A raw tea sensory classification and flavor critical threshold extraction method and system
By using multi-source data fusion and machine learning methods, the consistency and accuracy of flavor grading of raw tea for new tea drinks have been solved, achieving low-cost and efficient flavor detection and quantitative standard setting, which meets the needs of industrial production.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CETC BIGDATA RES INST CO LTD
- Filing Date
- 2026-04-30
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies for flavor grading and screening of raw tea for new tea drinks suffer from poor consistency and low repeatability due to reliance on human experience. Furthermore, the accuracy of prediction models based on single test data is limited, failing to meet the high precision and efficiency requirements of industrial production. Additionally, they lack the ability to reverse-engineer high-dimensional and complex physicochemical test data into quantitative standards.
By employing multi-source data fusion and machine learning methods, macroscopic sensor data is acquired through electronic nose and electronic tongue, and microscopic physicochemical component data is acquired by combining targeted omics technology. The random forest algorithm is used to screen core features, a limit gradient boosting tree multi-classification model is trained, and a SHAP tree interpreter is used to calculate local marginal contribution values, construct dependency function mapping, find the critical threshold of physical concentration, and realize the formulation of quantitative flavor standards.
It achieves low-cost and high-efficiency flavor detection of raw tea, overcomes the shortcomings of black-box models, provides visualized quantitative standards, provides precise flavor management basis for industrial production, and improves prediction accuracy and system robustness.
Smart Images
Figure CN122132928A_ABST
Abstract
Description
Technical Field
[0001] This invention discloses a method and system for sensory classification and flavor critical threshold extraction of raw tea, belonging to the field of artificial intelligence and machine learning technology. Background Technology
[0002] Currently, the flavor grading and selection of raw tea for new tea drinks mainly rely on manual evaluation experience or conventional prediction models based on single test data. Manual evaluation methods heavily depend on the sensory experience and subjective judgment of evaluators, resulting in inherent drawbacks such as poor consistency, low repeatability, and difficulty in large-scale implementation. Conventional prediction models based on single test data often only utilize macroscopic sensor data such as electronic noses and electronic tongues or a few physicochemical indicators for flavor discrimination, making it difficult to comprehensively capture the synergistic and antagonistic effects between the complex biochemical components that determine tea flavor. This leads to limited prediction accuracy and fails to meet the high-precision, high-efficiency quality control requirements of industrial production.
[0003] In existing intelligent sensory evaluation technologies, some solutions attempt to introduce high-throughput omics detection techniques to obtain more comprehensive physicochemical information. However, directly performing full-dimensional omics detection on massive samples is costly and time-consuming. There is a lack of effective mechanisms for efficient collaboration and dimensionality reduction screening of macroscopic sensory data and microscopic targeted omics data, making it difficult to achieve low-cost, high-efficiency industrial-scale surveys. Furthermore, existing intelligent sensory predictions mostly employ traditional neural networks or basic machine learning models. These models are often considered black boxes; the system can only output the final classification result or score, but cannot accurately explain which specific core chemical components dominate the flavor profile, resulting in a lack of credibility and traceability in the model output.
