Dove flock positive and negative feedback mechanism-simulated manned aerial vehicle / unmanned aerial vehicle pursuit deep learning control method

Abstract

The invention discloses a manned aircraft / unmanned aerial vehicle pursuit deep learning control method simulating a pigeon positive and negative feedback mechanism. The method comprises the following steps: step 1, giving mathematical models of a manned aircraft and an unmanned aerial vehicle and performing environment initialization; step 2, building a quaternion PID controller of the unmanned aerial vehicle; 3, a manned aerial vehicle / unmanned aerial vehicle hierarchical decision-making architecture is established; step 4, mathematical modeling of a pigeon positive and negative feedback mechanism; 5, designing a manned aerial vehicle / unmanned aerial vehicle cluster reward function simulating positive and negative feedback of the pigeon flock; 6, planning and designing an escape unmanned aerial vehicle path; step 7, predicting an escape unmanned aerial vehicle path; and 8, generating a manned aerial vehicle / unmanned aerial vehicle cluster pursuit learning strategy based on the actor-commentator network. According to the invention, an efficient, flexible and extensible bionic intelligent pursuit control framework is constructed, and the pursuit task execution efficiency and cooperative combat capability of the manned aerial vehicle / unmanned aerial vehicle heterogeneous cluster in a complex dynamic environment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to a human-in-the-loop / unmanned aerial vehicle (UAV) pursuit-evasion deep learning control method based on a pigeon-like flocking positive-negative feedback mechanism, and belongs to the field of UAV autonomous flight control. BACKGROUND

[0002] In recent years, the rapid development of UAV swarm technology has brought revolutionary changes to the fields of military confrontation, search and rescue, target tracking, etc. However, in the pursuit-evasion confrontation task facing complex dynamic environments, traditional UAV swarm control methods still have significant limitations. Existing multi-UAV pursuit-evasion methods are mainly based on game theory, control theory, and heuristic algorithms. Although game theory methods can model the pursuit-evasion problem as a mathematical game and develop strategies by solving the Nash equilibrium, they require complete environmental information and accurate mathematical models, which lack adaptability in random interactive environments, partially observable conditions, and dynamic changing scenarios. Control theory methods use optimal control theory to design pursuit-evasion trajectories, such as the Hamilton-Jacobi-Isaacs equation, but have high computational complexity and are difficult to handle high-dimensional state spaces and nonlinear dynamic constraints. Heuristic algorithms such as artificial potential field and particle swarm optimization are simple to implement but lack theoretical guarantees and are prone to local optima in complex environments. The common problem of these traditional methods is the inability to effectively integrate human wisdom and machine intelligence, lacking adaptive learning ability in complex confrontation environments. Simply relying on pre-set algorithms for UAV swarms often lacks sufficient intelligent decision-making ability and real-time adaptability in the face of highly uncertain and rapidly changing confrontation environments.

[0003] A human-in-the-loop / UAV heterogeneous system provides a more powerful solution for complex task execution by integrating the unique advantages of both platforms. UAV swarms have the advantages of low cost, flexible deployment, and the ability to perform dangerous tasks, enabling them to quickly cover large areas and perform repetitive tasks. In pursuit-evasion confrontation, they can quickly penetrate target areas for reconnaissance and surveillance, gathering basic intelligence such as target location and threat distribution. Human pilots, on the other hand, have complex decision-making capabilities, situational awareness advantages, and flexible response capabilities, enabling them to handle highly uncertain and complex scenarios. In pursuit tasks, they can quickly identify key targets based on their rich tactical knowledge, adjust combat plans flexibly, and maintain effective command and control in complex environments. In dynamic environments such as disaster rescue, relying solely on pre-set algorithms for UAVs often fails to cope with unexpected situations and complex decision-making requirements. Human pilots can provide critical tactical decision support, including identifying real and false targets, assessing threat levels, developing attack timing, and coordinating multi-UAV penetration paths, ensuring the coordinated combat effectiveness of heterogeneous swarms in high-intensity confrontation. Therefore, building a human-in-the-loop / UAV heterogeneous collaborative system can fully leverage the intelligent decision-making advantages of human pilots and the execution efficiency of UAVs, achieving optimal task execution results through "human-machine collaboration and complementary advantages."

