Multi-uav adaptive prediction cooperative hunting method in obstacle environment

Abstract

The application discloses a kind of multi-unmanned aerial vehicle adaptive prediction cooperative hunting methods under obstacle environment, this method will continuous acceleration control space be converted into the discrete candidate action set of polar coordinate sampling;Using kinematic prediction model deduces the future state corresponding to each candidate action;Using multi-dimensional heuristic scoring model covering convergence of surrounding situation, formation cooperation, smooth obstacle avoidance, candidate action parallel evaluation is completed, and the optimal acceleration instruction is screened.The application introduces tangential flow guiding mechanism, converts obstacle normal repulsion into smooth circumnavigation signal, realizes obstacle avoidance and target attraction logic decoupling, overcomes stagnation shock and local minimum problem caused by traditional potential field attraction-repulsion offset.The application can carry out online real-time hunting to maneuvering escape target, and has high robustness and hunting success rate in complex obstacle dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of aircraft guidance and control technology, specifically to a method for adaptive prediction and cooperative encirclement of multiple unmanned aerial vehicles (UAVs) in obstacle environments. Background Technology

[0002] With the rapid advancement of drone swarm technology, its applications in the civilian sector are becoming increasingly widespread and complex, placing higher demands on the swarm's autonomous decision-making and collaborative control capabilities in unknown dynamic environments, requiring greater real-time performance and robustness. Multi-drone collaborative encirclement, as an efficient group collaboration method, plays a core role in civilian tasks such as urban low-altitude security, protection of key areas and major events, public safety emergency response, and wildlife monitoring. Enhancing the collaborative encirclement capabilities of swarms is crucial for ensuring public airspace safety and improving the efficiency of comprehensive law enforcement and mission execution, and has broad application prospects in the field of civilian intelligent equipment.

[0003] Current methods for collaborative obstacle avoidance using drones can be broadly categorized into two types. One type is analytical control methods based on traditional geometric rules or artificial potential fields. These methods typically utilize the superposition of gravitational and repulsive fields to quickly generate control commands by solving for the potential field gradient. Many scholars have established mature theoretical frameworks in this field, capable of meeting the needs of real-time obstacle avoidance and collaboration in simple environments. However, limited by the predefined form of the potential field function, this method is prone to local minima leading to mission failure when facing multiple constraints, nonlinear dynamics, and complex obstacle distributions. Its adaptability in dynamic scenarios still needs improvement.

[0004] Another category is trajectory planning methods based on numerical optimization or heuristic search. Compared to traditional potential field methods, these methods offer significant advantages in handling complex boundary constraints and multi-machine collision avoidance due to the introduction of more advanced prediction models and evaluation functions. However, the computational cost of these methods typically increases exponentially with cluster size. Currently, most numerical methods are limited by cumbersome iterative optimization processes and high sensitivity to initial values, generally resulting in long computation times and difficulty in handling sudden target maneuvers, thus limiting their application in real-world online real-time containment tasks.

[0005] In particular, existing collaborative encirclement schemes rely on traditional potential field-based obstacle avoidance methods that directly superimpose the obstacle's normal repulsive force and the target's attractive force, resulting in direct collisional coupling between the attraction and repulsion forces. When a UAV approaches an obstacle while simultaneously moving towards the target, a zero-point resultant force is easily formed, causing the UAV to stagnate, its trajectory to oscillate at high frequencies, and it to become trapped in a local minimum deadlock. Furthermore, traditional schemes do not specifically address the vanishing potential field gradient problem during long-distance tracking and lack adaptive weight adjustment logic for the tracking-encirclement task phase. Numerical optimization methods, on the other hand, are limited by computational complexity and struggle to handle online real-time control under complex obstacles, target maneuvers, and multi-UAV formation constraints. These shortcomings collectively hinder the practical implementation of collaborative encirclement missions involving multiple UAVs in obstacle-prone environments. Summary of the Invention

[0006] In view of this, the present invention provides a multi-UAV adaptive prediction and cooperative encirclement method in obstacle environment. It introduces a hybrid potential field obstacle avoidance algorithm based on tangential flow guidance, which transforms the obstacle repulsion force into a smooth detour guidance signal, effectively decoupling the environmental constraints from the target attraction logic, avoiding oscillations or stagnation of UAVs during obstacle avoidance, and ensuring that UAVs can maintain accurate encirclement configuration and stable real-time response speed under complex geographical constraints.

[0007] To solve the above-mentioned technical problems, the present invention is implemented as follows.

