Abstract: To address the issue that X-ray detection alone is insufficient to fully capture the complex material changes during the flotation clean coal process, an intelligent flotation system based on multi-source heterogeneous information fusion, integrating X-ray detection and machine vision technologies, was designed. A prediction model for clean coal ash content was established using Federated Learning (FED) and Convolutional Neural Networks (CNN). Elemental analysis of flotation clean coal slurry was conducted using an X-ray ash analyzer. 1D-CNN was applied to process the elemental content data to extract temporal features. Meanwhile, the ash content of flotation clean coal was detected using flotation froth vision technology, and 2D-CNN was used to process tailings image information to extract spatial features. An attention mechanism was adopted to fuse the temporal and spatial features derived from multi-source heterogeneous information, and through a fully connected layer to conduct regression prediction of ash content in flotation clean coal. The FED model effectively addressed privacy protection and collaborative modeling in multi-source heterogeneous information fusion through a modular aggregation method and a dynamic weighting strategy. Experimental results showed that the FED-CNN model achieved a maximum error of 4.44% and a coefficient of determination (R²) of 0.94. The prediction accuracy was higher than that of the 2D-CNN model based on tailings images and the 1D-CNN model based on X-ray data.