Process optimization of chain-die forming for asymmetric channels by an image-based machine learning method

Abstract

The coupled springback of twist and angle is a major shape error for high-strength asymmetric channels in the automotive industry. In the process design stage of sheet metal forming, scalar-based surrogate models such as artificial neural networks, are the common optimization method to predict springback and reduce shape error. However, such models fail to learn and utilize the location information of the formed part, which will result in poor prediction accuracy and compensation effect for the coupled springback in asymmetric channels. Therefore, in this paper, based on the U-net convolutional neural network (CNN), an image-based surrogate model is established to predict the dimensional deviation full field of the formed part. An attention-based ResNet layer is proposed and replaced with the bottleneck of the U-net-style base network to enhance the sensitivity of the network for different regions of die surface in the sheet metal forming. An automatic data preprocessing method is developed to convert the design parameters and simulation results into unified image data. After training, the proposed CNN surrogate model shows a highly accurate prediction for the dimensional deviation full field in real-time. Furthermore, an image-based optimization architecture, in conjunction with a differential evolution algorithm with self-adaptive factors, the automatic data preprocessing method, and the CNN surrogate model, is developed to minimize the dimensional deviations over the entire region of the formed part. Finally, the validation experiments of chain-die forming are carried out by using optimized design parameters and the forming accuracy has been significantly improved.