Multi-agent cooperative control system based on adaptive learning

Abstract

The invention discloses a multi-agent cooperative control system based on adaptive learning, and belongs to the technical field of multi-agent system control, and the system specifically comprises the steps: collecting and analyzing the wind field data of the position of each agent; distinguishing two typical working conditions of a global uniform wind field and a local disturbance wind field based on a consistency analysis result; for a global uniform wind field, a model prediction control method is adopted to generate a wind field compensation strategy for keeping formation stability; for a local disturbance wind field, an accurate airflow disturbance state is obtained through environment three-dimensional reconstruction and fluid simulation, then the expected track deviation of each agent is predicted, and a collaborative avoidance instruction set is generated in combination with a relative position relationship. According to the method, adaptive cooperative control of multiple agents under different wind field working conditions is realized, the hidden conflict problem caused by local wind field disturbance is effectively solved, and the flight safety and task reliability of an agent cluster in a complex environment are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of multi-agent system control technology, specifically to a multi-agent cooperative control system based on adaptive learning. Background Technology

[0002] With the rapid development of the low-altitude economy, multi-agent systems are widely used in scenarios such as logistics distribution, urban patrol, and emergency rescue. In densely built-up urban environments, intelligent agent clusters need to cooperate in operations between tall buildings and narrow valleys, which places extremely high demands on flight safety. Traditional cooperative control systems mainly rely on GPS positioning and inertial navigation data, sharing location information through communication networks to achieve basic obstacle avoidance functions. This technical approach has good reliability in environments with stable weather conditions.

[0003] In recent years, researchers have begun to focus on the impact of environmental factors on agent swarms. Existing solutions employ wind field sensors mounted on agents to detect local airflow changes and adaptive control algorithms to adjust the flight attitude of individual agents in real time. Some advanced systems have also established data-sharing mechanisms among agents, allowing swarm members to exchange wind field measurement data for their respective locations, thereby achieving a preliminary understanding of the environmental wind field. These methods have improved the stability of single agents in wind-disturbed environments to some extent.

[0004] However, existing technical solutions still have significant limitations: First, they can only passively respond to existing wind field disturbances and cannot predict the propagation patterns of complex airflow phenomena such as regional flow around and vortex shedding within the cluster; second, they lack the ability to model the airflow interactions between agents. When multiple agents operate in confined airspace, the wake vortex generated by the leading agent may severely affect the flight trajectory of the following agents, and this implicit conflict risk caused by wind field transmission has not been effectively resolved; finally, existing solutions mainly focus on stability control at the individual level and fail to establish a collaborative avoidance mechanism for group wind disturbance risks at the system level, which limits the operational safety and reliability of multi-agent systems in complex wind field environments. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-agent cooperative control system based on adaptive learning, which solves the following technical problems: Existing solutions mainly focus on stability control at the individual level, failing to establish a collaborative avoidance mechanism for group wind disturbance risks at the system level. This limits the operational safety and reliability of multi-agent systems in complex wind field environments.

[0006] The objective of this invention can be achieved through the following technical solutions: A multi-agent cooperative control system based on adaptive learning, comprising: The data sensing module is used to collect wind field vector data at the location of the intelligent agent and transmit the wind field vector data to the shared database; The data analysis module is used to call the wind field vector data in the shared database and perform consistency analysis. When the wind field data of each agent is consistent within the preset threshold range, it is determined to be a globally uniform wind field; otherwise, it is determined that there is a local disturbance wind field. The collaborative control module is used to generate a global wind field compensation strategy based on wind field data if the wind field is determined to be globally uniform, and to start the simulation control module if the wind field is determined to be locally disturbed. The simulation control module is used to acquire environmental image sequences through a stereo vision sensor and generate an environmental point cloud model, extract fluid simulation boundaries from the environmental point cloud model, use the current wind speed data as the flow field initialization parameter, and use the numerical solution method of differential equations to calculate the airflow disturbance state on the current navigation path. The instruction generation module is used to acquire the airflow disturbance state and the real-time spatial coordinates measured by each agent through the positioning system, calculate the expected trajectory deviation of each agent based on the airflow disturbance state, construct a relative position relationship matrix by combining the real-time spatial coordinates of each agent, and generate a cooperative avoidance instruction set containing heading adjustment amount and speed control parameters based on the expected trajectory deviation and the relative position relationship matrix.

[0007] As a further aspect of the present invention: the specific process of performing consistency analysis on the wind field vector data in the shared database in the data analysis module is as follows: Obtain the wind field vector dataset collected by all agents within the same time window from the shared database, perform normalization preprocessing on the wind speed measurement values ​​of each agent in the dataset, and calculate the mean and standard deviation of the magnitude of the wind speed vectors of all agents. A consistency test model based on Gaussian distribution is established, and the wind speed vector of each agent is matched with the statistical features of the overall dataset for similarity. The deviation of the wind field data of each agent from the overall distribution is calculated by Mahalanobis distance. When the deviation of the wind field data of all agents from the overall distribution is less than the preset threshold, it is determined that the wind field data of each agent is consistent within the preset threshold range.

[0008] As a further aspect of the present invention: the specific process of generating the global wind field compensation strategy in the collaborative control module is as follows: The system acquires the GPS positioning coordinates, inertial measurement unit attitude data, and wind field sensor measurements of each agent. It uses an extended Kalman filter to fuse multi-source sensor data and establish motion state estimates for each agent. Based on a preset task track point sequence, it generates a reference track using a cubic spline interpolation algorithm. It calculates the lateral and longitudinal deviations between the actual position of each agent and the reference track and establishes a cost function that includes position error, heading error, and velocity error. Model predictive control algorithm is used to construct an optimization objective function that includes trajectory tracking accuracy and energy consumption indicators. Dynamic constraints based on agent mass distribution and rotational inertia, as well as actuator constraints based on the maximum thrust output of propulsion motor and the range of control surface deflection angle are set. The objective function and constraints are transformed into a second-order cone programming problem, and the optimal control input sequence in the multi-step prediction time domain is solved by a sequential quadratic programming algorithm. The first element of the optimal control input sequence is extracted as the thrust allocation command and control surface command of the propulsion system in the current control cycle. The control commands of each agent are weighted and averaged with those of neighboring agents by a distributed consensus algorithm. The coordinated control commands are then distributed to the motor controllers and servo motor drivers of each agent through a time-division multiple access communication protocol.

