Study on nonlinear modeling for quality evaluation and analysis of constituent of tea using near infrared spectroscopy and chemometrics

Tea is one of the three major non alcohol beverages in the world. We have a long history in tea drinking in Japan. It also is widely produced in over 20 prefectures include Shizuoka prefecture as the certer of production area in Japan. However tea is agricultural product, It belongs to the technical crop with the fabrication. Therefore, the fabriction technology in the factory is very important as same as the cultivation one in the tea field. While the improvement of the production basis advances likes other industry, the production system and the evaluation custom for tea as outdate state which depends on the experiene and the intuition have still remained in various scenes of tea. This is a big issue to rationalize and improve the efficiency for the production and the circulation. Especially, the problems are pointed out in the scene of the quality evaluation of crude tea, the moisture assessment in tea process and the classification of fresh tea leaf. Then, the experimental research which was carried out to confirm the practicability and the possibility of the application of the technique of near infrared (NIR) spectroscopy method and the chemometrics mainly on the artificial neural network (ANN) for the quality evaluation and process control of tea in various scene depend on analysis of constituents. In this study, the features of the absorbance data obtained using the near infrared (NIR) spectroscopic method for quality evaluation of tea were clarified that the high correlation in the variables mutually exists and the explanatory variables are multivariate. In addition, the new modeling technique was proposed in order to show the practical answer to both of the nonlinear problem which exists between the constituent of tea and absorbance, and the quality score of tea and absorbance, and the problem of the multiple collinearity in the multiple linear regression (MLR) analysis used for evaluation model of tea until now. The effectiveness of the technique proposed in this study were discussed through the comparison of the conventional modeling technique. The summary is arranged in the following. 1) First, the technique of the chemometrics mainly on the neural network was applied for absorbance in near infrared region and the colorimetry value in visible region in order to establish the quality assessment system for crude tea which can replace sensory test. Large number of crude tea samples which prevailed in the market were collected, and the quality of them were evaluated by the sensory test. Simultaneously, the absorbance in near infrared range and the colorimetry value in visible range were measured, and they were used to be the experimental data. There are multi-collinearity problem for calibration, because they are high correlated mutually. The relationship between sensory test score and the measured value using the NIR instrument and the colorimeter, was analyzed and the latent variable which existed there was extracted by principal component analysis (PCA). Multivariate nonlinear model by neural network was constructed on the basis of the latent variables extracted, and the feasibility of the model was examined on the unknown sample. The result shows that the nonlinear relation exists between the quality assessment score by sensory test and the absorbance value of near infrared spectroscopy (NIR) and colorimeter and the high correlation mutually also exists between the inspection items of sensory test and between the variables of the value obtained by NIR and colorimeter. It was verified that the nonlinear model which combined principal component analysis (PCA) with artificial neural network (ANN) would be able to be applied for quantitative analysis instead of the multiple linear regression (MLR) model which is linear coupled model. The 14 new principal components which were orthogonalized to each other from the values measured using NIR and colorimeter, extracted by PCA and the back 7th principal component include noise as a disturbance factor was removed in order to establish the stability model. After that, the 7-7-1 hierarchical neural network was constructed considering the nonlinear relationship which existed between the principal component score and sensory test score. As a result, the predictivity of the model for quality of crude tea by numerical analysis, the predictive value by multivariate nonlinear regression model (PCA-ANN) showed the fair accuracy on the unknown real samples. The trained neural network performed fairly satisfactorily when tested with inspection data since the correlation factor between actual and estimated data was 0.9290. It became clear that multivariate nonlinear regression model (PCA-ANN) based on near infrared absorbance and visible calorimetric value was effective for the quality assessment of crude tea. 2) Second, near infrared spectroscopic method (NIR) and the chemometrics mainly on neural network were applied in order to establish the method of measuring for water content of tea leaf in the manufacturing process to replace the sensory assessment by veteran engineer. Large number of process tea samples were collected in the actual tea manufacture factory, and the absorbance of near infrared region were measured using the interference fixed filter typed equipment. Simultanously, the moisture of the samples were measured by drying air method, and were used as the reference value for modeling. After analyzing the relation between the moisture and the absorbance, the latent variables were extracted by principal component analysis (PCA). The multivariate nonlinear model using the neural network was constructed on the basis of latent variables and the moisture as reference value, and the feasibility of the model was examined on the validation sample set. The analysis shows that the relation which is nonlinear between water content in tea leave and near infrared absorbance as same as the quality evaluation model for crude tea described in 1 section, and the high