Micro-meteorological perception and multi-agent collaborative decision method and system for low-altitude scene

Abstract

The application discloses a micro-meteorological perception and multi-agent collaborative decision-making method and system for low-altitude scenes, and belongs to the technical field of low-altitude emergency command and unmanned aerial vehicle cluster scheduling. The system comprises a perception layer, a decision-making layer and an execution layer. The perception layer generates a micro-scale three-dimensional wind field through cascading downscaling, and constructs a voxelized risk cost field based on the wind resistance of unmanned aerial vehicles. The decision-making layer adopts a heterogeneous multi-agent collaborative decision-making engine, realizes the binding of a large language model and a domain calculation model through function interface reasoning double channels, combines a four-stage specialized workflow arrangement, and introduces an unoverlappable meteorological veto mechanism. The execution layer generates simulation verification files and supports dynamic strategy correction. The application solves the problems of insufficient micro-scale meteorological perception accuracy of existing systems, lack of meteorological hard constraints in multi-agent collaboration, and low flight safety resilience, and can be widely applied to low-altitude emergency scenes such as medical material delivery and disaster rescue.

Description

Technical Field

[0001] This invention relates to the field of low-altitude emergency command and UAV swarm scheduling technology, and in particular to a micro-meteorological perception and multi-agent collaborative decision-making method and system for low-altitude scenarios. Background Technology

[0002] With the rapid development of drone technology, drones, with their advantages of maneuverability, rapid deployment, and lack of terrain limitations, have been increasingly widely used in low-altitude emergency scenarios such as medical supply delivery, earthquake rescue, forest fire monitoring, and flood disaster relief. However, existing low-altitude emergency command and decision-making systems still face the following core technical challenges in practical applications: (1) The accuracy of microscale meteorological perception is seriously insufficient: the spatial resolution of traditional mesoscale meteorological forecasts is only at the kilometer level and the temporal resolution is at the hour level, which cannot accurately predict local microscale meteorological phenomena such as canyon turbulence, low-level wind shear, and urban heat island circulation. These microscale meteorological factors have a great impact on the safety of low-altitude UAVs and are one of the main causes of UAV flight accidents.

[0003] (2) Lack of hard meteorological constraints in multi-agent collaboration: Existing multi-agent decision-making frameworks are mostly general-purpose frameworks (such as LangGraph, CrewAI, AutoGen, etc.), which mainly focus on task allocation and collaboration efficiency, and do not deeply integrate meteorological constraints into each stage of the decision-making process. The system lacks an unhinderable meteorological safety veto mechanism, and when meteorological conditions change suddenly, it cannot terminate dangerous tasks in time, making it difficult to guarantee the safety of decision-making.

[0004] (3) Low safety of path planning and formation control: Existing path planning algorithms are mostly based on static maps and simple obstacle avoidance, without considering the dynamic risk cost field of voxels, and cannot dynamically adjust the flight path according to real-time weather conditions. At the same time, UAV formation control lacks a real-time crosswind compensation mechanism, which can easily lead to safety problems such as formation disorder and collision in complex wind field environments.

[0005] (4) It is difficult to balance the efficiency and security of the decision-making process: The general workflow orchestration architecture cannot meet the dual requirements of real-time performance and security in low-altitude emergency scenarios. Pure serial process decision-making is inefficient and cannot meet the timeliness requirements of emergency tasks; pure parallel process is prone to decision conflicts, requiring a lot of time to resolve conflicts, and it is difficult to guarantee the priority of meteorological constraints. Summary of the Invention

[0006] The purpose of this invention is to propose a micro-meteorological perception and multi-agent collaborative decision-making method and system for low-altitude scenarios, which solves the problems of insufficient micro-scale meteorological perception accuracy, lack of meteorological hard constraints for multi-agent collaboration, low safety of path planning and formation control, and difficulty in balancing efficiency and safety in decision-making processes in existing low-altitude emergency decision-making systems.

[0007] To achieve the above objectives, this invention provides a micro-weather perception and multi-agent collaborative decision-making system for low-altitude scenarios, comprising: The perception layer is used to acquire mesoscale meteorological data and perform cascaded downscaling to generate microscale meteorological digital twins, and to construct a voxelized risk cost field based on the physical wind resistance capability of UAVs. The decision layer, which communicates with the perception layer, includes a heterogeneous multi-agent collaborative decision engine and a four-stage hybrid workflow orchestration module. The heterogeneous multi-agent collaborative decision engine achieves the binding of abstract reasoning of large language models and domain-specific computing models through a dual-channel function-grounded reasoning. The four-stage hybrid workflow orchestration module adopts a serial + parallel + iterative low-altitude emergency specialization architecture to generate the final execution plan. The execution layer communicates with the decision-making layer to receive and execute the final execution plan, while simultaneously feeding back flight status data to the decision-making layer in real time for dynamic strategy assimilation and correction.

