Multi-agent cooperative control method based on graph theory

Abstract

The invention provides a multi-agent cooperative control method based on a graph theory, and the method comprises the steps: constructing a communication topology of each agent according to the graph theory, and generating a neighbor set; generating a weight matrix according to the neighbor set; generating a control signal according to the weight matrix and the state data of the neighbor agent; acquiring environment data and shared state data according to the control signal, and generating a state feature set; adjusting the motion trail of each agent according to the state feature set, and generating trail data; according to the trajectory data, all agents are coordinated to complete a collaborative task, and a consistency result is generated; and adjusting control parameters according to the consistency result, and generating a grouping consistency result.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a multi-agent collaborative control method based on graph theory. Background Art

[0002] Coordinated control of multi-agent systems in complex environments presents significant challenges. The system must achieve distributed consensus and formation control in a dynamically changing environment, while ensuring adaptability and robustness. A major technical challenge lies in achieving efficient coordination among multiple agents within limited communication resources and computing power. Specifically, agents must construct a global communication topology based on local information, but this topology can frequently change, making control signal calculation difficult. Furthermore, there is a conflict between agent state consistency and predetermined geometric formations, necessitating a balance between individual and global objectives. In complex environments, agents also need to process large amounts of environmental data and share state information, posing a significant challenge in extracting key features and optimizing motion trajectories. There is also a trade-off between leaderless consensus control and event-triggered communication strategies, balancing control effectiveness while reducing communication overhead. These challenges are particularly acute in practical applications such as drone swarms and robot formations, significantly impacting the system's collaborative performance and mission success. Resolving these technical conflicts and achieving efficient coordinated control of multi-agent systems in complex environments remains a critical and pressing issue. Summary of the Invention

[0003] The present invention provides a multi-agent collaborative control method based on graph theory, which mainly includes:

[0004] According to graph theory, the communication topology of each intelligent agent is constructed to generate a neighbor set; a weight matrix is generated based on the neighbor set; a control signal is generated based on the weight matrix and the state data of the neighbor intelligent agent; environmental data and shared state data are obtained based on the control signal to generate a state feature set; the motion trajectory of each intelligent agent is adjusted according to the state feature set to generate trajectory data; according to the trajectory data, each intelligent agent is coordinated to complete the collaborative task and generate a consistency result; according to the consistency result, the control parameters are adjusted to generate a group consistency result.

[0005] Furthermore, the communication topology of each intelligent agent is constructed according to graph theory to generate a neighbor set, including: representing the communication topology in the form of a directed graph, wherein nodes represent intelligent agents and edges represent communication links; determining the connection relationship of neighboring intelligent agents according to the communication range of each intelligent agent; and generating a neighbor set of each intelligent agent according to the connection relationship.

[0006] Furthermore, generating a weight matrix based on the neighbor set includes: determining the weight of each agent's neighbor based on the signal strength of the communication link; processing the weight using a normalization method to ensure that the sum of the weights is a preset value; and generating a weight matrix based on the weights.

[0007] Furthermore, the generating of the control signal based on the weight matrix and the state data of the neighboring agent includes: obtaining the state data of the neighboring agent; calculating the difference between the neighboring state and its own state by weighted average; and generating a first control signal based on the difference.

[0008] Furthermore, the obtaining of environmental data and shared status data according to the control signal and generating a status feature set includes: obtaining environmental data through local sensors, wherein the environmental data includes obstacle positions; obtaining shared status data from neighboring intelligent agents, wherein the shared status data includes neighbor positions; generating an environmental feature set based on the environmental data and the shared status data; and determining the current state and environmental features of each intelligent agent based on the environmental feature set to generate a status feature set.

[0009] Furthermore, the motion trajectory of each intelligent agent is adjusted according to the state feature set to generate trajectory data, including: determining the position and speed of each intelligent agent according to the state feature set; optimizing the motion trajectory through the control signal; generating speed and direction instructions according to the optimized motion trajectory; and generating first trajectory data according to the speed and direction instructions.

[0010] Furthermore, the method of coordinating the intelligent agents to complete collaborative tasks based on the trajectory data and generating a consistency result includes: detecting whether the communication topology switches; if the communication topology switches, regenerating the weight matrix and generating a second control signal; coordinating the intelligent agents to form a predetermined geometric shape based on the second control signal and the trajectory data; and determining whether the states of the intelligent agents converge to a common value based on the coordination result to generate a consistency result.

[0011] Furthermore, adjusting the control parameters according to the consistency result to generate a group consistency result includes: adjusting the control parameters according to the consistency result, and the control parameters are adaptively updated through state errors; generating a third control signal according to the adjusted control parameters; determining the fixed group to which each intelligent agent belongs according to the third control signal; judging whether the state difference within the same group tends to zero, and whether the states of different groups converge to different values, to generate a group consistency result.

