Unmanned aerial vehicle cluster conflict decoupling and flight path reconstruction method and system
By constructing a heterogeneous relationship graph between UAVs, missions, and airspace and using graph neural networks for distributed decision-making, the problem of flight path conflict in multi-UAV swarms in urban low-altitude environments was solved, achieving safe and efficient collaborative flight and energy consumption management, and improving mission completion rate.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FOSHAN UNIVERSITY
- Filing Date
- 2026-01-20
- Publication Date
- 2026-05-12
AI Technical Summary
In urban low-altitude environments, the flight paths of multiple drone swarms are prone to intersection, entanglement, and conflict, leading to mission interruption or delay, increased energy consumption, and impacting endurance and mission completion rate.
Construct a heterogeneous relationship graph of UAV-mission-airspace, use graph neural networks for distributed decision-making and optimization, achieve safe and efficient trajectory planning through attention mechanism and multi-layer stacking, trigger decoupling mechanism to avoid conflicts, and manage energy consumption constraints.
It enables safe and efficient collaborative flight of UAV swarms in complex low-altitude environments, improving mission completion rate and reliability while reducing energy consumption.
Smart Images
Figure CN122015847A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of UAV cooperative control and path planning technology, and in particular to a method and system for UAV swarm conflict decoupling and trajectory reconstruction. Based on an energy-constrained graph neural network, it integrates airspace structure, task association and energy efficiency for multi-objective optimization, which can realize safe, efficient and cooperative flight of UAV swarms. Background Technology
[0002] With the accelerated digitalization of urban airspace and the diversified expansion of drone application scenarios, multi-drone swarm collaborative task execution has become a crucial support for the development of urban low-altitude economy. In typical application scenarios such as logistics delivery, urban inspection, and emergency response, drone swarms can significantly improve task efficiency through collaborative operations. However, in typical urban environments with dense buildings and narrow airspace, drone swarm operation faces severe challenges. Dense building clusters constitute a complex three-dimensional airspace structure, with limited and variable flight paths, making it easy for multiple drone tracks to intersect, become entangled, or even collide. This spatial congestion not only causes interruptions or delays in task execution but also poses serious safety hazards. Simultaneously, the energy competition among drones is becoming increasingly prominent. The additional maneuvers required to avoid collisions lead to a sharp increase in energy consumption, thus affecting the overall swarm's endurance and task completion rate. Therefore, the industry urgently needs an intelligent decision-making method that can uniformly model airspace structure, task relationships, and energy consumption constraints, and possess online optimization and adaptive capabilities. Summary of the Invention
[0003] This invention aims to address the shortcomings of existing technologies by providing a method and system for UAV swarm conflict decoupling and trajectory reconstruction based on energy-constrained graph neural networks. Its core objective is to uniformly encode and learn the topological relationships, airspace constraints, and energy consumption characteristics of UAV swarms through graph neural networks, achieving safe, efficient, and energy-saving multi-UAV collaborative trajectory planning, ultimately improving the overall task completion rate and reliability of the swarm system. The specific technical solution is as follows: A method for decoupling conflicts and reconstructing flight paths in a drone swarm includes the following steps: Obtain UAV nodes, task nodes, airspace nodes, task-related edges, airspace conflict edges, and energy consumption constraint edges. Construct a UAV-task-airspace heterogeneous relationship graph based on UAV nodes, task nodes, airspace nodes, task-related edges, airspace conflict edges, and energy consumption constraint edges. Based on graph neural networks, neighbor sampling and aggregation are performed on nodes to obtain preliminary updated features. When aggregating neighbor information, an attention mechanism is introduced to assign different importance weights to different neighbors, obtain attention coefficients, and update node features based on attention coefficients. Multi-layer stacking and global information fusion are performed to aggregate the final features of all nodes and obtain a global feature vector; Obtain the control command vector of each drone node and the conflict risk between any two drones, determine whether the conflict risk exceeds the risk threshold, and if so, trigger the decoupling mechanism.
[0004] The proposed UAV swarm conflict decoupling and trajectory reconstruction method constructs a heterogeneous relationship graph of UAV-mission-airspace and uses a graph neural network with fused energy consumption constraints for distributed decision-making and optimization. This enables conflict decoupling and trajectory collaborative planning for multi-UAV swarm missions in complex low-altitude environments, thereby achieving safe, efficient, and collaborative flight of UAV swarms.
[0005] Preferably, the UAV swarm conflict decoupling and trajectory reconstruction method further includes the following steps: Obtain the airspace conflict edge weights and the distance between the UAV and the obstacle, and obtain the safety conflict cost based on the airspace conflict edge weights and the distance between the UAV and the obstacle; Obtain the edge weights of the energy consumption constraints and the power consumption of the motor, and obtain the energy consumption constraint cost based on the edge weights of the energy consumption constraints and the power consumption of the motor. Obtain the task-related edge weights and drone status, and based on the task-related edge weights and drone status, obtain the task coordination cost used to ensure that drones for high-priority tasks reach their targets; The trajectory reconstruction optimization strategy is implemented based on the costs of security conflicts, energy consumption constraints, and mission coordination.
[0006] Preferably, the specific method for implementing the trajectory reconstruction optimization strategy includes the following steps: Construct dynamic constraints, control quantity constraints, and electrical quantity constraints; Construct a total cost function based on security conflict cost, energy consumption constraint cost, and task coordination cost; Based on the constructed dynamic constraints, control quantity constraints, and electrical quantity constraints, the total cost function is minimized to obtain the optimal control sequence.
