Wine grape quality producing area tracing method and system

Abstract

The invention provides a wine grape quality producing area tracing method and system, and relates to the technical field of food quality and safety detection. The method comprises the following steps: acquiring wine grape samples with known production place labels, and obtaining quality information and environment information of the wine grape samples to construct a basic sample set; carrying out multi-source feature fusion on the basis, extracting key quality features and environment features capable of distinguishing producing areas, and generating quality producing area fingerprint parameters; and training a target producing area discrimination model to construct a quality-producing area associated fingerprint database. And after multi-source fusion features of a to-be-detected sample are extracted, high-precision production area identification is realized through joint matching of the discrimination probability and the feature distance. According to the method, the traceability stability and reliability are improved through collaborative modeling of the quality features and the wind-soil features.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of food quality and safety detection, in particular to a wine grape quality and origin traceability method and system. BACKGROUND

[0002] With the rapid development of the wine industry, the planting area and the number of wine grape production areas continue to increase, and the proportion of grape raw materials in cross-regional circulation and deep processing continues to rise. In order to ensure the brand image and product quality stability of the production area, the industry has gradually introduced information technology such as bar code identification, radio frequency tag, two-dimensional code anti-fake and block chain storage, and recorded and tracked the process of wine grape harvesting, transportation and processing. At the same time, some researches began to use stable isotopes, trace elements and conventional physicochemical indicators to identify the origin of wine or wine grape, and explored the origin discrimination method based on physicochemical fingerprint. To some extent, the above technologies have realized the information chain record from "production area to wine bottle", providing preliminary technical support for reducing label counterfeiting and production area fraud.

[0003] Under the background of fine agriculture and digital vineyard construction, the wine grape industry's attention to "production area characteristics" and "quality grading" continues to increase, and the origin traceability technology gradually evolves from simply tracking the logistics process to "combining origin identification with quality evaluation based on inherent quality characteristics". Using grape fruit sugar acid balance, phenolic and anthocyanin accumulation, aroma precursor composition and other quality indicators, combined with plot-scale soil, climate and cultivation management data, a multi-dimensional fingerprint reflecting the characteristics of the land has become an important demand of wine enterprises and regulatory authorities. At the same time, the introduction of intelligent algorithms such as machine learning and pattern recognition makes it possible to comprehensively analyze multi-source data, which is conducive to mining the difference patterns of production areas from multi-dimensional feature space, improving the accuracy, efficiency and interpretability of traceability discrimination, and providing a technical foundation for establishing production area grading standards, brand protection and quality supervision.

[0004] However, the existing wine grape and wine origin traceability technology still has the following shortcomings: first, many schemes mainly rely on logistics identification and process record, focusing on "source tracking of information flow", and the quality characteristics of wine grape are not fully utilized. Once the label is counterfeited, tampered or the information chain is broken, it is difficult to achieve reliable traceability; second, some physicochemical fingerprint methods mainly target a single or a few indicators, lacking systematic integration of multi-dimensional quality characteristics such as sugar acid balance, phenolic structure, aroma precursor and sensory quality, making it difficult to accurately depict the subtle quality differences between different production areas; third, the existing origin discrimination research based on physicochemical indicators mainly inputs physicochemical indicators and production area labels into general classification models, lacking the design of collaborative modeling of quality characteristics and production environment factors, and failing to form a "quality-origin fingerprint" for explaining the correlation between production land and quality. The stability and interpretability of the model are insufficient in cross-year and cross-climate conditions.

[0005] Therefore, it is urgent to develop a method of integrating the internal quality indicators of wine grapes with environmental factors of the production area, constructing a quality-geographical origin correlation fingerprint library and a corresponding discrimination model, and realizing the traceability of wine grapes based on quality characteristics, so as to improve the discrimination accuracy of the production area, enhance the quantitative expression ability of the characteristic quality of the production area, and meet the actual needs of brand protection and quality supervision. SUMMARY

[0006] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a wine grape quality and origin traceability method and system, which utilizes the collaborative modeling of key quality characteristics and environmental characteristics to accurately identify the differences between production areas, so that the wine grape samples to be tested can obtain more reliable results of the production area determination under the complex terroir background.

[0007] To achieve the above-mentioned purpose, the present application provides the following scheme:

[0008] A wine grape quality and origin traceability method, comprising:

[0009] Collecting wine grape samples with known geographical origin labels in different production areas, obtaining corresponding plot positions, meteorological, soil and cultivation management information, and forming a basic sample set containing quality information and environmental information;

[0010] Detecting the wine grape samples in the basic sample set for physicochemical indicators and sensory indicators, pre-processing and normalizing, and constructing a multi-source fusion feature vector corresponding to the environmental information;

[0011] Based on the multi-source fusion feature vector and the geographical origin label, key quality characteristics and environmental characteristics that distinguish the production areas are extracted by using a feature selection algorithm, and quality and origin fingerprint parameters of each production area are generated;

[0012] Training a classification model with the multi-source fusion feature vector and the geographical origin label to obtain a target production area discrimination model, and correlating the target production area discrimination model with the quality and origin fingerprint parameters to form a quality-geographical origin correlation fingerprint library;

[0013] Obtaining a to-be-tested multi-source fusion feature vector of a to-be-tested wine grape sample;

[0014] Inputting the to-be-tested multi-source fusion feature vector into the target production area discrimination model to obtain a production area discrimination result, and matching the production area discrimination result with the fingerprint parameters in the quality-geographical origin correlation fingerprint library to determine a target production area according to the discrimination probability and the feature distance, and outputting a traceability result containing the target production area and the confidence.

[0015] Preferably, the wine grape samples with known geographical origin labels are collected in different production areas, the corresponding plot positions, meteorological, soil and cultivation management information are obtained, and a basic sample set containing quality information and environmental information is formed, comprising:

[0016] According to the plot distribution of the production area, sample layers are set, and wine grape samples with known production site labels are collected from each sample layer, each wine grape sample is bound with its plot location parameter to form a sample record;

[0017] According to the plot location parameter, the data fields of temperature, precipitation, solar radiation, organic matter content and pH value are retrieved from the meteorological database and the soil database respectively, and the data fields are associated with the corresponding sample records to form the environmental information;

[0018] The sugar content, total acid, pH value, phenolic content and soluble solids content of each wine grape sample are detected to obtain the corresponding quality detection data field, and the quality detection data field is integrated with the sample-level environmental information to generate the basic sample set containing quality information and environmental information.

