Multi-agent distributed coordination control method driven by heterogeneous sensing neural network

Abstract

The invention provides a multi-agent distributed coordination control method driven by a heterogeneous perception neural network, relates to the technical field of multi-agent distributed coordination control, and defines a heterogeneous distributed nonlinear system composed of a plurality of heterogeneous agents. Establishing a distributed heterogeneous multi-agent system model by establishing a heterogeneous agent kinetic equation and a communication topological structure; constructing a heterogeneous depth controller for environment perception, carrying out fusion processing on multi-source heterogeneous information by dynamically adjusting neighbor weights, and outputting agent state information and environment perception information; the communication connection is dynamically adjusted and the communication topology is reconstructed by using the intelligent agent state information and the environment perception information output by the controller; a reconstructed stable communication topology is used to realize distributed coordination control of heterogeneous multi-agents, a control strategy is optimized, a convergence theory framework is established, multiple constraints are processed in a unified manner, controller design is simplified, and the heterogeneous formation realizes zero-collision coordination motion in a complex obstacle environment.

Description

Technical Field

[0001] This invention relates to the field of multi-agent distributed coordination control technology, specifically a multi-agent distributed coordination control method driven by heterogeneous perceptual neural networks. Background Technology

[0002] With the rapid development of artificial intelligence technology, heterogeneous multi-agent systems present unprecedented challenges to the coordinated control of these systems. The differences in capabilities among agents render traditional control methods based on the isomorphism assumption ineffective, necessitating the establishment of new heterogeneous modeling and control frameworks. Furthermore, multi-constraint control systems in complex dynamic environments are required to possess strong environmental adaptability and real-time decision-making capabilities. Safety-critical applications place stringent demands on the stability and reliability of control systems, and traditional control theory suffers from difficulties in modeling and high computational complexity when dealing with high-dimensional nonlinear systems. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a multi-agent distributed coordination control method driven by heterogeneous perceptual neural networks, comprising: Step 1: Define a heterogeneous distributed nonlinear system composed of multiple heterogeneous agents, and establish a distributed heterogeneous multi-agent system model by establishing the dynamic equations of the heterogeneous agents and the communication topology. Step 2: Construct a heterogeneous depth controller for environmental perception. By dynamically adjusting the neighbor weights, multi-source heterogeneous information is fused and processed to output agent state information and environmental perception information. Step 3: Utilize the agent's state information and environmental perception information output by the controller to dynamically adjust the communication connection and reconstruct the communication topology; Step 4: Utilize the reconstructed stable communication topology to achieve distributed coordinated control of heterogeneous multi-agent systems and optimize the control strategy.

[0004] Preferably, in step 1, the agent dynamics equation at time t+1 is expressed as: ; in, For including subsystems The set of nodes of the subsystem and its neighbors, at time t+1 each subsystem The state vector is The control input at time t is , For unknown process noise at time t, Given a heterogeneous set of parameters, different subsystems i have different masses. Geometric radius Maximum thrust and drag coefficient ; The global system dynamics equations at time t+1, consisting of all subsystems, are as follows: ; Wherein, the state vectors of the global system at time t and time t+1 are respectively , The control input of the global system at time t is The unknown process noise and heterogeneous parameter set of the global system at time t are respectively , .

[0005] Preferably, in step 2, the intelligent agent The heterogeneous sensory neural controller is defined as a dynamic system: ; ; ; in, For the agent at time t Perceived information, This is a function for fusion of heterogeneous sensing information. For agents at time t and t+1 The internal state of the controller, Let be the state vector of agent j. For intelligent agents The state update function, For intelligent agents The control function, It is a set of obstacles and environments.

[0006] Preferably, define an intelligent agent Layered perception information for: , These represent individual tracking error, relative state of neighbors, formation configuration deviation, and environmental perception information, respectively.

[0007] Preferably, the heterogeneous perception and control network of the entire heterogeneous depth controller adopts a distributed architecture, where each agent only needs to communicate with its neighbors: ; Controller update equation: ; in This is relative state information. For the agent at time t and Heterogeneous weights between them.

[0008] Preferably, in step 3, the network topology is controlled based on environmental perception information. Dynamically adjust to adapt to environmental constraints; Environmental weight for: ; Communication feasibility function for: ; For intelligent agents and Between moments Environmental weights; For safety constraint functions; For intelligent agents and The position vector; A collection of obstacle environments; For intelligent agents and The physical radius; LOS is the line-of-sight reachability determination function; For attenuation parameters; The distance function is the distance to the obstacle. Based on security constraints, define the security constraint function: ; The Euclidean distance between the two agents; For an additional safe distance; For the smoothing parameters of safety constraints.

