Artificial intelligence-based unmanned aerial vehicle task intelligent scheduling management method

Abstract

This invention discloses an artificial intelligence-based intelligent scheduling and management method for unmanned aerial vehicle (UAV) missions, belonging to the field of intelligent UAV scheduling and control technology. The method includes collecting environmental perception data streams and mission command streams from a UAV swarm to be scheduled, converting them into heterogeneous data frames with a unified timestamp, parsing and extracting airspace obstacle distribution and mission urgency information, inputting this information into a spatiotemporal grid construction engine, dynamically adjusting the grid resolution to generate an adaptive spatiotemporal grid matrix, calculating the mission density value and risk weight value of each grid unit to form a global situation quantification table, and using a graph neural network model to deduce the node adaptability score of each UAV unit. Based on this, task assignment is completed, an initial scheduling command set is generated, and after verifying and eliminating path conflicts, a UAV swarm collaborative operation scheme is output. This method improves the adaptability of UAV scheduling to the airspace environment and mission requirements through adaptive spatiotemporal grid modeling and graph neural network deduction.

Description

Technical Field

[0001] This invention belongs to the field of intelligent scheduling and control technology for unmanned aerial vehicles (UAVs), specifically an intelligent scheduling and management method for UAV missions based on artificial intelligence. Background Technology

[0002] Current UAV swarm mission scheduling and management mostly employs fixed-resolution spatiotemporal grids to model the operational airspace, processing environmental perception data and mission command data independently. Task allocation and path planning rely on traditional algorithms, with only basic verification of flight paths performed during scheduling. This approach fails to integrate heterogeneous data using unified timestamps, the grid parameters for airspace modeling cannot be adjusted to changes in the operational environment, mission status quantification relies solely on a single-dimensional parameter, and scheduling decisions do not incorporate an analysis of the compatibility between UAV units and the airspace situation.

[0003] Fixed-resolution spatiotemporal grids cannot adapt to the dynamic distribution of airspace obstacles. Complex operational airspaces are prone to insufficient grid accuracy, while simple airspaces suffer from grid resource redundancy. Independent processing of heterogeneous data leads to reduced spatiotemporal matching between airspace and task information, resulting in deviations in global situation quantification. Traditional algorithms cannot deduce the fit between UAV unit nodes and operational situations, resulting in insufficient rationality in task assignment and potential path conflicts in initial scheduling commands. Therefore, it is necessary to dynamically adjust the grid resolution based on airspace obstacle distribution to construct an adaptive spatiotemporal grid matrix, quantify the global situation through task density and risk weights of grid units, and rely on graph neural networks to deduce the node fit score of UAV units to complete task assignment. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art;

[0005] Therefore, this invention proposes an intelligent scheduling and management method for unmanned aerial vehicle (UAV) missions based on artificial intelligence, including:

[0006] Collect environmental perception data streams and task command streams around the drone swarm to be scheduled, and convert the environmental perception data streams and task command streams into heterogeneous data frames with a unified timestamp.

[0007] The heterogeneous data frames are parsed to extract spatial obstacle distribution information and task urgency information, which are then input into the spatiotemporal grid construction engine.

[0008] The spatiotemporal grid construction engine dynamically adjusts the grid resolution based on the airspace obstacle distribution information to generate an adaptive spatiotemporal grid matrix covering the operation area of ​​the drone swarm to be scheduled.

[0009] Traverse the adaptive spatiotemporal grid matrix, calculate the task density value and risk weight value in each grid cell, and summarize them to form a global situation quantification table;

[0010] Read the global situation quantification table and use a graph neural network model to deduce the node adaptability score of each UAV unit in the UAV swarm to be scheduled.

[0011] Based on the node adaptability score, tasks are assigned to the drone swarm to be scheduled, and an initial scheduling instruction set containing flight path points and execution order is generated.

[0012] Verify the path conflicts in the initial scheduling instruction set, eliminate the conflicts, and output the final UAV swarm collaborative operation scheme.

[0013] Further, environmental perception data streams and task command streams around the drone swarm to be scheduled are collected, and the environmental perception data streams and task command streams are converted into heterogeneous data frames with a unified timestamp, including:

[0014] The multimodal sensor array deployed on the airborne end of the unmanned aerial vehicle swarm to be scheduled is activated. The multimodal sensor array simultaneously captures visible light image data, infrared thermal imaging data and radar point cloud data, which are then converged to form the environmental perception data stream.

[0015] Receive task dispatch messages from remote command center, parse the target coordinate sequence and payload operation parameters in the task dispatch messages, and organize them into the task instruction stream;

[0016] A high-precision clock synchronization mechanism is established to assign an absolute time tag to each frame of sensing data in the environmental sensing data stream and each instruction in the task instruction stream.

[0017] The sensed data and instruction data with the absolute time stamp are encapsulated into a fixed-length data packet. The data packet contains a data source identifier field, a data type field, and a payload field. The payload field stores the original sampled value, thereby forming the heterogeneous data frame.

[0018] Further, the heterogeneous data frames are parsed to extract spatial obstacle distribution information and task urgency information. This spatial obstacle distribution information and task urgency information are then input into the spatiotemporal raster construction engine, including:

[0019] Read the data type field in the heterogeneous data frame. If the data type field indicates that it is sensing data, call the three-dimensional reconstruction algorithm to process the radar point cloud data in the payload field, identify the geometric contours and spatial coordinates of airborne obstacles, and generate the airspace obstacle distribution information.

[0020] If the data type field indicates instruction data, then the target coordinate sequence in the payload field is parsed, the task creation time and execution deadline are extracted, the reciprocal of the remaining time window is calculated as the basic urgency, and weighted in combination with the task type weight in the payload operation parameters to generate the task urgency information.

[0021] Align the obstacle coordinate set in the airspace obstacle distribution information with the task coordinates and urgency value in the task urgency information according to the absolute time label;

[0022] The time-aligned obstacle coordinates set, task coordinates, and urgency value are packaged into an engine input structure. The header of the engine input structure contains the current timestamp, and the main body contains an obstacle list and a task list. The engine input structure is then sent to the receiving interface of the spatiotemporal grid construction engine.

[0023] Furthermore, the spatiotemporal grid construction engine dynamically adjusts the grid resolution based on the airspace obstacle distribution information to generate an adaptive spatiotemporal grid matrix covering the operational area of ​​the unmanned aerial vehicle (UAV) swarm to be scheduled, including:

[0024] The spatiotemporal grid construction engine reads the list of obstacles in the engine input structure and counts the number and density of obstacles per unit space volume.

[0025] When the number density of obstacles exceeds a preset dense threshold, a high-resolution mode is triggered, reducing the physical size of the grid cells and increasing the number of subdivision layers in the grid hierarchy.

[0026] When the number density of obstacles is lower than a preset sparse threshold, a low-resolution mode is triggered, increasing the physical size of the grid cells and reducing the number of subdivision layers in the grid hierarchy.

[0027] Using the geographical center of the operation area of ​​the drone swarm to be scheduled as the origin, the latitude and longitude grid is divided according to the current physical size of the grid unit, and each grid unit is assigned a unique index code.

[0028] The obstacle list and task list in the engine input structure are mapped to the corresponding grid cells, and the obstacle occupancy flag and task urgency accumulation value in each grid cell are recorded, thereby constructing the adaptive spatiotemporal grid matrix.