[0019] Preferably, the wine grape samples in the basic sample set are detected for physicochemical indicators and sensory indicators, preprocessed and normalized to construct a multi-source fusion feature vector corresponding to the environmental information, including:

[0020] The sugar content, total acid, pH value, phenolic content and soluble solids content of each wine grape sample in the basic sample set are detected to form the corresponding physicochemical indicator data field, and the color, aroma intensity and taste structure are recorded to form the sensory indicator data field;

[0021] The physicochemical indicator data field and the sensory indicator data field are subjected to missing value elimination and abnormal value identification according to a predetermined format to obtain preprocessed quality data records;

[0022] Interval scaling normalization is performed on the quality data records to make the numerical scales of the quality data records and the environmental information consistent, forming normalized quality features;

[0023] The normalized quality features and the corresponding environmental information are spliced to obtain the multi-source fusion feature vector.

[0024] Preferably, based on the multi-source fusion feature vector and the production site label, a feature selection algorithm is used to extract key quality features and environmental features that distinguish the production sites to generate quality and production site fingerprint parameters for each production site, including:

[0025] The multi-source fusion feature vector is grouped according to the production site label to form a feature data set distinguished by production area;

[0026] A correlation analysis was performed on the feature data set to filter out quality and environmental features that were not sufficiently relevant to the place of origin label, thus obtaining a candidate feature set.

[0027] The feature selection algorithm is applied to the candidate feature set to calculate the feature importance score, thereby obtaining the key quality features and key environmental features that can distinguish different production areas. Based on the key quality features and key environmental features, the feature distribution parameters of each production area are calculated to form a parameter set used to characterize the differences in features of each production area.

[0028] The parameter set is organized according to the place of origin label to generate the quality place of origin fingerprint parameters.

[0029] Preferably, a feature selection algorithm is applied to the candidate feature set to calculate feature importance scores, thereby obtaining key quality features and key environmental features that can distinguish different production areas. Based on these key quality features and key environmental features, feature distribution parameters for each production area are calculated, forming a parameter set used to characterize the differences in features among different production areas, including:

[0030] For each candidate feature in the candidate feature set, calculate the sample mean and sample variance of each origin on the candidate feature according to the origin label, and calculate the population mean and population variance of all origins on the candidate feature;

[0031] Based on the sample mean, sample variance, population mean, and population variance, the feature importance score for each candidate feature is calculated using the following formula: ;in, For the first Feature importance scores for each candidate feature; Place of origin In the The sample mean on each candidate feature; For all places of origin in the first The overall mean over each candidate feature; Place of origin In the Sample variance over each candidate feature; Place of origin The corresponding number of samples; These are positive real numbers used to prevent the denominator from being zero and to suppress numerical fluctuations. This indicates a summation operation across all production locations;

[0032] Candidate features with feature importance scores greater than a preset threshold are identified as key quality features and key environmental features. Feature distribution parameters for each production area are calculated based on each key quality feature and each key environmental feature, using the following formula: ;in, Place of origin In the Feature distribution parameters on key features; For all places of origin in the first Overall variance over key features;

[0033] The characteristic distribution parameters of the same place of origin on all key quality characteristics and key environmental characteristics are arranged in the order of characteristic index to form a parameter set corresponding to the place of origin, and the parameter set is used as the quality fingerprint parameter of the place of origin.

[0034] Preferably, a classification model is trained using the multi-source fusion feature vector and the origin label to obtain a target origin discrimination model, and the target origin discrimination model is associated with the quality origin fingerprint parameters to form a quality-origin association fingerprint database, including:

[0035] The multi-source fusion feature vector is divided into a training dataset and a validation dataset according to the origin label, and the training dataset is used as input to build a classification model to perform origin discrimination learning.

[0036] During the training of the classification model, the verification accuracy of each iteration cycle is recorded, and the training round with stable verification accuracy and reaching a preset threshold is selected based on the changes in the verification accuracy, so as to obtain a target origin discrimination model that meets the origin identification accuracy requirements.

[0037] The category output structure of the target origin discrimination model is matched one-to-one with the origin index of the quality origin fingerprint parameter to establish a mapping relationship between the classification output result and the feature distribution parameter.

[0038] The target origin discrimination model and its mapping relationship with the quality origin fingerprint parameters are organized according to the origin label and stored as a quality-origin association fingerprint database.

[0039] Preferably, the multi-source fusion feature vector to be tested is input into the target origin discrimination model to obtain the origin discrimination result, and matched with the fingerprint parameters in the quality-origin association fingerprint database. The target origin is determined according to the discrimination probability and feature distance, and the traceability result containing the target origin and confidence level is output, including:

[0040] The multi-source fusion feature vector to be tested is input into the target origin discrimination model to obtain the discrimination probability distribution for each origin, thus forming the origin discrimination result;

[0041] Based on the candidate origin labels appearing in the origin discrimination results, the corresponding fingerprint parameters are retrieved from the quality-origin association fingerprint database, and the feature distance relative to each fingerprint parameter is calculated based on the values ​​of the multi-source fusion feature vector to be tested on key quality features and key environmental features, thus forming the feature distance result.

[0042] For each candidate origin, the discrimination probability in the origin discrimination result is combined with the feature distance result to obtain a comprehensive discrimination score, and the candidate origin that meets the preset conditions is selected as the target origin.

[0043] The target origin and its corresponding comprehensive discrimination score are converted into a confidence level, and a traceability result containing the target origin and the confidence level is generated.

[0044] Preferably, the preset conditions include: the discrimination probability is greater than a first probability threshold, the feature distance is less than a first distance threshold, and the comprehensive discrimination score is ranked highest among all candidate origins; the first probability threshold is determined based on the median value of the discrimination probability distribution of the target origin discrimination model on the verification dataset, and the first distance threshold is determined based on the lower quartile of the feature distance distribution between the fingerprint parameters of each origin in the quality-origin association fingerprint database.

