Work

Uncertainty-aware quantification of the predictive-information gap between control-side signals and post-stroke thermography in industrial deep drawing

Year: 2026
Type: article
Abstract: In series deep drawing, the realised component geometry can vary from cycle to cycle while nominal machine settings remain unchanged. This study investigates how the predictive information contained in routinely recorded control-side signals compares with that contained in post-stroke infrared observations. The analysis uses 46,048 production cycles from an industrial automotive deep-drawing line. Seven uncertainty-aware regression approaches were first compared to qualify a common predictive reference with respect to point accuracy, empirical interval coverage, interval width, and computational effort. Controlled group-wise ablation then showed that machine-state, blank-holder-force, and setpoint variables alone achieved R 2 = 0.544 for cycle-wise flange-length variation, whereas 24 absolute infrared descriptors achieved R 2 = 0.842 at near-nominal prediction-interval coverage. The infrared contribution was spatially non-uniform, with Area-Of-Interest (AOI) 1, 7, 2, and 8 showing the largest performance losses when removed. These results show that, within the observed production regime, routine control-side records contain substantially less predictive information about cycle-wise geometric variation than the post-stroke thermal observation. The main contribution is an industrial-scale quantification of the predictive-information gap between routinely recorded control-side signals and post-stroke thermal observations with respect to the same cycle-wise geometric response.
Language: English
Open Access status: hybrid
Estimated APC paid: $3,520
OpenAlex ID: W7234919633
FWCI:
Cites: 25
Cited by:
Related to: 10