[0012] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0013] The present invention discloses a multi-agent collaborative formation control method, which constructs a communication topology through graph theory, generates a neighbor set and a weight matrix, and calculates control signals to achieve agent state consistency. Based on environmental data and shared states, an environment and state feature set is generated to optimize the motion trajectory. The communication topology switch is detected and the control signal is recalculated to coordinate the agents to complete the formation task of a predetermined geometric shape. The state convergence is judged to achieve leaderless consistency, and an event-triggered control strategy is designed to adjust the communication frequency. The present invention solves the problem of collaborative control of multi-agent systems in complex environments, realizes distributed consistency and formation control, improves the adaptability and robustness of the system, and can be applied to scenarios such as drone clusters and robot formations, with important theoretical and practical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a flow chart of a multi-agent collaborative control method based on graph theory of the present invention. DETAILED DESCRIPTION

[0015] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0016] like Figure 1 In this embodiment, a multi-agent collaborative control method based on graph theory may specifically include:

[0017] Step S101: construct a multi-agent communication topology based on graph theory. The communication topology is represented in the form of a directed graph, where nodes represent agents and edges represent communication links, and a neighbor set is generated.

[0018] Obtain the initial position data and communication range parameters of the multi-agent system and construct a communication topology represented by a directed graph with agents as nodes. Define communication links based on the edges of the directed graph, calculate the outgoing and incoming edges of each agent node, and generate a neighbor set. Extract the state feature set of each agent from the neighbor set and determine the state feature matrix. Use the state feature matrix to calculate the relative position and velocity differences between agents to obtain a weighted communication matrix. If the weighted communication matrix meets the connectivity conditions, extract the global features of the topology through matrix decomposition to generate a topology feature vector. Based on the topology feature vector and the state feature set, optimize and generate a first control signal to determine the speed and direction instructions. Update the motion trajectory of each agent using the speed and direction instructions to generate first trajectory data. Obtain the first trajectory data, calculate the coordination deviation between agents, and obtain a deviation correction vector. Use the deviation correction vector to adjust the weight matrix of the communication topology to generate an updated control signal.

[0019] Specifically, the initial position data of the multi-agent system is obtained. For example, the coordinates of agents A, B, and C are (0, 0), (2, 1), and (1, 3), respectively. The communication range parameter is set to 3 meters. A directed graph with the agents as nodes is constructed. If the distance between A and B is 2.2 meters and less than the communication range, a communication link from A to B is established. Communication links are defined based on the edges of the directed graph. The outgoing and incoming edges of each agent node are calculated. For example, the outgoing edge of agent A is A→B, and the incoming edge is empty. This generates a neighbor set S106 = {B}. The state feature set of each agent is extracted from the neighbor set. For example, the state features of agent A include position (0, 0), velocity 0.5 m / s, and direction 30°. The state feature matrix M is determined as [0, 0, 0.5, 30; 2, 1, 0.8, 45; 1, 3, 0.6, 60]. The state feature matrix is used to calculate the relative position and velocity differences between agents. For example, the position difference between A and B is (2, 1) and the velocity difference is 0.3 m / s. This yields a weighted communication matrix W = [0, 0.7, 0; 0, 0, 0.6; 0.5, 0, 0]. If the weighted communication matrix satisfies the connectivity condition (i.e., the algebraic connectivity of the matrix is greater than 0), the global topological features are extracted through singular value decomposition, generating a topological feature vector V = [0.3, 0.5, 0.2]. Based on the topological feature vector and the state feature set, gradient descent is used to optimize the generation of the first control signal, determining a speed command of 0.7 m / s and a direction command of 40°. The speed and direction commands are used to update the trajectory of each agent. For example, if the new position of agent A is (0.35, 0.23), the first trajectory data T = [(0.35, 0.23), (2.4, 1.5), (1.2, 3.6)] is generated. Obtain the first trajectory data and calculate the coordination deviation between agents. For example, if the trajectory deviation between agents A and B is 0.2 meters, obtain the deviation correction vector D = [0.1, 0.2, 0.15]. Use the deviation correction vector to adjust the weight matrix of the communication topology. For example, adjust the weight of A→B in W from 0.7 to 0.8, generating the updated control signal U = [0, 0.8, 0; 0, 0, 0.6; 0.5, 0, 0].

[0020] Step S102: Calculate the weight of each agent's neighbors based on the neighbor set. The weight is determined based on the communication link signal strength, and a normalization method is used to generate a weight matrix.

