Detection of ash content in clean coal based on deep feature fusion and residual distribution modeling

Near-real-time determination of clean coal ash is essential for quality stabilization and optimal control of flotation processes. To address delayed laboratory feedback, pronounced fluctuations in operating conditions, and the limited adaptability of conventional models to complex nonlinearities, this study proposes an X-ray fluorescence (XRF)-based method integrating feature selection, deep feature fusion, and uncertainty quantification. SHAP is used to identify key input variables, while a cascaded ResTCN–BiLSTM–multi-head attention network captures local feature combinations, bidirectional dependencies, and salient information. Adaptive-bandwidth kernel density estimation (ABKDE) is further employed to model the residual distribution and quantify prediction uncertainty. Validation on 352 industrial samples yielded a test RMSE of 0.4910, R2 of 0.9150, and MAPE of 3.7136%. During continuous operation at a coal preparation plant, the system achieved a mean absolute error of 0.15, with more than 96% of predictions exhibiting errors no greater than 0.30. These results demonstrate high accuracy, stability, and engineering applicability, supporting near-real-time monitoring and intelligent control of flotation processes.