Deep learning-based prediction of the mechanical properties of cold-rolled strip steel has generally overlooked the issue of data imbalance that exists in real-world production environments. Moreover, the production process of cold-rolled steel is highly nonlinear, with its mechanical properties influenced by multi-scale couplings between chemical compositions and process parameters. Existing studies have insufficiently explored the complex interactions between these factors and their impact on mechanical behavior. To address this issue, this paper proposes a composite loss function that integrates effective label density and per-sample error, aiming to mitigate the effects of data imbalance in steel rolling tasks. Additionally, a Multi-scale Gated Attention-Driven Network (MGAD-Net) based on the composite loss is introduced for predicting the mechanical properties of cold-rolled strip steel. Dilated convolutions are employed for hierarchical extraction of heterogeneous process-composition features. A sparse parallel self-attention mechanism is also applied to precisely model the cross-process interaction effects in the rolling procedure, enabling comprehensive extraction of complex feature relationships. Compared with existing networks, the proposed MGAD-Net achieves prediction accuracies of 99.45%, 97.89%, and 99.02% for yield strength, tensile strength, and elongation, respectively, with corresponding improvements in R 2 scores of 1.46%, 1.31%, and 3.82%.
Zhang et al. (Wed,) studied this question.