Zhang Bao-hua, Jiang Yong-cheng, Sha Wen, Zhang Xian-yi, Cui Zhi-feng
Guang Pu Xue Yu Guang Pu Fen Xi. 2015 Feb;35(2):390-3.
Three feature extraction algorithms, such as the principal component analysis (PCA), the discrete cosine transform (DCT) and the non-negative factorization (NMF), were used to extract the main information of the spectral data in order to weaken the influence of the spectral fluctuation on the subsequent quantitative analysis results based on the SERS spectra of the pesticide thiram. Then the extracted components were respectively combined with the linear regression algorithm--the partial least square regression (PLSR) and the non-linear regression algorithm--the support vector machine regression (SVR) to develop the quantitative analysis models. Finally, the effect of the different feature extraction algorithms on the different kinds of the regression algorithms was evaluated by using 5-fold cross-validation method. The experiments demonstrate that the analysis results of SVR are better than PLSR for the non-linear relationship between the intensity of the SERS spectrum and the concentration of the analyte. Further, the feature extraction algorithms can significantly improve the analysis results regardless of the regression algorithms which mainly due to extracting the main information of the source spectral data and eliminating the fluctuation. Additionally, PCA performs best on the linear regression model and NMF is best on the non-linear model, and the predictive error can be reduced nearly three times in the best case. The root mean square error of cross-validation of the best regression model (NMF+SVR) is 0.0455 micormol x L(-1) (10(-6) mol x L(-1)), and it attains the national detection limit of thiram, so the method in this study provides a novel method for the fast detection of thiram. In conclusion, the study provides the experimental references the selecting the feature extraction algorithms on the analysis of the SERS spectrum, and some common findings of feature extraction can also help processing of other kinds of spectroscopy.
采用主成分分析（PCA）、离散余弦变换（DCT）和非负矩阵分解（NMF）这三种特征提取算法来提取光谱数据的主要信息，以减弱光谱波动对基于农药福美双表面增强拉曼光谱（SERS）的后续定量分析结果的影响。然后，将提取的成分分别与线性回归算法——偏最小二乘回归（PLSR）和非线性回归算法——支持向量机回归（SVR）相结合，建立定量分析模型。最后，采用五折交叉验证法评估不同特征提取算法对不同回归算法的影响。实验表明，由于SERS光谱强度与分析物浓度之间存在非线性关系，SVR的分析结果优于PLSR。此外，无论采用哪种回归算法，特征提取算法都能显著改善分析结果，这主要是因为它提取了源光谱数据的主要信息并消除了波动。另外，PCA在线性回归模型上表现最佳，NMF在非线性模型上表现最佳，在最佳情况下预测误差可降低近三倍。最佳回归模型（NMF+SVR）的交叉验证均方根误差为0.0455微摩尔·升⁻¹（10⁻⁶摩尔·升⁻¹），达到了福美双的国家检测限，因此本研究方法为福美双的快速检测提供了一种新方法。总之，本研究为SERS光谱分析中特征提取算法的选择提供了实验参考，一些特征提取的常见发现也有助于其他类型光谱的处理。
