基于粒子群算法的最小二乘支持向量机在红花提取液近红外定量分析中的应用

提出一种基于粒子群算法的最小二乘支持向量机(PSO-LS-SVM)方法,用于建立红花提取过程关键质控指标的定量分析模型.近红外光谱数据经波段选择、预处理和主成分分析(降维)后,利用粒子群优化(PSO)算法对最小二乘支持向量机算法中的参数进行优化,然后使用最优参数建立固含量和羟基红花黄色素A(HSYA)浓度的定量校正模型.将校正结果与偏最小二乘法回归(PLSR)和BP神经网络(BP-ANN)比较,并将所建的3个模型用于红花提取过程未知样本的预测.结果表明,BP-ANN校正结果优于PSO-LS-SVM和PLSR,但是对验证集和未知样品集的预测能力较差,而PSO-LS-SVM和PLSR模型的校正、验证结果相近,相关系数均大于0.987,RMSEC和RMSEP值相近且小于0.074,RPD值均大于6.26,RSEP均小于5.70％.对于未知样品集,pSO-LS-SVM模型的RPD值大于8.06,RMSEP和RSEP值分别小于0.07％和5.84％,较BP-ANN和PLSR模型更低.本研究所建立的PSO-LS-SVM模型表现出较好的模型稳定性和预测精度,具有一定的实践意义和应用价值,可推广用于红花提取过程的近红外光谱定量分析.

Application of Particle Swarm Optimization Based Least Square Support Vector Machine in Quantitative Analysis of Extraction Solution of SafflowerUsing Near-infrared Spectroscopy

A novel particle swarm optimization(PSO) based least squares support vector machine(LS-SVM) method was investigated for quantitative analysis of extraction solution of safflower using near-infrared(NIR) spectroscopy.The usable spectral region(5400-6500 cm-1) was identified,spectral preprocessing of Norris derivative smoothing was employed,and spectral dimension was also reduced through principal component analysis(PCA).In this paper,the PSO algorithm was applied to select the LS-SVM hyper-parameters(including the regularization and kernel parameters).The calibration models of total solid content and hydroxysafflor yellow A(HSYA) were established using the optimum hyper-parameters of LS-SVM.The performance of LS-SVM models was compared with partial least squares regression(PLSR) and back-propagation artificial neural networks(BP-ANN).The feasibility of these three methods was examined on the unknown sample set.Experimental results showed that the calibration results of BP-ANN were superior to PSO-LS-SVM and PLSR,however,the prediction accuracy of validation and unknown sample set was inferior.For PSO-LS-SVM and PLSR models,the correlation coefficients of the calibration and validation set were above 0.987,the RMSEC and RMSEP values were close to each other and less than 0.074,residual predictive deviation(RPD) values were all above 6.26,and the RSEP values were controlled within 5.70%.For the unknown sample set,the RPD values of PSO-LS-SVM models were above 8.06,the RMSEP and relative standard errors of prediction(RSEP) values were less than 0.07 and 5.84% respectively,which were much lower than BP-ANN and PLSR models.The PSO-LS-SVM algorithm employed in this paper exhibited excellent model robustness and prediction accuracy,which has definite practice significance and application value.