Multi-uav cooperative search method based on memory-enhanced artificial honeybird algorithm

Abstract

The application discloses a multi-unmanned aerial vehicle cooperative search method based on a memory-enhanced artificial honeybird algorithm, and the method comprises the following steps: a mixed space environment model containing a continuous dynamics domain and a discrete probability perception domain is constructed; environment information of neighbor nodes is acquired through an established local data link, a local map fusion strategy based on a clearance priority principle is adopted, and a local discrete probability grid map is updated; the output local discrete probability grid map, output distributed dynamic communication topology and current unmanned aerial vehicle state are taken as inputs, a multi-target fitness function is constructed, a memory-enhanced MA-AHA path planning algorithm is executed, and an optimal navigation target point is generated; the optimal navigation target point is converted into a speed control instruction to drive the unmanned aerial vehicle to move; and when it is detected that the unmanned aerial vehicle appears in a motion deadlock, a multimode deadlock recovery mechanism is used to correct the speed instruction. The application can significantly reduce the average flight path length of the unmanned aerial vehicle group and effectively save the energy consumption of maneuvering.

Description

Technical Field

[0001] This invention relates to the field of multi-UAV cooperative control and path planning technology, and in particular to a multi-UAV cooperative search method based on a memory-enhanced artificial hummingbird algorithm. Background Technology

[0002] Unmanned Aerial Vehicle (UAV) swarm technology, with its high flexibility, low cost, and strong survivability, has become a key force in modern emergency rescue, military reconnaissance, and environmental monitoring. Its core advantage lies in achieving rapid coverage and precise target search over large, unknown areas through collaborative perception and decision-making among multiple UAVs. However, in real-world applications, such as battlefield environments with strong electromagnetic warfare or disaster relief sites where communication infrastructure has been damaged, UAV swarms often face severe communication limitations. These limitations manifest not only in limited communication distance and insufficient bandwidth but also in the highly dynamic changes in communication topology and frequent network disconnections. Maintaining swarm collaboration efficiency and mission consistency in the absence of a central node for unified scheduling and with severely hampered information exchange remains a pressing problem to be solved.

[0003] Current UAV cooperative path planning and search methods mainly rely on two major categories of technologies in spatial modeling and decision-making mechanisms, but both have significant limitations when dealing with the aforementioned complex and constrained scenarios. The first category is grid-based methods, such as cover search algorithms and A / B algorithm. Algorithms and their variants. These methods discretize the continuous environment into a grid, effectively reducing the dimensionality of the environment representation and making information fusion computation between multiple drones simple and easy to implement. However, their fundamental drawback lies in the loss of motion constraints. The discretized paths typically consist of a series of broken line segments or right-angle turns, ignoring the drone's own aerodynamic characteristics and kinematic constraints, such as minimum turning radius and inertia. This makes it difficult to smoothly track the planned paths in actual flight, and frequent acceleration, deceleration, and large-angle turns significantly increase energy consumption and reduce operation time. The second category is based on continuous space methods, such as the Artificial Potential Field (APF) method and the Rapid Expanding Random Tree (RRT). These methods solve in a continuous domain, generating smooth flight trajectories that conform to physical properties. However, in multi-drone collaborative scenarios, their bottleneck lies in the complexity of information interaction. Describing environmental obstacles and exploration states in continuous space usually requires complex geometric parameters or high-dimensional vectors, which generates a huge amount of data when sharing information between multiple drones, placing extremely high demands on communication bandwidth. In communication-constrained environments, this high bandwidth requirement often cannot be met, leading to collaborative failure.

[0004] Furthermore, at the decision-making algorithm level, existing swarm intelligence algorithms such as Particle Swarm Optimization (PSO) and Grey Wolf Algorithm (GWO), while performing excellently in solving optimization problems, suffer from systemic defects when directly applied to UAV cooperative search under communication-constrained conditions. This application finds that traditional algorithms typically assume the existence of globally ideal communication or a central server, allowing individuals to obtain the globally optimal solution in real time. However, in decentralized scenarios with limited communication, individuals can only obtain local neighbor information, and UAVs cannot synchronize the environmental map in real time, leading to discrepancies in their environmental state assessments of the same area. This inconsistency in multi-UAV environmental perception easily causes false target noise to spread erroneously in the swarm network, causing traditional algorithms to stagnate or become chaotic due to the information silo effect. Simultaneously, most swarm intelligence algorithms rely on reactive decision-making based on the current state, lacking an effective memory mechanism for historical exploration paths and a deadlock escape mechanism. In complex obstacle or maze terrain, UAV swarms are prone to inefficiently repeating searches in already explored areas or getting trapped in local optima from which they cannot escape. Moreover, in the event of communication delays or interruptions, traditional algorithms lack an effective short-range dynamic separation mechanism, and multiple drones are prone to spatial congestion or even collisions when converging on the same target point.

[0005] In summary, existing technologies face problems such as fragmented spatial modeling, fragile communication mechanisms, and blind decision-making algorithms when dealing with multi-UAV cooperative search problems in communication-constrained environments. Ultimately, this leads to low efficiency, easy collisions, and difficulty in maintaining motion smoothness in multi-UAV cooperative search under communication-constrained environments. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-UAV cooperative search method based on a memory-enhanced artificial hummingbird algorithm. This method utilizes the efficiency of grid-based distributed map fusion while perfectly preserving the physical reality of continuous domain flight, thereby significantly reducing the average flight path length of the UAV swarm.

[0007] To achieve the above objectives, the present invention provides the following solution: A multi-UAV cooperative search method based on a memory-enhanced artificial hummingbird algorithm includes the following steps: S1. Based on the search area environment, construct a hybrid spatial environment model that includes a continuous dynamic domain and a discrete probability perception domain, and obtain the continuous coordinates and discrete probability grid map of the UAV based on the hybrid spatial environment model. S2. Based on the preset communication radius and discrete communication period, construct a distributed dynamic communication topology under adaptive spatiotemporal constraints based on the continuous coordinates of the UAV. The distributed dynamic communication topology establishes local data links between UAVs when the spatiotemporal constraints are met. S3. Obtain environmental information of neighboring nodes through the established local data links, and update the local discrete probability raster map by adopting a local map fusion strategy based on the principle of clearing priority. S4. Using the local discrete probability grid map output by S3, the distributed dynamic communication topology output by S2, and the current UAV status as input, construct a multi-objective fitness function and execute the memory-enhanced MA-AHA path planning algorithm to generate the optimal navigation target point. S5. Convert the optimal navigation target point into speed control commands to drive the drone's movement; when a motion deadlock is detected in the drone, correct the speed commands through a multimodal deadlock recovery mechanism to guide the drone out of trouble.

