基于机器学习的密集烘烤过程烟叶失水率预测模型对比

为准确预测密集烘烤过程烟叶失水率，以精准调控烘烤工艺参数，基于机器学习建立烟叶失水率预测模型。以翠碧一号中部叶为材料，实时采集烘烤过程中烟叶图像和失水率；利用图像处理技术提取烟叶的10种颜色特征和10种纹理特征，通过变量聚类和皮尔逊相关性分析优选出2种颜色特征(a^*/b^*、R)和2种纹理特征(梯度熵、梯度分布不均匀性)；将训练集图像的4种优选特征和烟叶失水率作为输入变量，分别对建立的网格式支持向量机(GS-SVM)、遗传算法优化的BP神经网络(GA-BP)、极限学习机(ELM)3种预测模型进行训练。利用3种预测模型对测试集图像进行烟叶失水率预测并与实际失水率比较。结果表明，3种预测模型均能够较为准确地预测密集烘烤过程烟叶失水率(均方根误差RMSE≤0.014 0，决定系数R2≥0.996 1)，对变黄期(0～40 h)和定色期(40～100 h)的预测误差小于干筋期(100～140 h)。该技术可为烟叶烘烤智能调控系统的研发提供支持。

Machine learning-based models for predicting dehydration rate of tobacco leaf during bulk curing and comparisons thereof

To accurately predict the dehydration rate of tobacco leaf during bulk curing and precisely control curing process parameters, three models for predicting the dehydration rate of tobacco leaf were established based on machine learning. Taking the middle leaves of cv. CB-1 as the material, the images and dehydration rates of tobacco leaves were realtimely collected during curing process. Image processing technology was used to extract 10 color features and 10 textural features of the leaves, and 2 color features (a*/b*, R) and 2 textural features (gradient entropy, nonuniformity of gradient distribution) were selected through variable clustering and Pearson correlation analysis. The three established prediction models, the grid search support vector machine (GS-SVM), genetic algorithm optimized BP neural network (GA-BP), extreme learning machine (ELM) models, were subject to training with the four selected features of the images in the training set and the dehydration rates of tobacco leaves. The dehydration rates of tobacco leaves predicted by the three models were compared with the actual dehydration rates. The results showed that all the three prediction models could accurately predict the dehydration rates of tobacco leaves during bulk curing with the root mean square error (RMSE)≤0.014 0 and coefficient of determination (R2)≥0.996 1. The prediction errors for the leaves at yellowing stage (0-40 h) and color-fixing stage (40-100 h) were lower than those at stem-drying stage (100-140 h). This technology provides a support for the development of intelligent control system for tobacco leaf curing.