Intelligent sensing and monitoring method for whole fermentation process of congou black tea

Abstract

The invention discloses an intelligent sensing and monitoring method for the whole fermentation process of Congou black tea, and the method employs THz-TDS to measure the active excitation THz transmission spectrum of Congou black tea, employs an LDA dimension reduction method to extract the THz absorbance spectrum feature matrix of Congou black tea at different fermentation times, inputs the THz absorbance spectrum feature matrix into an HDBSCAN clustering device and a two-layer convolutional neural network DCNN regression device, and carries out the detection of the THz absorbance spectrum feature matrix. And finally obtaining a clustering three-dimensional effect picture and a regression result picture. According to the method, the LDA-HDBSCAN model and the LDA-DCNN model are comprehensively utilized, so that the change of the tea polyphenol content in the fermentation process of the congou black tea can be successfully predicted, and rapid, environment-friendly, high-precision and lossless intelligent sensing and monitoring of the fermentation degree of the congou black tea are realized; a reliable technical means with practical application potential is provided for the field of intelligent monitoring of the whole fermentation process of the congou black tea, and the method has wide application prospects in congou black tea fermentation moderation judgment and related fields due to high accuracy and efficiency.

Description

Technical Field

[0001] The invention relates to intelligent sensing and monitoring technology for the whole process of black tea fermentation, and in particular to an LDA-DCNN model based on THz-TDS to predict changes in tea polyphenols in Gongfu black tea during the fermentation process, and specifically to an intelligent sensing and monitoring method for the whole process of Gongfu black tea fermentation. Background Art

[0002] Tea is the most popular beverage in the world, not only because of its pleasant flavor, but also because of its health benefits. According to the different fermentation degrees of tea polyphenols, it can generally be divided into unfermented tea, semi-fermented tea, and fully fermented tea. As an unfermented tea, green tea retains many natural vitamins and other substances during the processing process. Therefore, the flavor of green tea is generally characterized by "clear soup, green leaves, and rich flavors". However, as a fully fermented black tea, fermentation plays a decisive role in the flavor. At present, there have been many studies on the influence of fermentation conditions on the quality of black tea at home and abroad. Studies have shown that too long or too short fermentation time of black tea has an adverse effect on the formation of black tea quality. However, the quality of black tea that changes with fermentation time has not been evaluated. Therefore, it is of great significance to quickly and accurately identify the fermentation degree of black tea with fermentation time without loss and evaluate them to promote the development of the black tea industry.

[0003] The frequency of THz waves is between 0.1 and 10 THz. All substances in nature are composed of molecules. The weak interactions between molecules have obvious characteristic absorption peaks in the THz band and show strong absorption and dispersion characteristics. The absorption and dispersion characteristics can be used as the "fingerprint" spectrum of the substance. Based on this, THz-TDS can be applied to material detection, and it is highly favored because of its low energy and strong penetration and can be used for non-destructive, rapid and accurate detection of food.

[0004] THz spectral data is high-dimensional nonlinear redundant data. The linear discriminant algorithm LDA (Linear discriminant analysis, referred to as LDA) can project high-dimensional THz spectral data into the optimal discriminant vector space, extract classification information and compress the dimension of the feature space. After projection, the distance between classes can be maximized and the distance within classes can be minimized at the same time. The analytical solution based on the generalized eigenvalue problem can be obtained, which can avoid the local minimum problem often encountered in the construction of nonlinear algorithms. The HDBSCAN algorithm is based on the density space clustering algorithm DBSCAN combined with the hierarchical clustering algorithm. It can introduce a stability metric that can automatically select reliable clustering parameters and evaluate the quality of clustering results, so that HDBSCAN can effectively identify diverse clustering structures and significantly improve the robustness of clustering when processing THz spectral data after LDA dimensionality reduction. At present, there is a blank in the identification of the quality changes of black tea with fermentation time using the linear discriminant algorithm-deep convolutional neural network regression model LDA-DCNN (Linear Discriminant Analysis-Double Convolutional Neural Network, LDA-DCNN) based on the terahertz time-domain spectroscopy system THz-TDS (THz-TDS, Terahertz time-domain spectroscopy, referred to as THz-TDS) to identify the quality changes of black tea with fermentation time. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent perception and monitoring method for the entire fermentation process of Gongfu black tea in view of the deficiencies in the prior art. This method improves the detection efficiency and accuracy for the intelligent monitoring of the entire fermentation process of Gongfu black tea, and has broad application prospects in the field of Gongfu black tea fermentation.