[0008] A multi-UAV adaptive prediction and cooperative encirclement method in an obstacle environment includes: Step 1: Divide the acceleration amplitude and azimuth angle of the UAV into multiple energy levels to generate a candidate action set containing K candidate acceleration vectors. ;in, Candidate action set The first in Candidate acceleration vectors, and They are respectively The acceleration amplitude and azimuth angle; Step 2: For each drone executing the coordinated capture, based on the drone's current control cycle position... and speed Each candidate acceleration vector is determined using a kinematic prediction model. Predicted position in the next control cycle ; Step 3: The drone bases its actions on the candidate action set. Each candidate acceleration vector in and the predicted position generated in step 2 A multi-dimensional heuristic scoring model encompassing encirclement situation convergence, formation coordination, and smooth obstacle avoidance is employed to determine each candidate acceleration vector. The overall score; among which, the obstacle tangential flow guidance item corresponding to the smooth obstacle avoidance. The design is based on the predicted velocity generated by the candidate acceleration vector and the tangent vector when the UAV approaches the obstacle. The inner product; Step 4: Select the candidate acceleration vector with the highest comprehensive score as the optimal acceleration command for the current control cycle. ; Optimal acceleration command Substitute the data into the UAV dynamics model to drive the UAV to evolve to the next state; repeat steps 2-4.

[0009] Preferably, in step 1, dividing the acceleration amplitude and azimuth angle of the UAV into multiple energy levels involves: dividing the acceleration amplitude energy level between the maximum and minimum values ​​of the acceleration amplitude; and dividing the acceleration amplitude energy level between 0 and 10 ... The energy levels are divided into azimuth angles.

[0010] Preferably, in step 2, the kinematic prediction model is:

[0011] in, To apply acceleration vector The predicted position of the UAV in the next control cycle is then determined. To discretize the control time step.

[0012] Preferably, in step 3, the multidimensional heuristic scoring model is:

[0013] in, Candidate acceleration vector Overall score; These are the weighting coefficients of each potential field; The target capture guide term guides the drone to converge to the enclosing circle; As a cooperative interval term, the inter-machine repulsion force is adjusted by an exponential decay function, forcing the drone swarm to spontaneously form a geometrically symmetrical, equally spaced distribution formation during the encirclement phase; For the long-range heuristic guidance term targeting the target, the greater the absolute difference between the distance of the UAV's predicted position from the target and the expected encirclement circle, the smaller the long-range heuristic guidance term becomes, but it is never zero, providing a continuous and stable directional gradient when the UAV is far away from the target. Tangential flow scoring of obstacles is used to guide the drone to smoothly navigate around obstacles; when the drone approaches an obstacle... When calculating the normal vector tangent vector ; Tangential flow guidance for obstacles Predicted velocities generated for candidate acceleration vectors Approaching obstacles with drones Time tangent vector Inner product:

[0014] .

[0015] Preferably, the target capture guide item A Gaussian potential field is used to evaluate the radial deviation between the predicted position of the UAV and the ideal capture radius, guiding the UAV to converge to the encircling circle.

[0016] in, The UAV predicts its position in the next control cycle based on candidate acceleration vector predictions. The position vector of the target; The ideal encirclement radius; The standard deviation of the Gaussian function determines the range and steepness of the attraction field's influence.

[0017] Preferably, the cooperative interval term for:

[0018] in, The UAV predicts its position in the next control cycle based on candidate acceleration vector predictions. The position vector of the nearest neighbor drone; Indicates the predicted location of the drone Euclidean distance between the nearest neighbor drone; ideal inter-drone distance , For the ideal encirclement radius, The size of a drone swarm assigned to a single target for a capture mission.

[0019] Preferably, the long-range heuristic guide item for:

[0020] in, The UAV predicts its position in the next control cycle based on candidate acceleration vector predictions. The target's position vector; For the heuristic term gain coefficient, The ideal encirclement radius; The absolute difference between the distance of the drone's predicted position from the target and the expected encirclement circle.

[0021] Preferably, the method further includes: Based on the current control cycle position of the drone With the target location relative distance Dynamically adjust the co-interval term in the multidimensional heuristic scoring model Weighting coefficients ;if If the distance exceeds the set threshold, then Use the set low weighting coefficient To switch to the tracking phase, attraction takes precedence, guiding the drone swarm to converge rapidly toward the target; if If the distance is less than or equal to the set distance threshold, then Use the set high weighting coefficient In order to switch to the encirclement phase, overcome the local extreme value of target attraction, and force the drone to spread evenly along the encirclement circumference; The set distance threshold is , For the ideal encirclement radius, Distance buffer margin for determining the phase switching between the tracking and encirclement phases.