[0007] Preferably, the specific method for implementing the trajectory reconstruction optimization strategy further includes the following steps: When the risk of conflict is too high, the battery is severely depleted, or the environment changes abruptly, a global trajectory reconstruction is triggered.
[0008] Preferably, the specific method for implementing the trajectory reconstruction optimization strategy further includes the following steps: after triggering reconstruction, only the affected subset of UAVs is replanned, rather than the entire cluster.
[0009] Preferably, the UAV swarm conflict decoupling and trajectory reconstruction method further includes the following steps: Determine whether the decoupling mechanism has been triggered. If so, use the drone as a node and a similarity matrix jointly constructed based on task relevance and airspace conflict. Divide the cluster into multiple subgroups based on the similarity matrix and proceed to the next step. Otherwise, maintain the status quo. Based on model predictive control, a corresponding primary objective is assigned to each subgroup, and the cluster center reference track is calculated, where the primary objective is to minimize the local cost of the subgroup. An artificial potential field is introduced for online fine-tuning between coarse-grained track points generated by model predictive control to handle sudden obstacles; Once the risk of conflict between subswarms is resolved or the mission phase is completed, the merging and recovery of the drone subswarms is triggered.
[0010] Preferably, the specific method for determining whether the decoupling mechanism has been triggered includes the following steps: The system detects whether the energy status of drones within the cluster is unbalanced, whether the current task modularity meets the standard, and whether the conflict risk exceeds the risk threshold. If any of these conditions are met, the decoupling mechanism is triggered; otherwise, the status quo is maintained.
[0011] A UAV swarm conflict decoupling and trajectory reconstruction system, used to implement the aforementioned UAV swarm conflict decoupling and trajectory reconstruction method, includes: The heterogeneous relationship graph construction module is used to obtain UAV nodes, task nodes, airspace nodes, task-related edges, airspace conflict edges, and energy consumption constraint edges, and construct a UAV-task-airspace heterogeneous relationship graph based on UAV nodes, task nodes, airspace nodes, task-related edges, airspace conflict edges, and energy consumption constraint edges. The node feature update module is used to sample and aggregate the neighbors of nodes based on graph neural networks to obtain preliminary updated features. When aggregating neighbor information, an attention mechanism is introduced to assign different importance weights to different neighbors, obtain attention coefficients, and update node features based on attention coefficients. The feature vector acquisition module is used to perform multi-layer stacking and global information fusion to aggregate the final features of all nodes and obtain the global feature vector. The decoupling module is used to obtain the control command vector of each UAV node and the conflict risk between any two UAVs, determine whether the conflict risk exceeds the risk threshold, and if so, trigger the decoupling mechanism.
[0012] Preferably, the UAV swarm conflict decoupling and trajectory reconstruction system further includes: The security conflict cost acquisition module is used to obtain the airspace conflict edge weights and the distance between the UAV and the obstacle, and to obtain the security conflict cost based on the airspace conflict edge weights and the distance between the UAV and the obstacle. The energy consumption constraint cost acquisition module is used to obtain the energy consumption constraint edge weights and motor power consumption, and to obtain the energy consumption constraint cost based on the energy consumption constraint edge weights and motor power consumption. The task coordination cost acquisition module is used to obtain the task association edge weights and UAV status, and to obtain the task coordination cost to ensure that the UAVs of high-priority tasks reach the target based on the task association edge weights and UAV status. The trajectory reconstruction optimization strategy module is used to implement trajectory reconstruction optimization strategies based on safety conflict costs, energy consumption constraints, and mission coordination costs.
[0013] Preferably, the trajectory reconstruction optimization strategy module includes: The constraint construction unit is used to construct dynamic constraints, control constraints, and electrical constraints. The total cost function construction unit is used to construct the total cost function based on the security conflict cost, energy consumption constraint cost, and task coordination cost. The control strategy optimization unit is used to minimize the total cost function and obtain the optimal control sequence based on the constructed dynamic constraints, control quantity constraints, and electrical quantity constraints. Attached Figure Description
[0014] The invention will be further understood from the following description taken in conjunction with the accompanying drawings. The components in the drawings are not necessarily drawn to scale, but rather the emphasis is on illustrating the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.
[0015] Figure 1 This is a schematic diagram of the overall process of a method for decoupling and reconstructing flight paths in a drone swarm according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the hardware embedded system of the UAV airborne perception and control system in one embodiment of the present invention; Figure 3 This is a schematic diagram of the graph neural network optimization process in one embodiment of the present invention; Figure 4 This is a flowchart illustrating a method for decoupling and reconstructing flight paths in a drone swarm, as described in another embodiment of the present invention. Figure 1 ; Figure 5 This is a schematic diagram of the overall process of the trajectory reconstruction mechanism in one embodiment of the present invention; Figure 6 This is a flowchart illustrating a specific method for implementing a trajectory reconstruction optimization strategy in one embodiment of the present invention. Figure 7 This is a flowchart illustrating a method for decoupling and reconstructing drone swarm conflicts in one embodiment and another embodiment of the present invention. Figure 2 ; Figure 8This is an overall process architecture diagram of a UAV swarm conflict decoupling and trajectory reconstruction system according to an embodiment of the present invention. Detailed Implementation
[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to its embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.
[0017] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "upper," "lower," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.
[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The terms "and" and "or" as used herein include any and all combinations of one or more of the associated listed items.
