SE-PDS enhanced NIR spectral transfer learning: A machine learning approach for cross-instrument jet fuel property...

This study proposes an improved Shared-Embedding-Enhanced Piecewise Direct Standardization (SE-PDS) algorithm to achieve near-infrared (NIR) spectral transfer across different instruments. Combined with a Random Forest (RF) model, successfully enables cross-instrument quantitative prediction of 16 physicochemical properties of jet fuels. The validation was conducted using 86 jet fuel samples collected from two different instruments. The results demonstrate that the spectra calibrated by SE-PDS exhibit superior performance compared to those calibrated by Direct Standardization (DS) and Piecewise Direct Standardization (PDS) methods, with an average correlation coefficient of 0.9999 and a minimum correlation coefficient of 0.9997 between the two instruments. In terms of predictive performance, the RF model achieved mean coefficients of determination (R²) of 0.8102, 0.7613, and 0.9231 under DS, PDS, and SE-PDS, respectively, demonstrating superior performance over the Partial Least Squares (PLS) model (DS: 0.7174, PDS: 0.6855, SE-PDS: 0.8878). These results confirm the RF model's stronger capability in nonlinear modeling. Notably, the SE-PDS-RF combination exhibited the best predictive performance, with R² values ranging between 0.8858 and 0.9851 across all target properties, alongside consistently lower relative error rates (RER) compared to traditional transfer methods. This approach integrates spectral matrices and window indices from all segments, enabling a single model to leverage shared information across different characteristics while automatically identifying and eliminating irregular fluctuations and abrupt variations in spectral data. The study provides a robust solution for cross-instrument model transfer of complex mixtures, highlighting the synergistic advantages of nonlinear modeling and deep learning-enhanced transfer strategies.