A food supply chain large model knowledge graph identification and treatment analysis system and method

Abstract

The application discloses a food supply chain large model knowledge graph identification governance analysis system and method, and the system comprises a large model semantic processing module, a knowledge graph storage and graph calculation module, an identification abnormal intelligent risk control module, a natural language analysis interactive module and an algorithm adaptive iteration optimization module. Through the division and cooperation of large model semantic intelligent processing and knowledge graph structured storage, the application realizes food supply chain scene automatic coding template generation, heterogeneous data automatic registration, food product global identification association networking; through a double-dimensional fusion risk control mechanism, accurate abnormal identification is realized; through natural language ambiguity resolution and interactive completion, the use threshold of food traceability query is reduced; through a complete set of algorithm adaptive optimization and incremental iteration mechanism, the dynamic evolution of model parameters, judgment threshold and feature weight is realized, and the automation level, security protection capability, interactive convenience and scene generalization capability of food supply chain identification governance are improved.

Description

Technical Field

[0001] This invention relates to the fields of food supply chain identifier resolution, knowledge graph construction, and supply chain data governance technology, specifically to a food supply chain large-scale model knowledge graph identifier governance and resolution system and method. Background Technology

[0002] The current food supply chain identification and resolution system has fully covered the entire chain of food raw materials, semi-finished products, finished products, cold chain storage, logistics and circulation, and quality testing. Relying on identification technology, it realizes traceability and control of food throughout the entire process, which is an important foundation for food safety supervision and supply chain risk management.

[0003] Existing identifier resolution and governance technologies still have many key technical shortcomings. First, the traditional food supply chain identifier coding, matching, and registration process relies heavily on manual intervention. The data formats of various business operations such as ingredients, processing, and cold chain logistics are messy, requiring manual identification of entity types, customization of coding templates, and completion of identifier binding and registration. This results in high labor costs, low implementation efficiency, and poor standardization, and cannot meet the needs of large-scale network governance of massive amounts of ingredients and goods in the food supply chain.

[0004] Secondly, the existing food supply chain labeling risk control system has a low level of intelligence, and mostly adopts a fixed rule hard verification mode. It can only identify basic anomalies such as format errors, but cannot identify complex risk behaviors such as counterfeit food labels, invalid labels, malicious batch registration, high-frequency crawler scanning, and unauthorized abnormal parsing. This can easily lead to security risks such as false labels entering the network, label data leakage, and parsing link congestion. The credibility and stability of the food labeling system are insufficient.

[0005] Finally, traditional food labeling resolution and interaction has a high barrier to entry, requiring users to be proficient in professional coding rules, standardized search commands and platform operation procedures. Ordinary business personnel and regulatory personnel cannot quickly complete label traceability and data query, resulting in poor human-computer interaction adaptability and low utilization rate of label data.

[0006] In addition, existing AI-enabled food supply chain labeling governance technologies generally suffer from problems such as ambiguous module division of labor, static and fixed algorithm models, inability to iterate parameters, single judgment logic, and insufficient rigor of formulas. They also have weak model generalization ability, poor scenario adaptability, high long-term operation and maintenance costs, and lack closed-loop evolution capabilities for self-learning, self-optimization, and self-adaptation for the food supply chain. Summary of the Invention

[0007] The purpose of this invention is to provide a knowledge graph identification, governance, and parsing system and method for a large-scale food supply chain model, in order to solve the technical problems in the existing technology, such as high cost of manual governance of the food supply chain, weak risk control capabilities, high interaction threshold, fixed and non-iterative models, poor logical self-consistency, and insufficient algorithm rigor.

[0008] To address the aforementioned issues, this invention discloses a food supply chain large-scale model knowledge graph identification governance and parsing system, comprising: a large-scale model semantic processing module, a knowledge graph storage and graph computation module, an intelligent risk control module for identifying anomalies, a natural language parsing and interaction module, and an algorithm adaptive iterative optimization module;

[0009] The large model semantic processing module is used to perform semantic layer processing operations such as entity recognition, semantic extraction, intent understanding and relation annotation, and output standardized entity relation triples;

[0010] The knowledge graph storage and graph computation module is used to receive standardized entity relationship triples output by the large model semantic processing module, perform triple structured storage, node link management, graph traversal query and association verification processing, and construct an identifier knowledge graph for the food supply chain.

[0011] The intelligent risk control module for identifying abnormalities is used to perform abnormal identification processing on food supply chain labels and output the corresponding risk control results.

[0012] The natural language parsing and interaction module is used to receive natural language queries input by users, complete semantic processing, convert the user's natural language query into standardized identifier parsing instructions, call the knowledge graph storage and graph computing module to obtain query results, and feed the query results back to the user.

[0013] The algorithm adaptive iterative optimization module is used to perform dynamic iterative optimization of model weights, decision thresholds, and feature weights based on the running samples generated during system operation.

[0014] Preferably, the large-scale model semantic processing module is a large-scale model that has been fine-tuned for the industry. This module integrates multiple identification standards, industry business specifications, and scenario adaptation requirements, adaptively generating hierarchical and standardized identification coding templates based on Handle, OID, and Ecode specifications. For structured, semi-structured, and unstructured heterogeneous data in the food supply chain scenario, it automatically matches the optimal coding template through entity recognition, attribute extraction, and type determination. This completes the automatic generation of unique entity identifiers for the food supply chain, automatic entry of attribute information, and automatic registration of parsing nodes, enabling seamless coding and automatic network access for food supply chain entities. Furthermore, it performs semantic relationship extraction and attribute annotation processing on the entire network's food supply chain identification data, outputting standardized entity relationship triples.

