A multi-agent cooperative control method based on graph theory

Abstract

This invention belongs to the field of intelligent agent collaboration, and mainly involves a multi-agent collaborative control method based on graph theory. It collects environmental data and agent operating states from a warehousing scenario to construct a time-varying road network map; generates a predicted conflict map based on agent trajectory deduction; establishes a coupling analysis mechanism between conflict location, edge state, and agent passage relationships; identifies local high-conflict areas with conflict propagation risk by calculating edge-level conflict propagation potential and edge elastic capacity threshold, and clusters edges and nodes within these areas to generate subgraph-level collaborative control units; generates collaborative control commands for associated agents in each subgraph-level collaborative control unit to achieve refined collaborative scheduling within a local range; effectively suppresses the cascading propagation of conflicts to neighboring edges and nodes while avoiding high-frequency full-graph replanning, reducing local queue backlog and deadlock tendencies, while simultaneously considering overall passage efficiency and continuous control stability.

Description

Technical Field

[0001] This invention belongs to the field of intelligent agent cooperation, and mainly relates to a multi-agent cooperative control method, electronic device and storage medium based on graph theory. Background Technology

[0002] In human-machine co-operation scenarios such as e-commerce forward warehouses or pharmaceutical warehouses, multiple autonomous mobile robots (AMRs) are typically deployed in the warehouse to perform handling tasks between shelving areas, picking areas, buffer areas, and inbound / outbound aisles. In such scenarios, aisle width is limited, intersections are densely distributed, and there are situations such as random human movement, temporary obstruction of aisles, and dynamic changes in task priorities. When multiple AMRs are running in narrow aisles, intersections, and near picking points, path overlap, vehicle collisions, waiting accumulation, and local congestion are likely to occur.

[0003] To address the aforementioned issues, existing technologies typically first abstract the internal road network into a graph structure, and then combine path planning, local obstacle avoidance, intersection management, first-come-first-served, integer programming scheduling, or hierarchical control to allocate paths, avoid conflicts, and prioritize passage for each robot. When environmental disturbances occur, operation is restored through local replanning or priority reallocation.

[0004] However, most existing technologies only respond to the current point of conflict, and the control scope is usually limited to the current intersection, the current path segment, or the current conflicting robot. It is difficult to identify the cascading propagation of the conflict to adjacent road segments, adjacent agents, and downstream nodes. In high-density warehousing scenarios such as e-commerce forward warehouses or pharmaceutical warehouses, yielding, stopping, or detouring at an intersection can easily lead to further conflict spread, queue backlog, or even deadlock. At the same time, existing technologies often focus on ensuring safe passage, making it difficult to simultaneously consider overall throughput, long-term waiting balance, and reducing the frequency of replanning. Fixed priority, first-come-first-served, or frequent global path recalculation methods can easily lead to some robots repeatedly yielding, increased local waiting time differences, and decreased system throughput when manual temporary passage, changes in task urgency, and repeated local congestion occur. Summary of the Invention

[0005] To address the problem that existing technologies struggle to simultaneously achieve conflict propagation suppression, overall traffic efficiency assurance, long-term traffic balance control, and low-frequency replanning in high-density warehousing scenarios, this invention provides a graph theory-based multi-agent cooperative control method, electronic device, and storage medium to solve one or more of the problems mentioned in the background.

[0006] To solve the above problems, the present invention employs the following technology: Firstly, this invention proposes a multi-agent cooperative control method based on graph theory: Collect basic data information from the target warehouse, including environmental data and agent data; Graph structure modeling is performed based on the environmental data and agent data to obtain a time-varying road network map; future trajectory inference is performed based on the time-varying road network map to generate a predicted conflict map; the predicted conflict map is mapped onto the time-varying road network map to obtain a conflict mapping map; conflict analysis is performed on the conflict mapping map to obtain the edge-level conflict propagation potential and edge elastic capacity threshold. Clustering is performed on the conflict mapping graph based on the edge-level conflict propagation potential and edge elastic capacity threshold to obtain locally high-conflict subgraphs; key nodes of the locally high-conflict subgraphs are screened to obtain subgraph-level cooperative control units; instruction calculation is performed on the subgraph-level cooperative control units to obtain a set of cooperative control instructions.

[0007] As a preferred implementation, graph structure modeling is performed based on the environmental data and agent data to obtain a time-varying road network diagram, including: Based on environmental data, establish the edges and nodes in the graph structure, label the attributes of the graph structure, and create an initial graph; Collect and statistically analyze the occupancy rate, average passage time, frequency of manual passage, and passage quality of each edge in the initial graph within the control period, and generate time-varying edge weights; The time-varying edge weights are added to the attributes of the corresponding edges and updated on a rolling basis according to a preset control cycle to form a time-varying road network map.

[0008] As a preferred implementation, future trajectory extrapolation is performed based on a time-varying road network map to generate a predicted conflict map, including: The A* path search algorithm with direction constraints is used to generate a complete movement path from the current position to the target node for each agent. The movement path is divided into several simulation stages based on adjacent nodes. Combining the agent's speed and path attributes, the expected entry and exit times for each stage are calculated to generate the trajectory map of each agent. The trajectory maps of all agents are compared pairwise to identify the conflict structure and conflict type, and a predicted conflict map containing conflict information is generated.

[0009] As a preferred embodiment, the A* path search algorithm with directional constraints includes: Based on the time-varying edge weights in the time-varying road network diagram, edges are classified into traversable edges, poor edges, or impassable edges. During the path search process, the number of bad edges and bad nodes is accumulated, and states with smaller accumulated numbers are expanded first. The path that satisfies the passage constraints and has the minimum cumulative number of bad path elements is output as the agent's movement path.

[0010] As a preferred implementation, conflict analysis is performed on the conflict mapping graph to obtain the edge-level conflict propagation potential and the edge elastic capacity threshold, including: The predicted conflict map is mapped to the time-varying road network map to generate a conflict mapping map. For each conflict edge in the conflict mapping diagram, analyze its predicted time occupancy rate, reservation queue length, attributes of high-frequency manual passage areas, and impact of neighborhood backlog. Based on the analysis results, determine the conflict propagation potential level of the edge. Among them, the impact of neighborhood backlog starts from the target edge segment and performs a k-hop neighborhood search along the connection relationship in the conflict mapping graph. The impact is scored according to the structural attributes and edge segment attributes. The impact of neighborhood backlog is divided into different levels according to the score. The k value is calibrated according to the average edge length in the compartment, braking distance and intersection spacing. Based on historical data calibration or offline stress testing, determine the edge elasticity capacity threshold for each edge segment.

[0011] As a preferred implementation, clustering is performed on the conflict mapping graph based on the edge-level conflict propagation potential and the edge elastic capacity threshold to generate locally high-conflict subgraphs, including: Based on the relationship between the edge-level conflict propagation potential level and the predicted input intensity and edge elastic capacity threshold, high-conflict seed edges and edges to be merged are screened. Using the high-conflict seed edge as the cluster center, search for adjacent edge segments along the connection relationship in the conflict map, and merge the edge segments that meet the conditions into the current cluster range; Repeat the clustering process until no more edges can be merged, forming several locally high-conflict subgraphs.

