基于烟叶外观特征与打叶工艺参数的片烟叶片结构预测模型研究

Prediction Model of Leaf Structures of Tobacco Strips Based on the Appearance Features of Tobacco Leaves and Threshing Technical Parameters

[目的]探究烟叶的外观特征、工艺参数与片烟叶片结构之间的关系,为提升打叶质量提供理论依据。[方法]以片烟叶片结构预测模型为研究对象,选取870条烟叶外观特征、打叶工艺参数与对应叶片结构数据作为训练集,构建了包括支持向量机、随机森林、多层感知机等机器学习回归模型,并基于训练集交叉验证平均MAE进行模型选择。以97条烟叶外观特征、打叶工艺参数与对应叶片结构数据作为测试集来评估所选回归模型的泛化性能。[结果]片烟大片率最佳预测模型为SVR,其在测试集上的相对分析误差和拟合优度分别为1.685 8和0.648 1,预测值与真实值间的相关系数为0.806 2。片烟中片率最佳预测模型为Random Forest,其在测试集上的相对分析误差和拟合优度分别为1.590 8和0.604 9,预测值与真实值间的相关系数为0.780 4。[结论]基于烟叶外观特征和打叶工艺参数,通过构建SVR和Random Forest模型并选取适当的超参数,能够较为准确地预测所得片烟大片率和中片率。

[Objective]To explore the relationship between appearance features of tobacco leaves,threshing technical parameters and leaf structure of tobacco strips,to provide theoretical basis for improving the quality of the threshing and redrying process.[Method]Taking the leaf structure prediction model as the research object,870 tobacco leaf appearance features,threshing technical parameters and corresponding leaf structure data were selected as the training set,and machine learning regression models were constructed including support vector machine,random forest,multi-layer perceptron.Model selection was based on the cross-validation MAE of the training set.The generalisation performance of the selected regression models was evaluated using 97 tobacco appearance features,threshing technical parameters and corresponding leaf structure data as the test set.[Result]The best model for predicting the percentage of strips with>25.4 mm was SVR,with relative percentage difference and goodness of fit of 1.6858 and 0.6481 on the test set,respectively,and the correlation coefficient between the predicted values and the true values of 0.8062.The best model for predicting the percentage of strips of 12.7-25.4 mm was Random Forest,with relative percentage difference and goodness of fit of 1.5908 and 0.6049 on the test set,respectively,and the correlation coefficient between the predicted and true values was 0.7804.[Conclusion]Based on the appearance features of tobacco leaves and the threshing technical parameters,the SVR and Random Forest models were constructed and appropriate hyperparameters were selected,which could accurately predict the the percentage of strips with>25.4 mm and the percentage of strips of 12.7-25.4 mm.