Unmanned aerial vehicle complex airspace cooperative passing method based on graph neural network

Abstract

The invention discloses an unmanned aerial vehicle complex airspace cooperative passage method based on a graph neural network, and the method comprises the steps: collecting the airspace data of an unmanned aerial vehicle complex airspace, and generating standardized airspace data; constructing a mixed queuing network, and forming arrival, queue length and service capability state of each service node and service link; mapping to obtain graph nodes, graph edges and graph features, and constructing an airspace graph; inputting the airspace graph into the graph neural network, and carrying out constraint updating on the mixed queuing network; integrating the service capability state, the queue length state and the congestion propagation parameters to form a network passing cost relationship; and executing a Frank-Wolfe algorithm, and generating a cooperative passing instruction for each unmanned aerial vehicle. According to the invention, by introducing the mixed queuing network and the Frank-Wolfe algorithm, efficient and stable cooperative passage scheduling of multiple unmanned aerial vehicles in a complex airspace environment is realized.

Description

Technical Field

[0001] This invention relates to the field of UAV airspace management technology, and in particular to a method for collaborative passage of UAVs in complex airspace based on graph neural networks. Background Technology

[0002] With the gradual opening of low-altitude airspace and the large-scale application of drones in logistics, inspection and monitoring, and emergency response, the operational density of drones continues to increase, making the issue of collaborative passage in complex airspace environments increasingly prominent. Traditional airspace management and drone scheduling technologies are mostly based on static route planning or local obstacle avoidance strategies, typically simplifying the airspace into discrete segments or grid structures and controlling passage according to fixed rules or instantaneous states. These methods can meet basic operational requirements under low-density, regular airspace conditions, but in complex airspace with multiple segment intersections, sector boundary constraints, and dynamic environmental disturbances, problems such as local congestion, queue accumulation, and the propagation of passage conflicts are prone to occur.

[0003] Existing technologies attempt to optimize drone traffic flow by introducing network modeling or intelligent algorithms. For example, they use traffic flow models, queuing models, or graph structures to abstractly describe airspace traffic relationships and combine these with optimization algorithms to allocate paths or traffic volumes. While these fundamental technologies can reflect the airspace topology and traffic load to some extent, they typically employ fixed parameters or static constraints, making it difficult to dynamically characterize changes in service capacity and congestion propagation effects. Furthermore, they do not adequately consider the differences between various types of traffic nodes and links. Moreover, existing methods often lack an effective closed-loop mechanism between network state updates and traffic decisions, making it difficult to achieve refined and highly coordinated traffic scheduling under continuously changing airspace conditions. This limits their effectiveness in high-density, complex airspaces.

[0004] Therefore, how to provide a method for collaborative passage of UAVs in complex airspace based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for collaborative passage of unmanned aerial vehicles (UAVs) in complex airspace based on graph neural networks. This invention fully utilizes the modeling capabilities of graph neural networks, queuing network theory, and continuous network optimization methods to systematically model and solve the collaborative passage problem of UAVs in complex airspace. By introducing a hybrid queuing network, the passage relationships of flight segments, intersection nodes, and sector entrances and exits are expressed in a service-oriented manner. Arrival status, queue length status, and service capacity status are mapped to airspace graph features. Graph neural networks are used to perform constraint learning and dynamic correction of airspace operation status, forming a network passage cost relationship that reflects congestion propagation and changes in passage capacity. Combined with the Frank-Wolfe algorithm, continuous flow and path joint allocation is performed for UAV passage demand, generating collaborative passage instructions for multiple UAVs. This invention effectively characterizes the evolution of passage load and congestion propagation mechanisms in complex airspace, possessing advantages such as high passage efficiency, strong collaboration, strong adaptability to dynamic airspace, and good stability of scheduling results. It is suitable for high-density, highly constrained UAV operation scenarios in complex airspace.

[0006] A method for cooperative passage of unmanned aerial vehicles in complex airspace based on graph neural networks according to an embodiment of the present invention includes:

[0007] Collect airspace data from complex airspaces of UAVs, preprocess the airspace data, and generate standardized airspace data;

[0008] A hybrid queuing network is constructed based on standardized airspace data to map the passage relationships of flight segments, intersection nodes, and sector entrances and exits, forming the arrival status, queue length status, and service capacity status of each service node and service link.

[0009] In the hybrid queuing network, service nodes are mapped to graph nodes, service links are mapped to graph edges, and arrival status, queue length status, and service capacity status are used as graph features to construct a spatial graph.

[0010] The spatial graph is input into the graph neural network to obtain the effective service capacity parameters, capacity correction parameters and congestion propagation parameters corresponding to each service node and service link, and the constraints of the hybrid queuing network are updated.

[0011] Based on the updated hybrid queuing network, the travel cost information of flight segments and intersection nodes is determined by integrating service capacity status, queue length status and congestion propagation parameters, thus forming network travel cost relationships.

[0012] Based on the network access cost relationship, the Frank-Wolfe algorithm is used to iteratively calculate the access requirements of UAVs, obtain the segment traffic allocation results and flight path allocation results, and generate cooperative access instructions for each UAV.

[0013] Optionally, the airspace data specifically includes airspace structure and route network data, airspace rules and operational constraints data, airspace restrictions and obstacles data, environmental and risk field data, and airspace operational status data.

[0014] Optionally, the preprocessing of the spatial data specifically includes data cleaning and consistency verification, coordinate and benchmark unification, time alignment, topology and indexing.

[0015] Optionally, the process of forming the arrival status, queue length status, and service capability status corresponding to each service node and service link includes:

[0016] Based on standardized airspace data, a hybrid queuing network is constructed to describe the service nodes, service links, and queuing relationships of UAVs in complex airspace. Depending on the type of UAV mission, the corresponding passage behavior is mapped to different queuing operation mechanisms. The hybrid queuing network structurally includes a heterogeneous service node layer and a directional link expression layer. Furthermore, the operational state of the queuing network is modeled and updated through a state fusion representation layer, a multi-dimensional topology perception layer, and a dynamic temporal update layer.

[0017] Based on airspace structure and airway network data, service units are divided into segments, junction nodes, sector entrances and exits. These service units are mapped to service nodes in a hybrid queuing network, forming a set of service nodes that includes segment service nodes, junction service nodes, and sector entrance and exit service nodes.

[0018] Based on airspace structure and airway network data, according to the connection relationship of air segments, the merging and diverging relationship of converging nodes and the adjacency relationship between sectors, the passage association relationship between service nodes is established and mapped into service links with clear passage directions, forming a set of service links;

[0019] Based on the airspace operation status data, the arrival status, queue length status and service capability status of each service node and service link are determined for the service node set and service link set.