[0012] As a preferred implementation, the key nodes of the locally high-conflict subgraph are screened to obtain a subgraph-level collaborative control unit, including: Nodes that are in conflict in the local high-conflict subgraph, nodes that are directly connected to the high-conflict seed edge, and convergence nodes that connect multiple conflict edges are selected as key nodes of the subgraph. The number of node conflicts, the number of high propagation potential edges, and the number of associated agents of each key node in the subgraph are counted, and then classified and filtered to obtain core key nodes and general key nodes. Agents that are directly associated with core critical nodes or whose movement paths pass through highly conflicting edges connected to them are identified as core collaborative objects. The core collaborative object, auxiliary collaborative objects, and their corresponding local high-conflict subgraphs are encapsulated together to generate a subgraph-level collaborative control unit.

[0013] As a preferred implementation, instruction calculation is performed based on the subgraph-level cooperative control unit to obtain a set of cooperative control instructions, including: For each subgraph-level collaborative control unit, a collaborative control strategy is formulated with core critical nodes as the priority control positions and core collaborative objects as the priority adjustment objects. The collaborative control strategies corresponding to each subgraph-level collaborative control unit are summarized to generate a collaborative control instruction set, which is then sent to the control terminals of each intelligent agent for execution.

[0014] As a preferred implementation, the cooperative control strategy includes traffic order adjustment, speed adjustment, waiting for release, and local rerouting; Among them, the passage order adjustment is based on the time overlap of each associated agent in the same subgraph-level collaborative control unit at key nodes, edges or conflict positions. The passage sequence is determined according to the priority rule, and the passage order adjustment instructions of the corresponding agents are generated according to the passage sequence. Speed ​​adjustment: Based on the conflict state in front of the agent and the propagation potential level of the edge segment it is located in, the running speed of the agent is adjusted and a speed adjustment command is generated; Waiting to release: Based on the conflict state before the agent enters the high propagation potential region, guide the agent to the waiting position and generate a waiting release instruction according to the release conditions of the conflict region; Local rerouting: Based on the propagation potential level of subsequent edge segments in the original path and the relationship between its input strength and edge elastic capacity threshold, the affected edge segments are avoided, and the local path is replanned after avoidance to generate local rerouting instructions.

[0015] As a preferred implementation, after the set of collaborative control commands is sent to the control terminals of each intelligent agent for execution, a closed-loop evaluation and parameter correction step is also included: Collect the actual execution results of each agent and write them back to the time-varying road network diagram, conflict mapping diagram and subgraph-level cooperative control unit; The current control effectiveness is evaluated based on indicators such as the number of conflict resolutions, changes in waiting time, changes in propagation potential, and changes in capacity utilization. Based on the evaluation results, the parameters of the collaborative control strategy in subsequent control cycles are modified.

[0016] In a second aspect, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, characterized in that: when the processor executes the computer program, it implements the steps of the multi-agent cooperative control method in the first aspect.

[0017] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that: when the computer program is executed by a processor, it implements the steps of the multi-agent cooperative control method in the first aspect.

[0018] The beneficial effects of this invention are: 1. By coupling analysis of time-varying road network maps and predicted conflict maps, the propagation potential and elastic capacity threshold of edge-level conflicts are calculated. The spread trend of conflicts can be identified and intervention can be carried out in advance before they occur, effectively suppressing cascading propagation, reducing the risk of queue backlog and deadlock, and improving the system stability in high-density environments. 2. Based on the local high-conflict subgraph, a collaborative control unit is constructed to output instructions such as passage time window, yielding order, waiting release and local rerouting. This will converge the control to the local area, reduce invalid waiting and repeated conflicts, avoid control jitter caused by global replanning, ensure the passage rhythm of intersections and narrow alley areas, and significantly improve the overall throughput efficiency. 3. Conduct closed-loop evaluation based on actual operating data, dynamically adjust the conflict propagation potential parameters and fairness compensation coefficients, avoid long-term repeated yielding by the agent, reduce local waiting differences, and improve control balance and long-term stability under continuous operation. Attached Figure Description

[0019] Figure 1 This is an exemplary method flowchart of a multi-agent cooperative control method based on graph theory provided in an embodiment of the present invention; Figure 2 This is a comparison diagram showing the effect of a graph theory-based multi-agent cooperative control method provided in this embodiment of the invention compared with traditional inventions. Detailed Implementation

[0020] To make the technical means, creative features, and achieved objectives and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention. Unless otherwise specified, the experimental methods in the following embodiments are conventional methods, and the materials and reagents used in the following embodiments are commercially available unless otherwise specified.

[0021] Example 1 Figure 1 As shown in the exemplary method flowchart of the present invention, this embodiment provides a multi-agent cooperative control method based on graph theory, specifically including the following steps: Step S1: Collect environmental data and agent data of the warehousing scenario, and obtain a time-varying road network map based on the environmental data; Specifically, environmental data of the target warehouse is collected, including warehouse aisles, intersections, picking stations, and agent waiting areas. The specific acquisition method is as follows: an electronic map of the warehouse (such as CAD drawings) is constructed, the above key locations are marked on the map, and the dimensions of the warehouse are marked, mainly including the length of the aisles and the dimensions of the areas (intersections, picking stations, and agent waiting areas). The system then parses and generates structured road network data. Information on high-frequency personnel traffic areas is obtained as follows: the warehouse is divided into corresponding areas based on aisles, intersections, picking stations, and agent waiting areas. Several surveillance cameras are installed in the warehouse to cover each area, and each area is marked in the monitoring screen of the corresponding cameras by setting up electronic fences. Subsequently, the surveillance cameras automatically count the frequency of personnel passage and dwell time in each area within a continuous time window. The continuous time window is set according to the frequency of personnel passage in the warehouse. When the frequency of personnel passage in the warehouse is high, a shorter continuous time window is used; when the frequency of personnel passage in the warehouse is low, a longer continuous time window is used. In this embodiment, the continuous time window is set to 5 minutes. For personnel passage frequency, two thresholds are set: a convergence threshold and a through threshold. Both thresholds are specifically set according to the size of the continuous time window. A shorter time window uses a lower threshold, and vice versa. The concepts of the two thresholds are explained using the settings in this embodiment. The convergence threshold is used in areas where people tend to converge, such as intersections and picking stations. Within a continuous time window, when the total frequency of passage in an area is higher than 12 times, the corresponding area is marked as a high-frequency human passage area. The through-passage threshold is used in areas where people mainly pass through, such as alleyways and intelligent waiting areas. Within a continuous time window, when the frequency of passage in an area is higher than 5 times, the corresponding area is marked as a high-frequency human passage area. For dwell time, a dwell time threshold is set. Within a continuous time window, when there are more than 2 people in the same area at the same time, and the continuous dwell time of each person in the area exceeds 90 seconds, the corresponding area is marked as a high-frequency human passage area. Connect to the intelligent agent management system to collect intelligent agent data in the target warehouse, including the intelligent agent's ID, current position, target position, current speed, running direction, and task-related data such as task identifier and task time; classify the obtained intelligent agent data according to the ID to form a data set for each intelligent agent in the warehouse; Next, a graph structure model is performed on the environmental data. Intersections, picking stations, and agent waiting areas are used as graph nodes, and passable lane sections are used as edges. Based on the actual connection relationships, the relationships between graph nodes are abstracted. Corresponding size data are labeled as attributes, and high-frequency human passage areas are labeled as attributes on the nodes and edges of the graph. The data of the agent at the current moment is identified by ID to form marker points. The agent is marked on the graph according to its current position to form an initial graph. Subsequently, based on the data from the intelligent agent, the occupancy rate, average passage time, frequency of manual passage, and passage quality of each edge in the initial graph were statistically analyzed. The occupancy rate is calculated based on the actual traffic capacity of the corresponding lane, including forward and reverse occupancy rates. First, it's determined whether the lane has bidirectional traffic capacity, based on the lane's effective width, the agent's body width, and the safe passing distance. If the lane's effective width is greater than or equal to twice the sum of the agent's body width and the safe passing distance, the lane is recorded as bidirectional; otherwise, it's recorded as unidirectional. Then, a standard passage threshold is determined based on the lane's effective length, the agent's body length, and the minimum following distance. The effective length of the lane is divided by the sum of the agent's body length and the minimum following distance, rounded down to obtain the standard passage threshold for the corresponding lane. This implementation... In the example, assuming that the aisle lengths in the warehouse are uniform and the spacing between each agent is consistent, the standard passage threshold is set to 4. When the bidirectional passage attribute is that bidirectional passage is allowed, the number of forward and reverse agents in the aisle are counted separately, and the forward occupancy rate and reverse occupancy rate are calculated by combining the number of agents in the same direction and the number of agents in the opposite direction in the target direction. When the bidirectional passage attribute is that unidirectional passage is allowed, the current direction of travel of the existing agents is taken as the forward direction, and the forward occupancy rate is calculated based on the number of forward agents in the aisle. As long as the forward occupancy rate of the current aisle is not 0, the reverse occupancy rate is fixed at 100%, which corresponds to the physical meaning that in a real environment, as long as there are agents in a one-way street, the entry of reverse agents will inevitably cause a blockage. The average passage time is obtained by recording the time when each agent enters the current lane and the time when it leaves the current lane, and averaging the passage time of each agent through the current lane. It is used to represent the reference time for an agent to pass through the lane. The frequency of manual traversal is set based on the attribute of the high-frequency manual traversal area. When the area corresponding to an edge is marked as a high-frequency manual traversal area, the frequency of manual traversal of that edge is set to high frequency; otherwise, it is set to low frequency. Traffic quality indicates whether the current traffic capacity of a roadway is affected by other factors. It uses monitoring cameras to detect whether there are any abnormal situations in the current roadway, such as temporary obstacles, abnormal personnel, temporary items, or local blockages. When an abnormal situation is detected, the current traffic quality is set to abnormal; otherwise, it is set to normal. The occupancy rate, average passage time, frequency of manual passage, and passage quality are concatenated in sequence to form time-varying edge weights; Time-varying edge weights are added as attributes of corresponding edges in the initial graph. Each device in the warehouse uploads its information according to the control cycle. The control cycle is set based on the frequency of changes in traffic status and the density of agent operation in the warehouse. When the agent operation density is high and the conflict relationship changes rapidly, a shorter control cycle is used; when the agent operation density is low and the conflict relationship changes slowly, a longer control cycle is used. In this embodiment, the control cycle is 10 seconds. In each control cycle, the time-varying edge weights are recalculated, the edge attributes in the graph are updated, and a time-varying road network graph is formed.

