Border hierarchical intervention control method based on multi-agent system

Abstract

The present application relates to computer vision, space-time data mining and multi-agent system technical field, specifically to a boundary hierarchical intervention control method based on multi-agent system, comprising: obtaining preset physical constraint boundary information and initial communication topology weight matrix, collecting kinematic perception data of each agent node and outputting smooth state vector; solving boundary distance function; dividing the running state of the current agent node into a safety zone, a warning intervention zone and a limit intervention zone; generating a dynamic communication topology weight matrix in the warning intervention zone, reducing the weight of received neighbor information and increasing the weight of sent own state information; in the limit intervention zone, constructing a dissipative control term, and fusing a nominal cooperative control law, a dynamic communication topology weight matrix and a control barrier function to generate an adaptive control instruction; output to the execution unit and broadcast to the neighbor; the present application realizes the consideration of boundary safety, formation continuity and execution smoothness.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, spatiotemporal data mining and multi-agent systems, specifically a boundary hierarchical intervention control method based on multi-agent systems. Background Technology

[0002] In existing multi-agent collaborative control systems, there are usually multiple agent nodes, communication topology and nominal collaborative control law. Multiple agent nodes form a distributed system after state interaction through preset communication relationships. This distributed system is often used in the security management and control of large passenger flow in smart parks.

[0003] For security management of large crowds in smart parks, traditional grid management and manual intercom dispatching models have the following technical shortcomings: uneven spatial distribution of system control nodes, unable to dynamically respond to time-varying fluctuations in target density within the work area, leading to node failure or system overload in congested areas, while control units in open areas are not effectively utilized; defects in spatial discretization modeling, with traditional orthogonal matrix grid division having inherent errors in calculating crowd diffusion vectors, failing to consider the amplifying effect of terrain factors such as passage width, slope, and distance to hazard sources on risks; heterogeneous multimodal features and prediction lag, with monitoring video, wireless probe, and physical gate data being independent and lacking data interaction mapping, resulting in a lack of flow direction prediction capabilities based on physical topology; and discontinuous dispatching response, with response speed relying on manual command unable to keep up with the speed of risk evolution, and a lack of automated coordination mechanisms for air and ground security resources.

[0004] Existing systems typically determine proximity to physical boundaries based on node location or simple distance thresholds. In complex terrain, if the system directly executes the control logic of manual intervention upon anomaly detection to collect the status of nearby targets, it increases the risk of collisions in the group's operational trajectories and makes it difficult to consider the physical flow constraints within the operational airspace. Furthermore, the single threshold switching method can easily cause abrupt changes in control commands and incoordination in the neighborhood response, leading to increased risk of boundary crossings near the boundary, decreased formation stability, and cascading oscillations caused by local interventions. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a boundary-level hierarchical intervention control method based on a multi-agent system. This method is applied to a multi-agent system comprising multiple agent nodes. Specifically, the technical solution of this invention includes: Step 1: Obtain the preset physical constraint boundary information, initial communication topology weight matrix, preset nominal cooperative control law and pre-built control obstacle function, collect multimodal perception data of each agent node and perform state filtering, and output the smooth state vector of each agent node. Step 2: Taking any one of the multiple agent nodes as the current agent node, calculate the output value of the boundary distance function constructed by the current agent node based on the smoothed state vector; Step 3: Based on the output value of the boundary distance function, the current operating state of the agent node is divided into a safe zone, a warning intervention zone, and a limit intervention zone. When in the safe zone, the initial communication topology weight matrix is ​​maintained. When in the warning intervention zone, the initial communication topology weight matrix is ​​adjusted to generate a dynamic communication topology weight matrix. When in the limit intervention zone, a dissipative control term is constructed based on the fractional derivative of the boundary distance function, and the nominal cooperative control law, the dynamic communication topology weight matrix, and the control barrier function are fused to generate an adaptive control command. Step four: Output the adaptive control command to the execution unit configured in the current agent node, add a system clock timestamp to the dynamic communication topology weight matrix, and broadcast the dynamic communication topology weight matrix with timestamp and the smoothed state vector of the current agent node to neighboring nodes.

[0006] Preferably, when in the early warning intervention zone, adjusting the initial communication topology weight matrix to generate a dynamic communication topology weight matrix includes: adjusting the initial communication topology weight matrix through continuous mapping based on the output value of the boundary distance function to generate the dynamic communication topology weight matrix.

[0007] Preferably, the process of collecting multimodal perception data from each agent node and performing state filtering to output a smooth state vector includes: pre-constructing a spatiotemporal topology model based on a hexagonal discrete global grid; acquiring the multimodal perception data, which includes video surveillance target detection data, physical node traffic density perception data, and wireless probe density data from each agent node; spatiotemporally aligning the multimodal perception data under a unified timestamp and then inputting it into a spatiotemporal graph convolutional network and an extended Kalman filter algorithm for data fusion; and estimating state parameters using the extended Kalman filter algorithm to output a smooth state vector containing the motion rate and heading angle of dynamic agent nodes and the congestion state of static agent nodes.

[0008] Preferably, step two, which calculates the output value of the boundary distance function constructed by the current agent node, includes: extracting the motion rate and heading angle of the dynamic agent node itself from the smoothed state vector; substituting the motion rate and heading angle of the agent node itself into the spatiotemporal topology model based on a hexagonal discrete global grid to construct forward and lateral dynamic risk pressure prediction vectors; and calculating the shortest distance from the endpoint of the dynamic risk pressure prediction vector to the preset physical constraint boundary as the output value.

[0009] Preferably, step three divides the operating state into a safe zone, a warning intervention zone, and a limit intervention zone, including: setting a first distance threshold and a second distance threshold, wherein the first distance threshold is greater than the second distance threshold; when the output value is greater than the first distance threshold, it is determined that the system is in the safe zone; when the output value is less than or equal to the first distance threshold and greater than the second distance threshold, it is determined that the system is in the warning intervention zone; when the output value is less than or equal to the second distance threshold, it is determined that the system is in the limit intervention zone.

[0010] Preferably, the first distance threshold is determined based on the current agent node's movement rate, time penalty term, maximum communication delay error margin, and safety buffer distance; the second distance threshold is determined based on the maximum scheduling response distance of the multi-agent system and the limit diffusion distance caused by hardware response delay.

[0011] Preferably, adjusting the initial communication topology weight matrix through continuous mapping includes: using the output value of the boundary distance function of the current agent node as the independent variable, adjusting the initial communication topology weight matrix using a preset continuous attenuation and gain mapping relationship, obtaining attenuation weights and gain weights, and generating the dynamic communication topology weight matrix.

[0012] Preferably, when in the extreme intervention zone, generating an adaptive control command includes: constructing the dissipative control term based on the fractional derivative of the boundary distance function with respect to time, wherein the dissipative control term includes a preset gain coefficient; and, under the non-negative evolution constraint of the control barrier function, using a quadratic programming algorithm to fuse the nominal cooperative control law and the dissipative control term to obtain the adaptive control command.

