A multi-source data fusion method for constructing a dynamic graph of cattle and sheep by-products industry

Abstract

The application discloses a kind of multi-source data fusion's cattle and sheep by-product industry dynamic graph construction method, comprising the following steps: constructing bioactivity-oriented whole industry chain dynamic ontology model, and obtaining multi-source heterogeneous data in parallel and carrying out pretreatment and cleaning to it;Cross-modal entity alignment and feature fusion are carried out to the data after cleaning, and initial knowledge graph is constructed;Based on biological activity attenuation coefficient, the entity and relationship edge in initial knowledge graph are updated with dynamic weight and ladder type pruning;In response to external risk events, topological reconstruction is carried out to knowledge graph, and compliance mark and relationship path are updated;Based on the value constraint of graph retrieval enhancement generation, the optimal utilization path is retrieved from the current knowledge graph, and natural language decision output is generated.The application can realize that graph weight is updated in real time and topological dynamic reconstruction, and generate optimal decision, with the advantages of timeliness dynamic representation, active safety risk control and value maximization decision.

Description

Technical Field

[0001] This invention relates to the fields of smart agriculture and big data technology, and in particular to a method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion. Background Technology

[0002] By-products from cattle and sheep slaughter (such as bones, blood, internal organs, glands, etc.) have extremely high development and utilization value and can be used in many high value-added fields such as biopharmaceuticals, food additives, feed, and cosmetics.

[0003] However, the development and utilization of by-products from cattle and sheep slaughter currently faces several technological bottlenecks: First, a data gap of "resources available, but no information." From upstream farming and midstream slaughtering and processing to downstream market applications, the entire industry chain suffers from severe data silos. Information is not shared among farmers, slaughterhouses, processing plants, research institutions, and the market, making it difficult for ordinary practitioners to easily access in-depth knowledge such as "the specific extraction process, equipment requirements, and market value of a certain by-product in the biopharmaceutical field." Second, existing knowledge representations lack a time-sensitive dimension. Traditional agricultural databases or static knowledge graphs cannot effectively depict the key characteristics of cattle and sheep by-products (especially glands and blood)—their bioactivity decays exponentially over time. The value recorded in the graph is fixed, unable to dynamically adjust the recommended processing paths and priorities based on the actual time the by-product has been removed from the body, leading to delayed decision-making and resource devaluation. Third, a lack of dynamic risk response mechanisms. Existing static information systems cannot respond to such events in real time. Once a risk occurs, they cannot automatically and quickly cut off the flow of contaminated or non-compliant byproducts to sensitive areas such as food and medicine, posing a huge risk to biosafety and compliance. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion. This invention enables real-time updating of map weights and dynamic topology reconstruction, and generates optimal decisions based on value constraints, offering advantages such as timely dynamic representation, proactive safety risk control, and value-maximizing decision-making.

[0005] The technical solution of this invention: A method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion, comprising the following steps: Step S1: Construct a dynamic ontology model of the entire industry chain for bioactivity. Based on the dynamic ontology model of the entire industry chain, acquire multi-source heterogeneous data in parallel and preprocess and clean it. Step S2: Perform cross-modal entity alignment and feature fusion on the cleaned data to construct an initial knowledge graph; Step S3: Based on the bioactivity decay coefficient, perform dynamic weight updates and step-by-step pruning on the entities and relation edges in the initial knowledge graph; Step S4: In response to external risk events, reconstruct the knowledge graph topology and update compliance tags and relationship paths; Step S5: Enhanced graph retrieval generation based on value constraints, retrieving the optimal utilization path from the current knowledge graph and generating natural language decision output.

[0006] The above-mentioned method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion, in step S1, includes constructing a dynamic ontology model of the entire industry chain oriented towards bioactivity, which includes: Define core entity classes including by-product entities, biochemical components, processing technology, downstream products, and compliance status; Add a bioactivity attenuation coefficient and a compliance mark to the properties of the byproduct entity.