[0045] A system for tracing the origin of wine grapes, comprising:

[0046] The basic sample construction unit is used to collect wine grape samples with known origin labels in different production areas, obtain information on the location, weather, soil and cultivation management of the corresponding plots, and form a basic sample set containing quality information and environmental information;

[0047] The quality feature extraction unit is used to detect the physicochemical and sensory indicators of wine grape samples in the basic sample set, and after preprocessing and normalization, construct a multi-source fusion feature vector corresponding to the environmental information.

[0048] The key feature identification unit is used to extract key quality features and environmental features that distinguish the place of origin based on the multi-source fusion feature vector and the place of origin label, and generate quality place of origin fingerprint parameters for each place of origin.

[0049] The discrimination model training unit is used to train a classification model with the multi-source fusion feature vector and the origin label to obtain a target origin discrimination model, and associate the target origin discrimination model with the quality origin fingerprint parameters to form a quality-origin association fingerprint database.

[0050] The feature acquisition unit is used to acquire the multi-source fusion feature vector of the wine grape sample to be tested;

[0051] The traceability discrimination output unit is used to input the multi-source fusion feature vector to be tested into the target origin discrimination model, obtain the origin discrimination result, match it with the fingerprint parameters in the quality-origin association fingerprint database, determine the target origin based on the discrimination probability and feature distance, and output the traceability result containing the target origin and confidence level.

[0052] The present invention discloses the following technical effects:

[0053] (1) This invention collects wine grape samples with known origin labels in different production areas and binds the location, weather, soil and cultivation management information with the quality test results to construct a basic sample set containing quality information and environmental information. This allows the terroir information of the production area and the quality characteristics of the grapes themselves to enter the data system at the same time, avoiding the problem of broken traceability caused by relying solely on logistics identification or label information, and improving the reliability of origin traceability.

[0054] (2) In this invention, the physicochemical and sensory indicators are preprocessed and normalized in a unified manner in the basic sample set, and a multi-source fusion feature vector is constructed. This realizes the quantitative expression of multi-dimensional differences in quality, effectively improves the shortcomings of traditional methods in that the quality characteristics are insufficient and it is difficult to form a stable origin fingerprint, and improves the characterization ability of origin differences.

[0055] (3) Based on multi-source fusion feature vectors and place of origin labels, this invention uses feature selection algorithm to automatically identify and distinguish key quality features and key environmental features of the place of origin, and generates quality place of origin fingerprint parameters, so that the differences in the place of origin are presented in a structured numerical way, which overcomes the problem of relying on a single indicator and making it difficult to capture subtle differences in the place of origin in the existing methods, and improves the discrimination stability across years and climate conditions.

[0056] (4) This invention trains a classification model with multi-source fusion feature vectors and place of origin labels to obtain a target place of origin discrimination model, and associates it with quality place of origin fingerprint parameters to form a quality-place of origin association fingerprint database. This realizes the deep integration of data-driven model and local fingerprint, overcomes the shortcomings of model black box and uninterpretable results in the prior art, and improves the interpretability and credibility of traceability results.

[0057] (5) This invention inputs the multi-source fusion feature vector to be tested into the target origin discrimination model and combines it with the fingerprint parameters in the quality-origin association fingerprint database to determine the target origin based on the discrimination probability and feature distance. This makes the traceability process rely on both the model results and terroir characteristics, making the final traceability results more stable and accurate. It overcomes the defect of the single discrimination model being unstable for extreme samples and achieves high-precision and high-robust traceability of wine grape quality origin. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 A flowchart of the method provided in an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram of the system structure provided in an embodiment of the present invention. Detailed Implementation

[0061] 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.

[0062] The purpose of this invention is to provide a method and system for tracing the origin of wine grapes. By constructing a multi-source fingerprint system that integrates quality characteristics and terroir information and combining it with a discriminant model to achieve dual verification, the accuracy, stability and interpretability of tracing the origin of wine grapes are significantly improved.

[0063] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0064] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a method for tracing the origin of wine grapes, comprising:

[0065] Step 100: Collect wine grape samples with known origin labels from different production areas, obtain information on the location, weather, soil and cultivation management of the corresponding plots, and form a basic sample set containing quality information and environmental information;

[0066] Step 200: Detect physicochemical and sensory indicators of wine grape samples in the basic sample set, and construct a multi-source fusion feature vector corresponding to environmental information after preprocessing and normalization.

[0067] Step 300: Based on the multi-source fusion feature vector and origin label, use the feature selection algorithm to extract the key quality features and environmental features that distinguish the origin, and generate the quality origin fingerprint parameters for each origin.

[0068] Step 400: Train a classification model with multi-source fusion feature vectors and origin labels to obtain a target origin discrimination model, and associate the target origin discrimination model with quality origin fingerprint parameters to form a quality-origin association fingerprint database;

[0069] Step 500: Obtain the multi-source fusion feature vector of the wine grape sample to be tested;

[0070] Step 600: Input the multi-source fusion feature vector to be tested into the target origin discrimination model to obtain the origin discrimination result, and match it with the fingerprint parameters in the quality-origin association fingerprint database. Determine the target origin based on the discrimination probability and feature distance, and output the traceability result containing the target origin and confidence level.

[0071] Specifically, in step 100 of this embodiment, firstly, based on the boundaries and spatial distribution of the target wine grape producing area, continuous or adjacent plots are divided into several sampling strata according to topographic features, planting scale, and management methods. Representative plots are selected within each sampling stratum, and wine grape samples with clearly recorded origin labels are collected from each plot during the harvest season. During collection, this embodiment simultaneously acquires the geographical location parameters of each wine grape sample's location and stores these parameters in a one-to-one correspondence with the corresponding wine grape sample, forming a sample record containing sample identification, origin label, and plot location parameters. Subsequently, based on the plot location parameters in each sample record, this embodiment automatically retrieves daily or hourly temperature, precipitation, and solar radiation information covering the plot from a pre-configured meteorological database, and retrieves soil organic matter content and pH indicators corresponding to the plot from a soil database. The numerical information obtained from the meteorological and soil databases is then associated with the corresponding sample record in a field format, forming environmental information corresponding to each sample record. Next, this embodiment conducts quality testing on each collected wine grape sample, using conventional physicochemical analysis methods to determine sugar content, total acid, pH, phenolic content, and soluble solids content. The test results are then archived as quality test data fields according to a fixed field order. Based on this, the quality test data fields corresponding to each wine grape sample are integrated with the aforementioned environmental information in the same data structure, so that each record simultaneously contains the quality information of the wine grape sample and its growing environment information. This constructs a basic sample set containing both quality and environmental information, providing complete data support for subsequent multi-source feature construction, feature selection, and origin discrimination modeling.

