Research on calibration transfer methods in coal quality spectral analysis across instruments and coal types.

Background: In coal quality analysis, spectroscopic techniques such as near-infrared spectroscopy (NIRS) and X-ray fluorescence (XRF) offer rapid and non-destructive measurements, but differences between instruments and coal types may lead to the failure of established models. Given the complexity of coal, it is important to systematically evaluate the applicability of traditional calibration transfer methods, such as the Slope/Bias (S/B) method and Piecewise Direct Standardization (PDS). Meanwhile, machine learning algorithm such as Domain-Adversarial Neural Networks (DANN), have shown great potential in addressing data distribution differences between source and target domains, offering new possibilities for calibration transfer.

Results: To address the challenges of cross-instrument and cross-coal-type model transfer in coal quality analysis, this study systematically compares S/B method, PDS, and DANN for calibration transfer by using two self-developed NIRS-XRF coal quality analyzers and 264 gas coal and fat coal samples. For cross-instrument transfer, DANN achieved a coefficient of determination for prediction (R) of 0.92, a root mean square error of prediction (RMSE) of 0.68 % and a mean absolute relative deviation (MARD) of 6.19 %., outperforming S/B method (0.84, 0.92 %, 8.95 %) and PDS (0.88, 0.82 %, 6.70 %). In cross-coal-type transfer, traditional methods failed to provide reliable prediction results (e.g., S/B method with R = 0.22, RMSE = 1.10 %, and MARD = 13.61 %), whereas DANN demonstrated transfer capability with R = 0.75 and RMSE = 0.64 %, and MARD = 7.89 %. Notably, DANN required only 18 cross-instrument or 14 cross-coal-type calibration samples to reach optimal performance. Spectral comparisons and elemental analysis further explain the limitations of traditional methods in coal quality analysis.

Significance: The results highlight the superior adaptability of DANN in spectroscopic calibration transfer across instruments and coal types, offering significant accuracy improvements with minimal calibration effort. This study provides an innovative solution for calibration transfer in coal spectroscopic analysis and holds substantial significance for promoting the practical application of spectroscopic technologies in industrial settings.