Multi-unmanned aerial vehicle search and rescue task planning method based on ABC-DQN hierarchical optimization

Abstract

The invention discloses a multi-unmanned aerial vehicle search and rescue task planning method based on ABC-DQN hierarchical optimization, and belongs to the technical field of multi-unmanned aerial vehicle cooperative search and rescue. According to the method, task planning is carried out on a search and rescue area applying multiple unmanned aerial vehicles based on an artificial bee colony algorithm and a deep Q network, and adaptive cooperative search and rescue of the multiple unmanned aerial vehicles is completed. According to the method, global planning of a search and rescue task and path planning of a local area are both constrained and optimized, scene analysis of a search and rescue area and selection of the number of unmanned aerial vehicles with different performances are firstly carried out, and the unmanned aerial vehicle of a central control node is determined; and then the central unmanned aerial vehicle completes static global region division by using an ABC algorithm, then unmanned aerial vehicles in a local region perform adaptive search and rescue path planning based on a DQN algorithm, and in a local search and rescue process, the local unmanned aerial vehicles and the central unmanned aerial vehicle perform information interaction. Once the target is determined to appear, the central unmanned aerial vehicle further reduces the global area and performs ABC algorithm division again until the search and rescue target is determined to be within the required range.

Description

Technical Field

[0001] The present invention belongs to the technical field of multi-UAV collaborative search and rescue, and specifically relates to a multi-UAV search and rescue mission planning method based on ABC-DQN hierarchical optimization. Background Art

[0002] Multi-UAV collaborative search and rescue can be broadly divided into two parts: global planning and local planning of the search and rescue area. For search and rescue missions in complex terrain, the multi-UAV collaborative search and rescue process primarily involves the following steps. First, a preliminary understanding of the search and rescue area is required, such as the search target and scope, terrain characteristics, and the difficulty and distribution of the search and rescue area. This information can then be used to select drones with appropriate performance and the configuration of the central drone. Finally, each drone is numbered, and its fixed parameters are known in advance and stored in the central drone's data pool. This allows the central drone to group drones into clusters and adapt them to local areas based on the search and rescue area. To improve search and rescue efficiency, an artificial bee colony algorithm (ABC) can be used to plan search and rescue missions. The ABC algorithm is an optimization algorithm based on swarm intelligence that simulates the foraging behavior of bees to solve optimization problems. It offers significant advantages in multi-UAV collaborative search and rescue missions. This method utilizes three stages of collaboration: employed bees, observer bees, and scout bees, enabling a comprehensive global search within the solution space, avoiding local optima. In drone search and rescue missions, the ABC algorithm can effectively divide the global search and rescue area, ensuring comprehensive mission coverage. Furthermore, it has low requirements for the continuity and differentiability of the objective function, making it suitable for complex nonlinear optimization problems. In multi-drone collaborative search and rescue, the ABC algorithm can adapt to complex constraints such as varying terrain difficulty and drone performance. Furthermore, the algorithm has few parameters and is easy to implement, requiring only a few parameters such as the population size and maximum number of iterations, making it easy to implement and adjust.

[0003] However, the ABC algorithm is less efficient in local fine search, which may lead to slower convergence speed. Its shortcomings such as weak local search ability and slow convergence speed limit its application in real-time and high-dimensional problems, resulting in the algorithm not being very effective in search and rescue tasks with high real-time requirements. Therefore, in complex multi-UAV collaborative search and rescue scenarios, the ABC algorithm may need to be combined with other dimensionality reduction or divide-and-conquer strategies. Therefore, selecting the DQN algorithm to optimize the path planning of local areas and adaptively completing the search and rescue planning of local areas will help to make up for the shortcomings of the ABC algorithm, and when the ABC algorithm has already been allocated, the computational complexity of the DQN algorithm will also be reduced. Accordingly, the present invention proposes a solution for multi-UAV collaborative search and rescue tasks that combines the ABC algorithm with the DQN algorithm. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-UAV search and rescue mission planning method based on ABC-DQN hierarchical optimization, which can comprehensively improve the real-time collaborative search and rescue combat capability of multi-UAVs mainly in three aspects: global planning, local planning and path optimization.

[0005] Technical solution: A multi-UAV search and rescue mission planning method based on ABC-DQN hierarchical optimization includes the following steps:

[0006] S1. Analyze the terrain characteristics and data targets of the target area based on the requirements of the search and rescue mission, determine the number of drones required to participate in the search and rescue, assign different roles to each drone to handle different tasks, determine the deployment location of the central drone and input preset data in advance to ensure that the fixed configuration information of each drone is understood by the central drone;

[0007] S2. The central drone uses the ABC algorithm to perform static global area division. After the division, the task load of drones in different clusters in the search and rescue area is balanced. Then, drones in the local area perform adaptive search and rescue path planning based on the DQN algorithm. During the local search and rescue process, the local drones and the central drone will maintain information exchange. When the search and rescue target appears, the central drone will treat the local area as the global area and make global planning decisions again through the ABC algorithm. This iteration is repeated until the search and rescue target appears within the specified range.