[0020] Optionally, constructing the spatial domain graph includes:

[0021] Based on the hybrid queuing network, the service node set and service link set are extracted to determine the graph nodes and graph edges;

[0022] Map each service node in the service node set to a graph node in the spatial domain graph, map each service link in the service link set to a graph edge in the spatial domain graph, and establish the topological relationship between graph nodes connected by graph edges.

[0023] Based on arrival status, queue length status, and service capability status, configure corresponding node feature information for each graph node and corresponding edge feature information for each graph edge.

[0024] Based on graph nodes, graph edges, node feature information, and edge feature information, a spatial graph is constructed that uses arrival status, queue length status, and service capability status as graph features.

[0025] Optionally, the constraint update of the hybrid queuing network includes:

[0026] Extract the node feature information corresponding to each graph node and the edge feature information corresponding to each graph edge in the spatial domain graph. Use a graph neural network to perform feature propagation and feature aggregation on the node feature information and edge feature information to obtain the updated feature information of each graph node and each graph edge.

[0027] Based on the updated feature information, the effective service capacity parameters corresponding to each service node, the capacity correction parameters corresponding to each service link, and the congestion propagation parameters corresponding to each service node and each service link are mapped to obtain the effective service capacity parameters corresponding to each service node and each service link.

[0028] Based on effective service capacity parameters and capacity correction parameters, the service capacity status of each service node and each service link in the hybrid queuing network is updated, and the queue length status of each service node and each service link in the hybrid queuing network is updated based on congestion propagation parameters.

[0029] Write the updated service capacity status and the updated queue length status into the hybrid queuing network to obtain the constrained updated hybrid queuing network.

[0030] Optionally, the formation of network access cost relationships includes:

[0031] The heterogeneous service node layer extracts segment service nodes, convergence service nodes, and sector entry and exit service nodes, and adds corresponding structural role identifiers to each service node. At the same time, it obtains the service capability status, queue length status, and congestion propagation parameters of each service node to form a set of structured state representations.

[0032] The directional link expression layer is based on a structured state representation set. It extracts the service links associated with each service node and sets different directional weights and coupling interference weights for each pair of forward and reverse links. Logical jump links are introduced to form an asymmetric weighted link structure set.

[0033] The state fusion representation layer jointly associates the structured state representation set with the asymmetric weighted link structure set, and introduces a state feature stack to organize and store the service capability status, queue length status and congestion propagation parameters of each service node and each service link, forming a fused state set.

[0034] The multidimensional topology perception layer performs topology association processing on the fused state set based on the connection relationship and logical jump link between service nodes and service links, determines the adjacency relationship and propagation association range of each service node and each service link, and performs topology propagation correction on the state feature stack structure according to the structural role identifier and congestion propagation parameters to form a topology corrected state set.

[0035] The dynamic time-series update layer constructs an auxiliary occupancy graph based on the topology correction state set, records the occupancy interval of each service node and each service link in the time dimension, and performs time-series updates on the passage status of service nodes and service links in combination with the state feature stack, generating the node passage status and link passage status of the current scheduling period.

[0036] Based on the node passage status and link passage status, the passage cost of intersecting nodes and the passage cost of flight segments are determined, and the network passage cost relationship is obtained by combining the connection relationship between service nodes and service links.

[0037] Optionally, generating cooperative passage instructions for each drone includes:

[0038] Based on the network traffic cost relationship, the set of flight segments and the set of intersection nodes in the airspace network are determined, and the traffic cost information corresponding to each flight segment is obtained. At the same time, the set of UAVs participating in cooperative traffic is determined, and the corresponding traffic demand for each UAV is determined and a hierarchical candidate path set is constructed.

[0039] Based on the set of flight segments, the set of intersection nodes, and the information on passage costs, the overall passage cost is taken as the network passage optimization target. Consistent relationship constraints between flight segment passage flow and UAV path passage allocation, as well as conservation constraints on passage demand, are established.

[0040] Under the conditions of satisfying the consistency relationship constraint and the passage demand conservation constraint, the Frank-Wolfe algorithm is executed to generate an initial UAV path passage allocation scheme, obtain the corresponding segment passage flow allocation status, and establish a state stack for the segment and the intersection node to store the passage cost status, segment passage flow status, queue length status and service capacity status, forming a set of state sequences.

[0041] Based on the segment traffic flow allocation status and status stack structure, the traffic cost corresponding to each segment and each intersection node is updated and calculated. In the hierarchical candidate path set, the current minimum traffic cost path set is determined for each UAV, and a path-resource heterogeneous two-layer graph is constructed.

[0042] Based on the minimum passage cost path set and the path-resource heterogeneous two-layer graph, path filtering and priority adjustment are performed to build a corresponding auxiliary passage allocation scheme for each UAV, and the corresponding auxiliary segment passage flow allocation status is formed to determine the search direction of passage allocation.

[0043] Based on the search direction of traffic allocation, the update ratio is determined by comparing the overall traffic cost within a preset step size. The traffic flow allocation status of the current flight segment and the UAV path traffic allocation scheme are updated according to the update ratio, and the iteration process is terminated when the preset convergence condition is met.

[0044] Based on the segment traffic flow allocation results and UAV path traffic allocation results obtained at the end of the iteration, a corresponding flight path is determined for each UAV, and a coordinated passage instruction containing the segment passage order, the passing order of intersection nodes, and the passage time arrangement is output.

[0045] The beneficial effects of this invention are:

[0046] This invention proposes a collaborative UAV airspace passage method based on graph neural networks. Addressing the problems of coarse characterization of passage capacity, insufficient modeling of congestion propagation mechanisms, and poor stability of collaborative scheduling in existing UAV airspace passage technologies, this invention systematically describes the operational state of complex airspace by introducing a modeling approach combining hybrid queuing networks and graph neural networks. Air segments, intersection nodes, and sector boundaries in the airspace are abstracted into service nodes and service links with service semantics. An airspace operation network model reflecting arrival intensity, queue length, and service capacity is constructed. Graph neural networks are used to learn the passage status and congestion propagation relationships between multiple nodes and links, dynamically correcting the effective service capacity and passage cost parameters of the airspace.

[0047] The Frank-Wolfe algorithm is used to jointly optimize the path and traffic flow for the passage requirements of multiple UAVs, generating a cooperative passage scheme that meets airspace constraints. This invention accurately reflects the evolution of traffic load under dynamic airspace conditions, effectively suppresses the spread of local congestion to the global level, improves the consistency and stability of multi-UAV passage decisions, and balances passage efficiency and safety. It offers advantages such as strong adaptability to complex airspace, good cooperative scheduling effect, and high operational reliability. Attached Figure Description

[0048] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0049] Figure 1 This is a flowchart of a method for collaborative passage in complex airspace by unmanned aerial vehicles based on graph neural networks, as proposed in this invention.

