Cross-Section Bead Image Prediction in Laser Keyhole Welding of AISI 1020 Steel Using Deep Learning Architectures

Abstract

A deep learning model was applied for predicting a cross-sectional bead image from laser welding process parameters. The proposed model consists of two successive generators. The first generator produces a weld bead segmentation map from laser intensity and interaction time, which is subsequently translated into an optical microscopic (OM) image by the second generator. Both generators exhibit an encoder-decoder structure based on a convolutional neural network (CNN). In the second generator, a conditional generative adversarial network (cGAN) was additionally employed with multiscale discriminators and residual blocks, considering the size of the OM image. For a training dataset, laser welding experiments with AISI 1020 steel were conducted on a large process window using a 2 KW fiber laser, and a total of 39 process conditions were used for the training. High-resolution OM images were successfully generated, and the predicted bead shapes were reasonably accurate (R-Squared: 89.0% for penetration depth, 93.6% for weld bead area).

Keywords

• Laser welding
• weld-bead prediction
• deep learning
• image-to-image translation