Application of Recurrent Neural Networks in Uncertainty Analysis of Sheet Metal Forming

The quality of deep-drawn sheet metal components can be strongly influenced by different sources of uncertainty, such as variations in process conditions, deviations in tool geometry, and variations in material properties between coils. Identifying the underlying causes of forming defects remains a challenging and time-consuming task due to the complexity of the forming process. This study presents a machine learning (ML) framework for tracing sources of uncertainty in the forming of a cylindrical cup. By analysing key outputs from a standardized cylindrical cup test, including force-displacement sequences, earing evolution, and thickness distribution in different sections of the cup, the ML model aims to predict multiple sources of uncertainty. After identifying the principal sources of variation, numerical simulations using finite element analysis were performed to create a comprehensive dataset for the development and training of the ML model. The results from the simulated tests can be considered as sequential data, allowing their evaluation using recurrent neural networks (RNNs), which are particularly suited for modelling temporal or ordered datasets. The RNNs demonstrated strong performance in leveraging the temporal features of the forming data, achieving high accuracy in identifying the origins of uncertainty.