[0015] The large-scale model semantic processing module includes an adaptive scoring model for identifier encoding and a heterogeneous entity classification and matching probability model; the calculation formula for the adaptive scoring model for identifier encoding is:

[0016] ;

[0017] In the formula, The overall adaptation score for the encoding rules; For multi-identification system standard compliance score; Score based on industry scenario adaptation; Scoring based on entity type attributes; These are dynamic weighting coefficients, which are updated in real time by the algorithm's adaptive iterative tuning module.

[0018] The formula for calculating the heterogeneous entity classification matching probability model is as follows:

[0019] ;

[0020] In the formula, For entity multidimensional feature vectors; The encoding template category for the i-th type of entity; The posterior probability of matching the corresponding template to the entity is determined; the large model semantic processing module selects the template corresponding to the maximum posterior probability to complete automatic coding and registration.

[0021] Preferably, the knowledge graph storage and graph computing module constructs an identifier knowledge graph. The nodes of the identifier knowledge graph cover all types of food supply chain entity identifiers, including food raw materials, semi-finished ingredients, finished food products, processing work orders, cold chain warehousing, food logistics, and food quality inspection records. The node edge relationships of the identifier knowledge graph include multi-dimensional relationships such as entity affiliation, temporal flow, operation and maintenance association, upstream and downstream of the supply chain, and business coupling. For identifier lifecycle change scenarios, the knowledge graph storage and graph computing module adopts a management mechanism of logical deletion and timestamp expiration. When an identifier is cancelled, scrapped, deactivated, or changed, historical nodes and associated links are not physically deleted; only an expiration timestamp is marked, historical traceability and audit data are retained, and new graph nodes are generated for new business entities. The knowledge graph storage and graph computing module uses the standardized entity relationship triples output by the large model semantic processing module to enable real-time incremental updates of graph nodes and edge relationships. The knowledge graph storage and graph computing module includes an identifier entity association degree calculation model. The calculation formula for the identifier entity association degree calculation model is:

[0022] ;

[0023] In the formula, The degree of business relevance between the two identified entities; Temporal similarity; For business semantic similarity; Link-level similarity; These are the normalized weighting coefficients.

[0024] Preferably, the intelligent risk control module for identifying abnormalities constructs a two-dimensional abnormality identification system that combines static detection of the authenticity of food supply chain identifiers with dynamic detection of user behavior, and executes a hierarchical fusion judgment strategy. The static detection of identifier authenticity verifies the compliance and authenticity of a single food supply chain identifier, and comprehensively judges whether the identifier is invalid, counterfeit, or illegal by combining coding standards, entity association relationships, and data source credibility. The dynamic detection of user behavior monitors the registration, parsing, and access behavior of food supply chain identifiers in real time, and identifies malicious behaviors such as batch registration, high-frequency scanning, crawler traversal, and abnormal cross-domain access. The hierarchical fusion judgment strategy includes: when both the identifier authenticity abnormality score and the behavior abnormality score exceed the corresponding threshold, it is judged as a high-risk malicious behavior, and the intelligent risk control module for identifying abnormalities directly intercepts the registration or parsing request, freezes the operating entity, and reports an alarm; when only a single dimension score exceeds the corresponding threshold, it is judged as a suspected abnormality, and traffic restriction control, log retention, and push for manual review are executed without directly interrupting the business; when both dimension scores are below the threshold, it is judged as a normal request and is directly allowed to flow.

[0025] Preferably, the intelligent risk control module for identifying anomalies includes an identification authenticity tolerance scoring model and a behavior anomaly normalization scoring model; the calculation formula for the identification authenticity tolerance scoring model is:

[0026] ;

[0027] In the formula, To identify the confidence level of anomalies; This is a data source credibility factor used to weaken the correlation penalty for data registered in the trusted whitelist; This is an entity lifecycle maturity factor used to reduce the association verification weight for newly registered entities; These are the dimension weight coefficients;

[0028] The calculation formula for the behavioral anomaly normalization scoring model is as follows:

[0029] ;

[0030] In the formula, Request frequency per unit time; To request the identification of dispersion; The frequency statistics are the mean and standard deviation; The mean and standard deviation are the statistical values ​​for the dispersion.

[0031] Preferably, the natural language parsing and interaction module is used to construct a semantic ambiguity resolution and interactive information completion mechanism;

[0032] The natural language parsing and interaction module performs intent recognition, keyword extraction, and semantic matching on the natural language queries related to food traceability input by the user. When the semantic matching degree is lower than a preset threshold, it automatically determines that the query statement has missing information or semantic ambiguity, and actively triggers interactive clarification, prompting the user to supplement missing fields such as time range, entity type, batch information, and business link. At the same time, it outputs the top K highly matched candidate parsing tasks for the user to select and confirm. After the semantic matching is successful, the natural language requirement is automatically converted into standardized identifier parsing instructions, and the knowledge graph query interface of the knowledge graph storage and graph computing module is called to aggregate the full-link related data of food supply chain entities and generate structured and visualized query results to be fed back to the user.

