Cross components calibration transfer of NIR spectroscopy Model through PCA and weighted ELM-based TrAdaBoost algorithm

摘要

With the rapid development of NIR spectroscopy technology and chemometrics, many previous studies have focused on calibration transfer of quantitative analysis model and lots of effectively methods have been proposed, such as slope and bias correction (SBC), piecewise direct standardization (PDS) etc., by which we can implement calibration transfer between different spectrometers. Furthermore, whether it is possible to realize calibration transfer cross different components or not? To answer this question, this paper proposed a novel method which combines principal component analysis (PCA), weighted extreme learning machine (ELM) and TrAdaBoost algorithm. Two public NIR spectroscopy datasets (Corn and Gasoline) are applied to validate the possibility and effectiveness of proposed algorithm through four different experimental protocols. The experimental results show that while the objects of source and target domains are same, whatever calibration transfer between different instruments or components (experimental protocol #2, #3 and #4), the generalization performance of target domain model will improve a lot, especially while target domain contains fewer samples. Particularly, compared with experimental protocol #2 and #3 (only instruments or components between source and target domains are different), there is a significant improvement while the instruments and components are all different (experimental protocol #4). However, while the objects, components and instruments between source and target domains are all different, the generalization performance of quantitative analysis model can not be improved after calibration transfer. The experimental results indicate that in the area of NIR spectroscopy calibration transfer area, the assumption of original TrAdaBoost algorithm can be relaxed so that the labels between source and target domains can different (cross components), but the features must be same.