[0009] Preferably, considering the topology reconstruction process of a heterogeneous multi-agent system in an environment, the topology reconstruction process maintains system connectivity if the heterogeneous weights satisfy the following condition: Connectivity preservation conditions: ; It is the second smallest eigenvalue. It is the smallest eigenvalue; Heterogeneous compatibility conditions: ; For heterogeneous metrics, This is the critical value; Reconstruction smoothness condition: ; Network topology at time t+1 Network topology at time t The rate of topological change, The rate of change threshold; Environmental adaptability conditions: ; The minimum weight threshold, And converges to the optimal topology for environmental adaptation. satisfy: ; in, .

[0010] Preferably, a heterogeneous sensing distributed state estimator is designed, where each agent... Maintaining local state estimation: ; in This represents the posterior state estimate of agent i based on observations at time t. For heterogeneous Kalman estimation functions, For prior state estimation, For observation vectors; Neighbor state estimation through heterogeneous weight fusion : ; Fusion weights Adaptive adjustment based on estimation accuracy: ; in, This is the posterior state estimate of agent j based on observations at time t; Prior covariance matrix The traces.

[0011] Preferably, the objective function for heterogeneous multi-agent coordinated learning is... Defined as: ; in, For comprehensive performance indicators, This is an enhancement for heterogeneous collaboration; Converging to the global optimum : ; Existing for a limited time Make: ; Among them, the exact convergence rate is .

[0012] Preferably, a coordinated control system for the heterogeneous multi-agent system is established, which unifies the processing of multi-level constraints and constructs a comprehensive constraint function. : ; The constraints are numbered as follows: ,use To indicate the first The constraints are in functional form, and the constraint weights are... .

[0013] Compared with the prior art, the present invention has the following beneficial effects: This invention pioneers a heterogeneous perceptual neural network architecture, breaking through the limitations of the traditional homogeneous assumption and achieving adaptive configuration of parameter dimensions. It enables personalized controller design based on the physical characteristics (mass, thrust, radius) of different agents. Simultaneously, it establishes an environment-driven topology reconstruction mechanism, achieving intelligent reconstruction and stability assurance of the communication topology through a triple heterogeneous weighting mechanism (capability similarity, environmental adaptability, and task weight). Furthermore, it innovatively constructs a multi-scale coordinated unified framework, effectively solving the multi-level coupling problem of micro-level individual control, meso-level neighbor coordination, and macro-level formation maintenance.

[0014] This invention achieves a significant improvement in computational efficiency, with an inference time of only 21,550 microseconds, a 5-fold improvement compared to the traditional MPC method (106,982 microseconds), meeting real-time control requirements. In terms of safety, the system achieves 100% safety constraint satisfaction under 30% heterogeneous parameter differences, ensuring zero-collision safety. It exhibits excellent learning convergence characteristics, reducing formation error by 31.6% within 500 training rounds, achieving a final convergence accuracy of 98.6%, and possesses strong environmental adaptability, capable of autonomously traversing obstacle apertures of 1.8m and achieving automatic formation compression and reconstruction.

[0015] This invention establishes a complete theoretical system, including the topological stability reconstruction theorem guaranteeing connectivity maintenance and reconstruction smoothness, and the coordination control convergence theorem providing a mathematical proof of exponential convergence rate. Employing a distributed architecture design, the communication complexity is only O(|Ni|), far lower than the centralized O(N) complexity. 2 This laid the theoretical foundation for large-scale system applications. Attached Figure Description

[0016] Figure 1 A schematic diagram of the overall framework of a heterogeneous depth controller for environmental perception; Figure 2 To determine the optimal trajectory for four heterogeneous unmanned aerial vehicles to perform coordinated formation flight in a geometrically constrained environment; Figure 3 This is a correlation matrix; Figure 4 The convergence trajectory of formation error during training; Figure 5 This represents the training evolution trajectory of the distance between agents. Detailed Implementation

[0017] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0018] Example 1 I. Establishing a Distributed Heterogeneous Multi-Agent System Model 1. Heterogeneous distributed nonlinear systems This invention is considered to be by A distributed system consisting of interconnected heterogeneous quadrotor UAVs. The coupling network between the subsystems in the system is represented by an undirected communication graph. Description, in which Represents the N subsystems in the network, and the edge set It includes subsystems that can communicate with each other. .

[0019] Each subsystem The state vector is The control input is Its dynamic equation at time t+1 is expressed as: ; in, For including subsystems and the set of its neighbors' nodes, For unknown process noise, each subsystem at time t+1 The state vector is The control input at time t is , For unknown process noise at time t, This is a heterogeneous parameter set. The heterogeneity is reflected in the different qualities of different subsystems. Geometric radius Maximum thrust and drag coefficient .