[0033] The natural language parsing and interaction module includes a natural language semantic mapping matching model, and the calculation formula for the natural language semantic mapping matching model is as follows:

[0034] ;

[0035] In the formula, For semantic matching degree; Improve the accuracy of user intent recognition; For text semantic vector similarity; For core keyword matching degree; These are the weighting coefficients.

[0036] Preferably, the algorithm adaptive iterative tuning module is used to perform offline batch optimization, dynamic threshold update, incremental learning, and positive and negative sample feedback closed-loop processing;

[0037] The algorithm's adaptive iterative optimization module updates the weight vectors of each model using a daily offline batch gradient descent iterative method, and configures an adaptive learning rate and a loss function specific to each model; it dynamically updates the judgment threshold based on the temporal distribution characteristics of food supply chain business data flow; it continuously collects new food entity types, new food supply chain business scenarios, new risk behaviors, and diversified query sentences to build an incremental sample pool, and completes model parameter updates through incremental fine-tuning; it collects misjudged, missed, and mismatched samples after manual review to form a supervised learning closed loop, and back-optimizes the model weights, semantic vector space, and risk feature library.

[0038] Preferably, the algorithm adaptive iterative tuning module employs gradient descent weight adaptive iteration, and the iterative formula used for gradient descent weight adaptive iteration is:

[0039] ;

[0040] In the formula, This is the current weight vector; Adaptive learning rate; The gradient of the loss function; The weights are updated iteratively;

[0041] Coding adaptation model with The loss function is used; the knowledge association model uses... The loss function is used in the anomaly risk control model. The loss function is used; the semantic mapping model uses the cosine distance of semantic vectors. This is the loss function.

[0042] Preferably, the algorithm adaptive iterative tuning module is used to perform dynamic threshold adaptive updates, and the calculation formula used for the dynamic threshold adaptive update is:

[0043] ;

[0044] In the formula, For time series smoothing coefficients; This represents the mean of the recent normal sample data distribution. , The threshold values ​​are the judgment thresholds before and after the update. The algorithm adaptive iterative optimization module completes adaptive balancing through the calculation formula. During peak business hours, a lenient threshold is used to ensure business access, while during quiet periods, a strict threshold is used to ensure system security.

[0045] This application also discloses a knowledge graph identifier governance and parsing method for a large-scale food supply chain model, including:

[0046] S1: Multi-source heterogeneous food supply chain business data and user query data are connected to the system mentioned above to complete data cleaning and format normalization preprocessing.

[0047] S2: The preprocessed data is sent to the large model semantic processing module to complete semantic extraction, entity classification, encoding matching and output of standardized entity relation triples;

[0048] S3: Write the standardized entity relationship triples into the knowledge graph storage and graph computation module to complete the structured storage and networking process;

[0049] S4: Call the intelligent risk control module for abnormal identification to complete the detection of the authenticity of food supply chain identification and abnormal user behavior, and realize risk control through the integrated decision-making mechanism;

[0050] S5: The user's natural language query is converted into a standardized identifier parsing instruction by the natural language parsing and interaction module through ambiguity resolution and semantic mapping. The knowledge graph query capability is then invoked to aggregate the full-link related data of the food supply chain entities and fed back to the user.

[0051] S6: The entire process operation log is precipitated as iterative samples and sent to the algorithm adaptive iterative optimization module for offline algorithm optimization and incremental learning, so as to realize the continuous iterative upgrade of the system capabilities.

[0052] Beneficial effects:

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] First, this application replaces the manual coding, matching, and registration operations in the traditional food supply chain scenario with a large model semantic processing module, which greatly reduces the cost of human operation and maintenance. It can meet the large-scale network governance of massive goods in the food supply chain, and can adapt to multiple identification systems such as Handle, OID, and Ecode as well as the private business rules of food enterprises, thus meeting the unified management and control needs of heterogeneous identifications in multi-category food supply chains.

[0055] Secondly, this application uses a two-dimensional anomaly identification approach, combining static detection of label authenticity with dynamic detection of user behavior, along with a hierarchical fusion judgment strategy and fault-tolerant correction logic. This approach completely solves the problems of false detection, missed detection, and judgment conflicts in traditional risk control models, effectively intercepting risky behaviors such as counterfeit labels, malicious batch registration, and abnormal parsing scanning, and comprehensively improving the security and stability of the food supply chain label resolution system.

[0056] Third, this application adopts a management mechanism of logical deletion plus timestamp expiration through the knowledge graph storage and graph computing module. While meeting the dynamic update of food supply chain identification throughout its entire life cycle, it also retains complete historical business data of the entire food chain, meeting the actual business needs of food safety auditing, quality traceability, and liability determination.

[0057] Fourth, this application addresses the practical problems of semantic ambiguity and incomplete information in colloquial food queries by using the semantic ambiguity resolution, interactive field completion, and candidate task recommendation mechanisms of the natural language parsing interaction module. This enables universal food label queries without professional barriers and significantly improves the utilization rate of label data.