[0008] Preferably, in S1, the hybrid spatial environment model includes smooth motion modeling in the continuous dynamics domain and Gaussian perception modeling in the discrete probability perception domain; the smooth motion modeling in the continuous dynamics domain includes velocity recursion equations and position update equations: Velocity recurrence relation:

[0009] Position update equation:

[0010] in, and The first A drone in Time and The actual flight velocity vector at any given moment, for The desired speed control command at the underlying level at any given moment. The inertial smoothing coefficient, and The first A drone in Time and The continuous spatial coordinate vector at time t. This represents the discrete control step size of the model.

[0011] Preferably, Gaussian perception modeling in the discrete probability sensing domain includes: rasterizing the continuous search region into discrete grids of equal size, and performing multiplicative updates on the grid probabilities within the UAV's sensing radius based on a Gaussian kernel function.

[0012] in, For the updated Time of the first In the perception map maintained by the drone itself, the center location is The probability of the existence of a grid target. Before the update The probability that a grid target exists at this location at any given time. This represents the sensor's maximum detection confidence level. The Gaussian kernel width coefficient, For the first A drone in Continuous spatial position coordinates at any given time. The center coordinates of the discrete raster. This represents the Euclidean distance between the drone's current position and the center of the grid.

[0013] Preferably, in S2, a distributed dynamic communication topology under adaptive spatiotemporal constraints is constructed based on the continuous coordinates of the UAV, specifically including: By simulating communication radius constraints and communication time slot constraints, a UAV communication link criterion based on spatial filtering and temporal sparsity is constructed. This satisfies the spatiotemporal constraints of the communication link criterion, establishing a distributed communication link. The UAV communication link criterion based on spatial filtering and temporal sparsity is as follows:

[0014] in, For time k, the first The drone and the first The communication link status of the drone is 1 for established link and 0 for disconnected link. For the first A drone in Continuous spatial position coordinates at any given time. For the first A drone in Continuous spatial position coordinates at any given time. To limit the communication radius, it is used for spatial filtering. For discrete communication periods, used to implement time sparsity. This is the modulo operation.

[0015] Preferably, in S3, the local map fusion strategy based on the clearing priority principle is as follows:

[0016] in, For the fusion Time of the first The drone is used for local maintenance, and its central location is... The final update value is the probability of target existence in the discrete raster. For a set of nodes that are effectively connected, and These are the minimum-dominant weight and the mean-smoothing weight coefficients, respectively, and they satisfy... ; The feature term representing the neighborhood minimum is extracted from... Time-local communication set All drones within the area are in coordinates The probability of the minimum target being assigned; This is the mean smoothing auxiliary term. Time-local communication set All drones within the area are in coordinates The arithmetic mean of probability observations.

[0017] Preferably, in S4, the constructed multi-objective fitness function is as follows:

[0018] in:

[0019]

[0020]

[0021]

[0022] in, As candidate navigation target points, , , , As weight, For the target gain term, This represents the probability of the target existing in the local discrete probability map corresponding to the candidate waypoint; As a historical penalty item, The preset threshold for determining whether an area has been explored; For distance-based energy consumption, This is the current location of the drone; To separate the pressure term, For safe distance threshold, The coordinates of neighboring nodes obtained based on the communication link.

[0023] Preferably, in S4, a multi-objective fitness function is constructed and a memory-enhanced MA-AHA path planning algorithm is executed to generate the optimal navigation target point, specifically including: The optimization objective is to minimize the multi-objective fitness function, with candidate waypoints in continuous space as optimization variables. The population is randomly initialized, and the multi-objective fitness function of the initial population is calculated. A dual-mode search strategy is triggered by random probability. Guided foraging pattern: Utilizing the group's optimal state information achievable under current communication conditions to apply long-range attraction, driving individuals to escape their current local convergence region and achieving position updates: when random probability... When entering global exploration mode, the location update method is as follows:

[0024] in, For new destinations, For the first The current waypoint of the drone. To guide the scaling factor of foraging stride, To guide waypoints to the target For nondeterministic perturbation vectors that follow a standard normal distribution; Territorial foraging mode: This mode utilizes a nondeterministic perturbation vector following a standard normal distribution to sample the neighborhood of the current solution, achieving position updates: when the random probability... When entering partial development mode, the location update method is as follows:

[0025] in, Scaling factor for territorial foraging step size; Each iteration of the guided foraging mode or the territory foraging mode calculates the fitness value of candidate waypoints. Finally, after multiple iterations and convergence, the waypoint with the minimum fitness is output. As the optimal navigation target point for the next moment.

[0026] Preferably, in S5, the velocity control command for converting the optimal navigation target point is composed of the target's gravitational field velocity vector, the static obstacle's repulsive field velocity vector, and the dynamic neighbor's repulsive field velocity vector; the velocity control command is as follows:

[0027] in, The target gravitational field velocity vector, Let $\mathbf{ ... The magnitude of gravity; The velocity vector of the repulsive force field of the static obstacle. For the normal repulsive component, This is the tangential repulsive force component; The velocity vector of the dynamic neighbor repulsion field. This is the direction vector of the current drone moving away from neighboring drones; This represents the magnitude of the repulsive force.

[0028] Preferably, in S5, the speed command is corrected through a multimodal deadlock recovery mechanism, specifically including: The drone maintains a fixed-length historical trajectory buffer in real time and calculates the Euclidean displacement between the current position and the starting point of the buffer. If the Euclidean displacement is less than the set stagnation threshold, the drone is determined to have motion deadlock. When motion deadlock is detected, a preset amplitude escape velocity vector is applied for a fixed duration. The speed commands actually executed by the drone are corrected.

[0029] Preferably, the collaborative search method also includes a confidence accumulation confirmation mechanism based on Bayesian principles: To simulate the sensor detection characteristics of a drone in a real-world physical scenario, the observation confidence level is calculated and accumulated in real time based on the relative distance between the drone and the target and the environmental occlusion status. The discrete-time update formula is as follows:

[0030] in, for Time of the first The cumulative observation confidence level of the drone on the target. The upper limit is the cutoff value. This refers to the step size in discrete simulation. The rate of change function of the confidence level at the current moment is calculated using the following formula:

[0031] in, For the continuous spatial coordinates of the UAV, For the target coordinates, This is the maximum effective detection radius of the sensor. Based on the observation gain coefficient, This is the confidence decay coefficient when the field of view is lost.

