Multi-unmanned aerial vehicle task allocation and path planning dual-stage joint optimization method and system

Abstract

The invention provides a multi-unmanned aerial vehicle task allocation and path planning dual-stage joint optimization method and system, and belongs to the technical field of unmanned aerial vehicle path planning. According to the invention, by constructing a novel three-dimensional multi-unmanned aerial vehicle task allocation and path planning dual-stage joint optimization framework, collaborative optimization of task allocation and path planning is realized; wherein the RL-TPTA algorithm in the task allocation stage dynamically optimizes a task access sequence by means of K-means clustering and reinforcement learning, the total flight distance of the unmanned aerial vehicle can be effectively reduced, and the adaptability in a large-scale task scene is higher; the MSM-EABC algorithm in the path planning stage is higher in convergence speed through multi-strategy fusion, and can efficiently jump out of local optimum and avoid three-dimensional environment collision risks; the two-stage framework combines the advantages of the two, the path length can be remarkably shortened in different-scale task scenes, the convergence efficiency is improved, the operation time is shortened, the adaptability to the complex three-dimensional environment is better, and the task allocation and path planning coupling optimization problem in the three-dimensional unmanned aerial vehicle cluster search and rescue after disaster is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle path planning, in particular to a multi-unmanned aerial vehicle task allocation and path planning two-stage joint optimization method and system. BACKGROUND

[0002] In post-earthquake search and rescue operations, unmanned aerial vehicle search and rescue technology has become the core force to break through the traditional rescue bottleneck due to its unique advantages of high efficiency and strong environmental adaptability. However, the complex rescue environment derived from earthquake disasters poses a serious challenge to search and rescue efficiency. First, the terrain barrier problem, the disaster area often forms an "island effect" due to landslides and road collapses, making it difficult for ground rescue teams to quickly reach the core trapped area. Second, the influence of extreme climate, low temperature environment not only accelerates the loss of body temperature of trapped personnel, compresses the "golden 72 hours" rescue window, but also makes rescue personnel face the risk of frostbite, and the operation efficiency is greatly reduced. Third, geographical location restriction, the epicenter is often located in remote areas, and professional rescue forces are difficult to assemble in time due to traffic congestion, delaying the best rescue opportunity. Fourth, the threat of secondary disasters, the risk of secondary collapse of buildings caused by aftershocks poses a continuous threat to the safety of rescue site personnel. Fifth, the communication system is paralyzed, the earthquake destroys the ground communication facilities, and the low temperature environment exacerbates the performance degradation of electronic equipment, resulting in serious obstacles to command and dispatch and on-site information transmission. In the face of the above difficulties, unmanned aerial vehicles, with their advantages of high efficiency and safety, flexibility, and convenient deployment, provide innovative solutions to overcome the response delay, poor environmental adaptability, and high disposal cost of traditional emergency rescue methods, and are deeply promoting the technological revolution in the field of emergency rescue.

[0003] The UAV swarm task allocation problem can be considered as a high-order variant of the MTSP in a three-dimensional post-disaster scenario: the core connection is that both need to allocate discrete task points to multiple agents and optimize the access path to minimize the total cost; but the key is to avoid dynamic obstacles (such as mountains, ruins) in complex three-dimensional terrain, strictly meet the physical constraints of UAVs, and solve the problem of swarm collaborative obstacle avoidance and generate continuous and smooth flight trajectories (non-simple point sequence path), which significantly surpasses the two-dimensional static environment and combinatorial optimization category of traditional MTSP. There are many difficulties in the UAV swarm task allocation problem at present, and many scholars have conducted in-depth modeling and algorithm research on this problem, aiming to improve the efficiency and optimization performance of task allocation. Mao et al. constructed a multi-UAV task allocation model as a Markov decision process, and designed a policy network based on the transformer style for the task allocation of UAV swarm. Kurdi et al. proposed an algorithm based on the foraging behavior of bacteria for the optimal task allocation problem of multi-UAV system. Some scholars have proposed centralized and distributed task allocation methods for this problem. The centralized method realizes task allocation through a global optimization model, but faces the problem of high computational complexity. For example, Liu et al. proposed a centralized task allocation strategy for the UAV swarm task allocation problem. Although this method can be globally optimized, it may cause bottlenecks when the number of UAVs is large. The decentralized task allocation model proposed by Petrenko et al. is to make decisions independently by each UAV. This method improves the robustness of the system and reduces the burden of central control, but may lead to a non-optimal global solution. In addition to centralized and decentralized task allocation models, some scholars have also proposed task allocation models based on task requirements and UAV capabilities. Skaltsis et al. assigned UAVs to specific tasks according to their different height maneuverability or sensor capabilities based on the specific needs of some tasks. Rinaldi et al. proposed a greedy auction algorithm, which allocates heterogeneous tasks to heterogeneous UAVs through a bidding mechanism. The current research on UAV swarm task allocation planning algorithms faces many challenges. It is worth noting that most algorithms can effectively solve the problem of reasonably allocating tasks to multiple UAVs in the environment, but none of them have detailed processing of the subsequent flight path of each UAV. Future research directions will focus on improving the intelligent real-time performance and adaptability of algorithms. At the same time, combining emerging technologies such as deep learning and reinforcement learning, it will be possible to provide more powerful support for the task allocation and subsequent flight path planning of UAVs, and promote the application of UAV swarm in complex tasks.

[0004] In recent years, with the deepening research in the field of UAV path planning, scholars at home and abroad have summarized many efficient UAV path planning algorithms. In past research, exact algorithms, including A* algorithm and Dijkstra's algorithm, have been widely used as classic path planning algorithms to solve UAV path planning problems. However, exact algorithms require traversing the entire solution space to find the shortest path, resulting in high computational complexity, making them unsuitable for solving large-scale, highly complex problems. Furthermore, they suffer from the limitation of unreachable target nodes during the planning process. Therefore, exact algorithms are not suitable for large-scale, complex terrain. Compared to exact algorithms, approximate algorithms, with their excellent global search capabilities in complex environments, are increasingly widely used in the field of path planning.

