HATR–FTIR wavenumber selection for predicting biodiesel/diesel blends flash point

• We propose a new framework for selecting predictive wavenumbers in HATR–FTIR.
• The method tests four wavenumber importance indices, leading to good results.
• Our propositions are applied to biodiesel/diesel blends.

A novel HATR–FTIR wavenumber selection framework is proposed to predict the flash point of biodiesel/diesel blends. Partial Least Squares (PLS) regression is applied to spectra and four wavenumber importance indices are derived from PLS parameters. Noisy and irrelevant wavenumbers are then iteratively removed from the HATR–FTIR spectra according to the order suggested by each index following a backward procedure, and the Root Mean Square Error (RMSE) of the PLS model assessed. Two approaches are then suggested to select the recommended wavenumber subset once the iterative elimination procedure is finished. Using the recommended wavenumber importance index, the proposed method retained only average 5.13% of original wavenumbers, while reducing the average RMSE 21.6%, from 1.302 to 1.021. The method is then compared to flash point prediction with Principal Component Regression (PCR) when wavenumbers are selected using importance indices derived from Principal Component Analysis (PCA) parameters.