[0009] As a further aspect of the present invention: the specific process of generating the environmental point cloud model in the simulation control module is as follows: Image sequences from both left and right perspectives are acquired synchronously using a stereo vision sensor. The image sequences are then time-stamped and distortion-corrected. A scale-invariant feature transform algorithm is used to extract feature points from the corrected images and calculate their feature descriptors. A random sampling consensus algorithm is used to perform stereo matching on feature points of adjacent frames, and the three-dimensional spatial coordinates of the matching feature points are calculated based on the triangulation principle. The three-dimensional spatial coordinates are combined with the corresponding color information to generate initial point cloud data. Voxel grid downsampling is applied to the initial point cloud data to reduce the point cloud density. A statistical outlier removal algorithm is used to filter out noise points. Finally, the point cloud data of consecutive time periods are registered and fused based on the iterative nearest point algorithm to construct an environmental point cloud model with a unified coordinate system.

[0010] As a further aspect of the present invention: the specific process of extracting the fluid simulation boundary in the simulation control module is as follows: The system receives an environmental point cloud model and divides the point cloud data into continuous surface segments based on a region growing algorithm. It then performs normal vector analysis and curvature calculation on each surface segment to identify regions with similar geometric features. Based on the Poisson reconstruction algorithm, the segmented point cloud fragments are transformed into continuous mesh surfaces. Feature lines are extracted from the mesh surfaces to preserve the geometric features at curvature abrupt changes. Based on the curvature distribution and normal vector consistency of the mesh surfaces, key geometric boundaries affecting fluid motion are identified, including protruding edges, sharp corners, and discontinuous areas of the surface. Topology optimization is performed on the identified key geometric boundaries, and the Laplace smoothing algorithm is used to eliminate mesh distortion. Local mesh refinement is implemented in areas with large curvature. The optimized mesh surface is converted into boundary conditions of the fluid computation domain. No-slip boundaries, slip boundaries, and pressure outlet boundaries are set according to the geometric feature type to obtain the fluid simulation boundary model.

[0011] As a further aspect of the present invention: the specific process of calculating the airflow disturbance state on the current navigation path in the simulation control module is as follows: The generated fluid simulation boundary model is used as the basis for defining the computational domain. The computational domain is divided into interconnected grid cells using a spatial discretization method. Grid densification is implemented in the area around the agent's navigation path. Using the currently collected wind speed data as the initial condition of the flow field, the basic equations of fluid motion are solved through an iterative method, and no-slip boundary conditions are set on the surface of the fluid simulation boundary model. The fluid motion control equations are solved iteratively within each computation time step to obtain the airflow velocity distribution and pressure distribution data of each grid cell in the computational domain. The rotational component of the velocity field is extracted along the intelligent agent's navigation path to identify the spatial distribution of the vortex structure and calculate the spatial distribution of the airflow disturbance intensity. By analyzing the gradient change of the velocity field, vorticity transport analysis is performed to track the spatiotemporal evolution of the vortex structure and obtain a set of airflow disturbance state parameters that include the disturbance intensity distribution, vorticity field characteristics, and pressure gradient changes.

[0012] As a further aspect of the present invention: the specific process for calculating the expected trajectory deviation of each intelligent agent in the instruction generation module is as follows: The real-time spatial coordinates and airflow disturbance state parameter set of each agent are obtained. The turbulence intensity distribution and pressure gradient data in the airflow disturbance state are converted into aerodynamic disturbance torque acting at the center of mass of the agent. The aerodynamic disturbance torque is decomposed into pitching torque, rolling torque and yaw torque components along the three axes of the agent's body coordinate system. The mass distribution characteristics and rotational inertia parameters of each intelligent agent are obtained, and the angular acceleration generated by each torque component is calculated through the rigid body dynamics equation. The angular acceleration is integrated over time to obtain the change in the attitude angle of the intelligent agent. Combined with the current flight velocity vector, the change trend of the spatial position of the intelligent agent is deduced through kinematic relationships. The predicted change in position is compared with the preset reference track, and the lateral offset, longitudinal offset and altitude offset of the track are calculated respectively to generate a sequence of expected track deviation data with timestamps.

[0013] As a further aspect of the present invention: the specific process of generating a cooperative avoidance instruction set including heading adjustment and speed control parameters in the instruction generation module is as follows: The Euclidean distance between each agent is calculated based on their spatial coordinates. An envelope sphere with a radius of safety threshold is established with each agent as the center. The spatial motion trajectory of each envelope sphere in the future time period is inferred based on the expected trajectory deviation data sequence. The minimum distance between any two enveloping spheres on their trajectory is detected. When the minimum distance is less than a safety threshold, they are marked as potential conflict pairs. The non-penetration constraint condition for the enveloping spheres is set as the distance between the centers of any two agents is continuously greater than the sum of their safety radii. The dynamic constraint condition for the agents is set as the turning angular rate does not exceed a preset maximum value and the acceleration is maintained within the limit range. A distributed optimization algorithm is used to solve the heading angle sequence and velocity sequence of each agent in the prediction time domain under the above constraints. The first term of the obtained heading angle sequence is used as the heading adjustment amount, and the first term of the velocity sequence is used as the velocity control parameter. The control parameters between the agents are coordinated through a consensus protocol and a time-synchronized cooperative avoidance instruction set is generated.

[0014] The beneficial effects of this invention are: 1) This invention achieves accurate identification and adaptive response to complex wind field environments by constructing an intelligent wind field situational awareness and decision-making mechanism. Through a data perception module, wind field vector data from multiple agents is collected in real time, and consistency analysis is performed by a data analysis module, effectively distinguishing between two fundamentally different meteorological conditions: a globally uniform wind field and a locally disturbed wind field. Under globally uniform wind field conditions, this invention generates a unified wind field compensation strategy through a collaborative control module, enabling the agent cluster to resist wind field influences in a coordinated manner, maintaining formation stability and trajectory accuracy. When a locally disturbed wind field is detected, this invention immediately activates a simulation control module, achieving accurate identification and modeling of the local wind field disturbance source through 3D environmental reconstruction and fluid simulation. This adaptive decision-making mechanism based on wind field characteristic identification allows agents to collaboratively select the optimal response strategy according to actual meteorological conditions, avoiding unnecessary complex calculations under uniform wind field conditions while ensuring the activation of high-level response measures under complex wind field environments, significantly improving the system's environmental adaptability and decision-making intelligence.

[0015] 2) This invention achieves a technological leap from passive obstacle avoidance to active obstacle avoidance by establishing a forward-looking wind field disturbance prediction and trajectory extrapolation capability. Through a simulation control module, environmental point cloud data perceived by stereo vision is transformed into fluid simulation boundaries. Combined with real-time wind speed data, numerical calculation methods are used to extrapolate the airflow disturbance state along the intelligent agent's flight path, enabling early prediction of potential wind field risks. The airflow disturbance state is converted into specific aerodynamic disturbance torques, and the expected trajectory deviations of each intelligent agent under wind field action are extrapolated through rigid body dynamics equations. This allows potential risks to be identified before actual conflicts occur, providing sufficient reaction time for subsequent collaborative decision-making and achieving a revolution in safety protection modes.

