Unmanned aerial vehicle cluster cooperative control system and method

Abstract

The invention discloses an unmanned aerial vehicle cluster cooperative control system and method, and relates to the technical field of unmanned aerial vehicle cluster control, and the system comprises a multi-source sensing unit, an integrated laser radar, an RGB-D camera and a millimeter wave radar array, and is used for generating a centimeter-level precision three-dimensional map and tracking a dynamic obstacle in real time; the collaborative decision-making unit comprises an artificial intelligence prediction module which is used for predicting a collision point based on the movement state of the obstacle; the safety space construction module is used for generating a dynamic safety sphere according to the formation envelope size and the obstacle volume; the artificial intelligence correction module is used for constructing a safety section and generating an optimization path; and the dynamic communication unit is used for carrying out cluster data interaction by adopting hybrid networking of 5G and LoRa adaptive switching. The unmanned aerial vehicle cluster can be used as a whole to avoid obstacles, the original forms of the unmanned aerial vehicles are synchronously kept, and it is guaranteed that multiple unmanned aerial vehicles in the cluster do not collide.

Description

Technical Field

[0001] The present invention relates to the technical field of UAV swarm control, and particularly to a UAV swarm collaborative control system and method. Background Art

[0002] Currently, UAVs have shown broad application prospects in many application fields such as infrastructure inspection, underground mineral exploration, accident search and rescue, mapping, and precision agriculture. In general scenarios, the flight tasks of UAVs do not consider the influence of dynamic obstacles. However, the actual situation is exactly the opposite, and flight accidents are often caused by dynamic obstacles. Therefore, in complex environments, multiple UAVs need to fly in formation collaboratively to jointly deal with unidentified flying objects, and the UAV formation needs to determine the movement trajectory of dynamic obstacles.

[0003] A Chinese patent with the publication number "CN113625762B" discloses a UAV obstacle avoidance method and system, and a UAV swarm obstacle avoidance method and system. It obtains the position information of dynamic obstacles; based on the position information and the quasi-linear parameter-varying model, inversely predicts the flight trajectory of the dynamic obstacles; based on the flight trajectory, determines the collision time interval between the UAV and the dynamic obstacles; based on the collision time interval, determines whether the UAV enters the collision area, and when the UAV enters the collision area, selects a corresponding obstacle avoidance scheme according to the relative position between the UAV and the dynamic obstacles, so that it can inversely predict the flight trajectory of dynamic obstacles only based on the position information and the quasi-linear parameter-varying model, achieving the purpose of accurately predicting the movement trajectory of dynamic obstacles and accurately avoiding obstacles.

[0004] However, when UAV swarms avoid obstacles, it is often necessary to disrupt the swarm formation. During the obstacle avoidance process, not only does it need to maintain a certain distance from the obstacles, but also it needs to take into account the collision risk with other UAVs within the swarm. When this patent selects a corresponding obstacle avoidance scheme according to the relative position between the UAV and the dynamic obstacles, it not only disrupts the swarm formation but also has difficulty taking into account the collision risk with multiple UAVs within the swarm.

[0005] Therefore, how to prevent collisions between multiple UAVs within the swarm while achieving obstacle avoidance has become an urgent technical problem to be solved. Summary of the Invention

[0006] The technical problem solved by the present invention is that it is difficult in the prior art to prevent collisions between multiple UAVs within the swarm while achieving obstacle avoidance.

[0007] To solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, a UAV swarm cooperative control system includes: a multi-source perception unit, integrating a lidar, an RGB-D camera, and a millimeter-wave radar array, for generating a three-dimensional map with centimeter-level accuracy and real-time tracking of dynamic obstacles; a cooperative decision-making unit, including: an artificial intelligence prediction module, for predicting a collision point based on the motion state of an obstacle; a safety space construction module, for generating a dynamic safety sphere according to the formation envelope size and the volume of the obstacle; an artificial intelligence correction module, for constructing a safety section and generating an optimized path; a dynamic communication unit, for performing cluster data interaction by using a hybrid networking with adaptive switching between 5G and LoRa.

[0008] Preferably, the working modes of the multi-source perception unit include: the lidar and the RGB-D camera, for generating a three-dimensional map with centimeter-level accuracy through calibration parameter fusion; the millimeter-wave radar array, for continuously tracking dynamic obstacles within a preset detection distance and estimating their motion speed and direction; meteorological sensors arranged on each UAV, for real-time collecting wind speed, rainfall intensity, and electromagnetic interference intensity to form an environmental disturbance parameter set.

[0009] Preferably, the operation mechanism of the dynamic communication unit includes: when the signal-to-noise ratio of the 5G channel is lower than a threshold or the bit error rate exceeds a preset limit value, automatically switching to the LoRa communication link; adopting a security protocol compliant with the FIPS140-2 standard for data transmission, dynamically generating an encryption key and updating it periodically.

[0010] Preferably, the cooperative decision-making unit includes a heterogeneous computing architecture: an airborne FPGA chip runs a lightweight object detection model for real-time obstacle recognition; a cloud GPU cluster deploys a deep reinforcement learning model to optimize the obstacle avoidance strategy parameters in combination with historical flight data.

[0011] Preferably, the training process of the lightweight object detection model is as follows: inputting a UAV obstacle training data set into the original YOLOv7 model; removing redundant channels in the backbone network with a contribution degree lower than a preset channel threshold based on the activation contribution degree ranking of the feature map; constructing a teacher-student architecture in the model after removing redundant channels, and forcing the student model to imitate the spatial attention distribution of the teacher model in the feature pyramid layer through a multi-scale feature alignment loss function.

