基于多尺度扩张卷积神经网络的近红外光谱定量分析模型研究  

Quantitative Analysis Model of Near Infrared Spectroscopy Based on Multi-Scale Dilated Convolutional Neural Networks

作　　者：李强[1] 陈蓓[1] 张芳[1] LI Qiang;CHEN Bei;ZHANG Fang(School of Electrical and Control Engineering,Shaanxi University of Science&Technology,Xi′an 710021,China)

机构地区：[1]陕西科技大学电气与控制工程学院,西安710021

出　　处：《分析化学》2025年第3期451-463,共13页Chinese Journal of Analytical Chemistry

基　　金：国家自然科学基金项目(No.62203285)资助。

摘　　要：近红外光谱(Near infrared spectroscopy,NIRS)技术广泛应用于医药、食品和化工行业的定量分析中。本研究提出了一种基于多尺度扩张卷积神经网络的近红外光谱定量分析模型(Multi-scale dilated convolutional spectral network,MDCSpecNet),该模型由一维卷积层、批归一化层、最大池化层、多尺度扩张卷积神经网络和全连接层组成,其中,一维卷积层和最大池化层对原始光谱做初步特征提取和特征降维,批归一化层加快模型收敛,多尺度扩张卷积神经网络对光谱特征进行提取与融合,全连接层对特征信息进行线性表示,增加模型的预测精度和泛化能力。利用公开的药品、谷物、小麦、牛奶、汽油与三聚氰胺的近红外光谱数据集建立MDCSpecNet预测模型,并与一维卷积神经网络(One dimensional convolution neural network,1D-CNN)、偏最小二乘法(Partial least squares,PLS)、支持向量机(Support vector regression,SVR)和极限学习机(Extreme learning machine,ELM)建模方法的预测结果进行对比分析。结果表明,在药品活性成分(Active pharmaceutical ingredient,API)含量、谷物葡萄糖含量、谷物乳酸盐含量、谷物水分含量、小麦蛋白质含量、汽油辛烷值以及三聚氰胺浊点预测中,相较于其它4种建模方法,MDCSpecNet模型的精度分别提升了16.0%、36.7%、25.1%、22.6%、34.2%、15.2%、22.6%(1D-CNN),46.9%、66.7%、73.2%、65.8%、16.6%、15.9%、13.7%(PLS),68.1%、70.6%、81.7%、73.9%、69.2%、77.9%、56.0%(SVR)和62.0%、20.4%、48.9%、85.6%、50.4%、13.0%、44.6%(ELM)。基于多尺度扩张卷积神经网络的MDCSpecNet模型解决了传统近红外光谱建模方法精度低和泛化能力差等问题,利用MDCSpecNet模型进行多种物质的近红外光谱定量分析是可行的。Near infrared spectroscopy(NIRS)technology has been widely applied in quantitative analysis of pharmaceuticals,food,and chemical industries.In this study,a NIRS quantitative analysis model(MDCSpecNet)based on a multi-scale dilated convolutional neural network was proposed.The model consisted of a onedimensional convolutional layer,a batch normalization layer,a max-pooling layer,a multi-scale dilated convolutional neural network,and a full-connected layer.Among which,the one-dimensional convolutional layer and the max-pooling layer performed preliminary featured extraction and dimensionality reduction on the original spectra,the batch normalization layer accelerated the convergence of the model,the multi-scale dilated convolutional neural network extracted and fused spectral features,and the fully-connected layer linearly represented the feature information,enhancing the model′s prediction accuracy and generalization ability.MDCSpecNet prediction models were established using publicly available NIRS datasets of pharmaceuticals,grains,wheat,milk,and gasoline.The prediction resultswere compared and analyzed with those of onedimensional convolution neural network(1D-CNN),partial least squares(PLS),support vector regression(SVR),and extreme learning machine regression(ELM)modeling methods.The results showed that,in prediction of the content of active pharmaceutical ingredient(API)in pharmaceuticals,the glucose content in grains,the lactate content in grains,the moisture content in grains,the protein content in wheat,the octane number in gasoline and the cloud point of melamine,the accuracy of the MDCSpecNet model increased by 16%,36.7%,25.1%,22.6%,34.2%,15.2%and 22.6% compared to 1D-CNN,46.9%,66.7%,73.2%,65.8%,16.6%,15.9% and 13.7% compared to PLS,68.1%,70.6%,81.7%,73.9%,69.2%,77.9% and 56% compared to SVR,and 62%,20.4%,48.9%,85.6%,50.4%,13%and 44.6% compared to ELM,respectively.The MDCSpecNet model based on the multi-scale dilated convolutional neural network addressed the issues of low accuracy and poor generaliza

关 键 词：近红外光谱 定量分析 扩张卷积 多尺度特征融合 

分 类 号：O657.3[理学—分析化学]