[0021] Obtain the communication link signal strength data between each agent and other agents in the neighbor set and determine the signal strength value. Based on the signal strength value, calculate the original weights between each agent and its neighbors to obtain an initial weight set. Use a normalization method to process the initial weight set, generate normalized weight values, and ensure that the sum of the weights is 1. Based on the normalized weight values, construct a weight matrix between each agent and its neighbors to obtain a weight matrix. Using the weight matrix, analyze the communication topology and determine the connection relationship between each agent in the topology. If there are zero values in the weight matrix, introduce a default minimum weight value to fill it, and obtain an updated weight matrix. Based on the updated weight matrix, calculate the collaborative control parameters of each agent and generate control signal data. Using the control signal data, adjust the motion state of each agent and determine the real-time position of the collaborative formation. Obtain the adjusted position data, update the communication link signal strength of the neighbor set, and obtain a new signal strength value.

[0022] Specifically, the signal strength data of the communication link between each agent and other agents in the neighbor set is obtained to determine the signal strength values. For example, the signal strengths between agent A and neighbors B, C, and D are -60dBm, -70dBm, and -80dBm, respectively. Based on the signal strength values, the original weights between each agent and its neighbors are calculated to obtain an initial weight set. For example, using the inverse of the signal strength as the weight, the initial weights between agent A and neighbors B, C, and D are 1 / 60, 1 / 70, and 1 / 80, respectively. A normalization method is used to process the initial weight set to generate normalized weight values, and the sum of the weights is determined to be 1. For example, by dividing the initial weights by the sum of the weights, the normalized weights between agent A and neighbors B, C, and D are 0.45, 0.38, and 0.17, respectively. Based on the normalized weight values, a weight matrix is constructed between each agent and its neighbors to obtain a weight matrix. For example, in the matrix, the weights between agent A and neighbors B, C, and D are 0.45, 0.38, and 0.17, respectively. The weight matrix is used to analyze the communication topology and determine the connection relationships between each agent in the topology. For example, the connection strength between agent A and neighbor B is the highest, while the connection strength between agent C and neighbor D is the lowest. If there are zero values in the weight matrix, the default minimum weight value is introduced to fill it, resulting in an updated weight matrix. For example, the zero value is replaced with 0.01. Based on the updated weight matrix, the collaborative control parameters of each agent are calculated, and control signal data is generated. For example, the target position of agent A is calculated using the weighted average method. The motion state of each agent is adjusted based on the control signal data to determine the real-time position of the collaborative formation. For example, agent A moves to the target position according to the control signal. The adjusted position data is obtained, and the communication link signal strength of the neighbor set is updated to obtain a new signal strength value. For example, the signal strength between agent A and neighbor B is updated to -55dBm.

[0023] Step S103: Generate a first control signal for each agent based on the weight matrix and the state data of the neighboring agents. The first control signal is calculated by weighted averaging the difference between the neighboring agents' states and the agent's own state.

[0024] The state data of each agent and the state data of neighboring agents are obtained to obtain an agent state set. The state data of neighboring agents are weighted according to a weight matrix to determine a weighted neighbor state value. The difference between the agent's own state and the weighted neighbor state value is calculated to obtain a state difference set. The state difference set is processed using a weighted average method to generate a first control signal. If the value of the first control signal exceeds a preset range, it is normalized to obtain an adjusted first control signal. Based on the adjusted first control signal, combined with the weight matrix, the control input of each agent is calculated to determine a control input set. The state data of each agent is updated using the control input set to obtain a new state set. If the difference between the new state set and the previous state set is less than a threshold, the state convergence is confirmed and the final control input is determined. Based on the final control input, execution instructions for each agent are generated to obtain an instruction set.

[0025] Specifically, the state data of each agent and its neighboring agents are acquired, using a distributed sensor network to collect the agent's position, velocity, and acceleration information. For example, the state data of agent A is [1.2, 0.5, 0.1], and the state data of neighboring agent B is [1.5, 0.6, 0.2]. The neighboring agent's state data is weighted according to a weight matrix W = [0.6, 0.4]. The weighted neighboring state value is calculated as 1.5 × 0.6 + 0.6 × 0.4 = 1.14. The difference between the agent's own state and the weighted neighboring state value is calculated. For example, the difference in the x-direction state of agent A is 1.2 - 1.14 = 0.06. A weighted average method is used to process the state difference set. If the state difference set is [0.06, -0.1, 0.05] and the weights are [0.5, 0.3, 0.2], the first control signal is 0.06 × 0.5 + (-0.1) × 0.3 + 0.05 × 0.2 = 0.01. If the value of the first control signal exceeds the preset range [-0.5, 0.5], normalization is performed. For example, if the control signal is 0.6, it is adjusted to 0.5. Based on the adjusted first control signal and the weight matrix W = [0.6, 0.4], the control input is calculated as 0.5 × 0.6 + 0.4 × 0.4 = 0.46. The control input set is used to update the state data of each agent. For example, the new state of agent A is [1.2 + 0.46 × 0.1, 0.5 + 0.46 × 0.1, 0.1 + 0.46 × 0.1] = [1.246, 0.546, 0.146]. If the Euclidean distance difference between the new state set and the previous state set is less than the threshold of 0.01, the state converges and the final control input is determined to be 0.46. Based on this final control input, execution instructions are generated for each agent. For example, the instruction for agent A is [accelerate 0.46, turn 0.1].

