Multi-agent space cooperative treatment method and system

Abstract

The invention relates to the technical field of space governance, and discloses a multi-agent space collaborative governance method which comprises the following steps: processing multi-source heterogeneous data such as satellite images and sensor readings, establishing cross-type semantic association through a geographic space data embedding technology, generating a unified structured text after optimizing the satellite images through vLLM, and synchronizing the unified structured text to a central database; an agent role portrait is dynamically generated by the central server large language model based on a preset Prompt template, and generation does not depend on a fixed rule; then, selecting a target node in the edge-center architecture, disassembling a total task into sub-tasks, establishing semantic mapping, calculating a matching probability, and performing optimal distribution by a reward borrowing function; generating a governance scheme in a perception layer-decision layer-execution layer framework, and outputting a coded operation instruction; based on an execution feedback updating strategy, a multi-level mechanism is set, roles are automatically redistributed, and the governance continuity is guaranteed. According to the invention, the overall efficiency and reliability of space governance can be improved in the face of dynamic scenes or emergency situations.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of space governance, in particular to a multi-agent space collaborative governance method and system. BACKGROUND

[0002] In the field of space governance, satellite images, sensor real-time readings, policy texts and geographic coordinates and other types of information are needed to carry out work, and these information sources are scattered and have great differences in format, which are the core basis for supporting governance decision-making and agent scheduling. When dealing with such multi-source heterogeneous information, the current technology often lacks effective correlation means, making it difficult to establish semantic connections between different types of information, resulting in low information integration efficiency, and the generated governance foundation data is difficult to meet the unified and standardized application requirements.

[0003] At the same time, the environmental state and governance needs in the space governance scene are prone to dynamic changes, for example, in the event of sudden disasters or abnormal regional environmental parameters, agents need to quickly adapt to scene adjustments roles and tasks; but in the existing method, when facing dynamic scenes or emergency situations, response lags are prone to occur, affecting the overall efficiency and reliability of space governance.

[0004] From the above, how to improve the overall efficiency and reliability of space governance when facing dynamic scenes or emergency situations still needs to be solved. SUMMARY

[0005] In order to be able to improve the overall efficiency and reliability of space governance when facing dynamic scenes or emergency situations, the application provides a multi-agent space collaborative governance method and system.

[0006] In a first aspect, the application provides a multi-agent space collaborative governance method, which adopts the following technical solution: A multi-agent space collaborative governance method, comprising: Processing multi-source heterogeneous data composed of satellite images, sensor real-time readings, policy texts and geographic coordinate data, establishing corresponding cross-type semantic association based on geographic spatial data embedding technology, satellite images are optimized by vLLM dynamic memory management, geographic coordinates are associated with regional environmental parameters and governance needs to generate spatial attribute semantic expression, generate unified structured text, and transmit to the central database through the distributed consistency synchronization mechanism; Based on the spatial environment information, governance target constraints and agent capability characteristics in the structured text, when the environment changes or the demand adjusts, the central server large language model generates an agent role image including name, skill, knowledge boundary and spatial work range based on a preset Prompt template, and generates an agent role image that does not depend on fixed rules, wherein the knowledge boundary includes policy and regulation adaptation instructions, and the spatial work range includes associated geographic coordinates. Obtain the total task information of the role image and spatial governance, and in the edge-center collaborative architecture, select the target computing node based on the total task characteristics combined with the role semantic attributes; the target node uses the Divide-and-Conquer algorithm to split the total task information into subtasks containing dependency relationships, establishes a semantic mapping between the subtasks and the role image, calculates the matching probability, and optimizes the distribution rules according to the scene to obtain the corresponding splitting and distribution results; In the perception layer-decision layer-execution layer architecture, based on the splitting and distribution results, generate the corresponding multi-dimensional governance scheme combined with real-time spatial data, determine the agent task set and output the governance operation instructions, wherein the perception layer transmits the real-time data to the decision layer after being processed by the edge node, the decision layer integrates the agent suggestions based on the large language model and processes the conflicts according to the semantic priority, and the execution layer encodes the geographic metadata and instructions based on the spatial perception protocol and transmits them to the corresponding agent; Obtain the execution feedback information of the agent, update the role generation strategy and distribution rules of the large language model based on the execution feedback information; set up a multi-level trigger mechanism, wherein the sensor data threshold value starting condition trigger and the sudden disaster event trigger are automatically redistributed.

[0007] Optionally, in the process of processing the multi-source heterogeneous information, the method further comprises: The sensor real-time readings are first aligned with the time sequence records by the Kalman filtering algorithm for outlier rejection; Before the vLLM dynamic memory management optimization, a multi-band image fusion technology is used to enhance the spatial environment feature extraction accuracy of the satellite image; Based on the enhanced satellite image and the aligned sensor readings, a semantic association is established to reduce the error rate of the spatial environment information in the generated structured text.

[0008] Optionally, in the process of generating the agent role image, the method further comprises: The pre-set Prompt template further includes role capability evaluation indicators, wherein the role capability evaluation indicators include spatial governance task proficiency, policy and regulation matching degree, and cross-regional collaboration response speed; After the center server large language model generates the role image, it will also combine the real-time regional environment change information transmitted back by the perception layer to update the role capability evaluation indicators once a time, and dynamically adjust the skill weight and spatial working range boundary in the role image.

[0009] Optionally, in the process of obtaining the corresponding splitting and distribution results, the method further comprises: Based on the geographical coordinate associated region of the subtask, add a subtask spatiotemporal constraint determination step; If the geographical range of adjacent sub-tasks overlaps and the execution time conflicts, the sub-task with high governance demand urgency is retained preferentially and the execution time of another sub-task is adjusted; When calculating the matching probability, in addition to the comprehensive sub-task query, the agent capability and the historical allocation result, the agent real-time task load rate is introduced, so that the matching probability calculation result is more consistent with the actual bearing capacity of the agent.

[0010] Optionally, the method further comprises: Before integrating the agent suggestions, the decision layer calls a pre-built spatial governance scenario knowledge base, which includes historical governance cases and corresponding scheme effect records; By comparing the real-time spatial information with the historical information of similar scenarios in the knowledge base through a large language model, a corresponding scheme feasibility evaluation report is generated; When the pass rate of the scheme in the evaluation report is ≥85%, the corresponding final governance scheme is generated based on the agent suggestions, and before the execution layer outputs the governance operation instruction, the geographic metadata in the instruction is checked for coordinates.

[0011] Optionally, during the automatic role redistribution process, the method further comprises: The feedback information is classified according to spatial governance type and regional complexity, including ecological monitoring, disaster emergency, urban control, and regional complexity including core urban area, remote mountainous area, etc. The weight initial matrix is preset to replace the fixed dimension weighting method; The feedback information is semantically reflected by a large language model to obtain corresponding reflection results, which are stored as scenario memories and injected into subsequent prompts. Based on the scenario memories, the weight matrix is dynamically corrected. If the role adaptation degree of the feedback information is less than the threshold for three consecutive times, the tool evolution process is automatically started. The large language model generates tool invocation strategies based on governance needs, and updates the skill boundaries of the role portrait.

[0012] Optionally, in addition to the condition trigger and the event trigger, a capability decay trigger is added, and the method further comprises: Based on the scenario memory, if the success rate of the agent decreases by more than 30% with each task round, it is determined that the capability is decaying, and the role redistribution is automatically triggered. The resource threshold trigger uses a dynamic calculation model to adjust in real time according to the current governance task urgency; When the extreme scenario starts the standby agent, the historical optimal role-task matching trajectory in the scenario memory is migrated to the standby agent, and the role parameters are fine-tuned in combination with the real-time spatial information of the perception layer. At the same time, the lightweight model snapshot of the core role is pre-stored in the edge node, and the standby agent can directly call the snapshot to complete the initialization when it is started, shortening the switching delay to seconds; After the standby agent runs, it performs capability coordination verification with other node agents through a space awareness protocol to ensure that its skills cover the current unfinished sub-tasks.

