Chemically Informed Targeted Multiblock Data Fusion for Coal Analysis Using Laser-Induced Breakdown Spectroscopy and Near-Infrared Reflectance Spectroscopy

Accurate and comprehensive characterization of coal quality is essential for improving combustion efficiency and reducing emissions in coal-fired power plants. Combining Laser-induced Breakdown Spectroscopy (LIBS) and Near-infrared Reflectance Spectroscopy (NIRS) provides complementary atomic and molecular information, but effective data fusion remains challenging. This study proposes a targeted multiblock analysis framework guided by the intrinsic chemical correlations among coal properties. First, a supervised feature extraction method is applied to the LIBS and NIRS spectra to obtain feature blocks corresponding to specific coal property. For each target property, primary prediction models are then constructed on all feature blocks. Their regression coefficient matrices are combined into a final prediction model by assigning weights to each block according to its predictive accuracy. To address the fact that a universal model cannot accommodate all coal ranks, a coal type classification of unknown samples is further developed using spectral features. Compared with single block models, the targeted multiblock model reduces the average absolute errors (AAE) to 0.313 MJ kg–1, 0.426%, 0.491%, and 0.204% for calorific value, volatile matter, ash content, and moisture content, respectively, versus 0.920 MJ kg–1, 0.847%, 0.939%, and 0.336% for the baseline method. Furthermore, distance correlations among coal properties show a high agreement with the multiblock weight distribution. These results demonstrate that the proposed targeted multiblock strategy improves both the accuracy and chemical interpretability of LIBS-NIRS coal analysis, and provides a clear route to integrating multispectral and chemical information.