Optimizing multi-UAV reconnaissance mission allocation method based on unsupervised learning discrete pigeon flock

Abstract

The present invention discloses an unsupervised learning discrete pigeon flock optimization method for multi-UAV reconnaissance mission allocation: Step 1: Establish a UAV swarm model U total ={U1, U2,..., U Nv}; Step 2: Simplify the UAV model based on the Dubins curve assumption; Step 3: Establish a reconnaissance target and a UAV sensor model; Step 4: Flexibly group the data samples T of the reconnaissance target through cosine similarity clustering; Step 5: Design a discrete pigeon flock algorithm based on global coding crossover and mutation; Step 6: Use the discrete pigeon flock algorithm to optimize the fitness function for the multi-UAV Dubins model; Step 7: Output the result graph of the unsupervised learning discrete pigeon flock optimization for multi-UAV reconnaissance mission allocation. The present invention can reduce the track coincidence degree, reduce the calculation time and the number of adjustable variables, and improve the robustness; support task pre-allocation and re-allocation, and has a certain environmental adaptability; has the characteristics of flexible partitioning, enabling the UAV swarm to improve the execution efficiency when performing reconnaissance mission allocation for targets.

Description

Technical Field

[0001] The invention discloses a method for allocating multiple unmanned aerial vehicle (UAV) reconnaissance tasks based on unsupervised learning discrete pigeon group optimization, and belongs to the field of autonomous control of unmanned aerial vehicles. Background Art

[0002] With the rapid development of drone technology and the increasingly complex working environment of drones, drones are playing an increasingly important role in military and civilian applications, such as intelligent search, regional monitoring, environmental detection and rescue missions. However, due to its own size and capabilities, it is difficult for a single drone to complete some complex tasks. Therefore, the collaborative work of multiple drones to complete tasks has become a hot topic. In order to achieve collaborative work, task allocation is a necessary behavior to take the best order when implementing tasks and maximize the benefits of the drone group.

[0003] Usually, the task allocation problem is regarded as a simple traveling salesman problem, which only finds the shortest Euclidean distance between multiple targets, and does not take into account the kinematic constraints of the UAV itself and the reconnaissance range of the sensors carried by the UAV. Considering the kinematic constraints of the UAV itself, the Dubins curve model is introduced into the task allocation model. The Dubins curve is the shortest path connecting two two-dimensional planes under the conditions of satisfying the curvature constraint and the specified start and end directions. It can design a reconnaissance path with the minimum turning radius for the UAV. Considering the reconnaissance range of the sensors carried by the UAV, a sensor model is established. Only when the reconnaissance target is within the reconnaissance range of the sensor can the reconnaissance be considered successful.

[0004] The algorithms for allocating reconnaissance tasks for multiple UAVs can be divided into two categories: distributed algorithms and centralized algorithms. For tasks with a large number of targets and a large UAV cluster, UAVs usually use greedy strategies or local information methods, which will result in a high degree of overlap in task assignment tracks and often face challenges in information consistency and task consistency. Based on the introduction of the Dubins model, the minimum turning radius and maximum range restrictions are considered, and an unsupervised learning strategy is adopted to propose the idea of ​​flexible grouping. In order to minimize the overlap of tracks, a clustering method based on cosine similarity is used to partition the targets to be reconnaissanced by the UAV into different combinations. Then, task assignment and path optimization are performed on multiple UAVs.

[0005] Commonly used methods in task allocation problems include genetic algorithms and particle swarm algorithms, but they are not intelligent enough when achieving target selection, often easily cause overlapping of task tracks, and have disadvantages such as long calculation time, many parameters that need to be adjusted, and low robustness.