[0029] Furthermore, the adaptive spatiotemporal grid matrix is ​​traversed, and the task density value and risk weight value within each grid cell are calculated and summarized to form a global situation quantification table, including:

[0030] The adaptive spatiotemporal grid matrix is ​​scanned row by row. For each grid cell with a non-empty task list, the total number of task entries in the task list is counted and divided by the physical area of ​​the grid cell to obtain the task density value.

[0031] Read the obstacle occupancy flag within the grid cell. If the obstacle occupancy flag is true, query the pre-stored obstacle threat database to obtain the collision risk assessment coefficient of the obstacle.

[0032] The risk weight value is obtained by multiplying the collision risk assessment coefficient by a preset environmental penalty factor and then combining it with the meteorological turbulence index of the airspace where the grid cell is located.

[0033] A two-dimensional table structure is constructed as the global situation quantification table. The row index of the two-dimensional table structure corresponds to the index code of the raster cell, and the column fields include the task density value and the risk weight value.

[0034] The task density value and risk weight value of each grid cell obtained by traversal calculation are filled into the corresponding row of the two-dimensional table structure to complete the filling of the global situation quantization table.

[0035] Furthermore, the global situation quantization table is read, and the node adaptability score of each UAV unit in the UAV swarm to be scheduled is deduced through a graph neural network model, including:

[0036] Load a pre-trained graph neural network model, wherein the input layer of the graph neural network model receives the grid cell node features from the global situation quantization table;

[0037] Each row of data in the global situation quantization table is converted into a graph node, and the node feature vector contains the task density value, the risk weight value, and the geographic location code of the grid cell;

[0038] Based on the spatial connectivity between adjacent grid cells, edge connection relationships are constructed to form the spatial situation map data structure to be processed;

[0039] The spatial situation map data structure is input into the graph neural network model for forward propagation calculation. The graph neural network model updates the hidden state of the central node by aggregating neighbor node information.

[0040] The node features corresponding to the current positions of each UAV unit in the UAV swarm to be scheduled are extracted from the output layer of the graph neural network model. After mapping transformation by a fully connected layer, the node fitness score is output. The node fitness score represents the degree of task execution matching of the adaptive spatiotemporal grid unit corresponding to the current real-time spatial position of each UAV unit in the UAV swarm to be scheduled.

[0041] Furthermore, based on the node adaptability score, tasks are assigned to the swarm of drones to be scheduled, generating an initial scheduling instruction set containing flight path points and execution order, including:

[0042] Collect the node fitness scores of all UAV units in the unmanned aerial vehicle swarm to be scheduled at the output layer of the graph neural network model, and construct a score matrix.

[0043] The Hungarian algorithm is used to process the scoring matrix, solve for the optimal matching relationship between the UAV unit and the task target point, and determine the target task list for each UAV unit.

[0044] For each UAV unit's target task list, combined with the risk weight values ​​in the global situation quantification table, a heuristic search algorithm is used to plan a continuous trajectory from the current position to each task target;

[0045] On the continuous flight path, locations with significant terrain variations and mission execution locations are selected as key nodes, and the geographic coordinates of the key nodes are extracted as the flight path points.

[0046] According to the order of task execution and arrival time constraints, the order of the flight path points is arranged, and the UAV unit number, flight path point sequence and expected action are encapsulated into a scheduling record;

[0047] The scheduling records of all UAV units are aggregated and encoded according to a unified instruction format to generate the initial scheduling instruction set.

[0048] Furthermore, the path conflicts in the initial scheduling instruction set are verified, and after conflict elimination, the final UAV swarm collaborative operation scheme is output, including:

[0049] Analyze each scheduling record in the initial scheduling instruction set and extract the flight path point sequence corresponding to the UAV unit number;

[0050] A four-dimensional spatiotemporal trajectory model is constructed, and the flight path point sequence of each UAV unit is converted into a position coordinate function that varies with time.

[0051] Check if there is an intersection between the spatiotemporal trajectories of any two drone units. If there is an intersection and the time interval is less than the safe avoidance threshold, it is determined to be a path conflict.

[0052] For drone unit pairs that are determined to have path conflicts, adjust the flight path point sequence of at least one drone unit, specifically by shifting the takeoff time along the time axis or adding intermediate collision avoidance path points by detouring.

[0053] Repeat the path conflict checking and adjustment steps until there are no path conflicts in the initial scheduling instruction set;

[0054] All scheduling records after conflict resolution are repackaged, digital signatures and verification codes are added, and the final drone swarm collaborative operation scheme is generated and sent to the drone swarm to be scheduled.

[0055] Furthermore, the construction method of the graph neural network model includes:

[0056] Define the graph structure input specification, set the node feature dimension of the spatial situation map to a fixed length, and the node feature dimension shall at least cover the three-dimensional feature vector of task density value, risk weight value and geographic location code. At the same time, define the adjacency matrix as an undirected edge connection relationship based on the spatial connectivity of grid cells.

[0057] A multi-layer graph convolutional network architecture is constructed by stacking graph convolutional layers, batch normalization layers, and activation function layers in sequence. The graph convolutional layers use a message passing mechanism to aggregate neighbor node features, and the activation function layers use linear units with leakage correction to enhance the expressive power of sparse features.

[0058] A graph attention mechanism module is configured and embedded after the graph convolutional layer. By calculating the attention coefficients of the feature vectors between nodes, the contribution weights of different neighboring nodes to the feature update of the central node are dynamically allocated. The calculation of the attention coefficients is based on the learnable linear transformation and softmax normalization operation between the feature vectors of the nodes.

[0059] The model output mapping head is designed to input the node hidden state vector output by the last layer of graph convolutional network into two layers of fully connected neural network. The first fully connected layer compresses the feature dimension into the intermediate hidden layer space, and the second fully connected layer outputs a single value as the node fitness score. A sigmoid activation function is then applied after the output layer to normalize the score to the range of 0 to 1.

[0060] The model training process involves collecting a labeled dataset of historical UAV scheduling tasks. The labeled dataset includes ground truth labels for node fitness generated by human experts. The mean squared error loss function is used to measure the difference between the model prediction score and the ground truth labels. The backpropagation algorithm combined with the Adam optimizer is used to iteratively update the network weight parameters until the loss on the validation set converges to below a preset threshold. The trained network parameter file is then saved as the pre-trained graph neural network model.

[0061] Furthermore, it also includes a dynamic rescheduling triggering mechanism, specifically including:

[0062] During the execution of the final drone swarm collaborative operation scheme, real-time environmental feedback data from the multimodal sensor array is continuously monitored.

[0063] The real-time environmental feedback data is compared with a preset safety threshold. If a sudden obstacle intrusion or a change in the state of the mission target is detected, a rescheduling trigger signal is generated.

[0064] In response to the rescheduling trigger signal, the currently executing scheduling process is interrupted, and the status of the completed tasks is frozen.

[0065] The latest environmental perception data stream and task instruction stream are re-acquired, and the process jumps to the step of parsing the heterogeneous data frame, generating a new adaptive spatiotemporal raster matrix based on the latest data.

[0066] Based on the new adaptive spatiotemporal grid matrix, the global situation quantization table and node adaptability score are recalculated to generate incremental scheduling adjustment instructions, which replace the original instructions and are then issued for execution.