[0022] Preferably, the method further includes introducing a vectorized collision handling mechanism based on a position correction algorithm during the encirclement process: Real-time calculation of the Euclidean distance between drones, when any two drones Euclidean distance Below the safety threshold At that time, calculate the two drones Overlapping vectors ; The safe collision radius for a single drone;

[0023] in, and drones and Position vector in the current control cycle; For drones Apply displacement compensation:

[0024] in," " indicates an assignment operation; This indicates obtaining all information related to drones. Drones with a distance below the safety threshold Overlapping vectors , these overlapping vectors Add them together and divide by 2 to get the result for the drone. Displacement compensation applied.

[0025] Preferably, step 4 involves determining the optimal acceleration command. Then, further refine the optimal acceleration command. Make corrections:

[0026] in, The final acceleration command, after smoothing correction, is substituted into the UAV dynamics model; For instructions The amplitude; This is a direction smoothing factor used to control the direction of the command towards the current velocity. The degree of deflection is used to simulate real physical constraints.

[0027] Beneficial effects: (1) This invention introduces a hybrid potential field obstacle avoidance algorithm based on tangential flow guidance, which transforms the normal repulsive force of the obstacle into a smooth detour guidance signal, thereby decoupling the obstacle avoidance logic from the target attraction logic. This can prevent the UAV from stalling or oscillating in front of the obstacle, improve the smoothness of UAV swarm maneuvering in obstacle environments, and solve the problem of local minimum deadlock caused by the repulsive-attraction collision of the traditional potential field method.

[0028] (2) The present invention constructs a multidimensional heuristic scoring model that includes target capture guidance, cooperative interval, long-range heuristic guidance, and obstacle tangential flow scoring, and completes parallel prediction search by sampling discrete candidate action sets in polar coordinates. On the one hand, it uses the Gaussian potential field to suppress velocity overshoot and trajectory oscillation when approaching the target, and relies on the long-range heuristic guidance to eliminate the problem of gradient vanishing at long distances. On the other hand, it uses the cooperative interval with local distance perception to realize the spontaneous symmetrical layout of the cluster with equal spacing, without the need for complex global communication, and takes into account both the capture convergence accuracy and formation coordination performance.

[0029] (3) In a preferred scheme, physical flight safety during dense encirclement by overlapping vector displacement compensation is ensured.

[0030] (4) In a preferred scheme, the weight of the cooperative interval term is dynamically adjusted according to the relative distance between the UAV and the target to achieve a smooth transition from rapid convergence in the tracking phase to uniform configuration in the encirclement phase.

[0031] (5) In a preferred embodiment, the acceleration command is further smoothed to fit the real dynamic constraints of the UAV, and the control command can be output online in real time. It has strong robustness against maneuvering escape targets and is suitable for online capture missions in real scenarios. Attached Figure Description

[0032] Figure 1 This is a flowchart of the adaptive prediction-cooperative encirclement algorithm of the present invention; Figure 2 This is a schematic diagram of a multi-drone encirclement and capture operation, a specific example of the present invention; Figure 3 This is a convergence curve of the distance error between a swarm of multiple UAVs and the target in a specific embodiment of the present invention; Figure 4 This is a configuration symmetry curve diagram of a multi-UAV swarm in a specific embodiment of the present invention; Figure 5 This is a graph showing the maximum escape angle of the target in a specific embodiment of the present invention; Figure 6 This is a two-dimensional trajectory diagram of multi-team collaborative encirclement and capture in a dynamic multi-obstacle environment in a specific embodiment of the present invention. Detailed Implementation

[0033] This invention provides an adaptive predictive cooperative encirclement method for multiple UAVs in obstacle-prone environments. Addressing the deadlock and convergence challenges in multi-UAV cooperative encirclement missions under dynamic, non-cooperative, and complex obstacle avoidance environments, this invention proposes a multi-dimensional constrained adaptive predictive cooperative encirclement method based on an evaluation function-driven heuristic decision-making framework. This method first establishes a discrete dynamic model of the UAVs, transforming the continuous control space into a discrete set of candidate acceleration actions sampled in polar coordinates. It then predicts the future state of the UAVs under each candidate action and constructs a multi-dimensional heuristic scoring model that integrates target acquisition, formation coordination, long-range guidance, and tangential flow obstacle avoidance, evaluating and selecting the optimal control command in parallel. Furthermore, this invention employs vectorized dynamic collision detection displacement compensation and adaptive weight adjustment in a two-stage encirclement-tracking phase to complete configuration correction, iteratively achieving stable cooperative encirclement of maneuvering escape targets in obstacle-prone environments.