[0008] Preferably, the perception layer includes a microscale meteorological simulation engine and a multi-source data fusion module; The microscale meteorological simulation engine drives a large eddy simulation model based on mesoscale meteorological forecast data, generating three-dimensional wind field grid data with a preset resolution, covering a preset spatial and altitude range. The multi-source data fusion module is used to access third-party emergency databases and real-time turbulence observation data, and to assimilate and correct the wind field grid data. The current system uses a single NetCDF wind field data file.

[0009] Preferably, the perception layer also includes a five-level risk quantification assessment module, used to establish a five-level risk level mapping relationship oriented towards the safety target level, and to construct a voxelized risk cost field based on three-dimensional wind field grid data; the five risk levels are divided into safe level, low risk level, medium risk level, high risk level, and no-fly level based on the physical wind resistance threshold of the UAV; the cost function of the voxelized risk cost field is: ; in, , , These are the altitude, latitude, and longitude coordinates of the voxel, respectively. The risk level corresponds to the voxel; Normalized composite wind speed corresponding to the voxel; This indicates invalid boundary data.

[0010] Preferably, the heterogeneous multi-agent collaborative decision-making engine includes a weather warning agent, an airspace management agent, a UAV cluster agent, and a report integration agent. The four agents resolve conflicts through a sequential feedback loop mechanism. The conflict resolution strategies are, in order, adjusting the formation type, simplifying the path, and abandoning the task. The Weather Warning Agent possesses an unassailable veto power over weather data, used to analyze wind field data and output multi-altitude risk assessment reports. It implements multi-altitude weather analysis, simultaneously performing wind field statistical calculations at four altitudes: 100m, 150m, 200m, and 250m. Statistical indicators include maximum wind speed, average wind speed, data coverage, and safe zone percentage. The multi-altitude profile data is then analyzed by a large language model, which recommends the optimal flight altitude. The system defaults to adopting the recommended altitude and displays the recommended altitude range of 100-250m to the user. The airspace management agent is used to generate flight corridors of preset width and perform triple route verification of airspace, weather, and obstacles; the route is segmented and visualized at the front end according to risk level: safe area is green, low-risk area is yellow, medium-risk area is orange, high-risk area is red, and no-data area is gray dashed line. Risk labels are calculated point by point by the path planning module during the planning stage, and invalid data areas at the boundary are marked as no-data areas instead of no-fly zones; UAV cluster agent is used for formation allocation, waypoint calculation and first-order crosswind compensation; The report integrates agents to summarize the analysis outputs of weather warning agents, airspace management agents, and UAV cluster agents. It generates a comprehensive Chinese decision report through a large language model. The report includes a comprehensive summary, decision chain, comprehensive risk rating, operational recommendations, risk warnings, and execution summary.

[0011] Preferably, the function-grounded inference dual-channel architecture includes an LLM inference channel and a computation channel; the LLM inference channel is used for policy inference and natural language decision report generation, and adopts the DeepSeek V4 Flash large language model inference engine; the computation channel is used to execute numerical calculation tools for accurate numerical calculation. The system adopts a hierarchical reasoning strategy using the DeepSeek V4 Flash large language model: the report generation task uses JSON mode with a temperature of 0.2; the decision reasoning task uses thinking mode with a high reasoning intensity and a temperature of 0.3; the system has a built-in automatic degradation and retry mechanism for timeouts, automatically retrying if a single large language model call exceeds 180 seconds. Function grounding inference supports report generation mode: After the numerical calculation tool pre-calculates the data, it injects prompt words, and the LLM generates a decision report with a preset timeout.

[0012] Preferably, the workflow of the four-stage hybrid workflow orchestration module includes: Meteorological preliminary stage: Sequentially execute wind field data analysis, risk cost field construction and meteorological veto check, output risk base map, and directly terminate the process if meteorological limits are exceeded. Parallel computing phase: Airspace management agents perform flight corridor planning and triple route verification in parallel, while UAV cluster agents perform formation feasibility assessment and endurance assessment in parallel; Conflict resolution phase: Iterations are performed for a preset number of rounds, sequentially executing no-fly zone crossing checks, formation spacing verification, and endurance verification. The flight path and formation scheme are dynamically adjusted based on conflict feedback. Formation selection introduces a wind direction adaptive mechanism. The system automatically recommends formation types based on the average east-west and north-south wind components of the flight path area: V-shaped formations are recommended to enhance lateral stability when strong crosswinds are greater than 7 m / s; straight formations are recommended to reduce wind resistance when headwinds are greater than 5 m / s and crosswinds are less than 4 m / s; and triangular formations are recommended to obtain optimal sensor coverage under low wind speed conditions. This suggestion is passed to the UAV cluster agent, and the large language model combined with the task context confirms or adjusts the final formation type. Decision output stage: Output structured decision documents, standard geographic format flight routes, complete decision log chains, and visualized data. Approval reports and formation execution reports are generated by a large language model. The front-end visualization system implements 3D layered wind speed heat map rendering, supporting the overlay display of heat maps at five altitude levels: 100m, 150m, 200m, 250m, and 300m. At the same time, a wind field particle animation system is configured, using a self-developed Canvas 2D renderer, supporting multi-layer wind field particle overlay visualization. It achieves latitude and longitude spatial particle movement through 4 corner point projections per frame + bilinear interpolation, supporting real-time rendering of 2000-3000 particles.

