multi−source information prediction of coal slime flotation index based on feature level fusion and XGBoost

Fine coal flotation is a critical operation in coal preparation, and the accurate prediction of flotation performance indices is essential for process parameter optimization, product quality stabilization, and production efficiency improvement. However, the flotation process is characterized by multivariable coupling, strong nonlinearity, and insufficient state representation, making it difficult for a single information source to comprehensively describe variations in flotation conditions. To address these issues, a multisource information−based prediction method integrating feature−level fusion and extreme gradient boosting (XGBoost) was proposed. First, flotation froth images were acquired using an industrial camera and subjected to grayscale conversion, noise reduction, and watershed segmentation. Structural features, including the mean bubble diameter, standard deviation of bubble diameter, bubble count, mean bubble area, and circularity, were extracted. Texture features, including contrast, energy, entropy, and homogeneity, were subsequently calculated from the gray−level co−occurrence matrix. The extracted visual features were then integrated with six key operating variables, namely impeller speed, air flow rate, feed flow rate, feed concentration, collector dosage, and frother dosage, through feature−level concatenation. A unified multisource feature vector was thereby constructed. An XGBoost−based model was subsequently developed to characterize the nonlinear relationships between the multisource information and the flotation performance indices. Furthermore, low−contribution variables were eliminated according to feature importance, and the fused feature space was reconstructed to reduce information redundancy and improve model generalization. The experimental results showed that, after feature selection, the coefficients of determination (R²) for the prediction of flotation yield, clean coal ash content, and combustible recovery reached 0.631, 0.562, and 0.956, respectively, while the corresponding root mean square errors (RMSEs) were 1.09, 0.86, and 1.26. The highest prediction performance was obtained for combustible recovery. Feature importance analysis indicated that collector dosage, air flow rate, mean bubble diameter, and froth circularity were the principal variables affecting the prediction results. The proposed method enables the complementary representation of froth morphology and operating conditions and provides a practical modeling approach for online state perception, soft−sensor development, and process optimization in fine coal flotation.