Convolutional neural networks for prediction of geometrical errors in incremental sheet metal forming

Abstract

Die-less single point incremental forming (SPIF) is a highly versatile process that can form sheet metal blanks into desired geometries. A significant drawback of this process is the poor geometric accuracy largely attributed to material spring-back. This paper presents the use of Convolutional Neural Networks-Forming Prediction (CNN-FP) to predict this error, allowing users to better understand the expected distribution of geometric error prior to actual forming. The reported work differs from other published approaches in that a CNN was used as an automatic and flexible method for quantifying local geometries. The CNN-FP model was trained using a set of SPIF geometries with varying wall angles and corner radii. The performance of the trained model was validated using two SPIF geometries: one consisting of untrained wall angles and the other combining various features to create a complex geometry. The CNN-FP model achieved an RMSE (Root mean squared error) of 0.381 mm at 50 mm depth for the untrained wall angle. For the untrained complex geometry, the CNN-FP performance was found to be 0.391 mm at 30 mm depth. However, a significant deterioration was observed at 50 mm depth of the complex geometry, where the model’s prediction had an RMSE of 0.903 mm. While the model was shown to be efficacious in most of the validation tests, limitations in the breadth of training samples was attributed to the model’s degraded performance in some instances.