[0026] Step S104: According to the first control signal, environmental data is acquired through local sensors, and shared state data is acquired from neighboring intelligent agents, wherein the environmental data includes obstacle positions, and an environmental feature set is generated.

[0027] Based on the first control signal, local sensors are activated to acquire environmental data, including obstacle locations. Shared state data is obtained from neighboring agents via a communication interface, including neighbor locations. A data fusion algorithm is used to merge the environmental data and shared state data to generate an initial environmental feature set. If there are missing data in the initial environmental feature set, an interpolation algorithm is used to fill in the missing data to obtain a complete environmental feature set. Based on the complete environmental feature set, a feature extraction algorithm is applied to extract key features of obstacle and neighbor locations, determining key feature vectors. A clustering algorithm is used to analyze the spatial distribution characteristics of the key feature vectors to obtain the spatial structure of the environment. If the spatial structure information contains outliers, a filtering algorithm is used to remove them to obtain optimized spatial structure information. Based on the optimized spatial structure information, an environmental feature topology map is constructed to generate a topological representation of the environmental feature set. This topological representation is encoded using a deep learning model to obtain the final environmental feature set.

[0028] Specifically, based on the first control signal, the local LiDAR sensor is activated to scan the surrounding environment at a frequency of 10Hz, acquiring environmental data including obstacle coordinates (x1 = 2.3m, y1 = 1.5m). Shared status data packets sent by neighboring agents are received via the 5G communication interface and parsed to obtain the neighbor's location coordinates (x2 = 4.1m, y2 = 0.8m). A Kalman filter algorithm is used to fuse the LiDAR data with the neighbor's location data to generate an initial environmental feature set containing the coordinates of obstacles and neighbors. If any coordinates are missing in the initial feature set, a cubic spline interpolation algorithm is used based on five adjacent frames of data to fill in the missing values. Principal component analysis is applied to the complete feature set to extract the relative distance (d = 2.4m) and azimuth (θ = 35°) between the obstacle and the neighbor as key feature vectors. The DBSCAN clustering algorithm is used with a neighborhood radius of ε = 1.2m, and the spatial distribution of the feature vectors is analyzed to form three clusters. If the clustering results contain outliers that deviate by 2σ from the mean, median filtering is used to replace the outlier coordinate values. Based on the optimized spatial data, an environmental topology graph is constructed with agents as nodes and relative distances as edges, and the connection weights between nodes are set to 1 / d 2 The topology map is input into the graph convolutional neural network, and after processing through three hidden layers, a 128-dimensional environment feature vector is output.

[0029] Step S105 , determining the current state and environmental features of each agent based on the environmental feature set, wherein the current state includes position and speed, and generating a state feature set.

[0030] The system obtains environmental feature data and extracts raw scene information for each agent, generating an initial dataset containing position, velocity, and obstacle distribution. Based on this initial dataset, the position and velocity information of each agent is separated to determine the current state of each agent. Based on the current state, the relative position of each agent relative to obstacles is calculated to obtain spatial relationship data between each agent and environmental obstacles. If the obstacle distribution density in the spatial relationship data exceeds a preset threshold, cluster analysis is performed on the obstacles to determine the cluster characteristics of each agent's neighboring obstacles. Based on these neighboring obstacle cluster characteristics, the environmental features of each agent are updated to obtain a subset of environmental features that includes the dynamic obstacle distribution. This subset of environmental features is combined with the position and velocity of each agent to generate a state feature set for each agent. If there are missing data in the state feature set, an interpolation algorithm is used to fill in the missing data to obtain a complete state feature set. Based on the complete state feature set, feature normalization is performed to determine a unified state feature representation for each agent. This unified state feature representation is then integrated with time series information to generate the final state feature set.

[0031] Specifically, the system extracts raw information about the agent's scene from the environmental feature data set. Position, velocity, and obstacle distribution data are collected using a lidar or visual sensor with a sampling frequency of 10Hz, resulting in an initial dataset consisting of 3D coordinates, velocity vectors, and obstacle point clouds. Based on this initial dataset, the Kalman filter algorithm is used to separate the position and velocity information of each agent. The state vector is set to [x, y, z, vx, vy, vz] to determine the current motion state of each agent. Based on the current state, the Euclidean distance between the agent and the obstacle is calculated. A KD-Tree algorithm is used to accelerate spatial queries, obtaining spatial relationship data for obstacles within a 5-meter radius around each agent. If the obstacle density exceeds 0.1 per cubic meter, the DBSCAN clustering algorithm is used with a neighborhood radius of ε = 1.5 meters and a minimum sample size of min_samples = 3 to determine the cluster center and boundary features of each agent's neighboring obstacles. Based on the clustering results, a subset of environmental features is updated, the motion trends of dynamic obstacles are marked, and linear regression is used to predict their position changes within the next 2 seconds. A state feature vector is constructed by combining a subset of environmental features with the agent's position and velocity. This vector contains 12 features, including the agent's coordinates, velocity, azimuth, and distance to the nearest obstacle. If a state feature contains missing values, cubic spline interpolation is used to fill in the missing time steps to ensure feature continuity. Based on the complete feature set, the Z-score normalization method is used to scale the data in each dimension to a distribution with a mean of 0 and a variance of 1, unifying the feature representations across different dimensions. The normalized features are then fused with historical data from the past five frames, and an LSTM network is used to extract temporal dependencies, producing the final state feature set.