[0008] S3: Adopt the DQN algorithm to perform adaptive search and rescue path planning for the local area. During this process, the drone dynamically adjusts its heading based on its own endurance, current location, unsearched areas, and nearby obstacles to optimize the search and rescue path.

[0009] This process establishes a dynamic feedback mechanism for the mission execution process. The drones communicate regularly, and the central drone will dispatch drones based on the real-time situation to ensure the continuous progress of the search and rescue mission. After the search and rescue target is determined, the central drone in the area notifies each drone to stop working. At this time, each drone will conduct a final communication to inform its own performance status for the next global planning and task division.

[0010] In the above scheme, the static global region partition modeling process using the ABC algorithm includes constructing a fitness function and overlapping penalty to optimize the global region partition and task allocation of multi-UAV collaborative search and rescue. The fitness function is used to quantify the advantages and disadvantages of each region partition scheme to achieve a match between terrain difficulty and UAV performance. The overlapping penalty is used to achieve the independence of different local regions.

[0011] The mathematical expression of the fitness function is:

[0012]

[0013] where λ ij represents the terrain difficulty coefficient of grid cell (i, j), |Bk| is the total number of grid cells in the kth local area;

[0014] The mathematical expression of overlap penalty is:

[0015]

[0016] The numerator |Bk∩Bl| represents the overlapping area of regions k and l, the denominator |Bk∪Bl| represents the union area of the two regions, and the penalty coefficient η is used to adjust the impact of overlap on the fitness value. The function overlap(x) aims to calculate the ratio of the intersection to the union of all different subset pairs, thereby measuring the degree of element overlap. X represents all solutions, corresponding to the region partitioning scheme.

[0017] Furthermore, the specific process of using the ABC algorithm to plan tasks in the global area is as follows:

[0018] S21. Construct a scene model of the search and rescue area, divide the search and rescue area into an M×N grid map, and define each unit with a two-dimensional coordinate system (i, j). The terrain difficulty coefficient is λ ij ∈[0,1], 0 means simple, 1 means complex;

[0019] S22, based on the UAV battery life T m , detection radius R m , climbing speed v m Considering the above, we set the weights α related to the three i , the performance of the UAV is quantified as:

[0020] P m =α1T m +α2R m +α3v m

[0021]

[0022] S23. Determine the constraints for the first global planning, including:

[0023] Hard constraints: The area is divided based on the number of drones and the search and rescue range. The area of a single area does not exceed the total coverage of the drones operating in the area.

[0024] Soft constraint: Optimizing the matching between regional difficulty and drone performance. The calculation is defined as maximizing the sum of the products of regional difficulty and drone performance in the search and rescue scenario. This process is solved by the ABC algorithm.

[0025] Furthermore, the calculation process of the artificial bee colony algorithm includes:

[0026] (11) Food source coding stage

[0027] The food source in the ABC algorithm is represented by X, and each solution X represents the area division and drone allocation scheme:

[0028] X={(B1,D1),…,(B K ,D K )}

[0029] The region boundary B k Represents the Kth local area:

[0030] B k =[X k1 ,Y k1 ,X k2 ,Y k2 ]

[0031] Drone Allocation D k Indicates the number of low, medium, and high performance drones that should be allocated to the K-th local area:

[0032]

[0033] Hard constraints:

[0034] The global area must be divided into no fewer than 4 local areas, and the number of drones in each local area must be no fewer than 3;

[0035] For the above process, the fitness function and overlap penalty are set to optimize the global area division and task allocation of multi-UAV collaborative search and rescue;

[0036] (12) Employed bee stage

[0037] Through local adjustments, we gradually optimize the terrain difficulty and drone performance matching of each sub-area and improve the fitness value. Specifically:

[0038] First randomly select a local area B k , adjust its boundaries, where ΔB is the maximum adjustment step, U(-1,1) is a uniformly distributed random number, and then the drone is redistributed, randomly selecting two areas k, l to exchange drones, where δ∈{1,2}, and must satisfy That is, the number of the adjusted drones of this type can at least meet the adjusted quantity requirements;

[0039] (13) Observation Bee Stage

[0040] Probabilistic selection and depth search are performed based on the fitness value of the solution, where the fitness value of each solution is used to calculate the probability of selection P(X i )for:

[0041]

[0042] Where f(X i ) is the fitness value of the solution Xi, SN is the total number of solutions, It is the sum of the fitness values of all solutions. This formula indicates that the probability of each solution being selected is proportional to its fitness value. The higher the fitness value, the greater the probability of the solution being selected, thus guiding the algorithm to search for a better solution.

[0043] For the selected solution X i Perform neighborhood search and generate a new solution X new , the formula is:

[0044] X new =X best +ω·(X best -X worst )

[0045] where X new is the generated new solution, X best is the current optimal solution, X worst is the current worst solution, ω is the learning rate, and its value range is [0.5, 0.8];

[0046] (14) Scouting Bee Stage

[0047] This stage solves the reset stagnant solution, if the solution is X i If there is no improvement after L consecutive iterations, regenerate:

[0048]

[0049] The optimal solution of the output results includes: region partition matrix {Bk}, drone allocation table {Dk}, and global fitness value f(X*);

[0050] According to the drone allocation table {Dk}, each drone will be dynamically organized into multiple collaborative clusters and fly to the preset local search and rescue area. Each drone cluster uses the DQN network algorithm to realize intelligent collaborative search and rescue planning in the target area.

