Multi-task allocation method for heterogeneous UAV clusters based on decomposition learning particle swarm

Through the method based on decomposition of learning particle swarm, the complexity of multi-objective and strong constraints in the task allocation of drone clusters is solved, efficient rescue task allocation is achieved, the utilization of drone resources is optimized, and the allocation quality and efficiency are improved. .

CN118466582BActive Publication Date: 2024-12-06CHINA UNIV OF MINING & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410539170.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-30
Publication Date
2024-12-06
Estimated Expiration
2044-04-30

AI Technical Summary

Technical Problem

In the allocation of drone cluster missions, existing algorithms are difficult to effectively deal with complex scenarios of multi-objective, multi-preference and strong constraints, especially in post-disaster rescue, which cannot fully utilize the characteristics and resources of different types of drones, resulting in task allocation Inefficient.

Method used

Using a heterogeneous drone cluster multi-task allocation method based on decomposition learning particle swarm, a multi-objective allocation model is established through the optimization goals of minimum rescue punishment and minimum resource occupation, and a particle swarm initialization strategy, a decomposition learning particle update strategy and internal Cross-mutation mechanism optimizes drone types and task execution order, and improves algorithm efficiency and diversity.

Benefits of technology

It has achieved efficient allocation of rescue tasks under multi-objective and strong constraints, taking into account the drone's endurance and capacity constraints, improving the quality and efficiency of task allocation, reducing the invalid calculation of the algorithm, and solving the repetition and partial failure of rescue tasks. Issue of execution.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118466582B_ABST
    Figure CN118466582B_ABST
Patent Text Reader

Abstract

The present invention discloses a multi-task allocation method for heterogeneous unmanned aerial vehicle clusters based on decomposition learning particle swarms, which belongs to the field of unmanned aerial vehicles. It takes the minimum rescue penalty and the minimum resource occupation as the optimization objective function, and establishes a multi-objective allocation model for unmanned aerial vehicle cluster rescue tasks under constraints; and uses a particle swarm optimization algorithm to obtain the optimal solution for the objective function. The present invention adopts the above-mentioned multi-task allocation method for heterogeneous unmanned aerial vehicle clusters based on decomposition learning particle swarms, takes minimizing rescue penalties and resource occupation as evaluation indicators, and establishes a multi-objective rescue task allocation model, which fully considers the characteristics of the problem itself and can generate a set of feasible and high-quality particles; can use the excellent solutions in the learning library to continuously update the inferior components in the particles, and more effectively complete the update of the particle position; can locally develop the sparse areas on the Pareto frontier, and further improve the distribution of the Pareto optimal solution set obtained by the algorithm.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles, and in particular to a multi-task allocation method for a heterogeneous unmanned aerial vehicle cluster based on decomposition learning particle swarm. Background Art

[0002] Unmanned Aerial Vehicle (UAV) is an aircraft that can complete tasks without human control. Due to its advantages such as high efficiency, flexibility and low cost, it has been widely used in military reconnaissance, disaster monitoring, post-disaster rescue, logistics distribution and other fields in recent years. At the same time, with the development of technologies such as artificial intelligence, network communication and autonomous systems, UAV technology has gradually developed towards cluster collaboration. Among them, task allocation is the key technology for UAV clusters to efficiently perform tasks. With the continuous expansion of the task field of UAV clusters, the scope of task allocation continues to expand, and the scale and complexity of the problem have also increased exponentially, which has brought new challenges to the task allocation technology of UAV clusters. In the use of UAV clusters to perform post-disaster rescue material distribution tasks, it is often necessary to consider the number of trapped people in the disaster area and the amount of rescue materials required, and make trade-offs in the performance of different types of UAVs such as capacity and speed, and then select the most appropriate UAV type and task execution order. In scenarios such as multiple indicators / goals, multiple preferences and strong constraints, it is difficult for manual decision-making to fully explore the best match between different types of UAVs and rescue points with different attributes.

[0003] In terms of model solving, heuristic search algorithms are widely used as solving methods. However, for the problem of task allocation for drone swarms, which has many rescue tasks with complex distribution, strong environmental constraints, and limited disposable resources, existing classic heuristic algorithms such as genetic algorithms, simulated annealing algorithms, taboo search algorithms, and ant colony algorithms cannot obtain ideal solution results. In view of this, scholars have made improvements on these algorithms in terms of initialization quality, convergence speed, and solution accuracy.

[0004] Particle swarm optimization (PSO), as a classic group intelligence optimization algorithm, has the advantages of being simple to implement and fast in convergence, and has been widely used in various fields. In drone swarm rescue missions, different rescue mission points often have different rescue values, that is, different rescue weights. However, existing studies assume that all mission points have equal weights. In addition, when allocating drone swarm tasks, the drones involved often have different properties. Some drones may have a faster speed, but their endurance is relatively short; on the contrary, some drones may have a larger capacity, but their endurance is relatively long. However, most of the above studies assume that all drones have the same capacity and endurance. Summary of the invention

[0005] The purpose of the present invention is to provide a multi-task allocation method for heterogeneous UAV clusters based on decomposition learning particle swarm to solve the problems mentioned in the background technology.

[0006] To achieve the above object, the present invention provides a method for allocating multi-tasks of heterogeneous UAV clusters based on decomposition learning particle swarm, comprising the following steps:

[0007] S1. Taking the minimum rescue penalty and the minimum resource occupation as the optimization objective function, a multi-objective allocation model for UAV swarm rescue tasks is established under constraints;

[0008] S2. Use the particle swarm optimization algorithm to obtain the optimal solution of the objective function.

[0009] Preferably, the two objective functions in S1 are defined as:

[0010]

[0011] Among them, f1 is the rescue penalty function, Punish i is the rescue penalty value of drone i; f2 is the resource occupation value of K drones and M base vehicles, Cost u Cost C are the resource usage values ​​of a single UAV and an unmanned vehicle respectively; set Nu and Nc are the total number of drones and base vehicles owned by the rescue center;

[0012] The calculation formula for the rescue penalty value of drone i in f1 is as follows:

[0013]

[0014] in, To start from the starting point, UAV i reaches the task O i,j time, For Task O i,j The rescue value; Indicates that drone i completes mission O i,j The rescue penalty received when

[0015] The calculation of is divided into two cases:

[0016] When the target point O i,j-1 With the target point O i,j When it is a rescue mission, The calculation formula is as follows:

[0017]

[0018] in, is the time to unload supplies after arriving at the rescue point j, The target point O i,j-1 With the target point O i,j The distance between type is the flight speed of UAV i;

[0019] When the target point O i,j-1 For rescue point, target point O i,j When it is a base vehicle, the UAV is performing the next rescue mission. i,j+1 Previously, we went to the nearby base to replenish supplies or replace batteries. The calculation formula is as follows:

[0020]

[0021] in, The time to replace the battery after arriving at the base vehicle.