[0022] Step S2: Based on the time-varying road network map, perform future trajectory extrapolation and generate a predicted conflict map; Specifically, for each agent in the time-varying road network map at the current moment, future trajectory projection is performed. For each agent's data recorded in the time-varying road network map, the complete movement path from the current position to the target position is first determined. The path determination method is as follows: Using the node corresponding to the target location as the target node, an A* path search with directional constraints is performed on the time-varying road network graph. During the search, the current running direction of the target agent, the bidirectional passage attributes of each edge, the forward occupancy rate, the reverse occupancy rate, and the movement direction of existing agents within the edge are combined to determine the edge state of each candidate edge: when a candidate edge meets the current direction entry condition and does not trigger the current agent's same-direction occupancy rate being less than 50%, the frequency of manual crossing being low, and the passage quality being normal, the candidate edge is determined to be a passable edge; when a candidate edge meets the current direction entry condition but there is a situation where the corresponding direction occupancy rate reaches a high occupancy state (above 50%, below 100%), the frequency of manual crossing is high, or the passage quality is abnormal, the candidate edge is determined to be a bad edge; when the reverse occupancy rate of a one-way passable edge is 100% and is in the same direction as the current agent's movement direction, the candidate edge is determined to be a non-passable edge; For the current path searched by the algorithm, record the cumulative number of bad edges, the cumulative number of bad nodes, and the path passage records in the path; during the path search process, prioritize the search states with the smaller sum of the cumulative number of bad edges and the cumulative number of bad nodes for expansion; after the target node is found, output the searched path as the current agent's movement path; After determining the current agent's movement path, a phased trajectory deduction is performed on the agent according to the order of edges and nodes in the movement path. Each phase consists of an edge segment between two adjacent nodes in the movement path, and multiple phases are formed until the corresponding endpoint node. The alley length corresponding to the edge and the node size (intersection diameter) corresponding to the node are read from the time-varying road network diagram. Using the agent's current movement speed as the base speed, the estimated travel time is calculated for each phase. When the endpoint of a phase is a normal node, the base speed remains unchanged. When the endpoint of a phase is a problematic node, a speed reduction correction is applied to the base speed, based on the triggered problematic conditions. For each triggered abnormal travel condition, the speed is reduced by 15%. When the edge segment corresponding to a phase contains a curve, the base speed is adjusted accordingly. Perform a speed reduction correction on curves, reducing the speed by 15%. When both defective nodes and curves exist in the same stage, perform the speed reduction correction sequentially. Then, use the sum of the corresponding lane length and node size as the expected travel distance for that stage, and divide the expected travel distance by the corrected stage speed to generate the expected travel time for that stage. Then, use the expected departure time of the previous stage as the expected entry time of the next stage, and generate the expected entry and departure times for the current agent in each stage. Write the edge identifier, node identifier, expected entry time, and expected departure time of each stage into the trajectory record to generate the trajectory map of the current agent. Perform trajectory deduction for each agent in the time-varying road network map to obtain the trajectory map for all agents. After generating the trajectory graphs of all agents, the trajectory graphs of any two agents are compared to determine the conflict type. When two agents enter the same side within a similar timeframe and their corresponding running time intervals overlap, the conflict type is set to edge conflict. When two agents enter the same node within a similar timeframe and their corresponding running time intervals overlap, the conflict type is set to node conflict. When two agents travel in opposite directions on the same unidirectional edge and their corresponding running time intervals overlap, the conflict type is set to opposing conflict. Subsequently, using agents as graph nodes, the complete path graph of agents reaching the target location is preserved. The conflict relationships identified between agents are used as additional annotations for the conflict graph edges, and conflict information, including conflict type, conflict location, and conflict time interval, is added to the corresponding graph edges to generate a predicted conflict graph. The trajectory graphs are processed pairwise. When multiple agents have conflicts on the same side or at the same node, multiple agents are used as graph nodes, and the corresponding predicted conflict graphs are stitched together to remove duplicate agents. The predicted conflict map is updated on a rolling basis according to the control cycle, and is regenerated based on the complete path simulation results in each control cycle.

