A state self-driven multi-agent crop growth report generation method and system

Abstract

The application discloses a state self-driven multi-agent crop growth report generation method and system, first, a multi-agent cooperative communication mechanism is constructed, a unified event-driven model and a communication pipeline are built, information routing and state synchronization between heterogeneous agents such as planning, querying and data analysis are realized, and asynchronous cooperation and data intercommunication of a multi-agent network under complex logic are guaranteed. Secondly, a state self-driven agent architecture is realized, that is, a self-driven core based on a finite state machine is implanted into the agent, so that each node can autonomously complete state flow, action decision and local task execution according to real-time context environment and task feedback mechanism. Finally, automatic generation of a structured document is realized, special report agents are used to logically aggregate and template rendering of multi-dimensional analysis data and knowledge fragments output by the cooperative network, and finally, a structured analysis report with standard layout and rigorous content is output.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing, and agricultural remote sensing technology, specifically to a state-driven multi-agent crop growth report generation method and system. Background Technology

[0002] With the rapid development of agricultural remote sensing technology, large language models, and artificial intelligence agent technology, automatically generating crop growth analysis reports based on multi-source remote sensing data has become an important research direction in the field of smart agriculture. Currently, agricultural production management widely utilizes information such as UAV remote sensing imagery, satellite multispectral data, crop indices (e.g., NDVI), meteorological data, and historical planting records of plots to comprehensively analyze crop growth, growth differences, pest and disease risks, and yield predictions. However, structuring the aforementioned multi-source data and automatically generating logically clear and professionally standardized growth analysis reports still heavily relies on human expert experience, resulting in low efficiency and inconsistent standards. In recent years, large language models and intelligent agent technology have been gradually introduced into agricultural remote sensing analysis scenarios to achieve automated generation of crop growth reports. However, in practical applications, existing intelligent agents generally exhibit technical defects such as contextual redundancy and logical disorder when performing long-process, highly structured tasks like agricultural growth report generation. Specifically, the crop growth report generation process often includes several sub-tasks: generating a summary of remote sensing image preprocessing results; analyzing crop index trends; identifying and inferring the causes of abnormal plots; assessing the impact of meteorological factors; generating field management recommendations; and evaluating the overall growth level. These sub-tasks are both data-dependent and logically progressive. However, in the existing single-agent architecture, the local analysis data required for the current sub-task (such as the change in the average NDVI of a certain plot) is often mixed with the global task objective (such as the whole-season growth evaluation framework) in the unified prompting context. This leads to: the agent easily ignoring the overall report structure when analyzing specific indicators; potentially deviating from actual monitoring data when generating management recommendations; redundant contextual information causing distraction; and logical hierarchy disorder during the generation of long reports.

[0003] While existing task decomposition methods formally segment the report generation process, they fail to build a refined workflow for scenarios like agricultural growth analysis, which involve "strong data constraints and a strong logical structure." Therefore, they struggle to guarantee the professional consistency and structural stability of the reports. To alleviate the limitations of single agents in handling complex agricultural analysis tasks, some existing technologies attempt to adopt a multi-agent collaborative architecture, such as setting up separate data analysis agents, text generation agents, and decision-making suggestion agents. However, while using a multi-agent architecture achieves some degree of functional decoupling of the various modules in agricultural growth reports, it introduces new technical bottlenecks: in agricultural remote sensing analysis scenarios, global information (such as crop type, growth stage, and regional climate background) and local analysis data (such as anomalies in specific plot indices) are strongly interdependent. Without a precise state management mechanism, information transmission between multiple agents can easily lead to: loss of global planting background information; redundancy and duplication of data for the same plot across multiple agents; inconsistencies between current growth stage judgments and management suggestions; and inferences about the causes of anomalies deviating from previous analysis results. In the process of heterogeneous multi-agent collaborative generation of crop growth reports, due to the lack of dynamic and precise context isolation and state-driven management mechanisms, global agricultural background information is prone to information pollution or redundant accumulation when it is transmitted between multiple nodes, which seriously affects the professionalism and logical stability of the generated report.

[0004] In summary, in the specific application scenario of automated crop growth report generation, existing technologies struggle to simultaneously address: the unified constraints of the overall agricultural planting context; the precise focus of local data analysis for each subtask; the structured information flow between multiple agents; and the state stability during the long report generation process. Summary of the Invention

[0005] To address the problems in existing technologies, such as disordered agent context, interference between global information and local task execution, and difficulty in controlling context distribution during multi-agent collaboration, this invention provides a state-driven multi-agent crop growth report generation method. This invention achieves effective isolation and precise execution of global guidance and local details by reconstructing the agent's context structure and combining it with a unified event communication mechanism.