[0005] In the field of approximate algorithms, metaheuristic algorithms have gradually become a research focus due to their unique optimization mechanisms and superior performance. These algorithms construct heuristic search frameworks by simulating natural phenomena or biological behavior, providing innovative solutions to the UAV path planning problem. Zheng et al. modeled the multi-UAV search path optimization problem as a distributed constrained optimization problem in a finite time domain and proposed an enhanced genetic algorithm based on a distributed stochastic algorithm. Liu et al. proposed an adaptive genetic algorithm and an improved artificial bee colony algorithm for UAV swarms performing post-disaster medical rescue missions, and developed a balanced search strategy to improve the algorithm's search efficiency. Chen et al. applied the ant colony algorithm to solve the coverage path planning problem in a multi-UAV cooperative search system. Hu et al. proposed an enhanced multi-strategy dolphin optimization algorithm. By introducing fish swarm aggregation strategy, random mixed mutation strategy, and chaotic oppositional learning strategy to balance the algorithm's exploration and development capabilities, the algorithm's comprehensive ability to solve the UAV swarm cooperative operation path planning problem was effectively improved.

[0006] In the field of metaheuristic algorithms, the Artificial Bee Colony Algorithm (ABC) has become a focus of attention due to its unique swarm intelligence mechanism, excellent balance between global exploration and local exploitation, and strong robustness to solutions to complex problems. In recent years, with the increasing complexity of UAV path planning problems, the ABC algorithm has been increasingly applied to UAV trajectory generation due to its adaptive search strategy, low parameter dependence, and efficient convergence characteristics. Xu et al. proposed an improved ABC algorithm based on chaos theory (Chaotic ABC) to solve the path planning problem of unmanned combat aerial vehicles (UCAVs) in complex battlefield environments. Han et al. proposed an improved ABC algorithm based on multi-strategy evolutionary learning (MSEL-ABC) for the path planning problem of unmanned autonomous helicopters (UAHs). This method effectively solves the problems of traditional ABC algorithms easily getting trapped in local optima and having low search efficiency under complex constraints, improving the autonomy and intelligence of path planning. Lin proposed an improved Artificial Bee Colony Algorithm (IABC) based on multi-strategy synthesis. This method effectively solves problems such as high-dimensional optimization, susceptibility to local optima, and path discontinuity in UAV path planning in complex urban environments by integrating multiple improved strategies, including tent chaotic mapping and Pareto principle initialization of the population. Although ABC plays an important role in multivariate heuristic algorithms, its inherent defects are becoming increasingly prominent in complex engineering optimization: First, in the face of high-dimensional and strongly constrained optimization scenarios, the algorithm's convergence speed is prone to a significant decrease due to the dramatic increase in solution space complexity, resulting in limited search efficiency; second, the quality of the initial population has a great impact on the final solution. If the initial solution distribution deviates from the optimal region, it is difficult to effectively correct the search direction even through iterative optimization; more importantly, standard ABC relies on a single search equation during the development phase, and is prone to getting trapped in local optima in multi-peaked or rugged solution spaces, leading to premature convergence. These problems may directly affect the collaborative efficiency and mission success rate of UAV swarms in tasks with stringent real-time requirements, such as post-disaster search and rescue.

[0007] With the rapid development of machine learning technology, reinforcement learning, as an important branch, has gradually emerged in the field of UAV path planning. Sonny et al. proposed a UAV path planning method based on reinforcement learning algorithms, which can simultaneously avoid both static and dynamic obstacles, achieving more efficient UAV path planning. Wu et al. proposed an improved particle swarm optimization algorithm based on deep reinforcement learning for multi-UAV collaborative search and rescue missions. By fully utilizing the powerful expressive capabilities of reinforcement learning and the global search capabilities of particle swarm optimization, they effectively improved the effectiveness and performance of path planning. Yu et al. proposed a multi-objective differential evolution algorithm based on reinforcement learning, dynamically adjusting relevant parameters in the algorithm through reinforcement learning. Puente-Castro et al. combined reinforcement learning algorithms and artificial neural networks to control the flight paths of UAV swarms on maps with obstacles. This combined approach can improve the autonomy and adaptability of UAV swarms, effectively saving costs, improving time efficiency, and enhancing fault tolerance.

[0008] The limitations of the existing technologies mentioned above in handling the task allocation and path planning problem of 3D UAV swarm disaster search and rescue include: traditional mathematical programming algorithms are difficult to handle large-scale dynamic constraints due to the curse of dimensionality; classical heuristic algorithms are prone to getting trapped in suboptimal solutions; centralized scheduling frameworks lack the ability to adapt to environmental changes; task allocation and path planning are often processed separately, resulting in insufficient global optimality; traditional path planning algorithms are difficult to cope with the high complexity and dynamic obstacle avoidance of 3D disaster scenarios and do not adequately consider multi-UAV cooperative collision avoidance; metaheuristic algorithms are prone to getting trapped in local optima due to poor initial population quality and single search strategy, making it difficult to meet the requirements of real-time performance and robustness. Summary of the Invention

[0009] The purpose of this invention is to provide a two-stage joint optimization method and system for multi-UAV task allocation and path planning. It aims to construct a novel two-stage joint optimization framework for three-dimensional multi-UAV task allocation and path planning. Through the collaborative optimization of a three-stage algorithm based on reinforcement learning in the task allocation stage and an efficient artificial bee colony algorithm based on multi-strategy fusion in the path planning stage, it solves the problem of coupled optimization of task allocation and path planning in UAV swarm disaster search and rescue in a three-dimensional environment, realizes efficient search and rescue mission execution in complex three-dimensional environments, and improves path length optimization, convergence efficiency and adaptability to complex environments.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] In a first aspect, the present invention provides a two-stage joint optimization method for multi-UAV task allocation and path planning, comprising:

[0012] The K-means clustering algorithm is used to initially assign search and rescue tasks, ensuring that task points are reasonably allocated to each UAV according to spatial distribution; an initial access order is generated based on the task location to plan a preliminary task execution sequence for each UAV; and the task access order is optimized through reinforcement learning algorithm.

[0013] Based on the optimized task access order, and combined with the artificial bee colony algorithm using a linear interpolation perturbation strategy, a bee colony aggregation strategy is introduced to perform more effective global search and local exploration in the solution space. Based on the global exploration mechanism of Levy flight, the task flight path scheme is determined.

