Froth-tailings fusion image-based concentrate ash content prediction in coal flotation using stacking model

Concentrate ash soft measurement based on coal flotation process images is a promising approach. However, using a single model to predict ash content based solely on froth or tailings images often results in poor accuracy and instability. To address these limitations, a soft-sensing scheme for predicting concentrate ash in coal flotation is proposed, utilizing froth-tailings fusion images and a stacking model. This stacking model combines knowledge from multiple convolutional neural networks (CNNs). The froth and tailings image dataset was divided into six intervals, and pixel-level fusion was performed using the Laplacian pyramid fusion algorithm. Three CNNs (Alexnet, Resnet34, VGG16) and four stacking models were trained on the fusion images. The ablation study shows that the stacking model with Alexnet and VGG16 (called Va-stacking) as base learners and the Random Forest model as meta-learner achieves the optimal prediction accuracy of 93.63%. Feature engineering, a Froth-stacking model, a Tailings-stacking model, and a Vision Transformer (VIT) were explored and compared with Va-stacking. The results indicate that the froth-tailings fusion images outperform single froth/tailings images and feature engineering. In the image-based coal ash prediction task, the ensemble learning framework surpasses the sequence-based VIT.