[0016] 3) This invention achieves a balance between overall safety and efficiency in multi-agent systems under wind disturbance by constructing a collaborative optimization decision-making system. A relative positional relationship matrix between agents is built through an instruction generation module. Combined with expected trajectory deviation data, an optimization model is established that includes non-penetrating envelope constraints and agent dynamic constraints. A distributed optimization algorithm is used to solve for the optimal heading adjustment sequence and speed control sequence for each agent. This collaborative decision-making mechanism comprehensively considers the states and constraints of all agents. The generated collaborative avoidance instruction set ensures safe intervals between agents while also considering trajectory tracking accuracy and energy consumption. Furthermore, a consensus protocol guarantees the time synchronization and coordination of the instruction set. This avoids decision-making conflicts and inefficiencies caused by individual agents acting independently, achieving a unified improvement in system-level safety, efficiency, and coordination. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a schematic diagram of a multi-agent cooperative control system based on adaptive learning according to the present invention. Detailed Implementation

[0019] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 As shown, the present invention is a multi-agent cooperative control system based on adaptive learning, comprising: The data sensing module is used to collect wind field vector data at the location of the intelligent agent and transmit the wind field vector data to the shared database; The data analysis module is used to call the wind field vector data in the shared database and perform consistency analysis. When the wind field data of each agent is consistent within the preset threshold range, it is determined to be a globally uniform wind field; otherwise, it is determined that there is a local disturbance wind field. Each agent collects wind field vector data in real time from its location using onboard wind field sensors. For example, the wind pressure sensor on a drone measures the pressure difference generated by airflow based on fluid dynamics principles. A calibration circuit converts the pressure signal into digital wind speed and direction parameters, thus obtaining wind field vector data. This data is transmitted to a shared database via a wireless communication protocol, enabling centralized storage and synchronous updates of data from all agents. The shared database retrieves wind field vector datasets from all agents within the same time window. First, the data undergoes normalization preprocessing to eliminate dimensional differences. Then, the magnitude of the wind speed vectors for all agents is calculated. The mean and standard deviation are used to reflect the statistical characteristics of the overall wind field. Then, a consistency analysis is performed based on the Gaussian distribution model. The degree of deviation is quantified by calculating the Mahalanobis distance between each agent's wind field data and the overall distribution. This is because the Mahalanobis distance takes into account the data covariance and can more accurately assess the similarity between multidimensional vectors. When the deviation of all agents is less than the preset threshold, it indicates that the wind field is spatially uniform and is thus judged as a globally uniform wind field. Otherwise, it indicates that there is a local disturbance wind field. For example, near a group of buildings, some agents may report abnormal wind speeds due to vortex effects, causing the consistency analysis to fail and triggering the local disturbance judgment.

[0021] By acquiring and analyzing real-time wind field data, the type of environmental wind field can be dynamically identified, thereby adaptively selecting control strategies. This avoids unnecessary complex calculations under uniform wind field conditions, thus saving system resources. At the same time, when local disturbances are detected, advanced simulation and control mechanisms are triggered in a timely manner, ensuring that the intelligent agent cluster maintains coordination and safety in variable wind fields, laying the foundation for subsequent collaborative obstacle avoidance and trajectory optimization.

[0022] The collaborative control module is used to generate a global wind field compensation strategy based on wind field data if the wind field is determined to be globally uniform, and to start the simulation control module if the wind field is determined to be locally disturbed. If the wind field is determined to be globally uniform, it means that the wind field data collected by all agents are highly consistent in statistical characteristics, indicating that the wind field is spatially uniform and evolves slowly over time. In this case, adopting a global compensation strategy based on wind field data has clear physical significance: since all agents are similarly affected by the wind field, coordinated compensation can be achieved through a unified control law. This compensation strategy utilizes an extended Kalman filter to fuse multi-source sensor data, establishes a motion state estimation model for the system, and solves for the optimal control sequence based on a model predictive control framework. It is understandable that the system dynamics under a uniform wind field can be described by a relatively simple linear or weakly nonlinear model. Adopting a centralized optimization-distributed execution architecture can both guarantee control performance and significantly reduce computational complexity.

[0023] When localized wind disturbances are detected, it indicates a significant spatial gradient in the wind field, which may change rapidly over time. In this situation, simple global compensation strategies are no longer applicable because the wind field impact on agents at different locations differs fundamentally. Localized wind disturbances are typically caused by specific environmental structures (such as buildings and terrain), and their flow field characteristics are complex, including nonlinear phenomena such as separated flow and vortex shedding. By reconstructing a 3D environmental model using stereo vision and performing fluid simulations based on this model, an accurate wind field-environment coupling model can be established. This provides the necessary foundational data for subsequent precise control.

[0024] The simulation control module is used to acquire environmental image sequences through a stereo vision sensor and generate an environmental point cloud model, extract fluid simulation boundaries from the environmental point cloud model, use the current wind speed data as the flow field initialization parameter, and use the numerical solution method of differential equations to calculate the airflow disturbance state on the current navigation path. The essence of localized wind field disturbance is a complex flow phenomenon generated by the interaction between airflow and environmental structures such as buildings. Its impact can only be accurately predicted through detailed environmental modeling and physical simulation. Reconstructing a three-dimensional environmental model using stereo vision can realistically reflect the key geometric features affecting airflow motion. Extracting the fluid simulation boundary can ensure that numerical calculations accurately capture the obstruction and guidance effect of solid walls on airflow. Using numerical solutions of differential equations can simulate the nonlinear behaviors such as separation and vortexing of airflow when passing through complex environments based on first principles. Compared with empirical formulas or simplified models, this physical simulation-based method can more accurately predict the spatiotemporal evolution of localized wind fields, providing reliable input data for subsequent trajectory prediction and collaborative avoidance. It ensures that the hidden wind field risks caused by environmental disturbances can be anticipated and addressed, ultimately helping to achieve safe navigation of intelligent agents in complex urban environments and avoiding trajectory deviations and collisions caused by wind field disturbances.

[0025] The instruction generation module is used to acquire the airflow disturbance state and the real-time spatial coordinates measured by each agent through the positioning system, calculate the expected trajectory deviation of each agent based on the airflow disturbance state, construct a relative position relationship matrix by combining the real-time spatial coordinates of each agent, and generate a cooperative avoidance instruction set containing heading adjustment amount and speed control parameters based on the expected trajectory deviation and the relative position relationship matrix.