[0004] In nature, pigeon groups have developed unique positive and negative feedback learning mechanisms over long-term evolution, exhibiting exceptional wisdom in collective flight. The positive feedback mechanism of pigeon groups mainly reflects the excitation and amplification of environmental information. When beneficial information is discovered, group members quickly respond through information transmission and behavior imitation, exciting the entire group to gather and act in the target direction, thereby achieving rapid exploration and utilization of favorable environmental factors. The negative feedback mechanism, on the other hand, serves to converge and regulate. When group behavior becomes dispersed or deviates, individual interactions and adjustments bring the group back to a stable state, maintaining the tightness and consistency of the flight formation. The organic combination of this positive and negative feedback mechanism enables pigeon groups to both acutely perceive and respond to environmental changes, while maintaining the stability and coordination of group behavior. Applying the positive and negative feedback mechanisms of pigeon groups to pursuit and evasion tasks of manned / unmanned aircraft can improve the consistency and formation tightness of heterogeneous groups in pursuit and evasion tasks at the micro level through the negative feedback learning mechanism, and improve the exploration ability of the group in complex unknown environments and the utilization ability of known information at the macro level through the positive feedback mechanism, thereby effectively combining group intelligence and individual decision-making, and providing a new biological intelligent inspiration for improving the pursuit and evasion capabilities of manned / unmanned aircraft heterogeneous systems in complex environments. SUMMARY

[0005] The purpose of the present application is to provide a kind of manned / unmanned aircraft pursuit and evasion deep learning control method of pigeon group positive and negative feedback mechanism, to provide a kind of intelligent efficient manned aircraft and unmanned aircraft cooperative pursuit and evasion control method, to improve the task execution efficiency and safety of unmanned aircraft cluster in complex dynamic environment. The system introduces the advantages of intelligent decision-making of manned aircraft and autonomous control of unmanned aircraft, and the characteristics of pigeon group positive and negative feedback mechanism in biological world, to build an efficient, flexible and scalable heterogeneous cooperative control framework.

[0006] A kind of manned / unmanned aircraft pursuit and evasion deep learning control method of pigeon group positive and negative feedback mechanism, the implementation steps are as follows:

[0007] In the present method, the following concepts are defined as follows:

[0008] Manned / unmanned aircraft cluster: refers to manned aircraft and unmanned aircraft in a task scene at the same time, by formulating task target priority, adjusting reward function according to battlefield situation change, changing the strategy of 3 or more unmanned aircraft cluster.

[0009] Unmanned aircraft cluster: refers to a cluster formed by 3 or more unmanned aircraft, and the dynamic model of each unmanned aircraft is formula (1)-(3).

[0010] Step one: given manned aircraft and unmanned aircraft mathematical model and environment initialization

[0011] The manned vehicle and the unmanned vehicle model are the same, both are six degrees of freedom quadrotor unmanned vehicle model, referred to as unmanned vehicle model. Given the unmanned vehicle model, the linear acceleration and angular acceleration are shown in equation (2) and equation (3) respectively:

[0012]

[0013]

[0014] wherein, is the position of the unmanned vehicle in the world coordinate system, is the velocity vector of the unmanned vehicle, g is the gravity acceleration, R is the rotation matrix from the body coordinate system to the world coordinate system, f j is the force generated by the jth rotor, m is the mass of the unmanned vehicle, is the angular velocity vector of the unmanned vehicle, is the roll angle, θ is the pitch angle, ψ is the yaw angle, I = diag(I x ,I y ,I z ) is the moment of inertia of the unmanned vehicle, the moment acts on the body coordinate system, and is expressed as wherein τ j is the moment generated by the jth rotor, r pos,j is the position of the jth rotor in the body coordinate system. The rotor speed is modeled as a first-order system based on a time constant T m , and is expressed as Ω cmd,j is the desired speed of the jth rotor. is the orientation quaternion of the unmanned vehicle, and the state of the unmanned vehicle is expressed as

[0015] Step two: building of the unmanned vehicle quaternion PID controller

[0016] The attitude of the quadrotor is represented by the quaternion q = [q0, q1, q2, q3] T , wherein q0 is the scalar part, q1, q2, q3 are the vector parts, and satisfy the normalization constraint: q0 2 + q1 2 + q2 2 + q3 2 = 1. The conversion relationship between the quaternion and the Euler angle ω is:

[0017]

[0018] When the desired thrust T d and the desired Euler angle are given, the Euler angle error e = ω d- ω, 4 motor commands are obtained by PID control.

