基于随机森林利用差分拉曼光谱对塑料食品包装瓶的分类研究

Classification of Plastic Food Packaging Bottles by Differential Raman Spectroscopy Based on Random Forest

中文关键词: 差分拉曼光谱  塑料食品包装瓶  人工神经网络  随机森林算法

英文关键词: differential Raman spectroscopy  plastic food packaging bottles  artificial neural network  random forest algorithm

中文摘要:

      目的 建立一种快速无损的检验塑料食品包装瓶的分析方法，提供一种快速分类模型。方法 利用差分拉曼光谱对100个塑料食品包装瓶样品进行检验，根据样品的差分拉曼特征峰可以对样品进行分类，样品可被分成聚对苯二甲酸乙二醇酯和聚丙烯两大类，对其中数目较多的第I类继续根据样品中所含填料的不同进行分类。利用贝叶斯判别、多层感知器和随机森林算法分别构建分类模型对继续分类结果进行分析验证。结果 第I类样本可继续被分为4类，贝叶斯判别结合留一交叉验证法分类正确率为71.7%，多层感知器神经网络分类模型的训练集和测试集分类正确率分别为100%和86.2%，随机森林分类模型的训练集和测试集分类正确率分别为100%和96.5%。通过比较发现，差分拉曼光谱与随机森林算法相结合可以对塑料食品包装瓶实现有效的分类。结论 该方法简单快速，样品用量少且无损样品，可为塑料食品包装品的物证鉴定提供科学依据。

英文摘要:

      The work aims to establish a fast and non-destructive analysis method for inspecting plastic food packaging bottles and provide a fast classification model. 100 plastic food packaging bottle samples were tested by differential Raman spectroscopy. The samples were classified based on their differential Raman characteristic peaks and divided into two categories of polyethylene terephthalate and polypropylene. The Class I samples in a larger number were further classified based on the different fillers contained. The classification model was constructed by Bayesian discriminant analysis, multi-layer perceptron, and random forest algorithm to analyze and verify the continued classification results.The Class I samples were further divided into four categories. The classification accuracy of Bayesian discriminant combined with left one cross validation method was 71.7%, the classification accuracy of the training and testing sets of the multi-layer perceptron neural network classification model was100% and 86.2%, respectively, and the classification accuracy of the random forest classification model0020was 100% and 96.5%. Through comparison, it was found that the combination of differential Raman spectroscopy and random forest algorithm could effectively classify plastic food packaging bottles.This method is simple and fast, requiring a small sample size but not damaging samples, which can provide scientific basis for the identification of physical evidence in plastic food packaging products.