[0013] In a second aspect, the application provides a multi-agent space coordination governance system, which adopts the following technical solution: A multi-agent space coordination governance system comprises: A multi-source heterogeneous data processing and storage module processes multi-source heterogeneous data composed of satellite images, sensor real-time readings, policy texts, and geographic coordinate data, establishes corresponding cross-type semantic associations based on geographic space data embedding technology, optimizes satellite images through vLLM dynamic memory management, associates regional environmental parameters with governance requirements to generate spatial attribute semantic expressions, generates a unified structured text, and transmits it to the central database through a distributed consistency synchronization mechanism; A role image dynamic generation module dynamically generates an agent role image including name, skill, knowledge boundary, and spatial working range based on the space environment information in the structured text, governance target constraints, and agent capability characteristics, and generates an image that does not depend on fixed rules when the environment changes or the demand adjusts, wherein the knowledge boundary includes policy and regulation adaptation instructions, and the spatial working range includes associated geographic coordinates; A total task decomposition and intelligent distribution module obtains total task information of the role image and space governance, selects a target computing node based on total task characteristics and role semantic attributes in an edge-center collaborative architecture; the target node splits the total task information into sub-tasks with dependencies using a Divide-and-Conquer algorithm, establishes semantic mapping between the sub-tasks and the role image, calculates matching probabilities, and constructs a reward function according to the scene to optimize the distribution rules and obtain the corresponding decomposition and distribution results; A governance scheme generation and instruction output module generates a corresponding multi-dimensional governance scheme based on the decomposition and distribution results in combination with real-time spatial data in a perception layer-decision layer-execution layer architecture, determines an agent task set and outputs governance operation instructions, wherein the perception layer transmits real-time data to the decision layer after processing by the edge node, the decision layer integrates agent suggestions based on a large language model and processes conflicts according to semantic priority, and the execution layer encodes geographic metadata and instructions based on a space awareness protocol and transmits them to the corresponding agent; A feedback optimization and multi-level trigger guarantee module obtains execution feedback information of the agent, updates the role generation strategy and distribution rules of the large language model based on the execution feedback information, and sets up a multi-level trigger mechanism, wherein the sensor data threshold value start condition trigger and the sudden disaster event trigger are automatically redistributed.

[0014] In a third aspect, the present application provides a multi-agent space collaborative governance system, which adopts the technical scheme as follows: A multi-agent space collaborative governance system, comprising a processor, wherein the processor runs a program of the multi-agent space collaborative governance method according to any one of the above.

[0015] In a fourth aspect, the present application provides a storage medium, which adopts the technical scheme as follows: A storage medium, which stores a program of the multi-agent space collaborative governance method according to any one of the above.

[0016] In summary, the present application includes at least one of the following beneficial technical effects: In a dynamic scene, the multi-source heterogeneous data processing link is subjected to outlier rejection, image feature enhancement and vLLM dynamic memory management, which not only ensures the accuracy of the spatial environment information but also speeds up the data processing speed, and in combination with a distributed synchronization mechanism, real-time data support is provided for subsequent decision-making; the agent role portrait can update the ability index and working range based on the regional environment changes, and when the task is allocated, the time and space constraint judgment is used to avoid task conflict, the real-time load rate is introduced to optimize the matching accuracy, the decision layer quickly generates a highly feasible scheme with the help of the scene knowledge base, and the dynamic adaptation capability of each link greatly reduces resource waste and decision delay, thereby realizing efficient connection of the governance process.

[0017] In the face of emergency situations, a multi-level triggering mechanism (sensor threshold, sudden disaster, agent capability attenuation) can capture abnormalities in real time and automatically start role redistribution to avoid governance gaps; in extreme scenarios, backup agents rely on historical optimal matching trajectories and lightweight model snapshots to achieve second-level startup and parameter fine-tuning, while cross-node capability collaboration verification ensures skill coverage of incomplete tasks, in combination with feedback-driven agent tool evolution, effectively dealing with possible capability gaps in the emergency process, ensuring continuous and stable operation of the governance system under sudden conditions, and significantly reducing the risk of governance failure caused by system interruption or insufficient capacity. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flowchart of a multi-agent space collaborative governance method according to an exemplary embodiment.

[0019] Figure 2 is a structural block diagram of a multi-agent space collaborative governance system according to an exemplary embodiment. DETAILED DESCRIPTION

[0020] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0021] In the description of the present specification, the description referring to the terms "certain embodiments", "one embodiment", "some embodiments", "illustrative embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the described embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0022] The embodiments of the present application disclose a multi-agent space collaborative governance method, referring to Figure 1 , comprising: S100, processing multi-source heterogeneous data composed of satellite images, sensor real-time readings, policy texts and geographic coordinate data, establishing corresponding cross-type semantic association based on geographic spatial data embedding technology, satellite images are optimized by vLLM dynamic memory management, geographic coordinate associated regional environment parameters and governance demand generate spatial attribute semantic expression, generate unified structured text, through distributed consistency synchronization mechanism to the central database.

[0023] Among them, S100 specifically includes the following steps: Step 1, classification and initial arrangement of multi-source heterogeneous data: In the initial stage of S100, first of all, the four types of core multi-source heterogeneous data of satellite images, sensor real-time readings, policy texts and geographic coordinate data are classified and sorted. Because the formats, dimensions and information densities of different types of data are significantly different, for example, satellite images are high-dimensional visual data, sensor readings are time series numerical data, policy texts are unstructured text data, and geographic coordinates are spatial position data, independent preprocessing channels need to be established according to data types first, and obvious invalid information (such as blank readings of sensors offline, extreme noise area in satellite images) is removed, which lays a foundation for subsequent targeted processing and avoids processing deviation caused by mixed data.

[0024] Step 2, vLLM dynamic memory management optimization of satellite images: To solve the problem of large amount of satellite image data and low efficiency of processing, the vLLM dynamic memory management technology is used for optimization. This technology dynamically adjusts the memory allocation mechanism of large language model (LLM) when loading satellite image data. According to the resolution, band number and regional coverage of the image, the memory resources are allocated adaptively: for high-resolution, multi-band images of key areas, the memory supply is prioritized to ensure that the detailed information is not lost; for low-priority background area images, the memory reuse mechanism is used to reduce resource occupation. At the same time, the pre-compilation and batch processing capability of vLLM can improve the feature extraction speed of satellite images, avoiding the impact of image processing delay on the timeliness of the overall management process.

[0025] Step 3: Geographical coordinate association and spatial attribute semantic expression generation: Taking geographical coordinate data as the core link, it is deeply associated with regional environmental parameters and management needs to generate spatial attribute semantic expression. Specifically, first, through geocoding technology, discrete geographical coordinates are mapped to actual management areas (such as streets, watersheds, and ecological protection areas), and the specific spatial range corresponding to each coordinate point is determined; then the environmental parameters (such as air quality index, water quality indicators, and vegetation coverage) of the area are retrieved, combined with the current stage of management needs (such as pollution control, ecological restoration, and disaster prevention), and the association of "coordinate position + environmental state + management target" is converted into a structured semantic description. For example, in a certain river basin, the current water quality COD value is 45 mg / L, and the management demand is to reduce the COD value to below 30 mg / L, making the originally single coordinate data have practical significance in the context of management.

[0026] Step 4: Cross-type semantic association establishment (based on geographical spatial data embedding technology): With the help of geographical spatial data embedding technology, the semantic barriers between different types of data are broken down, and a unified cross-type semantic association is established. This technology maps various types of data to the same geographical spatial semantic dimension: for satellite images, their pixel-level spatial features are extracted and converted into semantic vectors (such as "image feature vector of vegetation coverage area in a certain region"); for sensor real-time readings, combined with the geographical coordinates of their collection locations, "location + time series value" semantic vectors are generated (such as "PM2.5 value vector of sensor at 30° N XX' for 24 hours"); for policy texts, keywords related to spatial governance (such as "watershed non-discharge range" and "ecological red line area") are extracted and associated with corresponding geographical coordinates to generate "policy provisions + spatial range" semantic vectors. Through vector similarity calculation, different types of data form semantic associations around the same geographical spatial scene, such as automatically associating "abnormal water area" in satellite images with "water quality exceeding standard readings" of sensors and "water governance provisions" in policy texts, realizing the fusion of data information.

