Non-destructive quantification of tobacco blend components using FT-NIR spectroscopy coupled with multivariate machine learning

Résumé fourni par la source

Ensuring the stability of blending quality in formulated tobacco is critical for maintaining cigarette quality and market reputation. In this study, FT-NIR spectroscopy combined with multivariate machine learning was employed to predict the proportions of tobacco components in formulated tobacco. Spectral pre-processing (MSC, SNV, L2-normalization) and feature selection methods (SPA, CARS, GA) were applied before establishing regression models using techniques such as PLSR, SVR, GPR, BRR, and CNN. Full-spectra-based models generally exhibited the highest performance, with R 2 p exceeding 0.95, and RMSEP lower than 1.21 % for all components. Although CARS-based models were slightly less accurate than the full-spectra-based models, they still achieved comparable predictive performance, while drastically reducing the number of input variables. This reduction not only lowers computational cost but also minimizes redundant spectral information. For instance, the CARS-SVR model for tobacco silk achieved an R 2 p of 0.982 and RMSEP of 0.99 %, and the CARS-SVR model for expanded tobacco silk reached an R 2 p of 0.988 and RMSEP of 0.65 %. All optimal CARS-based models maintained RPD values above 4, indicating reliable predictive capability. These results demonstrate that FT-NIR spectroscopy can accurately determine tobacco blend proportions, providing robust theoretical and technical support for industrial quality control. • FT-NIR spectroscopy enables rapid, non-destructive prediction of tobacco blend ratios. • Accurate predictions were achieved for all four tobacco components, with the best models for each attaining an RMSEP below 1.3 %. • CARS-SVR models reduced variables while maintaining high accuracy. • The approach offers robust support for industrial tobacco blending quality control.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.