A multi-agent cooperative control method based on graph theory

Abstract

The application provides a multi-agent cooperative control method based on graph theory, comprising: constructing a communication topology of each agent according to graph theory to generate a neighbor set; generating a weight matrix according to the neighbor set; generating a control signal according to the weight matrix and state data of a neighbor agent; obtaining environment data and shared state data according to the control signal to generate a state feature set; adjusting a motion trajectory of each agent according to the state feature set to generate trajectory data; coordinating each agent to complete a cooperative task according to the trajectory data to generate a consistency result; and adjusting a control parameter according to the consistency result to generate a grouping consistency result.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and particularly relates to a multi-agent collaborative control method based on graph theory. BACKGROUND

[0002] The collaborative control of multi-agent systems in complex environments faces great challenges. The system needs to achieve distributed consistency and formation control in a dynamically changing environment, while ensuring adaptability and robustness. The main technical contradiction is how to achieve efficient collaboration between multiple agents under limited communication resources and computing power. Specifically, agents need to construct a global communication topology based on local information, but the topology structure may frequently switch, making it difficult to calculate control signals. In addition, there is a contradiction between agent state consistency and pre-determined geometric formation, which needs to balance individual and overall goals. In complex environments, agents also need to process a large amount of environmental data and shared state information, and how to extract key features and optimize motion trajectories is a major difficulty. There is also a trade-off between leaderless consensus control and event-triggered communication strategies, which needs to ensure control effectiveness while reducing communication burden. These problems are particularly prominent in practical application scenarios such as UAV clusters and robot formations, and have a significant impact on the collaborative performance and task completion effect of the system. How to effectively solve these technical contradictions and achieve efficient collaborative control of multi-agent systems in complex environments is a key problem that needs to be solved. SUMMARY

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

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

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

[0006] Further, the generating the weight matrix according to the neighbor set comprises: determining the weight of each agent neighbor according to the signal strength of the communication link; processing the weight by using a normalization method to ensure that the weight sum is a preset value; and generating the weight matrix according to the weight.

[0007] Further, the generating the control signal according to the weight matrix and the state data of the neighbor agent comprises: obtaining the state data of the neighbor agent; calculating the difference between the neighbor state and the self state by weighted average; and generating the first control signal according to the difference.

[0008] Further, the obtaining the environment data and the shared state data according to the control signal, and generating the state feature set, comprises: obtaining the environment data by a local sensor, wherein the environment data comprises obstacle positions; obtaining the shared state data from the neighbor agent, wherein the shared state data comprises neighbor positions; generating the environment feature set according to the environment data and the shared state data; and determining the current state and the environment feature of each agent according to the environment feature set, and generating the state feature set.

[0009] Further, the adjusting the motion trajectory of each agent according to the state feature set, and generating the trajectory data, comprises: determining the position and the speed of each agent according to the state feature set; optimizing the motion trajectory by the control signal; generating the speed and direction instruction according to the optimized motion trajectory; and generating the first trajectory data according to the speed and direction instruction.

[0010] Further, the coordinating each agent to complete the cooperative task according to the trajectory data, and generating the consistency result, comprises: detecting whether the communication topology is switched; if the communication topology is switched, generating the weight matrix and the second control signal again; coordinating each agent to form a predetermined geometric shape according to the second control signal and the trajectory data; and judging whether the state of each agent converges to a common value according to the coordination result, and generating the consistency result.

[0011] Further, the adjusting the control parameter according to the consistency result, and generating the grouping consistency result, comprises: adjusting the control parameter according to the consistency result, wherein the control parameter is updated adaptively according to the state error; generating the third control signal according to the adjusted control parameter; determining the fixed group to which each agent belongs according to the third control signal; and judging whether the state difference in the same group tends to zero and the state of different groups converges to different values, and generating the grouping consistency result.

[0012] The technical scheme provided by the embodiment of the application can have the following beneficial effects:

[0013] The application discloses a multi-agent cooperative formation control method, a communication topology is constructed through graph theory, a neighbor set and a weight matrix are generated, and a control signal is calculated to realize agent state consistency. According to environmental data and shared states, environmental and state feature sets are generated, and motion trajectories are optimized. The communication topology switching is detected and the control signal is recalculated to coordinate the agents to complete a formation task of a predetermined geometric shape. State convergence is judged to realize leaderless consistency, and an event-triggered control strategy is designed to adjust the communication frequency. The application solves the cooperative control problem of a multi-agent system in a complex environment, realizes distributed consistency and formation control, improves the adaptability and robustness of the system, and can be applied to scenes such as unmanned aerial vehicle clusters and robot formation, and has important theoretical and practical significance. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 A flowchart of a multi-agent cooperative control method based on graph theory. DETAILED DESCRIPTION

[0015] The technical solutions of the application will be described clearly and completely below in conjunction with embodiments. Obviously, the described embodiments are only some of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0016] As Figure 1 , the multi-agent cooperative control method based on graph theory specifically can include the following steps.

