An AttnGAN-based Macro Image Synthesis Model for Predicting the Internal Structure of Billets

The internal structure of casting billets directly affects the performance of steel products, but in practice, it can not be detected online. Macro images (low-magnification snapshots of the casting billet sample cross-sections) can visualize the structure, but obtaining them is lagging and costly. Numerical simulation methods are time-consuming to calculate, and only corrected for a few steel grades, making accuracy difficult to guarantee under other working conditions. Thus, we propose a model based on generative adversarial networks (GANs) to quickly predict the internal structure of billets, which can synthesize macro images from production parameters. However, the macro images are less visually varied and have small and dense features. Conventional GANs struggle to generate macro images of casting billets. Therefore, this model is based on the attentional GAN (AttnGAN) to learn different level features through the supervision of multi-resolution images. The generators apply three-branch residual blocks to predict new pixels by multi-scale information, which improves the realism of details. Additionally, considering the ordering property of continuous casting, a series model of the deep neural network and recurrent neural network is designed to encode input parameters. It is combined with image information pretraining to simplify the learning of the mapping relationship in the macro image synthesis model. After comparison with other text-to-image generation models, our model performs better in terms of evaluation metrics and visual effects. The prediction results are verified using two samples from SCM435-M steel. The obtained results have a high consistency with the real data, with the absolute error of the equiaxed crystal rate being 0.84% and 0.6%, respectively. Moreover, the inference speed of our model is fast, which is of reference value for the optimization of the continuous casting process.