[0019] The core objective of this invention is to uniformly encode and learn the topological relationships, airspace constraints, and energy consumption characteristics of UAV swarms through Graph Neural Network (GNN), thereby achieving safe, efficient, and energy-saving multi-UAV collaborative trajectory planning and ultimately improving the overall task completion rate and reliability of the swarm system.
[0020] To achieve the above objectives, this invention provides a method for UAV swarm conflict decoupling and trajectory reconstruction based on an energy-constrained graph neural network. Its core lies in constructing an integrated graph model of "airspace-UAV-mission" and utilizing a graph neural network incorporating energy constraints for distributed decision-making and optimization. Through dynamic decoupling, incremental updates, and a distributed computing architecture, this method effectively meets the real-time requirements of UAV swarm control in complex environments while ensuring optimization performance.
[0021] like Figure 1 As shown, the UAV swarm conflict decoupling and trajectory reconstruction method of the present invention includes the following steps: S1. Obtain UAV nodes, task nodes, airspace nodes, task-related edges, airspace conflict edges, and energy consumption constraint edges. Construct a UAV-task-airspace heterogeneous relationship graph based on UAV nodes, task nodes, airspace nodes, task-related edges, airspace conflict edges, and energy consumption constraint edges.
[0022] Specifically, UAV Nodes represent each physical UAV entity. Each node is initialized with a feature vector containing: [longitude, latitude, altitude, speed vector, remaining battery power, and task payload]. Task Nodes represent logical tasks to be executed. Each node is initialized with a feature vector containing: [task type ID, target coordinates, deadline timestamp, and priority weight]. Space Nodes represent key structures or areas in the airspace. Each node is initialized with a feature vector containing: [channel width, wind direction interference coefficient, passage cost, and safe distance threshold].
[0023] Task-related edge (drone) The task can be understood as connecting the drone with the task assigned to it, and its weight is determined by the task urgency and the drone's status, as shown in Equation 1:
[0024] The above formula It represents the "association strength / assignment tendency / urgency weight between the drone and a certain task", and is used as the edge weight of directed edges (drone → task) when constructing a graph, or as the edge weight factor in GNN message passing. This reflects the urgency of the task. This represents the remaining energy state of the drone. Here, B is the time scaling factor, used to adjust the sensitivity of the sigmoid function.
[0025] Airspace conflict border (drones) (Airspace) is an undirected edge connecting the UAV to the airspace node it is currently in or about to enter. The weights represent the conflict risk, as shown in Equation 2:
[0026] This represents the intensity / weight of the conflict risk between the drone and the airspace node it is currently in or about to enter, where k is the control weight function for the distance difference ( Sensitivity coefficient.
[0027] Energy Constraint Edge (UAV) (Unmanned aerial vehicles) can be understood as connecting undirected edges that may have mutual influence in terms of energy or trajectory. The weights are determined by state similarity, as shown in Equation 3:
[0028] in, and These represent the remaining energy levels of drone i and drone j at the current moment, respectively, reflecting the differences in endurance and energy consumption. and This indicates the current motion state of the drone and is used to measure the difference in motion between the two drones.
[0029] S2, based on a graph neural network, performs neighbor sampling and aggregation on nodes to obtain preliminary updated features. When aggregating neighbor information, an attention mechanism is introduced to assign different importance weights to different neighbors, obtain attention coefficients, and update node features based on attention coefficients.
[0030] This step involves distributed decision-making and feature learning based on graph neural networks (GNNs). It performs message passing and aggregation within the graph neural network and employs an architecture that combines GraphSAGE (Graph Sample and Aggregate) and GAT (Graph Attention Network) for feature learning.
[0031] Specifically, step S2 includes the following: Step 1: GraphSAGE Neighbor Sampling and Aggregation. First, a fixed number of neighbors (e.g., k=5) are sampled for each node to handle dynamic network structures. Then, mean aggregation or max pooling aggregation is performed on the neighbor features of each node to generate its neighborhood summary vector h_N(i). Finally, the node's own features are... Concatenate with the neighborhood summary h_N(i) and pass through a fully connected layer (such as...) This is combined to generate the initial update features for the node.
[0032] Step 2: GAT Attention Mechanism Integration. When aggregating neighbor information, an attention mechanism is introduced to assign different importance weights to different neighbors. The attention coefficient is calculated as shown in Equation 4:
[0033] in Let be the edge feature, representing the "importance score" that node i gives to neighbor node j when aggregating neighbor information. Let W be the attention vector, and let W be the linear transformation matrix. We perform a linear projection on the node features, mapping the original node features to the attention space / latent space, which facilitates comparison and aggregation. and This represents the mapping of the features of node i and node j to the same latent space, and is used to calculate attention relevance.
[0034] Step 3: Normalize the attention coefficient using softmax, as shown in Equation 5:
[0035] This represents the relative importance weight of neighbor node j when node i aggregates its neighbor node information. This represents the original attention score of node i to its neighbor node j, reflecting the degree of correlation between the two under the constraints of the feature space and edge features.
[0036] Step 4: Update node features by weighted summation, see Equation 6:
[0037] This represents the new feature vector of node i after completing the aggregation and nonlinear transformation of neighbor information in the current GAT layer. This represents the representation after mapping the original features of neighbor node j to the latent space.
[0038] S3 performs multi-layer stacking and global information fusion to aggregate the final features of all nodes and obtain the global feature vector.
[0039] Here, by stacking multiple layers and fusing global information, it is ensured that the model not only focuses on local interactions, but also makes optimization decisions from a global perspective.