[0020] Through operator form, subsystem Represented as: ; in , for The operator form defines the subsystem dynamics considering heterogeneous parameters.

[0021] Let m be the control input space for agent i. iIt refers to the dimension of the control input (for heterogeneous quadrotor dynamics, the control input is a 3-dimensional thrust vector). Let n be the state space of agent i. i It is the state vector dimension (in the dynamics of heterogeneous quadrotors, the state vector dimension is 6-dimensional: position + velocity).

[0022] By combining the local system dynamics, we can obtain the global system dynamic equations: ; Wherein, the state vectors of the global system at time t and time t+1 are respectively , The control input of the global system at time t is The unknown process noise and heterogeneous parameter set of the global system at time t are respectively , ; , , , This is a global heterogeneous parameter set.

[0023] n-dimensional real space; m-dimensional real space; : Real number space.

[0024] Similarly, the system can be rewritten in operator form: ; x is the operator form of the global state vector, u is the operator form of the global control input, and w is the operator form of the global unknown process noise.

[0025] 2. Heterogeneous Quadrotor Dynamics For a quadcopter unmanned aerial vehicle system, the state vector of each agent is: ,in and These represent the position and velocity vectors, respectively. Control input. The thrust vector is a three-dimensional vector that satisfies the constraints. .

[0026] The continuous-time dynamic equations of the heterogeneous quadrotor are as follows: ; ; Among them, gravity term Resistance term The drag coefficient is Coupling terms This represents the interaction forces between intelligent agents.

[0027] For a controlled heterogeneous system, the following assumptions are made: Assumption 1: Causal Operator Make the input To system response The mapping belongs to Space, process noise According to the unknown distribution Distribution. The heterogeneity of parameters ensures that the system has sufficient diversity to validate the effectiveness of the proposed control method.

[0028] II. Heterogeneous Depth Controller for Environmental Perception Based on the distributed heterogeneous multi-agent system model established in Part 1, this part will utilize the heterogeneous parameters (mass, geometric radius, maximum thrust, drag coefficient) of each agent to design a targeted heterogeneous depth controller. Traditional control methods are based on the isomorphism assumption and cannot effectively handle the physical differences between agents. However, the heterogeneous sensing depth controller proposed in this invention can be personalized according to the specific physical characteristics of each agent, achieving true heterogeneous adaptive control.

[0029] This heterogeneous deep controller takes the heterogeneous parameter set defined in the system model as design input and constructs differentiated control network architectures for agents with different capabilities through an adaptive parameter dimension configuration mechanism. Simultaneously, the controller integrates environmental awareness capabilities, enabling it to respond in real-time to complex environmental constraints and dynamically adjust control strategies to ensure the coordinated control performance of the heterogeneous multi-agent system under various environmental conditions.

[0030] The overall framework of the heterogeneous depth controller for environmental perception is as follows: Figure 1 As shown, this controller achieves adaptive coordination among agents through a heterogeneous sensing neural network and dynamically adjusts the control strategy based on environmental information, while ensuring system stability. Unlike existing methods, the proposed controller can handle the heterogeneity of agent capabilities and realize distributed coordinated control based on environmental perception.

[0031] To address the complexity of heterogeneous multi-agent systems, this step designs a heterogeneous perception control network architecture. The core innovation of this architecture lies in fusing the heterogeneous characteristics of the agents with environmental information to construct an adaptive coordination control network.

[0032] To address the heterogeneous characteristics of intelligent agents, a heterogeneous perceptual neural network controller is designed. Unlike traditional unified architectures, each intelligent agent... The controller is based on its physical parameters Personalized design.

[0033] Definition 1 (Heterogeneous Sensory Neural Controller): Intelligent Agent The heterogeneous sensory neural controller is defined as a dynamic system: ; ; ; in Let t be the internal state of the controller. For information perceived at time t, This is a function for fusion of heterogeneous sensing information. For intelligent agents The control function, For intelligent agents The state update function.

[0034] intelligent agent Dimensions of the controller's internal state for: ; in, This serves as the reference dimension for all intelligent agents.

[0035] Heterogeneous adjustment factors of mass and thrust and for: ; For reference quality, For reference thrust.

[0036] To address the problem of multi-scale spatial coordination and unification, the input information of heterogeneous perceptual neural networks... Structured design is carried out according to three scales: micro, meso, and macro. ; Microscale information (individual dynamic space) ): ; Mesoscale information (inter-agent interaction space) ): ; Macro-scale information (formation geometry space) ): ; For the desired formation configuration; The formation projection function is as follows: ; For intelligent agents In desired formation configuration The target location.

[0037] Environmental perception information : ; Environmental constraints.