[0013] Preferably, outputting the adaptive control command to the execution unit configured in the current agent node includes: converting the adaptive control command into a drive signal for the execution unit; encapsulating the smoothed state vector of the current agent node with the dynamic communication topology weight matrix, and adding a timestamp from the system clock to generate a broadcast data packet; and sending the broadcast data packet to the neighboring nodes within the communication topology range through a preset local communication network.

[0014] Preferably, the method further includes: obtaining the desired state vector output by the multi-agent system under a preset nominal cooperative control law, and calculating a global state error vector in combination with the smoothed state vector; constructing a global Lyapunov candidate function based on the global state error vector and the dynamic communication topology weight matrix; when the time derivative of the candidate function is greater than zero, positively adjusting the gain coefficient of the dissipative control term, and limiting the rate of change of the dynamic communication topology weight matrix to not exceed a preset smoothed derivative threshold; when the time derivative is less than or equal to zero, stopping the upward adjustment of the gain coefficient and maintaining the current gain coefficient.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention acquires preset physical constraint boundary information corresponding one-to-one with the actual terrain factors and channel vulnerability of a smart park, establishes an initial communication topology weight matrix, and performs cross-modal spatiotemporal alignment and spatiotemporal graph convolutional network and extended Kalman filter fusion on video surveillance target detection data, physical node traffic density perception data, and wireless probe density data under a unified timestamp. The output is a smooth state vector containing the motion rate and heading angle of dynamic agent nodes and the congestion state of static agent nodes. This can effectively reduce the impact of disturbances such as abnormal positioning jumps, lack of interactive mapping of multi-source data, and multipath reflection on system judgment in wide-area complex terrain environments, providing a continuous and reliable state basis for subsequent boundary risk identification, realizing quantitative risk assessment and predictive deployment, and significantly reducing the incidence of safety accidents. 2. This invention continuously reconstructs the initial communication topology weight matrix within the early warning intervention zone using a negative exponential mapping function based on the boundary distance function. This reduces the weight of risk nodes receiving state information from neighboring nodes and increases the weight of them sending their own state information to neighboring nodes. This enables nodes approaching the boundary to break free from the continuous following pull under the fixed communication topology, while also allowing local risk avoidance intentions to spread to neighboring nodes more quickly. This improves the formation configuration mismatch and response mismatch problems caused by fixed communication relationships in existing systems. 3. This invention introduces a dissipative control term based on the fractional derivative of the boundary distance function within the extreme intervention zone, integrates it with the nominal cooperative control law using quadratic programming, and generates adaptive control commands by combining the hard constraints of the control obstacle function. This enables continuous coordination of drone takeoff guidance, security personnel interception, and IoT gate flow restriction without exceeding the preset physical constraint boundaries. It avoids the sudden changes in execution unit control and corresponding security risks caused by directly executing traditional centralized open-loop intervention commands, solves the technical contradiction between local node computing power overload and cooperative control response delay, and realizes the transformation of system control logic to predictive proactive intervention. 4. By issuing adaptive control commands to the execution unit and simultaneously broadcasting a smooth state vector carrying a timestamp and a dynamic communication topology weight matrix to neighboring nodes, this invention enables neighboring nodes to accurately identify the control cycle to which the information belongs and adjust their following strategies in a timely manner, thus avoiding the amplification of local communication delays and cooperative instability caused by outdated states being followed incorrectly. 5. This invention limits the rate of change of weights throughout the entire process of updating the dynamic communication topology weight matrix and generating control instructions. It also constructs a global Lyapunov candidate function based on the continuously mapped dynamic communication topology weight matrix and adjusts the gain of the dissipative control term in a coordinated manner so that its derivative is always less than or equal to zero. This allows communication layer reconstruction and control layer intervention to be incorporated into a unified closed-loop stability framework, overcoming the potential transient instability caused by dynamic topology changes. This achieves a balance between boundary security, formation continuity, and global stability, effectively reducing the data load of communication interaction between system control terminals and improving the accuracy and response speed of resource scheduling. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and are used to explain the invention. They do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating the boundary hierarchical intervention control method based on a multi-agent system provided in this application embodiment. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0018] A boundary-level intervention control method based on multi-agent systems is applied to multi-agent systems containing multiple agent nodes. The specific steps include: Step 1: Obtain the preset physical constraint boundary information, initial communication topology weight matrix, nominal cooperative control law and control obstacle function, collect multimodal perception data of each intelligent agent node and perform state filtering, and output a smooth state vector; Step 2: Taking any one of the multiple agent nodes as the current agent node, calculate the output value of the boundary distance function constructed by the current agent node based on the smoothed state vector; Step 3: Based on the output value, divide the current operating state of the agent node into a safe zone, a warning intervention zone, and a limit intervention zone; when in the safe zone, maintain the initial communication topology weight matrix; when in the warning intervention zone, adjust the initial communication topology weight matrix to generate a dynamic communication topology weight matrix; when in the limit intervention zone, construct a dissipative control term based on the fractional derivative of the boundary distance function, and integrate the nominal cooperative control law, the dynamic communication topology weight matrix, and the control obstacle function to generate an adaptive control command; Step four: Output the adaptive control command to the execution unit, and broadcast the timestamped dynamic communication topology weight matrix and the smoothed state vector to neighboring nodes.

