Pyrazine-flavor-based multi-mode intelligent identification method for Maotai-flavor liquor process

Abstract

The invention discloses a pyrazine-flavor-based multi-mode intelligent identification method for a Maotai-flavor liquor process. The method comprises the following steps: pretreatment: pretreating Maotai-flavor liquor samples of different brewing processes; collecting sensory data of electronic noses and electronic tongues of Maotai-flavor baijiu of different brewing processes; establishing a liquid phase detection method for the characteristic components of the alkylpyrazine in the Maotai-flavor liquor; verifying a component analysis method through precision, stability and repeatability experiments; component quantitative analysis based on liquid chromatography fluorescence detection; performing clustering and difference evaluation on Maotai-flavor liquor samples of different brewing processes by adopting multivariate statistical analysis; establishing a Maotai-flavor liquor process identification model: utilizing VIPgt in multivariate statistics; establishing different machine learning multi-classification models for different brewing processes of Maotai-flavor liquor according to the common characteristics of 1, and performing multi-model comparison; and based on the optimal performance classification model, establishing an SHAP model of the Maotai-flavor liquor brewing process identification model, and globally explaining the identification model by adopting an SHAP algorithm.

Description

Technical Field

[0001] Specifically, this involves a method for identifying the brewing process of Maotai-flavor liquor based on high-performance liquid chromatography-fluorescence detection (HPLC-FLD) combined with intelligent sensory technology and machine learning. This method is used to quickly and accurately identify Maotai-flavor liquors with different brewing processes and evaluate their quality. Background Technology

[0002] As an important category of traditional Chinese baijiu, Maotai-flavor baijiu is deeply loved by consumers for its unique flavor and brewing techniques. Its brewing process is complex, involving multiple stages such as raw material selection, fermentation time, and distillation methods. Different processes directly affect the flavor composition and quality of the baijiu. Currently, Maotai-flavor baijiu sold commercially mainly employs three brewing processes: Kunsha, Suisha, and Fansha. Baijiu produced using different processes varies in quality and market value. Traditional baijiu quality evaluation and process identification primarily rely on sensory evaluation and single chemical component analysis, which suffers from strong subjectivity, low differentiation, and an inability to systematically reflect the complex differences in flavor compounds.

[0003] Alkylpyrazine compounds (such as tetramethylpyrazine, trimethylpyrazine, and 2,6-dimethylpyrazine) have been proven to be key flavor compounds and process markers in Maotai-flavor liquor. Their content is significantly correlated with the fermentation cycle, and their content and distribution directly reflect the complexity and quality of the brewing process. However, traditional detection methods for alkylpyrazines (such as GC-MS) suffer from problems such as complex pretreatment and low sensitivity. This invention utilizes their inherent fluorescence properties to establish a dedicated HPLC-FLD detection method, and combines electronic sensing and machine learning; multimodal fusion can improve the accuracy and efficiency of identification.

[0004] High-performance liquid chromatography-fluorescence (HPLC-FLD) boasts advantages such as high sensitivity, high separation efficiency, and good repeatability, enabling effective separation and quantitative analysis of various flavor compounds (e.g., esters, alcohols, aldehydes, acids) in baijiu. However, existing technologies are largely limited to the detection of specific compounds, lacking research on the differences in certain components or classes of components related to the brewing process of Maotai-flavor baijiu, and failing to incorporate machine learning for intelligent classification and process identification. Therefore, a comprehensive method integrating HPLC-FLD detection with intelligent sensory processing and machine learning models is needed. This method should utilize multimodal analysis in various forms, including chemometrics, intelligent taste and olfaction, and machine learning, to improve the accuracy and efficiency of identifying the brewing process and evaluating the quality of Maotai-flavor baijiu. This technology can effectively distinguish traditional Maotai-flavor baijiu processes, resolving the current issue of process confusion in the Maotai-flavor baijiu market. Summary of the Invention

[0005] This invention proposes a multimodal intelligent identification method for the process of Maotai-flavor liquor based on pyrazine flavor. It provides a more objective and intelligent method for identifying and evaluating the process of Maotai-flavor liquor based on alkylpyrazine characteristic components, and solves the shortcomings of traditional methods in terms of accuracy, identification efficiency and automation.

