关键词: 梁类零件, 冲压回弹, 深度学习, 拉丁超立方采样, 高斯混合聚类

Abstract: A method was proposed herein for springback prediction of automobile beam parts based on deep learning. Based on the two-dimensional springback theory, the three-dimensional beam parts were discretized into several sections, and the cross-section curve of the beam parts was converted into a dual-channel image that could be recognized by neural network model using the method of double-plane projection and the image binarization method. Based on Latin hypercube sampling method, the stamping process parameters and sheet material parameters of the beam structure parts were sampled, and the subsequent deep learning network training samples were obtained by CAE springback simulation. In order to study the springback of beam structures under different geometric cross-sections, material parameters, and processing parameters, convolutional neural network models based on LeNet-5, AlexNet and NiN were used as geometric cross-section recognition models. Meanwhile, the fully connected neural network model was used to couple material parameters and processing parameters to obtain the springback algorithm model of beam parts. Car beam structural parts were taken as the research object. Based on Gaussian mixture clustering, the springback samples were divided into three types: small springback, medium springback, and large springback. Each type of springback sample was verified by the springback algorithm model. Verification results show that the deep learning model has the highest accuracy based on AlexNet, and the algorithms robustness is also stronger than that of the other two, which is more suitable for springback prediction of beam parts.

Key words: beam part, stamping springback, deep learning, Latin hypercube sampling, Gaussian mixture clustering