[0022] Preferably, the constraints in S1 include:

[0023] The distance that drone i travels from the base vehicle to the base vehicle is not greater than the battery limit of this type of drone:

[0024]

[0025] Where, l' is the number of tasks that UAV i needs to complete before it leaves the base vehicle and returns to the base vehicle next time;

[0026] The total amount of materials unloaded by drone i after it departs from the current base vehicle is not greater than the capacity of the type of drone:

[0027]

[0028] Each mission is assigned to a drone, and all missions are assigned to drones for execution:

[0029]

[0030] The drone i reserves a certain amount of power:

[0031]

[0032] In formula (8), back = max(min(dc m,n ,m=1,2,…,M),n=1,2,…,N), is the maximum value of the distance from all rescue points to their nearest base vehicle; l” is the distance of the base vehicle in the actual flight sequence O i The position in

[0033] When the drone i completes the mission Oi,j If the remaining power If the requirements of the next task are not met, that is, the remaining power is not enough to reach the next rescue point, then return to the nearest base vehicle to replace the battery; embedding the above constraints into the problem encoding, we have:

[0034]

[0035] Among them, O i,j+1 is the base vehicle sequence value embedded in the current task execution sequence of UAV i, and m is the real index value of the embedded base vehicle;

[0036] When the drone i completes the mission O i,j When the remaining materials do not meet the requirements of the next task, it is necessary to return to the nearest base vehicle to replenish the materials. The above constraints are embedded in the problem encoding, and we have:

[0037]

[0038] In formula (10), Complete the rescue mission for the drone i,j The amount of relief supplies remaining at the time of the

[0039] Each drone starts from the base vehicle:

[0040] O i,1 >100 (11)

[0041] Each drone i returns to the base vehicle:

[0042] O i,end >100 (12)

[0043] Each drone performs at least one mission:

[0044]

[0045] Before and after a mission, the number of drones carried or recovered by each base vehicle shall not exceed its maximum capacity:

[0046]

[0047] In formula (14), space max The maximum hangar capacity of the base vehicle.

[0048] Preferably, the specific steps of S2 are as follows:

[0049] S21. Establish appropriate particle encoding and decoding strategies;

[0050] S22, combining clustering and single-step greedy selection sorting operators to initialize the particle swarm;

[0051] S23, using a decomposition-based learning algorithm to update particles, including a learning phase and a correction phase;

[0052] S24. Use a local development mechanism based on internal crossover mutation to expand the diversity of the optimal solution frontier.

[0053] Preferably, the specific content of S21 is:

[0054] In the encoding stage, the encoding matrix X of the drone task allocation scheme is established according to the type of drone, the number of drone dispatches K, and the number of rescue points N. In the matrix X, the first column is the drone type, and its value range is {1,2}; when the value is 1, it means that type1 drone is selected; when the value is 2, it means that type2 drone is selected; the second to N+1 columns are the task sequences assigned to the drones, and their values ​​are {1,2,…,N}; each row in X contains the type of drone called and the rescue task sequence assigned to it; the number of drone dispatches K determines the number of rows in X;

[0055] In the decoding stage, for any particle X, the decoding operation is performed on each row of it in turn. First, the type of the i-th UAV is obtained from the value of its first column; when the value is 1, it means that the type 1 UAV is selected; when the value is 2, it means that the type 2 UAV is selected; secondly, the rescue mission order of the i-th UAV is obtained from its 2nd to N+1th columns, set to Then, from the starting task First, check whether the current endurance of the i-th drone can reach O i The next rescue mission, or whether the drone's current supplies can meet O i If not, the i-th UAV must first go to the nearest base vehicle to replace the battery and replenish supplies. At this time, the base vehicle will be inserted before the next rescue mission as the flight node that the i-th UAV must pass through. i After completing all rescue tasks, the decoding result of the i-th row element of particle X is obtained.

[0056] Preferably, the specific content of S22 includes:

[0057] In the clustering-based initialization phase, first, all rescue tasks are divided into K subclusters using K-means clustering technology; second, the total amount of materials required for the rescue tasks in each subcluster is calculated, and the K subclusters are sorted in descending order according to the total amount of materials; finally, the K rescue drones are sorted in descending order according to the material capacity, and the rescue tasks contained in the kth subcluster are assigned to the kth rescue drone in turn, k = 1, 2, ..., K;

[0058] After initialization, the rescue missions of each drone in particle X are sorted using a single-step greedy selection sorting algorithm. The current position of the i-th drone is pos, select The comprehensive benefits of the next rescue mission are:

[0059]

[0060] Among them, d p,j For rescue mission Distance from the drone, v j For the task The rescue value;

[0061] The specific method is as follows: (1) Use formula (15) to calculate Task k The comprehensive benefit value of all rescue tasks in the , select the task with the largest comprehensive benefit value as the first task that the drone needs to rescue, set as task' k ;

[0062] (2) From Task k Delete task' k , and set the robot's position to task' k ;

[0063] (3) Continue to use formula (15) to calculate Task i The comprehensive benefit value of all rescue tasks in Task i Select the task with the largest comprehensive benefit value as the second task that the drone needs to rescue;

[0064] (4) Continue to execute steps (2) and (3) until the task is determined. k The order in which all tasks are executed.

[0065] Preferably, the content of the learning phase is as follows:

[0066] Introducing the learning library and the concept of task execution efficiency, for any particle X i , the task execution efficiency of the kth UAV is:

[0067]

[0068] In the formula, y k is the rescue penalty of drone k in particle; v j Assemble Tasks for rescue missions k The rescue value of the jth task in ;

[0069] For each particle to be updated, the task execution efficiency of all drones it contains is calculated using the task execution efficiency, and the task rescue sequence with the worst execution efficiency is selected; in the learning stage, each particle updates the worst rescue sequence it contains, and the specific update formula is as follows:

[0070]

[0071] in, is the i-th particle in the t-th generation population, for The execution order of the rescue mission assigned to the kth UAV in the particle encoding matrix is ​​the kth row of the particle encoding matrix, where K is the scale of UAVs called by the particle; target1 is the particle The index value of the updated row, that is The row with the smallest value; The best facilitator is selected from the learning pool; for The kth UAV is assigned the execution order of the rescue mission, K' is the scale of UAVs called by the optimal particle; for a particle containing K UAVs and M base vehicles Select the element with the same [K,M] value from the learning library as the optimal guide, that is,

[0072] Preferably, the contents of the revision stage are as follows:

[0073] a. Redistribute and re-rank the rescue tasks; define the formula for the task burden of drone k as:

[0074]

[0075] Where, vs j Assemble Tasks for rescue missions k The amount of emergency supplies required for the jth rescue mission; S k is the material capacity of drone k;

[0076] For the rescue task j, let its position be The center position of the set of tasks performed by drone k is center_p k =(center_x k ,center_y k ), the transfer distance of rescue mission j relative to drone k is:

[0077]

[0078] Normalize the task burden and the transfer distance value, and define the weighted value of the two as the benefit value of selecting drone k for rescue task j, that is:

[0079]

[0080] Where ω1 and ω2 are the weight values ​​of task burden and transfer distance respectively, and ω1+ω2=1, ω1≥0, ω2≥0.

[0081] The update formula of the particle correction stage is as follows:

[0082]

[0083] In the formula, is the candidate position particle, Sc is the set of drones whose rescue missions have changed, which corresponds to the particle The updated row index value in; Indicates for The reordering result of the tasks assigned to the kth UAV in ;

[0084] b. Correct the drone type. The specific steps are as follows:

[0085] (1) Sequentially extract the rescue mission execution sequence of the type 2 UAV in the current particle;

[0086] (2) Calculate the total amount of rescue materials and total electricity consumed to complete the rescue sequence;

[0087] (3) If the total amount of rescue materials and the total power are both within the range of type 1 drones, the type of drone in the rescue sequence of the mission is changed to type 1;

[0088] (4) Repeat (1)-(3) until all rescue mission sequences corresponding to type 2 drones in the particle are checked.

[0089] Preferably, the specific content of S24 is:

[0090] Suppose the candidate solution selected from the learning base is According to formula (19), we can get The task burden value task_burden of all drones in the task_burden is used to select the drone with the largest task burden and its corresponding rescue mission execution plan The UAV with the smallest mission burden and its corresponding rescue mission execution plan Then, select X a1,max and X a1,min As two chromosomes, a crossover or mutation operation is performed to generate a new solution, which is used to replace the worst particle in the current population.

