中文摘要：

为进一步提高FTIR光谱法实现特征吸收光谱严重重叠的甲烷、乙烷、丙烷、异丁烷、正丁烷、异戊烷以及正戊烷七组分混合气体定量分析的精度和速度，提出一种核偏最小二乘（Kernel Partial Least Square，KPLS）特征提取耦合支持向量回归机（Support Vector Regression Machine，SVR）的红外光谱定量分析新方法．首先采用KPLS方法对上述七组分混合气体的FTIR光谱进行特征提取，然后将特征提取得到的特征组分作为SVR的输入建立混合气体的定量分析模型．对标准混合气体进行定量分析的结果显示：KPLS—SVR模型的预测精度高于未进行特征提取SVR模型预测的精度，同时预测时间也减少了一半．研究表明，KPLS法可以很好地提取隐含在混合气体FTIR光谱数据与其组分浓度之间的非线性特征并有效地消除光谱数据噪声，大幅度降低数据维数，与SVR耦合可以提高红外光谱分析的精度和速度，是一种有效的红外光谱定量分析方法．

英文摘要：

A new method for FTIR spectral quantitative analysis was presented. The new method couples kernel partial least squares （KPLS） feature extraction with support vector regression machine（SVR） to improve the quantitative analysis accuracy and speed of seven-component alkane gas mixtures composedoof methane, ethane, propane, iso-butane, n-butane, isopentane, and n-pentane, whose feature absorption spectra are cross each other and overlapped seriously. Firstly, the KPLS was employed to extract feature components from the FTIR spectra of above-mentioned seven-component gas mixtures. And then, the extracted feature components were fed into SVR to create the quantitative analysis model of seven component gases. The quantitative analysis results of calibration gas mixtures show that the prediction accuracy by KPLS-SVR model is higher than that by SVR model without feature extraction processing. Meanwhile, the predicting time by KPLS-SVR model is only half of that by SVR model. The study indicates that KPLS approach can effectively extract the latent nonlinear features implied in the spectra and component concentration, eliminate the noise of FTIR spectral data, and reduce the dimension of the spectral data. Coupling with SVR, KPLS feature extraction can improve the accuracy of FTIR spectral analysis, shorten the predicting time. KPLS-SVR is a very effective method for infrared spectral quantitative analysis.