[0013] Preferably, the execution layer includes a simulation package export module, which is used to automatically generate Gazebo simulation verification files according to the final execution scheme. The simulation verification files include SDF world files, MAVROS waypoint YAML files, and ROS Launch files.

[0014] A micro-weather perception and multi-agent collaborative decision-making method for low-altitude scenarios includes the following steps: S1: The sensing layer acquires mesoscale meteorological data and third-party emergency data, generates three-dimensional microscale wind field data with a preset resolution through cascaded downscaling, and constructs a voxelized risk cost field based on a five-level risk quantification system. S2: The decision-making level initiates a four-stage hybrid workflow. First, the meteorological preliminary stage is executed to conduct a meteorological veto check. If the check passes, the parallel computing stage begins. S3: In the heterogeneous multi-agent collaborative decision-making engine, the weather warning agent, airspace management agent, and UAV cluster agent execute their respective tasks in parallel, and resolve conflicts through multiple rounds of iteration via a sequential feedback loop mechanism to reach the final execution plan; S4: The execution layer generates simulation verification files based on the final execution plan, and also supports receiving flight status data. The decision layer performs dynamic strategy assimilation and correction based on the feedback data.

[0015] Preferably, in step S2, the specific rules for the meteorological veto check are as follows: if the wind speed in any area reaches or exceeds the preset no-fly wind speed threshold, the task is terminated directly; based on the different preset thresholds for the proportion of no-fly zones and the proportion of high-risk zones, corresponding suggestions are output for canceling the flight, adjusting the flight altitude or path, finely planning the flight path to avoid the risk, and reducing the flight altitude.

[0016] Preferably, step S3 further includes multi-objective constrained path planning based on a three-dimensional heuristic search algorithm and dynamic formation control that integrates first-order crosswind compensation; the three-dimensional heuristic search algorithm searches for the optimal path in a voxelized risk cost field, only accessing voxels with a risk level lower than the no-fly zone and an altitude not exceeding the maximum flight altitude of the UAV, and using Euclidean distance as a heuristic function. During the route planning phase of the airspace management agent, the system implements a dual-route comparison strategy: Route A is the shortest path, using the shortest path cost grid with a risk weight of 3, prioritizing distance; Route B is the safety-priority path, using the safety-priority cost grid with a risk weight of 20, strictly avoiding high-risk areas; after both candidate routes are independently planned by the A* algorithm, they are compared and evaluated by a large language model to select the optimal route. Dynamic formation supports various geometric configurations, allows setting preset safety distances, and the crosswind compensation formula is as follows: ; ; in, This is the north-south position compensation amount. This is the east-west position compensation amount. This represents the north-south wind speed component. For the east-west wind speed component, The nominal flight speed of the drone.

[0017] Therefore, the micro-meteorological sensing and multi-agent collaborative decision-making method and system for low-altitude scenarios described above have the following advantages: (1) Significantly improve the accuracy of microscale meteorological perception: Through the cascaded downscaling technology of mesoscale meteorological forecasting and large eddy simulation, it is possible to generate three-dimensional microscale wind fields with meter-level resolution, accurately predict meteorological phenomena such as local turbulence and wind shear that have a significant impact on low-altitude flight, and provide precise meteorological constraints for UAV flight. (2) Significantly enhances decision-making security: Introduces an unavoidable meteorological veto mechanism, making meteorological safety the highest priority in decision-making. Simultaneously, this invention proposes a four-layered defense-in-depth safety design concept, ensuring flight safety from multiple dimensions and significantly reducing the accident rate; (3) Achieving efficient multi-agent collaboration: A dual-channel architecture based on function-grounded inference is adopted to fully leverage the abstract reasoning capabilities of the large language model and the precise numerical computation capabilities of the domain computing model, achieving complementary advantages between the two. The sequential feedback loop mechanism ensures that multiple agents can quickly reach a consensus decision; (4) Balancing decision-making efficiency and safety: The specialized workflow orchestration of four stages of serial + parallel + iterative processes ensures the priority of meteorological constraints through the meteorological preliminary stage, improves decision-making efficiency through parallel computing, and ensures the correctness of decisions through iterative conflict resolution, perfectly adapting to the needs of low-altitude emergency scenarios. (5) Improve flight stability under complex wind fields: Three-dimensional heuristic search path planning based on voxelized risk cost field can automatically avoid high-risk areas. Dynamic formation control with first-order crosswind compensation can adjust the UAV position according to the real-time wind field and maintain formation stability.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is an overall architecture diagram of the micro-meteorological perception and multi-agent collaborative decision-making system for low-altitude scenarios in an embodiment of the present invention. Figure 2 This is a flowchart of a four-stage hybrid workflow in an embodiment of the present invention; Figure 3 This is an architecture diagram of the heterogeneous multi-agent collaborative decision-making engine in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0021] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0022] Example like Figure 1-3 As shown, this embodiment provides a micro-weather perception and multi-agent collaborative decision-making system for low-altitude scenarios, including: The perception layer is used to acquire mesoscale meteorological data and perform cascaded downscaling to generate microscale meteorological digital twins, and to construct a voxelized risk cost field based on the physical wind resistance capability of UAVs. The decision layer, which communicates with the perception layer, includes a heterogeneous multi-agent collaborative decision engine and a four-stage hybrid workflow orchestration module. The heterogeneous multi-agent collaborative decision engine achieves the binding of abstract reasoning of large language models and domain-specific computing models through a dual-channel function-grounded reasoning. The four-stage hybrid workflow orchestration module adopts a serial + parallel + iterative low-altitude emergency specialization architecture to generate the final execution plan. The execution layer communicates with the decision-making layer to receive the final execution plan and generate simulation verification files. It also supports receiving flight status data for dynamic policy assimilation and correction.