[0072] Sampling stratification refers to dividing plots within a production area into several relatively homogeneous strata based on geographical location, planting conditions, and management models, without altering the original planting layout. Samples within each stratum can represent the overall characteristics of that stratum, thereby improving sample representativeness and reducing sampling bias. Sample records refer to data entries established using individual wine grape samples as the basic unit. Each sample record includes at least a sample identifier, origin label, and corresponding plot location parameters, serving as the foundational index for subsequent linking of environmental and quality information. Environmental information refers to environmental characteristics such as temperature, precipitation, solar radiation, organic matter content, and pH, retrieved from meteorological and soil databases using plot location parameters and associated with the sample records. The characteristics are used to reflect the climate and soil conditions of the origin of the wine grape sample; the quality test data field refers to the quality characteristics such as sugar content, total acid, pH, phenolic content and soluble solids content stored in a unified format after completing the physicochemical test for a single wine grape sample, which are used to characterize the intrinsic quality of the wine grape; the basic sample set refers to the data set composed of multiple sample records, each of which contains both quality test data fields and environmental information. This embodiment uses this basic sample set to carry out subsequent feature construction and model training, so that the multi-source fusion feature vectors and origin labels used in subsequent steps can be directly extracted from this basic sample set, ensuring the consistency and repeatability of the method execution.

[0073] Specifically, in step 200 of this embodiment, after obtaining the basic sample set, each wine grape sample in the basic sample set is sequentially subjected to quality testing to simultaneously obtain physicochemical and sensory indicators. On the one hand, conventional physicochemical analysis methods are used to determine the sugar content, total acid, pH, phenolic content, and soluble solids content of each wine grape sample, and these are recorded as physicochemical indicator data fields according to a fixed field order. These physicochemical indicator data fields are structured data that characterizes the physicochemical quality of a sample using multiple numerical entries, with each sample as a unit. On the other hand, trained sensory evaluators score the color, aroma intensity, and taste structure of each wine grape sample under unified evaluation conditions, and quantify these scores according to preset levels or score ranges, recording them as sensory indicator data fields. These sensory indicator data fields are a set of numerical expressions of the originally subjective sensory impressions. Through the above processing, this embodiment ensures that each wine grape sample corresponds to both a set of physicochemical indicator data fields and a set of sensory indicator data fields, providing a foundation for subsequent unified preprocessing.

[0074] Next, in this embodiment, the aforementioned physicochemical index data fields and sensory index data fields are merged in the same data structure. Missing value removal and outlier identification are performed on the merged quality data to obtain preprocessed quality data records. The preprocessed quality data records refer to the set of quality data retained within the same sample record after performing completeness and reasonableness checks on values ​​such as sugar content, total acidity, pH, phenolic content, soluble solids content, color, aroma intensity, and mouthfeel structure. Missing value removal can be achieved by deleting samples lacking key indicators or by using neighboring sample imputation. Outlier identification can be achieved by setting a reasonable value range or by identifying extreme values ​​based on historical distribution characteristics and replacing or marking them. After obtaining the preprocessed quality data records, this embodiment performs interval scaling normalization processing on each value according to a preset interval, mapping indicators with different dimensions and different value ranges to a unified numerical interval on an equal scale, so that the numerical scale of physicochemical indicators and sensory indicators is consistent with the numerical scale of the environmental information of the corresponding sample. The result after normalization processing is recorded as normalized quality feature in this embodiment, which is the direct input for subsequent feature fusion.

[0075] Finally, in this embodiment, the normalized quality characteristics corresponding to each wine grape sample are concatenated with the environmental information already established for that sample in the basic sample set to obtain a multi-source fusion feature vector. Feature concatenation refers to the sequential arrangement and combination of normalized sugar content, total acidity, pH, phenolic content, soluble solids content, color, aroma intensity, and mouthfeel structure with corresponding temperature, rainfall, solar radiation, organic matter content, pH, and cultivation management information within the same sample record, forming a continuous numerical sequence. This numerical sequence is the multi-source fusion feature vector in this embodiment. The significance of the multi-source fusion feature vector lies in representing the intrinsic quality characteristics of the wine grape sample and its growing environment characteristics in a unified structured form. This allows subsequent feature selection algorithms and classification model training to be directly based on this vector, ensuring that quality information and environmental information are utilized synchronously during the modeling stage, thereby improving the accuracy and stability of origin identification.

[0076] Specifically, in step 300 of this embodiment, after constructing the multi-source fusion feature vector and assigning a corresponding origin label to each sample record, this embodiment first groups the multi-source fusion feature vectors according to the origin label, dividing the multi-source fusion feature vectors of all samples belonging to the same origin into the same set, forming a feature data set divided by origin. Here, the feature data set refers to a data structure that centrally stores the multi-source fusion feature vectors of all samples under a given origin. Each feature data set contains multiple sample records, and each record contains both quality features and environmental features, used for subsequent comparison and analysis of the same feature dimension within the same origin and between different origins.

[0077] After obtaining the feature data set categorized by production region, this embodiment performs a correlation analysis on the relationship between each feature dimension and the origin label. Specifically, this embodiment examines the numerical differences between different production regions and the stability within the same production region for each feature dimension, including sugar content, total acidity, pH, phenolic content, soluble solids content, color, aroma intensity, taste structure, temperature, precipitation, solar radiation, organic matter content, and soil pH. Features that show significant differences between different production regions but relatively small variations within the same production region are marked as features with strong correlation to the origin label. Features whose trends are not clearly related to the origin label or fluctuate greatly are marked as features with insufficient correlation and are removed from subsequent calculations. The remaining feature set is called the candidate feature set in this embodiment. The meaning of the candidate feature set is to screen feature dimensions that have potential contributions to distinguishing production regions from all quality and environmental features, providing a narrowed range of candidates for further refined feature selection.