[0019] This embodiment provides a boundary-level intervention control mechanism based on a multi-agent system, such as... Figure 1 As shown; specifically, this method is deployed in a collaborative air-ground operation scenario involving multiple drones, security personnel, and IoT devices in a smart park's large-scale passenger flow management task; the operation airspace has complex physical terrain limitations such as narrow bridges, stairs, and areas near water, and each drone, security personnel terminal, or IoT gate acts as an intelligent agent node; this type of scenario also presents risks of crowding and stampedes, formation deformation, and local communication delays, thus making it suitable as the main application line of this solution; Furthermore, the operational airspace scheduling system issues preset physical constraint boundary information; this boundary information is not an abstract mathematical region, but an operational boundary that corresponds one-to-one with actual facilities, such as obtaining the safety red line of terrain vulnerability requirements through elevation maps, and establishing the minimum safe lateral distance from narrow bridges and stairs through radar scanning. At the same time, the system establishes an initial communication topology weight matrix for each machine, reflecting the information interaction relationship between nodes under normal collaborative conditions, such as master-slave following and priority reference. For example, three security agents that are adjacent to each other can form a one-hop adjacency relationship. By default, the state reference weight of the middle node to the preceding and following nodes is similar. During the operation phase, each node continuously collects sensing data, which is then filtered to form a smooth state vector. The smoothing process here aims to eliminate measurement noise caused by engineering disturbances such as random disturbances in the environment and multipath effects, and to avoid amplification of misjudgments by unprocessed raw data, which could lead to unnecessary and rapid corrections of security resources near the boundary. The smooth state vector includes at least the current agent node's own motion rate and heading angle, and can also be extended to the heading angle, angular velocity, etc., depending on the control configuration. After obtaining the smooth state, the system further calculates the boundary distance function to characterize the remaining available safety margin between the agent and the current relevant physical boundary. The distance does not only refer to the geometric straight-line distance, but also characterizes the comprehensive approximation degree of the agent node as it evolves to the dangerous boundary along the current motion trend. Taking UAV inspection as an example, when the system is preparing to perform reverse interception and schedule resources for evidence collection based on historical clues such as slight disturbance trajectories found in the previous inspection cycle, it allocates preset high-priority control weights to the reverse and lateral margins. When the crowd spreads rapidly towards the narrow bridge area, although the geometric position has not yet crossed the boundary, due to the speed direction and inertia, its effective boundary margin will decay at a rate higher than the preset rate of change. The system performs hierarchical judgment based on the boundary distance function; if the agent is still in the safe zone, the original formation and communication relationship are maintained to ensure the system's collaborative stability; if it enters the early warning intervention zone, a dynamic communication topology weight matrix is ​​generated by continuously mapping and adjusting the initial communication topology weight matrix, specifically as follows: Based on the output value of the boundary distance function, the initial communication topology weight matrix is ​​adjusted by continuous mapping to generate the dynamic communication topology weight matrix. Under this logic, instead of immediately taking a step maximum braking action, the communication topology is reconstructed by continuous mapping: the current agent node approaching the boundary is controlled to reduce the weight of receiving state information from neighboring nodes according to a preset decay function, while increasing the influence of its own state propagation to surrounding agents. The purpose is that nodes approaching the boundary have the highest priority for local risk perception and should be less constrained by the cooperative following of other nodes under the fixed communication topology. Instead, they should temporarily become the information source for local risk avoidance trajectories. If they further enter the extreme intervention zone, it means that smooth adjustment of the communication topology weight matrix alone is not enough to ensure safety. At this time, a dissipative control term is introduced to suppress high-frequency dynamics caused by large gradient turning, emergency braking or queue traction. This control term is then fused with the original cooperative control law and combined with the control barrier function to apply an insurmountable safety evolution constraint, thereby outputting continuously changing adaptive control commands. The obtained adaptive control commands are sent to the execution unit and converted into the adjustment amount of drive motor torque, steering mechanism angle or braking pressure. At the same time, the dynamic communication topology weight matrix with timestamps and the current smooth state vector are broadcast to neighboring intelligent agents. The timestamp is used by neighboring nodes to identify the control cycle of this state information to prevent erroneous following of historical failure state information. At the anomaly handling level, if the positioning of a certain drone is briefly lost, but the inertial data is still continuous, the system allows short-term prediction based on the most recent smooth state within a limited time window, and marks the external broadcast of that node as a state of reduced confidence; if the communication interruption exceeds the preset time, the neighboring intelligent agent will no longer increase the reference weight of the information sent to it, but will instead execute a local conservative strategy; if multiple boundaries are close at the same time, such as being close to a narrow bridge, stairs and water area, the boundary with the strictest safety constraints will be selected as the main boundary, and the other boundaries will participate in the control generation as additional constraints to avoid control direction conflicts; In the task of security management of large crowds in a smart park, three drones perform grid-based patrols. The forward node discovers a suspected abnormal crowd gathering target due to the evolution of historical clues and prepares to adaptively change its trajectory and perform air-ground coordinated scheduling to conduct multimodal refined evidence collection, gradually approaching the three-dimensional safety constraint boundary. At this time, the forward node no longer operates completely according to the formation rhythm of the following aircraft, but first increases the broadcast influence of its own risk avoidance status, so that the middle node and the rear node can reduce their insistence on maintaining the original formation in time. If the leading node continues to approach the limit, the system further introduces smooth dissipative intervention on the leading node side, so that its turning and deceleration, and automatic scheduling are released continuously, without causing sudden fuselage sway; after the middle node and the rear node receive the new topology weight and state packet, they synchronously adjust the following strategy, thereby avoiding cascading instability where the rear node maintains the original speed while the leading node has decelerated to avoid it. The purpose of this step is to transform boundary risks into information flow reconstruction and then into control layer intervention without relying on centralized emergency scheduling. This allows local dangers to be quickly perceived and collaboratively handled by neighboring nodes in a distributed manner, thereby achieving a balance between boundary security, formation continuity, and execution smoothness.

[0020] Furthermore, the specific steps for outputting a smooth state vector through state filtering include: pre-constructing a spatiotemporal topology model based on a hexagonal discrete global grid; acquiring multimodal sensing data, which includes video surveillance target detection data of each agent node, traffic density sensing data of physical nodes, and wireless probe density data; spatiotemporally aligning the multimodal sensing data under a unified timestamp, and then inputting it into a spatiotemporal graph convolutional network and an extended Kalman filter algorithm for data fusion; estimating state parameters through the extended Kalman filter algorithm, and outputting a smooth state vector containing the motion rate and heading angle of dynamic agent nodes, as well as the congestion state of static agent nodes; In the context of the smart park large-scale passenger flow management operation scenario in this embodiment, dynamic intelligent agent nodes mainly refer to execution units with spatial mobility capabilities within the operational airspace, such as drones and security personnel terminals. Their operational characteristics need to be characterized by their movement speed and heading angle. Static intelligent agent nodes mainly refer to IoT devices with fixed positions and responsible for local area state perception, such as IoT turnstiles. Their state is mainly characterized by the personnel passage density and congestion status at the physical location of the node.

[0021] This embodiment provides a fusion output mechanism for node state perception; specifically, continuing the aforementioned smart park security management operation scenario, an internal multimodal perception unit and an external absolute positioning unit are deployed on each intelligent agent, wherein the former continuously outputs video surveillance target detection data, and the latter outputs the absolute position of the intelligent agent in the operation spatial coordinate system and the gate passage volume; It should be noted that the preset nominal cooperative control law mentioned in this embodiment is a conventional multi-agent formation control algorithm based on a consensus protocol in the art; the pre-constructed control obstacle function is a continuously differentiable function that satisfies the non-negative evolution characteristic based on the geometric relationship of the security defense line. Its specific mathematical form needs to be derived by solving the kinematic differential equations of the current agent nodes to ensure that the constraint conditions include the system affine input terms, thereby supporting the feasibility of solving the quadratic programming algorithm; both of the above are mature algorithm models known to those skilled in the art, and can be tuned and directly called by those skilled in the art according to specific application scenarios. Specifically, relying solely on external absolute positioning has significant shortcomings; narrow bridges, stairs, steel structure roofs, and areas near water in wide and complex terrain environments are prone to signal reflection, causing occasional jumps in external positioning; relying solely on single-modal data is also unreliable, as drift accumulates after long-term operation; therefore, this embodiment performs spatiotemporal alignment and fusion of the two types of data. Spatiotemporal alignment means transforming data with different sampling frequencies, timestamps, and installation locations into a unified control cycle and reference system. For example, an inertial device may output data every 10 milliseconds, while an external positioning system may output data every 50 milliseconds. The system needs to first align the clock, then convert the sensor installation offset to a unified reference point, and then perform fusion. Specifically, a timestamp interpolation algorithm is used to synchronize sensor data with different sampling frequencies, and a preset coordinate system transformation matrix is ​​used to eliminate the geometric offset of the sensor installation to complete spatiotemporal alignment. The role of extended Kalman filtering here is to perform weighted fusion of multi-source data based on the short-term motion continuity of nodes; combined with spatiotemporal graph convolutional networks, video and gate data are more suitable for describing short-term dynamic changes, especially during turns and acceleration / deceleration; absolute positioning is more suitable for correcting long-term drift to determine the global coordinate positioning of the agent in the working airspace. Specifically, the network structure and fusion process of the spatiotemporal graph convolutional network and the extended Kalman filter algorithm are as follows: Spatiotemporally aligned video and multi-source sensing data such as gate access are input into the spatiotemporal graph convolutional network. This network structure contains alternating cascaded spatial graph convolutional layers and temporal convolutional layers to extract spatial interaction features and temporal evolution features among multiple agents. After mapping through fully connected layers, short-time dynamic pseudo-observation vectors are output. Construct a state prediction model using an extended Kalman filter: in, This is the smoothed state vector from the previous time step. Let be the nonlinear kinematic state transition function of the agent node. This refers to system process noise. The global coordinate observation data output by the external absolute positioning unit is concatenated with the short-time dynamic pseudo-observation vector to construct a joint observation vector. ; Calculate the Jacobian observation matrix and Kalman gain based on the observation equation. The state is then updated using the joint observation vector. In the formula, For nonlinear observation functions, This is the smoothed state vector containing motion rate and heading angle, which is the output after the current control cycle is fused.

