Prediction of transverse thickness difference in medium- and high-carbon steels based on transformer and optimisation of rolling schedules

The cross-sectional profile quality of medium- and high-carbon steels is crucial in high-end manufacturing. Machine learning algorithms are employed to predict transverse thickness differences in cold rolling. Facing variable-length sequences in single-stand mills, the predictive abilities of multi-layer perceptron, random forest, long short-term memory and transformer models are compared, with the transformer model showing the highest accuracy, with an R 2 of 0.9783. Then, machine learning is integrated with particle swarm optimisation, a cold rolling mechanism model and stable rolling requirements to create a rolling schedule optimisation model aiming to minimise transverse thickness differences. A finite-element model of the S6-High cold rolling mill (CRM) is established via ABAQUS to interpret optimisation results. Field applications reveal that this research reduces the average transverse thickness difference in rolling medium- and high-carbon steels from 26.6 to 17.5 μm, greatly improving the control level in the S6-High CRM.