Unmanned aerial vehicle ad hoc network anti-collision route planning method based on graph neural network

Abstract

The invention relates to the technical field of unmanned aerial vehicle anti-collision path planning, in particular to an unmanned aerial vehicle ad hoc network anti-collision path planning method based on a graph neural network, and the method comprises the steps: obtaining the topographic data of a task airspace, recording the initial positions, target positions and motion constraint conditions of all unmanned aerial vehicles, and constructing unmanned aerial vehicle graph structure data; according to the unmanned aerial vehicle graph structure data, taking each unmanned aerial vehicle as a unit node, and representing a potential collision relationship between two adjacent unmanned aerial vehicles by a side between the nodes; initializing features of the nodes and the edges, and completing training of an unmanned aerial vehicle graph neural network model based on the input features of the nodes and the edges as input of the preset unmanned aerial vehicle graph neural network model; and performing collision risk prediction based on the unmanned aerial vehicle graph neural network model, and generating and outputting an optimal flight path. According to the invention, through the graph neural network and multi-objective optimization, intelligent cooperative flight of the unmanned aerial vehicle group is realized, the collision risk and energy consumption are significantly reduced, and the network connectivity and task efficiency are ensured at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) anti-collision path planning, and in particular to a UAV self-organizing network anti-collision route planning method based on a graph neural network. Background Art

[0002] With the rapid development of drone technology, multi-UAV collaborative missions are becoming increasingly common in areas such as military reconnaissance, emergency rescue, and logistics. During these missions, UAVs must maintain communication connectivity to enable information sharing and collaborative decision-making, while also avoiding collisions with other UAVs. Path planning for a swarm of UAVs involves a trade-off between maintaining communication connectivity, avoiding collision risks, and optimizing energy consumption.

[0003] Traditional UAV path planning methods rely primarily on manually pre-defined rules and heuristic algorithms, which have significant limitations in cooperative UAV flight. For one thing, centralized control relies on frequent updates of global information, which not only incurs high communication overhead but also makes it difficult to respond in real time in dynamic environments.

[0004] On the other hand, rule-based distributed strategies lack the ability to model complex interactive relationships, which can easily lead to local optimality or intensified conflicts.

[0005] Furthermore, traditional approaches often separate communication, collision avoidance, and energy optimization, making it difficult to achieve coordinated optimization of multiple objectives. This results in low computational efficiency and often fails to meet the requirements of high precision and low latency in practical applications, particularly in complex terrain or dense drone swarms. Summary of the Invention

[0006] To address the above technical issues, a graph neural network-based anti-collision route planning method for UAV ad hoc networks is proposed. The method includes: based on the UAV graph structure data, each UAV is treated as a unit node, and the edges between nodes represent the potential collision relationship between two adjacent UAVs;

[0007] Initialize node and edge features, including drone flight status features and environmental features. These input node and edge features are used as input for the pre-set drone graph neural network model. The model uses a loss function to measure communication connectivity, collision risk, and energy consumption. The model's weights are iteratively updated to complete training.

[0008] Based on the trained UAV graph neural network model, the collision risk between each UAV and its neighboring nodes is predicted, the optimal flight path is generated and output, and the UAV flight is executed until the flight mission is completed.

[0009] As a preferred solution of the UAV self-organizing network anti-collision route planning method based on graph neural network of the present invention, the motion constraint conditions of the UAV include:

[0010] According to the UAV model and flight mission requirements, set the flight speed upper limit constraint, acceleration upper limit constraint, turning radius constraint, and climb / descent rate constraint as the UAV kinematic constraint parameters;

[0011] The flight path planning of the UAV is limited by the altitude limit of the mission airspace, the minimum safe distance between obstacles, the boundaries of the no-fly zone and temporary restrictions;

[0012] The UAV's motion constraint parameters are mathematically converted, and collision risk penalty weights are introduced to form the UAV's motion constraint conditions. Based on the real-time collected UAV operation data, it is determined whether the constraint correction should be triggered.

[0013] As a preferred solution of the UAV self-organizing network anti-collision route planning method based on graph neural network of the present invention, the construction of UAV graph structure data includes:

[0014] Each drone is represented as a node, and each node is bound to the drone's current position, kinematic data, remaining battery power, and airspace environment data;

[0015] Create edges based on the relative positions and communication connectivity between drones in the mission airspace. If the distance between any two drones is less than the preset communication radius, an edge representing a potential collision relationship is established.

