基于多方法优势组合的烟叶配方模块组合分类  

Classification of tobacco leaf formula modules based on a combination of multiple methodological advantages

作　　者：蒋佳磊[1] 廖付 郝贤伟[1] 汤晓东[1] 陈晓水[1] 朱书秀[1] 赵振杰 JIANG Jialei;LIAO Fu;HAO Xianwei;TANG Xiaodong;CHEN Xiaoshui;ZHU Shuxiu;ZHAO Zhenjie(Technology Centre,China Tobacco Zhejiang Industrial Co.,Ltd,Hangzhou 310024,China)

机构地区：[1]浙江中烟工业有限责任公司技术中心,杭州市310024

出　　处：《中国烟草学报》2024年第6期29-38,共10页Acta Tabacaria Sinica

基　　金：浙江中烟揭榜挂帅项目:基于优化烟叶资源使用效率及提升普一类产品竞争力的配方关键技术研究与应用(ZJZY2022A001);“利群”品牌主要烟叶产区烤烟化学成分特征与可用性评价(ZJZY2019C003)。

摘　　要：为深入研究卷烟品牌烟叶配方分组方法和原则,基于457个烟叶样品的12种感官评吸指标,比较了4种判别分析和4种机器学习方法对4种烟叶分类的建模集正确率(R)、验证集正确率(r)和平均正确率(m)的影响,并基于分类方法选择和权重分配构建了一种高精度的组合分类方法。结果表明:(1)与判别分析相比,机器学习的R显著提高,而r显著下降,且LS-SVM的R最高(92.8%),FDA和F-BDA的r最高(80.2%),但m无显著性差异;(2)优化选择M-BDA、FDA、ANN和KNN四种方法,按精度赋权建立的组合分类方法同时提高了R(95.3%)和r(89.0%),且m由低于84%提高到92.2%,并通过理论计算和实际结果验证了组合分类方法的普遍有效性;(3)组合分类方法Kappa系数均大于0.8,方法可靠,一致性程度高,验证集m-F1度量显著提升21.2%,模型泛化能力大为增强;(4)优雅感、杂气、余味、润感和清晰度5项指标对分类起主要作用,符合利群品牌的风格特征;(5)误判样品(6.5%)指标评分与其模块真实类别的不匹配归因于对库存、成本和质量的平衡,基本符合烟叶配方的调整空间。To deeply study grouping methods and principles of cigarette brand tobacco leaf formulas,based on the 457 tobacco samples of 12 kinds of sensory evaluation indexes,this study compared four discriminant analysis and four machine learning methods for modeling set accuracy(R),validation set accuracy(r),and average accuracy(m)across four tobacco leaf classifications.A high-precision composite classification method was constructed based on method selection and weight allocation.Results showed that:(1)Compared with discriminant analysis,machine learning significantly improved R,while r significantly decreased,with LS-SVM having the highest R(92.8%),and FDA and F-BDA having the highest r(80.2%),but there was no significant difference in m;(2)Optimized selection of M-BDA,FDA,ANN,and KNN methods and the composite classification method established by accuracy weighting simultaneously improved R(95.3%)and r(89.0%),and increased m from below 84%to 92.2%,validating the general effectiveness of the composite classification method through theoretical calculations and practical results;(3)The Kappa coefficient of the composite classification method was greater than 0.8,indicating reliability,high consistency,and a significant improvement in validation set m-F1 measure by 21.2%,greatly enhancing the model's generalization ability;(4)Five indicators,namely elegance,off-flavor,aftertaste,moistness,and clarity,played a major role in classification,aligning with the style characteristics of the Liqun brand;(5)Misjudged samples(6.5%)with indicator scores not matching their real module categories were attributed to the balance of stock,cost,and quality,generally conforming to the adjustment space of tobacco leaf formulas.

关 键 词：烟叶 组合分类 判别分析 机器学习 感官评吸 配方模块 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TS452[自动化与计算机技术—控制科学与工程]