Chemically informed spectro-structural feature augmentation for interpretable NIR chemometric classification of fuel grades

Rapid and interpretable near-infrared discrimination of fuel grades remains challenging because adjacent fuel grades usually exhibit highly overlapped C-H overtone and combination bands, while the structural origin of model decisions is often insufficiently described. To address this problem, this study proposes a knowledge-guided spectro-structural feature augmentation framework, termed SSFA, for interpretable NIR-based fuel classification. Unlike conventional variable selection or simple feature-weighting methods, SSFA uses changes in the carbon-chain length of n-alkanes as chemical prior knowledge, converts the "carbon chain length-NIR response" relationship into structure-augmented features with directional constraints, and embeds structure-sensitive bands into the model input space as learnable descriptors. These chemically meaningful descriptors are fused with full-spectrum NIR variables, enabling classifiers to simultaneously capture data-driven spectral patterns and structural discriminative cues. A total of 292 fuel samples were collected, covering jet fuel, gasoline, and diesel from 26 representative refineries. The effectiveness of SSFA was systematically evaluated under different spectral preprocessing methods, feature selection strategies, and machine-learning classification models. SSFA improved classification performance across multiple linear, nonlinear, and tree-based models. The optimal model, based on SG second-derivative preprocessing and BP-ANN, achieved a fuel-grade recognition accuracy of 95.6% and an average positive predictive value of 96.30%. In addition, all 12 real fuel samples were correctly identified, providing preliminary evidence for practical applicability. The results indicate that SSFA is an interpretable feature augmentation framework for NIR-based fuel quality screening and may provide a methodological reference for spectral classification of other complex chemical mixtures.