光谱分析中Elastic Net 变量选择与降维方法

在利用红外光谱进行多组分混合气体定量分析建模中，须根据各目标气体成分的光谱特点进行光谱维数降维和特征变量选择。以甲烷、乙烷、丙烷、异丁烷、正丁烷、异戊烷和正戊烷等7种气体为分析目标，采用最小绝对收缩和选择算子（LASSO）与弹性网络（ElasticNet）方法进行目标气体数据预处理。针对LASSO 和ElasticNet方法参数优化选择的问题，采用均方误差和预测偏差最小两个准则进行参数的优化选取。对4cm-1的实测光谱数据，采用LASSO和ElasticNet方法分别在0.0019和0.0021均方误差条件下使得维度从2542维分别降为2 维和3维，LASSO 的交叉灵敏度最大和最小为10.2718%和1.420 5%，ElasticNet分别为5.4945%和0.7493%。结果表明:Elastic Net在用于光谱定量分析的数据预处理中具有一定的优势，为准确建立定量分析模型奠定了基础。

Spectral wavelength selection and dimension reduction using Elastic Net in spectroscopy analysis

In the use of Fourier transform infrared spectroscopy to build the multi-component gases quantitative analysis model, it is necessary to reduce the dimensions and select characteristics wavelength according to the target gas spectral. Through the regularization algorithm analysis, least absolute shrinkage and selection operator (LASSO) and Elastic Net method were used to do these for seven kinds of mixed gases of methane, ethane, propane, iso-butane, n-butane, iso-pentane and n-pentane. The minimum mean square error (MSE) and prediction deviation were used as the criteria to select LASSO and Elastic Net parameters. Finally, the resolution of 4cm-1 measured spectral data was analyzed. The dimension of spectra were reduced from 2 542 d to 2d and 3d respectively by using LASSO and Elastic Net method under the condition of the MSE of 0.001 9 and 0.002 1. The cross sensitivity of maximum and minimum were 10.271 8% and 1.420 5% by LASSO method. The cross sensitivity of maximum and minimum were 5.494 5% and 0.749 3% by Elastic Net. Results show that the Elastic Net method was better in the characteristic variable selection and the spectral dimension reduction for gas spectral quantitative analysis，and it was foundation to establish the accurate quantitative analysis model.