[0019] Step three: manned / unmanned hierarchical decision architecture

[0020] A hierarchical decision architecture is constructed, in which the manned aircraft performs situation awareness, is responsible for threat assessment and target priority determination, and the unmanned aircraft cluster is responsible for specific pursuit path planning and execution. The target threat assessment function is expressed as:

[0021] Thr i = w1 x Dis i + w2 x Vel i + w3 x Eva i + w4 x Val i (5)

[0022] where Dis i is the distance from target i to the boundary, Vel i is the escape speed of target i, Eva i is the maneuverability assessment of target i, Val i is the strike value of target i, and w1, w2, w3, w4 are weight coefficients.

[0023] The target priority sorting algorithm is expressed as:

[0024] Prio i = a x Thr i + b x Urg i + g x SucRate i (6)

[0025] where Urg i is the urgency of target i (based on the escape progress), SucRate i is the expected success rate under the current unmanned aircraft configuration, a, b, g are dynamic adjustment factors.

[0026] According to the number and characteristics of targets, more unmanned aircraft are allocated to high-threat targets, while considering the remaining energy and current position of the unmanned aircraft. The manned aircraft divides the unmanned aircraft cluster into two groups according to the above two indexes of threat assessment and target priority determination.

[0027] Step four: mathematical modeling of pigeon positive and negative feedback mechanism

[0028] The positive feedback mechanism of pigeons mainly reflects the excitation and amplification of environmental information. When beneficial information is found, group members will quickly respond through information transmission and behavior imitation, excite the whole group to gather and act towards the target direction, so as to realize the rapid exploration and utilization of beneficial environmental factors, and establish the mathematical expression of positive feedback mechanism:

[0029]

[0030] where β i is the positive feedback amplification coefficient, I i (t) is the environmental favorable information.

[0031] The negative feedback mechanism of pigeon group plays a role of convergence and adjustment. When the group behavior appears dispersion or deviation, through the interaction and adjustment among individuals, the pigeon group re-converges to the stable state, maintains the compactness and consistency of the flight formation, and establishes the mathematical expression of negative feedback mechanism:

[0032] F negative = -γ n ×D(t) (8)

[0033] where γ n is the negative feedback adjustment coefficient, and D(t) is the individual dispersion degree in pigeon group.

[0034] Step five: design of human-in-the-loop / UAV swarm reward function based on pigeon group positive and negative feedback

[0035] The mathematical expression (7) of positive feedback mechanism is mapped into the human-in-the-loop / UAV swarm to design the macro positive feedback reward function of pigeon group simulation:

[0036] R e = ε1×A x / A total + ε2×exp(-C risk ) (9)

[0037] where ε1, ε2 are exploration incentive coefficients, A x is the area of new exploration region, A total is the area of the whole region, and C risk is the collision risk.

[0038] The center position of the kth group of UAVs is calculated as:

[0039]

[0040] The maximum radius r mk of the UAVs from the center position is calculated. In order to make the radius not exceed the threshold r d , the mathematical expression (8) of negative feedback mechanism is mapped into the human-in-the-loop / UAV swarm to design the micro negative feedback reward function of pigeon group simulation:

[0041] R c = -λ1||r mk -r d || 2 (11)

[0042] Where λ1 is the negative feedback coefficient.

[0043] When manned and unmanned aircraft are in the same group, design a coordinated reward function for manned / unmanned aircraft:

[0044] R o =-ν1×|V manned -V unmanned_avg |-ν2×sync de (12)

[0045] sync de =|p manned -p unmanned_avg | (13)

[0046] Among them, V manned For manned aircraft speed, V unmanned_avg sync is the average speed of the drone swarm. de For synchronization deviation measurement, ν1 and ν2 are coordination adjustment coefficients, p manned For the location of the manned aircraft, p unmanned_avg The average position of the drone swarm.

[0047] The capture reward function is set as follows:

[0048]

[0049] Where N is the total number of drones, a ai The reward for capturing the i-th drone is a, where a is the reward for capturing an escaped drone. ai =1, otherwise 0.