[0012] Preferably, the operations performed by the artificial intelligence correction module include: analyzing the relative position weights of each UAV in the formation based on the attention mechanism to generate an optimal path candidate set for maintaining the formation; adopting a hybrid optimization algorithm to perform multi-objective trade-offs on the flight trajectory under the constraint of the safety section, and the multi-objectives include minimum energy consumption, shortest time, and flight path smoothness.

[0013] Preferably, the safety space construction module is specifically configured to: calculate the radius of the formation envelope sphere according to the real-time distribution density of the UAV formation, and determine the safety margin in combination with the predicted value of the obstacle volume; take the collision time window as the time dimension and construct a safety space that shrinks or expands over time.

[0014] Preferably, the artificial intelligence correction module further includes: a speed adaptive adjustment component that dynamically adjusts the flight speed according to the length difference between the corrected path and the original path, satisfying: when the corrected path is extended, the speed is increased at the maximum allowable acceleration within the remaining flight time; when the corrected path is shortened, the overall formation synchronization is maintained for uniform speed or deceleration adjustment.

[0015] Preferably, the system further includes an energy optimization module, and its working method is: calculating the power distribution scheme of each UAV through a convex optimization algorithm to minimize the total energy consumption of the cluster; combining a genetic algorithm to search for a local optimal solution that meets the endurance requirements under power constraints.

[0016] In a second aspect, the present invention provides a method for collaborative control of a UAV cluster. Based on the system described in any of the above embodiments, the method includes: establishing an environmental dynamic model through multi-sensor fusion; performing real-time collision risk detection at an edge node to identify safety threat events reaching a threshold; generating an obstacle avoidance trajectory correction scheme through cloud-edge collaborative computing; and synchronously updating the flight control instructions of all UAVs in the cluster using hybrid networking communication.

[0017] The beneficial effects of the present invention are as follows: By generating a dynamic safety sphere according to the formation envelope size and the obstacle volume, constructing a safety section, and generating an optimized path, the UAV cluster can be used as a whole to avoid obstacles, so that when the cluster avoids obstacles, the original shape of the UAVs can be maintained synchronously, and collisions between multiple UAVs in the cluster can be prevented; Through the multi-modal sensor fusion of the multi-source perception unit, the high-precision ranging of the lidar, the stereo vision capture of the RGB-D camera, and the motion tracking ability of the millimeter-wave radar work together, so that the obstacle recognition resolution reaches the centimeter level, thus realizing the real-time reconstruction of the three-dimensional trajectory of dynamic obstacles and the update of the millimeter-wave level motion state in a complex environment; Through the anthropomorphic trajectory prediction algorithm of the collaborative decision-making unit, the prediction time domain of the obstacle trajectory can be extended, so as to improve the accuracy of the collision time interval prediction; Through the adaptive switching architecture of 5G and LoRa dual-mode communication, seamless switching is triggered by real-time detection of the link quality, so as to improve the communication availability in complex scenarios, and thus realize the synchronous update of the obstacle avoidance path and the morphological reorganization of a large-scale cluster in a short time. Description of the Drawings

[0018] Figure 1 It is a structural block diagram of a UAV cluster collaborative control system provided by an embodiment of the present invention; Figure 2Schematic flowchart of the UAV swarm cooperative control method provided by another embodiment of the present invention; Figure 3 Schematic diagram of the application scenario of the UAV swarm cooperative control method provided by an embodiment of the present invention; Figure 4 Schematic diagram of the application scenario of another UAV swarm cooperative control method provided by an embodiment of the present invention. Detailed implementation manners

[0019] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the detailed implementation manners of the present invention with reference to the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments.

[0020] Embodiment 1, referring to Figure 1 , which is an embodiment of the present invention, provides a UAV swarm cooperative control system, including: a multi-source perception unit, integrating a lidar, an RGB-D camera, and a millimeter-wave radar array, for generating a three-dimensional map with centimeter-level accuracy and real-time tracking of dynamic obstacles; a cooperative decision-making unit, including: an artificial intelligence prediction module, for predicting the collision point based on the movement state of the obstacle; a safety space construction module, for generating a dynamic safety sphere according to the formation envelope size and the obstacle volume; an artificial intelligence correction module, for constructing a safety section and generating an optimized path; a dynamic communication unit, for adopting a hybrid networking with adaptive switching between 5G and LoRa for swarm data interaction.

[0021] The fusion technology of LiDAR and RGB-D camera achieves spatial alignment through a checkerboard calibration board, with the calibration error controlled within 0.3 cm. During the calibration process, by minimizing the reprojection error, the LiDAR point cloud (coordinate system {X_L, Y_L, Z_L}) is precisely matched with the pixel coordinates ({u, v, d}) of the RGB-D camera. When fusing data, the LiDAR generates a point cloud P = {p_i | p_i = (x_i, y_i, z_i)} with a density of 500 points per square meter, and the RGB-D camera outputs a color image C and a depth map D. Through the calibration matrix T, the point cloud P is projected onto the pixel coordinate system of C to generate a dense point cloud P' with RGB colors. Finally, the ICP (Iterative Closest Point) algorithm is used to optimize the local matching between P' and D, and the fused 3D map is output. In the accuracy verification, a target with known coordinates (accuracy ±0.1 cm) was placed in a 10 m × 10 m standard test field, and the measured reconstruction error was ±1.8 cm, meeting the design index (±2 cm). In the building spraying scenario, the LiDAR scans the building facade at a frequency of 20 Hz to obtain the structural outline, and the RGB-D camera captures the surface texture (such as cracks, window edges) at 30 fps. After fusion, a 3D model with color information is generated for spraying path planning, avoiding the window frame area with an error less than 2 cm.