[0022] The smooth state vector obtained after fusion includes at least the motion rate and heading angle, so that the current intelligent agent node's operating state estimate will not change drastically due to a single instantaneous jump point; for example, if the external positioning of a certain period changes frequently and deviates from the direction of a narrow bridge or stairs by a distance exceeding the preset deviation threshold set based on the sensor calibration error, but the UAV's airspeed, acceleration and the trajectory of the previous period all show that it is actually still in the center of the reverse interception, then the anomaly will not be directly regarded as the real boundary crossing evidence, but will be weakened after filtering; To facilitate understanding, a simplified deduction example can be used: Consider an agent node. In three consecutive control cycles, the inertial data showed that it maintained a stable forward movement, while the external positioning in the second cycle suddenly shifted to the right. Before fusion, the original position sequence showed discreteness. After fusion, the output smooth state still maintained a continuous transition and gave a lateral deflection component less than the set displacement threshold, instead of directly determining that the agent had crossed the preset physical constraint boundary. The motion rate and heading angle of the generated agent node are more consistent with the real physical motion inertia. At the anomaly handling level, if the external absolute positioning is completely lost within a short period of time, the system can maintain operation mainly based on inertial measurement and historical filtering status for a preset duration, and simultaneously reduce the confidence level of the node's status; if the inertial measurement unit saturates or drifts abnormally, the system increases the correction priority of the external absolute positioning and restricts high-dynamic control actions; if both types of data are abnormal at the same time, the node exits the collaborative master control link, retains only the minimum safety actions, such as deceleration, parking, or exiting the congested area flow restriction, and broadcasts the anomaly flag to its neighbors. During nighttime transport, when the front-end node passes near the metal gantry, a reflection deviation occurs in the positioning, and the original position reading momentarily approaches the outer boundary of the security control. If this original value is used directly, the system will make a logical misjudgment and assume that the security resource is about to collide with the narrow bridge or stairs, thereby triggering unnecessary emergency avoidance. After using the fusion processing of this embodiment, since the continuous output of the smooth sway and longitudinal acceleration change is referenced at the same time, the system judges that the deviation is inconsistent with the actual body movement, so the output is still a smooth approach but not touching the edge. The purpose of this step is to provide a usable, continuous, and reliable state basis for subsequent boundary distance assessments, thereby enabling the identification of real risks rather than an overreaction to measurement noise.

[0023] Further, step two specifically includes: extracting the motion rate and heading angle of the dynamic agent node itself from the smoothed state vector; substituting the motion rate and heading angle of the current agent node itself into the pre-built spatiotemporal topology model, which is used to construct forward and lateral dynamic risk pressure prediction vectors based on the terrain physical vulnerability, channel width, slope and hazard source distance of the hexagonal grid nodes; calculating the shortest distance from the endpoint of the dynamic risk pressure prediction vector to the preset physical constraint boundary; and using the shortest distance as the output value of the boundary distance function.

[0024] This embodiment provides a boundary distance quantification mechanism; specifically, after obtaining the smooth state of the agent node, it does not directly use whether the current position crosses the line as the sole criterion for judgment, but instead substitutes the movement speed and heading angle into the spatiotemporal topology model corresponding to the operational airspace facility to form a boundary distance function that is more in line with the actual operational risks. Specifically, if only position is considered and speed is ignored, a judgment lag will occur in highly dynamic operation scenarios. For example, although the agent has not yet encountered the safety barrier edge, if the angle between its migration velocity vector and the boundary normal direction is less than the preset safety threshold, and the current turn is insufficient to return to the correct position within the remaining distance, then its effective safety margin has actually decreased. Conversely, if the angle is greater than or equal to the preset safety threshold, or the current turn is sufficient to return to the correct position within the remaining distance, then it is determined that the effective safety margin has not substantially decreased. The role of the spatiotemporal topology model is to combine the vulnerability of grid nodes, the turning sweep range, the braking extension distance, and the boundary geometry to obtain a shortest distance that can be used for control decisions. For the straight-flight cruise area, this distance is more of a residual lateral margin from the edge to the side boundary. For three-dimensional spatially constrained edges, this distance is a forward margin along the direction of travel to the edge protection line. For the turning area, the attitude tilt angle sweep space during yaw maneuvers should be considered. The spatiotemporal topology model is used to construct forward and lateral dynamic risk pressure prediction vectors based on the terrain physical vulnerability, channel width, slope, and hazard source distance of hexagonal grid nodes. The spatiotemporal topology model is pre-constructed offline based on the geometric polyhedral expansion algorithm. Specifically, the working space is divided into hexagonal grids, the channel width and slope parameters of each grid are extracted, and vulnerability weights are assigned according to a preset mapping table. Euclidean distance penalty terms for hazard sources are superimposed to generate a risk cost map with spatial topological associations, which is stored in the static storage area of ​​the agent node. The shortest distance is the minimum geometric Euclidean distance from the boundary of the dynamic risk pressure prediction vector to the preset physical constraint boundary. A simplified scenario can be described as follows: There are nodes... Its smooth state vector contains the current position. and speed The spatiotemporal topology model defines three types of boundaries, namely, terrain vulnerability side boundaries. High voltage edge and dynamic forensic security envelopes generated based on historical clues The system will evaluate separately. Relative to the remaining margins of these three types of boundaries, the smallest one is taken as the current main boundary distance output; if If the vehicle moves at a speed greater than a preset first speed threshold in the middle of the straight section, then... This may not be the most dangerous boundary; if it is about to enter the restricted work area, then... This could become the minimum distance; the resulting output value represents the most pressing physical risk source at present. At the fault tolerance level, if multiple boundary distances are approximately the same after substituting both position and velocity, they can be prioritized according to the boundary hazard level. For example, personnel isolation boundaries take precedence over ordinary safety defense line boundaries, and three-dimensional no-fly isolation boundaries take precedence over regular operation boundaries. If a certain type of boundary model is temporarily unavailable, such as when construction or renovation causes the electronic map to not be updated, the system will temporarily replace it with a conservatively expanded general safety envelope to avoid false releases due to missing boundaries. If the agent is in a low-speed parking state, the impact of velocity on the effective boundary margin can be appropriately reduced to prevent false judgments triggered by static jitter. When the forward node detects abnormal clustering clues and prepares to break away from the cruise formation for low-altitude evidence collection, its nose position has not yet approached the ground obstacle. However, because the turn has not been fully corrected and the dive speed is higher than the recommended value for close-to-evidence collection, the movement trend is obviously moving towards the safety protection boundary. At this time, this embodiment does not wait for the physical position to reach the preset second distance threshold before alarming. Instead, it provides a dynamic boundary distance output based on the spatial envelope formed by the position and speed. The purpose of this step is to transform the potential danger of crowds that have not yet materialized into a predictable and controllable level of risk, thereby enabling proactive and targeted intervention at the boundary.