[0050] Figure 2 This is a schematic diagram of the hybrid queuing network structure of a UAV complex airspace cooperative passage method based on graph neural network proposed in this invention;

[0051] Figure 3 This is a flowchart illustrating the execution of the Frank-Wolfe algorithm, which is proposed in this invention as a method for cooperative passage of unmanned aerial vehicles in complex airspace based on graph neural networks. Detailed Implementation

[0052] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0053] refer to Figure 1 , Figure 2 and Figure 3 A method for cooperative passage of unmanned aerial vehicles (UAVs) in complex airspace based on graph neural networks includes:

[0054] Collect airspace data from complex airspaces of UAVs, preprocess the airspace data, and generate standardized airspace data;

[0055] A hybrid queuing network is constructed based on standardized airspace data to map the passage relationships of flight segments, intersection nodes, and sector entrances and exits, forming the arrival status, queue length status, and service capacity status of each service node and service link.

[0056] In the hybrid queuing network, service nodes are mapped to graph nodes, service links are mapped to graph edges, and arrival status, queue length status, and service capacity status are used as graph features to construct a spatial graph.

[0057] The spatial graph is input into the graph neural network to obtain the effective service capacity parameters, capacity correction parameters and congestion propagation parameters corresponding to each service node and service link, and the constraints of the hybrid queuing network are updated.

[0058] Based on the updated hybrid queuing network, the travel cost information of flight segments and intersection nodes is determined by integrating service capacity status, queue length status and congestion propagation parameters, thus forming network travel cost relationships.

[0059] Based on the network access cost relationship, the Frank-Wolfe algorithm is used to iteratively calculate the access requirements of UAVs, obtain the segment traffic allocation results and flight path allocation results, and generate cooperative access instructions for each UAV.

[0060] In this embodiment, the airspace data specifically includes airspace structure and route network data, airspace rules and operational constraints data, airspace restrictions and obstacles data, environmental and risk field data, and airspace operational status data.

[0061] In this embodiment, the preprocessing of spatial data specifically includes data cleaning and consistency verification, coordinate and benchmark unification, time alignment, topology and indexing.

[0062] In this embodiment, the process of forming the arrival status, queue length status, and service capability status corresponding to each service node and service link includes:

[0063] Based on standardized airspace data, a hybrid queuing network is constructed to describe the service nodes, service links, and queuing relationships of UAVs in complex airspace. According to different UAV mission types, corresponding passage behaviors are mapped to different queuing operation mechanisms. The hybrid queuing network structurally includes a heterogeneous service node layer and a direction link expression layer. The operational state of the queuing network is modeled and updated through a state fusion representation layer, a multi-dimensional topology perception layer, and a dynamic temporal update layer. Specifically, mapping corresponding passage behaviors to different queuing operation mechanisms involves:

[0064] During the airspace modeling process, the type of UAV passage mission is identified to determine whether it belongs to periodic operation, continuous passage across multiple segments, or one-time or temporary passage mission. For UAV missions that operate periodically or continuously between multiple segments, they continue to participate in airspace resource competition after completing passage through a single segment and node, maintaining backflow characteristics in the hybrid queuing network. For one-time and temporary passage missions that leave the current airspace after completing passage and no longer participate in subsequent resource competition, they no longer participate in airspace resource competition after completing passage through the corresponding service node and service link.

[0065] Based on airspace structure and route network data, service units are divided into segments, junction nodes, and sector ingress and egress nodes. These service units are mapped to service nodes in a hybrid queuing network, forming a service node set that includes segment service nodes, junction service nodes, and sector ingress and egress service nodes. The formation of this service node set is specifically as follows:

[0066] Based on airspace structure and airway network data, the topology of the airspace airway network is analyzed. The segments connecting adjacent intersections are treated as linear passage units with continuous passage capability constraints and mapped as segment service nodes. The locations where multiple segments merge and diverge are identified as intersection nodes and mapped as intersection service nodes. The sector boundary entry and exit positions that control UAVs entering and leaving different sectors are mapped as sector entry service nodes and sector exit service nodes, respectively, forming a service node set including segment service nodes, intersection service nodes, and sector entry and exit service nodes.

[0067] Based on airspace structure and route network data, according to segment connectivity, convergence and divergence relationships of intersecting nodes, and adjacency relationships between sectors, a service link set is established and mapped to service links with clearly defined directions. Specifically, forming the service link set involves:

[0068] Based on airspace structure and airway network data, the service node set is analyzed for topological relationships. The sequential traffic relationship between service nodes of air segments is determined according to the spatial connection order of air segments. The traffic relationship between converging service nodes and adjacent air segment service nodes is determined according to the merging and diverging directions of air segments at converging nodes. The traffic relationship between sector entry service nodes, sector exit service nodes and adjacent service nodes is determined according to the cross-sector connection of airways at sector boundaries. The above traffic relationships are mapped into directed service links with clear traffic directions. Each service link is labeled with a starting service node and a target service node, forming a service link set.

[0069] Based on airspace operational status data, the arrival status, queue length status, and service capability status of each service node and service link are determined for the service node set and service link set. Specifically, the determination of the arrival status, queue length status, and service capability status of each service node and service link is as follows:

[0070] Based on airspace operation status data, the operation status of each service unit in the service node set and service link set is dynamically analyzed. The number of UAVs entering each service node and traveling along each service link within the time window is counted to determine the arrival status of each service node and service link. The number of UAVs that have entered each service node but have not yet completed passage is accumulated and counted. Combined with the waiting time of UAVs at the service nodes, the queue length status of each service node and service link is obtained. Based on airspace operation status, segment geometry characteristics, intersection node conflict handling capabilities, and sector boundary passage control requirements, the maximum passage rate and passage capacity supported by each service node and service link are estimated to determine the corresponding service capability status.

[0071] In this embodiment, constructing the spatial domain graph includes:

[0072] Based on the hybrid queuing network, the service node set and service link set are extracted to determine the graph nodes and graph edges. Specifically, determining the graph nodes and graph edges involves:

[0073] Based on the set of service nodes and service links already constructed in the hybrid queuing network, each service node is mapped to a graph node in the spatial domain graph. The passage relationship established between service nodes through service links is extracted, and each service link is mapped to a graph edge in the spatial domain graph according to the starting service node, the target service node, and the passage direction.

[0074] Map each service node in the service node set to a graph node in the spatial domain graph, map each service link in the service link set to a graph edge in the spatial domain graph, and establish the topological relationship between graph nodes connected by graph edges.