[0007] In the aforementioned method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion, step S1 involves acquiring multi-source heterogeneous data in parallel, including: Structured data including by-product type, slaughter time, origin and quarantine status, acquired through NFC reader / writer device; Unstructured images and spectral data characterizing the appearance and internal composition of by-products were acquired using a multispectral vision device that integrates visible light and near-infrared sensors. Semi-structured market data and unstructured scientific research literature and regulatory text data are obtained from the Internet and authorized interfaces through the communication module.

[0008] The aforementioned method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion, in step S1, includes preprocessing and cleaning the parallel-acquired multi-source heterogeneous data, including: Perform time logic verification and outlier removal on structured data acquired via NFC; Denoising and illumination correction are performed on unstructured images and spectral data acquired through multispectral vision devices. The system analyzes, normalizes, and integrates semi-structured market data acquired through the communication module, and performs noise reduction, stop word filtering, and risk keyword pre-marking on unstructured scientific research literature and regulatory text data.

[0009] In the aforementioned method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion, step S2, the cross-modal entity alignment and feature fusion, includes: A pre-trained visual model is used to process image data from multi-source heterogeneous data to extract high-dimensional feature vectors characterizing the biochemical activity of by-products. Natural language processing models are used to perform named entity recognition on text data in multi-source heterogeneous data, and triple knowledge including ingredients, processes and uses is extracted. Using batch number and timestamp as anchors, the high-dimensional feature vector is aligned and fused with the triple knowledge to generate dynamic entity nodes containing real-time status information.

[0010] The aforementioned method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion, specifically the method for dynamic weight updating in step S3, is as follows: Periodically calculate the difference between the current time of the entity node and its slaughter time. ; Based on the difference between the current time of the entity node and its slaughter time Calculate and update the weights of relation edges pointing to entity nodes for high-value purposes. .

[0011] In the aforementioned method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion, step S3, the tiered pruning includes: When the updated weight Below the first activity threshold When this happens, the relation edge pointing to the entity node that is restricted by the first activity threshold is cut off; When the updated weight Below the second activity threshold When this happens, the relation edge pointing to the entity node that is subject to the second activity threshold is severed; And so on until the updated weights are applied. Below the nth activity threshold When, the relation edges pointing from the entity node to the use restricted by the nth activity threshold are cut off, wherein, , < < .

[0012] In the aforementioned method for constructing a dynamic graph of the cattle and sheep by-product industry based on multi-source data fusion, step S4, the topological reconstruction of the knowledge graph, includes: Collect and identify external events containing specific risk trigger words and associated region codes; Locate all by-product entity nodes in the knowledge graph that match the origin attribute with the risk area code; Reset the relational edge weights of the entity nodes pointing to the relevant industry sectors and their value applications to zero or delete them; Add a relation edge to the entity node pointing to the harmless treatment or industrial non-skin-friendly use.

[0013] In the aforementioned method for constructing a dynamic graph of the cattle and sheep by-product industry based on multi-source data fusion, step S5, the value-constrained graph retrieval enhancement generation, includes: Based on the user's query intent, retrieve all feasible utilization path subgraphs in the knowledge graph updated in steps S3 and S4; Calculate the expected net profit for each path by combining real-time market price data; We use a large language model to organize and visualize the path with the highest expected net profit and its calculation results in natural language.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. By introducing a bioactivity decay coefficient, this invention enables the knowledge graph to reflect the activity state of by-products in real time, avoiding resource devaluation and waste caused by information lag, and achieving accurate quantification of the value of by-products.

[0015] 2. This invention uses an event-driven graph topology reconstruction mechanism to automatically trigger a "circuit breaker" during an epidemic or regulatory change, cutting off the sensitive flow of risky products and ensuring that decision outputs comply with biosafety and food compliance requirements.