[0091] Therefore, the present invention adopts the above-mentioned heterogeneous UAV cluster multi-task allocation method based on decomposition learning particle swarm, which has the following beneficial effects:

[0092] (1) Considering multiple optimization objectives such as rescue penalty and resource occupancy rate, and taking into account constraints such as drone endurance and maximum capacity, a multi-objective allocation model for drone swarm rescue tasks is established, which is mainly based on different rescue value mission points, multiple types of drones and multiple base vehicles;

[0093] (2) A hybrid particle encoding and decoding method is proposed that takes into account the type and scale of UAVs, the order of task execution, and the performance of UAVs, which improves the efficiency of the algorithm operation and reduces the invalid calculations in the algorithm operation;

[0094] (3) A particle swarm initialization strategy combining clustering and single-step greedy selection sorting operators is proposed to improve the quality of the initial allocation scheme;

[0095] (4) A two-stage particle updating strategy based on decomposition learning is proposed. By continuously updating the worst rescue sequence contained in each particle, the learning phase can generate new positions with better index values; by correcting the infeasible positions generated in the learning phase, the correction phase can solve problems such as repeated execution of rescue tasks and failure to execute some rescue tasks;

[0096] (5) A local development mechanism based on internal crossover mutation is proposed to further improve the diversity of the Pareto optimal solution set obtained by the algorithm.

[0097] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] Figure 1 It is a basic framework diagram of a multi-objective particle swarm algorithm based on solution decomposition according to an embodiment of the present invention;

[0099] Figure 2 is an example diagram of an initial task allocation strategy according to an embodiment of the present invention;

[0100] Figure 3 is a diagram of an example of correction of the type of drone with redundant capabilities according to an embodiment of the present invention;

[0101] Figure 4 is an example diagram of a particle update strategy according to an embodiment of the present invention;

[0102] Figure 5 is an example diagram of crossover and mutation operations according to an embodiment of the present invention;

[0103] Figure 6 is a set of task allocation solutions obtained by the multi-objective particle swarm algorithm of an embodiment of the present invention;

[0104] Figure 7It is a specific task allocation scheme obtained by the algorithm proposed in the embodiment of the present invention for the case where the minimum rescue penalty (f1) is the minimum;

[0105] Figure 8 It is a specific task allocation scheme obtained by the algorithm proposed in the embodiment of the present invention for the case where the resource occupancy (f2) is minimal;

[0106] Fig. 9 This is a specific task allocation solution obtained by the algorithm proposed in the embodiment of the present invention for the case where 0.5f1+0.5f2 is minimum. DETAILED DESCRIPTION

[0107] Example

[0108] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0109] The present invention discloses a multi-task allocation method for heterogeneous UAV clusters based on decomposition learning particle swarm, the specific contents of which are as follows:

[0110] S1. Taking minimum rescue penalty and minimum resource occupation as the optimization objective function, a multi-objective allocation model for UAV swarm rescue tasks is established under constraints.

[0111] 1.1 Description of actual application scenarios

[0112] The specific application scenario is described as follows: After a disaster occurs, there are multiple locations within a certain area that are worth investigating and dispatching rescue drones, such as schools, hospitals, supermarkets, residential areas, etc. Different types of rescue locations have different corresponding rescue values. The rescue level is divided according to the flow of people at the rescue location. Assume that there are N rescue locations in the area, represented by the target set Task = {task 1 ,task 2 ,…,task N}, their corresponding rescue value can be expressed as V = [v 1 ,v 2 ,…,v N ], where v i Indicates task task i rescue value.

[0113] The rescue scene includes different types of rescue drones and multiple base vehicles. Rescue drones can be divided into two types: high-speed, small-payload, short-range drone type 1 and low-speed, large-payload, long-range drone type 2; the two types of drones have different ranges, capacities, and speeds. There are M base vehicles in the scene, which can not only serve as base vehicles to release drones, but also as resource supply stations to provide rescue supplies or replace batteries for drones. In the event of a disaster, the rescue center will use the K-means clustering algorithm to calculate the optimal parking positions of these base vehicles based on the distribution of N rescue missions. Let the parking position set of the base vehicles be PC = {pc1, pc2,…, pc M}, where pc i Represents the coordinate position of the i-th base vehicle.

[0114] The number of drones dispatched for the rescue mission is K, and their types are TYP = [type 1 ,type 2 ,…,type K ]. The rescue center will place each drone on the most suitable base vehicle according to the execution sequence of the assigned tasks. After a rescue drone departs from the base vehicle, it can go to any base vehicle to replenish resources on the way, and can go to any base vehicle to complete the recovery after completing the assigned task. When all tasks in the Task are completed, the rescue mission is successfully completed. Table 1 shows the specific symbols involved in the model and the corresponding descriptions.

[0115] Table 1 Symbols and descriptions

[0116]

[0117]

[0118] 1.2 Objective Function

[0119] The main goal of the rescue mission is to save more lives in the shortest possible time. Considering that the sooner the rescue material needs of the trapped people in each rescue point after the disaster are met, the greater the possibility of their survival, the purpose is to maximize the probability of survival of more people; at the same time, considering that there may be more than one area to be rescued, the proportion of rescue resources in a certain area must take into account the balance of the overall rescue resource allocation. Therefore, this embodiment uses the minimum rescue penalty and the minimum resource occupation as two objective functions. The minimum rescue penalty is the sum of the rescue penalties of all drones participating in the mission, and the minimum resource occupation is the sum of the resource occupation of the drones and base vehicles participating in the rescue mission in this area.

[0120] X K×(N+2)represents the generalized task assignment matrix of all K drones, where the first column is the type of the K drones, the second column is the initial position of the drone (i.e. the number of the base vehicle), and the remaining 3 to N columns are the task lists assigned to the drones. When j ≥ 3, x i,j = 1, indicating that the j-2 rescue mission is assigned to drone i, x i,j =0 means the j-2 rescue mission is not assigned to UAV i; i =[O i,1 ,O i,2 ,…,O i,l ] indicates that drone i completes the assigned task set X i The flight path at that time is the base vehicle, rescue point and recovery point that the UAV passes through in sequence from the starting point, where O i,j is the real index value of the rescue mission or the base vehicle passing through, l is O i The number of nodes included. Except for the base car at the recovery point, path O i Other base vehicles in the system replenish supplies and replace batteries for the drones. The variable to be optimized in this embodiment is the task allocation matrix X K×(N+2) and the order in which each drone completes the assigned tasks. Based on this, the above two objective functions can be defined as:

[0121]

[0122] In formula (1), f1 is the rescue penalty function, Punish i is the rescue penalty value of UAV i; f2 is the resource occupation value of K UAVs and M base vehicles, where Cost u and Cost C are the resource usage values ​​of a single UAV and an unmanned vehicle respectively. Without loss of generality, we set Among them, Nu and Nc are the total number of drones and base vehicles owned by the rescue center.

[0123] The calculation formula for the rescue penalty value of drone i in f1 is as follows:

[0124]

[0125] In formula (2), To start from the starting point, UAV i reaches the task O i,j time, For Task O i,j The rescue value; Indicates that drone i completes mission O i,j The rescue penalty received when the The calculation of is divided into two cases:

[0126] When the target point O i,j-1 With the target point O i,j When it is a rescue mission, The calculation formula is as follows:

[0127]

[0128] in, is the time to unload supplies after arriving at the rescue point j, The target point O i,j-1 With the target point O i,j The distance between type is the flight speed of UAV i.

[0129] 2) When the target point O i,j-1 For rescue point, target point O i,j When it is a base vehicle, the UAV is performing the next rescue mission. i,j+1 Previously, it was necessary to go to a nearby base vehicle to replenish supplies or replace batteries. The calculation formula is as follows:

[0130]

[0131] in, The time to replace the battery after arriving at the base vehicle.