[0025] Furthermore, step three specifically includes: setting a first distance threshold and a second distance threshold, wherein the first distance threshold is greater than the second distance threshold; The first distance threshold is determined based on the motion rate, time penalty term, maximum communication delay error margin, and safety buffer distance in the smooth state vector of the current agent node. The time penalty term is a distance margin reduction compensation coefficient derived from historical motion data that has repeatedly approached the physical boundary. Specifically, the average decrease rate of the boundary distance function within the previous preset number of control cycles is extracted and multiplied by the preset motion inertia time constant to calculate the time penalty term.

[0026] The second distance threshold is determined jointly based on the maximum resource scheduling distance of the multi-agent system and the extreme risk diffusion distance caused by hardware response delay; When the value of the boundary distance function is greater than the first distance threshold, it is determined that the area is in a safe zone, and the preset nominal cooperative control law is invoked. When the value of the boundary distance function is less than or equal to the first distance threshold and greater than the second distance threshold, it is determined that the area is in the early warning intervention zone, and the step of generating a dynamic communication topology weight matrix is ​​executed. When the value of the boundary distance function is less than or equal to the second distance threshold, it is determined that the area is in the extreme intervention zone, and the extreme intervention zone control step is triggered.

[0027] This embodiment provides a boundary state classification and determination mechanism; specifically, after obtaining the boundary distance output, instead of using a single threshold for discrete control switching between normal and dangerous conditions, a first distance threshold and a second distance threshold are set to divide the state into a safe zone, a warning intervention zone, and a limit intervention zone. Specifically, relying solely on a threshold has a technical flaw in system stability: once an agent makes a slight deviation exceeding the threshold, it will instantly switch from normal coordination to the highest priority intervention state, which can easily lead to abrupt changes in control commands and exceed the dynamic response bandwidth of neighboring nodes. The core mechanism of the three-tiered classification is to divide the evolution process of boundary risks into three stages: normal operation, relationship reconstruction, and action intervention. The first distance threshold corresponds to the physical boundary at which the cooperative relationship needs to be changed. At this time, the agent still has available margin, and the focus is to weaken the pull of the external formation on it in advance. The second distance threshold corresponds to the boundary at which the dynamic output must be directly constrained. At this time, it is not enough to ensure safety by adjusting the information relationship alone. Within the safe zone, the system invokes the nominal cooperative control law to maintain the queue's cooperative operation state and control rhythm; upon entering the early warning intervention zone, the system does not directly output braking commands exceeding the preset braking threshold, but instead activates the topology reconstruction logic, allowing risk nodes to gradually transform from followers to local leaders; upon entering the extreme intervention zone, the extreme intervention zone control steps are triggered, suppressing the trend of continuing to approach the boundary through additional damping and obstacle constraints; To facilitate understanding, a microscopic state transition diagram can be given: Suppose that the boundary distance of a node gradually decreases from being significantly greater than the first threshold to between the two thresholds, the system changes its communication influence relationship; if the distance continues to decrease to no higher than the second threshold, then the forced protection of the dynamic layer is superimposed; the whole process is reflected as a hierarchical progressive increase of the control weight, rather than a single step to the maximum braking torque; At the level of anomaly tolerance, if the boundary distance fluctuates around the threshold, a hysteresis band or a minimum holding period can be set to avoid frequent switching between adjacent regions. If the region judgments corresponding to different boundaries are inconsistent, for example, if the relative safety defense boundary is in the warning zone and the relative spatial constraint edge is in the extreme intervention zone, it should be handled according to a higher risk level. If the credibility of the upstream state decreases, the threshold can be appropriately expanded to allow the system to enter the warning or extreme intervention state earlier. When the leading node is operating normally on the main inspection route, the boundary margin is sufficient and the system maintains nominal formation control. As it flies away from the complex airflow area, the right safety margin drops to the warning range. At this time, the information influence relationship between it and its neighbors is changed first. If the leading node continues to drift outward and approaches the outer envelope of the narrow bridge and staircase protection, the system further enters the extreme intervention state and directly applies smooth constraints to its deceleration and turning. The purpose of this step is to break down boundary intervention into a hierarchical and progressive response chain, thereby achieving a smooth transition from the information layer to the control layer and reducing the execution impact caused by mode switching.

[0028] Further, the step of generating a dynamic communication topology weight matrix is ​​performed, specifically including: extracting the initial communication topology weight matrix and the boundary distance function of the current agent node; Using the output value of the boundary distance function of the current agent node as the independent variable, the initial communication topology weight matrix is ​​adjusted using a preset continuous attenuation and gain mapping relationship to obtain attenuation weights and gain weights, so as to dynamically update the overall communication topology weight matrix of the system.

[0029] This embodiment provides a topology reconstruction mechanism based on boundary approximation degree; specifically, within the early warning intervention zone, the basic adjacency relationship between agents is not changed, but the reference weights and influence hierarchy between nodes are continuously adjusted; Specifically, hierarchical judgment alone is insufficient. If an agent is close to the boundary but still receives neighbor opinions and outputs its own state with the same weight, the neighbor's intention to maintain the formation will continue to influence the risk node, causing a conflict between its control objectives of risk avoidance and following the formation. Therefore, this embodiment introduces two types of dynamic adjustment coefficients: one type is used to reduce the weight of the current risk node in receiving neighbor state information, and the other type is used to increase the weight of its sending its own state information to neighbors. This mechanism aims to encourage risk nodes to reduce the weight of receiving state information from neighboring nodes and simultaneously increase the weight limit of its sending its own state information to neighboring nodes. The engineering significance of negative exponential mapping is that the adjustment magnitude of the topology weight matrix increases exponentially as it approaches the boundary; the further away from the boundary, the closer it is to the original state; the entire change process is continuous without step jumps; this design can avoid the topology jump defects caused by single weight step jumps or connection abrupt changes. Through the continuous mapping of weights with the degree of boundary approximation, the smooth migration of information flow direction and intensity within the local communication network is achieved. The calculation logic of the negative exponential mapping function: First dynamic adjustment coefficient With the second dynamic adjustment coefficient The calculation formula based on the negative exponential mapping is as follows: In the formula, This is the current boundary distance function value, and its dimension is length. This is the first distance threshold, and its dimension is length; This is a preset proportional adjustment constant, to ensure that the exponent is a dimensionless pure number. Its dimension is defined as the reciprocal of its length; It is a natural constant; Then, the system updates the received attenuation weight. With transmit gain weight : In the formula, is the dimensionless initial communication topology weight matrix.

