目的 探究人工智能感官与多源信息融合技术用于中药五味药性二分类辨识方法的可行性，为中药药性评价提供新的方法借鉴。方法 选取122种仅含单一味不含兼味的5类代表性中药饮片（源自《中国药典》2020年版）及14种常用的食品类样本，使用PEN3型电子鼻及ASTREE、SA402B型电子舌采集136种样本的智能感官信息，以得到的信息矩阵作为自变量（X），药典项下的性味描述等作为标杆信息（Y），利用主成分分析-判别分析（principal component analysis-discriminant analysis，PCA-DA）、最小二乘-支持向量机（least squares-support vector machine，LS-SVM）2种化学计量学方法，分别基于单一型号智能感官设备（单源）和多智能感官信息融合（多源）建立五味二分类（酸/非酸、咸/非咸、辛/非辛、甘/非甘、苦/非苦）辨识模型Y＝F(X)，以交互验证的正判率作为模型优选指标。结果 经留一法交互验证，基于单源信息的五味二分类模型中最大正判率分别是98.53%（ASTREE/PCA-DA、LS-SVM）、97.06%（ASTREE/PCA-DA）、84.56%（ASTREE/LS-SVM）、89.71%（ASTREE/LS-SVM）、84.56%（ASTREE/LS-SVM），基于多源信息的五味二分类辨识模型的最大正判率分别是99.26%［(ASTREE＋SA402B)/PCA-DA］、99.26%［(ASTREE＋SA402B、PEN3＋ASTREE＋SA402B)/PCA-DA］、88.97%［(PEN3＋ASTREE＋SA402B)/LS-SVM］、91.91%［(PEN3＋ASTREE＋SA402B)/PCA-DA］、86.76%［(ASTREE＋SA402B)/PCA-DA］，多源信息融合后模型的正判率有所提高，平均提高2.35%（P＜0.01）。结论ASTREE型电子舌对于中药五味二分类辨识方面表现良好；多源信息融合后模型的正判率较单独使用任一电子鼻或电子舌均有所提高，为中药五味药性评价提供一定参考。

Objective To explore the feasibility of artificial intelligence sensory and multi-source information fusion technology for the binary classification identification method of five flavors of traditional Chinese medicine (TCM), and provide a new method for the evaluation of drug properties of TCM. Methods A total of 122 representative TCM decoction pieces with only one flavor (derived from the 2020 edition of Chinese Pharmacopoeia) and 14 commonly used food samples were selected. The intelligent sensory information of 136 samples was collected by PEN3 electronic nose, ASTREE electronic tongue and SA402B electronic tongue. The obtained information matrix was used as the independent variable X, and the description of the nature and flavor under the items in pharmacopoeia was used as the benchmark information Y. Two chemometric methods, principal component analysis-discriminant analysis (PCA-DA) and least squares support vector machine (LS-SVM), were used. Based on a single type of intelligent sensory equipment (single source) and multi-intelligent sensory information fusion (multi-source), the five flavors two classification (acid/non-acid, salty/non-salty, spicy/non-spicy, sweet/non-sweet, bitter/non-bitter) identification model Y = F(X) was established, and the positive rate of cross-validation was used as the model optimization index. Results Through the leave-one-out cross-validation method, the maximum correct judgment rates in the five-flavor two-classification model based on single-source information were 98.53% (ASTREE/PCA-DA, LS-SVM), 97.06% (ASTREE/PCA-DA), 84.56% (ASTREE/LS-SVM), 89.71% (ASTREE/LS-SVM), 84.56% (ASTREE/LS-SVM). The maximum correct rate of the two-classification identification model based on multi-source information was 99.26% [(ASTREE + SA402B)/PCA-DA], 99.26% [(ASTREE + SA402B, PEN3 + ASTREE + SA402B)/PCA-DA], 88.97% [(PEN3 + ASTREE + SA402B)/LS-SVM], 91.91% [(PEN3 + ASTREE + SA402B)/PCA-DA], 86.76% [(ASTREE + SA402B)/PCA-DA], respectively. The correct rate of the model after multi-source information fusion was improved, with an average increase of 2.35% (P < 0.01). Conclusion ASTREE electronic tongue performs well in the identification of five flavors and two classifications of TCM. The correct judgment rate of the model after multi-source information fusion is higher than that of any single electronic nose or electronic tongue, which provides a reference for the evaluation of five flavors of TCM.

R283.6

国家自然科学基金资助项目（81774452）；国家自然科学基金资助项目（81001646）；河南省中医药拔尖人才培养项目资助（2019ZYBJ07）；国家重点研发计划中医药现代化重点专项（2017YFC1703402）；河南省高层次人才特殊支持“中原千人计划”—“中原青年拔尖人才”项目（ZYQR201912158）；河南省卫生健康中青年学科带头人专项（HNSWJW-2020014）；国家中医临床研究基地科研专项（2021JDZY106）；河南省中医科学研究基地专项（2021JDZY104）；河南中医药大学2022年度研究生科研创新能力提升计划项目（2022KYCX027）；河南省科技攻关项目（222102310377）