A large model knowledge graph-based penaeus vannamei intelligent feeding decision method

Abstract

The application discloses a kind of based on big model knowledge graph's Penaeus vannamei intelligent feeding decision method, belong to aquaculture technical field. Including: obtain Penaeus vannamei breeding document, feeding record, water quality record, feed instruction, production procedure and expert experience data, based on big language model constructs Penaeus vannamei intelligent feeding field knowledge graph;Intelligent feeding decision model is constructed based on Penaeus vannamei intelligent feeding field knowledge graph fusion big language model;Real-time acquisition breeding factor related data and user question or automatic feeding task, and after preprocessing input the intelligent feeding decision model, generate structured feeding plan;Structured feeding plan is converted into feeding machine control instruction, and feeding machine is executed and issued. The present application constructs "sensor perception-knowledge graph-big model decision-feeding machine execution" closed loop architecture, solves the problem that traditional feeding relies on artificial experience, data utilization is insufficient and execution is not continuous.

Description

Technical Field

[0001] This invention relates to the field of aquaculture technology, and in particular to an intelligent feeding decision-making method for Litopenaeus vannamei based on a large model knowledge graph. Background Technology

[0002] In the farming of Litopenaeus vannamei, feeding management directly affects its growth rate, survival rate, feed conversion ratio, farming cycle, and economic benefits. In existing farming scenarios, farmers typically make manual judgments on the total daily feed amount, feeding frequency, and feeding time based on water temperature, dissolved oxygen, pH, weather, farming stage, uneaten feed levels, and experience observations.

[0003] However, existing technologies have the following drawbacks: (1) The input dimensions are limited, usually covering only a small number of basic environmental parameters, making it difficult to integrate multiple information such as salinity, ammonia nitrogen, nitrite, body length, weight and residual bait level.

[0004] (2) The fixed rule judgment method is difficult to adapt to the real breeding scenario with multiple factors coupled and dynamic changes.

[0005] (3) Knowledge bases and expert experience are difficult to be uniformly called upon as a basis for real-time decision-making.

[0006] (4) The output results mostly remain at the level of text suggestions, and it is difficult to directly form structured instructions that can be executed by the automatic feeder.

[0007] (5) The lack of a layered design between the front-end acquisition interface and the back-end execution interface is not conducive to the access of real devices.

[0008] (6) Lack of a complete closed loop from perception, reasoning, reporting, analysis to execution. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention constructs a closed-loop architecture of "sensor perception - knowledge graph - large model decision-making - feeding machine execution," enabling water quality data, aquaculture status, feeding knowledge, and equipment control to work collaboratively in the same link. This solves the problems of traditional feeding relying on human experience, insufficient data utilization, and discontinuous execution.

[0010] To achieve the above objectives, this invention provides an intelligent feeding decision-making method for Litopenaeus vannamei based on a large-scale model knowledge graph, comprising the following steps: (1) Obtain documents on Litopenaeus vannamei farming, feeding records, water quality records, feed instructions, production procedures and expert experience data, and construct a knowledge graph of intelligent feeding of Litopenaeus vannamei based on a large language model; (2) Construct an intelligent feeding decision model based on the knowledge graph of the field of intelligent feeding of Litopenaeus vannamei and the fusion of a large language model; collect relevant data of aquaculture factors and user questions or automatic feeding tasks in real time, and input them into the intelligent feeding decision model after preprocessing to generate a structured feeding plan; (3) Convert the structured feeding plan into feeding machine control instructions and send them to the feeding machine for execution.

[0011] Furthermore, step (1) specifically includes: (1.1) Preprocess the unstructured text and embedded tables in the documents on Litopenaeus vannamei farming, feeding records, water quality records, feed instructions, production procedures and expert experience data, and convert them into natural language descriptions; (1.2) Use a large language model to extract entities that include at least environmental factors, breeding stage, feed type, feeding strategy, risk factors, detection indicators, control measures, equipment, nutritional indicators and disease symptoms; (1.3) Constructing knowledge triples: in: For a collection of entities, For a set of entity relations, A collection of attributes; (1.4) Write the knowledge triples into a graph database or local graph structure storage, and create text block vector, entity vector and relation vector indexes; (1.5) Perform entity alias merging, relation deduplication and quality verification on the knowledge graph to obtain a standardized knowledge graph for the intelligent feeding of Litopenaeus vannamei.