[0132] Furthermore, the generalized distribution matrix Taking the corresponding rescue mission sequence or flight path O = {[1022102], [1014102], [10231011101]} as an example, the calculation process of the above objective function value is explained. 3×6 It can be seen that the current scene contains 3 drones, 2 base vehicles and 4 rescue missions. According to the first row

[10100] and the actual flight path of the corresponding drone O1 = [1022102], the type of drone 1 is type1. Before performing the mission, it was placed on base vehicle 2 and the assigned mission was rescue point 2. The mission rescue sequence corresponding to drone 1 is O1 = [1022102], indicating that the actual flight path is base vehicle 2 → rescue point 2 → base vehicle 2 (无人机回收) , where O i,j When >100, it means the target point O of drone i i,j For the base vehicle, O i,j -100 is the number of the base vehicle; i,j <100, it means the target point O of drone i i,jWhen replenishing supplies, replacing batteries, or recycling, the drone chooses the base vehicle closest to it. Similarly, drone 3 is type 2 and is placed on base vehicle 2 before performing the mission. The assigned task set is {rescue point 1, rescue point 3}, and the order of performing rescue tasks is: rescue point 3 → rescue point 1. Its corresponding task rescue sequence is O3 = [1023 101 1 101], indicating that its actual flight path is base vehicle 2 → rescue point 3 → base vehicle 1. (补充物资、更换电池) →Rescue point 1→Base vehicle 1 (无人机回收) .

[0133] It should be noted that the model of this embodiment ignores the differences caused by the order of releasing drones, and assumes that all drones are released together at the beginning of the mission; the model of this embodiment assumes that the base vehicle can replenish supplies and replace batteries for multiple drones at the same time, and there is no situation where drones wait in line.

[0134] 1.3 Constraints

[0135] In actual situations, the range and payload of rescue drones are limited, and the performance parameters of different types of rescue drones are also different. In order to make the model more realistic and maximize the use of drone resources, the following constraints are given:

[0136] 1) The distance traveled by drone i from the base vehicle to the base vehicle must not exceed the battery (distance) limit of the type of drone:

[0137]

[0138] Among them, l' is the number of tasks that UAV i needs to complete before it leaves the base vehicle and returns to the base vehicle next time.

[0139] 2) The total amount of materials unloaded by drone i after it departs from the current base vehicle shall not be greater than the capacity of the type of drone:

[0140]

[0141] 3) Each task can only be assigned to one drone, and all tasks need to be assigned to drones for execution:

[0142]

[0143] 4) The drone i will always reserve a certain amount of power to ensure that it can return to the nearest base vehicle smoothly after completing the mission:

[0144]

[0145] In formula (7), back = max(min(dc m,n,m=1,2,…,M),n=1,2,…,N), is the maximum value of the distance from all rescue points to their nearest base vehicles. For the convenience of calculation, the reserved power of each UAV is deducted in advance when executing constraint 5). l” is the base vehicle in the actual flight sequence O i The position in which, for example, O i =[101 2 10 103 5 102], l”=[1 4 6].

[0146] 5) When the drone i completes the mission O i,j If the remaining power If the requirements of the next task are not met, that is, the remaining power is not enough to reach the next rescue point, it is necessary to return to the nearest base vehicle to replace the battery. Embedding the above constraints into the problem encoding, we have:

[0147]

[0148] Among them, O i,j+1 is the base vehicle sequence value embedded in the current task execution sequence of UAV i, and m is the real index value of the embedded base vehicle.

[0149] 6) When the drone i completes the mission O i,j If the remaining supplies do not meet the requirements of the next task, you must first return to the nearest base vehicle to replenish supplies. Embedding the above constraints into the problem encoding, we have:

[0150]

[0151] In formula (9), Complete the rescue mission for the drone i,j The amount of relief supplies remaining.

[0152] 7) Each drone i must start from the base vehicle:

[0153] O i,1 >100 (10)

[0154] 8) Each drone i must return to the base vehicle:

[0155] O i,end >100 (11)

[0156] 9) Each drone performs at least one mission:

[0157]

[0158] 10) Before and after the mission, the number of drones carried or recovered by each base vehicle shall not exceed its maximum capacity:

[0159]

[0160] In formula (13), space max The maximum hangar capacity of the base vehicle.

[0161] S2. Use the particle swarm optimization algorithm to obtain the optimal solution of the objective function.

[0162] The above-mentioned UAV rescue task allocation model not only needs to optimize the execution order of rescue tasks, but also needs to optimize the number, type and number of UAVs used for rescue. It also contains multiple types of constraints, which makes the solution space of the problem very complicated. To deal with the above problems, a multi-objective particle swarm optimization algorithm based on solution decomposition (SdLMOPSO) is proposed.

[0163] Figure 1 The basic framework of the algorithm is presented. Compared with the standard PSO algorithm, SdLMPSO introduces three new strategies or operators. Among them, the purpose of the "particle swarm initialization strategy guided by problem characteristics" is to generate a set of high-quality and diverse initial particle swarms with the help of the characteristics of the problem itself. The main purpose of the "particle update strategy based on decomposition learning" is to continuously update the worst rescue sequence in each particle using the historical optimal allocation scheme stored in the learning library, thereby generating new particles with better index values; the purpose of the "local mutation mechanism based on the learning library" is to use the mutation mechanism to deeply develop some potential particles in the population, thereby improving the distribution of the obtained Pareto optimal solution set; in addition, a learning library LearningSet is introduced to save the best rescue sequence found by the population during the iteration process.

[0164] 2.1 Particle Encoding and Decoding

[0165] Before using the particle swarm optimization algorithm to solve the problem, we must first establish a suitable particle encoding and decoding strategy. According to the characteristics of the problem, a mixed integer encoding method is used to represent a task allocation scheme, i.e., a particle; at the same time, considering the flight constraints of the UAV, a particle decoding strategy embedded in the base vehicle is given.

[0166] 2.1.1 Coding stage

[0167] According to the type of drone, the number of drone dispatches K, and the number of rescue points N, the encoding matrix X of the drone task allocation scheme is established. In the matrix X, the first column is the drone type, and its value range is {1,2}. When the value is "1", it means that type1 drone is selected; when the value is "2", it means that type2 drone is selected. The second to N+1 columns are the task sequences assigned to the drone, and their values ​​are {1,2,…,N}. Each row in X contains the type of drone called and the rescue task sequence assigned to it. The number of drone dispatches K determines the number of rows in X.

[0168] Taking the problem involving 3 drones and 7 rescue missions as an example, Table 2 shows an example of encoding. The particle corresponding to this encoding is The first line indicates that the first drone selects type 1 and its assigned rescue mission is {task 2 , task 3}, the rescue order is task 3 →task 2 .

[0169] Table 2 Particle coding examples

[0170]

[0171] 2.1.2 Decoding stage

[0172] Due to the constraints of the drone's endurance and the amount of rescue materials it carries, the mission rescue sequence given in the above encoding stage is not completely feasible. For example, when a drone needs to go from rescue mission 1 to rescue mission 2, if its endurance cannot complete the flight mission, the drone must first go to the nearest base vehicle to replace the battery or charge. In view of this, a decoding strategy embedded in the base vehicle is given.

[0173] For any particle X, the following decoding operation is performed on each row of the particle X (taking the i-th row as an example): First, the type of the i-th drone is obtained from the value of the first column. As mentioned above, when the value is "1", it means that the type 1 drone is selected; when the value is "2", it means that the type 2 drone is selected. Then, the rescue mission order of the i-th drone is obtained from the 2nd to the N+1th columns, which can be set as Then, from the starting task First, check whether the current endurance of the i-th drone can reach O i The next rescue mission, or whether the drone's current supplies can meet O i If not, the i-th UAV needs to go to the nearest base vehicle to replace the battery and replenish supplies. At this time, the base vehicle will be inserted before the next rescue mission as the flight node that the i-th UAV must pass through. Based on the above method, after checking O i After completing all rescue tasks, the decoding result of the i-th row element of particle X can be obtained.