[0016] The UAV kinematic data and airspace environment data are written into the feature vectors of nodes and edges respectively to form the UAV graph structure data.

[0017] As a preferred solution of the UAV self-organizing network anti-collision route planning method based on graph neural network of the present invention, the UAV graph structure data includes:

[0018] During the flight of the UAV, the feature vectors of nodes and edges are periodically updated according to the motion constraints of the UAV.

[0019] If a new drone joins or an existing drone exits, nodes and edges are added or deleted to complete the dynamic adjustment of the drone graph structure data.

[0020] As a preferred solution of the UAV self-organizing network anti-collision route planning method based on graph neural network of the present invention, the UAV graph neural network model includes:

[0021] It consists of a graph attention network layer based on drone graph structure data, a time series LSTM structure, and a collision risk and flight path decision output layer.

[0022] As a preferred solution of the UAV self-organizing network anti-collision route planning method based on graph neural network of the present invention, the graph attention network layer includes:

[0023] Calculate the attention coefficient of each node in the drone graph structure data relative to all its neighboring nodes;

[0024] Multiply each group of attention coefficients by the feature vectors of the corresponding neighboring nodes element by element and add them up to get the multi-head attention aggregation result;

[0025] Splice the multi-head aggregation results to obtain the hidden representation of the node in the current layer;

[0026] The node hidden representation is passed as input to the next layer of GAT;

[0027] After multi-layer aggregation, the final hidden representation of the node is input into the time series LSTM structure.

[0028] As a preferred solution of the UAV self-organizing network anti-collision route planning method based on graph neural network described in the present invention, the time series LSTM structure includes:

[0029] Arrange the final hidden representation of the node according to the time step to serve as the input at the current moment;

[0030] The final hidden representation of the node is input into the LSTM unit together with the LSTM hidden state obtained in the previous time step. It is updated through the input gate, forget gate, and output gate of the LSTM unit to obtain a new LSTM hidden state, which represents the temporal characteristics of the node at this moment.

[0031] For each node, after traversing all time steps, the hidden state representation of the last output of the LSTM unit is represented as the final temporal representation of each node.

[0032] As a preferred solution of the UAV self-organizing network anti-collision route planning method based on graph neural network of the present invention, the collision risk and flight path decision output layer includes:

[0033] For the final time series representation of each node, the collision risk of the drone and the connectivity of adjacent drones are output through the fully connected layer;

[0034] If the UAV collision risk exceeds the set threshold or the connectivity of adjacent UAVs is lower than the set value, the flight path adjustment module is triggered to generate a new UAV flight path;

[0035] Deliver new drone flight paths to drones, perform drone emergency avoidance, or change drone flight paths.

[0036] As a preferred solution of the UAV self-organizing network collision avoidance route planning method based on graph neural network of the present invention, wherein: the communication connectivity, collision risk and energy consumption indicators are measured by loss function, and the weight of the UAV graph neural network model is iteratively updated, including:

[0037] Construct a connectivity matrix based on the nodes and edges in the current drone graph structure data;

[0038] Calculate the ratio of the actual number of available links to the theoretical maximum number of links in the connectivity matrix to obtain the communication connectivity evaluation value;

[0039] Based on the distance, relative speed and predicted flight path between drones, the probability of collision between any two drones at time t is calculated, which is represented as collision risk.

[0040] Establish a UAV energy consumption model and calculate the energy consumption at time t based on the UAV's flight speed, acceleration, attitude adjustment, and flight path turns;

[0041] Accumulate energy consumption at each moment to obtain the overall energy consumption index;

[0042] The communication connectivity evaluation value, collision risk and overall energy consumption index are defined as corresponding loss terms respectively, and each loss term is integrated into the total loss function through weighted linear combination.

