An On-Line Hierarchical Decomposition Based Bayesian Model for Quality Prediction during Hot Strip Rolling

Abstract

Mathematical models have been widely used for the prediction of the microstructure and mechanical properties in hot rolling of strip. However, their accuracy is insufficient for quality control purposes. To accurately predict these characteristics, it is necessary to create models which can replicate the thermomechanical state of the material and its evolution during processing. In addition, these models should be able to capture the uncertainties introduced in the processing by the dynamics of the casting process and the subsequent rolling. These uncertainties lead to considerable variations in the material and mechanical state of the rolled strip. This paper presents the development of a hybrid model (MICROL) which uses the mills setting and the real time plant data such as chemical composition; forces and temperatures; and integrates them a Bayesian format to predict the desired quality attributes as well as microstructural features. This information is combined into Bayesian Hierarchical models to create an on-line tool that predicts the properties of each individual rolled coil, as well as provide information on the batch-to-batch and heat-to-heat variations. Case study from a steel Plant is presented which illustrates the implementation, calibration and validation of this model across different materials grades. Model results are found to be within the 5% tolerance of the measured values for many steel grades and rolling conditions.