[0078] After obtaining the candidate feature set, this embodiment applies a feature selection algorithm to the candidate feature set to calculate the feature importance score of each candidate feature in distinguishing different origins. The feature selection algorithm can adopt a tree-based feature evaluation method, that is, when constructing a classification process with candidate features as input and origin labels as output, the contribution of different candidate features in dividing samples and improving classification purity is statistically analyzed. Features with higher contribution are assigned higher importance scores, and features with lower contribution are assigned lower importance scores. Here, the feature importance score refers to a numerical index that quantitatively evaluates the influence of a single feature on the final origin discrimination result, considering that all candidate features participate in origin differentiation. In this embodiment, it can be normalized within an interval between zero and one. For example, the importance score of a highly important feature is close to one, and the importance score of a less important feature is close to zero. In this embodiment, features with importance scores exceeding a preset threshold are selected from the candidate feature set as key quality features and key environmental features based on feature importance scores. For each key feature dimension, the central tendency and dispersion of each production area on that feature are calculated. For example, the typical level and fluctuation range of the production area on that feature can be characterized by calculating the average value and distribution range of samples within the production area. These values ​​used to characterize the typical values ​​and variation characteristics of the production area on the key features are collectively referred to as feature distribution parameters.

[0079] Finally, this embodiment organizes the feature distribution parameters corresponding to each key quality characteristic and key environmental characteristic according to the place of origin label, forming a parameter set to characterize the differences in characteristics of each place of origin, and defines the parameter set of each place of origin as the quality place of origin fingerprint parameter. The quality place of origin fingerprint parameter is a parameter vector formed by arranging the feature distribution parameters of a place of origin on a set of key quality characteristics and key environmental characteristics in a fixed order. This parameter vector not only retains the quality and environmental differences of the place of origin relative to other places of origin, but also avoids the interference of redundant or invalid features. In subsequent implementation steps, this embodiment uses the quality place of origin fingerprint parameter as a compact representation to describe the comprehensive terroir and quality characteristics of each place of origin. When the wine grape sample to be tested is used for place of origin tracing and identification, it can be compared with these quality place of origin fingerprint parameters to intuitively reflect the degree of similarity between the sample to be tested and the typical characteristics of different places of origin, thereby improving the accuracy and interpretability of place of origin identification.

[0080] In this exemplary embodiment, the process of applying a feature selection algorithm to calculate the feature importance score for the candidate feature set is as follows: First, for each candidate feature in the candidate feature set, the samples are divided into different origin groups according to the origin label. Within each origin group, the sample mean and sample variance of the candidate feature are calculated. Simultaneously, the overall mean and overall variance of the candidate feature are calculated across all samples from all origins. To determine the importance of the feature, this embodiment constructs a feature importance score, which is determined by two parts: one part represents the degree to which the average of each origin group deviates from the overall average, and the other part represents the degree of fluctuation within each origin group. A larger former indicates a more significant difference between different origins, while a smaller latter indicates a more stable feature within the same origin. In this embodiment, the former is weighted and summed according to the number of samples from each origin, and the latter is weighted and summed according to the number of samples, then a very small positive constant is added (this constant is derived from the algorithm setting to avoid division by zero and smooth the value; a typical value is 0.01). The two are then divided to obtain the feature importance score. For example, when a candidate feature is sugar content, assuming there are 3 production areas with sample sizes of 10, 12, and 8 respectively, the average sugar content of each production area is 18, 20, and 22 respectively, the overall average is approximately 20, and the internal sugar content variances of each production area are 1, 2, and 1.5 respectively, then in this embodiment, the importance score of the sugar content feature can be obtained by weighting and calculating in the above manner. The higher the score, the better the feature can distinguish the differences between production areas.

[0081] After obtaining the feature importance scores of all candidate features, this embodiment sets a preset threshold, designating candidate features with scores higher than this threshold as key quality features and key environmental features. The preset threshold can be determined by the distribution of candidate feature scores; for example, the median or 75th percentile of all scores can be selected, ensuring that features with higher scores enter the key feature set. After determining the key features, this embodiment calculates the feature distribution parameters for each origin for each key feature. The calculation method is as follows: first, calculate the difference between the average value of a particular origin for that feature and the average value of all origin samples; then, standardize the feature using the overall variance of all origins for that feature and the aforementioned constant as a scaling factor; multiply the standardized result by the importance score of that feature to obtain the feature distribution parameters for that origin. This parameter simultaneously reflects the direction and magnitude of the origin's deviation from the overall average level in that feature dimension, and, combined with the feature importance score, reflects the contribution of that feature to the differences in origins. For example, in the aforementioned sugar content characteristics, if the average sugar content of a certain production area is 22, the overall mean is 20, and the variance is moderate, then the characteristic distribution parameter of the production area in terms of sugar content characteristics may be a positive value (e.g., 1.6), while the production area with lower sugar content may get a negative value (e.g., -1.4). In this way, the differences in taste characteristics between different production areas can be distinguished by the magnitude and sign of the numerical value.

[0082] Finally, in this embodiment, the feature distribution parameters of the same origin across all key quality characteristics and key environmental characteristics are arranged sequentially according to a fixed feature index order, forming a parameter set corresponding to the origin. This set is defined as the quality origin fingerprint parameter of that origin. The quality origin fingerprint parameter is essentially an array structure, where each position corresponds to a key feature, and each value represents the origin's performance and degree of difference in that feature, thus preserving the distinguishability between origins while compressing dimensionality. By generating corresponding quality origin fingerprint parameters for each origin, this embodiment presents the differences in key characteristics between different origins in a structured and computable manner, providing clear numerical basis for subsequent origin matching and origin identification of samples to be tested, and improving the accuracy and interpretability of traceability results.