[0012] Furthermore, the data related to the aquaculture factors include one or more of the following: water temperature, dissolved oxygen, pH value, salinity, ammonia nitrogen, nitrite, aquaculture stage, stocking density, body length, body weight, and residual feed level. The data related to aquaculture factors are obtained through any one of the following methods: real-time push via WebSocket, reporting via REST API, collection via MQTT gateway, or manual input via the front end.

[0013] Furthermore, step (2) specifically involves: (2.1) Obtain real-time data on aquaculture factors and convert it into a context that can be understood by a large language model; The key-value pairs of each field of the aquaculture factor-related data are converted into natural language descriptive text according to the preset Chinese fishery terminology label mapping rules to form an aquaculture environment context; (2.2) Receive user questions or automatic feeding task trigger signals, and perform query intent identification and classification; Initial intent determination is performed using keyword heuristic matching: if the query contains explicit emergency keywords, it is directly determined to be an emergency response query; if the query contains feeding task keywords, it is determined to be a feeding suggestion query; for fuzzy queries containing sensor indicator names and abnormal semantic combinations, a large language model is further invoked in conjunction with the knowledge graph context of the intelligent feeding domain of Litopenaeus vannamei to enhance risk determination and output classification results; if the intent is emergency response, proceed to step (2.3); if the intent is feeding suggestion, proceed to step (2.4). (2.3) For emergency response type queries, the relevant context of emergency handling is retrieved directly from the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei in hybrid mode, the large language model is called to generate an emergency plan report, the feeding plan generation process is not executed, and the processing ends; (2.4) For feeding suggestion type queries, the context that the large language model can understand in step (2.1) and the intent category identified in step (2.2) are concatenated into an enhanced query statement according to a preset template, and then proceed to step (2.5). (2.5) For the enhanced query statement, the large language model is called to extract high-level keywords and low-level keywords, and the two sets of keywords are rewritten to obtain the rewritten high-level query statement and low-level query statement. Based on the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei, parallel retrieval of four retrieval modes (Naïve, Local, Global, and Hybrid) is performed simultaneously through LightRAG to obtain multi-mode retrieval results. On this basis, special retrieval is performed on six dimensions: weather conditions, breeding stage, breeding density, nutritional principles, risk factors, and practical experience. Based on the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei, multi-dimensional special query results are generated. (2.6) Calculate the environmental correction factor based on water temperature, dissolved oxygen, pH value, ammonia nitrogen, nitrite and residual feed level; in: Water temperature correction factor, For dissolved oxygen correction factor, pH correction factor For ammonia nitrogen correction factor, For nitrite correction factor, This is the correction factor for uneaten bait; If any of the following factors—water temperature, dissolved oxygen, pH, ammonia nitrogen, nitrite, and residual feed level—is not present in the aquaculture factor data, then the corresponding correction factor in the above formula is 1. (2.7) Determine the basic feeding coefficient based on the breeding stage and calculate the total daily feeding amount; in: This refers to the total daily feeding amount; Basic feeding coefficient; This refers to the remaining number of fish in the pond; Average weight; If the average weight cannot be obtained, the feeding plan generation process will not be executed, and the process will end. (2.8) Distribute the total daily feed amount to multiple feeding periods according to a preset ratio to obtain: in: For the first Feeding amount per feeding Allocate proportions according to the corresponding time periods; (2.9) Perform comprehensive reasoning and structured output using a large language model; The multi-modal retrieval results and multi-dimensional specialized query results from step (2.5), along with the calculation results from steps (2.7) and (2.8), are merged into an enhanced knowledge context. The large language model is then invoked to perform the following steps: (2.9.1) Quality inspection and consistency verification: Based on the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei, cross-validation is performed to enhance the consistency between the knowledge context and the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei, and contradictory conclusions are marked and confidence assessments are given. (2.9.2) Comprehensive synthesis of feeding scheme: The total daily feeding amount and the amount of food given each time By integrating with the knowledge graph in the field of intelligent feeding of Litopenaeus vannamei, which provides feeding strategy suggestions, environmental risk analysis, comparison of practical experience, and explanation of nutritional principles, a decision report is generated that includes the feeding amount calculation process, key judgment criteria, feed type suggestions, pellet size suggestions, and precautions. (2.9.3) Construction of the traceability appendix: Automatically add traceability and verification chapters with the structure of input source, parameter source, calculation link, evidence extract and executability statement to ensure that the decision-making process is traceable; The final output includes a decision report and a structured feeding plan; the structured feeding plan includes: total daily feeding amount, feed type, protein content, and feeding records, wherein each feeding record includes the feeding time period and feeding amount.

