Optimising predictive accuracy in sheet metal stamping with advanced machine learning: A LightGBM and neural network ensemble approach

Abstract

This article presents an innovative ensemble model that integrates advanced machine learning techniques to enhance the precision of sheet metal stamping processes. By combining a light gradient boosting machine (LightGBM) with deep neural networks (DNNs), the model achieves high accuracy in predicting the final geometry of stamped sheet metal parts, and proactively identifies potential deviations to guarantee strict compliance to geometrical tolerances. In a comprehensive evaluation based on diverse performance metrics, the ensemble model demonstrates substantial improvements over the individual models, achieving a high coefficient of determination R² of 0.951. Significantly, an extensive dataset derived from finite element method simulations is found to facilitate the training of our models in a variety of stamping scenarios, giving superior generalisability and reliability in terms of predictions. In addition, the integration of the ensemble model into an interactive web platform for real-time predictive analytics underscores its practical application in manufacturing settings, as it can optimise decision-making and operational efficiency. The predictive power of the ensemble model and its integration into a real-time framework provide a solid foundation for further advancements in developing a digital twin of the sheet metal stamping process. Our findings highlight the transformative potential of combining diverse machine learning techniques to revolutionise manufacturing processes, thus ensuring higher quality, adaptability, and cost efficiency.