[0006] To achieve the above objectives, the technical solution adopted by this invention is a state-driven multi-agent crop growth report generation method, comprising the following steps: Initialize the multi-agent collaborative operating environment and communication topology, establish a global event manager, communication pipelines and global sessions to provide a unified foundation for asynchronous collaboration among multiple agents; A state-driven intelligent agent kernel is constructed, and the working context of each heterogeneous intelligent agent is reconstructed. The intelligent agent context is divided into a state management module, a global guidance module, and a dynamic state region. The planning agent decomposes the crop growth report generation task based on the global context, forming fine-grained subtasks with clear input-output boundaries and dependencies; After receiving the corresponding subtask, each executing agent completes local closed-loop processing of retrieval, analysis, diagnosis or reasoning within the dynamic state region, and undergoes state evolution under execution feedback; Each executing agent encapsulates local task results and state changes into standard events, and completes cross-node state synchronization and data routing through the global event manager; The report-generating agent aggregates, logically organizes, and renders the structured results output by each upstream agent to generate a well-formatted and clearly structured crop growth report.

[0007] Furthermore, a global event manager and communication pipeline are first instantiated, and an underlying message routing mechanism is constructed to establish a global session for maintaining the execution cycle of this report generation task. The global session is used to record task identifier, crop type, monitoring area, growth period, report template requirements, and common constraint information during the execution process. During the construction of the underlying message routing mechanism, to ensure the uniqueness and traceability of communication routes, batch name verification is performed on the agents to be registered under mutual exclusion control conditions; let the set of agent objects to be registered be:

[0008] Extract its corresponding set of names and denote it as:

[0009] If and only if the cardinality of the deduplicated name set satisfies:

[0010] The system determines that all agent names are unique and writes each agent instance into the local routing table to complete the address binding required for subsequent event delivery; if equation (3) is not satisfied, it means that at least two agents have the same identifier, and the system triggers an initialization failure exception. In equations (1) to (3), A represents the set of registered intelligent agent objects. Name represents the i-th agent instance, m represents the total number of instantiated agents, and Name( ) represents the unique identifier name of the i-th agent. Represents the set consisting of all agent names, | | indicates the cardinality of the set.

[0011] Furthermore, the state management module is used to perceive the current lifecycle state and state transition conditions, including the necessary perception information to guide state switching, enabling the agent to autonomously determine its current lifecycle. Each agent executing a specific task is controlled by a state management module to make task decisions, controlling a set of runtime state variables related to the specific task, the lifecycle of the specific task execution from start to finish, and the state switching strategy bound to the task. The global guidance module, i.e., the main prompt word area, is used to guide and constrain the behavioral boundaries of the agent in automated tasks. It includes a set of static rules defining task specifications, a clear and explicit list of overall goals, and template constraints for formatted output in the task execution process. The dynamic state area serves as the core working area for fine-grained task execution. When the agent focuses on processing a sub-task, all local detailed instructions for that sub-task are dynamically gathered here and are tightly bound to the current execution state. The dynamic state area is the agent's context used to store the above information and define a set of operations for information transformation within the area. The runtime state is ultimately divided into three aspects: sub-task instructions, intermediate results of task execution, and feedback increments formed by self-evaluation of task execution. During the construction of the dynamic state region, the system first merges the current session state, task dependencies, and additional metadata to form a joint context variable space, the calculation formula of which is:

[0012] In equation (4), This represents the dictionary of merged union context state variables. This represents the current set of running session states of the system. This represents the underlying task execution dependency set. This indicates an additional meta-dataset.

[0013] Furthermore, the full working context space reconstruction expression of the agent at time t is:

[0014] Its state evolution equation is:

[0015] In equations (5) and (6), Indicates that the intelligent agent is in The full working context space at any given moment; Represents non-overlapping structured splicing operations across different contextual dimensions; This represents a static, global guidance area that characterizes the overall requirements of the task. express A separate state-aware region used to guide state machine transitions at all times; express Moment-based intelligent agents centrally process the first The dynamic state area for each sub-task; and These represent the intelligent agent in Time and The lifecycle running status at any given moment; This indicates the received external trigger event stream or execution feedback signal; This represents the self-driven state transition function of the agent kernel.

[0016] Furthermore, let the global overall natural language task be... The context of the global prompt is Then the task parsing mapping function inside the planning agent is expressed as:

[0017] In equation (7), This represents the task parsing mapping function within the planning agent. This represents n fine-grained subtasks decoupled from the global task; the subtasks include task units required for different execution nodes such as information retrieval, data processing, growth diagnosis, anomaly analysis, suggestion generation, and report writing.