[0040] Specifically, step S3 includes: Step 1: Stack multiple GNN layers, repeating Step 2 multiple times (e.g., 2 or 3 layers), with the output of each layer serving as the input to the next. By stacking multiple layers, each node can receive information from multi-hop neighbors, thereby gaining a broader global perspective and understanding the overall operational status of the cluster.
[0041] Step 2: After passing through all GNN layers, the final features of all nodes are aggregated to generate a global feature vector representing the state of the entire graph / cluster.
[0042] S4: Obtain the control command vector of each drone node and the conflict risk between any two drones, determine whether the conflict risk exceeds the risk threshold, and if so, trigger the decoupling mechanism.
[0043] Specifically, the hardware embedded system framework diagram and graph neural network optimization process of the UAV airborne perception and control system are as follows: Figure 2 , Figure 3 As shown.
[0044] Figure 2 In this context, the Livox MID-360 LIDAR is a Livox LiDAR model that uses laser ranging to acquire 3D point clouds of the environment. The Jetson TX2 is an embedded high-performance computing platform from NVIDIA, used to process the LiDAR point cloud data and output GNN inference instructions. The STM32F767 is a 32-bit microcontroller (MCU) from STMicroelectronics, the core control unit of the entire system, responsible for receiving various data, executing instructions, and outputting control signals.
[0045] The MPU9250 IMU is a sensor that measures the device's attitude, acceleration, and angular velocity. It is responsible for acquiring attitude angles (the device's attitude information, such as pitch, roll, and heading) and transmitting them to the STM32. The DRV8305 driver chip is a motor driver chip from TI (Texas Instruments). It is responsible for receiving PWM signals from the STM32, driving the motor, and simultaneously feeding back three-phase current to the STM32. The T-MotorMN5208 is a brushless motor, which is the system's actuator. It is driven by the driver chip and transmits its own status to the STM32 via Hall effect feedback.
[0046] Step S4 mainly includes: Step 1, Node-level Output: Output a control command vector for each UAV node, such as [Δ velocity, Δ heading], and output a state prediction for each task node.
[0047] Step 2: Predict the probability or strength of the existence of a certain edge, which can be used for conflict prediction. For example, output the probability that there is a conflict between any two drones, and trigger a decoupling mechanism if the probability exceeds a threshold.
[0048] When the "edges" corresponding to two drones meet the following conditions, they can be judged as having a high risk of conflict: 1. The probability of the edge's existence exceeds a preset threshold (indicating that the two drones are likely to interact); 2. The strength of the edge exceeds a preset threshold (indicating that the interaction between the two drones is sufficiently close).
[0049] In summary, the proposed UAV swarm conflict decoupling and trajectory reconstruction method constructs a heterogeneous relationship graph between UAVs, missions, and airspace, and utilizes a graph neural network with fused energy consumption constraints for distributed decision-making and optimization. This enables conflict decoupling and collaborative trajectory planning for multi-UAV swarm missions in complex low-altitude environments, thereby achieving safe, efficient, and collaborative flight of UAV swarms.
[0050] In one embodiment, such as Figure 4 As shown, the UAV swarm conflict decoupling and trajectory reconstruction method further includes the following steps: S5 obtains the airspace conflict edge weights and the distance between the UAV and the obstacle, and obtains the safety conflict cost based on the airspace conflict edge weights and the distance between the UAV and the obstacle.
[0051] S6, obtain the edge weights of the energy consumption constraint and the power consumption of the motor, and obtain the energy consumption constraint cost based on the edge weights of the energy consumption constraint and the power consumption of the motor.
[0052] S7: Obtain the task-related edge weights and drone status, and obtain the task coordination cost to ensure that drones for high-priority tasks reach their targets based on the task-related edge weights and drone status.
[0053] S8 implements a trajectory reconstruction optimization strategy based on the costs of security conflicts, energy consumption constraints, and mission coordination.
[0054] Specifically, this embodiment provides a trajectory reconstruction optimization strategy design that integrates energy consumption constraints and airspace conflict conditions. This part is a downstream optimization layer of the graph neural network model, and its function is to map the abstract features output by the GNN into specific, safe, and energy-efficient UAV trajectory control commands. The core flowchart of this embodiment is as follows. Figure 5 As shown, a continuous optimization closed loop incorporating multi-objective constraints is included, specifically comprising the following points: Step 1: Based on the spatial conflict edge weights in GNN Calculate the security conflict cost term See equation 7:
[0055] Where: H is the prediction time domain length. Let be the set of all obstacle neighbor nodes of UAV i in the current airspace model. This is an indicator function that is activated when the distance between drone i and obstacle j is less than the safe distance.
[0056] Step 2: Energy-constrained edge weights in GNN The mathematical expression for the energy consumption constraint cost term calculated using the motor power consumption model is shown in Equation 8:
[0057] In the above formula, This represents a sub-term of the objective function that cumulatively values the energy consumption of swarm drones within the prediction time domain. It is used to suppress high-energy-consuming strategies and reflect energy constraints when optimizing / reconstructing trajectories. The output of the motor power consumption model for the i-th UAV represents the motor power consumption (or propulsion system rate) of UAV i at time i, which is calculated from state variables such as speed and rotational speed through the power consumption model. This is the energy constraint penalty coefficient, used to adjust the relative weights of the "power consumption item" and the "electricity constraint penalty item" in the total energy cost. It reflects the drone's current battery / energy status and is used to determine whether it is approaching the low battery threshold. The specified battery safety threshold is used. When the remaining energy is below this threshold, it is considered that there is a risk to the battery's endurance or that the battery is not feasible. In such cases, stronger penalties should be imposed or the return-to-home / exit-mission logic should be triggered. This is the fully charged battery level.