[0051] In the above scheme, step S3 is for the UAV to use the DQN algorithm to adaptively plan the search and rescue path of the mission area. The DQN decision target is the search and rescue area coverage rate, the state is its own endurance, current position and obstacles, and the action is the next flight direction. The purpose is to optimize the path planning and thus achieve the purpose of reducing redundant calculations. The specific decision design is as follows:

[0052] Drone state vector: Among them E t Indicates the remaining battery life. represents the coordinate system of the t-th UAV in the current K-th local area, O t represents the obstacle distribution mask matrix, C t Indicates the heat map of the covered area;

[0053] The discrete action set at∈{N,S,E,W,NE,NW,SE,SW} represents the eight directions of east, south, west, north, southeast, southwest, northeast, and northwest;

[0054] Reward function:

[0055] ΔC represents the new coverage rate, d obstacle represents the distance to the nearest obstacle, OverlapPenalty represents the path duplication penalty term, α, β, γ, η are all weight coefficients;

[0056] DQN network update: Q(s t ,a t )←Q(s t ,a t )+α[r t +γmax a' Q(s t +1,a')-Q(s t ,a t )], where: Q(s t ,a t ) represents the state-action value function, which is in state s t Next take action a t The expected cumulative reward, α represents the learning rate, which controls the influence of new information on the current Q value at each update, r t is to perform action a at time step t t The immediate reward obtained after γ is the discount factor used to measure the current value of future rewards; γmax a' Q(s t +1,a') is in the next state s t The maximum Q value among all possible actions a' under +1 represents the expected cumulative reward of taking the best action in the next state, Q(s t+1,a') is in the next state s t +1 for the Q-value of taking action a';

[0057] The loss function is used to measure the difference between the current Q value and the target Q value. The parameters of the Q network are updated by minimizing the loss function. Together, they constitute the update rule in Q learning, which is used to gradually optimize the agent's strategy to maximize the cumulative reward. Its mathematical representation is as follows:

[0058] L=E[(r+γmax a” Q target (s',a')-Q online (s,a)) 2 ]

[0059] where Q target (s',a') is the Q value of the target network, Q online (s,a) is the Q value of the online network, which represents the Q value estimate under the current strategy.

[0060] Furthermore, each UAV maintains periodic data interaction with the central control node within the time step Δt, and the transmitted data packets adopt a simplified triple structured format:

[0061] The remaining battery status of the drone is expressed as a percentage, T m ∈[0%,100%];

[0062] The current position coordinates of the drone are expressed in the local area coordinate system, and the Kth partition is expressed as

[0063] The target detection state is encoded using Boolean values and is represented as B k ∈{0,1}, where 1 indicates the target exists and 0 indicates the target does not exist;

[0064] The communication mechanism composed of this triplet minimizes the amount of data transmission and optimizes the communication bandwidth utilization while ensuring task reliability.

[0065] Furthermore, when the local area is within the preset time threshold T threshold If no effective search and rescue results are obtained within 30 days, the central control node will activate the dynamic task reallocation mechanism, which will give priority to dispatching drones that have completed their assigned tasks to the target area for collaborative search.

[0066] When a search and rescue target is detected, the central control node will broadcast a termination command to all drones and, based on the last state exchange data, use the artificial bee colony optimization algorithm to perform global task replanning. The only thing that needs to be updated is the drone parameter performance quantification formula: m =α1Tm +α2R m +α3v m , is the weight coefficient, where T m The corresponding value is the battery life and detection radius R. m With climbing speed v m is a fixed parameter, which remains unchanged, and then α i Adjust the weights.

[0067] Compared with the prior art, the present invention has the following essential features and significant improvements:

[0068] (1) The proposed method achieves a synergistic integration of global and local optimization. Through a hierarchical architecture system, the artificial bee colony (ABC) algorithm is used to perform global region division and task load balancing, while the deep Q network (DQN) is used to achieve dynamic optimization of local paths. The ABC algorithm ensures comprehensive global task coverage and balanced resource allocation, while the DQN algorithm reduces local redundant searches and obstacle collision risks through adaptive path planning. The two complement each other and significantly improve search and rescue efficiency and real-time performance.

[0069] (2) The method described in this invention improves the ability to dynamically adapt and adjust the path in real time. In a local area, the drone uses the DQN algorithm to perceive the environment (such as endurance, obstacle distribution, and covered area) in real time, dynamically adjusts the flight path and speed, and optimizes coverage and safety through a reward function. This mechanism significantly reduces the repeated searches and computational redundancy in traditional fixed path planning, improving the flexibility and success rate of search and rescue in complex terrain.