[0006] The technical solution of this invention provides a method for identifying the brewing process of Maotai-flavor liquor based on HPLC-FLD combined with a machine learning model, comprising eight steps:

[0007] Step 1: Perform pretreatment on the liquor sample based on HPLC and electronic nose analysis, including four processes: dilution, vortexing, sonication, and filtration.

[0008] Step 2: Collect electronic nose sensory data of sauce-flavored baijiu produced using different brewing processes;

[0009] Step 3: Collect electronic tongue sensory data of sauce-flavored baijiu produced using different brewing processes;

[0010] Step 4: Establishment of a liquid phase detection method for characteristic alkylpyrazine components in Maotai-flavor liquor: The reliability of the component analysis method was verified through precision, stability, and repeatability experiments.

[0011] Step 5: Quantitative analysis of components based on liquid chromatography fluorescence detection: Collect chromatograms of sauce-flavored baijiu from different brewing processes, import the chromatograms of baijiu samples from different brewing processes into the software, and use the area normalization method to quantitatively analyze the characteristic components of alkylpyrazine in baijiu based on the standard curve established in Step 3.

[0012] Step 6: Use multivariate statistical analysis to cluster and evaluate the differences among samples of Maotai-flavor liquor produced using different brewing processes;

[0013] Step 7: Establishment of the identification model for the brewing process of Maotai-flavor liquor: Using the common feature of VIP>1 in multivariate statistics, different machine learning multi-classification models are established for different brewing processes of Maotai-flavor liquor, and the multiple models are compared.

[0014] Step 8: Based on the best-performing classification model, establish the SHAP model for the identification model of the brewing process of Maotai-flavor liquor. Use the SHAP algorithm to perform a global interpretation of the identification model: calculate the SHAP value of HPLC-FLD features and intelligent sensory response values; clarify the contribution of each feature to the process identification results and the threshold conditions through feature importance ranking and dependency graph analysis.

[0015] Furthermore, in step one, the specific steps include: accurately taking 10 ml of the liquor sample, placing it in a 100 mL volumetric flask, adding ultrapure water to make up to 100 mL, vortexing at room temperature for 2 min, sonicating for 15 min, diluting and letting it stand for more than 30 min, filtering through a 0.22 μm filter membrane, placing it in a sample bottle, and then performing HPLC-FLD and electronic nose determination.

[0016] Furthermore, in step two, the electronic nose device used in the experiment includes 10 metal oxide gas sensors: W1C (aromatic components), W5S (nitrogen oxides), W3C (ammonia and aromatic components), W6S (hydrogen), W5C (alkanes and aromatic components), W1S (short-chain alkanes), W1W (inorganic sulfides), W2S (alcohols, ethers, aldehydes, and ketones), W2W (aromatic components and organic sulfides), and W3S (long-chain alkanes). Before detection, 15 mL of the prepared liquor sample is placed in a 60 mL headspace vial, the cap is tightened, and the vial is kept in place for 30 minutes before detection begins. Each liquor sample is measured in triplicate under the following conditions: analysis sampling time of 80 s, sensor cleaning time of 120 s, sensor zeroing time of 5 s, sample preparation time of 5 s, and injection flow rate of 400 ml / min.

[0017] Furthermore, in step three, the electronic tongue device used in the experiment was the TS-5000Z taste analysis system (abbreviated as electronic tongue, Insent, Japan). It evaluates five basic tastes (sour, astringent, bitter, salty, and umami) and sweetness by detecting changes in membrane potential generated by electrostatic or hydrophobic interactions between various taste substances and artificial lipid membranes. To obtain the sample potential, 90 ml of diluted baijiu (10% vol) sample was precisely pipetted into the corresponding small measuring cup, and then the sensor was immersed in each sample for 30 seconds. The experiment was conducted at room temperature, and the detection probe was cleaned with distilled water before and after each test. All analyses were performed three times. Test solutions: reference solution (artificial saliva) was KCl + tartaric acid; negative electrode cleaning solution was water + ethanol + HCl; positive electrode cleaning solution was KCl + water + ethanol + KOH.