[0014] Furthermore, if the collected data on aquaculture factors includes weight information, then the average weight is the given weight information. When the collected data on aquaculture factors does not contain weight information but does contain body length information, the weight value is calculated based on the body length as a basic reference value and determined comprehensively based on the knowledge graph of the intelligent feeding field of Litopenaeus vannamei. If there is no reference information on body weight and length in the collected data on breeding factors, the system will determine whether the relevant body weight can be obtained based on the knowledge graph. If it cannot be obtained, the system will inform the user truthfully in the report and provide relevant breeding information for reference, but will not feed the animal.

[0015] Furthermore, the feeding task keywords include at least one of "oxygen deficiency", "rainstorm", and "typhoon"; the feeding task keywords include at least one of "feeding" and "feed"; and the abnormal semantics include at least one of "high", "exceeding standard", and "deterioration".

[0016] Furthermore, step (3) specifically involves: (3.1) Design the interface for configuring the feeder, testing connectivity, and querying status; (3.2) Convert the structured feeding plan into feeder control instructions, which include the total daily feeding amount, feed type, feeding frequency, feeding time, and feeding amount per feeding; (3.3) Issue control commands according to the operating mode and record the decision history and equipment execution results.

[0017] This invention also provides an intelligent feeding decision-making system for Litopenaeus vannamei based on a large model knowledge graph, comprising: The knowledge building and data management layer is used to acquire documents on Litopenaeus vannamei farming, feeding records, water quality records, feed instructions, production procedures and expert experience texts, and to build a knowledge graph for the intelligent feeding of Litopenaeus vannamei based on a large language model; The data acquisition layer is used to collect real-time data on the aquaculture environment, user questions, or automatic feeding tasks, and to perform preprocessing. The collaborative decision-making computing layer is used to construct an intelligent feeding decision model based on a knowledge graph fusion with a large language model in the field of intelligent feeding of Litopenaeus vannamei; the pre-processed aquaculture environment data and user questions or automatic feeding tasks are input into the intelligent feeding decision model to generate a structured feeding plan; The cloud-edge physical control layer is used to receive aquaculture environment data through sensor interfaces, convert structured feeding plans into executable instructions for automatic feeders, send them to the automatic feeders, and receive feedback from the equipment execution.

[0018] Furthermore, the cloud-edge physical control layer specifically includes: (1) Design a sensor data access interface to receive data uploaded by water quality sensors, aquaculture gateways, edge devices or front-end simulation devices; (2) Perform field mapping, device number binding, timestamp recording, and integrity verification on sensor data; (3) Design the interface for the feeder configuration, connectivity test, and status query; (4) Convert the structured feeding plan into feeder control instructions, which include the total daily feeding amount, feed type, feeding frequency, feeding time, and feeding amount per feeding. (5) Issue feeding instructions according to the operating mode, and record sensor history, decision history and equipment execution results.

[0019] The beneficial effects of this invention are: 1. A closed-loop architecture of "sensor perception - knowledge graph - big model decision-making - feeder execution" was constructed for the feeding scenario of Litopenaeus vannamei. This architecture enables water quality data, aquaculture status, feeding knowledge and equipment control to work collaboratively in the same link, solving the problems of traditional feeding relying on human experience, insufficient data utilization and discontinuous execution.

[0020] 2. Based on a large language model, a knowledge graph for intelligent feeding of Litopenaeus vannamei is constructed, which structurally links the farming stage, environmental factors, feed type, feeding strategy and risk control measures, so that feeding decisions have clear knowledge basis and reduce the illusion risk when general large models directly generate solutions.