[0075] Based on arrival status, queue length status, and service capability status, corresponding node feature information is configured for each graph node, and corresponding edge feature information is configured for each graph edge, wherein:

[0076] The specific steps for configuring corresponding node feature information for each graph node are as follows:

[0077] For each graph node, the arrival status, queue length status, and service capability status of the service node are used as node feature information, based on the corresponding service node type and current running status.

[0078] The specific steps for configuring corresponding edge feature information for each graph edge are as follows:

[0079] For each graph edge, the arrival status, queue delivery status, and link connectivity related to the service link are used as edge feature information, in combination with the corresponding service link attributes.

[0080] Based on graph nodes, graph edges, node feature information, and edge feature information, a spatial domain graph is constructed, using arrival status, queue length status, and service capability status as graph features. The construction of the spatial domain graph specifically involves:

[0081] The graph nodes mapped from the service nodes are used as the node set of the spatial graph, and the graph edges mapped from the service links are used as the edge set of the spatial graph. The connection topology between the graph nodes is established according to the relationship between the starting service node and the target service node corresponding to each graph edge. The node feature information and edge feature information are written into the corresponding graph nodes and graph edges respectively to form the spatial graph.

[0082] In this embodiment, the constraint update of the hybrid queuing network includes:

[0083] Node feature information corresponding to each graph node and edge feature information corresponding to each graph edge are extracted from the spatial domain graph. A graph neural network is then used to perform feature propagation and feature aggregation on the node and edge feature information to obtain updated feature information for each graph node and edge. The feature propagation and feature aggregation processes are specifically performed as follows:

[0084] The node feature vectors of each graph node and the edge feature vectors of each graph edge are read from the spatial domain graph. Based on the adjacency relationship of the spatial domain graph, the set of neighboring nodes and the set of associated graph edges of each graph node are determined. In the graph neural network, the graph edges are used as information transmission channels. The source node features and the corresponding graph edge features are jointly mapped to form a message representation transmitted along the directed graph edges. The messages received by the target node from the neighboring nodes are accumulated according to the connection direction of the graph to obtain the aggregated features of the target node.

[0085] After feature aggregation is completed, the target node’s own features are fused and updated with the aggregated features, and the graph edge features are jointly updated based on the updated features of the two end nodes and the original edge features, finally obtaining the updated feature information of each graph node and each graph edge.

[0086] Based on the updated feature information, the effective service capacity parameters corresponding to each service node, the capacity correction parameters corresponding to each service link, and the congestion propagation parameters corresponding to each service node and each service link are mapped to obtain the following:

[0087] Based on the updated feature information of each graph node and each graph edge output by the graph neural network, semantic parsing and parameter mapping are performed on the updated features. The parameter mapping relationship between the updated feature information and the graph node is established to obtain the effective service capability parameters corresponding to each service node.

[0088] The updated feature information corresponding to the graph edges is mapped to capacity adjustment to obtain the capacity correction parameters for each service link.

[0089] By combining the load change trends and mutual influence relationships reflected in the update features of graph nodes and graph edges, the state propagation intensity is extracted through the congestion impact mapping relationship, and the congestion propagation parameters corresponding to each service node and each service link are determined.

[0090] Based on effective service capacity parameters and capacity correction parameters, the service capacity status of each service node and each service link in the hybrid queuing network is updated, and the queue length status of each service node and each service link in the hybrid queuing network is updated based on congestion propagation parameters, wherein:

[0091] The update of the service capability status of each service node and each service link in the hybrid queuing network specifically involves:

[0092] Based on the operational status of service nodes and the impact of the surrounding environment on traffic capacity, the maximum carrying capacity of each service node and service link is adjusted in conjunction with capacity correction parameters, and this is updated as the new service capacity status.

[0093] The update of the queue length status corresponding to each service node and each service link in the hybrid queuing network based on congestion propagation parameters is specifically as follows:

[0094] Based on congestion propagation parameters, combined with node and edge characteristics, the queue length state of each service node and service link in the hybrid queuing network is updated. Specifically, the queue length state of the service node and service link is adjusted by calculating the congestion propagation effect and the influence of neighboring nodes.

[0095] Write the updated service capacity status and the updated queue length status into the hybrid queuing network to obtain the constrained updated hybrid queuing network.

[0096] In this embodiment, the formation of network access cost relationships includes:

[0097] The heterogeneous service node layer extracts segment service nodes, convergence service nodes, and sector entry and exit service nodes, and adds corresponding structural role identifiers to each service node. At the same time, it obtains the service capability status, queue length status, and congestion propagation parameters of each service node to form a set of structured state representations.

[0098] The directional link expression layer, based on a structured state representation set, extracts the service links associated with each service node and sets different directional weights and coupling interference weights for each pair of forward and reverse links. Logical jump links are introduced to form an asymmetric weighted link structure set, where:

[0099] A forward-moving link is the path a drone takes from the starting service node to the target service node along the service link. A reverse-moving link is the link opposite to the forward-moving link, representing the path a drone takes from the target service node back to the starting node, which is usually the return path.

[0100] The specific steps involve setting different directional weights and coupling interference weights for each pair of forward and reverse passage links, as follows:

[0101] Forward and reverse links are assigned different directional weights based on their roles and traffic priorities in the airspace network. Forward links carry a larger volume of traffic and are assigned a weight higher than the weight threshold, while reverse links carry a smaller volume of traffic and are assigned a weight lower than the weight threshold.

[0102] The coupling interference weight represents the interference intensity between paths. For two highly overlapping links, the coupling interference weight will be higher than the interference weight threshold. For links that do not overlap and are far apart, the interference weight will be lower than the interference weight threshold.

[0103] Logical hop links refer to a connection path in an airspace network where two service nodes are not directly connected, but through specific airspace configurations or flight plans, there exists a connection path that spans multiple service nodes or flight segments. UAVs can quickly travel from one area to another in the airspace through logical hop links, bypassing possible congestion and complex areas.

[0104] The state fusion representation layer jointly associates the structured state representation set with the asymmetric weighted link structure set, and introduces a state feature stack to organize and store the service capability state, queue length state, and congestion propagation parameters of each service node and each service link, forming a fused state set, in which:

[0105] The state feature stack is a data structure that organizes and stores state information from multiple rounds. The state feature stack saves the state data of each round in chronological order in the form of a stack, including service capacity status, queue length status and congestion propagation parameters, and dynamically records the state changes of each service node and service link in different iterations.

[0106] The formation of the fusion state set specifically refers to:

[0107] The service capability status will be stored as a value based on the actual passage capacity of each service node, representing the available service capacity of the node in the current time window. The queue length status will record the number of drones waiting to pass on each service node and the time, and store it in the form of queue length. The congestion propagation parameters will be stored according to the congestion status and propagation effect between nodes. All status information is stored sequentially in the status feature stack.