[0050] The total reward function is then:

[0051] R total =R c +R o +R e +R a (15)

[0052] Step Six: Escape Drone Path Planning and Design

[0053] The escape path of the unmanned aerial vehicle (UAV) is constructed using the artificial potential field method. The total potential force is calculated as follows:

[0054] F total =F pursuer +F boundary +F obstacle (16)

[0055] Among them, F pursuer The repulsive force F generated by the pursuer boundary F is the repulsive force generated by the boundary constraints. obstacleRepulsion force generated for obstacles.

[0056] When the pursuers are within the detection range of the escaping drone and are not blocked by obstacles, the pursuer repulsion force is generated, which is calculated as follows:

[0057]

[0058] target rposi = p dronei -p target (18)

[0059] where ni is the number of pursuers, target rposi is the distance between the pursuer and the escaping drone, p dronei is the position of pursuer i, p target is the position of the escapee, and ε is the protection from zero, ε = 1e-5.

[0060] The boundary force is calculated as follows:

[0061] F boundary = [F xy , F zup + F zdown ] (19)

[0062]

[0063] force xy = tar xy / (||tar xy || + ε) (21)

[0064]

[0065]

[0066] where S is the diagonal length of the set scene, is the horizontal position of the escaping drone, h max is the set height, tar z is the height of the escaping drone.

[0067] The obstacle repulsion force is calculated as follows:

[0068] F obstacle = [F obs_xy , 0] (24)

[0069]

[0070] where nk is the total number of obstacles within the detection range, is the position vector of the kth obstacle and the escaping drone, dist boundTo determine the distance between the escape drone and the surface of the obstacle.

[0071] The speed command for the escaping drone is:

[0072] velo cmd =k vel ×F total / ||F total || (26)

[0073] Where, k vel is the velocity constant.

[0074] Step 7: Predicting the Path of the Escaped Drone

[0075] A Long Short-Term Memory (LSTM) network is used to predict the path of an escaping drone. The network's historical input consists of t time steps, with each time step's input features including the positions of all pursuing drones, the position and velocity of the escaping drone, and the positions of obstacles. The network outputs m prediction time steps, including the positions of the escaping drones.

[0076] Step 8: Generation of Manned / Unmanned Swarm Chase Learning Strategy Based on Actor-Critic Network

[0077] First, an observation encoder is designed based on a self-attention mechanism. The observation data for UAV i includes: its own state information o self Other drone information other Obstacles and environmental information ob The trajectory prediction for the escaping drone is included in its own state information. These three sets of data are independently encoded using a multilayer perceptron with 128 dimensions to generate feature vectors h. self h other h ob Next, h self Feature vector h from other drones other and the embedding of obstacles h ob They are fed together into a multi-head self-attention module to capture the correlation features h between these embeddings. attn To further enhance its embedded information h self Feature fusion is performed using a multilayer perceptron:

[0078] h enhanced =h attn +h self (27)

[0079] h final =MLP fusion (h enhanced (28)

[0080] In the actor network, h finalParametrized by a Gaussian distribution, in the critic network, h final The input is generated into a multi-layer perception to generate a final state value estimate.

[0081] The present application provides a pigeon flock positive and negative feedback mechanism manned / unmanned aerial vehicle pursuit and escape deep learning control method. The main advantages of the heterogeneous cooperative control system are: first, the present application provides a complete pigeon flock positive and negative feedback manned / unmanned aerial vehicle cluster pursuit and escape control workflow, including pigeon flock positive and negative feedback mechanism reward function design and deep learning strategy optimization, further improving the pursuit efficiency of manned / unmanned aerial vehicles; second, the present application provides a manned / unmanned aerial vehicle heterogeneous pursuit and escape cooperative control method, which organically integrates human high-level tactical decision-making intelligence and unmanned aerial vehicle precise bottom-level execution capability, realizes the complementary layered cooperative architecture, and has intelligent pursuit and escape situation awareness, adaptive strategy generation and real-time dynamic adjustment capability; third, the present application designs a manned / unmanned aerial vehicle heterogeneous decision learning method based on a deep reinforcement learning framework, supports multi-agent cooperative learning and online strategy optimization, and has good scalability, modular configuration capability and environmental adaptability. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 is a manned / unmanned aerial vehicle cluster pursuit and escape control flowchart.

