发酵作为影响红茶品质形成的重要流程, 发酵品质程度的判断主要基于人工经验, 难以实现准确客观的评价。 该研究主要针对于工夫红茶发酵工序, 以不同发酵时序下的样品为对象, 利用高光谱检测技术并结合化学计量学方法, 对制备的不同发酵程度的样本进行无损检测和智能判别。 首先利用高光谱成像仪（400～1 000 nm）采集工夫红茶发酵样品的高光谱数据, 并根据气温、 茶叶嫩度、 萎凋情况、 揉捻过程、 发酵叶颜色及香气等现场生产信息, 将6个不同发酵时序下的样本, 根据发酵程度依次划分为3类（轻度发酵、 适度发酵、 过度发酵）。 为了降低采集高光谱信息时因培养皿中发酵叶的不平整而产生的散射现象对光谱数据的影响, 选取标准正态变量变换算法（standard normal variate, SNV）与多元散射校正算法（multiplicative scatter correction, MSC）对全波段光谱进行预处理, 将预处理后的光谱数据进行主成分分析（principal components analysis, PCA）, 分别得到前3个主成分的三维载荷图, 根据样本在图中的空间分布特征, 因而选择效果较好的SNV预处理方法。 以全波段光谱最优主成分作为模型输入量, 建立邻近算法(K-nearest neighbor, KNN)、 随机森林（random forests, RF）、 极限学习机（extreme learning machine, ELM）判别模型, 识别率分别为63.89%, 94.44%和86.11%, 结果表明, 非线性模型（RF、 ELM）识别率较高, 其中RF模型性能优于ELM模型。 为比较基于全波段与特征波长建立的工夫红茶发酵品质程度模型判别效果, 采用连续投影算法（successive projections algorithm, SPA）提取31个特征波长进行PCA降维处理, 以特征波长最优主成分作为模型输入量, 构建SPA-KNN, SPA-RF和SPA-ELM判别模型, 识别率分别为83.33%, 91.67%和91.67%。 通过SPA对变量筛选后, SPA-KNN和SPA-ELM模型性能明显提高, SPA-RF模型识别准确度略有下降。 与特征波长建立的模型相比, 全波段建立的RF模型性能最佳, 对工夫红茶轻度发酵、 适度发酵、 过度发酵的判别率分别达到了100%, 83.33%和83.33%。 研究结果为推进红茶智能化、 数字化加工的实现, 提供了理论基础和科学依据。

Fermentation is a crucial process that affects the quality of black tea. The quality of fermentation is mainly judged on artificial experience, hence making it difficult to accomplish an accurate and objective evaluation. The objective of this study is to investigate the fermentation quality of Congou at different times and temperatures using non-destructive techniques and intelligent discrimination methods (chemometrics). For this purpose, different congou fermentation samples were prepared at different fermentation timings. Thereafter, these samples under went testing by hyperspectral detection technology and chemometrics methods. First, the hyperspectral imager (400~1 000 nm) was used to collect the hyperspectral data of Congou fermentation samples. Next, according to the on-site production information, such as temperature, tea tenderness, withering condition, rolling process, fermentation leaf color, aroma, and so on, six different fermentation samples under different time series were divided into three categories according to the degree of fermentation (light, moderate, and excessive fermentation). Standard normal variate (SNV) and multiple scatter correction (MSC) were selected to preprocess the full-band spectrum.Principal components analysis (PCA) was applied to the preprocessed spectral data to obtain the three-dimensional load maps of the first three principal components (PCs). Thereafter, a better SNV preprocessing method was selected according to the spatial distribution characteristics of the samples in the map. The k-nearest neighbor (KNN), random forests (RF), and extreme learning machine (ELM) discriminant models were established by using the optimal PCs of the full-band spectrum as the model input.The recognition rates of KNN, RF, and ELM were 63.89%, 94.44%, and 86.11%, respectively. The results showed that the recognition rate of the non-linear model (RF, ELM) was higher, and the performance of the RF model was better than that of the ELM model. 31 characteristic wavelengths were extracted by successive projections algorithm (SPA) for PCA dimension reduction.The SPA-KNN, SPA-RF and SPA-ELM discriminant models were constructed, and their recognition rates were 83.33%, 91.67%, and 91.67%, respectively. After the variables were screened by SPA, the performances of SPA-KNN and SPA-ELM models were found to be significantly improved, and the recognition accuracy of the SPA-RF model was slightly decreased. As compared with the model established by the characteristic wavelength, the RF model established in the whole band showed the best performance, and the discrimination rates of light fermentation, moderate fermentation and excessive fermentation of Congoureached at 100%, 83.33%, and 83.33%, respectively. The research results provided a theoretical and scientific basis for advancing the realization of intelligent and digital processing of black tea.