[0018] Furthermore, in step five, the chromatographic conditions for collecting the baijiu samples by HPLC are as follows: a Sunfire C18 column (100A, 5μm, 4.6mm×250mm); column temperature 40℃; mobile phase A (0.1% trifluoroacetic acid + 0.1% formic acid): phase B (acetonitrile) = 93:7; flow rate 1mL / min; total run time for each sample 38min; gradient elution program: 0-17min, 93%A and 7%B; 17-18min, 93%-85%A and 7%-15%B; 18-35min, 85%-93%A; 35-38min, 93%A; chromatograms are recorded at emission wavelengths of 340nm and excitation wavelengths of 271nm.

[0019] Furthermore, step six specifically includes the following steps:

[0020] (a) Construction of a classification model based on principal component analysis; (b) Construction of a classification model based on partial least squares method.

[0021] Furthermore, step six is ​​divided into the following operations:

[0022] (a) Classification model construction using Logistic Regression; (b) Classification model construction using CatBoost; (c) Classification model construction using Artificial Neural Network (ANN); (d) Classification model construction using Decision Tree (DT); (e) Classification model construction using Random Forest (RF); (f) Classification model construction using Support Vector Machine (SVM); (g) Classification model construction using K-Nearest Neighbors (KNN); (h) Classification model construction using XGBoost; (i) Classification model construction using LightGBM.

[0023] Furthermore, the datasets for the nine classification models are divided into training and validation sets. Based on the scientific classification ratio of model training, we divide the training and validation sets in an 8:2 ratio, with 158 data points used for training and 40 data points used for testing and validation.

[0024] Furthermore, in the Catboost and XGBoost models, Bayesian optimization is used to find the optimal combination of hyperparameters for the classifier through the Tree-structured Parzen Estimators algorithm from the Hyperopt library. The search space covers three key parameters: tree depth (range 3-10), learning rate (range 0.01-0.3), and number of iterations (range 50-300). The optimization objective is to minimize 1-accuracy (equivalent to maximizing classification accuracy), and 5-fold cross-validation is used to ensure the model's generalization ability.

[0025] Furthermore, the soy sauce-flavored liquor includes representative brewing processes such as "fan sha," "sui sha," and "kun sha."

[0026] The beneficial effects of this invention after adopting the above technical solutions are as follows: This study systematically analyzed Maotai-flavor liquor produced using different brewing processes by combining high-performance liquid chromatography-fluorescence detection (HPLC-FLD) with intelligent sensory technology, multivariate statistical analysis, and machine learning. This enabled the identification of Maotai-flavor liquor production processes. Through partial least squares discriminant analysis (PLS-DA), key alkylpyrazine markers in Maotai-flavor liquor were successfully identified. Simultaneously, a Maotai-flavor liquor production process evaluation model based on principal component analysis (PCA) was constructed, clarifying the differences in chemical characteristics of Maotai-flavor liquor produced using different brewing processes, and feature mining was performed through cluster analysis (K-means). Furthermore, a Maotai-flavor liquor production process identification model was established using machine learning algorithms such as XGBoost, Catboost, and RF, achieving an accuracy rate of up to 100%. This not only deepened the understanding of the key flavor formation mechanism of Maotai-flavor liquor but also provided technical support and theoretical basis for the standardization of Maotai-flavor liquor quality and the identification of genuine and counterfeit products. This achieved objective, accurate, and efficient identification of Maotai-flavor liquor production processes, effectively curbing counterfeit products and protecting market share and profits. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0028] Figure 1 This is an overview diagram of the technical flow of the present invention;

[0029] Figure 2 This is a graph showing the analysis results of the electronic nose for different brewing processes of Maotai-flavor liquor according to the present invention.

[0030] Figure 3 The results of electronic tongue analysis of sauce-flavored Baijiu produced by different brewing processes according to the present invention are shown in the figure.

