Recognition of Coal-Slurry Flotation Working Conditions by Fusing Froth Images and Tailings Ash Content Data

Manual judgment in coal-slurry flotation suffers from dynamic hysteresis and sensitivity to illumination and dust. In contrast, froth images characterize macroscopic flotation appearances, while tailings ash content reflects the internal pulp quality, and the two sources provide complementary information. This paper proposes a flotation condition recognition method based on the fusion of froth images and tailings ash content data. Visualization via violin plots shows the high correlation between ash content and flotation states. In this framework, froth images are designated as primary inputs and ash content as auxiliary signals, with an optimal fusion-weight ratio determined through ablation experiments. Multimodal models based on ResNet18, MobileNetV2, and MobileNetV3 are compared, and image-only and tailings-ash-only models are used to assess the value of multimodal fusion. Experiments demonstrate that Multimodal MobileNetV3 achieves an accuracy of 94.65%, outperforming Multimodal MobileNetV2 at 87.98%, Multimodal ResNet18 at 86.13%, the image-only model at 86.43%, and the tailings-ash-only model at 81.63%. t-SNE visualization indicates that multimodal fusion enhances feature discrimination and clustering. This method effectively improves classification accuracy and provides a new approach for intelligent working condition recognition in coal preparation plants.