Concentrate ash content is a core indicator for evaluating the flotation process and regulating operational parameters. Most soft measurement studies focus on mapping ash content using a single data source. It neglects the multi-parameter nature of the process. To address this gap, this work proposes a multi-source heterogeneous data framework integrating froth image features and process parameters. A Deformable Attention Neural Network (DANet) is designed to extract froth image features, where deformable convolutional kernels exhibit an “effective dilation” positively correlated with froth size. After the fully connected layers in DANet-V6 (version 6) were replaced with Bayesian-optimized eXtreme Gradient Boosting (BOXGBoost), the R² increased by 8.31% to 0.851 on industrial data. Through relative marginal contribution analysis, 20 image feature vectors are selected, reducing computational costs by 81.16% with only 0.51% performance drop. The vectors are fused with process parameters to construct multi-source data. The DANetV6-BOXGBoost achieves state-of-the-art performance on the test set: R² = 0.933, RMSE = 0.302, and MAE ± Std = 23.4 ± 1.9%, significantly outperforming single-source data approaches. This work provides a novel framework for multi-source data fusion in mineral processing, its practical implementation will enhance the efficiency and intelligence of coal production.

Coal flotation,Image recognition,Multi-source heterogeneous data,Deformable convolution,Attention mechanism