An agricultural decision-making method based on multi-agent cooperation
By combining multi-agent collaboration and a large language model, unified processing and cross-domain decision-making of multi-source heterogeneous data are achieved, solving the problems of multi-source heterogeneous data processing and cross-task collaboration in existing technologies, and providing efficient and intelligent agricultural decision support.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN AEROSPACE STAR DATA SYST TECH CO LTD
- Filing Date
- 2025-12-08
- Publication Date
- 2026-07-21
AI Technical Summary
Existing intelligent decision-making methods for agriculture cannot efficiently process multi-source heterogeneous data, lack cross-task collaboration and dynamic scheduling capabilities, and are unable to meet the comprehensive decision-making needs in complex agricultural scenarios.
A multi-agent collaborative approach is adopted, which uses a large language model to parse user needs, breaks them down into multiple sub-tasks, dynamically schedules specialized agents to perform data processing and decision-making, and uses a data pipeline mechanism to achieve seamless collaboration between different agents, ultimately generating a natural language agricultural decision report.
It enables unified processing of multi-source heterogeneous data and cross-domain integrated decision-making, providing efficient, intelligent, and interpretable agricultural decision support, and improving the scientific and intelligent level of agricultural production.
Smart Images
Figure CN121436586B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of agricultural artificial intelligence technology, specifically relating to an agricultural decision-making method based on multi-agent collaboration. Background Technology
[0002] With the development of agricultural modernization, the sources of data in agricultural production processes are becoming increasingly diversified, including field images, meteorological monitoring, soil nutrient sensors, agricultural machinery operation records, and farmers' natural language processing needs. These multi-source, heterogeneous data provide rich information support for agricultural production; however, due to significant differences in data types, how to effectively integrate them and formulate actionable agricultural decisions remains a core issue facing current agricultural intelligence.
[0003] Most existing intelligent decision-making methods in agriculture focus on single tasks, such as pest and disease detection based on image recognition, disaster early warning based on meteorological data, or fertilization recommendations based on statistical models. While these methods have made breakthroughs in certain aspects, they lack cross-task collaborative capabilities and struggle to meet the comprehensive decision-making needs in complex agricultural scenarios. For example, when farmers ask how to ensure the safety of winter wheat during the grain-filling stage when a cold wave arrives, they often need to combine information from multiple aspects, including crop phenological forecasting, meteorological disaster analysis, soil fertility assessment, and agricultural machinery scheduling. Existing single-task models cannot handle such comprehensive, multi-stage reasoning and decision-making.
[0004] In recent years, the development of large language models has provided new ideas for intelligent agriculture. Large language models possess powerful natural language understanding and knowledge reasoning capabilities, enabling them to transform complex agricultural needs into executable sub-tasks. However, large language models themselves lack the ability to directly analyze images, time-series meteorological data, or structured nutrient data; therefore, they must rely on external specialized intelligent agents to perform specific calculations and reasoning. The capabilities of a single intelligent agent are limited and cannot cover the complex tasks of the entire agricultural process, necessitating a new method that enables multi-agent collaborative work.
[0005] Therefore, designing an intelligent decision-making method that can uniformly process multi-source heterogeneous data and complete complex agricultural tasks through multi-agent collaboration has become a key issue in agricultural artificial intelligence research and application. Summary of the Invention
[0006] The problem this invention aims to solve is that existing agricultural intelligent decision-making methods cannot efficiently process multi-source heterogeneous data and lack cross-task collaboration and dynamic scheduling capabilities. It proposes an agricultural decision-making method based on multi-agent collaboration.
[0007] To achieve the above objectives, the present invention provides the following technical solution:
[0008] An agricultural decision-making method based on multi-agent cooperation includes the following steps:
[0009] S1. Obtain agricultural decision requests submitted by users, standardize the multi-source heterogeneous data in the agricultural decision requests, and obtain standardized data;
[0010] S2. Perform task semantic parsing on the standardized data obtained in step S1 to divide it into subtasks, then establish a set of candidate agents for each subtask, and calculate the scheduling score for the set of subtask-agent combinations. Based on the scheduling score results, obtain a dynamic scheduling scheme for agents.
[0011] S3. Establish a data pipeline mechanism to automatically convert the upstream agent output to the downstream agent input of the agent dynamic scheduling scheme obtained in step S2, complete cross-agent collaboration, and have each agent process synchronously based on a unified interface specification to obtain the parallel execution result of multiple agents.
