基于卷烟常规化学检测数据及近红外信息的卷烟牌号识别

发布时间：2018-09-04 09:37

【摘要】：本文以不同牌号卷烟的近红外光谱和常规理化化学检测指标数据为基础,提出了基于逆序法的近红外数据稳定性检验和基于信息增益的特征抽提方法。分别以卷烟的9个常规检测化学指标、基于经验选择的6个常规化学特征、近红外光谱的主成分得分及5个抽提的主成分、12个抽提的近红外光谱与化学指标的综合指标为卷烟牌号分类特征,采用(改进)与简化的KNN模式识别方法建立了一系列卷烟牌号判别模型。以272个取自6个不同牌号的卷烟为样本,对每个判别模型均随机选取4组检验集,采用4组检验集的平均预测正确率考察模型的预测性能。根据模型的结果确定影响成品卷烟品牌风格的重要特征,为卷烟品牌维护、配方设计与维护提供理论依据与指导。结果表明无论是对两个牌号卷烟的两类分类还是对三个牌号卷烟的三类分类及六个牌号卷烟的六类分类问题,以前16个近红外光谱主成分及12个抽提的综合特征为分类特征所建立的判别模型均可给出92.65%～98.86%的平均预测正确率；9个化学指标与6个常规化学指标为分类特征时,判别模型对两个牌号及三个牌号卷烟的平均预测正确率也能达到95%以上,对6个牌号卷烟的平均预测正确率为83.09%～84.56%。基于近红外光谱对卷烟牌号识别的准确率要高于基于常规化学的识别准确率,模式识别模型预测准确率随样本类别复杂程度提高而降低。综上,近红外信息可以很好反映不同品牌卷烟的主要风格特征,根据近红外光谱可对卷烟牌号进行准确鉴别。提示可利用近红外信息进行成品卷烟品牌管理、叶组配方设计与维护。
[Abstract]:Based on the NIR spectra of different brands of cigarettes and the conventional physicochemical detection index data, this paper presents a method based on reverse sequence method for the stability test of NIR data and the feature extraction method based on information gain. According to the 9 routine chemical indexes of cigarette and the 6 conventional chemical characteristics based on experience, The principal component scores of NIR spectra and 5 principal components extracted from NIR spectra, and the comprehensive indexes of NIR spectra and chemical indexes of 12 extracted cigarettes were classified as cigarette grades. A series of cigarette brand discrimination models were established by using (improved) and simplified KNN pattern recognition method. Based on 272 cigarettes of 6 different grades, 4 groups of test sets were randomly selected for each discriminant model, and the prediction performance of the model was investigated with the average prediction accuracy of the four groups of test sets. According to the results of the model, the important characteristics affecting the brand style of finished cigarette are determined, which provides theoretical basis and guidance for cigarette brand maintenance, formula design and maintenance. The results show that the classification of two types of cigarettes, three categories of three brands and six categories of six kinds of cigarettes are not only related to the classification of two brands of cigarettes, but also to the classification of six types of cigarettes. The discriminant models based on 16 principal components of NIR spectra and 12 extracted principal components are classified features, and the average prediction accuracy is 92.65% or 98.86%, and when 9 chemical indexes and 6 conventional chemical indexes are classified characteristics, the average prediction accuracy is 92.65% or 98.86%, while 9 chemical indexes and 6 conventional chemical indexes are classified characteristics. The average prediction accuracy of the discriminant model for two brands and three grades of cigarettes was over 95%, and the average accuracy rate for six brands of cigarettes was 83.09 and 84.56. The accuracy rate of cigarette brand recognition based on near infrared spectrum is higher than that based on conventional chemistry, and the prediction accuracy of pattern recognition model decreases with the increase of sample class complexity. In conclusion, NIR information can well reflect the main style characteristics of different brands of cigarettes, according to NIR spectrum can accurately identify cigarette brands. It is suggested that brand management, leaf formulation design and maintenance can be carried out by using near infrared information.
【学位授予单位】：华东理工大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TS47

【参考文献】