State Key Laboratory of Tea Plant Biology and Utilization, Anhui Agricultural University, Hefei, P. R. China.
J Sci Food Agric. 2020 Aug;100(10):3950-3959. doi: 10.1002/jsfa.10439. Epub 2020 May 14.
Grading represents an essential criterion for the quality assurance of black tea. The main objectives of the study were to develop a highly robust model for Chinese black tea of seven grades based on cognitive spectroscopy.
Cognitive spectroscopy was proposed to combine near-infrared spectroscopy (NIRS) with machine learning and evolutionary algorithms, selected feature information from complex spectral data and show the best results without human intervention. The NIRS measuring system was used to obtain the spectra of Chinese black tea samples of seven grades. The spectra acquired were preprocessed by standard normal variate transformation (SNV), multiplicative scatter correction (MSC) and minimum/maximum normalization (MIN/MAX), and the optimal pretreating method was implemented using principal component analysis combined with linear discriminant analysis algorithm. Three feature selection evolutionary algorithms, which were a genetic algorithm (GA), simulated annealing (SA) and particle swarm optimization (PSO), were compared to search the best preprocessed characteristic wavelengths. Cognitive models of Chinese black tea ranks were constructed using extreme learning machine (ELM), K-nearest neighbor (KNN) and support vector machine (SVM) methods based on the selected characteristic variables. Experimental results revealed that the PSO-SVM model showed the best predictive performance with the correlation coefficients of prediction set (R ) of 0.9838, the root mean square error of prediction (RMSEP) of 0.0246, and the correct discriminant rate (CDR) of 98.70%. The extracted feature wavelengths were only occupying 0.18% of the origin.
The overall results demonstrated that cognitive spectroscopy could be utilized as a rapid strategy to identify Chinese black tea grades. © 2020 Society of Chemical Industry.
分级是红茶质量保证的重要标准。本研究的主要目的是开发一种基于认知光谱学的中国红茶七级的高稳健模型。
提出了认知光谱学，将近红外光谱（NIRS）与机器学习和进化算法相结合，从复杂的光谱数据中选择特征信息，在无需人工干预的情况下显示最佳结果。使用 NIRS 测量系统获取中国红茶七个等级的样本光谱。采用标准正态变量变换（SNV）、乘法散射校正（MSC）和最小/最大归一化（MIN/MAX）对获得的光谱进行预处理，并采用主成分分析与线性判别分析算法相结合的方法实施最佳预处理方法。比较了三种特征选择进化算法，即遗传算法（GA）、模拟退火（SA）和粒子群优化（PSO），以搜索最佳预处理特征波长。基于所选特征变量，使用极限学习机（ELM）、K-最近邻（KNN）和支持向量机（SVM）方法构建中国红茶等级的认知模型。实验结果表明，PSO-SVM 模型具有最佳的预测性能，预测集的相关系数（R）为 0.9838，预测均方根误差（RMSEP）为 0.0246，正确判别率（CDR）为 98.70%。提取的特征波长仅占原始波长的 0.18%。
总的来说，认知光谱学可以作为一种快速识别中国红茶等级的策略。 © 2020 英国化学学会。