[0018] Furthermore, after task decomposition, a context isolation constraint objective function is constructed:

[0019] In equation (8), This represents the target value for the constraint of task parsing and context isolation. This represents the local context assigned to the i-th subtask. This represents the local context assigned to the j-th subtask. This function represents the calculation function for mutual information redundancy between different local contexts. Subtasks in multi-agent cooperative networks and The data dependency penalty term between them, where λ represents the balance coefficient.

[0020] Furthermore, the specific expression of the mutual information redundancy calculation function is as follows:

[0021] The first item is the semantic repetition rate of the field. These are the weighting coefficients. These are semantic vector representations of the local context of the i-th and j-th subtasks, respectively. The data dependency penalty is specifically represented as follows:

[0022] Indicates a subtask For the subtask The set of required dependent fields for the output, Indicates the set of dependent fields that are actually retained currently.

[0023] Furthermore, the planning agent further disassembles using the grouped observation slicing method. Let the length of the input sequence to be processed be L, and the benchmark parallel grouping threshold be G. Then the number of standard parallel task groups n and the remainder r are expressed as:

[0024] The i-th slice subtask sequence Is expressed as:

[0025] In equations (9) and (10), n represents the number of standard parallel task groups that can be distributed, and r represents the remainder that cannot be evenly divided, Represents the i-th set of fine-grained observation sequences with strict input and output boundaries. i satisfies 0 ≤ i < n, Inputs represents the input sequence to be disassembled, and G represents the benchmark parallel grouping threshold.

[0026] Furthermore, after each execution agent completes its local task, it does not directly splice and send its own complete context to other nodes. Instead, it encapsulates the execution result and status change as a standard asynchronous communication event, publishes it to the global event manager uniformly, and then the event manager accurately delivers the valid data payload to the downstream nodes according to the routing rules; The standard asynchronous communication event circulated in the system is defined as:

[0027] In equation (11), E represents the standard communication event, ID represents the globally unique event identification serial number, Type represents the discrete type of the communication event, and Source and Target respectively represent the network topology node addresses of the message sender agent and the target receiver agent, Represents the data payload formed after structuring the original output data x of the agent, Represents the absolute timestamp; The routing distribution result of the global event manager for any output event Is expressed as:

[0028] In equation (12), Represents the output event of the source node by the global event manager The routing distribution result, where N represents the set of all cooperative agent nodes in the network, and ψ represents the data feature cleaning and core parameter extraction function. This represents the j-th collaborating node. Represents a node For event types The subscription status, where I(·) represents the indicator function, takes the value of 1 when the target node has pre-subscribed to the communication event type, and takes the value of 0 otherwise; To prevent circular calls or deadlocks during asynchronous collaboration, the system also maintains a global communication activity record mapping table. For any sending node s, receiving node r, and content c, if the current record table already contains... This indicates that the receiving node is waiting for or calling the sending node. In this case, the event control logic is as follows:

[0029] If the current call is valid, the system updates the global active record mapping table:

[0030] In equations (13) and (14), Indicates time The global communication activity record mapping table, whose key-value pairs are used to represent the current call and callee relationship, Error indicates an illegal circular call exception thrown, and Target(r).execute(c) indicates that the call content c is handed over to the target node r for execution; Indicates the timing of event scheduling. This indicates the addition of a new call relationship record, meaning a node. Current processing is handled by node The initiated call.

[0031] The present invention also provides a state-driven multi-agent crop growth report generation system, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the state-driven multi-agent crop growth report generation method as described in the above technical solution.

[0032] Compared with the prior art, the present invention has the following advantages and beneficial effects: This invention proposes a state-driven multi-agent crop growth report generation method, which consists of three core modules: a multi-agent collaborative communication mechanism, a state-driven context reconstruction mechanism, and an event-driven data routing mechanism, forming a complete technical closed loop for automated agricultural remote sensing report generation scenarios. First, at the architecture level, this invention constructs a unified asynchronous event communication topology based on a publish / subscribe model, introducing a global event manager to structure and schedule data interactions between multiple agents. Compared to traditional point-to-point context splicing, this invention effectively avoids the disordered diffusion and redundant superposition of global context among multiple nodes through standardized event encapsulation and precise routing, significantly reducing the risk of information pollution and logical confusion during multi-agent collaboration. This mechanism is particularly suitable for cross-module scenarios involving multi-source remote sensing data, meteorological data, and historical planting data in crop growth report generation. Second, regarding state evolution and self-driven execution mechanisms, this invention uses a state transition function Φ to implement closed-loop control of the agent's lifecycle, enabling the agent to automatically perform state transitions and policy adjustments based on execution feedback and external events. This mechanism enables the crop growth report generation process to have stage-aware capabilities. For example, it automatically enters a supplementary query state when data is missing, and automatically enters a result summary state after analysis is completed, thereby improving the overall system's robustness and fault tolerance. Finally, in terms of data routing and event standardization, this invention achieves precise data delivery and load pruning among multiple agents by constructing a multi-dimensional tuple structure that includes event types, data payload feature extraction results, and a subscription matrix. By using the feature cleaning function ψ to structurally purify the data payload, it is possible to significantly reduce the interference of irrelevant prompts on the downstream agent's decision-making process, prevent global context expansion, and improve the long-term operational stability of the system.