[0012] S4. The results of the multi-agent parallel execution obtained in step S3 are weighted and fused with confidence scores to calculate the confidence score of the final agricultural decision scheme, and a natural language agricultural decision report is generated based on the large language model.
[0013] Furthermore, the specific implementation method of step S1 includes the following steps:
[0014] S1.1. Obtain agricultural decision requests submitted by users, including one or more of the following: natural language instructions, crop growth images, and structured environmental data;
[0015] S1.2. Standardize the multi-source heterogeneous data in agricultural decision-making requests, including word segmentation, stop word filtering, and UTF-8 encoding conversion for text instructions; resolution normalization and histogram equalization for image data; and missing value imputation and unit unification conversion for structured data.
[0016] S1.3. Output the data after standardization in step S1.2 as a standardized data packet in JSON format.
[0017] Furthermore, the specific implementation method of step S2 includes the following steps:
[0018] S2.1. Perform task semantic parsing on the standardized data obtained in step S1. Use the BERT-based agricultural domain classification model to perform task semantic parsing and output the task labels and confidence scores of the sub-tasks.
[0019] S2.2. Establish a set of candidate agents for each subtask obtained in step S2.1. The agents include one or more of the following: phenological prediction agent, meteorological analysis agent, pest and disease diagnosis agent, soil nutrient analysis agent, agricultural machinery scheduling agent, and agricultural decision-making agent. This results in a set of subtask-agent combinations.
[0020] S2.3. Calculate the scheduling score for the set of subtask-agent combinations obtained in step S2.2, using the following formula:
[0021]
[0022] in, This represents the scheduling score of agent i on subtask j. This represents the historical accuracy of agent i on subtask j. This represents the relevance matching degree of agent i on subtask j. This represents the real-time load rate of agent i. The delay of agent i in executing subtask j is represented by the four indicators, all of which are normalized to the interval [0,1]. , , , The weighting coefficients are for historical accuracy, relevance matching, real-time load rate, and latency. , , , The sum is 1;
[0023] S2.4. Select the agent with the highest scheduling score from the candidate agent set to perform the corresponding sub-task operation, and obtain the agent dynamic scheduling scheme.
[0024] Furthermore, the specific implementation method for establishing the data pipeline mechanism in step S3 includes the following steps:
[0025] S3.1. Data Transformation Based on Mapping Tables: When the output of the upstream agent in the agent dynamic scheduling scheme is discrete, the upstream result 𝑥 is mapped to the input parameters of the downstream agent according to a preset function. The expression is:
[0026]
[0027] in, The input function is used to transform the downstream agent. Represents a mapping table function;
[0028] S3.2. Upstream and downstream agents transmit data based on a rule engine;
[0029] When the input satisfies the condition set C={ , ,…, When}, the corresponding rule action set A={ , ,…, Its execution logic is defined as follows:
[0030]
[0031] in, The set of actions output by the rules engine. Let k be the k-th indicator function, where k is any one of m. For the k-th condition, For the k-th rule action, when the input satisfies the condition The value is 1 if the condition is met, and 0 otherwise.
[0032] S3.3. When a downstream agent receives data from an upstream agent and performs continuous input, the expression for generating the data using a regression model or neural network function is as follows:
[0033]
[0034] in, To generate data, For regression models or neural network functions, These are model parameters;
[0035] The upstream result 𝑥 is predicted and generated, and the output is a numerical input that can be directly used by the downstream agent.
[0036] Furthermore, the specific implementation method of step S4 includes the following steps:
[0037] S4.1. Perform confidence-weighted fusion of the multi-agent parallel execution results obtained in step S3 to calculate the confidence of the final solution. The expression for calculating the confidence of the final solution is as follows:
[0038]
[0039] in, Assuming confidence level for the final solution, Let i be the weight of agent i. The output confidence of agent i;
[0040] S4.2. Natural language generation is based on a large language model in the agricultural field. The results of the multi-agent parallel execution obtained in step S3 are used as structured input data. After being processed by the semantic understanding and content generation module, the agricultural decision report in natural language form is output, realizing the human-computer readable expression of agricultural decision results.