[0014] As a further limitation of the first aspect of the present invention, the initial task allocation model is based on the K-means clustering algorithm, representing all search and rescue task points as a set of coordinate points P = {p1, p2, ..., p...} in three-dimensional space. n}, where p i =(x i ,y i ,z i ) represents the coordinates of the i-th task point; the search and rescue task point data are clustered into K clusters according to the number of drones, and the task points in each cluster are assigned to the corresponding drones.

[0015] As a further limitation of the first aspect of the present invention, after assigning specific tasks to each drone, an initial access order is generated according to the principle of proximity to distance, including: assuming the drone's flight starting point is S = (x s ,y s ,z s The set of task points is P = {p1, p2, ..., p...} n}; where p i =(x i ,y i ,z i ) represents the coordinates of the i-th task point; for each task point p i Calculate the Euclidean distance d between it and the UAV's flight starting point S. i The formula is All task points are arranged according to distance d i Sort the tasks from smallest to largest to obtain an ordered sequence of task points P' = {p'1, p'2, ..., p'}. n}, where p'1 is the task point closest to the starting point, p' n The task point is the one furthest from the starting point. Finally, based on the sorted task point sequence P', the initial visiting order of the drone is generated, that is, the drone first visits the task point p'1 closest to the starting point, and then visits p'2, p'3, ..., p'' in sequence. n .

[0016] As a further definition of the first aspect of this invention, the global exploration mechanism of Lévy flight includes: utilizing the long step size jump allowed by the heavy-tailed characteristic of the Lévy distribution to guide the scout bee out of the local optimum region. Specifically, the widely used Mantegna algorithm is used to approximate the Lévy step size s, which is calculated by combining two independent standard normal distributed random variables u and v. The Lévy step size approximately follows a Lévy distribution. In the scout bee phase of the artificial bee colony algorithm, the Lévy step size is used to reset the solution to the number of judgments required to reach a local optimum, thereby enhancing global exploration. If the new solution exceeds the solution space boundary, reflection processing is performed.

[0017] As a further limitation of the first aspect of the present invention, the bee swarming strategy includes: dynamically controlling the switching frequency of bees between collective cooperative search and random wandering by adjusting the constant FP, effectively balancing the exploration and development process, improving the algorithm's ability to escape local optima and the diversity of solutions; in each iteration, the bees integrate the search strategy guided by the global optimum and the behavioral strategy inspired by bee swarming to generate dual candidate solutions in parallel; by directly comparing the fitness values ​​of the two sets of candidate solutions, the better individual is selected and retained to enter the next generation of the population.

[0018] As a further limitation of the first aspect of the present invention, the linear interpolation perturbation strategy is to generate intermediate points on the line connecting the starting point and the ending point, and then apply a small-range random perturbation to the intermediate points; specifically, it includes: firstly, uniformly sampling N intermediate points on the line connecting the starting point S and the ending point T. A baseline path P0 is constructed, and the coordinates of its intermediate points are calculated using linear interpolation. Subsequently, a three-dimensional random perturbation is applied to each baseline point, and an offset δ is generated by sampling in a spherical coordinate system. i Generate perturbed path points To dynamically balance the quality and diversity of the initial solution, the strategy introduces a linear adaptive decay mechanism for the perturbation radius, which means that the perturbation radius gradually decreases as the number of iterations increases.

[0019] Secondly, the present invention provides a two-stage joint optimization system for multi-UAV task allocation and path planning, comprising:

[0020] The task allocation module is used to initially allocate search and rescue tasks using the K-means clustering algorithm, ensuring that task points are reasonably allocated to each UAV according to spatial distribution; it generates an initial access order based on task location, planning a preliminary task execution sequence for each UAV; and it optimizes the task access order through reinforcement learning algorithms.

[0021] The path planning module is used to determine the mission flight path scheme based on the optimized mission access order, combined with the artificial bee colony algorithm with a straight line interpolation perturbation strategy, and the bee colony aggregation strategy. It also incorporates a bee colony aggregation strategy to perform more effective global search and local development in the solution space, based on the global exploration mechanism of Levy flight.

[0022] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the two-stage joint optimization method for multi-UAV task allocation and path planning as described in the first aspect.

[0023] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the two-stage joint optimization method for multi-UAV task allocation and path planning as described in the first aspect.

[0024] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the two-stage joint optimization method for multi-UAV task allocation and path planning as described in the first aspect.

[0025] Terminology Explanation:

[0026] NP-hard problems: Optimization problems with extremely high computational complexity. B-spline path fitting: A core technology for transforming discrete waypoints into continuous, smooth flight trajectories, generating continuously differentiable smooth paths through mathematical interpolation. Greedy strategy: A decision-making method that maximizes immediate gains based on current information. Its core idea is "choosing the best option at each step." Crossover operator: A local search operator in reinforcement learning that optimizes the task access order, used to generate new task sequences. Three-stage task allocation algorithm based on reinforcement learning: The core algorithm in the task allocation stage, achieving efficient task-UAV matching through three-stage collaboration. Efficient artificial bee swarm algorithm based on multi-strategy fusion: The core algorithm in the path planning stage, improving 3D path planning performance by fusing three improved strategies: straight-line interpolation perturbation strategy, bee swarm aggregation strategy, and Lévy flight-guided scout bee mechanism. Maximum climb angle: The maximum pitch angle threshold limited by the UAV's power performance. Lévy flight: A random walk with a heavy-tailed probability distribution of step size, meaning there is a relatively high probability of large steps during random walks.

[0027] The beneficial effects of this invention are as follows: By constructing a novel two-stage joint optimization framework for 3D multi-UAV task allocation and path planning, the collaborative optimization of task allocation and path planning is achieved, overcoming the problem of insufficient global optimality caused by the independent solution of the two in traditional methods. In the task allocation stage, the RL-TPTA algorithm dynamically optimizes the task access order by using K-means clustering and reinforcement learning, which can effectively reduce the total flight distance of UAVs and is more adaptable to large-scale task scenarios. In the path planning stage, the MSM-EABC algorithm achieves faster convergence speed through multi-strategy fusion, can efficiently escape local optima and avoid collision risks in the 3D environment. The two-stage framework combines the advantages of both, which can significantly shorten the path length, improve convergence efficiency and running time in task scenarios of different scales, and has better adaptability to complex 3D environments, effectively solving the problem of coupled optimization of task allocation and path planning in 3D UAV swarm disaster search and rescue.