[0004] More critically, existing technologies generally lack the ability to reverse engineer high-dimensional and complex physicochemical testing data into quantitative standards that can guide actual industrial production. Specifically, existing methods cannot dynamically attribute and accurately extract the core chemical components and their critical physical concentration thresholds that trigger specific sensory attributes such as high aroma and rich flavor. This leads to a serious disconnect between AI predictions and tea companies' actual procurement, blending, and flavor component standardization practices. This deficiency makes it difficult for intelligent sensory evaluation technology to be truly implemented in industrial production, and it cannot provide accurate and quantitative scientific basis for the standardized management of raw tea flavor. Summary of the Invention
[0005] This application provides a method for sensory classification and flavor critical threshold extraction of raw tea, applied to a computer system, including the following steps: Step S1: Obtain multi-source data from multiple raw tea samples. The multi-source data includes macroscopic sensing data collected using an electronic nose and electronic tongue, absolute concentration data of microscopic physicochemical components collected using targeted omics technology, and corresponding artificial sensory evaluation labels. Step S2: Based on the random forest algorithm, the absolute concentration data of the microscopic physicochemical components are evaluated for feature importance, and the top N physicochemical components with the highest contribution are selected as the core feature subset. The concentration of each component in the core feature subset is normalized to generate standard radar fingerprint maps corresponding to different sensory categories. Step S3: Using the subset of core features as input and the artificial sensory evaluation labels as output, train the extreme gradient boosting tree multi-classification model, and use the SHAP tree interpreter to calculate the local marginal contribution value of each core feature to the model prediction result. Step S4: For the target sensory category, construct a dependency function mapping between the absolute physical concentration of the target core feature and its SHAP local marginal contribution value, and traverse the mapping through an algorithm to find the zero-value crossover point that causes the model prediction result to reverse. The absolute physical concentration corresponding to the crossover point is established as the physical concentration critical threshold for the target core feature to trigger the target sensory category. Step S5: Receive the physicochemical test data of the raw tea sample to be tested, extract only the component concentrations corresponding to the core feature subset, input the trained limit gradient boosting tree multi-classification model, obtain the sensory classification prediction results and their confidence levels, and use the physical concentration critical threshold as the judgment basis to output a quantitative flavor standard for guiding industrial production.
[0006] Furthermore, step S2 also includes: Construct a random forest classification model containing multiple decision trees, and calculate the average decrease in Gini impurity caused by each physicochemical component as a splitting node in all decision trees; The average decrease is used as the feature importance score of the physicochemical component, and the components are sorted from high to low scores to select the top N physicochemical components to form the core feature subset.
[0007] Furthermore, step S2 also includes: The concentrations of each component in the core feature subset are normalized, and the absolute concentrations of each component are mapped to the interval of 0, 10, 1 using range transformation to eliminate the dimensional differences between different physicochemical components. The process involves generating standard radar fingerprint maps corresponding to different sensory categories, calculating the normalized concentration mean of each component of the core feature subset of all samples under the same sensory category, and plotting the mean sequence as a closed polygon in polar coordinates to form the standard radar fingerprint map for that sensory category.
[0008] Furthermore, step S3 also includes: The extreme gradient boosting tree multi-class model is trained using an additive model framework. By iteratively adding decision trees, the overall objective function, which includes a multi-class log loss function and a tree complexity regularization term, is minimized. This learns the complex nonlinear mapping relationship between a subset of core features and human sensory evaluation labels.
[0009] Furthermore, step S3 also includes: The SHAP tree interpreter is used to calculate the local marginal contribution of each core feature to the model prediction result; For each sample to be predicted, the SHAP tree interpreter calculates the contribution value of each core feature to predicting the sample as a specific sensory category. The sum of the contribution values of all core features is equal to the difference between the predicted value of the sample and the baseline predicted value of the category.
[0010] Furthermore, step S4 also includes: Construct a dependency function mapping between the absolute physical concentration of the target's core features and its local marginal contribution value of SHAP; Using the absolute physical concentration of the target core feature as the x-axis and its corresponding SHAP local marginal contribution value as the y-axis, a scatter plot or fitted curve is plotted to visualize and quantify the nonlinear dependency between the two.
[0011] Furthermore, in step S4, the algorithm traverses the mapping to find the zero-value crossover point that causes the model prediction result to reverse. Specifically, in the dependency function mapping, the SHAP value is set to 0 as the baseline for reversing the prediction direction. The algorithm automatically searches for and locates the intersection point of the dependency function curve and the baseline. The horizontal coordinate value corresponding to the intersection point is the physical concentration critical threshold.
[0012] Furthermore, S1 also includes: The absolute concentrations of volatile aroma components were collected using solid-phase microextraction-gas chromatography-mass spectrometry, and the absolute concentrations of non-volatile flavor components were collected using ultra-high performance liquid chromatography-tandem mass spectrometry. The multi-source data acquisition and preprocessing also includes feature splicing of the absolute concentrations of the volatile aroma components and the absolute concentrations of the non-volatile taste components to form a full-dimensional physicochemical feature matrix.