[0021] 3. A feeding decision-making mechanism integrating knowledge graph retrieval results and real-time sensor data is proposed. The daily feed amount is calculated based on the stock volume, average body weight, basal feeding coefficient, and environmental correction coefficient. The environmental correction coefficient comprehensively considers water temperature, dissolved oxygen, pH, ammonia nitrogen, nitrite, and residual feed level, enabling the feed amount to be dynamically adjusted according to the aquaculture environment, reducing the accumulation of residual feed and water quality deterioration caused by overfeeding. 4. An IoT interface for sensors and feeders was designed to directly convert the structured feeding plan generated by the intelligent decision-making model into feeder control commands, realizing automatic connection from generation and confirmation to execution of the feeding plan, and improving the real-time performance, accuracy and engineering feasibility of feeding control. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the intelligent feeding decision-making method for Litopenaeus vannamei based on a large model knowledge graph, according to an embodiment of the present invention.

[0023] Figure 2 This is an example of the knowledge graph portion of an embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of the large model enhancement and inference process based on the graph in an embodiment of the present invention. Detailed Implementation

[0025] The present invention will be further explained and described below with reference to the accompanying drawings and embodiments.

[0026] This invention provides an intelligent feeding decision-making method for Litopenaeus vannamei based on a large-scale knowledge graph model. Model training and testing in this embodiment were conducted on a Windows 10 64-bit system. The experimental computer consisted of an Intel(R) Core(TM) i9-14900K 3.20 GHz CPU, an NVIDIA GeForce RTX 4090D GPU, and 64GB of memory. The software environment included Python 3.12, CUDA 12.4, and the Neo4j graph database.

[0027] For example Figure 1 As shown, the method specifically includes the following steps: S101. Obtain documents on Litopenaeus vannamei farming, feeding records, water quality records, feed instructions, production procedures, and expert experience data, and construct a knowledge graph for intelligent feeding of Litopenaeus vannamei based on a large language model.

[0028] (1) Preprocess the unstructured text and embedded tables in the documents on Litopenaeus vannamei farming, feeding records, water quality records, feed instructions, production procedures and expert experience data, and convert them into natural language descriptions.

[0029] (2) Use a large language model to extract entities including at least environmental factors, breeding stage, feed type, feeding strategy, risk factors, detection indicators, control measures, equipment, nutritional indicators and disease symptoms.

[0030] like Figure 2 As shown.

[0031] (3) Constructing knowledge triples: Among them: For a collection of entities, For a set of entity relations, It is a collection of attributes.

[0032] (4) Write the knowledge triples into the graph database or local graph structure storage, and establish text block vector, entity vector and relation vector indexes.

[0033] (5) Merge entity aliases, remove duplicate relationships and perform quality checks on the knowledge graph to obtain a standardized knowledge graph for the intelligent feeding of Litopenaeus vannamei.

[0034] S102. Construct an intelligent feeding decision model based on a knowledge graph fusion with a large language model in the field of intelligent feeding of Litopenaeus vannamei; collect aquaculture environment data and user questions or automatic feeding tasks in real time, and input them into the intelligent feeding decision model after preprocessing to generate a structured feeding plan.

[0035] Aquaculture environmental data includes one or more of the following: water temperature, dissolved oxygen, pH, salinity, ammonia nitrogen, nitrite, rearing stage, stocking density, body length, weight, and uneaten feed level. Data related to aquaculture factors is obtained through any of the following methods: real-time push via WebSocket, reporting via REST API, collection via MQTT gateway, or manual input at the front end.

[0036] like Figure 3 As shown, the specific process is as follows: (1) Obtain real-time data on aquaculture factors and convert it into a context that can be understood by a large language model.

[0037] The key-value pairs of each field in the aquaculture factor-related data are converted into natural language descriptive text according to the preset Chinese fisheries terminology label mapping rules, forming an aquaculture environment context.

[0038] (2) Receive user questions or automatic feeding task trigger signals, and perform query intent identification and classification.