[0070] (3) The method described in this invention enhances iterative optimization and efficient communication mechanisms: the central UAV and each cluster maintain periodic communication through streamlined structured data, integrating search and rescue status (such as battery level, location, and target detection) in real time, and iteratively calling the ABC algorithm based on feedback to re-divide the global area. This mechanism enables the system to quickly respond to target discovery or environmental changes, dynamically narrowing the potential search and rescue range, avoiding the waste of resources caused by global repeated planning, and significantly shortening the overall search and rescue time. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is a flow chart of the method of the present invention;

[0072] Figure 2 This is a schematic diagram of the allocation of local areas after the global area is divided based on the ABC algorithm of the present invention;

[0073] Figure 3 This is a schematic diagram of the allocation of drone clusters after global area division based on the ABC algorithm of the present invention;

[0074] Figure 4 Schematic diagram of the multi-UAV search and rescue system based on ABC-DQN hierarchical optimization of the present invention. DETAILED DESCRIPTION

[0075] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0076] Combine Figure 1-4 The present invention provides a multi-UAV search and rescue mission planning method based on ABC-DQN hierarchical optimization, comprising the following steps:

[0077] Step S1, use satellite remote sensing images, geographic information system (GIS) data or drone aerial images to obtain detailed terrain information of the search and rescue area. Divide the area into areas of different difficulty levels based on terrain complexity, obstacle density, and target distribution characteristics (such as hot spots where targets may exist). Then select drones that are suitable in performance and reasonable in number. Determine the central drone based on the terrain and number of drones in the search and rescue area. The configuration of the central drone must meet the requirements of high performance and high endurance to ensure that it can continuously communicate with each drone. The specific selection criteria are as follows:

[0078] Prioritize performance and get P m The highest and longest-lasting drone (Tm ≥ 90% of similar drones) is selected as the central node. The deployment location is preferably the geometric center of the search and rescue area to ensure balanced communication distance. If the terrain is complex, the deployment location is optimized using the signal strength model.

[0079] Each drone's fixed parameters (battery life, detection radius, and speed) and weight coefficients are pre-entered into the central drone database. A drone performance table and a search and rescue area difficulty classification table are also established to provide accurate data support for subsequent ABC algorithm calculations.

[0080] Step S2: Use the ABC algorithm to plan tasks in the global area:

[0081] Based on the pre-set data from step S1, the task load is balanced by considering the number of local regions and the number of drones allocated to each local region. Preferably, there are at least four local regions, and at least three drones per region. These constraints can be adjusted dynamically as the global region shrinks.

[0082] Step S201: problem modeling and initial global planning based on the ABC algorithm:

[0083] During the initial global planning process, the search and rescue area's terrain complexity and the fixed parameters of the drones have already been provided to the central drone in step S1. The number of drones assigned to each local area must not be less than three. Furthermore, the global area must be divided into at least four independent sub-areas to ensure balanced mission coverage and resource allocation. The following is the problem modeling for the first use of the ABC algorithm:

[0084] (1) Search and rescue scenario model

[0085] The search and rescue area is divided into an M×N grid map. Each unit is defined by a two-dimensional coordinate system (i, j). The terrain difficulty coefficient is defined according to the data in step S1 and is expressed as follows:

[0086] λ ij ∈[0, 1] (0 = simple, 1 = complex)

[0087] In order to better adapt the local area to the UAV load, the area is also divided and constrained, that is, the area of a single sub-area ≤ A max ;

[0088] The preset partitioning constraints are as follows:

[0089] The area of a single sub-region ≤ A max ,

[0090] Difficulty difference between adjacent areas ≤ Δλ threshold .

[0091] (2) The drones are divided into different performance categories based on their fixed parameters (battery life, detection radius, and navigation speed) and weight coefficients. This step is to simplify the algorithm parameters and improve the calculation speed. The three parameters are converted into one parameter, which is expressed as:

[0092] P m =α1T m +α2R m +α3v m

[0093] Weight coefficient:

[0094] According to the performance of drones, drones are divided into three categories: low, medium and high performance.

[0095] (3) Each time the ABC algorithm is performed, the premise constraints are dynamically adjusted. During the first global planning, the following constraints are set:

[0096] Hard constraints: arbitrary region k, N k ≥3, where N k Indicates the number of drones in a local area;

[0097] k≥4, K represents the number of local regions;

[0098] Soft constraint: Maximize ∑(regional difficulty × drone performance).

[0099] (4) Computational solution based on artificial bee colony algorithm.

[0100] 1) Food source encoding stage: First, the artificial bee colony algorithm determines the possible food sources. Here, the food sources represent possible area divisions and drone allocation plans, represented by X:

[0101] X={(B1,D1),…,(B K ,D K )}

[0102] The Kth local area is B k Indicates that B k =[X k1 ,Y k1 ,X k2 ,Y k2 ];

[0103] The number of low-, medium-, and high-performance drones that should be allocated in the Kth local area is D k Indicates that the expression is:

[0104] Hard constraints: That is, the global area cannot be divided into less than 4 local areas, and the number of drones in each local area cannot be less than 3.

[0105] 2) Set the fitness function and overlap penalty.

[0106] In the ABC algorithm, a fitness function and overlap penalty are set to optimize the global region division and task allocation of multi-UAV collaborative search and rescue. The fitness function is used to quantify the pros and cons of each region division scheme. Its core goal is to achieve a match between terrain difficulty and UAV performance. The specific expression is:

[0107]

[0108] where λ ij represents the terrain difficulty coefficient of grid cell (i, j) (0 is simple, 1 is complex), |Bk| is the total number of grid cells in the kth local area. The higher the fitness value, the more suitable the comprehensive terrain difficulty of the area is for the assigned UAV performance. For example, in a complex terrain area (λ k High performance UAVs should be assigned to the simple areas (λ k Lower) can allocate low-performance drones to maximize resource utilization.