[0058] Step 3: Calculate the task coordination cost. This step ensures that drones performing high-priority tasks can reach their targets more accurately. See Equation 9:
[0059] In the formula, This represents the total cost of task collaboration; a smaller value indicates a higher degree of task completion. This is the core weighting factor. The larger the value, the more important the task of drone i is, and the greater the penalty for its trajectory deviation, which is given priority during optimization. It is the core cost item, which directly measures the difference between the current position of the drone and the target position, and is the direct basis for optimizing the trajectory tracking accuracy. This is a time constraint penalty function used to handle task time urgency. Typically, the penalty increases sharply after a timeout, forcing the optimizer to complete the planning before the deadline. H represents the time penalty coefficient. H is used to define the time step range for forward optimization of Model Predictive Control (MPC).
[0060] In one embodiment, such as Figure 6 As shown, the specific method for implementing the trajectory reconstruction optimization strategy includes the following steps: S81, construct dynamic constraints, control constraints, and electrical constraints.
[0061] S82 constructs the total cost function based on the security conflict cost, energy consumption constraint cost, and task coordination cost.
[0062] S83, based on the constructed dynamic constraints, control quantity constraints, and electrical quantity constraints, minimizes the total cost function to obtain the optimal control sequence.
[0063] Specifically, a total cost function can be constructed based on dynamic constraints, control quantity constraints, and energy constraints, taking into account safety conflict costs, energy consumption constraints, and task coordination costs. Then, rolling optimization is performed based on Model Predictive Control (MPC), using numerical optimization algorithms (such as Sequential Quadratic Programming (SQP) or the interior-point method) to solve the optimization problem in real time, minimizing the total cost function to obtain the optimal control sequence. And only the first control variable in the sequence is executed. At the next time step, based on the new state measurement / estimate, the above optimization is executed again in a rolling manner, forming a closed-loop feedback.
[0064] In one embodiment, the specific method for implementing the trajectory reconstruction optimization strategy further includes the following steps: triggering global trajectory reconstruction when the conflict risk is too high, the power is severely insufficient, or the environment changes abruptly.
[0065] The specific methods for implementing the trajectory reconstruction optimization strategy also include the following steps: after triggering reconstruction, only the affected subset of UAVs is replanned, rather than the entire cluster.
[0066] Specifically, dynamic triggering and incremental updates are implemented. Global trajectory reconstruction is triggered when any of the following conditions are met: excessively high conflict risk, severe power shortage, or a sudden change in the environment. After reconstruction is triggered, only the affected subset of drones (determined by high attention weight nodes in the graph neural network) is replanned, rather than the entire cluster, to reduce computational load.
[0067] In one embodiment, such as Figure 7 As shown, the UAV swarm conflict decoupling and trajectory reconstruction method further includes the following steps: S9. Determine whether the decoupling mechanism is triggered. If so, use the drone as a node and a similarity matrix jointly constructed by task relevance and airspace conflict degree. Divide the cluster into multiple subgroups based on the similarity matrix and proceed to the next step. Otherwise, maintain the status quo.
[0068] S10, based on model predictive control, assigns a corresponding primary objective to each subgroup and calculates the cluster center reference track, where the primary objective is to minimize the local cost of the subgroup.
[0069] S11 introduces an artificial potential field for online fine-tuning between the coarse-grained track points generated by model predictive control to handle sudden obstacles.
[0070] S12, when the risk of conflict between subgroups is resolved or the mission phase is completed, triggers the fusion and recovery of the drone subgroup.
[0071] Specifically, one of the initial purposes of dividing into subgroups is to avoid airspace / resource conflicts among drones within the cluster. However, during the operation of the subgroups, there may still be "conflict risks" between the subgroups due to airspace overlap and resource competition. Once this risk (such as clearing the airspace overlap area and eliminating resource conflicts) disappears, the subgroups no longer need to maintain an independent separation state.
[0072] Drone swarm missions are typically divided into phases (e.g., "inspecting area A" is one phase, and "searching and rescuing in area B" is the next phase). Once the current mission phase of a sub-swarm is completed, the independent execution status of the sub-swarm becomes meaningless and needs to be reorganized.
[0073] Therefore, in this embodiment, fusion refers to merging the previously separated subgroups back into a complete UAV cluster (or reorganizing the cluster according to new mission requirements), ending the "local independent control" state of the subgroups. Restoration refers to restoring the global collaborative control mode of the cluster (replacing the "local MPC optimization" mode of the subgroups), and re-planning the cluster's missions, tracks, and resource allocation.
[0074] The specific method for determining whether to trigger the decoupling mechanism includes the following steps: detecting whether the energy state of the drones in the cluster is unbalanced, whether the current task modularity meets the standard, and whether the conflict risk exceeds the risk threshold. If any of the conditions are met, the decoupling mechanism is triggered; otherwise, the status quo is maintained.
[0075] Specifically, drone energy is a core resource for mission execution. Energy imbalance detection refers to detecting significant differences in the remaining energy status of drones within the cluster (e.g., some drones are nearly depleted while others have sufficient power). When this difference in energy distribution exceeds a preset threshold, it is considered an energy imbalance. This imbalance slows down the overall efficiency of the cluster (drones with low power cannot keep up with those with high power), thus requiring decoupling (splitting the cluster into sub-clusters). Subsequently, adaptive tasks can be assigned to sub-clusters with different energy states (e.g., the low-power sub-cluster performs short-distance / low-energy tasks).