[0013] Furthermore, step S2 also includes: Lasso regression screening based on L1 regularization, feature importance screening based on LightGBM, or linear dimensionality reduction methods based on principal component analysis; The step S3 uses CatBoost, LightGBM gradient boosting tree frameworks, or deep neural networks that introduce attention mechanisms. The LIME locally interpretable model is used in step S3.
[0014] According to a second aspect of the present invention, the present invention claims protection for a sensory classification and flavor critical threshold extraction system for raw tea, comprising: One or more processors; A memory that stores one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the method for sensory classification and flavor critical threshold extraction of raw tea.
[0015] Compared with the prior art, the beneficial results of the present invention are: (1) Low detection cost and high screening efficiency: Through a hierarchical filtering mechanism of macroscopic sensing coarse screening, targeted omics fine screening and random forest dimensionality reduction, the biochemical data of dozens of dimensions is accurately reduced to at least a few core features, which greatly reduces the high-throughput physicochemical detection cost of massive raw tea samples and the computing power consumption of online systems.
[0016] (2) Breaking through the black box to achieve standard quantification, the introduction of XGBoost combined with SHAP interpretable attribution algorithm not only achieves accurate flavor prediction, but also innovatively extracts the critical threshold of physicochemical concentration that triggers specific sensory judgments such as high aroma and rich mellowness, filling the technical gap of AI directly guiding the formulation of industrial-grade flavor component standards.
[0017] (3) Empowering visualized intelligent quality control, based on the normalization processing of core features, a standard radar fingerprint spectrum of raw tea of different sensory categories was constructed, eliminating the dimensional differences of physicochemical components, and providing an intuitive geometric comparison template for quality monitoring and raw material blending in the new tea beverage production line.
[0018] (4) High prediction accuracy and strong system robustness. The Limiting Gradient Boosting Tree (XGBoost) can effectively capture the nonlinear synergistic effect between complex chemical components, realize end-to-end objective online intelligent grading, and significantly overcome the subjectivity and batch-to-batch fluctuations caused by the reliance on human experience in traditional sensory evaluation. Attached Figure Description
[0019] Figure 1 This is a flowchart illustrating the process of a method for sensory classification and flavor critical threshold extraction of raw tea disclosed in an embodiment of this application; Figure 2 This application discloses a method for sensory classification and flavor threshold extraction of raw tea, which provides fingerprint profiles of core flavor components from raw tea of different sensory types. Figure 3 This is an analysis diagram of the local dependence and physical critical threshold extraction of a raw tea sensory classification and flavor critical threshold extraction method disclosed in the embodiments of this application; Figure 4 This is a global attribution distribution map of the core flavor components of a raw tea sensory classification and flavor critical threshold extraction method disclosed in the embodiments of this application for a specific sensory type; Figure 5 This is a bar chart showing the importance ranking of the physicochemical characteristics of raw tea based on the random forest algorithm, which is a method for sensory classification and flavor critical threshold extraction of raw tea disclosed in this application. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0021] The technical solutions disclosed in the various embodiments of this application are described in detail below with reference to the accompanying drawings.
[0022] According to a first embodiment of the present invention, the present invention claims protection for a method for sensory classification and flavor critical threshold extraction of raw tea, applied to a computer system, with reference to... Figure 1 This includes the following steps: Step S1: Obtain multi-source data from multiple raw tea samples. The multi-source data includes macroscopic sensing data collected using an electronic nose and electronic tongue, absolute concentration data of microscopic physicochemical components collected using targeted omics technology, and corresponding artificial sensory evaluation labels. Step S2: Based on the random forest algorithm, the absolute concentration data of the microscopic physicochemical components are evaluated for feature importance, and the top N physicochemical components with the highest contribution are selected as the core feature subset. The concentration of each component in the core feature subset is normalized to generate standard radar fingerprint maps corresponding to different sensory categories. Step S3: Using the subset of core features as input and the artificial sensory evaluation labels as output, train the extreme gradient boosting tree multi-classification model, and use the SHAP tree interpreter to calculate the local marginal contribution value of each core feature to the model prediction result. Step S4: For the target sensory category, construct a dependency function mapping between the absolute physical concentration of the target core feature and its SHAP local marginal contribution value, and traverse the mapping through an algorithm to find the zero-value crossover point that causes the model prediction result to reverse. The absolute physical concentration corresponding to the crossover point is established as the physical concentration critical threshold for the target core feature to trigger the target sensory category. Step S5: Receive the physicochemical test data of the raw tea sample to be tested, extract only the component concentrations corresponding to the core feature subset, input the trained limit gradient boosting tree multi-classification model, obtain the sensory classification prediction results and their confidence levels, and use the physical concentration critical threshold as the judgment basis to output a quantitative flavor standard for guiding industrial production.