[0043] As a preferred solution of the UAV self-organizing network anti-collision route planning method based on graph neural network of the present invention, wherein: the probability of collision between any two UAVs at time t is solved, including:

[0044] For each drone, based on the drone’s current position and operating data, at each discrete moment, the spatial distance between the two drones is calculated and compared with the preset safety distance;

[0045] If the distance between two drones is less than the preset safety distance, it is marked as a risky moment, and the relative speed and direction of the drones at that moment are recorded;

[0046] The marking results of all risk moments are summarized within the prediction time window to form a comprehensive degree metric that characterizes the collision probability of two drones, and the comprehensive degree metric is converted into collision probability based on empirical rules.

[0047] Beneficial effects of the present invention: The present invention significantly improves the intelligence level of UAV collaborative flight through the integration of graph neural networks and time series models. The introduction of graph structure data enables the model to efficiently characterize the spatial topology and dynamic interaction relationship of the UAV swarm, providing accurate predictions for real-time path planning and anti-collision control. The design of the multi-objective loss function enables the model to achieve a flexible balance between communication connectivity, collision risk and energy consumption, meeting the multi-objective optimization requirements in complex mission scenarios. The present invention can not only effectively reduce collision risk and energy consumption, but also significantly improve the collaborative efficiency and task completion of the UAV swarm while ensuring network connectivity. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0049] Figure 1 An overall flow chart of a method for anti-collision route planning in a UAV ad hoc network based on a graph neural network, provided as an embodiment of the present invention.

[0050] Figure 2 A logical schematic diagram of constructing drone graph structure data for a method for anti-collision route planning in a drone self-organizing network based on a graph neural network provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0051] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0052] Example 1, with reference to Figure 1 , which is the first embodiment of the present invention, provides a method for anti-collision route planning in a UAV self-organizing network based on a graph neural network, comprising:

[0053] like Figure 1 As shown, the present invention provides one embodiment: a method for anti-collision route planning in a UAV self-organizing network based on a graph neural network, the method comprising the following steps:

[0054] Obtain terrain data for the mission airspace, record the initial positions, target positions, and motion constraints of all drones, and construct drone graph structure data;

[0055] Based on the drone graph structure data, each drone is regarded as a unit node, and the edges between nodes represent the potential collision relationship between two adjacent drones;

[0056] Initialize node and edge features, including drone flight status features and environmental features. These input node and edge features are used as input for the pre-set drone graph neural network model. The model uses a loss function to measure communication connectivity, collision risk, and energy consumption. The model's weights are iteratively updated to complete training.

[0057] Based on the trained UAV graph neural network model, the collision risk between each UAV and its neighboring nodes is predicted, the optimal flight path is generated and output, and the UAV flight is executed until the flight mission is completed.

[0058] In the specific application of this embodiment, in a coordinated UAV flight mission, terrain data for the mission airspace must first be acquired, including information such as terrain height, obstacle location, and weather conditions. Data collection is performed using remote sensing technology, a geographic information system (GIS), or sensors onboard UAVs to ensure accuracy and real-time performance. After data collection, data processing is performed to ensure uniform data format and extract features relevant to UAV flight, such as terrain undulation, obstacle distribution, wind speed, and direction. The initial position, target position, and motion constraints of all UAVs are recorded. The initial position is acquired using GPS or other positioning systems and stored in a data structure for subsequent calculation and analysis. The target position is set based on mission requirements and recorded in the system to ensure a clear flight target for each UAV. Furthermore, UAV motion constraints are defined, including maximum flight speed, minimum safe distance, and turning radius. This information is linked to the UAV's state information to ensure compliance with safety and performance standards during flight. A UAV graph structure is constructed. Each UAV is treated as a node in the graph, and node attributes include the UAV's state characteristics (such as position, speed, and heading). Based on the relative positions and potential collision risks between drones, edges are established between nodes. Edge attributes represent the collision relationships between adjacent drones. Edge weights can be dynamically adjusted based on factors such as the distance and relative speed between drones, more accurately reflecting potential collision risks. Node and edge features are initialized, including the drone's flight state characteristics (such as speed, acceleration, and heading) and environmental characteristics (such as the distance to surrounding obstacles and weather conditions). These features are normalized to ensure a relatively balanced influence in model training, preventing certain features from dominating the training due to their large numerical range. Initialized node and edge features are used as input to construct an input tensor for training the graph neural network. A multi-objective loss function is designed to measure communication connectivity, collision risk, and energy consumption. Model weights are iteratively updated using a backpropagation algorithm. Multiple rounds of training are performed using the training dataset until the model converges or reaches the preset performance indicators, completing the training. Based on the trained drone graph neural network model, the collision risk between each drone and its neighboring nodes is predicted in real time. By combining the drone's target position and motion constraints, an optimal flight path is generated and optimized to ensure safety and efficiency. The path generation algorithm should be able to adjust in real time to address unexpected environmental changes or mission requirements. The generated optimal flight path and related control instructions are sent to the UAV execution unit, which monitors the UAV's status and environmental changes in real time to ensure flight safety and adjust the path as needed. The UAV then flies along the planned path until the mission is completed, recording flight data for subsequent analysis and optimization. These steps provide strong data support and technical assurance for the safety and efficiency of UAV collaborative flight missions.