[0030] This ensures that the closer the distance to the boundary, the closer the first dynamic adjustment coefficient is to the attenuation coefficient of the preset lower limit, while the second dynamic adjustment coefficient is closer to a multiplier greater than 1. Taking quantization as an example: Suppose that the initial receiving and transmitting communication topology weights related to a certain node are both 0.5, the first distance threshold is 2.0 meters, and the current boundary distance function evaluation is 1.0 meters. After substituting into the mapping logic, the first dynamic adjustment coefficient is calculated to be 0.6, so the attenuation weight is updated to 0.5 × 0.6 = 0.3, indicating that its influence from neighbors has decreased; the second dynamic adjustment coefficient is calculated to be 1.67, so the gain weight is updated to 0.5 × 1.67 = 0.835, indicating that its own state broadcast has increased its control influence on neighbors; the system completes the smooth update of weights by directly updating the row and column elements corresponding to the node in the topology weight matrix, thereby realizing the adaptive allocation of information flow in the local communication network.

[0031] Furthermore, the specific steps for triggering the limit intervention zone control include: the dissipative control term is constructed based on the fractional derivative of the boundary distance function with respect to time; Under the non-negative evolution constraint of the pre-constructed control barrier function, the continuously changing adaptive control command is obtained by using a quadratic programming algorithm to fuse the preset nominal cooperative control law and the dissipative control term.

[0032] This embodiment provides a dissipation control generation mechanism within the extreme intervention zone; specifically, when a risk node has entered the extreme intervention zone, relying solely on topological relationship adjustment is insufficient to prevent the tendency to cross the boundary, and at this time it is necessary to directly apply smooth and constrained intervention to the dynamic output; Specifically, if a discrete control mode that triggers an acceleration exceeding a preset threshold as soon as a threshold is reached is directly adopted, although the system response time is short, it will cause new system control instability problems: on the one hand, the lift and attitude angle adjustment limits may be triggered instantaneously, causing instability; on the other hand, the sudden change in control commands will be amplified by neighbors through the communication link, forming a chain of oscillations; therefore, this embodiment introduces a dissipative control term based on the time evolution characteristics of the boundary distance function in the limit intervention zone. The fractional derivative is a dynamic quantity that is sensitive to both the current approach trend and recent historical changes. It is better than the ordinary instantaneous rate of change at reflecting whether a persistent boundary shock is forming. The dissipative term constructed based on this quantity does not simply brake the agent to a standstill, but rather superimposes a damping effect on it to suppress high-frequency and intense actions. Based on this, the system uses quadratic programming to integrate the original nominal cooperative control law with the dissipative control term; within the discrete control period, numerical approximation is used to calculate the fractional derivative, specifically using the Grünwald-Letnikov difference scheme to obtain the discrete-time approximation of the fractional calculus operator.

[0033] Its purpose is to: prioritize the execution of the nominal collaborative control instructions required by the original task, and ensure that the generated adaptive control instructions all meet the safety evolution constraints of the control barrier function; the role of the control barrier function is to limit the further evolution of the intelligent agent system towards the preset risk boundary by applying mathematical hard constraints; the final control instructions are not commands of a single component, but the most suitable comprehensive actions under the condition of safety. To facilitate understanding, a simplified process deduction can be given: Suppose that at a certain node there are two candidate action directions, one of which is more conducive to maintaining the formation rhythm, and the other of which is more conducive to moving away from the boundary; if the former will cause the boundary margin to continue to deteriorate, then the obstacle constraint does not allow it to be used directly; the system will solve for the optimal continuous adaptive control command between maintaining the preset cooperative mission objective and blocking the dangerous evolution trend. Therefore, the result is usually to first smoothly decelerate and correct the course at a rate less than the preset angular velocity threshold, rather than suddenly performing reverse maneuvering; At the fault tolerance level, if the control solution within the extreme intervention zone becomes infeasible due to input saturation, the system directly switches to the minimum risk safety backup intervention action, such as restricted braking and directional docking, while broadcasting an infeasibility flag; if the value of the boundary distance function rises back to above the second distance threshold, the system will not immediately cancel the dissipation term, but will gradually exit in a continuous decay manner to avoid secondary risks caused by suddenly resuming aggressive control after just escaping danger; if the control obstacle function involves multiple boundary constraints at the same time, priority is given to ensuring the insurmountable physical boundary, and then the cooperative motion target is taken into consideration. Pre-node As the vehicle continues to veer towards the narrow bridge and stairs at the exit of the slippery curve, it has entered the limit intervention zone. At this point, simply issuing a step-up maximum steering command may cause instability; while issuing only a step-up maximum braking command may trigger a rear-end collision risk. This embodiment suppresses high-frequency yaw by dissipation terms and integrates the original cooperative driving intention under obstacle constraints, ultimately outputting adaptive control commands with limited deceleration and continuous return to center. At the same time, since its topology weights have been adjusted in advance during the warning stage, the center node and the rear node have actively adjusted their following strategies and appropriately increased their spacing, thus giving the front node sufficient space to execute safety control. The purpose of this mechanism is to achieve executable, continuous, and non-overshooting dynamic intervention near the limit boundary, thereby avoiding execution unit shock and neighbor cascading instability caused by discrete control switching. To further clarify the integration logic of dissipative control terms and quadratic programming, dissipative control terms The mathematical expression is as follows: In the formula, For fractional derivative calculus operators, For time The changing output value of the current boundary distance function. Let the order of the fractional calculus operator be . ; This is a preset gain coefficient; it is used to enable the dissipative control term to be directly integrated with the nominal cooperative control law. It internally includes a dimensional compensation factor, the equivalent dimension of which is the dimension of the input control signal multiplied by . .

[0034] In the quadratic programming solution, the hard constraints of the control barrier function established by the system are expressed as follows: In the formula, The pre-constructed control barrier function is built based on the difference between the shortest geometric distance from the center of the agent node to the physical boundary and the preset safe braking distance, to ensure continuous differentiability when the multi-agent system approaches the physical boundary; The first-order time derivative of the control barrier function along the system's evolution trajectory; A smooth state vector; The adaptive control command to be determined; The preset evolution coefficient is strictly greater than zero, and its dimension is the reciprocal of time.

[0035] The system performs the convex optimization solution in each control cycle. If the candidate original control action attempts to violate the hard constraints, the quadratic programming will automatically find the solution closest to the objective function on the safe boundary of the control input space, ensuring that the final output instruction does not exceed the physical bottom line and is smooth and continuous.