[0022] The dynamic tracking technology of millimeter-wave radar array uses 4 77GHz millimeter-wave radars (Arbe Phoenix chips), which are arranged at the four corners of the drone to form a 360° coverage. The detection algorithm includes range-Doppler processing, where the target range and radial velocity are extracted through FFT, with resolutions of 0.5 m and 0.1 m / s respectively; multi-target tracking uses the JPDA (Joint Probability Data Association) algorithm to associate multi-radar data and outputs a list of obstacle trajectories; motion estimation is based on the Extended Kalman Filter (EKF) to predict the position of obstacles within the next 5 seconds, with a confidence level exceeding 95%. This system can track 32 dynamic targets simultaneously, with a maximum detection range of 500 m and an azimuth accuracy of ±0.1°. In the logistics distribution obstacle avoidance scenario, the millimeter-wave radar detects a truck cutting in laterally at a speed of 120 km / h, at a distance of 350 m. The system predicts that the truck will enter the formation flight path in 8 seconds, triggers the obstacle avoidance instruction, and the artificial intelligence correction module generates a detour path, causing the drone swarm to move up 20 m to avoid the truck.

[0023] The real-time monitoring technology of meteorological sensors uses a Gill WindMaster Pro anemometer (range 0 - 60 m / s, accuracy ±0.1 m / s), a Vaisala WXT530 rainfall sensor (able to distinguish rain / snow, resolution 0.01 mm), and an R&S FPL1003 spectrum analyzer (detection range 10 MHz - 6 GHz). Data processing includes decomposing the wind speed vector into three-axis components (V_x, V_y, V_z) in the UAV coordinate system, mapping the rainfall intensity to a dynamic system compensation coefficient (for example, the motor power increases by 15% during heavy rain), and switching to an anti-interference communication mode when the electromagnetic interference exceeds the threshold (>10 V / m). During the inspection of offshore wind turbines, the meteorological sensors detect an instantaneous wind speed of 25 m / s (level 10 wind), and the cooperative decision-making unit activates the anti-wind mode. The formation switches to a triangular dense formation to reduce the wind resistance area, and at the same time, the motor power is increased to 120% to ensure stable attitude.

[0024] The UAV swarm cooperative control system is a complex technical architecture based on multi-source perception, cooperative decision-making, and dynamic communication, aiming to achieve the efficient and safe operation of UAV swarms in complex environments. The system integrates lidar, RGB-D cameras, millimeter-wave radar arrays, and meteorological sensors to perceive the environmental state in real time; conducts path planning, obstacle avoidance, and formation adjustment through a cooperative decision-making unit; and realizes internal data interaction and instruction synchronization within the swarm through a dynamic communication unit.

[0025] The core modules of the system include a multi-source perception unit, a cooperative decision-making unit, and a dynamic communication unit. The multi-source perception unit generates a high-precision 3D map through the fusion technology of lidar and RGB-D cameras, and at the same time, the millimeter-wave radar array dynamic tracking technology monitors the movement state of obstacles in real time. The cooperative decision-making unit predicts potential collision risks based on the perception data and generates a dynamic safety sphere through a safety space construction module to ensure the safe flight of UAV swarms in complex environments. The artificial intelligence correction module generates an optimal path candidate set for maintaining formation based on the attention mechanism and a hybrid optimization algorithm, and conducts multi-objective trade-offs on the flight trajectory, including minimizing energy consumption, minimizing time, and flight path smoothness. The dynamic communication unit automatically switches to the LoRa communication link when the signal-to-noise ratio of the 5G channel is lower than the threshold or the bit error rate exceeds the preset limit to ensure the reliability of data transmission. At the same time, a security protocol compliant with the FIPS140-2 standard is adopted for data transmission, and an encryption key is dynamically generated and updated periodically to ensure the security of swarm communication.

[0026] The fusion technology of lidar and RGB-D camera achieves spatial alignment through a checkerboard calibration board, and the calibration error is controlled within 0.3 cm. During the calibration process, the LiDAR point cloud and the pixel coordinates of the RGB-D camera are precisely matched by minimizing the reprojection error. When fusing data, the LiDAR generates a point cloud with a density of 500 points per square meter, and the RGB-D camera outputs a color image and a depth map. Through the calibration matrix, the point cloud is projected onto the pixel coordinate system to generate a dense point cloud with RGB colors, and the ICP algorithm is used to optimize the local matching to output the fused three-dimensional map. In the accuracy verification, a target with known coordinates was placed in a 10 m × 10 m standard test field, and the measured reconstruction error was ±1.8 cm, meeting the design specifications.

