S110100120孙鸿烨.doc

分类号： 密 级： U D C ： 编 号： 近红外光谱建模中的变量选择方法研究 Near-infrared Spectrum Modeling Method of Variable Selection Research 学位授予单位及代码： 长春理工大学 （10186） 学科专业名称及代码： 物理电子学 （08001） 指 导 教 师： 石晓光教授 研 究 生： 孙鸿烨 论文起止时间：2012.10—2013.12 摘 要 近红外光谱技术以其分析效率高、检测速度快、成本较低等等诸多优点在多个领域都有着非常广泛的应用。但是由于光谱中的某些波长点,不仅对于分析和建模没有任何贡献, 甚至会影响模型质量, 导致更加复杂的模型及其预测能力下降。因此选择代表样本信息的重要变量一直是是近红外光谱建模分析中的重要内容。 本文主要在复杂的近红外光谱信息中采用移动窗口偏最小二乘法，无信息变量消除法，遗传算法和连续投影算法进行了光谱特征变量的选择，一定程度上消除了冗余信息，并结合支持向量机方法进行建模。通过对比四种变量选择方法，可以得到采用连续投影算法进行变量选择后，一定程度上简化了模型，并提高了模型的预测精度。 BSTRACT There are many applications of Near-Infrared Spectroscopy in various fields for its high efficiency, rapid detection, low cost and so unique advantages. However, the Wavelength points in some certain wavelengths not only have any contribution to the analysis and modeling and even affect the quality of the model. It will lead to more complex models and the decline of predictive ability. Therefore choose to represent an important variable sample information has been the important content in modeling and analysis of Near Infrared Spectroscopy. In this paper, in the complex four methods were used: moving window partial least square, uninformative variables elimination, genetic algorithms, successive projections algorithm, extracting useful information and with support vector machine modeling. Comparing with four kinds of variable selection methods, the method of continuous projection algorithm variable selection could enhance quantitative spectral calibration model prediction accuracy and modeling efficiency and the NIR calibration models with good predictive ability and stability could also be acquired. Key words:near-infrared spectroscopy variable selection support vector regression 目 录 摘 要 I ABSTRACT II 目 录 III 第一章 绪 论 1 1.1 引言 1 1.2 近红外光谱技术的发展 1 1.3 近红外光谱的化学及物理基础 2 1.4 近红外光谱技术的特点及其分析步骤 3 1.5 本文研究的目的意义和研