[0058] Fifth, this application achieves weight gradient iteration, dynamic threshold optimization, incremental learning, and positive and negative sample feedback closed-loop optimization through an algorithm adaptive iterative optimization module, breaking the limitations of static and fixed traditional algorithm models; it can continuously optimize model accuracy according to changes in business scenarios, data distribution, and risk types, and significantly improves scenario generalization ability and long-term operation and maintenance stability. Attached Figure Description

[0059] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0060] Figure 1 This is a structural block diagram of a knowledge graph identification, governance, and parsing system for a large-scale food supply chain model disclosed in this invention.

[0061] Figure 2 This is a flowchart of a knowledge graph identification governance and parsing method for a large-scale food supply chain model disclosed in this invention. Detailed Implementation

[0062] It should be understood that in the description of this invention, the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0063] It should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "set," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0064] The technical solution of the present invention will be further illustrated below through specific embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The content of the embodiments does not constitute a limitation on the present invention. 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.

[0065] like Figure 1 As shown, to achieve the above-mentioned objectives, this invention provides a food supply chain large-scale model knowledge graph identification governance and parsing system, including: a large-scale model semantic processing module, a knowledge graph storage and graph computation module, an identification anomaly intelligent risk control module, a natural language parsing and interaction module, and an algorithm adaptive iterative optimization module.

[0066] The large model semantic processing module is configured to perform semantic layer processing operations such as entity recognition, semantic extraction, intent understanding, and relation annotation, and output standardized entity relation triples.

[0067] Furthermore, the large-scale model semantic processing module is a large-scale model that has been fine-tuned for the industry. This module is configured to integrate multiple identification standards, industry business specifications, and scenario adaptation requirements, adaptively adapting to Handle, OID, and Ecode specifications, intelligently generating hierarchical and standardized identification coding templates, and supporting personalized fine-tuning of enterprise-specific business rules. For structured, semi-structured, and unstructured heterogeneous data in the food supply chain scenario, it automatically matches the optimal coding template through entity recognition, attribute extraction, and type determination. This completes the automatic generation of unique identifiers for food supply chain entities, automatic entry of attribute information, and automatic registration of parsing nodes, achieving seamless coding and automatic network access for food supply chain entities. It also performs semantic relationship extraction and attribute annotation processing on the entire network's food supply chain identification data, outputting standardized entity relationship triples.

[0068] Among them, the large model that has been fine-tuned by the industry refers to the large model that is based on a general pre-trained large model and fine-tuned by using labeled samples in the field of food supply chain identification, multi-system identification standard text, and food supply chain business description dataset. By slightly updating some weight parameters of the model, the large model can master the ability of food entity recognition, attribute extraction, and entity relationship labeling, and is used for food supply chain semantic processing tasks of Handle, OID, and Ecode multi-identification systems.

[0069] The large model semantic processing module includes an identifier encoding adaptive scoring model and a heterogeneous entity classification matching probability model.

[0070] The formula for the adaptive scoring model of identifier coding is as follows:

[0071] ;

[0072] In the formula, The overall adaptation score for the encoding rules; For multi-identification system standard compliance score; Score based on industry scenario adaptation; Scoring based on entity type attributes; These are dynamic weighting coefficients, which are adaptively and iteratively updated in real time by the algorithm.

[0073] The formula for calculating the heterogeneous entity classification matching probability model is as follows:

[0074] ;

[0075] In the formula, For entity multidimensional feature vectors; The encoding template category for the i-th type of entity; The posterior probability of matching the corresponding template to the entity is determined; the large model semantic processing module selects the template corresponding to the maximum posterior probability to complete the automatic coding and registration.

[0076] The large-scale model semantic processing module handles semantic layer processing, transforming multi-source, heterogeneous, unstructured, and semi-structured raw information from the food supply chain into standardized entity relationship triples that can be recognized by machines. The identifier coding adaptive scoring model performs weighted calculations from multiple dimensions, including standard compliance, scenario adaptation, and entity type, to achieve objective quantitative screening of coding templates. The heterogeneous entity classification and matching probability model automatically matches entities with templates based on probability calculations, enabling large-scale, seamless coding and network integration of food supply chain entities. This effectively reduces manual maintenance costs and provides a semantic data source for knowledge graph construction.

[0077] The knowledge graph storage and graph computation module receives standardized entity relationship triples output by the large model semantic processing module, performs structured storage of triples, node link management, graph traversal query, and relationship verification processing, and constructs an identifier knowledge graph for the food supply chain.

[0078] Furthermore, the knowledge graph storage and graph computing module is also used to construct an identifier knowledge graph. The nodes of this identifier knowledge graph cover all types of food supply chain entity identifiers, including food raw materials, semi-finished ingredients, finished food products, processing work orders, cold chain warehousing, food logistics, and food quality inspection records. The node edge relationships in the identifier knowledge graph include multi-dimensional relationships such as entity affiliation, temporal flow, operational association, upstream and downstream supply chain relationships, and business coupling. For changes throughout the identifier's lifecycle, the knowledge graph storage and graph computing module employs a management mechanism of logical deletion plus timestamp expiration. When an identifier is cancelled, scrapped, deactivated, or changed, historical nodes and associated links are not physically deleted; only an expiration timestamp is marked, historical traceability and audit data are retained, and new graph nodes are generated for new business entities. Based on the standardized entity relationship triples output by the large model semantic processing module, the knowledge graph storage and graph computing module achieves real-time incremental updates of graph nodes and edge relationships.