[0032] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: (1) The method of this invention proposes a continuous-discrete coupled hybrid spatial environment model, which couples the continuous dynamic domain with the discrete probability perception domain in a closed loop, effectively solving the inherent defects of traditional discrete grid methods and continuous spatial methods. In specific implementation, this invention introduces a one-way dimensionality reduction query mechanism, so that path optimization decision-making is always carried out in the continuous domain to ensure the smooth execution of the generated trajectory; while the discrete grid is only used as the query data source after dimensionality reduction of high-dimensional environmental features. This hybrid modeling method not only utilizes the efficiency of the grid method in processing distributed map fusion, but also perfectly preserves the physical reality of continuous domain flight (satisfying C2 continuity), thereby significantly reducing the average flight path length of the UAV swarm and effectively saving maneuvering energy consumption.

[0033] (2) This invention proposes a Memory-Enhanced Artificial Hummingbird Algorithm (MA-AHA): Based on the traditional artificial hummingbird algorithm, a memory mechanism of historical visit potential field is introduced. A multi-objective fitness evaluation function containing target gain, historical penalty, energy consumption constraint, and anti-collision repulsion force is constructed. A dual-mode search strategy of guided foraging and territorial foraging is designed to adapt to scenarios where only local neighbor information can be obtained. This solves the problems of traditional swarm intelligence algorithms, such as "no memory and easy repetition, reliance on global communication and easy stagnation, and lack of constraints and easy collision", effectively avoiding repeated visits to already explored areas. Simulation experiments show that compared with traditional swarm intelligence algorithms, the average task completion time (TTD) of this method is effectively shortened under the same environment, and all targets can be discovered faster. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart illustrating the multi-UAV cooperative search method based on the memory-enhanced artificial hummingbird algorithm provided by the present invention. Figure 2 This is a logical structure diagram of the search strategy based on memory and repulsion guidance in this invention; Figure 3 This is a box plot of task completion time (TTD) under multiple Monte Carlo simulation experiments in an embodiment of the present invention; Figure 4 This is a frequency distribution histogram of the minimum safe distance between UAVs during the simulation process in this embodiment of the invention; Figure 5 is a simulation diagram of the multi-UAV cooperative search trajectory in an obstacle environment according to an embodiment of the present invention; Where (a) is t (a) Simulation diagram of multi-UAV cooperative search trajectory at 22.8s; (b) Simulation diagram of multi-UAV cooperative search trajectory at t≤46.0s; (c) Simulation diagram of multi-UAV cooperative search trajectory at t≤92.4s; (d) Simulation diagram of multi-UAV cooperative search trajectory at t≤138.8s; (e) Simulation diagram of multi-UAV cooperative search trajectory at t≤185.2s; (f) Simulation diagram of multi-UAV cooperative search trajectory at t≤231.6s. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] like Figure 1 As shown, the present invention provides a multi-UAV cooperative search method based on a memory-enhanced artificial hummingbird algorithm, comprising the following steps: S1. Based on the search area environment, construct a hybrid spatial environment model that includes a continuous dynamic domain and a discrete probability perception domain, and obtain the continuous coordinates and discrete probability grid map of the UAV based on the hybrid spatial environment model. S2. Based on the preset communication radius and discrete communication period, construct a distributed dynamic communication topology under adaptive spatiotemporal constraints based on the continuous coordinates of the UAV. The distributed dynamic communication topology establishes local data links between UAVs when the spatiotemporal constraints are met. S3. Obtain environmental information of neighboring nodes through the established local data links, and update the local discrete probability raster map by adopting a local map fusion strategy based on the principle of clearing priority. S4. Using the local discrete probability grid map output by S3, the distributed dynamic communication topology output by S2, and the current UAV status as input, construct a multi-objective fitness function and execute the memory-enhanced MA-AHA path planning algorithm to generate the optimal navigation target point. S5. Convert the optimal navigation target point into speed control commands to drive the drone's movement; when a motion deadlock is detected in the drone, correct the speed commands through a multimodal deadlock recovery mechanism to guide the drone out of trouble.

[0039] Specifically, the method of the present invention includes: S1: Based on the search area environment, construct a hybrid spatial environment model that includes a continuous dynamic domain and a discrete probability perception domain, and obtain the continuous coordinates and discrete probability raster map of the UAV based on the hybrid spatial environment model.

[0040] The core of this step is to resolve the modeling contradiction between the loss of motion details in traditional discrete grid methods and the excessive amount of communication data required by continuous space methods. A two-layer hybrid model is constructed, consisting of a bottom-layer discrete probabilistic sensing domain and an upper-layer continuous dynamics domain. The bottom-layer discrete probabilistic sensing domain reduces the dimensionality of the environment representation to accommodate limited computational resources, while the upper-layer continuous state space ensures the physical feasibility of the UAV's motion. This invention does not employ discrete solution followed by back-mapping; instead, it constructs a unidirectional dimensionality reduction update and query mechanism for both domains using Gaussian kernel functions and coordinate indexes. That is, the UAV's path optimization decision-making is always performed in the continuous domain, with the bottom-layer discrete probabilistic map serving only as a query data source after dimensionality reduction of the high-dimensional environment features. This achieves a closed-loop coupling between dimensionality reduction of the environment representation and the continuity of motion control.

[0041] In this embodiment, the hybrid spatial environment model includes smooth motion modeling in a continuous dynamic domain and Gaussian perception modeling in a discrete probability perception domain.

[0042] In this embodiment, the smooth motion modeling in the continuous dynamics domain is as follows: To simulate the aerodynamic delay and inertial characteristics of a rotary-wing UAV, avoid abrupt changes in control quantities caused by discrete decision-making, i.e., avoid discontinuous large-angle broken lines or high-frequency jitter in the planned path, ensure that the decision results conform to the inertial constraints of the UAV, and guarantee the C2 continuity of the flight trajectory, the first... The velocity recursive smooth kinematic equation of the drone: In this embodiment, the smooth motion modeling in the continuous dynamics domain is as follows: to simulate the aerodynamic delay and inertial characteristics of a rotary-wing UAV, avoid abrupt changes in control quantities caused by discrete decision-making, and ensure the C2 continuity of the flight trajectory, the first... The state update equations for the UAV include the velocity recursive smooth kinematic equations and the position update equations: Velocity recurrence relation:

[0043] Position update equation:

[0044] in, and The first A drone in Time and The actual flight velocity vector at any given moment; for The desired speed control command at the underlying level at any given moment; The inertial smoothing coefficient (first-order damping element coefficient) is taken as 0.8 in this embodiment; and The first A drone in Time and A continuous spatial coordinate vector at any given time; This represents the discrete control step size of the system. This position update equation provides a continuous and smooth base coordinate input for subsequent calculations of the communication topology distance, the sensing Gaussian kernel function, and the multi-target fitness.

