Leveraging Machine Learning Insights at the Hot Strip Rolling Mill, Tata Steel, Port Talbot

The Hot Strip Mill at Tata Steel, Port Talbot is a long-established, heavily instrumented, steel strip rolling mill. The mechanical properties of the finished strip depend on the cooling of the strip as it passes along the run-out table, decided before the coil-run from a physics-based simulation. Working with the MATLAB machine learning toolbox, this investigation details the process of developing an accurate (best final Pearson Coefficient of 0.8780) data-driven prediction of the two-phase cooling to final Coiling Temperature, independent of the physical model. Though ultimately trained on practical events resulting from the physics-driven control system, data-driven insights are isolated from the direct assumptions made within that existing simulation – thus affirming or providing insights into those established assumptions. The insights gained inform new visualizations of the complex datasets, for clearer illustration to industrial partners and process experts and highlight where further instrumentation would be most usefully invested. The finding that the cooling profile can be approximated as a two-phase linear process has implications for the assumptions made about the completeness and location of the metallurgical transition point in the current physical model. In addition to the parameters decided by the process model, it is the gauge of the strip that is the most important intrinsic factor affecting coiling temperature. Labelling by gauge highlights coils for which CT is over and under predicted, to aid further root cause analysis and the most efficient deployment of new instrumentation.