[0023] The perception layer includes a microscale meteorological simulation engine and a multi-source data fusion module. The microscale meteorological simulation engine drives a large eddy simulation model based on mesoscale meteorological forecast data to generate three-dimensional wind field grid data with meter-level resolution, covering a preset spatial and altitude range. The multi-source data fusion module is a reserved interface for this invention, reserved for accessing third-party emergency databases and meter-second real-time turbulence observation data to assimilate and correct the wind field grid data. The current system uses a single NetCDF wind field data file.

[0024] The perception layer also includes a five-level risk quantification assessment module, used to establish a five-level risk level mapping relationship oriented towards safety target levels, and to construct a voxel-based risk cost field based on three-dimensional wind field grid data; the five risk levels are: L0 safe level (wind speed < 5.0 m / s), L1 low risk level (5.0 m / s ≤ wind speed < 7.0 m / s), L2 medium risk level (7.0 m / s ≤ wind speed < 8.5 m / s), L3 high risk level (8.5 m / s ≤ wind speed < 10.0 m / s), and L4 no-fly level (wind speed ≥ 10.0 m / s); the cost function of the voxel-based risk cost field is: ; in, , , These are the altitude, latitude, and longitude coordinates of the voxel, respectively. The risk level corresponds to the voxel; The normalized composite wind speed for the corresponding voxel is calculated as follows: , The actual wind speed synthesized from voxels. For the east-west wind speed component, This represents the north-south wind speed component. The maximum wind resistance speed for the drone is set to 10.0 m / s in this embodiment; when the risk level is L4 (no-fly zone), the cost is directly set to the maximum value. Marked as impassable area; This indicates invalid boundary data, corresponding to a data-free region. The cost is set to a moderate value of 22.0, allowing the A* algorithm to pass but giving it some leeway.

[0025] The heterogeneous multi-agent collaborative decision-making engine includes a weather warning agent, an airspace management agent, a UAV cluster agent, and a report integration agent. The four agents resolve conflicts through a sequential feedback loop mechanism. The conflict resolution strategies are, in order, adjusting the formation type, simplifying the path, and abandoning the task. The Weather Warning Agent possesses an unassailable veto power over weather data, used to analyze wind field data and output multi-altitude risk assessment reports. It implements multi-altitude weather analysis, simultaneously performing wind field statistical calculations at four altitudes: 100m, 150m, 200m, and 250m. Statistical indicators include maximum wind speed, average wind speed, data coverage, and safe zone percentage. The multi-altitude profile data is then analyzed by a large language model, which recommends the optimal flight altitude. The system defaults to adopting the recommended altitude and displays the recommended altitude range of 100-250m to the user. The airspace management agent is used to generate a 15m wide flight corridor and perform triple route verification of airspace, weather, and obstacles. The route is segmented and visualized at the front end according to the risk level: the safe zone is green, the low-risk zone is yellow, the medium-risk zone is orange, the high-risk zone is red, and the no-data zone is a gray dashed line. The risk label is calculated point by point by the path planning module during the planning stage. Invalid data areas at the boundary are marked as no-data zones instead of no-fly zones, and are passed to the front end rendering through the flight corridor. UAV cluster agent is used for formation allocation, waypoint calculation and first-order crosswind compensation; The report integrates the analysis outputs of three agents: a weather warning agent, an airspace management agent, and a drone swarm agent. It generates a comprehensive Chinese decision report using the DeepSeek V4 Flash model. The inputs are the LLM report texts of the first three agents, formation data, and task context. The outputs include a comprehensive summary, decision chain, comprehensive risk rating, operational suggestions, risk warnings, and execution summary.

