结合PLS-DA与SVM的近红外光谱软测量方法

为了提高近红外光谱分析精度,提出结合偏最小二乘判别分析(PLS-DA)与支持向量机(SVM)的软测量方法(PLS-DA-SVM).该方法利用一组由不同类别组成的训练样本,引入二叉树进行多重分类,节点分类器由PLS-DA方法建立;利用偏最小二乘支持向量机(PLS-SVM)建立每类样本的定量模型.预测时,用PLS-DA分类树对待测样本进行分类,选择相应的PLS-SVM模型进行定量分析.实验利用PLS-DA-SVM方法和近红外光谱数据建立汽油的研究法辛烷值软测量模型,针对2个批次共计57个成品汽油样本进行蒙特卡洛交叉检验.结果表明,对汽油牌号进行识别,平均分类错误率为0.07%,低于其他常用分类方法;对研究法辛烷值进行预测,均方误差达到0.243,复相关系数达到0.991,较PLS、LS-SVM等方法有显著提高.

Near-infrared spectroscopy soft-sensing method by combining partial least squares discriminant analysis and support vector machine

To improve the performance of near-infrared spectral analysis,this paper proposes a soft-sensing method(PLS-DA-SVM) which combines partial least squares discriminant analysis(PLS-DA) and support vector machine(SVM).Based on training samples with several classes,firstly,a binary tree built by PLS-DA is introduced for multiple classification;secondly,sub-models for quantitative analy are constructed by partial least squares support vector machine(PLS-SVM).For a test sample,PLS-DA classification tree serves to determine its class,and the corresponding PLS-SVM sub-model is selected for quantitative analysis.A PLS-DA-SVM model with near-infrared spectroscopy data was established to determine the research octane number of gasoline samples.Monte Carlo cross validation was preformed with 57 product gasoline samples from 2 oil refineries.Results show that mean classification error rate for the recognition of gasoline brands is 0.07%,which is lower than other pattern recognition methods.Root mean square error of prediction(RMSEP) is reduced to 0.243 and correlation coefficient(R2) is up to 0.991,which show great improvement upon PLS,LS-SVM and other modeling methods.