[0027] The dynamic tracking technology of millimeter-wave radar arrays uses 4 77-GHz millimeter-wave radars, which are arranged at the four corners of the UAV to form a 360° coverage. The detection algorithm includes range-Doppler processing, and the target range and radial velocity are extracted through FFT, with resolutions of 0.5 m and 0.1 m / s respectively; multi-target tracking uses the JPDA algorithm to associate multi-radar data and outputs a list of obstacle trajectories; motion estimation is based on the extended Kalman filter to predict the position of obstacles within the next 5 seconds, with a confidence level exceeding 95%. This system can track 32 dynamic targets simultaneously, with a maximum detection range of 500 m and an azimuth accuracy of ±0.1°. In the obstacle avoidance scenario of logistics distribution, the millimeter-wave radar detected a truck cutting in laterally at a speed of 120 km / h, at a distance of 350 m. The system predicted that the truck would enter the formation flight path after 8 seconds, triggering the obstacle avoidance instruction, and the artificial intelligence correction module generated a detour path, causing the UAV swarm to move up 20 m to avoid the truck.

[0028] The real-time monitoring technology of meteorological sensors uses a Gill WindMaster Pro anemometer, a Vaisala WXT530 rainfall sensor, and an R&S FPL1003 spectrum analyzer. Data processing includes decomposing the wind speed vector into three-axis components in the UAV coordinate system, mapping the rainfall intensity to the dynamic system compensation coefficient, and switching to the anti-interference communication mode when the electromagnetic interference exceeds the threshold. During the inspection of an offshore wind turbine generator, the meteorological sensor detected an instantaneous wind speed of 25 m / s, and the cooperative decision-making unit activated the anti-wind mode. The formation switched to a triangular dense formation to reduce the wind resistance area, and at the same time, the motor power was increased to 120% to ensure attitude stability.

[0029] The collaborative decision-making unit includes an artificial intelligence prediction module, a safety space construction module, and an artificial intelligence correction module. The artificial intelligence prediction module predicts the future trajectory of an obstacle based on its motion state and calculates the potential collision points with the UAV cluster. The safety space construction module calculates the radius of the formation envelope sphere according to the real-time distribution density of the UAV formation and determines the safety margin in combination with the predicted value of the obstacle volume. The artificial intelligence correction module analyzes the relative position weights of each UAV in the formation based on the attention mechanism, generates an optimal path candidate set for maintaining the formation, and uses a hybrid optimization algorithm to perform multi-objective trade-offs on the flight trajectory.

[0030] The dynamic communication unit adopts a hybrid networking of adaptive switching between 5G and LoRa to ensure the stability and security of data interaction. When the signal-to-noise ratio of the 5G channel is lower than the threshold or the bit error rate exceeds the preset limit, it automatically switches to the LoRa communication link. At the same time, an encryption protocol compliant with the FIPS140-2 standard is adopted to dynamically generate encryption keys and update them periodically to ensure the security of data transmission.

[0031] The energy optimization module calculates the power distribution scheme of each UAV through a convex optimization algorithm to minimize the total energy consumption of the cluster. The genetic algorithm is combined to search for a local optimal solution that meets the endurance requirements under power constraints. In long-term tasks, the energy optimization module significantly extends the endurance time of the cluster.

[0032] The system performs excellently in scenarios such as inspection of offshore wind turbine generators, obstacle avoidance in logistics distribution, and building spraying. In the inspection of offshore wind turbine generators, meteorological sensors collect real-time wind speed, rainfall intensity, and electromagnetic interference intensity to form an environmental disturbance parameter set. In obstacle avoidance in logistics distribution, the millimeter-wave radar array detects dynamic obstacles and triggers an obstacle avoidance instruction, and the artificial intelligence correction module generates a detour path. In building spraying, the fusion technology of lidar and RGB-D camera generates a high-precision three-dimensional map for spraying path planning.

[0033] The system performance indicators include that the three-dimensional map reconstruction error ≤ ±2 cm, the millimeter-wave radar array can track 32 dynamic targets simultaneously, the maximum detection distance is 500 m, the communication stability ensures that the data transmission interruption time ≤ 0.5 s when switching between 5G and LoRa, the total energy consumption of the cluster is reduced by 20%, and the endurance time is extended by 15%.

[0034] The UAV cluster collaborative control system realizes efficient and safe operation in complex environments through the deep integration of multi-source perception, collaborative decision-making, and dynamic communication.

[0035] In a second aspect, the present invention provides a method for collaborative control of an unmanned aerial vehicle (UAV) swarm. Based on the system described in any of the above embodiments, the method includes: establishing an environmental dynamic model through multi-sensor fusion; performing real-time collision risk detection at edge nodes to identify safety threat events reaching a threshold; generating an obstacle avoidance trajectory correction scheme through cloud-edge collaborative computing; and synchronously updating the flight control instructions of all UAVs in the swarm using hybrid networking communication.

[0036] By generating a dynamic safety sphere according to the formation envelope size and the obstacle volume, constructing a safety section plane and generating an optimized path, the UAV swarm can be regarded as a whole to avoid obstacles, so that when the swarm avoids obstacles, the original form of the UAVs can be synchronously maintained and collisions between multiple UAVs in the swarm can be prevented; through the multi-modal sensor fusion of the multi-source perception unit, the high-precision ranging of the lidar, the stereo vision capture of the RGB-D camera, and the motion tracking ability of the millimeter-wave radar work together, so that the obstacle recognition resolution reaches the centimeter level, thereby realizing the real-time reconstruction of the three-dimensional trajectory of dynamic obstacles and the update of the millimeter-wave level motion state in a complex environment; through the anthropomorphic trajectory prediction algorithm of the collaborative decision-making unit, the prediction time domain of the obstacle trajectory can be extended, thereby improving the accuracy of the collision time interval estimation; through the adaptive switching architecture of 5G and LoRa dual-mode communication, seamless switching is triggered by real-time detection of the link quality, so that the communication availability in complex scenarios is improved, thereby realizing the synchronous update of the obstacle avoidance path and the form reorganization of the large-scale swarm in a short time.