[0023] Step S3: Map the predicted conflict map to the time-varying road network map to obtain the conflict mapping map; perform conflict analysis on the conflict mapping map to obtain the edge-level conflict propagation potential and edge elastic capacity threshold. Specifically, the time-varying road network map of the current control period and all the predicted conflict maps are taken as the graph structure state of the current control period; all the predicted conflict maps are combined with the corresponding agent ID and conflict information, and each conflict information is mapped to the corresponding edge segment and node in the time-varying road network map according to the corresponding conflict position to obtain the conflict mapping map. In the conflict mapping diagram, each conflict edge segment is analyzed one by one. For any target edge segment, the predicted time occupancy rate, reservation queue length, attributes of high-frequency manual passage area and the impact of neighborhood back congestion are analyzed and predicted. Among them, the predicted time occupancy rate refers to: for each side, obtaining its occupancy rate in the time-varying edge weight at the predicted conflict time, including the forward occupancy rate and the reverse occupancy rate, calculated using the number of agents at the current predicted conflict time and the standard number threshold; The reservation queue length refers to: reading the estimated entry time of each agent entering the target segment and the estimated departure time of each agent leaving the target segment from the trajectory map; recording the latest estimated departure time of the agent currently occupying the target segment as the release time of the target segment; then, within the time window corresponding to the current control cycle, filtering out agents whose estimated entry time of the target segment is earlier than the release time, sorting them according to their estimated entry time, and recording the number of sorted agents waiting to enter as the reservation queue length of the target segment; The artificial high-frequency crossing zone attribute refers to using the artificial high-frequency crossing zone attribute of the current edge segment as an artificial disturbance factor, i.e., high frequency or low frequency; Neighborhood congestion impact refers to: starting from the target edge segment, searching for adjacent edges outward along the connection relationship within the conflict mapping using a k-hop neighborhood approach (k is preferably 2 or 3, which can be calibrated based on the average edge length within the compartment, braking distance, and intersection spacing. When all three distance parameters are large, the range of conflict impact propagation is relatively limited, and a smaller k value can be selected; conversely, when the lane is short, the braking distance is small, or the intersections are dense, the conflict is more likely to spread to a more distant neighborhood, so a larger k value is selected). For adjacent edges with high occupancy rates (same-direction occupancy rate greater than 50%), reservation queue lengths greater than 0, or... Existing conflict edges are statistically analyzed. A one-hop edge directly connected to the target edge is scored 3 points; a two-hop edge separated from the target edge by one edge is scored 2 points; and a three-hop edge separated from the target edge by two edges is scored 1 point. These scores are accumulated to obtain the neighborhood blocking score for the target edge. When the neighborhood blocking score is greater than or equal to 6, the neighborhood blocking impact of the target edge is set to high; when the neighborhood blocking score is greater than or equal to 3 and less than 6, the neighborhood blocking impact of the target edge is set to medium; and when the neighborhood blocking score is less than 3, the neighborhood blocking impact of the target edge is set to low. Based on the predicted time slot occupancy rate, reservation queue length, attributes of high-frequency manual passage areas, and the impact of neighborhood congestion, the degree of conflict propagation of the target edge segment is classified and determined, thereby generating the edge-level conflict propagation potential corresponding to the target edge segment. The specific determination rules are as follows: When the target segment simultaneously satisfies the following conditions: the predicted time occupancy rate reaches a high occupancy state, the reservation queue length is greater than 0, the manual high-frequency passage area attribute is high and the neighboring back congestion impact is high, the target segment is set as a high propagation potential segment. When any one of the following conditions is met, the target edge segment is set as the mid-propagation potential edge segment. The conditions that can be met include: the predicted time slot occupancy rate reaches a high occupancy state and the reservation queue length is greater than 0; the predicted time slot occupancy rate reaches a high occupancy state and the manual high-frequency crossing area attribute is high-frequency; the reservation queue length is greater than 0 and the impact of neighboring backlog is high; the manual high-frequency crossing area attribute is high-frequency and the impact of neighboring backlog is high. Except for high propagation potential edge segments and medium propagation potential edge segments, the remaining target edge segments are set as low propagation potential edge segments; the propagation potential level corresponding to the target edge segments is recorded in the edge state record of the conflict mapping diagram and the time-varying road network diagram to obtain the edge-level conflict propagation potential corresponding to each edge segment; After obtaining the edge-level conflict propagation potential of each target edge segment, the edge elasticity capacity threshold of each target edge segment is further determined. The edge elasticity capacity threshold is used to represent the upper limit of passage that the target edge segment can withstand under the condition of maintaining basic stable passage. It is obtained as follows: For any target edge segment, its historical operation record is read, and the historical operation record is grouped according to the number of agents entering the target edge segment per unit time to obtain the operation samples corresponding to different input intensities; then, the average delay time and average queue length under each input intensity are statistically analyzed; when the input intensity increases, if the average delay time and average queue length remain stable, the input intensity is further increased; when the average delay time and average queue length increase significantly with the increase of input intensity, the number of agents corresponding to the previous level of input intensity is registered as the edge elasticity capacity threshold of the target edge segment; if the historical operation record of the target edge segment is insufficient, offline stress testing is performed on the target edge segment, and the average delay time and average queue length under each level of input intensity are recorded in a step-by-step manner, and the edge elasticity capacity threshold of the target edge segment is determined according to the same rules.

[0024] Step S4: Cluster the conflict map based on the edge-level conflict propagation potential and edge elastic capacity threshold to generate local high-conflict subgraphs; screen the key nodes of the local high-conflict subgraphs to obtain the subgraph-level collaborative control unit. Specifically, based on the edge-level conflict propagation potential and the edge elastic capacity threshold, the edge segments in the conflict mapping graph are used as clustering objects to filter the current conflict state of each edge segment. The specific process is as follows: When the edge-level conflict propagation potential of the target edge segment is high, or when the input intensity of the target edge segment at the prediction time reaches the corresponding edge elastic capacity threshold, the target edge segment is marked as a high-conflict seed edge segment. When the edge-level conflict propagation potential of the target edge segment is medium, and the current input intensity reaches the preset proportion of the corresponding edge elastic capacity threshold, the target edge segment is marked as an edge segment to be merged. The preset ratio is used to characterize the extent to which the edge segment is close to the capacity limit. It is specifically set according to the average capacity of the warehouse. In this embodiment, it is set to 0.8. Subsequently, using the high-conflict seed edge as the initial cluster center, the search is conducted along the connection relationships in the conflict map to find edges that are directly connected to it. When an adjacent edge is an edge to be merged, or an adjacent edge does not meet the merging condition but shares the same conflict node with the current high-conflict seed edge, the adjacent edge is merged into the current cluster range. The same adjacent search and merging process is then performed on the newly merged edge until no more mergeable edges appear in the current cluster range, thus forming a local high-conflict subgraph. After constructing a local high-conflict subgraph, select new initial cluster centers from the high-conflict seed edges that have not yet been assigned to any local high-conflict subgraph, and repeat the above process until all high-conflict seed edges in the conflict map are assigned to the corresponding local high-conflict subgraphs. After generating each local high-conflict subgraph, the nodes in each local high-conflict subgraph are filtered for key nodes. Specifically, for any local high-conflict subgraph, all nodes are read first, and then the following three types of nodes are filtered as key nodes of the subgraph: first, the node itself that is in conflict; second, the node that is directly connected to the high-conflict seed edge; and third, the convergence node that connects two or more conflicting edges. Subsequently, the number of node conflicts, the number of high propagation potential edges, and the number of associated agents were counted for each key node in the subgraph. Among them, the number of node collisions is used to indicate the number of times the node is recorded as a collision location in the current control period, the number of connected high propagation potential edges is used to indicate the number of edges connected to the node that are registered as high propagation potential edges, and the number of associated agents is used to indicate the number of agents that pass through the node or are expected to enter the node in the current control period. After completing the above statistics, the key nodes of each subgraph are classified and screened. The specific process is as follows: when a key node of a subgraph simultaneously meets the following conditions: the number of node conflicts reaches a preset threshold, the number of connected high propagation potential edges reaches a preset threshold, and the number of associated agents reaches a preset scale threshold, the key node of the subgraph is recorded as a core key node; when a key node of a subgraph does not meet the conditions of a core key node, but meets at least two of the conditions, the key node of the subgraph is recorded as a candidate key node; the other key nodes of the subgraph, excluding core key nodes and candidate key nodes, are recorded as general key nodes. The preset number threshold, preset quantity threshold, and preset scale threshold are all calibrated based on historical conflict records within the warehouse. The number of conflicts, the number of high-propagation potential edges connected, and the number of associated agents for all key nodes in the subgraphs are statistically analyzed over multiple historical control cycles, and their distribution curves are plotted. The mean or median of each indicator is used as the basic threshold, and then fine-tuned by combining the system's sensitivity to core key node identification with actual control resource constraints. Finally, a threshold combination that can effectively capture high-impact nodes while avoiding over-labeling is determined. In this embodiment, the node conflict number threshold is 3 times, the number of high-propagation potential edges connected is 2, and the number of associated agents is 4. After obtaining the core key nodes, candidate key nodes, and general key nodes, the association of agents in the local high-conflict subgraph is processed by assigning them to other agents. The specific method is as follows: Agents that are directly associated with core critical nodes or whose movement paths pass through high-conflict edges connected by core critical nodes are designated as core collaborative objects. Agents that are directly associated with candidate key nodes or whose movement paths are expected to pass through candidate key nodes after entering the boundary of a local high-conflict subgraph are identified as auxiliary cooperation objects. After the attribution process is completed, the core collaborative object, the auxiliary collaborative object, and the corresponding local high-conflict subgraph are encapsulated together to generate the subgraph-level collaborative control unit corresponding to the local high-conflict subgraph. The subgraph-level collaborative control unit includes: subgraph identifier, set of included edge segments, set of included subgraph key nodes, core key node identifier, candidate key node identifier, set of core collaborative objects, and set of auxiliary collaborative objects.