[0076] Task modularity describes whether the overall task of the current cluster can be broken down into multiple relatively independent and complete sub-task modules (for example, if the overall task is "area A inspection + area B search and rescue," these two sub-tasks are independent modules). Task modularity assessment detects the degree of modularity of the current task: if the task can be clearly broken down into multiple non-interfering and independently complete sub-task modules, it is determined that the task modularity meets the standard. At this point, decoupling (splitting the cluster into subgroups) is triggered, allowing different subgroups to be responsible for a sub-task module, thereby improving the parallel execution efficiency of the task.
[0077] This embodiment provides a mechanism for decoupling group behavior and optimizing local trajectories under multi-task objectives. The core of this mechanism is the construction of a hierarchical decision-making and control system. Based on global optimization using graph neural networks, it dynamically decouples a large cluster into highly cohesive, loosely coupled cooperative subgroups and performs distributed fine-tuning, such as... Figure 8 As shown, the specific points include the following: Step 1: Determine whether decoupling needs to be initiated based on dynamic decoupling trigger conditions, which are divided into conflict risk threshold judgment, energy state imbalance judgment, and task modularity judgment. If any condition is met, output the instruction "Execute Decoupling"; otherwise, output "Maintain Status quo".
[0078] Step 2, Global Layer - Decoupling of Group Behavior. Once triggered, this step is responsible for intelligently dividing the large cluster into multiple collaborative subgroups.
[0079] Step 3: Subgrouping based on Spectral Clustering: Using UAVs as nodes, a similarity matrix S is constructed using task relevance and airspace conflict. See Equation 10.
[0080]
[0081] This represents the similarity between UAV i and UAV j when performing spectral clustering subgroup division, and is used to construct the similarity matrix required for spectral clustering. Used to measure the degree of correlation between drones i and j at the mission level. This formula is used to measure the degree of airspace conflict between drones i and j. It ensures that drones with closely coordinated missions and lower airspace conflict are more likely to be grouped into the same subgroup.
[0082] Step 4: Assign a primary target to each sub-swarm (e.g., Sub-SwarmA is responsible for inspecting area A), and calculate its cluster center reference track. After sub-swarm division, each sub-swarm performs internal self-organizing optimization in parallel.
[0083] Step 5: Model Predictive Control (MPC) within the subgroup. The optimization objective is to minimize the local cost function of the subgroup, as shown in Equation 11:
[0084] The above formula gives the local optimization objective of MPC within a subgroup, where M is the subgroup size. and These represent the drone's position and the reference position, respectively. The first term is the squared tracking error, used to drive the drone to the reference trajectory / target point; the second term is the safety distance penalty term. The distance between the two aircraft. The minimum safe distance threshold, This is a penalty function used to quantify and penalize unsafe close-in behavior between drones. This is a tradeoff coefficient. By minimizing... While ensuring a safe distance within the subgroup, it improves the tracking accuracy of the subgroup reference trajectory and high-priority task targets.
[0085] Step 6: Between the coarse-grained track points generated by MPC, an artificial potential field is introduced for online fine-tuning to handle sudden obstacles.
[0086] In autonomous navigation systems, Model Predictive Control (MPC) is typically used to generate coarse-grained waypoints for a future timeframe. These waypoints are planned based on current environmental information and kinematic models, providing global path guidance. However, MPC planning can be lengthy and relies on pre-defined models. Therefore, during path execution, if unexpected obstacles are encountered, paths generated solely by MPC may not respond in time, posing a collision risk.
[0087] To address these issues, the Artificial Potential Field (APF) method can be introduced for online fine-tuning. APF models the environment as a potential field: an gravitational potential field points towards the target point, attracting the UAV to move towards it; a repulsive potential field is generated by obstacles, causing the UAV to move away from potential collision sources. Between the coarse-grained points of the MPC trajectory, APF calculates the local potential field in real time and generates a smooth, fine-tuned trajectory through optimization methods such as gradient descent, dynamically avoiding sudden obstacles.
[0088] Specifically, the workflow of this method includes: 1. State prediction and potential field calculation: The system monitors the UAV state and environment in real time, uses the UAV motion model to predict the future state, and calculates the artificial potential field that incorporates collision avoidance rules (such as COLREGs).
[0089] 2. Online fine-tuning and feedback control: Between the waypoints generated by MPC, APF adjusts its heading according to the current potential field to generate a locally optimal path; at the same time, it continuously monitors environmental changes through feedback control and dynamically corrects the track to cope with emergencies.
[0090] Here, the introduction of APF enhances the adaptability and robustness of the planning system, especially in complex encounter scenarios (such as multiple boats meeting), enabling rapid response to obstacles. Local minima traps (such as drones getting stuck due to the combined force of obstacles) can be optimized by improving the potential field function or introducing virtual target points.
[0091] This method combines the global planning capability of MPC with the local real-time response capability of APF, enabling fine-grained adjustment of the trajectory and enhancing the safety of autonomous navigation.
[0092] Step 7: Merging and Restoration. Merging conditions are triggered when the risk of conflict between subgroups is resolved or the task phase is completed.
[0093] By constructing a highly adaptive and collaborative distributed network composed of heterogeneous nodes (computing units), centralized intelligence is decomposed into generalizable edge intelligence, and its deployment architecture is as follows: Figure 3As shown, the system decomposes complex global computing tasks to different nodes through a three-level collaboration of "cloud-edge-device", and finally achieves large-scale deployment through a lightweight model and efficient communication protocol.
[0094] In general, the general process of this invention can be referred to Figure 1 as well as Figure 4 .
