基于1D-CNN的近红外光谱分类算法研究  

Research on A Classification Algorithm of Near-Infrared Spectroscopy Based on 1D-CNN

作　　者：蒲姗姗 郑恩让[1] 陈蓓[1] PU Shan-shan;ZHENG En-rang;CHEN Bei(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,陕西西安710021

出　　处：《光谱学与光谱分析》2023年第8期2446-2451,共6页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(31670596)资助。

摘　　要：近红外光谱(NIR)技术应用广泛,但其建模精度容易受到光谱预处理算法的影响,传统近红外光谱分析过程中,预处理方法的选取主要靠人为经验,有时会遗漏一些光谱特征。提出了一种无需光谱预处理的一维卷积神经网络(1D-CNN)近红外光谱分类方法。为了比较BP神经网络(BP)、支持向量机(SVM)、极限学习机(ELM)三种传统近红外光谱分析模型与一维卷积神经网络(1D-CNN)建模方法的分类效果,在不同等级药品、啤酒和不同种类芒果、葡萄NIR数据集中进行了对比实验,实验结果显示采用1D-CNN模型的分类准确率最高,其中药品4分类、啤酒2分类、芒果10分类、葡萄19分类的准确率分别为96.77%、93.75%、96.45%、88.75%。最后讨论了均值中心化(MC)、标准化、多元散射校正(MSC)、标准正态变量变换(SNV)、一阶差分、二阶差分、小波变换(WT)7种光谱预处理方法对不同模型的影响。经过预处理后,BP神经网络、SVM和ELM的分类准确率有明显的变化,而1D-CNN模型的分类效果在预处理前和预处理后基本无变化,且分类准确率依旧最高。结果表明,相比传统近红外光谱分类方法,所提出的1D-CNN方法可实现对食品和药品NIR快速准确分类,不需要任何的光谱预处理,说明深度学习方法在近红外光谱处理领域中具有广阔的应用前景和研究价值。Near-infrared(NIR)spectroscopy technology has been widely used in many fields,but spectral pretreatment algorithm easily affects its modelling accuracy.In traditional near-infrared spectroscopy analysis,the selection of pretreatment methods mainly depends on human experience,and sometimes some spectral information will be lost.Therefore,a near-infrared spectrum classification method of a one-dimensional convolution neural network(1D-CNN)without spectral pretreatment is proposed in this paper.In order to compare the classification effects of three traditional near-infrared spectral analysis models of BP neural network(BP),support vector machines(SVM)and extreme learning machine(ELM)with one-dimensional convolutional neural network(1D-CNN)modeling method,comparative experiments were carried out on NIR data sets of different grades of drug,beer,mango and grape.The experimental results show that the classification accuracy of the 1D-CNN model is the highest,among which the accuracy of drug 4 classification is 96.77%,beer 2 classification is 93.75%,mango 10 classification is 96.45%,and grape 19 classification is 88.75%.Finally,the effects of seven different spectral pretreatment methods,such as mean centralization(MC),standardization,multiple scatter correction(MSC),standard normal variable transformation(SNV),first-order difference,second-order difference and wavelet transform(WT)on different models are also discussed.After pretreatment,the classification accuracy of the BP neural network,SVM and ELM changes significantly,while the classification effect of the 1D-CNN model has no change before and after pretreatment,and the classification accuracy is still the highest.The results show that compared with the traditional NIR spectral classification methods,the 1D-CNN method proposed in this paper can realize the rapid and accurate NIR classification of food and drugs and does not need any spectral pretreatment.It shows that the deep learning method has broad application prospects and research value in NIR spectral proc

关 键 词：近红外光谱 预处理 一维卷积神经网络 分类 

分 类 号：O657.3[理学—分析化学]