[0038] Define intelligent agents The hierarchical sensing information is , These represent individual tracking error, relative neighbor state, formation configuration bias, and environmental perception information, respectively. Neighbor weights Dynamically adjust based on the similarity of agent capabilities and environmental conditions: ; Task weight is Environmental weight is Ability similarity weight Consider heterogeneous parameter matching: ; Here, is the standard deviation parameter for heterogeneity similarity, where heterogeneity measure Defined as follows: The method supports heterogeneous metrics: Any combination of parameters within the range.

[0039] The entire heterogeneous sensing and control network adopts a distributed architecture, where each agent only needs to communicate with its neighbors: ; Controller update equation: ; in, This is relative state information. For the agent at time t and The heterogeneous weights between them. This architecture has the advantage of low communication complexity. The communication complexity of the proposed distributed control architecture is only... ,in For intelligent agents The number of neighbors is far lower than that of a centralized approach. Complexity. Control laws for each agent. By relying solely on neighbor states, the deployability of the actual system is ensured. Intelligent agent control input vector, For intelligent agents Control functions; Neighbor set Status information; Intelligent agent The internal controller state.

[0040] III. Environmental Perception and Heterogeneous Topological Stability Reconstruction Based on the design of a heterogeneous deep controller, this step utilizes the agent state information and environmental perception information output by the controller to achieve intelligent reconfiguration of the communication topology. In heterogeneous multi-agent systems, since each agent has different physical capabilities and task roles, a fixed communication topology cannot adapt to dynamically changing environments and task requirements. Therefore, it is necessary to dynamically adjust the communication connections between agents based on the real-time state information output by the controller.

[0041] During operation, the heterogeneous depth controller outputs state information such as the position and velocity of each agent, as well as perception results of environmental obstacles and safety constraints. This information provides crucial decision-making basis for topology reconfiguration: by analyzing the relative positional relationships between agents, the establishment and disconnection of communication connections can be optimized; by utilizing environmental perception information, communication interruptions caused by obstacle occlusion can be predicted and avoided; and by combining heterogeneous parameter differences, more reasonable cooperative relationships can be established. This step uses this state information as input to design an environment-driven topology reconfiguration mechanism.

[0042] To address the previously mentioned problem of adaptive reconfiguration of heterogeneous topologies, this step establishes an environment-aware topology reconfiguration theory. By designing a topology stability mechanism, the connectivity and reconfiguration stability of heterogeneous control networks in complex environments are ensured.

[0043] 1. Environment-driven topology reconfiguration mechanism Based on environmental perception information, control network topology Dynamically adjust to adapt to complex environmental constraints.

[0044] Environmental weight Considering communication feasibility and the impact of obstacles: ; Communication feasibility function: ; For intelligent agents and Between moments Environmental weights; This is a communication feasibility function; For safety constraint functions; For intelligent agents and The position vector; A collection of obstacle environments; For intelligent agents and The physical radius. LOS is the line-of-sight reachability determination function; The attenuation parameter controls the degree to which obstacles affect communication; This is an obstacle distance function that calculates the degree to which the path is blocked by obstacles.

[0045] Based on security constraints, define a security constraint function. : ; The Euclidean distance between the two agents; For additional safety distance (safety margin); The smoothing parameter for safety constraints controls the steepness of the sigmoid function; the sigmoid is an S-shaped function that achieves a smooth transition from hard constraints to safety constraints.

[0046] Task weight Adjustments will be made based on current mission requirements and formation objectives: ; The variance parameter for task weights controls the degree of influence of formation deviation on the weights; , respectively intelligent agents The expected position of j in the formation.

[0047] 2. Heterogeneous compatibility and connectivity analysis To ensure the stability of the topology reconstruction process, a heterogeneous compatibility metric is defined. First, a heterogeneous weight matrix is ​​established. and its corresponding Laplace matrix .

[0048] Lemma 1: Heterogeneous Weighted Networks Connectivity is maintained if and only if the second smallest eigenvalue of its Laplacian matrix satisfies .

[0049] Maintaining connectivity requires considering the impact of heterogeneity metrics on network robustness. When heterogeneity metrics are too high, the coordination ability between agents decreases, necessitating the introduction of heterogeneity compatibility constraints.

[0050] Lemma 2: For heterogeneous metrics There is a critical value. , making when At that time, heterogeneous networks can maintain effective coordination and control.

[0051] 3. Stability Guarantee for Topology Reconfiguration Establish a stability theory for the topology reconfiguration process. Define the topology at time t+1. Topology at time t topological change rate : ; To avoid frequent topology oscillations, the reconstruction speed needs to be limited.

[0052] Lemma 3: If the rate of topological change satisfies This ensures a smooth refactoring process and avoids drastic fluctuations in control performance.

