Machine learning-based development of chromium (Cr) content control process of RH furnace for high-purity iron

In order to accurately control the residual composition content of high-purity iron to meet the standards for FePO4 preparation, this article has investigated influence factors of chromium (Cr) elemental in Ruhrstahl–Heraeus vacuum refining (RH furnace). Furthermore we have developed a process which control Cr element by machine learning (ML) algorithm which may provide new ideas of the control of residual composition in pure iron. For this purpose, we screened the data from more than 300 furnaces of mainly pure iron and low CrNi steels from the plant, containing the main process parameters in the RH furnace. The isolation forest algorithm (IF algorithm) was used for outlier processing, combined with feature engineering methods such as Pearson correlation analysis, recursive feature elimination (RFE) and exhaustive methods to filter the greatest impacting features on Cr content prediction at the numerical level. Furthermore, four regression models, support vector regression (SVR), K nearest neighbour (KNN), random forest (RF) and extreme gradient boosting (XGBoost), were selected for modeling, and the impact of features was evaluated by Shapley additive explanations (SHAP) value analysis. Ultimately, each model achieved a 90% prediction effect for the Cr element, while the RF model was less effective in predicting Cr elements. The best performance of Cr content prediction validated the rationality of the screened characteristics. Subsequently, based on the metallurgical analysis of the characteristics identified through model screening, this article proposes three metallurgically viable methods for controlling the Cr content within the RH furnace.