[0034] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0035] Figure 1 A flowchart of the multi-UAV adaptive prediction cooperative encirclement method under obstacle conditions of the present invention is shown. The method includes the following steps: Step 1: Construct a discrete dynamics model of the UAV.

[0036] In this invention, the UAV adopts a second-order discrete dynamics model, and its state update formula is as follows:

[0037] in, , , These represent the velocity vector, position vector, and acceleration vector of the UAV in the current control cycle, respectively. , These represent the velocity and position vectors for the next control cycle, respectively. This represents the discretized control time step of the system. This is an amplitude limiting function to ensure that the rate does not exceed [the specified value]. .

[0038] To prevent physically impossible instantaneous return of the drone at high speeds, the acceleration vector preferred as a command in this invention... The correction is made, and the correction formula is:

[0039] in, The preferred control command for the collaborative capture method of the present invention has the physical meaning of an acceleration vector, which is used to control the acceleration of the UAV. The final acceleration command, after smoothing correction, is substituted into the UAV dynamics model to drive the UAV to evolve to the next state. For expected instructions The magnitude of the amplitude; This is a direction smoothing factor used to control the direction of the command towards the current velocity. The degree of deflection is used to simulate real physical constraints.

[0040] Step 2: Construct the action space.

[0041] This step constructs a discretized candidate action set based on polar coordinate sampling: the acceleration amplitude and azimuth angle of the UAV are divided into multiple energy levels, generating a candidate action set containing K candidate acceleration vectors.

[0042] in, Candidate action set The first in Candidate acceleration vectors, and They are respectively The acceleration amplitude and azimuth angle.

[0043] The acceleration amplitude and azimuth angle of the UAV are divided into multiple energy levels: between the maximum and minimum values ​​of the acceleration amplitude, the energy levels are further divided into... Each acceleration amplitude energy level; in 0~ Divided by angular resolution Each azimuth energy level generates a series of energy levels. Candidate action set of action vectors This replaces the traditional continuous action generation network.

[0044]

[0045] In the formula, Represents the range of acceleration energy levels. The specified angular resolution.

[0046] Step 3: Predict future states.

[0047] In this step, each drone performing the coordinated capture is positioned according to its current control cycle. and speed Each candidate acceleration vector is determined using a kinematic prediction model. Predicted position in the next control cycle This enables a paradigm shift from continuous optimization to parallel search.

[0048] The kinematic prediction model is as follows:

[0049] in, Candidate action set The first in Candidate acceleration vectors, To apply acceleration vector The predicted position of the drone in the next control cycle. To discretize the control time step. and These are the actual position and velocity vectors of the UAV in the current control cycle, calculated using the UAV's discrete dynamics model.

[0050] Step 4: The drone uses the candidate action set Each candidate acceleration vector in and the predicted position generated in step three A multi-dimensional heuristic scoring model encompassing encirclement situation convergence, formation coordination, and smooth obstacle avoidance is employed to determine each candidate acceleration vector. The overall score.

[0051] To guide UAV swarms to achieve efficient target acquisition and maintain a stable encirclement posture in complex and dynamic environments, the heuristic scoring function is designed as a composite structure that aims to balance multiple objectives such as target acquisition accuracy, formation symmetry, flight safety constraints, and dynamic non-cooperative robustness.

[0052] Therefore, the multidimensional heuristic scoring model is designed to encompass multiple potential fields and structures, including encirclement situation convergence, formation coordination, long-range continuous guidance, and smooth obstacle avoidance. The specific model design is as follows: 1. Target Acquisition Guide : Target capture guide Used to guide the drone to converge to the surrounding circle.

[0053] To avoid velocity overshoot and trajectory oscillations that occur when approaching a target in traditional linear gravity models, this invention innovatively employs a Gaussian potential field to assess the predicted position and the ideal encirclement radius. The radial deviation guides the UAV to converge within the enclosing circle. A target acquisition guidance item is designed based on this radial deviation. for:

[0054] in, The UAV predicts its position in the next control cycle based on candidate acceleration vector predictions. The target's position vector; The ideal encirclement radius; The standard deviation of the Gaussian function determines the range and steepness of the attraction field's influence.

[0055] 2. Coordination Interval Item : Cooperative interval term It is used to maintain the symmetry of the formation.