[0023] In this embodiment, during the multi-level sensing and targeted omics data fusion acquisition, for a massive amount of initial raw tea samples, macroscopic clustering and elimination are performed using an electronic nose / tongue. The absolute concentrations of physicochemical components are then extracted from the remaining candidate samples using SPME-GC-MS and UPLC-MS / MS. The full-dimensional physicochemical feature vector of the i-th tea sample is defined as follows: Where D represents the total dimensions of the detected substance, such as 23 dimensions; the corresponding artificial sensory evaluation label is defined as follows: These correspond to the regular type, the highly aromatic type, the rich and mellow type, and the highly aromatic and rich and mellow type, respectively.
[0024] When constructing the core flavor profile dimensionality reduction and fingerprint map, feature screening based on Gini impurity was performed, and a random forest model was established for fitting. The mapping is as follows. For any feature j, the decrease in Gini impurity resulting from node splits in all decision trees t is calculated to obtain the feature importance score. : Where T is the total number of decision trees, for example, 200. Let j be the set of nodes that split in the t-th tree. According to... Sort the features in descending order and extract the top N features (e.g., N=8) with the highest scores to form the core feature subset after dimensionality reduction. Spectral mapping and normalization: To eliminate the dimensional differences in the absolute concentrations of various chemical substances, such as linalool and theanine, in the polar coordinate system, a range transformation is used to map the core feature subset, generating spectral coordinate values. :
[0025] Utilizing each sensory category The mean is used to generate a standard radar fingerprint in a polar coordinate polygon.
[0026] When extracting critical thresholds based on interpretable machine learning, XGBoost multi-class target fitting is performed, and the dimensionality-reduced core matrix is obtained. Input a limit gradient boosting tree XGBoost model. Minimize the overall objective function Obj using an additive model framework to learn the complex nonlinear boundary between physicochemical concentration and flavor category:
[0027] Where L is the multi-class log loss function mlogloss. The regularization term penalizes the tree complexity, where n represents the total number of tea samples participating in model training, k represents the number of the current decision tree in the XGBoost additive model, and K represents the total number of decision trees in the model.
[0028] SHAP local attribution and dependency analysis is employed, introducing a SHAP tree interpreter to calculate local marginal contributions. For the category c to be determined, such as the high-aroma and rich-tasting type, and the core feature j, its SHAP contribution value to the predicted output is calculated. Satisfies the summation property:
[0029] The physical threshold is used to calculate the absolute physical concentration of feature j. Its SHAP contribution value Dependency function mapping The algorithm iterates through the dependency mapping curve to find the coordinates of the zero-value crossover point that reverses the sensory judgment category, i.e., to determine the critical threshold that satisfies the following judgment condition. :
[0030] Extract the obtained x-coordinate Such as linalool concentration This was established as the physical concentration threshold for triggering this type of flavor determination.
[0031] When applying refined grading and dynamic standards, the critical thresholds calculated for each core component are used. This data is combined into a standardized database of quantified flavor components. During online inference, the quality control system can automatically match and determine the category and confidence level by inputting the physicochemical indicators of a new sample.
[0032] Data acquisition platforms, such as physicochemical testing laboratories, collect data from target samples that have undergone initial screening by human sensory evaluation. Table 1 Data Collection Table
[0033] Furthermore, step S2 also includes: Construct a random forest classification model containing multiple decision trees, and calculate the average decrease in Gini impurity caused by each physicochemical component as a splitting node in all decision trees; The average decrease is used as the feature importance score of the physicochemical component, and the components are sorted from high to low scores to select the top N physicochemical components to form the core feature subset.