[0031] Figure 4 HPLC-FLD chromatogram of the mixed standard solution of the present invention;

[0032] Figure 5 This is a K-means clustering analysis diagram of the Maotai-flavor liquor samples of the present invention;

[0033] Figure 6 Cluster heatmap of alkylpyrazine compound content in soy sauce-flavored Baijiu produced using different brewing processes according to the present invention;

[0034] Figure 7The image shows the PLS-DA analysis results of sauce-flavored baijiu samples produced using different brewing processes according to the present invention.

[0035] Figure 8 The confusion matrix and ROC curve of the CatBoost multi-class classification model of this invention are shown below.

[0036] Figure 9 The confusion matrix and ROC curve of the random forest multi-classification model of this invention are shown below.

[0037] Figure 10 The confusion matrix and ROC curve of the XGBoost multi-class classification model of this invention are shown below.

[0038] Figure 11 The confusion matrix and ROC curve of the SVM multi-classification model of this invention are shown below.

[0039] Figure 12 Here is a confusion matrix diagram of the nine machine learning models of this invention;

[0040] Figure 13 Here are the ROC curves for the nine machine learning models of this invention;

[0041] Figure 14 This is a comparison chart of multiple performance indicators for the nine machine learning models of this invention;

[0042] Figure 15 A honeycomb diagram illustrating the importance of shap characteristics for different brewing processes of sauce-flavored baijiu according to the present invention.

[0043] Figure 16 This is a feature dependency diagram of different brewing processes for soy sauce-flavored baijiu according to the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] like Figure 1-16 As shown, this invention provides a multimodal intelligent identification method for the brewing process of Maotai-flavor liquor based on pyrazine flavor. In this embodiment, Maotai-flavor liquors from three different brewing processes in different production areas across the country are randomly selected as research objects.

[0046] Example 1

[0047] A multimodal intelligent identification method for the production process of Maotai-flavor liquor based on pyrazine flavor includes the following steps (three groups of samples are processed and tested in the same batch):

[0048] (1) Accurately take 10 ml of the liquor sample and place it in a 100 mL volumetric flask. Add ultrapure water to make up to 100 mL. Sonicate at room temperature for 15 min. After dilution, let it stand for more than 30 min. Before detection, place 15 mL of the prepared liquor sample in a 60 mL headspace vial, tighten the cap and keep it for 30 min before starting the detection. Each liquor sample is measured in parallel three times. The measurement conditions are as follows: analysis sampling time is 80 s, sensor cleaning time is 120 s, sensor zeroing time is 5 s, and sample preparation time is 5 s. The sample is pumped into the sensor array at an injection flow rate of 400 ml / min as shown in Table 1.

[0049] Table 1 Sensor performance of PEN3.5 electronic nose

[0050]

[0051]

[0052] (2) Import the detection results of the 10 sensors of the electronic nose into the software and analyze them using principal component analysis. The results are as follows: Figure 2 As shown.

[0053] (3) Establishment of a model for identifying the brewing process of Maotai-flavor liquor, including the establishment of a CatBoost machine learning classification model, such as... Figure 8 As shown, the test set accuracy is 0.986, the test set precision is 0.988, the test set recall is 0.961, the test set F1 score is 1.0, the test set Kappa coefficient is 1.0, the 5-fold cross-validation score accuracy is averaged 0.911, and the accuracy standard deviation is 0.083. The overall model performance is excellent.

[0054] Example 2

[0055] A multimodal intelligent identification method for the production process of Maotai-flavor liquor based on pyrazine flavor includes the following steps (three groups of samples are processed and tested in the same batch):

[0056] (1) Accurately take 10 ml of the liquor sample, place it in a 100 mL volumetric flask, add ultrapure water to make up to 100 mL, sonicate at room temperature for 15 min, dilute and let stand for more than 30 min, filter through a 0.22 μm filter membrane, place in a sample bottle, and enter HPLC-FLD for determination.

[0057] (2) Validity of the liquid chromatography method and spectral analysis: The reliability of the analytical method for alkylpyrazine content in baijiu was verified through precision, stability, and repeatability experiments, such as... Figure 3Chromatographic data were recorded, and tetramethylpyrazine (peak 1) was used as the reference peak in the chromatographic data analysis. The relative retention time, relative peak area, and peak area of ​​the common peaks were calculated. The results are summarized in Table 2.