[0028] The advantages of additional aspects of the invention will be set forth more clearly in the following description or will be learned by practice of the invention. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of 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.

[0030] Figure 1 This is a flowchart of the three-stage task allocation algorithm based on reinforcement learning as described in an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of the three crossover operators described in the embodiments of the present invention.

[0032] Figure 3 This is a flowchart of the efficient artificial bee colony algorithm based on multi-strategy fusion as described in an embodiment of the present invention. Detailed Implementation

[0033] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0034] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0035] It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as here.

[0036] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, and / or groups thereof.

[0037] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0038] To facilitate understanding of the present invention, the present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0039] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0040] Example 1

[0041] In this embodiment 1, a two-stage joint optimization system for multi-UAV task allocation and path planning is first provided, including: a task allocation module, used to initially allocate search and rescue tasks using the K-means clustering algorithm to ensure that task points are reasonably allocated to each UAV according to spatial distribution; generate an initial access order based on task location to plan a preliminary task execution sequence for each UAV; and optimize the task access order using a reinforcement learning algorithm. The path planning module, based on the optimized task access order, combines an artificial bee colony algorithm with a straight-line interpolation perturbation strategy, introduces a bee colony aggregation strategy, and performs more effective global search and local exploration in the solution space, determining the task flight path scheme based on the Levy flight global exploration mechanism.

[0042] In this embodiment, the aforementioned system is used to implement a two-stage joint optimization method for multi-UAV task allocation and path planning, including: a practical task allocation module, which uses the K-means clustering algorithm to initially allocate search and rescue tasks, ensuring that task points are reasonably allocated to each UAV according to spatial distribution; generates an initial access order based on task location, and plans a preliminary task execution sequence for each UAV; and optimizes the task access order through a reinforcement learning algorithm. The practical path planning module, based on the optimized task access order, combines an artificial bee colony algorithm with a linear interpolation perturbation strategy, introduces a bee colony aggregation strategy, and performs more effective global search and local exploration in the solution space. Based on the Levy flight global exploration mechanism, it determines the task flight path scheme.

[0043] The initial task allocation model is based on the K-means clustering algorithm, representing all search and rescue task points as a set of coordinate points P = {p1, p2, ..., p...} in three-dimensional space. n}, where p i =(x i ,y i ,z i ) represents the coordinates of the i-th task point; the search and rescue task point data are clustered into K clusters according to the number of drones, and the task points in each cluster are assigned to the corresponding drones.

[0044] After assigning specific tasks to each UAV, determining the order in which the UAV visits each task is a crucial step in path planning. A reasonable visit order can significantly reduce the UAV's flight path length and improve task execution efficiency. This paper proposes a method for generating an initial visit order based on task location. This method generates an initial visit order from closest to furthest from each task point's distance to the UAV's flight starting point, laying the foundation for subsequent visit order optimization.

[0045] The core idea of ​​this method is to use the Euclidean distance between the task point and the UAV's flight starting point to generate the initial visiting order according to the principle of proximity to distance. The specific steps of this method are as follows: First, let the UAV's flight starting point be S = (x... s ,y s ,z s The set of task points is P = {p1, p2, ..., p...} n}. Where p i =(x i ,y i ,z i Let p represent the coordinates of the i-th task point. Then, for each task point p... i Calculate the Euclidean distance d between it and the UAV's flight starting point S. i The formula is Then, arrange all task points according to distance d.i Sort the tasks from smallest to largest to obtain an ordered sequence P. ' ={p ' 1,p ' 2,...,p ' n}, where p ' 1 is the task point closest to the starting point, p ' n It is the task point farthest from the starting point; finally, according to the sorted task point sequence P ' The initial access order of the drones is generated, that is, the drones first visit the task point p that is closest to the starting point. ' 1, then visit p in sequence ' 2,p ' 3,…,p ' n This process effectively reduces the total flight path length of the drone and improves mission execution efficiency.

[0046] The global exploration mechanism of Lévy flight includes: utilizing the heavy-tailed characteristic of the Lévy distribution to allow for long-step jumps, guiding scout bees out of local optima. Specifically, the widely used Mantegna algorithm is used to approximate the Lévy step size s, which is calculated by combining two independent standard normally distributed random variables u and v. The Lévy step size approximately follows a Lévy distribution. In the scout bee phase of the artificial bee colony algorithm, the Lévy step size is used to reset the solution to the number of times the scout bee has been trapped in a local optimum, thus enhancing global exploration. If a new solution exceeds the solution space boundary, reflection processing is performed.

[0047] Specifically, addressing the low efficiency of the traditional random reset strategy for scout bees in artificial bee colony algorithms, this embodiment proposes a global exploration mechanism based on Lévy flight. Utilizing the heavy-tailed characteristic of the Lévy distribution, which allows for long-step jumps, the scout bees are guided out of local optima. The implementation employs the widely used Mantegna algorithm to approximate the Lévy step size s, which is calculated by combining two independent standard normal distributed random variables u and v, as shown in the following formula:

[0048]

[0049] in, σ v =1. Where Γ is the gamma function, and the distribution parameter β is typically chosen to be 1.5. The formula for calculating the Lévy step size is as follows:

[0050]

[0051] The step size s approximately follows a Lévy distribution and exhibits heavy-tailed characteristics. In the scout bee phase of the artificial bee colony algorithm, the Lévy step size is used to reset the solution to minimize the number of decisions required to avoid getting trapped in local optima, thus enhancing global exploration. The specific formula is:

[0052] x i,j =x i,j +α×s×(x best -x i,j )

[0053] Where α is the step size factor, controlling the jump amplitude. s is the Lévy step size, generated using the Mantegna algorithm. best This is the current globally optimal solution, used to guide the search direction.

[0054] If the new solution exceeds the boundary of the solution space, reflection processing is performed. The specific formula is:

[0055]

[0056] The swarming strategy includes: dynamically controlling the switching frequency of bees between collective cooperative search and random wandering by adjusting the constant FP, effectively balancing the exploration and development process, and improving the algorithm's ability to escape local optima and the diversity of solutions; in each iteration, the bees combine the search strategy guided by the global optimum with the behavioral strategy inspired by swarming to generate dual candidate solutions in parallel; by directly comparing the fitness values ​​of the two sets of candidate solutions, the better individuals are selected and retained to enter the next generation of the population.