[0026] The function-grounded inference dual-channel architecture includes an LLM inference channel and a computation channel; the LLM inference channel is used for policy inference and natural language decision report generation, and adopts the DeepSeek series inference engine; the computation channel is used to execute Python tools for precise numerical calculations. The system employs a hierarchical strategy using the same model but different inference modes: Report generation mode: JSON forced output (response_format=json_object, temperature=0.2); Decision reasoning mode: Thinking chain enabled (thinking=enabled, reasoning_effort=high, temperature=0.3). It integrates with the DeepSeek V4 Flash model via an OpenAI-compatible API. The system has a built-in automatic timeout degradation mechanism: automatically switching to a degraded model after a single call exceeds 180 seconds, ensuring uninterrupted decision-making. Function-based grounded inference supports the report generation mode: Python pre-computes data and injects prompts, then LLM generates the decision report with a 30-second timeout.

[0027] The workflow of the four-stage hybrid workflow orchestration module includes: Phase 1 Meteorological Pre-test: PALM wind field data parsing, five-level risk cost field construction and meteorological veto check are executed sequentially, and a risk base map is output. If the meteorological limits are exceeded, the process is terminated directly. Phase 2 Parallel Computing Stage: Airspace Management Agents perform flight corridor planning and triple route verification in parallel, while UAV Cluster Agents perform formation feasibility assessment and endurance assessment in parallel. Phase 3, the conflict resolution phase, involves three rounds of iterative checks: no-fly zone crossing, formation spacing verification, and endurance verification. Based on conflict feedback, the flight path and formation scheme are dynamically adjusted. Formation selection incorporates a wind-adaptive mechanism: the system automatically recommends formation types based on the average U / V wind components of the flight path area—a V-shape is recommended for strong crosswinds (>7m / s) to enhance lateral stability, a straight line is recommended for headwinds (>5m / s and crosswinds <4m / s) to reduce wind resistance, and a triangle is recommended for low-wind conditions to achieve optimal sensor coverage. This recommendation is passed to the UAV cluster agent, which is then confirmed or adjusted by the LLM (Local Management Module) in conjunction with the task context. Phase 4 Decision Output Stage: Outputs structured JSON decision files, GeoJSON format flight routes, complete decision log chains, and visualized data. The LLM then generates approval reports and formation execution reports. The front-end visualization system implements a 3D layered wind speed heatmap rendering function (WindHeatmapRenderer), supporting overlay display of heatmaps at five altitude levels: 100m, 150m, 200m, 250m, and 300m. Simultaneously, the wind field particle animation system (WindParticleRenderer) uses a self-developed Canvas2D renderer, supporting multi-layered wind field particle overlay visualization. Each frame uses four corner projections plus bilinear interpolation to achieve particle movement in latitude and longitude space, with real-time rendering of 2000-3000 particles.

[0028] The system backend implements the Gazebo / ROS Noetic simulation package export function: automatically generates SDF world files (including quadcopter models), MAVROS waypoint YAML files (GLOBAL_RELATIVE_ALT waypoint sequences + takeoff + return instructions) and ROS Launch files (Gazebo environment + MAVROS node configuration) based on the decision results, and packages them into ZIP for download.

[0029] A micro-meteorological sensing and multi-agent collaborative decision-making method for low-altitude scenarios includes the following steps: S1: The sensing layer acquires mesoscale meteorological data and third-party emergency data, generates three-dimensional microscale wind field data with a preset resolution through cascaded downscaling, and constructs a voxelized risk cost field based on a five-level risk quantification system. S2: The decision-making level initiates a four-stage hybrid workflow. First, the meteorological preliminary stage is executed to conduct a meteorological veto check. If the check passes, the parallel computing stage begins. The specific rules for the meteorological veto inspection are as follows: if the wind speed in any area is ≥10.0m / s, the mission will be terminated directly; if the proportion of no-fly zones is ≥50%, it is strongly recommended to cancel or postpone the flight; if the proportion of no-fly zones is 30%~50%, it is recommended to adjust the flight altitude or path; if the proportion of no-fly zones is 10%~30%, it is necessary to carefully plan the flight route to avoid them; if the proportion of high-risk areas is ≥20%, it is recommended to reduce the flight altitude.

[0030] S3: In the heterogeneous multi-agent collaborative decision-making engine, the weather warning agent, airspace management agent, and UAV cluster agent execute their respective tasks in parallel, and resolve conflicts through multiple rounds of iteration via a sequential feedback loop mechanism to reach the final execution plan; It also includes multi-objective constrained path planning based on the 3DA heuristic search algorithm and dynamic formation control incorporating first-order crosswind compensation. The 3DA heuristic search algorithm searches for the optimal path in a voxelized risk cost field, only visiting voxels with a risk level <4 and an altitude not exceeding the maximum flight altitude of the UAV, using Euclidean distance as the heuristic function. In the route planning stage of the airspace management agent, the system implements a dual-route comparison strategy: Route A (shortest path, using the shortest path cost grid, risk weight ×3, distance priority) and Route B (safety priority, using the safety priority cost grid, risk weight ×20, strictly avoiding high-risk areas). Both candidate routes are independently planned by the A* algorithm and then compared and evaluated by LLM to select the optimal route. Dynamic formation supports three geometric configurations: triangle, straight line, and V-shape. The formation spacing is 5m, and the crosswind compensation formula is: ； ; in, This is the north-south position compensation amount. This is the east-west position compensation amount. This represents the north-south wind speed component. For the east-west wind speed component, The nominal flight speed of the drone.