[0039] Initial intent judgment is made through keyword heuristic matching: if the query contains explicit emergency keywords (including "feeding", "feed", etc.), it is directly judged as an emergency response query; if the query contains feeding task keywords, it is judged as a feeding suggestion query; for fuzzy queries containing sensor indicator names and abnormal semantic combinations (such as "too high", "exceeding the standard", "deteriorating"), the large language model is further called to combine the knowledge graph context of the intelligent feeding domain of Litopenaeus vannamei to enhance risk judgment and output the classification result; if the intent is an emergency response, proceed to step (3); if the intent is a feeding suggestion, proceed to step (4).

[0040] The feeding task keywords must include at least one of "oxygen deficiency", "rainstorm", and "typhoon"; the feeding task keywords must include at least one of "feeding" and "feed"; the abnormal semantics must include at least one of "too high", "exceeding the standard", and "deterioration".

[0041] (3) For emergency response queries, the relevant context of emergency handling is retrieved directly from the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei in hybrid mode, and the emergency plan report is generated by calling the large language model. The feeding plan generation process is not executed, and the processing ends.

[0042] (4) For feeding suggestion type queries, the context that the large language model can understand in step (1) and the intent category identified in step (2) are concatenated into an enhanced query statement according to a preset template, and then proceed to step (5).

[0043] (5) For the enhanced query statement, the large language model is called to extract high-level keywords and low-level keywords, and the two sets of keywords are rewritten to obtain the rewritten high-level query statement and low-level query statement. Based on the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei, parallel retrieval of four retrieval modes (Naïve, Local, Global, and Hybrid) is performed simultaneously through LightRAG to obtain multi-mode retrieval results. On this basis, special retrieval is performed on six dimensions: weather conditions, breeding stage, breeding density, nutritional principles, risk factors and practical experience. Based on the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei, multi-dimensional special query results are generated.

[0044] (6) Calculate the environmental correction factor based on water temperature, dissolved oxygen, pH value, ammonia nitrogen, nitrite and residual feed level; in: Water temperature correction factor, For dissolved oxygen correction factor, pH correction factor For ammonia nitrogen correction factor, For nitrite correction factor, This is the correction factor for residual bait.

[0045] The correction factor is determined based on the input values ​​for water temperature, dissolved oxygen, pH, ammonia nitrogen, nitrite, and uneaten feed level. If any of these factors is not present in the aquaculture factor data, the corresponding correction factor in the above formula is 1.

[0046] (7) Determine the basic feeding coefficient according to the breeding stage and calculate the total daily feeding amount.

[0047] in: This refers to the total daily feeding amount; Basic feeding coefficient; This refers to the remaining number of fish in the pond; This represents average body weight.

[0048] When the collected data on aquaculture factors includes weight information, the average weight is the weight information provided.

[0049] When the collected data on aquaculture factors does not contain weight information but does contain body length information, the weight value is calculated based on the body length as a basic reference value and then comprehensively determined based on the knowledge graph of intelligent feeding of Litopenaeus vannamei.

[0050] Where: W represents weight, L represents body length, and the coefficients 0.0106 and 3.066 are relatively accurate median values ​​obtained by querying data on the correlation between weight and body length of Litopenaeus vannamei farmed shrimp.

[0051] If there is no reference information on body weight and length in the collected data on breeding factors, the system will determine whether the relevant body weight can be obtained based on the knowledge graph. If it cannot be obtained, the system will inform the user truthfully in the report and provide relevant breeding information for reference, but will not feed the user and will end the process.

[0052] (8) Distribute the total daily feed amount to multiple feeding periods according to a preset ratio to obtain: in: For the first Feeding amount per feeding The proportion is allocated to the corresponding time period.

[0053] (9) Perform comprehensive reasoning and structured output of large language model.

[0054] The multi-modal retrieval results, multi-dimensional special query results, and calculation results from steps (7) and (8) are merged into an enhanced knowledge context, and the large language model is called to perform the following steps: (9.1) Quality inspection and consistency verification: Based on the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei, cross-validation is used to enhance the consistency between the knowledge context and the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei, and contradictory conclusions are marked and confidence assessments are given.

[0055] (9.2) Comprehensive synthesis of feeding program: The total daily feeding amount and the amount of food given each time By integrating with the knowledge graph in the field of intelligent feeding of Litopenaeus vannamei, which provides feeding strategy suggestions, environmental risk analysis, comparison of practical experience, and explanation of nutritional principles, a decision report is generated that includes the feeding amount calculation process, key judgment criteria, feed type suggestions, pellet size suggestions, and precautions.

