论文部分内容阅读
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。
Model transfer is the key technique to solve the commonality of analytical instruments or analytical methods. Near-infrared spectroscopy is greatly affected by measuring instruments or measuring conditions, and model transfer is particularly important for near-infrared spectroscopy. In this paper, the model transfer algorithms, which are widely used and newly proposed in recent years, are reviewed. The relations and differences between standard and non-standard algorithms are reviewed from the perspective of computational principle. The standard sample algorithm mainly introduces the model transfer method based on multivariate calibration, factor analysis, artificial neural network and multitasking. The standardless sample algorithm mainly introduces the model transfer method based on spectral correction, model parameter calibration and robust modeling. The characteristics and transfer effects of various model transfer methods are analyzed from the perspective of algorithms, and the further development of model transfer algorithms is prospected. Segment direct normalization and its variants are still the gold standard for model transfer in many of the reviewed methods, however, algorithms based on factor analysis are becoming popular and methods based on neural networks and multitasking have also attracted more in recent years More attention. However, in practice, it is more difficult or even impossible to obtain a standard sample to measure its spectrum on the host and the slave, and the standard-based model transfer is more practical. In addition, with the development of instrument miniaturization, imaging and hyperspectral imaging, model transfer will become even more necessary in the future.