[0031] S4: The execution layer receives and executes the final execution plan through a web service architecture of FastAPI backend + Vue frontend (future expansion to cloud-edge-device architecture), while providing real-time feedback on flight status data. The decision layer then dynamically assimilates and corrects the strategy based on the feedback data.

[0032] One specific implementation method is as follows: This embodiment is used for simulation verification of emergency medical supply delivery missions within urban areas. The specific implementation method is as follows: 1. Data processing at the perception layer: The perception layer first acquires mesoscale meteorological forecast data, drives the large eddy simulation engine to perform downscaling, and generates 3D wind field grid data with a resolution of 10m×10m×10m. The grid size is 399×399×52 layers, covering a spatial range of approximately 4km×4km and a height range of 0~505m, containing continuous wind field data for 13 time frames. At the same time, meter-second turbulence observation data from ground meteorological stations are integrated to assimilate and correct the wind field grid data.

[0033] The five-level risk quantification assessment module establishes a five-level risk level mapping relationship based on the physical wind resistance threshold (maximum wind resistance 10.0 m / s) of the quadcopter UAV used in this embodiment: L0 safe level (wind speed < 5.0 m / s), L1 low risk level (5.0 m / s ≤ wind speed < 7.0 m / s), L2 medium risk level (7.0 m / s ≤ wind speed < 8.5 m / s), L3 high risk level (8.5 m / s ≤ wind speed < 10.0 m / s), and L4 no-fly zone (wind speed ≥ 10.0 m / s). The normalized composite wind speed for each voxel is calculated. Substitute the values ​​into the cost function to construct a voxelized risk cost field. For example, when the actual wind speed of a voxel is 6.0 m / s, the risk level is L1, and the cost value is 1×10 + 6 / 10×2 = 11.2; when the wind speed is 10.5 m / s, the risk level is L4, marked as a no-fly zone, and the cost value is 10. 6 .

[0034] 2. Decision-making level workflow execution: The decision-making level initiates a four-stage hybrid workflow. In this embodiment, a maximum of three rounds of conflict resolution iterations are performed, and the width of the flight corridor is set to 15m. Phase 1: Meteorological Preliminary Stage. This phase involves analyzing PALM wind field data, constructing a risk map, and conducting a meteorological veto check. In this example, a significant increase in wind speed with altitude was detected: a maximum wind speed of 5.09 m / s at 50m, 6.84 m / s at 200m, and a maximum wind speed of 13.91 m / s at 300m (partial no-fly zone). There are no L4 no-fly zones within the recommended altitude range of 100-250m, and the overall no-fly zone percentage is <10%. The check is passed, and the process proceeds to the next stage. The meteorological warning agent simultaneously completes wind field statistics at four altitude levels: 100m, 150m, 200m, and 250m. After LLM analysis, 150m is recommended as the optimal flight altitude, with an average wind speed of 3.2 m / s and a safe zone percentage of 92%.

[0035] Phase 2 Parallel Computing Stage: The airspace management agent generates a 15m wide flight corridor based on the risk cost field, performs triple verification of airspace, weather, and obstacles, and generates two candidate routes, the shortest path and the safety priority path, through a dual-route comparison strategy; the UAV cluster agent evaluates the formation feasibility and endurance of three quadcopter drones, and the system automatically recommends a triangular formation based on the average wind speed in the route area (3.2m / s, crosswind 1.1m / s).

[0036] Phase 3 Conflict Resolution: Three rounds of iterative conflict checks were conducted, revealing that high-risk areas comprised more than 20% of the original flight path. Based on rule recommendations, the flight altitude was lowered to below 50m, and the adjusted flight path was conflict-free. Considering the mission's timeliness requirements, the LLM ultimately selected a safety-priority path and confirmed the use of a triangular formation.

[0037] Phase 4 Decision Output Stage: Heterogeneous multi-agent systems reach a consensus on the execution plan through a sequential feedback loop mechanism, outputting a structured JSON decision file and a GeoJSON format flight route. The report integrates the agents to generate an approval report containing a comprehensive risk rating and operational recommendations. The front end simultaneously renders a wind speed heatmap and wind field particle animation at a height of 50m, visually displaying the risk distribution of the flight route.