[0057] Specifically, a swarming strategy is introduced during the bee collection phase to simulate group cooperation and random exploration mechanisms. This strategy dynamically controls the switching frequency between collective cooperative search (global exploration) and random wandering (local exploration) by adjusting the constant FP, effectively balancing the exploration and development processes, and improving the algorithm's ability to escape local optima and the diversity of solutions. The specific formula is as follows:

[0058]

[0059] Among them, FX i J(t) represents the position of the i-th bee in the candidate population generated after the t-th iteration. J(t) can adaptively control the bee's exploration behavior based on the number of iterations. i(t) represents the position of the i-th bee in the initial population at iteration t. R is a random number between 0 and 1. FP is the probability of the bees performing both collective cooperative search and random exploration behaviors in the solution space, which is taken as a constant of 0.2 in this study. r1 and r2 are two random numbers between 0 and 1. U is a binary array that compares FP with an array of random numbers in (0, 1), where values ​​less than FP are represented as 1, and the rest as 0. l1 and l2 are random indices of two individuals.

[0060] In each iteration, the foraging bees combine a globally optimal search strategy with a swarm-aggregation-inspired behavioral strategy to generate two candidate solutions in parallel. The algorithm directly compares the fitness values ​​of the two candidate solutions and selects the better individual to enter the next generation of the population. The specific formula is as follows:

[0061]

[0062] BX i (t)=x i +φ*(x i -x k )+Ψ*(x best -x i )

[0063] Where i represents the current foraging bee, and k is the ID of a foraging bee other than i. best is the optimal individual in the current population. φ is a random number between [0,1], representing the degree of perturbation of the i-th nectar source by the k-th nectar source. Ψ represents the influence of the globally optimal individual on the generation of a new solution. Here, f represents the fitness function corresponding to the current optimization problem.

[0064] Example 2

[0065] To address the challenge of coupled optimization of task allocation and path planning in UAV swarm disaster search and rescue in 3D environments, this embodiment constructs a novel two-stage joint optimization framework. This framework achieves efficient search and rescue mission execution in complex 3D environments through the collaborative design of task allocation and path planning. Its core consists of two interrelated algorithmic components. Specifically, the framework first constructs a three-stage task allocation algorithm based on reinforcement learning to achieve spatial clustering and dynamic sequence optimization of task points; then, it designs an efficient artificial bee colony algorithm based on multi-strategy fusion to solve the problems of safe path generation and local optimum avoidance in complex 3D environments. Through information interaction and collaborative optimization, the two stages overcome the problem of insufficient global optimality caused by the independent solution of task allocation and path planning in traditional methods, achieving path length reduction, improved convergence efficiency, and optimized runtime in task scenarios of different scales.

[0066] An efficient and reasonable task allocation model is a key foundation for the successful implementation of UAV swarm search and rescue missions after disasters. This embodiment proposes a three-stage task allocation method based on reinforcement learning. First, the method uses the K-means clustering algorithm to initially allocate search and rescue tasks, ensuring that task points are reasonably distributed to each UAV according to spatial distribution. Second, an initial access order is generated based on task locations, planning a preliminary task execution sequence for each UAV. Finally, a reinforcement learning algorithm is used to optimize the task access order, further improving the efficiency and coordination of task execution. These three stages are interconnected and together constitute a complete task allocation optimization framework, providing theoretical support and practical guidance for efficient search and rescue by UAV swarms in complex post-disaster environments. The flowchart of the three-stage task allocation algorithm based on reinforcement learning is shown below. Figure 1 As shown.

[0067] The initial task allocation model proposed in this embodiment is based on the K-means clustering algorithm. First, all search and rescue task points are represented as a set of coordinate points P = {p1, p2, ..., p...} in three-dimensional space. n}, where p i =(x i ,y i ,z i Let represent the coordinates of the i-th task point. Then, the search and rescue task point data is clustered into K clusters according to the number of drones. Finally, the task points in each cluster are assigned to the corresponding drones. In this way, the task points of each drone form a dense cluster around them, thereby reducing the flight distance of the drones.

[0068] After assigning specific tasks to each drone, determining the order in which the drones visit these tasks is a crucial step in path planning. A reasonable visit order can significantly reduce the drone's flight path length and improve task execution efficiency. This embodiment proposes a method for generating an initial visit order based on task location. This method generates an initial visit order from closest to furthest from each task point and the drone's flight starting point, laying the foundation for subsequent visit order optimization.

[0069] The core idea of ​​this method is to use the Euclidean distance between the task point and the UAV's flight starting point to generate the initial visiting order according to the principle of proximity to distance. The specific steps of this method are as follows: First, let the UAV's flight starting point be S = (x... s ,y s ,z s The set of task points is P = {p1, p2, ..., p...} n}. Where p i =(x i ,y i ,z iLet p represent the coordinates of the i-th task point. Then, for each task point p... i Calculate the Euclidean distance d between it and the UAV's flight starting point S. i The formula is Then, arrange all task points according to distance d. i Sort the tasks from smallest to largest to obtain an ordered sequence P. ' ={p ' 1,p ' 2,...,p ' n}, where p ' 1 is the task point closest to the starting point, p ' n It is the task point farthest from the starting point; finally, according to the sorted task point sequence P ' The initial access order of the drones is generated, that is, the drones first visit the task point p that is closest to the starting point. ' 1, then visit p in sequence ' 2,p ' 3,…,p ' n This process effectively reduces the total flight path length of the drone and improves mission execution efficiency.

[0070] Reinforcement learning (RL) is a machine learning method that learns optimal policies through interactions between an agent and its environment. Its core elements include the agent, environment, state, action, reward, and policy: the agent performs actions that influence the environment, the environment provides feedback with a new state and reward, and the agent adjusts its policy accordingly to maximize cumulative reward. This interaction process is typically modeled as a Markov Decision Process (MDP), where the agent selects an action based on its current state and policy, and the environment transitions to a new state and returns an immediate reward. As an unsupervised learning method, RL can autonomously learn policies that maximize cumulative reward through continuous interaction.

[0071] The Q-value table for reinforcement learning is shown in Table 1.

