Classification-based process parameter recommendation in sheet metal forming

Abstract

For the production of sheet metal parts for car bodies, continuous adjustment of process parameters is required to maintain the desired part quality in the presence of scattering blank properties. The digital transformation enables data-driven methods for finding process parameters instead of a time-consuming experience-driven trial-and-error approach. However, it is still hard to measure quality for every part due to cost and technical limitations. Removing data points of low-quality parts helps in recommending proper process parameters. In this paper, we propose a classification-based solution enhanced by a particular data preprocessing for recommending cushion cylinder forces. The solution utilizes anomaly detection and knowledge-based methods to remove potential low-quality data points without quality measures in data prepossessing. On the processed data, a classification model is trained to predict process parameters according to blank properties. Our solution detects 32.46% low-quality parts and gives competitive performance (94.30% prediction accuracy) compared to a model trained on data comprising quality measures.