基于近红外光谱和变分自编码建模鉴别多类药物  

Identifying Multi-Class Drugs by Using Near-Infrared Spectroscopy and Variational Auto-Encoding Modeling

作　　者：郑安兵[1] 杨辉华[1,2] 潘细朋 尹利辉 冯艳春[3] ZHENG An-bing;YANG Hui-hua;PAN Xi-peng;YIN Li-hui;FENG Yan-chun(School of Automation,Beijing University of Posts and Telecommunications,Beijing 100876,China;School of Computer and Information Security,Guilin University of Electronic Science and Technology,Guilin 541004,China;China Institute for Food and Drug Control,Beijing 100050,China)

机构地区：[1]北京邮电大学自动化学院,北京100876 [2]桂林电子技术大学计算机科学与信息安全学院,广西桂林541004 [3]中国食品药品检定研究院,北京100050

出　　处：《光谱学与光谱分析》2020年第12期3946-3952,共7页Spectroscopy and Spectral Analysis

基　　金：National Natural Science Foundation of China (21365008,61906050);Guangxi Technology R&D Program (2018AD11018)

摘　　要：不同厂商(品牌)的药品仍存在一定的差异,价格不同,有销售商家用低廉药物产品换上假的大品牌包装在市场上高价销售。无专利药品或无生产、销售(如走私进口药)许可资质的药品也有可能贴上伪造的正规品牌包装在市场上出售。这些药品逃避药物监管和审批程序,损害消费者利益并给整个药物市场带来重大危害。因此,准确鉴别不同来源的药品在药品质量监管中具有重要意义。近红外光谱分析(NIR)具有仪器成本低、可直接测量、可无损检测、可现场检测等优点,特别适合药品的快速建模分析。采用近红外光谱直接鉴别出多个厂商、品种的药品,有重要应用价值同时又存在重大技术挑战,主要体现在需要有效的提取特征器和合适的分类器。自编码是深度学习方法中一个重要分支,它主要用于数据的非线性降维特征提取。变分自编码(VAE)是近年来最为流行的自编码算法,它通过变分法学习输入数据的一族欠完备的单变量正态分布特征,用以表示盲源因素对数据施加的影响,具有较强的特征提取能力,广泛应用于计算机视觉、语音识别等领域,在NIR分析方面未见报道。基于VAE,充分利用VAE既是特征提取器,又是数据生成器的优点,通过特殊设计的人工神经网络结构和损失函数,构建面向多品种、多厂商药品NIR分类模型。以29个厂商生产的4种药品(盐酸二甲双胍片,盐酸氯丙嗪片,马来酸氯苯那敏片,头孢呋辛酯片)的1 721个样本为实验对象,建立药品的多品种、多厂商分类鉴别实验。对比SVM, BP-ANN, PLS-DA等传统化学计量学算法及稀疏自编码(SAE)、深度信念网络(DBN)、深度卷积网络(CNN)等深度学习算法,其分类性能优良,同时具有良好的鲁棒性和可扩展性。With the expansion of online pharmacies,more and more counterfeit drugs without drug patents or licenses will appear in the markets with forged brand packaging.It is inevitable that the low-cost drug products will be sold at a high price if there are no methods to identify the source.These drugs evade drug supervision and approval procedures,harm the interests of consumers and bring great risks to the whole drug market.Near infrared spectroscopy(NIR)has the advantages of low cost,direct measurement,non-destructive testing and on-site testing.It is especially suitable for the rapid modeling and analysis of drugs in the condition that there are effective feature extraction and appropriate classifiers.Meanwhile,Auto-encoding is an important branch of deep learning method,which is mainly used for extracting non-linear dimensional reduction feature of data,and Variational Auto-encoding(VAE)is the most popular Auto-encoding algorithm in recent years,it has strong feature extraction ability and is widely used in computer vision,speech recognition and other fields,yet there is no report on the NIR analysis.Based on VAE,through a specially designed artificial neural network structure and loss function,this paper constructs NIR classification model for multi-category and multi-manufacturer drugs.Four kinds of drugs(metformin hydrochloride tablets,chlorpromazine hydrochloride tablets,chlorphenamine maleate tablets,cefuroxime ester tablets)produced by 29 manufacturers were used as the experimental objects to establish the multi-class classification and identification experiments.Compared with SVM,BP-ANN,PLS-DA and sparse Auto-coding(SAE),deep belief network(DBN),deep convolution network(CNN),etc.,the algorithm has excellent classification performance,good robustness and scalability.

关 键 词：近红外光谱 药品鉴别 多分类 深度学习VAE 

分 类 号：O434.3[机械工程—光学工程]