基于迁移学习与TrAdaBoost-SVM方法的XRF中药重金属超标研究  

Research on Classification of Heavy Metal Pb in Honeysuckle Based on XRF and Transfer Learning

作　　者：马骞 杨婉琪 李福生 程惠珠 赵彦春 MA Qian;YANG Wan-qi;LI Fu-sheng;CHENG Hui-zhu;ZHAO Yan-chun(Research Center for Intelligent Equipment,School of Automation Engineering,University of Electronic Science and Technology of China,Chengdu 611731,China;Yangtze Delta Region Institute(Huzhou),University of Electronic Science and Technology of China,Huzhou 313001,China)

机构地区：[1]电子科技大学自动化工程学院,四川成都611731 [2]电子科技大学长三角研究院(湖州),浙江湖州313001

出　　处：《光谱学与光谱分析》2023年第9期2729-2733,共5页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(62075028)资助。

摘　　要：中药材重金属超标问题日趋严重,将阻碍中药产业的未来高质量发展,因此研究高效、准确、便捷的超标鉴定方法对于了解中药的安全性具有重要意义。X射线荧光(XRF)光谱分析具有无损检测、快速准确、样品制备方便等优势,在元素分析领域获得广泛应用。由于中药材重金属超标阈值低(如中国药典2020年版规定铅超标为5 mg·kg-1),中药的种类繁多,基体复杂,国家标准样本匮乏,常规的分类算法难以准确鉴定超标问题。将迁移学习与多分类支持向量机(TrAdaBoost-SVM)方法结合,以金银花为例,采用与金银花相似的国家土壤标准样品的光谱特征信息用于数据增强,将土壤标准样品和少量中药样本混合建立迁移学习和支持向量机分类模型。通过实验验证,迁移学习和TrAdaBoost-SVM结合的分类优化方法,与传统SVM、AdaBoost分类算法相比,鉴定重金属元素铅(Pb)的超标准确率有显著提高。通过测试数据集的预测验证,TrAdaBoost-SVM模型的预测准确率为96.7%,高于传统SVM及AdaBoost分类模型的准确率。所提出的迁移学习与TrAdaBoost-SVM结合的方法,可在小样本条件下建立分类模型,并对中药的重金属超标准确预测,具有一定的理论意义和应用价值。The problem of heavy metals exceeding the standard in Chinese medicinal materials is becoming increasingly serious,which will hinder the high-quality development of the Chinese medicine industry in the future.Therefore,research on efficient,accurate and convenient methods for the identification of excessive heavy metals is of great value for understanding the safety of traditional Chinese medicine.X-ray fluorescence spectrometry(XRF)instruments have the advantages of non-destructive testing,fast and accurate,and convenient sample preparation,and are widely used in elemental analysis.Due to the low threshold of heavy metals in traditional Chinese medicinal materials(for example,the 2020 edition of the Chinese Pharmacopoeia stipulates that the lead exceeds the standard at 5 mg·kg^(-1)),there are many types of traditional Chinese medicines,complex matrices,and lack of national standard samples.Conventional classification algorithms are difficult to identify excessive problems accurately.This paper combines transfer learning with a multi-class support vector machine(TrAdaBoost SVM)method.The spectral feature information of national soil standard samples similar to honeysuckle is used for data enhancement,and the standard soil sample and a small amount of traditional Chinese medicine samples are mixed with establish Transfer learning and support vector machine classification models.Through the experimental verification,the classification optimization method combining transfer learning and TrAdaBoost-SVM,compared with the traditional SVM and AdaBoost classification algorithm,the accuracy rate of identifying the heavy metal element lead(Pb)exceeding the standard has been significantly improved.Through the prediction verification of the test dataset,the prediction accuracy of the TrAdaBoost-SVM model is 96.7%,which is higher than that of the traditional SVM and AdaBoost classification models.The method of combining transfer learning and TrAdaBoost-SVM proposed in this paper can establish a classification model under the

关 键 词：X荧光光谱分析技术 迁移学习 支持向量机 中药重金属超标分类 

分 类 号：O434.13[机械工程—光学工程]