[0017] In step S101, a communication topology of multi-agents is constructed according to graph theory, the communication topology is represented in the form of a directed graph, a node represents an agent, an edge represents a communication link, and a neighbor set is generated.

[0018] The initial position data and the communication range parameters of the multi-agent system are acquired, and a communication topology represented in the form of a directed graph with agents as nodes is constructed. The communication link is defined according to the edges of the directed graph, the out-edge and the in-edge of each agent node are calculated, and a neighbor set is generated. The state feature set of each agent is extracted through the neighbor set, and a state feature matrix is determined. The relative position and the speed difference between agents are calculated by using the state feature matrix, and a weighted communication matrix is obtained. If the weighted communication matrix meets the connectivity condition, the global features of the topology are extracted by using a matrix decomposition method, and a topology feature vector is generated. According to the topology feature vector and the state feature set, a first control signal is optimized and generated, and speed and direction instructions are determined. The motion trajectory of each agent is updated through the speed and direction instructions, and first trajectory data is generated. The first trajectory data is acquired, the cooperative deviation between agents is calculated, and a deviation correction vector is obtained. The weight matrix of the communication topology is adjusted by using the deviation correction vector, and an updated control signal is generated.

[0019] Specifically, the initial position data of the multi-agent system is obtained, such as 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, the directed graph with agents as nodes is constructed, if the distance between A and B is 2.2 meters and less than the communication range, the communication link A→B is established. According to the edge definition of the directed graph, the out-edge and in-edge of each agent node are calculated, such as the out-edge of agent A is A→B and the in-edge is empty, and the neighbor set S106={B} is generated. The state feature set of each agent is extracted through the neighbor set, such as the state features of agent A include position (0, 0), speed 0.5 m / s and direction 30°, and the state feature matrix M=[0, 0, 0.5, 30; 2, 1, 0.8, 45; 1, 3, 0.6, 60] is determined. The relative position and speed difference between agents are calculated by using the state feature matrix, such as the position difference (2, 1) and the speed difference 0.3 m / s between A and B, and the weighted communication matrix W=[0, 0.7, 0; 0, 0, 0.6; 0.5, 0, 0] is obtained. If the weighted communication matrix satisfies the connectivity condition, i.e. the algebraic connectivity of the matrix is greater than 0, the global features of the topology are extracted by singular value decomposition, and the topology feature vector V=[0.3, 0.5, 0.2] is generated. According to the topology feature vector and the state feature set, the first control signal is generated by using gradient descent method, and the speed command 0.7 m / s and the direction command 40° are determined. The motion trajectory of each agent is updated by using the speed and direction commands, such as the new position of agent A is (0.35, 0.23), and the first trajectory data T=[(0.35, 0.23), (2.4, 1.5), (1.2, 3.6)] is generated. The coordination deviation between agents is calculated by using the first trajectory data, such as the trajectory deviation between A and B is 0.2 meters, and the deviation correction vector D=[0.1, 0.2, 0.15] is obtained. The weight matrix of the communication topology is adjusted by using the deviation correction vector, such as the weight of A→B in W is corrected from 0.7 to 0.8, and the updated control signal U=[0, 0.8, 0; 0, 0, 0.6; 0.5, 0, 0] is generated.

[0020] In step S102, the weight of each agent neighbor is calculated according to the neighbor set, and the weight is determined based on the communication link signal strength, and the weight matrix is generated by using the normalization method.

[0021] The communication link signal strength data of each agent and other agents in the neighbor set is obtained, and signal strength values are determined. According to the signal strength values, the original weights between each agent and the neighbors are calculated, and an initial weight set is obtained. The initial weight set is processed by a normalization method to generate normalized weight values, and the weight sum is determined to be 1. According to the normalized weight values, a weight matrix between each agent and the neighbors is constructed, and a weight matrix is obtained. Through the weight matrix, the communication topology structure is analyzed, and the connection relationship of each agent in the topology is determined. If there is a zero value in the weight matrix, a default minimum weight value is introduced to fill in, and an updated weight matrix is obtained. According to the updated weight matrix, the cooperative control parameters of each agent are calculated, and control signal data is generated. Through the control signal data, the motion state of each agent is adjusted, and the real-time position of the cooperative formation is determined. The adjusted position data is obtained, the communication link signal strength of the neighbor set is updated, and new signal strength values are obtained.