[0033] In summary, this invention achieves precise separation and efficient coordination of global constraints and local details in complex agricultural growth report generation tasks through a state self-driven mechanism and a structured context reconstruction method, demonstrating good engineering practical value and promising prospects for promotion. Attached Figure Description

[0034] Figure 1 This is a flowchart of the method of the present invention; Figure 2 A schematic diagram of a three-dimensional context reconstruction and state self-driving kernel; Figure 3 This is a schematic diagram illustrating the report generation method of the present invention. Detailed Implementation

[0035] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0036] This invention proposes a state-driven multi-agent method for generating crop growth reports. This method innovatively restructures the agent context into three parts: the first part is state-driven information, used to guide the stage switching in the crop growth report generation process, enabling the agent to autonomously manage its current agricultural analysis stage (such as the index analysis stage, anomaly diagnosis stage, or management suggestion generation stage); the second part is global agricultural task prompt information, used to anchor the overall report structure, including global constraints such as crop type, growth stage, monitoring area, and report format; the third part is the state region, used to concentrate the detailed analysis content of the current subtask. For example, in the crop index trend analysis stage, all statistical data, anomaly thresholds, and spatial distribution results related to NDVI changes are dynamically converged in this region and strongly bound to the current state, avoiding interference from information from other stages. Through the dynamic isolation and coordination mechanism of the above three parts of the context, in the complex agricultural application scenario of automatic crop growth report generation, the following can be achieved: precise separation of global agricultural background constraints and local data analysis; structured collaboration among multiple agents; stage-driven advancement of the report generation logic; and stable output of professional agricultural report structures. This effectively solves the problems of context disorder and low efficiency of multi-agent collaboration in existing agricultural intelligent reporting systems.

[0037] Please see Figure 1 The flowchart illustrates a state-driven multi-agent crop growth report generation method provided by this invention, comprising the following steps: Step 1: Initialize the multi-agent collaborative operating environment and communication topology, establish a global event manager, communication pipelines and global sessions to provide a unified foundation for asynchronous collaboration among multiple agents.

[0038] Step 2: Construct the state-driven intelligent agent kernel and reconstruct the working context of each heterogeneous intelligent agent, making it consist of a state management module, a global guidance module, and a dynamic state region.

[0039] Step 3: The planning agent decomposes the crop growth report generation task based on the global context, forming fine-grained subtasks with clear input-output boundaries and dependencies.

[0040] Step 4: After receiving the corresponding subtask, each executing agent completes local closed-loop processing such as retrieval, analysis, diagnosis or reasoning within the dynamic state region, and performs state evolution under execution feedback.

[0041] Step 5: Each executing agent encapsulates the local task results and state changes into standard events, and completes cross-node state synchronization and data routing through the global event manager.

[0042] Step 6: The report generation agent aggregates, logically organizes, and renders the structured results output by each upstream agent to generate a well-formatted and clearly structured crop growth report.

[0043] Furthermore, the specific implementation method of step 1 is as follows: In step 1, the system (the program system running the intelligent agent) first instantiates the global event manager and communication pipeline, and establishes a global session to maintain the execution cycle of this report generation task. The global session is used to record task identifier, crop type, monitoring area, growth period, report template requirements, and common constraint information during execution.

[0044] During the construction of the underlying message routing mechanism, to ensure the uniqueness and traceability of communication routes, the system performs batch name verification on agents to be registered under mutual exclusion control conditions. Let the set of agent objects to be registered be:

[0045] Extract its corresponding set of names and denote it as:

[0046] If and only if the cardinality of the deduplicated name set satisfies:

[0047] The system determines that all agent names are unique and writes each agent instance into the local routing table to complete the address binding required for subsequent event delivery; if equation (3) is not satisfied, it means that at least two agents have the same identifier, and the system triggers an initialization failure exception.