[0006] The technical solution for achieving the purpose of the present invention is:

[0007] An intelligent sensing and monitoring method for the whole fermentation process of Gongfu black tea, comprising the following steps:

[0008] 1) Pretreatment of Gongfu black tea samples: Nine kinds of black tea samples with fermentation time separated by 1 hour were mixed with high-density polyethylene at a mass ratio of 1:3, and then repeatedly ground with an agate mortar and filtered with a 200-mesh sieve. The samples were accurately weighed with a FA2004B electronic analytical balance, and each sample was 200 mg. Then, the sample powder was pressed with a mold at a pressure of 10 MPa for 2 to 3 minutes using an FM-4A tablet press. The diameter of the obtained sample was about 13 mm and the thickness was 1 mm. Then, the temperature was set to 50 degrees Celsius, and the sample was dried in a YB-IA constant temperature drying oven for 30 to 40 minutes;

[0009] 2) THz-TDS was used to irradiate the nine types of processed Gongfu black tea to obtain the transmission source domain spectrum data of each sample, where the THz spectrum range is 0.3~2.0THz;

[0010] 3) performing feature extraction on the source domain THz spectrum data obtained in step 2) by using LDA dimensionality reduction to obtain an input spectrum feature matrix, wherein the dimension of the original spectrum data is reduced from thousands of digits to single digits, and the key features of the original spectrum data are retained;

[0011] 4) The HDBSCAN clustering algorithm and DCNN regression algorithm are used to process and analyze the feature matrix obtained in step 3). The HDBSCAN algorithm uses a hierarchical density clustering method to automatically identify the clustering structure and noise points in the data set. The inter-reachable distance of the HDBSCAN algorithm using a hierarchical density clustering method is:

[0012] d mreach-k (a,b)=max{core k (a),core k (b),d(a,b)},

[0013] where d(a,b) represents the "normal" distance according to the chosen metric, i.e., the Euclidean distance,

[0014] The most stable clusters using the hierarchical density clustering method are:

[0015]

[0016] For a given cluster, if λ birth is the λ value of its parent cluster to be divided into this cluster, and λ dearth The λ value when dividing the cluster into its subclusters, then the stability of cluster M is λ M It is expressed as:

[0017] λ M =λ birth -λ dearth ;

[0018] The convolution of the DCNN regression algorithm using the convolutional neural network method is:

[0019]

[0020] The ReLU activation layer of the convolutional neural network is:

[0021]

[0022] 5) Based on the clustering and regression results, the relationship between the changes in THz sensitive components during the whole process of black tea fermentation was obtained. The regression model with tea polyphenols content as output performed best, indicating that THz waves are most sensitive to changes in tea polyphenols content in black tea during the fermentation process.

[0023] This technical solution is based on the fact that the biological macromolecules of Gongfu black tea have obvious absorption in THz-TDS, so as to obtain the original data of the transmission spectrum, and then reduce the dimension through LDA, and retain the important features in the original spectrum data; then use THz to be sensitive to the changes in the content of tea polyphenols in the fermentation process of Gongfu black tea, use the data after LDA dimension reduction as the input of the DCNN regressor to predict the content of tea polyphenols in the fermentation process of black tea, and finally output the regression result, and finally obtain the result graph of predicting the content of tea polyphenols in different fermented black teas.

