基于高光谱成像的加料烟叶丙二醇含量无损检测与可视化分析

Non-destructive Detection and Visualization Analysis of Propylene Glycol Content in Processed Tobacco Leaves Using Hyperspectral Imaging

摘要:

烟叶加料作为烟草加工制丝工艺中的关键环节，对改善烟叶的物理和化学特性，以及提升产品品质具有重要意义，但现有加料精度检测主要集中在用量监控，缺乏加料后效果的评估。本文针对加料后烟叶的微量添加物含量的无损检测及可视化分析，构建了基于高光谱成像和卷积神经网络（CNN）方法的烟叶微量添加物含量检测模型、光谱预处理方法与特征波长选择技术优化开展建模探究。通过高光谱成像系统采集添加不同比例丙二醇烟叶样本的光谱数据，分别采用标准正态变换（SNV）、多元散射校正（MSC）、Savitzky-Golay滤波平滑3种数据预处理方法对比，并通过竞争性自适应重加权算法（CARS）、主成分分析（PCA）筛选特征波长以及光谱曲线波谷点对应波长，确定了1146、1614、2511、2517、2522、1941nm 6个共同的一致关键波长。分别构建CNN、随机森林（RF）、偏最小二乘回归（PLSR）模型进行加料烟叶微量添加物丙二醇含量的检测。结果表明，SNV-PCA-CNN模型在训练集和测试集中的检测效果最佳，取前4个主成分数量累计贡献率可达99%，训练集决定系数R2C为0.9880、均方根误差RMSE为0.0020kg/kg，测试集决定系数R2P为0.9896、均方根误差RMSE为0.0021kg/kg，具备优良的拟合与泛化能力，深度学习CNN模型在测试集上的表现显著优于机器学习RF和PLSR方法。因此基于高光谱成像的CNN模型能够对加料烟叶微量添加物丙二醇含量及可视化进行准确检测及评估。

Abstract:

The application of trace amounts of sugar solution additives to tobacco leaves is a critical step in the tobacco processing and cutting technology, significantly impacting the physical and chemical properties of the leaves and enhancing cigarette quality. However, current precision detection methods for sugar solution additives primarily focus on dosage monitoring, lacking an evaluation of the post-application effects.The hyperspectral imaging technology and deep learning methods were utilized to perform non-destructive detection and visualization analysis of trace additives in tobacco leaves after sugar solution application. A prediction system based on a deep learning convolutional neural network (CNN) model was developed, incorporating multiple spectral preprocessing methods and feature band selection techniques to optimize model performance and improve the detection accuracy of additive content in tobacco leaves. Spectral data from tobacco samples with varying proportions of propylene glycol were collected by using a hyperspectral imaging system. The data were preprocessed by using three methods: standard normal variate (SNV), multiplicative scatter correction (MSC), and Savitzky-Golay filtering, which were employed for data preprocessing respectively. Feature bands were selected through competitive adaptive reweighted sampling (CARS), principal component analysis (PCA), and identification of spectral trough points, resulting in six common consistent key wavelengths at 1146nm, 1614nm,2511nm, 2517nm, 2522nm, and 1941nm. CNN, random forest (RF), and partial least squares regression (PLSR) models were constructed to predict the additive content, with hyperspectral data visualization conducted by using the CNN method. The results showed that the SNV-PCA-CNN model achieved the best predictive performance for both the training set (R2C was 0.9880, RMSE was 0.0020kg/kg) and the test set (R2P was 0.9896, RMSE was 0.0021kg/kg), and the cumulative contribution rate was close to 99% by taking the first-four principal components, demonstrating excellent fitting and generalization capabilities. The predictive ability of the deep learning CNN model significantly outperformed the performance of traditional machine learning methods RF and PLSR, reflecting the CNN model sufficient generalization capabilities for hyperspectral data of tobacco samples with sugar solution additives. The combination of hyperspectral imaging and the CNN model showed great potential for detecting trace additives in tobacco leaves, providing technical support for non-destructive testing and precise control in the tobacco processing industry.