[0034] Furthermore, step S2 also includes: The concentrations of each component in the core feature subset are normalized, and the absolute concentrations of each component are mapped to the interval of 0, 10, 1 using range transformation to eliminate the dimensional differences between different physicochemical components. The process involves generating standard radar fingerprint maps corresponding to different sensory categories, calculating the normalized concentration mean of each component of the core feature subset of all samples under the same sensory category, and plotting the mean sequence as a closed polygon in polar coordinates to form the standard radar fingerprint map for that sensory category.
[0035] In this embodiment, feature normalization employs max-min normalization to map the absolute content of the core chemical substance to a relative intensity range from 0 (lowest) to 1 (highest). Radar chart mapping, averaging by sensory type, plotting it as a closed polygon in polar coordinate system, constructing characteristic flavor component fingerprints for each flavor type, such as red representing high aroma and rich mellow type, blue representing high aroma type, etc., as a visual comparison benchmark; Reference Figure 2 The polar coordinate radar chart shows the relative concentration distribution of core chemical substances such as linalool and theanine under different sensory types, such as high aroma and strong alcohol, high aroma, and strong alcohol, after maximum and minimum normalization. This chart eliminates the dimensional differences between trace aroma components and macroscopic taste components, forming a geometric feature profile that can be used for industrial quality control and comparison.
[0036] Furthermore, step S3 also includes: The extreme gradient boosting tree multi-class model is trained using an additive model framework. By iteratively adding decision trees, the overall objective function, which includes a multi-class log loss function and a tree complexity regularization term, is minimized. This learns the complex nonlinear mapping relationship between a subset of core features and human sensory evaluation labels.
[0037] Reference Figure 3 This figure illustrates the nonlinear mapping between the actual physical concentration of a single core component, such as linalool, and its contribution to the predicted category, such as high aroma. The horizontal axis represents the absolute physical concentration of the component, and the vertical axis represents its corresponding SHAP value. The horizontal dashed line in the figure represents the zero-value baseline SHAP=0, where the model's prediction direction reverses. The horizontal axis value corresponding to the intersection of this baseline and the data distribution curve is the critical physical concentration threshold for triggering this flavor category, extracted by the system through reverse engineering.
[0038] Furthermore, step S3 also includes: The SHAP tree interpreter is used to calculate the local marginal contribution of each core feature to the model prediction result; For each sample to be predicted, the SHAP tree interpreter calculates the contribution value of each core feature to predicting the sample as a specific sensory category. The sum of the contribution values of all core features is equal to the difference between the predicted value of the sample and the baseline predicted value of the category.
[0039] Reference Figure 4 This diagram illustrates the global influence mechanism of core features extracted using the SHAP tree interpreter on a specific target category, such as the high-aroma and high-alcohol type. The vertical axis is sorted according to the overall importance of the features, while the horizontal axis represents the SHAP attribution contribution value. Different colored data points represent the actual absolute concentration of that physicochemical feature. This diagram visually reveals which high- or low-concentration biochemical components positively or negatively drive the final sensory classification result.
[0040] Furthermore, step S4 also includes: Construct a dependency function mapping between the absolute physical concentration of the target's core features and its local marginal contribution value of SHAP; Using the absolute physical concentration of the target core feature as the x-axis and its corresponding SHAP local marginal contribution value as the y-axis, a scatter plot or fitted curve is plotted to visualize and quantify the nonlinear dependency between the two.
[0041] Furthermore, in step S4, the algorithm traverses the mapping to find the zero-value crossover point that causes the model prediction result to reverse. Specifically, in the dependency function mapping, the SHAP value is set to 0 as the baseline for reversing the prediction direction. The algorithm automatically searches for and locates the intersection point of the dependency function curve and the baseline. The horizontal coordinate value corresponding to the intersection point is the physical concentration critical threshold.
[0042] In this embodiment, cross-modal classification and critical threshold extraction employ nonlinear fitting. The dimensionality-reduced features are input into the XGBoost multi-class prediction engine parameters set as follows: n_estimators=150, max_depth=5, learning_rate=0.05, to learn the nonlinear mapping between high-dimensional features and sensory rating.