[0027] Step 5, Unified Structured Text Generation and Distributed Synchronous Storage: Based on the above processing results, a unified structured text is generated. The structured text needs to include four core fields: "data source type, geographic spatial range, core information content, semantic association identifier". For example, "Data source: satellite image; Geospatial range: North Latitude - , East Longitude - ; Core information: There are 2 vegetation withering areas in this area; Semantic association identifier: Associated sensor readings (number S001) show that soil humidity has been below the threshold for nearly 7 days, and associated policy text (number P005) is 'Vegetation Protection and Repair Ordinance'". After completing the generation of structured text, the text is transmitted to the central database through a distributed consistency synchronization mechanism: this mechanism ensures the accuracy of text transmission through multi-node data verification (edge node and central node bidirectional check), while using incremental synchronization strategy (only transmitting updated or new text data) to reduce bandwidth occupation, ensuring that the central database can obtain unified multi-source data processing results in real time and reliably, avoiding the deviation of subsequent decision-making caused by asynchronous data. Through classification and targeted optimization (such as satellite image vLLM memory management), the efficiency and accuracy of data processing are improved; through geographic coordinate association and cross-type semantic association, data barriers are broken down, and information fusion is achieved; through unified structured text and distributed synchronization, a solid data foundation is laid for subsequent S200 intelligent agent role portrait generation (providing accurate spatial environment information and governance demand data), S300 task decomposition and allocation (providing unified data semantic logic), and S400 governance scheme generation (providing real-time spatial data support), ensuring the accuracy, real-time and collaboration of the entire spatial governance process from the source, avoiding governance decision-making errors or inefficiencies caused by data problems. S200, based on the spatial environment information in the structured text, governance target constraints, and intelligent agent capability characteristics, when the environment changes or demand adjusts, the central server large language model dynamically generates an intelligent agent role portrait including name, skill, knowledge boundary, and spatial working range based on the pre-set Prompt template. The generated portrait does not rely on fixed rules, and the knowledge boundary includes policy and regulation adaptation instructions, and the spatial working range includes associated geographic coordinates.

[0028] Specifically, S200 includes the following steps: Step 1, extraction and integration of input information: In the initial stage of S200, first extract the core information related to role image generation from the unified structured text generated in S100, including spatial environment information, governance target constraints, and combining with the basic ability characteristics of the agent to form a complete input data set.

[0029] Among them, the spatial environment information is obtained from the "geospatial range" and "core information content" fields of the structured text, such as the vegetation coverage state and water quality of a certain area; the governance target constraint corresponds to the governance demand in the structured text, such as "improve the water quality of a certain river basin to the standard level" and "control the dust pollution in a certain area"; the basic ability characteristics of the agent include the existing hardware support of the agent (such as data processing rate, communication coverage), existing functional modules (such as environment monitoring module, instruction execution module), etc. Through the classification extraction and cross verification of the three types of information, it is ensured that the input data can accurately reflect the actual demand of the current governance scene and the ability basis of the agent, avoiding the disconnection between the image and the scene due to information missing.

[0030] Step 2, dynamic generation of trigger condition judgment and model calling: Based on the extracted input information, real-time judgment of "environmental change" or "demand adjustment" is made, and then the central server large language model is started. For the judgment of "environmental change", by comparing the current extracted spatial environment information with the historical same period or the structured text information of the previous period, if the key environmental indicators (such as the PM2.5 concentration in a certain area, the dissolved oxygen content of water body) change by more than the preset threshold, it is judged as environmental change; for the judgment of "demand adjustment", according to the update of the governance target constraint, such as the original target is "short-term disaster emergency monitoring", and now it is adjusted to "long-term ecological restoration tracking", the generation process is triggered. When any trigger condition is met, the central server large language model is automatically called, and the real-time data interaction channel between the model and the input data set is established, providing computing support for subsequent image generation.

[0031] Step 3, parameter injection of preset Prompt template: After calling the large language model, the preset Prompt template is used as the generation framework, and the input information extracted in step 1 is parameterized according to the template structure. The preset Prompt template includes four core modules: "scene positioning", "target matching", "ability adaptation", and "space binding". In the "scene positioning" module, the spatial environment information is injected to clearly define the governance area and environmental background corresponding to the image. In the "target matching" module, the governance target constraints are injected to define the specific governance tasks supported by the image. In the "ability adaptation" module, the agent's basic ability characteristics are injected to ensure that the generated skills are compatible with the agent's hardware and existing modules. In the "space binding" module, the geographic coordinate association interface is reserved to prepare for subsequent space work range generation. Through parameter injection, the large language model can generate images around specific scenarios, avoiding non-targeted generalization output, while ensuring the compatibility of the image with actual governance needs and agent capabilities.

[0032] Step 4: Dynamic generation of role image core elements: Based on the injected parameter Prompt template, the large language model dynamically generates the four core elements of the role image, and the generation process does not rely on fixed rules. Among them, "name" is generated by combining scene positioning and target matching results, and needs to reflect the association between governance function and area, such as "certain river basin water quality monitoring agent" and "certain urban dust control agent". "Skills" are dynamically matched according to governance target constraints and agent capability characteristics, such as "water quality index real-time analysis" and "governance measure effect feedback" for the "water quality improvement" target, rather than using a fixed skill list. "Knowledge boundary" includes policy and regulation adaptation instructions, which clearly define the policy guidance and operation standards that the agent needs to follow when conducting governance in the corresponding area, such as "adapt to regional water resources protection related policy requirements, and clearly define the compliance operation standards for water intake and drainage links". "Space work range" is generated by associating the geographic coordinates in the structured text, accurately defining the agent's governance coverage area, such as "certain river basin range in North Latitude - , East Longitude - ", ensuring that the agent only performs tasks within the specified space, avoiding cross-regional resource waste.

[0033] Step 5: Verification and associated storage of role image: After the role image is generated, multi-dimensional verification is needed to ensure the completeness and adaptability of the elements, to check whether the "skills" cover the core needs of the governance target, whether the "knowledge boundary" is consistent with the regional policy orientation, and whether the "spatial working range" is accurately matched with the geographic coordinates. If there is a deviation, return to the large language model to adjust again; after the verification, associate the role image with the structured text of S100 (through geographic coordinate indexing binding), and store it in the image exclusive module of the center database. When storing, add a dynamic update identifier to the image, record the generation time and trigger condition type, so that subsequent monitoring of new environmental changes or demand adjustments can quickly retrieve historical images for iterative optimization, while ensuring that S300 can directly associate the corresponding spatial environment information and governance target when obtaining the role image, reducing data query time.

[0034] By dynamically triggering conditions, the image can be updated in real time with environmental and demand changes, avoiding "capability disconnection" of fixed images after scenario changes; through Prompt template parameter injection and core element dynamic generation, it ensures the high adaptability of the image to governance needs and agent capabilities, and improves scenario adaptability through the "independent of fixed rules" feature, especially in complex spatial governance scenarios, it can generate more targeted role functions; through verification and associated storage, it ensures that the image quality is reliable and easy to call later. Overall, S200, as the core link connecting "data processing" and "task allocation", provides accurate "agent capability templates" for the total task decomposition and intelligent allocation of S300, and through the binding of spatial working range and geographic coordinates, the adaptation of knowledge boundary and policy orientation, the task execution of the agent in the subsequent governance process is more spatially accurate and compliant, ensuring the efficiency and reliability of spatial governance from the perspective of "agent adaptability", avoiding waste of governance resources or task execution deviation due to ambiguous agent roles and mismatched capabilities.

[0035] S300, obtains the role image and the total task information of spatial governance, and selects the target computing node based on the total task characteristics and the semantic attributes of the role in the edge-center collaborative architecture; the target node uses the Divide-and-Conquer algorithm to split the total task information into subtasks with dependencies, establishes semantic mapping between the subtasks and the role image, calculates the matching probability, and constructs a reward function according to the scene to optimize the allocation rules and obtain the corresponding splitting and allocation results.