[0109] In addition, the overlap penalty is used to emphasize the independence of local areas. If the areas overlap, multiple drone swarms may repeatedly search the same area, wasting time and energy. Non-overlapping area division can ensure balanced task distribution and avoid overloading some areas due to multiple coverage. The expression is:

[0110]

[0111] The numerator |Bk∩Bl| represents the overlapping area of regions k and l, the denominator |Bk∪Bl| represents the union area of the two regions, and the penalty coefficient η (recommended value 0.2) is used to adjust the impact of overlap on the fitness value.

[0112] 3) Local adjustments and drone redistribution are carried out during the hired bee stage.

[0113] The matching degree between the terrain difficulty and the UAV performance of each sub-area is optimized iteratively to maximize the soft constraint target ∑(region difficulty × UAV performance). The specific implementation steps are as follows:

[0114] Randomly select a region B from the current set of divided local regions k , adjust its boundaries, where ΔB is the maximum adjustment step (for example, ΔB = 2 grid cells). Then generate a uniformly distributed random number U(-1,1) and calculate the boundary offset: ΔX = ΔB·U(-1,1), ΔY = ΔB·U(-1,1), and set B k The boundary coordinates (X k1 ,Y k1 ) and (X k2 ,Y k2 ) are offset by (ΔX, ΔY) to generate a new boundary B k '. Ensure that the area of the adjusted region is ≤ Amax (preset maximum value), and if the new boundary causes region overlap or violates the area constraint, abandon the adjustment and regenerate the offset. Then redistribute the drones and randomly select two regions k, l to exchange drones, where δ∈{1,2}, and must satisfy That is, the number of drones of the adjusted type at least meets the adjustment requirements.

[0115] 4) The observation bee stage is the key link for probabilistic selection and deep search based on the fitness value of the solution. Its core goal is to further improve the global search efficiency by focusing on optimizing high-quality solutions.

[0116] Calculate the selection probability P(X i ), the formula is:

[0117]

[0118] Where f(Xi ) is the solution to X i The fitness value, SN is the total number of solutions, is the sum of the fitness values of all solutions. This formula indicates that the probability of each solution being selected is proportional to its fitness value. Solutions with higher fitness values are more likely to be selected, thus guiding the algorithm to search for more optimal solutions.

[0119] For the selected solution X i Perform neighborhood search and generate a new solution X new , the formula is:

[0120] X new =X best +ω·(X best -X worst )

[0121] where X new is the generated new solution, X best is the current optimal solution, X worst is the current worst solution, ω is the learning rate, and its value range is [0.5, 0.8].

[0122] (5) The scout bee stage solves the reset stagnation solution.

[0123] If the solution is X i If there is no improvement after L consecutive iterations, a new solution is generated. Specifically, when a solution stagnates, we will randomly re-divide the area (satisfying the minimum number of sub-areas) and evenly distribute drones in each sub-area to ensure that there are at least 3 drones in each sub-area. The expression is as follows:

[0124]

[0125] (6) The optimal solution for the output results includes: the region partition matrix {Bk}, the drone allocation table {Dk}, and the global fitness value f(X*). Assume that the search and rescue area is divided into an M×N=6×6 grid map and divided into 4 sub-regions (satisfying the hard constraint K≥4). The resulting region partition matrix {Bk} can be as shown in the figure, and the drone allocation table {Dk} can be as shown in the figure. Through performance-terrain coupled optimization and constraint-driven search, efficient search and rescue in complex scenarios is ensured.

[0126] Step S202: Based on the division results, the drones go to the assigned local area and communicate with the central drone regularly:

[0127] Specifically, based on the drone allocation table {Dk}, drones with matching performance (e.g., low, medium, and high performance) are dynamically divided into multiple collaborative clusters. Each cluster flies to the target location according to pre-defined subregion boundaries (defined by the region partition matrix {Bk}). Within the target region, each drone cluster uses the Deep Q-Network (DQN) algorithm to implement intelligent collaborative search and rescue planning. Specifically, the DQN algorithm is used to adaptively partition the regions within the cluster, allowing the search and rescue mission to be executed in parallel. Within each time step Δt, the drone cluster maintains periodic communication with the central control node to update the mission status in real time and optimize the global path planning strategy. This mechanism effectively ensures efficient execution of the search and rescue mission and adaptability to dynamic environments.

[0128] Step S203: Based on the search and rescue results, the ABC algorithm is used for global planning again until the search and rescue target is found.

[0129] Specifically, after determining the next global area, the central drone will comprehensively consider the endurance and performance parameters of the currently available drones, and appropriately adjust the environmental modeling and drone performance weights and constraints. The first thing is to appropriately adjust the weights of the three drone parameters (endurance, detection radius, and navigation speed). Because at the beginning, the battery level is 100%, now the performance of the drone is re-divided according to the level of battery so that the allocation plan can be more accurate. Then the artificial bee colony (ABC) algorithm is used again for global path planning to further narrow the potential area of the search and rescue target. This process will continue to iterate until the search and rescue target is detected in the next local area. If the target does not appear, the system will repeat step S203 until the search and rescue target is finally located and confirmed.