[0056] (9.3) Construction of traceability appendix: Automatically add traceability and verification chapters with the structure of input source, parameter source, calculation link, evidence extract and executability statement to ensure that the decision-making process is traceable.

[0057] The final output includes a decision report and a structured feeding plan; the structured feeding plan includes: total daily amount of feed (total_daily_amount_grams, in grams), feed type (feed_type), protein content, and feeding records, wherein each feeding record includes the feeding time period and the amount of feed.

[0058] S103. Convert the structured feeding plan into feeding machine control instructions and send them to the feeding machine for execution.

[0059] (1) Design the interface for feeder configuration, connectivity testing and status query; (2) Convert the structured feeding plan into feeder control instructions, which include the total daily feeding amount, feed type, feeding frequency, feeding time, and feeding amount per feeding; (3) Issue control commands according to the operating mode and record the decision history and equipment execution results.

[0060] This invention also provides an intelligent feeding decision-making system for Litopenaeus vannamei based on a large-scale model knowledge graph, comprising: The knowledge building and data management layer is used to acquire documents on Litopenaeus vannamei farming, feeding records, water quality records, feed instructions, production procedures, and expert experience texts. Based on a large language model, a knowledge graph for the intelligent feeding of Litopenaeus vannamei is constructed.

[0061] The data acquisition layer is used to collect real-time data on the aquaculture environment, user questions, or automatic feeding tasks, and to perform preprocessing. The collaborative decision-making computing layer is used to construct an intelligent feeding decision model based on a knowledge graph fusion with a large language model in the field of intelligent feeding of Litopenaeus vannamei; the pre-processed aquaculture environment data and user questions or automatic feeding tasks are input into the intelligent feeding decision model to generate a structured feeding plan.

[0062] The cloud-edge physical control layer is used to receive aquaculture environment data through sensor interfaces, convert structured feeding plans into executable instructions for automatic feeders, send them to the automatic feeders, and receive feedback from the equipment execution.

[0063] (1) Design a sensor data access interface to receive data uploaded by water quality sensors, aquaculture gateways, edge devices or front-end simulation devices.

[0064] (2) Perform field mapping, device number binding, timestamp recording, and integrity verification on sensor data; (3) Design the interface for feeder configuration, connectivity testing and status query.

[0065] (4) Convert the structured feeding plan into feeder control instructions. The control instructions include the total daily feeding amount, feed type, feeding frequency, feeding time, and feeding amount per feeding.

[0066] (5) Issue feeding instructions according to the operating mode, and record sensor history, decision history and equipment execution results.

[0067] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart feeding decision-making method for Litopenaeus vannamei based on a large-scale model knowledge graph, characterized in that, The steps include the following: (1) Obtain documents on Litopenaeus vannamei farming, feeding records, water quality records, feed instructions, production procedures and expert experience data, and construct a knowledge graph of intelligent feeding of Litopenaeus vannamei based on a large language model; (2) Construct an intelligent feeding decision model based on the knowledge graph of the field of intelligent feeding of Litopenaeus vannamei and the fusion of a large language model; collect relevant data of aquaculture factors and user questions or automatic feeding tasks in real time, and input them into the intelligent feeding decision model after preprocessing to generate a structured feeding plan; (3) Convert the structured feeding plan into feeding machine control instructions and send them to the feeding machine for execution.

2. The intelligent feeding decision-making method for Litopenaeus vannamei based on a large-scale knowledge graph as described in claim 1, characterized in that, The specific steps (1) are as follows: (1.1) Preprocess the unstructured text and embedded tables in the documents on Litopenaeus vannamei farming, feeding records, water quality records, feed instructions, production procedures and expert experience data, and convert them into natural language descriptions; (1.2) Use a large language model to extract entities that include at least environmental factors, breeding stage, feed type, feeding strategy, risk factors, detection indicators, control measures, equipment, nutritional indicators and disease symptoms; (1.3) Constructing knowledge triples: in: For a collection of entities, For a set of entity relations, For a collection of attributes; (1.4) Write the knowledge triples into a graph database or local graph structure storage, and create text block vector, entity vector and relation vector indexes; (1.5) The knowledge graph is merged with entity aliases, deduplicated with relations and quality checked to obtain a standardized knowledge graph for the intelligent feeding of Litopenaeus vannamei.