[0095] The present invention also provides a UAV swarm conflict decoupling and trajectory reconstruction system for implementing the aforementioned UAV swarm conflict decoupling and trajectory reconstruction method, which includes a heterogeneous relationship graph construction module, a node feature update module, a feature vector acquisition module, and a decoupling module.
[0096] The heterogeneous relationship graph construction module is used to obtain UAV nodes, task nodes, airspace nodes, task-related edges, airspace conflict edges, and energy consumption constraint edges. Based on the UAV nodes, task nodes, airspace nodes, task-related edges, airspace conflict edges, and energy consumption constraint edges, it constructs a UAV-task-airspace heterogeneous relationship graph. The node feature update module is used to sample and aggregate the neighbors of nodes based on graph neural networks to obtain preliminary updated features. When aggregating neighbor information, an attention mechanism is introduced to assign different importance weights to different neighbors, obtain attention coefficients, and update node features based on attention coefficients.
[0097] The feature vector acquisition module is used to perform multi-layer stacking and global information fusion to aggregate the final features of all nodes and obtain global feature vectors; the decoupling module is used to obtain the control command vector of each UAV node and the conflict risk between any two UAVs, and determine whether the conflict risk exceeds the risk threshold. If so, the decoupling mechanism is triggered.
[0098] The UAV swarm conflict decoupling and trajectory reconstruction system is based on a heterogeneous relationship graph construction module, a node feature update module, a feature vector acquisition module, and a decoupling module. By constructing a heterogeneous relationship graph of UAV-mission-airspace and using a graph neural network with fused energy consumption constraints for distributed decision-making and optimization, it can achieve conflict decoupling and trajectory collaborative planning for multi-UAV swarm missions in complex low-altitude environments, thereby enabling safe, efficient, and collaborative flight of UAV swarms.
[0099] As a preferred technical solution, the UAV swarm conflict decoupling and trajectory reconstruction system further includes a security conflict cost acquisition module, an energy consumption constraint cost acquisition module, a task coordination cost acquisition module, and a trajectory reconstruction optimization strategy module.
[0100] The safety conflict cost acquisition module is used to obtain the airspace conflict edge weights and the distance between the UAV and the obstacle, and obtains the safety conflict cost based on the airspace conflict edge weights and the distance between the UAV and the obstacle; the energy consumption constraint cost acquisition module is used to obtain the energy consumption constraint edge weights and the motor power consumption, and obtains the energy consumption constraint cost based on the energy consumption constraint edge weights and the motor power consumption.
[0101] The task coordination cost acquisition module is used to obtain the task association edge weights and UAV status, and to obtain the task coordination cost to ensure that UAVs for high-priority tasks reach their targets based on the task association edge weights and UAV status; the trajectory reconstruction optimization strategy module is used to implement the trajectory reconstruction optimization strategy based on the security conflict cost, energy consumption constraint cost and task coordination cost.
[0102] The trajectory reconstruction optimization strategy module includes a constraint construction unit, a total cost function construction unit, and a control strategy optimization unit.
[0103] The constraint construction unit is used to construct dynamic constraints, control quantity constraints, and energy constraints; the total cost function construction unit is used to construct the total cost function based on the safety conflict cost, energy consumption constraint cost, and task coordination cost; the control strategy optimization unit is used to minimize the total cost function and obtain the optimal control sequence based on the constructed dynamic constraints, control quantity constraints, and energy constraints.
[0104] Here, dynamic constraints, control constraints, and electrical constraints are constructed using a constraint construction unit, and a total cost function is constructed using a total cost function construction unit. Finally, the total cost function is minimized using the constructed dynamic constraints, control constraints, and electrical constraints to obtain the optimal control sequence. The optimization can be re-executed based on new state measurements / estimates to form a closed-loop feedback.
[0105] In summary, the core of this invention lies in constructing an integrated graph model of "airspace-UAV-mission" and using a graph neural network with fused energy consumption constraints for distributed decision-making and optimization. Through dynamic decoupling, incremental updates, and a distributed computing architecture, it can effectively meet the real-time requirements of UAV swarm control in complex environments while ensuring optimized performance.
[0106] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0107] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.
Claims
1. A method for decoupling conflicts and reconstructing flight paths in a drone swarm, characterized in that, The method for decoupling and reconstructing drone swarm conflicts includes the following steps: Obtain UAV nodes, task nodes, airspace nodes, task-related edges, airspace conflict edges, and energy consumption constraint edges. Construct a UAV-task-airspace heterogeneous relationship graph based on UAV nodes, task nodes, airspace nodes, task-related edges, airspace conflict edges, and energy consumption constraint edges. Based on graph neural networks, neighbor sampling and aggregation are performed on nodes to obtain preliminary updated features. When aggregating neighbor information, an attention mechanism is introduced to assign different importance weights to different neighbors, obtain attention coefficients, and update node features based on attention coefficients. Multi-layer stacking and global information fusion are performed to aggregate the final features of all nodes and obtain a global feature vector; Obtain the control command vector of each drone node and the conflict risk between any two drones, determine whether the conflict risk exceeds the risk threshold, and if so, trigger the decoupling mechanism.