[0032] Step S106: adjusting the motion trajectory of each agent according to the state feature set, wherein the motion trajectory generates speed and direction instructions by optimizing the first control signal to generate first trajectory data.

[0033] Acquire state feature set data, extract the position, velocity, and environmental parameters of the agent, and obtain feature vectors. Based on the feature vectors, calculate the initial motion trajectory of each agent and determine the preliminary direction and velocity values. Optimize the initial motion trajectory through the first control signal to generate adjusted speed and direction instructions. Using the optimized speed and direction instructions, calculate the first trajectory data to obtain the trajectory point sequence of the agent. If there is a conflict between the first trajectory data and the environmental data, adjust the control signal using the weight matrix to determine the new trajectory constraints. Based on the adjusted control signal, recalculate the motion trajectory of the agent to obtain an updated trajectory point sequence. Obtain the updated trajectory point sequence, perform collision detection in combination with the environmental data, and determine the safety of the trajectory. Based on the safety test results, optimize the parameters of the weight matrix and generate a new control signal. Using the new control signal, generate the final motion trajectory data to obtain a stable trajectory sequence of the agent.

[0034] Specifically, the multi-sensor fusion system acquires state feature data, including the agent's GPS coordinates (e.g., longitude 118.78°, latitude 32.04°), speed (2.5 m / s), and obstacle distance (3.2 m) detected by lidar. Principal component analysis (PCA) is used to reduce the dimensionality of the agent to generate a feature vector [0.45, -0.12, 0.67]. Based on the feature vector, the A* algorithm is used to calculate the initial motion trajectory, setting the initial speed threshold to 1.8 m / s and the heading angle to 45°. A PID controller is used to generate the first control signal (proportional coefficient Kp = 1.2, integral coefficient Ki = 0.05), optimizing the speed and direction commands to 2.1 m / s and 38°. The first trajectory data is generated using a cubic spline interpolation algorithm, outputting a trajectory point sequence [(x1, y1), (x2, y2)]. If the environmental data detects that the distance between a trajectory point (x1, y1) and an obstacle is less than 1 meter, the weight matrix W = [0.6, 0.3; 0.2, 0.8] is used to recalculate the control signal, constraining the curvature radius of the new trajectory to ≥ 2 meters. Model Predictive Control (MPC) is used to replan the trajectory, outputting the updated sequence [(x1', y1'), (x2', y2')]. Collision detection is performed in conjunction with the grid map. If the Euclidean distance between all trajectory points and obstacles is greater than 0.5 meters, the trajectory is deemed safe. Based on the safety detection results, the weight matrix W is updated to [0.55, 0.35; 0.25, 0.75] using gradient descent to generate the final control signal. Kalman filtering is used for smoothing, outputting the stable trajectory sequence [(x1'', y1''), (x2'', y2'')].

[0035] Step S107 : detecting a communication topology switch based on the first trajectory data, wherein the switching condition is determined based on a change in the communication range. If a switch occurs, recalculating the weight matrix and generating a second control signal.

[0036] Obtain first trajectory data and, by parsing the location information of the agents in the data, obtain a relative distance matrix between the agents. Based on the relative distance matrix, compare the distances between the agents with a preset communication range threshold to determine whether the communication topology has changed. If the communication topology has changed, reconstruct the adjacency matrix based on the new distance matrix to determine a new communication topology. Using the new communication topology, calculate the connection weights between the agents to obtain an updated weight matrix. Based on the updated weight matrix and the velocity and acceleration information of the first trajectory data, generate preliminary control signal parameters. Obtain preliminary control signal parameters, adjust the signal parameters through state error analysis, and obtain a second control signal. Based on the second control signal, detect the nonlinear dynamic response of the agent to determine whether a state deviation exists. If a state deviation exists, update the control parameters through an adaptive algorithm to obtain an adjusted control parameter set. Generate a third control signal using the adjusted control parameter set in combination with the consistency result.

