plunge grinding of piston rings: Requirements for reliable digital-twin development under production noise

Cylindrical plunge grinding of piston rings is a precision finishing operation, and its dimensional accuracy governs engine sealing and emission compliance. It is widely assumed that such accuracy can be controlled through fixed machine settings; this study tests this assumption under realistic production noise. An industrial dataset of repeated dimensional measurements collected across 30 experimental runs with 2 recorded noise sources (mandrel type and ring position) was analysed. Three predictive models were evaluated using a grouped validation scheme that kept all repeats of a condition together, avoiding the optimistic accuracy that ordinary random splitting produces. All models predicted the held-out conditions no better than a simple average, showing that unmeasured disturbances rather than adjustable settings dominate variability. Therefore, a model-free robust design framework was developed, ranking the conditions directly based on their observed performance and scatter without a fitted model. The recommended shallow-dressing setting achieved a capable performance on two of the three dimensions, while the third indicated an irreducible variability floor associated with unmeasured disturbances. This study contributes by examining the parameter-only predictive performance under a leakage-free validation protocol and by outlining the in-process sensing layer that a reliable grinding digital twin would require.