[0025] Step S5: Based on the subgraph-level cooperative control unit, perform instruction calculation to obtain the cooperative control instruction set, and issue and execute the cooperative control instruction set; Specifically, based on the obtained subgraph-level cooperative control units, instruction calculation is performed on each subgraph-level cooperative control unit. The specific process includes: taking the core key nodes in each subgraph-level cooperative control unit as priority control positions and the core cooperative objects as priority adjustment objects, and formulating cooperative control strategies for each edge segment and each associated agent in the local high-conflict subgraph; wherein, the cooperative control strategies include passage order adjustment, speed adjustment, waiting release, and local rerouting, which are used to adjust the passage order, passage speed, waiting position, and subsequent path of the agent in advance before the agent actually enters the high-conflict area; The passage order adjustment refers to: reading the expected entry time, expected departure time, edge-level conflict propagation potential, and edge elastic capacity threshold of each associated agent within the same subgraph-level collaborative control unit; when multiple agents have time overlap at the same key node, the same edge segment, or the same conflict location, a passage order is generated according to preset priority rules; the preset priority rules include: core collaborative objects take precedence over auxiliary collaborative objects, agents that have entered on high propagation potential edges take precedence over agents that have not yet entered, and agents that are expected to leave the conflict area earlier take precedence over agents that are expected to leave the conflict area later; when the priority rules are the same, the passage order is determined according to the order of expected entry time into the conflict location; based on the determined passage order, yield control instructions are generated for agents that need to yield, and priority passage control instructions are generated for agents that can pass first. Speed ​​adjustment refers to: reading the current movement speed, the conflict position ahead, and the propagation potential level of the corresponding edge segment of each associated agent in the local high-conflict subgraph; when there is a conflict position ahead of the target agent, and the time interval between its expected entry time and the expected departure time of the agent ahead is less than a preset safety time interval, generating a deceleration control command for the target agent; when the conflict position ahead of the target agent has been released, and the propagation potential level of its edge segment has decreased from high propagation potential to medium or low propagation potential, generating a speed recovery control command for the target agent. Waiting for release refers to: reading the connection relationship between the current position of each associated agent and the nearest waiting position in the local high-conflict subgraph; when the target agent continues to move forward at the current moment and will enter a high propagation potential edge segment, and the conflict position in front of it has not yet been released, guiding the target agent to the waiting position closest to the current node and not blocking the main passage edge segment, forming a waiting control instruction; the waiting control instruction includes a waiting position identifier and a waiting duration condition; the waiting duration condition is used to indicate that the target agent remains in a waiting state until the conflict position in front is not released or the propagation potential level of the corresponding conflict edge segment is not reduced; when the corresponding conflict position is released and the target edge segment returns to a passable state, a release control instruction is generated, causing the target agent to exit the waiting state and re-enter the passage path; Local rerouting refers to: reading the target agent's original movement path and current position in a local high-conflict subgraph; when a subsequent edge segment in the original movement path is registered as a high-propagation potential edge segment, and the current input strength of this edge segment reaches the edge elasticity capacity threshold, the edge segment is marked as a temporary avoidance edge segment; subsequently, in the conflict mapping graph corresponding to the local high-conflict subgraph, using the node corresponding to the target agent's current position as the starting node and the node corresponding to the original target position as the target node, after excluding the temporary avoidance edge segment, the path search is re-executed to generate a local alternative path; then, the local alternative path is written into the target agent's control path record to generate a local rerouting control command; if no feasible path is found after excluding the temporary avoidance edge segment, the original movement path remains unchanged, and a waiting control command is generated for the target agent; After the collaborative control strategy is formulated, the control results corresponding to each subgraph-level collaborative control unit are summarized to generate a collaborative control instruction set. The collaborative control instruction set includes: target agent identifier, instruction type, target edge or target node identifier, target speed value, waiting position identifier, rerouting path record, and instruction effective time. Subsequently, the collaborative control instruction set is sent to the control terminal of each agent and drives each agent to perform passage adjustment according to the corresponding control instruction.

[0026] Example 2: After formulating and issuing the set of cooperative control instructions, the running status of each agent is collected, and the status is written back. The control effect is optimized based on the status writing back results: After issuing and executing the set of cooperative control instructions, the execution results of each agent are written back. The status writing back refers to: collecting the actual entry time of each agent into the edge / node, the actual departure time, the actual movement speed, the actual waiting time, the actual rerouting path, the actual conflict resolution result, and the execution failure flag; for agents that execute successfully, the corresponding execution results are written back to the status records of the time-varying road network graph, the conflict mapping graph, and the subgraph-level cooperative control unit to update the edge status, node status, and agent status of the current control cycle; for agents that fail to execute, the reason for failure is written into the abnormal status record, and the path search or waiting release process is retried to generate new cooperative control instructions; After completing the state write-back, the control effect within the current control cycle is evaluated. The control effect includes: the number of conflict resolutions, changes in average waiting time, changes in propagation potential within local high-conflict subgraphs, and changes in edge elastic capacity utilization. When the propagation potential level of the corresponding edge segment decreases, the average waiting time shortens, and the number of conflict resolutions increases after control, the current control strategy is marked as an effective control strategy. When the propagation potential level does not decrease or the waiting time increases after control, the corresponding control strategy is marked as an inefficient control strategy, and the passage order adjustment rules, deceleration ratio, or waiting release conditions in the next control cycle are modified, thereby completing the control effect optimization based on state write-back.