[0045] The technical advantage of smooth motion modeling in continuous dynamics domains lies in the fact that by constructing a first-order damping element through an inertial smoothing coefficient, the velocity vector at the next moment is forced to retain 80% of the current motion state. This mechanism simulates the aerodynamic delay characteristics of a physical UAV at the algorithm level, eliminating the abrupt changes in control quantities that may be caused by discrete decision-making, thereby ensuring that the generated trajectory satisfies second-order smoothness (C2 continuity), and ensuring that the planning results conform to the aerodynamic delay characteristics and physical control constraints of a rotorcraft UAV during high-speed maneuvers.

[0046] In this embodiment, the Gaussian sensing model in the discrete probability sensing domain is as follows: the continuous search area of ​​1000m×1000m is rasterized into a discrete grid of equal size. Utilizing the radial decay characteristic of the Gaussian kernel function, the non-uniform detection accuracy of the sensor in the continuous space (high confidence at the center, low confidence at the edges) is accurately mapped to the discrete grid, achieving lossless dimensionality reduction from the continuous physical space to the discrete decision space. When the UAV is located in continuous coordinates... At that time, a multiplicative update is performed on the grid targets within the drone's perception radius based on probability, using the following formula:

[0047] in, For the updated Time of the first In the perception map maintained by the drone itself, the center location is The probability of the existence of grid targets is determined, and this probability gradually decreases as the exploration level increases after the drone detection, thus achieving rapid elimination of environmental uncertainty; Before the update The probability of a grid target existing at this location at any given time is initially set to a uniform distribution of 0.5. The maximum detection confidence level of the sensor is set to 0.95 in this embodiment; The Gaussian kernel width coefficient is determined by the sensor's detection radius. and The two work together to ensure the physical consistency of environmental modeling; For the first A drone in Continuous spatial position coordinates at any given time; The center coordinates of the discrete raster; This represents the Euclidean distance between the drone's current position and the center of the grid.

[0048] S2: Based on the preset communication radius and discrete communication cycle, construct an adaptive spatiotemporal constraint-based distributed dynamic communication topology based on the continuous coordinates of the UAV. When the spatiotemporal constraint conditions are met, the distributed dynamic communication topology establishes local data links between UAVs.

[0049] This invention addresses communication bandwidth-constrained scenarios by abandoning the idealized assumption of traditional "global fully connected communication." Instead, it internalizes the constraints of limited communication bandwidth and dynamic topology changes into design conditions. By constructing a distributed dynamic communication topology through spatiotemporal constraints, it utilizes a clearing-priority local map fusion strategy to achieve consistency in environmental perception among multiple UAVs, thus solving the problems of information silos and false target noise propagation under communication constraints.

[0050] In this embodiment, a distributed dynamic communication topology under adaptive spatiotemporal constraints is constructed based on the continuous coordinates of UAVs. This includes: constructing a UAV-to-UAV communication link criterion based on spatial filtering and temporal sparsity by simulating communication radius limitations and communication time slot constraints; a distributed communication link is established only when the spatiotemporal constraints of the communication link criterion are met. The criterion for inter-UAV communication links based on spatial filtering and temporal sparsity is:

[0051] in, For time k, the first The drone and the first The communication link status of the drone is 1 for established link and 0 for disconnected link. To limit the communication radius, this embodiment uses 350m for spatial filtering, allowing only nearby nodes to interact. For discrete communication cycles, this embodiment uses a 5-step size to implement time sparsity. To enable modulo operations, the UAV only interacts with data at steps 5, 10, and 15. This spatiotemporal coupling constraint reduces the instantaneous throughput and frequency of the distributed network, ensuring the engineering feasibility of the algorithm in weak communication environments.

[0052] S3: The UAV with established data link acquires environmental information from neighboring nodes and updates the local discrete probability grid map using a local map fusion strategy based on the principle of prioritizing clearing. Under the condition of limited communication bandwidth, it prioritizes suppressing the spread of false target noise to achieve consistency in environmental cognition among multiple UAVs.

[0053] In this embodiment, a local discrete probabilistic raster map is updated using a local map fusion strategy based on the principle of prioritizing clearing. This includes: when communication conditions are met, addressing the information silos and cognitive conflicts among distributed nodes in communication-constrained environments, this invention proposes a fusion strategy based on the principle of conservative consistency. In this embodiment, a local discrete probabilistic raster map is updated using a local map fusion strategy based on the clearing priority principle. This includes: when communication conditions are met, addressing the information silos and cognitive conflicts among distributed nodes in communication-constrained environments, this invention proposes a local map fusion strategy based on the conservative consistency principle, with the following update formula:

[0054] in, For the fusion Time of the first The drone is used for local maintenance, and its central location is... The final update value is the probability of the target's existence in the discrete raster. Based on step S2, in Time and the The drone successfully established a set of valid connected nodes for a local data link (including the first...) (The drone itself), this set mathematically defines the local spatial topological constraints under distributed communication constraints; and These are the minimum-dominant weight and the mean-smoothing weight coefficients, respectively, and they satisfy... ; The feature term representing the neighborhood minimum is extracted from... Time-local communication set All drones within the area are in coordinates The minimum probability of the existence of the target is assigned (physically representing the strongest certainty in the local network that the area has been explored and confirmed to have no target). This is the mean smoothing auxiliary term. Time-local communication set All drones within the area are in coordinates The arithmetic mean of probabilistic observations is designed to avoid systematic misjudgments caused by the failure of a single sensor node.

[0055] The technical advantage of the local map fusion strategy based on the principle of prioritizing clearing is that, when insufficient bandwidth prevents real-time full synchronization of the global map, it can... The minimum value feature item prioritizes the adoption of "no target / explored" information (i.e., low probability values) within the local neighborhood. This "purge-first" fusion mechanism (in this embodiment, the minimum value is preferably set as the dominant weight) Mean smoothing weight This technology can give absolute authority to low-probability security intelligence with certainty, quickly suppress the spread of false high-probability unknown noise in distributed networks, force the swarm to quickly reach a local consensus that "the current area has been explored", thereby guiding drones to migrate to unexplored areas and fundamentally reducing the rate of repeated searches under communication-restricted conditions.