[0059] Example 2 is the second embodiment of the present invention. In the implementation of this embodiment, the motion constraints of the drone are specifically as follows:

[0060] According to the UAV model and flight mission requirements, set the flight speed upper limit constraint, acceleration upper limit constraint, turning radius constraint, and climb / descent rate constraint as the UAV kinematic constraint parameters;

[0061] The flight path planning of the UAV is limited by the altitude limit of the mission airspace, the minimum safe distance between obstacles, the boundaries of the no-fly zone and temporary restrictions;

[0062] The UAV's motion constraint parameters are mathematically converted, and collision risk penalty weights are introduced to form the UAV's motion constraint conditions. Based on the real-time collected UAV operation data, it is determined whether the constraint correction should be triggered.

[0063] In this embodiment, the upper limit constraint of the UAV’s flight speed is: V max =min(V model ,V task ), where V max is the upper limit of the UAV’s flight speed, V model is the maximum flight speed of the drone model, V task The maximum flight speed required by the mission. Acceleration limit constraint: A max =A model , where A max is the upper limit of the UAV’s acceleration, A model The maximum acceleration and turning radius constraints of the drone model are: Among them, R min is the minimum turning radius of the drone, V is the flight speed of the drone, g is the acceleration of gravity, and θ is the tilt angle when turning. Climb / descent rate constraint: R climb =min(R climb,model ,R climb,task ), where R climb R is the upper limit of the climb / descent rate of the UAV. climb,model is the maximum climb / descent rate of the drone model, R climb,task The setting of each constraint in this embodiment is based on the technical specifications of the UAV and the specific mission requirements, ensuring that the UAV can fly stably in complex environments and complete the predetermined mission.

[0064] Example 3, as Figure 2 As shown, this is the third embodiment of the present invention. In the specific application of this embodiment, the construction of drone graph structure data is specifically as follows:

[0065] Each drone is represented as a node, and each node is bound to the drone's current position, kinematic data, remaining battery power, and airspace environment data;

[0066] Create edges based on the relative positions and communication connectivity between drones in the mission airspace. If the distance between any two drones is less than the preset communication radius, an edge representing a potential collision relationship is established.

[0067] The UAV kinematic data and airspace environment data are written into the feature vectors of nodes and edges respectively to form the UAV graph structure data.

[0068] In implementation, this embodiment treats each drone as a node in the graph and obtains the drone's current position, records its real-time coordinates within the mission airspace, obtains the drone's kinematic data, including speed, acceleration, heading, and climb / descent rate, and records the drone's remaining battery life (typically expressed as a percentage, reflecting the drone's endurance). It also collects environmental data for the drone's mission airspace, including the location of surrounding obstacles, meteorological conditions (such as wind speed and direction), and airspace altitude restrictions. All of this feature data is integrated into a feature vector to form a drone node. Edges are created based on the relative positions and communication connectivity between drones in the mission airspace. For each pair of drones, the distance between them is calculated. If the distance between the two drones is less than a preset communication radius, an edge is established between the two nodes, indicating a potential collision between them. The length of each edge is recorded, along with the communication status. In this embodiment, any existing edge is marked as connected. The potential collision risk is assessed based on the relative speed and distance between the drones. All drone nodes and their feature data are integrated into a node set, and all established edges and their feature data are integrated into an edge set. Through the node set and edge set, a complete drone graph structure data is constructed.

[0069] Furthermore, the drone graph structure data also includes:

[0070] During the flight of the UAV, the feature vectors of nodes and edges are periodically updated according to the motion constraints of the UAV.