[0016] 3. This invention combines graph retrieval enhancement generation technology, economic value calculation, and large language model, which can directly guide farmers or enterprises to choose the most profitable and compliant processing method, thus reducing the professional knowledge threshold. Attached Figure Description

[0017] Figure 1 This is a hardware topology diagram of the operating environment for the method of this invention; Figure 2 This is the overall logic flowchart of the method of the present invention; Figure 3 This is a flowchart of the step-by-step automatic pruning method based on time decay in this invention. Figure 4 This is a schematic diagram of the event-driven risk blocking and graph reconstruction method of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0019] Example: A method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion. The hardware of this method adopts a three-layer architecture: acquisition layer, processing layer, and output layer. The hardware functions and selection of each layer must meet the requirements of multi-source heterogeneous data acquisition, edge computing, map visualization, and interaction. Figure 1 As shown, the specific configuration is as follows: The acquisition layer is the core of parallel acquisition of multi-source heterogeneous data, including an NFC reader / writer, a multispectral vision device integrating visible light and near-infrared sensors, a high-precision weighing module, and a 5G / wireless communication module. The NFC reader / writer needs to support the reading of information from livestock product traceability tags, the multispectral vision device needs to have the hardware foundation of Gaussian filtering and illumination correction, and the communication module needs to support API interface calls and web data crawling under the HTTPS protocol.

[0020] Processing layer: The core implementation includes data cleaning, feature fusion, graph construction and dynamic updates. It includes edge computing terminals with AI computing power modules (supporting GPU acceleration to meet the inference needs of visual models and NLP models), databases (Neo4j is recommended, which is adapted to the entity-relationship edge storage and retrieval of knowledge graphs), and structured / semi-structured data storage servers.

[0021] Output layer: The core implementation of graph visualization and natural language decision output includes industrial touch screens / interactive terminals, supporting graph topology display, user natural language query, and visualization of decision results.

[0022] like Figure 2 As shown, the method of the present invention specifically includes the following steps: Step S1: Construct a dynamic ontology model of the entire industry chain oriented towards bioactivity. Based on the dynamic ontology model of the entire industry chain, acquire multi-source heterogeneous data in parallel and preprocess and clean it.

[0023] This step is fundamental to graph construction. Its core tasks include defining the top-level framework of the industry logic and collecting and cleaning multi-source heterogeneous data to eliminate data silos and contaminated data, providing a high-quality data foundation for subsequent feature fusion. Specifically: 1.1 Constructing a dynamic ontology model of the entire industry chain oriented towards bioactivity The Web Ontology Language (OWL) is used to define the ontology framework for the entire industrial chain of cattle and sheep by-products. The core work includes defining core entity classes and expanding entity attributes. Specific operations are as follows: Define core entity classes: Identify five core entity classes, namely by-product entities (such as specific slaughter by-products such as sheep blood, yak spleen, cattle bones, glands, etc.), biochemical components (such as active ingredients contained in by-products such as thrombin, bilirubin, proteins, etc.), processing technology (such as ultrasonic crushing, enzymatic hydrolysis, extraction, drying, etc.), downstream products (such as biopharmaceutical raw materials, food additives, feed, cosmetic raw materials, etc.), and compliance status (normal, warning, high risk).

[0024] Extended byproduct entity attributes: In addition to traditional entity attributes, a bioactivity attenuation coefficient is added. The two core attributes are ( ) and compliance mark; among which The values ​​are assigned and stored based on industry standards or experimental data for different byproducts. The compliance flag is assigned the default value of "normal" and will be dynamically updated as risk events occur.

[0025] 1.2 Parallel acquisition of multi-source heterogeneous data Based on the ontology model built in 1.1, structured, semi-structured, and unstructured data are acquired in parallel through hardware acquisition layers. The acquisition methods and content of each type of data strictly match the entity class definitions of the ontology model. Specific operations are as follows: Structured data collection: The NFC reader reads the traceability tag information of cattle and sheep by-products to obtain structured data including by-product variety, slaughter time, place code, quarantine status and batch number. The data is stored in JSON format, and each data entry is stamped with a timestamp and place index.

[0026] Unstructured data acquisition: RGB images characterizing the appearance of by-products and near-infrared spectral data characterizing their intrinsic biochemical components and activities are acquired through multispectral vision devices; unstructured text data such as scientific research literature, regulatory texts, and epidemic notices in PDF / Word format are crawled from the Internet / authorized platforms through communication modules.