[0174] Taking the third row in Table 2 (i.e. the third drone) as an example, the type of the drone is type1, and the rescue tasks that need to be performed in sequence are {task 6 ,task 1First, the drone flies from the starting position (assuming base vehicle 1) to the rescue mission task. 6 . Assume that we reach task 6 The remaining power of the drone cannot reach the task 1 , then it needs to fly to the nearest base vehicle (assuming it is base vehicle 2) to replace the battery. Then, the drone will fly directly from base vehicle 2 to task 1 Finally, following the principle of proximity, the nearest base vehicle (assuming it is base vehicle 2) is found to complete the recovery of the drone. The decoding result of the third line is: base vehicle 1 → task 6 →Base vehicle 2→task 1 →Base car 2, that is [101 6 102 1 102].

[0175] Table 3 Particle decoding example

[0176]

[0177]

[0178] 2.2 Problem feature-guided particle swarm initialization strategy

[0179] As mentioned above, the variables to be optimized in this embodiment include the type of drone, the rescue sequence of each drone, that is, the number of rows of the matrix X, the value of the first column in X, and the values ​​of the second to N-1 columns in X. When initializing a particle, the number of rows of the corresponding encoding matrix should be determined first, that is, the number of drones called. This embodiment uses a random strategy to determine the number of drones to be dispatched within the allowed number range. Subsequently, the type of each drone needs to be determined. Similarly, this embodiment randomly initializes the type of drone within the optional drone type. Finally, the rescue sequence of each drone in X is initialized using the methods in Sections 2.2.1 and 2.2.2.

[0180] 2.2.1 Clustering-based initialization allocation strategy

[0181] In the UAV task allocation problem, the flight distance of the UAV is an important evaluation indicator. Directly assigning each rescue task to the nearest UAV usually results in a better task allocation result. This embodiment proposes a task initial allocation strategy based on K-means clustering to initialize the values ​​of the 2nd to K+1 columns in particle X, that is, to determine the initial rescue task set for each UAV in particle X.

[0182] Taking particle X with K drones as an example, first, all rescue tasks are divided into K subclusters using K-means clustering technology; then, the total amount of materials required for the rescue tasks in each subcluster is calculated, and the K subclusters are sorted in descending order according to the total amount of materials; finally, the K rescue drones are sorted in descending order according to the material capacity, and the rescue tasks contained in the kth subcluster are assigned to the kth rescue drone in turn, k = 1, 2, ..., K.

[0183] When determining the initial rescue mission set for each drone, the above method takes into account the maximum capacity of each drone. This can effectively prevent drones from frequently returning to the base vehicle to replenish rescue supplies when performing missions. Furthermore, Algorithm 1 gives the pseudo code of the clustering-based initial task allocation strategy. In particular, if there are multiple particles in the initial population that call the same number and type of drones, then except for the first particle that uses the above clustering strategy, the other particles all use a random method to complete the allocation of the initial rescue mission.

[0184] Assume that a particle contains 5 drones, and their types are TYP = 22111], Figure 2 An example of the initial task allocation strategy is given. First, the 16 rescue tasks are divided into 5 sub-clusters, among which the total amount of materials required by sub-cluster 1 is the largest, which is 150; and the total amount of materials required by sub-cluster 5 is the smallest, which is 40. Then, according to the material capacity of the drone and the total amount of materials required by each sub-cluster, the 5 sub-clusters are allocated to the corresponding 5 drones in turn. The allocation results are shown in the figure. Figure 2 As shown in the table on the right.

[0185] 2.2.2 Single-step greedy selection sort operator

[0186] After completing the initial task allocation, it is necessary to sort the rescue tasks of each drone in particle X. On the one hand, since the penalty intensity of the rescue point is linearly related to its own rescue value and the time of unloading materials, in order to obtain the minimum rescue penalty, a drone will often give priority to the task point with high rescue value; on the other hand, when determining the most suitable rescue task, the distance from the drone to these rescue tasks must also be considered, and the rescue tasks closest to them must be given priority. Taking both of the above factors into consideration, a rescue task initial sorting strategy based on single-step greedy selection is proposed.

[0187] First, a rescue mission evaluation index that takes into account both rescue value and distance is proposed to determine the comprehensive benefits of selecting a rescue mission. The current position of the i-th drone is pos, so choose The comprehensive benefits of the next rescue mission are:

[0188]

[0189] Among them, d p,j For rescue mission Distance from the drone, v j For the task rescue value.

[0190] Taking the kth drone in X as an example, assuming that the assigned task set is The proposed initial sorting strategy for rescue tasks is as follows: (1) Using formula (14) to calculate Task k The comprehensive benefit value of all rescue tasks in the rescue task list is selected as the first task that the drone needs to rescue. Let's set it as task' k ; (2) From Task k Delete task' k , and set the robot's position to task' k ; (3) Continue to use formula (14) to calculate Task i The comprehensive benefit value of all rescue tasks in Task i Select the task with the largest comprehensive benefit value as the second task that the drone needs to rescue; (4) Continue to execute steps (2) and (3) until Task k The execution order of all tasks in . Furthermore, Algorithm 1 gives the pseudo code of the sorting strategy based on single-step greedy selection.

[0191] Table 4 Pseudo code of sorting strategy based on single-step greedy selection

[0192]

[0193] 2.3 Particle Update Strategy Based on Decomposition Learning

[0194] In the standard particle swarm optimization algorithm, each particle updates its position according to the speed value jointly determined by the global leader and the individual leader. However, the decision variables of the problem dealt with in this embodiment are not only discrete in value, but also have strong dependencies between some components. The traditional particle update method is difficult to meet the requirements. A particle update strategy based on decomposition learning is proposed.

[0195] This strategy consists of two stages: the learning stage and the correction stage. In the learning stage, each particle uses the historical optimal allocation plan saved in the learning library to update the worst rescue sequence it contains, thereby generating a new position with a better index value; in the correction stage, the strategy will correct the infeasible positions generated in the learning stage to solve problems such as repeated execution of rescue tasks and failure to execute some rescue tasks. It should be noted that a learning library LearningSet is introduced to save the best rescue sequence found by the population during the iteration process.

[0196] 2.3.1 Learning Phase

[0197] In the learning phase, we first introduce the concept of task execution efficiency, starting from the particle X to be updated. i Select the worst rescue sequence, that is, the row elements that need to be updated.

[0198] Definition 1: Task execution efficiency. For any particle X i , the task execution efficiency of the kth UAV is:

[0199]

[0200] In the formula, y k is the rescue penalty of drone k in particle; v j is the rescue value of the jth task in the rescue task set Taskk.

[0201] For each particle to be updated, the task execution efficiency of all drones it contains can be calculated using Definition 1, and the task rescue sequence with the worst execution efficiency can be selected. As mentioned above, in the learning phase, each particle only updates the worst rescue sequence it contains. The specific update formula is as follows:

[0202]

[0203] in, is the i-th particle in the t-th generation population, for The execution order of the rescue mission assigned to the kth UAV in the particle encoding matrix is ​​the kth row of the particle encoding matrix, where K is the scale of UAVs called by the particle; target1 is the particle The index value of the row to be updated, that is The row with the smallest value. is the best guide selected from the learning library. for The kth UAV is assigned the execution order of the rescue mission, and K' is the scale of UAVs called by the optimal particle. For a particle containing K UAVs and M base vehicles This embodiment selects elements with the same [K, M] value from the learning library as its optimal guide, that is,

[0204] 2.3.2 Correction Phase

[0205] For each particle to be updated After executing formula (16), the new position There may be situations where some rescue tasks are repeated or not executed. To this end, a rescue task redistribution and secondary sorting mechanism is proposed to correct these infeasible positions. At the same time, after completing the redistribution and secondary sorting, the particles may also store the situation of redundant drone capabilities, that is, the material capacity and maximum range of a drone are significantly greater than the total material demand and total range of the assigned task set. In view of this, a correction mechanism for redundant drones is also proposed.