[0056] To address the problem of local clustering and cluster deadlock that easily occurs in traditional artificial potential field methods in multi-machine collaboration, this invention designs a distance-aware adaptive collaboration interval term. By adjusting the inter-drone repulsion force through an exponential decay function, the drone swarm is forced to spontaneously form a geometrically symmetrical, equally spaced distribution during the encirclement phase.

[0057] Define ideal machine distance .in, The size of the drone swarm allocated to perform a capture mission against a single target. The ideal encirclement radius.

[0058] Cooperative interval term The specific design is as follows:

[0059] in, The position vector of the nearest neighbor drone; Indicates the predicted location of the drone The Euclidean distance to the nearest neighbor drone.

[0060] This design focuses only on local repulsion at extremely close distances, which avoids computational redundancy in the global communication topology and effectively breaks the deadlock of local minima caused by symmetrical forces.

[0061] 3. Long-term heuristic facilitation items : Because Gaussian potential fields exhibit a "gradient vanishing" phenomenon at long distances, drones lose directional guidance in the initial stages of a mission and fall into blind, random searches. To address this, this invention specifically designs a global long-range guidance term for the target. The greater the absolute difference between the distance of the drone's predicted position from the target and the expected encirclement circle, the smaller the long-range heuristic guidance term becomes, but it is never zero, thus providing a continuous and stable directional gradient as the drone moves away from the target.

[0062] Long-term heuristics The specific design is as follows:

[0063] in, The target's position vector; For the heuristic term gain coefficient, The ideal encirclement radius; The absolute difference between the distance of the drone's predicted position from the target and the expected encirclement circle.

[0064] 4. Obstacle Tangential Flow Scoring : In complex obstacle environments, traditional potential field methods, due to the direct clash between repulsive forces and attractive forces, easily cause UAVs to become stuck in front of obstacles. Therefore, this invention proposes an innovative obstacle avoidance assessment mechanism based on tangential flow guidance.

[0065] When the drone approaches the obstacle When calculating the normal vector tangent vector :

[0066] Obstacle tangential flow guidance item The predicted velocity is designed to be generated from the candidate acceleration vector. Approaching obstacles with drones Time tangent vector Inner product:

[0067] The core innovation of this design lies in transforming the normal repulsive force of rigid obstacles into a smooth detour signal, enabling drones to glide along the edge of obstacles without vibration. This achieves complete decoupling of obstacle avoidance logic and target attraction logic, significantly improving the maneuverability of the swarm in confined spaces.

[0068] The multidimensional heuristic scoring model couples multiple potential field terms into a single scalar value through weighted summation:

[0069] in, Candidate acceleration vector Overall score; These are the weighting coefficients of various potential fields, used to balance the accuracy of encirclement, the coordination distance, and obstacle avoidance performance.

[0070] Based on comprehensive scoring, global parallel scanning and optimization of the pre-simulation state can be achieved, enabling real-time filtering of the optimal capture command and deadlock escape.

[0071] Step 5: Calculate the action set The overall score of all candidate actions Select comprehensive score The largest candidate acceleration vector is taken as the optimal acceleration command for the current control cycle. .

[0072]

[0073] Then the optimal acceleration command Substitute the smoothing correction formula and dynamic model constructed in step one into the driving force for the UAV to evolve to the next state.

[0074] Step 6: Dynamic collision detection and cooperative configuration correction based on vectorized operations.

[0075] To address the requirements for flight safety and formation uniformity during tight encirclement operations, a vectorized collision handling mechanism based on a position correction algorithm is introduced.

[0076] First, the Euclidean distance matrix between drones is calculated in real time. When any two drones... Euclidean distance Below the safety threshold At that time, calculate the two drones Overlapping vectors :

[0077] in, and drones and The actual physical position vector in the current control cycle The safe collision radius for a single drone.

[0078] After calculating the overlap vector, displacement compensation is applied to the UAV along the opposite direction of the connecting line using a position correction algorithm. (UAV) The final safe location update formula is as follows:

[0079] in," " indicates an assignment operation; This indicates obtaining all information related to drones. Drones with a distance below the safety threshold Overlapping vectors , these overlapping vectors Add them together and divide by 2 to get the result for the drone. Displacement compensation applied.

[0080] This mechanism ensures absolute physical safety for multiple machines in extremely dense environments by directly modifying physical coordinates.

[0081] At the same time, based on the current control cycle position of the drone With the target location relative distance Dynamically adjust the collaborative distance term in the multidimensional heuristic scoring model Weighting coefficients ;if If the distance exceeds the set threshold, then Use the set low weighting coefficient To switch to the tracking phase; if If the distance is less than or equal to the set distance threshold, then Use the set high weighting coefficient This allows switching to the encirclement phase. Weighting coefficients. Expressed as:

[0082] In the above formula, the distance threshold is set to... ,in To determine the distance buffer margin for phase switching; and These represent the co-exclusion weight constants for the tracking and encirclement phases, respectively.