[0022] Specifically, the communication link signal strength data of each agent and other agents in the neighbor set is obtained, and signal strength values are determined, for example, the signal strengths of agent A and neighbors B, C, and D are-60 dBm, -70 dBm, and-80 dBm, respectively. According to the signal strength values, the original weights between each agent and the neighbors are calculated, and an initial weight set is obtained, for example, the inverse of the signal strength is used as the weight, and the initial weights of agent A and neighbors B, C, and D are 1 / 60, 1 / 70, and 1 / 80, respectively. The initial weight set is processed by a normalization method to generate normalized weight values, and the weight sum is determined to be 1, for example, the initial weights are divided by the total weight sum, and the normalized weights of agent A and neighbors B, C, and D are 0.45, 0.38, and 0.17, respectively. According to the normalized weight values, a weight matrix between each agent and the neighbors is constructed, and a weight matrix is obtained, for example, the weights of agent A and neighbors B, C, and D in the matrix are 0.45, 0.38, and 0.17, respectively. Through the weight matrix, the communication topology structure is analyzed, and the connection relationship of each agent in the topology is determined, for example, the connection strength between agent A and neighbor B is the highest, and the connection strength between agent C and neighbor D is the lowest. If there is a zero value in the weight matrix, a default minimum weight value is introduced to fill in, and an updated weight matrix is obtained, for example, the zero value is replaced by 0.01. According to the updated weight matrix, the cooperative control parameters of each agent are calculated, and control signal data is generated, for example, the target position of agent A is calculated by using the weighted average method. Through the control signal data, the motion state of each agent is adjusted, and the real-time position of the cooperative formation is determined, for example, agent A moves to the target position according to the control signal. The adjusted position data is obtained, the communication link signal strength of the neighbor set is updated, and new signal strength values are obtained, for example, the signal strength between agent A and neighbor B is updated to-55 dBm.

[0023] In step S103, a first control signal of each agent is generated according to the weight matrix and the state data of the neighbor agents, and the first control signal is calculated by weighted average of the difference between the neighbor state and the self state.

[0024] The state data of each agent and the state data of the neighbor agents are obtained to obtain an agent state set. The state data of the neighbor agents is weighted according to the weight matrix to determine a weighted neighbor state value. The difference between the self state and the weighted neighbor state value is calculated to obtain a state difference set. The state difference set is processed by a weighted average method to generate a first control signal. If the value of the first control signal exceeds a preset range, the first control signal is normalized to obtain an adjusted first control signal. The control input of each agent is calculated according to the adjusted first control signal and the weight matrix to determine a control input set. The state data of each agent is updated by 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 value, it is confirmed that the state converges, and the final control input is determined. The execution instruction of each agent is generated according to the final control input to obtain an instruction set.

[0025] Specifically, the state data of each agent and the state data of neighbor agents are acquired, and the position, speed and acceleration information of the agent are collected by a distributed sensor network, for example, the state data of agent A is [1.2, 0.5, 0.1], and the state data of neighbor agent B is [1.5, 0.6, 0.2]. According to the weight matrix, the state data of the neighbor agent is weighted, the weight matrix W = [0.6, 0.4], and the weighted neighbor state value is calculated as 1.5 x 0.6 + 0.6 x 0.4 = 1.14. By the state of the agent itself and the weighted neighbor state value, the difference between the two is calculated, for example, the x direction state difference of agent A is 1.2-1.14 = 0.06. The state difference set is processed by using the weighted average method, if the state difference set is [0.06, -0.1, 0.05], the weight is [0.5, 0.3, 0.2], then the first control signal is 0.06 x 0.5 + (-0.1) x 0.3 + 0.05 x 0.2 = 0.01. If the value of the first control signal exceeds the preset range [-0.5, 0.5], normalization processing is performed, for example, the control signal is adjusted to 0.5 when the control signal is 0.6. According to the adjusted first control signal, combined with the weight matrix W = [0.6, 0.4], the control input is calculated as 0.5 x 0.6 + 0.4 x 0.4 = 0.46. The state data of each agent is updated through the control input set, for example, the new state of agent A is [1.2 + 0.46 x 0.1, 0.5 + 0.46 x 0.1, 0.1 + 0.46 x 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 value 0.01, it is confirmed that the state converges, and the final control input is determined as 0.46. According to the final control input, the execution instruction of each agent is generated, for example, the instruction of agent A is [accelerate 0.46, turn 0.1].

[0026] In step S104, according to the first control signal, the environment data is acquired by the local sensor, and the shared state data is acquired from the neighbor agent, the environment data includes the position of the obstacle, and the environment feature set is generated.