[0038] 3. Execution layer simulation verification: The execution layer automatically generates Gazebo simulation verification files through the simulation package export module, including an SDF world file containing terrain and obstacles, a MAVROS YAML file containing waypoint sequences, and a ROS Launch file. Users can download this simulation package and run UAV swarm mission simulations in their local Gazebo environment to verify the feasibility and safety of the decision-making scheme. During the simulation, the UAVs collect real-time status data such as wind speed and position and feed it back to the decision-making layer. The decision-making layer dynamically updates the risk cost field based on the feedback data, and if a sudden wind shear is detected, a flight path replanning is immediately triggered.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A micro-meteorological sensing and multi-agent collaborative decision-making system for low-altitude scenarios, characterized in that, include: The perception layer is used to acquire mesoscale meteorological data and perform cascaded downscaling to generate microscale meteorological digital twins, and to construct a voxelized risk cost field based on the physical wind resistance capability of UAVs. The decision layer, which communicates with the perception layer, includes a heterogeneous multi-agent collaborative decision engine and a four-stage hybrid workflow orchestration module. The heterogeneous multi-agent collaborative decision engine achieves the binding of abstract reasoning of large language models and domain-specific computing models through a dual-channel function-grounded reasoning. The four-stage hybrid workflow orchestration module adopts a serial + parallel + iterative low-altitude emergency specialization architecture to generate the final execution plan. The execution layer communicates with the decision-making layer to receive and execute the final execution plan, while simultaneously feeding back flight status data to the decision-making layer in real time for dynamic strategy assimilation and correction.

2. The micro-meteorological sensing and multi-agent collaborative decision-making system for low-altitude scenarios according to claim 1, characterized in that: The perception layer includes a microscale meteorological simulation engine and a multi-source data fusion module; The microscale meteorological simulation engine drives a large eddy simulation model based on mesoscale meteorological forecast data, generating three-dimensional wind field grid data with a preset resolution, covering a preset spatial and altitude range. The multi-source data fusion module is used to access third-party emergency databases and real-time turbulence observation data, and to assimilate and correct the wind field grid data. The current system uses a single NetCDF wind field data file.

3. The micro-meteorological sensing and multi-agent collaborative decision-making system for low-altitude scenarios according to claim 2, characterized in that: The perception layer also includes a five-level risk quantification assessment module, used to establish a five-level risk level mapping relationship oriented towards safety target levels, and to construct a voxelized risk cost field based on three-dimensional wind field grid data; the five risk levels are divided into safe level, low risk level, medium risk level, high risk level, and no-fly level based on the physical wind resistance threshold of the UAV; the cost function of the voxelized risk cost field is: ； in, , , These are the altitude, latitude, and longitude coordinates of the voxel, respectively. The risk level corresponds to the voxel; Normalized composite wind speed corresponding to the voxel; This indicates invalid boundary data.

4. The micro-meteorological sensing and multi-agent collaborative decision-making system for low-altitude scenarios according to claim 1, characterized in that: The heterogeneous multi-agent collaborative decision-making engine includes a weather warning agent, an airspace management agent, a UAV cluster agent, and a report integration agent. The four agents resolve conflicts through a sequential feedback loop mechanism. The conflict resolution strategies are, in order, adjusting the formation type, simplifying the path, and abandoning the task. The Weather Warning Agent possesses an unassailable veto power over weather data, used to analyze wind field data and output multi-altitude risk assessment reports. It implements multi-altitude weather analysis, simultaneously performing wind field statistical calculations at four altitudes: 100m, 150m, 200m, and 250m. Statistical indicators include maximum wind speed, average wind speed, data coverage, and safe zone percentage. The multi-altitude profile data is then analyzed by a large language model, which recommends the optimal flight altitude. The system defaults to adopting the recommended altitude and displays the recommended altitude range of 100-250m to the user. The airspace management agent is used to generate flight corridors of a preset width and to perform triple route verification of airspace, weather, and obstacles; The flight path is segmented and visualized at the front end according to the risk level: the safe zone is green, the low-risk zone is yellow, the medium-risk zone is orange, the high-risk zone is red, and the no-data zone is a gray dashed line. The risk label is calculated point by point by the path planning module during the planning stage. The boundary invalid data area is marked as a no-data zone instead of a no-fly zone. UAV cluster agent is used for formation allocation, waypoint calculation and first-order crosswind compensation; The report integrates agents to summarize the analysis outputs of weather warning agents, airspace management agents, and UAV cluster agents. It generates a comprehensive Chinese decision report through a large language model. The report includes a comprehensive summary, decision chain, comprehensive risk rating, operational recommendations, risk warnings, and execution summary.

5. The micro-meteorological sensing and multi-agent collaborative decision-making system for low-altitude scenarios according to claim 4, characterized in that: The function-grounded inference dual-channel architecture includes an LLM inference channel and a computation channel; the LLM inference channel is used for policy inference and natural language decision report generation, and adopts the DeepSeek V4 Flash large language model inference engine. The calculation channel is used to execute numerical calculation tools for precise numerical calculations; The system adopts a hierarchical reasoning strategy using the DeepSeek V4 Flash large language model: the report generation task uses JSON mode with a temperature of 0.2; the decision reasoning task uses thinking mode with a high reasoning intensity and a temperature of 0.3; the system has a built-in automatic degradation and retry mechanism for timeouts, automatically retrying if a single large language model call exceeds 180 seconds. Function grounding inference supports report generation mode: After the numerical calculation tool pre-calculates the data, it injects prompt words, and the LLM generates a decision report with a preset timeout.