[0056] S4: Using the updated local discrete probability grid map output from step S3, the distributed dynamic communication topology output from step S2, and the current UAV state as input, a multi-objective fitness function is constructed, and a memory-enhanced MA-AHA path planning algorithm is executed to generate the optimal navigation target point. Specifically, when evaluating candidate waypoints in continuous space: the probability of target existence in the corresponding area of ​​the waypoint is queried using the local probability map from S3, serving as the input for calculating the target gain term in the fitness function; the coordinate information of the current connected neighbors is obtained using the dynamic communication topology from S2, serving as the input for calculating the separation pressure term (collision avoidance) in the fitness function; simultaneously, during the global exploration phase of the MA-AHA algorithm, the optimal waypoint in the effective neighbor set is extracted based on the topological relationship from S2, serving as the gravity source input for guiding the foraging mode. Through the deep coupling of the above multi-source information, the optimal navigation target point for the next time step is iteratively generated.

[0057] This step is the core decision-making layer of the invention. Based on the traditional artificial hummingbird algorithm, a memory enhancement mechanism (achieved through the historical visit potential field) is introduced to construct a multi-objective fitness evaluation function that includes target gain term, historical visit potential field, distance energy consumption, and formation separation repulsion force. Then, through the MA-AHA algorithm's guided foraging (global exploration) and territory foraging (local development) dual-mode search strategy, candidate waypoints are optimized using only local neighbor information to generate the optimal search waypoint that balances search efficiency, no duplication, low energy consumption, and collision avoidance.

[0058] In this embodiment, the planner evaluates candidate waypoints in a continuous space, and the logic for constructing its fitness function is as follows: Figure 2 As shown, it is essentially a multi-objective fitness function that is linearly weighted by positive gain terms and negative cost terms, aiming to minimize the following: The multi-objective fitness function is obtained by weighting objective gain terms, history penalty terms, distance energy consumption terms, and separation pressure terms:

[0059] in, As candidate navigation target points, , , , As weight, For the target gain term, As a historical penalty item, For distance-based energy consumption, This is to separate the pressure term.

[0060] in, As candidate navigation target points, to The weights for each item are determined. To achieve a closed-loop connection with the aforementioned environmental perception and communication results, the four indicators are specifically based on the local probability map. The following calculations are performed based on the real-time status of the drone: Target gain term (Task-driven): Utilizing a coordinate mapping mechanism, the probability of target existence corresponding to candidate waypoints in the local discrete probability map is directly extracted, i.e.:

[0061] This reflects the potential mission value of the candidate region; the higher the value, the stronger the positive gravitational pull.

[0062] Historical Penalties (Exploration Constraint): The exploration status of a region is determined based on a probability map. When a candidate point is located in a low-probability region that has already been explored, a penalty is activated, i.e.:

[0063] In the formula, This is a preset threshold for determining "explored". If the probability of a point is below the threshold, it is considered a searched area, and a high penalty is imposed to drive the drone to break away from localized, repetitive searches.

[0064] Distance energy consumption (Physical overhead): Calculate the current position of the drone. The physical span between the candidate waypoints, i.e.:

[0065] This study aims to optimize flight energy efficiency by constraining the maneuver range of a single-step plan using Euclidean distance.

[0066] Separation pressure term (Cooperative Security): Neighbor node coordinates obtained based on communication links Construct a strong repulsive field, that is:

[0067] In the formula, This is a safe distance threshold. It is triggered only when a candidate point is too close to its neighbor, forcing the cluster to maintain safe formation.

[0068] In the evaluation system of this multi-objective fitness function, each parameter term and Figure 2The logic boxes are mapped one-to-one, achieving a dynamic balance between "exploration desire" and "physical constraints" through weighting coefficients. Specifically, the target gain term... Propelling the drone towards a high-probability, unknown region, it serves as the sole positive gravitational source; counterbalanced by this are three costly penalties: historical penalties. As a spatiotemporal constraint, it effectively prevents the algorithm from getting stuck in local repetitive coverage; distance energy consumption term Used to constrain planning step size to improve the economy of long-endurance missions; separate pressure terms. When the distance between drones is less than a safety threshold, a strong repulsion field is introduced to ensure that the cluster avoids physical collisions while expanding its search area. The above four indicators are coupled together under a weighted summation mechanism, jointly guiding the drones to maximize information acquisition while taking into account obstacle avoidance safety and mission efficiency, thus converging to the global optimum.

[0069] In this embodiment of the application, the memory-enhanced MA-AHA path planning algorithm is executed based on a multi-objective fitness function to generate the optimal navigation target point, including: First, with the minimum multi-objective fitness function as the optimization objective, candidate waypoints in continuous space are used as optimization variables (when evaluating continuous waypoints, the probability value of the underlying discrete map is directly queried by rounding down the coordinates, avoiding the trajectory non-smoothness problem caused by discrete solution). The population is randomly initialized, and the multi-objective fitness function of the initial population is calculated. A dual-mode search strategy is triggered by random probability. Waypoint updates are only achieved by using the best waypoint among the effective neighbors of the current UAV with established data links in the distributed dynamic communication topology (i.e., the waypoint with the minimum multi-objective fitness value among the neighbors with which the UAV has established communication). Global optimal information is not required, which is suitable for decentralized scenarios with limited communication.

[0070] The dual-mode search strategy includes guided foraging mode and territorial foraging mode; The dual-mode search strategy includes guided foraging mode and territorial foraging mode; The foraging pattern is guided by applying long-range attraction using information about the group's optimal state achievable under current communication conditions. This drives individuals to escape their current local convergence zone and achieve large-scale state migration to higher-value regions. Specifically: When random probability When this happens, the algorithm enters global exploration mode, and its position update follows the following rules:

[0071] in, For new destinations; For the first The drone (the first in the population) The current waypoint of each individual; To guide the foraging step size scaling factor, a value of 0.5 is used in this embodiment; The waypoint for the target is dynamically selected based on the local memory access table, with the candidate waypoint having the lowest historical access frequency. It is a nondeterministic perturbation vector that follows a standard normal distribution.

[0072] The territorial foraging mode utilizes a nondeterministic perturbation vector that follows a standard normal distribution to sample the neighborhood of the current solution, achieving refined coverage and confirmation of the discovered target area. Specifically: When random probability When the algorithm enters local development mode, its position update follows:

[0073] in, This is the scaling factor for the territory foraging step size, which is set to 0.5 in this embodiment.

[0074] Secondly, the fitness value of the candidate waypoints is calculated in each iteration. The algorithm is set to iterate 200 times, and finally converges and outputs the waypoint with the smallest fitness. As the optimal navigation target point for the next moment.