[0024] This technical solution adopts HDBSCAN clustering and DCNN regression algorithms, which can simultaneously process a large number of Gongfu black teas with different fermentation times, and can quickly, environmentally friendly, highly accurate and non-destructively judge the fermentation status of Gongfu black tea, and ultimately complete intelligent perception and monitoring of the entire Gongfu black tea fermentation process.

[0025] This method improves the detection efficiency and accuracy for intelligent monitoring of the entire Gongfu black tea fermentation process, and has broad application prospects in the field of Gongfu black tea fermentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of the principle of the DCNN regression algorithm in the embodiment;

[0027] Figure 2 It is a schematic diagram of the qualitative perception effect of the changes in the ingredients of Gongfu black tea at different fermentation times in the embodiment;

[0028] Figure 3 It is a schematic diagram of the quantitative monitoring effect of predicting the tea polyphenol content of Gongfu black tea in the test set regression results in the embodiment. DETAILED DESCRIPTION

[0029] The contents of the present invention are further described below in conjunction with the drawings and embodiments, but the present invention is not limited thereto.

[0030] Example:

[0031] An intelligent sensing and monitoring method for the whole fermentation process of Gongfu black tea, comprising the following steps:

[0032] 1) Nine kinds of black tea samples with fermentation time separated by 1 hour were mixed with high-density polyethylene in a mass ratio of 1:3, and after repeated grinding in an agate mortar, filtered through a 200-mesh sieve, accurately weighed using an FA2004B electronic analytical balance, each sample was 200 mg, and then the sample powder was pressed with a mold at a pressure of 10 MPa for 2 to 3 minutes using an FM-4A tablet press. The diameter of the obtained sample was about 13 mm and the thickness was 1 mm. Then the temperature was set to 50 degrees Celsius, and the sample was dried in a YB-IA constant temperature drying oven for 30-40 minutes;

[0033] 2) THz-TDS was used to irradiate the nine types of processed Gongfu black tea to obtain the transmission source domain spectrum data of each sample, where the THz spectrum range was 0.3-2.0 THz, and 5400 Gongfu black tea samples were irradiated by THz-TDS;

[0034] 3) LDA dimension reduction is used to perform feature extraction on the source domain THz spectrum data obtained in step 2) to obtain an input spectrum feature matrix, wherein the dimension of the original spectrum data is reduced from thousands of digits to single digits, and the key features of the original spectrum data are retained. When the spectrum data is subjected to dimension reduction processing, the target dimension is set to 3 in this example;

[0035] 4) Using HDBSCAN clustering algorithm and Figure 1 The DCNN regression algorithm shown in the figure processes the feature matrix obtained in step 3) of the analysis. The HDBSCAN algorithm uses a hierarchical density clustering method to automatically identify the clustering structure and noise points in the data set. The inter-reachable distance of the hierarchical density clustering method used by the HDBSCAN algorithm is:

[0036] d mreach-k (a,b)=max{core k (a),core k (b),d(a,b)},

[0037] where d(a,b) represents the "normal" distance according to the chosen metric, i.e., the Euclidean distance,

[0038] The most stable clusters using the hierarchical density clustering method are:

[0039]

[0040] For a given cluster, if λ birth is the λ value of its parent cluster to be divided into this cluster, and λ dearth The λ value when dividing the cluster into its subclusters, then the stability of cluster M is λ M It is expressed as:

[0041] λ M =λ birth-λ dearth ;

[0042] The convolution of the DCNN regression algorithm using the convolutional neural network method is:

[0043]

[0044] The ReLU activation layer of the convolutional neural network is:

[0045]

[0046] 5) Based on the clustering and regression results, the relationship between the changes in THz sensitive components during the whole process of black tea fermentation was obtained. The regression model with tea polyphenols content as output performed best, indicating that THz waves are most sensitive to changes in tea polyphenols content in black tea during the fermentation process.