[0036] S300 specifically includes the following steps: Step 1, role image and total task information integration acquisition: First, the agent role image set generated in S200 is retrieved from the central database, and the total task information issued by the space governance system is received. The role image needs to include the skill type, knowledge boundary, spatial work range (associated geographic coordinates), and real-time load status of the agent. The total task information needs to clearly define the governance target (such as ecological monitoring in a certain area, environmental remediation after a disaster), spatial coverage, execution time limit, and core requirements (such as data accuracy, response speed). Preliminary verification is performed on the two types of information to ensure that the role image has no information missing and the total task parameters have no logical contradictions (such as avoiding the total task spatial range exceeding the work range of all roles). This provides complete and accurate basic information for subsequent node selection and task decomposition, avoiding process stalls caused by incomplete information. Step 2, target computing node selection under edge-center collaborative architecture: Based on the resource distribution characteristics of the edge-center collaborative architecture, the total task characteristics, and the role semantic attributes, the target computing node is determined. The total task characteristics focus on three aspects: one is the urgency (such as high urgency for sudden disaster governance, and regular urgency for daily environmental monitoring), two is the spatial coverage (such as cross-city regional tasks requiring central node coordination, and single street tasks can be handled by edge nodes), and three is the computational complexity (such as multi-region data correlation analysis requiring high-power nodes, and simple data statistics can be handled by edge lightweight nodes). The role semantic attributes focus on skill matching degree (such as when the total task contains image recognition requirements, preferentially associate roles with image analysis skills), and spatial work range overlap degree (such as when the role work range and total task area overlap degree exceeds 80%, preferentially consider the edge node where it is located). The finally selected node needs to meet the requirements of "total task processing capacity adaptation" and "role collaboration convenience", for example, high-urgency cross-regional tasks select central nodes, and regular local tasks select edge nodes, achieving efficient utilization of architecture resources. Step 3, Divide-and-Conquer algorithm to realize total task decomposition: The target computing node invokes the Divide-and-Conquer algorithm to break down the overall task into subtasks with dependencies, following a logic of "first splitting large modules, then refining sub-modules." First, the overall task is decomposed into several independent core modules. For example, the "overall task of watershed management" is broken down into three modules: "watershed water quality monitoring," "pollution source location," and "management measure implementation." Then, for each core module, subtasks are further broken down based on the role's spatial work scope. For instance, the "watershed water quality monitoring" module is broken down by geographical coordinates into "upstream monitoring subtasks," "midstream monitoring subtasks," and "downstream monitoring subtasks." The key is to clearly define the dependencies between subtasks: for example, the "pollution source location" subtask must be completed before the "management measure implementation" subtask can be initiated; the "upstream monitoring subtask" must be completed first, and then the midstream and downstream monitoring subtasks can be implemented simultaneously to avoid management failures due to disordered task order. Simultaneously, the decomposition process must ensure that the spatial scope of the subtasks does not overlap, the time nodes do not conflict, and the complexity of each subtask is appropriate for the agent's processing capabilities. Step 4: Semantic mapping and matching probability calculation between subtasks and character profiles: Based on the core requirements of subtasks and the semantic attributes of role profiles, a bidirectional semantic mapping relationship is established. For example, the requirement for the "upstream water quality monitoring subtask" is "possessing the ability to analyze water quality parameters and having a work scope covering the upstream of the basin." This corresponds to a role possessing "water quality analysis skills" and whose spatial work scope includes "upstream coordinates of the basin," forming a semantic association of "subtask requirement - role ability." The matching probability between the two is then calculated, taking into account three factors: first, the subtask query matching degree (the semantic overlap between the subtask requirement and the role's skills, such as the matching degree between "water quality monitoring" and "water quality analysis"); second, the agent's ability suitability (the suitability between the role's current ability level and the subtask difficulty, such as a senior role suited for a high-difficulty monitoring task); and third, historical allocation performance (the success rate of the role in handling similar subtasks in the past; if the past success rate exceeds 90%, the matching probability weight is increased). This multi-dimensional calculation ensures that the matching probability truly reflects the suitability of the role in handling the subtask, avoiding the problem of "ability mismatch." Step 5, Optimize the construction and allocation rules of the scenario-based reward function: A reward function is constructed based on the characteristics of the current governance scenario to optimize the final task allocation rules. The core parameters of the reward function are dynamically adjusted according to the scenario: for example, in emergency governance scenarios (such as post-earthquake environmental assessment), "task completion timeliness" accounts for 40% of the weight, "result accuracy" accounts for 30%, and "resource consumption" accounts for 30%; in routine governance scenarios, "result accuracy" accounts for 40%, "resource consumption" accounts for 30%, and "timeliness" accounts for 30%. The function sets reward and penalty items: when the role matching probability is high and meets the scenario weight requirements (such as a role with high matching probability and fast response speed in an emergency scenario), a high reward value is given; when the role matching probability is low or does not meet the scenario requirements (such as a role with high matching probability but excessive resource consumption in a routine scenario), a penalty value is given. Based on the calculation results of the reward function, the allocation rules are adjusted to prioritize assigning subtasks to the role with the highest reward value, while ensuring that no role undertakes subtasks beyond its own capacity. The final output is a clear "subtask-role" correspondence, i.e., the decomposition and allocation results. The edge-center collaborative point selection and divide-and-conquer algorithm decomposition not only adapts to the processing needs of different types of tasks, but also ensures the orderly progress of tasks through dependency analysis, avoiding resource waste and process chaos. The combination of semantic mapping and scenario-based reward functions significantly improves the matching accuracy of sub-tasks and roles, reducing governance efficiency losses caused by "capability mismatch". At the same time, the decomposition and allocation results output by this step directly provide a "sub-task list" and "execution role positioning" for the S400 governance solution, ensuring that subsequent solutions can accurately correspond to specific tasks and agents. From the task allocation level, this guarantees the efficiency and orderliness of spatial governance, avoiding governance delays or failures caused by unreasonable task decomposition or improper role allocation.

[0037] In the S400 architecture, which consists of a perception layer, a decision layer, and an execution layer, a multi-dimensional governance scheme is generated based on the decomposition and allocation results combined with real-time spatial data. The system determines the agent's task set and outputs governance operation instructions. The perception layer processes real-time data through edge nodes and transmits it to the decision layer. The decision layer integrates agent suggestions based on a large language model and handles conflicts according to semantic priority. The execution layer encodes geographic metadata and instructions based on a spatial perception protocol and transmits them to the corresponding agent.

[0038] Specifically, S400 includes the following steps: Step 1: Real-time data processing and transmission at the perception layer In the initial phase of S400, the perception layer first completes the acquisition and preprocessing of real-time spatial data, and then performs localized processing through edge nodes. The real-time data acquired by the perception layer covers a variety of dynamic information, such as regional environmental indicators updated in real time by sensors (e.g., soil moisture, air quality values), the latest surface images transmitted by satellites, and spatial state changes recorded by mobile monitoring devices (e.g., population density, traffic flow in a certain area).

[0039] Edge nodes perform targeted processing on this data: noise reduction and time alignment are performed on time-series data (such as continuous sensor readings) to avoid misjudgments caused by data fluctuations; local feature extraction is performed on image data (such as satellite updated images) to focus on areas related to the governance task (such as ecological restoration areas and disaster-prone areas); at the same time, all processed data are associated with corresponding geographic coordinates to mark the specific spatial range to which the data belongs (such as a street or a watershed).

[0040] After processing, the edge nodes transmit standardized real-time spatial data to the decision-making level through low-latency transmission links, ensuring that the decision-making level can obtain the latest and most accurate spatial status information, and providing dynamic data support for the generation of governance solutions.

[0041] Step 2: Generation of multi-dimensional governance solutions and determination of task sets for decision-makers: After receiving real-time spatial data transmitted from the perception layer, the decision-making layer, combined with the task decomposition and allocation results output by S300, initiates the generation process of a multi-dimensional governance plan. First, the decision-making layer calls the large language model, taking "real-time spatial data + sub-task allocation list" as input, and simultaneously collects governance suggestions proposed by each agent based on its own role profile (generated by S200). For example, the agent responsible for ecological restoration may suggest "supplementing irrigation in areas where soil moisture is below the threshold", and the agent responsible for traffic control may suggest "optimizing traffic routes in densely populated areas".

[0042] The large language model integrates these suggestions at the semantic level: first, it identifies the matching degree between the suggestions and subtasks and real-time data (such as whether irrigation suggestions correspond to areas with abnormal soil moisture data), and then handles suggestion conflicts according to semantic priority. For example, when "disaster emergency evacuation suggestions" conflict with "regular traffic control suggestions", emergency suggestions are retained first; when there are differences in similar suggestions (such as different agents having different suggestions on irrigation volume), the optimal solution is determined by combining historical governance effect data and real-time environmental parameters (such as the current water retention capacity of the soil).

[0043] Ultimately, the decision-making level generates a multi-dimensional governance plan that covers "spatial scope, governance objectives, implementation steps, and responsible intelligent agents," and clarifies the specific task set of each intelligent agent based on the plan (such as "intelligent agent A is responsible for irrigation operations in area XX, and intelligent agent B is responsible for pedestrian flow management in section XX").

[0044] Step 3, Execution layer instruction encoding and precise transmission: After the decision-making layer determines the task set for the intelligent agent, the execution layer completes the encoding and transmission of governance operation instructions. The execution layer first standardizes and encodes the geographic metadata and operation instructions in the governance plan based on the spatial perception protocol: the geographic metadata encoding needs to clearly define the precise spatial information corresponding to the task, including the latitude and longitude range of the target area, boundary coordinates, and key point markers (such as the valve location in the irrigation area and the entrance and exit of the drainage area); the operation instruction encoding needs to transform the abstract task into specific action instructions that the intelligent agent can execute. For example, "supplementary irrigation" is encoded as "start the XX model irrigation equipment, set the flow rate to XX cubic meters / hour, and the duration to XX minutes", and "human flow guidance" is encoded as "play guidance voice at XX location, and update the human flow density broadcast every 10 minutes".