[0083] Specifically, in step 400 of this embodiment, firstly, based on the already constructed multi-source fusion feature vectors and corresponding origin labels, this embodiment divides all samples into hierarchical groups according to the origin labels, with one part serving as the training dataset and the other as the validation dataset. Each record in the training dataset consists of a multi-source fusion feature vector and its corresponding origin label, while the validation dataset uses the same data organization method. Based on this, this embodiment selects a supervised classification model, using the multi-source fusion feature vectors from the training dataset as input and the origin labels as the output targets. Through iterative training, the classification model learns the correspondence between the multi-source fusion feature vectors and different origin labels. Here, the classification model refers to a model structure that can accept fixed-length feature vectors and output the probabilities or scores of each candidate origin, such as tree models, ensemble learning models, or neural network models, but is not limited to a specific model form.

[0084] During the training of the classification model, this embodiment evaluates the current model using a validation dataset after each training iteration. The proportion of samples correctly identified as the true origin in the validation dataset is calculated and recorded as the validation accuracy for that iteration. Subsequently, this embodiment tracks the changes in validation accuracy over multiple iterations. When the validation accuracy gradually increases and remains above a preset threshold (e.g., greater than 0.90) for several iterations, the model is considered to have good generalization ability. If training continues but the validation accuracy no longer improves or even decreases, overfitting has occurred. This embodiment selects the most suitable training epoch based on the stable range of the validation accuracy and confirms the classification model obtained in the corresponding epoch as the target origin discrimination model. The target origin discrimination model refers to the final classification model that, given a multi-source fusion feature vector, can output the confidence level or probability distribution of each candidate origin, used to perform origin discrimination on unknown samples.

[0085] After obtaining the target origin discrimination model, this embodiment establishes a correspondence between the model's category output structure and the quality origin fingerprint parameters of each of the aforementioned origins. Specifically, each category index in the model output layer corresponds to a specific origin label, and each origin label has already been associated with a set of quality origin fingerprint parameters in the previous steps. This embodiment establishes a one-to-one correspondence mapping between the category index, the origin label, and the corresponding quality origin fingerprint parameters, and stores this mapping relationship together with the target origin discrimination model itself. The data is then organized according to the origin labels to form a quality-origin association fingerprint database. The quality-origin association fingerprint database refers to a comprehensive data structure containing the target origin discrimination model and the mapping relationship between each category output and the corresponding quality origin fingerprint parameters. When any multi-source fusion feature vector to be tested is input, the target origin discrimination model can provide candidate origins and their probabilities, while the fingerprint parameters can interpret the typical characteristics of each candidate origin. This allows for a comprehensive judgment combining probability discrimination and fingerprint matching in subsequent traceability steps.

[0086] Further, step 500 of this embodiment is as follows: After completing the modeling of the basic sample set, this embodiment selects the wine grape samples to be tested for origin traceability. The wine grape samples to be tested are analyzed using a testing scheme completely consistent with the basic sample set to obtain quality information such as sugar content, total acid, pH, phenolic content, soluble solids content, color, aroma intensity, and mouthfeel structure. Simultaneously, the location of the corresponding plot, meteorological information, and soil information are collected or retrieved, and missing value removal, outlier identification, and interval scaling normalization are performed in the same manner as the aforementioned steps. Subsequently, this embodiment concatenates the normalized quality features with the corresponding environmental information in the same data structure to obtain a multi-source fusion feature vector to be tested, whose structure and dimensions are consistent with the basic sample set. The multi-source fusion feature vector to be tested refers to a set of numerical sequences formed by arranging the quality characteristics and environmental characteristics of a single wine grape sample in a predetermined order. This numerical sequence is completely consistent in format with the multi-source fusion feature vector used when training the target origin discrimination model, thus ensuring that the sample to be tested can be directly fed into the trained discrimination model for inference.

[0087] Furthermore, in step 600, this embodiment first inputs the aforementioned multi-source fusion feature vector to be tested into the target origin discrimination model. The target origin discrimination model provides a discrimination probability for each candidate origin, and the discrimination probabilities of all candidate origins together constitute the origin discrimination result. The origin discrimination result refers to a set of probability values ​​indexed by candidate origins obtained on the model output layer for the wine grape sample to be tested. Each value reflects the probability that the sample belongs to the corresponding origin. Subsequently, this embodiment selects several candidate origin labels with high discrimination probabilities from the origin discrimination result and retrieves the corresponding fingerprint parameters from the aforementioned quality-origin association fingerprint database based on these candidate origin labels. The fingerprint parameters are the quality origin fingerprint parameters extracted for each origin, including the feature distribution parameters of the origin on key quality features and key environmental features. This embodiment further calculates the feature distance of the sample to be tested relative to the fingerprint parameters of each origin based on the difference between the values ​​of the multi-source fusion feature vector to be tested on these key features and the corresponding feature distribution parameters in the fingerprint parameters of each candidate origin. Feature distance refers to the difference measure obtained by comparing the features of the sample under test with the fingerprint parameters of a certain place of origin in key feature dimensions. It can reflect the degree of closeness between the sample under test and the typical features of the place of origin by combining the magnitude of the difference of each feature. The smaller the value, the closer the sample under test is to the place of origin in terms of overall features.

[0088] After obtaining the origin discrimination result and feature distance, this embodiment combines the corresponding discrimination probability and feature distance for each candidate origin to calculate a comprehensive discrimination score. The comprehensive discrimination score is a numerical indicator reflecting the combined effect of the model's probability discrimination result and fingerprint similarity. In this embodiment, the comprehensive discrimination score increases with increasing discrimination probability and decreases with increasing feature distance, thus preserving the model's statistical judgment ability based on large-sample training while introducing the constraint of quality origin fingerprint parameters on the degree of matching of local characteristics. This embodiment pre-sets a set of preset conditions, including: the discrimination probability corresponding to the candidate origin is greater than a first probability threshold, the corresponding feature distance is less than a first distance threshold, and the comprehensive discrimination score of the candidate origin is ranked highest among all candidate origins. The first probability threshold is determined based on the median of the discrimination probability distribution of the target origin discrimination model on the validation dataset. That is, during the model validation phase, the correct class probability of all validation samples is statistically calculated, and the median is used as the lower limit for selecting reasonable confidence levels. For example, when most correct discrimination probabilities are concentrated above 0.85, the first probability threshold can be set to 0.85. The first distance threshold is determined based on the lower quartile of the feature distance distribution between fingerprint parameters of each origin in the quality-origin association fingerprint database. That is, the feature distance is calculated between each pair of fingerprint parameters of all origins to obtain a set of distance values, and the 25th quartile is used as the maximum allowable distance. For example, when the feature distance between most similar origins is concentrated below 0.5, the first distance threshold can be set to 0.5.