[0027] Semi-structured data acquisition: Market inventory and price data in XML format are obtained by calling the authorized supplier's API interface through the communication module. Market data with fixed HTML tags from industry websites and threshold tables of physicochemical components (such as veterinary drug residue limits) issued by relevant departments are crawled in a targeted manner as the basis for subsequent compliance verification and value calculation.

[0028] 1.3 Preprocessing and Cleaning of Multi-Source Heterogeneous Data Based on the characteristics of the three types of data, a differentiated cleaning method is adopted to remove outliers, noise, and invalid information, thereby achieving data standardization. Specific operations are as follows: Structured data cleaning: Perform time logic verification and outlier removal; verify whether the slaughter time is earlier than the current time and remove data with time logic errors; set constraint thresholds such as weight and batch number (e.g., individual weight > 0, batch number not empty) to remove meaningless dirty data, and retain the field-based valid data after cleaning.

[0029] Unstructured data cleaning includes: Image / spectral data: Gaussian filtering was used for noise reduction, adaptive histogram equalization (CLAHE) was used for illumination correction, and baseline correction was performed on near-infrared spectral data to eliminate environmental interference factors; Text data: Regular expressions are used to remove HTML noise tags, garbled characters, and pop-up ads. A livestock-specific stop word list is loaded to filter meaningless function words (such as "of" and "about"). Risk keywords (such as "epidemic", "embargo", and "heavy metal exceedance") in epidemic notices and regulatory texts are pre-marked and stored.

[0030] Semi-structured data cleaning: ETL tools are used to parse XML / HTML format data, remove tags and extract core key-value pairs; unit normalization is performed, converting market prices (yuan / ton, yuan / kg) and physicochemical indicators (mg / kg, PPM) from different sources into preset standard units, and data integration is completed.

[0031] Step S2: Perform cross-modal entity alignment and feature fusion on the cleaned data to construct an initial knowledge graph.

[0032] This step, based on the high-quality multi-source data cleaned in step S1, maps the multi-source heterogeneous data into dynamic entity nodes and relation edges of the graph through visual feature extraction, text entity extraction, and cross-modal fusion. Finally, an initial knowledge graph is constructed in the graph database. Specific operations are as follows: 2.1 Visual Feature Vectorization: Input the by-product images / spectral data cleaned in step S1 into the pre-trained Vision Transformer (ViT, a model that applies the Transformer architecture to computer vision tasks) visual model, use the AI ​​computing power module of the processing layer to perform model inference, extract high-dimensional feature vectors that characterize the biochemical activity and freshness of the by-products, the vector dimension is consistent with the model preset, and after extraction, they are bound and stored with the corresponding batch number.

[0033] 2.2 Text Semantic Entity Extraction: Input the cleaned scientific literature, process documents, and regulatory texts from step S1 into the pre-trained BERT natural language processing model to perform named entity recognition (NER) and extract the triple knowledge of "component-process-use" (such as "sheep blood-extraction-thrombin" and "bovine bone-enzymatic hydrolysis-food additive"). The triple knowledge must strictly match the core entity class defined in step S1 and be bound and stored with the corresponding batch number after extraction.

[0034] 2.3 Cross-modal entity alignment and fusion: Using batch number and timestamp as unique anchors, the high-dimensional feature vector of step S2.1 is associated and matched with the triple knowledge of step S2.2 to generate dynamic entity nodes containing real-time status information for each by-product batch; the entity nodes must contain basic information of by-products, visual features, text triples, bioactivity decay coefficient, and full information of compliance markers.

[0035] 2.4 Constructing the Initial Knowledge Graph: In the graph database (Neo4j), dynamic entity nodes are used as graph nodes, and "contains (byproducts - biochemical components)", "uses (byproducts - processing technology)", and "points to (byproducts - downstream products)" are used as graph relation edges; initial weights are assigned to all relation edges. (Based on the priority assignment of high-value uses according to the by-product theory), the compliance mark is set to "normal" by default, and the initial knowledge graph is constructed and stored.

[0036] Step S3: Based on the bioactivity decay coefficient, perform dynamic weight updates and step-by-step pruning on the entities and relation edges in the initial knowledge graph.