[0045] After encoding, the execution layer accurately transmits the encoded instructions to the corresponding agent through a dedicated communication link established by the spatial awareness protocol. This link can optimize the transmission path based on the agent's spatial location (associated with geographic coordinates) to avoid instruction transmission delays or missends. At the same time, a checksum is added during instruction transmission to ensure that the instructions received by the agent are complete and tamper-proof, thus ensuring the consistency of subsequent execution actions with the governance plan.

[0046] Edge node processing in the perception layer ensures the validity of real-time spatial data, providing dynamic basis for solution generation and preventing solutions from becoming unrealistic due to data lag. The decision layer's large language model integration and semantic priority conflict handling fully leverage the professional capabilities of each agent while ensuring solution uniformity and emergency adaptability, avoiding deviations in governance direction due to conflicting suggestions. Standardized coding and precise transmission in the execution layer ensure that governance commands are accurately and efficiently transmitted to the corresponding agents, preventing execution errors caused by ambiguous or mistransmitted commands. Overall, the S400 inherits the task allocation results from the S300 and connects the actual execution actions of the agents. Through the division of labor and collaboration of the three-layer architecture, it significantly improves the scientific rigor, relevance, and implementation efficiency of spatial governance solutions, providing crucial support for the subsequent precise execution of governance tasks by agents.

[0047] S500 acquires the execution feedback information corresponding to the intelligent agent and updates the role generation strategy and allocation rules of the large language model based on the execution feedback information; it sets up a multi-level triggering mechanism, in which the triggering condition of sensor data exceeding the threshold and the triggering event of detecting a sudden disaster are both triggered to automatically reassign roles.

[0048] Specifically, S500 includes the following steps: Step 1: The agent performs multi-dimensional collection and classification of feedback information: In the initial phase of S500, comprehensive collection and classification of execution feedback information from each agent is conducted. This feedback information covers core status data of the agent's task execution, such as task completion progress (e.g., "Environmental monitoring task in a certain area is 85% complete"), resource consumption during execution (e.g., CPU utilization, data transmission latency), matching degree of governance effect (e.g., "Actual water quality improvement value deviates from target value by 10%)), and abnormal situations encountered during execution (e.g., "Sensor temporary offline caused data acquisition interruption"). Since different types of feedback information have significantly different uses (e.g., completion progress is used to evaluate task effectiveness, resource consumption is used to optimize allocation rules), a classified storage directory needs to be established according to "task effectiveness," "resource consumption," and "abnormal events." Simultaneously, the corresponding agent ID and task number are associated to ensure that each piece of feedback information can be traced back to a specific role and task, laying the foundation for subsequent accurate analysis and avoiding policy update deviations caused by chaotic feedback information.

[0049] Step 2, semantic parsing and key feature extraction of feedback information: For the collected categorized feedback information, a large language model is used for semantic parsing and key feature extraction, transforming unstructured feedback descriptions into quantifiable and correlated feature data. For example, for the text feedback "A certain agent's governance effect failed to meet standards due to insufficient skills," the model will parse out the core problem of "low role suitability" and extract key features such as "the type of skill of the agent involved, the governance indicators that failed to meet standards, and the deviation value." For the abnormal feedback "The sensor readings in a certain area exceeded the threshold for 5 consecutive minutes," features such as "the location of the abnormal sensor, the type of parameter exceeding the threshold, and the duration of exceeding the threshold" will be extracted. At the same time, through semantic association technology, the feedback features are bound to the role profile attributes (such as skills and spatial working range) generated by S200 and the task allocation results (such as sub-task type and matching probability) of S300, clarifying the correlation between the feedback problem and the previous role generation and task allocation stages, and providing accurate directional basis for policy updates.

[0050] Step 3: Update the role generation strategy of the large language model based on feedback features: By combining the extracted key feedback features, the role generation strategy of the large language model is dynamically updated to optimize the subsequent role profile generation logic. For example, if multiple feedback messages indicate that "the spatial working range of a certain type of agent is too wide, causing response delays," the parameter of "spatial working range division" in the role generation Prompt template is adjusted, optimizing the original rule of dividing by large areas into a more refined division based on "core governance points + radiation range." If feedback indicates that "some agents lack the ability to handle specific environmental parameters," the processing ability weight of the corresponding parameter is increased in the role skill generation stage to ensure that the subsequently generated role profiles are more in line with actual governance needs. During the update process, it does not rely on fixed rules but is based on the statistical regularity of feedback features. For example, when the proportion of "insufficient skills" feedback exceeds 30%, the skill generation logic is automatically triggered to iterate, avoiding the disconnect between roles and needs caused by fixed strategies.

[0051] Step 4: Optimize task allocation rules based on feedback features: Based on resource consumption and task performance characteristics in the feedback information, the task allocation rules in S300 are optimized to improve the rationality and efficiency of subsequent task allocation. For example, if the feedback shows that "a certain computing node is overloaded and experiencing execution delays due to excessive subtask allocation," the selection logic of the target computing node is adjusted, adding "historical load rate of the node" as a selection criterion, prioritizing nodes with loads below the threshold. If the feedback shows that "although the matching probability between the subtask and the role is high, the effect is poor because the real-time load of the role is not considered," a "real-time load coefficient of the agent" is added when calculating the matching probability to correct the original matching probability, making the allocation result more in line with the actual carrying capacity of the agent. At the same time, in combination with the task dependency issues in the feedback (such as "delay of the preceding subtask causes subsequent task lag"), the decomposition logic of the Divide-and-Conquer algorithm is optimized, adding "subtask time priority" judgment to ensure more reasonable dependencies.

[0052] Step 5, Deployment of multi-level triggering mechanism and automatic role reallocation: A multi-level mechanism combining "condition triggering" and "event triggering" is deployed, and role reassignment is automatically executed when the triggering conditions are met to ensure the continuity of the governance process. For condition triggering: real-time monitoring of environmental parameter data transmitted by sensors (such as air quality, water quality, and temperature) is implemented. When a parameter exceeds a preset threshold three times consecutively (e.g., PM2.5 concentration exceeding 150), the triggering mechanism will trigger an event. When an abnormal parameter is detected, the system automatically triggers role reassignment, retrieves the updated role generation strategy from S200, and quickly generates a temporary role adapted to manage the abnormal parameter. Simultaneously, based on the optimized allocation rules from S300, the relevant sub-tasks are reassigned to the new role. For event-triggered events: real-time monitoring modules on edge nodes (such as disaster monitoring cameras and vibration sensors) capture sudden disasters (such as floods and landslides). Once a disaster signal is detected, the emergency reassignment process is immediately triggered, suspending non-critical tasks, prioritizing the management role resources in the disaster area, calling on roles with disaster emergency handling skills to quickly take over the sub-tasks in the disaster area, and simultaneously disabling the task allocation permissions of faulty agents in the disaster area to avoid ineffective execution.

[0053] By classifying, collecting, and accurately analyzing feedback information, the shortcomings in the early stages of role generation and task allocation are identified, providing data support for strategy optimization. Role generation strategies and allocation rules are updated based on feedback characteristics to achieve a closed-loop iteration of "governance-feedback-optimization," continuously improving the adaptability between roles and tasks. The deployment of multi-level triggering mechanisms enables rapid response in the event of environmental anomalies or sudden disasters, ensuring uninterrupted governance processes through automatic role reassignment.

[0054] Overall, S500 not only solves the static problems that may exist in the early process, but also provides emergency support capabilities for the governance system, ensuring the long-term effectiveness and stability of multi-agent spatial collaborative governance from the two dimensions of "continuous optimization" and "risk response".