[0067] Compared with the prior art, the beneficial effects of the present invention are:

[0068] By dynamically adjusting the grid resolution based on the airspace obstacle distribution information, an adaptive spatiotemporal grid matrix covering the operation area of ​​the drone swarm to be scheduled is generated. This allows the spatial accuracy of the grid to be adapted to the distribution density of obstacles in the operation airspace. The grid division in dense obstacle areas is more refined, while the grid division in sparse obstacle areas is more extensive. This makes the structure of the spatiotemporal grid matrix fit the actual spatial characteristics of the drone operation airspace, avoiding the problem of insufficient adaptability between fixed resolution grids and the operation environment. This ensures that the results of airspace modeling are consistent with the actual operation scenario, and that the shape of the grid matrix matches the spatial requirements of drone operations.

[0069] The adaptive spatiotemporal grid matrix is ​​traversed to calculate the task density and risk weight values ​​within each grid cell, which are then aggregated to form a global situation quantification table. By using a graph neural network model to deduce the node adaptability score of each UAV unit in the UAV swarm to be scheduled, the task distribution and risk parameters in the airspace can be transformed into standardized quantitative data. Relying on the graph neural network's ability to analyze the correlation between airspace topology and UAV nodes, the correspondence between UAV units and operational grids and task requirements is accurately matched. Task assignment is completed based on the node adaptability score. The generated initial scheduling instruction set can fit the global airspace situation, reducing the probability of path conflicts in the instructions. This makes the task assignment and flight path planning of the UAV swarm more in line with the actual needs of the operational scenario, and optimizes the generation and execution logic of scheduling instructions. Attached Figure Description

[0070] Figure 1This is a flowchart illustrating the steps of the AI-based intelligent scheduling and management method for unmanned aerial vehicle (UAV) missions described in this invention.

[0071] Figure 2 A flowchart for generating an adaptive spatiotemporal raster matrix;

[0072] Figure 3 Heatmap of drone-mission node compatibility score;

[0073] Figure 4 The training loss curve for the graph neural network model;

[0074] Figure 5 The graph neural network output is the distribution of drone node fitness scores. Detailed Implementation

[0075] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] See Figure 1 This invention provides an intelligent scheduling and management method for unmanned aerial vehicle (UAV) missions based on artificial intelligence. The method includes: upon startup, collecting environmental perception data streams and mission command streams around the UAV swarm to be scheduled, and converting these two types of data streams into heterogeneous data frames with a unified timestamp. These heterogeneous data frames are parsed to extract airspace obstacle distribution information and mission urgency information, and this information is input into a spatiotemporal grid construction engine. The spatiotemporal grid construction engine dynamically adjusts its grid resolution based on the airspace obstacle distribution information, generating an adaptive spatiotemporal grid matrix covering the operating area of ​​the UAV swarm to be scheduled. The system traverses this adaptive spatiotemporal grid matrix, calculates the mission density value and risk weight value within each grid cell, and summarizes them into a global situation quantification table. This global situation quantification table is read, and a pre-trained graph neural network model is used to extrapolate and calculate the node adaptability score of each UAV unit in the UAV swarm to be scheduled. Based on these node adaptability scores, mission assignment is performed on the UAV swarm to be scheduled, thereby generating an initial scheduling command set containing flight path points and execution order. The initial scheduling instruction set is checked for path conflicts. After eliminating all identified conflicts, the final UAV swarm collaborative operation scheme is output.

[0077] In one embodiment of the invention, a multimodal sensor array deployed on the airborne end of a swarm of unmanned aerial vehicles (UAVs) to be scheduled is activated. This multimodal sensor array simultaneously captures visible light image data, infrared thermal imaging data, and radar point cloud data, which converge to form an environmental perception data stream. Simultaneously, a task dispatch message from a remote command center is received, and the target coordinate sequence and payload operation parameters in the message are parsed to form a task command stream. To fuse these two types of data streams, a high-precision clock synchronization mechanism is established, assigning an absolute time tag to each frame of perception data in the environmental perception data stream and each command in the task command stream. The perception data and command data with absolute time tags are encapsulated into fixed-length data packets. Each data packet contains a data source identifier field, a data type field, and a payload field. The payload field stores the original sampled values, thus forming a heterogeneous data frame with a unified timestamp. When parsing the heterogeneous data frame, the data type field is read. If the data type field indicates perception data, a 3D reconstruction algorithm is invoked to process the radar point cloud data in the payload field, identifying the geometric contours and spatial coordinates of aerial obstacles and generating airspace obstacle distribution information. If the data type field indicates instruction data, the target coordinate sequence in the payload field is parsed, the task creation time and execution deadline are extracted, the reciprocal of the remaining time window is calculated as the basic urgency, and weighted in conjunction with the task type weight in the payload operation parameters to generate task urgency information. Next, the obstacle coordinate set in the airspace obstacle distribution information and the task coordinates and urgency values ​​in the task urgency information are aligned according to absolute time labels. The time-aligned obstacle coordinate set, task coordinates, and urgency values ​​are packaged into an engine input structure. The header of this engine input structure contains the current timestamp, and the body contains the obstacle list and task list. Finally, this engine input structure is sent to the receiving interface of the spatiotemporal raster construction engine.

[0078] In specific implementation, a multimodal sensor array deployed on the airborne end of the unmanned aerial vehicle (UAV) swarm to be scheduled is activated. The multimodal sensor array synchronously captures visible light image data, infrared thermal imaging data, and radar point cloud data, which are then aggregated to form an environmental perception data stream. At the same time, it receives task dispatch messages from the remote command center, parses the target coordinate sequence and payload operation parameters in the task dispatch messages, and organizes them into a task instruction stream. A high-precision clock synchronization mechanism is established to assign an absolute time tag to each frame of perception data in the environmental perception data stream and each instruction in the task instruction stream. It can be understood that the absolute time tag ensures the time consistency of different data sources. In some embodiments, the high-precision clock synchronization mechanism uses a network time protocol to achieve microsecond-level synchronization. The perception data and instruction data with absolute time tags are encapsulated into fixed-length data packets. The data packets contain a data source identifier field, a data type field, and a payload field. The payload field stores the original sampled value, thus forming a heterogeneous data frame under a unified timestamp. Optionally, the data source identifier field is used to distinguish the sensor type or instruction source.

[0079] In specific implementation, the process of parsing heterogeneous data frames includes reading the data type field in the heterogeneous data frame. If the data type field indicates sensing data, a 3D reconstruction algorithm is invoked to process the radar point cloud data in the payload field, identify the geometric contours and spatial coordinates of airborne obstacles, and generate airspace obstacle distribution information. If the data type field indicates command data, the target coordinate sequence in the payload field is parsed, the task creation time and execution deadline are extracted, the reciprocal of the remaining time window is calculated as the basic urgency, and weighted in conjunction with the task type weight in the payload operation parameters to generate task urgency information. In some embodiments, the task urgency information is calculated using a formula:

[0080]

[0081] in: This indicates the urgency level of the task. This indicates the task type weight in the load operation parameters. This represents the remaining time window from the current moment to the execution deadline. The obstacle coordinate set from the airspace obstacle distribution information and the task coordinates and urgency values ​​from the task urgency information are aligned using absolute timestamps. Optionally, a timestamp matching algorithm is used for the alignment operation to ensure data time synchronization. The time-aligned obstacle coordinate set, task coordinates, and urgency values ​​are packaged into an engine input structure. The header of the engine input structure contains the current timestamp, and the main body contains the obstacle list and the task list. The engine input structure is sent to the receiving interface of the spatiotemporal grid construction engine. It can be understood that the engine input structure provides a standardized data interface.