[0206] 1) Redistribution and secondary sorting of rescue tasks

[0207] For particles and its candidate locations First compare and The element value is obtained to obtain the set of repeated rescue tasks task_overlap and the set of unexecuted rescue tasks task_redistri; then, the rescue tasks in task_overlap are erased from the original UAV execution task sequence, and the rescue tasks of the UAV are reordered; then, for each rescue task in task_redistri, a suitable UAV is selected for it according to formula (19) (that is, it is inserted into a more suitable row), and the rescue tasks contained in the UAV / row are reordered; finally, the candidate position particles are calculated by decoding The rescue penalty y'.

[0208] Definition 2: Mission burden. The formula for defining the mission burden of drone k is:

[0209]

[0210] Where, vs j is the amount of emergency supplies required for the jth rescue task in the rescue task set Taskk; S k is the material capacity of drone k.

[0211] Definition 3: Transfer distance. For the task j to be rescued, assume that its location is The center position of the set of tasks performed by drone k is center_p k =(center_x k ,center_y k ), then the transfer distance of rescue mission j relative to drone k is:

[0212]

[0213] Normalize the task burden and the transfer distance value, and define the weighted value of the two as the benefit value of selecting drone k for rescue task j, that is:

[0214]

[0215] Where ω1 and ω2 are the weight values ​​of task burden and transfer distance respectively, and ω1+ω2=1, ω1≥0, ω2≥0.

[0216] Based on this, the update formula of the particle correction stage is as follows:

[0217]

[0218] In the formula, Sc is the set of drones whose rescue missions have changed, which corresponds to the particle The row index value that needs to be updated; Indicates for The reordering result of the kth UAV being assigned tasks. In order to maintain the diversity of the particle swarm, a hybrid method of single-step greedy sorting strategy and random sorting strategy is used to sort The kth UAV is assigned tasks for reordering. The probability greedy_pro value is used to control the probability of the particle adopting a single-step greedy strategy. The specific update formula of greedy_pro is shown in formula (21):

[0219]

[0220] In the formula, greedy is the set initial probability, ger is the current number of iterations, and G is the set total number of iterations. It can be seen that the greedy_pro value decreases as the number of iterations increases, which can increase the probability of jumping out of the local optimum while ensuring the convergence of the population.

[0221] 2) UAV type correction operator

[0222] In the model of this embodiment, the type 1 drone is better than the type 2 drone in speed, and the type 2 drone is better than the type 1 drone in material capacity and range; this naturally produces a calling trend: when the drone is assigned tasks and executed in order, if the type 1 drone can meet its total material demand and power requirements without returning to the base vehicle, then the rescue penalty obtained by the type 1 drone for executing this task sequence must be lower than the rescue penalty obtained by the type 2 drone. At this point, it is obvious that we should give priority to the type 1 drone. In order to reduce the computational burden, this mechanism only corrects the non-inferior particles in the current population.

[0223] The specific execution steps are as follows: (1) Sequentially extract the rescue mission execution sequence of the type 2 UAV in the current particle; (2) Calculate the total amount of rescue materials and the total power consumed to complete the rescue sequence; (3) If the total amount of rescue materials and the total power are both within the range of the type 1 UAV, then change the UAV type of the mission rescue sequence to type 1; (4) Repeat (1)-(3) until all the rescue mission sequences corresponding to the type 2 UAVs in the particle are checked.

[0224] Figure 3 An example of the proposed capability redundant drone correction mechanism is shown. First, the particle decoding matrix is ​​extracted The first one belongs to the "2" type of UAV and its rescue mission sequence The rescue mission sequence is [101,3,2,101]; then, calculate the total amount of rescue materials sum_sup and the total distance sum_dis required for the drone to execute the mission sequence. Assuming that the total amount of rescue materials and the total distance are both within the range of type 1 drones, the drone type corresponding to the rescue mission sequence is changed to "1".

[0225] 2.3.3 Execution steps of particle update strategy

[0226] The particles in the population As an example, Algorithm 3 gives the execution steps of the proposed particle update strategy. First, in lines 1-2, the particle to be updated is selected according to formula (16-2): The row corresponding to the drone with the lowest task execution efficiency (i.e. ), and the optimal guide of the particle Secondly, lines 3-4 update the particle by equation (16-1) ; then, lines 5-8 are based on the method in Section 2.3.2. Correction is performed to generate candidate locations Finally, in lines 9-14, simulated annealing is used to complete the final update of the particle position, and we get

[0227] It should be noted that when the simulated annealing method is used to determine Replace When the algorithm is iterated, there may be a phenomenon of particle degradation. To prevent the population from losing the optimal solution, this embodiment uses the learning library LearningSet to continuously save the optimal particles obtained during the population iteration process. At the same time, this embodiment sets the probability of accepting a poor solution. In the early stage of the algorithm iteration, particles will accept the poor solution with a greater probability, which can expand the diversity of the population; in the later stage of the algorithm iteration, particles will accept the poor solution with a smaller probability, which can improve the convergence of the population.

[0228] Assume that the worst rescue sequence that needs to be updated in the particle has been found through equation (15), that is, the task execution order of drone target1 is [102, 2, 5, 103, 7, 103], Figure 4 An example of particle update strategy is given. Figure 4 As can be seen from ①, the particle The task allocation plan corresponding to the target1 drone to be updated includes 3 base vehicles, 3 drones, and the rescue penalty is 156.75. First, according to the number of base vehicles, the number of drones, the scale of the rescue mission, etc., find the particles at the corresponding position in the learning library LearningSet and learn them, such as Figure 4 As shown in ②; and the rescue tasks (repeated tasks) that need to be transferred to this drone are set task_overlap = [8,3] and the rescue tasks (unexecuted tasks) that need to be transferred to other drones are set task_redistri = [2,5], as shown in Figure 4 As shown in ③; then, the rescue tasks in task_overlap are erased from the rescue sequence of the original UAV k1 and reordered, as shown in Figure 4 As shown in ④; According to formula (17-19), the new target drones k2 and k3 for each rescue task in task_redistri are determined, these rescue tasks are inserted into the rescue task set of the target drone, and the rescue task sequence of the target drone is reordered, as shown in Figure 4 As shown in ⑤, for example, [103,9,10,103] is the original rescue sequence of target drone k2, and [103,9,5,10,101] is the reordering result after inserting rescue task task5. Then perform the type correction operation in part 2) of Section 2.3.2, as follows Figure 4 As shown in ⑥; Finally, by Figure 4 ⑦ in the figure obtains the candidate position of the particle

[0229] Table 5 Particle update strategy based on decomposition learning

[0230]

[0231] 2.4 Local development mechanism based on internal crossover mutation

[0232] In order to further expand the diversity of the Pareto frontier obtained by the algorithm, a local development mechanism based on internal crossover mutation is proposed. First, the candidate solutions of the learning library LearningSet are non-dominated sorted to obtain its non-inferior solution set Ar1; then, the candidate solution with the sparsest distribution on the target space f2 is found from the non-inferior solution set Ar1; then, the local development mechanism is executed with the candidate solution as the object to generate a new solution; finally, the new solution is used to replace the worst particle in the current population.

[0233] Suppose the candidate solution selected from LearningSet is First, according to formula (17), we can get The task burden value task_burden of all drones in the task_burden can then be selected to select the drone with the largest task burden and its corresponding rescue mission execution plan The UAV with the smallest mission burden and its corresponding rescue mission execution plan Then, select X a1,max and X a1,min As two chromosomes, a crossover or mutation operation is performed to generate a new solution.

