Abstract
To improve the accuracy and reliability of coal quality assessment via near-infrared spectroscopy, this study proposes a multi-method analysis framework for robust spectral feature selection. A core challenge is reconciling the trade-offs between different analytical approaches: statistical methods often yield smooth but diffuse results, while machine learning models can identify sharp, localized features that may lack stability. Our framework addresses this by integrating diverse analytical perspectives, including statistical correlations, SHAP-interpreted machine learning models, and latent-variable regression. We then introduce a novel fusion strategy that synthesizes the importance profiles from these methods based on inter-method consistency, curve smoothness, and local concentration. Experimental results demonstrate this fusion yields more interpretable and physicochemically coherent wavelength importance profiles for both Moisture (Mad) and Volatile Matter (Vad). The selected features consistently achieve superior prediction performance across various regression models, showing particular robustness with limited training data. This work offers a structured methodology for identifying compact and informative spectral features, facilitating the development of efficient models for online monitoring and contributing to improved process control.