[0083] This mechanism enables efficient convergence during the tracking phase and a smooth switching between uniform configurations during the encirclement phase.

[0084] Repeat steps three through six in a loop.

[0085] At this point, the coordinated capture and control process has ended.

[0086] Next, the Monte Carlo test is used to evaluate the multi-dimensional interception effectiveness and verify the robustness of the algorithm.

[0087] A Monte Carlo simulation framework was constructed, and the algorithm performance was quantitatively evaluated using recorded index data by conducting batch experiments under different random seeds.

[0088] To make the dynamic game between multiple UAVs and the target more closely resemble a real high-intensity non-cooperative environment, the escaping target is designed with the intelligent ability to actively find weak points in the encirclement and break through the force balance deadlock. The target's dynamic model is consistent with the UAV's dynamic model, and the target's motion command planning is divided into two dimensions: optimal escape gap calculation and adaptive deadlock breaking based on velocity fall. First, the target real-time sensing warning radius Encircle the drones within the area. When an interception threat exists, extract the maximum angular gap between the azimuth angles of any two adjacent drones. Set the direction of the bisector of this gap as the breakout direction. Combined with the avoidance acceleration of static obstacles Generate basic escape acceleration command :

[0089] In the formula, This is for gap escape gain. For the target of the first Acceleration to avoid a static obstacle. When the target is in contact with the first... Distance of a static obstacle Less than the security perception threshold When this evasion action is triggered, its calculation expression is:

[0090] in, This is the obstacle avoidance gain constant, used to control the strength of the repulsive force; and These represent the position vectors of the target and the obstacle, respectively.

[0091] Secondly, to prevent the target from becoming stuck in a local blind spot, an escape disturbance factor that dynamically accumulates with low speed is introduced. The dynamic evolution logic of this factor is based on the target's current velocity amplitude and the low-speed determination threshold. Discretize and iteratively update the relationship:

[0092] In the formula, This is the disturbance factor value from the previous control cycle; A low-speed threshold for determining whether a target has stalled; and These are the fixed step size constants for factor accumulation and decay, respectively. This asymmetric smooth update mechanism ensures a smooth increase in the target's intention to escape while it is continuously trapped, and a slow decay of its state after escaping, effectively suppressing high-frequency oscillations in control commands.

[0093] Target's final acceleration command Based on the basic escape command and heuristic maneuver compensation acceleration Linear interpolation generation:

[0094] Furthermore, when the sensing radius When there are no intercepted threats, the target will follow random movement until threat detection is triggered again.

[0095] The following three core performance indicators are defined for evaluation: Distance error To measure the stability of the containment radius:

[0096] in, This represents the position vector of drone i. In this article, when it is necessary to distinguish between different drones, the subscript 'i' is used for the position vector.

[0097] Configurational symmetry To measure the uniformity of drone distribution:

[0098] Maximum escape angle To measure the degree of containment within the encirclement, calculate the angular difference between adjacent drones relative to the target. Take its maximum value The core criterion for successful capture is that the target does not possess a physical opening for a straight-line breakout, that is:

[0099] To ensure that the drone remains within the effective containment zone and does not collide, a distance effectiveness function for a single drone is defined. :

[0100] Define time Binary decision function :

[0101] Let the total number of simulation steps be To eliminate non-steady-state interference during the initial tracking phase, statistics are collected starting from the later stages of the simulation. Overall interception success rate. Represented as:

[0102] in, The initial number of simulation steps to enter the effective statistical phase after excluding the initial unsteady tracking phase; Index for discrete simulation steps; To control the time step; denominator This represents the total number of valid frames used in the evaluation.

[0103] This formula quantitatively evaluates the overall interception performance of the algorithm in complex and dynamic environments by calculating the proportion of frames that meet the successful encirclement state within the effective statistical period to the total number of evaluation frames.

[0104] Based on the above model and settings, in one instance, such as Figure 2 As shown, four drones surround and capture a target. The parameters of the drones and the target are shown in Table 1.

[0105] Table 1 Initial values ​​of UAVs and targets

[0106] To address the real-time decision-making requirements in highly maneuverable environments, a continuous acceleration command space is provided. Convert to polar coordinate discrete action evaluation set By dividing the acceleration amplitude into three energy levels and the azimuth angle into 36 directions with angular resolution, a candidate action set containing 108 action vectors is generated. .