[0071] If a new drone joins or an existing drone exits, nodes and edges are added or deleted to complete the dynamic adjustment of the drone graph structure data.

[0072] When a new drone joins a mission, a new node is created for it and its initial feature data is bound. The distance between the new drone and existing drones is calculated. If the distance is less than the preset communication radius, an edge is established between the new drone and the corresponding existing drone to represent the potential collision relationship, and the edge features of the existing drone are updated. When an existing drone exits a mission, its node is first deleted from the node set, and all edges associated with it are deleted to ensure the integrity of the graph structure data. The edge features of the remaining drones are updated, and the communication status and collision risk between them are recalculated to reflect the changes in the drone network.

[0073] Example 4 is the fourth embodiment of the present invention. In the further implementation of this embodiment, the drone graph neural network model is specifically composed of a graph attention network layer based on drone graph structure data, a time series LSTM structure, and a collision risk and flight path decision output layer.

[0074] The graph attention network layer is specifically:

[0075] Calculate the attention coefficient of each node in the drone graph structure data relative to all its neighboring nodes;

[0076] Multiply each group of attention coefficients by the feature vectors of the corresponding neighboring nodes element by element and add them up to get the multi-head attention aggregation result;

[0077] Splice the multi-head aggregation results to obtain the hidden representation of the node in the current layer;

[0078] The node hidden representation is passed as input to the next layer of GAT;

[0079] After multi-layer aggregation, the final hidden representation of the node is input into the time series LSTM structure.

[0080] The time series LSTM structure is specifically as follows:

[0081] Arrange the final hidden representation of the node according to the time step to serve as the input at the current moment;

[0082] The final hidden representation of the node is input into the LSTM unit together with the LSTM hidden state obtained in the previous time step. It is updated through the input gate, forget gate, and output gate of the LSTM unit to obtain a new LSTM hidden state, which represents the temporal characteristics of the node at this moment.

[0083] For each node, after traversing all time steps, the hidden state representation of the last output of the LSTM unit is represented as the final temporal representation of each node.

[0084] The collision risk and flight path decision output layer is specifically:

[0085] For the final time series representation of each node, the collision risk of the drone and the connectivity of adjacent drones are output through the fully connected layer;

[0086] If the UAV collision risk exceeds the set threshold or the connectivity of adjacent UAVs is lower than the set value, the flight path adjustment module is triggered to generate a new UAV flight path;

[0087] Deliver new drone flight paths to drones, perform drone emergency avoidance, or change drone flight paths.

[0088] In the specific implementation, the feature vector of each drone node contains its position, speed, acceleration, remaining power and other information. Assume that the feature vector of node i is h i , the characteristic vector of neighbor node j is h j , for each node i, calculate its attention score with all neighbor nodes j, and map the node features through the weight matrix W, z i =Wh i , z j =Wh j , where z i 、z j The feature representations of node i and node j after linear transformation are calculated respectively, and the attention score e of node i and neighbor node j is calculated. ij , e ij =LeakyReLU(a T [z i ||z j ]), where a is the weight vector, and then the attention score is normalized by the softmax function to obtain the attention coefficient α ij The feature vectors of neighboring nodes are weighted summed by the attention coefficient to obtain the aggregation result In order to enhance the expressiveness of the model, this embodiment uses multiple attention heads. Each head independently calculates the attention coefficient and aggregation results, and finally splices the results of all heads. The multi-head aggregation results are linearly transformed to generate the hidden representation of the node in the current layer. The hidden representation of the current layer is used as input and passed to the GAT of the next layer for further feature extraction and aggregation. In addition, in the collision risk and flight path decision output layer, the final time series representation of each node is output through the fully connected layer to output the collision risk of the drone and the connectivity of adjacent drones. The specific implementation is as follows: Among them, y i is the output of node i, including the predicted value of collision risk and connectivity, W out is the weight matrix of the fully connected layer, is the final temporal representation of node i after traversing all time steps, which integrates the feature changes of the node in multiple time steps, b out It is the bias of the fully connected layer. If the collision risk of the drone exceeds the set threshold or the connectivity of adjacent drones is lower than the set value, the system will trigger the flight path adjustment module. The flight path adjustment module generates a new drone flight path based on the current drone status and environmental data, and transmits the new drone flight path to the drone to perform emergency avoidance or change the flight path to ensure the effectiveness and safety of the path adjustment.