[0027] like Figure 2 The comparison chart of the effects of this invention and traditional inventions is shown in the figure. The black bars in the figure represent the effects of this invention, and the gray bars represent the effects of traditional inventions. This invention effectively improves the conflict propagation suppression rate by constructing a predictive conflict graph and edge-level conflict propagation potential, identifying the diffusion trend of conflicts and formulating processing strategies before they occur. By adopting a local high-conflict subgraph and a subgraph-level cooperative control unit, the control range is converged to the area that truly needs intervention, preventing a large number of irrelevant agents from being forced to adjust their paths, thus achieving a better passage efficiency improvement rate than traditional inventions. Furthermore, a closed-loop correction mechanism is introduced, resulting in a higher dynamic disturbance adaptation rate compared to traditional inventions.

[0028] Example 3: This example provides an electronic device, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the multi-agent cooperative control method proposed in the above examples.

[0029] The electronic device can be a terminal, comprising a processor, memory, communication interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0030] This embodiment provides a storage medium on which a computer program is stored. When executed by a processor, the program implements the multi-agent cooperative control method proposed in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0031] Example 4: Taking the morning replenishment period in a pharmaceutical warehousing scenario as an example, the target warehouse includes 8 internal aisles, 4 intersections, 6 picking stations, and 2 agent waiting areas. Currently, there are 9 agents performing handling tasks. Agent A1 is located in the west waiting area, with its target location being the east picking station; Agent A2 is located in the south aisle, with its target location being the north picking station; Agent A3 is located on the west side of the central intersection, with its target location being the southeast picking station; Agent A4 is located in the north aisle, with its target location being the central waiting area; Agent A5 is located in the south waiting area, with its target location being the northwest picking station; the remaining agents are distributed around the aisles and intersections. Structured road network data is obtained by parsing the warehouse electronic map, where the length of the main passageway is set to 2. The dimensions of the roadway are set to 2 meters, the branch roadway length is set to 14 meters, the intersection size is set to 3.6 meters, the picking station size is set to 2.4 meters, and the intelligent agent waiting area size is set to 3.2 meters. Surveillance cameras cover each area, and electronic fences are used to mark the areas. Personnel passage frequency and dwell time are counted within a continuous time window, which is set to 4 minutes. Within this continuous time window, the central intersection recorded a total passage frequency of 16 times, the east picking station recorded a total passage frequency of 14 times, the north-south main roadway recorded a passage frequency of 7 times, and the south waiting area recorded a passage frequency of 6 times. Simultaneously, the east picking station also recorded two people continuously dwelling in the same area for 98 seconds. Therefore, the central intersection, the east picking station, the north-south main roadway, and the south waiting area are marked as high-frequency human passage areas. Simultaneously, the system collects the ID identifier, current position, target position, current speed, direction of travel, task identifier, and task deadline for each agent. Specifically, A1's current speed is 1.1 meters per second, its direction of travel is west to east, and its task deadline is 09:18:00; A2's current speed is 1.0 meter per second, its direction of travel is south to north, and its task deadline is 09:16:40; A3's current speed is 0.9 meters per second, its direction of travel is west to east, and its task deadline is 09:17:10; A4's current speed is 1.2 meters per second, its direction of travel is north to south, and its task deadline is 09:19:20; A5's current speed is 1.0 meter per second, its direction of travel is south to west, and its task deadline is... The deadline is 09:17:50. Subsequently, the intersection, picking station, and agent waiting area are used as graph nodes, and each passable lane segment is used as an edge to form an initial graph. Then, the bidirectional passage attribute is determined based on the effective lane width, agent body width, and safe distance between vehicles. The effective width of the main passage is 2.2 meters, which is recorded as bidirectional passage, and the effective width of the branch lane is 1.3 meters, which is recorded as unidirectional passage only. Then, the standard passage threshold is determined based on the effective lane length, agent body length, and minimum following distance. The agent body length is 0.8 meters, and the minimum following distance is 0.5 meters. Therefore, the standard passage threshold corresponding to the main passage is 16, and the standard passage threshold corresponding to the branch lane is 10. Taking the central main lane as an example, the current number of forward-moving agents is 6, and the number of reverse-moving agents is 2. Therefore, the forward occupancy rate of this side is 37.5%, and the reverse occupancy rate is 12.5%. Taking the southeast branch lane as an example, since this side has a two-way passage attribute that allows only one-way passage and there are already 2 agents moving in the predetermined direction, the forward occupancy rate of this side is 20%, and the reverse occupancy rate is fixed at 100%. Then, based on the entry and exit times, the average passage time is calculated. The average passage time for the central main lane is 18 seconds, and the average passage time for the southeast branch lane is 13 seconds. Next, the frequency of manual passage is determined based on the marking of the high-frequency manual passage area, and temporary items and abnormal personnel are identified through the monitoring camera. Among them, there are temporary items in front of the picking station on the east side, and the passage quality is recorded as abnormal. The other main sections are recorded as normal passage. Finally, the occupancy rate, average passage time, frequency of manual passage and passage quality are concatenated in sequence to form time-varying edge weights, which are added to the corresponding edge attributes. Considering that the intelligent agent operation density is high during this period, the control cycle is set to 8 seconds, thereby generating a time-varying road network map that changes with time at the current control time. The future trajectory of each of the nine agents in the time-varying road network map is predicted. Taking A1 as an example, with its target location corresponding to the node as the target node, an A* path search with directional constraints is performed on the time-varying road network map. During the search, the candidate edge for A1 to enter the central main roadway from the western waiting area meets the current direction entry conditions, and the same-direction occupancy rate is 37.5%, the frequency of manual passage is low, and the passage quality is normal. Therefore, this candidate edge is determined to be a passable edge. The section in front of the eastern picking station that A1 intends to enter is determined to be a defective edge because the frequency of manual passage is high and the passage quality is abnormal. A1 intends to detour. The southeast branch road lane is designated as a non-travelable lane because it is a one-way passage lane with a 100% reverse occupancy rate. The cumulative number of defective edges, cumulative number of defective nodes, and path passage records for each candidate path are recorded. The path with the smaller sum of cumulative defective edges and cumulative defective nodes is selected as the movement path for A1. Subsequently, a phased trajectory deduction is performed according to the order of edges and nodes in the movement path. In the first phase, A1 passes through a 22-meter-long main passage lane and enters a 3.6-meter-wide intersection. The endpoint of this phase is a normal node, and the base speed remains constant at 1.1 meters per second. The expected travel distance for this phase is... The first stage is 25.6 meters long, with an estimated travel time of 23.3 seconds. The second stage involves entering the section before the eastern picking station and its corresponding endpoint. The endpoint of this stage is a defective node, triggering a traffic anomaly. The base speed is reduced to 0.935 meters per second. If the section contains curves, a curve reduction is applied, resulting in a speed of 0.79475 meters per second. The estimated entry and exit times are then calculated based on the corresponding side length and node dimensions. Similarly, A2, A3, A4, and A5 generate their own trajectory diagrams. The trajectory diagram comparison shows that A1 and A3 are at 0... Between 9:12:28 and 09:12:36, A2 and A4 enter the central intersection, with overlapping operating time intervals, resulting in node conflict; between 09:12:31 and 09:12:42, A2 and A4 enter the northbound main roadway, with overlapping operating time intervals, resulting in edge conflict; A3 and A5 are expected to enter the same unidirectional edge in opposite directions on the southeast branch roadway, resulting in opposing conflict; using agents as graph nodes and conflict relationships as graph edges, and writing conflict type, conflict location, and conflict time interval into the graph edges, a predicted conflict graph is generated; within each control cycle, the predicted conflict graph is regenerated based on the complete path deduction results; Using the time-varying road network map and predicted conflict map of the current control cycle as the graph structure state, each conflict information is mapped to the corresponding edge segment and node