茶叶傅里叶红外光谱的可能模糊K调和均值聚类分析

茶叶的品种不同，其有机化学成分含量往往不同，其功效也是不尽相同的，因此，研究出一种简单、高效、识别率高的茶叶品种鉴别技术方法是十分有必要的。中红外光谱技术是一种快速检测技术，在用中红外光谱仪采集得到的茶叶中红外光谱中含有噪声信号。为了对含噪声茶叶中红外光谱的准确分类以实现茶叶品种分类，将可能模糊C-均值聚类(PFCM)思想应用到K调和均值(KHM)聚类，设计出一种可能模糊K调和均值(PFKHM)聚类算法，计算出PFKHM的模糊隶属度、典型值和聚类中心。可能模糊K调和均值聚类能有效解决K调和均值聚类的噪声敏感性问题。用傅里叶红外光谱分析仪(FTIR-7600型)分别对三种茶叶(优质乐山竹叶青、劣质乐山竹叶青和峨眉山毛峰)进行扫描以获取它们的傅里叶中红外光谱。光谱波数区间是4 001.569～401.121 1 cm-1。先采用主成分分析法(PCA)将光谱数据压缩到20维，再采用线性判别分析(LDA)将光谱数据压缩到两维并提取鉴别特征信息。最后分别用K调和均值聚类和可能模糊K调和均值聚类实现茶叶品种分类。实验结果：当权重指数m=2，q=2和p=2时，KHM具有91.67%的聚类准确率，PFKHM聚类准确率达到94.44%；KHM迭代12次达到收敛，而PFKHM迭代11次就可以达到收敛。采用傅里叶红外光谱技术检测茶叶，用主成分分析和线性判别分析压缩光谱数据，再用可能模糊K调和均值聚类进行品种分类可快速、准确地实现茶叶品种的鉴别。

Possibilistic Fuzzy K-Harmonic Means Clustering of Fourier Transform Infrared Spectra of Tea

Different variety of tea often has diversified organic chemical components, and their effects are not the same. Therefore, it is very necessary to develop a simple, efficient, high recognition rate method in classifying tea varieties. Mid-infrared spectroscopy is a rapid detection technology, and there is noise signal in the mid-infrared spectra of tea samples collected by spectrometer. With a view to identifying tea varieties through the classification of the mid-infrared spectra of tea samples with noise, possibilistic fuzzy c-means clustering was applied in K-harmonic means clustering (KHM) and a novel clustering, called possibilistic fuzzy K-harmonic means clustering (PFKHM), was proposed. PFKHM can produce both fuzzy membership value and typicality value and solved the noise sensitivity problem of KHM. First of all, we used FTIR-7600 spectrometer to scan three varieties of tea samples (i. e. Emeishan Maofeng, high quality Leshan trimeresurus and low quality Leshan trimeresurus) for their Fourier transform infrared spectroscopy (FTIR) data. The wave number of FTIR data ranged from 4 001.569 to 401.121 1 cm-1. Secondly, we employed principal component analysis (PCA) to compress spectral data into 20-dimensional data which were compressed into two-dimensional data by linear discriminant analysis (LDA). Lastly, we used KHM and PFKHM to classify the tea varieties respectively. The experimental results indicated that when the weight index m=2, q=2 and p=2 the clustering accuracy rates of KHM and PFKHM achieved 91.67% and 94.44%, respectively. KHM was convergent after 12 iterations and PFKHM was convergent after 12 iterations. Tea varieties could be quickly and accurately classified by testing tea with FTIR technology, compressing spectral data with PCA and LDA, and classifying tea varieties with PFKHM.