[0083] Figure 2 is an escape unmanned aerial vehicle potential field method trajectory diagram. DETAILED DESCRIPTION

[0084] See Figure 1 As shown in the figure, the effectiveness of the pigeon flock positive and negative feedback mechanism manned / unmanned aerial vehicle pursuit and escape deep learning control method is verified by a specific example. The example uses an NVIDIA RTX 6000 Ada graphics card, based on an Ubuntu 20.04 system, and an Isaac Sim 2022.2.0 version. The implementation flowchart of the pigeon flock positive and negative feedback mechanism manned / unmanned aerial vehicle pursuit and escape deep learning control method is as shown in Figure 1 The specific practice steps are as follows:

[0085] Step 1: Given the mathematical model of manned / unmanned aerial vehicles and environment initialization

[0086] According to formulas (1)-(3), 4 unmanned aerial vehicles are set, of which 1 is manned, 1 is an escape unmanned aerial vehicle, and the other 2 are pursuer unmanned aerial vehicles. The speed of the manned and pursuer unmanned aerial vehicles is 1 m / s, and the speed of the escape unmanned aerial vehicle is 1.2 m / s. The mass of all unmanned aerial vehicles is 1 kg. The initial positions of the manned and pursuer unmanned aerial vehicles are [0.6, 0, 0.5] m, [0.8, 0, 0.5] m, [0.8, -0.2, 0.5] m, and [0.8, 0.2, 0.5] m. The position of the escape unmanned aerial vehicle is [-0.8, 0, 0.5] m. The obstacles are cylinders, the number of which is 4, the radius of each cylinder is 0.1 m, the height is 1.2 m, and the positions of the cylinders are [0, 0.15, 0.6] m, [0, -0.15, 0.6] m, [0, 0.45, 0.6] m, and [0, -0.45, 0.6] m.

[0087] The radius of the field S is 0.9 m. When the distance between any manned and pursuer unmanned aerial vehicle and the escape unmanned aerial vehicle is less than 0.3 m, it indicates that the pursuit is successful. The detection radius of all unmanned aerial vehicles is 50 m, and the collision radius is 0.07 m. The acceleration of gravity g = 9.81 m / s 2 , and the simulation step is 0.01 s.

[0088] Step two: Building of the quaternion PID controller of the unmanned aerial vehicle

[0089] The control parameters in the PID are set as K p =[250, 250, 120], K I =[500, 500, 16.7], and K D =[2.5, 2.5, 0]. After the desired thrust T d and the desired Euler angle are given, the Euler angle error is calculated, and 4 motor commands are obtained through PID control.

[0090] Step three: Hierarchical decision-making architecture of manned / unmanned aerial vehicles

[0091] The hierarchical decision-making architecture is built, in which the manned aerial vehicle is responsible for situation awareness, threat assessment, and target priority determination, and the unmanned aerial vehicle cluster is responsible for specific pursuit and escape path planning and execution. The target threat assessment is performed according to formula (5), the target priority is sorted according to formula (6), and more unmanned aerial vehicles are allocated to high-threat targets according to the number and characteristics of the targets, while considering the remaining energy and current position of the unmanned aerial vehicles. Since there is only one escape unmanned aerial vehicle in this scenario, all unmanned aerial vehicles of our side are designated to pursue the escape unmanned aerial vehicle.

[0092] Step four: Mathematical modeling of pigeon positive and negative feedback mechanism

[0093] The positive feedback mechanism of pigeon group mainly reflects the excitation and amplification of environmental information. When beneficial information is found, group members will quickly respond through information transmission and behavior imitation, excite the whole group to gather and act towards the target direction, so as to realize the rapid exploration and utilization of beneficial environmental factors. The negative feedback mechanism of pigeon group plays a role of convergence and regulation. When the group behavior appears to be scattered or deviated, through the interaction and regulation between individuals, the pigeon group re-converges to a stable state, maintains the compactness and consistency of the flight formation, and follows equations (7)-(8) to establish the mathematical expressions of the positive and negative feedback mechanisms of pigeon group.

[0094] Step five: design of manned / unmanned aerial vehicle cluster reward function based on pigeon group positive and negative feedback

[0095] According to equations (7)-(8), the pigeon group positive and negative feedback mechanisms are mapped into the design of manned / unmanned aerial vehicle cluster reward function, and equations (9)-(10) are obtained. The center position of all chasing unmanned aerial vehicles is calculated, and the maximum radius of the unmanned aerial vehicles from the center position is calculated The reward function is calculated according to equations (9)-(15).