[0037] This step addresses the exponential decay of bioactivity in bovine and ovine by-products over time by dynamically updating the knowledge graph. Through weight adjustments and step-by-step pruning, it reflects the actual utilization value of by-products in real time, avoiding resource waste caused by information lag. Specific operations are as follows: 3.1 Decay Time Calculation: Set up a periodic scanning mechanism for the system (recommended scan cycle is 10 minutes, which can be adjusted according to actual needs). By periodically scanning all dynamic entity nodes in the graph database through the processing layer computer, the time difference between the current time and the slaughter time is calculated. Time difference The time of byproducts outside the body is used as a core parameter for calculating activity decay.

[0038] 3.2 Dynamic weight update of relation edges: Based on the first-order dynamic model, using the bioactivity decay coefficient ( ) and time difference Update the current weight of the relation edge pointing to the entity node for high-value purposes. The calculation formula is: ,in It is a natural constant. The initial weights set for step S2, The system assigns a bioactivity attenuation coefficient to the by-product entity (the specific threshold and coefficient can be dynamically configured by the system based on the by-product type, target product standard, and real-time industry data; this embodiment does not impose limitations on the specific values); the system automatically completes the formula calculation and transfers the relation edge weights from the graph database. Updated to This enables real-time dynamic adjustment of weights.

[0039] 3.3, Step-by-step pruning of relational edges: based on a preset first activity threshold ( ) and second activity threshold ( () < The updated weighted relation edges are automatically pruned in a step-by-step manner, with the pruning logic strictly following the value hierarchy, such as... Figure 3 As shown, the specific steps are as follows: when < When the byproduct loses its value for use restricted by the first activity threshold (e.g., value for biomedical extraction), the relation edge pointing from the entity node to the use restricted by the first activity threshold is severed (the weight of the relation edge is reset to 0 and marked as "invalid" in the graph database). when < When the byproduct loses its value for use restricted by the second activity threshold (e.g., food / feed processing value), the relation edge pointing from the entity node to the use restricted by the second activity threshold is cut off (the weight of the relation edge is reset to 0 and marked as "invalid" in the graph database). And so on until the updated weights are applied. Below the nth activity threshold When this happens, the entity node pointing to the use restricted by the nth activity threshold is disconnected, and the relation edge pointing to the entity node pointing to the use restricted by the nth activity threshold is disconnected (the weight of this relation edge is reset to 0, and it is marked as "invalid" in the graph database). , < < ; After pruning, the map retains only the edges related to the compliant high-value applications of by-products under their current activity, thus achieving dynamic optimization of the map structure.

[0040] Step S4: In response to external risk events, reconstruct the knowledge graph topology and update compliance tags and relationship paths.

[0041] This step constructs an event-driven risk circuit breaker mechanism to achieve real-time response to external sudden risk events (such as animal epidemics, regulatory changes, and origin contamination). Through graph topology reconstruction, it cuts off risk flows, adds compliance paths, and ensures the compliance of the graph. Figure 4 As shown, the specific steps are as follows: 4.1 External Risk Event Collection and Identification: The system collects external information released by the Internet, the Ministry of Agriculture and Rural Affairs, and the Center for Disease Control and Prevention in real time through the communication module. It deploys risk keyword and regional coding identification algorithms to automatically identify external risk events containing risk trigger words (such as "epidemic", "embargo", "veterinary drug residue exceeding the standard") and associated risk regional codes. After identification, the system automatically triggers the map topology reconstruction instruction.

[0042] 4.2 Risk Entity Node Location: Based on the identified risk area codes, the system performs precise searches in the knowledge graph of the graph database to locate all dynamic entity nodes of by-products whose origin attributes match the risk area codes. The system then "marks" these nodes with risks to achieve rapid identification of risk entities.