[0055] Taking the emergency scenario of "sudden rainstorm causing water pollution in a river basin in an city" as an example, the solution first responds quickly through the multi-level triggering mechanism of the S500: one hour after the rainstorm, if the water quality sensor in the basin detects that the COD value exceeds 120mg / L three times consecutively (the threshold is 80mg / L), the condition triggering mechanism is automatically triggered; at the same time, the disaster monitoring camera at the edge node captures garbage from the riverbank being washed into the water, triggering the event. At this time, the S100 immediately processes multi-source data. Satellite images are optimized through vLLM dynamic memory management to quickly extract the pollution diffusion range of the river basin; real-time sensor readings are aligned through Kalman filtering to clarify the pollution concentration gradient; geographic coordinates are associated with the river basin zoning (upstream tributaries, midstream urban section, downstream wetlands) and the governance requirement of "controlling pollution diffusion within 24 hours" to generate structured text and synchronize it to the central database. Data integration is completed in just 20 minutes, providing accurate and real-time data support for emergency governance and avoiding the expansion of the pollution range due to data lag. The solution then uses S200 and S300 to achieve role adaptation and efficient task allocation: the central server's large language model is based on structured text and dynamically generates three types of roles: "basin water quality monitoring intelligent agent", "pollution source interception intelligent agent" and "emergency dredging intelligent agent". The former's spatial work scope is bound to the coordinates of the upstream tributary, and its skills focus on real-time COD analysis; the latter two cover the middle-stream urban section and the downstream wetland respectively, and their knowledge boundaries include emergency operation specifications. In the edge-center collaborative architecture, three edge nodes (handling local monitoring and interception tasks) and one center node (coordinating dredging and scheduling) are selected around the watershed. The target node uses the Divide-and-Conquer algorithm to break down the "total pollution control task" into three sub-tasks: "real-time monitoring of pollution range", "activation of interception facilities along the river", and "dredging of key areas". The matching probability is calculated by combining the real-time load rate of the agent (e.g., "monitoring of high-concentration pollution areas" is assigned to water quality monitoring agents with a load of less than 40%). Emergency tasks are prioritized through a reward function (the reward weight for "interception activation" is set to 50%). Task allocation is completed within 15 minutes to avoid delays in treatment due to role mismatch or task backlog. Finally, the reliability of the governance is ensured through S400 execution and S500 feedback optimization: the edge nodes of the perception layer transmit pollution concentration update data to the decision layer every 10 minutes. The decision layer integrates the suggestions of the intelligent agents through the big language model. When the feedback from the "water quality monitoring intelligent agent" that "pollution in the middle reaches is still spreading" conflicts with the suggestion from the "pollution source interception intelligent agent" that "increase the number of interception points", the latter is retained according to the "emergency priority" and a governance plan of "adding 2 temporary interception dams" is generated. The execution layer encodes the "interception dam coordinates + start-up sequence" into instructions based on the spatial perception protocol and transmits them to the corresponding intelligent agents. After the execution feedback showed "low downstream dredging efficiency", the S500 immediately updated the role generation strategy (added the "large equipment collaboration" skill to the dredging agent) and optimized the allocation rules (prioritize agents with historical dredging efficiency of over 90% for the next similar task); during this period, a sudden failure of an edge node triggered the S500 role reassignment, calling the backup dredging agent (loading the historical best parameter snapshot and completing initialization within 3 seconds) to take over the task. Ultimately, the COD value of the watershed was reduced to within the threshold within 48 hours, which is 3 times more efficient than traditional manual treatment, and there was no interruption in treatment, fully demonstrating the efficiency and reliability of the solution in emergency scenarios.

[0056] In this embodiment of the application, the method further includes the following steps in processing multi-source heterogeneous information: Step 1: After S100 "Classification and Initial Processing of Multi-Source Heterogeneous Data" and before "Establishment of Cross-Type Semantic Association", the Kalman filter algorithm is executed on the real-time readings of the sensors (such as watershed water quality pH value and turbidity time series data). First, the algorithm constructs a "state prediction - observation update" loop, predicts the theoretical value at the current moment based on the historical reading patterns of the sensors, and then corrects the prediction deviation by combining the actual collected readings, and removes invalid data caused by sensor fluctuations (such as instantaneous abnormal readings caused by rainstorm impact). At the same time, the corrected readings are time-series aligned according to the timestamp to ensure that data from sensors at different locations in the same time period can be compared horizontally, providing a regular time series data foundation for subsequent association analysis.

[0057] Step 2: Before the "vLLM dynamic memory management optimization of satellite images" in S100, multi-band image fusion technology is used for satellite images (such as visible light and near-infrared images of the watershed): key bands related to watershed pollution monitoring are selected (such as visible light bands reflecting water turbidity and near-infrared bands identifying pollutant types). The feature information of multi-band images is superimposed through pixel-level fusion algorithms, preserving the advantageous features of each band (such as watershed boundary details in the visible light band and pollutant distribution features in the near-infrared band), weakening the noise interference of single bands, and generating a fused image with high feature recognition, laying the foundation for accurate extraction of environmental features during subsequent vLLM optimization processing.

[0058] Step 3: After completing sensor reading alignment and satellite image enhancement, a targeted association operation is added to the "Cross-type Semantic Association Establishment" step in S100: Using geographic coordinates as the link, spatial features such as "location and area of ​​watershed pollution patches" extracted from the enhanced satellite image are bound to temporal features such as "pollutant concentration and change trend at the corresponding location" in the aligned sensor readings. A semantic vector of "spatial feature - temporal data" association is generated through geospatial data embedding technology. This association operation can reduce the error interference of single data types (such as blurred satellite image boundaries and local sensor data deviations), and ultimately reduce the error rate of spatial environmental information such as watershed pollution range and pollutant concentration in the structured text generated by S100.

[0059] In this embodiment of the application, the method for generating the intelligent agent character profile further includes: Step 1: Inject role ability assessment metrics from the preset Prompt template: In the "Parameter Injection of Preset Prompt Template" step of S200, a new configuration of role capability assessment indicators is added: In the "Capability Adaptation" module of the template, three types of indicators are added: spatial governance task proficiency (such as the historical pass rate of the agent in completing similar governance tasks), policy and regulation matching degree (such as the fit between the agent's operation and the regional governance policy orientation), and cross-regional collaboration response speed (such as the response time of the agent's collaborative instructions with other regional agents). The indicator definitions and calculation logic (such as task proficiency = number of historically completed tasks / total number of tasks) are embedded into the template parameters, so that when the large language model generates the role profile, it can simultaneously output the capability assessment results containing the three types of indicators, providing a quantitative basis for subsequent profile optimization.

[0060] Step 2: Regularly update character ability assessment indicators and adjust character profiles: After the "Verification and Association Storage of Role Profiles" in S200, a timed update process is initiated: a fixed period (e.g., every hour) is set, and the regional environmental change information transmitted back in real time from the perception layer (e.g., a change in the type of governance task in a certain region, or an increase in cross-regional collaboration needs) is retrieved and compared with the current role capability assessment indicators; if environmental changes cause indicator deviations (e.g., an increase in cross-regional collaboration needs causes the "cross-regional collaboration response speed" indicator to fail to meet the standard), the role profile skill weights (e.g., increasing the weight of "collaborative communication skills") and spatial work range boundaries (e.g., expanding the coverage of the collaboration area) are dynamically adjusted; after the update, the new indicators and the adjusted profiles are synchronized to the central database to ensure that the role capabilities are always adapted to the real-time governance scenario.

[0061] In this embodiment of the application, the method further includes the following steps in obtaining the corresponding disassembly and allocation results: Step 1: After "Divide-and-Conquer algorithm for overall task decomposition" and before "Semantic mapping of subtasks and role profiles" in S300, a spatiotemporal constraint determination step is added: Based on the geographical coordinates of each subtask's associated region (e.g., the geographical scope of a street's garbage classification supervision subtask and the traffic management subtask of the same street), first compare the geographical boundaries of adjacent subtasks to determine if there is overlap; then check the execution time windows of the overlapping subtasks to confirm if there is a time conflict. If both "geographical overlap + time conflict" are satisfied, then retrieve the urgency level of the subtask's governance needs (e.g., the urgency level of an emergency disaster relief subtask is higher than that of a routine patrol), prioritize the execution of the subtask with higher urgency according to its original sequence, and postpone the execution time of the other subtask to a non-conflicting period to avoid interference between tasks in the same area. Step 2: In the S300 "Semantic Mapping and Matching Probability Calculation of Subtasks and Role Profiles" step, a new real-time load rate parameter is added. When calculating the matching probability, in addition to the original subtask query, agent capabilities, and historical allocation results, the current real-time task load rate of each agent is additionally retrieved (e.g., the number of unfinished tasks and the proportion of computing resources occupied by the agent). The load rate is converted into a weighting factor (e.g., if the load rate exceeds 80%, the matching probability weight is reduced) and incorporated into the matching probability calculation formula. This makes the results more closely match the actual carrying capacity of the agent. For example, if an agent has a high skill matching degree but a real-time load rate of 90%, its final matching probability will be lowered to avoid task execution delays due to overload and improve the fit between the matching results and the agent's carrying capacity.

[0062] In this embodiment of the application, the method further includes: Step 1: The decision-making level accesses the spatial governance scenario knowledge base: Before the "Integrated Intelligent Agent Suggestion" in the S400's "Generation of Multi-Dimensional Governance Schemes and Determination of Task Sets at the Decision-Making Level," the decision-making level first calls upon a pre-built spatial governance scenario knowledge base. This knowledge base stores historical governance cases (such as past cases of pollution control in a certain watershed or emergency management in a certain urban area) and corresponding records of the effectiveness of the solutions (such as the implementation results of the solutions in the cases and the reasons for not achieving the expected results). By calling upon the knowledge base, historical experience is provided for the generation of current governance solutions, avoiding unfounded solution design.