[0130] Step S3: Adopt the DQN algorithm to perform adaptive search and rescue path planning for the local area:

[0131] Step S301: The drones in the local area are divided into adaptive regions based on the DQN algorithm and perform search and rescue tasks in parallel:

[0132] First, set the DQN's decision objective: search and rescue area coverage. Next, set the state to be considered in this decision. To avoid redundant computation, this state only considers the drone's flight time, current location, obstacles, and the search and rescue level in the area. Finally, the action to be determined is the next flight direction. This reduces redundant computation by adaptively optimizing the drone's path planning.

[0133] State Space Design (State S): UAV state vector

[0134] Among them E t Indicates the remaining battery life. represents the coordinate system of the t-th UAV in the current K-th local area, O t represents the obstacle distribution mask matrix (0 / 1 indicates the existence of obstacles), C t Indicates the heat map of the covered area (cumulative coverage).

[0135] Action Space Design (ActionA): To reduce the computational dimensionality of decision-making and better match the drone's computing capabilities, actions here only consider discrete directions, without making directional decisions. The discrete action set is represented as at∈{N,S,E,W,NE,NW,SE,SW}, representing the eight directions: east, south, west, north, southeast, southwest, northeast, and northwest.

[0136] Reward function design (RewardR): The reward function is used to quantify the immediate feedback of the drone after each action, and guide the drone's behavior through multi-objective trade-offs. The specific formula is:

[0137]

[0138] Where ΔC represents the new coverage rate, d obstacle represents the distance to the nearest obstacle, OverlapPenalty represents the path duplication penalty, and α, β, γ, and η are weight coefficients.

[0139] Where α·ΔC is the coverage rate reward, which encourages drones to explore uncovered areas first. When the proportion of α is high, it means that the coverage rate dominates the reward, and vice versa. To reward obstacle avoidance, the drone is encouraged to stay away from obstacles by counting the distance inversely. β and α have the same function (the subsequent γ and η have the same function and will not be repeated here). The η·OverlapPenalty path penalty is used to suppress redundant paths and reduce ineffective searches.

[0140] The present invention uses the Q network to learn and update the strategy through temporal difference (TD), incorporates long-term benefits into current decisions, and gradually optimizes the strategy. The expression is as follows:

[0141] Q(s t ,a t )←Q(s t ,a t )+α[r t +γmax a' Q(s t +1,a')-Q(s t ,a t )]

[0142] Where Q(st ,a t ) means in state s t Next take action a t The expected cumulative reward. α represents the learning rate, which controls the impact of new information on the Q value (such as α = 0.01 means slow update). t Indicates that action a is executed at time step t t γ represents the discount factor, which is used to measure the current value of future rewards (e.g. γ = 0.9 means emphasizing long-term benefits). max a' Q(s t +1,a') means in the next state s t +1, the maximum Q value among all possible actions a' represents the optimal future benefit.

[0143] The loss function is used to train the Q network and is as follows:

[0144] L=E[(r+γmax a” Q target (s',a')-Q online (s,a)) 2 ],

[0145] where Q target (s',a') This is the Q value of the target network, Q online (s, a) is the Q-value of the online network, representing the estimated Q-value under the current policy. E stands for mean squared error (MSE). The entire formula aims to minimize the difference between the target Q-value and the current Q-value, thereby improving prediction accuracy.

[0146] Therefore, through DQN decision-making, the drone ultimately accomplishes one thing: when the drone approaches an obstacle, the obstacle avoidance reward will significantly reduce the Q value, prompting the network to choose a detour; at the same time, the coverage reward encourages it to explore adjacent uncovered areas, allowing the drone to learn to prioritize high-reward actions in complex environments, and ultimately find a balance between safety and efficiency.

[0147] In step S302, each UAV maintains periodic data interaction with the central control node within a time step Δt.

[0148] The transmitted data packet adopts a simplified triplet structure format: First, the remaining battery status of the drone, quantified in percentage form (T m ∈[0%,100%]); Second, the current position coordinates of the drone are expressed in the local area coordinate system ( Where K represents the Kth partition); third, the target detection state is encoded using Boolean values (B k∈{0,1}, where 1 indicates the target exists and 0 indicates the target does not exist). This communication mechanism optimizes the communication bandwidth utilization while ensuring task reliability by minimizing the amount of data transmitted.

[0149] In step S303, the central control node dynamically adjusts the drone allocation mechanism to keep the mission moving forward.

[0150] The central node receives real-time data from all clusters, updates the global coverage heat map and obstacle map, and performs dynamic replanning: if a target is found in a certain area, the ABC algorithm is triggered to re-divide the global area and narrow the search range; if a cluster is low on power (for example, Tm < 20%), the adjacent idle drone is dispatched to take over the task. When a drone detects a target, it immediately reports to the central node, triggering a termination command. The central node broadcasts a termination signal, and all drones stop searching and return. After the mission is completed, each drone uploads the final status data (such as remaining power, coverage area) for evaluating mission efficiency. The central node records the mission log and optimizes the weight parameters of the next ABC-DQN collaborative strategy to improve the execution effect of future tasks. The only thing that needs to be updated is the drone parameter performance quantification formula:

[0151] P m =α1T m +α2R m +α3v m

[0152] is the weight coefficient

[0153] Where T m The battery life and detection radius R reported by the last communication m With climbing speed v m is a fixed parameter and α i Adjust the proportions appropriately.