[0079] The knowledge graph storage and graph computation module includes an entity association degree calculation model. The calculation formula for the entity association degree calculation model is as follows:

[0080] ;

[0081] In the formula, The degree of business relevance between the two identified entities; Temporal similarity; For business semantic similarity; Link-level similarity; These are the normalized weighting coefficients.

[0082] The knowledge graph storage and graph computation module completes the structured networking of identifiable entities and business relationships in the food supply chain. Utilizing logical deletion and timestamp expiration management mechanisms, it avoids physically deleting historical business data, thus fully preserving historical records required for food lifecycle traceability and auditing. Through graph traversal query capabilities, it can quickly complete chain-like traceability queries of food entities' upstream and downstream processes and temporal flows. The entity association calculation model integrates temporal, semantic, and link dimensions to quantify the strength of business relationships between different food entities, outputting calculable association values. This provides quantifiable evidence of association relationships for subsequent verification of the authenticity of food supply chain identifiers.

[0083] The intelligent risk control module for identifying anomalies in food supply chain labels is used to perform anomaly identification processing and output corresponding risk control results.

[0084] Furthermore, the intelligent risk control module for identifying anomalies constructs a two-dimensional anomaly identification system that combines static detection of the authenticity of food supply chain identifiers with dynamic detection of user behavior, and executes a hierarchical and integrated judgment strategy. Static identifier authenticity detection is configured to verify the compliance and authenticity of individual food supply chain identifiers, comprehensively determining whether an identifier is invalid, counterfeit, or in violation of regulations based on coding standards, entity relationships, and data source credibility. Dynamic user behavior detection is configured to monitor the registration, parsing, and access behavior of food supply chain identifiers in real time, identifying malicious behaviors such as batch registration, high-frequency scanning, web crawling, and abnormal cross-domain access.

[0085] The tiered and integrated judgment strategy includes: when both the authenticity score and the behavior score exceed the corresponding thresholds, it is judged as high-risk malicious behavior. The intelligent risk control module for identifying anomalies directly intercepts the registration or parsing request, freezes the operating entity, and reports an alert. When only one dimension score exceeds the corresponding threshold, it is judged as suspected anomaly, and rate limiting control is implemented, logs are retained, and the request is pushed to manual review without directly interrupting business. When both dimension scores are below the threshold, it is judged as a normal request, and the request is directly allowed to proceed.

[0086] Furthermore, the intelligent risk control module for identifying anomalies includes an identification authenticity tolerance scoring model and a behavior anomaly normalization scoring model.

[0087] The formula for calculating the authenticity tolerance scoring model is as follows:

[0088] ;

[0089] In the formula, To identify the confidence level of anomalies; The data source credibility factor is a penalty for weakening the relevance of data registered in the trusted whitelist. This is the entity lifecycle maturity factor, which reduces the association verification weight for newly registered entities. These are the dimension weight coefficients.

[0090] In addition, the formula for calculating the behavioral anomaly normalized scoring model is as follows:

[0091] ;

[0092] In the formula, Request frequency per unit time; To request the identification of dispersion; The frequency statistics are the mean and standard deviation; The mean and standard deviation are the statistical values ​​for the dispersion.

[0093] The intelligent risk control module for identifying anomalies abandons the traditional single, fixed-rule verification method. It conducts joint detection from two dimensions: the authenticity of the food supply chain identifiers themselves and user access behavior. Combined with a tiered and integrated judgment strategy, it distinguishes between high-risk malicious behavior, suspected abnormal behavior, and normal requests, balancing system security and business continuity. The identifier authenticity fault-tolerant scoring model introduces data source trust factors and entity lifecycle maturity factors to perform fault-tolerant corrections for trusted source data and newly registered food entities, reducing the probability of newly registered legitimate isolated food entities being misjudged. The behavior anomaly normalization scoring model standardizes request frequency and request identifier dispersion, eliminating differences in different feature dimensions and enabling quantitative assessment of access behavior risks. It effectively identifies risks such as forged identifiers, bulk malicious registration, and web crawler scanning, improving the security and anti-attack capabilities of the entire food supply chain identifier resolution system.

[0094] The Natural Language Parsing Interaction Module receives user-input natural language queries for food traceability, performs semantic processing, converts the user's natural language query into standardized identifier parsing instructions, calls the knowledge graph storage and graph computing module to obtain the query results, and then feeds the query results back to the user.

[0095] Furthermore, the natural language parsing and interaction module constructs a semantic ambiguity resolution and interactive information completion mechanism, and performs intent recognition, keyword extraction, and semantic matching on the user's input natural language query for food traceability. When the semantic matching degree is lower than a preset threshold, it automatically determines that the query statement has missing information or semantic ambiguity, and proactively triggers interactive clarification, prompting the user to supplement missing fields such as time range, entity type, batch information, and business chain. At the same time, it outputs the top K highly matched candidate parsing tasks for the user to select and confirm. After the semantic matching is successful, the natural language requirement is automatically converted into standardized identifier parsing instructions, calling the knowledge graph query interface of the knowledge graph storage and graph computing module, aggregating the full-chain related data of food supply chain entities, and generating structured and visualized query results to be fed back to the user.

