卷积神经网络和支持向量机算法在塑料近红外光谱分类中的模型应用  

Convolutional Neural Network and Support Vector Machine Models for Plastic Classification by Near-infrared Spectroscopy

作　　者：张文杰 焦安然 田静 王晓娟 王斌 徐晓轩 ZHANG Wen-jie;JIAO An-ran;TIAN Jing;WANG Xiao-juan;WANG Bin;XU Xiao-xuan(The Key Laboratory of Weak-Light Nonlinear Photonics,Ministry of Education,School of Physics,Nankai University,Tianjin 300071,China;School of Food and Biological Engineering,Jiangsu University,Zhenjiang 212013,China;Ningbo Customs Technology Center,Ningbo 315048,China)

机构地区：[1]南开大学物理科学学院弱光非线性光子学教育部重点实验室,天津300071 [2]江苏大学食品与生物工程学院,江苏镇江212013 [3]宁波海关技术中心,浙江宁波315048

出　　处：《分析测试学报》2021年第7期1062-1067,共6页Journal of Instrumental Analysis

基　　金：天津市科技支撑计划重点项目(15ZCZDGX00780);国家重点研发计划(2016YFC0400709);江苏省高层次创新创业人才引进计划“2019双创博士”&“2020双创人才”。

摘　　要：机器学习算法的应用使得塑料自动分类成为可能,而废旧塑料的分类回收对保护环境、节约资源有重要意义。该文结合近红外光谱分析技术,比较了使用一维卷积神经网络(1D CNN)和多元散射处理后支持向量机算法(MSC-SVM)建模的效果,及对PP新生料、PP再生料、PE新生料、PE再生料4种塑料分类的准确率。基于100个塑料样本近红外光谱数据的分类结果表明,验证集上1D CNN模型准确率为91.5%,MSCSVM模型准确率为90.8%。1D CNN模型用于识别PP和PE新生料时,准确率可达100%。证明1D CNN建模方法在小数据集上进行准确塑料分类是可行的。Nowadays it is possible to automatically classify plastic waste by machine learning algorithms,which is of great significance for protecting the natural environment and saving resources.To establish better plastic classification models,the performances of multiplicative scatter correctionsupport vector machines(MSC-SVM)model and one-dimensional convolutional neural network(1D CNN)model were compared in identifying 4 types of plastic in this paper,as well as the accuracies of NIRS technique for classifying PP new raw material,PP recycled material,PE new raw material and PE recycled material,respectively.Based on the spectra data of 100 plastic samples,the experiment results showed that in validation set,the accuracy for MSC-SVM model is 90.8%while that for 1D CNN model is 91.5%.Particularly,1D CNN model provided excellent classification results in identifying PE and PP new raw material samples with the accuracies reached up to 100%,which indicated that 1D CNN model is efficient to classify different types of plastic on small dataset.

关 键 词：近红外光谱 卷积神经网络 支持向量机 塑料分类 

分 类 号：O657.3[理学—分析化学]