[0041] The beneficial effects of this invention are:
[0042] This invention discloses an agricultural decision-making method based on multi-agent collaboration. By introducing a large language model as the core for task parsing and natural language generation, and combining it with a multi-agent collaboration mechanism, it achieves full-process automation from user natural language requirements to structured agricultural decisions. Specifically, this method is implemented through the following technical solutions: First, users submit agricultural decision requests through mobile terminals or web interfaces. The input data includes natural language instructions, field images, and environmental parameters. The system standardizes multi-source inputs to ensure that different types of data can be accessed within a unified framework. Subsequently, the scheduling center parses user requirements based on the large language model, decomposes complex tasks into multiple atomic-level agricultural sub-tasks, and assigns a set of candidate agents to each sub-task. By designing a scheduling scoring formula, the "sub-task-agent combination" is evaluated, comprehensively considering indicators such as historical accuracy, task relevance, load rate, and latency, and dynamically selecting the optimal agent to execute the task. Then, through a predefined data pipeline, the results are transferred and automatically converted between different agents, including various forms such as mapping tables, rule engines, and prediction models, ensuring seamless connection between upstream outputs and downstream inputs, and supporting parallel and collaborative computing among multiple agents. Finally, the system performs confidence-weighted fusion of the results from each agent to calculate the reliability of the final solution, and uses an agricultural big language model to generate a natural language report, outputting intuitive and easy-to-understand agricultural technology decision-making suggestions.
[0043] The present invention discloses an agricultural decision-making method based on multi-agent collaboration, which innovatively combines semantic parsing of a large language model with professional computation of multiple agents, realizing task decomposition, dynamic scheduling and result fusion in complex agricultural scenarios, and can provide efficient, intelligent and interpretable decision support for agricultural production.
[0044] This invention introduces a large language model and a multi-agent collaborative mechanism to uniformly process heterogeneous data from multiple sources, such as natural language, images, meteorology, and soil, breaking through the limitations of traditional agricultural decision-making methods that can only be applied to a single task, and realizing cross-domain integrated decision-making. Attached Figure Description
[0045] Figure 1 This is a flowchart of an agricultural decision-making method based on multi-agent collaboration as described in this invention. Detailed Implementation
[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.
[0047] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.
[0048] To further understand the invention's content, features, and effects, the following specific embodiments are provided, along with accompanying drawings. Figure 1 Detailed explanation is as follows:
[0049] Example 1:
[0050] An agricultural decision-making method based on multi-agent cooperation includes the following steps:
[0051] S1. Obtain agricultural decision requests submitted by users, standardize the multi-source heterogeneous data in the agricultural decision requests, and obtain standardized data;
[0052] Furthermore, the specific implementation method of step S1 includes the following steps:
[0053] S1.1. Obtain agricultural decision requests submitted by users, including one or more of the following: natural language instructions, crop growth images, and structured environmental data;
[0054] S1.2. Standardize the multi-source heterogeneous data in agricultural decision-making requests, including word segmentation, stop word filtering, and UTF-8 encoding conversion for text instructions; resolution normalization and histogram equalization for image data; and missing value imputation and unit unification conversion for structured data.
[0055] S1.3. Output the data after standardization in step S1.2 as a standardized data packet in JSON format.
[0056] S2. Perform task semantic parsing on the standardized data obtained in step S1 to divide it into subtasks, then establish a set of candidate agents for each subtask, and calculate the scheduling score for the set of subtask-agent combinations. Based on the scheduling score results, obtain a dynamic scheduling scheme for agents.
[0057] Furthermore, the specific implementation method of step S2 includes the following steps:
[0058] S2.1. Perform task semantic parsing on the standardized data obtained in step S1. Use the BERT-based agricultural domain classification model to perform task semantic parsing and output the task labels and confidence scores of the sub-tasks.
[0059] S2.2. Establish a set of candidate agents for each subtask obtained in step S2.1. The agents include one or more of the following: phenological prediction agent, meteorological analysis agent, pest and disease diagnosis agent, soil nutrient analysis agent, agricultural machinery scheduling agent, and agricultural decision-making agent. This results in a set of subtask-agent combinations.
[0060] S2.3. Calculate the scheduling score for the set of subtask-agent combinations obtained in step S2.2, using the following formula:
[0061]
[0062] in, This represents the scheduling score of agent i on subtask j. This represents the historical accuracy of agent i on subtask j. This represents the relevance matching degree of agent i on subtask j. This represents the real-time load rate of agent i. The delay of agent i in executing subtask j is represented by the four indicators, all of which are normalized to the interval [0,1]. , , , The weighting coefficients are for historical accuracy, relevance matching, real-time load rate, and latency. , , , The sum is 1;
[0063] S2.4. Select the agent with the highest scheduling score from the candidate agent set to perform the corresponding sub-task operation, and obtain the agent dynamic scheduling scheme.
