摘要
针对目前烟叶加料工序中糖料液加料效果无法进行无损检测的问题,基于高光谱成像及机器学习构建了烟叶糖料液施加量判别模型。首先,利用可见光-短波红外高光谱成像系统获取不同糖料液施加量烟叶样品的高光谱数据,采用标准正态变换(SNV)进行预处理;然后,分别使用全频域数据和主成分分析(PCA)降维数据,结合支持向量机(SVM)、逻辑回归(LR)、多层感知机(MLP)、随机森林(RF)构建4种判别模型并进行验证。结果表明:SNV预处理显著增强了高光谱数据的特征集中度;在使用全频域波段数据建模时,短波红外波段内模型的预测准确率均显著高于可见光波段,短红外波段的LR模型准确率最高(为98.23%);相较于全频域数据建模,使用PCA降维后的前10个主成分数据建模时,短红外波段的模型预测准确率无显著变化,而可见光波段的RF模型预测准确率提升较为明显(达71.43%);在可见光波段内,PCA降维后4种判别模型的最高准确率对应的主成分数量分别为217个、55个、47个、59个,在短波红外波段内,则分别为13个、11个、117个、46个。整体上,LR和RF模型表现出较优异的预测性能,在短波红外波段内,基于PCA降维数据的LR模型在使用较少主成分时仍能获得高准确率,具有快速、无损、精准地判别烟叶糖料液施加量的能力。
To address the challenge of non-destructive detection of sucrose solution application in the tobacco leaf processing stage,a discrimination model for sucrose solution application based on hyperspectral imaging and machine learning had been developed.Hyperspectral data of tobacco leaf samples with varying sucrose solution applications were first acquired using a visible-shortwave infrared hyperspectral imaging system and preprocessed with standard normal variate(SNV).Four discrimination models for sucrose solution application were then constructed and validated using full-spectrum data and principal component analysis(PCA)reduced data,in conjunction with support vector machine(SVM),logistic regression(LR),multilayer perceptron(MLP),and random forest(RF).The results showed that SNV preprocessing significantly enhanced the feature concentration of the hyperspectral data.When modeling with full-spectrum data,the models in the shortwave infrared band demonstrated significantly higher prediction accuracy compared to those in the visible light band,with the LR model in the shortwave infrared band achieving the highest accuracy of 98.23%.Compared to full-spectrum data modeling,the prediction accuracy of models using the top 10 principal components from PCA reduced data showed little change in the shortwave infrared band,while the RF model′s accuracy in the visible light band improved significantly to 71.43%.In the visible light band,the highest accuracy for PCA-reduced data models corresponded to 217,55,47,and 59 principal components,while in the shortwave infrared band,the numbers were 13,11,117,and 46,respectively.Overall,LR and RF models exhibited superior predictive perf ormance,with the LR model based on PCA-reduced data in the shortwave infrared band maintaining high accuracy with fewer principal components,demonstrating the capability for rapid,non-destructive,and precise determination of sucrose solution application on tobacco leaves.
作者
张建栋
杨忠泮
吴恋恋
徐大勇
朱萍
张雯晶
堵劲松
ZHANG Jiandong;YANG Zhongpan;WU Lianlian;XU Dayong;ZHU Ping;ZHANG Wenjing;DU Jinsong(Technology R&D Center,Gansu Tobacco Industrial Co.,Ltd.,Lanzhou 730050,China;School of Mechanical and Electrical Engineering,Lanzhou University of Technology,Lanzhou 730050,China;Zhengzhou Tobacco Research Institute of CNTC,Zhengzhou 450001,China)
出处
《轻工学报》
CAS
北大核心
2024年第5期86-94,共9页
Journal of Light Industry
基金
中国烟草实业发展中心科技项目计划“青年人才”项目(ZYSYQ-2023-09)
甘肃烟草工业有限责任公司科技项目(KJXM-2023-09)
烟草行业烟草工艺重点实验室引领计划专项科技重点项目(202023AWCX01)。
关键词
高光谱成像
机器学习
烟叶加料工序
糖料液施加量
逻辑回归
hyperspectral imaging
machine learning
tobacco leaf processing stage
sucrose solution application
logistic regression