[0108] The multi-dimensional topology awareness layer, based on the connection relationships and logical jump links between service nodes and service links, performs topology association processing on the fused state set to determine the adjacency relationships and propagation association ranges of each service node and each service link. Furthermore, based on structural role identifiers and congestion propagation parameters, it performs topology propagation correction on the state feature stack structure, forming a topology-corrected state set, where:

[0109] The determination of the adjacency relationship and propagation range between each service node and each service link specifically includes:

[0110] If two service nodes are connected through a service link, they are marked as connected in the adjacency relationship. For logical hop links, even if the service nodes are not directly connected, they can establish an indirect connection through logical hops and are marked as adjacent in the adjacency relationship. The propagation association range refers to the range of influence of the state of each service node on the state of adjacent nodes, which depends on the distance between nodes, the congestion propagation effect, and the service capacity.

[0111] The topology propagation correction of the state feature stack structure specifically involves:

[0112] Based on the structural role identifiers of each service node and the current congestion propagation parameters, topology propagation is performed on each node in the state feature stack. The impact is transmitted to adjacent service nodes based on the current node's state information. During the propagation process, the node's state information is corrected according to the connection relationship and the state of adjacent nodes. The service capacity, queue length, and congestion propagation parameters of each node are updated in a weighted manner according to the state of adjacent nodes during the propagation process, and the data in the state feature stack is adjusted accordingly. All corrected node states are summarized to form a topology corrected state set.

[0113] The dynamic time-series update layer constructs an auxiliary occupancy graph based on the topology correction state set, records the occupancy interval of each service node and each service link in the time dimension, and performs time-series updates on the access status of service nodes and service links in conjunction with the state feature stack, generating the node access status and link access status for the current scheduling period, where:

[0114] The construction of the auxiliary occupancy map specifically involves:

[0115] For each service node, the throughput, queue length and congestion propagation status are recorded within a specific time period. For the service link, the traffic, congestion level and resource consumption in time sequence are recorded and transformed into graph nodes and graph edges of the occupancy graph. An auxiliary occupancy graph is constructed to reflect the occupancy status of each service node and service link in the time dimension within the scheduling cycle.

[0116] The time-series update of the access status of service nodes and service links by combining the state feature stack is as follows:

[0117] Read the historical state data of each node and link in the state feature stack, and combine it with the current topology correction state set to update the real-time state of each service node and service link. By comparing the historical state and the current state, and combining the occupancy time interval of each node and link in the auxiliary occupancy graph, adjust the passage status of the node and link. If the queue length of the service node is long in the current time period, the state will be adjusted to a higher level of congestion.

[0118] Based on node traffic status and link traffic status, the traffic cost of intersecting nodes and the traffic cost of flight segments are determined. Furthermore, the network traffic cost relationship is obtained by combining the connection relationship between service nodes and service links, where:

[0119] The determination of the passage cost at the intersection node and the passage cost for the flight segment is specifically as follows:

[0120] The passage cost of a convergence node is determined by evaluating the node's airspace passage capacity, queuing status, and congestion propagation. The passage cost of a convergence node can be calculated based on the node's queue length, arrival rate, and congestion propagation parameters, and then weighted and summed to obtain the passage cost of the convergence node that reflects the difficulty of the node in handling drone traffic at the current moment.

[0121] The segment passage cost is assessed based on the segment's passage capacity, drone traffic within the segment, and potential resource constraints. The segment passage cost is closely related to the segment's load status, congestion level, and whether there are any special passage constraints. The segment passage cost can be calculated by weighting the segment's service capacity, traffic load, and resource constraints.

[0122] The network access cost relationship is obtained by combining the connection relationship between service nodes and service links, specifically as follows:

[0123] By determining the connection relationship between service nodes and service links, the travel cost between service links is obtained, and thus the cost of the entire network is obtained. The travel cost of service nodes propagates and accumulates in the network. By calculating the travel cost of each service link and the influence of the service node, the total travel cost in the network is obtained by summing them.

[0124] In this embodiment, generating cooperative passage instructions for each UAV includes:

[0125] Based on network traffic cost relationships, the set of flight segments and the set of intersection nodes in the airspace network are determined, and the traffic cost information corresponding to each flight segment is obtained. Simultaneously, the set of UAVs participating in coordinated traffic is determined, and the corresponding traffic demand is assigned to each UAV, and a hierarchical candidate path set is constructed, where:

[0126] The determination of the set of flight segments and the set of intersection nodes in the airspace network specifically involves:

[0127] Based on the spatial layout of the airspace and the data of the airway network, the set of air segments within the airspace is determined. The set of air segments consists of various airway segments in the airspace. Each air segment represents the flight path of the UAV from one service node to another. Each air segment is assigned passage cost information.

[0128] The set of intersection nodes consists of various intersection nodes in the airspace. Intersection nodes are usually the intersection points of two or more flight segments, representing key nodes for UAVs to merge or diverge in the airspace.

[0129] The process of determining the corresponding traffic demand for each drone and constructing a hierarchical candidate path set specifically involves:

[0130] Based on the drone's mission planning and objectives, the starting point, destination, and expected flight path are determined, and then the traffic flow and time requirements for the drone are calculated to obtain the drone's traffic demand.

[0131] Based on the starting and target positions of the UAV, different paths are selected in the airspace network. The candidate path set is divided into multiple main candidate path sets, secondary candidate path sets, and buffer candidate path sets. The main candidate path set contains the optimal path, the secondary candidate path set contains the backup path, and the buffer candidate path set is a set of paths reserved to deal with emergencies. Each path consists of a sequence of flight segments and a sequence of intersection nodes, and an index is established with the relationship between the path and the network passage cost.

[0132] Based on the set of flight segments, the set of intersecting nodes, and the information on travel costs, the overall travel cost is used as the network travel optimization objective. Consistent relationship constraints are established between flight segment traffic flow and UAV path travel allocation, along with conservation constraints on travel demand. Specifically:

[0133] Consistency constraint refers to the requirement in network optimization that the path allocation of UAVs and the traffic flow of flight segments remain consistent. During the optimization process, the path of each UAV must pass through the corresponding flight segment, and the traffic flow of each flight segment must be consistent with the traffic flow of its corresponding UAV.

[0134] The conservation constraint of traffic demand means that during the optimization process of the entire airspace network, the traffic demand of each UAV must be reasonably allocated on every node and link in the network, and the entry and exit states of the UAV should be consistent.