[0053] Environmental complexity has a significant impact on topological stability. An environmental complexity index is defined. : ; Let i be the distance from agent i to the nearest obstacle. This is a numerically stable term to prevent the denominator from being zero.

[0054] Lemma 4: For a weighted Laplace matrix, its smallest non-zero eigenvalue satisfies: ; in It is the unweighted Laplace matrix corresponding to the topological structure.

[0055] Based on the above analysis, a stability theorem for topological reconstruction is given: Theorem 1 (Environment-Aware Heterogeneous Topology Stable Reconstruction Theorem): Considering heterogeneous multi-agent systems in the context of environmental perception... In the topology reconstruction process, if the heterogeneous weights satisfy the following conditions: Connectivity preservation conditions: ; Heterogeneous compatibility conditions: ; Reconstruction smoothness condition: ; Environmental adaptability conditions: , The minimum weight threshold is the critical value that guarantees connectivity. It is the smallest eigenvalue. This is the threshold for the rate of change.

[0056] The topology reconstruction process preserves system connectivity and converges to the optimal topology adapted to the environment. ,satisfy: ; in To determine the optimal topology considering environmental constraints, The objective function is the topology optimization function.

[0057] IV. Distributed Coordination Control of Heterogeneous Multi-Agent Systems After completing the stable reconstruction of the heterogeneous topology for environmental perception in the previous step, this step will utilize the reconstructed stable communication topology to achieve distributed coordinated control of heterogeneous multi-agent systems. The stable topology provides a reliable communication foundation for distributed coordinated control, ensuring that each agent can effectively exchange information and coordinate decision-making.

[0058] The reconstructed communication topology has the following key characteristics: First, topological connectivity is guaranteed, enabling the distributed control algorithm to achieve global consistency; second, the heterogeneous weight mechanism enables stronger communication connections between agents with similar physical capabilities, improving coordination efficiency; and third, environmental adaptability ensures the robustness of the topology in complex environments, avoiding the impact of communication interruptions caused by environmental changes on control performance.

[0059] Based on this stable topology, this step designs a distributed coordination control algorithm, making full use of the optimized communication structure provided by topology reconstruction. Through topological adjacency relationships, each agent can obtain necessary neighbor information, achieving an organic combination of local decision-making and global coordination. Simultaneously, heterogeneous weight information provides an important basis for trade-offs between distributed estimation and control, enabling the coordination control algorithm to adapt to the specific needs of heterogeneous systems.

[0060] Based on the stable topology reconstruction guaranteed by Theorem 1, this step designs a heterogeneous multi-agent distributed coordination control algorithm for the multi-scale spatial coordination and unification problem and the environmental perception and global state reconstruction problem. The multi-scale coordination mechanism achieves the unification of micro-level individual control, meso-level neighbor coordination, and macro-level formation maintenance through hierarchical perception information.

[0061] 1. Distributed state estimation and information fusion Under limited communication conditions, each agent needs to reconstruct global state information based on local observations. A heterogeneous sensing distributed state estimator is designed, where each agent... Maintaining local state estimation: ; in This represents the posterior state estimate based on the observations at time t. For heterogeneous Kalman estimation functions, For prior state estimation, This is the observation vector.

[0062] Heterogeneous Kalman gain considers differences in agent characteristics : ; in, The prior covariance matrix, Observation matrix transpose, This is the heterogeneous observation noise covariance matrix.

[0063] Among them, heterogeneous observation noise covariance : ; This is the observation noise covariance matrix for the reference agent or standard configuration.

[0064] Neighbor state estimation through heterogeneous weight fusion : ; This is the posterior state estimate of agent j based on observations at time t; Fusion weights Adaptive adjustment based on estimation accuracy: ; Prior covariance matrix The traces.

[0065] 2. Heterogeneous Coordination Learning and Control To align with the established control objectives, the objective function of heterogeneous multi-agent coordinated learning... Defined as: ; in For comprehensive performance indicators, Enhancements for heterogeneous collaboration: ; The desired relative positions of the formation.

[0066] 3. Coordinate and control convergence and performance guarantees Establish coordinated control for heterogeneous multi-agent systems, unify the handling of multi-level constraints, and construct a comprehensive constraint function. : ; The constraint conditions are numbered as follows: ,use To indicate the first The functional form of the constraints: Heterogeneous dynamic constraints ; Heterogeneous performance constraints , ; Heterogeneous security constraints ; Environmental constraints , ; Constraint weights Adaptive adjustment based on heterogeneous characteristics For maximum thrust, For maximum speed, Let be the velocity vector of agent i at time t. Let be the geometric radii of agents i and j, respectively. The sum of the safety margins for agents i and j: To maintain a safe distance from obstacles.