[0072] Table 1

[0073]

[0074] In 3D UAV task allocation, the initial access order is generated based on the distance between the task point and the starting point. To optimize this order, this embodiment designs three crossover operators for local sequence optimization. A new sequence is generated by simulating the crossover operation of a genetic algorithm, and its total flight path length (the sum of distances for each segment) is iteratively evaluated. The task execution sequence is gradually optimized to the optimum, significantly improving planning efficiency and shortening flight distance. The specific crossover operators are as follows:

[0075] Multiple Position Crossover (MPX): From the current individual's task access sequence, randomly select L consecutive gene positions (i.e., task points) and completely reverse the order of these L task points.

[0076] Position Recombination by Order Adjustment Crossover (PRX): From the current individual's task access sequence, L gene positions (i.e., task points) are randomly selected, and these task points are extracted in the order they appear in the original sequence to form an independent gene fragment. Then, an insertion position is randomly selected in the task sequence, and the gene fragment is inserted into that position.

[0077] Position Recombination by Reverse Order Adjustment Crossover (PRRX): From the current individual's task access sequence, L gene positions (i.e., task points) are randomly selected and rearranged in the reverse order of the original sequence to form a new gene fragment. Then, an insertion position is randomly selected in the task sequence, and the reversed gene fragment is inserted into that position.

[0078] In classic reinforcement learning, agents often select actions based on a greedy strategy. The core idea is to choose the action with the highest Q-value in the current state at each decision step, fully utilizing known high-reward paths to accelerate algorithm convergence. However, this strategy can easily lead to the agent over-reliance on local optima, neglecting the exploration of other actions. Therefore, this embodiment employs a generationally decaying semi-random greedy strategy. By dynamically adjusting the greedy selection coefficient, it prioritizes the selection of the optimal action while randomly exploring other actions with a certain probability, balancing utilization and exploration, and effectively avoiding the algorithm getting trapped in local optima. The greedy selection coefficient is shown below:

[0079]

[0080] The greedy selection coefficient a∈(0,1) and the random number b∈(0,1) are used. When b is greater than a, the action with the largest Q value in state S is selected; when b is less than a, an action in state S is randomly selected.

[0081] The calculation of the generation-wise decaying greedy selection coefficient 'a' is as follows:

[0082]

[0083] Among them, a start and a endThe greed coefficients at the beginning and end are represented by γ, episode represents the current iteration number, and γ represents the greed coefficient at the beginning and end. decay This indicates the number of decay steps, and decay_steps indicates the size of the decay step.

[0084] To solve the UAV 3D path planning problem, this embodiment improves upon the traditional artificial bee colony algorithm framework, proposing a high-efficiency artificial bee colony algorithm based on multi-strategy fusion. Addressing the low efficiency of the initial search phase in traditional artificial bee colony algorithms, a novel linear interpolation perturbation strategy is proposed to heuristically generate the initial population. Furthermore, to overcome the inefficiency caused by overly simplistic search equations in standard artificial bee colony algorithms, a swarm aggregation strategy is introduced to guide the bees in more efficient global search and local exploration within the solution space. Finally, to guide scout bees to more efficiently escape local optima and generate new solutions, a global exploration mechanism based on Lévy flight is proposed. The specific flowchart of the algorithm is shown below. Figure 3 As shown.

[0085] In the standard artificial bee colony algorithm, the initial nectar source location is randomly generated in the solution space. This can lead to inefficient initial solutions that are far from the optimal path, potentially resulting in more circuitous or colliding paths and requiring more iterations for optimization. To address this issue, the Linear Interpolation Perturbation Method (LIPM) is proposed to generate the initial population. This strategy generates intermediate points on the line connecting the start and end points and then applies small-scale random perturbations to these intermediate points. This strategy can ensure that the planned path is as close as possible to the optimal path while satisfying the UAV collision avoidance constraint, thereby accelerating the algorithm's convergence speed.

[0086] Specifically, firstly, N intermediate points are uniformly sampled along the line connecting the starting point S and the ending point T. The baseline path P0 is constructed, and the coordinates of the intermediate points are calculated using linear interpolation as follows:

[0087]

[0088] Among them, (x s ,y s ,z s ) and (x t ,y t ,z t The coordinates of the starting and ending points are represented by , respectively. Then, a three-dimensional random perturbation is applied to each reference point, and an offset δ is generated by sampling in spherical coordinates. i =(Δx) i ,Δy i ,Δz i ), satisfying ∥δi ||2≤∈。 Generate perturbed path points. The specific formula for calculating the random offset is as follows.

[0089] Δx i =∈·sinφcosψ,Δy i =∈·sinφsinψ,Δz i =∈·cosφ φ~[0,π],ψ~[0,2π]

[0090] The perturbation amplitude is limited by a preset maximum perturbation radius to ensure that new solutions are explored near the baseline path. To dynamically balance the quality and diversity of the initial solutions, the strategy introduces a linear adaptive decay mechanism for the perturbation radius, meaning the perturbation radius gradually decreases as the iteration number (k) increases. The formula for calculating the perturbation radius is as follows:

[0091]

[0092] Where, ∈ k is the perturbation radius of the k-th iteration; ∈0 is the initial perturbation radius, set to 0.1·∥TS∥2 in this study. K is the maximum number of iterations. This mechanism allows for larger perturbations in the early stages of the algorithm to enhance diversity, while in the later stages it focuses on refining the high-quality solution region near the baseline path, effectively improving the overall quality of the initial population and laying a good foundation for subsequent optimization. The pseudocode of the linear interpolation perturbation strategy is shown in Table 2. This strategy first generates the initial nectar source location based on the coordinates of the starting and ending points, then adds random three-dimensional perturbations to the N intermediate points on the connecting line, while continuously calculating the population fitness function and iteratively optimizing.

[0093] Table 2

[0094]

[0095] Traditional artificial bee colony optimization (ABC) algorithms, relying on a single search equation, limit the diversity of the population's exploration in the solution space and are prone to premature convergence, especially when dealing with high-dimensional and complex optimization problems, where efficiency and accuracy are limited. Inspired by the intelligent behavior of shark swarms, this embodiment introduces a swarm aggregation strategy during the bee collection phase to simulate a collaborative and random exploration mechanism. This strategy dynamically controls the switching frequency of bees between collective collaborative search (global exploration) and random wandering (local exploration) by adjusting the constant FP, effectively balancing the exploration and development processes, and improving the algorithm's ability to escape local optima and the diversity of solutions.