[0135] Under the constraints of consistency and traffic demand conservation, the Frank-Wolfe algorithm is executed to generate an initial UAV path allocation scheme, obtain the corresponding segment traffic flow allocation status, and establish a state stack for segments and intersection nodes to store the traffic cost status, segment traffic flow status, queue length status, and service capacity status, forming a set of state sequences, where:

[0136] The initial drone path allocation scheme is generated as follows:

[0137] Based on the mission requirements and target locations of the UAVs, multiple feasible flight paths are calculated for each UAV. The paths are usually selected based on the set of flight segments, the set of intersection nodes and the corresponding passage cost information in the airspace network. It is necessary to ensure that the passage cost of the path is below the threshold. Multiple predetermined paths are assigned to each UAV to obtain the initial UAV path passage allocation scheme.

[0138] The establishment of a state stack for flight segments and rendezvous nodes is specifically as follows:

[0139] By recording the state changes of each flight segment and meeting node in different scheduling cycles and iterations, and storing them in chronological order, a state stack is obtained. In each scheduling cycle, the state of the flight segment and meeting node is updated according to the current network state. For each flight segment and meeting node, the current passage cost, traffic status, queue length status and service capacity status are recorded and pushed into the state stack. In each iteration, new state information is pushed to the top of the state stack, while historical state information is stored in the stack, forming an ordered set of state sequences.

[0140] Based on the segment traffic flow allocation state and state stack structure, the traffic cost corresponding to each segment and each intersection node is updated and calculated. Then, the set of paths with the current minimum traffic cost is determined for each UAV in the hierarchical candidate path set, and a path-resource heterogeneous two-layer graph is constructed, where:

[0141] The update calculation of the passage cost corresponding to each flight segment and each intersection node is as follows:

[0142] Based on the segment traffic flow allocation status and historical data in the status stack, the actual passage cost of the segment is determined. Segments with high traffic density may have a passage cost higher than the threshold, while idle and less congested segments are assigned a cost lower than the threshold. For intersecting nodes, the passage cost is mainly calculated based on the node's waiting time, queue length, and congestion propagation status, and then weighted and summed to obtain the current passage cost of each segment and intersecting node.

[0143] The step of determining the current minimum passage cost path set for each UAV in the hierarchical candidate path set specifically involves:

[0144] Based on the starting node and target node of each UAV, all possible paths are found in the airspace network. The passage cost of each path is obtained by calculating the sum of the passage costs of all segments and intersection nodes on the path. By comparing the passage costs of all paths, the path with the minimum passage cost is selected as the optimal path set for the UAV. If multiple paths have similar minimum costs, all paths with minimum costs are selected as the candidate path set for the UAV.

[0145] The construction path – resource heterogeneity two-layer graph – is specifically as follows:

[0146] The path layer contains candidate paths for all UAVs. Each path consists of a series of segments and intersection nodes. The resource layer contains service nodes and service links in the airspace. Each service node and service link is represented as an airspace resource. By analyzing the connection relationship between service nodes and service links, cross-layer edges of the graph are established. The service links and service nodes corresponding to each path are connected as edges in the graph, and each edge is assigned a weight. The weight is calculated and summed based on the passage cost, resource consumption, and congestion level. The interaction between the path layer and the resource layer is modeled as a heterogeneous two-layer graph structure, resulting in the path-resource heterogeneous two-layer graph.

[0147] Based on the minimum passage cost path set and the path-resource heterogeneous two-layer graph, path filtering and priority adjustment are performed to construct a corresponding auxiliary passage allocation scheme for each UAV, and to form the corresponding auxiliary segment passage flow allocation status, thus determining the search direction for passage allocation, wherein:

[0148] The specific steps for path filtering and priority adjustment are as follows:

[0149] Based on the edge weight information in the path-resource graph and the passage cost of each path, the feasibility of each path is evaluated. If the passage cost of a path is higher than the threshold and the resource consumption is high, it will be regarded as a low-priority path. Paths that do not meet the requirements will be eliminated first. Links with path conflicts higher than the threshold will be assigned a priority lower than the priority benchmark. Paths with low cost, high resource utilization and less interference will be retained and sorted as high-priority paths for drones to choose from.

[0150] The formation of the corresponding auxiliary segment traffic flow allocation status is specifically as follows:

[0151] By utilizing the traffic allocation of each segment along each path in the selected optimal path set, and based on the traffic allocation of the path and the service capacity status of the segment, the actual traffic flow of each segment within the current time window is calculated. The auxiliary traffic flow allocation status of all segments forms an overall traffic allocation scheme.

[0152] The search direction for determining the passage allocation is specifically as follows:

[0153] Based on the selected set of minimum passage cost paths and the status of auxiliary segment traffic allocation, the difference between the current path's traffic distribution and the expected optimization objective is evaluated. If the current path's traffic distribution does not meet the optimization objective or the path's passage cost is higher than the threshold, the traffic allocation is adjusted to reduce the traffic of high-cost paths and increase the traffic of low-cost paths. In high-conflict areas, the path traffic is reduced, and the search direction guides the traffic redistribution until the preset convergence condition is reached.

[0154] Based on the search direction of traffic allocation, the update ratio is determined by comparing the overall traffic cost within a preset step size. The traffic flow allocation status of the current flight segment and the UAV path traffic allocation scheme are then updated according to this update ratio. The iteration process terminates when a preset convergence condition is met.

[0155] The preset step size range is 0.1 to 0.5 steps;

[0156] The comparison of the overall passage cost within the preset step size range specifically includes:

[0157] Calculate the current overall passage cost, adjust the overall passage cost according to the preset step size range, evaluate whether the overall cost will improve if the current path or traffic allocation status is changed, if the passage cost is reduced after adjusting the step size, continue to adjust in the current search direction, otherwise reduce the adjustment range, and compare the overall passage cost through repeated comparison and calculation.

[0158] The update of the current traffic flow allocation status and UAV path traffic allocation scheme according to the update ratio is specifically as follows:

[0159] Based on the overall passage cost comparison results, the update ratio is calculated in combination with the preset step size range. If the difference between the current cost and the target cost is greater than the threshold, and the step size is greater than the threshold, the update ratio is adjusted to be greater than the ratio threshold. Based on the update ratio, the flow allocation status of the flight segment and the passage allocation scheme of the UAV path are adjusted. If the passage cost of the path is too high, the update ratio will reduce the flow of the path and correspondingly increase the flow of the low-cost path, and gradually adjust the flow allocation status of each flight segment and UAV path.

[0160] The preset convergence condition means that during the iteration process, if the change in the travel cost converges to a stable value, the update will be terminated and the optimization will be completed.

[0161] Based on the segment traffic flow allocation results and UAV path traffic allocation results obtained at the end of the iteration, a corresponding flight path is determined for each UAV, and a coordinated passage instruction containing the segment passage order, the passing order of intersection nodes, and the passage time arrangement is output.