[0063] Step 2: Generate a feasibility assessment report for large language model comparison information: After accessing the knowledge base, the decision-making layer uses a large language model to semantically compare the real-time spatial information (such as the current regional environmental status and governance task requirements) transmitted from the perception layer with historical information in the knowledge base about similar scenarios (such as historical cases and solution effects under similar environmental conditions). This analysis examines the fit between the current solution approach and historically effective solutions, as well as potential risk points. Based on the comparison results, a feasibility assessment report is generated, which must clearly state the probability of the solution's approval and the supporting evidence.

[0064] Step 3, Scheme pass rate determination and execution layer coordinate verification: When the approval rate of the proposed solution in the evaluation report is ≥85%, the decision-making layer generates the final governance solution based on the agent's suggestions. Before entering the S400 "execution layer instruction encoding and precise transmission" stage, the execution layer needs to perform coordinate verification on the geographic metadata in the instruction (such as the latitude and longitude of the task area and the coordinates of key operation points). By comparing it with the standard geographic coordinates stored in the central database, the coordinate deviation is corrected to ensure that the instruction accurately points to the target governance area and avoids execution deviation caused by coordinate errors.

[0065] In this embodiment of the application, the method further includes the following steps during the automatic role reassignment process: Step 1, Preset the initial matrix for feedback information classification and scene adaptation weights: Following the "Multi-dimensional Collection and Classification of Feedback Information from Intelligent Agents" in S500, a new classification dimension has been added: feedback information is further subdivided according to spatial governance type (ecological monitoring, disaster emergency response, urban management) and regional complexity (core urban area, remote mountainous area, etc.), such as "feedback information of a disaster emergency response task in a core urban area" and "feedback information of an ecological monitoring task in a remote mountainous area". At the same time, a scenario-adaptive weight initial matrix is ​​preset, in which different governance types and regional complexities correspond to different weights (such as disaster emergency response type has a higher weight than ecological monitoring, and core urban area has a higher weight than remote mountainous area). This matrix replaces the traditional fixed-dimensional weighting method, making the feedback analysis more in line with the characteristics of the scenario.

[0066] Step 2: Feedback information semantic reflection, contextual memory storage, and weight adjustment: After the S500's "semantic parsing and key feature extraction of feedback information," the large language model performs semantic reflection on the parsed feedback information. For example, it analyzes whether "low role suitability is due to insufficient skills in disaster emergency scenarios or poor adaptability to remote mountainous environments," generating reflection results. The reflection results are stored as contextual memory and injected into the prompt for subsequent role generation. At the same time, the initial weight matrix of step 1 is dynamically corrected based on contextual memory. If multiple reflections show that "the feedback deviation in disaster emergency response in remote mountainous areas is large," the weight ratio of this scenario is increased to optimize the accuracy of feedback analysis.

[0067] Step 3, Character compatibility assessment and tool evolution process initiation: In the S500's "role generation strategy based on feedback features to update the large language model", a new adaptation monitoring is added: the role adaptation in the feedback information is statistically analyzed in real time. If it is lower than the preset threshold for three consecutive times (e.g., adaptation < 60%), the tool evolution process is automatically started. The large language model generates a tool calling strategy based on the current governance needs (e.g., the "rapid analysis of drone aerial photography data" tool is urgently needed in disaster relief in remote mountainous areas), which clarifies the tool type and calling trigger conditions. At the same time, the skill boundaries of the corresponding role profile are updated to include the operation skills of the new tool in the role skill range, thereby improving the adaptation ability of the role to the scene.

[0068] In this embodiment of the application, in addition to condition triggering and event triggering, a new capability decay triggering method is added, and the method further includes: Step 1: Determining Capacity Decay Trigger and Dynamically Adjusting Resource Thresholds: In the S500 "Multi-level Trigger Mechanism Deployment and Automatic Role Reassignment Execution", a new capability decay triggering process has been added: Based on the previously stored scenario memory, the execution success rate of the agent is monitored in real time. If the execution success rate of an agent drops by more than 30% with each task round, it is immediately determined to be capability decay, and role reassignment is automatically triggered (the updated role generation strategy is retrieved to match the new role). At the same time, the resource threshold trigger adopts a dynamic calculation model, which adjusts the threshold according to the urgency of the current governance task (e.g., the CPU utilization threshold is relaxed to 90% for disaster emergency tasks and kept at 80% for daily tasks) to avoid unnecessary load migration.

[0069] Step 2, rapid initialization of the backup agent in extreme scenarios: When a backup agent is activated in extreme scenarios (such as the collective ability of the original characters decaying or a sudden disaster causing multiple characters to fail), the "historical best character-task matching trajectory" (such as the character skills and task allocation logic with the highest suitability in similar tasks in the past) is first extracted from the context memory and migrated to the backup agent; then, the role parameters of the backup agent are fine-tuned (such as adjusting skill weights and spatial working range) by combining real-time spatial information from the perception layer (such as the regional environment of the currently unfinished task and the type of sub-task); at the same time, the core character lightweight model snapshot pre-stored on the edge node is called and directly loaded to complete the initialization, reducing the backup agent switching delay to the second level.

[0070] Step 3, Collaborative verification of backup agent capabilities: After the backup agent starts running, it establishes communication with other node agents through the spatial perception protocol and carries out capability collaboration verification: compares the backup agent's skill list (such as environmental monitoring and task execution skills) with the requirements of the currently unfinished sub-tasks (such as water quality monitoring and pollution control sub-tasks in a certain area) to confirm whether the skills are fully covered; if there are skill gaps, it will immediately link with other node agents to supplement the adaptation capabilities to ensure that there are no execution gaps in the unfinished sub-tasks. This application discloses a multi-agent spatial collaborative governance system, referring to... Figure 2 ,include: The multi-source heterogeneous data processing and storage module 001 processes multi-source heterogeneous data consisting of satellite images, real-time sensor readings, policy texts, and geographic coordinate data. Based on geospatial data embedding technology, it establishes corresponding cross-type semantic associations. Satellite images are optimized through vLLM dynamic memory management. Geographic coordinates are associated with regional environmental parameters and governance needs to generate spatial attribute semantic expressions, generating unified structured text, which is transmitted to the central database through a distributed consistency synchronization mechanism. The dynamic generation module 002 for character profiles, based on spatial environment information, governance goal constraints, and agent capability characteristics in structured text, dynamically generates agent character profiles, including names, skills, knowledge boundaries, and spatial working ranges, using a large language model on the central server based on a preset Prompt template when the environment changes or needs are adjusted. The generation does not rely on fixed rules. The knowledge boundaries include policy and regulation adaptation instructions, and the spatial working range includes associated geographic coordinates. The overall task decomposition and intelligent allocation module 003 obtains the overall task information of role profile and space governance. In the edge-center collaborative architecture, it selects target computing nodes based on the overall task features and the semantic attributes of the roles. The target nodes use the Divide-and-Conquer algorithm to decompose the overall task information into sub-tasks with dependencies, establish a semantic mapping between sub-tasks and role profiles, calculate the matching probability, and construct a reward function according to the scenario to optimize the allocation rules and obtain the corresponding decomposition and allocation results. The governance scheme generation and instruction output module 004, in the perception layer-decision layer-execution layer architecture, generates corresponding multi-dimensional governance schemes based on the decomposition and allocation results combined with real-time spatial data, determines the intelligent agent task set and outputs governance operation instructions. Among them, the perception layer processes real-time data through edge nodes and transmits it to the decision layer. The decision layer integrates intelligent agent suggestions based on a large language model and handles conflicts according to semantic priority. The execution layer encodes geographic metadata and instructions based on spatial perception protocol and transmits them to the corresponding intelligent agent. The feedback optimization and multi-level trigger guarantee module 005 obtains the execution feedback information corresponding to the intelligent agent and updates the role generation strategy and allocation rules of the large language model based on the execution feedback information; a multi-level trigger mechanism is set up, in which the sensor data exceeding the threshold trigger condition and the detection of a sudden disaster trigger event are automatically reassigned roles.

[0071] This application also discloses a multi-agent spatial collaborative governance system, including a processor, wherein the processor runs a program of any one of the multi-agent spatial collaborative governance methods described above.

[0072] This application also discloses a storage medium storing a program of the multi-agent spatial collaborative governance method described in any one of the above embodiments.