[0037] Specifically, the first trajectory data is obtained, and by analyzing the location information of the agents in the data, the Euclidean distance formula is used to calculate the relative distance matrix between each agent. For example, the distance between agent A and agent B is 5 meters, and the distance between agent A and agent C is 8 meters. Based on the relative distance matrix, the distance between the agents is compared with the preset communication range threshold of 10 meters to determine whether the communication topology has changed. For example, if the distance between agent A and agent B is less than the threshold, the communication connection is maintained, while if the distance between agent A and agent C exceeds the threshold, the communication connection is disconnected. If the communication topology has changed, the adjacency matrix is reconstructed based on the new distance matrix. For example, the connection weight between agent A and agent C is set to 0, and the connection weight between agent A and agent B remains at 1. Based on the new communication topology structure, the Laplace matrix calculation method is used to calculate the connection weights between each agent to obtain an updated weight matrix. For example, the weight of agent A is 0.6, and the weight of agent B is 0.4. Based on the updated weight matrix and the velocity and acceleration information from the first trajectory data, a PID control algorithm is used to generate preliminary control signal parameters. For example, the control signal parameters for agent A are a velocity of 2 m / s and an acceleration of 1 m / s². After obtaining the preliminary control signal parameters, the state error analysis is performed and the least squares method is used to adjust the signal parameters to obtain a second control signal. For example, the adjusted control signal parameters are a velocity of 2.5 m / s and an acceleration of 0.8 m / s². Based on the second control signal, the nonlinear dynamic response of the agent is detected, and the Lyapunov stability theory is used to determine whether there is a state deviation. For example, the state deviation of agent A is 0.3. If a state deviation exists, the control parameters are updated using an adaptive algorithm, and the gradient descent method is used to obtain the adjusted control parameter set. For example, the updated control parameters are a velocity of 2.7 m / s and an acceleration of 0.7 m / s². Using this adjusted control parameter set and the consistency results, a distributed control algorithm is used to generate a third control signal. For example, the final control signal for agent A is a velocity of 2.7 m / s and an acceleration of 0.7 m / s².

[0038] Step S108: Coordinate the agents to complete a collaborative formation task based on the second control signal and the first trajectory data, wherein the formation task maintains a predetermined geometric shape and generates second trajectory data.

[0039] The second control signal and first trajectory data are obtained. By analyzing the signal content and trajectory parameters, the current state and expected motion instructions of each agent are determined. Based on the analyzed agent states and combined with the constraints of the predetermined geometric shape, the relative positions of each agent in the formation are calculated to determine the target formation configuration. If the target formation configuration deviates from the current state, the position parameters of each agent are adjusted using a geometric transformation algorithm to obtain preliminary formation coordination data. Based on the preliminary formation coordination data, the speed and direction instructions in the first trajectory data are integrated to generate motion adjustment vectors for each agent and determine the coordinated motion plan. The coordinated motion plan is obtained and the motion adjustment vectors are smoothed using an optimization algorithm to obtain continuous trajectory control instructions. If the continuous trajectory control instructions meet the predetermined geometric error threshold, they are assigned to each agent, generating local trajectory data for each agent. Based on the local trajectory data of each agent, a trajectory consistency check is performed, and the overall coordination of the formation is determined by comparing the trajectory parameters of adjacent agents. The verified local trajectory data are then used to synthesize the global motion paths of each agent to obtain the initial set of second trajectory data. According to the initial set, the data fusion algorithm is applied to refine the global motion path and generate the final second trajectory data.

[0040] Specifically, the system obtains the second control signal and first trajectory data. By analyzing the signal content and trajectory parameters, Fourier transform is used to extract the signal frequency characteristics. Trajectory noise is then processed using a Kalman filter algorithm to determine the current state and expected motion instructions for each agent. Based on the analyzed agent states and the constraints of the predetermined geometric shape, the Euclidean distance formula is used to calculate the relative position of each agent in the formation. The target formation configuration is then fitted using the least squares method. If the target formation configuration deviates from the current state, the position parameters of each agent are adjusted using an affine transformation algorithm. The coordinates are corrected using a rotation matrix and translation vectors to obtain preliminary formation coordination data. Based on this preliminary formation coordination data, the speed and direction instructions from the first trajectory data are integrated, and a PID controller is used to generate motion adjustment vectors for each agent. The coordinated motion plan is then determined using the Newton iteration method. Once the coordinated motion plan is obtained, the motion adjustment vectors are smoothed using a gradient descent optimization algorithm, and continuous trajectory control instructions are generated using cubic spline interpolation. If continuous trajectory control commands meet the predetermined geometric error threshold, they are distributed to each agent using a distributed algorithm, and local trajectory data for each agent is generated using Lagrange interpolation. Trajectory consistency checks are performed based on each agent's local trajectory data. By calculating the differences in trajectory parameters between adjacent agents, the covariance matrix is used to determine the overall coordination of the formation. The verified local trajectory data are then used to synthesize the global motion paths of each agent using the Dijkstra algorithm. A weighted average method is then used to generate an initial set of secondary trajectory data. Based on this initial set, the global motion paths are refined using the Kalman filter data fusion algorithm, and the final secondary trajectory data are generated using Gaussian smoothing.