[0082] In one embodiment of the present invention, see [reference] Figure 2 The spatiotemporal grid construction engine reads the obstacle list from the engine input structure and calculates the obstacle density per unit volume. When the obstacle density exceeds a preset density threshold, a high-resolution mode is triggered, reducing the physical size of the grid cells and increasing the number of subdivisions in the grid hierarchy. When the obstacle density is below a preset sparse threshold, a low-resolution mode is triggered, increasing the physical size of the grid cells and reducing the number of subdivisions in the grid hierarchy. Using the geographic center of the drone swarm's operating area as the origin, a latitude and longitude grid is divided according to the current physical size of the grid cells, and each grid cell is assigned a unique index code. The obstacle list and task list in the engine input structure are mapped to the corresponding grid cells, recording the obstacle occupancy flag and the accumulated task urgency value within each grid cell, thus constructing an adaptive spatiotemporal grid matrix. The adaptive spatiotemporal grid matrix is ​​traversed, scanning row by row. For each grid cell with a non-empty task list, the total number of task entries in its task list is counted and divided by the physical area of ​​the grid cell to obtain the task density value of that grid cell. The obstacle occupancy flag within the grid cell is read. If the flag is true, the pre-stored obstacle threat database is queried to obtain the collision risk assessment coefficient for that obstacle. This coefficient is multiplied by a preset environmental penalty factor, and then summed with the meteorological turbulence index of the airspace where the grid cell is located to obtain the risk weight value for that grid cell. A two-dimensional table structure is constructed as the global situation quantification table. The row index of this table structure corresponds to the index code of the grid cell, and the column fields contain the task density value and risk weight value. The task density value and risk weight value of each grid cell obtained through iteration are filled into the corresponding row of the two-dimensional table structure, completing the global situation quantification table filling.

[0083] In practical implementation, the spatiotemporal grid construction engine reads the obstacle list from the engine input structure and counts the obstacle density per unit volume. When the obstacle density exceeds a preset density threshold, a high-resolution mode is triggered, reducing the physical size of the grid cells and increasing the number of subdivisions in the grid hierarchy. When the obstacle density is below a preset sparsity threshold, a low-resolution mode is triggered, increasing the physical size of the grid cells and decreasing the number of subdivisions in the grid hierarchy. In some embodiments, the density threshold is set to 5 obstacles per cubic kilometer, and the sparsity threshold is set to 1 obstacle per cubic kilometer. The physical size of the grid cell in high-resolution mode is 10 meters × 10 meters, and the physical size of the grid cell in low-resolution mode is... The grid is 50m x 50m in size, with the geographic center of the area where the UAV swarm to be scheduled operates as the origin. It is divided into latitude and longitude grids according to the physical size of the current grid cells. Each grid cell is assigned a unique index code. Optionally, the index code adopts a three-segment numerical combination of "longitude level - latitude level - altitude level". The obstacle list and task list in the engine input structure are mapped to the corresponding grid cells. The obstacle occupancy flag and the cumulative task urgency value in each grid cell are recorded. The obstacle occupancy flag is a Boolean value, indicating whether there is an obstacle in the grid cell. The cumulative task urgency value is the arithmetic sum of the urgency values ​​of all tasks in the grid cell. Thus, an adaptive spatiotemporal grid matrix is ​​constructed.

[0084] In practical implementation, the process of traversing the adaptive spatiotemporal grid matrix involves row-by-row scanning. For each grid cell with a non-empty task list, the total number of task entries in the task list is counted and divided by the physical area of ​​the grid cell to obtain the task density value of the grid cell. The obstacle occupancy flag within the grid cell is read. If the obstacle occupancy flag is true, the pre-stored obstacle threat database is queried to obtain the collision risk assessment coefficient of the obstacle. It can be understood that the collision risk assessment coefficient is preset based on the type, motion state, and size of the obstacle. The collision risk assessment coefficient is multiplied by a preset environmental penalty factor and then combined with the meteorological turbulence index of the airspace where the grid cell is located to obtain the risk weight value of the grid cell. In some embodiments, the risk weight value is calculated using a formula:

[0085]

[0086] in: This represents the risk weight value. This represents the collision risk assessment coefficient. Indicates environmental penalty factor, To represent the meteorological turbulence index, a two-dimensional table structure is constructed as the global situation quantification table. The row index of the two-dimensional table structure corresponds to the index code of the raster cell, and the column fields contain the task density value and risk weight value. The task density value and risk weight value of each raster cell obtained by traversing and calculating are filled into the corresponding row of the two-dimensional table structure to complete the filling of the global situation quantification table. Optionally, the global situation quantification table is stored in the form of a key-value pair database table or a two-dimensional array in memory.

[0087] In one embodiment of the present invention, a pre-trained graph neural network model is loaded. The input layer of this graph neural network model receives the grid cell node features from a global situation quantization table. Each row of data in the global situation quantization table is converted into a graph node, and the node feature vector contains the task density value, risk weight value, and the geographic location code of the grid cell. Edge connections are constructed based on the spatial connectivity between adjacent grid cells to form the airspace situation map data structure to be processed. The airspace situation map data structure is input into the graph neural network model for forward propagation calculation. The graph neural network model updates the hidden state of the central node by aggregating neighbor node information. Node features corresponding to the current positions of each UAV unit in the UAV swarm to be scheduled are extracted from the output layer of the graph neural network model. After mapping transformation by a fully connected layer, a node fitness score is output. The node fitness score represents the degree of task execution matching of the adaptive spatiotemporal grid cell corresponding to the current real-time spatial position of each UAV unit in the UAV swarm to be scheduled. The node fitness scores corresponding to all UAV units in the UAV swarm to be scheduled at the output layer of the graph neural network model are collected to construct a score matrix. The Hungarian algorithm is used to process the scoring matrix, solving for the optimal matching relationship between UAV units and mission target points, and determining the target mission list for each UAV unit. For each UAV unit's target mission list, combined with the risk weight values ​​in the global situation quantification table, a heuristic search algorithm is used to plan a continuous trajectory from the current position to each mission target. Locations with significant terrain variations and mission execution positions are selected as key nodes on this continuous trajectory, and their geographic coordinates are extracted as flight path points. The flight path points are ordered according to the sequence of mission execution and arrival time constraints, and the UAV unit number, flight path point sequence, and expected execution action are encapsulated into a scheduling record. All UAV unit scheduling records are summarized and encoded according to a unified instruction format to generate an initial scheduling instruction set.

[0088] Each scheduling record in the initial scheduling instruction set is analyzed to extract the flight path point sequence corresponding to the UAV unit number. A four-dimensional spatiotemporal trajectory model is constructed, converting the flight path point sequence of each UAV unit into a position coordinate function that varies with time. The spatiotemporal trajectories of any two UAV units are checked for intersection. If an intersection exists and the time interval is less than the safe avoidance threshold, a path conflict is identified. For UAV unit pairs identified as having path conflicts, the flight path point sequence of at least one UAV unit is adjusted, specifically by shifting the takeoff time along the time axis or adding intermediate collision avoidance path points. The path conflict checking and adjustment steps are repeated until there are no path conflicts in the initial scheduling instruction set. All scheduling records after conflict resolution are repackaged, digital signatures and verification codes are added, and the final UAV swarm collaborative operation plan is generated and distributed to the UAV swarm to be scheduled.

