Development of Neural Network for Improving Forming Accuracy of Axisymmetric Shape in Incremental Forming

抄録

A method using a convolutional neural network (CNN) to predict and correct inaccuracies in incremental forming. By training the CNN with height data of truncated cone shapes, the network can generate more accurate tool paths reducing errors in forming height, wall angle, initial diameter, shear droop height and gap volume. Three models were created with resolutions of 50×50 pixels, 100×100 pixels and 200×200 pixels for training data. The results demonstrated that the CNN significantly improved forming accuracy, with the 100x100 pixel model achieving optimal performance. This model showed the best balance between computational efficiency and detail resolution, achieving a height matching rate of 75.351%, an angle error of 0.852%, an initial diameter accuracy of 91.426%, and reducing gap volume and shear droop height by 39.359% and 23.386% respectively. The findings highlight the potential of CNNs in improving the precision of incremental forming, suggesting further research into complex shapes.