correlation exists between the near infrared absorbance mutually. lt was verified that the application of nonlinear model which combined neural network (ANN) with principal component analysis (PCA) were desirable for modeling in stead of linear regression (LR) method used in previous research. Extracting latent variables using the principal component analysis (PCA), even if the data with high correlation mutually like the absorbance of near infrared spectroscopy, the problem of multiple collinearity was solved. The 3-3-1 hierarchy neural network model was constructed for the nonlinear relation between the moisture and the PCA score calculated by principal component analysis (PCA). Through the comparison with other standard prediction models, the predictivity of this new model for water content was examined. It showed that the value predicted by multivariate nonlinear model (PCA-ANN) was sufficiently equivalent to actual moisture content on the unknown samples. This model performed with a high accuracy of prediction in prediction set. The efficiency of the combined model of PCA-ANN regression becomes clear as the modeling for multi-collinearity in variables and non-linearity between the moisture content covering a wide range in tea processing and the absorbance provided with NIRS. The PCA-ANN regression model which has modeled the nonlinear relationship between the water content in wide range and absorbance in near infrared region, which has solved the multi-collinearity problem in variables mutually, showed the effectiveness for water measurement of tea leaf in the manufacturing tea process. 3) Finally, Near infrared (NIR) reflectance method and Chemometrics were applied in order to establish the quality assessment system at the raw material stage before process. It is very important to analyze the amino acid content, because they influences the quality of tea. Therefore the assessment system by which the fresh tea leaves were classified according to the amino acid content was demanded to develop in factory. The raw tea leaf of which the quality differed were collected in great numbers, and their 7 main amino acids content were determined quantitatively using the high performance liquid chromatography (HPLC), while their absorbance were measured by near infrared spectroscopy (NIR). Their values were used as the experimental data for modeling. Three chemometrics procedure ; Genetic algorithm (GA) Partial Least Squares (PLS) and Neural Network (NN), were applied for modeling the nonlinear relation of the amino acid content by HPLC with the absorbance by NIR, and the prediction model for the 4 main amino acids content of raw tea leaf was constructed with the combined method (GA-PLS-ANN), the predictivity of this model on validation sample set was examined. The results showed that the 700 multivariate data obtained by the scanning typed near infrared spectrometer included the inform mutually overlapped and had high correlation each other, and the selection of the optimum variabl and the contraction of information were indispensable. It was shown that the relation between PLS score and amino acid content analyzed by HPLC is nonlinear and the multivariate nonlinear regression model (GA-PLS-ANN) is adequate to this relation. Then the genetic algorithm (GA) was applied for large number of near infrared absorbance in order to choose the variable of the effective decimal used in prediction model. Next, the latent variables were extracted from the 15 optimal variables by the PLS method. Considering the nonlinear relationship which existed between PLS score and amino acid content, the 4-4-4 hierarchy ANN regression model (GA-PLS-ANN) were constructed for prediction of amino acid. The acceptability of this multivariate nonlinear regression model (GA-PLS-ANN) was verified on the unknown samples. The estimation of the ANN regression model obtained was very precise. The multivaiate nonlinear regression model (GA-PLS-ANN) which combined genetic algorithm (GA) and Partial Least Square (PLS) with neural network, applied for near infrared absorbance, could estimate the multiple amino acid content in fresh tea leaf simultaneously. This model is effective as the quality assessment system of the raw material, because it is more reproducible and objective than sensory classification. The GA-PLS-ANN regression model, when compared with the conventional multiple linear regression (MLR) model, greatly reduced the number of essential variables, and was more adaptable for the prediction of other unknown samples. In addition, the GA-PLS-ANN regression model was also suited for a nonlinear relation between variables and amino acids. These results are useful to develop the evaluation management system in various scene ; quality assessment of tea, production control according to water content and quality assessment of raw material of the tea, because it also has the rapidity which is equivalent to ordinary sensory test method. Using genetic algorithm (GA), principal component analysis (PCA), partial least square (PLS) and neural network (ANN), the multivariate nonlinear model in this study, which combined the chemometrics| variously with near infrared reflectance (NIR) method were concluded to have the high accuracy of predictability, the excellent contractive function of much information and the flexible adaptability for the nonlinear relation besides with rapidity and non destruction. With the features, this methodology cann be sufficiently utilized for not only tea section : evaluation of quality and constituent of raw leaf and crude tea in process or market stage, but also quality assessment and analysis of component in other agricultural products and food. In present, the practical machine are developed after research which already have reported and some operation system work actually for assessment of constituent in crude tea or fresh tea leaf. These performance and knowledge obtained in this study which are certainly useful for the improvement in the accuracy of evaluation and measurement for constituent and quality in tea at the real tea production and process scenes.