[0026] First, the system acquires airflow disturbance state data from the simulation control module, including turbulence intensity distribution, pressure gradient, and vorticity field characteristics. Simultaneously, it receives real-time spatial coordinates measured by each agent via GPS and inertial navigation systems. When calculating the expected trajectory deviation based on the airflow disturbance state, the system converts the turbulence intensity distribution data into unsteady aerodynamic forces acting on the agent, and the pressure gradient data into additional torques. Based on the agent's mass characteristics and rotational inertia parameters, it calculates the resulting angular acceleration using rigid body dynamics equations, then obtains the attitude angle change through time integration. Finally, it combines the current motion state to deduce the spatial position change over a future period, forming the expected trajectory deviation dataset. This calculation method is based on the fact that airflow disturbances essentially affect the trajectory by changing the forces and torques acting on the agent; a complete dynamic chain is necessary to accurately predict its trajectory. Final Results: When constructing the relative position relationship matrix using real-time spatial coordinates, the system calculates the Euclidean distance and relative velocity between every two agents, forming an adjacency matrix describing the spatial topology of the group. It is understandable that the collision risk between agents depends on their relative positions and motion trends, requiring precise mathematical descriptions of this spatial relationship. When generating a collaborative avoidance instruction set based on the expected track deviation and the relative position relationship matrix, the system first identifies potential safety distance violations caused by the expected track deviation. Then, it establishes an optimization problem including collision avoidance constraints and dynamic constraints. A distributed solution method is used to calculate the required heading adjustment sequence and velocity control sequence for each agent. Finally, the solution results are converted into specific control commands. Group collaborative avoidance requires comprehensive consideration of the states and constraints of all agents, and optimization methods can find the overall optimal avoidance strategy.

[0027] The trajectory deviation of a single agent can affect the safety of the entire group through relative positional relationships. Cooperative optimization at the system level is necessary. By accurately calculating the expected trajectory deviation, potential trajectory conflicts can be predicted in advance. By constructing a relative positional relationship matrix, the degree of mutual influence between agents can be quantified. Based on this information, generating cooperative avoidance instructions can ensure that the group maintains a safe distance during the avoidance process while minimizing interference with the original task. This enables proactive safety protection for multi-agent systems in windy disturbance environments. By replacing traditional reactive collision avoidance with predictive avoidance, the agent cluster can maintain both operational efficiency and operational safety in complex windy environments.

[0028] In a preferred embodiment of the present invention, the specific process of performing consistency analysis on the wind field vector data in the shared database in the data analysis module is as follows: Obtain the wind field vector dataset collected by all agents within the same time window from the shared database, perform normalization preprocessing on the wind speed measurement values ​​of each agent in the dataset, and calculate the mean and standard deviation of the magnitude of the wind speed vectors of all agents. A consistency test model based on Gaussian distribution is established, and the wind speed vector of each agent is matched with the statistical features of the overall dataset for similarity. The deviation of the wind field data of each agent from the overall distribution is calculated by Mahalanobis distance. When the deviation of the wind field data of all agents from the overall distribution is less than the preset threshold, it is determined that the wind field data of each agent is consistent within the preset threshold range.

[0029] In another preferred embodiment of the present invention, the specific process of generating the global wind field compensation strategy in the collaborative control module is as follows: The system acquires the GPS positioning coordinates, inertial measurement unit attitude data, and wind field sensor measurements of each agent. It uses an extended Kalman filter to fuse multi-source sensor data and establish motion state estimates for each agent. Based on a preset task track point sequence, it generates a reference track using a cubic spline interpolation algorithm. It calculates the lateral and longitudinal deviations between the actual position of each agent and the reference track and establishes a cost function that includes position error, heading error, and velocity error. Model predictive control algorithm is used to construct an optimization objective function that includes trajectory tracking accuracy and energy consumption indicators. Dynamic constraints based on agent mass distribution and rotational inertia, as well as actuator constraints based on the maximum thrust output of propulsion motor and the range of control surface deflection angle are set. The objective function and constraints are transformed into a second-order cone programming problem, and the optimal control input sequence in the multi-step prediction time domain is solved by a sequential quadratic programming algorithm. The first element of the optimal control input sequence is extracted as the thrust allocation command and control surface command of the propulsion system in the current control cycle. The control commands of each agent are weighted and averaged with those of neighboring agents by a distributed consensus algorithm. The coordinated control commands are then distributed to the motor controllers and servo motor drivers of each agent through a time-division multiple access communication protocol.

[0030] First, the GPS coordinates, inertial measurement unit attitude data, and wind field sensor measurements of each agent are acquired. These three types of data provide position, attitude, and external environment information, respectively. The state at the next moment is predicted using a system dynamics model, and then the prediction results are corrected using actual measurements to obtain a more accurate motion state estimate. A reference track is generated based on a preset task track point sequence using a cubic spline interpolation algorithm. Cubic spline interpolation ensures that the generated trajectory curve passes through all track points and has a continuous second derivative, which makes the reference track both meet path requirements and have smooth characteristics. The lateral and longitudinal deviations of each agent's actual position from the reference track are calculated, and a cost function including position error, heading error, and velocity error is established. This multi-dimensional error comprehensive evaluation can fully reflect the tracking performance of the agent. A model predictive control algorithm is used to construct an optimization objective function that includes track tracking accuracy and energy consumption indicators. By solving the optimal control problem in the finite time domain in each control cycle, only the first control variable is implemented, and then the solution is re-solved at the next moment, timely correction of deviations is achieved. Dynamic constraints based on the agent's mass distribution and moment of inertia, as well as actuator constraints based on the propulsion motor's maximum thrust output and the control surface deflection angle range, are established. These constraints ensure that the generated control commands are physically feasible and meet equipment safety requirements. The objective function and constraints are transformed into a second-order cone programming problem and solved using a sequential quadratic programming algorithm. Second-order cone programming can efficiently handle this type of optimization problem with quadratic constraints. Sequential quadratic programming decomposes the original problem into a series of quadratic programming subproblems to gradually approach the optimal solution. The first element of the optimal control input sequence is extracted as the propulsion system thrust allocation command and control surface command for the current control cycle. This is a standard practice in model predictive control, ensuring both optimality and real-time control. The control commands of each agent are weighted and averaged with those of neighboring agents using a distributed consensus algorithm. The distributed consensus algorithm achieves global consistency across the entire system through local information exchange between agents, ensuring the coordination of group behavior. Finally, the coordinated control commands are distributed to the motor controllers and servo motor drivers of each agent via a time-division multiple access (TDMA) communication protocol. TDMA avoids signal conflicts through time division, ensuring reliable transmission of control commands.

