由于分子结构的高相似性, 烃类气体混合物中各组分红外光谱谱峰重叠严重, 导致浓度的精确监测一直是化学计量学的难题。 为了应对这一挑战, 提出一种粗精选策略二进制灰狼优化（RSBGWO）算法, 用于优选红外光谱特征, 建立高精度定量分析模型。 该方法以交叉验证下光谱定量分析模型的均方根误差（RMSECV）平均值作为适应度函数值。 在粗选阶段, 进行第一次全局迭代, 更新α狼、 β狼和δ狼所选特征变量的位置信息； 在精选阶段, 结合α狼所选的特征变量以及剔除α狼未选中特征变量位置后的β狼和δ狼特征变量, 更新狼群位置信息, 逐步降低RMSECV值, 提取为全局最优特征波长, 并引入非线性收敛因子加快收敛速度。 该算法在采集的359个混合烷烃气体样本的红外光谱数据集上进行了实验测试并验证了所提算法的效果。 与bGWO和bPSO特征提取算法比较, 基于本文提出的RSBGWO算法建立的MLR模型在分析甲烷、 乙烷、 丙烷和二氧化碳气体浓度时, 特征选择数量均降低了96%以上, 预测均方根误差（RMSEP）均低于数据采集过程中所使用的配气系统的仪器误差, 相对预测偏差（RPD）均提高了15以上。 相对于全谱建模的MLR模型和PLS模型, 基于RSBGWO算法建立的MLR模型和PLS模型的预测精度有显著增高, 预测效果对定量分析模型的依赖性降低了。 实验结果表明, 提出的方法具有优秀的红外光谱特征提取能力, 能够明显提高定量分析模型的预测效果。 该方法能够促进光谱检测技术在生物制药、 食品化工、 油气勘探等领域的应用, 尤其是在含同系有机物混合物的应用场合。

Due to the seriously overlapped infrared spectral peaks of each component in hydrocarbon gas mixtures, which is caused by the high similarity of molecular structures, it has always been a difficult problem in stoichiometry to precisely monitor the concentration. A rough and fine selection strategy binary gray wolf optimization (RSBGWO) algorithm is proposed to optimize infrared spectral features and establish a high-precision quantitative analysis model to address this challenge. It takes the mean value of root mean square error (RMSECV) of the spectral quantitative analysis model based on cross-validation as the fitness function. In the rough selection stage, the first global iteration is carried out to update the location information of the selected characteristic variables for α wolf, β wolf and δ wolf. In the fine selection stage, combining the characteristic variables for α wolf, the characteristic variables for β wolf and δ wolf after eliminating the corresponding characteristic variables in which position are not selected for α wolf, are used to update the location information of wolves, in order to reduce the RMSECV value gradually and make sure that the extracted characteristic wavelength is globally optimal. In addition, a nonlinear convergence factor is introduced to accelerate the convergence speed.The algorithm is tested on the infrared spectral data set of 359 mixed alkane gas samples, and the effect of the proposed algorithm is verified. Compared with bGWO and bPSO feature extraction algorithms, the MLR model based ontheRSBGWO algorithm proposed in this paper reduces the number of the selected feature by more than 96% and increases the relative prediction deviation (RPD) by more than 15. The root mean square error of prediction (RMSEP) is lower than the instrument error of gas distribution system used for data acquisition when analyzing the concentrations of methane, ethane, propane and carbon dioxide. Compared with the MLR model and PLS model of full spectrum modeling, the prediction accuracy of the MLR model and PLS model based on the RSBGWO algorithm proposed in this paper is significantly improved, and the dependence of prediction effect on the quantitative analysis model is reduced. The experimental results show that the method proposed in this paper can significantly improve the analysis effect of the quantitative analysis model of infrared spectroscopy. The method can promote the application of spectral detection technology in biopharmaceuticals, the food chemical industry, oil and gas exploration, etc., especially in the application occasions containing homologous organic compounds.