What question did this study set out to answer?

This study aims to improve the accuracy of machine learning predictions for mechanical properties in sheet metal forming by evaluating feature selection.

June 3, 2026Open Access

Feature Engineering and Data Scaling for Spatially Resolved Post-Springback Mechanical State Predictions in Sheet Metal Forming

Key Points

This study aims to improve the accuracy of machine learning predictions for mechanical properties in sheet metal forming by evaluating feature selection.
Evaluated spatially resolved predictions of residual thickness, strain, and stress using an input-output dataset from a validated simulation model.
Compared machine learning features derived from force-displacement curves with material and process parameters.
Assessed learning curves to examine prediction accuracy as training set size increases.
Baseline descriptors were sufficient for effective plastic strain predictions.
Adding force-displacement curve features achieved modest reductions in prediction error for post-springback von Mises stress and maximum principal strain in hot spots.
Learning curves showed a stepwise reduction in prediction error with larger training budgets, particularly in stress state targets.

Abstract

Abstract Data-driven sheet metal forming relies on fast, reliable predictions of quality indicators to avoid tolerance violations in downstream assembly. This study evaluates spatially resolved machine learning predictions of residual thickness, strain and stress for an S -rail forming process. An input–output dataset generated with a validated simulation model is used to compare features derived from the force–displacement curve (FDC) with material and process parameters. Results are target dependent, and baseline descriptors are sufficient for effective plastic strain, whereas adding FDC features yields modest but consistent error reductions for post-springback von Mises stress and maximum principal strain, also in critical hot spots. Learning curves assess how accuracy scales with training set size. Across targets, they show a steplike reduction in prediction error at higher training budgets, while stress state targets exhibit a smaller decrease. Overall, practical guidance is provided on feature selection and minimum sample sizes for robust, data-driven sheet metal process chains.

AI에게 질문

Bookmark

View Full Paper