[0031] It is understandable that the control problem under uniform wind field conditions is relatively regular, making it suitable for a centralized decision-making-distributed execution architecture based on optimization theory. This architecture obtains accurate state estimates through multi-source sensor fusion, ensures path smoothness through trajectory interpolation, achieves feedforward compensation through model predictive control, ensures group coordination through distributed consensus, and guarantees reliable command transmission through communication protocols. The aim is to maximize system efficiency while ensuring control performance. This approach effectively resists wind field disturbances and maintains formation stability while avoiding unnecessary consumption of complex computational resources. Ultimately, it helps achieve efficient collaborative operation of the intelligent agent cluster in a uniform wind field environment, providing an economical and reliable solution for adaptive control of the entire system under different wind field conditions. This design allows the system to maintain good performance at a low cost under normal operating conditions and reserves sufficient computational resources to cope with complex operating conditions.

[0032] In another preferred embodiment of the present invention, the specific process of generating the environmental point cloud model in the simulation control module is as follows: Image sequences from both left and right perspectives are acquired synchronously using a stereo vision sensor. The image sequences are then time-stamped and distortion-corrected. A scale-invariant feature transform algorithm is used to extract feature points from the corrected images and calculate their feature descriptors. A random sampling consensus algorithm is used to perform stereo matching on feature points of adjacent frames, and the three-dimensional spatial coordinates of the matching feature points are calculated based on the triangulation principle. The three-dimensional spatial coordinates are combined with the corresponding color information to generate initial point cloud data. Voxel grid downsampling is applied to the initial point cloud data to reduce the point cloud density. A statistical outlier removal algorithm is used to filter out noise points. Finally, the point cloud data of consecutive time periods are registered and fused based on the iterative nearest point algorithm to construct an environmental point cloud model with a unified coordinate system.

[0033] Image sequences from both left and right perspectives are simultaneously acquired using a stereo vision sensor, similar to the principle of human stereoscopic vision. Two cameras capture the same scene from different angles, and depth information can be obtained by comparing the differences between the two images. The image sequences undergo timestamp alignment and distortion correction. Timestamp alignment ensures that the left and right images were captured at the same time, while distortion correction eliminates errors caused by lens optical distortion, making the images closer to the real scene. A scale-invariant feature transform algorithm is used to extract feature points from the corrected images and calculate their feature descriptors. This algorithm can stably identify key points in images under different scales and rotation conditions, such as the corners of buildings and the edges of windows, and generates a vector describing the surrounding image features for each feature point. Based on the principle of triangulation, the three-dimensional spatial coordinates of matching feature points are calculated. Using the optical center positions of the two cameras and the pixel coordinates of the feature points in the two images, the feature descriptors are calculated through geometric relationships. The system accurately locates the points in 3D space; it combines 3D spatial coordinates with corresponding color information to generate initial point cloud data, with each point containing position coordinates and color values, forming a preliminary 3D description of the environment; it performs voxel mesh downsampling on the initial point cloud data to reduce point cloud density by dividing the space into uniform cubic meshes and retaining a representative point in each mesh, thus significantly reducing the amount of data while maintaining the geometric features of the scene; it employs a statistical outlier removal algorithm to filter out noise points, which analyzes the distance distribution between each point and its neighbors to identify and remove outliers that are too far from the mean, such as floating points caused by sensor noise; finally, it uses an iterative nearest-point algorithm to register and fuse point cloud data from consecutive time periods, which iteratively calculates to find the optimal rigid body transformation between two point clouds, aligning point cloud data collected at different times to the same coordinate system, and gradually constructing a complete environmental point cloud model.

[0034] Acquiring depth information through stereo vision is the foundation for building environmental models. Timestamp alignment and distortion correction ensure data quality, feature point extraction and matching realize the transformation from two-dimensional images to three-dimensional space, triangulation provides the theoretical basis for three-dimensional coordinate calculation, point cloud downsampling improves processing efficiency while maintaining accuracy, outlier removal enhances model accuracy, and point cloud registration and fusion construct a complete environmental model. This provides accurate computational boundaries for fluid simulation, enabling it to realistically reflect the impact of environmental geometry on airflow and ensuring the reliability of simulation results. Ultimately, it helps the system accurately predict local wind field disturbances, providing a reliable basis for collaborative obstacle avoidance decisions by intelligent swarms. This refined environmental modeling enables the system to cope with various wind field anomalies in complex urban environments, improving the environmental adaptability and safety of the entire system.

[0035] In another preferred embodiment of the present invention, the specific process of extracting the fluid simulation boundary in the simulation control module is as follows: The system receives an environmental point cloud model and divides the point cloud data into continuous surface segments based on a region growing algorithm. It then performs normal vector analysis and curvature calculation on each surface segment to identify regions with similar geometric features. Based on the Poisson reconstruction algorithm, the segmented point cloud fragments are transformed into continuous mesh surfaces. Feature lines are extracted from the mesh surfaces to preserve the geometric features at curvature abrupt changes. Based on the curvature distribution and normal vector consistency of the mesh surfaces, key geometric boundaries affecting fluid motion are identified, including protruding edges, sharp corners, and discontinuous areas of the surface. Topology optimization is performed on the identified key geometric boundaries, and the Laplace smoothing algorithm is used to eliminate mesh distortion. Local mesh refinement is implemented in areas with large curvature. The optimized mesh surface is converted into boundary conditions of the fluid computation domain. No-slip boundaries, slip boundaries, and pressure outlet boundaries are set according to the geometric feature type to obtain the fluid simulation boundary model.

[0036] First, the environmental point cloud model is received. Based on a region growing algorithm, the point cloud data is divided into continuous surface segments. This algorithm starts from a seed point and compares the normal vector direction and curvature value of adjacent point clouds, grouping points with similar geometric features together to form continuous surfaces. The principle behind this is that the surfaces of objects in the real environment typically have a continuous normal vector distribution, while abrupt changes in normal vectors occur at the boundaries between different objects. Normal vector analysis and curvature calculation are performed on each surface segment. The normal vector reflects the orientation of the surface. Principal component analysis is used to calculate the covariance moments of the local neighborhood of the point cloud. The eigenvectors are obtained, and curvature characterizes the degree of surface bending, calculated through eigenvalues. These two parameters can effectively distinguish different geometric features such as planes, cylinders, and spheres. Based on the Poisson reconstruction algorithm, the segmented point cloud fragments are transformed into continuous mesh surfaces. This algorithm reconstructs implicit surfaces by treating the point cloud as the gradient field of a probability field and solving the Poisson equation. Its advantage lies in generating closed and smooth surface models, effectively filling holes in the point cloud. Feature lines are extracted from the mesh surface, preserving the geometric features at points of curvature abrupt change. The Gaussian curvature and average curvature at the mesh vertices are calculated. The curvature is used to connect the extreme points of curvature to form characteristic lines. These characteristic lines usually correspond to geometric features that have an important impact on airflow, such as building edges and corners. Based on the curvature distribution and normal vector consistency of the mesh surface, key geometric boundaries affecting fluid motion are identified, including prominent edges, sharp corners, and discontinuous areas of the surface. These geometric features can change the development of the boundary layer, leading to airflow separation and vortex generation. Topology optimization is performed on the identified key geometric boundaries, and the Laplace smoothing algorithm is used to eliminate mesh distortion. This algorithm improves mesh quality by adjusting the vertex positions to make them tend towards the centroid of neighboring vertices. Local mesh refinement is implemented in areas with high curvature. The resolution of these areas is improved by subdividing the mesh cells, because high curvature areas usually require denser meshes to accurately capture flow details. The optimized mesh surface is converted into boundary conditions of the fluid computational domain. No-slip boundaries, slip boundaries, and pressure outlet boundaries are set according to the geometric feature type. No-slip boundaries are used to simulate the viscous effect of fluid on solid surfaces, slip boundaries are used to simplify the handling of flow regions far from objects, and pressure outlet boundaries are used to set the pressure conditions at the flow outlet.

