Research On Machine Learning Algorithms For Tea Blending Process

China is a big country in tea production,consumption and export.There are many ways of tea processing.In the processing of tea such as Keemun black tea and Roasted green tea,blending technology is usually used,that is,according to the quality requirements of the product,the process of combining two or more semi-finished products in a certain proportion to form the finished tea.In this process,tea quality review and blending proportion design are two key technologies.At present,tea blending still relies on experts,which is highly subjective,lacks quantifiable technical means,and makes it difficult to train technical personnel.Focusing on the problem of grade identification and blending proportion design in the process of tea blending,this paper has carried out the following main work:(1)Tea sample data collection and feature extraction.In order to characterize the quality of tea samples,the shape,color,texture and spectral data of tea samples were collected by machine vision device and near-infrared spectrometer,and the shape,color,texture and spectral features were extracted respectively.(2)Tea grade identification.Taking Keemun black tea with seven grades as the research object,aiming at the problem of grade recognition,a high-precision grade recognition model based on decision level data fusion is established.Firstly,the single feature classification results are obtained by using BP and SVM algorithms;Secondly,based on Naive Bayesian,Decision tree,DS evidence theory and SVM algorithm,a decision level data fusion model is established;In order to compare the effects of different kinds of data fusion,a feature level data fusion model is further established based on Decision tree and SVM.The experimental results show that the test set accuracy of DS evidence theory and Naive Bayes algorithm in the decision level data fusion method can reach more than 99%,and the classification accuracy is significantly better than the single feature classification model and the two feature level data fusion models in this paper.The above research results will provide a reference method for the quality evaluation of Keemun black tea in processing and marketing.(3)Design of tea blending proportion.Aiming at the design of tea blending proportion,the blending proportion calculation model based on optimized parameters was established,which can calculate the blending proportion with high accuracy.This paper plans to build a multi-output regression model to calculate the tea blending proportion.Firstly,three tea blending proportion calculation models of Decision Tree,Random Forest and Gradient Boosting Decision Tree were constructed,and the model data dimensions and some key parameters are optimized.Secondly,the calculated tea blending proportion is compared with the preset blending proportion,and the model performance is analyzed.The coefficient of determination R~2 and the root mean square error RMSE of the validation set of the Gradient Boosting Decision Tree are 0.9422 and 0.0586 respectively,which are better than the Decision tree and the Random Forest algorithm.The results show that the tea blending proportion prediction model based on tree model established in this paper can solve the blending proportion with high accuracy,which has reference significance for the design of tea blending proportion.Aiming at the problems of grade identification and blending proportion design in the process of tea blending,this paper builds a grade identification model based on the data fusion method and a blending proportion calculation model based on the tree model method.It provides data support and reference methods for the development of related technologies,and has certain application value for realizing the digitalization and standardized processing of tea.