Xu Sai, Sun Xiuxiu, Lu Huazhong, Zhang Qianqian
Public Monitoring Center for Agro-Product of Guangdong Academy of Agricultural Sciences, Guangzhou 510640, China.
Guangdong Academy of Agricultural Sciences, Guangzhou 510640, China.
Sensors (Basel). 2019 May 22;19(10):2359. doi: 10.3390/s19102359.
The objective of this study was to find an intelligent and fast method to detect the type, blended ratio, and mixed ratio of ancient Pu'er tea, which is significant in maintaining order in the Pu'er tea industry. An electronic nose (E-nose) and a visible near infrared spectrometer (VIS/NIR spectrometer) were applied for tea sampling. Feature extraction was conducted using both the traditional method and a convolutional neural network (CNN) technique. Linear discriminant analysis (LDA) and partial least square regression (PLSR) were applied for pattern recognition. After sampling while using the traditional method, the analysis of variance (ANOVA) results showed that the mean differential value of each sensor should be selected as the optimal feature extraction method for E-nose data, and raw data comparison results showed that 19 peak/valley values and two slope values were extracted. While the format of E-nose data was in accord with the input format for CNN, the VIS/NIR spectrometer data required matrixing to meet the format requirements. The LDA and PLSR analysis results showed that CNN has superior detection ability, being able to acquire more local features than the traditional method, but it has the risk of mixing in redundant information, which can act to reduce the detection ability. Multi-source information fusion (E-nose and VIS/NIR spectrometer fusion) can collect more features from different angles to improve the detection ability, but it also contains the risk of adding redundant information, which reduces the detection ability. For practical detection, the type of Pu'er tea should be recognizable using a VIS/NIR spectrometer and the traditional feature extraction method. The blended ratio of Pu'er tea should also be identifiable by using a VIS/NIR spectrometer with traditional feature extraction. Multi-source information fusion with traditional feature extraction should be used if the accuracy requirement is extremely high; otherwise, a VIS/NIR spectrometer with traditional feature extraction is preferred.
本研究的目的是找到一种智能且快速的方法来检测古代普洱茶的种类、拼配比例和混合比例，这对于维护普洱茶行业的秩序具有重要意义。采用电子鼻（E-nose）和可见近红外光谱仪（VIS/NIR光谱仪）对茶叶进行采样。使用传统方法和卷积神经网络（CNN）技术进行特征提取。应用线性判别分析（LDA）和偏最小二乘回归（PLSR）进行模式识别。在采用传统方法采样后，方差分析（ANOVA）结果表明，应选择每个传感器的平均差值作为电子鼻数据的最佳特征提取方法，原始数据比较结果显示提取了19个峰/谷值和两个斜率值。虽然电子鼻数据的格式符合CNN的输入格式，但VIS/NIR光谱仪数据需要进行矩阵化以满足格式要求。LDA和PLSR分析结果表明，CNN具有优越的检测能力，能够比传统方法获取更多局部特征，但存在混入冗余信息的风险，这可能会降低检测能力。多源信息融合（电子鼻和VIS/NIR光谱仪融合）可以从不同角度收集更多特征以提高检测能力，但也存在添加冗余信息从而降低检测能力的风险。对于实际检测，使用VIS/NIR光谱仪和传统特征提取方法应能够识别普洱茶的种类。使用VIS/NIR光谱仪和传统特征提取方法也应能够识别普洱茶的拼配比例。如果精度要求极高，则应使用传统特征提取的多源信息融合；否则，优先选择使用传统特征提取的VIS/NIR光谱仪。