[0234] Assumptions and Figure 5 An example of crossover and mutation operations is shown. The crossover operation selects chromosome segments of the same position and length on two target chromosomes for exchange. Figure 5 As shown on the left, first, all non-zero values ​​on the chromosome are set to 1, and the length of the chromosome crossover segment is fixed to half of the total number of rescue tasks. The result after the crossover operation is and When performing a mutation operation, randomly select Previous non-zero element Negate, or select randomly The element with value 0 Negate, such as Figure 5 After completing this step, similar to the update strategy in Section 2.3, the tasks of the UAVs whose assigned tasks have changed need to be reordered, which will not be repeated here.

[0235] 2.5 Learning Library Update Strategy

[0236] This embodiment introduces a learning library LearningSet to store the optimal solution obtained by the algorithm. max ×K max The cell group, where M max and K maxare the maximum size of base vehicles and the maximum size of drones set according to the total number of rescue missions N. Unlike the external archive set commonly used in multi-objective evolutionary algorithms, LearningSet only retains one optimal particle for each combination of (number of base vehicles, number of drones), which is the optimal task allocation solution obtained by the algorithm so far. Because the number of combinations of (number of base vehicles, number of drones) is limited, there is no need to limit the maximum capacity of LearningSet.

[0237] Assume that the current particle swarm is POPt, and the non-inferior solution set selected from it is AR2. For each candidate solution in AR2 (in terms of candidate solution Take this as an example), and perform the following operations in sequence: (1) The decoding matrix Determine the number of base vehicles M and the number of drones K it uses; (2) Extract the solution with the same (M, K) value in the learning library, which may be set as Element (M, K), where (M, K) is the position index of the element of the solution; (2) Compare and the rescue penalty f1 value of Element(m,k). If If it wins, it is retained in LearningSet and Element(m,k) is deleted.

[0238] Simulation experiment

[0239] Consider the following scenario: there are 48 rescue missions, the lower and upper limits of the number of dispatched drones are 8 and 20 respectively, and the lower and upper limits of the number of dispatched base vehicles are 2 and 6 respectively. The rescue value of each rescue mission point is randomly divided into 1-4 levels; the amount of rescue materials required for each rescue point is linearly related to its rescue value, and the relationship between the two is: rescue material quantity = rescue value × 10; the time required for drones to unload rescue materials is linearly related to the amount of rescue materials required by the rescue point, and the relationship between the two is: unloading rescue material time = rescue material quantity / 2; the specific performance of the drone is shown in Table 6. The placement of the base vehicle is determined in advance according to the number of required base vehicles and the location of all rescue mission points, and the base vehicle is stationary. The total number of base vehicles and rescue drones held by the rescue center is 10 and 20 respectively, then the resource occupation of each base vehicle and each drone is 0.05 and 0.025 respectively.

[0240] In the simulation experiment, each drone takes off from the base vehicle and completes the assigned rescue mission in sequence. When performing the mission, it is necessary to ensure that the drone replenishes supplies and power in time during the mission, and returns to the base vehicle for recovery after completing the mission.

[0241] Table 6 UAV performance parameters

[0242]

[0243] The algorithm parameters are set as follows: the population size N is 30, and the maximum number of iterations is 200. The hypervolume (HV) index is used to evaluate the performance of each algorithm. For the three scenarios, the reference points are set to (1000, 0.35), (2000, 0.6), and (3500, 0.9).

[0244] Select the experimental result with the HV value closest to the median value among the 100 experiments. Figure 6 A set of task allocation solutions obtained by the algorithm proposed in this embodiment is shown.

[0245] The solutions obtained by the algorithm proposed in this embodiment are respectively based on three different tendencies: pursuing the minimum rescue penalty (f1), pursuing the minimum resource occupation (f2), and balancing the two. Figure 7-Figure 9 The task allocation schemes obtained by the proposed algorithm in three typical cases are shown. The dotted line in the figure is the flight trajectory of type 1 rescue drone, and the solid line is the flight trajectory of type 2 rescue drone; the arrows indicate the flight order of the drones; the black dots represent the base vehicles, such as MCV2 represents base vehicle 2; the red dots represent the rescue mission locations, and the annotations under the red dots represent (rescue mission number, rescue value), for example (15, 3) means that the rescue value of rescue mission 15 is level 3; the specific rescue scheme values ​​are shown in Table 7.

[0246] Table 7 Target values ​​obtained by the proposed algorithm for different typical cases

[0247]

[0248]

[0249] Therefore, the present invention adopts the above-mentioned heterogeneous UAV cluster multi-task allocation method based on decomposition learning particle swarm, takes minimizing rescue penalty and resource occupation as evaluation indicators, and establishes a multi-objective rescue task allocation model. Aiming at the established task allocation model, a multi-objective particle swarm algorithm based on solution decomposition learning is proposed. The proposed problem feature-guided particle swarm initialization strategy fully considers the characteristics of the problem itself and can generate a set of feasible and high-quality particles; the particle update strategy based on decomposition learning can continuously update the inferior components of the particles using the excellent solutions in the learning library, and more effectively complete the update of the particle position; the designed local development mechanism based on internal crossover mutation can locally develop the sparse areas on the Pareto frontier, further improving the distribution of the Pareto optimal solution set obtained by the algorithm. An optimal solution set with better convergence and distribution can be obtained, which is a highly competitive UAV cluster rescue task allocation solution method.

[0250] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A multi-task allocation method for heterogeneous UAV clusters based on decomposition learning particle swarm, It is characterized in that The following steps are involved: S1. Taking the minimum rescue penalty and the minimum resource occupation as the optimization objective function, under the constraint conditions, a multi-objective allocation model of UAV swarm rescue tasks is established. The rescue UAVs are divided into high-speed, small-load, short-range UAV type 1 and low-speed, large-load, long-range UAV type 2. The two objective functions are defined as: Among them, f 1 To rescue the penalty function, Punish i is the rescue penalty value of drone i; f 2 is the resource usage of K drones and M base vehicles, Cost u and Cost C are the resource usage values ​​of a single UAV and an unmanned vehicle respectively; set Nu and Nc are the total number of drones and base vehicles held by the rescue center, respectively; f 1 The calculation formula of the rescue penalty value of drone i is as follows: in, To start from the starting point, UAV i reaches the task O i,j time, For Task O i,j The rescue value; Indicates that drone i completes mission O i,j The rescue penalty received when in The calculation of is divided into two cases: When the target point O i,j-1 With the target point O i,j When it is a rescue mission, The calculation formula is as follows: in, is the time to unload supplies after arriving at the rescue point j, The target point O i,j-1 With the target point O i,j The distance between type is the flight speed of UAV i; When the target point O i,j-1 For rescue point, target point O i,j As the base vehicle, the drone is performing the next rescue mission. i,j+1 Previously, we went to the nearby base to replenish supplies or replace batteries. The calculation formula is as follows: in, The time to replace the battery after arriving at the base vehicle; S2. Use the particle swarm optimization algorithm to obtain the optimal solution of the objective function.