[0048] In equations (1) to (3), A represents the set of registered intelligent agent objects. Name represents the i-th agent instance, m represents the total number of instantiated agents, and Name( ) represents the unique identifier name of the i-th agent. Represents the set consisting of all agent names, | | indicates the cardinality of the set. Through these constraints, the system can eliminate target ambiguity in subsequent event routing during the initialization phase.

[0049] Furthermore, the specific implementation method of step 2 is as follows: In step 2, to avoid mixing global task information and local execution details in the same prompting context, the system implants a state-driven kernel based on a finite state machine into each heterogeneous intelligent agent and reconstructs its working context into three parts: a state management module, a global guidance module, and a dynamic state region. The system comprises several modules: a state management module, a global guidance module, and a dynamic state region. The former is responsible for perceiving the current lifecycle state and state transition conditions, containing necessary information to guide state switching, enabling the agent to autonomously determine its current lifecycle stage (e.g., initialization, running, waiting, completion, error). Each agent executing a specific task is controlled by a state management module, which manages a set of runtime state variables related to the task, the lifecycle of the task from start to finish, and the state switching strategy bound to the task. The latter, the main prompting area, guides and constrains the agent's behavior in automated tasks, including a set of static rules defining task specifications, a clear and explicit list of overall goals, and template constraints for formatted output in the task execution process. The latter, the core working area for fine-grained task execution, dynamically gathers all local detailed instructions related to a subtask when the agent focuses on that subtask, and is tightly bound to the current execution state. The dynamic state region is the agent's context used to store the aforementioned information and define a set of operations for information transformation within that region. The runtime state is ultimately divided into three aspects: subtask instructions, intermediate results of task execution, and feedback increments formed by self-evaluation of task execution.

[0050] During the construction of the dynamic state region, the system first merges the current session state, task dependencies, and additional metadata to form a joint context variable space, the calculation formula of which is:

[0051] In equation (4), This represents the dictionary of merged union context state variables. This represents the current running session state set of the system, including communication information between agents and interaction information between agents and the user. This represents the underlying task execution dependency set, including the constraint set set by the global bootstrapping module and the execution state of each intelligent entity. This represents the additional metadata dataset, which serves as extended signals and includes non-essential, extensible information such as events and execution environment. The system replaces placeholder variables in local execution instructions and performs template evaluation and rendering, thereby constructing a dynamic state workspace bound to the current task stage without introducing irrelevant information.

[0052] Furthermore, the full working context space reconstruction expression of the agent at time t is:

[0053] Its state evolution equation is:

[0054] In equations (5) and (6), Indicates that the intelligent agent is in The full working context space at any given moment; Represents non-overlapping structured splicing operations across different contextual dimensions; A unified representation of the shared context corresponding to the global guidance module, used to store the overall task objective and cross-step shared constraint information. express A separate state-aware region used to guide state machine transitions at all times; express Moment-based intelligent agents centrally process the first The dynamic state area for each sub-task; and These represent the intelligent agent in Time and The lifecycle running status at any given moment; This indicates the received external trigger event stream or execution feedback signal; This represents the self-driven state transition function of the agent kernel, which is used to calculate the state at the next moment based on the current state, external triggering events, and state awareness information. Its implementation depends on a set of predetermined state transition rules and a large language model.

[0055] Furthermore, the specific implementation method of step 3 is as follows: In step 3, the planning agent receives the user's request to generate a crop growth report and, combined with information from the global guidance module such as crop type, growth stage, monitoring area, data source, and report structure requirements, performs a structured analysis of the overall task. The planning agent breaks down the complex task into multiple fine-grained subtasks with clear input and output boundaries and distributes them to the corresponding execution agents according to task type and dependencies.

[0056] Let the global natural language task be... Then the task parsing mapping function inside the planning agent is expressed as:

[0057] In equation (7), The task parsing mapping function within the planning agent can be implemented using a large language model. This represents n fine-grained subtasks decoupled from the global task. These subtasks may include task units required for different execution nodes, such as information retrieval, data processing, growth diagnosis, anomaly analysis, suggestion generation, and report writing.

[0058] To reduce context redundancy between different subtasks while maintaining necessary data dependencies between tasks, the system further constructs a context isolation constraint objective function:

[0059] In equation (8), This represents the target value for the constraint of task parsing and context isolation. Let λ represent the local context assigned to the i-th subtask, and let λ represent the balance coefficient. The function representing the mutual information redundancy calculation between different local contexts is specifically expressed as follows:

[0060] The first item is the field semantic repetition rate, calculated by the embedding model. These are the weighting coefficients. These represent the semantic vector representations of the subtask context.