[0037] Referring to Figure 2 , which is another embodiment of the present invention. Specifically, the method includes S110 to S180: S110, real-time collecting the first position information of the UAV swarm and the second position information of the obstacles.

[0038] The airborne FPGA chip runs a lightweight target detection model to achieve real-time obstacle recognition at no less than 30 frames per second.

[0039] Among them, the lightweight target detection model is an optimized YOLOv7 model, and its optimization process includes a dynamic channel pruning stage, a mixed precision quantization stage, and a cross-model knowledge distillation stage.

[0040] Dynamic Channel Pruning Stage: Based on the L1 norm, the output feature maps of the convolutional layer are sorted according to channel importance, and redundant channels with activation contributions lower than the preset threshold are removed to generate the pruned model topology. The generated model topology (such as the number of channels and layer connection relationships) directly determines the sensitivity analysis object in the mixed-precision quantization stage. Based on the pruned model topology, perform sparse fine-tuning, freeze the weight updates of the pruned channels through the gradient masking technique, and restore the object detection accuracy to over 98% of the initial model. Verify the average precision loss (ΔmAP) of the pruned model on the UAV typical obstacle dataset. If ΔmAP > 2%, iteratively adjust the pruning threshold until the constraint is satisfied.

[0041] Mixed-Precision Quantization Stage: According to the layer sensitivity analysis results of the pruned model, divide the network quantization priorities, perform 8-bit fixed-point quantization on the feature extraction backbone network, and retain 16-bit floating-point calculations for the detection head network. Based on the activation distribution statistics of the pruned model, adopt an asymmetric quantization calibration strategy to dynamically adjust the truncation range of the activation function output to retain edge feature responses. Deploy a quantization-aware training mechanism, simulate the gradient update error under the combined action of pruning and quantization in backpropagation, and generate a low-bit compatible model. The quantization gradient error statistics results (such as the proportion of lost edge features) become the weight allocation basis for the multi-scale loss function in the cross-model knowledge distillation stage.

[0042] Cross-Model Knowledge Distillation Stage: Based on the structural characteristics of the low-bit compatible model, construct a teacher-student dual-network architecture, where the teacher network is the original YOLOv7 floating-point model, and the student network is the pruned and quantized low-bit model. According to the accuracy defect distribution of the quantized model, design a multi-scale feature alignment loss function to force the student network to imitate the spatial attention distribution of the teacher network in the feature pyramid layer. Soften the output probability distribution of the teacher network through the temperature scaling strategy, and strengthen the knowledge transfer effect for small target categories in the UAV scenario, so that the small target detection AP is increased by ≥4.

[0043] When deploying on FPGA hardware, based on the final model structure after knowledge distillation, convert it into a computation graph for the programmable gate array chip instruction set, and merge the convolution and batch normalization operations through layer fusion technology. According to the parallelism requirements of the model computation graph, dynamically configure the operation core array of the programmable gate array chip, and map the feature map slices to independent processing units. Based on the timing constraints of the real-time inference pipeline, verify the frame rate stability of the model at an input resolution of 640×640. If it does not reach 30fps, iteratively adjust the pruning rate and quantization strategy in reverse. The frame rate verification results trigger the reverse adjustment of the pruning rate and quantization strategy, forming a closed-loop optimization.

[0044] Three-dimensional scene reconstruction is performed through a Light Detection and Ranging (LiDAR) and an RGB-D camera to obtain map data, and the first motion trajectory is determined based on the map data. For example, in scenarios such as disaster relief, precision agriculture, logistics distribution, and building spraying construction, the first motion trajectory of the unmanned aerial vehicle needs to be determined based on the map data.

[0045] The millimeter-wave radar is used to dynamically track obstacles and collect the second position information in real time.

[0046] S120. Determine the first geometric data of the unmanned aerial vehicle cluster according to the distribution data of the unmanned aerial vehicle cluster, and determine the second geometric data of the obstacle according to the second position information of the obstacle.

[0047] Specifically, the first distribution sphere of the unmanned aerial vehicle cluster can be determined according to the distribution data of the unmanned aerial vehicle cluster; the second distribution sphere of the obstacle can be determined according to the second position information of the obstacle.

[0048] The first distribution sphere is consistent with the envelope size of the unmanned aerial vehicle formation, and the second distribution sphere is consistent with the envelope size of the obstacle.

[0049] Regarding the geometric distributions of the unmanned aerial vehicle cluster and the obstacle as spheres can facilitate the calculation of the dangerous space and make the dangerous space as accurate as possible.

[0050] S130. Obtain the first motion trajectory of the unmanned aerial vehicle cluster.

[0051] The first motion trajectory of the unmanned aerial vehicle cluster is a preset trajectory before executing the obstacle avoidance algorithm.

[0052] S140. Predict the second motion trajectory of the obstacle according to the position information of the obstacle.

[0053] The prediction of the second motion trajectory of the obstacle is realized based on the LSTM network.

[0054] S150. Determine the collision point according to the first motion trajectory, the second motion trajectory, the first geometric data, and the second geometric data.

[0055] Convert the spatio-temporal distribution parameters of the unmanned aerial vehicle cluster into a first dynamic envelope sphere, and convert the motion state parameters of the obstacle into a second predicted envelope sphere. Based on the extended Kalman filter algorithm, perform spatio-temporal intersection analysis on the motion trajectories of the first dynamic envelope sphere and the second predicted envelope sphere to determine the collision point coordinates and the collision time window.