[0075] S5: Converts the optimal navigation target point into a speed control command, executes continuous domain dynamics control to drive the UAV's actions, and corrects the actual speed of the UAV through a multimodal deadlock recovery mechanism when the UAV experiences motion deadlock, guiding the UAV to get out of motion deadlock.

[0076] This step achieves a closed-loop transformation from "high-level algorithm decision-making" to "low-level physical execution," converting the optimal navigation target point into an executable desired velocity vector for the UAV. It achieves smooth maneuvers for obstacle avoidance and collision prevention through the superposition of multiple field forces. At the same time, it integrates a multimodal deadlock recovery mechanism to solve the problem of the UAV getting stuck in local extreme points in complex obstacle environments. It also introduces nondeterministic perturbation and dynamic obstacle verification methods to enhance the robustness of the algorithm in complex environments.

[0077] In this embodiment of the application, the optimal navigation target point is converted into a low-level velocity control command. The velocity control command is composed of the target gravitational field velocity vector, the static obstacle repulsive field velocity vector, and the dynamic neighbor repulsive field velocity vector. Specific speed control commands for:

[0078] in, The velocity vector within the target gravitational field plays a crucial role in guiding the UAV to the optimal waypoint. Its calculation formula is:

[0079] In the formula, This is the direction vector from the UAV's current position to the optimal navigation target point; The magnitude of gravity is proportional to the distance between the drone and the target. This is the velocity vector of the repulsive force field of a static obstacle. This invention introduces a tangential component into this repulsive force field, which is composed of the superposition of the normal repulsive force perpendicular to the obstacle surface and the tangential repulsive force along the obstacle edge. Its calculation formula is:

[0080] In the formula, This is the normal repulsive force component, directed perpendicularly to the obstacle, used to prevent the drone from physically colliding with the obstacle; The tangential repulsive force component is parallel to the edge of the obstacle and is used to guide the drone to fly smoothly around the obstacle's outline, avoiding hard-hitting or sudden changes of direction.

[0081] This is the velocity vector of the dynamic neighbor repulsion field, designed to achieve dynamic collision avoidance between machines. Its calculation formula is:

[0082] In the formula, This is the direction vector of the current drone moving away from neighboring drones; This represents the magnitude of the repulsive force. This repulsive force is triggered only when the distance between the machines is less than a set safety threshold, and... The size of the force is inversely proportional to the distance between the machines (i.e., the smaller the distance between the machines, the more exponentially the repulsive force increases), thus forcing the cluster to maintain a safe collision avoidance distance.

[0083] In this embodiment of the application, to address the motion deadlock problem of a drone getting stuck in a local extreme point or obstacle trap, i.e., in a non-convex environment, the present invention introduces a multimodal deadlock recovery mechanism based on a sliding window of historical positions. Specifically: The drone maintains a historical trajectory buffer with a fixed length of 70 steps in real time and calculates the Euclidean displacement between the current position and the starting point of the buffer. In this embodiment, the stagnation threshold is set to 6.0m. If the Euclidean displacement is less than the set stagnation threshold, the drone is determined to be trapped in a local extreme point or obstacle trap, that is, motion deadlock occurs.

[0084] When a motion deadlock is detected, control of the drone is temporarily transferred to the escape module, and a preset amplitude escape velocity vector is applied for a certain duration. (Escape Velocity Vector) This vector is a random direction that moves away from obstacles / local extreme regions, and lasts for 10 steps. It corrects the actual speed commands executed by the UAV. , This guides the drone to gradually escape the influence of the current gravitational potential field, thereby enhancing the robustness of the algorithm in complex obstacle environments.

[0085] Based on the above steps, this invention constructs a 1000m×1000m simulation environment, sets up 5 randomly distributed targets, and performs 200 independent Monte Carlo simulations within an area containing complex circular and rectangular obstacles. Furthermore, to further verify the feasibility and robustness of this invention in actual physical systems, this embodiment, based on the above algorithm framework, refines and constrains the engineering implementation details of the simulation experiment as follows: First, regarding the visualization and representation of the search task's state, this invention constructs a color-coded interaction mechanism based on a state machine. In the simulation scenario, the target to be searched is initialized to a red state, representing "active / not found"; once the algorithm logic determines that a target meets the capture conditions, the system will lock onto the target in real time and flip its state to green, representing "cleared". This mechanism allows the UAV swarm to perceive the global task progress in real time, avoiding repeated searches of the same captured target by multiple UAVs, and further improving the efficiency of collaborative search.

[0086] Secondly, regarding the final determination logic for target acquisition, this invention designs a confidence accumulation and line-of-sight (LOS) attenuation confirmation mechanism based on Bayesian principles. It abandons the idealized "instantaneous visibility equals acquisition" rule and rigorously simulates the distance attenuation characteristics of UAV sensors and the line-of-sight (LOS) effect in real-world physical scenarios. The UAV calculates and accumulates the observation confidence in real time based on the relative distance to the target and the environmental occlusion status. Its discrete-time update formula is as follows:

[0087] in, for Time of the first The cumulative observation confidence level of the UAV for the target, with an initial value of 0 and an upper limit cutoff value. This embodiment is set to 20.0; This refers to the step size in discrete simulation. The confidence rate of change function at the current moment is calculated using the following logic:

[0088] In the above formula, For the continuous spatial coordinates of the UAV; The target coordinates; This is the maximum effective detection radius of the sensor (35.0m in this embodiment); The basic observation gain coefficient (10.0 in this embodiment); This is the confidence decay coefficient (i.e., the forgetting mechanism) when the field of view is lost; in this embodiment, it is set to 0.2. This mechanism perfectly realizes the physical simulation that "the closer the distance and the faster the confidence accumulates when there is no obstruction," and the confidence will decrease reasonably when the target is lost.

[0089] This embodiment sets a confidence threshold for target confirmation. When the confidence level of any node in the drone swarm meets the condition... When the target is successfully captured, the system determines that the target has been successfully captured, triggering a target state reversal signal (marked as cleared). All drones clear their tracking confidence for the target and synchronize the cleared target status to the local network via the distributed communication link in step S2. This mechanism fundamentally avoids false target locking caused by a single false alarm from the sensor.