[0089] Specifically, the communication connectivity, collision risk, and energy consumption indicators are measured through the loss function, and the weights of the UAV graph neural network model are iteratively updated. Specifically,

[0090] Construct a connectivity matrix based on the nodes and edges in the current drone graph structure data;

[0091] According to the graph structure data of the drones at the current moment, the node set of the drones is defined to represent the identities of all drones. At the same time, the edge set between drones is defined to represent the direct communication links between drones. Construct an adjacency matrix. In the connectivity matrix, if there is a direct communication link between two drones, the corresponding matrix element is 1; if not, it is 0. By traversing the edge set between drones, the adjacency matrix is ​​filled, and then a connectivity matrix of the same size as the adjacency matrix is ​​created. For each pair of drones, a graph traversal algorithm is used to check whether there is a path from one drone to another. If so, the corresponding position in the connectivity matrix is ​​set to one, indicating that the two drones are connected. After the traversal is completed, the connectivity matrix will reflect the connectivity between all drones, and each element in the matrix indicates whether there is a connected path between the corresponding drones;

[0092] Calculate the ratio of the actual number of available links to the theoretical maximum number of links in the connectivity matrix to obtain the communication connectivity evaluation value;

[0093] Based on the distance, relative speed and predicted flight path between drones, the probability of collision between any two drones at time t is calculated, which is represented as collision risk.

[0094] For each drone, based on the drone’s current position and operating data, at each discrete moment, the spatial distance between the two drones is calculated and compared with the preset safety distance;

[0095] If the distance between two drones is less than the preset safety distance, it is marked as a risky moment, and the relative speed and direction of the drones at that moment are recorded;

[0096] The marking results of all risk moments are summarized within the prediction time window to form a comprehensive degree metric that characterizes the collision probability of two drones. The comprehensive degree metric is then converted into a collision probability based on empirical rules.

[0097] Establish a UAV energy consumption model and calculate the energy consumption at time t based on the UAV's flight speed, acceleration, attitude adjustment, and flight path turns;

[0098] Accumulate energy consumption at each moment to obtain the overall energy consumption index;

[0099] The communication connectivity evaluation value, collision risk and overall energy consumption index are defined as corresponding loss terms respectively, and each loss term is integrated into the total loss function through weighted linear combination.

[0100] Example 5 is the fifth embodiment of the present invention, which differs from the first two embodiments in that:

[0101] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0102] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0103] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0104] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0105] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A graph neural network-based anti-collision route planning method for drones in self-organizing networks, characterized by: include, Based on the drone graph structure data, each drone is regarded as a unit node, and the edges between nodes represent the potential collision relationship between two adjacent drones; Initialize node and edge features, including drone flight status features and environmental features. These input node and edge features are used as input for the pre-set drone graph neural network model. The model uses a loss function to measure communication connectivity, collision risk, and energy consumption. The model's weights are iteratively updated to complete training. Based on the trained UAV graph neural network model, the collision risk between each UAV and its neighboring nodes is predicted, the optimal flight path is generated and output, and the UAV flight is executed until the flight mission is completed.

2. The anti-collision route planning method for a UAV ad hoc network based on a graph neural network according to claim 1, characterized in that: The motion constraints of the UAV include: According to the UAV model and flight mission requirements, set the flight speed upper limit constraint, acceleration upper limit constraint, turning radius constraint, and climb / descent rate constraint as the UAV kinematic constraint parameters; The flight path planning of the UAV is limited by the altitude limit of the mission airspace, the minimum safe distance between obstacles, the boundaries of the no-fly zone and temporary restrictions; The UAV's motion constraint parameters are mathematically converted, and collision risk penalty weights are introduced to form the UAV's motion constraint conditions. Based on the real-time collected UAV operation data, it is determined whether the constraint correction should be triggered.

3. The anti-collision route planning method for a UAV ad hoc network based on a graph neural network according to claim 2, characterized in that: The construction of drone graph structure data includes: Each drone is represented as a node, and each node is bound to the drone's current position, kinematic data, remaining battery power, and airspace environment data; Create edges based on the relative positions and communication connectivity between drones in the mission airspace. If the distance between any two drones is less than the preset communication radius, an edge representing a potential collision relationship is established. The UAV kinematic data and airspace environment data are written into the feature vectors of nodes and edges respectively to form the UAV graph structure data.