[0056] S160. Determine the dangerous space according to the first geometric data, the second geometric data, and the collision point.

[0057] Generate a dangerous sphere space based on the sum of the radii of two spheres. The mathematical expression is: S_safe = { (x,y,z)| (x - x_c)^2 + (y - y_c)^2 + (z - z_c)^2 ≤ (r1 + r2 + δ)^2}, where (x_c, y_c, z_c) are the coordinates of the collision point, r1 and r2 are the radii of the two spheres respectively, and δ is the safety margin coefficient.

[0058] Furthermore, the third distribution sphere of the UAV at the collision point can be obtained according to the position of the collision point, the fourth distribution sphere of the obstacle at the collision point can be predicted according to the position of the collision point and the second position information, the dangerous radius can be determined according to the radius of the third distribution sphere and the radius of the fourth distribution sphere, and the sphere obtained with the collision point as the center of the sphere and the dangerous radius as the radius of the sphere is the dangerous sphere, and the space inside the dangerous sphere is the dangerous space.

[0059] S170. Modify the first motion trajectory according to the dangerous space to obtain the third motion trajectory of the UAV cluster.

[0060] Specifically, the two intersection points of the first motion trajectory and the dangerous sphere are respectively determined as the first intersection point and the second intersection point. The first intersection point and the second intersection point respectively form a first tangent plane and a second tangent plane with the dangerous sphere. The planes tangent to the sphere in the vertical plane of the first tangent plane and the vertical plane of the second tangent plane are respectively determined as the first vertical plane and the second vertical plane. Modify the trajectory segment of the first motion trajectory in the dangerous space according to the first vertical plane or the second vertical plane to obtain the third motion trajectory.

[0061] If the time for the UAV cluster to reach the second intersection point lags behind that of the first intersection point; then the intersection point of the second vertical plane and the first motion trajectory can also be determined as the correction starting point; perform an optimization operation according to the correction starting point, the dangerous sphere and the second intersection point to determine the third motion trajectory.

[0062] S180. Control the UAV cluster to fly according to the third motion trajectory.

[0063] Specifically, determine the flight time and the first flight distance of the UAV cluster from the correction starting point to the second intersection point in the first motion trajectory, determine the second flight distance of the UAV cluster from the correction starting point to the second intersection point in the third motion trajectory, and adjust the flight speed of the UAV cluster according to the difference between the second flight distance and the first flight distance, as well as the flight time, to ensure that the UAV flies according to the preset time schedule.

[0064] Refer to Figure 3 and Figure 4 , combined with specific examples, illustrate the processes of S1110~S180 above. Among them, Figure 3 and Figure 4These are schematic diagrams of application scenarios of the UAV swarm cooperative control method. In the embodiments of this application Figure 3 and Figure 4 are top views on the horizontal plane.

[0065] In one example, as shown in Figure 3 or Figure 4 After generating a high-precision 3D map using the lidar and RGB-D camera fusion technology, the server in the command center plans the movement route of the UAV swarm according to the 3D map, generates the first movement trajectory, and sends the first movement trajectory to the UAV swarm. Among them, Figure 3 or Figure 4 The circle at point A represents the UAV swarm, which is the first distribution sphere obtained by the safety space construction module calculating the radius of the formation envelope sphere according to the real-time distribution density of the UAV formation. The dotted line drawn from the circle at point A is part of the first movement trajectory; during the flight of the UAV swarm, based on the acquisition of the millimeter-wave radar and the operation of the lightweight YOLOv7 target detection model (frame rate ≥ 30fps) on the airborne FPGA, it is recognized that there are obstacles ahead. The obstacles are as shown in Figure 3 or Figure 4 The circle at point B in. It is the second distribution sphere obtained by the safety space construction module calculating the radius of the obstacle envelope sphere according to the volume of the obstacle; the artificial intelligence prediction module predicts the movement path of the obstacle based on the movement state of the obstacle through a pre-trained dynamic threat prediction model (based on the LSTM network and game theory) to obtain the second movement trajectory. Among them, the second movement trajectory can be referred to as Figure 3 or Figure 4 The dotted line drawn from the circle at point B in; the safety space construction module further performs spatio-temporal intersection analysis on the movement trajectories of the first dynamic envelope sphere and the second predicted envelope sphere based on the extended Kalman filter algorithm, determines the collision point coordinates and the collision time window, and generates a dangerous sphere according to the sum of the radii of the two spheres. Among them, the collision point is as shown in Figure 3 or Figure 4 At point C in, and the dangerous sphere is as shown in Figure 3 or Figure 4 The circle at point D in; the two intersection points of the first movement trajectory and the dangerous sphere D are respectively represented as the first intersection point a1 and the second intersection point a2 in Figure 3 , where the time for the UAV swarm to fly to the first intersection point a1 is earlier than the time to fly to the second intersection point a2.

[0066] When correcting the first movement trajectory according to the dangerous space to obtain the third movement trajectory of the UAV swarm, Figure 3 It is shown in that the second intersection point a2 and the dangerous sphere form a second tangent plane. The second tangent plane is as shown in Figure 3 The dotted line passing through point a2 in ( Figure 3 is a plan view. In Figure 3The second cross-section is represented by a straight line), and a perpendicular plane is made to the second cross-section to obtain Figure 3 Another dotted line perpendicular to the dotted line passing through point a3 in, the dotted line passing through point a3 is the second perpendicular plane, the intersection point a3 of the second perpendicular plane and the first motion trajectory is the starting point for correcting the first motion trajectory, and the second intersection point a2 is the ending point for correcting the first motion trajectory.

[0067] When correcting the first motion trajectory according to the dangerous space to obtain the third motion trajectory of the UAV cluster, Figure 4 It is shown in that the first intersection point a1 and the dangerous sphere form a first cross-section, and the first cross-section is as Figure 3 shown by the dotted line passing through point a1 in ( Figure 4 is a plan view, in Figure 4 the first cross-section is represented by a straight line), and a perpendicular plane is made to the first cross-section to obtain Figure 4 Another dotted line perpendicular to the dotted line passing through point a3 in, the dotted line passing through point a3 is the first perpendicular plane, the intersection point a3 of the first perpendicular plane and the first motion trajectory is the ending point for correcting the first motion trajectory, and the first intersection point a1 is the starting point for correcting the first motion trajectory. Take a1 as the starting point for correction.

[0068] Among them, the cross-section direction is determined by the tangency characteristics of the dangerous sphere surface, representing the direction of the shortest safe path around the obstacle; the perpendicular plane direction is determined by the orthogonal offset of the original UAV trajectory, representing the maximum safe distance adjustment benchmark; point a3 is the intersection point of the perpendicular plane direction and the first motion trajectory, serving as the starting point for high-safety space obstacle avoidance or the ending point for emergency obstacle avoidance; points a1 / a2 are the intersection points of the cross-section direction and the original trajectory, serving as the starting point for emergency obstacle avoidance or the path return anchor point.

[0069] In the obstacle avoidance scenario, the space around the dangerous sphere is divided into two categories according to the safety level: low-safety space, the area adjacent to the sphere surface (near the cross-section direction). When directly determining the correction trajectory along the cross-section, the UAV cluster has a risk of being unable to cope with the sudden change of the obstacle trajectory (such as acceleration, turning) due to insufficient response time; high-safety space, the area far from the sphere (the offset range of the perpendicular plane direction). If the path is corrected along the perpendicular plane direction, the redundant detour distance may lead to the loss of mission timeliness. This solution constructs a trajectory correction logic that takes into account both safety and timeliness through the orthogonal geometric constraints of the cross-section and the perpendicular plane. The specific technical implementation is as follows.

[0070] Solution 1, Progressive obstacle avoidance in high-safety space (corresponding to Figure 3), the correction starting point a3 can trigger path adjustment from a high - safety - level space based on the orthogonal offset rule in the vertical plane direction. The vertical plane direction ensures that the volume extension of the UAV cluster maintains the maximum safety distance from the dangerous sphere. The correction ending point a2 can bypass the sphere along the tangent plane direction, generate the shortest geometric path using the tangency characteristic between the tangent plane and the sphere, and smoothly return to the original trajectory at point a2 to ensure the continuity of the formation structure. That is, the a3 point in Figure 3 can be determined as the correction starting point of the first motion trajectory, and a2 as the correction ending point, so that the UAV cluster can timely adjust the orthogonal offset in the vertical plane direction at point a3, realize geometric obstacle avoidance along the vertical plane approaching the tangent plane path from a high - safety - level space, facilitate reserving a fault - tolerance space for obstacle trajectory prediction using the tangency characteristic between the tangent plane and the sphere, and enable the cluster to smoothly return to the original trajectory with the minimum curvature radius after bypassing the sphere along the tangent plane at point a2, maintaining the rigid stability of the formation structure.

[0071] Solution 2: Emergency truncation in a low - safety - level space (corresponding to Figure 4 ). The correction starting point a1 can trigger an emergency turn in the tangent plane direction in the collision - approaching area, force path truncation using the tangency characteristic of the tangent plane, and generate the shortest bypass path. The correction ending point a3 can terminate the boundary through orthogonal constraint in the vertical plane direction to ensure a quick realignment with the original flight path after leaving the dangerous area and avoid secondary path deviation. That is, the a1 in Figure 4 can be determined as the correction starting point of the first motion trajectory, and a3 as the correction ending point, so as to quickly truncate and reconstruct the path by triggering an emergency turn in the tangent plane direction in a low - safety - level space, facilitate using the orthogonality between the vertical plane direction and the tangent plane to constrain the obstacle - avoidance termination boundary (point a3), and enable the UAV cluster to realign with the preset flight path based on the vertical plane direction after leaving the dangerous area at point a3, avoiding secondary trajectory conflicts caused by excessive path correction.

[0072] The correction starting point a3 can force the UAV cluster to start obstacle avoidance from a high - safety - level space through orthogonal offset in the vertical plane direction. The orthogonality of the vertical plane ensures the maximization of the volume safety distance. The correction ending point a2 can enable the UAV cluster to bypass the sphere along the tangent plane direction. The geometric shortest - path characteristic of the tangent plane guarantees the bypass efficiency. The correction starting point a1 can trigger an emergency turn in the tangent plane direction in a low - safety - level space. The tangency characteristic of the tangent plane forces the path length to be the shortest, avoiding mission timeout. The correction ending point a3 can terminate the boundary through the vertical plane direction constraint. The orthogonality between the vertical plane and the tangent plane prevents excessive path deviation and ensures a quick return to the original flight path after obstacle avoidance.

[0073] Therefore, when it is found that the obstacle B is far from the UAV cluster A, a3 can be used as the starting point for correction, and a2 can be used as the ending point for correction, so as to collect as much motion data of the obstacle B as possible, update the second motion trajectory, and update the dangerous sphere C, thereby ensuring the accuracy of obstacle avoidance; when it is found that the obstacle B is close to the UAV cluster A, a1 can be used as the starting point for correction, and a3 can be used as the ending point for correction to avoid obstacles as soon as possible.

[0074] In another alternative way, the point a3 in Figure 3 can also be used as the starting point for correcting the first motion trajectory, and the point a3 in Figure 4 can be used as the ending point for correcting the first motion trajectory.

[0075] By treating the UAV cluster as a whole and realizing obstacle avoidance for moving obstacles as a whole, not only can obstacle avoidance be achieved, but also the original shape of the UAV can be maintained synchronously and collisions between multiple UAVs in the cluster can be prevented.

[0076] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. Among them, the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, abbreviated as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, abbreviated as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, abbreviated as EPROM), programmable read-only memory (Programmable Red-Only Memory, abbreviated as PROM), read-only memory (Read-Only Memory, abbreviated as ROM), magnetic memory, flash memory, magnetic disk or optical disk. These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured article including an instruction device, and the instruction device implements the functions specified in Figure 1 one process or multiple processes and / or boxes Figure 1 one box or multiple boxes.

[0077] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A drone swarm cooperative control system, characterized in that, Comprising: A multi-source perception unit, integrating lidar, RGB-D camera and millimeter-wave radar array, for generating a three-dimensional map with centimeter-level accuracy and real-time tracking of dynamic obstacles; A collaborative decision-making unit, including: An artificial intelligence prediction module, for predicting the collision point based on the motion state of the obstacle; A safety space construction module, for generating a dynamic safety sphere according to the formation envelope size and the obstacle volume; An artificial intelligence correction module, for constructing a safety section and generating an optimized path; A dynamic communication unit, for adopting a hybrid networking with adaptive switching between 5G and LoRa for cluster data interaction.

2. The system according to claim 1, wherein The working modes of the multi-source perception unit include: Lidar and RGB-D camera, for generating a three-dimensional map with centimeter-level accuracy through calibration parameter fusion; Millimeter-wave radar array, for continuously tracking dynamic obstacles within a preset detection distance and estimating their motion speed and direction; Weather sensors installed on each drone, for real-time collecting wind speed, rainfall intensity and electromagnetic interference intensity to form an environmental disturbance parameter set.

3. The system according to claim 2, wherein The operating mechanism of the dynamic communication unit includes: When the signal-to-noise ratio of the 5G channel is lower than the threshold or the bit error rate exceeds the preset limit, automatically switch to the LoRa communication link; Adopt a security protocol compliant with the FIPS140-2 standard for data transmission, dynamically generate encryption keys and update them periodically.

4. The system according to claim 1, characterized in that, The collaborative decision-making unit contains a heterogeneous computing architecture: An airborne FPGA chip runs a lightweight object detection model for real-time obstacle recognition; A cloud GPU cluster deploys a deep reinforcement learning model to optimize the obstacle avoidance strategy parameters in combination with historical flight data.

5. The system according to claim 4, characterized in that, The training process of the lightweight object detection model is as follows: Input the drone obstacle training data set into the original YOLOv7 model; Based on the contribution degree ranking of feature map activation, remove the redundant channels in the backbone network with contribution degree lower than the preset channel threshold; Build a teacher-student architecture in the model after removing redundant channels, and through a multi-scale feature alignment loss function, force the student model to imitate the spatial attention distribution of the teacher model in the feature pyramid layer.

6. The system according to claim 4, wherein The operations performed by the artificial intelligence correction module include: Analyze the relative position weights of each drone in the formation based on the attention mechanism to generate an optimal path candidate set for maintaining the formation; Adopt a hybrid optimization algorithm to perform multi-objective trade-offs on the flight trajectory under the constraint of the safety section, and the multi-objectives include minimum energy consumption, shortest time and flight path smoothness.

7. The system according to claim 6, wherein The safety space construction module is specifically used for: Calculate the formation envelope sphere radius according to the real-time distribution density of the drone formation, and determine the safety margin in combination with the predicted value of the obstacle volume; Construct a safety space that shrinks or expands over time with the collision time window as the time dimension.

8. The system according to claim 7, wherein The artificial intelligence correction module also includes: A speed adaptive adjustment component, which dynamically adjusts the flight speed according to the length difference between the corrected path and the original path, satisfying: When the corrected path is extended, increase the speed at the maximum allowable acceleration within the remaining flight time; When the corrected path is shortened, maintain the overall synchronization of the formation for uniform or deceleration adjustment.

9. The system according to claim 1, wherein The system also includes an energy optimization module, and its working method is: Calculate the power distribution scheme of each UAV through a convex optimization algorithm to minimize the total energy consumption of the cluster; Combine the genetic algorithm to search for a local optimal solution that meets the endurance requirements under power constraints.

10. A method for collaborative control of an unmanned aerial vehicle (UAV) cluster, characterized in that, Based on the system described in any one of claims 1-8, the method includes: Establish an environmental dynamic model through multi-sensor fusion; Perform real-time collision risk detection at the edge node to identify safety threat events that reach the threshold; Generate an obstacle avoidance trajectory correction scheme through cloud-edge collaborative computing; Use hybrid networking communication to synchronously update the flight control instructions of all UAVs in the cluster.

Images