Study on slurry mass concentration prediction method driven by multi-source data fusion

In coal preparation processes, slurry mass concentration is a critical parameter affecting key operations such as dense medium separation, flotation, and thickening/filtration. Variations in slurry mass concentration directly impact equipment efficiency and product quality, making accurate prediction vital for intelligent coal preparation and process optimization control. However, traditional offline detection methods are constrained by lengthy measurement cycles and sluggish response times, which fail to satisfy the requirements of real-time monitoring. Existing online detection techniques primarily measure slurry volumetric concentration but exhibit reduced accuracy owing to variations in particle true density and physical characteristics, leading to inadequate robustness. Therefore, integrating multiple detection modalities for online prediction of slurry mass concentration has become an effective approach to enhance measurement accuracy and prediction stability. This article proposes a multivariate-fusion, data-driven method for slurry mass concentration prediction based on optical intensity (5/50/200 mW) and differential pressure. An Adaptive Regularization and Initialization-based Incremental Constructive Algorithm (ARI-ICA) model with improved nonlinear feature mapping is employed, in which an adaptive node-generation mechanism and gradually decaying regularized regression are introduced to predict the slurry volumetric concentration and true particle density from the optical-intensity and differential-pressure signals, after which the slurry mass concentration is computed via the governing physical relation. Parameter ranges settings: residual-driven weight sampling interval [0.05E, 1.0E] (best); regularization-decay grid [0.90, 0.95, 0.99] (best 0.99). The study comprises >160 samples (70/30 train/test). This fusion not only enhances feature representation and robustness against sensor noise but also ensures physical consistency through the intrinsic correlations between optical absorption, flow resistance, and slurry density. The model leverages adaptive feature node generation and regularized regression techniques to capture nonlinear input-output relationships under limited sample conditions, thereby mitigating overfitting risk while preserving model sparsity and interpretability. In this study, the proposed method achieves a root mean square error (RMSE) of 0.121, a coefficient of determination (R 2 ) of 0.930, and a mean absolute error (MAE) of 0.080 for slurry mass concentration prediction, all outperforming the comparative models and greatly enhancing the accuracy and real-time capability of slurry mass concentration soft sensing, indicating significant potential for industrial applications.