[0058] Table 2. Relative Standard Deviation (RSD) values ​​for method validation

[0059]

[0060] (3) Establishment of a method for identifying the process of Maotai-flavor liquor: The relative peak areas of each common peak are imported into the software for cluster analysis, and the results are as follows. Figure 5 As shown.

[0061] (4) Establishment of a model for identifying the production process of sauce-flavored baijiu, including the establishment of random forest machine learning classification models, such as... Figure 9 As shown in the figure. The results show that the test set accuracy is 0.936, the test set precision is 0.954, the test set recall is 0.947, the test set F1 score is 0.949, the test set Kappa coefficient is 0.859, the mean 5-fold cross-validation score accuracy is 0.808, and the standard deviation of accuracy is 0.184. The model performs well overall.

[0062] Example 3

[0063] A multimodal intelligent identification method for the production process of Maotai-flavor liquor based on pyrazine flavor includes the following steps (three groups of samples are processed and tested in the same batch):

[0064] (1) Accurately take 10 ml of the liquor sample, place it in a 100 mL volumetric flask, add ultrapure water to make up to 100 mL, sonicate at room temperature for 15 min, dilute and let stand for more than 30 min, filter through a 0.22 μm filter membrane, place in a sample bottle, and enter HPLC-FLD for determination.

[0065] (2) Validity of the liquid chromatography method and spectra: The reliability of the fingerprint analysis method was verified through precision, stability, and repeatability experiments. Standard curves for 13 alkylpyrazine standards were established using X: concentration (mg / L) and Y: peak area (AU), and the detection limit and quantitation limit of the HPLC-FLD detection method were calculated. The results are shown in Table 2.

[0066] Table 2. Standard curve equation and R value for APZs analysis using HPLC-FLD. 2 Limit of detection and limit of quantitation, linear range

[0067]

[0068]

[0069] (3) Quantitative analysis of alkylpyrazines in Baijiu: The content of alkylpyrazines in Maotai-flavor Baijiu produced by different brewing processes was detected. The HPLC-FLD detection results were imported into the software to establish a cluster heatmap of APZs content. The results are as follows: Figure 5 As shown.

[0070] (4) Establishment of a model for identifying the brewing process of sauce-flavored baijiu: A random forest machine learning classification model was established, and the results are as follows. Figure 9 As shown in the figure. The results show that the test set accuracy is 0.905, the test set precision is 0.919, the test set recall is 0.919, the test set F1 score is 0.919, the test set Kappa coefficient is 0.848, the mean 5-fold cross-validation score accuracy is 0.708, and the standard deviation of accuracy is 0.171. The model performs well overall.

[0071] Example 4

[0072] A multimodal intelligent identification method for the production process of Maotai-flavor liquor based on pyrazine flavor includes the following steps (three groups of samples are processed and tested in the same batch):

[0073] (1) Accurately take 10 ml of the liquor sample, place it in a 100 mL volumetric flask, add ultrapure water to make up to 100 mL, sonicate at room temperature for 15 min, dilute and let stand for more than 30 min, filter through a 0.22 μm filter membrane, and place it in a sample bottle for testing.

[0074] (2) Electronic nose detection of Maotai-flavor liquor: Before testing, the sample was transferred into a headspace vial and allowed to incubate at 25°C for 10 minutes. The self-cleaning time of the sample was set to 100 s. The sample was diluted 10 times and pumped into the sensor array at 400 mL / min for 80 s. The stability data of the curve from 57 s to 59 s were used for PCA analysis.

[0075] (3) Quantitative analysis of alkylpyrazines in baijiu: The content of alkylpyrazines in sauce-flavored baijiu with different brewing processes was detected by HPLC-FLD.

[0076] (4) Establishment of identification method for sauce-flavored Baijiu production process: The above-mentioned test indicators were imported into the software, and a model based on partial least squares method was established. The results are as follows: Figure 5 As shown.