[0162] Example 1:

[0163] To verify the feasibility of this invention in practice, it was applied to the integrated low-altitude logistics and emergency inspection airspace of a coastal city. This airspace covers port and warehouse areas, the airspace above main urban roads, and riverbank areas. Multiple types of flight missions, including logistics delivery drones, inspection drones, and emergency response drones, occur simultaneously within this airspace. The area has densely intersecting air routes, with a significant increase in drone arrivals at some nodes during morning and evening peak hours. Traditional methods based on static air routes or local obstacle avoidance frequently result in segmental queues, the spread of localized congestion to adjacent segments, and multiple drones repeatedly yielding to each other, leading to a decrease in overall traffic efficiency and significant delays for some missions.

[0164] When applying the method of this invention in this scenario, the airspace is divided into a service-oriented airspace structure containing segment nodes, intersection nodes, and sector entry / exit nodes. Each segment corresponds to a service link with service capability parameters, and each intersection node corresponds to a service node with queue status, thus constructing a hybrid queuing network model. During actual operation, the arrival rate of UAVs, the current queue length, and the instantaneous passage capacity of each segment are continuously collected, and these state features are mapped to node features and edge features of the airspace graph structure. These features are then input into a graph neural network for operational state learning and parameter correction. The effective service capability correction value and congestion propagation coefficient output by the graph neural network are written back to the airspace network model to dynamically update the passage cost relationship. The Frank-Wolfe algorithm is used to jointly allocate the passage paths and traffic of multiple UAVs, generate cooperative passage instructions, and issue them to each UAV for execution.

[0165] In actual testing, the average number of UAVs operating in this airspace daily was approximately 180, with up to 62 UAVs simultaneously in the air during peak hours. After adopting the method of this invention, the average queuing time during the morning peak period decreased from 92 seconds to 41 seconds, the maximum queue length at intersection nodes decreased from an average of 8.3 flights to 3.9 flights, and the overall average completion time for UAV passage was shortened by approximately 27%. Under high-density operating conditions, no chain-reaction congestion caused by localized congestion occurred, and the on-time arrival rate of emergency inspection tasks increased from 86.5% to 96.8%. The results indicate that the method of this invention can effectively suppress congestion propagation, improve the stability and operational efficiency of multi-UAV collaborative passage in complex airspace, and has good practical application effects.

[0166] Table 1. Comparison of operational performance of UAV cooperative passage methods under complex airspace conditions

[0167] Comparison indicators Traditional static flight path / local obstacle avoidance methods The collaborative passage method of the present invention Improvement range Number of drones in the air during peak hours (number of drones) 62 62 — Average queuing time per flight segment (seconds) 92 41 ↓ Approximately 55.4% Maximum queuing time for a flight segment (seconds) 168 73 ↓ Approximately 56.5% Average queue length (number of flights) at intersection nodes 8.3 3.9 ↓ Approximately 53.0% Maximum queue length (number of flights) at the intersection node 14 6 ↓ Approximately 57.1% Average time for drones to complete passage (seconds) 486 355 ↓ Approximately 27.0% Number of times the route was repeatedly adjusted during the passage process (per aircraft) 2.6 0.9 ↓ Approximately 65.4% Number of times (per day) that localized congestion spreads to adjacent flight segments 6.4 1.2 ↓ Approximately 81.3% On-time arrival rate of emergency inspection tasks (%) 86.5 96.8 ↑10.3 percentage points Number of times / day that traffic failed or was rolled back during peak hours 4.8 0.7 ↓ Approximately 85.4%

[0168] As can be seen from the data in Table 1, under the condition that the number of UAVs in the air remains the same during peak hours, the cooperative passage method proposed in this invention shows significant advantages in several key operational indicators. In terms of segment queuing, the average queuing time of a segment under the traditional method is 92 seconds, and the maximum queuing time reaches 168 seconds. However, after adopting the method of this invention, the average queuing time of a segment decreases to 41 seconds, and the maximum queuing time decreases to 73 seconds. This indicates that by introducing a hybrid queuing network and combining it with a graph neural network for service capacity correction, it is possible to effectively alleviate segment passage pressure and reduce the waiting time of UAVs in key segments.

[0169] Regarding the operational status at the junction nodes, the traditional method results in an average queue length of 8.3 flights and a maximum queue length of 14 flights, which can easily cause local congestion and spread to adjacent flight segments. The method of this invention controls the average queue length at the junction nodes to within 3.9 flights and reduces the maximum queue length to 6 flights. This shows that the method of this invention can accurately characterize the changes in traffic capacity at the nodes and suppress queue accumulation through a collaborative allocation mechanism.

[0170] In terms of overall traffic efficiency, the average completion time for drone passage has been reduced from 486 seconds to 355 seconds, and the number of path adjustments during passage has been reduced from 2.6 times per flight to 0.9 times. This reflects that the present invention has better stability in the joint allocation of passage paths and traffic flow, and avoids the time loss caused by repeated path adjustments due to local conflicts.

[0171] Regarding congestion propagation and operational reliability, traditional methods, under high-density operating conditions, result in an average of 6.4 instances per day of localized congestion spreading to adjacent flight segments. The method of this invention reduces this number to 1.2 times per day, effectively suppressing the chain reaction of congestion. In terms of mission support capabilities, the on-time arrival rate of emergency inspection missions has increased from 86.5% to 96.8%, and the number of passage failures or backtrackings during peak hours has decreased from 4.8 times per day to 0.7 times per day. This demonstrates that the present invention can balance passage efficiency and safety under complex airspace conditions, significantly improving the overall stability and reliability of multi-UAV collaborative operations.

[0172] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for complex airspace cooperative passing of unmanned aerial vehicles based on a graph neural network, characterized in that, include: Collect airspace data from complex airspaces of UAVs, preprocess the airspace data, and generate standardized airspace data; A hybrid queuing network is constructed based on standardized airspace data to map the passage relationships of flight segments, intersection nodes, and sector entrances and exits, forming the arrival status, queue length status, and service capability status of each service node and service link. In the hybrid queuing network, service nodes are mapped to graph nodes, service links are mapped to graph edges, and arrival status, queue length status, and service capacity status are used as graph features to construct a spatial graph. The spatial graph is input into the graph neural network to obtain the effective service capacity parameters, capacity correction parameters and congestion propagation parameters corresponding to each service node and service link, and the constraints of the hybrid queuing network are updated. Based on the updated hybrid queuing network, the travel cost information of flight segments and intersection nodes is determined by integrating service capacity status, queue length status and congestion propagation parameters, thus forming network travel cost relationships. Based on the network access cost relationship, the Frank-Wolfe algorithm is used to iteratively calculate the access requirements of UAVs, obtain the segment traffic allocation results and flight path allocation results, and generate cooperative access instructions for each UAV.

