随机赋予类别与一对多分类策略判别不同类型烟用爆珠

Randomly Assigning Labels and “Leave-One-Class-Out” Strategies to Distinguish Different Types of Tobacco Capsules

采用近红外光谱法结合PLS-DA模式识别算法建立快速鉴别不同类型烟用爆珠的方法。收集蜜甜型、清甜型和薄荷型3种类型爆珠不同批次共计27个样品,并采集其近红外光谱数据,合计408张光谱,建立烟用爆珠类型快速识别模型,并对预测集样品进行分类预测。如何选择代表性样本以及选择的数目是研究的关键。数据处理结果表明采用PLS-DA法结合校正集样本选择算法K-S,在两种策略条件下,识别准确度呈现较好结果:识别准确度基本在95%以上,同时一对多分类策略的结果优于随机赋予类别策略。表明选择代表性样本的重要性,另外该模型的建立为烟用爆珠分类提供一种新颖的、快速的、无损鉴别分析方法。

Near-infrared spectroscopy combined with PLS-DA was used to establish rapid discriminationmodels for different types of tobacco capsules.A total of 27 samples were collected from three kinds of capsules:fresh sweet,sweet and mint,and their near-infrared spectral data were collected,408 spectra in all.Rapid discrimination models of capsules for tobacco were established,and tested samples were predicted.How to select representative samples and number of samples is the key in this research.The result of data processing shows that PLS-DA combined with K-S,recognition accuracy shows good performance:recognition accuracy reaches more than 95%under two strategies;mean while discrimination performance of"leave-one-class-out"is better than that of randomly assigned labels.It shows the importance of selecting representative samples.In addition,established models can provide new,fast and nondestructive method to classify tobacco bead.