[0090] The specific experimental setup of this invention, including the experimental environment and supplementary experimental parameters, is as follows: Detailed parameter settings for the simulation experimental environment and underlying collaborative control logic. To fully verify the feasibility, robustness, and algorithm performance of this invention in a real physical system, this embodiment constructs a highly realistic multi-UAV collaborative search test environment based on a real physics engine. Its core engineering implementation parameters are constrained in great detail as follows: (1) Physical space and discretization modeling parameters: The dimension of the continuous search physical space is set to To achieve a one-way dimensionality reduction query mapping between continuous and discrete dimensions, the space is discretized into... The probabilistic grid network (i.e., the actual physical size of a single grid cell is) Five targets are randomly distributed within the area to be searched (the target generation logic must satisfy the distance from the boundary). and distance from the obstacle (Safety collision avoidance constraints). The environment contains 7 circular static obstacles of different radii and 6 rectangular static obstacles of different sizes, with a total obstacle coverage of approximately 15%.

[0091] (2) Kinematics and initial configuration of the unmanned aerial vehicle platform: The collaborative cluster consists of three rotary-wing UAVs, whose initial continuous physical coordinates are configured as follows: , and They are widely dispersed. The maximum maneuvering speed of the drones is limited to... The physical step size for discrete simulation of the control system is strictly set to... .

[0092] (3) Parameters sensed by airborne sensors: Maximum effective physical detection radius of the sensor The value is set to 35.0m. When performing multiplicative updates on the local map using the Gaussian kernel function, the maximum detection confidence decay coefficient of the kernel function is set to 0.95, and the Gaussian kernel width coefficient is adaptively calculated from the sensing radius (i.e., ...). In the confidence accumulation logic of target acquisition, the single-observation confidence gain coefficient... Natural decay coefficient of lost field of vision The cumulative confidence physical limit is 20.0, and the threshold for successful capture is 15.0.

[0093] (4) Constraints and fusion parameters of weak communication networks: To rigorously simulate a restricted communication environment, the maximum effective physical interaction radius of a single machine is set to [value]. (Much smaller than the diagonal length of the environment). The discrete communication interaction cycle is strictly limited to 5 control steps (i.e., every...) (One interaction). When performing local map fusion, a "clean-up priority" strategy is adopted, and the minimum value is set as the dominant weight. Mean smoothing weight .

[0094] (5) Memory-Enhanced Planner (MA-AHA) core parameters: To balance real-time computing power and optimization accuracy, the size of the candidate waypoint population maintained by the planner in a single iteration is set to [size missing]. The maximum number of iterations for real-time optimization is set to 30. In the dynamic search radius mechanism, the initial exploration radius is set to 200.0m, the upper limit of exploration is set to 650.0m, and the adaptive growth step of the radius when there is no convergence in a single attempt is 60.0m.

[0095] (6) Simulation termination and verification conditions: The maximum step size for a single Monte Carlo simulation is rigidly limited to 9000 steps (i.e., 1800 seconds of physical time), and the patience value for early system termination is set at 3500 steps with no progress. Under the above-mentioned stringent and detailed parameter boundaries, this invention conducted 200 independent Monte Carlo simulation experiments to comprehensively evaluate the statistical performance indicators of the algorithm.

[0096] The results show that: 1. Convergence speed: such as Figure 3 As shown, the average task completion time (TTD) of this method is 306.85 seconds, which is about 6.2% shorter than the benchmark AHA algorithm; 2. Security: such as Figure 4 As shown, it maintained a 0% collision rate in a complex environment with 15% obstacle coverage, while the comparison algorithm had a 0.5% collision risk. 3. Energy consumption: As shown in Figures 5(a)-(f), the average flight path length of the UAV swarm is reduced, which effectively improves the efficiency of operation coverage.

[0097] The present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements a multi-UAV cooperative search method based on a memory-enhanced artificial hummingbird algorithm as described above.

[0098] The present invention designs an adaptive spatiotemporal constrained distributed dynamic communication topology and a clearing-priority local map fusion strategy: abandoning the assumption of fully connected communication, it reduces the amount of data transmission from the source by using spatial communication radius and temporal communication cycle as dual constraints; and proposes a fusion rule that prioritizes the adoption of low-probability information to solve the problems of information silos and cognitive conflicts under communication constraints, and quickly suppress the spread of false target noise.

[0099] The method of this invention transforms the optimal navigation target point generated by the MA-AHA algorithm in continuous space into continuous velocity vector control commands. By introducing a tangential component into the obstacle repulsion field through an improved artificial potential field obstacle avoidance, smooth flight around obstacles is achieved. Combined with a dynamic neighbor separation mechanism (i.e. separation pressure term) under communication constraints, inter-machine collisions are effectively avoided while maintaining the search coverage. At the same time, a deadlock detection and escape mechanism based on historical trajectory sliding window is designed to solve the problem of UAVs getting stuck in local extrema in complex obstacle environments and improve the robustness of the algorithm.

[0100] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0101] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A multi-UAV cooperative search method based on a memory-enhanced artificial hummingbird algorithm, characterized in that, Includes the following steps: S1. Based on the search area environment, construct a hybrid spatial environment model that includes a continuous dynamic domain and a discrete probability perception domain, and obtain the continuous coordinates and discrete probability grid map of the UAV based on the hybrid spatial environment model. S2. Based on the preset communication radius and discrete communication period, construct a distributed dynamic communication topology under adaptive spatiotemporal constraints based on the continuous coordinates of the UAV. The distributed dynamic communication topology is a local data link established between UAVs when the spatiotemporal constraints are met. S3. Obtain environmental information of neighboring nodes through the established local data links, and update the local discrete probability raster map by adopting a local map fusion strategy based on the principle of clearing priority. S4. Using the local discrete probability grid map output by S3, the distributed dynamic communication topology output by S2, and the current UAV status as input, construct a multi-objective fitness function and execute the memory-enhanced MA-AHA path planning algorithm to generate the optimal navigation target point. S5. Convert the optimal navigation target point into a speed control command to drive the UAV to move; when a motion deadlock is detected in the UAV, correct the speed command through a multimodal deadlock recovery mechanism to guide the UAV to get out of trouble.

2. The multi-UAV cooperative search method based on the memory-enhanced artificial hummingbird algorithm according to claim 1, characterized in that, In S1, the hybrid spatial environment model includes smooth motion modeling in a continuous dynamic domain and Gaussian perception modeling in a discrete probability perception domain; the smooth motion modeling in the continuous dynamic domain includes velocity recursion equations and position update equations: Velocity recurrence relation: Position update equation: in, and The first A drone in Time and The actual flight velocity vector at any given moment, for The desired speed control command at the underlying level at any given moment. The inertial smoothing coefficient, and The first A drone in Time and The continuous spatial coordinate vector at time t. This represents the discrete control step size of the model.