[0067] Based on the above analysis, the main theoretical results of heterogeneous multi-agent distributed coordinated control are presented as follows: Theorem 2 (Convergence Theorem for Heterogeneous Multi-Agent Distributed Coordination Control) Consider by A multi-agent system composed of heterogeneous intelligent agents, employing a heterogeneous perceptual neural controller, with the control objective defined by the heterogeneous coordination learning and control part: ; in For comprehensive performance indicators, This is an enhancement for heterogeneous collaboration.

[0068] Assumptions: Based on Theorem 1, heterogeneous coordinated learning and control design are satisfied, the environmental complexity index is defined according to the environmental-driven topology reconstruction mechanism, and the heterogeneous metric satisfies the compatibility condition according to Lemma 2.

[0069] The distributed coordination control algorithm for heterogeneous multi-agent systems has the following convergence property: Heterogeneous Coordination Learning Algorithm Converges to Global Optimum : ; This is the set of system parameters at the k-th iteration.

[0070] Existing for a limited time Make: ; Among them, the precise convergence rate for: ; This is the minimum learning rate.

[0071] Example 2 Simulation experiment and result analysis: Table 1 shows the simulation experiment parameter configuration.

[0072] Table 1

[0073] Note: In the table The parameter values ​​represent the corresponding values ​​for agents 1 to 4, reflecting the heterogeneous nature of the system. All agents perform point-to-point formation flight missions from a stationary state to a stationary state.

[0074] To verify the effectiveness of the proposed heterogeneous multi-agent distributed deep control method, this embodiment designs a simulation experiment based on the parameter configuration in Table 1. The experimental task requires four heterogeneous quadrotor UAVs to cross obstacles in the x=0 plane from their initial formation position to reach the target position, achieving high-precision coordinated control while satisfying safe distance constraints.

[0075] The system reached the convergence criterion within a thousand iterations, and subsequent optimizations focused on refining the trajectory. Figure 3 As shown, five sets of experiments based on optimal loss verified the superior performance of the heterogeneous formation: four heterogeneous UAVs with 30% parameter differences achieved precise coordination in a complex constraint environment—autonomously compressing and traversing a 1.8m obstacle aperture from a standard 4m×4m formation, and then accurately reconstructing the target configuration. The heterogeneous perception controller effectively compensated for physical differences: the heavy-load unit utilized its inertia advantage to achieve a smooth trajectory, while the light-load unit maintained synchronization through an agile strategy, demonstrating excellent heterogeneous adaptability.

[0076] Figure 2 The study demonstrated the optimal trajectory of four heterogeneous UAVs performing coordinated formation flight in a geometrically constrained environment, verifying the effectiveness of autonomous reconfiguration capability and heterogeneous perception control.

[0077] Correlation matrix ( Figure 3 Quantitative analysis reveals the intrinsic control mechanism of heterogeneous multi-agent systems. Formation error shows a strong positive correlation with control effort (r=0.86), confirming the theoretical cost expectation of precise formation control and verifying the system's physical consistency. A key finding is the effective decoupling of target tracking and safety constraints: the correlation between target distance and minimum distance is only 0.09, breaking through the inherent limitation of target-obstacle avoidance conflict in traditional methods. The negative correlation between minimum distance and performance indicators reflects the essential trade-off between safety constraints and system optimization.

[0078] Figure 3 The correlation matrix was used to quantitatively analyze the intrinsic relationship between key performance indicators in a heterogeneous formation control system, and the effective decoupling mechanism between target tracking and safety constraints was verified.

[0079] To further verify the theoretical effectiveness of the proposed method, this paper conducts an in-depth analysis of the learning dynamics during the training process. Figure 4 The convergence trajectory of the formation error is shown, with the system rapidly converging from an initial error of 146.07m to the optimal state. Within the first 500 training rounds, the error sharply decreased to below 100m, representing an improvement of 31.6%. Subsequently, it entered a stable convergence phase, eventually stabilizing at 1.98m, achieving an overall improvement rate of 98.6%. This convergence pattern fully conforms to the exponential decay law predicted by Theorem 2, verifying the theoretical convergence guarantee of the heterogeneous coordinated control algorithm and demonstrating the rapid adaptability of the proposed method in complex constraint environments.

[0080] Figure 4 The convergence trajectory of the formation error during training is shown, verifying the exponential convergence characteristics and fast learning ability of the heterogeneous coordinated control algorithm.

