基于可扩展的自表示学习波段选择算法在近红外光谱回归建模中的影响研究  

Effects of Scalable One-pass Self-representation Learning on Near Infrared Spectroscopy Regression Modeling

作　　者：郭拓 梁小娟 马晋芳 袁凯 葛发欢 肖环贤 GUO Tuo;LIANG Xiao-juan;MA Jin-fang;YUAN Kai;GE Fa-huan;XIAO Huan-xian(School of Electrical Information and Artificial Intelligence,Shaanxi University of Science and Technology,Xi'an 710021,China;Department of Electro-Optical Engineering,Jinan University,Guangzhou 510632,China;Nansha Research Institute,Sun Yat-sen University,Guangzhou 511458,China;Jiangxi Poly Pharmaceutical Co.Ltd.,Ganzhou 341900,China)

机构地区：[1]陕西科技大学电子信息与人工智能学院,陕西西安710021 [2]暨南大学光电工程系,广东广州510632 [3]中山大学南沙研究院,广东广州511458 [4]江西保利制药有限公司,江西赣州341900

出　　处：《分析测试学报》2022年第8期1214-1220,共7页Journal of Instrumental Analysis

基　　金：国家自然科学基金重点项目(62031021);陕西省教育厅-科学研究计划项目(20JK0533)。

摘　　要：该文提出了一种基于可扩展的自表示学习(SOP-SRL)波段选择与偏最小二乘(PLS)建模的定量模型分析方法,以安胎丸指标含量阿魏酸、黄芩苷和汉黄芩苷为研究对象,通过SOP-SRL选取代表性波段,采用PLS建立近红外光谱回归模型,并与相关系数法(CC)、正则化自表示学习算法(RSR)和稀疏子空间聚类法(SSC)3种波段选择算法的建模结果进行对比,以校正决定系数(R_(c)^(2))、校正均方根误差(RMSECV)、预测决定系数(R_(p)^(2))和预测均方根误差(RMSEP)为评价标准,对回归模型的预测性能进行评估。结果显示,SOP-SRL在3种数据集上均取得了较好的结果,建模波段从全波长的800分别减少到70、67、87;RMSEP分别从0.0801、6.3495、0.7425下降到0.0653、3.6208、0.4073,分别下降了18%、43%、45%;相应的R_(p)^(2)分别从0.9119、0.8794、0.9158提高到0.9388、0.9526、0.9701,分别提高了3%、8%、6%。结果表明,经SOP-SRL波长选择后模型的预测能力相比于其他几种算法得到显著提升,基于SOP-SRL的PLS模型可以实现安胎丸指标含量的快速检测。Near-infrared spectroscopy is widely applied in the quality monitoring process of tradition⁃al Chinese medicine since it features with rapid detection,and making no damage to the samples and no pollution to the environment in the meantime.In order to realize the rapid prediction of the target ingredients of Antai Pills,a new near-infrared spectroscopy modeling method was proposed,which combines scalable one-pass self-representation learning(SOP-SRL)with partial least-squares(PLS).Taking ferulicacid,baicalin and wogonoside in Antai pills as the research objects,the rep⁃resentative bands selected by SOP-SRL were compared with three band selection algorithms,such as correlation coefficient method(CC),regularized self-representation algorithm(RSR)and sparse subspace clustering(SSC).Then,the quantitative model was established by PLS.The evaluation criteria of the model are root mean squares error of cross validation(RMSECV),corrected determina⁃tion coefficient(R_(c)^(2)),predicted root mean square error(RMSEP)and predicted determination coeffi⁃cient(R_(p)^(2)).Results indicated that the SOP-SRL had good results on all three datasets.Compared with all bands,the selected bands of SOP-SRL were reduced from 800(FULL)to 70,67 and 87,respectively.The RMSEP decreased from 0.0801,6.3495,0.7425 to 0.0653,3.6208,0.4073,decreased by 18%,43%and 45%,respectively.The R_(p)^(2) increased from 0.9119,0.8794,0.9158 to 0.9388,0.9526,0.9701,respectively,increased by 3%,8%and 6%.Therefore,the results of the SOP-SRL algorithm were significantly better than other comparison algorithms.The SOP-SRL algorithm could improve the accuracy of quantative model.The model combining SOP-SRL with PLS could rapidly detect the target ingredients of Antai pills.

关 键 词：近红外光谱 波段选择 可扩展的自表示学习方法(SOP-SRL) 偏最小二乘法(PLS) 指标含量测定 

分 类 号：O657.7[理学—分析化学] R286[理学—化学]