[0107] The multidimensional heuristic scoring model is as follows:

[0108] in, Dynamic adjustment. To achieve efficient convergence during the tracking phase and a smooth transition to a uniform configuration during the encirclement phase, during the tracking phase, Configuration The value is 2. At this point, attraction dominates, guiding the drone swarm to converge rapidly to the target area; while in the encirclement phase, Configuration The value is 15, which significantly increases the repulsion weight to overcome the local extrema of the target attraction term and forces the drone to spread out evenly along the encirclement circumference.

[0109] Through simulation, the core collaborative effectiveness evaluation curves reflecting the quality of the encirclement are as follows: Figures 3 to 5 As shown. The distance error convergence curve between the multi-UAV swarm and the target in this invention is shown in the figure. )like Figure 3 As shown, guided by the target attraction score, the relative distance between the swarm and the target decreases rapidly and monotonically, eventually converging smoothly near the ideal encirclement radius. The configuration symmetry curve of the multi-UAV swarm of this invention (…) )like Figure 4 As shown, under the adaptive adjustment of the cooperative repulsion mechanism, the UAV spontaneously adjusts its position within the encirclement, significantly reducing the symmetry error of the entire formation and maintaining it at a low and stable level. The maximum escape angle curve of the target in this invention (…). )like Figure 5 As shown, as the encirclement tightens, the maximum gap angle of the encirclement rapidly falls below the safety threshold of 180°, achieving a complete topological lock-on of the maneuvering target. Finally, the two-dimensional trajectory diagram of multi-team cooperative encirclement in a dynamic multi-obstacle environment of this invention is shown below. Figure 6 As shown in the figure, the dynamic evolution of the cluster from long-distance tracking and smoothly bypassing obstacles without oscillation using the tangential flow evaluation mechanism to finally forming a stable encirclement of the target.

[0110] To further verify the robustness of the algorithm under random and unknown environments, 100 independent repeated experiments were conducted according to the Monte Carlo simulation framework constructed in step five. In each experiment, random perturbations were introduced into the initial positions of the target and interceptor missile, as well as the distribution coordinates of static obstacles. Statistical results show that the decision-making algorithm proposed in this paper, which integrates tangential flow obstacle avoidance and adaptive cooperative potential field, achieves a high overall success rate in encirclement and capture under complex and highly dynamic environments. The accuracy rate reached 100%. This proves that the present invention can effectively overcome the extreme value deadlock problem of traditional methods when facing sudden target maneuvers and dense obstacle interference, demonstrating extremely high fault tolerance and robustness.

[0111] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for adaptive prediction and cooperative encirclement of multiple UAVs in an obstacle environment, characterized in that, include: Step 1: Divide the acceleration amplitude and azimuth angle of the UAV into multiple energy levels to generate a candidate action set containing K candidate acceleration vectors. ;in, For candidate action sets The first in One candidate acceleration vector, and They are respectively The acceleration amplitude and azimuth angle; Step 2: For each drone executing the coordinated capture, based on the drone's current control cycle position... and speed Each candidate acceleration vector is determined using a kinematic prediction model. Predicted position in the next control cycle ; Step 3: The drone bases its actions on the candidate action set. Each candidate acceleration vector in and the predicted position generated in step 2 A multi-dimensional heuristic scoring model encompassing encirclement situation convergence, formation coordination, and smooth obstacle avoidance is employed to determine each candidate acceleration vector. The overall score; among which, the obstacle tangential flow guidance item corresponding to the smooth obstacle avoidance. The design is based on the predicted velocity generated by the candidate acceleration vector and the tangent vector when the UAV approaches the obstacle. The inner product; Step 4: Select the candidate acceleration vector with the highest comprehensive score as the optimal acceleration command for the current control cycle. ; Optimal acceleration command Substitute the data into the UAV dynamics model to drive the UAV to evolve to the next state; repeat steps 2-4.

2. The multi-UAV adaptive prediction and cooperative encirclement method in an obstacle environment as described in claim 1, characterized in that, In step 1, dividing the acceleration amplitude and azimuth angle of the UAV into multiple energy levels involves: dividing the acceleration amplitude energy level between the maximum and minimum values ​​of the acceleration amplitude; and dividing the acceleration amplitude energy level between 0 and 10 ... The energy levels are divided into azimuth angles.

3. The multi-UAV adaptive prediction and cooperative encirclement method in an obstacle environment as described in claim 1, characterized in that, In step 2, the kinematic prediction model is: in, To apply acceleration vector The predicted position of the UAV in the next control cycle is then determined. To discretize the control time step.