[0096] The natural language parsing interaction module includes a natural language semantic mapping matching model. The calculation formula for the natural language semantic mapping matching model is as follows:

[0097] ;

[0098] In the formula, For semantic matching degree; Improve the accuracy of user intent recognition; For text semantic vector similarity; For core keyword matching degree; These are the weighting coefficients.

[0099] The natural language parsing and interaction module solves the problem of traditional food traceability label parsing requiring mastery of professional coding and retrieval commands. Users can conduct food traceability queries using everyday natural language. The natural language semantic mapping and matching model integrates intent, vectors, and keywords to calculate semantic matching degree, quantifying the degree of matching between user questions and standard parsing tasks. This ensures the accuracy of natural language to machine standard commands and effectively improves the utilization rate of label data.

[0100] The algorithm adaptive iterative optimization module is used to perform dynamic iterative optimization of model weights, decision thresholds, and feature weights based on the running samples generated during system operation.

[0101] Furthermore, the algorithm adaptive iterative tuning module is used to perform offline batch optimization, dynamic threshold updates, incremental learning, and positive and negative sample feedback closed-loop processing. The module updates the weight vectors of each model using a daily offline batch gradient descent iterative approach, configuring adaptive learning rates and model-specific loss functions; dynamically updates judgment thresholds based on the temporal distribution characteristics of food supply chain business data streams; continuously collects newly added food entity types, new food supply chain business scenarios, new risk behaviors, and diverse query formats to build an incremental sample pool, and updates model parameters through incremental fine-tuning. It collects misjudged, missed, and incorrectly matched samples after manual review, forming a supervised learning closed loop, and inversely optimizes model weights, semantic vector space, and risk feature library.

[0102] The algorithm adaptive iterative tuning module uses gradient descent weight adaptive iteration. The iterative formula used in gradient descent weight adaptive iteration is as follows:

[0103] ;

[0104] In the formula, This is the current weight vector; Adaptive learning rate; The gradient of the loss function; The weights are the weights after the iterative update.

[0105] Coding adaptation model with As the loss function; the knowledge association model uses As a loss function; the anomaly risk control model uses As the loss function; the semantic mapping model uses the cosine distance of the semantic vectors. As a loss function.

[0106] The algorithm adaptive iterative tuning module performs dynamic threshold adaptive updates. The calculation formula used for dynamic threshold adaptive updates is as follows:

[0107] ;

[0108] In the formula, For time series smoothing coefficients; This represents the mean of the recent normal sample data distribution. , These are the judgment thresholds before and after the update; the algorithm adaptive iterative optimization module completes adaptive balancing through calculation formulas, using a lenient threshold during peak business hours to ensure business operation, and a strict threshold during quiet periods to ensure system security.

[0109] Unlike traditional systems where algorithm parameters are statically fixed, the adaptive iterative optimization module employs daily offline gradient descent iterations, ensuring no interference with online business operations. Different business models are configured with dedicated loss functions for targeted optimization. Dynamic thresholds are adaptively updated to accommodate fluctuations in food supply chain business traffic. Relying on an incremental sample pool and positive / negative samples obtained through manual review to form a closed-loop learning process, the system can adapt to new food entity types, novel food supply chain business scenarios, and new risk management methods without requiring retraining on the entire dataset. This results in increasingly higher recognition accuracy over time, reducing long-term maintenance and transformation costs in the food supply chain and enhancing generalization capabilities across various food sub-scenarios.

[0110] This application also discloses a knowledge graph identifier governance and parsing method for a large-scale food supply chain model, such as... Figure 2 As shown, it includes the following steps:

[0111] S1: Multi-source heterogeneous food supply chain data and user query data are integrated into the system to complete data cleaning and format normalization preprocessing;

[0112] S2: The preprocessed data is sent to the large model semantic processing module to complete semantic extraction, entity classification, encoding matching and output of standardized entity relation triples;

[0113] S3: Write standardized entity relationship triples into the knowledge graph storage and graph computation module to complete structured storage and networking processing;

[0114] S4: Call the intelligent risk control module for abnormal identification to complete the detection of the authenticity of food supply chain identification and abnormal user behavior, and realize risk control through the integrated decision-making mechanism;

[0115] S5: User natural language queries are resolved and semantically mapped by the natural language parsing interaction module, and converted into standardized identifier parsing instructions. The knowledge graph query capability is then invoked to aggregate the full-link related data of entities in the food supply chain and fed back to the user.

[0116] S6: The entire process operation log is stored as iterative samples and sent to the algorithm adaptive iterative optimization module to support offline algorithm optimization and incremental learning, so as to realize the continuous iterative upgrade of system capabilities.

[0117] The aforementioned data flow forms a complete closed loop. After preprocessing, multi-source heterogeneous food supply chain data undergoes semantic parsing, graph storage networking, and risk control in sequence. User natural language food traceability queries, after semantic conversion, invoke graph queries to obtain results. Log samples generated during system operation are fed back to the algorithm adaptive iterative optimization module, driving continuous iteration and evolution of various models, thus achieving a complete closed loop of data input-business processing-result output-sample feedback.

[0118] The applicant declares that the above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Those skilled in the art should understand that any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention fall within the scope of protection and disclosure of the present invention.

