基于近红外光谱技术的茶红酸汤质量快速无损评价

Rapid and non-destructive evaluation of the quality of tea red sour soup based on near-infrared spectroscopy technology

中文关键词:  茶红酸汤  质量评价  近红外光谱  主成分分析  人工神经网络

英文关键词:tea red sour soup  quality evaluation  near infrared spectroscopy  principal component analysis  artificial neural network.

中文摘要:

      目的 尝试应用近红外光谱技术（near infrared spectroscopy, NIRS）对茶红酸汤的质量进行快速无损评价。方法 收集三个质量等级的茶红酸汤样品，扫描获得近红外光谱后进行光谱预处理，剔除部分噪声信息，再通过反向区间偏最小二乘法筛选特征光谱区间并进行主成分分析，最后应用反向传播人工神经网络（back propagation artificial neural network, BP-ANN）建立预测模型。结果 最佳光谱预处理方法为（多元散射校正+一阶导数）；筛选出四个特征光谱区间，前三个主成分的累积贡献率为97.85%。当使用tanh传递函数建立BP-ANN-NIRS模型时结果最佳，交互验证均方根误差和决定系数（determination coefficient of cross validation, Rc2）分别为0.953和0.031，预测集均方根误差和决定系数（determination coefficient of prediction set, Rp2）分别为0.942和0.041，验证集样品的预测|偏差| <0.08。结论 应用近红外光谱技术实现了茶红酸汤的质量地快速准确评价，为茶红酸汤质量评价提供一种新的参考方法。

英文摘要:

      ABSTRACT: Objective This study aims to investigate the potential of near infrared spectroscopy (NIRS) for the rapid and non-destructive evaluation of the tea red sour soup quality. Methods Three quality levels of tea red sour soup samples were collected and scanned to obtain near-infrared spectra for spectral preprocessing to eliminate some noise information. Then the characteristic spectral intervals were screened by the backward interval partial least squares method and subjected to principal component analysis, and finally back propagation artificial neural network (BP-ANN) was applied to build a prediction model. Results The optimal spectral preprocessing method was (multivariate scattering correction (MSC) + first derivative (FD)); Four characteristic spectral intervals were selected, and the cumulative contribution rate of the first three principal components was 97.85%. When the tanh transfer function was used to establish the BP-ANN-NIRS model, the results were optimal, with root mean square error and determination coefficient of cross validation (Rc2) of 0.953 and 0.031, respectively. The root mean square error and determination coefficient of prediction set (Rp2) were 0.942 and 0.041, respectively. The absolute values of prediction deviations