[0043] SHAP attribution and dependency analysis was performed, extracting the underlying Booster to construct a TreeExplainer, and calculating the SHAP value matrix corresponding to each sensory label. For target categories such as high-aroma and rich-bodied types, SHAP dependency maps for specific ingredients such as linalool were plotted.
[0044] Zero-value crossover threshold extraction: In the dependency graph, the algorithm automatically searches for the critical point where the SHAP contribution value changes from negative to positive, i.e., the crossover point with the SHAP=0 horizontal reference line. The system extracts the actual physical concentration corresponding to this crossover point, such as linalool concentration. This was established as the red line for physicochemical concentration that triggers this type of sensory judgment.
[0045] Furthermore, S1 also includes: The absolute concentrations of volatile aroma components were collected using solid-phase microextraction-gas chromatography-mass spectrometry, and the absolute concentrations of non-volatile flavor components were collected using ultra-high performance liquid chromatography-tandem mass spectrometry. The multi-source data acquisition and preprocessing also includes feature splicing of the absolute concentrations of the volatile aroma components and the absolute concentrations of the non-volatile taste components to form a full-dimensional physicochemical feature matrix.
[0046] In this embodiment, feature splicing is performed during heterogeneous feature dimensionality reduction extraction, and the extracted aroma feature vector and taste feature vector are spliced together to form a multidimensional panoramic omics feature matrix X, such as 23-dimensional. Random forest dimensionality reduction constructs a random forest classification model containing 200 decision trees (n_estimators=200) to fit X with sensory labels y; When the core features are locked, feature importance scores are extracted, and they are sorted in descending order of contribution. The top 8 core features that determine flavor classification are selected, such as: linalool, geraniol, ionone, theanine, EGCG, EGC, soluble sugar, and caffeine, which greatly reduces the feature dimensionality. Reference Figure 5 This figure presents intermediate results from dimensionality reduction of high-dimensional omics data. The horizontal axis represents the Gini importance score, and the vertical axis represents various physicochemical components. This figure intuitively reflects the contribution of each component to sensory classification, and based on this, the top N items with the highest cumulative contribution, such as the top 8 core flavor feature components, are selected to remove redundant noise for subsequent models.
[0047] Furthermore, step S2 also includes: Lasso regression screening based on L1 regularization, feature importance screening based on LightGBM, or linear dimensionality reduction methods based on principal component analysis; The step S3 uses CatBoost, LightGBM gradient boosting tree frameworks, or deep neural networks that introduce attention mechanisms. The LIME locally interpretable model is used in step S3.
[0048] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.
Claims
1. A method for sensory classification and flavor threshold extraction of raw tea, characterized in that, When applied to a computer system, it includes the following steps: Step S1: Obtain multi-source data from multiple raw tea samples. The multi-source data includes macroscopic sensing data collected using an electronic nose and electronic tongue, absolute concentration data of microscopic physicochemical components collected using targeted omics technology, and corresponding artificial sensory evaluation labels. Step S2: Based on the random forest algorithm, the absolute concentration data of the microscopic physicochemical components are evaluated for feature importance, and the top N physicochemical components with the highest contribution are selected as the core feature subset. The concentration of each component in the core feature subset is normalized to generate standard radar fingerprint maps corresponding to different sensory categories. Step S3: Using the subset of core features as input and the artificial sensory evaluation labels as output, train the extreme gradient boosting tree multi-classification model, and use the SHAP tree interpreter to calculate the local marginal contribution value of each core feature to the model prediction result. Step S4: For the target sensory category, construct a dependency function mapping between the absolute physical concentration of the target core feature and its SHAP local marginal contribution value, and traverse the mapping through an algorithm to find the zero-value crossover point that causes the model prediction result to reverse. The absolute physical concentration corresponding to the crossover point is established as the physical concentration critical threshold for the target core feature to trigger the target sensory category. Step S5: Receive the physicochemical test data of the raw tea sample to be tested, extract only the component concentrations corresponding to the core feature subset, input the trained limit gradient boosting tree multi-classification model, obtain the sensory classification prediction results and their confidence levels, and use the physical concentration critical threshold as the judgment basis to output a quantitative flavor standard for guiding industrial production.
