Variable speed rolling force prediction with theoretical and data-driven models

The variations in thickness at the head and tail significantly impact the product quality in tandem cold rolling. Improving rolling force calculation accuracy under speed-up and speed-down conditions is essential to enhance thickness control. This paper proposes a sequential coupled model combining data-driven and theoretical models to predict rolling force under speed variation conditions. To enhance accuracy and overcome the limitations of traditional velocity field, a linear cross-section dynamic velocity field taking into account the effects of roll gap changes on the deformation zone is established. Moreover, a long short-term memory (LSTM) network is constructed to capture sequential information from production data and compensate for changes in friction state caused by speed variation and roughness attenuation in the theoretical model. 12 coils, totaling 12,000 samples, are selected from the process data acquisition system (PDA) and utilized as the dataset after normalization. Mutual information (MI) theory is employed for feature selection to improve the prediction accuracy and reduce the computing consumption. Learning rate (LR) adjustment is applied to enhance the stability of initial network and the convergence speed. Compared to evaluated models, the proposed model has relative error within +/- 2.5 %, R2 of it is 0.9935 and the prediction time is only around 2.5 ms. The results suggest that the proposed model demonstrates the most favorable overall performance. Additionally, some key rolling parameters whose physical relationship with rolling force remains uncertain are investigated.