近红外光谱模型转移新算法

New Algorithms for Calibration Transfer in Near Infrared Spectroscopy

模型转移是解决分析仪器或分析方法通用性的关键技术。近红外光谱受测量仪器或测量条件的影响较大,模型转移对近红外光谱技术的实际应用尤为重要。本文综述了近年来近红外光谱分析中被广泛应用和新提出的模型转移算法,从计算原理角度梳理了有标样和无标样算法的联系和区别。有标样算法重点介绍了基于多元校正、因子分析、人工神经网络、多任务学习的模型转移方法,无标样算法重点介绍了基于光谱校正、模型参数校正和稳健建模的模型转移方法。从算法的角度分析了各种模型转移方法的特点和转移效果,并展望了模型转移算法的进一步发展。在综述的众多方法中分段直接标准化及其变体仍是模型转移的黄金标准,但是,基于因子分析的算法正变得受欢迎且基于神经网络和多任务学习的方法近年来也吸引了越来越多的注意。但是,在实际应用中,获得标准样品以在主机和子机上测得其光谱比较困难甚至是不可能的,无标样模型转移则更加实用。此外,随着仪器小型化、成像及超光谱成像的发展,模型转移在未来会变得愈加必不可少。

Calibration transfer is a key technique to ensure the consistency of instruments or analytical methods. Near infrared spectra is strongly influenced by the status of instrument or the environment of measurement. Therefore, calibration transfer is essential for practical applications of near infrared spectroscopy. This paper provides an overview of the state-of-the-art pertaining to calibration transfer methods, including those with and without standard samples, and the focus is on the methods based on multivariate calibration, factor analysis, neural network and multi-task learning for algorithms with standard samples and on the methods based on spectral correction, model coefficient correction and robust multivariate calibration for algorithms without standard samples. The formulation and efficiency of different calibration transfer methods are analyzed from the viewpoint of the algorithms. Among the summarized methods, piecewise direct standardization （PDS） and its variants are still the gold standard for calibration transfer, but algorithms based on factor analysis is becoming popular and those based on neural network and multi-task learning has gradually attracted more attention in recent years. In practical applications, however, it is difficult or even impossible to obtain the standard samples for measuring their spectra on both master and slave instruments, calibration transfer without standard samples is more practical. Furthermore, with the development of the instrumentation in miniaturization, imaging and hyper-spectral imaging, calibration transfer will be more and more essential in the future.