in the time-varying road network map according to the conflict location, resulting in a conflict mapping map. Taking the edge segment of the central main roadway as the target edge segment for analysis, its current positive occupancy rate is 37.5%. Within the time window corresponding to the current control cycle, the latest expected departure time of the agent that has occupied this edge segment is 09:12:40. The two agents that are expected to enter this edge segment at 09:12:34 and 09:12:38 are both earlier than this release time. Therefore, the reservation queue length of this edge segment is recorded as 2. This edge segment connects to the central intersection, which has been marked as a high-frequency manual crossing area. Therefore, the attribute of the high-frequency passage area is recorded as high frequency; then, a k-hop neighborhood search is performed starting from this edge segment, and in this embodiment, a 2-hop neighborhood is selected; among the one-hop edges directly connected to it, one edge segment has a high occupancy rate, which is scored as 3 points, and another edge segment has a reservation queue length greater than 0, which is also scored as 3 points; among the two-hop edges separated by one edge, one edge segment has a conflict edge, which is scored as 2 points; the cumulative neighborhood backlog score is 8, so the neighborhood backlog impact is recorded as high; since this edge segment simultaneously satisfies the conditions of a reservation queue length greater than 0, the attribute of the high-frequency passage area being high frequency, and the neighborhood backlog impact being high, although the occupancy rate has not reached a high occupancy state, it forms a continuous conflict association with the adjacent edge segment, so it is adjacent to it. The section in front of the eastern picking station is analyzed as a key target section. Currently, the occupancy rate of this section in the same direction is 60%, the queue length is 3, the high-frequency manual passage zone attribute is high, and the impact of neighboring congestion is high; therefore, this section is classified as a high-propagation-potential section. The northbound section of the central intersection currently has a occupancy rate of 54%, a queue length of 1, a low-frequency manual passage zone attribute, and a medium impact of neighboring congestion; therefore, this section is classified as a medium-propagation-potential section. The remaining sections are classified as low-propagation-potential sections. Subsequently, the edge elasticity capacity threshold is determined, the historical operation records of the section in front of the eastern picking station are read, and groups are formed according to the number of agents entering this section per unit time. When the input intensity is 3, 4, and 5 agents per minute, the average delay time is 5 seconds, 7 seconds, and 8 seconds, respectively, and the average queue length is 1, 1, and 2, respectively. When the input intensity increases to 6 agents per minute, the average delay time jumps to 15 seconds, and the average queue length jumps to 4. Therefore, the input intensity before the jump, i.e., 5 agents per minute, is registered as the edge elasticity capacity threshold for this edge segment. For the northbound edge segment, historical records show that the delay and queue length remain stable before 4 agents per minute, and jump significantly at 5 agents per minute. Therefore, 4 agents per minute is registered as its edge elasticity capacity threshold. Thus, the edge-level conflict propagation potential and edge elasticity capacity threshold of each target edge segment are obtained. Using the edges in the conflict mapping graph as clustering objects, the current conflict state of each edge is filtered. The edge in front of the eastern picking station is marked as a high-conflict seed edge due to its high edge-level conflict propagation potential. The northward edge is marked as a pending-merge edge because its edge-level conflict propagation potential is medium and its current input intensity reaches 0.85 of its edge elasticity capacity threshold. Although the westward edge of the central intersection does not meet the pending-merge condition, it shares the same conflict node as the edge in front of the eastern picking station, so it is included in the clustering scope. After continuing the search along the connection relationship, a local high-conflict subgraph is formed, which includes the edge in front of the eastern picking station, the northward edge of the central intersection, and the westward edge of the central intersection. Subsequently, the nodes in this local high-conflict subgraph are filtered for key nodes. Among them, the central intersection was recorded as a conflict location 4 times in the current control cycle, with 2 connected high-propagation potential edges and 5 associated agents. Therefore, this node meets all three screening conditions and is recorded as a core critical node. The eastern picking station entrance node was recorded as a conflict location 2 times, with 2 connected high-propagation potential edges and 3 associated agents. It meets two of the conditions and is therefore recorded as a candidate critical node. The remaining nodes are recorded as general critical nodes. Then, the associated agents are assigned. The movement paths of A1, A2, and A3 all pass through the high-conflict edges connected by the core critical node, so they are recorded as core collaborative objects. A4 and A5 are expected to pass through the candidate critical node after entering the boundary of the local high-conflict subgraph, so they are recorded as auxiliary collaborative objects. Finally, the local high-conflict subgraph, the core critical node identifier, the candidate critical node identifier, the core collaborative object set, and the auxiliary collaborative object set are encapsulated together to generate the corresponding subgraph-level collaborative control unit. Based on the subgraph-level collaborative control unit, instruction calculation is performed first. The expected entry time, expected departure time, edge-level conflict propagation potential, and edge elasticity capacity threshold of A1, A2, A3, A4, and A5 are read. Specifically, A3 is expected to enter the central intersection at 09:12:27, A1 at 09:12:28, A2 at 09:12:30, A4 at 09:12:31, and A5 at 09:12:33. Since A1 and A3 are both core collaborative objects, and A3 is expected to leave the conflict area earlier, a priority passage control instruction is generated for A3, and a yield control instruction is generated for A1. A2 and A4 have an edge conflict, and A2's expected entry time is earlier than A4's; therefore, a priority passage control instruction is generated for A2, and a yield control instruction is generated for A4. Next, speed adjustment is performed. The conflict position ahead of A1 has not yet been released, and the time interval between its expected entry time and A3's expected departure time is only 1 second, which is less than the preset safety interval of 3 seconds. Therefore, a deceleration control command is generated for A1, reducing the target speed from 1.1 m / s to 0.7 m / s. After A3 passes, a speed recovery control command is generated for A1, restoring the target speed to 1.0 m / s, and then the waiting release is executed. If A4 continues to advance, it will enter a high propagation potential segment, and the conflict position ahead has not been released. Therefore, A4 is guided to the southern waiting area, which is closest to the current node and does not block the main passage segment. A waiting control command is generated, the waiting position is recorded as the southern waiting area, and the waiting duration condition is recorded as the conflict position ahead being released and the propagation potential level of the target segment decreasing. Then, local re-execution is executed. The routing process is as follows: The subsequent edge segment in A5's original movement path is the high-propagation potential edge segment in front of the eastern picking station. Since the current input intensity of this edge segment has reached the edge elasticity capacity threshold, it is marked as a temporary avoidance edge segment. Subsequently, using the node corresponding to A5's current position as the starting node and the node corresponding to the original target position as the target node, the path search is re-executed after excluding this temporary avoidance edge segment. This yields a local alternative path that bypasses the western edge segment in the middle, and this local alternative path is written into A5's control path record, forming a local rerouting control instruction. Finally, the control results, such as A1 slowing down and subsequently restoring speed, A3 prioritizing passage, A4 waiting for release, A5 local rerouting, and A2 prioritizing passage, are summarized to generate a set of collaborative control instructions, which are then sent to the control terminals of each intelligent agent for execution. After the execution of this control cycle, the node conflicts, edge conflicts, and opposing conflicts that were originally concentrated at the central intersection and the entrance of the picking station on the east side were all resolved, with a total of 5 conflicts resolved. The number of high propagation potential edge segments corresponding to the central intersection decreased from 2 to 0, the northbound edge segment was adjusted from medium propagation potential to low propagation potential, and the reservation queue length of the edge segment in front of the picking station on the east side decreased from 3 to 1. The actual waiting time for A1 was 6 seconds, the actual waiting time for A4 was 11 seconds, and the detour for A5 increased the passage time by 9 seconds. However, the average waiting time of each agent in the local high-conflict subgraph decreased from the predicted value of 18 seconds before control to the actual value of 7 seconds after control. During the same statistical period, the input intensity entering the edge segment in front of the picking station on the east side decreased from 6 agents per minute to 4 agents per minute, falling back to within its edge elasticity capacity threshold.