[0041] Step S109: Based on the second trajectory data, determine whether the states of each agent have converged. If the state difference approaches zero and there is no leader, it is determined to be leaderless consistency, and a consistency result is generated.

[0042] Based on the second trajectory data, the state vectors of each agent are extracted. The differences between the state vectors are calculated to obtain a state difference set. If the absolute values of all differences in the state difference set are less than a preset threshold, the state differences are determined to be approaching zero, and a preliminary convergence judgment result is generated. Based on the preliminary convergence judgment result, the system configuration information is queried. If there is no preset leader identifier, the candidate is determined to be leaderless consistency and a consistency candidate flag is generated. If the consistency candidate flag is leaderless consistency, a time series analysis is performed on the second trajectory data, and the trend of the state differences is examined to obtain a time series stability result. Based on the time series stability result, if the state differences consistently approach zero within a preset time window, leaderless consistency is confirmed and a consistency result is generated. Based on the consistency result, the state update frequency of each agent is extracted and the frequency distribution is calculated to obtain a state synchronization metric. If the state synchronization metric is greater than the preset synchronization threshold, the state difference set is weighted averaged to generate a consistency strength index. Based on the consistency strength index and the sampling time interval of the second trajectory data, an interpolation algorithm is used to complete the missing state points, resulting in an optimized consistency dataset. Based on the optimized consistency data set, a structured report of the consistency results is generated, and the consistency judgment is output through the data encapsulation protocol.

[0043] Specifically, the state vectors of each agent are extracted from the second trajectory data. The state differences between adjacent agents are calculated using the Euclidean distance formula. For example, if the state vector of agent A is [1.2, 0.8] and the state vector of agent B is [1.1, 0.9], the state difference is √((1.2-1.1)²+(0.8-0.9)²)=0.1414. The differences between all agents are stored in a state difference set. If all values in the state difference set are less than the preset threshold of 0.2, the state differences are considered to be approaching zero, and a preliminary convergence judgment result is generated. Based on this result, the system configuration file is queried to check whether the leader_id field exists. If not, the candidate is marked as leaderless consistency. A sliding window analysis is performed on the second trajectory data with a window size of 5 seconds. The standard deviation of the state differences within the window is calculated. If the standard deviation is less than 0.05 for three consecutive windows, leaderless consistency is confirmed. The state update intervals of each agent are extracted, and the dominant frequency component is calculated using Fourier transform. If the dominant frequency difference of more than 90% of the agents is less than 0.1Hz, the state synchronization metric is determined to have met the standard. Time-attenuated weights are applied to the state difference set, with a weight of 0.6 for the last 3 seconds, 0.3 for data from 3-6 seconds, and 0.1 for data from 6 seconds ago. The weighted average is calculated as the consistency strength indicator. A cubic spline interpolation algorithm is used to fill in missing points in the trajectory data with a sampling interval of 0.5 seconds to ensure time series continuity. The consistency determination results are finally encapsulated in JSON format, including fields such as convergence type, strength indicator, and timing analysis data, and transmitted to the control center via the TCP protocol.

[0044] Step S1010: designing an event trigger control strategy based on the consistency result, wherein the trigger function is determined based on the state error and a preset threshold, generating a third control signal, and controlling the communication frequency to achieve group consistency.

[0045] 1. Based on the group consistency results, obtain the state data of each agent within a fixed group, determine the state difference within the same group and the state convergence value between different groups. 2. Based on the state difference and convergence value, calculate the state error of each agent relative to the group target state, and obtain a state error set. 3. If the error of any agent in the state error set exceeds a preset threshold, trigger the event detection process to determine whether the event triggering condition is met. 4. Based on the event triggering condition, generate a trigger function and determine the trigger function output value. 5. Based on the trigger function output value, generate a third control signal and obtain the control signal parameters. 6. Based on the third control signal parameters, adjust the communication frequency of each agent and determine the communication frequency configuration. 7. Based on the communication frequency configuration, update the state information exchange between the agents and obtain a new state data set. 8. Based on the new state data set, recalculate the state difference within the group and the convergence value between different groups to determine whether group consistency has been achieved. 9. If group consistency has not been achieved, return to the state error calculation process, iteratively generate the next round of third control signals, and determine the updated control strategy.

