摘要:
【目的】解决烟叶感官质量评价中存在的主观性强、数据获取困难等问题,实现基于数字化分析对烟叶感官质量的精准定量预测。【方法】以河南、湖南、云南和贵州4个典型风格产区的264份烟叶为研究对象,开展化学成分检测与感官质量评价,经化学指标间相关性分析剔除冗余后,采用基于随机森林的交叉验证递归特征消除方法(RFECV-RF)对各感官指标筛选最优特征子集,再分别采用极端梯度提升(XGBoost)、分类梯度提升(CatBoost)和轻量级梯度提升机(LightGBM)3种经典梯度提升(Boosting)算法,经五折交叉验证优化超参数后建立9项感官指标预测模型。【结果】1)化学指标相关性分析剔除总糖、糖碱比、钾氯比和棕榈酸4项化学成分指标,保留总糖、还原糖、烟碱等25项化学成分指标用于后续建模。2) RFECV-RF特征筛选选取各感官指标对应的最佳特征组,明确总氮、还原糖、钾和烟碱是影响烟叶感官质量的关键化学成分,其交叉验证的均方根误差(RMSE)均低于全特征集模型,有效降低模型复杂度,显著提升预测精度。3)最优算法下各感官指标决定系数(R2)为0.711 3~0.894 0,RMSE为0.084 5~0.140 4,平均绝对百分比误差(MAPE)为1.06%~1.70%,均取得了良好且稳定的预测效果。【结论】本文预测模型框架可实现烟叶感官质量高精度的量化预测,为卷烟产品数字化配方设计与品质控制提供了参考。
Abstract:
【Objective】 To address the challenges of subjectivity and data acquisition difficulties in tobacco leaf sensory quality evaluation, and to achieve precise quantitative prediction of tobacco leaf sensory quality based on digital analysis. 【Methods】 A total of 264 tobacco leaf samples from four typical producing regions in China—Henan, Hunan, Yunnan, and Guizhou—were selected for chemical composition analysis and sensory quality evaluation. After removing redundant variables through correlation analysis of chemical indicators, RFECV-RF was used to select the optimal feature subsets for each sensory attribute. Three classic boosting algorithms—XGBoost, CatBoost, and LightGBM—were applied, with hyperparameters optimized using five-fold cross-validation within the Optuna framework to build prediction models for nine sensory attributes. 【Results】 1) Correlation analysis of chemical indices removed four chemical constituent indices, namely total sugar, sugar-to-nicotine ratio, potassium-to-chlorine ratio, and palmitic acid, and retained 25 chemical composition indices, including reducing sugar and nicotine, for subsequent modeling. 2) RFECV-RF feature selection identified the optimal feature subset for each sensory attribute, and further demonstrated that total nitrogen, reducing sugar, potassium, and nicotine were the key chemical constituents affecting tobacco leaf sensory quality. Except for “impact”, the root mean square error ( RMSE) obtained by cross-validation was lower than that of the full-feature model, indicating that feature selection effectively reduced model complexity and improved prediction accuracy. 3) Under the optimal algorithm, the coefficient of determination (R2) for the sensory attributes ranged from 0. 711 3 to 0. 894 0, the RMSE ranged from 0. 084 5 to 0. 140 4, and the mean absolute percentage error (MAPE) ranged from 1. 06% to 1. 70%, indicating good and stable predictive performance. 【Conclusion】 The prediction model framework enables high-precision quantification of tobacco leaf sensory quality. The research result provide aa reference for the digital formulation design and quality control of cigarette products.
点击查看大图
下载:
