Abstract
Accurate modeling of near-infrared spectroscopy (NIRS) data is not easy, especially for automatic and optimal analysis, since it includes multiple steps and many algorithms and parameters in each step. Even the order of the methods for pretreatment affects the outcomes of models. In this work, a novel strategy is reported to discover the suitable parameters for data pre-processing and variable selection with the help of particle swarm optimization (PSO). Using the optimization process of 14 parameters for pre-processing and two parameters for variable selection as examples, it reduces the experience required for NIRS modeling and helps to construct partial least squares (PLS) quantitative models with significant improvement. This approach promotes the use of NIRS after simplifying the process to construct ideal models. In this study, three NIRS spectra of flavors and fragrances and 13 corresponding physical and chemical indices were applied to characterize the strategy. Simultaneous modeling of PSO global optimization and traditional methods were comprehensively investigated for comparison, in which the indices R2, Q2, RMESC and RMSECV were employed for evaluation. The average values were 0.89, 0.88, 0.29 and 0.47 by the former, and 0.86, 0.66, 0.36 and 0.57 by the latter. Almost all the former modeling results were better than the latter. This fully demonstrates that the proposed method is powerful for NIRS data analysis.
Acknowledgments
This work was supported by the foundations (Contract No. 202043000834036) of the China Tobacco Hunan Industrial Company and the foundation (2019XX03) from China Tobacco Yunnan Industrial Company.