Neural-Net–Based Predictive Modeling of Spout Eye Size in Steelmaking

Gas stirring is commonly used in pyrometallurgical vessels to enhance mass and heat transfer and to promote impurity removal. In the case of secondary steelmaking, the spout eye area is caused by the escape of the gas from the top of the smelt where the liquid steel is directly exposed to the air, and oxygen can be picked up through the spout eye area that can reduce the quality of steel. Thus, controlling the size of the spout eye area is very important to improving the quality of the steel and to keeping the consistency of the product. The set of prevailing models to predict spout eye size are based on specific practically difficult variables, e.g., height of slag in hot upper layer of vessels and gas flow rate at nozzle exit. Recently, the cold model results showed that the stirring process can be conveniently monitored by the signals such as (1) the image signal from the top of the vessel, (2) the sound of the stirring process, and (3) the vibration on the wall of the vessel. This article outlines the key details of a novel research investigation using neural-network–based predictive modeling such as general regression neural networks (GRNN) with genetic adaptive calibrations. Predictive capacities and generalization potentials of five model constructs (i.e., with different sets of input parameters) were explored, and the neural net modeling yielded encouraging outcomes, e.g., (1) excellent goodness-of-fit generalization measures including high values of correlation and R2 validation parameters (e.g., r = 0.921 and R2 = 0.845 in a model validation), and (2) low values of root mean square of errors (e.g., 3.034). Overall, the research outlined in this article demonstrates that the spout eye size can be effectively predicted by predictive neural net modeling with convenient and practically measurable variables such as sound and vibration observations on the steelmaking vessels. These results have only been demonstrated for a cold model of the process, and further work is required to show that this approach can be extended to industrial operations.