[0096] The specific formula is as follows:

[0097]

[0098] Among them, FX iJ(t) represents the position of the i-th bee in the candidate population generated after the t-th iteration. J(t) can adaptively control the bee's exploration behavior based on the number of iterations. i (t) represents the position of the i-th bee in the initial population at iteration t. R is a random number between 0 and 1. FP is the probability of the bees performing both collective cooperative search and random exploration behaviors in the solution space, which is taken as a constant of 0.2 in this study. r1 and r2 are two random numbers between 0 and 1. U is a binary array that compares FP with an array of random numbers in (0, 1), where values ​​less than FP are represented as 1, and the rest as 0. l1 and l2 are random indices of two individuals.

[0099] In each iteration, the foraging bees combine a globally optimal search strategy with a swarm-aggregation-inspired behavioral strategy to generate two candidate solutions in parallel. The algorithm directly compares the fitness values ​​of the two candidate solutions and selects the better individual to enter the next generation of the population. The specific formula is as follows:

[0100]

[0101] BX i (t)=x i +φ*(x i -x k )+Ψ*(x best -x i )

[0102] Where i represents the current foraging bee, and k is the ID of a foraging bee other than i. best is the optimal individual in the current population. φ is a random number between [0,1], representing the degree of perturbation of the i-th nectar source by the k-th nectar source. Ψ represents the influence of the globally optimal individual on the generation of a new solution. Here, f represents the fitness function corresponding to the current optimization problem.

[0103] To address the low efficiency of the traditional random reset strategy for scout bees in artificial bee colony algorithms, this embodiment proposes a global exploration mechanism based on Lévy flight. Utilizing the heavy-tailed characteristic of the Lévy distribution, which allows for long-step jumps, the mechanism guides scout bees out of local optima. Specifically, the widely used Mantegna algorithm is used to approximate the Lévy step size s, which is calculated by combining two independent standard normal distributed random variables u and v, as shown in the following formula:

[0104]

[0105] in, σ v =1. Where Γ is the gamma function, and the distribution parameter β is typically chosen to be 1.5. The formula for calculating the Lévy step size is as follows:

[0106]

[0107] The step size s approximately follows a Lévy distribution and exhibits heavy-tailed characteristics. In the scout bee phase of the artificial bee colony algorithm, the Lévy step size is used to reset the solution to minimize the number of decisions required to avoid getting trapped in local optima, thus enhancing global exploration. The specific formula is:

[0108] x i,j =x i,j +α×s×(x best -x i,j )

[0109] Where α is the step size factor, controlling the jump amplitude. s is the Lévy step size, generated using the Mantegna algorithm. best This is the current globally optimal solution, used to guide the search direction.

[0110] If the new solution exceeds the boundary of the solution space, reflection processing is performed. The specific formula is:

[0111]

[0112] In this embodiment, the proposed efficient artificial bee colony algorithm based on multi-strategy fusion improves upon the traditional artificial bee colony algorithm by modifying the initial population generation and the search for foraging bees and scout bees, specifically addressing the characteristics of the 3D UAV swarm disaster search and rescue path planning problem. This enhances the algorithm's search efficiency and performance. The pseudocode of the efficient artificial bee colony algorithm based on multi-strategy fusion is shown in Table 3. In the initialization phase, the proposed improved algorithm uses a linear interpolation perturbation strategy to generate the initial population. In the foraging bee phase, a swarm aggregation strategy is introduced to guide the foraging bees in exploring the entire solution space. Finally, in the scout bee phase, a Lévy flight strategy is used to guide the scout bees out of local optima and into global exploration.

[0113] Table 3

[0114]

[0115]

[0116] Example 3

[0117] This embodiment 3 provides a non-transitory computer-readable storage medium for storing computer instructions. When these computer instructions are executed by a processor, they implement the two-stage joint optimization method for multi-UAV task allocation and path planning as described above. This method includes:

[0118] The K-means clustering algorithm is used to initially assign search and rescue tasks, ensuring that task points are reasonably allocated to each UAV according to spatial distribution; an initial access order is generated based on the task location to plan a preliminary task execution sequence for each UAV; and the task access order is optimized through reinforcement learning algorithm.

[0119] Based on the optimized task access order, and combined with the artificial bee colony algorithm using a linear interpolation perturbation strategy, a bee colony aggregation strategy is introduced to perform more effective global search and local exploration in the solution space. Based on the global exploration mechanism of Levy flight, the task flight path scheme is determined.

[0120] Example 4

[0121] This embodiment 4 provides a computer device, including a memory and a processor, wherein the processor and the memory communicate with each other, and the memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute the two-stage joint optimization method for multi-UAV task allocation and path planning as described above, the method including:

[0122] The K-means clustering algorithm is used to initially assign search and rescue tasks, ensuring that task points are reasonably allocated to each UAV according to spatial distribution; an initial access order is generated based on the task location to plan a preliminary task execution sequence for each UAV; and the task access order is optimized through reinforcement learning algorithm.

[0123] Based on the optimized task access order, and combined with the artificial bee colony algorithm using a linear interpolation perturbation strategy, a bee colony aggregation strategy is introduced to perform more effective global search and local exploration in the solution space. Based on the global exploration mechanism of Levy flight, the task flight path scheme is determined.

[0124] Example 5

[0125] This embodiment 5 provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes instructions to implement the two-stage joint optimization method for multi-UAV task allocation and path planning as described above. The method includes:

[0126] The K-means clustering algorithm is used to initially assign search and rescue tasks, ensuring that task points are reasonably allocated to each UAV according to spatial distribution; an initial access order is generated based on the task location to plan a preliminary task execution sequence for each UAV; and the task access order is optimized through reinforcement learning algorithm.

[0127] Based on the optimized task access order, and combined with the artificial bee colony algorithm using a linear interpolation perturbation strategy, a bee colony aggregation strategy is introduced to perform more effective global search and local exploration in the solution space. Based on the global exploration mechanism of Levy flight, the task flight path scheme is determined.