[0043] 4.3 Graph Topology Reconstruction: For locked risky entity nodes, dynamically adjust the relation edges to complete the graph topology reconstruction. Specific operations are as follows: Cut off risky relationship edges: Set the weight of the relationship edges pointing to sensitive areas such as food processing, pharmaceutical extraction, and cosmetic raw materials of risky entities to zero or delete them directly, so as to completely cut off the sensitive flow of risky by-products; Add a new compliance relationship edge: Add a relationship edge to the risk entity node pointing to harmless treatment (high temperature incineration, deep burial) or industrial non-skin-friendly use, and assign reasonable weights to the new relationship edge to realize the compliance diversion of risk by-products; 4.4 Update compliance flags: Update the compliance flags of risk entity nodes in the graph database from "normal" to "warning" or "high risk", and synchronously store risk event information (such as risk type, region, trigger time) to achieve full recording of compliance information.

[0044] Step S5: Enhanced graph retrieval generation based on value constraints, retrieving the optimal utilization path from the current knowledge graph and generating natural language decision output.

[0045] This step is the final application of the knowledge graph. Based on the effective knowledge graph after dynamic updates in step S3 and risk reconstruction in step S4, and combined with GraphRAG (Graph-based Retrieval-Augmented Generation, a retrieval enhancement and generation technique combining knowledge graphs and large language models) technology and the large language model, it responds to user query intent, retrieves the optimal utilization path, and generates natural language decision output, realizing the transformation from "graph data" to "decision suggestions." Specific operations are as follows: 5.1 User Query Intent Parsing and Subgraph Retrieval: Users input natural language queries (such as "How to process this batch of sheep blood to maximize profits?" or "Feasible processing paths for beef spleen from a certain production area") through the interactive terminal of the output layer. The system parses the query intent and retrieves all feasible utilization path subgraphs in the graph database. Subgraph retrieval must strictly filter invalid relation edges with a weight of 0 and risk entities marked as "warning / high risk" for compliance, retaining only compliant, effective, and high-value utilization paths under the current activity.

[0046] 5.2 Calculation of Expected Net Profit for Utilization Paths: Combining the real-time market price data collected and cleaned in step S1, the expected net profit is calculated for each feasible utilization path. The calculation logic is as follows: Expected net profit = final downstream product market price - processing cost - logistics cost - raw material cost.

[0047] Among them, the processing cost and logistics cost are preset based on industry standards or actual enterprise data. The system automatically completes the profit calculation and sorting of all paths and selects the optimal utilization path with the highest expected net profit.

[0048] Natural Language Decision Generation and Output: The optimal utilization path, profit calculation results, and real-time status of byproducts (such as in vitro time, bioactivity, and compliance status) are input into a pre-trained large language model (such as GPT, Wenxin Yiyan, etc.). The model organizes the information into natural language to generate easily understandable and actionable decision suggestions. Simultaneously, the output layer terminal visualizes the knowledge graph topology, optimal path, and profit data. An example of a decision suggestion is: "Given that this batch of sheep blood has been in vitro for more than 4 hours, the bioactivity weight..." < Furthermore, since there is no risk of disease outbreak in the production area, it is recommended to abandon the extraction of thrombin for biopharmaceutical purposes and switch to the production of blood meal protein feed, with an estimated net profit of XX yuan per ton. In summary, this invention enables the integration and fusion of structured, semi-structured, and unstructured data across the entire industry chain, providing comprehensive data support for industrial decision-making. Through a bioactivity decay mechanism, this invention reflects the actual utilization value of by-products in real time, avoiding insufficient utilization of high-value resources due to information lag. Through an event-driven topology reconstruction mechanism, this invention achieves automatic risk identification and mitigation, ensuring that by-product utilization paths comply with biosafety and food safety requirements. The natural language decision output of this invention directly guides farmers / enterprises to choose the most profitable and compliant processing method, requiring no specialized graph analysis knowledge, thus enhancing the practicality and accessibility of the solution.