[0089] In specific implementation, a pre-trained graph neural network model is loaded. The input layer of the graph neural network model receives the grid cell node features from the global situation quantization table, converting each row of data in the global situation quantization table into a graph node. The node feature vector contains the task density value, risk weight value, and the geographic location code of the grid cell. Edge connections are constructed based on the spatial connectivity between adjacent grid cells, forming the airspace situation map data structure to be processed. It can be understood that the edge connection relationship is defined based on the adjacency relationship of the grid cells in three-dimensional space. In some embodiments, if two grid cells share a face, edge, or vertex, they are considered adjacent and a connection is established. The airspace situation map data structure is input into the graph neural network model for forward propagation calculation. The graph neural network model updates the hidden state of the central node by aggregating neighbor node information. The node features corresponding to the current position of each UAV unit in the UAV swarm to be scheduled are extracted from the output layer of the graph neural network model. After mapping transformation by the fully connected layer, the node fitness score is output. The node fitness score represents the degree of task execution matching of the adaptive spatiotemporal grid cell corresponding to the current real-time spatial position of each UAV unit in the UAV swarm to be scheduled. The calculation formula for the node fitness score is:

[0090]

[0091] in: Indicates unmanned aerial vehicle (UAV) unit Node adaptability score, Indicates unmanned aerial vehicle (UAV) unit The feature vector corresponding to the grid node, This represents the weight vector of the fully connected layer. This represents the bias term of the fully connected layer.

[0092] In practice, the node fit scores of all UAV units in the UAV swarm to be scheduled are collected at the output layer of the graph neural network model, and a score matrix is ​​constructed. The rows of the score matrix correspond to UAV units, and the columns correspond to task target points or grid units. Optionally, the task target points are determined by the coordinates in the task list and associated with specific grid units. See Table 1 for a node fit score matrix containing three UAV units and three task target points.

[0093] Table 1: Node Fitting Rating Matrix

[0094] Unmanned Aerial Vehicle Unit Task objective point A rating Task objective point B rating Task objective point C score UAV1 0.85 0.72 0.91 UAV2 0.78 0.88 0.65 UAV3 0.92 0.68 0.79

[0095] The Hungarian algorithm is used to process the scoring matrix, solve for the optimal matching relationship between UAV units and task target points, and determine the target task list for each UAV unit. For each UAV unit's target task list, combined with the risk weight values ​​in the global situation quantification table, a heuristic search algorithm is used to plan a continuous trajectory from the current position to each task target. On the continuous trajectory, locations with significant terrain undulations and task execution locations are selected as key nodes, and the geographical coordinates of the key nodes are extracted as flight path points. The order of the flight path points is arranged according to the sequence of task execution and arrival time constraints. The UAV unit number, flight path point sequence, and expected execution action are encapsulated into a scheduling record. The scheduling records of all UAV units are summarized and encoded according to a unified instruction format to generate an initial scheduling instruction set.

[0096] In specific implementation, each scheduling record in the initial scheduling instruction set is parsed, the flight path point sequence corresponding to the UAV unit number is extracted, a four-dimensional spatiotemporal trajectory model is constructed, and the flight path point sequence of each UAV unit is converted into a position coordinate function that changes with time. It is checked whether there is an intersection between the spatiotemporal trajectories of any two UAV units. If there is an intersection and the time interval is less than the safety avoidance threshold, it is determined to be a path conflict. In some embodiments, the safety avoidance threshold is dynamically set according to the minimum safe interval and speed of the UAV. For UAV unit pairs that are determined to have path conflicts, the flight path point sequence of at least one UAV unit is adjusted. Specifically, the takeoff time is shifted along the time axis or an intermediate collision avoidance path point is added. The path conflict checking and adjustment steps are repeated until there are no path conflicts in the initial scheduling instruction set. It can be understood that the path conflict checking adopts an iterative optimization algorithm until the conflict-free condition is met. All scheduling records after conflict resolution are repackaged, digital signatures and verification codes are attached, and the final UAV cluster collaborative operation scheme is generated and sent to the UAV cluster to be scheduled. Optionally, the digital signature adopts an asymmetric encryption algorithm to ensure the trustworthiness of the instruction source.

[0097] See Figure 3In the AI-based intelligent scheduling and management method for UAV missions, the UAV-task node fit score heatmap is the core visualization output of the graph neural network model. It visually presents the degree of task execution matching between each UAV unit in the UAV swarm to be scheduled and different task target points. Specifically, the data source for this heatmap is a node fit score matrix. Its row dimension corresponds to the UAV units to be scheduled (UAV1, UAV2, UAV3), and its column dimension corresponds to the task target points to be executed (Task A, Task B, Task C). The values ​​within the matrix are the normalized fit scores (range 0-1) output by the graph neural network model. A higher score indicates a better match between the UAV unit and the corresponding task target point. The heatmap uses a blue-white gradient color scheme, mapping the fit score values ​​by color depth: dark blue areas correspond to high scores (approaching 1.00), representing extremely strong fit between the UAV unit and the task target point, making it the optimal assignment candidate; light blue to white areas correspond to low scores (approaching 0.60), representing weaker fit, and are not recommended for priority assignment. Taking the data in the figure as an example, the fit score between UAV3 and task A is 0.92, the highest value in the entire matrix, indicating that this UAV unit is most suitable for executing task A; the fit score between UAV2 and task C is 0.65, the lowest value in the entire matrix, indicating that this combination has the worst task execution matching degree. The heatmap is the core input basis for the subsequent task assignment process: based on the scoring matrix, the Hungarian algorithm is used to solve for the optimal UAV-task matching relationship, determine the target task list of each UAV unit, provide quantitative support for subsequent trajectory planning and scheduling instruction generation, and achieve optimal allocation of UAV swarm task resources.

[0098] In one embodiment of the present invention, the graph neural network model is constructed as follows: A graph structure input specification is defined, setting the node feature dimension of the spatial situation map to a fixed length. The node feature dimension at least covers three-dimensional feature vectors containing task density values, risk weight values, and geographic location codes. Simultaneously, the adjacency matrix is ​​defined as an undirected edge connection relationship based on the spatial connectivity of grid cells. A multi-layer graph convolutional network architecture is constructed, sequentially stacking graph convolutional layers, batch normalization layers, and activation function layers. The graph convolutional layers use a message-passing mechanism to aggregate neighbor node features, and the activation function layers use linear units with leakage correction to enhance the expressive power of sparse features. A graph attention mechanism module is configured and embedded after the graph convolutional layers. By calculating the attention coefficients between node feature vectors, the contribution weights of different neighbor nodes to the feature update of the central node are dynamically allocated. The calculation of the attention coefficients is based on learnable linear transformations and softmax normalization operations between node feature vectors. The model output mapping head is designed to input the hidden state vectors of the nodes output from the last layer of the graph convolutional network into a two-layer fully connected neural network. The first fully connected layer compresses the feature dimension into the intermediate hidden layer space, and the second fully connected layer outputs a single numerical value as the node fitness score. A sigmoid activation function is then applied after the output layer to normalize the score to the 0-1 range. The model training process involves collecting a labeled dataset of historical drone scheduling tasks. This dataset includes ground truth labels for node fitness generated by human experts. The mean squared error loss function is used to measure the difference between the model's predicted score and the ground truth labels. The backpropagation algorithm combined with the Adam optimizer is used to iteratively update the network weight parameters until the loss on the validation set converges below a preset threshold. The trained network parameter file is then saved as a pre-trained graph neural network model.