[0096] Step six: design of escape unmanned aerial vehicle path planning

[0097] As shown in Figure 2 , the artificial potential field method is used to construct the escape unmanned aerial vehicle path. The total potential field force is calculated according to equation (16). When the pursuer is within the detection range of the escape unmanned aerial vehicle and is not blocked by the obstacle, the repulsive force of the pursuer is generated, which is calculated according to equations (17)-(18). At the same time, the boundary force is calculated according to equations (19)-(23), the obstacle repulsive force is calculated according to equations (24)-(25), and the speed command of the escape unmanned aerial vehicle is calculated according to equation (26) through the total potential field force. The escape unmanned aerial vehicle is simplified to move directly according to the speed command.

[0098] Step seven: escape unmanned aerial vehicle path prediction

[0099] The long short-term memory network is used to predict the path of the escape unmanned aerial vehicle. The hidden layer is set to 1 layer with 64 hidden units, the network history input is 10 time steps, and the input features of each time step include the positions of all pursuer unmanned aerial vehicles, the position and speed of the escape unmanned aerial vehicle, and the positions of obstacles. The network outputs 5 predicted time steps, including the position of the escape.

[0100] Step eight: generation of manned / unmanned aerial vehicle cluster pursuit and escape learning strategy based on actor-critic network

[0101] First, an observation encoder based on a self-attention mechanism is designed. The observation data for UAV i includes its own state information, information about other UAVs, and obstacle environment information. The trajectory prediction for the escaping UAV is included in its own state information. These three sets of data are independently encoded using a multilayer perceptron with 128 dimensions to generate a feature vector h. self h other h ob Next, h self Feature vector h from other drones other and the embedding of obstacles h ob They are fed together into a multi-head self-attention module to capture the correlation features h between these embeddings. attn To further enhance its embedded information h self Using a multilayer perceptron, feature fusion is performed according to equations (27)-(28) to obtain h. final In the actor network, h final Parameterization is performed using a Gaussian distribution in the critic network, h final The input is fed into a multilayer perceptron to generate the final state value estimate.

Claims

1. A deep learning control method for manned / unmanned aerial vehicle (UAV) pursuit and escape based on a positive and negative feedback mechanism mimicking pigeon flocks, characterized in that: The steps of this method are as follows: Step 1: Given the mathematical models of manned and unmanned aircraft and the environment initialization; Step 2: Building a quaternion PID controller for the drone; Step 3: Hierarchical decision-making architecture for manned / unmanned aircraft; A hierarchical decision-making architecture is constructed, in which manned aircraft are responsible for situational awareness, threat assessment, and target priority determination, while drone swarms are responsible for specific pursuit and escape route planning and execution. Based on the number and characteristics of the targets, more drones are allocated to high-threat targets. At the same time, considering the remaining energy and current position of the drones, the manned aircraft divide the drone swarm into two groups based on the above threat assessment and target priority determination. Step 4: Mathematical Modeling of Positive and Negative Feedback Mechanism in Pigeon Flocks Mathematical expression for establishing a positive feedback mechanism: Where, β i I is the positive feedback amplification factor. i (t) represents environmentally favorable information; Mathematical expression for establishing a negative feedback mechanism: F negative =-γ n ×D(t) (8) Where, γ n t is the negative feedback adjustment coefficient, and D(t) is the individual dispersion in the pigeon flock; Step 5: Design of reward function for manned / unmanned swarms with positive and negative feedback similar to pigeon flocks; Step Six: Escape Drone Path Planning and Design; Step 7: Predict the escape route of the drone; A Long Short-Term Memory (LSTM) network is used to predict the path of an escaping drone. The network's historical input consists of t time steps, and the input features at each time step include the positions of all pursuing drones, the position and velocity of the escaping drone, and the positions of obstacles. The network output consists of m prediction time steps, including the positions of the escaping drones. Step 8: Generation of manned / unmanned swarm pursuit and escape learning strategies based on actor-critic network.