[0061] Subtasks in multi-agent cooperative networks and The data dependency penalty term is specifically expressed as follows:

[0062] Subtasks Pair Task The set of required dependency fields for the output. This represents the set of dependency fields that are currently retained. All of the above results were obtained by structured extraction from a large language model.

[0063] When the input demand is large, the planning agent further decomposes it using a grouped observation slice approach. Let the length of the input sequence to be processed be L, and the baseline parallel grouping threshold be G. Then the standard number of parallel task groups n and the margin r are expressed as:

[0064] in, This represents the modulo operation, and the sequence of subtasks for the i-th slice. Represented as:

[0065] In equations (9) and (10), n represents the number of groups of standard parallel tasks that can be distributed, and r represents the non - divisible remainder. represents the i - th set of fine - grained observation sequences with strict input - output boundaries, where i satisfies 0 ≤ i < n, Inputs represents the input sequence to be disassembled, and G represents the benchmark parallel grouping threshold.

[0066] Furthermore, the specific implementation method of step 4 is as follows: In step 4, each vertical - domain execution agent is awakened after hearing a sub - task event that matches its subscribed topic, and writes the local input information of the current sub - task into the dynamic state area. During the execution process, the execution agent only retrieves, analyzes, diagnoses, or reasons around the content in this dynamic state area, without directly loading other global details that are irrelevant to the current stage.

[0067] During the task execution, the agent evolves its self - driving state according to the state transition function Φ shown in equation (6). When receiving a new external event or intermediate feedback, the agent automatically transfers to the running, waiting, completed, or error state according to the current state and local execution results, enabling a closed - loop processing flow for a single sub - task. This mechanism enables the execution agent to automatically trigger supplementary queries, re - analysis, or exception reporting when encountering input missing, constraint conflicts, or result inconsistencies, without the need for explicit manual intervention.

[0068] Furthermore, the specific implementation method of step 5 is as follows: In step 5, after each execution agent completes its local task, instead of directly splicing and sending its complete context to other nodes, it encapsulates the execution result and state change as a standard asynchronous communication event and publishes it to the global event manager uniformly. Then, the event manager accurately delivers the valid data payload to the downstream nodes according to the routing rules. The standard asynchronous communication event circulating in the system is defined as:

[0069] In equation (11), E represents the standard communication event, ID represents the globally unique event identification serial number, Type represents the discrete type of the communication event, Source and Target respectively represent the network topology node addresses of the message - sending agent and the target receiving agent. represents the data payload formed after structuring the original output data x of the agent. represents the absolute timestamp.

[0070] The routing and distribution result of the global event manager for any output event is expressed as:

[0071] In equation (12), This indicates that the global event manager outputs events to the source node. The routing distribution result, where N represents the set of all cooperative agent nodes in the network, and ψ represents the data feature cleaning and core parameter extraction function. This represents the j-th collaborating node. Represents a node For event types The subscription status is represented by I(·), which is an indicator function. It takes the value 1 when the target node has pre-subscribed to the communication event type, and otherwise takes the value 0. Represents the raw output data of the source agent.

[0072] To prevent circular calls or deadlocks during asynchronous collaboration, the system also maintains a global communication activity record mapping table. For any sending node s, receiving node r, and content c, if the current record already exists... This indicates that the receiving node is waiting for or calling the sending node. In this case, the event control logic is as follows:

[0073] If the current call is valid, the system updates the global active record mapping table:

[0074] In equations (13) and (14), Indicates time The global communication activity record mapping table uses key-value pairs to represent the current caller-caller relationship. Error indicates an illegal circular call exception thrown, and Target(r).execute(c) means that the call content c is handed over to the target node r for execution. Indicates the timing of event scheduling. This indicates the addition of a new call relationship record, meaning a node. Current processing is handled by node The initiated call.

[0075] Furthermore, the specific implementation method of step 6 is as follows: In step 6, the report generation agent continuously listens for task completion events from the retrieval agent, analysis agent, diagnostic agent, and other upstream nodes, and aggregates, cleans, and logically arranges the received structured results. The structured results include basic information about the monitoring area, crop growth index analysis results, abnormal plot identification results, abnormal cause diagnosis results, comprehensive growth assessment results, and management recommendations.