[0089] When a candidate origin simultaneously meets the following conditions: its discrimination probability is higher than a first probability threshold, its feature distance is less than a first distance threshold, and its comprehensive discrimination score is the highest among all candidate origins, this embodiment determines the candidate origin as the target origin and further converts the comprehensive discrimination score corresponding to the target origin into a confidence level. The confidence level is an indicator used in this embodiment to quantify the reliability of the traceability results. It can be normalized to map the comprehensive discrimination score to a numerical range of 0 to 1, or mapped to a percentage. For example, when the comprehensive discrimination score of a target origin is significantly higher than that of other candidate origins, the confidence level can be close to 1 or 100%. Finally, this embodiment encapsulates the target origin and the corresponding confidence level into the traceability result output. The traceability result clearly gives both the determined origin of the wine grape sample to be tested and the degree of credibility of this determination, facilitating regulatory authorities, production enterprises, and researchers to comprehensively utilize the model output and terroir fingerprint for decision-making, thereby improving the accuracy and interpretability of the traceability of the origin of wine grape quality.

[0090] Corresponding to the above methods, such as Figure 2 As shown, this embodiment also provides a traceability system for the quality and origin of wine grapes, including:

[0091] The basic sample construction unit is used to collect wine grape samples with known origin labels in different production areas, obtain information on the location, weather, soil and cultivation management of the corresponding plots, and form a basic sample set containing quality information and environmental information;

[0092] The quality feature extraction unit is used to detect the physicochemical and sensory indicators of wine grape samples in the basic sample set, and after preprocessing and normalization, construct a multi-source fusion feature vector corresponding to the environmental information.

[0093] The key feature identification unit is used to extract key quality features and environmental features that distinguish the place of origin based on the multi-source fusion feature vector and the place of origin label, and generate quality place of origin fingerprint parameters for each place of origin.

[0094] The discrimination model training unit is used to train a classification model with the multi-source fusion feature vector and the origin label to obtain a target origin discrimination model, and associate the target origin discrimination model with the quality origin fingerprint parameters to form a quality-origin association fingerprint database.

[0095] The feature acquisition unit is used to acquire the multi-source fusion feature vector of the wine grape sample to be tested;

[0096] The traceability discrimination output unit is used to input the multi-source fusion feature vector to be tested into the target origin discrimination model, obtain the origin discrimination result, match it with the fingerprint parameters in the quality-origin association fingerprint database, determine the target origin based on the discrimination probability and feature distance, and output the traceability result containing the target origin and confidence level.

[0097] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0098] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for tracing the origin of wine grapes, characterized in that, include: Wine grape samples with known origin labels were collected from different production areas to obtain information on the location, weather, soil, and cultivation management of the corresponding plots, forming a basic sample set containing quality and environmental information; Physicochemical and sensory indicators were detected on the wine grape samples in the basic sample set. After preprocessing and normalization, a multi-source fusion feature vector corresponding to the environmental information was constructed. Based on the multi-source fusion feature vector and the origin label, the key quality features and environmental features that distinguish the origin are extracted using a feature selection algorithm to generate quality origin fingerprint parameters for each origin. A classification model is trained using the multi-source fusion feature vector and the place of origin label to obtain a target place of origin discrimination model. The target place of origin discrimination model is then associated with the quality place of origin fingerprint parameters to form a quality-place of origin association fingerprint database. Obtain the multi-source fusion feature vector of the wine grape sample to be tested; The multi-source fusion feature vector to be tested is input into the target origin discrimination model to obtain the origin discrimination result, and it is matched with the fingerprint parameters in the quality-origin association fingerprint database. The target origin is determined according to the discrimination probability and feature distance, and the traceability result containing the target origin and confidence level is output.

2. The method for tracing the origin of wine grapes according to claim 1, characterized in that, Grape samples with known appellation labels were collected from different production areas to obtain information on the location, climate, soil, and cultivation management of the corresponding plots, forming a basic sample set containing both quality and environmental information, including: Sampling strata are set up according to the distribution of production area plots, and wine grape samples with known production area labels are collected from each sampling stratum. Each wine grape sample is bound to its plot location parameters to form a sample record. Based on the location parameters of the plot, data fields of temperature, precipitation, solar radiation, organic matter content and pH value are retrieved from the meteorological database and the soil database, respectively. The data fields are associated with the corresponding sample records to form the environmental information. For each wine grape sample, the sugar content, total acid, pH value, phenolic content, and soluble solids content are measured to obtain corresponding quality test data fields. These quality test data fields are then integrated with the sample-level environmental information to generate the basic sample set containing both quality and environmental information.

3. The method for tracing the origin of wine grapes according to claim 1, characterized in that, Physicochemical and sensory indicators were detected in the wine grape samples from the basic sample set. After preprocessing and normalization, a multi-source fusion feature vector corresponding to the environmental information was constructed, including: For each wine grape sample in the basic sample set, the sugar content, total acid, pH value, phenolic content and soluble solids content are detected to form corresponding physicochemical index data fields, and the color, aroma intensity and taste structure are recorded to form sensory index data fields. The physicochemical index data field and the sensory index data field are processed according to a preset format to remove missing values ​​and identify outliers, resulting in preprocessed quality data records. The quality data records are subjected to interval scaling normalization to ensure that the numerical scale of the quality data records is consistent with that of the environmental information, thereby forming normalized quality features. The normalized quality features are concatenated with the corresponding environmental information to obtain the multi-source fusion feature vector.