[0037] In another preferred embodiment of the present invention, the specific process of calculating the airflow disturbance state on the current navigation path in the simulation control module is as follows: The generated fluid simulation boundary model is used as the basis for defining the computational domain. The computational domain is divided into interconnected grid cells using a spatial discretization method. Grid densification is implemented in the area around the agent's navigation path. Using the currently collected wind speed data as the initial condition of the flow field, the basic equations of fluid motion are solved through an iterative method, and no-slip boundary conditions are set on the surface of the fluid simulation boundary model. The fluid motion control equations are solved iteratively within each computation time step to obtain the airflow velocity distribution and pressure distribution data of each grid cell in the computational domain. The rotational component of the velocity field is extracted along the intelligent agent's navigation path to identify the spatial distribution of the vortex structure and calculate the spatial distribution of the airflow disturbance intensity. By analyzing the gradient change of the velocity field, vorticity transport analysis is performed to track the spatiotemporal evolution of the vortex structure and obtain a set of airflow disturbance state parameters that include the disturbance intensity distribution, vorticity field characteristics, and pressure gradient changes.

[0038] By spatial discretization, continuous problems are transformed into computable discrete problems. Mesh refinement ensures computational accuracy in key areas. Setting reasonable initial and boundary conditions guarantees the realism of the simulation. Iteratively solving the governing equations is the core step in obtaining flow field information. Analyzing the rotational components of the velocity field can identify vortex structures, and calculating disturbance intensity can assess flow stability. By constructing a complete flow field cognition system, wind field risks on the flight path can be predicted in advance, and vortex regions and strong disturbance regions that may affect flight can be accurately identified. Ultimately, this helps the system formulate more scientific and effective trajectory planning and control strategies. This wind field analysis method based on physical simulation greatly improves the safety and reliability of intelligent agents flying in complex wind field environments.

[0039] In another preferred embodiment of the present invention, the specific process of calculating the expected trajectory deviation of each intelligent agent in the instruction generation module is as follows: The real-time spatial coordinates and airflow disturbance state parameter set of each agent are obtained. The turbulence intensity distribution and pressure gradient data in the airflow disturbance state are converted into aerodynamic disturbance torque acting at the center of mass of the agent. The aerodynamic disturbance torque is decomposed into pitching torque, rolling torque and yaw torque components along the three axes of the agent's body coordinate system. The mass distribution characteristics and rotational inertia parameters of each intelligent agent are obtained, and the angular acceleration generated by each torque component is calculated through the rigid body dynamics equation. The angular acceleration is integrated over time to obtain the change in the attitude angle of the intelligent agent. Combined with the current flight velocity vector, the change trend of the spatial position of the intelligent agent is deduced through kinematic relationships. The predicted change in position is compared with the preset reference track, and the lateral offset, longitudinal offset and altitude offset of the track are calculated respectively to generate a sequence of expected track deviation data with timestamps.

[0040] First, the real-time spatial coordinates and airflow disturbance state parameter sets of each agent are obtained. These data come from the positioning system and the simulation control module, respectively, providing basic input for subsequent calculations. The turbulence intensity distribution and pressure gradient data in the airflow disturbance state are converted into aerodynamic disturbance torques acting on the center of mass of the agent. This conversion is based on the principle of aerodynamics. Turbulence intensity reflects the degree of irregular pulsation of airflow, while pressure gradient reflects the pressure change of airflow in space. The combined effect of the two will produce an uneven aerodynamic force distribution on the surface of the agent. These forces will form disturbance torques relative to the center of mass. The aerodynamic disturbance torque is decomposed into pitch, roll, and yaw torque components along the three axes of the agent's body coordinate system. This decomposition is because the agent's motion can be decomposed into rotations about the three coordinate axes, corresponding to pitch, roll, and yaw motions respectively, facilitating subsequent dynamic analysis. The mass distribution characteristics and moment of inertia parameters of each agent are obtained; these parameters reflect the agent's inertial response characteristics to rotational motion. The angular acceleration generated by each torque component is calculated using rigid body dynamics equations, which describe the relationship between torque and angular acceleration. The principle is that the torque acting on an object is equal to the product of the moment of inertia and the angular acceleration. The change in the agent's attitude angle is obtained by time integration of the angular acceleration. The integration process is essentially a cumulative process. The angular velocity change caused by angular acceleration over time is added, and then the angular change caused by angular velocity over time is accumulated. Combined with the current flight velocity vector, the trend of the agent's spatial position change is deduced through kinematic relationships. This is based on kinematic principles, combining attitude changes with velocity vectors to calculate the agent's trajectory in space. The predicted position change is compared with a preset reference trajectory, which is the ideal path that the agent expects to follow. The lateral, longitudinal, and altitude deviations of the trajectory are calculated separately. These three components completely describe the agent's deviation from the reference trajectory. Finally, a data sequence of expected trajectory deviations with timestamps is generated. The timestamps record the time corresponding to each deviation value, which is convenient for subsequent analysis of the deviation development trend.

[0041] By converting airflow disturbances into torque effects, the fundamental impact of wind fields on agent motion can be reflected. Torque decomposition allows for the analysis of motion responses in various directions. Rigid body dynamics equations ensure the physical accuracy of motion prediction. Time integration enables the derivation from instantaneous states to continuous motion. Combining kinematic relationships transforms attitude changes into spatial position changes. Comparison with reference tracks quantifies the actual degree of deviation, while timestamp sequences record the evolution of the deviation. By predicting the agent's motion deviation under wind field conditions in advance, accurate predictive information can be provided for subsequent coordinated avoidance, allowing for proactive track corrections rather than reactive corrections after actual deviations occur. Ultimately, this helps the agent swarm maintain its predetermined flight formation and track under wind-disturbed conditions, ensuring flight safety and mission execution accuracy.