3. The multi-UAV cooperative search method based on the memory-enhanced artificial hummingbird algorithm according to claim 2, characterized in that, The discrete probability sensing domain Gaussian sensing modeling includes: rasterizing the continuous search region into discrete grids of equal size, and performing multiplicative updates on the grid probabilities within the UAV's sensing radius based on a Gaussian kernel function. in, For the updated Time of the first In the perception map maintained by the drone itself, the center location is The probability of the existence of a grid target. Before the update The probability that a grid target exists at this location at any given time. This represents the sensor's maximum detection confidence level. The Gaussian kernel width coefficient, For the first A drone in Continuous spatial position coordinates at any given time. The center coordinates of the discrete raster. This represents the Euclidean distance between the drone's current position and the center of the grid.

4. The multi-UAV cooperative search method based on the memory-enhanced artificial hummingbird algorithm according to claim 1, characterized in that, In step S2, the construction of a distributed dynamic communication topology under adaptive spatiotemporal constraints based on the continuous coordinates of the UAV specifically includes: By simulating communication radius constraints and communication time slot constraints, a UAV inter-UAV communication link criterion based on spatial filtering and temporal sparsity is constructed. This satisfies the spatiotemporal constraints of the communication link criterion, establishing a distributed communication link. The UAV inter-UAV communication link criterion based on spatial filtering and temporal sparsity is as follows: in, For time k, the first The drone and the first The communication link status of the drone is 1 for established link and 0 for disconnected link. For the first A drone in Continuous spatial position coordinates at any given time. For the first A drone in Continuous spatial position coordinates at any given time. To limit the communication radius, it is used for spatial filtering. For discrete communication periods, used to implement time sparsity. This is the modulo operation.

5. A multi-UAV cooperative search method based on a memory-enhanced artificial hummingbird algorithm according to claim 1, characterized in that, In S3, the local map fusion strategy based on the principle of prioritizing clearing is defined by the following formula: in, For the fusion Time of the first The drone is used for local maintenance, and its central location is... The final update value is the probability of target existence in the discrete raster. For a set of nodes that are effectively connected, and These are the minimum-dominant weight and the mean-smoothing weight coefficients, respectively, and they satisfy... ; The feature term representing the neighborhood minimum is extracted from... Time-local communication set All drones within the area are in coordinates The probability of the minimum target being assigned; This is the mean smoothing auxiliary term. Time-local communication set All drones within the area are in coordinates The arithmetic mean of probability observations.

6. The multi-UAV cooperative search method based on memory-enhanced artificial hummingbird algorithm according to claim 1, characterized in that, In S4, the constructed multi-objective fitness function is as follows: in: in, As candidate navigation target points, , , , As weight, For the target gain term, This represents the probability of the target existing in the local discrete probability map corresponding to the candidate waypoint; As a historical penalty item, The preset threshold for determining whether an area has been explored; For distance-based energy consumption, This is the current location of the drone; To separate the pressure term, For safe distance threshold, The coordinates of neighboring nodes obtained based on the communication link.

7. A multi-UAV cooperative search method based on a memory-enhanced artificial hummingbird algorithm according to claim 1, characterized in that, In step S4, a multi-objective fitness function is constructed and a memory-enhanced MA-AHA path planning algorithm is executed to generate the optimal navigation target point, specifically including: The optimization objective is to minimize the multi-objective fitness function, with candidate waypoints in continuous space as optimization variables. The population is randomly initialized, and the multi-objective fitness function of the initial population is calculated. A dual-mode search strategy is triggered by random probability. Guided foraging pattern: Utilizing the group's optimal state information achievable under current communication conditions to apply long-range attraction, driving individuals to escape their current local convergence region and achieving position updates: when random probability... When entering global exploration mode, the location update method is as follows: in, For new destinations, For the first The current waypoint of the drone. To guide the scaling factor of foraging stride, To guide waypoints to the target For nondeterministic perturbation vectors that follow a standard normal distribution; Territorial foraging mode: This mode utilizes a nondeterministic perturbation vector following a standard normal distribution to sample the neighborhood of the current solution, achieving position updates: when the random probability... When entering partial development mode, the location update method is as follows: in, Scaling factor for territorial foraging step size; Each iteration of the guided foraging mode or the territory foraging mode calculates the fitness value of candidate waypoints. Finally, after multiple iterations and convergence, the waypoint with the minimum fitness is output. As the optimal navigation target point for the next moment.

8. A multi-UAV cooperative search method based on a memory-enhanced artificial hummingbird algorithm according to claim 1, characterized in that, In step S5, the velocity control command for the optimal navigation target point conversion is composed of the target gravitational field velocity vector, the static obstacle repulsive field velocity vector, and the dynamic neighbor repulsive field velocity vector; the velocity control command is as follows: in, The target gravitational field velocity vector, Let $\mathbf{ ... The magnitude of gravity; The velocity vector of the repulsive force field of the static obstacle. For the normal repulsive component, This is the tangential repulsive force component; The velocity vector of the dynamic neighbor repulsion field. This is the direction vector of the current drone moving away from neighboring drones; This represents the magnitude of the repulsive force.

9. A multi-UAV cooperative search method based on a memory-enhanced artificial hummingbird algorithm according to claim 1, characterized in that, In S5, the speed command is corrected through a multimodal deadlock recovery mechanism, specifically including: The drone maintains a fixed-length historical trajectory buffer in real time and calculates the Euclidean displacement between the current position and the starting point of the buffer. If the Euclidean displacement is less than the set stagnation threshold, the drone is determined to have motion deadlock. When motion deadlock is detected, a preset amplitude escape velocity vector is applied for a fixed duration. The speed commands actually executed by the drone are corrected.

10. A multi-UAV cooperative search method based on a memory-enhanced artificial hummingbird algorithm according to claim 1, characterized in that, The collaborative search method also includes a confidence accumulation confirmation mechanism based on Bayesian ideas: To simulate the sensor detection characteristics of a drone in a real-world physical scenario, the observation confidence level is calculated and accumulated in real time based on the relative distance between the drone and the target and the environmental occlusion status. The discrete-time update formula is as follows: in, for Time of the first The cumulative observation confidence level of the drone on the target. The upper limit is the cutoff value. This refers to the step size in discrete simulation. The rate of change function of the confidence level at the current moment is calculated using the following formula: in, For the continuous spatial coordinates of the UAV, For the target coordinates, This is the maximum effective detection radius of the sensor. Based on the observation gain coefficient, This is the confidence decay coefficient when the field of view is lost.

Images