2. The method for sensory classification and flavor threshold extraction of raw tea according to claim 1, characterized in that, Step S2 further includes: Construct a random forest classification model containing multiple decision trees, and calculate the average decrease in Gini impurity caused by each physicochemical component as a splitting node in all decision trees; The average decrease is used as the feature importance score of the physicochemical component, and the components are sorted from high to low scores to select the top N physicochemical components to form the core feature subset.
3. The method for sensory classification and flavor threshold extraction of raw tea according to claim 1, characterized in that, Step S2 further includes: The concentrations of each component in the core feature subset are normalized, and the absolute concentrations of each component are mapped to the interval of 0, 10, 1 using range transformation to eliminate the dimensional differences between different physicochemical components. The process involves generating standard radar fingerprint maps corresponding to different sensory categories, calculating the normalized concentration mean of each component of the core feature subset of all samples under the same sensory category, and plotting the mean sequence as a closed polygon in polar coordinates to form the standard radar fingerprint map for that sensory category.
4. The method for sensory classification and flavor threshold extraction of raw tea according to claim 1, characterized in that, Step S3 further includes: The extreme gradient boosting tree multi-class model is trained using an additive model framework. By iteratively adding decision trees, the overall objective function, which includes a multi-class log loss function and a tree complexity regularization term, is minimized. This learns the complex nonlinear mapping relationship between a subset of core features and human sensory evaluation labels.
5. The method for sensory classification and flavor threshold extraction of raw tea according to claim 1, characterized in that, Step S3 further includes: The SHAP tree interpreter is used to calculate the local marginal contribution of each core feature to the model prediction result; For each sample to be predicted, the SHAP tree interpreter calculates the contribution value of each core feature to predicting the sample as a specific sensory category. The sum of the contribution values of all core features is equal to the difference between the predicted value of the sample and the baseline predicted value of the category.
6. The method for sensory classification and flavor critical threshold extraction of raw tea according to claim 1, characterized in that, Step S4 further includes: Construct a dependency function mapping between the absolute physical concentration of the target's core features and its local marginal contribution value of SHAP; Using the absolute physical concentration of the target core feature as the x-axis and its corresponding SHAP local marginal contribution value as the y-axis, a scatter plot or fitted curve is plotted to visualize and quantify the nonlinear dependency between the two.
7. The method for sensory classification and flavor threshold extraction of raw tea according to claim 6, characterized in that, In step S4, the algorithm traverses the mapping to find the zero-value crossover point that causes the model prediction result to reverse. Specifically, in the dependency function mapping, the SHAP value is set to 0 as the baseline for reversing the prediction direction. The algorithm automatically searches for and locates the intersection point of the dependency function curve and the baseline. The horizontal coordinate value corresponding to the intersection point is the physical concentration critical threshold.
8. The method for sensory classification and flavor critical threshold extraction of raw tea according to claim 1, characterized in that, S1 further includes: The absolute concentrations of volatile aroma components were collected using solid-phase microextraction-gas chromatography-mass spectrometry, and the absolute concentrations of non-volatile flavor components were collected using ultra-high performance liquid chromatography-tandem mass spectrometry. The multi-source data acquisition and preprocessing also includes feature splicing of the absolute concentrations of the volatile aroma components and the absolute concentrations of the non-volatile taste components to form a full-dimensional physicochemical feature matrix.
9. The method for sensory classification and flavor threshold extraction of raw tea according to claim 1, characterized in that, Step S2 further includes: Lasso regression screening based on L1 regularization, feature importance screening based on LightGBM, or linear dimensionality reduction methods based on principal component analysis; The step S3 uses CatBoost, LightGBM gradient boosting tree frameworks, or deep neural networks that introduce attention mechanisms. The LIME locally interpretable model is used in step S3.
10. A sensory classification and flavor critical threshold extraction system for raw tea, characterized in that, include: One or more processors; A memory having stored one or more programs that, when executed by one or more processors, cause the one or more processors to implement a method for sensory classification and flavor critical threshold extraction of raw tea according to any one of claims 1 to 9.