核密度估计算法结合近红外光谱技术鉴别三叶青产地

三叶青是我国珍稀中药材，具有多种疗效，但不同产地的三叶青有效成分含量差异悬殊，为防止三叶青以次充好，其产地鉴别尤为重要。以浙江、云南、安徽、广西和湖北五个产地的三叶青为研究对象，利用傅里叶变换近红外光谱分析仪(Fourier transform near infrared spectroscopy, FT-NIR)收集4 000~10 000 cm-1范围内的近红外光谱，由于三叶青近红外光谱数据还未完善，因此在其产地鉴别上，应对鉴别算法提出更高的要求，即在实现三叶产地鉴别的同时，还要能够有效地识别出其他或未知新产地的三叶青。针对这一问题，本文结合三叶青近红外光谱数据的特点，对算法共做了三方面改进：①从距离的角度估计样本的概率密度；②以训练样本可信度的方式计算带宽参数；③在未知新产地的识别上，提出一种基于训练集样本的概率密度函数的识别方法。结果表明，该算法对训练集样本的识别精度达到了100%，且在140组预测集样本中，只有3组样本识别出错，并能够100%地识别出未知新产地的三叶青，说明基于核密度估计的改进算法在三叶青产地鉴别上，不仅鉴别精度高，且能够有效识别出其他或未知新产地的三叶青。

Identification of Tetrastigma hemsleyanum from Different Places with FT-NIR Combined with Kernel Density Estimation Algorithm

Tetrastigma hemsleyanum, a rare medicinal herbs in China, contains many kinds of curative effects. However, the content of active ingredients of T. hemsleyanum from different places is remarkablely different. So, it is necessary to discriminate this promising medicinal T. hemsleyanum from different places. In this work, spectra of T. hemsleyanum collected from Zhejiang, Yunnan, Anhui, Guangxi and Hubei provinces were recorded with Fourier transform near infrared spectroscopy, ranging from 10 000 to 4 000 cm-1. And the identification algorithm was applied to effectively identify the T. hemsleyanum from the known origin and other new places because the spectral data of T. hemsleyanum is not sufficient. Hence, in this study, three improvements of kernel density estimation algorithm have been achieved to identify T. hemsleyanum: (1) estimate the probability density of the samples via the perspective of distance; (2) calculate the bandwidth parameters by training the credibility of samples; (3) propose a recognition method based on probability density function of training set samples to recognize unknown origin. The identifying accuracy of training set sample and prediction set by the algorithm were reached 100% and 97.8%, respectively. Additionally, the new places of T. hemsleyanum can be accurately identified used the algorithm. The results show that the improved algorithm based on kernel density estimation can effectively identify T. hemsleyanum, and recognize the unknown origin samples.