近红外光谱结合极限学习机用于红参产地鉴别研究

Study on the Origin Identification of Talinum paniculatum by Using Near Infrared Spectroscopy Combined with Extreme Learning Machine

中药是中华民族的文化瑰宝,其质量的鉴别研究与中药现代化进程关系密切,具有深远意义。本文利用近红外漫反射光谱技术结合主成分分析(PCA)、K-近邻算法(KNN)、极限学习机(ELM)方法建立预测模型,对7种不同来源的红参药材样品进行产地鉴别研究,以期为其他中药材的快速、准确鉴别和质量控制提供参考。本文利用NIRS采集来自不同产地的红参光谱信息(900~2200 nm),通过光谱预处理手段消除噪声和样品组内差异的影响,进而建立主成分分析、K-近邻算法(KNN)、极限学习机预测模型,对不同来源的红参进行产地鉴别。结果表明:7种红参(1种高丽红参+6种国产红参)在主成分分析模型中表现出了较为明显的分类聚集特征,但仍存在误判的可能;在K-近邻算法(KNN)模型中,当K值为5时,具有较强的识别能力,准确率达到96%;相比KNN模型,极限学习机模型中的21个红参样本产地识别正确率达100%。近红外漫反射光谱技术结合ELM方法建立的预测模型能够快速、准确、无损地鉴别不同产地的红参,可以解决传统检测方法很难克服的中药材鉴别规模庞大、分类多、建模速度慢等问题,且使用方便,具有很高的推广价值。

Traditional Chinese medicine is the cultural treasure of the Chinese nation.The study on its quality identification is related to the modernization of traditional Chinese medicine closely.This paper aimed to establish a prediction model to identify Talinum paniculatum from 7 different sources,through Near Infrared Reflectance spectroscopy(NIRS)combined with Principal Component Analysis(PCA),K-nearest neighbor algorithm model(KNN)and Extreme Learning Machine Method(ELM),which provided a reference for the rapid,accurate identification and quality control of Chinese medicinal materials.The results showed that seven types of T.paniculatum(one Korean ginseng and six kinds of domestic radix ginseng Rubra)showed obvious classification clustering characteristics in the principal component analysis,but there was still the possibility of miscalculation.In the K-Nearest Neighbor algorithm(KNN)model,when the K value was equal to 5,it has strong recognition ability,with the accuracy rate of 96%.Compared with the KNN,the accuracy of origin identification of 21 T.paniculatum prediction samples in the extreme learning machine model reached 100%.In words,near-infrared diffuse reflectance spectroscopy combined with the ELM method could quickly,accurately and non-destructively identify red ginseng from different origins.Meanwhile,it can solve the problems of large-scale,multi-classification and slow modeling,that are difficult to overcome by traditional detection methods.This method is easy to use,has the very high popularization value.