Intelligent real-time ash content detection for coal flotation concentrate using multi-source data fusion

Accurate and real-time online detection of ash content in flotation concentrate is paramount for achieving intelligent optimization and closed-loop control of the flotation process. This capability is critical because coal remains one of the most vital mineral resources in the global energy sector, and ash content is a fundamental indicator of coal quality, directly governing the product quality of fine coal flotation and the overall economic efficiency of coal preparation operations. To address the limitations of delayed system response and inefficient multi-source data fusion in existing methods, this paper proposes an intelligent detection approach for ash content in flotation concentrate based on multi-source information fusion. The core framework integrates heterogeneous data sources, including X-ray fluorescence (XRF) spectra, key process parameters, and tailings image features. First, spectral feature selection and dimensionality reduction are performed using the Successive Projections Algorithm combined with Multiple Linear Regression (SPA-MLR), effectively reducing data dimensionality while preserving critical information. Then, a differentiable Soft Dynamic Time Warping (Soft-DTW) algorithm is employed to align asynchronous time series, enhancing temporal consistency across data sources. Finally, an Interpretable Configuration Algorithm with Response-Weight Mechanism (ICA-RW) strategy is proposed, jointly driven by Node Response Change Value (NRCV) and Output Weight Norm (OWN). This strategy enables dynamic pruning and recovery of hidden nodes, allowing the single-hidden-layer neural network to be adaptively compressed and restructured for improved generalization and model compactness. Field experiments conducted in an industrial coal preparation plant demonstrate that the proposed method significantly outperforms conventional models in both prediction accuracy and robustness. The system reliably meets the stringent process requirement of maintaining ash absolute error within ±0.3 %, providing a practical and effective solution for intelligent closed-loop control in flotation operations and contributing significantly to data-driven smart mining development.