Across different instruments about tobacco quantitative analysis model of NIR spectroscopy based on transfer learning

With the development of near-infrared (NIR) spectroscopy, various calibration transfer algorithms have been proposed, but such algorithms are often based on the same distribution of samples. In machine learning, calibration transfer between types of samples can be achieved using transfer learning and does not need many samples. This paper proposed an instance transfer learning algorithm based on boosted weighted extreme learning machine (weighted ELM) to construct NIR quantitative analysis models based on different instruments for tobacco in practical production. The support vector machine (SVM), weighted ELM, and weighted ELM-AdaBoost models were compared after the spectral data were preprocessed by standard normal variate (SNV) and principal component analysis (PCA), and then the weighted ELM-TrAdaBoost model was built using data from the other domain to realize the transfer from different source domains to the target domain. The coefficient of determination of prediction (R-2) of the weighted ELM-TrAdaBoost model of four target components (nicotine, Cl, K, and total nitrogen) reached 0.9426, 0.8147, 0.7548, and 0.6980. The results demonstrated the superiority of ensemble learning and the source domain samples for model construction, improving the models' generalization ability and prediction performance. This is not a bad approach when modeling with small sample sizes and has the advantage of fast learning.

Automated summary

This study proposed an instance transfer learning algorithm based on weighted extreme learning machine for constructing NIR quantitative analysis models for tobacco under different instruments. The results demonstrated the importance of ensemble learning and source domain samples in improving the models' generalization ability and prediction performance.