[0027] According to the first control signal, the local sensor is activated to obtain environmental data containing obstacle positions. Through the communication interface, shared state data containing neighbor positions is obtained from the neighbor agent. A data fusion algorithm is used to merge the environmental data and the shared state data to generate an initial environmental feature set. If there is 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. According to the complete environmental feature set, a feature extraction algorithm is applied to extract key features of the obstacle positions and the neighbor positions, and a key feature vector is determined. Through a clustering algorithm, the spatial distribution characteristics in the key feature vector are analyzed to obtain spatial structure information of the environment. If the spatial structure information contains abnormal points, a filtering algorithm is used to remove the abnormal points to obtain optimized spatial structure information. According to the optimized spatial structure information, an environmental feature topology graph is constructed to generate a topological representation of the environmental feature set. Through a deep learning model, the topological representation is encoded to obtain the final environmental feature set.

[0028] Specifically, according to the first control signal, the local laser radar sensor is activated to scan the surrounding environment at a frequency of 10Hz to obtain environmental data containing obstacle coordinates (x1=2.3m, y1=1.5m). Through the 5G communication interface, the shared state data packet sent by the neighbor agent is received and parsed to obtain the neighbor position coordinates (x2=4.1m, y2=0.8m). The Kalman filter algorithm is used to fuse the laser radar data and the neighbor position data to generate an initial environmental feature set containing obstacle and neighbor coordinates. If there is missing coordinates in the initial feature set, the cubic spline interpolation algorithm is used to complete the missing values based on the adjacent 5 frames of data. The principal component analysis algorithm is applied to the complete feature set to extract the relative distance (d=2.4m) and the azimuth angle (θ=35°) between the obstacle and the neighbor as the key feature vector. Through the DBSCAN clustering algorithm, the neighborhood radius ε=1.2m is set to analyze the spatial distribution of the feature vector to form 3 clusters. If there are abnormal points deviating from the mean value 2σ in the clustering result, the median filter is used to replace the abnormal coordinate values. Based on the optimized spatial data, an environmental topology graph is constructed with the agent as the node and the relative distance as the edge, and the connection weight between nodes is set to 1 / d 2 . The topology graph is input into the graph convolutional neural network, and after 3 layers of hidden layer processing, a 128-dimensional environmental feature vector is output.

[0029] Step S105, according to the environmental feature set, the current state and environmental feature of each agent are determined, the current state includes position and velocity, and a state feature set is generated.

[0030] The environment feature set data is obtained, the original information of the scene where each agent is located is extracted, and the initial data set containing position, velocity and obstacle distribution is obtained. According to the initial data set, the position and velocity information of each agent is separated, and the current state of each agent is determined. Through the current state, the relative position of each agent relative to the obstacle is calculated, and the spatial relationship data of each agent and the environmental obstacle is obtained. If the obstacle distribution density in the spatial relationship data is higher than the preset threshold, the obstacle is clustered and analyzed, and the neighboring obstacle cluster characteristics of each agent are judged. According to the neighboring obstacle cluster characteristics, the environment features of each agent are updated to obtain the environment feature subset containing dynamic obstacle distribution. Through the environment feature subset, the position and velocity of each agent are combined to generate the state feature set of each agent. If there is missing data in the state feature set, an interpolation algorithm is used to fill the missing data to obtain a complete state feature set. According to the complete state feature set, a feature standardization process is performed to determine the unified state feature representation of each agent. Through the unified state feature representation, time series information is fused to generate the final state feature set.

[0031] Specifically, the original information of the scene where the agent is located is extracted from the environment feature set data, and the position, velocity and obstacle distribution data are collected by laser radar or visual sensor, the sampling frequency is set to 10Hz, and the initial data set containing three-dimensional coordinates, velocity vector and obstacle point cloud is obtained. According to the initial data set, the Kalman filter algorithm is used to separate the position and velocity information of each agent, the state vector is set as [x, y, z, vx, vy, vz], and the current motion state of each agent is determined. Through the current state, the Euclidean distance between the agent and the obstacle is calculated, and the KD-Tree is used to accelerate spatial query to obtain the obstacle spatial relationship data within 5 meters around each agent. If the obstacle distribution density exceeds 0.1 / m3, the DBSCAN clustering algorithm is used, the neighborhood radius ε is set to 1.5 meters, and the minimum sample number min_samples is set to 3, the clustering center and boundary characteristics of the neighboring obstacle of each agent are judged. According to the clustering result, the environment feature subset is updated, the motion trend of the dynamic obstacle is marked, and the linear regression is used to predict the position change within the next 2 seconds. Through the environment feature subset, the position and velocity of the agent are combined to construct the state feature vector, which contains 12-dimensional features such as agent coordinates, velocity magnitude, direction angle and nearest obstacle distance. If there is missing value in the state feature, the cubic spline interpolation is used to fill the missing data at the missing time step to ensure the continuity of the feature. According to the complete feature set, the Z-score standardization method is used to scale the data in each dimension to a distribution with a mean of 0 and a variance of 1, and the feature representation of different dimensions is unified. Through the standardized features, the past 5 frames of historical data are fused, the LSTM network is used to extract the time series dependence, and the final state feature set is output.