4. The anti-collision route planning method for a UAV ad hoc network based on a graph neural network according to claim 3, characterized in that: The drone graph structure data includes: During the flight of the UAV, the feature vectors of nodes and edges are periodically updated according to the motion constraints of the UAV. If a new drone joins or an existing drone exits, nodes and edges are added or deleted to complete the dynamic adjustment of the drone graph structure data.

5. The anti-collision route planning method for a UAV ad hoc network based on a graph neural network according to claim 4, characterized in that: The UAV graph neural network model includes: It consists of a graph attention network layer based on drone graph structure data, a time series LSTM structure, and a collision risk and flight path decision output layer.

6. The anti-collision route planning method for a UAV ad hoc network based on a graph neural network according to claim 5, characterized in that: The graph attention network layer includes: Calculate the attention coefficient of each node in the drone graph structure data relative to all its neighboring nodes; Multiply each group of attention coefficients by the feature vectors of the corresponding neighboring nodes element by element and add them up to get the multi-head attention aggregation result; Splice the multi-head aggregation results to obtain the hidden representation of the node in the current layer; The node hidden representation is passed as input to the next layer of GAT; After multi-layer aggregation, the final hidden representation of the node is input into the time series LSTM structure.

7. The anti-collision route planning method for a UAV ad hoc network based on a graph neural network according to claim 6, characterized in that: The time series LSTM structure includes: Arrange the final hidden representation of the node according to the time step to serve as the input at the current moment; The final hidden representation of the node is input into the LSTM unit together with the LSTM hidden state obtained in the previous time step. It is updated through the input gate, forget gate, and output gate of the LSTM unit to obtain a new LSTM hidden state, which represents the temporal characteristics of the node at this moment. For each node, after traversing all time steps, the hidden state representation of the last output of the LSTM unit is represented as the final temporal representation of each node.

8. The anti-collision route planning method for a UAV ad hoc network based on a graph neural network according to claim 7, characterized in that: The collision risk and flight path decision output layer includes: For the final time series representation of each node, the collision risk of the drone and the connectivity of adjacent drones are output through the fully connected layer; If the UAV collision risk exceeds the set threshold or the connectivity of adjacent UAVs is lower than the set value, the flight path adjustment module is triggered to generate a new UAV flight path; Deliver new drone flight paths to drones, perform drone emergency avoidance, or change drone flight paths.

9. The anti-collision route planning method for a UAV ad hoc network based on a graph neural network according to claim 8, characterized in that: The loss function is used to measure communication connectivity, collision risk, and energy consumption indicators, and the weights of the UAV graph neural network model are iteratively updated, including: Construct a connectivity matrix based on the nodes and edges in the current drone graph structure data; Calculate the ratio of the actual number of available links to the theoretical maximum number of links in the connectivity matrix to obtain the communication connectivity evaluation value; Based on the distance, relative speed and predicted flight path between drones, the probability of collision between any two drones at time t is calculated, which is represented as collision risk. Establish a UAV energy consumption model and calculate the energy consumption at time t based on the UAV's flight speed, acceleration, attitude adjustment, and flight path turns; Accumulate energy consumption at each moment to obtain the overall energy consumption index; The communication connectivity evaluation value, collision risk and overall energy consumption index are defined as corresponding loss terms respectively, and each loss term is integrated into the total loss function through weighted linear combination.

10. The anti-collision route planning method for UAV ad hoc networks based on graph neural networks according to claim 9, characterized in that: The method for solving the probability of collision between any two UAVs at time t includes: For each drone, based on the drone’s current position and operating data, at each discrete moment, the spatial distance between the two drones is calculated and compared with the preset safety distance; If the distance between two drones is less than the preset safety distance, it is marked as a risky moment, and the relative speed and direction of the drones at that moment are recorded; The marking results of all risk moments are summarized within the prediction time window to form a comprehensive degree metric that characterizes the collision probability of two drones, and the comprehensive degree metric is converted into collision probability based on empirical rules.

Images