[0032] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-agent collaborative control method based on graph theory, characterized in that, include: Collect basic data information from the target warehouse, including environmental data and agent data; Graph structure modeling is performed based on the environmental data and agent data to obtain a time-varying road network map; future trajectory prediction is performed based on the time-varying road network map to generate a predicted conflict map; the predicted conflict map is mapped onto the time-varying road network map to obtain a conflict mapping map. Conflict analysis is performed on the conflict mapping graph to obtain the edge-level conflict propagation potential and the edge elastic capacity threshold; Clustering is performed on the conflict mapping graph based on the edge-level conflict propagation potential and edge elastic capacity threshold to generate locally high-conflict subgraphs; key nodes of the locally high-conflict subgraphs are screened to obtain subgraph-level cooperative control units; instruction calculation is performed on the subgraph-level cooperative control units to obtain a set of cooperative control instructions.

2. The multi-agent collaborative control method of claim 1, wherein, Based on the environmental data and agent data, graph structure modeling is performed to obtain a time-varying road network graph, including: Based on environmental data, establish the edges and nodes in the graph structure, label the attributes of the graph structure, and create an initial graph; Collect and statistically analyze the occupancy rate, average passage time, frequency of manual passage, and passage quality of each edge in the initial graph within the control period, and generate time-varying edge weights; The time-varying edge weights are added to the attributes of the corresponding edges and updated on a rolling basis according to a preset control cycle to form a time-varying road network map.

3. The multi-agent collaborative control method of claim 1, wherein, Future trajectory projection is performed based on time-varying road network maps, generating a predicted conflict map, including: The A* path search algorithm with direction constraints is used to generate a complete movement path from the current position to the target node for each agent. The movement path is divided into several simulation stages based on adjacent nodes. Combining the agent's speed and path attributes, the expected entry and exit times for each stage are calculated to generate the trajectory map of each agent. The trajectory maps of all agents are compared pairwise to identify the conflict structure and conflict type, and a predicted conflict map containing conflict information is generated.

4. The multi-agent cooperative control method according to claim 3, characterized in that: The A* path search algorithm with direction constraints includes: Based on the time-varying edge weights in the time-varying road network diagram, edges are classified into traversable edges, poor edges, or impassable edges. During the path search process, the number of bad edges and bad nodes is accumulated, and states with smaller accumulated numbers are expanded first. The path that satisfies the passage constraints and has the minimum cumulative number of bad path elements is output as the agent's movement path.

5. The multi-agent collaborative control method of claim 1, wherein, Conflict analysis is performed on the conflict mapping graph to obtain the edge-level conflict propagation potential and edge elasticity capacity threshold, including: The predicted conflict map is mapped to the time-varying road network map to generate a conflict mapping map. For each conflict edge in the conflict mapping diagram, analyze its predicted time occupancy rate, reservation queue length, attributes of high-frequency manual passage areas, and impact of neighborhood backlog. Based on the analysis results, determine the conflict propagation potential level of the edge. Among them, the impact of neighborhood backlog starts from the target edge segment and performs a k-hop neighborhood search along the connection relationship in the conflict mapping graph. The impact is scored according to the structural attributes and edge segment attributes. The impact of neighborhood backlog is divided into different levels according to the score. The k value is calibrated according to the average edge length in the compartment, braking distance and intersection spacing. Based on historical data calibration or offline stress testing, determine the edge elasticity capacity threshold for each edge segment.

6. The multi-agent collaborative control method of claim 1, wherein, Clustering is performed on the conflict mapping graph based on the edge-level conflict propagation potential and edge elastic capacity threshold to generate locally high-conflict subgraphs, including: Based on the relationship between the edge-level conflict propagation potential level and the predicted input intensity and edge elastic capacity threshold, high-conflict seed edges and edges to be merged are screened. Using the high-conflict seed edge as the cluster center, search for adjacent edge segments along the connection relationship in the conflict map, and merge the edge segments that meet the conditions into the current cluster range; Repeat the clustering process until no more edges can be merged, forming several locally high-conflict subgraphs.

7. The multi-agent collaborative control method of claim 1, wherein, The subgraph key nodes of the locally high-conflict subgraph are filtered to obtain a subgraph-level collaborative control unit, including: Nodes that are in conflict in the local high-conflict subgraph, nodes that are directly connected to the high-conflict seed edge, and convergence nodes that connect multiple conflict edges are selected as key nodes of the subgraph. The number of node conflicts, the number of high propagation potential edges, and the number of associated agents of each key node in the subgraph are counted, and then classified and filtered to obtain core key nodes and general key nodes. Agents that are directly associated with core critical nodes or whose movement paths pass through highly conflicting edges connected to them are identified as core collaborative objects. The core collaborative object, auxiliary collaborative objects, and their corresponding local high-conflict subgraphs are encapsulated together to generate a subgraph-level collaborative control unit.

8. The multi-agent cooperative control method according to claim 1, characterized in that, Based on the subgraph-level cooperative control unit, instruction calculation is performed to obtain a set of cooperative control instructions, including: For each subgraph-level collaborative control unit, a collaborative control strategy is formulated with core critical nodes as the priority control positions and core collaborative objects as the priority adjustment objects. The collaborative control strategies corresponding to each subgraph-level collaborative control unit are summarized to generate a collaborative control instruction set, which is then sent to the control terminals of each intelligent agent for execution.

9. The multi-agent cooperative control method according to claim 8, characterized in that, Cooperative control strategies include traffic order adjustment, speed adjustment, waiting for release, and local rerouting; Among them, the passage order adjustment is based on the time overlap of each associated agent in the same subgraph-level collaborative control unit at key nodes, edges or conflicting positions. The passage sequence is determined according to the priority rules, and the passage order adjustment instructions of the corresponding agents are generated according to the passage sequence. Speed ​​adjustment: Based on the conflict state in front of the agent and the propagation potential level of the edge segment it is located in, the running speed of the agent is adjusted and a speed adjustment command is generated; Waiting to release: Based on the conflict state before the agent enters the high propagation potential region, guide the agent to the waiting position and generate a waiting release instruction according to the release conditions of the conflict region; Local rerouting: Based on the propagation potential level of subsequent edge segments in the original path and the relationship between its input strength and edge elastic capacity threshold, the affected edge segments are avoided, and the local path is replanned after avoidance to generate local rerouting instructions.

10. The multi-agent cooperative control method according to claim 8, characterized in that, After the set of collaborative control commands is sent to the control terminals of each intelligent agent for execution, the process also includes closed-loop evaluation and parameter correction steps: Collect the actual execution results of each agent and write them back to the time-varying road network diagram, conflict mapping diagram and subgraph-level cooperative control unit; The current control effectiveness is evaluated based on indicators such as the number of conflict resolutions, changes in waiting time, changes in propagation potential, and changes in capacity utilization. Based on the evaluation results, the parameters of the collaborative control strategy in subsequent control cycles are modified.

Images