不同品种茶叶因其所含的有机化学成分不同, 其效果也会有差别。 所以, 寻找出一种能准确迅速的鉴别茶叶品种的技术方法是非常重要的。 近红外光谱(NIR)分析是一种无损检测技术, 能很好的鉴别茶叶品种。 使用NIR光谱仪采集茶叶的NIR数据。 为了对包含噪声信号的茶叶近红外光谱进行准确鉴别, 提出了一种模糊线性判别QR分析的新方法, 可以对茶叶近红外光谱进行准确分类。 通过使用模糊线性判别分析(FLDA)将由主成分分析(PCA)压缩的茶叶近红外光谱数据进行降维, 由模糊线性判别分析得出的特征向量构建鉴别向量矩阵, 对鉴别向量矩阵进行矩阵的QR分解, 得到新的鉴别向量矩阵。 经过模糊线性判别QR分析后使用K近邻算法进行分类, 具有准确率高等优点。 以岳西翠兰、 六安瓜片、 施集毛峰和黄山毛峰四种茶叶为研究样本, 每类65个, 茶叶样本总数为260个。 采集茶叶近红外光谱数据的仪器为AntarisⅡ型傅里叶近红外光谱仪对光谱数据进行预处理, 采用多元散射校正, 由于采集到的茶叶光谱数据存在散射干扰。 以此得到的近红外光谱数据的维数为1557维, 通过主成分分析压缩数据集的维数, 使得光谱数据集的维数达到7维。 经压缩过后的光谱数据集中的鉴别信息再通过模糊线性判别QR分析进行提取, 使得光谱数据的维数降低到3维。 利用K近邻算法对茶叶样本进行分类, 实现对茶叶品种的准确分类。 最后进行三种算法分析结果的比较, 分别是主成分分析结合K近邻算法、 主成分分析和线性判别分析结合K近邻算法、 主成分分析和模糊线性判别QR分析结合K近邻算法。 在权重指数 m=2, K=1条件下, 最后的分类准确率分别为83.89%, 87.78%和98.33%。 实验结果显示: 模糊线性判别QR分析可以实现茶叶近红外光谱的准确鉴别分析, 其展现出来的效果比主成分分析和线性判别分析表现的效果更好。

The effects of different varieties of tea are different because of their different organic chemical components. Therefore, it is essential to find a technical method that can accurately and quickly identify tea varieties. Near-infrared (NIR) spectroscopy is a nondestructive detection technology correctly identifying tea varieties. Due to noise signals in the NIR spectra of tea samples collected by the NIR spectrometer, a fuzzy linear discriminant QR analysis method was proposed to accurately identify the NIR spectra of tea samples containing noise signals. After the dimensionality of NIR spectra was compressed by principal component analysis (PCA), it was reduced using fuzzy linear discriminant analysis (FLDA). The discriminant vector matrix was constructed from the eigenvectors obtained by FLDA. The discriminant vector matrix was decomposed by QR decomposition to obtain a new discriminant vector matrix. Then, the K-nearest neighbor (KNN) algorithm was used for classification, which has the advantage of high accuracy. Four kinds of tea samples, namely Yuexi Cuilan, Lu'an Guapian, Shiji Maofeng and Huangshan Maofeng, were taken as the experimental samples. There were 65 tea samples in each category, and the total number of tea samples was 260. Firstly, the NIR spectral data of tea samples were collected by the Fourier NIR spectrometer Antaris Ⅱ. Secondly, the obtained NIR spectral data of tea were preprocessed, and the scattering effect of spectral data was reduced through multiple scattering correction. Thirdly, the dimensionality of NIR data is 1 557, so PCA was used to reduce the dimensionality of the spectra to 7. Then, fuzzy linear discriminant QR analysis was performed to extract the identification information from the compressed NIR spectra, and the dimensionality of the data was further reduced to 3 dimensions. Finally, KNN was used to classify tea samples and achieved the accurate classification of tea varieties. Furthermore, the experimental results were compared including three algorithms, which are PCA combined with KNN, PCA and linear discriminant analysis (LDA) combined with KNN, PCA and fuzzy linear discriminant QR analysis combined with KNN. Under the weight index m=2 and K=1, the final classification accuracies of the three algorithms were 83.89%, 87.78% and 98.33%, respectively. The experimental results showed that fuzzy linear discriminant QR analysis provided a method for the identification of NIR spectra of tea, and its effect was better than PCA and LDA.