Enhancing multi-type coal quality prediction accuracy with fusion spectra and classification models using NIRS and XRF techniques

Abstract

The various analytical indices of coal are important criteria for evaluating the quality of commercial coal. Coals of different qualities exhibit different physical and chemical characteristics in their utilization. In the case of multiple coal types, the spectral characteristics of different coals may overlap within certain wavelength ranges, or be affected by interference or noise from other coal types, leading to low accuracy in coal quality prediction. Rapid and accurate coal quality testing is of great significance for improving industrial production efficiency and enhancing corporate profitability. This study employs near-infrared spectroscopy (NIRS) and X-ray fluorescence spectroscopy (XRF) combined techniques to explore the accuracy and feasibility of predicting coal quality based on coal type classification models. In terms of classification algorithms, coal samples are identified and classified using Support Vector Machine (SVM) based on fusion spectra. Regarding the modeling approach, Partial Least Squares (PLS) is utilized to establish both an overall model for all coal samples and individual classification models corresponding to each coal type. The results show that the precision, accuracy, recall, and F₁ score of this classification algorithm reached 96.49%, 97.50%, 95.83%, and 96.41%, respectively. The determination coefficients (R²) for the classification model's predictions of ash, volatile matter, and sulfur in coal quality indicators reached 0.992, which represents improvements of 1.85%, 5.31%, and 10.10% over the overall model. The root mean square errors of prediction (RMSE_P) for these indicators were 0.062, 0.080, and 0.008, showing reductions of 0.24%, 0.68%, and 0.05% compared to the overall model. It indicates that the method of first identifying the coal type and then predicting coal quality indicators using the corresponding classification model can significantly improve the accuracy of coal quality detection in complex coal type scenarios.