[0036] Further, outputting the adaptive control command to the execution unit configured in the current agent node includes: converting the adaptive control command into a drive signal for the execution unit; encapsulating the smoothed state vector of the current agent node with the dynamic communication topology weight matrix, and adding a system clock timestamp to generate a broadcast data packet; and sending the broadcast data packet to neighboring nodes within the communication topology range through a preset local communication network.

[0037] This embodiment provides a control execution and neighborhood broadcast linkage mechanism; specifically, after generating adaptive control instructions, they need to be output to the execution unit of the local machine, and the associated state and topology changes are synchronized to neighboring nodes, so that local risk handling is upgraded from single-machine response to neighborhood collaborative response; Specifically, adaptive control commands are converted into drive signals that can be recognized by the execution unit; the purpose of the conversion is to match the upper-level control intent with the dynamic constraints of the lower-level mechanism. While completing the local control distribution, the system packages the current smooth state vector, dynamic communication topology weight matrix, and timestamp into a broadcast data packet, and sends it to neighboring nodes within the single-hop communication range through the local communication network; the state vector is used to synchronize its current actual pose and motion speed with neighboring nodes; the dynamic topology weight matrix is ​​used to instruct neighboring nodes to adaptively adjust the reference weights of the current node's state information; Timestamps are used to indicate which control cycle this information belongs to; the three types of data mentioned above are strongly coupled; if only the state is broadcast without updating the topology weights, neighboring nodes will still be limited by the initial reference weights and perform ineffective cooperative following after receiving the trajectory offset state of the preceding node; if only the topology is updated without broadcasting the smoothed state vector, neighboring nodes can only increase the reference weights of the preceding node, but cannot obtain its specific risk avoidance yaw trajectory; if timestamps are missing, old data and new data may be incorrectly mixed due to network latency; A simplified data packet structure can be used to illustrate this: A broadcast packet can contain a node identifier, location and velocity, current boundary level, neighbor-oriented weight adjustment value, and transmission time; after receiving the packet, the neighbor first checks whether the timing is newer than the local cache, and then updates the local follow policy; if the neighbor finds that the packet is outdated, it only retains the risk flag and does not directly correct the trajectory according to its status value. At the fault tolerance level, if local communication network is congested, the system can prioritize sending boundary level changes and key weight changes, while delaying the sending of non-critical state details; if a neighbor does not receive an update packet for a short period of time, it maintains the weight of the previous cycle and gradually switches to conservative following; if duplicate or out-of-order data packets are received, duplicates are removed and reordered based on timestamps and node sequence numbers; if the local node has entered a safe redundancy hovering mode, a flag indicating that formation traction will be stopped will be included in the broadcast to remind neighbors to actively detour or slow down. The purpose of this step is to transform local control results into information that is understandable, acceptable, and verifiable in the neighborhood, so that local interventions can quickly form a coordinated and consistent response within a communication range.

[0038] Furthermore, the method also includes: obtaining the desired state vector output by the multi-agent system under a preset nominal cooperative control law, calculating the global state error vector in combination with the smoothed state vector; and constructing a global Lyapunov candidate function based on the global state error vector and the dynamic communication topology weight matrix. When the time derivative of the candidate function is greater than zero, the gain coefficient of the dissipation control term is positively adjusted, and the rate of change of the dynamic communication topology weight matrix is ​​limited to not exceeding a preset smooth derivative threshold; when the time derivative is less than or equal to zero, the increase of the gain coefficient is stopped and the current gain coefficient is maintained.

[0039] This embodiment provides a smooth constraint mechanism for full-process closed-loop stability; specifically, in the entire process of dynamic communication topology weight matrix update and adaptive control command generation, weight changes and control interventions are not allowed to be independent of each other or aggressive on their own. Instead, the system evolution is constrained by weight change rate limits and global stability constraints. Specifically, although the aforementioned scheme has achieved topology reconstruction through continuous mapping, if the mapping changes too rapidly, it may still produce an effect similar to discrete control switching in engineering. For example, once a node approaches the boundary, its influence on its neighbors increases dramatically, while the neighbors' reference relationship with it decreases rapidly. Although the local network is not disconnected, the direction of information flow will suddenly reverse, causing significant incoordination in the group within one or several control cycles. Therefore, this embodiment further limits the rate of change of the dynamic communication topology weight matrix to not exceeding a preset smooth derivative threshold. Thus, even if the risk increases rapidly, the weight adjustment must proceed continuously along a controllable slope to ensure that the changes in the communication layer match the mechanical response capability. Based on this, the system uses the continuously mapped dynamic topological weights to construct a global Lyapunov candidate function, which is used to characterize whether the entire population evolves in a more stable direction under the current information structure and motion state. This candidate function is used as a criterion for global stability: as long as its derivative is not positive, it means that the system as a whole will not continuously amplify the error due to local intervention. To achieve this goal, the system will adjust the gain of the dissipation control term in a coordinated manner. In terms of specific control logic, this means that when the rate of change of the topology weight of the communication network reaches the preset reconstruction threshold, the dissipation intensity of the dynamic layer must also be matched accordingly. It is necessary to ensure that the boundary approach cannot be suppressed due to the damping coefficient being less than the critical value, nor can the system mobility be limited due to the damping coefficient being too large. For example: Set a preceding node Continuously increase its influence on the state of its neighbors; if the increase slope is too large, the central node... A sudden change in direction can occur within a single cycle to follow the avoidance trajectory of the preceding node, easily causing abrupt path detours; this embodiment limits the rate of change of weights to... The reference migration is completed smoothly over multiple cycles; meanwhile, if If the dissipation gain is increased too much, the local unit may be safe but may suddenly stall, which may trigger compression in the downstream unit. Therefore, the system will also make coordinated adjustments to the dissipation gain according to the global stability criterion to ensure that local safety and group stability are satisfied simultaneously. At the level of fault tolerance, if it is found that the current weight change rate limit cannot meet the extreme safety requirements in a timely manner within a certain period, the system can temporarily activate the higher priority entity boundary safety backup intervention action. However, this action should run in parallel with stability monitoring, and return to the smooth constraint track after leaving the most dangerous state. If the derivative of the candidate function shows signs of increasing, it indicates that the population error may be amplified. In this case, the system will prioritize slowing down the rate of topology change and appropriately increase the intensity of dissipation control. If communication interruption of some nodes leads to incomplete global criterion information, the remaining connected subgroups will each perform local stability constraints to prevent the entire system from abandoning stability control due to a single missing point. During continuous high-intensity transfers at night, the forward node The state changes repeatedly near the boundary, approaching, retreating, and then approaching again; if each change causes the weights to fluctuate significantly, the middle node... and subsequent nodes The following strategy will be changed repeatedly, resulting in high-frequency time-varying fluctuations in queue spacing and path oscillations; This embodiment limits the rate of change of weights, so that Even during periods of risk fluctuation, the impact of external information changes with a continuous slope; at the same time, the dissipation gain is adjusted according to the stability criteria of the entire fleet, so that the correction action of the front node will not be amplified into the cascading high-frequency oscillation of the queue; thus, it can be seen that this scheme overcomes the technical problem that dynamic topology inevitably leads to transient instability, because the topology here is not a jump rewrite, but a controlled evolution process jointly managed by continuous mapping and stability constraints. The purpose of this mechanism is to incorporate communication layer reconstruction and control layer intervention into a unified stability framework, thereby achieving closed-loop stability throughout the entire process of boundary intervention and preventing local security actions from evolving into global oscillations. To clarify the specific execution steps of the global stability constraints: the system constructs a quadratic global Lyapunov candidate function. As shown in the following formula: In the formula, The dimensionless global state error vector is obtained after dimension normalization. The global state error vector is obtained by subtracting the smoothed state vector of each agent node from the expected state vector output under the nominal cooperative control law and normalizing it. The expected state vector is generated offline by the work airspace scheduling system according to the global inspection task route and sent out, or it is calculated in real time by the current agent node's preset nominal cooperative control law according to the state of adjacent nodes. Transpose its matrix; This is the dynamic communication topology weight matrix within the current control cycle, and the dynamic communication topology weight matrix after diagonal loading is a symmetric positive definite matrix within the system control cycle; The system calculates the time derivative of the candidate function in each control cycle. When the time derivative is detected to be greater than zero, that is, the error shows a divergence trend under the current topology and control intervention, the system directly increases the gain coefficient of the dissipation control term proportionally in the local controller, thereby forcibly increasing the damping constraint on the high-frequency evolution of kinematics, thereby reducing the time derivative of the Lyapunov candidate function. When the gain is increased to the point that the time derivative returns to a range less than or equal to zero, the system stops increasing the gain and maintains this damped state. Through this quantization rule and dynamic compensation mechanism, the unpredictable problem of whether the system will diverge under distributed intervention is solved, and global stable convergence under multi-node collaborative operation is guaranteed.

