Prediction of pellet metallurgical properties based on multimodal data and attention enhancement mechanism

With the widespread application of iron ore pellets in modern steel production, traditional empirical models often struggle to account for the coupling effects between induration process conditions and internal microstructures. Consequently, the accurate prediction of pellet properties has become a critical challenge for enhancing production efficiency and product quality. To address this, this article proposes a pellet properties prediction model based on an attention mechanism. The model integrates roasting process parameters, pelletising raw material characteristics, and microstructural features extracted from industrial computed tomography (CT) images to achieve precise performance forecasting. Initially, image segmentation and feature extraction are performed on industrial CT slice data to obtain characterisation indicators such as porosity, liquid phase, and hematite content. Subsequently, the process data and CT features undergo standardisation and feature fusion. An attention mechanism is then incorporated into the XGBoost model to automatically learn the weighted relationships of different features regarding pellet properties. The model demonstrates high accuracy in predicting metallurgical properties, including compressive strength, RDI +6.3 , and RDI +3.15 , with coefficients of determination R 2 nearly all exceeding 0.9, significantly outperforming several traditional models. Furthermore, through the analysis of attention weights and SHAP feature importance, the dominant roles of key variables – such as porosity and roasting temperature – on performance are revealed. These research findings provide a novel intelligent analytical approach for the optimisation of pellet processes and quality control.