特征变量选取与模型的建立是光谱定量分析的两个主要研究内容.首先讨论了Tikhonov正则化特征光谱选取算法在多组分烷烃气体分析应用中的参数确定方法,然后针对甲烷、乙烷、丙烷、异丁烷、正丁烷、异戊烷和正戊烷七种烃烷的小浓度分析,从中红外吸收光谱中提取了七组特征光谱,并用这些特征光谱作为输出,用神经网络建立了七种烃烷气体的分析模型.为克服神经网络的过训练问题,提出根据误差处理的方式从多个训练好的网络中选择最优网络的网络最优参数选择法.最后给出了分析模型的标气检验结果表明,在各种烃烷气体1%范围内,提出的分析方法有效消除了各种烷烃之间的交叉敏感,交叉干扰小于0.5%;分辨率高,达20×10-6.%Feature variable selection and modeling are two of the most principal research contents in spectral analysis. In the present paper, beginning from the introduction of feature spectrum selection based on Tikhonov regularization and discussion on it's application in multi-component mixed alkane gas analysis, 7 sets of feature spectra were abstracted from the absorption spectra of 7 kinds of alkane gas, including methane, ethane, propane, iso-butane, n-butane, iso-pentane and n-pentane. In order to overcome the problem of over-training of neural network, a method called optimal parameter selection of neural netework (NN) was presented to build analysis model of analyte. Optimal parameters were selected from many trained networks with same architecture based on error process. And analysis models of spectral analysis for 7 kinds of alkane gas were built. Finally, the testing analysis results done with standard gases are given. The results show that the method presented in this paper can be used to reduce the cross-sensitivity between any two kinds of gas. The cross-sensitivity is less than 0.5％. The resolving power is as high as 20f 10-6.
展开▼