3. The intelligent feeding decision-making method for Litopenaeus vannamei based on a large-scale knowledge graph as described in claim 1, characterized in that, The data related to aquaculture factors include: water temperature, dissolved oxygen, pH value, salinity, ammonia nitrogen, nitrite, aquaculture stage, stocking density, body length, body weight, and one or more of the residual feed grade; The data related to aquaculture factors are obtained through any one of the following methods: real-time push via WebSocket, reporting via REST API, collection via MQTT gateway, or manual input via the front end.

4. The intelligent feeding decision-making method for Litopenaeus vannamei based on a large-scale knowledge graph as described in claim 3, characterized in that, Step (2) specifically involves: (2.1) Obtain real-time data on aquaculture factors and convert it into a context that can be understood by a large language model; The key-value pairs of each field of the aquaculture factor-related data are converted into natural language descriptive text according to the preset Chinese fishery terminology label mapping rules to form an aquaculture environment context; (2.2) Receive user questions or automatic feeding task trigger signals, and perform query intent identification and classification; Initial intent determination is performed using keyword heuristic matching: if the query contains explicit emergency keywords, it is directly determined to be an emergency response query; if the query contains feeding task keywords, it is determined to be a feeding suggestion query; for fuzzy queries containing sensor indicator names and abnormal semantic combinations, a large language model is further invoked in conjunction with the knowledge graph context of the intelligent feeding domain of Litopenaeus vannamei to enhance risk determination and output classification results; if the intent is emergency response, proceed to step (2.3); if the intent is feeding suggestion, proceed to step (2.4). (2.3) For emergency response type queries, the relevant context of emergency handling is retrieved directly from the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei in hybrid mode, the large language model is called to generate an emergency plan report, the feeding plan generation process is not executed, and the processing ends; (2.4) For feeding suggestion type queries, the context that the large language model can understand in step (2.1) and the intent category identified in step (2.2) are concatenated into an enhanced query statement according to a preset template, and then proceed to step (2.5). (2.5) For the enhanced query statement, the large language model is called to extract high-level keywords and low-level keywords, and the two sets of keywords are rewritten to obtain the rewritten high-level query statement and low-level query statement. Based on the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei, parallel retrieval of four retrieval modes (Naïve, Local, Global, and Hybrid) is performed simultaneously through LightRAG to obtain multi-mode retrieval results. On this basis, special retrieval is performed on six dimensions: weather conditions, breeding stage, breeding density, nutritional principles, risk factors, and practical experience. Based on the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei, multi-dimensional special query results are generated. (2.6) Calculate the environmental correction factor based on water temperature, dissolved oxygen, pH value, ammonia nitrogen, nitrite and residual feed level; in: Water temperature correction factor, For dissolved oxygen correction factor, pH correction factor For ammonia nitrogen correction factor, For nitrite correction factor, This is the correction factor for uneaten bait; If any of the following factors—water temperature, dissolved oxygen, pH, ammonia nitrogen, nitrite, and residual feed level—is not present in the aquaculture factor data, then the corresponding correction factor in the above formula is 1. (2.7) Determine the basic feeding coefficient based on the breeding stage and calculate the total daily feeding amount; in: This refers to the total daily feeding amount; Basic feeding coefficient; The remaining number in the pond; Average weight; If the average weight cannot be obtained, the feeding plan generation process will not be executed, and the process will end. (2.8) Distribute the total daily feed amount to multiple feeding periods according to a preset ratio to obtain: in: For the first Feeding amount per feeding Allocate proportions according to the corresponding time periods; (2.9) Perform comprehensive reasoning and structured output using a large language model; The multi-modal retrieval results and multi-dimensional specialized query results from step (2.5), along with the calculation results from steps (2.7) and (2.8), are merged into an enhanced knowledge context. The large language model is then invoked to perform the following steps: (2.9.1) Quality inspection and consistency verification: Based on the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei, cross-validation is performed to enhance the consistency between the knowledge context and the knowledge graph of the intelligent feeding domain of Litopenaeus vannamei, and contradictory conclusions are marked and confidence assessments are given. (2.9.2) Comprehensive synthesis of feeding scheme: The total daily feeding amount and the amount of food given each time By integrating with the knowledge graph in the field of intelligent feeding of Litopenaeus vannamei, which provides feeding strategy suggestions, environmental risk analysis, comparison of practical experience, and explanation of nutritional principles, a decision report is generated that includes the feeding amount calculation process, key judgment criteria, feed type suggestions, pellet size suggestions, and precautions. (2.9.3) Construction of the traceability appendix: Automatically add traceability and verification chapters with the structure of input source, parameter source, calculation link, evidence extract and executability statement to ensure that the decision-making process is traceable; The final output includes a decision report and a structured feeding plan; the structured feeding plan includes: total daily feeding amount, feed type, protein content, and feeding records, wherein each feeding record includes the feeding time period and feeding amount.

