A multi-unmanned aerial vehicle adaptive migration method for sudden no-fly zones

Abstract

The application provides a multi-unmanned aerial vehicle adaptive migration method for a sudden no-fly zone, and relates to the technical field of unmanned aerial vehicles. Fusion environment data is acquired through multi-unmanned aerial vehicle cooperative sensing, and a semantic boundary extraction algorithm based on deep learning is used to process the data to output a sudden no-fly zone map and unmanned aerial vehicle safety state coding. On this basis, the unmanned aerial vehicles are shunted and disposed according to the safety state coding and the sudden no-fly zone map, real-time obstacle avoidance trajectories are generated for unmanned aerial vehicles that have not intruded, and cluster cooperation is performed for unmanned aerial vehicles that have intruded. For unmanned aerial vehicles that have interrupted the task due to obstacle avoidance or escape, the application realizes task decomposition based on the sudden no-fly zone map and the space-time isolation degree, and realizes the continuous execution of cluster tasks through an energy consumption constraint auction mechanism, so that multiple unmanned aerial vehicles can better cope with complex and variable sudden no-fly zone environments.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) technology, and more specifically, to a multi-UAV adaptive migration method for sudden no-fly zones. Background Technology

[0002] In low-altitude operations in urban and jungle environments, unmanned aerial vehicle (UAV) systems face increasingly frequent challenges from unforeseen no-fly zones. These no-fly zones are typically temporary, established due to unforeseen events, posing a significant challenge to UAV systems performing missions, necessitating immediate course adjustments and mission interruptions.

[0003] Currently, the autonomous operation rate of unmanned aerial vehicle (UAV) systems is continuously increasing, but traditional UAV management methods mainly rely on manual notification and statically preset flight paths. This approach is difficult to meet the real-time response requirements when faced with sudden changes in airspace. When a no-fly zone suddenly appears, the UAV system often fails to obtain no-fly zone information in time and make effective track adjustments, causing the UAV to accidentally enter the no-fly zone, leading to safety accidents or mission failures.

[0004] Furthermore, in scenarios involving multiple drones operating collaboratively, a lack of dynamic coordination and adjustment capabilities at the task level can lead to a large-scale chain reaction of conflicts or even the failure of the entire swarm mission if a no-fly zone is suddenly established. In existing technologies, most path planning systems rely on predefined maps and centralized control, making it difficult to perform rapid local adjustments and path redeployment after a no-fly zone is temporarily set. This prevents multi-drone systems from achieving self-organizing capabilities—uninterrupted missions, no platform conflicts, and unrestricted responses—when faced with sudden no-fly zones.

[0005] There is currently no effective technical solution to the above problems. Summary of the Invention

[0006] The purpose of this application is to provide a multi-UAV adaptive migration method for sudden no-fly zones. When a sudden no-fly zone appears, multiple UAVs can autonomously perceive, make intelligent decisions, and adaptively migrate, while ensuring the continuous execution of the mission. This effectively solves the problems of delayed response, mission interruption, and cluster conflict in UAV systems when facing sudden no-fly zones in the prior art.

[0007] This application provides a multi-UAV adaptive migration method for sudden no-fly zones, including the following steps:

[0008] S1. Acquire fused environmental data through collaborative sensing of multiple drones;

[0009] S2. A deep learning-based semantic boundary extraction algorithm is used to process the fused environmental data to output a map of sudden no-fly zones and safety status codes for each drone.

[0010] S3. Based on the security status code and the map of the sudden no-fly zone, the drones are diverted and handled. Real-time obstacle avoidance trajectories are generated for drones that have not entered the no-fly zone, and drones that have entered the no-fly zone are drones that have entered the no-fly zone.

[0011] S4. For drones whose missions are interrupted due to obstacle avoidance or separation, mission decomposition is achieved based on the sudden no-fly zone map and spatiotemporal isolation, and the continuous execution of cluster missions is achieved through an energy consumption constraint auction mechanism.

[0012] Through this technical solution, this application enables multiple UAVs to autonomously perceive, make intelligent decisions, and adaptively migrate when a sudden no-fly zone appears, while ensuring the continuous execution of the mission. This effectively solves the problems of delayed response, mission interruption, and cluster conflict in existing UAV systems when facing sudden no-fly zones.

[0013] Optionally, step S1 includes:

[0014] Environmental data is collected in real time using lidar, vision cameras and millimeter-wave radar sensors on each drone.

[0015] The environmental data acquired by each drone is fused to obtain fused environmental data.

[0016] This technical solution enables the acquisition of more comprehensive and accurate environmental data through multi-sensor fusion.

[0017] Optionally, the deep learning-based semantic boundary extraction algorithm employs the U-Net model, which incorporates a dynamic boundary attention mechanism; step S2 includes:

[0018] S201. Receive fused environmental data, which includes a point cloud projection map reflecting distance information output by the LiDAR and an image reflecting texture and contour information output by the visual camera. Stitch the point cloud projection map and the image together by channel to form an input feature map with a size of H×W×3, where H is the image height and W is the image width.

[0019] S202. Five sequentially connected encoding stages are used to downsample and extract features from the input feature map layer by layer. Each encoding stage includes a convolutional layer, a batch normalization layer and a ReLU activation function, and residual connections are introduced. Finally, five sets of multi-scale feature maps are output.

[0020] S203. Based on the real-time shortest Euclidean distance between the UAV and the sudden no-fly zone, a spatial attention submodule is used to focus on the boundary of the sudden no-fly zone, generating real-time safety status codes for each UAV. Simultaneously, multi-scale feature maps are weighted and enhanced to obtain spatially enhanced features. ;

[0021] S204. Based on the UAV's safety status code, dynamically allocate weights for the LiDAR and visual channels, and use these weights to weightedly fuse the features of the LiDAR and visual channels to obtain channel fusion features. ; Enhance spatial features Features of channel fusion The concatenated outputs yield a comprehensive feature enhanced by dynamic attention. ;

[0022] S205. The segmentation mask is recovered from the comprehensive features using the decoding structure of the U-Net model, and a final feature map of the same size as the input feature map is output.

[0023] S206. Binarize the final feature map using the Sigmoid activation function, output the sudden no-fly zone semantic segmentation mask, and obtain the sudden no-fly zone map.

[0024] This technical solution, by introducing a U-Net model with a dynamic boundary attention mechanism, can more accurately identify the semantic boundaries of sudden no-fly zones and generate UAV safety status codes in real time, thereby improving the accuracy of no-fly zone identification and the timeliness of UAV response.

[0025] Optionally, after step S206, step S207 is also included:

[0026] A hybrid loss function is used to balance the classification accuracy of the security state coding with the consistency of the boundary areas of the sudden no-fly zone map. The expression of the hybrid loss function is as follows:

[0027] ;

[0028] in, ;

[0029] in, ;

[0030] in, ;

[0031] In the formula, The value of the loss function. These are the weighting coefficients. For cross-entropy loss, This is a loss for Dice. For pixels Authentic labels Predicting pixels in feature maps for the U-Net model The probability of being within the boundary of a sudden no-fly zone. Total number of pixels This is the smoothing coefficient.

[0032] This technical solution effectively balances the optimization of the model in terms of classification accuracy and consistency of no-fly zone boundary areas, resulting in no-fly zone maps that are not only accurately classified but also have smoother and more precise boundaries.

[0033] Optionally, step S203 includes: calculating the real-time shortest Euclidean distance between the UAV and the no-fly zone according to the following formula:

[0034] ;

[0035] In the formula, This represents the real-time shortest Euclidean distance between the drone and the sudden no-fly zone. This indicates traversing all from Find the minimum value among the results. Indicates the center of mass of the drone The first time to reach the sudden no-fly zone Boundary line segment The shortest Euclidean distance, if the drone's center of mass The vertical projection point falls on the first [number]th ... On the boundary line segment, then For the center of mass of the drone The straight-line distance to the projection point, if the projection point is not in the emergency no-fly zone, is the distance to the projection point. On the boundary line segment, then For the center of mass of the drone The first time to reach the sudden no-fly zone The minimum distance between the two ends of the boundary line segment;

[0036] when m, the drone's safety status code is 0;

[0037] when m, the drone's safety status code is 1;

[0038] when The safety status code for the drone is 2;

[0039] Normalize the real-time shortest Euclidean distance between the drone and the sudden no-fly zone to obtain the value of each pixel in the feature map. Spatial correlation with the boundary of the emergency no-fly zone ;

[0040] Calculate the spatial attention weights using the following formula:

[0041] ;

[0042] In the formula, Spatial attention weights, For the Sigmoid function, and All are learning parameters. Distance adjustment factor;

[0043] The spatially enhanced features are obtained by weighting the feature map according to the following formula:

[0044] ;

[0045] In the formula, For pixels Spatial enhancement features, For pixels The original eigenvalues.

[0046] This technical solution enables precise calculation of the real-time shortest Euclidean distance between the drone and the sudden no-fly zone, and, combined with a spatial attention mechanism, achieves refined focusing and feature enhancement of the no-fly zone boundary, thereby more accurately determining the safety status of the drone.

[0047] Optionally, step S3 includes:

[0048] The drones are handled according to their security status codes;

[0049] Based on the map of sudden no-fly zones, for unintruded drones, a real-time obstacle avoidance trajectory is generated using a pre-trained environment-adaptive motion primitive library. The environment-adaptive motion primitive library contains maneuvering flight templates and trajectory planning parameters for different types of sudden no-fly zones.

[0050] Based on the map of the sudden no-fly zone, an emergency disengagement mechanism is adopted for drones that have already intruded, enabling the drone swarm to disengage in a coordinated manner.

[0051] This technical solution enables precise triage of drones based on their safety status and maps of sudden no-fly zones. It provides real-time obstacle avoidance trajectories for drones that have not intruded and emergency escape mechanisms for drones that have intruded, thereby effectively ensuring the flight safety of drones.

[0052] Optionally, emergency evacuation mechanisms include:

[0053] A301. Immediately terminates the original mission and hovers the compromised drone, while transmitting a standardized alarm package through the cluster communication network. The standardized alarm package contains the real-time GPS location, remaining battery power, sensor data and mission priority of the compromised drone.

[0054] A302. Based on the map of the sudden no-fly zone, determine the shortest path direction from the intruding drone to the boundary of the sudden no-fly zone through geometric analysis, so as to minimize the time spent crossing the sudden no-fly zone;

[0055] A303. Nearby drones actively adjust their trajectories based on the shortest path direction, standardized alarm packages, and pre-built conflict prediction models to form an emergency passage and ensure the safe passage of intruding drones.

[0056] A304. Compare the remaining battery power of the intruded drone with its evacuation energy consumption. If the remaining battery power is greater than the evacuation energy consumption, evacuate in a straight line along the emergency exit to a safe point; if the remaining battery power is less than the evacuation energy consumption, perform a vertical descent and mark the crash site.

[0057] Optionally, in step S4, the steps of decomposing the task based on the sudden no-fly zone map and spatiotemporal isolation include:

[0058] The interrupted original task is split into multiple subtasks based on spatial continuity or temporal stages.

[0059] Spatial continuity splitting: The region corresponding to the complete original task is split into multiple non-overlapping sub-regions, each sub-region corresponds to a spatial sub-task, and the sub-regions corresponding to all spatial sub-tasks are spliced ​​together to equal the region corresponding to the complete original task.

[0060] Segment by time period: When the original task needs to be completed within a preset time period, the original task area remains unchanged, and the preset time period is divided into multiple time windows. Each time window corresponds to a time sub-task, and the sum of the time windows of all time sub-tasks is equal to the preset time period.

[0061] Based on the map of sudden no-fly zones, feasible sub-task sets are selected by spatiotemporal isolation, which is calculated according to the following formula:

[0062] Spatiotemporal isolation = •Time isolation+ • Spatial isolation;

[0063] Temporal isolation is the percentage of non-overlapping time windows between any two subtasks, calculated using the following formula:

[0064] Time isolation ;

[0065] In the formula, These are the weighting coefficients. This is the start time of the first subtask. This is the end time of the first subtask. This is the start time for the second subtask. This is the end time for the second subtask;

[0066] Spatial isolation is the percentage of non-overlapping execution regions between any two subtasks, calculated using the following formula:

[0067] Spatial isolation ;

[0068] In the formula, This represents the area of ​​the execution region for the first subtask. The area of ​​the execution region for the second subtask;

[0069] Determine the set of retained subtasks based on spatiotemporal isolation. Calculated according to the following formula:

[0070] ;

[0071] In the formula, For a set of sub-tasks, This refers to two subtasks The spatiotemporal isolation degree is obtained by weighting the temporal isolation degree and the spatial isolation degree. For isolation value, For the first subtask, This is the second subtask.

[0072] Optionally, in step S4, the steps for achieving continuous execution of cluster tasks through an energy-constrained auction mechanism include:

[0073] The drone that interrupted the task broadcasts key parameters of the subtask to the cluster. These key parameters include the execution area boundary coordinates and time window requirements. , ] , Estimated energy consumption demand Task priority coefficient ;

[0074] Other drones in the cluster calculate their bidding scores using the following formula, and the one with the highest score takes over the subtask:

[0075] ;

[0076] In the formula, To compete for the remaining battery power consumption of drones, To estimate the delay time for completing sub-tasks in order to bid for drones, This is the maximum allowable delay threshold;

[0077] The publisher assigns subtasks to the highest-scoring drone and generates a takeover trajectory certificate. The takeover trajectory certificate includes: handover timestamp, subtask parameters, identification of the drones before and after the task, and energy consumption commitment value. The subtask parameters include the sub-region and time window corresponding to the subtask. The takeover trajectory certificate is stored through a distributed blockchain.

[0078] Optionally, after step S4, step S5 is also included:

[0079] S5. Each drone collects all data from each obstacle avoidance and escape event to form a dataset for each drone, and continuously optimizes the environment-adaptive motion primitive library through a federated learning framework.

[0080] As can be seen from the above, the multi-UAV adaptive migration method for sudden no-fly zones provided in this application acquires fused environmental data through multi-UAV collaborative perception and processes the data using a semantic boundary extraction algorithm based on deep learning. This enables real-time output of sudden no-fly zone maps and UAV safety status codes, thus solving the problem in existing technologies where UAV systems struggle to obtain sudden no-fly zone information and effectively adjust their flight paths. Based on this, UAVs are triaged according to their safety status codes. Real-time obstacle avoidance trajectories are generated for non-intruding UAVs, while swarm collaborative disengagement is performed for intruding UAVs, effectively preventing safety accidents or mission failures caused by UAVs mistakenly entering sudden no-fly zones. Furthermore, for UAVs whose missions are interrupted due to obstacle avoidance or disengagement, task decomposition is achieved based on spatiotemporal isolation, and continuous execution of swarm tasks is achieved through an energy consumption constraint auction mechanism. This overcomes the shortcomings of existing multi-UAV systems in dealing with mission interruptions and platform conflicts when facing sudden no-fly zones. Finally, by collecting obstacle avoidance and escape event data and continuously optimizing the group decision-making strategy using a federated learning framework, the autonomous response and adaptive reconstruction of the task path of the multi-UAV system were realized, and the task was continuously executed under the dynamic changes of sudden no-fly zones. This effectively solved the technical problem of the lack of dynamic collaborative adjustment capability in the existing technology, which led to the failure of the entire cluster task, and significantly improved the safety, reliability and mission continuity of the UAV system.

[0081] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing embodiments of this application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings. Attached Figure Description

[0082] Figure 1 This is a flowchart illustrating the adaptive migration method for multiple unmanned aerial vehicles (UAVs) in response to sudden no-fly zones, as provided in an embodiment of this application. Detailed Implementation

[0083] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0084] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0085] Please refer to Figure 1 This application provides a method for adaptive migration of multiple unmanned aerial vehicles (UAVs) in the event of a sudden no-fly zone, including the following steps:

[0086] S1. Acquire fused environmental data through collaborative sensing of multiple drones;

[0087] S2. A deep learning-based semantic boundary extraction algorithm is used to process the fused environmental data to output a map of sudden no-fly zones and safety status codes for each drone.

[0088] S3. Based on the security status code and the map of the sudden no-fly zone, the drones are diverted and handled. Real-time obstacle avoidance trajectories are generated for drones that have not entered the no-fly zone, and drones that have entered the no-fly zone are drones that have entered the no-fly zone.

[0089] S4. For drones whose missions are interrupted due to obstacle avoidance or separation, mission decomposition is achieved based on the sudden no-fly zone map and spatiotemporal isolation, and the continuous execution of cluster missions is achieved through an energy consumption constraint auction mechanism.

[0090] "Sudden No-Fly Zones" refer to restricted airspace temporarily established due to unforeseen events, characterized by high dynamism and short warning times. "UAV Safety Status Coding" is an identifier used to quantify the safe distance or intrusion status between a UAV and a no-fly zone. For example, different coding values ​​can be set to indicate whether a UAV is in a safe area, approaching the no-fly zone boundary, or has intruded into the no-fly zone. "Fused Environmental Data" refers to environmental information collected collaboratively by sensors on multiple UAVs and then processed and integrated.

[0091] First, in step S1, multiple drones collaboratively perceive and acquire fused environmental data. The purpose of this step is to comprehensively and accurately obtain real-time information about the environment in which the drones are located. For example, each drone can independently carry multiple sensors, such as visual cameras, LiDAR, and millimeter-wave radar sensors, to collect environmental data within its respective field of view. This environmental data can be images, point clouds, distance information, etc. To achieve collaborative perception, wireless communication links can be established between the drones to transmit their raw environmental data to a central processing unit or perform distributed fusion within the cluster. The fusion processing can employ various technologies. For example, data from different sensors can be synchronized in time and registered spatially, and then heterogeneous data can be integrated into unified fused environmental data through data fusion algorithms (such as Kalman filtering, extended Kalman filtering, particle filtering, etc., all of which are existing technologies and will not be detailed here).

[0092] Secondly, in step S2, a deep learning-based semantic boundary extraction algorithm is used to process the fused environmental data to output a sudden no-fly zone map and safety status codes for each drone. This step is one of the core aspects of this application, aiming to achieve automatic identification of sudden no-fly zones and real-time judgment of drone status. For example, a U-Net model with a dynamic boundary attention mechanism is introduced to identify the semantic boundaries of sudden no-fly zones from the fused environmental data. The model receives the fused environmental data as input, performs multi-layer feature extraction and semantic segmentation, and finally outputs a binary map representing the boundary of the sudden no-fly zone, i.e., the sudden no-fly zone map. Simultaneously, the U-Net model can also calculate and output the safety status codes for each drone in real time based on the relative positional relationship between the drone and the identified no-fly zone boundary. For example, the safety status code can be defined based on the shortest distance between the drone and the sudden no-fly zone boundary: a greater distance is encoded as safe, a closer distance as a warning, and entering the sudden no-fly zone as an intrusion.

[0093] Next, in step S3, drones are triaged based on safety status codes and sudden no-fly zone maps. Real-time obstacle avoidance trajectories are generated for non-intruding drones, while swarm-coordinated disengagement is performed for intruding drones. This step aims to adopt differentiated response strategies based on the real-time status of the drones. Swarm-coordinated disengagement means that other drones will also adjust their trajectories accordingly based on the disengagement path of the intruding drones to avoid collisions and provide a safe disengagement route for the intruding drones.

[0094] Finally, in step S4, for drones whose missions are interrupted due to obstacle avoidance or escape, task decomposition is performed based on the sudden no-fly zone map and spatiotemporal isolation. The energy consumption constraint auction mechanism ensures the continued execution of cluster tasks. This step aims to solve the problem of drone mission interruption caused by sudden no-fly zones, ensuring the overall continuity of cluster tasks. For example, when a drone cannot continue its original mission due to obstacle avoidance or emergency escape, its mission is marked as interrupted. Based on the sudden no-fly zone map, the regional and temporal requirements of the interrupted mission are analyzed and decomposed into multiple sub-tasks. Task decomposition can be based on spatial continuity (decomposing large-area tasks into small-area sub-tasks) or temporal phases (decomposing long-term tasks into short-term sub-tasks). Subsequently, the spatiotemporal isolation between these sub-tasks is calculated, and a feasible set of sub-tasks is selected. Spatiotemporal isolation measures the degree of overlap between different sub-tasks in time and space to avoid task conflicts. For the selected sub-tasks, the energy consumption constraint auction mechanism is initiated. The drone with the interrupted mission, as the publisher, broadcasts the key parameters of the sub-task (such as execution area, time window, estimated energy consumption requirement, task priority, etc.) to other drones in the cluster. Other drones, acting as bidders, calculate their bidding scores based on factors such as their remaining energy consumption, mission load, and estimated delay time for completing sub-tasks. The drone with the highest score will take over the sub-task and generate a takeover trajectory certificate to ensure smooth handover and continuous execution of the mission.

[0095] In some implementations, step S1 includes:

[0096] Environmental data is collected in real time using lidar, vision cameras and millimeter-wave radar sensors on each drone.

[0097] The environmental data acquired by each drone is fused to obtain fused environmental data.

[0098] In practical applications, each drone can independently collect environmental data from its surroundings in real time. This heterogeneous environmental data, acquired from different sensors, is then transmitted to a data processing module, such as the drone's onboard processor or a ground control station, for fusion processing. This fusion processing can employ various techniques, such as Kalman filtering, extended Kalman filtering, or deep learning-based methods, to align and integrate data from different sensors in time and space, generating a unified, comprehensive, and accurate fused data set.

[0099] The solution presented in this application is able to more effectively acquire and fuse environmental data to improve the accuracy and real-time performance of identifying sudden no-fly zones because it achieves multi-dimensional and multi-modal environmental perception by configuring LiDAR, visual cameras, and millimeter-wave radar sensors on various UAVs. LiDAR provides precise distance information, visual cameras provide rich texture and contour information, and millimeter-wave radar provides reliable data under adverse weather conditions. It is precisely the synergistic effect of these heterogeneous sensor data that enables the system to overcome the limitations of single sensors in complex environments; for example, LiDAR is susceptible to rain and fog, and visual cameras experience performance degradation in low light conditions. By fusing this multi-source data, complementary and redundant environmental information can be obtained, thereby significantly enhancing the comprehensiveness and robustness of environmental perception.

[0100] In some implementations, the deep learning-based semantic boundary extraction algorithm employs the U-Net model, which incorporates a dynamic boundary attention mechanism; step S2 includes:

[0101] S201. Receive fused environmental data, which includes a point cloud projection map reflecting distance information output by the LiDAR and an image reflecting texture and contour information output by the visual camera. Stitch the point cloud projection map and the image together by channel to form an input feature map with a size of H×W×3, where H is the image height and W is the image width.

[0102] S202. Employs five sequentially connected encoding stages to downsample and extract features from the input feature map layer by layer. Each encoding stage includes a convolutional layer, a batch normalization layer, and a ReLU activation function, and introduces residual connections. Finally, the output... There are a total of 5 sets of multi-scale feature maps, among which, The resolution is H / 2×W / 2, the number of channels is 64, and it focuses on low-level detail features (such as sudden no-fly zone boundary textures). The resolution is H / 4×W / 4, the number of channels is 128, and it focuses on low- to mid-level combined features (such as continuous line segments at the boundary of sudden no-fly zones). The resolution is H / 8×W / 8, the number of channels is 256, and it focuses on mid-order structural features (such as the outline of local areas in sudden no-fly zones). The resolution is H / 16×W / 16, the number of channels is 512, and it focuses on mid-to-high-order semantic features (such as the approximate shape of sudden no-fly zones). The resolution is H / 32×W / 32, the number of channels is 1024, and the focus is on high-order global semantic features (such as the overall outline of sudden no-fly zones).

[0103] S203. Based on the real-time shortest Euclidean distance between the UAV and the no-fly zone, a spatial attention submodule is used to focus on the boundary of the no-fly zone, generating real-time safety status codes for each UAV. Simultaneously, multi-scale feature maps are weighted and enhanced to obtain spatially enhanced features. ;

[0104] S204. Based on the UAV's safety status code, dynamically allocate weights for the LiDAR and visual channels, and use these weights to weightedly fuse the features of the LiDAR and visual channels to obtain channel fusion features. , , The weights of the lidar channels, For the weights of the visual channel, Features of the lidar channel, Visual channel features; spatial enhancement features Features of channel fusion The output is concatenated (concatenation is a common feature fusion operation that involves concatenating features along a specified channel dimension) to obtain a comprehensive feature enhanced by dynamic attention. ;

[0105] S205. The comprehensive features are synthesized through the decoding structure of the U-Net model. Restore the segmentation mask and output a final feature map of the same size as the input feature map;

[0106] In step S205, the decoding structure of the U-Net model typically includes upsampling layers and convolutional layers, which can progressively recover the spatial resolution of the feature maps. Combined with skip connections in the encoding stage, it can also recover low-level detail information (such as...) Features in the middle) and higher-order semantic information (such as...) The features in the input feature map are combined to generate a fine-grained segmentation mask. Finally, a final feature map of the same size as the input feature map is output.

[0107] S206. Binarize the final feature map using the Sigmoid activation function, output the sudden no-fly zone semantic segmentation mask, and obtain the sudden no-fly zone map.

[0108] In step S206, the final feature map is binarized using the Sigmoid activation function. The Sigmoid activation function maps pixel values ​​in the final feature map to a range of 0 to 1. By setting a threshold, pixels can be classified as either no-fly zone boundaries or non-no-fly zone boundaries. This outputs a semantic segmentation mask for the no-fly zone, ultimately yielding a no-fly zone map. This map clearly indicates the precise boundaries of the no-fly zone, providing intuitive and accurate environmental information for subsequent UAV decision-making.

[0109] Specifically, the fused data includes a point cloud projection map reflecting distance information output by a LiDAR sensor, and an image reflecting texture and contour information output by a visual camera sensor. The point cloud projection map and the image are stitched together channel by channel to form an input feature map of size H×W×3. This stitching method allows the input feature map to simultaneously contain multi-dimensional information such as distance, texture, and contour, providing a rich data foundation for subsequent feature extraction.

[0110] This application's solution effectively addresses the problem of insufficient accuracy and real-time performance in no-fly zone boundary extraction caused by noise and redundant information in traditional methods when processing fused environmental data by introducing a U-Net model with a dynamic boundary attention mechanism. Specifically, firstly, in step S201, the point cloud projection map from the LiDAR and the image from the visual camera are stitched together by channel (for example, the point cloud projection map provides one depth channel, while the processed visual camera image provides two channels (e.g., one representing texture intensity and one representing contour information). Stacking these three channels forms an input feature map of size H×W×3. Here, H and W represent the height and width of the image, respectively, and "3" represents the number of feature channels ultimately input into the deep learning model), constructing an input feature map containing multi-dimensional information such as distance, texture, and contour. Secondly, step S202 utilizes a multi-layer encoding stage for feature extraction, ensuring the capture of complex features of the no-fly zone boundary from different scales. Furthermore, the core of this application lies in the dynamic boundary attention mechanism implemented in steps S203 and S204. In step S203, based on the real-time shortest Euclidean distance between the UAV and the no-fly zone, the spatial attention submodule focuses on the no-fly zone boundary. This allows the U-Net model to dynamically concentrate computational resources and attention on the most relevant boundary region according to the relative position of the UAV and the no-fly zone, significantly improving the accuracy of boundary recognition. Simultaneously, the real-time generated UAV safety status code provides immediate information for subsequent decisions. Subsequently, in step S204, the weights of the LiDAR and visual channels are dynamically allocated based on this safety status code, achieving adaptive weighted fusion of different sensor data sources. For example, when the safety status code is 0, the visual channel is emphasized, and the channel attention submodule allocates weights of 0.3 and 0.7 for the LiDAR and visual channels, respectively. When the safety status code is 1 or 2, the LiDAR channel is emphasized, and the channel attention submodule allocates weights of 0.7 and 0.3 for the LiDAR and visual channels, respectively. This dynamic adjustment allows the U-Net model to optimize feature representation according to the needs of the actual scenario, thereby enhancing robust perception of no-fly zone boundaries. Finally, the U-Net model's decoding structure (step S205) restores these dynamically attention-enhanced integrated features into a high-resolution segmentation mask, which is then binarized using a sigmoid activation function (step S206) to output an accurate no-fly zone semantic segmentation mask and generate a sudden no-fly zone map. This series of steps works synergistically to ensure the accuracy and real-time performance of boundary information during the conversion from the original fused data to the final no-fly zone map.

[0111] In some implementations, step S207 is included after step S206:

[0112] A hybrid loss function is used to balance the classification accuracy of the security state coding with the consistency of the boundary areas of the sudden no-fly zone map. The expression of the hybrid loss function is as follows:

[0113] ;

[0114] in, ;

[0115] in, ;

[0116] in, ;

[0117] In the formula, The value of the loss function. These are the weighting coefficients. For cross-entropy loss, This is a loss for Dice. For pixels Authentic labels Predicting pixels in feature maps for the U-Net model The probability of being within the boundary of a no-fly zone. Total number of pixels This is the smoothing coefficient.

[0118] Specifically, hybrid loss function This is a weighted sum of the cross-entropy loss and the Dice loss. The cross-entropy loss measures the probability that pixel i in the feature map predicted by the U-Net model belongs to the no-fly zone boundary. With pixels Authentic Labels The difference between the two methods aims to improve the classification accuracy of security state coding. Dice loss measures the overlap between the predicted results and the true labels, aiming to improve the consistency of boundary regions in sudden no-fly zone maps and reduce noise and holes in the segmentation mask. Weighting coefficients Used for balance and The contribution of the smoothing coefficient to the total loss function typically ranges from 0 to 1 and can be adjusted based on actual needs and model training performance. Setting it to 0.00001 is intended to avoid a denominator of zero and improve calculation stability. This is achieved through preset weighting coefficients. (For example, It can be set to 0.7, for example, when It is 0.2. When it is 0.1, = 0.7 * 0.2 + (1 - 0.7) * 0.1 = 0.14 + 0.03 = 0.17. This loss value This information is then used for backpropagation to update the weight parameters of the U-Net model. In a scenario where the U-Net model is trained to identify sudden no-fly zone boundaries and generate drone safety state codes, it is assumed that after one forward propagation, the model predicts a no-fly zone mask and a safety state code. A loss value L is calculated using a mixture loss function, for example, L = 0.17. This loss value L is then fed into the backpropagation algorithm. The backpropagation algorithm calculates the gradients of the weight parameters of all convolutional layers, batch normalized layers, and fully connected layers (if present) in the U-Net model based on the value of L. For example, the gradient of a convolutional kernel might be 0.01, indicating that the weights of that kernel need to be fine-tuned to reduce the loss. This gradient information is used to update the weight parameters of the U-Net model through an optimizer (such as the Adam optimizer or the SGD optimizer). Specifically, the optimizer adjusts each weight parameter according to the gradient direction and the learning rate, moving it in the direction that reduces the loss. For example, if the gradient of a weight parameter is positive, the optimizer will decrease the value of that weight parameter; if the gradient is negative, it will increase the value of that weight parameter. This process is repeated in each training iteration until the model converges or reaches the preset number of training epochs. In this way, the U-Net model is guided during training to simultaneously optimize classification accuracy and boundary region consistency. After sufficient training, the U-Net model can generate more clearly defined and more accurately classified maps of sudden no-fly zones.

[0119] In some implementations, step S203 includes: calculating the real-time shortest Euclidean distance between the UAV and the no-fly zone according to the following formula:

[0120] ;

[0121] In the formula, This represents the real-time shortest Euclidean distance between the drone and the sudden no-fly zone. This indicates traversing all from Find the minimum value among the results. This indicates the total number of boundary segments of the emergency no-fly zone. Indicates the center of mass of the drone The first time to reach the sudden no-fly zone Boundary line segment The Euclidean distance, if the center of mass of the drone The vertical projection point falls on the first [number]th ... On the boundary line segment, then For the center of mass of the drone The straight-line distance to the projection point, if the projection point is not in the emergency no-fly zone, is the distance to the projection point. On the boundary line segment, then For the center of mass of the drone The first time to reach the sudden no-fly zone The minimum distance between the two ends of the boundary line segment;

[0122] when m, the drone safety status code is 0; when the drone is more than 50 meters away from the boundary of the emergency no-fly zone, it means that the drone is far from the boundary of the no-fly zone and is in a relatively safe area;

[0123] when m, the drone safety status code is 1; when the drone is greater than 0 meters and less than or equal to 50 meters from the boundary of the emergency no-fly zone, it means that the drone has entered a warning area, is close to the boundary of the emergency no-fly zone, needs attention, and may need to take obstacle avoidance measures.

[0124] when The safety status code for drones is 2; when the drone is 0 meters away from the boundary of the emergency no-fly zone, it indicates the critical distance for the drone to enter the warning zone from the safe zone.

[0125] Normalize the real-time shortest Euclidean distance between the drone and the sudden no-fly zone to obtain the value of each pixel in the feature map. Spatial correlation with the boundary of the emergency no-fly zone ;

[0126] Calculate the spatial attention weights using the following formula:

[0127] ;

[0128] In the formula, Spatial attention weights, For the Sigmoid function, and All are learning parameters. Distance adjustment factor;

[0129] The spatially enhanced features are obtained by weighting the feature map according to the following formula:

[0130] ;

[0131] In the formula, For pixels Spatial enhancement features, For pixels The original eigenvalues.

[0132] The solution proposed in this application calculates the real-time shortest Euclidean distance between the drone and the no-fly zone using precise calculations. This provides a quantitative assessment of the safe distance between drones and no-fly zones. (By...) Normalized to spatial correlation, and dynamically generated spatial attention weights using a distance adjustment factor, this allows the deep learning model to adaptively focus more computational resources and attention on the no-fly zone boundary area. When a drone approaches the no-fly zone, The distance adjustment factor is reduced and the spatial attention weight is increased accordingly. As a result, when the original feature map is weighted and enhanced, the features of the no-fly zone boundary are significantly amplified. This effectively solves the problem that traditional methods are difficult to accurately and quickly determine the safe distance between the UAV and the no-fly zone in complex environments. It also ensures that this distance information can be effectively used to guide the UAV's obstacle avoidance behavior, thereby improving the UAV's perception accuracy and response speed to the no-fly zone boundary.

[0133] In some implementations, step S3 includes:

[0134] The drones are handled according to their security status codes;

[0135] "Diversion and handling" refers to classifying drones into different handling categories based on their security status codes, such as "non-intrusive drones" and "intrusive drones," and adopting differentiated response strategies for different categories.

[0136] When a drone is in a safe area (safety state code 0) or is approaching a no-fly zone (safety state code 1) but has not yet intruded, it is identified as a "non-intruding drone". At this time, based on the real-time acquired map of the sudden no-fly zone, a real-time obstacle avoidance trajectory is generated using a pre-trained "environment-adaptive motion primitive library".

[0137] When a drone is determined to have intruded into a sudden no-fly zone, it is identified as an "intruding drone." At this point, the "emergency disengagement mechanism" is immediately activated, causing the drone swarm to disengage in a coordinated manner.

[0138] Based on the map of sudden no-fly zones, for unintruded drones, a real-time obstacle avoidance trajectory is generated using a pre-trained environment-adaptive motion primitive library. The environment-adaptive motion primitive library contains maneuvering flight templates and trajectory planning parameters for different types of sudden no-fly zones.

[0139] Based on the map of the sudden no-fly zone, an emergency disengagement mechanism is adopted for drones that have already intruded, enabling the drone swarm to disengage in a coordinated manner.

[0140] The "Environment-Adaptive Motion Primitive Library" is a pre-built collection containing various maneuvering flight templates and trajectory planning parameters. "Maneuvering flight templates" refer to pre-defined flight modes or action sequences for specific environments or no-fly zone types, such as detours, climbs, and descents. "Track planning parameters" refer to numerical values ​​or rules used to adjust and optimize flight trajectories, such as speed limits, turning radii, and obstacle avoidance distances. The "environment-adaptive" characteristic of this primitive library is reflected in its ability to intelligently select or combine appropriate maneuvering flight templates and trajectory planning parameters based on the type, extent, and shape of the sudden no-fly zone as reflected in the sudden no-fly zone map, in order to generate the optimal real-time obstacle avoidance trajectory. Firstly, the sudden no-fly zone map is a semantic segmentation mask obtained by processing and fusing environmental data using deep learning semantic boundary extraction algorithms (such as the U-Net model). Those skilled in the art can extract the key geometric features of the no-fly zone from this binary semantic segmentation mask through image processing and geometric analysis techniques. For example, contour extraction algorithms can be used to identify the boundary pixels of a sudden no-fly zone, thereby fitting the "shape" of the no-fly zone (such as a circle, rectangle, polygon, or more complex irregular shape). Simultaneously, by calculating the minimum and maximum coordinates of the no-fly zone's boundary, the "range" (i.e., its spatial area and size) of the no-fly zone can be determined. As for the "type of the sudden no-fly zone," although the semantic segmentation mask itself may not directly provide a type label, it can be inferred from its shape, range, and other environmental context information (e.g., if the sudden no-fly zone is long and narrow, it may be inferred as a temporary flight restriction; if it is circular and has a heat source in the center, it may be inferred as a fire rescue zone) using pre-defined rules or classifiers, thus identifying the specific type of the sudden no-fly zone.

[0141] Secondly, once detailed information such as the type, extent, and shape of the sudden no-fly zone is obtained, the process enters the stage of "intelligently selecting or combining suitable maneuvering flight templates and trajectory planning parameters." The core of this stage is the "environmentally adaptive motion primitive library." This library pre-stores various "maneuvering flight templates" (e.g., a circling flight template for circular no-fly zones, a lateral translation template for linear no-fly zones, a low-altitude crossing template for high-altitude sudden no-fly zones, etc.) and corresponding "trajectory planning parameters" (e.g., minimum safe distance, maximum turning radius, climb / descent speed limits, obstacle avoidance path smoothness requirements, etc.) for different types and geometric features of sudden no-fly zones. "Intelligent selection" involves matching or retrieving the maneuvering flight template that best suits the current scenario's requirements from the environmentally adaptive motion primitive library based on the currently identified no-fly zone type, extent, and shape. For example, if the sudden no-fly zone is identified as a "circular fire rescue zone," the "large radius circling obstacle avoidance" template is preferentially selected. "Combination" means that for more complex or irregular sudden no-fly zones, multiple basic templates may be spliced ​​or merged to form a more complex obstacle avoidance strategy. Simultaneously, based on the specific range and shape of the sudden no-fly zone, corresponding trajectory planning parameters are retrieved or adjusted from the primitive library to finely customize the selected template. For example, based on the actual size of the sudden no-fly zone, the orbital radius or lateral translation distance is adjusted to ensure that the obstacle avoidance path is both safe and efficient.

[0142] Finally, the optimal real-time obstacle avoidance trajectory is generated using the maneuvering flight template and trajectory planning parameters obtained through the intelligent selection or combination described above. "Optimal" has multiple meanings; it typically refers to optimizing certain performance indicators as much as possible while meeting safety constraints (such as not entering no-fly zones or colliding with other drones), such as minimizing flight time, minimizing energy consumption, maximizing mission continuity, or maintaining trajectory smoothness. The trajectory generation algorithm (e.g., based on spline curves, polynomial interpolation, or optimization algorithms) uses the selected template and parameters, combined with the drone's own dynamic model and current state, to calculate a flight path that satisfies all constraints in real time. Because this process is performed "in real time," the drone can quickly respond to sudden changes in no-fly zones and adjust its flight path accordingly. The "real-time obstacle avoidance trajectory" refers to the flight path that the drone dynamically calculates and generates during flight based on perceived information about sudden no-fly zones, effectively avoiding them. This trajectory generation is real-time and can cope with dynamic changes in no-fly zone boundaries. The "emergency escape mechanism" aims to guide the drone to evacuate quickly, safely, and orderly after it inadvertently enters a no-fly zone.

[0143] Through the above technical solutions, this application can significantly improve the adaptive migration capability of multi-UAV systems when facing sudden no-fly zones. First, the triage based on safety state coding ensures precise management of UAVs in different states, avoiding a "one-size-fits-all" approach. Second, for unintruded UAVs, a real-time obstacle avoidance trajectory is generated using an environment-adaptive motion primitive library. This allows UAVs to flexibly select the optimal maneuvering flight template and trajectory planning parameters based on the specific type of no-fly zone and their own state, achieving fast, efficient, and safe obstacle avoidance, effectively preventing UAVs from accidentally entering no-fly zones and reducing the risk of safety accidents. Third, for intruded UAVs, an emergency disengagement mechanism and coordinated swarm disengagement ensure the safe evacuation of UAVs in emergency situations. This coordinated mechanism not only plans the shortest disengagement path but also coordinates the actions of other UAVs within the swarm, providing a safe disengagement channel for intruded UAVs, minimizing their dwell time in no-fly zones, and avoiding secondary collisions, thereby effectively ensuring the overall safety of the UAV swarm and the continuous execution of missions.

[0144] In some implementations, the emergency escape mechanism includes:

[0145] A301. Immediately terminates the original mission and hovers the compromised drone, while transmitting a standardized alarm package through the cluster communication network. The standardized alarm package includes the real-time GPS location, remaining battery power, sensor data, and mission priority of the compromised drone.

[0146] A302. Based on the map of the sudden no-fly zone, determine the shortest path direction from the intruding drone to the boundary of the sudden no-fly zone through geometric analysis, so as to minimize the time spent crossing the sudden no-fly zone;

[0147] A303. Neighboring drones proactively adjust their trajectories based on the shortest path direction, standardized alarm packets, and pre-built conflict prediction models to form emergency passages, ensuring the safe exit of intruding drones. Neighboring drones, based on the real-time location and exit path of intruding drones, use cluster communication and dynamic conflict prediction models to determine in real time other drones within the cluster that need to adjust their trajectories to ensure safe exit.

[0148] Among them, the shortest path direction indicates the shortest straight-line evacuation path from the current location of the intruding drone to the boundary of the sudden no-fly zone. The emergency passage is a safe physical passage jointly maintained by the cluster. It is a three-dimensional safe airspace dynamically formed around the "shortest path direction" to guide the intruding drone to a clear safe point while ensuring the safety of the drone.

[0149] A304. Compare the remaining battery power of the intruded drone with its evacuation energy consumption. If the remaining battery power is greater than the evacuation energy consumption, evacuate in a straight line along the emergency exit to a safe point; if the remaining battery power is less than the evacuation energy consumption, perform a vertical descent and mark the crash site.

[0150] Among them, the remaining power of the drone that has been compromised is directly provided by the drone's own system;

[0151] Among them, energy consumption is eliminated. ;

[0152] Among them, flight energy consumption , The length of the emergency access lane (in km) Cruise energy consumption coefficient (unit: Wh / km), which represents the energy consumed per kilometer by the UAV during cruise flight. Energy consumption for attitude adjustment (preset to 5Wh, usually based on statistical analysis of factory test data or historical flight data from drone manufacturers); To mitigate energy consumption redundancy, 10% of energy is reserved. .

[0153] This application's solution effectively addresses the issue of safe and efficient evacuation of intruding drones within sudden no-fly zones by introducing an emergency evacuation mechanism. Specifically, when a drone intrudes into a no-fly zone, step A301 immediately terminates the mission and broadcasts a standardized alarm packet, enabling rapid response to the emergency and sharing of critical information, laying the foundation for subsequent collaborative decision-making. Step A302 utilizes a map of the sudden no-fly zone for geometric analysis to determine the shortest path direction, thereby minimizing the drone's dwell time in the danger zone and reducing risk exposure. Step A303, through active trajectory adjustment of neighboring drones and a conflict prediction model, creates a safe evacuation channel for the intruding drone, effectively avoiding secondary collisions and ensuring the smoothness of the cluster's collaborative evacuation. Finally, step A304, by comparing and evaluating the remaining battery power with the evacuation energy consumption, provides the drone with a final disposal plan based on the actual situation. Whether it's a straight withdrawal to a safe point or a vertical landing, the aim is to maximize the drone's own safety and avoid unnecessary losses. It is precisely because of these collaborative and intelligent decision-making steps that this application's emergency evacuation mechanism can ensure the safe and efficient adaptive migration of intruding drones within sudden no-fly zones.

[0154] The conflict prediction model uses a two-dimensional spatiotemporal verification process to determine whether the trajectories of neighboring drones and intruding drones overlap spatially, or whether they will arrive at the same spatial area at the same time, in order to predict conflict risk and classify it into levels. Specifically:

[0155] Let the emergency passage that has been intruded into the drone be a line segment. ,in Current position This is the exit point of the no-fly zone. Nearby drones. The real-time location is Its pre-planned trajectory in the future The coordinates inside are , This indicates the maximum prediction duration. It is the shortest distance between the nearest drone and the exit path. ,in If the distance is equal to the wingspan, then a spatial conflict is determined. If the spatial distance... However, both may arrive at a certain spatial region at the same time, so it is necessary to further calculate the time difference threshold:

[0156] Set up an intrusion drone along The flight speed is Reach a point on the path The time is ,in This refers to the real-time location coordinates of the drone when it triggers an emergency escape, i.e., the starting point of the escape path. Nearby drones arrive. The estimated time is , This indicates the position of a nearby drone at the current moment (i.e., at time zero). The speed of the nearby drone. If , If the threshold value is set to a time threshold, then a spatiotemporal coupling conflict is determined to exist.

[0157] Calculate the risk value by combining spatial and temporal conflict parameters. And classify them into levels:

[0158]

[0159] in, To preset spatial weights, Set to 0.5. .

[0160] when 0.7 indicates a high-risk level, requiring immediate trajectory adjustments; the deviation distance needs to be greater than twice the wingspan.

[0161] When 0.3 0.7 indicates a medium risk level, requiring a trajectory adjustment within 1 second, with the deviation distance exceeding the wingspan.

[0162] when 0.3 is a low-risk level, which allows the vehicle to maintain its trajectory but reduce its speed to 60% of its original speed.

[0163] In some implementations, step S4, which involves decomposing the task based on the sudden no-fly zone map and spatiotemporal isolation, includes:

[0164] The interrupted original task is split into multiple subtasks based on spatial continuity or temporal stages.

[0165] Spatial continuity splitting: The region corresponding to the complete original task is split into multiple non-overlapping sub-regions, each sub-region corresponds to a spatial sub-task, and the sub-regions corresponding to all spatial sub-tasks are spliced ​​together to equal the region corresponding to the complete original task.

[0166] Segment by time period: When the original task needs to be completed within a preset time period, the original task area remains unchanged, and the preset time period is divided into multiple time windows. Each time window corresponds to a time sub-task, and the sum of the time windows of all time sub-tasks is equal to the preset time period.

[0167] Based on the map of sudden no-fly zones, feasible sub-task sets are selected by spatiotemporal isolation, which is calculated according to the following formula:

[0168] Spatiotemporal isolation = •Time isolation+ • Spatial isolation;

[0169] Temporal isolation is the percentage of non-overlapping time windows between any two subtasks, calculated using the following formula:

[0170] Time isolation ;

[0171] In the formula, These are the weighting coefficients. This is the start time of the first subtask. This is the end time of the first subtask. This is the start time for the second subtask. This is the end time for the second subtask;

[0172] Spatial isolation is the percentage of non-overlapping execution regions between any two subtasks, calculated using the following formula:

[0173] Spatial isolation ;

[0174] In the formula, This represents the area of ​​the execution region for the first subtask. The area of ​​the execution region for the second subtask;

[0175] Determine the set of retained subtasks based on spatiotemporal isolation. Calculated according to the following formula:

[0176] ;

[0177] In the formula, For a set of sub-tasks, This refers to two subtasks The spatiotemporal isolation degree is obtained by weighting the temporal isolation degree and the spatial isolation degree. For isolation value, For the first subtask, This is the second subtask.

[0178] Through the above technical solution, this application can effectively avoid temporal and spatial conflicts between subtasks, significantly improve the efficiency and safety of task execution, and ensure the smooth completion of cluster tasks. Specifically, by evaluating and filtering the spatiotemporal isolation of subtasks, the risk of collisions or interference when UAVs execute decomposed subtasks can be effectively reduced, improving the reliability of multi-UAV collaborative operations. Furthermore, this refined task decomposition and filtering mechanism enables the cluster to respond more flexibly to emergencies and optimize resource allocation, thereby maintaining task continuity and efficiency in complex and ever-changing environments.

[0179] In practical applications, the original area inspection task is first divided into multiple non-overlapping sub-regions based on spatial continuity. For example, a large rectangular inspection area is divided into four smaller rectangular sub-regions, each corresponding to a spatial sub-task. Simultaneously, if the original task has strict time requirements, it can also be divided into time-phased segments, dividing the total inspection time period into multiple time windows, each corresponding to a time sub-task. Subsequently, the spatiotemporal isolation between these newly generated sub-tasks is calculated. For example, for two adjacent spatial sub-tasks T1 and T2, if their time windows overlap but their spatial regions do not overlap at all, the spatial isolation is 1, and the temporal isolation is calculated based on the degree of overlap. If two sub-tasks T3 and T4 do not overlap at all in time but their spatial regions partially overlap, the temporal isolation is 1, and the spatial isolation is calculated based on the area of ​​the overlapping region. This is achieved through preset weighting coefficients. The overall spatiotemporal isolation between each pair of subtasks is calculated. Finally, these subtasks are filtered according to the set isolation value. For example, if the set isolation value is... If the spatiotemporal isolation is 0.1, then only subtasks with a spatiotemporal isolation greater than or equal to -0.1 will be retained in the final set of feasible subtasks. Subtasks with excessively low spatiotemporal isolation (i.e., high risk of conflict) will be excluded. Thus, the drone swarm can reallocate tasks based on the filtered set of subtasks, ensuring that each drone's subtasks are independent in time and space, thereby avoiding potential conflicts and guaranteeing the continuous, safe, and efficient completion of inspection tasks even after a sudden no-fly zone occurs.

[0180] In some implementations, step S4, which involves using an energy-constrained auction mechanism to ensure the continuous execution of cluster tasks, includes:

[0181] The drone that interrupted the task broadcasts key parameters of the subtask to the cluster. These key parameters include the execution area boundary coordinates and time window requirements. , ] , Estimated energy consumption demand Task priority coefficient ;

[0182] Other drones in the cluster calculate their bidding scores using the following formula, and the one with the highest score takes over the subtask:

[0183] ;

[0184] In the formula, To compete for the remaining battery power consumption of drones, To estimate the delay time for completing sub-tasks in order to bid for drones, This is the maximum allowable delay threshold;

[0185] The publisher assigns subtasks to the highest-scoring drone and generates a takeover trajectory certificate. The takeover trajectory certificate includes: handover timestamp, subtask parameters, identification of the drones before and after the task, and energy consumption commitment value. The subtask parameters include the sub-region and time window corresponding to the subtask. The takeover trajectory certificate is stored through a distributed blockchain.

[0186] In practical applications, after all bidding drones have calculated and submitted their bidding scores, the publisher assigns the subtask to the highest-scoring drone. To ensure the transparency, traceability, and immutability of the task handover process, the publisher generates a takeover trajectory certificate. This certificate records key information about the task handover in detail, including a handover timestamp to accurately record the moment of task transfer; subtask parameters, i.e., the sub-region and time window corresponding to the subtask, ensuring the clarity of the task content; the identification of the preceding and following drones to clarify the responsible party for the task; and the energy consumption commitment value, i.e., the energy that the takeover drone promises to consume to complete the subtask (i.e., the "estimated energy consumption requirement" initially broadcast for the subtask). To further enhance its security and reliability, the takeover trajectory certificate is stored using a distributed blockchain, ensuring that all participants can verify its authenticity and preventing any unauthorized modifications. This application's solution, by introducing an energy-constrained auction mechanism, effectively addresses the problem of how to efficiently and rationally allocate decomposed sub-tasks to ensure the continued execution of cluster tasks after mission interruption due to sudden no-fly zones in multi-UAV collaborative operations.

[0187] In some implementations, step S5 is included after step S4:

[0188] S5. Each drone collects all data from each obstacle avoidance and escape event to form a dataset for each drone, and continuously optimizes the environment-adaptive motion primitive library through a federated learning framework.

[0189] Specifically, after a drone performs obstacle avoidance or escape maneuvers, its sensor data, trajectory data, control commands, and the results of the obstacle avoidance or escape are recorded in real time and stored locally on the drone. This data forms the datasets for each drone, used for local model training. The federated learning framework is a distributed machine learning paradigm whose core principle is to allow multiple participants to collaboratively train a shared global model without directly sharing the original data. Under this framework, each drone acts as a local client, independently training its local model on its local dataset and only uploading model updates (such as gradient information) to the central server. The central server aggregates the received model updates, forming a global model update, and distributes it to each drone, thereby achieving continuous optimization of the environment-adaptive motion primitive library.

[0190] This application's solution addresses the issue of insufficient adaptability of traditional pre-trained models in dynamic environments by introducing a federated learning mechanism. Specifically, when a UAV performs obstacle avoidance or escape operations, its actual flight data, environmental perception data, and decision results are collected and formed into a local dataset. This dataset contains valuable experience from the various complex situations and coping strategies encountered by the UAVs in actual operation. Through the federated learning framework, each UAV trains its model on the local dataset and only uploads the updated model parameters to the central server. The central server aggregates the model parameter updates from different UAVs, thereby integrating the obstacle avoidance and escape experience of the entire UAV swarm in different scenarios to form a more robust and generalized global model. Subsequently, this global model update is distributed to each UAV to update its local environment-adaptive motion primitive library. This iterative process allows the environment-adaptive motion primitive library to continuously learn and optimize based on the data accumulated during actual operation, thereby continuously improving the UAV's obstacle avoidance and escape capabilities when facing sudden no-fly zones, effectively reducing task interruptions caused by insufficient model performance, and thus ensuring the continuous and efficient execution of swarm tasks.

[0191] In some implementations, step S5 includes:

[0192] Each drone collects all data from each obstacle avoidance and escape event to form a dataset;

[0193] Each drone continuously optimizes based on its local dataset, using gradient descent optimization according to the following formula:

[0194] ;

[0195] In the formula, For the first The previous local model parameters for each drone. For learning rate, For local loss function For local model parameters gradient, This is the weighted sum of the obstacle avoidance trajectory error and energy consumption loss of the drone.

[0196] The cloud collects the local model parameters of all drones, and calculates the global model parameters by weighting them according to the sample size of the dataset. The global model parameters are calculated according to the following formula:

[0197] ;

[0198] In the formula, These are global model parameters. For the first The percentage of each drone in the sample. The total number of drones, For the first The number of samples in the local event dataset owned by each drone;

[0199] The optimized global model parameters are injected into the environment-adaptive motion primitive library to update the maneuvering flight template and trajectory planning parameters, thereby enabling the continuous evolution of the cluster decision-making strategy.

[0200] In practical applications, after local optimization is completed, the cloud server is configured to collect all local model parameters uploaded by the drones. To fairly and effectively integrate this dispersed knowledge, the cloud server uses a weighted average to calculate the global model parameters. This weighting method ensures that drones with more effective training data contribute more to the global model, thereby ensuring that the global model more accurately reflects the overall experience of the cluster.

[0201] Furthermore, the optimized global model parameters are injected into an environment-adaptive motion primitive library. This library stores maneuvering flight templates and trajectory planning parameters for different no-fly zone types and environmental conditions. By updating these templates and parameters, the swarm's decision-making strategy can continuously evolve. This means that the drone swarm can adjust its obstacle avoidance and disengagement behavior based on the latest, globally aggregated experience, enabling it to better adapt to the complex and ever-changing no-fly zone environment.

[0202] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0203] The above are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for multi-UAV adaptive migration for burst no-fly zone, applied to a plurality of unmanned aerial vehicles (UAVs), comprising the steps of: The method comprises the steps of: S1. obtaining fused environment data through cooperative sensing of multiple unmanned aerial vehicles (UAVs); S2. processing the fused environment data using a deep learning-based semantic boundary extraction algorithm to output an emergent no-fly zone map and a safety state code of each UAV; S3. distributing the UAVs according to the safety state code and the emergent no-fly zone map, generating a real-time obstacle avoidance trajectory for a UAV that has not intruded into the emergent no-fly zone, and executing cluster cooperative separation for a UAV that has intruded into the emergent no-fly zone; S4. based on the emergent no-fly zone map and a time-space isolation degree, implementing task decomposition for a UAV whose task is interrupted due to obstacle avoidance or separation, and realizing continuous execution of cluster tasks through an energy consumption constraint auction mechanism; Based on the emergent no-fly zone map, a feasible sub-task set is screened through a time-space isolation degree, and the time-space isolation degree is calculated according to the following formula: spatial isolation temporal isolation spatial isolation The time isolation degree is the non-overlapping ratio of the time windows of any two sub-tasks, and is calculated according to the following formula: temporal isolation ; wherein is a weight coefficient, is a first subtask start time, is a first subtask end time, is a second subtask start time, is a second subtask end time; The space isolation degree is the non-overlapping ratio of the execution areas of any two sub-tasks, and is calculated according to the following formula: spatial isolation ; In the formula, is the execution area of the first subtask, is the execution area of the second subtask; Determining a set of subtasks to be retained based on spatiotemporal isolation is calculated according to the following equation: ； wherein is a set of subtasks, refers to two subtasks is a spatiotemporal isolation degree calculated by weighting the temporal isolation degree and the spatial isolation degree, is an isolation value, is a first subtask, is a second subtask.

2. The multi-UAV adaptive migration method for a burst-oriented no-fly zone according to claim 1, characterized in that, Step S1 comprises: Real-time collection of environment data through laser radars, visual cameras and millimeter wave radar sensors on each UAV; Fusion processing of the environment data obtained by each UAV to obtain fused environment data.

3. The multi-UAV adaptive migration method for a burst-oriented no-fly zone according to claim 1, characterized in that, The deep learning-based semantic boundary extraction algorithm adopts a U-Net model with a dynamic boundary attention mechanism; step S2 comprises: S201. receiving the fused environment data, wherein the fused environment data comprises a point cloud projection map reflecting distance information output by a laser radar and an image reflecting texture and contour information output by a visual camera, and the point cloud projection map and the image are spliced into an input feature map with a size of HxWx3, wherein H is the height of the image and W is the width of the image; S202. using five sequentially connected encoding stages to perform layer-by-layer downsampling and feature extraction on the input feature map, each encoding stage comprising a convolution layer, a batch normalization layer and a ReLU activation function, and a residual connection is introduced, and finally five groups of multi-scale feature maps are output; S203. Based on the real-time shortest Euclidean distance between the UAV and the sudden no-fly zone, the spatial attention submodule focuses on the boundary of the sudden no-fly zone, and real-time generates the safety state code of each UAV, while the multi-scale feature map is weighted and enhanced to obtain the spatial enhanced feature ; S204. According to the safety state code of the unmanned aerial vehicle, dynamically allocate the weight of the laser radar channel and the vision channel, and use the weight to fuse the features of the laser radar channel and the vision channel to obtain channel fusion features ; concatenate the spatial enhancement features with the channel fusion features to obtain the comprehensive features enhanced by dynamic attention ; S205. restoring the comprehensive features to a segmentation mask through the decoding structure of the U-Net model, and outputting a final feature map with the same size as the input feature map; S206. binarizing the final feature map through a Sigmoid activation function to output an emergent no-fly zone semantic segmentation mask, and obtaining an emergent no-fly zone map.

4. The multi-UAV adaptive migration method for a burst-oriented no-fly zone according to claim 3, characterized in that, After step S206, step S207 is further included: A hybrid loss function is used to balance the classification accuracy of the safety state code and the boundary region consistency of the emergent no-fly zone map, and the expression of the hybrid loss function is as follows: ； wherein ; wherein ; wherein ; In the formula, The value of the loss function. These are the weighting coefficients. For cross-entropy loss, For Dice's loss, For pixels Authentic labels Predicting pixels in feature maps for the U-Net model The probability of falling within the boundary of a sudden no-fly zone. Total number of pixels This is the smoothing coefficient.

5. The multi-UAV adaptive migration method for a burst-oriented no-fly zone according to claim 3, characterized in that, Step S203 comprises: ； In the formula, represents the real-time shortest Euclidean distance between the UAV and the emergent no-fly zone, represents traversing all From the minimum value in the results, represents the total number of boundary line segments of the emergent no-fly zone, represents the shortest Euclidean distance from the UAV centroid to the first boundary line segment of the emergent no-fly zone, if the vertical projection point of the UAV centroid falls on the first boundary line segment of the emergent no-fly zone, then is the straight-line distance from the UAV centroid to the projection point, if the projection point does not fall on the first boundary line segment of the emergent no-fly zone, then is the minimum value of the distances from the UAV centroid to the two ends of the first boundary line segment of the emergent no-fly zone; When m, the safety state of the UAV is coded as 0; When m, the safety state of the UAV is coded as 1; When The safety state of the UAV is coded as 2; The real-time shortest Euclidean distance between the unmanned aerial vehicle and the emergent no-fly zone is normalized to obtain each pixel in a feature map spatial correlation degree with the boundary of the emergent no-fly zone ; calculating the real-time shortest Euclidean distance between the UAV and the no-fly zone according to the following formula: ； wherein is a spatial attention weight, is a sigmoid function, and are learning parameters, is a distance adjustment factor; calculating the spatial attention weight according to the following formula: ； wherein, is a spatially enhanced feature of the pixel point , is an original feature value of the pixel point .

6. The multi-UAV adaptive migration method for a burst-oriented no-fly zone according to claim 1, characterized in that, enhancing the feature map through weighting according to the following formula to obtain a spatial enhanced feature: Step S3 comprises: Distributing the UAVs according to the safety state code; According to the sudden no-fly zone map, for the non-intruded UAV, a pre-trained environmental adaptive motion primitive library is used to generate a real-time obstacle avoidance trajectory, and the environmental adaptive motion primitive library includes maneuvering flight templates and trajectory planning parameters for different types of sudden no-fly zones; According to the sudden no-fly zone map, for the intruded UAV, an emergency escape mechanism is adopted to make the UAV cluster escape cooperatively.

7. The multi-UAV adaptive migration method for a burst-oriented no-fly zone according to claim 6, characterized in that, The emergency escape mechanism includes: A301. The intruded UAV immediately terminates the original task and hovers, while transmitting a standardized alarm package through the cluster communication network, the standardized alarm package including the real-time GPS position, remaining power, sensor data and task priority of the intruded UAV; A302. According to the sudden no-fly zone map, the shortest path direction of the intruded UAV to the boundary of the sudden no-fly zone is determined through geometric analysis to minimize the time of crossing the sudden no-fly zone; A303. The neighboring UAVs actively adjust the trajectory according to the shortest path direction, the standardized alarm package and the pre-constructed conflict prediction model to form an emergency channel, ensuring the safe passage of the intruded UAV; A304. The remaining power of the intruded UAV is compared with the energy consumption of escape, if the remaining power is greater than the energy consumption of escape, the intruded UAV escapes to a safe point along the emergency channel; if the remaining power is less than the energy consumption of escape, the intruded UAV performs vertical landing and marks the crash point.

8. The multi-UAV adaptive migration method for a burst-oriented no-fly zone according to claim 1, wherein, In step S4, the task decomposition based on the sudden no-fly zone map and the spatiotemporal isolation degree includes: The interrupted original task is split into multiple sub-tasks according to spatial continuity or time periodicity; According to spatial continuity, the area corresponding to the complete original task is split into multiple non-overlapping sub-areas, each sub-area corresponding to a spatial sub-task, and the sub-areas corresponding to all spatial sub-tasks are spliced to equal the area corresponding to the complete original task; According to time periodicity, when the original task needs to be completed within a preset time period, the original task area remains unchanged, and the preset time period is split into multiple time windows, each time window corresponding to a time sub-task, and the time windows of all time sub-tasks sum up to the preset time period.

9. The multi-UAV adaptive migration method for a burst-oriented no-fly zone according to claim 7, characterized in that, In step S4, the continuous execution of cluster tasks is realized through an energy consumption constrained auction mechanism, which includes: The key parameters of the sub-task are broadcasted to the cluster by the unmanned plane interrupting the task, the key parameters including execution area boundary coordinates, time window requirements , ], estimated energy consumption requirements , and task priority coefficient ; The other UAVs in the cluster calculate the bid scores according to the following formula, and select the UAV with the highest score to take over the sub-task: ； In the formula, a remaining power consumption of the bidding UAV, a delay time of the bidding UAV estimated to complete the sub-task, a maximum allowed delay threshold value; The publisher assigns the sub-task to the UAV with the highest score and generates a takeover trajectory certificate, which includes the handover timestamp, sub-task parameters, the identity of the previous and next task UAVs, and the energy consumption commitment value. The sub-task parameters include the sub-area and time window corresponding to the sub-task, and the takeover trajectory certificate is stored through a distributed blockchain.

10. The multi-UAV adaptive migration method for a burst-oriented no-fly zone according to claim 6, wherein, After step S4, step S5 is further included: S5. Each UAV collects all data of each obstacle avoidance and escape event to form a data set corresponding to each UAV, and continuously optimizes the environmental adaptive motion primitive library through a federated learning framework.

Images