基于白术FTIR的径向基函数神经网络鉴别研究

Identification of Rhizoma Atractylodes Based on FTIR Spectra and Radial Basis Function Network

为了鉴定白术及其伪品,采用径向基函数神经网络(RBF)分别测试了白术及其伪品的傅里叶变换红外光谱。采用36个样本作训练集,27个样本作检验集,用各种模式的BPF进行了监督性训练。当训练目标误差平方和定为0.01时,各类RBF对训练集中白术样本识别的正确率均为100%,但对检验集样本识别的结果各不相同,其识别的正确率与隐含层节点数S1有关。发现当S1较大时,识别正确率反而下降,可能此时网络的非线性程度过高,使其不适合于该类样本集的训练。线性—线性型RBF识别的结果随S1的变化不是很大,但识别的正确率不高,基本在85%左右。非线性—线性型RBF识别的结果最佳。当S1为3时,其识别正确率超过了97%。因此该法可用于简便、快速、准确地识别白术及其伪品。

In order to recognize the atractylodes macrocephala Koidz. （rhizoma atraetylodes） and its confusable varieties, three kinds of models of radial basis function network（RBF）, nonlinear-linear, linear-linear, and nonlinear-nonlinear model, were used combined with their Fourier transform infrared spectra （FTIR）. Rhizoma atractylodes models were collected by Fourier transform infrared spectra, 36 samples were gathered as a training target, and 27 samples as a test set, then their supervision training was performed using three models each. When the summation of error square of the training target was selected as 0.01, the correct rate for recognition of Fourier transform infrared spectra using each RBF was 100% for the training set, but was different for the test set, which depended on the number of mode in hidden layer, $1. It was found that with the increase of s1, the correct rate would decrease oppositely. This may be caused by the high degree of the nonlinearity of the networks, so that the models of networks were not fit for the training of this kind of sample set. When using linear-linear model of RBF, the correct rate varied with S1 to some extent, but was generally about 85%. Recognizing ability obtained using nonlinear-linear model of RBF was the best. Its correct rate of recognition was 〉97%. When S1=3, and so this method can be used to recognize atractylodes macrocephala Koidz. （rhizoma atraetylodes） and its confusable varieties simply, rapidlly and accurately.