Data-driven classification of the chemical composition of calcine in a ferronickel furnace oven using machine learning techniques

Calcines' chemical composition analysis is a key process in ferronickel smelting. These values allow for a clear understanding of the smelted product's expected quality, catering for any required chemical upgrading of the raw material or modification in the furnace's set-point if the calcine has undesired characteristics. Offline tests for calcines' chemical composition can take several days, potentially delaying the whole operation. A data-driven approach to chemical composition classification using on-line data is proposed by combining clustering classification through a mixed Principal Component Analysis (PCA) model, data processing and standardization process, with a Machine Learning classification algorithm, i.e. Extreme Gradient Boosting (XGBoost). This allows for an online prediction of calcines' chemical composition based on the furnace's current operating conditions. The proposed method's accuracy scored mean values between 82.1% and 85.9%, which is encouraging in comparison with other proposed methods. •Principal Component Analysis (PCA) clustering method is used to reduce the problem's high dimensionality.•Chemical composition classification is proposed by combining data preprocessing, PCA clustering and data-driven models.•Using industrial production process data bridges the gap between academic techniques and industrial applications.•Knowing the on-line calcine chemical composition brings different operational insights, improving the decision making process.•XGBoost was chosen for its iterative self correcting process, capacity to avoid over and under-fitting, among other remarks.