6. The micro-meteorological sensing and multi-agent collaborative decision-making system for low-altitude scenarios according to claim 5, characterized in that: The workflow of the four-stage hybrid workflow orchestration module includes: Meteorological preliminary stage: Sequentially execute wind field data analysis, risk cost field construction and meteorological veto check, output risk base map, and directly terminate the process if meteorological limits are exceeded. Parallel computing phase: Airspace management agents perform flight corridor planning and triple route verification in parallel, while UAV cluster agents perform formation feasibility assessment and endurance assessment in parallel; Conflict resolution phase: Iterations are performed for a preset number of rounds, sequentially executing no-fly zone crossing checks, formation spacing verification, and endurance verification. The flight path and formation scheme are dynamically adjusted based on conflict feedback. Formation selection introduces a wind direction adaptive mechanism. The system automatically recommends formation types based on the average east-west and north-south wind components of the flight path area: V-shaped formations are recommended to enhance lateral stability when strong crosswinds are greater than 7 m / s; straight formations are recommended to reduce wind resistance when headwinds are greater than 5 m / s and crosswinds are less than 4 m / s; and triangular formations are recommended to obtain optimal sensor coverage under low wind speed conditions. This suggestion is passed to the UAV cluster agent, and the large language model combined with the task context confirms or adjusts the final formation type. Decision output stage: Output structured decision documents, standard geographic format flight routes, complete decision log chains, and visualized data. Approval reports and formation execution reports are generated by a large language model. The front-end visualization system implements 3D layered wind speed heat map rendering, supporting the overlay display of heat maps at five altitude levels: 100m, 150m, 200m, 250m, and 300m. At the same time, a wind field particle animation system is configured, using a self-developed Canvas 2D renderer, supporting multi-layer wind field particle overlay visualization. It achieves latitude and longitude spatial particle movement through 4 corner point projections per frame + bilinear interpolation, supporting real-time rendering of 2000-3000 particles.

7. The micro-meteorological sensing and multi-agent collaborative decision-making system for low-altitude scenarios according to claim 1, characterized in that: The execution layer includes a simulation package export module, which is used to automatically generate Gazebo simulation verification files based on the final execution scheme. The simulation verification files include SDF world files, MAVROS waypoint YAML files, and ROS Launch files.

8. A micro-meteorological sensing and multi-agent collaborative decision-making method for low-altitude scenarios, implemented based on the micro-meteorological sensing and multi-agent collaborative decision-making system for low-altitude scenarios as described in any one of claims 1-7, characterized in that, Includes the following steps: S1: The sensing layer acquires mesoscale meteorological data and third-party emergency data, generates three-dimensional microscale wind field data with a preset resolution through cascaded downscaling, and constructs a voxelized risk cost field based on a five-level risk quantification system. S2: The decision-making level initiates a four-stage hybrid workflow. First, the meteorological preliminary stage is executed to conduct a meteorological veto check. If the check passes, the parallel computing stage begins. S3: In the heterogeneous multi-agent collaborative decision-making engine, the weather warning agent, airspace management agent, and UAV cluster agent execute their respective tasks in parallel, and resolve conflicts through multiple rounds of iteration via a sequential feedback loop mechanism to reach the final execution plan; S4: The execution layer generates simulation verification files based on the final execution plan, and also supports receiving flight status data. The decision layer performs dynamic strategy assimilation and correction based on the feedback data.

9. The micro-meteorological sensing and multi-agent collaborative decision-making method for low-altitude scenarios according to claim 8, characterized in that: In step S2, the specific rules for the meteorological veto check are as follows: if the wind speed in any area reaches or exceeds the preset no-fly wind speed threshold, the task is terminated directly; based on the different preset thresholds for the proportion of no-fly zones and the proportion of high-risk zones, corresponding suggestions are output, such as canceling the flight, adjusting the flight altitude or path, finely planning the flight path to avoid the risk, and reducing the flight altitude.

10. The micro-meteorological perception and multi-agent collaborative decision-making method for low-altitude scenarios according to claim 8, characterized in that: Step S3 also includes multi-objective constrained path planning based on a three-dimensional heuristic search algorithm and dynamic formation control with first-order crosswind compensation; the three-dimensional heuristic search algorithm searches for the optimal path in a voxelized risk cost field, only accessing voxels with a risk level lower than the no-fly zone and an altitude not exceeding the maximum flight altitude of the UAV, and using Euclidean distance as the heuristic function. During the route planning phase of the airspace management agent, the system implements a dual-route comparison strategy: Route A is the shortest path, using the shortest path cost grid with a risk weight of 3, prioritizing distance; Route B is the safety-priority path, using the safety-priority cost grid with a risk weight of 20, strictly avoiding high-risk areas; after both candidate routes are independently planned by the A* algorithm, they are compared and evaluated by a large language model to select the optimal route. Dynamic formation supports various geometric configurations, allows setting preset safety distances, and the crosswind compensation formula is as follows: ； ； in, This is the north-south position compensation amount. This is the east-west position compensation amount. This represents the north-south wind speed component. For the east-west wind speed component, This refers to the nominal flight speed of the drone.

Images