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

[0033] Obtain state feature set data, extract the position and velocity of the agent and the environmental parameters to obtain the feature vector. According to the feature vector, calculate the initial motion trajectory of each agent, and determine the preliminary direction and speed value. Through the first control signal, the initial motion trajectory is optimized to generate the adjusted speed and direction instruction. Using the optimized speed and direction instruction, the first trajectory data is calculated to obtain the trajectory point sequence of the agent. If there is a conflict between the first trajectory data and the environmental data, the weight matrix is used to adjust the control signal to judge the new trajectory constraint. According to the adjusted control signal, the motion trajectory of the agent is recalculated to obtain the updated trajectory point sequence. Obtain the updated trajectory point sequence, and combine the environmental data to perform collision detection to determine the safety of the trajectory. Through the safety detection result, the parameters of the weight matrix are optimized to generate a new control signal. Using the new control signal, the final motion trajectory data is generated to obtain the stable trajectory sequence of the agent.

[0034] Specifically, the state feature set data is obtained from the multi-sensor fusion system, including the GPS coordinates (such as longitude 118.78°, latitude 32.04°) of the agent, the speed (2.5 m / s) and the obstacle distance (3.2 m) detected by the laser radar. After dimension reduction by principal component analysis (PCA), the feature vector [0.45, -0.12, 0.67] is generated. Based on the feature vector, the A* algorithm is used to calculate the initial motion trajectory, and the initial speed threshold is set to 1.8 m / s and the heading angle is 45°. The PID controller is used to generate the first control signal (proportional coefficient Kp = 1.2, integral coefficient Ki = 0.05), and the adjusted speed and direction instruction are 2.1 m / s and 38°. The first trajectory data is generated by the cubic spline interpolation algorithm, and the trajectory point sequence [(x1, y1), (x2, y2)] is output. If the environmental data detects that the trajectory point (x1, y1) is less than 1 m from the obstacle, the weight matrix W = [0.6, 0.3; 0.2, 0.8] is called to recalculate the control signal, and the curvature radius of the new trajectory is constrained to be greater than or equal to 2 m. The model predictive control (MPC) is used to re-plan the trajectory, and the updated sequence [(x1', y1'), (x2', y2')] is output. Collision detection is performed in combination with the grid map, and if the Euclidean distance between all trajectory points and obstacles is greater than 0.5 m, it is determined to be safe. Based on the safety detection result, the weight matrix W is updated to [0.55, 0.35; 0.25, 0.75] using the gradient descent method to generate the final control signal. Through Kalman filtering smoothing processing, the stable trajectory sequence [(x1'', y1''), (x2'', y2'')] is output.

[0035] In step S107, a communication topology switching is detected according to the first trajectory data, and a switching condition is determined based on a communication range change. If the switching occurs, a weight matrix is recalculated, and a second control signal is generated.

[0036] The first trajectory data is obtained, the relative distance matrix between the agents is obtained by analyzing the position information of the agents in the data, and the distance between the agents is compared with the preset communication range threshold based on the relative distance matrix to determine whether the communication topology changes. If the communication topology changes, a new adjacency matrix is reconstructed based on the new distance matrix to determine a new communication topology structure. The connection weight between the agents is calculated through the new communication topology structure to obtain an updated weight matrix. The preliminary control signal parameters are generated based on the updated weight matrix and the speed and acceleration information of the first trajectory data. The preliminary control signal parameters are obtained, the signal parameters are adjusted through state error analysis to obtain a second control signal. The nonlinear dynamic response of the agent is detected based on the second control signal to determine whether there is a state deviation. If the state deviation exists, the control parameters are updated through an adaptive algorithm to obtain an adjusted control parameter set. The third control signal is generated based on the adjusted control parameter set and the consistency result.