[0077] (5) Establishment of a model for identifying the brewing process of sauce-flavored Baijiu, and establishment of a classification model using XGBoost learning. The results are as follows: Figure 9As shown. Test set accuracy: 0.905, test set precision: 0.933, test set recall: 0.926, test set F1 score: 0.921, test set Kappa coefficient: 0.851, folded cross-validation score accuracy mean: 0.918, accuracy standard deviation: 0.075. The model performs well overall.

[0078] Example 5

[0079] A multimodal intelligent identification method for the production process of Maotai-flavor liquor based on pyrazine flavor includes the following steps (three groups of samples are processed and tested in the same batch):

[0080] (1) Accurately take 10 ml of the liquor sample, place it in a 100 mL volumetric flask, add ultrapure water to make up to 100 mL, sonicate at room temperature for 15 min, dilute and let stand for more than 30 min, filter through a 0.22 μm filter membrane, and place it in a sample bottle for testing.

[0081] (2) Electronic tongue detection of Maotai-flavor liquor: The evaluation of five basic flavors (sour, astringent, bitter, salty, and umami) and sweetness is achieved by detecting changes in membrane potential generated by electrostatic or hydrophobic interactions between various flavor substances and the artificial lipid membrane. To obtain the sample potential, 90 ml of diluted Maotai-flavor liquor (10% vol) sample was precisely pipetted into the corresponding small measuring cup, and then the sensor was immersed in each sample for 30 seconds. The experiment was conducted at room temperature, and the detection probe was cleaned with distilled water before and after each test. The analysis was performed three times.

[0082] (3) Quantitative analysis of alkylpyrazines in baijiu: The content of alkylpyrazines in sauce-flavored baijiu with different brewing processes was detected by HPLC-FLD.

[0083] (4) Establishment of identification method for sauce-flavored Baijiu production process: The above-mentioned test indicators were imported into the software, and a model based on partial least squares method was established. The results are as follows: Figure 5 As shown.

[0084] (5) Establishment of a model for identifying the brewing process of sauce-flavored Baijiu, and establishment of a classification model using an SVM machine learning model. The results are as follows: Figure 11 As shown in the figure. The results show that the test set accuracy is 0.808, the test set precision is 0.816, the test set recall is 0.801, the test set F1 score is 0.801, the test set Kappa coefficient is 0.709, the mean of the 5-fold cross-validation score accuracy is 0.946, and the standard deviation of accuracy is 0.065. The overall performance of the model is good.

[0085] Example 6

[0086] A multimodal intelligent identification method for the production process of Maotai-flavor liquor based on pyrazine flavor includes the following steps (three groups of samples are processed and tested in the same batch):

[0087] (1) Accurately take 10 ml of the liquor sample, place it in a 100 mL volumetric flask, add ultrapure water to make up to 100 mL, sonicate at room temperature for 15 min, dilute and let stand for more than 30 min, filter through a 0.22 μm filter membrane, and place it in a sample bottle for testing.

[0088] (2) Electronic tongue detection of Maotai-flavor liquor: The evaluation of five basic flavors (sour, astringent, bitter, salty, and umami) and sweetness is achieved by detecting changes in membrane potential generated by electrostatic or hydrophobic interactions between various flavor substances and the artificial lipid membrane. To obtain the sample potential, 90 ml of diluted Maotai-flavor liquor (10% vol) sample was precisely pipetted into the corresponding small measuring cup, and then the sensor was immersed in each sample for 30 seconds. The experiment was conducted at room temperature, and the detection probe was cleaned with distilled water before and after each test. The analysis was performed three times.

[0089] (3) Electronic nose detection of Maotai-flavor liquor: Before testing, the sample was transferred into a headspace vial and allowed to incubate at 25°C for 10 minutes. The self-cleaning time of the sample was set to 100 s. The sample was diluted 10 times and pumped into the sensor array at 400 mL / min for 80 s. The stability data of the curve from 57 s to 59 s were used for PCA analysis.

[0090] (4) Quantitative analysis of alkylpyrazines in baijiu: The content of alkylpyrazines in sauce-flavored baijiu with different brewing processes was detected by HPLC-FLD.

