近红外光谱定量分析中三种新型波长选择方法研究

发布时间：2018-03-29 04:09

本文选题：近红外光谱　切入点：波长选择算法　出处：《中国农业大学》2017年博士论文

【摘要】：近红外光谱作为一种快速无损检测技术在农产品品质分析领域发挥着越来越重要的作用。然而近红外光谱中往往存在着大量无信息波长甚至是噪声波长,因此波长选择已经成为近红外光谱分析中一个关键的环节,目前已有多达几十种波长选择算法。本文系统分析和研究了现有波长选择算法原理的差异,并将近红外光谱分析中常用波长选择算法大体上划分为基于偏最小二乘模型参数、智能优化算法、连续投影策略、模型集群分析策略和波长区间选择等五类。为了解决现有方法存在的可靠性和稳定性问题,本文在近红外光谱谱学特点的基础上,分别借鉴模型集成、模型集群分析和串联等三种新思路开展了以下几项研究,并采用玉米、土壤、药片等三种代表性样品体系进行了算法验证。(1)首次采用移动窗口平滑集成策略(MWWS)对竞争性自适应加权采样算法(CARS)进行了改造,得到了一种名为MWWS-ECARS的新型波长选择算法。实验结果表明,MWS-ECARS算法通过移动窗口对CARS算法重复运行后各波长的累积被选频率进行平滑处理,不仅克服了 CARS算法选择波长的不稳定性,而且还可以通过调节移动窗口宽度和阈值的大小对最终选中波长区间的宽度进行优化。(2)在模型集群分析(MPA)的框架下提出了一种新型波长区间组合优化算法(ICO)。ICO算法首先采用软收缩的方式对波长区间组合进行优化,然后采用局部搜索的方式对最终入选区间的宽度进行自动优化。实验结果表明ICO算法不仅具有软收缩的特点,还具有收敛速度更快,参数设置较少等优点;ICO算法中采用波长区间替换波长点作为优化对象,既可以较好地降低软收缩策略在寻优过程中的计算负担,又可以降低优化算法出现过拟合的风险;其中采用的WBS方法被证明比WBMS方法更适合用于在模型集群分析框架下开发新型波长选择算法,因为其可以通过在MPA的随机采样环节引入更合适水平的随机因素来克服WBMS方法存在的缺陷。(3)基于算法串联的思路采用连续投影算法(SPA)对MWWS-ECARS和ICO算法选中的波长进行简化,结果表明SPA算法可以在保证所建模型的预测能力不出现显著下降的同时,进一步简化上述两种算法选中的波长点,但是样品体系越复杂,简化的力度就越小。(4)通过实验考察了不同预处理方法对ICO算法选择波长分布和建模效果的影响情况。结果表明,不同光谱预处理方法对于ICO算法选择波长的分布情况和建模效果均有着较强的影响。
[Abstract]:Near infrared spectroscopy as a rapid nondestructive detection technology plays an increasingly important role in the quality of agricultural products. However, the domain analysis in near infrared spectroscopy is the existence of a large number of information without wavelength even noise wavelength, therefore the wavelength selection has become a key link in near infrared spectroscopy analysis, there are dozens of wavelength selection research on the existing algorithm. The wavelength selection algorithm principle and difference analysis of this system, and the analysis of near infrared spectroscopy was used in wavelength selection algorithm is generally divided into partial least squares model parameters based on intelligent optimization algorithm, continuous projection strategy, model cluster analysis strategy and wavelength interval selection of five. In order to solve the problem of reliability and stability the existing methods, this paper in the near infrared spectrum based on the characteristics, respectively referring to the model integration, model and cluster analysis on With three new ideas in the following study, and the use of corn, soil, pills and other three kinds of representative sample system to verify the algorithm. (1) for the first time using moving window smoothing integration strategy (MWWS) sampling algorithm for competitive adaptive weighted (CARS) to get a choice of transformation a new algorithm for wavelength of MWWS-ECARS. The experimental results show that the MWS-ECARS algorithm by moving the window on the accumulation of each wavelength CARS algorithm was repeated after running smooth frequency rate, not only to overcome the CARS algorithm to select the wavelength of the instability, but also can be optimized by adjusting the window width and the width of the mobile threshold on the size of the final selected wavelength interval. (2) analysis in cluster model (MPA) framework proposed a new wavelength interval optimization algorithm (ICO).ICO algorithm adopts soft contraction on wavelength interval Combined optimum width and then uses a local search method was chosen to optimize the interval. The experimental results show that the ICO algorithm not only has soft shrinkage, but also has faster convergence speed, less parameter settings etc.; the ICO algorithm uses the wavelength interval to replace wavelength as the optimization object, which can reduce the computational burden in the optimization process of the soft contraction strategy, but also can reduce the risk of over fitting optimization algorithm; the WBS method proved more suitable for the development of new wavelength selection algorithm analysis framework in the model cluster than WBMS method, because it can overcome the defects of WBMS method by random factors into more appropriate level random sampling part of MPA. (3) the idea of using a continuous series algorithm based on the projection algorithm (SPA) to select the MWWS-ECARS and ICO algorithm in wavelength Simplified, results show that the prediction ability of SPA algorithm can model that does not appear significant decline at the same time, to further simplify the above two algorithms selected wavelengths, but the sample system is more complex, the simplified strength is small. (4) through the effects of the choice of the wavelength distribution and modeling the effect of ICO the algorithm of different pretreatment methods. The results show that different spectral pretreatment methods for the ICO algorithm to select the distribution and the effect of modeling wavelength has a stronger effect.

【学位授予单位】：中国农业大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：O657.33

【参考文献】