基于高光谱成像结合机器学习的半夏饮片鉴别  

Identification of Pinelliae Rhizoma Decoction Pieces by Hyperspectral Imaging Combined With Machine Learning

作　　者：李若彤 胡会强 曹诗宇 卢孟瑶 刘梦然 付嘉玥 毛晓波[2] 王海波 符玲[1,3] LI Ruo-tong;HU Hui-qiang;CAO Shi-yu;LU Meng-yao;LIU Meng-ran;FU Jia-yue;MAO Xiao-bo;WANG Hai-bo;FU Ling(College of Pharmacy,Zhengzhou University,Zhengzhou 450000,China;College of Electrical and Information Engineering,Zhengzhou University,Zhengzhou 450000,China;NMPA Key Laboratory for Quality Control of Traditional Chinese Medicine,Henan Institute for Drug and Medical Device Inspection,Zhengzhou 450018,China)

机构地区：[1]郑州大学药学院,河南郑州450000 [2]郑州大学电气与信息工程学院,河南郑州450000 [3]河南省药品医疗器械检验院,国家药品监督管理局中药材及饮片质量控制重点实验室,河南郑州450018

出　　处：《光谱学与光谱分析》2025年第5期1236-1242,共7页Spectroscopy and Spectral Analysis

基　　金：中央本级重大增减支项目(2060302-2101-26);国家药品监督管理局重点实验室开放课题(KF202104);河南省自然科学基金项目(232300421173);河南省省级科技研发计划联合基金(优势学科培育类)项目(232301420096)资助。

摘　　要：《中国药典》2020版关于“半夏”项下收载有:生半夏、清半夏、姜半夏和法半夏四种饮片。由于它们的外形相似、气味特征小,在生产管理、市场流通以及临床使用中极易混淆。常规化学检测受限于仪器试剂,且步骤繁琐,因此非常有必要探索建立一种准确、快速、无损的半夏饮片检测方法。尝试采用高光谱成像技术结合机器学习对四种半夏饮片进行了鉴别。采用主成分分析法(PCA)对高光谱数据进行特征提取,并基于全波段数据模型建立了支持向量机(SVM)、逻辑回归(LR)、多层感知机(MLP)、随机森林(RF)四种分类模型;通过四种分类模型对半夏饮片的训练集准确率和测试集准确率进行了考查,对四种模型在最佳性能下主成分占比进行了分析;此外,还对四种半夏饮片的高光谱数据进行了可视化降维分析(t-SNE)。基于PCA构建的SVM、LR、MLP和RF四种分类模型均能实现对四种半夏饮片的精准鉴别,其测试集精度分别为80.76%、96.45%、96.59%、86.77%,其主成分占比分别为60%、80%、70%、80%。t-SNE可视化降维分析结果说明了姜半夏和清半夏比较接近,和生半夏相比,成分有部分改变;法半夏经过炮制后,化学成分变化较大,和其他三种饮片差异非常明显。四种半夏的平均光谱反射率结果也与此一致。该研究首次将高光谱成像技术结合机器学习应用于半夏饮片预测建模,实现了准确、快速、无损鉴别,为其生产流通、临床正确使用提供了新的鉴别方法和科学依据。There are four kinds of decoction pieces for Pinelliae Rhizoma,namely Pinelliae Rhizoma,Pinelliae Rhizoma praeparatum cum alumine,Pinelliae Rhizoma praeparatum cum zingibere et alumine,and Pinelliae Rhizoma praeparatum,recorded in the current Chinese Pharmacopoeia(2020 edition).Due to their similar appearance and weak odor characteristics,it's easy to confuse their manufacturing management,market circulation,and clinical application.Because of the requirement for instruments,reagents,and intricate detection steps,exploring and establishing an accurate,rapid,and non-destructive detection method for Pinelliae Rhizoma decoction pieces is necessary.This paper used hyperspectral imaging combined with machine learning to identify the four kinds of Pinelliae Rhizom adecoction pieces.The principal component analysis(PCA)was utilized to extract features from the hyperspectral data,and support vector machine(SVM),logistic regression(LR),multi-layer perceptron(MLP),and random forest(RF)classification models were established based on the full-band data model.The accuracy of both training and test sets of four classification models was evaluated,along with an analysis ofthe principal component proportion of the four models under optimal performance.Additionally,the t-distributed stochastic neighbor embedding(t-SNE)visual dimensionality reduction analysiswas conducted on the hyperspectral data of the fourkinds of decoction pieces.The SVM,LR,MLP,and RF classification models based on PCA can achieve accurate identification for PinelliaeRhizoma decoction pieces.The accuracy of the test set is 80.76%,96.45%,96.59%,and 86.77%;in addition,the proportion of principal components is 60%,80%,70%,and 80%,respectively.The t-SNE analysis by dimensionality reduction showed that the components of Pinelliae Rhizoma praeparatum cum zingibere et alumine and Pinelliae Rhizoma praeparatum cum alumine were relatively close and partly changed compared with Pinelliae Rhizoma.However,the chemical composition of Pinelliae Rhizoma praeparatum changed g

关 键 词：高光谱 半夏饮片 机器学习 主成分分析 

分 类 号：O433.4[机械工程—光学工程]