[0099] In specific implementation, the construction method of the graph neural network model includes defining the graph structure input specification, setting the node feature dimension of the spatial situation map to a fixed length, and ensuring that the node feature dimension at least covers the three-dimensional feature vector of task density value, risk weight value, and geographic location encoding. At the same time, the adjacency matrix is ​​defined as an undirected edge connection relationship based on the spatial connectivity of grid cells. It can be understood that an undirected edge connection relationship means that if grid cell A and grid cell B are adjacent, the connection relationship is bidirectional. A multi-layer graph convolutional network architecture is constructed by stacking graph convolutional layers, batch normalization layers, and activation function layers in sequence. The graph convolutional layers use a message passing mechanism to aggregate the features of neighboring nodes, and the activation function layers use linear units with leakage correction to enhance the expressive power of sparse features. In some embodiments, the negative slope parameter of the linear units with leakage correction is set to 0.01. Refer to Table 2 to show a basic network architecture containing three layers of graph convolutional layers.

[0100] Table 2: A basic network architecture containing three graph convolutional layers

[0101] Layer number Layer type Output feature dimension Activation function 1 Convolutional layer 64 Linear unit with leakage correction 2 Batch Normalization Layer 64 none 3 Convolutional layer 32 Linear unit with leakage correction 4 Batch Normalization Layer 32 none 5 Convolutional layer 16 Linear unit with leakage correction 6 Model output mapping head 1 Sigmoid

[0102] A graph attention mechanism module is configured and embedded after the graph convolutional layer. It dynamically allocates the contribution weights of different neighboring nodes to the feature update of the central node by calculating the attention coefficients of the feature vectors between nodes. The calculation of the attention coefficients is based on a learnable linear transformation and softmax normalization operation between the node feature vectors. Optionally, the learnable linear transformation is implemented through a shared weight matrix. A model output mapping head is designed to input the node hidden state vectors output from the last graph convolutional network layer into a two-layer fully connected neural network. The first fully connected layer compresses the feature dimension into the intermediate hidden layer space, and the second fully connected layer outputs a single numerical value as the node fitness score. A sigmoid activation function is then applied after the output layer to normalize the score to the 0-1 range. The mapping from the node hidden state vector to the score can be expressed as:

[0103]

[0104] in: This represents the normalized node fitness score. Represents the hidden state vector of a node. and This represents the weights and biases of the first fully connected layer. This represents the activation function following the first fully connected layer. and This represents the weights and biases of the second fully connected layer. This represents the sigmoid activation function.

[0105] In practice, a model training process is implemented, and a labeled dataset of historical UAV scheduling tasks is collected. This labeled dataset includes ground truth labels for node fitness generated by human experts. The mean squared error loss function is used to measure the difference between the model's predicted score and the ground truth labels. The mean squared error loss function is defined as follows:

[0106]

[0107] in: Indicates the loss value. This indicates the number of samples in a batch. Indicates the first The ground truth label of the node fit of each sample. The graph neural network model represents the first... The predicted score for each sample is used to iteratively update the network weight parameters using the backpropagation algorithm combined with the Adam optimizer until the loss on the validation set converges to below a preset threshold. The trained network parameter file is then saved as a pre-trained graph neural network model. In some embodiments, the preset threshold is set to 0.001, and the initial learning rate of the Adam optimizer is set to 0.001.

[0108] See Figure 4 The horizontal axis represents the training epochs, and the vertical axis represents the loss value measured by mean squared error (MSE). The solid dotted line represents the training loss, the solid box line represents the validation loss, and the dashed line represents the preset convergence threshold of 0.15. Training loss: The training loss shows a continuous and stable decreasing trend with increasing training epochs, gradually decreasing from approximately 0.77 initially to 0.11 in the 20th epoch, and finally falling below the preset convergence threshold of 0.15. This indicates that the model can effectively learn node features such as task density and risk weights of grid cells in the airspace situation map on the training set. Through the message passing mechanism of the graph convolutional layer and the dynamic weight allocation of the graph attention mechanism, it gradually optimizes the predictive ability of UAV node fit scores, which is consistent with the training expectations of the mean squared error loss function and the Adam optimizer. Validation Loss: The validation loss decreased synchronously throughout, gradually decreasing from approximately 0.87 initially to 0.27 in round 20. This consistent downward trend mirrored the training loss, with no significant rebound in validation loss due to overfitting. This indicates the model possesses good generalization ability, effectively transferring the feature mapping patterns learned from the training set to unseen validation data, thus adapting to the complex and ever-changing airspace conditions in real-world drone dispatch scenarios. Convergence Threshold: A preset convergence threshold of 0.15 was used as the criterion for terminating model training. The training loss first fell below this threshold in round 19 and further decreased to 0.11 in round 20, satisfying the training termination condition of "until the loss on the validation set converges below the preset threshold." At this point, iteration could be stopped, network parameters saved, and the pre-training of the graph neural network model completed, providing reliable node fit scoring support for subsequent drone task assignment.

[0109] In one embodiment of the present invention, the method further includes a dynamic rescheduling trigger mechanism. During the execution of the final UAV swarm collaborative operation scheme, real-time environmental feedback data from a multimodal sensor array is continuously monitored. The real-time environmental feedback data is compared with a preset safety threshold. If a sudden obstacle intrusion or a change in the mission target state is detected, a rescheduling trigger signal is generated. In response to this rescheduling trigger signal, the currently executing scheduling process is interrupted, and the completed portion of the mission state is frozen. The latest environmental perception data stream and mission command stream are reacquired, and the process jumps to the step of parsing heterogeneous data frames, generating a new adaptive spatiotemporal raster matrix based on the latest data. Based on the new adaptive spatiotemporal raster matrix, the global situation quantization table and node adaptability score are recalculated, and incremental scheduling adjustment commands are generated to replace the original commands and are issued for execution.

[0110] In practice, the dynamic rescheduling triggering mechanism runs continuously during the execution of the final UAV swarm collaborative operation plan. It continuously monitors real-time environmental feedback data from the multimodal sensor array, including updated radar point clouds, visible light imagery, and aircraft status telemetry data. This real-time environmental feedback data is compared with preset safety thresholds. In some embodiments, the safety threshold includes a minimum distance threshold between obstacles and the UAV. And the threshold for the flag indicating a change in the mission objective state. If a sudden obstacle intrusion or a change in the mission objective state is detected, a rescheduling trigger signal is generated. It can be understood that the condition for determining a sudden obstacle intrusion is that the minimum spatial distance between the real-time sensed obstacle coordinates and any UAV's planned flight path is less than... The condition for determining a change in the status of a mission objective is that the objective validity flag in the received instruction update message changes.

[0111] In specific implementation, in response to the rescheduling trigger signal, the system interrupts the currently executing scheduling process, freezes the completed task status, and optionally, the task status freezing includes recording the reached path points, completed task actions, and the remaining energy and payload status of each UAV unit. It then re-acquires the latest environmental perception data stream and task command stream, jumps to the step of parsing heterogeneous data frames, generates a new adaptive spatiotemporal raster matrix based on the latest data, and recalculates the global situation quantization table and node adaptability score based on the new adaptive spatiotemporal raster matrix. In some embodiments, the recalculation process only performs incremental calculations on the spatial regions that have changed since the last scheduling and related tasks and UAV units, generating incremental scheduling adjustment commands to replace the original commands for execution. The incremental scheduling adjustment command generation process can be described as follows: Let the original scheduling command set be... The newly calculated scheduling instruction set is Then the incremental adjustment instruction is:

[0112]

[0113] Where: symbol This represents the set difference operation, i.e. Only includes with For different, newly added, or modified instruction parts, it can be understood that this method only transmits and updates the changed instruction parts, reducing communication overhead and processing latency.