[0091] (5) Establishment of identification method for sauce-flavored Baijiu production process: The above-mentioned test indicators were imported into the software, and a model based on partial least squares method was established. The results are as follows: Figure 5 As shown.

[0092] (6) Establishment of a model for identifying the brewing process of sauce-flavored Baijiu: Nine machine learning classification models were established, and the results are as follows. Figure 12 As shown.

[0093] (7) Comparison and screening of identification models for the brewing process of Maotai-flavor liquor, results are as follows: Figure 13 , Figure 14 As shown.

[0094] (8) A SHAP model for identifying the brewing process of Maotai-flavor liquor was established and globally interpreted. The results are as follows: Figure 15 , Figure 16 As shown.

[0095] Table 3. Prediction results of identification models for Maotai-flavor liquor with different brewing processes.

[0096]

[0097] This study employed electronic nose analysis and HPLC-FLD technology to conduct in-depth analysis of Maotai-flavor Baijiu produced using different brewing processes. Preliminary differences in volatile components and quantitative analysis of alkylpyrazines were achieved. Multivariate statistical analysis was used to screen for differentially expressed components in Maotai-flavor Baijiu produced using different brewing processes. Furthermore, this study constructed nine machine learning classification models for identifying the brewing processes of Maotai-flavor Baijiu, with the Catboost model achieving 100% accuracy. Further, based on Catboost, a SHAP model for identifying the brewing processes of Maotai-flavor Baijiu was established, determining the contribution of each detection index to the identification of Maotai-flavor Baijiu processes. This not only expands our understanding of the Maotai-flavor flavor of Maotai-flavor Baijiu but also provides theoretical basis and technical support for research on Maotai-flavor Baijiu.

[0098] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multimodal intelligent identification method for the production process of soy sauce-flavored baijiu based on pyrazine flavor, characterized in that, Includes eight steps: Step 1: Perform pretreatment on the liquor sample based on HPLC and electronic nose analysis, including four processes: dilution, vortexing, sonication, and filtration. Step 2: Collect electronic nose sensory data of sauce-flavored baijiu produced using different brewing processes; Step 3: Collect electronic tongue sensory data of sauce-flavored baijiu produced using different brewing processes; Step 4: Establishment of a liquid phase detection method for characteristic alkylpyrazine components in Maotai-flavor liquor: The reliability of the component analysis method was verified through precision, stability, and repeatability experiments. Step 5: Quantitative analysis of components based on liquid chromatography fluorescence detection: Collect chromatograms of sauce-flavored baijiu from different brewing processes, import the chromatograms of baijiu samples from different brewing processes into the software, and use the area normalization method to quantitatively analyze the characteristic components of alkylpyrazine in baijiu based on the standard curve established in Step 3. Step 6: Use multivariate statistical analysis to cluster and evaluate the differences among samples of Maotai-flavor liquor produced using different brewing processes; Step 7: Establishment of the identification model for the brewing process of Maotai-flavor liquor: Using the common feature of VIP>1 in multivariate statistics, different machine learning multi-classification models are established for different brewing processes of Maotai-flavor liquor, and the multiple models are compared. Step 8: Based on the best performing classification model, establish the SHAP model for the identification model of the brewing process of Maotai-flavor liquor, and use the SHAP algorithm to perform global interpretation of the identification model: calculate the SHAP value of HPLC-FLD characteristics and intelligent sensory response values; By ranking the importance of features and performing dependency graph analysis, the contribution of each feature to the process identification results and the threshold conditions are clarified.

2. The method according to claim 1, characterized in that, The specific steps in step one include: accurately taking 10 ml of the liquor sample, placing it in a 100 mL volumetric flask, adding ultrapure water to make up to 100 mL, vortexing at room temperature for 2 min, sonicating for 15 min, diluting and letting it stand for more than 30 min, filtering through a 0.22 μm filter membrane, placing it in a sample bottle, and then performing HPLC-FLD and electronic nose analysis.

