Cross-instrument coal quality spectral analysis via a PDS-assisted fine-tuning calibration transfer method

Area of Science:

• Analytical Chemistry
• Geochemistry
• Machine Learning

Background:

• X-ray fluorescence (XRF) instruments exhibit device-related variations causing spectral inconsistencies.
• These inconsistencies degrade quantitative coal quality models, particularly for ash content prediction.
• Existing calibration transfer and single-stage transfer learning methods struggle with significant inter-instrument spectral discrepancies, especially with limited samples.

Purpose of the Study:

• To propose a two-stage transfer learning framework, PDS-FT, for enhanced cross-instrument ash content prediction using XRF.
• To systematically evaluate the performance of PDS-FT against other methods like PDS, FT, DANN, and PDS-DANN.
• To demonstrate the framework's ability to handle inter-instrument spectral variations for robust coal quality analysis.

Main Methods:

• A two-stage transfer learning framework integrating Piecewise Direct Standardization (PDS) with Feature Fine-tuning (FT) was developed.
• Five approaches were evaluated: PDS, FT, Domain-Adversarial Neural Network (DANN), PDS-DANN, and the proposed PDS-FT.
• The study utilized 140 coal samples measured on two distinct XRF instruments.

Main Results:

• Standalone PDS and FT showed limited performance (R²=0.33, R²=0.51).
• DANN improved prediction (R²=0.74) but faced training instability.
• The proposed PDS-FT framework achieved the highest performance (R²=0.95, RMSE=0.59%), demonstrating superior accuracy and generalization.

Conclusions:

• The PDS-FT framework combines spectral-level physical correction with feature-level model adaptation for hierarchical domain alignment.
• This dual-stage strategy effectively mitigates inter-instrument discrepancies, enhancing model transferability and reliability.
• PDS-FT offers a practical and generalizable solution for coal quality analysis across diverse XRF instruments in industrial settings.