Error-diagnosis-guided cross-indicator adaptive calibration for near-infrared spectroscopic multi-property prediction of coal quality.

Near-infrared spectroscopy (NIRS) coal-quality datasets often contain many repeated spectra acquired from the same coal samples. Ignoring the coal-sample group structure and indicator-dependent error patterns may compromise the reliability of generalization assessment and calibration stability. To address this issue, this study proposes an error-diagnosis-guided cross-indicator adaptive calibration (ED-CIAC) method for NIRS-based multi-property prediction of coal quality. ED-CIAC performs coal-sample-group-level modeling by robustly aggregating repeated spectra and integrating complementary spectral representations to obtain property-specific base predictions. The remaining prediction errors are characterized across properties and value ranges to identify systematic residual structures. Cross-indicator features constructed from the multi-property base predictions are subsequently introduced as complementary calibration information. Candidate calibrators with direct and residual learning modes are evaluated within each training fold, and the calibrator, learning mode, and correction strength are jointly selected for each property according to internal validation performance. Experiments were conducted on 26 batches of NIRS coal-quality data, including 122,983 repeated spectra from 538 independent coal-sample groups, using coal-sample-group-level fivefold cross-validation. The proposed method achieved R2 values of 0.8717 ± 0.0227, 0.9077 ± 0.0120, 0.8112 ± 0.0460, and 0.9000 ± 0.0157 for total moisture (Mt), dry-basis ash (Ad), dry-basis volatile matter (Vd), and net calorific value as received (Qnet,ar), respectively, outperforming conventional chemometric, machine learning, and literature-based comparison methods. These results demonstrate that group-level multi-view base prediction, residual-guided property-specific calibration, and cross-indicator information utilization can improve the reliability of NIRS-based multi-property coal-quality prediction.