Claims

1. A knowledge graph identification, governance, and parsing system for a large-scale food supply chain model, characterized in that, include: The module includes a large-scale model semantic processing module, a knowledge graph storage and graph computation module, an anomaly identification and intelligent risk control module, a natural language parsing and interaction module, and an algorithm adaptive iterative optimization module. The large model semantic processing module is used to perform semantic layer processing operations such as entity recognition, semantic extraction, intent understanding and relation annotation, and output standardized entity relation triples; The knowledge graph storage and graph computation module is used to receive standardized entity relationship triples output by the large model semantic processing module, perform triple structured storage, node link management, graph traversal query and association verification processing, and construct an identifier knowledge graph for the food supply chain. The intelligent risk control module for identifying abnormalities is used to perform abnormal identification processing on food supply chain labels and output the corresponding risk control results. The natural language parsing and interaction module is used to receive natural language queries input by users, complete semantic processing, convert the user's natural language query into standardized identifier parsing instructions, call the knowledge graph storage and graph computing module to obtain query results, and feed the query results back to the user. The algorithm adaptive iterative optimization module is used to perform dynamic iterative optimization of model weights, decision thresholds, and feature weights based on the running samples generated during system operation.

2. The knowledge graph identification, governance, and parsing system for a large-scale food supply chain model according to claim 1, characterized in that, The large-scale model semantic processing module is a large-scale model that has been fine-tuned for the industry. This module integrates multiple identification standards, industry business specifications, and scenario adaptation requirements, adaptively generating hierarchical and standardized identification coding templates based on Handle, OID, and Ecode specifications. For structured, semi-structured, and unstructured heterogeneous data in the food supply chain scenario, it automatically matches the optimal coding template through entity recognition, attribute extraction, and type determination. This completes the automatic generation of unique entity identifiers for the food supply chain, automatic entry of attribute information, and automatic registration of parsing nodes, enabling seamless coding and automatic network access for food supply chain entities. Furthermore, it performs semantic relationship extraction and attribute annotation processing on the entire network's food supply chain identification data, outputting standardized entity relationship triples. The large-scale model semantic processing module includes an adaptive scoring model for identifier encoding and a heterogeneous entity classification and matching probability model; the calculation formula for the adaptive scoring model for identifier encoding is: ； In the formula, The overall adaptation score for the encoding rules; For multi-identification system standard compliance score; Score based on industry scenario adaptation; Scoring based on entity type attributes; These are dynamic weighting coefficients, which are updated in real time by the algorithm's adaptive iterative tuning module. The formula for calculating the heterogeneous entity classification matching probability model is as follows: ； In the formula, For entity multidimensional feature vectors; The encoding template category for the i-th type of entity; The posterior probability of matching the corresponding template to the entity is determined; the large model semantic processing module selects the template corresponding to the maximum posterior probability to complete automatic coding and registration.

3. The knowledge graph identification, governance, and parsing system for a large-scale food supply chain model according to claim 1, characterized in that, The knowledge graph storage and graph computation module constructs an identifier knowledge graph. The nodes of this knowledge graph cover all types of food supply chain entity identifiers, including food raw materials, semi-finished ingredients, finished food products, processing orders, cold chain warehousing, food logistics, and food quality inspection records. The node-edge relationships in the identifier knowledge graph include multi-dimensional relationships such as entity affiliation, temporal flow, operation and maintenance association, supply chain upstream and downstream, and business coupling. For identifier lifecycle changes, the knowledge graph storage and graph computation module employs a logical deletion and timestamp expiration management mechanism. When an identifier is cancelled, scrapped, deactivated, or changed, historical nodes and associated links are not physically deleted; only an expiration timestamp is marked, historical traceability and audit data are retained, and new graph nodes are generated for new business entities. Based on the standardized entity relationship triples output by the large model semantic processing module, the knowledge graph storage and graph computation module enables real-time incremental updates of graph nodes and edge relationships. The knowledge graph storage and graph computation module includes an identifier entity association degree calculation model; the calculation formula for the identifier entity association degree calculation model is: ； In the formula, The degree of business relevance between the two identified entities; For temporal similarity; For business semantic similarity; Link-level similarity; These are the normalized weighting coefficients.

4. The knowledge graph identification, governance, and parsing system for a large-scale food supply chain model according to claim 1, characterized in that, The intelligent risk control module for identifying abnormalities constructs a two-dimensional abnormality identification system that combines static detection of the authenticity of food supply chain labels with dynamic detection of user behavior, and executes a hierarchical fusion judgment strategy. The static detection of identifier authenticity verifies the compliance and authenticity of individual food supply chain identifiers, and comprehensively determines whether the identifier is invalid, counterfeit, or illegal by combining coding standards, entity relationships, and data source credibility. The dynamic detection of user behavior monitors the registration, parsing, and access behavior of food supply chain identifiers in real time, and identifies malicious behaviors such as batch registration, high-frequency scanning, crawler traversal, and abnormal cross-domain access. The hierarchical fusion judgment strategy includes: when both the identifier authenticity anomaly score and the behavior anomaly score exceed the corresponding threshold, it is judged as high-risk malicious behavior, and the identifier anomaly intelligent risk control module directly intercepts the registration or parsing request, freezes the operating entity, and reports an alarm; when only a single dimension score exceeds the corresponding threshold, it is judged as suspected anomaly, and traffic restriction control, log retention, and push for manual review are implemented without directly interrupting business; when both dimension scores are below the threshold, it is judged as a normal request and is directly allowed to flow.