[0081] Besides convergence performance, the ability to maintain safety constraints is a key indicator for heterogeneous formation control. Figure 5 The training evolution trajectory of the distance between agents is quantitatively demonstrated. In the early stages of training (0-500 rounds), the system exhibits significant distance control instability, with the minimum distance frequently violating the 0.8m safety threshold, highlighting the potential risks of an untrained system. As training progresses, the distance control mechanism gradually stabilizes, and the system autonomously learns the safety boundary constraints, with the average agent distance converging to the optimal range of 2.91m. Safety performance evaluation after training shows that although extreme distances of 0.285m may occur in local transients, the statistical safety violation rate is 0%, fully validating the robustness of the heterogeneous topology stable reconstruction mechanism in Theorem 1.

[0082] Figure 5 It demonstrates the dynamic evolution of the minimum distance between agents during training, which aligns with the learning ability and robustness of heterogeneous topology reconstruction mechanisms in maintaining security constraints.

[0083] Experimental results demonstrate that heterogeneous perceptual neural networks can automatically adapt to physical differences in intelligent agents. Under the same formation task, UAVs with different masses and thrust adopt differentiated control strategies: lighter UAVs respond more nimbly but require more precise control adjustments, while heavier UAVs move more stably but require advance planning. This adaptive characteristic is particularly evident in the obstacle crossing phase, where each UAV selects the optimal crossing trajectory based on its own capabilities, achieving coordinated unity within the heterogeneous system.

[0084] To comprehensively evaluate the performance advantages of the proposed method, this embodiment compares the heterogeneous sensing REN control method with two mainstream control methods: Model Predictive Control (MPC) and PID combined with artificial potential field method. These two methods were chosen because: MPC represents the mainstream technology of modern intelligent control, possessing strong prediction and optimization capabilities; PID and artificial potential field method are classic combinations of traditional control and obstacle avoidance techniques, widely used in engineering practice. Table 2 shows a comparison of the comprehensive performance of different control methods.

[0085] Table 2

[0086] Note: Task completion assessment is based on a comprehensive score of obstacle crossing success rate, formation maintenance accuracy, and target arrival accuracy; reasoning time is the calculation time for a single step of control; environmental adaptability refers to the ability to adapt to changes in obstacle configuration.

[0087] The comparison results show that each method exhibits different performance characteristics. While the MPC method possesses strong environmental adaptability and autonomous control capabilities, its inference time of 106,982 microseconds severely limits its real-time applications. This is mainly due to the fact that MPC requires solving complex constrained optimization problems in each control cycle, resulting in a heavy computational burden.

[0088] The PID and artificial potential field methods showed the best inference speed (1,090 microseconds) and were able to achieve obstacle avoidance and formation control under the specific environmental configuration of this experiment. However, these methods lacked autonomous control capabilities and required extensive manual parameter tuning for the current obstacle geometry and position. More importantly, the method faced significant challenges when the environment changed (such as obstacle relocation, hole size changes, or the addition of new obstacles): the potential field function needed to be redesigned, the PID parameters needed to be readjusted, and the method was prone to getting stuck in local optima, leading to formation jamming. This environmental dependence made the actual deployment cost of the method high and its adaptability poor.

[0089] The heterogeneous sensing neural controller method proposed in this invention achieves optimal task completion with an inference time of only 21,550 microseconds, a five-fold improvement over MPC, while maintaining strong environmental adaptability and high autonomous control capability. This performance advantage stems from two key factors: First, the recursive equilibrium structure of the network avoids online optimization, directly outputting control commands through neural network forward propagation; second, the heterogeneous sensing mechanism enables the controller to automatically adapt to the physical characteristics and environmental changes of different agents without manual parameter tuning. Although the method requires an offline training phase, once training is complete, it can adapt to various environmental configurations and exhibits good generalization ability.

[0090] Comprehensive analysis shows that the method of this invention achieves a good balance between real-time performance and adaptability while ensuring high task completion rates, making it particularly suitable for heterogeneous multi-agent coordinated control tasks with high real-time requirements. Compared with traditional methods, the method of this invention has significant advantages in autonomy, heterogeneous adaptability, and theoretical support, providing a more practical control solution for complex multi-agent systems.

[0091] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A multi-agent distributed coordination control method driven by heterogeneous perceptual neural networks, characterized in that, include: Step 1: Define a heterogeneous distributed nonlinear system composed of multiple heterogeneous agents, and establish a distributed heterogeneous multi-agent system model by establishing the dynamic equations of the heterogeneous agents and the communication topology. Step 2: Construct a heterogeneous depth controller for environmental perception. By dynamically adjusting the neighbor weights, multi-source heterogeneous information is fused and processed to output agent state information and environmental perception information. Step 3: Utilize the agent's state information and environmental perception information output by the controller to dynamically adjust the communication connection and reconstruct the communication topology; Step 4: Utilize the reconstructed stable communication topology to achieve distributed coordinated control of heterogeneous multi-agent systems and optimize the control strategy.