[0114] See Figure 5Based on 20 spatiotemporal grid units, the system is divided into four statistical intervals (grids 1-5, 6-10, 11-15, and 16-20) with five grid units per group. The vertical axis represents the node fit score normalized to the 0-1 interval, which indicates the degree of task execution matching between the current real-time spatial location of each UAV unit in the swarm to be scheduled and the adaptive spatiotemporal grid unit. In the derivation logic of the graph neural network model, each row of grid unit data in the global situation quantization table is converted into graph nodes. The node feature vector contains task density value, risk weight value, and grid unit geographic location encoding. After constructing a spatial situation map based on the spatial connectivity of adjacent grid units, the system performs forward propagation calculations through graph convolutional layers, graph attention mechanism modules, and fully connected output layers to finally output the node fit score corresponding to each grid unit. Figure 5 The distribution characteristics show that: the overall fitness scores of grid groups 11-15 are at the highest level, with a median of approximately 0.88 and interquartile range concentrated in the range of 0.80-0.91, indicating that the task density and risk weight combination of this group of grid cells is optimal, and the corresponding airspace is a high-fit area for UAV mission execution; the median score of grid groups 6-10 is approximately 0.64, with an interquartile range of 0.50-0.66, which is a medium fitness level; the median score of grid groups 1-5 is approximately 0.47, with an interquartile range of 0.38-0.62, which is the second highest fitness level; the overall scores of grid groups 16-20 are the lowest, with a median of approximately 0.43 and an interquartile range of 0.36-0.51, and there are a few high-scoring separation clusters, reflecting that the airspace environment or task load conditions of this group of grid cells are poor, and only local locations have mission execution fitness. The scoring distribution provides a quantitative decision-making basis for subsequent task assignment: the system uses the Hungarian algorithm based on the scoring matrix to solve the optimal matching relationship between UAV units and task target points, prioritizing the assignment of high-priority tasks to UAV units corresponding to high-scoring adaptation areas such as grid 11-15. Simultaneously, it plans flight paths by combining the risk weight values ​​from the global situation quantification table, ultimately generating a UAV swarm collaborative operation scheme without path conflicts. This distribution feature can also serve as a validation basis for the training effect of the graph neural network model. The concentration and hierarchical rationality between high-scoring partitions verify the model's effective learning of airspace situation characteristics and the accuracy of its adaptation assessment.

[0115] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. An intelligent scheduling and management method for unmanned aerial vehicle (UAV) missions based on artificial intelligence, characterized in that: include: Collect environmental perception data streams and task command streams around the drone swarm to be scheduled, and convert the environmental perception data streams and task command streams into heterogeneous data frames with a unified timestamp. The heterogeneous data frames are parsed to extract spatial obstacle distribution information and task urgency information, which are then input into the spatiotemporal grid construction engine. The spatiotemporal grid construction engine dynamically adjusts the grid resolution based on the airspace obstacle distribution information to generate an adaptive spatiotemporal grid matrix covering the operation area of ​​the drone swarm to be scheduled. Traverse the adaptive spatiotemporal grid matrix, calculate the task density value and risk weight value in each grid cell, and summarize them to form a global situation quantification table; Read the global situation quantification table and use a graph neural network model to deduce the node adaptability score of each UAV unit in the UAV swarm to be scheduled. Based on the node adaptability score, tasks are assigned to the drone swarm to be scheduled, and an initial scheduling instruction set containing flight path points and execution order is generated. Verify the path conflicts in the initial scheduling instruction set, eliminate the conflicts, and output the final UAV swarm collaborative operation scheme.

2. The AI-based intelligent scheduling and management method for unmanned aerial vehicle (UAV) missions as described in claim 1, characterized in that, Collect environmental perception data streams and task command streams around the drone swarm to be scheduled, and convert the environmental perception data streams and task command streams into heterogeneous data frames with a unified timestamp, including: The multimodal sensor array deployed on the airborne end of the unmanned aerial vehicle swarm to be scheduled is activated. The multimodal sensor array simultaneously captures visible light image data, infrared thermal imaging data and radar point cloud data, which are then converged to form the environmental perception data stream. Receive task dispatch messages from remote command center, parse the target coordinate sequence and payload operation parameters in the task dispatch messages, and organize them into the task instruction stream; A high-precision clock synchronization mechanism is established to assign an absolute time tag to each frame of sensing data in the environmental sensing data stream and each instruction in the task instruction stream. The sensed data and instruction data with the absolute time stamp are encapsulated into a fixed-length data packet. The data packet contains a data source identifier field, a data type field, and a payload field. The payload field stores the original sampled value, thereby forming the heterogeneous data frame.

3. The AI-based intelligent scheduling and management method for unmanned aerial vehicle (UAV) missions as described in claim 2, characterized in that, Parsing the heterogeneous data frames, extracting spatial obstacle distribution information and task urgency information, and inputting the spatial obstacle distribution information and task urgency information into the spatiotemporal raster construction engine, including: Read the data type field in the heterogeneous data frame. If the data type field indicates that it is sensing data, call the three-dimensional reconstruction algorithm to process the radar point cloud data in the payload field, identify the geometric contours and spatial coordinates of airborne obstacles, and generate the airspace obstacle distribution information. If the data type field indicates instruction data, then the target coordinate sequence in the payload field is parsed, the task creation time and execution deadline are extracted, the reciprocal of the remaining time window is calculated as the basic urgency, and weighted in combination with the task type weight in the payload operation parameters to generate the task urgency information. Align the obstacle coordinate set in the airspace obstacle distribution information with the task coordinates and urgency value in the task urgency information according to the absolute time label; The time-aligned obstacle coordinates set, task coordinates, and urgency value are packaged into an engine input structure. The header of the engine input structure contains the current timestamp, and the main body contains an obstacle list and a task list. The engine input structure is then sent to the receiving interface of the spatiotemporal grid construction engine.

4. The AI-based intelligent scheduling and management method for unmanned aerial vehicle (UAV) missions as described in claim 3, characterized in that, The spatiotemporal grid construction engine dynamically adjusts the grid resolution based on the airspace obstacle distribution information to generate an adaptive spatiotemporal grid matrix covering the operational area of ​​the drone swarm to be scheduled, including: The spatiotemporal grid construction engine reads the list of obstacles in the engine input structure and counts the number and density of obstacles per unit space volume. When the number density of obstacles exceeds a preset dense threshold, a high-resolution mode is triggered, reducing the physical size of the grid cells and increasing the number of subdivision layers in the grid hierarchy. When the number density of obstacles is lower than a preset sparse threshold, a low-resolution mode is triggered, increasing the physical size of the grid cells and reducing the number of subdivision layers in the grid hierarchy. Using the geographical center of the operation area of ​​the drone swarm to be scheduled as the origin, the latitude and longitude grid is divided according to the current physical size of the grid cell, and each grid cell is assigned a unique index code. The obstacle list and task list in the engine input structure are mapped to the corresponding grid cells, and the obstacle occupancy flag and task urgency accumulation value in each grid cell are recorded, thereby constructing the adaptive spatiotemporal grid matrix.

