利用电子鼻对6个不同贮藏时间下5个等级黄山毛峰茶进行检测。首先获取反映茶叶香气的原始特征向量，再通过主成分分析法（PCA）提取出前5个主成分作为主特征向量，然后以主特征向量作为BP神经网络（BPNN）的输入，建立黄山毛峰茶贮藏时间预测模型（称为PCA-BPNN）。通过对75个测试样本（每等级15个）实验测试表明：PCA-BPNN对于贮藏0 d的茶叶，最大预测误差为7 d，5个（6.67%）样本预测误差超过10 d；对于贮藏60 d的茶叶，最大预测误差为10 d，4个（5.33%）样本预测误差超过10 d；对于贮藏120 d的茶叶，最大预测误差为16 d，7个（9.33%）样本预测误差超过10 d；对于贮藏180 d的茶叶，最大预测误差为19 d，8个（10.67%）样本预测误差超过10 d；对于贮藏240 d的茶叶，最大预测误差为21 d，8个（10.67%）样本预测误差超过10 d；对于贮藏300 d的茶叶，最大预测误差为14 d，6个（8.00%）样本预测误差超过10 d。验证了PCA-BPNN预测模型用于检测黄山毛峰茶贮藏时间的可行性，同时与以原始特征变量作为输入的BPNN预测模型相比，性能更好。

Five grades of Huangshan Maofeng tea samples with six different storage times were analyzed by using an electronic nose. First, the original feature vectors representing the tea odor were acquired, and the first five principal components were extracted as the principal feature vectors. The principal feature vectors were used as the input of back propagation neural network (BPNN) to establish the prediction model for the storage time of Huangshan Maofeng tea (called PCA-BPNN). The test was carried out on 75 tea samples (15 samples of every grade). The results showed that for the tea at zero day of storage, the maximum prediction error was seven days and the prediction error of five samples exceeded ten days (6.67%). For the tea of 60 d of storage, the maximum prediction error was ten days, and the prediction error of four samples exceeded ten days (5.33%). For the tea of 120 d of storage, the maximum prediction error was 16 d and the prediction error of seven samples exceeded ten days (9.33%). For the tea of 180 d of storage, the maximum prediction error was 19 d and the prediction error of eight samples exceeded ten days (10.67%). For the tea of 240 d of storage, the maximum prediction error was 21 d and the prediction error of eight samples exceeded ten days (10.67%). For the tea of 300 d of storage, the maximum prediction error was 14 d and the prediction error of six samples exceeded ten days (8.00%). The feasibility of PCA-BPNN prediction model to determine the storage time of Huangshan Maofeng tea was verified. Moreover, the performance of PCA-BPNN prediction model was better than that of BPNN prediction model using the original feature vectors as the input.

安徽省高等学校省级自然科学研究项目(KJ2013Z195)；安徽省高等学校优秀青年人才基金项目(2012SQRL218)；国家级大学生创新创业训练计划项目(201511305023)