Coal combustion is a major source of CO2 emissions, making accurate carbon quantification essential for emission assessment and mitigation. This study proposes a trimodal fusion prediction system for carbon estimation, which integrates multienergy laser-induced breakdown spectroscopy (LIBS) and laser-induced plasma acoustic (LIPA) signals with machine learning for precise carbon analysis in coal. Using five standard coal samples, we established quantitative models via external and internal standard methods and evaluated low-level data fusion of LIBS and LIPA with various algorithms. Random forest demonstrated optimal performance, and we adopted feature importance ranking to enhance predictive capability. SHAP interpretability analysis revealed that although LIBS spectral features dominated the baseline model, carbon-related features were excluded, highlighting a lack of chemical interpretability. To address this, additional energy modalities were introduced, along with two novel feature extraction methods: targeted area-preserving PCA, which retains carbon-specific spectral regions during dimensionality reduction, and hybrid time-frequency alignment PCA, which enhances LIPA acoustic feature stability via time-frequency alignment. Trimodal data fusion of LIBS, LIPA, and multienergy information significantly improves model accuracy, reliability, and generalization capability, offering a promising tool for CO2 emission monitoring and coal quality assessment.
Gao et al. (Thu,) studied this question.