Predicting the blending ratio of Mee Tea based on near infrared spectroscopy

Abstract: A blending operation has been one of the refined processing procedures for the various types of tea. Since the quality of material teas varies greatly in different areas and seasons, it is a high demand for the standardized quality of commercial tea. There are some differences between the material teas and the commercial samples (called the standard sample). The tea blending can be widely used to maintain the product quality and increase output by blending different materials teas. Among them, the blending ratio can be determined by an experienced expert at present. A sample can be prepared after the quality evaluation on the various types of material teas, further to compare with the standard sample for an optimal blending ratio. However, the traditional blending depends mainly on the subjective senses of experts. Taking the roasted green tea (Mee tea) as the object, this study aims to rapidly and accurately predict the blending ratio using the Convolutional Neural Network (CNN). Near-infrared spectroscopy was also selected to effectively characterize the chemical components of tea. As such, the adjustment of flavor was achieved under the various ratios of raw teas. Four kinds of tea raw materials were used to prepare 25 tea samples, according to the preset ratio table, where 20 groups of sub-samples were prepared for each tea sample with different blending ratios. The spectrum of each sample was then collected. Moreover, the Standard Normal Variable (SNV) transformation was used to preprocess the spectrum. Then, four models of machine learning were constructed to predict the blending ratios. The performance of the model was evaluated to compare with the preset blending ratio. The models were: the automatic encoder with Softmax; CNN combined with Softmax; Principal Component Analysis (PCA) combined with Partial Least Squares (PLS); and PCA combined with Softmax. Subsequently, 3-fold cross validation was utilized to train the model. The feature dimension was verified from 10 to 100, with the step of 10. An optimal dimensionality was achieved to select the best performance of 3-fold cross validation sets. The results showed that the CNN combined with Softmax presented the best performance in the validation sets, with the coefficient of determination R2 of 0.964 3, and the Root Mean Squared Error (RMSE) of 0.047 2. The prediction data using the CNN combined with Softmax was significantly better than others, and the performance index in the test set was close to the validation, indicating the better generalization ability of the model. Additionally, the large convolution kernel in the first layer was more conducive to extracting the spectral features, in terms of the convolution kernels and activation functions. The finding can also provide theoretical data support to the digital and intelligent blending in mechanized tea production.