2. According to claim 1, the multi-task allocation method for heterogeneous UAV clusters based on decomposition learning particle swarm, It is characterized in that The constraints in S1 include: The distance that drone i travels from the base vehicle to the base vehicle is not greater than the battery limit of this type of drone: Where, l' is the number of tasks that UAV i needs to complete before it leaves the base vehicle and returns to the base vehicle next time; The total amount of materials unloaded by drone i after it departs from the current base vehicle is not greater than the capacity of the type of drone: Each mission is assigned to a drone, and all missions are assigned to drones for execution: The drone i reserves a certain amount of power: In formula (8), back = max(min(dc m,n ,m=1,2,…,M),n=1,2,…,N), is the maximum value of the distance from all rescue points to their nearest base vehicle; l” is the distance of the base vehicle in the actual flight sequence O i The position in When the drone i completes the mission O i,j If the remaining power If the battery is not enough to reach the next rescue point, return to the nearest base vehicle to replace the battery. Embedding the above constraints into the problem encoding, we have: Among them, O i,j+1 is the base vehicle sequence value embedded in the current task execution sequence of UAV i, and m is the real index value of the embedded base vehicle; When the drone i completes the mission O i,j When , if the remaining materials do not meet the requirements of the next task, we first return to the nearest base vehicle to replenish the materials. The above constraints are embedded in the problem encoding, and we have: In formula (10), Complete the rescue mission for the drone i,j The amount of relief supplies remaining at the time of the Each drone starts from the base vehicle: O i,1 >100 (11) Each drone i returns to the base vehicle: O i,end >100 (12) Each drone performs no less than 1 mission: Before and after a mission, the number of drones carried or recovered by each base vehicle shall not exceed its maximum capacity: In formula (14), space max The maximum hangar capacity of the base vehicle.

3. According to claim 2, the multi-task allocation method for heterogeneous UAV clusters based on decomposition learning particle swarm, It is characterized in that The specific steps of S2 are as follows: S21. Establish appropriate particle encoding and decoding strategies; S22, combining clustering and single-step greedy selection sorting operators to initialize the particle swarm; S23, using a decomposition-based learning algorithm to update particles, including a learning phase and a correction phase; S24. Use a local development mechanism based on internal crossover mutation to expand the diversity of the optimal solution frontier.

4. According to claim 3, the multi-task allocation method for heterogeneous UAV clusters based on decomposition learning particle swarm, It is characterized in that The specific contents of S21 are as follows: In the encoding stage, the encoding matrix X of the drone task allocation scheme is established according to the type of drone, the number of drone dispatches K, and the number of rescue points N. In the matrix X, the first column is the drone type, and its value range is {1,2}; when the value is 1, it means that type1 drone is selected; when the value is 2, it means that type2 drone is selected; the second to N+1 columns are the task sequences assigned to the drones, and their values ​​are {1,2,…,N}; each row in X contains the type of drone called and the rescue task sequence assigned to it; the number of drone dispatches K determines the number of rows in X; In the decoding stage, for any particle X, the decoding operation is performed on each row of it in turn, and the type of the i-th drone is obtained from the value of its first column; when the value is 1, it means that the type 1 drone is selected; when the value is 2, it means that the type 2 drone is selected; The rescue mission sequence of the i-th UAV is obtained from its 2nd to N+1th columns, and is set to From the starting task First, check whether the current endurance of the i-th drone can reach O i The next rescue mission, or whether the drone's current supplies can meet O i The need for the next rescue mission; If not, the i-th UAV will go to the nearest base vehicle to replace the battery and replenish supplies; at this time, the base vehicle will be inserted before the next rescue mission as the flight node that the i-th UAV must pass through. i After completing all rescue tasks, the decoding result of the i-th row element of particle X is obtained.

5. According to claim 4, the multi-task allocation method for heterogeneous UAV clusters based on decomposition learning particle swarm, It is characterized in that The specific contents of S22 include: In the clustering-based initialization phase, all rescue tasks are divided into K subclusters using K-means clustering technology; the total amount of materials required for the rescue tasks in each subcluster is calculated, and the K subclusters are sorted in descending order according to the total amount of materials; the K rescue drones are sorted in descending order according to the material capacity, and the rescue tasks contained in the kth subcluster are assigned to the kth rescue drone in turn, k = 1, 2, ..., K; After initialization, the rescue missions of each drone in particle X are sorted using a single-step greedy selection sorting algorithm. The current position of the i-th drone is pos, select The comprehensive benefits of the next rescue mission are: Among them, d p,j For rescue mission Distance from the drone, v j For the task The rescue value; The specific method is as follows: (1) Use formula (15) to calculate Task k The comprehensive benefit value of all rescue tasks in the rescue task list is selected as the first task of the drone rescue task, which is set as task' k ; (2) From Task k Delete task' k , and set the drone's position to task' k ; (3) Continue to use formula (15) to calculate Task i The comprehensive benefit value of all rescue tasks in Task i Select the task with the largest comprehensive benefit value as the second task that the drone needs to rescue; (4) Continue to execute steps (2) and (3) until the task is determined. k The order in which all tasks are executed.

6. According to claim 5, the method for multi-task allocation of heterogeneous UAV clusters based on decomposition learning particle swarm, It is characterized in that The learning phase includes the following: Introducing the learning library and the concept of task execution efficiency, for any particle X i , the task execution efficiency of the kth UAV is: In the formula, y k is the rescue penalty of drone k in particle; v j Assemble Tasks for rescue missions k The rescue value of the jth task in ; For each particle to be updated, the task execution efficiency of all contained drones is calculated using the task execution efficiency, and the task rescue sequence with the worst execution efficiency is selected; in the learning stage, each particle updates the worst rescue sequence it contains, and the specific update formula is as follows: in, is the i-th particle in the t-th generation population, for The kth UAV is assigned the execution order of the rescue mission. The kth row of the particle encoding matrix is ​​K, which is the scale of the UAVs called by the particle. target1 is the particle Update the index value of the row in The row with the smallest value; The best facilitator is selected from the learning pool; for The kth UAV is assigned the execution order of the rescue mission, K' is the scale of UAVs called by the optimal particle; for a particle containing K UAVs and M base vehicles To select elements with the same [K,M] value from the learning base as the optimal guide.

7. According to claim 6, the method for multi-task allocation of heterogeneous UAV clusters based on decomposition learning particle swarm, It is characterized in that The contents of the revision phase are as follows: a. Redistribute and re-rank the rescue tasks; define the formula for the task burden of drone k as: Where, vs j Assemble Tasks for rescue missions k The amount of emergency supplies required for the jth rescue mission; S k is the material capacity of drone k; For the rescue task j, let its position be The center position of the set of tasks performed by drone k is center_p k =(center_x k ,center_y k ), the transfer distance of rescue mission j relative to drone k is: Normalize the task burden and the transfer distance value, and define the weighted value of the two as the benefit value of selecting drone k for rescue task j. The formula is as follows: In the formula, ω 1 With ω 2 are the weights of task burden and transfer distance respectively, and ω 1 +ω 2 =1,ω 1 ≥0,ω 2 ≥0; The update formula of the particle correction stage is as follows: In the formula, is the candidate position particle, Sc is the set of drones whose rescue missions have changed, which corresponds to the particle The updated row index value in; Indicates for The reordering result of the tasks assigned to the kth UAV in ; b. Correct the drone type. The specific steps are as follows: (1) Sequentially extract the rescue mission execution sequence of the type 2 UAV in the current particle; (2) Calculate the total amount of rescue materials and total electricity consumed to complete the rescue sequence; (3) If the total amount of rescue materials and the total power are both within the range of type 1 drones, the type of drone in the rescue sequence of the mission is changed to type 1; (4) Repeat (1)-(3) until all rescue mission sequences corresponding to type 2 drones in the particle are checked.

8. According to claim 7, the method for multi-task allocation of heterogeneous UAV clusters based on decomposition learning particle swarm, It is characterized in that The specific contents of S24 are as follows: Suppose the candidate solution selected from the learning base is According to formula (19), we can get The task burden value task_burden of all drones in the task_burden is used to select the drone with the largest task burden and its corresponding rescue mission execution plan The UAV with the smallest mission burden and its corresponding rescue mission execution plan Then, select X a1,max and X a1,min As two chromosomes, a crossover or mutation operation is performed to generate a new solution, which is used to replace the worst particle in the current population.