[0037] Specifically, the first trajectory data is acquired, the relative distance matrix between each agent is calculated by analyzing the position information of the agents in the data using the Euclidean distance formula, 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. According to the relative distance matrix, the distance between the agents is compared with the preset communication range threshold 10 meters to determine whether the communication topology has changed, for example, the distance between agent A and agent B is less than the threshold, and the communication connection is maintained, while the distance between agent A and agent C exceeds the threshold, and the communication connection is disconnected. If the communication topology changes, 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 is maintained as 1. Through the new communication topology structure, the connection weight between each agent is calculated using the Laplacian matrix calculation method, and the updated weight matrix is obtained, for example, the weight of agent A is 0.6, and the weight of agent B is 0.4. According to the updated weight matrix, the speed and acceleration information of the first trajectory data is combined to generate preliminary control signal parameters using the PID control algorithm, for example, the control signal parameters of agent A are speed 2 meters / second and acceleration 1 meter / second2. The preliminary control signal parameters are acquired, the signal parameters are adjusted using the least squares method through state error analysis to obtain the second control signal, for example, the adjusted control signal parameters are speed 2.5 meters / second and acceleration 0.8 meters / second2. According to 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 the state deviation exists, the control parameters are updated through the adaptive algorithm, and the adjusted control parameter set is obtained using the gradient descent method, for example, the updated control parameters are speed 2.7 meters / second and acceleration 0.7 meters / second2. Through the adjusted control parameter set, the distributed control algorithm is used to generate the third control signal in combination with the consistency result, for example, the final control signal of agent A is speed 2.7 meters / second and acceleration 0.7 meters / second2.

[0038] In step S108, according to the second control signal and the first trajectory data, the agents are coordinated to complete the cooperative formation task, and the formation task maintains a predetermined geometric shape to generate second trajectory data.

[0039] The second control signal and the first trajectory data are acquired, and the current state and the expected motion instruction of each agent are obtained by analyzing the signal content and the trajectory parameter. According to the analyzed agent state, in combination with the constraint condition of the predetermined geometric shape, the relative position of each agent in the formation is calculated, and the target formation configuration is determined. If there is a deviation between the target formation configuration and the current state, the position parameter of each agent is adjusted through a geometric transformation algorithm to obtain preliminary formation coordination data. According to the preliminary formation coordination data, the speed and direction instructions in the first trajectory data are fused to generate the motion adjustment vector of each agent and determine the cooperative motion scheme. The cooperative motion scheme is acquired, and the motion adjustment vector is smoothed through an optimization algorithm to obtain continuous trajectory control instructions. If the continuous trajectory control instructions meet the error threshold of the predetermined geometric shape, they are distributed to each agent to generate the local trajectory data of each agent. According to the local trajectory data of each agent, trajectory consistency verification is performed, and the overall coordination of the formation is judged by comparing the trajectory parameters of adjacent agents. Through the verified local trajectory data, the global motion path of each agent is synthesized to obtain an initial set of the second trajectory data. According to the initial set, a data fusion algorithm is applied to refine the global motion path to generate the final second trajectory data.

[0040] Specifically, the second control signal and the first trajectory data are acquired, the signal content and the trajectory parameters are analyzed, the signal frequency characteristics are extracted by using Fourier transform, the trajectory noise is processed by using Kalman filtering algorithm, and the current state and the expected motion instruction of each agent are obtained. According to the analyzed state of the agent, the relative position of each agent in the formation is calculated by using the Euclidean distance formula combined with the constraint condition of the predetermined geometric shape, and the target formation configuration is fitted by using the least square method. If there is a deviation between the target formation configuration and the current state, the position parameters of each agent are adjusted by using the affine transformation algorithm, the coordinates are corrected by using the rotation matrix and the translation vector, and the preliminary formation coordination data are obtained. According to the preliminary formation coordination data, the speed and direction instructions in the first trajectory data are fused, the motion adjustment vector of each agent is generated by using the PID controller, and the cooperative motion scheme is determined by using the Newton iteration method. The cooperative motion scheme is acquired, the motion adjustment vector is smoothed by using the gradient descent optimization algorithm, and the continuous trajectory control instruction is generated by using the cubic spline interpolation. If the continuous trajectory control instruction meets the error threshold of the predetermined geometric shape, it is distributed to each agent by using the distributed algorithm, and the local trajectory data of each agent is generated by using the Lagrange interpolation. According to the local trajectory data of each agent, the trajectory consistency check is performed, the trajectory parameter difference of adjacent agents is calculated, the overall coordination of the formation is judged by using the covariance matrix. The global motion path of each agent is synthesized by using the Dijkstra algorithm according to the checked local trajectory data, and the initial set of the second trajectory data is obtained by using the weighted average method. According to the initial set, the Kalman filtering data fusion algorithm is applied to refine the global motion path, and the final second trajectory data is generated by using the Gaussian smoothing.

[0041] In step S109, according to the second trajectory data, it is judged whether the state of each agent converges. If the state difference tends to zero and there is no leader, it is determined that there is no leader consistency, and the consistency result is generated.

