From the perspective of combating fraud issues and examining keemun black tea properties, there was a contemporary urgent demand for a keemun black tea rankings identification system. Current rapid evaluation systems had been mainly developed for green tea grade evaluation, but there was space for improvement to establish a highly robust model. The present study proposed cognitive spectroscopy that combined near infrared spectroscopy (NIRS) with multivariate calibration and feature variable selection methods. We defined "cognitive spectroscopy" as a protocol that selects characteristic information from complex spectral data and showed optimal results without human intervention. 700 samples representing keemun black tea from seven quality levels were scanned applying an NIR sensor. To differentiate which wavelength variables of the acquired NIRS data carry key and feature information regarding keemun black tea grades, there were four different variables screening approaches, namely genetic algorithm (GA), successive projections algorithm (SPA), competitive adaptive reweighted sampling (CARS), and shuffled frog leaping algorithm (SFLA), were compared in this study. Cognitive models were developed using least squares support vector machine (LSSVM), back propagation neural network (BPNN) and random forest (RF) methods combined with the optimized characteristic variables from the above variables selection algorithms for the identification of keemun black tea rank quality. Experimental results showed that all cognitive models utilizing the SFLA approach achieved steady predictive results based on eight latent variables and selected thirteen characteristic wavelength variables. The CARS-LSSVM model with the best predictive performance was proposed based on selecting ten characteristic latent variables, and the best performance indicators of the model were as follows: the root mean square error of prediction (RMSEP) was 0.0413, the correlation coefficients of prediction set (Rp) was 0.9884, and the correct discriminant rate (CDR) was 99.01% in the validation process. This study demonstrated that cognitive spectroscopy represented a proper strategy for the highly identification of quality rankings of keemun black tea.

中文翻译：

---

基于认知光谱对基蒙红茶等级的高度识别：近红外光谱结合特征变量选择。

从打击欺诈问题和检查基蒙红茶特性的角度来看，当代迫切需要基蒙红茶等级识别系统。当前的快速评估系统主要是针对绿茶等级评估而开发的，但仍有改进空间以建立高度鲁棒的模型。本研究提出了将近红外光谱（NIRS）与多元校准和特征变量选择方法相结合的认知光谱。我们将“认知光谱法”定义为从复杂光谱数据中选择特征信息并显示最佳结果而无需人工干预的协议。使用NIR传感器扫描了700种代表七个质量等级的红茶的样品。为了区分所获取的NIRS数据中哪些波长变量携带有关基蒙红茶等级的关键和特征信息，有四种不同的变量筛选方法，即遗传算法（GA），连续投影算法（SPA），竞争性自适应加权采样（CARS） ，以及改组蛙跳算法（SFLA），在本研究中进行了比较。使用最小二乘支持向量机（LSSVM），反向传播神经网络（BPNN）和随机森林（RF）方法，结合上述变量选择算法中的优化特征变量，开发了识别模型，用于识别红茶等级质量。实验结果表明，基于SFLA方法的所有认知模型均基于八个潜在变量和选定的十三个特征波长变量获得了稳定的预测结果。在选择十个特征潜变量的基础上，提出了具有最佳预测性能的CARS-LSSVM模型，该模型的最佳性能指标为：预测的均方根误差（RMSEP）为0.0413，预测的相关系数为设定（Rp）为0.9884，在验证过程中正确的判别率（CDR）为99.01％。这项研究表明，认知光谱法是高度识别红茶质量等级的正确策略。在选择十个特征潜变量的基础上，提出了具有最佳预测性能的CARS-LSSVM模型，该模型的最佳性能指标为：预测的均方根误差（RMSEP）为0.0413，预测的相关系数为设定（Rp）为0.9884，在验证过程中正确的判别率（CDR）为99.01％。这项研究表明，认知光谱法是高度识别红茶质量等级的正确策略。在选择十个特征潜变量的基础上，提出了具有最佳预测性能的CARS-LSSVM模型，该模型的最佳性能指标为：预测的均方根误差（RMSEP）为0.0413，预测的相关系数为设定（Rp）为0.9884，在验证过程中正确的判别率（CDR）为99.01％。这项研究表明，认知光谱法是高度识别红茶质量等级的正确策略。验证过程中占01％。这项研究表明，认知光谱法是高度识别红茶质量等级的正确策略。验证过程中占01％。这项研究表明，认知光谱法是高度识别红茶质量等级的正确策略。