Predicting edge cracking in sheet metal forming: evaluating machine learning models and data transformations

Abstract

This work evaluates the performance of machine learning algorithms in predicting the strain values at which edge cracking occurs in sheet metal forming. Four regression models—Extreme Gradient Boosting, multilayer perceptron, support vector regression, and Gaussian processes—were tested, alongside two ensemble methods: majority voting and stacking. The models were trained and tested using a dataset of mechanical properties from 163 rolled steel sheet samples, derived from hole expansion and uniaxial tensile tests. The tensile test provided yield stress, tensile strength, deformation at maximum load, and elongation after fracture in both rolling and transverse directions, while the hole expansion test measured the deformation at which edge cracking occurs. The models were evaluated based on four metrics: root mean square relative error, maximum absolute error, mean absolute error, and R-squared value. Additionally, the impact of data transformations, including standardization, Box-Cox transformation, min-max normalization, and L2 normalization, was analyzed. The results demonstrate that Extreme Gradient Boosting and ensemble methods provide the most robust predictions, with significant performance improvements observed when data transformations are applied.