[0042] According to the second trajectory data, the state vectors of each agent are extracted, and the state difference set is obtained by calculating the difference between the state vectors. If all the absolute values of the state difference set are less than the preset threshold, it is determined that the state difference tends to zero, and a preliminary convergence judgment result is generated. According to the preliminary convergence judgment result, the system configuration information is queried, and if there is no preset leader identifier, it is determined as a leaderless consensus candidate, and a consensus candidate marker is generated. If the consensus candidate marker is leaderless consensus, the second trajectory data is analyzed in time sequence, and the time sequence stability result is obtained by checking the change trend of the state difference. According to the time sequence stability result, if the state difference continues to tend to zero within the preset time window, it is confirmed that there is no leader consensus, and a consensus result is generated. According to the consensus result, the state update frequency of each agent is extracted, and the state synchronization metric is obtained by calculating the frequency distribution. If the state synchronization metric is higher than the preset synchronization threshold, the state difference set is processed by weighted average, and a consensus strength indicator is generated. According to the consensus strength indicator, combined with the sampling time interval of the second trajectory data, the missing state points are completed by interpolation algorithm, and an optimized consensus data set is obtained. According to the optimized consensus data set, a structured report of the consensus result is generated, and the consensus judgment is output through the data encapsulation protocol.

[0043] Specifically, the state vectors of each agent are extracted from the second trajectory data, and the state difference between adjacent agents is calculated using the Euclidean distance formula. For example, the state vector of agent A is [1.2, 0.8], and the state vector of agent B is [1.1, 0.9], then the state difference is √((1.2-1.1)2+(0.8-0.9)2) = 0.1414. The difference between all agents is stored in the state difference set. If the values in the state difference set are all less than the preset threshold 0.2, it is determined that the state difference tends to zero, and a preliminary convergence judgment result is generated. According to the result, the system configuration file is queried to detect whether there is a leader_id field, and if there is no such field, it is marked as a leaderless consensus candidate. The second trajectory data is analyzed in a sliding window, and the window size is 5 seconds. The standard deviation of the state difference in the window is calculated, and if the standard deviation is continuously lower than 0.05 for 3 consecutive windows, it is confirmed that there is no leader consensus. The state update interval of each agent is extracted, and the main frequency component is calculated using Fourier transform. If the main frequency difference of more than 90% of the agents is less than 0.1 Hz, it is determined that the state synchronization metric meets the requirements. Time decay weights are applied to the state difference set, with the latest 3 seconds of data having a weight of 0.6, 3-6 seconds of data having a weight of 0.3, and 6 seconds or more of data having a weight of 0.1. The weighted average value is calculated as a consensus strength indicator. A cubic spline interpolation algorithm is used to complete the missing points in the trajectory data with a sampling interval of 0.5 seconds, ensuring the continuity of the time sequence. Finally, the consensus judgment result is encapsulated in JSON format, including fields such as convergence type, strength indicator, and time sequence analysis data, and transmitted to the control center through the TCP protocol.

[0044] Step S1010, according to the consistency result, design event trigger control strategy, the trigger function is determined based on state error and preset threshold, generate the third control signal, control the communication frequency to realize the grouping consistency.

[0045] 1According to the grouping consistency result, obtain the state data of each agent in the fixed grouping, determine the state difference value in the same grouping and the state convergence value between different groupings.2Through the state difference value and the convergence value, calculate the state error of each agent relative to the grouping target state to obtain a state error set.3If the error of any agent in the state error set exceeds the preset threshold, trigger the event detection process to determine whether the event trigger condition is met.4According to the event trigger condition, generate a trigger function to determine the trigger function output value.5Generate a third control signal through the trigger function output value to obtain a control signal parameter.6Adjust the communication frequency of each agent according to the third control signal parameter to determine the communication frequency configuration.7Update the state information interaction between each agent through the communication frequency configuration to obtain a new state data set.8Recalculate the state difference value in the grouping and the convergence value between different groupings according to the new state data set to determine whether the grouping consistency is achieved.9If the grouping consistency is not achieved, return to the state error calculation process to iteratively generate the next round of third control signal to determine the updated control strategy.

