针对混合气体建模过程中最小二乘支持向量机参数难以确定及红外光谱数据计算量过大的问题, 提出一种粒子群优化的最小二乘支持向量机方法, 用于建立基于主成分分析特征提取的红外光谱多组分气体定量分析模型。 首先对主吸收峰区域的550个红外光谱数据利用主成分分析技术进行了特征提取, 将降维得到的7个特征值作为模型的输入变量从而有效地降低了计算量。 混合气体主要由浓度范围分别是0.1%～1%的甲烷、 乙烷及0.1%～1.5%的丙烷三种组分气体组成。 采用最小二乘支持向量机技术分别建立了各组分气体的定量分析模型, 利用粒子群优化算法对最小二乘支持向量机算法中的参数进行了优化选取, 取代了传统的遍历优化方法, 然后利用取得的最优参数重建定量分析模型。 实验结果表明, 采用此方法离线建模所用时间比采用遍历优化方法节省40倍以上, 预测结果误差水平相当, 满足实测要求。 粒子群优化算法在全局优化及收敛速度方面具有较大优势。 粒子群优化算法与最小二乘支持向量机技术相结合用于混合气体定量分析是切实可行的, 具有一定的实际意义和应用价值。Swarm Optimization in Quantitative Analysis of Gas Mixture

According to the difficulty in selecting parameter of least squaresupport vector machine (LS-SVM) when modeling on the gas mixture, and the highcomputational complexity of the infrared spectrum data, LS-SVM optimized byparticle swarm optimization (PSO) algorithm was proposed to build an infraredspectrum quantitative analysis model with feature extracted by principalcomponent analysis (PCA). Firstly, seven feature variables were extracted by PCAas the input of the model from 550 infrared spectrum data of the main absorptionapex field, so the computational complexity was reduced. This model aimed atthree components of gas mixture, in which methane, ethane and propane gases areincluded. The concentration of each component ranged from 0.1% to 1%, 0.1% to 1%and 0.1% to 1.5% respectively. Each component quantitative analysis model wasbuilt by LS-SVM and the parameters were optimized by PSO algorithm, then theregression model would be reconstructed according to the optimal parameters. Thismethod replaced the traditional ergodic optimization. The experiment results showthat the time of offline modeling by PSOwas reduced to one fortieth of that ofergodic optimizing. The precision of the model was corresponsive. It can meet therequirement of the measure. PSO algorithm has more superior performance on globaloptimization and convergence speed. So it is feasible to combine PSO algorithmwith LS-SVM to create the infrared spectrum quantitative analysis model. It hasdefinite practice significance and application value.square support vector machine; Quantitative analysis; Gas mixture; Principalcomponent analysis