PR-DT1121-1 - 2021 Digital Transformations, November

Improved Prediction of Steel Hardness Through Neural Network Regression

C. Elkin, R. Bathla, T. Poplawski, S. Agashe, and V. Devabhaktuni

Download PDF

View Sample Page

Abstract:
Product hardness number is a vital input for finishing mill setup models in the calculation of load, torque and power in a hot strip mill. It is based on steel grade, whose value is updated from mill feedback. The goal of this work is to improve hardness prediction even if the steel grade is new to the hot mill. To do so, a robust neural network–based hardness prediction algorithm is proposed that inputs statistically significant features such as chemistry composition, aim gauge, aim width, rolling temperature and finishing temperature. Ultimately, this approach achieves mean absolute error below 2.1%, which substantially outperforms the baseline error of 2.75%.


Keywords: digital transformation, industry 4.0