2. The method for decoupling and reconstructing flight paths of a UAV swarm as described in claim 1, characterized in that, The UAV swarm conflict decoupling and trajectory reconstruction method also includes the following steps: Obtain the airspace conflict edge weights and the distance between the UAV and the obstacle, and obtain the safety conflict cost based on the airspace conflict edge weights and the distance between the UAV and the obstacle; Obtain the edge weights of the energy consumption constraints and the power consumption of the motor, and obtain the energy consumption constraint cost based on the edge weights of the energy consumption constraints and the power consumption of the motor. Obtain the task-related edge weights and drone status, and based on the task-related edge weights and drone status, obtain the task coordination cost used to ensure that drones for high-priority tasks reach their targets; The trajectory reconstruction optimization strategy is implemented based on the costs of security conflicts, energy consumption constraints, and mission coordination.
3. The method for decoupling and reconstructing flight paths of a UAV swarm as described in claim 2, characterized in that, The specific methods for implementing the trajectory reconstruction optimization strategy include the following steps: Construct dynamic constraints, control quantity constraints, and electrical quantity constraints; Construct a total cost function based on security conflict cost, energy consumption constraint cost, and task coordination cost; Based on the constructed dynamic constraints, control quantity constraints, and electrical quantity constraints, the total cost function is minimized to obtain the optimal control sequence.
4. The method for decoupling and reconstructing flight paths of a UAV swarm as described in claim 3, characterized in that, The specific methods for implementing the trajectory reconstruction optimization strategy also include the following steps: When the risk of conflict is too high, the battery is severely depleted, or the environment changes abruptly, a global trajectory reconstruction is triggered.
5. The method for decoupling and reconstructing flight paths of a UAV swarm as described in claim 4, characterized in that, The specific methods for implementing the trajectory reconstruction optimization strategy also include the following steps: after triggering reconstruction, only the affected subset of UAVs is replanned, rather than the entire cluster.
6. The method for decoupling and reconstructing flight paths of a UAV swarm as described in claim 5, characterized in that, The UAV swarm conflict decoupling and trajectory reconstruction method also includes the following steps: Determine whether the decoupling mechanism has been triggered. If so, use the drone as a node and a similarity matrix jointly constructed based on task relevance and airspace conflict. Divide the cluster into multiple subgroups based on the similarity matrix and proceed to the next step. Otherwise, maintain the status quo. Based on model predictive control, a corresponding primary objective is assigned to each subgroup, and the cluster center reference track is calculated, where the primary objective is to minimize the local cost of the subgroup. An artificial potential field is introduced for online fine-tuning between coarse-grained track points generated by model predictive control to handle sudden obstacles; Once the risk of conflict between subswarms is resolved or the mission phase is completed, the merging and recovery of the drone subswarms is triggered.
7. The method for decoupling and reconstructing flight paths of a UAV swarm as described in claim 6, characterized in that, The specific methods for determining whether the decoupling mechanism has been triggered include the following steps: The system detects whether the energy status of drones within the cluster is unbalanced, whether the current task modularity meets the standard, and whether the conflict risk exceeds the risk threshold. If any of these conditions are met, the decoupling mechanism is triggered; otherwise, the status quo is maintained.
8. A UAV swarm conflict decoupling and trajectory reconstruction system, used to implement the UAV swarm conflict decoupling and trajectory reconstruction method as described in any one of claims 1-7, characterized in that, The UAV swarm conflict decoupling and trajectory reconstruction system includes: The heterogeneous relationship graph construction module is used to obtain UAV nodes, task nodes, airspace nodes, task-related edges, airspace conflict edges, and energy consumption constraint edges, and construct a UAV-task-airspace heterogeneous relationship graph based on UAV nodes, task nodes, airspace nodes, task-related edges, airspace conflict edges, and energy consumption constraint edges. The node feature update module is used to sample and aggregate the neighbors of nodes based on graph neural networks to obtain preliminary updated features. When aggregating neighbor information, an attention mechanism is introduced to assign different importance weights to different neighbors, obtain attention coefficients, and update node features based on attention coefficients. The feature vector acquisition module is used to perform multi-layer stacking and global information fusion to aggregate the final features of all nodes and obtain the global feature vector. The decoupling module is used to obtain the control command vector of each UAV node and the conflict risk between any two UAVs, determine whether the conflict risk exceeds the risk threshold, and if so, trigger the decoupling mechanism.
9. The UAV swarm conflict decoupling and trajectory reconstruction system as described in claim 8, characterized in that, The UAV swarm conflict decoupling and trajectory reconstruction system also includes: The security conflict cost acquisition module is used to obtain the airspace conflict edge weights and the distance between the UAV and the obstacle, and to obtain the security conflict cost based on the airspace conflict edge weights and the distance between the UAV and the obstacle. The energy consumption constraint cost acquisition module is used to obtain the energy consumption constraint edge weights and motor power consumption, and to obtain the energy consumption constraint cost based on the energy consumption constraint edge weights and motor power consumption. The task coordination cost acquisition module is used to obtain the task association edge weights and UAV status, and to obtain the task coordination cost to ensure that the UAVs of high-priority tasks reach the target based on the task association edge weights and UAV status. The trajectory reconstruction optimization strategy module is used to implement trajectory reconstruction optimization strategies based on safety conflict costs, energy consumption constraint costs, and mission coordination costs.
10. The UAV swarm conflict decoupling and trajectory reconstruction system as described in claim 9, characterized in that, The trajectory reconstruction optimization strategy module includes: The constraint construction unit is used to construct dynamic constraints, control constraints, and electrical constraints. The total cost function construction unit is used to construct the total cost function based on the security conflict cost, energy consumption constraint cost, and task coordination cost. The control strategy optimization unit is used to minimize the total cost function and obtain the optimal control sequence based on the constructed dynamic constraints, control quantity constraints, and electrical quantity constraints.