[0128] In summary, the dual-stage joint optimization method and system for multi-UAV task allocation and path planning described in this invention constructs a dual-stage joint optimization framework for task allocation and path planning in post-disaster search and rescue of 3D UAV swarms. It proposes an integrated solution for UAV swarm task allocation and path planning in 3D post-disaster environments. This framework decomposes the problem into two interrelated stages, achieving coupled optimization of task allocation and path planning for UAV swarms. In the task allocation stage, a three-stage task allocation algorithm based on reinforcement learning (RL-TPTA) is proposed. This algorithm first uses K-means clustering to spatially cluster search and rescue task points, assigning neighboring tasks to the same UAV to reduce initial flight distance. Secondly, it generates an initial access order based on the Euclidean distance between the task point and the UAV's starting point. Finally, a semi-random greedy strategy with generational decay is introduced, dynamically adjusting the task execution order through reinforcement learning to achieve a balance between exploring new task sequences and utilizing existing optimal paths. To address the high collision risk and inefficient convergence of path planning in complex 3D terrain, an efficient artificial bee colony algorithm based on multi-strategy fusion is proposed. By introducing a heuristic initial population generation strategy, a swarm aggregation strategy, and a Lévy flight-guided scout bee strategy, the algorithm's exploration capabilities and robustness in highly complex environments are significantly enhanced.

[0129] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0130] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxesFigure 1 A device that provides the functions specified in one or more boxes.

[0131] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0132] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0133] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A two-stage joint optimization method for multi-UAV task allocation and path planning, characterized in that, include: The K-means clustering algorithm is used to initially assign search and rescue tasks, ensuring that task points are reasonably allocated to each UAV according to spatial distribution; An initial access order is generated based on the task location, and a preliminary task execution sequence is planned for each drone; the task access order is optimized through reinforcement learning algorithms. Based on the optimized task access order, and combined with the artificial bee colony algorithm using a linear interpolation perturbation strategy, a bee colony aggregation strategy is introduced to perform more effective global search and local exploration in the solution space. Based on the global exploration mechanism of Levy flight, the task flight path scheme is determined.

2. The two-stage joint optimization method for multi-UAV task allocation and path planning according to claim 1, characterized in that, The initial task allocation model is based on the K-means clustering algorithm, representing all search and rescue task points as a set of coordinate points P = {p1, p2, ..., p...} in three-dimensional space. n }, where p i =(x i ,y i ,z i ) represents the coordinates of the i-th task point; the search and rescue task point data are clustered into K clusters according to the number of drones, and the task points in each cluster are assigned to the corresponding drones.

3. The two-stage joint optimization method for multi-UAV task allocation and path planning according to claim 1, characterized in that, After assigning specific tasks to each drone, an initial access order is generated according to the principle of proximity to distance, including: Let the drone's flight starting point be S = (x s ,y s ,z s The set of task points is P = {p1, p2, ..., p...} n }; where p i =(x i ,y i ,z i ) represents the coordinates of the i-th task point; for each task point p i Calculate the Euclidean distance d between it and the UAV's flight starting point S. i The formula is All task points are arranged according to distance d i Sort the tasks from smallest to largest to obtain an ordered sequence of task points P' = {p'1, p'2, ..., p'}. n }, where p'1 is the task point closest to the starting point, p' n The task point is the one furthest from the starting point. Finally, based on the sorted task point sequence P', the initial visiting order of the drone is generated, that is, the drone first visits the task point p'1 closest to the starting point, and then visits p'2, p'3, ..., p'' in sequence. n .

4. The two-stage joint optimization method for multi-UAV task allocation and path planning according to claim 1, characterized in that, The global exploration mechanism of Lévy flight includes: utilizing the heavy-tailed characteristic of the Lévy distribution to allow for long-step jumps, guiding scout bees out of local optima. Specifically, the widely used Mantegna algorithm is used to approximate the Lévy step size s, which is calculated by combining two independent standard normally distributed random variables u and v. The Lévy step size approximately follows a Lévy distribution. In the scout bee phase of the artificial bee colony algorithm, the Lévy step size is used to reset the solution to the number of times the scout bee has been trapped in a local optimum, thus enhancing global exploration. If a new solution exceeds the solution space boundary, reflection processing is performed.

5. The two-stage joint optimization method for multi-UAV task allocation and path planning according to claim 1, characterized in that, The swarming strategy includes: dynamically controlling the switching frequency of bees between collective cooperative search and random wandering by adjusting the constant FP, effectively balancing the exploration and development process, and improving the algorithm's ability to escape local optima and the diversity of solutions; in each iteration, the bees combine the search strategy guided by the global optimum with the behavioral strategy inspired by swarming to generate dual candidate solutions in parallel; by directly comparing the fitness values ​​of the two sets of candidate solutions, the better individuals are selected and retained to enter the next generation of the population.

6. The two-stage joint optimization method for multi-UAV task allocation and path planning according to claim 1, characterized in that, The linear interpolation perturbation strategy involves generating intermediate points on the line connecting the start and end points, and then applying a small-range random perturbation to these intermediate points. Specifically, this includes: firstly, uniformly sampling N intermediate points on the line connecting the start point S and the end point T. A baseline path P0 is constructed, and the coordinates of its intermediate points are calculated using linear interpolation. Subsequently, a three-dimensional random perturbation is applied to each baseline point, and an offset δ is generated by sampling in a spherical coordinate system. i Generate perturbed path points To dynamically balance the quality and diversity of the initial solution, the strategy introduces a linear adaptive decay mechanism for the perturbation radius, which means that the perturbation radius gradually decreases as the number of iterations increases.

7. A two-stage joint optimization system for multi-UAV task allocation and path planning, characterized in that, include: The task allocation module is used to initially allocate search and rescue tasks using the K-means clustering algorithm, ensuring that task points are reasonably allocated to each UAV according to spatial distribution. An initial access order is generated based on the task location, and a preliminary task execution sequence is planned for each drone; the task access order is optimized through reinforcement learning algorithms. The path planning module is used to determine the mission flight path scheme based on the optimized mission access order, combined with the artificial bee colony algorithm with a straight line interpolation perturbation strategy, and the bee colony aggregation strategy. It also incorporates a bee colony aggregation strategy to perform more effective global search and local development in the solution space, based on the global exploration mechanism of Levy flight.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the two-stage joint optimization method for multi-UAV task allocation and path planning as described in any one of claims 1-6.

9. A computer device, characterized in that, The system includes a memory and a processor, which communicate with each other. The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the two-stage joint optimization method for multi-UAV task allocation and path planning as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions that implement the two-stage joint optimization method for multi-UAV task allocation and path planning as described in any one of claims 1-6.

Images