[0064] S3. Establish a data pipeline mechanism to automatically convert the upstream agent output to the downstream agent input of the agent dynamic scheduling scheme obtained in step S2, complete cross-agent collaboration, and have each agent process synchronously based on a unified interface specification to obtain the parallel execution result of multiple agents.
[0065] Furthermore, the specific implementation method for establishing the data pipeline mechanism in step S3 includes the following steps:
[0066] S3.1. Data Transformation Based on Mapping Tables: When the output of the upstream agent in the agent dynamic scheduling scheme is discrete, the upstream result 𝑥 is mapped to the input parameters of the downstream agent according to a preset function. The expression is:
[0067]
[0068] in, The input function is used to transform the downstream agent. Represents a mapping table function;
[0069] S3.2. Upstream and downstream agents transmit data based on a rule engine;
[0070] When the input satisfies the condition set C={ , ,…, When}, the corresponding rule action set A={ , ,…, Its execution logic is defined as follows:
[0071]
[0072] in, The set of actions output by the rules engine. Let k be the k-th indicator function, where k is any one of m. For the k-th condition, For the k-th rule action, when the input satisfies the condition The value is 1 if the condition is met, and 0 otherwise.
[0073] S3.3. When a downstream agent receives data from an upstream agent and performs continuous input, the expression for generating the data using a regression model or neural network function is as follows:
[0074]
[0075] in, To generate data, For regression models or neural network functions, These are model parameters;
[0076] The upstream result 𝑥 is predicted and generated, and the output is a numerical input that can be directly used by the downstream agent.
[0077] S4. The results of the multi-agent parallel execution obtained in step S3 are weighted and fused with confidence scores to calculate the confidence score of the final agricultural decision scheme, and a natural language agricultural decision report is generated based on the large language model.
[0078] Furthermore, the specific implementation method of step S4 includes the following steps:
[0079] S4.1. Perform confidence-weighted fusion of the multi-agent parallel execution results obtained in step S3 to calculate the confidence of the final solution. The expression for calculating the confidence of the final solution is as follows:
[0080]
[0081] in, Assuming confidence level for the final solution, Let i be the weight of agent i. The output confidence of agent i;
[0082] S4.2. Natural language generation is based on a large language model in the agricultural field. The results of the multi-agent parallel execution obtained in step S3 are used as structured input data. After being processed by the semantic understanding and content generation module, the agricultural decision report in natural language form is output, realizing the human-computer readable expression of agricultural decision results.
[0083] This embodiment introduces a large language model and a multi-agent collaborative mechanism to uniformly process heterogeneous data from multiple sources, such as natural language, images, meteorology, and soil. This breaks through the limitations of traditional agricultural decision-making methods that can only be applied to a single task, and enables cross-domain integrated decision-making.
[0084] The scheduling and scoring mechanism proposed in this embodiment can dynamically select the optimal agent for different sub-tasks and monitor the agent's load and latency in real time during task execution, thereby ensuring the rationality of task allocation and the efficiency of system operation.
[0085] The data pipeline mechanism designed in this embodiment supports multiple forms such as mapping tables, rule engines, and prediction models. It can automatically complete the data format conversion and logical mapping between different types of intelligent agents, improving the scalability and adaptability of multi-agent collaboration.
[0086] This embodiment utilizes confidence-weighted fusion and natural language generation mechanisms to not only provide structured decision results but also output natural language reports that conform to the language habits of agricultural practitioners, thereby improving the understandability and operability of the results.
[0087] Compared with traditional agricultural decision-making methods that rely on single-model reasoning or human experience, this invention can achieve intelligent, automated and interpretable decision support in complex agricultural scenarios (such as disaster response, crop management, and agricultural machinery scheduling), significantly improving the scientific and intelligent level of agricultural production.
[0088] The specific implementation plan of this embodiment is illustrated below:
[0089] The specific implementation method of step S1 includes the following steps:
[0090] S1.1. Obtain the agricultural decision request submitted by the user, input the natural language command: "Impact of cold wave on winter wheat grain filling period and coping solution", and upload the field RGB image of winter wheat in the middle of grain filling, as well as the CSV file containing temperature, precipitation and soil nutrient parameters.
[0091] S1.2. Perform standardization processing on the multi-source heterogeneous data contained therein:
[0092] (1) Perform word segmentation, stop word filtering, and UTF-8 encoding conversion on text instructions;
[0093] (2) Normalize the resolution of the image data to 224×224 pixels and perform histogram equalization.