[0049] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion, characterized in that, Includes the following steps: Step S1: Construct a dynamic ontology model of the entire industry chain for bioactivity. Based on the dynamic ontology model of the entire industry chain, acquire multi-source heterogeneous data in parallel and preprocess and clean it. Step S2: Perform cross-modal entity alignment and feature fusion on the cleaned data to construct an initial knowledge graph; Step S3: Based on the bioactivity decay coefficient, perform dynamic weight updates and step-by-step pruning on the entities and relation edges in the initial knowledge graph; Step S4: In response to external risk events, reconstruct the knowledge graph topology and update compliance tags and relationship paths; Step S5: Enhanced graph retrieval generation based on value constraints, retrieving the optimal utilization path from the current knowledge graph and generating natural language decision output.

2. The method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion according to claim 1, characterized in that, In step S1, constructing a dynamic ontology model for the entire bioactivity supply chain includes: Define core entity classes including by-product entities, biochemical components, processing technology, downstream products, and compliance status; Add a bioactivity attenuation coefficient and a compliance mark to the properties of the byproduct entity.

3. The method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion according to claim 1, characterized in that, In step S1, the multi-source heterogeneous data acquired in parallel includes: Structured data including by-product type, slaughter time, origin and quarantine status, acquired through NFC reader / writer device; Unstructured images and spectral data characterizing the appearance and internal composition of by-products were acquired using a multispectral vision device that integrates visible light and near-infrared sensors. Semi-structured market data and unstructured scientific research literature and regulatory text data are obtained from the Internet and authorized interfaces through the communication module.

4. The method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion according to claim 3, characterized in that, Step S1 involves preprocessing and cleaning the multi-source heterogeneous data acquired in parallel, including: Perform time logic verification and outlier removal on structured data acquired via NFC; Denoising and illumination correction are performed on unstructured images and spectral data acquired through multispectral vision devices. The system analyzes, normalizes, and integrates semi-structured market data acquired through the communication module, and performs noise reduction, stop word filtering, and risk keyword pre-marking on unstructured scientific research literature and regulatory text data.

5. The method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion according to claim 1, characterized in that, In step S2, the cross-modal entity alignment and feature fusion includes: A pre-trained visual model is used to process image data from multi-source heterogeneous data to extract high-dimensional feature vectors characterizing the biochemical activity of by-products. Natural language processing models are used to perform named entity recognition on text data in multi-source heterogeneous data, and triple knowledge including ingredients, processes and uses is extracted. Using batch number and timestamp as anchors, the high-dimensional feature vector is aligned and fused with the triple knowledge to generate dynamic entity nodes containing real-time status information.

6. The method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion according to claim 1, characterized in that, In step S3, the specific method for dynamic weight update is as follows: Periodically calculate the difference between the current time of the entity node and its slaughter time. ; Based on the difference between the current time of the entity node and its slaughter time Calculate and update the weights of relation edges pointing to entity nodes for high-value purposes. .

7. The method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion according to claim 5, characterized in that, In step S3, the stepped pruning includes: When the updated weight Below the first activity threshold When this happens, the relation edge pointing to the entity node that is restricted by the first activity threshold is cut off; When the updated weight Below the second activity threshold When this happens, the relation edge pointing to the entity node that is subject to the second activity threshold is severed; And so on until the updated weights are applied. Below the nth activity threshold When, the relation edges pointing from the entity node to the use restricted by the nth activity threshold are cut off, wherein, , < < .

8. The method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion according to claim 1, characterized in that, In step S4, the topological reconstruction of the knowledge graph includes: Collect and identify external events containing specific risk trigger words and associated region codes; Locate all by-product entity nodes in the knowledge graph that match the origin attribute with the risk area code; Reset the relational edge weights of the entity nodes pointing to the relevant industry sectors and their value applications to zero or delete them; Add a relation edge to the entity node pointing to the harmless treatment or industrial non-skin-friendly use.

9. The method for constructing a dynamic map of the cattle and sheep by-product industry based on multi-source data fusion according to claim 1, characterized in that, In step S5, the value-constrained graph retrieval enhancement generation includes: Based on the user's query intent, retrieve all feasible utilization path subgraphs in the knowledge graph updated in steps S3 and S4; Calculate the expected net profit for each path by combining real-time market price data; We use a large language model to organize and visualize the path with the highest expected net profit and its calculation results in natural language.

Images