Abstract: To address the slow detection of ash content in flotation clean coal and its inability to meet the requirements for rapid process regulation, an intelligent detection method integrating X−ray fluorescence (XRF) spectral information with a residual−driven, feature−weighted, regularized stochastic configuration network (RF−RSCN) is proposed. First, considering the high dimensionality, strong collinearity, and pronounced noise interference of online XRF spectral data, partial least squares (PLS) is employed for supervised dimensionality reduction to extract latent features closely associated with ash content. On this basis, a feature−weighting mechanism, a residual−driven node configuration strategy, and a dynamic regularization scheme are incorporated to enhance the model’s capability in capturing complex nonlinear relationships and improving generalization performance. The proposed method is validated using 924 samples collected from an online flotation clean coal system. Under a time−ordered training-testing split, the RF−RSCN model achieves an RMSE of 0.0134, an MAE of 0.0102, and an R² of 0.9965 on the test set, outperforming SCN, BSCN, IRVFLN, SVR, HDNN, PLSR, and BP models. In addition, the average testing time for a single sample is 3.5 × 10⁻⁵ s, indicating high computational efficiency. These results demonstrate that the proposed method can effectively establish the mapping relationship between XRF spectral information and clean coal ash signals, providing a reliable technical approach for rapid ash detection and process optimization in flotation systems.