[0047] In this example, 5400 groups of Gongfu black tea samples are used, including nine types of Gongfu black tea samples with different fermentation times. The number of samples for each fermentation time is 600, with a total of 600×9=5400 samples. In each sample, 93.3% of the samples are used as training sets and 6.7% of the samples are used as test sets, that is, each sample has 560 training samples and 40 test samples. The minimum cluster size in the HDBSCAN clusterer is set to 15, and the minimum number of samples is set to 20; in the DCNN regressor, the initial learning rate is set to 0.01, and after training 400 times, the learning rate is adjusted to 0.005 by a decrease factor of 0.5. The model is set to be trained 800 times, and the feature matrix after dimensionality reduction is used as input to construct the HDBSCAN clusterer model and the DCNN regressor model respectively. When evaluating the regression results, 5040 groups of training samples are used to train the model, and 360 groups of samples are used to test the regressor. According to the following Figure 2 As shown in the chart, the clustering accuracy of the THz spectrum features of Gongfu black tea at different fermentation times can reach 97.00%, which shows that the HDBSCAN clusterer can realize intelligent perception of the changes in various components of Gongfu black tea during the fermentation process. Figure 3 As shown in the chart, the determination coefficient R2 of the DCNN regression algorithm with tea polyphenol content as output when processing Gongfu black tea samples with different fermentation times reached 0.99391, which indicates that the DCNN regressor can achieve accurate quantitative monitoring of the tea polyphenol content in Gongfu black tea.

Claims

1. An intelligent sensing and monitoring method for the whole fermentation process of Gongfu black tea, characterized in that: The steps include: 1) Pretreatment of Gongfu black tea samples: Nine kinds of black tea samples with fermentation time separated by 1 hour were mixed with high-density polyethylene at a mass ratio of 1:3, and then repeatedly ground with an agate mortar and filtered with a 200-mesh sieve. The samples were accurately weighed with a FA2004B electronic analytical balance, and each sample was 200 mg. Then, the sample powder was pressed with a mold at a pressure of 10 MPa for 2 to 3 minutes using an FM-4A tablet press. The diameter of the obtained sample was about 13 mm and the thickness was 1 mm. Then, the temperature was set to 50 degrees Celsius, and the sample was dried in a YB-IA constant temperature drying oven for 30 to 40 minutes; 2) THz-TDS was used to irradiate the nine types of processed Gongfu black tea to obtain the transmission source domain spectrum data of each sample, where the THz spectrum range is 0.3~2.0THz; 3) performing feature extraction on the source domain THz spectrum data obtained in step 2) by using LDA dimensionality reduction to obtain an input spectrum feature matrix, wherein the dimension of the original spectrum data is reduced from thousands of digits to single digits, and the key features of the original spectrum data are retained; 4) The HDBSCAN clustering algorithm and DCNN regression algorithm are used to process and analyze the feature matrix obtained in step 3). The HDBSCAN algorithm uses a hierarchical density clustering method to automatically identify the clustering structure and noise points in the data set. The inter-reachable distance of the HDBSCAN algorithm using a hierarchical density clustering method is: d mreach-k (a,b)=max{core k (a),core k (b),d(a,b)}, where d(a,b) represents the "normal" distance according to the chosen metric, i.e., the Euclidean distance, The most stable clusters using the hierarchical density clustering method are: For a given cluster, if λ birth is the λ value of its parent cluster to be divided into this cluster, and λ dearth The λ value when dividing the cluster into its subclusters, then the stability of cluster M is λ M It is expressed as: l M =λ birth -l dearth ; The convolution of the DCNN regression algorithm using the convolutional neural network method is: The ReLU activation layer of the convolutional neural network is: 5) Based on the clustering and regression results, the relationship between the changes in THz sensitive components during the fermentation process of Gongfu black tea was obtained. The regression model with tea polyphenols content as output performed best, indicating that THz waves are most sensitive to changes in tea polyphenols content in black tea during the fermentation process.

Images