[0042] In another preferred embodiment of the present invention, the specific process of generating a cooperative avoidance instruction set including heading adjustment amount and speed control parameters in the instruction generation module is as follows: The Euclidean distance between each agent is calculated based on their spatial coordinates. An envelope sphere with a radius of safety threshold is established with each agent as the center. The spatial motion trajectory of each envelope sphere in the future time period is inferred based on the expected trajectory deviation data sequence. The minimum distance between any two enveloping spheres on their trajectory is detected. When the minimum distance is less than a safety threshold, they are marked as potential conflict pairs. The non-penetration constraint condition for the enveloping spheres is set as the distance between the centers of any two agents is continuously greater than the sum of their safety radii. The dynamic constraint condition for the agents is set as the turning angular rate does not exceed a preset maximum value and the acceleration is maintained within the limit range. A distributed optimization algorithm is used to solve the heading angle sequence and velocity sequence of each agent in the prediction time domain under the above constraints. The first term of the obtained heading angle sequence is used as the heading adjustment amount, and the first term of the velocity sequence is used as the velocity control parameter. The control parameters between the agents are coordinated through a consensus protocol and a time-synchronized cooperative avoidance instruction set is generated.

[0043] First, the Euclidean distance between each agent is calculated based on their spatial coordinates. This is obtained by taking the square root of the sum of the squares of the differences between the coordinates of two points in three-dimensional space, reflecting the actual spatial distance between agents. An envelope sphere with a radius equal to a safety threshold is established around each agent. This simplifies the agent into a spherical safety region, facilitating collision detection and spatial relationship analysis. The spatial trajectory of each envelope sphere in the future is inferred based on the expected trajectory deviation data sequence. This is done by superimposing the predicted position changes of each agent onto the current position of the envelope sphere, forming the sphere's movement path. The minimum distance between any two envelope spheres on their trajectories is detected by subtracting their respective radii from the distance between their centers at various times within the prediction period, finding the minimum value. When the minimum distance is less than the safety threshold, it is marked as a potential conflict pair, indicating a risk of collision between the two agents at some future time. A non-penetrating constraint condition for the envelope spheres is set: the distance between the centers of any two agents must continuously exceed the sum of their safety radii. Based on the fundamental principle of sphere collision detection, the system ensures that the two spheres remain separated at all times. Dynamic constraints on the agents are set: the turning angular rate does not exceed a preset maximum value, and the acceleration is maintained within a limit range. These constraints stem from the physical performance limitations of the agents, ensuring that the generated commands are physically feasible. A distributed optimization algorithm is used to solve for the heading angle and velocity sequences of each agent in the prediction time domain under the aforementioned constraints. The principle of distributed optimization is that each agent makes autonomous decisions based on local information, while information exchange ensures the consistency of group decisions. The first term of the solved heading angle sequence is used as the heading adjustment, and the first term of the velocity sequence is used as the velocity control parameter. This is a standard practice in model predictive control, ensuring both optimality and real-time control. A consensus protocol coordinates the control parameters among the agents. The consensus protocol uses parameter exchange and weighted averaging between adjacent agents to make the control parameters of the entire group tend to be consistent. Finally, a time-synchronized cooperative avoidance command set is generated to ensure that all agents execute the corresponding avoidance actions at the same time.

[0044] Collision avoidance in multi-agent systems requires consideration of group collaboration effects. Envelope sphere models simplify complex spatial relationship analysis, trajectory extrapolation based on anticipated track deviations can identify potential conflicts in advance, non-penetrating constraints ensure safe distances, dynamic constraints guarantee command feasibility, distributed optimization enables group decision-making, consensus protocols ensure action coordination, and time synchronization ensures execution consistency. The aim is to achieve safe collaborative flight of multiple agents in windy, disturbed environments. Its advantages include early warning of collision risks and generation of the overall optimal avoidance strategy, avoiding decision-making conflicts caused by individual agents acting independently. Ultimately, it helps the agent swarm maintain safe formation in complex windy environments while minimizing the impact on the original flight mission. This collaborative avoidance mechanism significantly improves the survivability and mission completion quality of multi-agent systems in disturbed environments.

[0045] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An adaptive learning based multi-agent cooperative control system, characterized in that, The method comprises the following steps: a data perception module is used to collect wind field vector data of the position of the agent and transmit the wind field vector data to a shared database; a data analysis module is used to call the wind field vector data in the shared database and perform consistency analysis, and when the wind field data of each agent is consistent within a preset threshold range, it is determined that it is a global uniform wind field, otherwise, it is determined that there is a local disturbance wind field; a cooperative control module is used to generate a global wind field compensation strategy based on the wind field data if it is determined that it is a global uniform wind field, and to start a simulation control module if it is determined that there is a local disturbance wind field; a simulation control module is used to collect environmental image sequences through a stereo vision sensor and generate an environmental point cloud model, extract a fluid simulation boundary from the environmental point cloud model, use current wind speed data as a flow field initialization parameter, and use a differential equation numerical solution method to calculate the air flow disturbance state on the current navigation path; an instruction generation module is used to obtain the air flow disturbance state and the real-time spatial coordinates of each agent measured by a positioning system, calculate the expected track deviation of each agent according to the air flow disturbance state, construct a relative position relationship matrix combining the real-time spatial coordinates of each agent, and generate a cooperative avoidance instruction set containing a heading adjustment amount and a speed control parameter based on the expected track deviation and the relative position relationship matrix.

2. The multi-agent collaborative control system based on adaptive learning according to claim 1, characterized in that, In the data analysis module, the specific process of consistency analysis on the wind field vector data in the shared database is as follows: obtain the wind field vector data set collected by all agents in the same time window from the shared database, normalize the wind speed measurement values of each agent in the data set, and calculate the mean and standard deviation of the module length of the wind speed vector of all agents; establish a consistency test model based on Gaussian distribution, match the wind speed vector of each agent with the statistical characteristics of the overall data set, and calculate the deviation degree of the wind field data of each agent from the overall distribution through Mahalanobis distance; when the deviation degrees of the wind field data of all agents from the overall distribution are all less than the preset threshold, it is determined that the wind field data of each agent is consistent within the preset threshold range.