[0076] Within its own dynamic state region, the report-generating agent maps the payloads output by each upstream node according to a preset document template and completes paragraph rendering in conjunction with the typesetting specifications. Since the upstream nodes output structured results filtered by Equations (11) and (12) rather than lengthy execution contexts, the report generation stage can complete content integration under low noise conditions, thereby ensuring the logical consistency and professionalism of the final report.

[0077] A schematic diagram illustrating the construction of a multi-agent interaction system and the generation of crop growth analysis reports using the method of this invention is shown below. Figure 3 As shown, after inputting image data, the multi-agent interaction effectively generates a multi-dimensional growth analysis report, effectively solving the problems of context disorder and low efficiency of multi-agent collaboration in existing agricultural intelligent reporting systems.

[0078] On the other hand, embodiments of the present invention also provide a state-driven multi-agent crop growth report generation system, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the state-driven multi-agent crop growth report generation method as described in the above technical solution.

[0079] It should be understood that any parts not described in detail in this specification belong to the prior art.

[0080] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A state-driven multi-agent crop growth report generation method, characterized in that, Includes the following steps: Initialize the multi-agent collaborative operating environment and communication topology, establish a global event manager, communication pipelines and global sessions to provide a unified foundation for asynchronous collaboration among multiple agents; A state-driven intelligent agent kernel is constructed, and the working context of each heterogeneous intelligent agent is reconstructed. The intelligent agent context is divided into a state management module, a global guidance module, and a dynamic state region. The planning agent decomposes the crop growth report generation task based on the global context, forming fine-grained subtasks with clear input-output boundaries and dependencies; After receiving the corresponding subtask, each executing agent completes local closed-loop processing of retrieval, analysis, diagnosis or reasoning within the dynamic state region, and undergoes state evolution under execution feedback; Each executing agent encapsulates local task results and state changes into standard events, and completes cross-node state synchronization and data routing through the global event manager; The report-generating agent aggregates, logically organizes, and renders the structured results output by each upstream agent to generate a well-formatted and clearly structured crop growth report.

2. The state-driven multi-agent crop growth report generation method according to claim 1, characterized in that: First, a global event manager and communication pipeline are instantiated, and an underlying message routing mechanism is built to establish a global session for maintaining the execution cycle of this report generation task. The global session is used to record task identifier, crop type, monitoring area, growth period, report template requirements, and common constraint information during the execution process. During the construction of the underlying message routing mechanism, to ensure the uniqueness and traceability of communication routes, batch name verification is performed on the agents to be registered under mutual exclusion control conditions; let the set of agent objects to be registered be: Extract its corresponding set of names and denote it as: If and only if the cardinality of the deduplicated name set satisfies: The system determines that all agent names are unique and writes each agent instance into the local routing table to complete the address binding required for subsequent event delivery; if equation (3) is not satisfied, it means that at least two agents have the same identifier, and the system triggers an initialization failure exception. In equations (1) to (3), A represents the set of registered intelligent agent objects. Name represents the i-th agent instance, m represents the total number of instantiated agents, and Name( ) represents the unique identifier name of the i-th agent. Represents the set consisting of all agent names, | | indicates the cardinality of the set.

3. The state-driven multi-agent crop growth report generation method according to claim 1, characterized in that: The state management module is used to perceive the current lifecycle state and state transition conditions, including the necessary perception information to guide state switching, enabling the agent to autonomously determine its current lifecycle. Each agent executing a specific task is controlled by a state management module, which controls a set of runtime state variables related to the specific task, the lifecycle of the specific task execution from start to finish, and the state switching strategy bound to the task. The global guidance module, also known as the main prompt word area, is used to guide and constrain the agent's behavior boundaries in automated tasks. It includes a set of static rules defining task specifications, a clear and explicit list of overall goals, and template constraints for formatted output in the task execution process. The dynamic state area serves as the core working area for fine-grained task execution. When the agent focuses on processing a sub-task, all local detailed instructions for that sub-task are dynamically gathered here and are tightly bound to the current execution state. The dynamic state area is the agent's context used to store the above information and define a set of operations for information transformation within the area. The runtime state is ultimately divided into three aspects: sub-task instructions, intermediate results of task execution, and feedback increments formed by self-evaluation of task execution. During the construction of the dynamic state region, the system first merges the current session state, task dependencies, and additional metadata to form a joint context variable space, the calculation formula of which is: In equation (4), This represents the dictionary of merged union context state variables. This represents the current set of running session states of the system. This represents the underlying task execution dependency set. This indicates an additional meta-dataset.