[0094] (3) The KNN algorithm is used to perform missing value imputation and unit conversion on structured data;
[0095] (4) Output a standardized JSON data packet with the following structure:
[0096] {
[0097] "task_id": "UUID",
[0098] "text_query": "Impact of Cold Wave on Winter Wheat Grain-Filling Stage and Countermeasures",
[0099] "image_data": "Base64 encoded image",
[0100] "env_data": {"temperature": [-5,-3,-2], "growth_stage": "Mid-stage grouting"}
[0101] }
[0102] The specific implementation method of step S2 includes the following steps:
[0103] S2.1. The scheduling center calls the BERT-based agricultural domain classification model to perform task semantic parsing, identifies the main task type as "meteorological disaster response", outputs the confidence score and decomposes the task into sub-tasks: phenological prediction, cold wave meteorological analysis, soil nutrient analysis, agricultural machinery scheduling, and agricultural decision generation.
[0104] S2.2. Establish a set of candidate agents for each subtask. The agents include: phenological prediction agent, cold wave analysis agent, soil nutrient analysis agent, agricultural machinery scheduling agent, and agricultural decision-making agent.
[0105] S2.3. Scheduling scoring is calculated for the subtask-agent combination, and the scoring formula is as follows:
[0106]
[0107] in, This represents the historical accuracy of agent i on subtask j. This indicates the degree of relevance and matching between the agent and the subtask. This indicates the real-time load rate of the intelligent agent. This indicates the delay in the agent executing the subtask, and all four indicators are normalized to the [0,1] interval; , , , These are the weighting coefficients, and their sum is 1.
[0108] S2.4. The scheduling center selects the highest-scoring agent from the candidate agent set to execute the corresponding subtask, and monitors it in real time during the task execution. If an increase in load or latency is detected, a re-scoring is triggered and an agent switch is performed.
[0109] The specific implementation method of step S3 includes the following steps:
[0110] S3.1. Data transformation based on mapping table: When the output of the upstream agent is a discrete level, it is transformed according to a preset function:
[0111]
[0112] The result 𝑥 is mapped to the input parameters of the downstream agent, where 𝑀 represents the mapping table function. The meteorological analysis agent outputs "Minimum temperature = -5℃, lasting 3 days", and the mapping table function... This is converted into "cold wave level = orange alert" and used as input for the agricultural decision-making intelligence agent;
[0113] S3.2. Data transmission based on a rule engine: When the input satisfies the condition set C={ , ,…, }
[0114] When, the corresponding rule action set A={ , ,…, Its execution logic is defined as follows:
[0115]
[0116] in, This is an indicator function that responds when the input meets the condition. The value is 1 if the crop's growth period equals the grain-filling period and the temperature is <0℃, the rule engine function R(x) automatically triggers the scheduling instruction to "delay irrigation for 3 days".
[0117] S3.3. Data Generation Based on Predictive Models: When downstream agents require continuous input, regression models or neural network functions are used.
[0118]
[0119] The upstream result 𝑥 is used to generate predictions, and the output is a numerical input that can be directly used by the downstream agent, where 𝜃 represents the model parameters. Soil nutrient analysis outputs "nitrogen content = 1.2%", and the prediction model function... Generate the corresponding "recommended fertilizer application rate = 5.0 kg / mu" and pass it to the fertilization decision subtask.
[0120] The specific implementation method of step S4 includes the following steps:
[0121] S4.1. The results from the multi-agent system are weighted and fused to calculate the confidence score of the final solution. The confidence score calculation for the final solution is as follows:
[0122]
[0123] in, Let i be the weight of agent i. Let be the output confidence of agent i.
[0124] S4.2. Using the finely tuned agricultural big language model LLaMA-Agri structured input, a natural language agricultural decision report is generated: "Based on phenological forecasts and meteorological analysis, winter wheat is in the middle of the grain-filling stage. The cold wave will last for 3 days, with a minimum temperature of -5℃. It is recommended to spray 5L / mu of antifreeze, delay irrigation for 3 days, and apply nitrogen fertilizer to mitigate the impact of the cold wave on yield."
[0125] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0126] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.