3. The multi-agent collaborative control system based on adaptive learning according to claim 1, characterized in that, In the cooperative control module, the specific process of generating a global wind field compensation strategy is as follows: obtain the GPS positioning coordinates, inertial measurement unit attitude data and wind field sensor measurement values of each agent, fuse the multi-source sensor data using an extended Kalman filter and establish the motion state estimation of each agent, generate a reference track based on a preset task track point sequence through a cubic spline interpolation algorithm, calculate the lateral deviation and longitudinal deviation of the actual position of each agent from the reference track, and establish a cost function containing position error, heading error and speed error; use a model predictive control algorithm to construct an optimization objective function containing track tracking accuracy and energy consumption indicators, set dynamic constraints based on the mass distribution and moment of inertia of the agent, and set actuator constraints of the maximum thrust output of the propulsion motor and the deflection angle range of the rudder surface; The target function and the constraint condition are converted into a second-order cone programming problem, and an optimal control input sequence in a multi-step prediction time domain is solved by a sequential quadratic programming algorithm, a first element of the optimal control input sequence is extracted as a propulsion system thrust distribution instruction and a rudder surface control instruction of a current control period, the control instructions of the agents are calculated by a distributed consistency algorithm, the coordinated control instructions are distributed to motor controllers and rudder drivers of the agents through a time division multiple access communication protocol.

4. The multi-agent collaborative control system based on adaptive learning according to claim 1, wherein, In the simulation control module, the specific process of generating the environment point cloud model is as follows: The left and right view image sequences are synchronously collected by a stereo vision sensor, the image sequences are timestamped and corrected for distortion, feature points in the corrected images are extracted by using a scale-invariant feature transform algorithm and feature descriptors of the feature points are calculated, the feature points of adjacent frames are matched by using a random sample consensus algorithm and three-dimensional space coordinates of the matched feature points are calculated based on a triangulation principle, initial point cloud data are generated by combining the three-dimensional space coordinates with corresponding color information, the initial point cloud data are down-sampled in a voxel grid to reduce point cloud density, noise points are filtered by using a statistical outlier removal algorithm, and point cloud data of continuous time periods are registered and fused based on an iterative closest point algorithm to construct an environment point cloud model with a unified coordinate system.

5. The multi-agent collaborative control system based on adaptive learning according to claim 1, wherein, In the simulation control module, the specific process of extracting the fluid simulation boundary is as follows: The environment point cloud model is received and the point cloud data are divided into continuous curved surface segments based on a region growing algorithm, normal vector analysis and curvature calculation are performed on each curved surface segment, and regions with similar geometric features are identified; The segmented point cloud segments are converted into continuous grid surfaces based on a Poisson reconstruction algorithm, feature lines of the grid surfaces are extracted, and geometric features at curvature mutation positions are retained, key geometric boundaries affecting fluid motion are identified according to curvature distribution and normal vector consistency of the grid surfaces, including protruding edges, sharp corners and discontinuous regions of curved surfaces; The identified key geometric boundaries are topologically optimized, grid distortion is eliminated by using a Laplace smoothing algorithm, and local grid densification is performed in regions with large curvature, the optimized grid surfaces are converted into boundary conditions of a fluid calculation domain, no-slip boundaries, slip boundaries and pressure outlet boundaries are set according to geometric feature types, and a fluid simulation boundary model is obtained.

6. The multi-agent collaborative control system based on adaptive learning according to claim 1, wherein, In the simulation control module, the specific process of calculating the airflow disturbance state on the current navigation path is as follows: The generated fluid simulation boundary model is defined as a calculation domain definition basis, the calculation domain is divided into interconnected grid cells by using a spatial discretization method, and grid densification is performed in a surrounding area of the agent navigation path; Wind speed data currently collected are used as initial conditions of a flow field, fluid motion basic equations are solved by using an iteration method, and no-slip boundary conditions are set on the surface of the fluid simulation boundary model; the fluid motion control equations are solved in each calculation time step, and airflow velocity distribution and pressure distribution data of each grid cell in the calculation domain are obtained; The rotation component of the velocity field along the intelligent agent flight path is extracted, the spatial distribution of vortex structure is identified, and the spatial distribution of air flow disturbance intensity is calculated; the vorticity transport analysis is performed by analyzing the gradient change of the velocity field, the spatiotemporal evolution law of the vortex structure is tracked, and a set of air flow disturbance state parameters including the disturbance intensity distribution, the vorticity field characteristics and the pressure gradient change are obtained.

7. The multi-agent collaborative control system based on adaptive learning according to claim 1, characterized in that, In the instruction generation module, the specific process of calculating the expected flight path deviation of each intelligent agent is as follows: The real-time spatial coordinates of each intelligent agent and the set of air flow disturbance state parameters are obtained, the turbulence intensity distribution and the pressure gradient data in the air flow disturbance state are converted into the aerodynamic disturbance torque acting on the intelligent agent centroid; the aerodynamic disturbance torque is decomposed into pitch torque, roll torque and yaw torque components along the three axes of the intelligent agent body coordinate system; The mass distribution characteristics and the moment of inertia parameters of each intelligent agent are obtained, and the angular acceleration generated by each torque component is calculated through the rigid body dynamics equation; The angular acceleration is time-integrated to obtain the intelligent agent attitude angle change, and the change trend of the intelligent agent spatial position is deduced through kinematic relationship combined with the current flight speed vector; the predicted position change is compared with the preset reference flight path, and the lateral deviation, longitudinal deviation and height deviation of the flight path are calculated respectively, and the expected flight path deviation data sequence with time stamp is generated.

8. The multi-agent collaborative control system based on adaptive learning according to claim 1, wherein, In the instruction generation module, the specific process of generating the cooperative avoidance instruction set including the heading adjustment amount and the speed control parameter is as follows: The Euclidean distance between each intelligent agent is calculated based on the spatial coordinates of each intelligent agent, the envelope sphere with a safety threshold radius is established around each intelligent agent, and the spatial motion trajectory of each envelope sphere in the future time period is deduced according to the expected flight path deviation data sequence; The minimum separation distance of any two envelope spheres on the motion trajectory is detected, and when the minimum separation distance is less than the safety threshold, it is marked as a potential conflict pair; the non-penetration constraint condition of the envelope sphere is set as the distance between the centers of any two intelligent agents being greater than the sum of their safety radii, the intelligent agent dynamics constraint condition is set as the steering angle rate not exceeding the preset maximum value and the acceleration being maintained within the limit range, and the distributed optimization algorithm is used to solve the heading angle sequence and the speed sequence of each intelligent agent in the predicted time domain under the constraint conditions; the first term of the solved heading angle sequence is taken as the heading adjustment amount, the first term of the speed sequence is taken as the speed control parameter, the control parameters between the intelligent agents are coordinated through the consistency protocol, and the time-synchronized cooperative avoidance instruction set is generated.

Images