2. The multi-agent distributed coordination control method driven by a heterogeneous perceptual neural network according to claim 1, characterized in that, In step 1, the agent dynamics equation at time t+1 is expressed as: ； in, For including subsystems The set of nodes of the subsystem and its neighbors, at time t+1 each subsystem The state vector is The control input at time t is , For unknown process noise at time t, Given a heterogeneous set of parameters, different subsystems i have different masses. Geometric radius Maximum thrust and drag coefficient ; The global system dynamics equations at time t+1, consisting of all subsystems, are as follows: ； Wherein, the state vectors of the global system at time t and time t+1 are respectively , The control input of the global system at time t is The unknown process noise and heterogeneous parameter set of the global system at time t are respectively , .

3. The multi-agent distributed coordination control method driven by a heterogeneous perceptual neural network according to claim 2, characterized in that, In step 2, the intelligent agent The heterogeneous sensory neural controller is defined as a dynamic system: ； ； ； in, For the agent at time t Perceived information, This is a function for fusion of heterogeneous sensing information. For agents at time t and t+1 The internal state of the controller, Let be the state vector of agent j. For intelligent agents The state update function, For intelligent agents The control function, It is a set of obstacles and environments.

4. The multi-agent distributed coordination control method driven by a heterogeneous perceptual neural network according to claim 3, characterized in that, Define intelligent agents Layered perception information for: ， These represent individual tracking error, relative state of neighbors, formation configuration deviation, and environmental perception information, respectively.

5. The multi-agent distributed coordination control method driven by a heterogeneous perceptual neural network according to claim 4, characterized in that, The entire heterogeneous depth controller's heterogeneous perception and control network adopts a distributed architecture, where each agent only needs to communicate with its neighbors: ； Controller update equation: ； in This is relative state information. For the agent at time t and Heterogeneous weights between them.

6. The multi-agent distributed coordination control method driven by a heterogeneous perceptual neural network according to claim 5, characterized in that, In step 3, the network topology is controlled based on environmental perception information. Dynamically adjust to adapt to environmental constraints; Environmental weight for: ； Communication feasibility function for: ； For intelligent agents and Between moments Environmental weights; For safety constraint functions; For intelligent agents and The position vector; A collection of obstacle environments; For intelligent agents and The physical radius; LOS is the line-of-sight reachability determination function; For attenuation parameters; The distance function is the distance to the obstacle. Based on security constraints, define the security constraint function: ； The Euclidean distance between the two agents; For an additional safe distance; For the smoothing parameters of safety constraints.

7. The multi-agent distributed coordination control method driven by a heterogeneous perceptual neural network according to claim 6, characterized in that, Consider the topology reconstruction process of a heterogeneous multi-agent system in an environment. If the heterogeneous weights satisfy the following condition, the topology reconstruction process maintains system connectivity: Connectivity preservation conditions: ; It is the second smallest eigenvalue. It is the smallest eigenvalue; Heterogeneous compatibility conditions: ; For heterogeneous metrics, This is the critical value; Reconstruction smoothness condition: ; Network topology at time t+1 Network topology at time t The rate of topological change, The rate of change threshold; Environmental adaptability conditions: ; The minimum weight threshold, And converges to the optimal topology for environmental adaptation. satisfy: 。 8. The multi-agent distributed coordination control method driven by a heterogeneous perceptual neural network according to claim 7, characterized in that, Design a distributed state estimator for heterogeneous sensing, where each agent... Maintaining local state estimation: ； in This represents the posterior state estimate of agent i based on observations at time t. For heterogeneous Kalman estimation functions, For prior state estimation, For observation vectors; Neighbor state estimation through heterogeneous weight fusion : ； Fusion weights Adaptive adjustment based on estimation accuracy: ； in, This is the posterior state estimate of agent j based on observations at time t; Prior covariance matrix traces, It is a numerically stable term.

9. The multi-agent distributed coordination control method driven by a heterogeneous perceptual neural network according to claim 8, characterized in that, Objective function of heterogeneous multi-agent coordinated learning Defined as: ； in, For comprehensive performance indicators, This is an enhancement for heterogeneous collaboration; Converging to the global optimum : ； Existing for a limited time Make: ； Among them, the exact convergence rate is .

10. The multi-agent distributed coordination control method driven by a heterogeneous perceptual neural network according to claim 9, characterized in that, Establish coordinated control for heterogeneous multi-agent systems, unify the handling of multi-level constraints, and construct a comprehensive constraint function. : ； The constraints are numbered as follows: ,use To indicate the first The constraints are in functional form, and the constraint weights are... .

Images