2. The method of claim 1, wherein, The airspace data specifically includes airspace structure and route network data, airspace rules and operational constraints data, airspace restrictions and obstacles data, environmental and risk field data, and airspace operational status data.

3. The method of claim 1, wherein, The preprocessing of spatial data specifically includes data cleaning and consistency verification, coordinate and benchmark unification, time alignment, topology and indexing.

4. The method of claim 1, wherein, The formation of arrival status, queue length status, and service capability status for each service node and service link includes: Based on standardized airspace data, a hybrid queuing network is constructed to describe the service nodes, service links, and queuing relationships of UAVs in complex airspace. According to the different types of UAV passage tasks, the corresponding passage behaviors are mapped to different queuing operation mechanisms. The hybrid queuing network includes a heterogeneous service node layer and a directional link expression layer in terms of structure. The operation state of the queuing network is modeled and updated through a state fusion representation layer, a multi-dimensional topology perception layer, and a dynamic temporal update layer. Based on airspace structure and airway network data, service units are divided into segments, junction nodes, sector entrances and exits. These service units are mapped to service nodes in a hybrid queuing network, forming a set of service nodes that includes segment service nodes, junction service nodes, and sector entrance and exit service nodes. Based on airspace structure and airway network data, according to the connection relationship of air segments, the merging and diverging relationship of converging nodes and the adjacency relationship between sectors, the passage association relationship between service nodes is established and mapped into service links with clear passage directions, forming a set of service links; Based on airspace operational status data, the arrival status, queue length status, and service capability status of each service node and service link are determined for the service node set and service link set.

5. The method of claim 1, wherein, The construction of the spatial domain graph includes: Based on the hybrid queuing network, the service node set and service link set are extracted to determine the graph nodes and graph edges; Map each service node in the service node set to a graph node in the spatial domain graph, map each service link in the service link set to a graph edge in the spatial domain graph, and establish the topological relationship between graph nodes connected by graph edges. Based on arrival status, queue length status, and service capability status, configure corresponding node feature information for each graph node and corresponding edge feature information for each graph edge. Based on graph nodes, graph edges, node feature information, and edge feature information, a spatial graph is constructed that uses arrival status, queue length status, and service capability status as graph features.

6. The method of claim 1, wherein, The constraint update for the hybrid queuing network includes: Extract the node feature information corresponding to each graph node and the edge feature information corresponding to each graph edge in the spatial domain graph. Use a graph neural network to perform feature propagation and feature aggregation on the node feature information and edge feature information to obtain the updated feature information of each graph node and each graph edge. Based on the updated feature information, the effective service capacity parameters corresponding to each service node, the capacity correction parameters corresponding to each service link, and the congestion propagation parameters corresponding to each service node and each service link are mapped to obtain the effective service capacity parameters corresponding to each service node and each service link. Based on effective service capacity parameters and capacity correction parameters, the service capacity status of each service node and each service link in the hybrid queuing network is updated, and the queue length status of each service node and each service link in the hybrid queuing network is updated based on congestion propagation parameters. Write the updated service capacity status and the updated queue length status into the hybrid queuing network to obtain the constrained updated hybrid queuing network.

7. The method of claim 1, wherein, The formation of network access cost relationships includes: The heterogeneous service node layer extracts segment service nodes, convergence service nodes, and sector entry and exit service nodes, and adds corresponding structural role identifiers to each service node. At the same time, it obtains the service capability status, queue length status, and congestion propagation parameters of each service node to form a set of structured state representations. The directional link expression layer is based on a structured state representation set. It extracts the service links associated with each service node and sets different directional weights and coupling interference weights for each pair of forward and reverse links. Logical jump links are introduced to form an asymmetric weighted link structure set. The state fusion representation layer jointly associates the structured state representation set with the asymmetric weighted link structure set, and introduces a state feature stack to organize and store the service capability status, queue length status and congestion propagation parameters of each service node and each service link, forming a fused state set. The multidimensional topology perception layer performs topology association processing on the fused state set based on the connection relationship and logical jump link between service nodes and service links, determines the adjacency relationship and propagation association range of each service node and each service link, and performs topology propagation correction on the state feature stack structure according to the structural role identifier and congestion propagation parameters to form a topology corrected state set. The dynamic time-series update layer constructs an auxiliary occupancy graph based on the topology correction state set, records the occupancy interval of each service node and each service link in the time dimension, and performs time-series updates on the passage status of service nodes and service links in combination with the state feature stack, generating the node passage status and link passage status of the current scheduling period. Based on the node passage status and link passage status, the passage cost of intersecting nodes and the passage cost of flight segments are determined, and the network passage cost relationship is obtained by combining the connection relationship between service nodes and service links.

8. The method of claim 1, wherein, The generation of cooperative passage instructions for each drone includes: Based on the network traffic cost relationship, the set of flight segments and the set of intersection nodes in the airspace network are determined, and the traffic cost information corresponding to each flight segment is obtained. At the same time, the set of UAVs participating in cooperative traffic is determined, and the corresponding traffic demand for each UAV is determined and a hierarchical candidate path set is constructed. Based on the set of flight segments, the set of intersection nodes, and the information on passage costs, the overall passage cost is taken as the network passage optimization target. Consistent relationship constraints between flight segment passage flow and UAV path passage allocation, as well as conservation constraints on passage demand, are established. Under the conditions of satisfying the consistency relationship constraint and the passage demand conservation constraint, the Frank-Wolfe algorithm is executed to generate an initial UAV path passage allocation scheme, obtain the corresponding segment passage flow allocation status, and establish a state stack for the segment and the intersection node to store the passage cost status, segment passage flow status, queue length status and service capacity status, forming a set of state sequences. Based on the segment traffic flow allocation status and status stack structure, the traffic cost corresponding to each segment and each intersection node is updated and calculated. In the hierarchical candidate path set, the current minimum traffic cost path set is determined for each UAV, and a path-resource heterogeneous two-layer graph is constructed. Based on the minimum passage cost path set and the path-resource heterogeneous two-layer graph, path filtering and priority adjustment are performed to build a corresponding auxiliary passage allocation scheme for each UAV, and form the corresponding auxiliary segment passage traffic allocation status to determine the search direction for passage allocation. Based on the search direction of traffic allocation, the update ratio is determined by comparing the overall traffic cost within a preset step size. The traffic flow allocation status of the current flight segment and the UAV path traffic allocation scheme are updated according to the update ratio, and the iteration process is terminated when the preset convergence condition is met. Based on the segment traffic flow allocation results and UAV path traffic allocation results obtained at the end of the iteration, a corresponding flight path is determined for each UAV, and a coordinated passage instruction containing the segment passage order, the passing order of intersection nodes, and the passage time arrangement is output.

Images