应用近红外光谱技术结合最小二乘支持向量机建立了国公酒中橙皮苷含量的模型。 利用Kernard-Stone法对训练集样本进行划分， 对光谱数据预处理方法进行了选择， 比较了平滑、 范围标度化、 自标度化、 一阶微分、 二阶微分以及这几种预处理相互结合的六种方法， 确定了以平滑、 一阶微分， 范围标度化作为国公酒近红外光谱的数据预处理方法， 采用组合的间隔偏最小二乘法筛选出有效波段8 211～8 312 cm－1及9 712～9 808 cm－1。 应用最小二乘支持向量机建立模型， 所建模型的交叉验证误差均方根为0.000 1， 预测误差均方根为0.004， 预测集的相对偏差小于5%。 与组合的间隔偏最小二乘法、 径向基-人工神经网络和支持向量机进行了比较。 该方法快速、 无损且可靠， 可作为国公酒中橙皮苷含量快速测定的手段。

Near-infrared spectroscopy (NIRS) combined with least squares support vector machines (LS-SVM) was used to establish a new method for the determination of the hesperidin content in guogongjiu medicinal wine. Firstly, training set was partitioned by Kernard-Stone (KS) algorithm. Secondly, spectral pretreatment methods were discussed in detail, comparing smoothing, rangescaling, autoscaling, first derivative, second derivative, along with those methods combined. Smoothing, first derivative and rangescaling were used for the pretreatment of the NIR spectra of guogongjiu medicinal wine. Thirdly, the effective interval was selected for 8 211-8 312 and 9 712-9 808 cm-1 by synergy interval partial least squares (siPLS). Finally, the model was established by LS-SVM, the root mean square error of cross validation (RMSECV) is 0.001, root mean square error of prediction (RMSEP) is 0.004, and relative deviation of predicting set is less than 5%. It was compared with siPLS, radial basis function neural network (RBF-NN), and SVM, The result shows that the method is rapid, non-destructive, and credible. It is an effective measurement for determining the hesperidin content in guogongjiu medicinal wine.