[0154] The present invention combines the ABC algorithm with the DQN algorithm in a hierarchical manner, processing global planning and local path optimization respectively, and realizing iterative optimization through a dynamic interaction mechanism. The global division of the ABC algorithm optimizes the computational load of the DQN algorithm. At the same time, the adaptive search and information feedback of the DQN algorithm give the ABC algorithm dynamics, making it more adaptable to dynamic environments and reducing the computational load of global planning to a certain extent. In addition, ABC is combined with DQN to make up for their respective shortcomings (ABC has global coverage but low local efficiency, and DQN is locally adaptive but has weak global planning). The above two algorithms complement each other and are optimized in layers, thereby reducing the computational dimension of each other and improving search efficiency.

Claims

1. A multi-UAV search and rescue mission planning method based on ABC-DQN hierarchical optimization, characterized by: The following steps are involved: S1. Deploy drones based on the target area and search and rescue mission, including determining the deployment location of the central drone and the number of other drones, and assigning different roles to each drone to handle different tasks; S2. Use the ABC algorithm to plan the missions of drones in the global area, and balance the mission load of drones in different clusters within the search and rescue area; The local drones and the central drone will maintain information exchange. When the search and rescue target appears, the central drone will treat the local area as the global area and make global planning decisions again through the ABC algorithm. This iteration will be repeated until the search and rescue target appears within the specified range. S3: Adopt the DQN algorithm to perform adaptive search and rescue path planning for the local area. During this process, the drone dynamically adjusts its heading based on its own endurance, current location, unsearched areas, and nearby obstacles to optimize the search and rescue path. A dynamic feedback mechanism is established for the mission execution process. The central drone dispatches drones based on real-time conditions to ensure the continuous advancement of the search and rescue mission. After the search and rescue target is determined, the central drone in the area notifies each drone to stop working. At this time, each drone will conduct a final communication to inform its own performance status for the next global planning and task division.

2. The multi-UAV search and rescue mission planning method according to claim 1, characterized in that: The static global region partitioning modeling process using the ABC algorithm includes constructing a fitness function and overlapping penalty to optimize the global region partitioning and task allocation of multi-UAV collaborative search and rescue. The fitness function is used to quantify the pros and cons of each region partitioning scheme, achieving a match between terrain difficulty and UAV performance, and overlapping penalty is used to achieve the independence of different local regions. The mathematical expression of the fitness function is: where λ ij represents the terrain difficulty coefficient of grid cell (i, j), |Bk| is the total number of grid cells in the kth local area; The mathematical expression of overlap penalty is: The numerator |Bk∩Bl| represents the overlapping area of regions k and l, the denominator |Bk∪Bl| represents the union area of the two regions, and the penalty coefficient η is used to adjust the impact of overlap on the fitness value.

3. The multi-UAV search and rescue mission planning method according to claim 1 or 2, characterized in that: The specific process of using the ABC algorithm to plan tasks in the global area is as follows: S21. Construct a scene model of the search and rescue area, divide the search and rescue area into an M×N grid map, and define each unit with a two-dimensional coordinate system (i, j). The terrain difficulty coefficient is λ ij ∈[0,1], 0 means simple, 1 means complex; S22, based on the UAV battery life T m , detection radius R m , climbing speed v m Considering the above, we set the weights α related to the three i , the performance of the UAV is quantified as: P m =α1T m +α2R m +α3v m S23. Determine the constraints for the first global planning, including: Hard constraints: The area is divided based on the number of drones and the search and rescue range. The area of a single area does not exceed the total coverage of the drones operating in the area. Soft constraint: Optimizing the matching between regional difficulty and drone performance. The calculation is defined as maximizing the sum of the products of regional difficulty and drone performance in the search and rescue scenario. This process is solved using the ABC algorithm.