5. The knowledge graph identification, governance, and parsing system for a large-scale food supply chain model according to claim 4, characterized in that, The intelligent risk control module for identifying anomalies includes an identification authenticity tolerance scoring model and a behavior anomaly normalization scoring model; the calculation formula for the identification authenticity tolerance scoring model is as follows: ； In the formula, To identify the confidence level of anomalies; This is a data source credibility factor used to weaken the correlation penalty for data registered in the trusted whitelist; This is an entity lifecycle maturity factor used to reduce the association verification weight for newly registered entities; These are the dimension weight coefficients; The calculation formula for the behavioral anomaly normalization scoring model is as follows: ； In the formula, Request frequency per unit time; To request the identification of dispersion; The frequency statistics are the mean and standard deviation; The mean and standard deviation are the statistical values ​​for the dispersion.

6. The knowledge graph identification, governance, and parsing system for a large-scale food supply chain model according to claim 1, characterized in that, The natural language parsing and interaction module is used to construct a mechanism for semantic ambiguity resolution and interactive information completion. The natural language parsing and interaction module performs intent recognition, keyword extraction, and semantic matching on the natural language queries related to food traceability input by the user. When the semantic matching degree is lower than a preset threshold, it automatically determines that the query statement has missing information or semantic ambiguity, and actively triggers interactive clarification, prompting the user to supplement missing fields such as time range, entity type, batch information, and business link. At the same time, it outputs the top K highly matched candidate parsing tasks for the user to select and confirm. After the semantic matching is successful, the natural language requirement is automatically converted into standardized identifier parsing instructions, and the knowledge graph query interface of the knowledge graph storage and graph computing module is called to aggregate the full-link related data of food supply chain entities and generate structured and visualized query results to be fed back to the user. The natural language parsing and interaction module includes a natural language semantic mapping matching model, and the calculation formula for the natural language semantic mapping matching model is as follows: ； In the formula, For semantic matching degree; Improve the accuracy of user intent recognition; For text semantic vector similarity; For core keyword matching degree; These are the weighting coefficients.

7. The knowledge graph identification, governance, and parsing system for a large-scale food supply chain model according to claim 1, characterized in that, The algorithm adaptive iterative tuning module is used to perform offline batch optimization, dynamic threshold update, incremental learning, and positive and negative sample feedback closed-loop processing. The algorithm's adaptive iterative optimization module updates the weight vectors of each model using a daily offline batch gradient descent iterative method, and configures an adaptive learning rate and a loss function specific to each model; it dynamically updates the judgment threshold based on the temporal distribution characteristics of food supply chain business data flow; it continuously collects new food entity types, new food supply chain business scenarios, new risk behaviors, and diversified query sentences to build an incremental sample pool, and completes model parameter updates through incremental fine-tuning; it collects misjudged, missed, and mismatched samples after manual review to form a supervised learning closed loop, and back-optimizes the model weights, semantic vector space, and risk feature library.

8. The knowledge graph identification, governance, and parsing system for a large-scale food supply chain model according to claim 7, characterized in that, The algorithm's adaptive iterative tuning module employs gradient descent weight adaptive iteration, and the iterative formula used for this gradient descent weight adaptive iteration is: ； In the formula, This is the current weight vector; Adaptive learning rate; The gradient of the loss function; The weights are updated iteratively; Coding adaptation model with The loss function is used in the knowledge association model. The loss function is used in the anomaly risk control model. The loss function is used; the semantic mapping model uses the cosine distance of semantic vectors. This is the loss function.

9. A knowledge graph identification, governance, and parsing system for a large-scale food supply chain model according to claim 7, characterized in that, The algorithm adaptive iterative tuning module is used to perform dynamic threshold adaptive updates, and the calculation formula used for the dynamic threshold adaptive update is: ； In the formula, For time series smoothing coefficients; This represents the mean of the recent normal sample data distribution. , The threshold values ​​are the judgment thresholds before and after the update. The algorithm adaptive iterative optimization module completes adaptive balancing through the calculation formula. During peak business hours, a lenient threshold is used to ensure business access, while during quiet periods, a strict threshold is used to ensure system security.

10. A knowledge graph identification governance and parsing method for a large-scale food supply chain model, characterized in that, include: S1: Multi-source heterogeneous food supply chain business data and user query data are accessed into the system described in any one of claims 1-9, and data cleaning and format normalization preprocessing are completed; S2: The preprocessed data is sent to the large model semantic processing module to complete semantic extraction, entity classification, encoding matching and output of standardized entity relation triples; S3: Write the standardized entity relationship triples into the knowledge graph storage and graph computation module to complete the structured storage and networking process; S4: Call the intelligent risk control module for abnormal identification to complete the detection of the authenticity of food supply chain identification and abnormal user behavior, and realize risk control through the integrated decision-making mechanism; S5: The user's natural language query is converted into a standardized identifier parsing instruction by the natural language parsing and interaction module through ambiguity resolution and semantic mapping. The knowledge graph query capability is then invoked to aggregate the full-link related data of the food supply chain entities and fed back to the user. S6: The entire process operation log is precipitated as iterative samples and sent to the algorithm adaptive iterative optimization module for offline algorithm optimization and incremental learning, so as to realize the continuous iterative upgrade of the system capabilities.

Images