4. The method for tracing the origin of wine grapes according to claim 1, characterized in that, Based on the multi-source fusion feature vector and the origin label, a feature selection algorithm is used to extract key quality features and environmental features that distinguish the origin, generating quality origin fingerprint parameters for each origin, including: The multi-source fusion feature vectors are grouped according to the origin label to form a feature data set divided by origin. A correlation analysis was performed on the feature data set to filter out quality and environmental features that were not sufficiently relevant to the place of origin label, thus obtaining a candidate feature set. The feature selection algorithm is applied to the candidate feature set to calculate the feature importance score, thereby obtaining the key quality features and key environmental features that can distinguish different production areas. Based on the key quality features and key environmental features, the feature distribution parameters of each production area are calculated to form a parameter set used to characterize the differences in features of each production area. The parameter set is organized according to the place of origin label to generate the quality place of origin fingerprint parameters.

5. The method for tracing the origin of wine grapes according to claim 4, characterized in that, A feature selection algorithm is applied to the candidate feature set to calculate feature importance scores, thereby obtaining key quality features and key environmental features that can distinguish different production areas. Based on these key quality features and key environmental features, feature distribution parameters for each production area are calculated, forming a parameter set used to characterize the differences in characteristics among production areas, including: For each candidate feature in the candidate feature set, calculate the sample mean and sample variance of each origin on the candidate feature according to the origin label, and calculate the population mean and population variance of all origins on the candidate feature; Based on the sample mean, sample variance, population mean, and population variance, the feature importance score for each candidate feature is calculated using the following formula: ;in, For the first Feature importance scores for each candidate feature; Place of origin In the The sample mean on each candidate feature; For all places of origin in the first The overall mean over each candidate feature; Place of origin In the Sample variance over each candidate feature; Place of origin The corresponding number of samples; These are positive real numbers used to prevent the denominator from being zero and to suppress numerical fluctuations. This indicates a summation operation across all production locations; Candidate features with feature importance scores greater than a preset threshold are identified as key quality features and key environmental features. Feature distribution parameters for each production area are calculated based on each key quality feature and each key environmental feature, using the following formula: ;in, Place of origin In the Feature distribution parameters on key features; For all places of origin in the first Overall variance over key features; The characteristic distribution parameters of the same place of origin on all key quality characteristics and key environmental characteristics are arranged in the order of characteristic index to form a parameter set corresponding to the place of origin, and the parameter set is used as the quality fingerprint parameter of the place of origin.

6. The method for tracing the origin of wine grapes according to claim 1, characterized in that, A classification model is trained using the multi-source fusion feature vector and the origin label to obtain a target origin discrimination model. The target origin discrimination model is then associated with the quality origin fingerprint parameters to form a quality-origin association fingerprint database, including: The multi-source fusion feature vector is divided into a training dataset and a validation dataset according to the origin label, and the training dataset is used as input to build a classification model to perform origin discrimination learning. During the training of the classification model, the verification accuracy of each iteration cycle is recorded, and the training round with stable verification accuracy and reaching a preset threshold is selected based on the changes in the verification accuracy, so as to obtain a target origin discrimination model that meets the origin identification accuracy requirements. The category output structure of the target origin discrimination model is matched one-to-one with the origin index of the quality origin fingerprint parameter to establish a mapping relationship between the classification output result and the feature distribution parameter. The target origin discrimination model and its mapping relationship with the quality origin fingerprint parameters are organized according to the origin label and stored as a quality-origin association fingerprint database.

7. The method for tracing the origin of wine grapes according to claim 1, characterized in that, The multi-source fusion feature vector to be tested is input into the target origin discrimination model to obtain the origin discrimination result, which is then matched with the fingerprint parameters in the quality-origin association fingerprint database. The target origin is determined based on the discrimination probability and feature distance, and the traceability result containing the target origin and confidence level is output, including: The multi-source fusion feature vector to be tested is input into the target origin discrimination model to obtain the discrimination probability distribution for each origin, thus forming the origin discrimination result; Based on the candidate origin labels appearing in the origin discrimination results, the corresponding fingerprint parameters are retrieved from the quality-origin association fingerprint database, and the feature distance relative to each fingerprint parameter is calculated based on the values ​​of the multi-source fusion feature vector to be tested on key quality features and key environmental features, thus forming the feature distance result. For each candidate origin, the discrimination probability in the origin discrimination result is combined with the feature distance result to obtain a comprehensive discrimination score, and the candidate origin that meets the preset conditions is selected as the target origin. The target origin and its corresponding comprehensive discrimination score are converted into a confidence level, and a traceability result containing the target origin and the confidence level is generated.

8. The method for tracing the origin of wine grapes according to claim 7, characterized in that, The preset conditions include: the discrimination probability is greater than a first probability threshold, the feature distance is less than a first distance threshold, and the comprehensive discrimination score is ranked highest among all candidate origins; the first probability threshold is determined based on the median value of the discrimination probability distribution of the target origin discrimination model on the verification dataset, and the first distance threshold is determined based on the lower quartile of the feature distance distribution between the fingerprint parameters of each origin in the quality-origin association fingerprint database.

9. A system for tracing the origin of wine grapes, characterized in that, include: The basic sample construction unit is used to collect wine grape samples with known origin labels in different production areas, obtain information on the location, weather, soil and cultivation management of the corresponding plots, and form a basic sample set containing quality information and environmental information; The quality feature extraction unit is used to detect the physicochemical and sensory indicators of wine grape samples in the basic sample set, and after preprocessing and normalization, construct a multi-source fusion feature vector corresponding to the environmental information. The key feature identification unit is used to extract key quality features and environmental features that distinguish the place of origin based on the multi-source fusion feature vector and the place of origin label, and generate quality place of origin fingerprint parameters for each place of origin. The discrimination model training unit is used to train a classification model with the multi-source fusion feature vector and the origin label to obtain a target origin discrimination model, and associate the target origin discrimination model with the quality origin fingerprint parameters to form a quality-origin association fingerprint database. The feature acquisition unit is used to acquire the multi-source fusion feature vector of the wine grape sample to be tested; The traceability discrimination output unit is used to input the multi-source fusion feature vector to be tested into the target origin discrimination model, obtain the origin discrimination result, match it with the fingerprint parameters in the quality-origin association fingerprint database, determine the target origin based on the discrimination probability and feature distance, and output the traceability result containing the target origin and confidence level.

Images