拉曼光谱背景扣除算法及其应用研究

【作者】 陈珊；

【导师】 李晓宁； 梁逸曾；

【作者基本信息】 中南大学 ， 分析化学， 2011， 硕士

【摘要】 拉曼光谱是一种散射光谱,是光通过介质时入射光与分子相互作用而引起的频率发生变化的散射,是一种利用分子振动-转动信息的光谱分析法。它作为一种新兴发展起来的的分析手段可提供快速、简便、可重复,无需对样品进行前处理的检测,通过光纤探头或石英器皿就可直接测量,从而对物质进行无损伤、快速的定性定量分析。但测定过程中,样品含有杂质或荧光吸收物质时会产生荧光干扰,且当激发光子提供了足够的能量以致产生荧光时,拉曼信号将变模糊甚至被掩盖。因此通过寻求避免荧光干扰的方法来改进其应用成了拉曼光谱学家们一直致力的问题。部分物质在测量时虽然有荧光干扰,但荧光辐射不足以全部湮灭拉曼信号时,仍可采集下拉曼光谱,此时光谱有一定的荧光背景,影响数据的进一步分析。本论文主要运用化学计量学的方法对拉曼背景扣除算法进行了研究,通过查找和阅读大量中外文献,对拉曼光谱原理、特点,拉曼光谱仪与拉曼光谱技术发展历程等内容做了概述。本文通过引入了惩罚最小二乘(Penalized Least Squares)和小波系数收缩(Wavelet Shrinkage)等平滑算法来滤除拉曼光谱中的高频噪声,与本文介绍的几种预处理方法相比具有一定的优势。同时也介绍了建立聚类,判别分析与回归分析模型的化学计量学方法。本文为扣除影响分析结果的荧光背景,我们提出了基于连续小波变换和惩罚最小二乘的算法baseline Wavelet,通过真实样本分析结果表明：所提出的背景扣除算法能在不丢失有效信息的前提下有效地将背景扣除。通过比较在滤除噪声和扣除背景前后建立聚类模型与判别分析模型,如PCA、Random Forest等模型,扣除背景后分离度和分类结果均有改善。另外分析了baseline Wavelet背景扣除算法中参数的设定对拟合背景平滑度的影响。本文通过提出自适应迭代重加权惩罚最小二乘airPLS算法对大量的样本进行背景扣除来建立回归模型。它利用稀疏矩阵技术,该方法能够快速有效扣除光谱中荧光背景。通过实例分析airPLS算法在定量方面的应用。通过比较几种背景扣除算法后建立的回归分析模型,说明airPLS算法在几种方法中不仅具有好的效果且在在运算时间上也有很好的优势,适合大批量的样本集。更多还原

【Abstract】 Raman spectroscopy is kind of scattering spectrum, which is caused by the light through the media and molecular interaction of the frequency changed by the scattering, using the information of vibrational rotational in a spectrometric method. It developed as a new analytical tools which provide fast, simple, repeatable, and samples need not pre-treatment, can measured directly by the fiber optic probe or through the quartz vessel. So it provides rapid and non destructive qualitative quantitative analysis. During the measurement procedure, as samples containing impurities or fluorescent material will bring the fluorescence interference and the excitation photon provides sufficient energy so that produce the fluorescence, the Raman signals will all be blurred or swamped by fluorescence.Although Raman spectra of some materials have fluorescent interference, but the fluorescence radiation is not sufficient to swamp the Raman signal, which can still be collected with Raman instruments. The Raman spectra have a certain background, which will affect further analysis of spectra using chemometrics method. Raman background correction algorithms are studied in this thesis, the brief introduction about domestic and foreign researching status in the Raman spectra, the purpose and significance of this thesis, and the main contents and methods are presented. By finding and reading a lot of literature, the theory and characteristics of Raman spectroscopy, instrument and development of Raman techniques were outlined.Penalized Least Squares and Wavelet Shrinkage are introduced in this paper to filter out high frequency noise in the Raman spectra. They have advantages when comparison with other soothing methods. Also we described the chemometric methods such as clustering, discriminant analysis and regression analysis model.Again, based on continuous wavelet transform and penalized least squares algorithms, the baselineWavelet is proposed and applied for baseline correction. The results show that background correction without missing important information. By comparion of models of cluster and discriminant analysis, such as PCA, Random Forest, before and after filtering noise and background correction, the separability and classification results are respectively improved. The influence of the lambda parameter from baselineWavelet on the smoothness of the fitted baseline has been clearly explained and clarified.We proposed the adaptive iterative re-weighted penalized least squares algorithm, which uses sparse matrix techniques to correct the baseline quickly and effectively. Since large number of samples is required for establishing the regression model. Example of quantitative analysis of airPLS algorithm is applied. By comparison of the built regression models which the spectra was pretreated with several baseline correction methods, the airpls baseline correction method has been demonstrated that it can fit the better baseline, have the smaller RMSECV values and require less computation time. The airpls method is very suitable for correcting the baseline of large amount calibration spectra when built the regression models with chemometrics algorithms.更多还原