4. The multi-UAV adaptive prediction and cooperative encirclement method in an obstacle environment as described in claim 1, characterized in that, In step 3, the multidimensional heuristic scoring model is as follows: in, Candidate acceleration vector Overall score; These are the weighting coefficients of each potential field; The target capture guide term guides the drone to converge to the enclosing circle; As a cooperative interval term, the inter-machine repulsion force is adjusted by an exponential decay function, forcing the drone swarm to spontaneously form a geometrically symmetrical, equally spaced distribution formation during the encirclement phase; For the long-range heuristic guidance term targeting the target, the greater the absolute difference between the distance of the UAV's predicted position from the target and the expected encirclement circle, the smaller the long-range heuristic guidance term becomes, but it is never zero, providing a continuous and stable directional gradient when the UAV is far away from the target. Tangential flow scoring of obstacles is used to guide the drone to smoothly navigate around obstacles; when the drone approaches an obstacle... When calculating the normal vector tangent vector ; Tangential flow guidance for obstacles Predicted velocities generated for candidate acceleration vectors Approaching obstacles with drones Time tangent vector Inner product: 。 5. The multi-UAV adaptive prediction and cooperative encirclement method in an obstacle environment as described in claim 4, characterized in that, The target capture guide A Gaussian potential field is used to evaluate the radial deviation between the predicted position of the UAV and the ideal capture radius, guiding the UAV to converge to the encircling circle. in, The UAV predicts its position in the next control cycle based on candidate acceleration vector predictions. The target's position vector; The ideal encirclement radius; The standard deviation of the Gaussian function determines the range and steepness of the attraction field's influence.

6. The multi-UAV adaptive prediction and cooperative encirclement method in an obstacle environment as described in claim 4, characterized in that, The coordinated interval item for: in, The UAV predicts its position in the next control cycle based on candidate acceleration vector predictions. The position vector of the nearest neighbor drone; Indicates the predicted location of the drone Euclidean distance between the nearest neighbor drone; ideal inter-drone distance , For the ideal encirclement radius, The size of a drone swarm assigned to a single target for a capture mission.

7. The multi-UAV adaptive prediction and cooperative encirclement method in an obstacle environment as described in claim 4, characterized in that, The long-range heuristic guide for: in, The UAV predicts its position in the next control cycle based on candidate acceleration vector predictions. The target's position vector; For the heuristic term gain coefficient, The ideal encirclement radius; The absolute difference between the distance of the drone's predicted position from the target and the expected encirclement circle.

8. The multi-UAV adaptive prediction and cooperative encirclement method in an obstacle environment as described in claim 4, characterized in that, The method further includes: Based on the current control cycle position of the drone With the target location relative distance Dynamically adjust the co-interval term in the multidimensional heuristic scoring model Weighting coefficients ;if If the distance exceeds the set threshold, then Use the set low weighting coefficient To switch to the tracking phase, attraction takes precedence, guiding the drone swarm to converge rapidly toward the target; if If the distance is less than or equal to the set distance threshold, then Use the set high weighting coefficient In order to switch to the encirclement phase, overcome the local extreme value of target attraction, and force the drone to spread evenly along the encirclement circumference; The set distance threshold is , For the ideal encirclement radius, Distance buffer margin for determining the phase switching between the tracking and encirclement phases.

9. The multi-UAV adaptive prediction and cooperative encirclement method in an obstacle environment as described in claim 4, characterized in that, The method further includes introducing a vectorized collision handling mechanism based on a position correction algorithm during the encirclement process: Real-time calculation of the Euclidean distance between drones, when any two drones Euclidean distance Below the safety threshold At that time, calculate the two drones Overlapping vectors ; The safe collision radius for a single drone; in, and drones and Position vector in the current control cycle; For drones Apply displacement compensation: in," " indicates an assignment operation; This indicates obtaining all information related to drones. Drones with a distance below the safety threshold Overlapping vectors , these overlapping vectors Add them together and divide by 2 to get the result for the drone. Displacement compensation applied.

10. The multi-UAV adaptive prediction and cooperative encirclement method in an obstacle environment as described in claim 1, characterized in that, Step 4: Determine the optimal acceleration command Then, further refine the optimal acceleration command. Make corrections: in, The final acceleration command, after smoothing correction, is substituted into the UAV dynamics model; For instructions The amplitude; This is a direction smoothing factor used to control the direction of the command towards the current velocity. The degree of deflection is used to simulate real physical constraints.

Images