[0046] Specifically, based on the group consistency results, the state data of each agent within a fixed group is obtained. For example, the state values of agents A, B, and C in group 1 are [1.2, 1.5, 1.3], respectively, and the state values of agents D and E in group 2 are [2.1, 2.3], respectively. The state differences within the same group are determined, such as the difference in group 1 is [0.3, 0.1], and the state convergence values between different groups are [1.0, 2.2]. Using the state differences and convergence values, the state error of each agent relative to the group's target state is calculated. For example, if agent A's target state is 1.0 and its state error is 0.2, the state error set is [0.2, 0.5, 0.3, 0.1, 0.1]. If the error of any agent in the state error set exceeds the preset threshold of 0.4, such as agent B's error is 0.5, the event detection process is triggered to determine whether the event triggering condition is met. For example, the conditional judgment formula |e_i|>δ is used, where δ is 0.4. Based on the event trigger condition, a trigger function is generated, for example, the trigger function f(e_i) = e_i - δ, and the trigger function output value is determined to be 0.1. A third control signal is generated based on the trigger function output value, for example, the control signal u_i = k*f(e_i), where k is the gain factor 1.5, resulting in a control signal parameter of 0.15. Based on the third control signal parameter, the communication frequency of each agent is adjusted. For example, the communication frequency of agent B is reduced from 10 Hz to 8 Hz, and the communication frequency configuration is determined as [10 Hz, 8 Hz, 10 Hz, 10 Hz]. Based on the communication frequency configuration, the state information exchange between the agents is updated. For example, agent B updates its state value to 1.4 at an 8 Hz frequency, resulting in a new state data set of [1.2, 1.4, 1.3, 2.1, 2.3]. Based on the new state data set, the state differences within the group and the convergence values between different groups are recalculated. For example, the difference of group 1 is updated to [0.2, 0.1]. The group consistency is determined, for example, if the difference is less than 0.3. If group consistency is not achieved, the process returns to the state error calculation process, iteratively generates the next round of third control signals, and determines the updated control strategy, such as further adjusting the communication frequency to 7 Hz.

[0047] It will be apparent to those skilled in the art that the present application is not limited to the details of the exemplary embodiments described above and that the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the present application is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A multi-agent collaborative control method based on graph theory, characterized in that: include: Construct the communication topology of each agent based on graph theory and generate a set of neighbors; generating a weight matrix according to the neighbor set; generating a control signal according to the weight matrix and the state data of the neighboring agents; Acquire environmental data and shared state data according to the control signal to generate a state feature set; Adjust the motion trajectory of each agent according to the state feature set to generate trajectory data; Coordinate the agents to complete collaborative tasks based on the trajectory data and generate consistent results; Adjust control parameters according to the consistency result to generate a group consistency result.

2. The method according to claim 1, wherein The communication topology of each intelligent agent is constructed according to graph theory to generate a neighbor set, including: The communication topology is represented in the form of a directed graph, where nodes represent agents and edges represent communication links; Determine the connection relationship between neighboring agents based on the communication range of each agent; A neighbor set of each agent is generated according to the connection relationship.

3. The method according to claim 1, wherein Generating a weight matrix according to the neighbor set includes: Determine the weight of each agent’s neighbors based on the signal strength of the communication link; The weights are processed using a normalization method to ensure that the sum of the weights is a preset value; A weight matrix is generated according to the weights.

4. The method according to claim 1, wherein Generating a control signal according to the weight matrix and the state data of the neighboring agents includes: Get the status data of neighboring agents; Calculate the difference between the neighbor state and the own state by weighted average; A first control signal is generated according to the difference.

5. The method according to claim 1, wherein The acquiring of environmental data and shared state data according to the control signal to generate a state feature set includes: Acquiring environmental data through local sensors, wherein the environmental data includes obstacle locations; Acquire shared state data from neighboring agents, wherein the shared state data includes neighbor locations; generating an environmental feature set according to the environmental data and the shared state data; The current state and environmental characteristics of each agent are determined according to the environmental characteristic set to generate a state characteristic set.

6. The method according to claim 1, wherein The step of adjusting the motion trajectory of each agent according to the state feature set to generate trajectory data includes: Determine the position and velocity of each agent according to the state feature set; Optimizing the motion trajectory by using the control signal; Generate speed and direction instructions based on the optimized motion trajectory; First trajectory data is generated according to the speed and direction instructions.

7. The method according to claim 1, wherein The step of coordinating the agents to complete the collaborative task based on the trajectory data and generating a consistent result includes: Detect whether the communication topology has switched; If the communication topology switches, the weight matrix is regenerated and a second control signal is generated; coordinate the agents to form a predetermined geometric shape according to the second control signal and the trajectory data; Based on the coordination results, it is determined whether the states of each agent converge to a common value and a consistency result is generated.

8. The method according to claim 1, wherein The adjusting the control parameters according to the consistency result to generate a group consistency result includes: adjusting a control parameter according to the consistency result, the control parameter being adaptively updated through a state error; generating a third control signal according to the adjusted control parameter; Determining the fixed group to which each agent belongs according to the third control signal; Determine whether the state difference within the same group tends to zero, and whether the states of different groups converge to different values, and generate group consistency results.

Images