5. The intelligent feeding decision-making method for Litopenaeus vannamei based on a large model knowledge graph according to claim 4, characterized in that: When the collected data on aquaculture factors includes weight information, the average weight is the weight information provided. When the collected data on aquaculture factors does not contain weight information but does contain body length information, the weight value is calculated based on the body length as a basic reference value and determined comprehensively based on the knowledge graph of the intelligent feeding field of Litopenaeus vannamei. If there is no reference information on body weight and length in the collected data on breeding factors, the system will determine whether the relevant body weight can be obtained based on the knowledge graph. If it cannot be obtained, the system will inform the user truthfully in the report and provide relevant breeding information for reference, but will not feed the animal.

6. The intelligent feeding decision-making method for Litopenaeus vannamei based on a large model knowledge graph as described in claim 4, characterized in that: The feeding task keywords include at least one of "oxygen deficiency", "rainstorm", and "typhoon"; the feeding task keywords include at least one of "feeding" and "feed"; the abnormal semantics include at least one of "too high", "exceeding the standard", and "deterioration".

7. The intelligent feeding decision-making method for Litopenaeus vannamei based on a large-scale knowledge graph as described in claim 4, characterized in that, The specific steps (3) are as follows: (3.1) Design the interface for configuring the feeder, testing connectivity, and querying status; (3.2) Convert the structured feeding plan into feeder control instructions, which include the total daily feeding amount, feed type, feeding frequency, feeding time, and feeding amount per feeding; (3.3) Issue control commands according to the operating mode and record the decision history and equipment execution results.

8. A smart feeding decision-making system for Litopenaeus vannamei based on a large-scale knowledge graph, characterized in that, include: The knowledge building and data management layer is used to acquire documents on Litopenaeus vannamei farming, feeding records, water quality records, feed instructions, production procedures and expert experience texts, and to build a knowledge graph for the intelligent feeding of Litopenaeus vannamei based on a large language model; The data acquisition layer is used to collect real-time data on the aquaculture environment, user questions, or automatic feeding tasks, and to perform preprocessing. The collaborative decision-making computing layer is used to construct an intelligent feeding decision model based on a knowledge graph fusion with a large language model in the field of intelligent feeding of Litopenaeus vannamei; the pre-processed aquaculture environment data and user questions or automatic feeding tasks are input into the intelligent feeding decision model to generate a structured feeding plan; The cloud-edge physical control layer is used to receive aquaculture environment data through sensor interfaces, convert structured feeding plans into executable instructions for automatic feeders, send them to the automatic feeders, and receive feedback from the equipment execution.

9. The intelligent feeding decision-making system for Litopenaeus vannamei based on a large-scale knowledge graph as described in claim 8, characterized in that, The cloud-edge physical control layer specifically includes: (1) Design a sensor data access interface to receive data uploaded by water quality sensors, aquaculture gateways, edge devices or front-end simulation devices; (2) Perform field mapping, device number binding, timestamp recording, and integrity verification on sensor data; (3) Design the interface for the feeder configuration, connectivity test, and status query; (4) Convert the structured feeding plan into feeder control instructions, which include the total daily feeding amount, feed type, feeding frequency, feeding time, and feeding amount per feeding. (5) Issue feeding instructions according to the operating mode, and record sensor history, decision history and equipment execution results.

Images