5. The AI-based intelligent scheduling and management method for unmanned aerial vehicle (UAV) missions as described in claim 4, characterized in that, The adaptive spatiotemporal grid matrix is ​​traversed, and the task density value and risk weight value within each grid cell are calculated. These values ​​are then summarized to form a global situation quantification table, including: The adaptive spatiotemporal grid matrix is ​​scanned row by row. For each grid cell with a non-empty task list, the total number of task entries in the task list is counted and divided by the physical area of ​​the grid cell to obtain the task density value. Read the obstacle occupancy flag within the grid cell. If the obstacle occupancy flag is true, query the pre-stored obstacle threat database to obtain the collision risk assessment coefficient of the obstacle. The risk weight value is obtained by multiplying the collision risk assessment coefficient by a preset environmental penalty factor and then combining it with the meteorological turbulence index of the airspace where the grid cell is located. A two-dimensional table structure is constructed as the global situation quantification table. The row index of the two-dimensional table structure corresponds to the index code of the raster cell, and the column fields include the task density value and the risk weight value. The task density value and risk weight value of each grid cell obtained by traversal calculation are filled into the corresponding row of the two-dimensional table structure to complete the filling of the global situation quantization table.

6. The AI-based intelligent scheduling and management method for unmanned aerial vehicle (UAV) missions as described in claim 5, characterized in that, Read the global situation quantization table, and use a graph neural network model to deduce the node adaptability score of each UAV unit in the UAV swarm to be scheduled, including: Load a pre-trained graph neural network model, wherein the input layer of the graph neural network model receives the grid cell node features from the global situation quantization table; Each row of data in the global situation quantization table is converted into a graph node, and the node feature vector contains the task density value, the risk weight value, and the geographic location code of the grid cell; Based on the spatial connectivity between adjacent grid cells, edge connection relationships are constructed to form the spatial situation map data structure to be processed; The spatial situation map data structure is input into the graph neural network model for forward propagation calculation. The graph neural network model updates the hidden state of the central node by aggregating neighbor node information. The node features corresponding to the current positions of each UAV unit in the UAV swarm to be scheduled are extracted from the output layer of the graph neural network model. After mapping transformation by a fully connected layer, the node fitness score is output. The node fitness score represents the degree of task execution matching of the adaptive spatiotemporal grid unit corresponding to the current real-time spatial position of each UAV unit in the UAV swarm to be scheduled.

7. The AI-based intelligent scheduling and management method for unmanned aerial vehicle (UAV) missions as described in claim 6, characterized in that, Based on the node adaptability score, tasks are assigned to the swarm of drones to be scheduled, generating an initial scheduling instruction set containing flight waypoints and execution order, including: Collect the node fitness scores of all UAV units in the unmanned aerial vehicle swarm to be scheduled at the output layer of the graph neural network model, and construct a score matrix. The Hungarian algorithm is used to process the scoring matrix, solve for the optimal matching relationship between the UAV unit and the task target point, and determine the target task list for each UAV unit. For each UAV unit's target task list, combined with the risk weight values ​​in the global situation quantification table, a heuristic search algorithm is used to plan a continuous trajectory from the current position to each task target; On the continuous flight path, locations with significant terrain variations and mission execution locations are selected as key nodes, and the geographic coordinates of the key nodes are extracted as the flight path points. According to the order of task execution and arrival time constraints, the order of the flight path points is arranged, and the UAV unit number, flight path point sequence and expected action are encapsulated into a scheduling record; The scheduling records of all UAV units are aggregated and encoded according to a unified instruction format to generate the initial scheduling instruction set.

8. The intelligent scheduling and management method for unmanned aerial vehicle (UAV) missions based on artificial intelligence as described in claim 7, characterized in that, Verify path conflicts in the initial scheduling instruction set, eliminate conflicts, and output the final UAV swarm collaborative operation scheme, including: Analyze each scheduling record in the initial scheduling instruction set and extract the flight path point sequence corresponding to the UAV unit number; A four-dimensional spatiotemporal trajectory model is constructed, and the flight path point sequence of each UAV unit is converted into a position coordinate function that varies with time. Check if there is an intersection between the spatiotemporal trajectories of any two drone units. If there is an intersection and the time interval is less than the safe avoidance threshold, it is determined to be a path conflict. For drone unit pairs that are determined to have path conflicts, adjust the flight path point sequence of at least one drone unit, specifically by shifting the takeoff time along the time axis or adding intermediate collision avoidance path points by detouring. Repeat the path conflict checking and adjustment steps until there are no path conflicts in the initial scheduling instruction set; All scheduling records after conflict resolution are repackaged, digital signatures and verification codes are added, and the final drone swarm collaborative operation scheme is generated and sent to the drone swarm to be scheduled.

9. The intelligent scheduling and management method for unmanned aerial vehicle (UAV) missions based on artificial intelligence as described in claim 8, characterized in that, The construction method of the graph neural network model includes: Define the graph structure input specification, set the node feature dimension of the spatial situation map to a fixed length, and the node feature dimension shall at least cover the three-dimensional feature vector of task density value, risk weight value and geographic location code. At the same time, define the adjacency matrix as an undirected edge connection relationship based on the spatial connectivity of grid cells. A multi-layer graph convolutional network architecture is constructed by stacking graph convolutional layers, batch normalization layers, and activation function layers in sequence. The graph convolutional layers use a message passing mechanism to aggregate neighbor node features, and the activation function layers use linear units with leakage correction to enhance the expressive power of sparse features. A graph attention mechanism module is configured and embedded after the graph convolutional layer. By calculating the attention coefficients of the feature vectors between nodes, the contribution weights of different neighboring nodes to the feature update of the central node are dynamically allocated. The calculation of the attention coefficients is based on the learnable linear transformation and softmax normalization operation between the feature vectors of the nodes. The model output mapping head is designed to input the node hidden state vector output by the last layer of graph convolutional network into two layers of fully connected neural network. The first fully connected layer compresses the feature dimension into the intermediate hidden layer space, and the second fully connected layer outputs a single value as the node fitness score. A sigmoid activation function is then applied after the output layer to normalize the score to the range of 0 to 1. The model training process involves collecting a labeled dataset of historical UAV scheduling tasks. The labeled dataset includes ground truth labels for node fitness generated by human experts. The mean squared error loss function is used to measure the difference between the model prediction score and the ground truth labels. The backpropagation algorithm combined with the Adam optimizer is used to iteratively update the network weight parameters until the loss on the validation set converges to below a preset threshold. The trained network parameter file is then saved as the pre-trained graph neural network model.

10. The AI-based intelligent scheduling and management method for unmanned aerial vehicle (UAV) missions as described in claim 9, characterized in that, It also includes a dynamic rescheduling trigger mechanism, specifically including: During the execution of the final drone swarm collaborative operation scheme, real-time environmental feedback data from the multimodal sensor array is continuously monitored. The real-time environmental feedback data is compared with a preset safety threshold. If a sudden obstacle intrusion or a change in the state of the mission target is detected, a rescheduling trigger signal is generated. In response to the rescheduling trigger signal, the currently executing scheduling process is interrupted, and the status of the completed tasks is frozen. The latest environmental perception data stream and task instruction stream are re-acquired, and the process jumps to the step of parsing the heterogeneous data frame, generating a new adaptive spatiotemporal raster matrix based on the latest data. Based on the new adaptive spatiotemporal grid matrix, the global situation quantization table and node adaptability score are recalculated to generate incremental scheduling adjustment instructions, which replace the original instructions and are then issued for execution.

Images