2. The method according to claim 1, characterized in that: In step three, the target threat assessment function is expressed as: Thr i =w1×Dis i +w2×Vel i +w3×Eva i +w4×Val i (5) Among them, Dis i Vel is the distance from target i to the boundary. i Let Eva be the escape velocity of target i. i For the assessment of the maneuverability of target i, Val i Let w1, w2, w3, and w4 be the strike value of target i, and w1, w2, w3, and w4 be the weighting coefficients. The target priority sorting algorithm is expressed as: Prio i =α×Thr i +β×Urg i +γ×SucRate i (6) Among them, Urg i SucRate represents the urgency of target i. i α represents the expected success rate under the current drone configuration, and β and γ are dynamic adjustment factors.

3. The method according to claim 1, characterized in that: The specific process of step five is as follows: Map the mathematical expression (7) of the positive feedback mechanism into the manned / unmanned aircraft cluster, and design a macroscopic positive feedback reward function similar to that of a pigeon flock: R e =ε1×A x / A total +ε2×exp(-C risk ) (9) Where ε1, ε2 are the exploratory incentive coefficients, and A x For the area of ​​the newly explored region, A total C represents the area of ​​the entire region. risk For collision risk; Find the center position of the k-th group of drones: And calculate the maximum radius of the group of drones from the center position. To ensure that the radius does not exceed the threshold r d By mapping the mathematical expression (8) of the negative feedback mechanism into a manned / unmanned aircraft swarm, a pigeon-like micro-negative feedback reward function is designed: Where λ1 is the negative feedback coefficient; When manned and unmanned aircraft are in the same group, design a coordinated reward function for manned / unmanned aircraft: R o =-ν1×|V manned -V unmanned_avg |-ν2×sync de (12) sync de =|p manned -p unmanned_avg | (13) Among them, V manned For manned aircraft speed, V unmanned_avg sync is the average speed of the drone swarm. de For synchronization deviation measurement, ν1 and ν2 are coordination adjustment coefficients, p manned For the location of the manned aircraft, p unmanned_avg The average position of the drone swarm; The capture reward function is set as follows: Where N is the total number of drones, a ai The reward for capturing the i-th drone is a, where a is the reward for capturing an escaped drone. ai =1, otherwise 0; The total reward function is then: R total =R c +R o +R e +R a (15)。 4. The method according to claim 1, characterized in that: The specific process of step six is ​​as follows: The escape path of the drone is constructed using the artificial potential field method; the total potential force is calculated as follows: F total =F pursuer +F boundary +F obstacle (16) Among them, F pursuer The repulsive force F generated by the pursuer boundary F is the repulsive force generated by the boundary constraints. obstacle The repulsive force generated by the obstacle; When the pursuer is within the detection range of the escaping drone and is not obstructed by obstacles, a repulsive force is generated, which is calculated as follows: target rposi =p dronei -p target (18) Where ni is the number of pursuers, and target rposi p is the distance between the pursuer and the escaping drone. dronei For the position of chaser i, p target The location of the escapee is given by ε, which is a division-by-zero protection factor, ε = 1e-5; The boundary force is calculated as follows: F boundary =[F xy ,F zup +F zdown ] (19) force xy =tar xy / (||tar xy ||+e) (21) Where S is the diagonal length of the scene. To escape the drone's horizontal position, h max To set the height, tar z To escape the drone's altitude; The repulsive force of the obstacle is calculated as follows: F obstacle =[F obs_xy ,0] (24) Where nk represents the total number of obstacles within the detection range. Let dist be the position vector between the k-th obstacle and the escaping drone. bound The distance between the escape drone and the surface of the obstacle; The speed command for the escaping drone is: bicycle cmd =k vel ×F total / ||F total || (26) Where, k vel is the velocity constant.

5. The method according to claim 1, characterized in that: The specific process of step eight is as follows: First, an observation encoder is designed based on a self-attention mechanism; the observation data of UAV i includes: its own state information o self Other drone information other Obstacles and environmental information ob The trajectory prediction of the escaping drone is included in its own state information; these three sets of data are independently encoded using a multilayer perceptron with 128 dimensions to generate a feature vector h. self h other h ob Next, h self Feature vector h from other drones other and the embedding of obstacles h ob They are fed together into a multi-head self-attention module to capture the correlation features h between these embeddings. attn To further enhance its embedded information h self Feature fusion is performed using a multilayer perceptron: h enhanced =h attn +h self (27) h final =MLP fusion (h enhanced ) (28) In the actor network, h final Parameterization is performed using a Gaussian distribution in the critic network, h final The input is fed into a multilayer perceptron to generate the final state value estimate.

Images