4. The state-driven multi-agent crop growth report generation method according to claim 1, characterized in that: The full working context space reconstruction expression of the agent at time t is: Its state evolution equation is: In equations (5) and (6), Indicates that the intelligent agent is in The full working context space at any given moment; Represents non-overlapping structured splicing operations across different contextual dimensions; This represents a static, global guidance area that characterizes the overall requirements of the task. express A separate state-aware region used to guide state machine transitions at all times; express Moment-based intelligent agents centrally process the first The dynamic state area for each sub-task; and These represent the intelligent agent in Time and The lifecycle running status at any given moment; This indicates the received external trigger event stream or execution feedback signal; This represents the self-driven state transition function of the agent kernel.

5. The state-driven multi-agent crop growth report generation method according to claim 1, characterized in that: Let the global natural language task be... The context of the global prompt is Then the task parsing mapping function inside the planning agent is expressed as: In equation (7), This represents the task parsing mapping function within the planning agent. This represents n fine-grained subtasks decoupled from the global task; the subtasks include task units required for different execution nodes such as information retrieval, data processing, growth diagnosis, anomaly analysis, suggestion generation, and report writing.

6. The state-driven multi-agent crop growth report generation method according to claim 1, characterized in that: After task decomposition, the context isolation constraint objective function is further constructed: In equation (8), This represents the target value for the constraint of task parsing and context isolation. This represents the local context assigned to the i-th subtask. This represents the local context assigned to the j-th subtask. This function represents the calculation function for mutual information redundancy between different local contexts. Subtasks in multi-agent cooperative networks and The data dependency penalty term between them, where λ represents the balance coefficient.

7. The state-driven multi-agent crop growth report generation method according to claim 6, characterized in that: The specific expression of the mutual information redundancy calculation function is as follows: The first item is the semantic repetition rate of the field. These are the weighting coefficients. These are semantic vector representations of the local context of the i-th and j-th subtasks, respectively. The data dependency penalty is specifically represented as follows: Subtasks Pair Task The set of required dependency fields for the output. This represents the set of dependency fields that are currently retained.

8. The state-driven multi-agent crop growth report generation method according to claim 1, characterized in that: The planning agent is further decomposed using a grouped observation slice method. Let the length of the input sequence to be processed be L, and the baseline parallel grouping threshold be G. Then the standard number of parallel task groups n and the margin r are expressed as: The i-th slice subtask sequence Represented as: In equations (9) and (10), n represents the number of groups of standard parallel tasks that can be distributed, and r represents the non - divisible remainder. represents the i - th set of fine - granularity observation sequences with strict input - output boundaries, where i satisfies 0 ≤ i < n, Inputs represents the input sequence to be disassembled, and G represents the benchmark parallel grouping threshold.

9. The state-driven multi-agent crop growth report generation method according to claim 1, characterized in that: After completing a local task, each agent does not directly send its complete context to other nodes. Instead, it encapsulates the execution result and state changes into standard asynchronous communication events, which are then uniformly published to the global event manager. The event manager then accurately delivers the valid data payload to downstream nodes according to routing rules. The standard asynchronous communication events circulating within the system are defined as follows: In Equation (11), E represents a standard communication event, ID represents a globally unique event identifier sequence number, Type represents the discrete type of the communication event, and Source and Target represent the network topology node addresses of the message sender and the target receiver, respectively. This represents the data payload formed after structured extraction of the agent's original output data x. Represents an absolute timestamp; The global event manager handles any output event. The routing distribution result is represented as follows: In equation (12), This indicates that the global event manager outputs events to the source node. The routing distribution result, where N represents the set of all cooperative agent nodes in the network, and ψ represents the data feature cleaning and core parameter extraction function. This represents the j-th collaborating node. Represents a node For event types The subscription status, where I(·) represents the indicator function, takes the value of 1 when the target node has pre-subscribed to the communication event type, and takes the value of 0 otherwise; To prevent circular calls or deadlocks during asynchronous collaboration, the system also maintains a global communication activity record mapping table. For any sending node s, receiving node r, and content c, if the current record table already contains... This indicates that the receiving node is waiting for or calling the sending node. In this case, the event control logic is as follows: If the current call is valid, the system updates the global active record mapping table: In equations (13) and (14), Indicates time The global communication activity record mapping table, whose key-value pairs are used to represent the current call and callee relationship, Error indicates an illegal circular call exception thrown, and Target(r).execute(c) indicates that the call content c is handed over to the target node r for execution; Indicates the timing of event scheduling. This indicates the addition of a new call relationship record, meaning a node. Current processing is handled by node The initiated call.

10. A state-driven multi-agent crop growth report generation system, characterized in that: It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the program instructions in the memory to execute the state-driven multi-agent crop growth report generation method as described in any one of claims 1-8.

Images