High accuracy raw coal analysis by LIBS via fusing feature aware and score-space anomaly detection

The primary objective of this study is to develop a robust ash quantification model for raw coal trucks via hierarchical grade classification and PCA-based anomaly detection. To achieve this, an integrated analytical strategy is proposed to address the persistent challenges of matrix effects and anomalous sample interference in laser induced breakdown spectroscopy (LIBS) for online analysis of complex, heterogeneous solids. The approach effectively combines feature-aware spectral classification, based on the cyano radical (CN) and diatomic carbon radical (C₂) molecular emission bands, with the anomaly detection algorithm in principal component space. The method first leverages the emission bands of CN and C₂ radicals to rapidly filter spectra dominated by non-coal impurities and classify coal spectra into three ash based grades, ultra-low ash (ULA, ≤10 wt%), low ash (LA, 10-20 wt%), and medium ash (MA, 20-30 wt%). Subsequently, a robust anomaly detection algorithm, based on the geometric distribution of data in principal component space, is employed to identify and eliminate residual spectral outliers arising from extreme heterogeneity or measurement error, and the proposed method is capable of capturing anomalous spectral structures as well as spectral data that cannot be adequately described by the PCA model representing the majority of coal samples. The established grade-specific partial least squares regression models achieved remarkable prediction accuracy, with the root mean square error of test (RMSET) reduced to 0.3592 wt% (ULA), 0.4318 wt% (LA), and 0.8290 wt% (MA), and coefficients of determination (R²) exceeding 0.92 for all categories. Field validation on moving coal trucks demonstrated excellent agreement with standard laboratory methods with measurement error less than 1.0 wt%, confirming the industrial robustness of the method. This work presents a generalizable LIBS analysis paradigm that transforms spectral awareness into analytical reliability, offering a potent solution not only for real-time coal quality assurance but also for the online compositional monitoring of other heterogeneous materials.