[0046] Specifically, according to the grouping consistency result, the state data of each agent in the 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, the state values of agents D and E in group 2 are [2.1, 2.3] respectively, the state difference value in the same group is determined, such as the difference value of group 1 is [0.3, 0.1], and the state convergence value between different groups is [1.0, 2.2]. Through the state difference value and the convergence value, the state error of each agent relative to the target state of the group is calculated, for example, the target state of agent A is 1.0, and the state error is 0.2, and the state error set [0.2, 0.5, 0.3, 0.1, 0.1] is obtained. If the error of any agent in the state error set exceeds the preset threshold 0.4, for example, the error of agent B is 0.5, the event detection process is triggered, and whether the event triggering condition is met is judged, for example, the conditional judgment formula |e_i|>δ is used, where δ is 0.4. According to the event triggering condition, the trigger function is generated, for example, the trigger function f(e_i)=e_i-δ, and the output value of the trigger function is determined as 0.1. Through the output value of the trigger function, the third control signal is generated, for example, the control signal u_i=k*f(e_i), where k is the gain coefficient 1.5, and the control signal parameter 0.15 is obtained. According to the third control signal parameter, the communication frequency of each agent is adjusted, for example, the communication frequency of agent B is reduced from 10Hz to 8Hz, and the communication frequency configuration [10Hz, 8Hz, 10Hz, 10Hz, 10Hz] is determined. Through the communication frequency configuration, the state information interaction between each agent is updated, for example, the state value of agent B is updated to 1.4 at a frequency of 8Hz, and a new state data set [1.2, 1.4, 1.3, 2.1, 2.3] is obtained. According to the new state data set, the state difference value in the group and the convergence value between different groups are recalculated, for example, the difference value of group 1 is updated to [0.2, 0.1], and whether the grouping consistency is achieved is judged, for example, the difference value is less than 0.3. If the grouping consistency is not achieved, the state error calculation process is returned, the next round of third control signal is iteratively generated, and the updated control strategy is determined, for example, the communication frequency is further adjusted to 7Hz.

[0047] It is apparent for those skilled in the art that the present application is not limited to the details of the foregoing exemplary embodiments, and the present application can be implemented in other concrete forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be considered in all aspects as illustrative and not restrictive, and the scope of the present application is defined by the appended claims rather than the foregoing description, and it is intended to embrace all changes falling within the meaning and range of equivalents of the claims. Any reference signs in the claims should not be considered as limiting the claims involved.

Claims

1. A multi-agent cooperative control method based on graph theory, characterized in that, include: Construct the communication topology of each agent based on graph theory, and generate a neighbor set; Generate a weight matrix based on the neighbor set; Control signals are generated based on the weight matrix and the state data of neighboring agents; Based on the control signals, environmental data and shared state data are acquired, and a state feature set is generated; Adjust the motion trajectory of each agent according to the state feature set to generate trajectory data; Based on the trajectory data, coordinate the various intelligent agents to complete collaborative tasks and generate consistent results; Adjust the control parameters based on the consistency results to generate group consistency results.

2. The method as described in claim 1, characterized in that, The step of constructing the communication topology of each agent based on graph theory and generating a neighbor set includes: The communication topology is represented in the form of a directed graph, where nodes represent agents and edges represent communication links; Determine the connection relationships between neighboring agents based on the communication range of each agent; Based on the connection relationships, a neighbor set is generated for each agent.

3. The method as described in claim 1, characterized in that, The step of generating a weight matrix based on the neighbor set includes: The weights of each agent's neighbors are determined 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; Generate a weight matrix based on the weights.

4. The method as described in claim 1, characterized in that, The step of generating control signals based on the weight matrix and the state data of neighboring agents includes: Obtain the state data of neighboring intelligent agents; The difference between the neighbor's state and its own state is calculated by weighted average. A first control signal is generated based on the difference.

5. The method as described in claim 1, characterized in that, The step of acquiring environmental data and shared state data based on the control signal and generating a state feature set includes: Environmental data is acquired through local sensors, including the location of obstacles. Obtain shared state data from neighboring intelligent agents, the shared state data including neighbor locations; An environmental feature set is generated based on the environmental data and the shared state data; Based on the environmental feature set, the current state and environmental characteristics of each agent are determined, and a state feature set is generated.

6. The method as described in claim 1, characterized in that, The step of adjusting the motion trajectory of each agent according to the state feature set and generating trajectory data includes: The position and velocity of each agent are determined based on the state feature set; The motion trajectory is optimized using the control signals; Generate speed and direction commands based on the optimized motion trajectory; First trajectory data is generated based on the speed and direction commands.

7. The method as described in claim 1, characterized in that, The step of coordinating the various 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 changes, the weight matrix is ​​regenerated and a second control signal is generated. Based on the second control signal and the trajectory data, the intelligent agents are coordinated to form a predetermined geometric shape; Based on the coordination results, it is determined whether the states of each agent have converged to a common value, and a consistent result is generated.

8. The method as described in claim 1, characterized in that, The step of adjusting the control parameters based on the consistency result to generate group consistency results includes: The control parameters are adjusted based on the consistency results, and the control parameters are adaptively updated through state errors. A third control signal is generated based on the adjusted control parameters; The fixed group to which each agent belongs is determined based on the third control signal; Determine whether the state difference within the same group approaches zero, and whether the states of different groups converge to different values ​​to generate a group consistency result.

Images