The springback is the main factor affecting the forming quality of hull plates in the cold forming process. To improve the forming quality, it is necessary to investigate springback prediction, obtain the appropriate springback control method and further guide the die design.

A fully convolutional network (FCN) is used to perform pixel-level calculations and regression calculation on the springback image so as to achieve springback prediction for each forming position on the sheet. In this study, a finite element (FE) model is established using ABAQUS 2019, and the numerical results are validated by the experimental results. The verified model is then applied to obtain the training sample set. The workpiece geometric information is used as the input of the neural network to retain all the information of the image, and the TensorFlow Core V2.2.0 platform is used to build the FCN based on different convolutional neural network models. Finally, the pros and cons of different neural networks are compared, and the optimal network is applied to the die design.

The results show that the maximum error of the predicted springback is 8.49%, where the constructed FCN32 has the highest accuracy. The proposed model can also realize one-time mould design with a calculation time of only 0.5 seconds and a maximum error of only 1.00%, significantly improving calculation efficiency.

The FCN-based algorithm proposed herein provides a springback prediction method for strips with high accuracy and efficiency, as well as offering a new approach to quick mould design.