[0073] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A multi-agent spatial collaborative governance method, characterized in that, include: The system processes multi-source heterogeneous data consisting of satellite images, real-time sensor readings, policy texts, and geographic coordinate data. Based on geospatial data embedding technology, it establishes corresponding cross-type semantic associations. Satellite images are optimized through vLLM dynamic memory management. Geographic coordinates are associated with regional environmental parameters and governance needs to generate spatial attribute semantic expressions. A unified structured text is generated and transmitted to the central database through a distributed consistency synchronization mechanism. Based on spatial environment information, governance goal constraints, and agent capability characteristics in structured text, when the environment changes or needs are adjusted, the central server's large language model dynamically generates agent role profiles, including names, skills, knowledge boundaries, and spatial work scope, based on a preset Prompt template. The generation does not rely on fixed rules. Among them, the knowledge boundary includes policy and regulation adaptation instructions, and the spatial work scope includes associated geographic coordinates. To obtain the overall task information of character profiling and spatial governance, in the edge-center collaborative architecture, target computing nodes are selected based on the overall task features and character semantic attributes. The target nodes use the Divide-and-Conquer algorithm to decompose the overall task information into sub-tasks with dependencies, establish a semantic mapping between sub-tasks and character profiling, calculate the matching probability, and construct a reward function according to the scenario to optimize the allocation rules and obtain the corresponding decomposition and allocation results. In the perception layer-decision layer-execution layer architecture, based on the decomposition and allocation results and real-time spatial data, a corresponding multi-dimensional governance scheme is generated, the intelligent agent task set is determined and the governance operation instructions are output. The perception layer processes real-time data through edge nodes and transmits it to the decision layer. The decision layer integrates intelligent agent suggestions based on a large language model and handles conflicts according to semantic priority. The execution layer encodes geographic metadata and instructions based on spatial perception protocol and transmits them to the corresponding intelligent agent. Obtain the execution feedback information corresponding to the intelligent agent, and update the role generation strategy and allocation rules of the large language model based on the execution feedback information; establish a multi-level triggering mechanism, in which the role is automatically reassigned when the sensor data exceeds the threshold triggering condition or when the sudden disaster is detected.

2. The multi-agent spatial collaborative governance method according to claim 1, characterized in that, In processing multi-source heterogeneous information, the method also includes: The real-time sensor readings are first processed using a Kalman filter algorithm to remove outliers and align with the time-series records. Before optimizing satellite imagery using vLLM dynamic memory management, multi-band image fusion technology is employed to enhance the accuracy of space environment feature extraction. Semantic associations are established between enhanced satellite imagery and aligned sensor readings, reducing the error rate of spatial environment information in the generated structured text.

3. The multi-agent spatial collaborative governance method according to claim 2, characterized in that, The method also includes the following steps in generating the agent role profile: The preset Prompt template also includes role capability assessment indicators, which include proficiency in spatial governance tasks, policy and regulation matching degree, and cross-regional collaboration response speed. After the central server generates a character profile using a large language model, it will also combine the regional environmental change information transmitted back in real time from the perception layer to update the character ability assessment indicators periodically and dynamically adjust the skill weights and spatial work range boundaries in the character profile.

4. The multi-agent spatial cooperative governance method according to claim 1, characterized in that, The method also includes the following steps in obtaining the corresponding disassembly and allocation results: Add a step to determine the spatiotemporal constraints of subtasks by associating geographic coordinates with regions based on subtasks. If the geographical scope of adjacent subtasks overlaps and their execution times conflict, the subtask with higher urgency of governance needs should be retained first, and the execution order of the other subtask should be adjusted. When calculating the matching probability, in addition to considering the subtask query, agent capabilities, and historical allocation results, the agent's real-time task load rate is also introduced to improve the consistency between the matching probability calculation results and the agent's actual carrying capacity.

5. The multi-agent spatial collaborative governance method according to claim 1, characterized in that, The method also includes: Before integrating the suggestions of the intelligent agent, the decision-making level calls the pre-built spatial governance scenario knowledge base, which includes historical governance cases and corresponding solution effect records. By comparing real-time spatial information with historical information of similar scenarios in the knowledge base using a large language model, a corresponding feasibility assessment report is generated. When the approval rate of the proposed solution in the evaluation report is ≥85%, the corresponding final governance solution is generated based on the agent's suggestion. Before the execution layer outputs the governance operation instruction, the geographic metadata in the instruction is verified by coordinates.

6. The multi-agent spatial cooperative governance method according to claim 1, characterized in that, The method also includes the following in the process of automatically reassigning roles: The feedback information is categorized according to spatial governance type and regional complexity. Spatial governance types include ecological monitoring, disaster emergency response, and urban management. Regional complexity includes levels such as core urban areas and remote mountainous areas. Among them, a preset scenario-adaptive initial weight matrix is ​​used to replace the fixed-dimensional weighting method. The large language model performs semantic reflection on the feedback information, obtains the corresponding reflection results, stores the reflection results as contextual memory and injects them into subsequent prompts, and dynamically adjusts the weight matrix based on the contextual memory; if the feedback information shows that the role suitability is lower than the threshold for three consecutive times, the tool evolution process is automatically started, the large language model generates a tool calling strategy based on governance needs, and updates the skill boundaries of the role profile at the same time.

7. The multi-agent spatial collaborative governance method according to claim 6, characterized in that, In addition to condition-triggered and event-triggered methods, a new capability decay trigger has been added, and the methods also include: Based on the aforementioned contextual memory, if the success rate of the intelligent agent decreases by more than 30% with each task round, it is determined to be a capability decay, and role reassignment is automatically triggered; and the resource threshold trigger adopts a dynamic calculation model, which is adjusted in real time according to the urgency of the current governance task. When the backup agent is activated in an extreme scenario, the historical best role-task matching trajectory in the scenario memory is migrated to the backup agent, and the role parameters are fine-tuned by combining the real-time spatial information of the perception layer; at the same time, a lightweight model snapshot of the core role is pre-stored at the edge node, and the backup agent can directly call the snapshot to complete the initialization when it is activated, reducing the switching delay to the second level. After the backup agent is running, it performs capability collaboration verification with other node agents through the spatial awareness protocol to ensure that its skills cover the currently unfinished sub-tasks.

8. A multi-agent spatial collaborative governance system, characterized in that, include: The multi-source heterogeneous data processing and storage module processes multi-source heterogeneous data consisting of satellite images, real-time sensor readings, policy texts, and geographic coordinate data. Based on geospatial data embedding technology, it establishes corresponding cross-type semantic associations. Satellite images are optimized through vLLM dynamic memory management. Geographic coordinates are associated with regional environmental parameters and governance needs to generate spatial attribute semantic expressions, generating unified structured text, which is transmitted to the central database through a distributed consistency synchronization mechanism. The dynamic generation module for character profiles, based on spatial environment information, governance goal constraints, and agent capability characteristics in structured text, dynamically generates agent character profiles, including names, skills, knowledge boundaries, and spatial working ranges, using a large language model on the central server based on a preset Prompt template when the environment changes or needs are adjusted. The generation does not rely on fixed rules. The knowledge boundaries include policy and regulatory adaptation instructions, and the spatial working range includes associated geographic coordinates. The overall task decomposition and intelligent allocation module obtains the overall task information of role profile and space governance. In the edge-center collaborative architecture, it selects target computing nodes based on the overall task features and the semantic attributes of the roles. The target nodes use the Divide-and-Conquer algorithm to decompose the overall task information into sub-tasks with dependencies, establish a semantic mapping between sub-tasks and role profiles, calculate the matching probability, and construct a reward function according to the scenario to optimize the allocation rules and obtain the corresponding decomposition and allocation results. The governance scheme generation and instruction output module, in the perception layer-decision layer-execution layer architecture, generates corresponding multi-dimensional governance schemes based on the decomposition and allocation results combined with real-time spatial data, determines the intelligent agent task set and outputs governance operation instructions. Among them, the perception layer processes real-time data through edge nodes and transmits it to the decision layer. The decision layer integrates intelligent agent suggestions based on a large language model and handles conflicts according to semantic priority. The execution layer encodes geographic metadata and instructions based on spatial perception protocol and transmits them to the corresponding intelligent agent. The feedback optimization and multi-level trigger guarantee module obtains the execution feedback information corresponding to the intelligent agent, updates the role generation strategy and allocation rules of the large language model based on the execution feedback information, and establishes a multi-level trigger mechanism, in which the roles are automatically reassigned when sensor data exceeds the threshold trigger condition or when a sudden disaster event is detected.

9. A multi-agent spatial collaborative governance system, characterized in that, Includes a processor, wherein the processor runs a program for the multi-agent spatial cooperative governance method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The program stores the multi-agent spatial collaborative governance method as described in any one of claims 1-7.

Images