Abstract
To address the bottleneck challenge of achieving transverse thickness difference (TTD) shape control in wide electrical steel for the tandem cold rolling mills, the TTD control strategy of electrical steel integrated data-driven (energy valley optimisation algorithm, convolutional neural network, multi-output long short-term memory, and attention mechanism [EVO-CNN-MLSTM-AM]) and a theoretical model is proposed in the hot-cold rolling process. A mathematical model for the hot-rolled section profile is established through rolling theory. The dataset for the hot-cold rolling process is constructed by combining the hot-rolled and cold-rolled parameters. In the EVO-CNN-MLSTM-AM model, CNN is employed to perform convolutional operations on the data and extract its feature components. These extracted feature components are then sequentially predicted using the multiple-output LSTM. The AM module automatically assigns weights to the hidden layer output vectors of the MLSTM at each time step, enabling multi-output prediction of edge drop, crown, and TTD. Multiple evaluation metrics are used to compare the SVR, LSTM, CNN-LSTM model, and CNN-MLSTM-AM model, demonstrating that the proposed model achieves higher prediction accuracies of 96.74% for the TTD and has better adaptability. Based on this, the influence of rolling parameters on the cold-rolled section profile is revealed using the Shapley interpretability method and association rules. An optimised TTD control strategy is proposed in the hot-cold whole rolling process, achieving fusion control of the cross-section shape in different rolling processes. Application of the proposed model in an industrial HSM-TCM production mill with millions of tons demonstrates a significant improvement, with the proportion of the TTD ≤ 7 μm increasing from 48.36% to 67.74%, in the shape quality for cold-rolled electrical steel.
Keywords
Get full access to this article
View all access options for this article.