3. The method according to claim 2, characterized in that, In step two, the baijiu sample is analyzed using an electronic nose, including analysis of aromatic components W1C, nitrogen oxides W5S, ammonia and aromatic components W3C, hydrogen W6S, alkanes and aromatic components W5C, short-chain alkanes W1S, inorganic sulfides W1W, alcohols, ethers, aldehydes and ketones W2S, aromatic components and organic sulfides W2W, and long-chain alkanes W3S; and 10 metal oxide gas sensors. Before detection, 15 mL of the prepared baijiu sample is placed in a 60 mL headspace vial and kept for 30 min before starting the analysis. Each sample is measured in triplicate under the following conditions: analysis sampling time of 80 s, sensor cleaning time of 120 s, sensor zeroing time of 5 s, sample preparation time of 5 s, and injection flow rate of 400 ml / min.

4. The method according to claim 3, characterized in that, In step three, electronic tongue analysis was performed on the baijiu samples: using the TS-5000Z taste analysis system (abbreviated as electronic tongue, Insent, Japan), the changes in membrane potential generated by electrostatic or hydrophobic interactions between various taste substances and artificial lipid membranes were used to evaluate five basic tastes (sour, astringent, bitter, salty, and umami) and sweetness. To obtain the sample potential, 90 ml of diluted baijiu (10% vol) sample was precisely pipetted into the corresponding small measuring cup, and then the sensor was immersed in each sample for 30 seconds. The experiment was conducted at room temperature, and the detection probe was cleaned with distilled water before and after each test. All analyses were performed three times. Test solutions: reference solution (artificial saliva) was KCl + tartaric acid; negative electrode cleaning solution was water + ethanol + HCl; positive electrode cleaning solution was KCl + water + ethanol + KOH.

5. The method according to claim 4, characterized in that, Step five consists of the following operations: Chromatographic conditions for HPLC-FLD analysis of Baijiu samples: Sunfire C18 column (100A, 5μm, 4.6mm×250mm); column temperature 40℃; mobile phase A (0.1% trifluoroacetic acid + 0.1% formic acid): phase B (acetonitrile) = 93:7; flow rate 1mL / min; total run time for each sample 38 min; gradient elution program: 0-17 min, 93%A and 7%B; 17-18 min, 93%-85%A and 7%-15%B; 18-35 min, 85%-93%A; 35-38 min, 93%A; chromatograms were recorded at fluorescence emission wavelengths of 340nm and excitation wavelengths of 271nm.

6. The method according to claim 5, characterized in that, Step six consists of the following operations: (a) model building based on K-means clustering analysis; (b) model building based on principal component analysis; and (c) model building based on partial least squares.

7. The method according to claim 6, characterized in that, Step seven consists of the following operations: (a) Classification model construction using Logistic Regression; (b) Classification model construction using CatBoost; (c) Classification model construction using Artificial Neural Network (ANN); (d) Classification model construction using Decision Tree (DT); (e) Classification model construction using Random Forest (RF); (f) Classification model construction using Support Vector Machine (SVM); (g) Classification model construction using K-Nearest Neighbors (KNN); (h) Classification model construction using XGBoost; (i) Classification model construction using LightGBM.

8. The method according to claim 7, characterized in that, The soy sauce-flavored liquor includes representative brewing techniques such as "fan sha," "sui sha," and "kun sha." 9. The method according to claim 8, characterized in that, The datasets for the nine classification models are divided into training and validation sets. Based on the scientific classification ratio of model training, we divide the training and validation sets in an 8:2 ratio, with 158 data points used for training and 40 data points used for testing the accuracy of the models.

10. The method according to claim 9, characterized in that, In the Catboost and XGBoost models, Bayesian optimization is used to find the optimal combination of hyperparameters for the classifier through the Tree-structured Parzen Estimators algorithm from the Hyperopt library. The search space covers three key parameters: tree depth (range 3-10), learning rate (range 0.01-0.3), and number of iterations (range 50-300). The optimization objective is to minimize 1 - accuracy (equivalent to maximizing classification accuracy), and 5-fold cross-validation is used to ensure the model's generalization ability.

Images