[0040] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A boundary-level intervention control method based on a multi-agent system, wherein the method is applied to a multi-agent system comprising multiple agent nodes, characterized in that, The specific steps include: Step 1: Obtain the preset physical constraint boundary information, initial communication topology weight matrix, preset nominal cooperative control law and pre-built control obstacle function, collect multimodal perception data of each agent node and perform state filtering, and output the smooth state vector of each agent node. Step 2: Taking any one of the multiple agent nodes as the current agent node, calculate the output value of the boundary distance function constructed by the current agent node based on the smoothed state vector; Step 3: Based on the output value of the boundary distance function, the current operating state of the agent node is divided into a safe zone, a warning intervention zone, and a limit intervention zone. When in the safe zone, the initial communication topology weight matrix is ​​maintained. When in the warning intervention zone, the initial communication topology weight matrix is ​​adjusted to generate a dynamic communication topology weight matrix. When in the limit intervention zone, a dissipative control term is constructed based on the fractional derivative of the boundary distance function, and the nominal cooperative control law, the dynamic communication topology weight matrix, and the control barrier function are fused to generate an adaptive control command. Step four: Output the adaptive control command to the execution unit configured in the current agent node, add a system clock timestamp to the dynamic communication topology weight matrix, and broadcast the dynamic communication topology weight matrix with timestamp and the smoothed state vector of the current agent node to neighboring nodes.

2. The method according to claim 1, characterized in that, When in the early warning intervention zone, adjusting the initial communication topology weight matrix to generate a dynamic communication topology weight matrix includes: adjusting the initial communication topology weight matrix through continuous mapping based on the output value of the boundary distance function to generate the dynamic communication topology weight matrix.

3. The method according to claim 1, characterized in that, The process involves collecting multimodal perception data from each agent node and performing state filtering to output a smooth state vector. This includes: pre-constructing a spatiotemporal topology model based on a hexagonal discrete global grid; acquiring the multimodal perception data, which includes video surveillance target detection data, physical node traffic density perception data, and wireless probe density data from each agent node; spatiotemporally aligning the multimodal perception data under a unified timestamp and then inputting it into a spatiotemporal graph convolutional network and an extended Kalman filter algorithm for data fusion; and estimating state parameters using the extended Kalman filter algorithm to output a smooth state vector containing the motion rate and heading angle of dynamic agent nodes, as well as the congestion state of static agent nodes.

4. The method according to claim 3, characterized in that, Step two, which calculates the output value of the boundary distance function constructed by the current agent node, includes: extracting the motion rate and heading angle of the dynamic agent node itself from the smoothed state vector; substituting the motion rate and heading angle of the agent node itself into the spatiotemporal topology model based on a hexagonal discrete global grid to construct forward and lateral dynamic risk pressure prediction vectors; and calculating the shortest distance from the endpoint of the dynamic risk pressure prediction vector to the preset physical constraint boundary as the output value.

5. The method according to claim 1, characterized in that, Step three divides the operating state into a safe zone, a warning intervention zone, and a limit intervention zone, including: setting a first distance threshold and a second distance threshold, wherein the first distance threshold is greater than the second distance threshold; when the output value is greater than the first distance threshold, it is determined that the system is in the safe zone; when the output value is less than or equal to the first distance threshold and greater than the second distance threshold, it is determined that the system is in the warning intervention zone; when the output value is less than or equal to the second distance threshold, it is determined that the system is in the limit intervention zone.

6. The method according to claim 5, characterized in that, The first distance threshold is determined based on the current agent node's movement rate, time penalty term, maximum communication delay error margin, and safety buffer distance; the second distance threshold is determined based on the maximum scheduling response distance of the multi-agent system and the limit diffusion distance caused by hardware response delay.

7. The method according to claim 2, characterized in that, Adjusting the initial communication topology weight matrix through continuous mapping includes: using the output value of the boundary distance function of the current agent node as the independent variable, adjusting the initial communication topology weight matrix using a preset continuous attenuation and gain mapping relationship, obtaining attenuation weights and gain weights, and generating the dynamic communication topology weight matrix.

8. The method according to claim 1, characterized in that, When in the extreme intervention zone, an adaptive control command is generated, including: constructing the dissipative control term based on the fractional derivative of the boundary distance function with respect to time, wherein the dissipative control term includes a preset gain coefficient; and under the non-negative evolution constraint of the control barrier function, using a quadratic programming algorithm to fuse the nominal cooperative control law and the dissipative control term to obtain the adaptive control command.

9. The method according to claim 1, characterized in that, Outputting the adaptive control command to the execution unit configured in the current agent node includes: converting the adaptive control command into a drive signal for the execution unit; encapsulating the smoothed state vector of the current agent node with the dynamic communication topology weight matrix and adding a timestamp from the system clock to generate a broadcast data packet; and sending the broadcast data packet to the neighboring nodes within the communication topology range through a preset local communication network.

10. The method according to claim 8, characterized in that, Also includes: Obtain the desired state vector output by the multi-agent system under a preset nominal cooperative control law, and calculate the global state error vector in combination with the smoothed state vector; A global Lyapunov candidate function is constructed based on the global state error vector and the dynamic communication topology weight matrix; when the time derivative of the candidate function is greater than zero, the gain coefficient of the dissipation control term is positively adjusted, and the rate of change of the dynamic communication topology weight matrix is ​​limited to not exceeding a preset smooth derivative threshold. When the time derivative is less than or equal to zero, stop increasing the gain coefficient and maintain the current gain coefficient.

Images