Claims
1. An agricultural decision-making method based on multi-agent collaboration, characterized in that, Includes the following steps: S1. Obtain agricultural decision requests submitted by users, standardize the multi-source heterogeneous data in the agricultural decision requests, and obtain standardized data; S2. Perform task semantic parsing on the standardized data obtained in step S1 to divide it into subtasks, then establish a set of candidate agents for each subtask, and calculate the scheduling score for the set of subtask-agent combinations. Based on the scheduling score results, obtain a dynamic scheduling scheme for agents. The specific implementation method of step S2 includes the following steps: S2.
1. Perform task semantic parsing on the standardized data obtained in step S1. Use the BERT-based agricultural domain classification model to perform task semantic parsing and output the task labels and confidence scores of the sub-tasks. S2.
2. Establish a set of candidate agents for each subtask obtained in step S2.
1. The agents include one or more of the following: phenological prediction agent, meteorological analysis agent, pest and disease diagnosis agent, soil nutrient analysis agent, agricultural machinery scheduling agent, and agricultural decision-making agent. This results in a set of subtask-agent combinations. S2.
3. Calculate the scheduling score for the set of subtask-agent combinations obtained in step S2.2, using the following formula: in, This represents the scheduling score of agent i on subtask j. This represents the historical accuracy of agent i on subtask j. This represents the relevance matching degree of agent i on subtask j. This represents the real-time load rate of agent i. The delay of agent i in executing subtask j is represented by the four indicators, all of which are normalized to the interval [0,1]. , , , The weighting coefficients are for historical accuracy, relevance matching, real-time load rate, and latency. , , , The sum is 1; S2.
4. Select the agent with the highest scheduling score from the candidate agent set to execute the corresponding sub-task operation, and obtain the agent dynamic scheduling scheme; S3. Establish a data pipeline mechanism to automatically convert the upstream agent output to the downstream agent input of the agent dynamic scheduling scheme obtained in step S2, complete cross-agent collaboration, and have each agent process synchronously based on a unified interface specification to obtain the parallel execution result of multiple agents. The specific implementation method for establishing the data pipeline mechanism in step S3 includes the following steps: S3.
1. Data Transformation Based on Mapping Tables: When the output of the upstream agent in the agent dynamic scheduling scheme is discrete, the upstream result 𝑥 is mapped to the input parameters of the downstream agent according to a preset function. The expression is: in, The input function for the transformed downstream agent is 𝑀( ) represents a mapping table function; S3.
2. Upstream and downstream agents transmit data based on a rule engine; When the input satisfies the condition set C={ , ,…, When}, the corresponding rule action set A={ , ,…, Its execution logic is defined as follows: in, The set of actions output by the rules engine. Let k be the k-th indicator function, where k is any one of m. For the k-th condition, For the k-th rule action, when the input meets the condition The value is 1 if the condition is met, and 0 otherwise. S3.
3. When a downstream agent receives data from an upstream agent and performs continuous input, the expression for generating the data using a regression model or neural network function is as follows: in, To generate data, For regression models or neural network functions, � represents the model parameters; The upstream result 𝑥 is predicted and generated, and the output is a numerical input that can be directly used by the downstream agent. S4. The results of the multi-agent parallel execution obtained in step S3 are weighted and fused with confidence scores to calculate the confidence score of the final agricultural decision scheme, and a natural language agricultural decision report is generated based on the large language model.
2. The agricultural decision-making method based on multi-agent collaboration according to claim 1, characterized in that, The specific implementation method of step S1 includes the following steps: S1.
1. Obtain agricultural decision requests submitted by users, including one or more of the following: natural language instructions, crop growth images, and structured environmental data; S1.
2. Standardize the multi-source heterogeneous data in agricultural decision-making requests, including word segmentation, stop word filtering, and UTF-8 encoding conversion for text instructions; resolution normalization and histogram equalization for image data; and missing value imputation and unit unification conversion for structured data. S1.
3. Output the data after standardization in step S1.2 as a standardized data packet in JSON format.
3. The agricultural decision-making method based on multi-agent collaboration according to claim 2, characterized in that, The specific implementation method of step S4 includes the following steps: S4.
1. Perform confidence-weighted fusion of the multi-agent parallel execution results obtained in step S3 to calculate the confidence of the final solution. The expression for calculating the confidence of the final solution is as follows: in, Assuming confidence level for the final solution, Let i be the weight of agent i. The output confidence of agent i; S4.
2. Natural language generation is based on a large language model in the agricultural field. The results of the multi-agent parallel execution obtained in step S3 are used as structured input data. After being processed by the semantic understanding and content generation module, the agricultural decision report in natural language form is output, realizing the human-computer readable expression of agricultural decision results.