4. The multi-UAV search and rescue mission planning method according to claim 3, characterized in that: The calculation process of the artificial bee colony algorithm includes: (11) Food source coding stage The food source in the ABC algorithm is represented by X, and each solution X represents the area division and drone allocation scheme: X={(B1,D1),…,(B K ,D K )} The region boundary B k Represents the Kth local area: B k =[X k1 ,Y k1 ,X k2 ,Y k2 ] Drone Allocation D k Indicates the number of low, medium, and high performance drones that should be allocated to the K-th local area: Hard constraints: The global area must be divided into no fewer than 4 local areas, and the number of drones in each local area must be no fewer than 3; For the above process, the fitness function and overlap penalty are set to optimize the global area division and task allocation of multi-UAV collaborative search and rescue; (12) Employed bee stage Through local adjustments, the terrain difficulty and drone performance matching of each sub-area are gradually optimized to improve the fitness value. Specifically: First randomly select a local area B k , adjust its boundaries, where ΔB is the maximum adjustment step, U(-1,1) is a uniformly distributed random number, and then the drone is redistributed, randomly selecting two areas k, l to exchange drones, where δ∈{1,2}, and must satisfy This constraint ensures that the number of drones of this type that are adjusted can at least meet the number requirements for processing tasks; (13) Observation Bee Stage Probabilistic selection and depth search are performed based on the fitness value of the solution, where the fitness value of each solution is used to calculate the probability of selection P(X i )for: Where f(X i ) is the fitness value of the solution Xi, SN is the total number of solutions, It is the sum of the fitness values of all solutions. This formula indicates that the probability of each solution being selected is proportional to its fitness value. The higher the fitness value, the greater the probability of the solution being selected, thus guiding the algorithm to search for a better solution. For the selected solution X i Perform neighborhood search and generate a new solution X new , the formula is: X new =X best +ω·(X best -X worst ) where X new is the generated new solution, X best is the current optimal solution, X worst is the current worst solution, ω is the learning rate, and its value range is [0.5, 0.8]; (14) Scouting Bee Stage This stage is used to solve the reset stagnant solution. If the solution X i If there is no improvement after L consecutive iterations, regenerate: The optimal solution of the output results includes: region partition matrix {Bk}, drone allocation table {Dk}, and global fitness value f(X*); According to the drone allocation table {Dk}, each drone will be dynamically organized into multiple collaborative clusters and fly to the preset local search and rescue area. Each drone cluster uses the DQN network algorithm to realize intelligent collaborative search and rescue planning in the target area.

5. The multi-UAV search and rescue mission planning method according to claim 1, characterized in that: In step S3, the UAV uses the DQN algorithm to adaptively plan the search and rescue path of the mission area. The DQN decision target is the search and rescue area coverage rate, the state is its own endurance, current position and obstacles, and the action is the next flight direction. This is to optimize the path planning and thus achieve the purpose of reducing redundant calculations. The specific decision design is as follows: Drone state vector: Among them E t Indicates the remaining battery life. represents the coordinate system of the t-th UAV in the current K-th local area, O t represents the obstacle distribution mask matrix, C t Indicates the heat map of the covered area; The discrete action set at∈{N,S,E,W,NE,NW,SE,SW} represents the eight directions of east, south, west, north, southeast, southwest, northeast, and northwest; Reward function: ΔC represents the new coverage rate, d obstacle Represents the distance to the nearest obstacle, OverlapPenalty represents the path duplication penalty term, and α, β, γ, and η are all weight coefficients; DQN network update: Q(s t ,a t )←Q(s t ,a t )+α[r t +γmax a' Q(s t +1,a')-Q(s t ,a t )], where: Q(s t ,a t ) represents the state-action value function, which is in state s t Next take action a t The expected cumulative reward, α represents the learning rate, which controls the influence of new information on the current Q value at each update, r t is to perform action a at time step t t The immediate reward obtained after γ is the discount factor used to measure the current value of future rewards; γmax a' Q(s t +1,a') is in the next state s t The maximum Q value among all possible actions a' under +1 represents the expected cumulative reward of taking the best action in the next state, Q(s t +1,a') is in the next state s t +1 state action value for taking action a'; The loss function is used to measure the difference between the current state action value and the target state action value. The parameters of the Q network are updated by minimizing the loss function. Together, they constitute the update rule in Q-learning, which is used to gradually optimize the agent's strategy to maximize the cumulative reward. Its mathematical representation is as follows: L=E[(r+γmax a' 'Q target (s',a')-Q online (s,a)) 2 ] where Q target (s',a') is the state action value of the target network, Q online (s,a) is the state-action value of the online network, which represents the state-action value estimate under the current strategy.

6. The multi-UAV search and rescue mission planning method according to claim 5, characterized in that: Each UAV maintains periodic data interaction with the central control node within the time step Δt, and the transmitted data packets adopt a simplified triple structured format: The remaining battery status of the drone is expressed as a percentage, T m ∈[0%,100%]; The current position coordinates of the drone are expressed in the local area coordinate system, and the Kth partition is expressed as The target detection state is encoded using Boolean values and is represented as B k ∈{0,1}, where 1 indicates the target exists and 0 indicates the target does not exist; The communication mechanism composed of this triplet minimizes the amount of data transmission and optimizes the communication bandwidth utilization while ensuring task reliability.

7. The multi-UAV search and rescue mission planning method according to claim 1 or 6, characterized in that: When the local area is within the preset time threshold T threshold If no effective search and rescue results are obtained within 30 days, the central control node will activate the dynamic task reallocation mechanism, which will give priority to dispatching drones that have completed their assigned tasks to the target area for collaborative search. When a search and rescue target is detected, the central control node will broadcast a termination command to all drones and, based on the last state exchange data, use the artificial bee colony optimization algorithm to perform global task replanning. The only thing that needs to be updated is the drone parameter performance quantification formula: m =α1T m +α2R m +α3v m , is the weight coefficient, where T m The corresponding value is the battery life and detection radius R. m With climbing speed v m is a fixed parameter, which remains unchanged, and then α i Adjust the weights.

Images