[0006] Pigeon Inspired Optimization (PIO) algorithm is an intelligent optimization algorithm designed to simulate the homing behavior of pigeons. It makes it easier for pigeons to return to their nests through three guidance tools: geomagnetic field information, solar altitude information, and landmark information. It can effectively solve a series of problems such as parameter optimization and numerical design, and usually has certain advantages in continuous optimization problems such as function extreme value problems. However, for complex combinatorial optimization problems, its continuity limits its scope of use. The compass operator and landmark operator are discretized and introduced into the combinatorial optimization problem of multi-UAV reconnaissance task allocation by using the method of global coding crossover and mutation. In summary, the present invention proposes a multi-UAV reconnaissance task allocation method based on unsupervised learning discrete pigeon group optimization, the purpose of which is to be able to perform task allocation on a multi-UAV group when determining the orientation and position of the target, and obtain the optimal coverage and occupancy of UAV reconnaissance, so as to scout the target under the condition of maximum effect-cost ratio. Summary of the invention

[0007] 1. Purpose of the invention:

[0008] The present invention provides a method for allocating multi-UAV reconnaissance tasks based on unsupervised learning by optimizing discrete pigeon groups. The purpose of the method is to allocate tasks to a multi-UAV group when determining the orientation and position of a target, so as to obtain the optimal coverage and occupancy of UAV reconnaissance, thereby reconnaissance the target with the maximum cost-effectiveness.

[0009] 2. Technical solution:

[0010] The present invention aims at the problem of multi-UAV reconnaissance task allocation and provides a multi-UAV reconnaissance task allocation method based on unsupervised learning discrete pigeon group optimization. The specific steps of the method are as follows:

[0011] Step 1: Establish the UAV swarm model U total = {U 1 ,U 2 ,...,U Nv}

[0012] It is assumed that the UAV U Nv Equipped with speed, heading angle and altitude, the drone model can be simplified to the following 6-state model:

[0013]

[0014] Also consider the practical constraints of drones:

[0015]

[0016] Where: is the position of UAV i, V t , ψi and λ i are the horizontal velocity, heading angle and altitude change rate respectively, and are the control inputs on the three loops of the autopilot, τ v , τ ψ and (τ λ ,τ h ) are three time constants, Nv represents the number of drones, V max ,V min ,n max and λ max are all greater than 0, which are the maximum speed, minimum horizontal speed, maximum lateral overload and maximum height change rate, respectively, min is the minimum height change rate, and is less than 0, g is the gravitational acceleration, which is 10m / s 2 .

[0017] Step 2: Simplify the UAV model based on the Dubins curve hypothesis

[0018] 1) The drone swarm has no combat damage while performing the mission;

[0019] 2) The altitude and speed of the drone swarm when performing reconnaissance missions are fixed;

[0020] 3) Each drone will operate at a different altitude;

[0021] 4) Each drone has sufficient payload reserves.

[0022] Based on the above assumptions, the drone in step 1 is simplified into three state quantities Q = (x, y, ψ). For each drone U Nv , its motion model can be expressed as:

[0023]

[0024] In the formula, x and y represent the position of the drone, v Nv It indicates the speed of the drone Nv. It represents the minimum turning radius, c represents the control input, |c|≤1, when c>0, it means the UAV turns counterclockwise; when c<0, it means the UAV turns clockwise; when c=0, it means the UAV goes straight in the original direction.

[0025] Set the Dubins curve to have the following motion:

[0026] D={LSL,RSR,RSL,LSR,RLR,LRL} (4)

[0027] In the formula, D represents the set of Dubins paths that can be selected, L represents counterclockwise rotation, R represents clockwise rotation, and S represents going straight in the original direction.

[0028] Step 3: Build reconnaissance target and drone sensor models

[0029] The target of this task assignment is a stationary target T total ={T 1 ,T 2 ,...,T Nu}, the radius of the target is r, the drone's sensor reconnaissance range can be assumed to be a circular range, and the field of view radius is R. When the target point is within the field of view, it means that the drone has performed reconnaissance on it. At this time, the following formula is true:

[0030] Rr<l (5)

[0031] In the formula, R and r are the reconnaissance radius of the drone sensor and the radius of the target point respectively, and l represents the allowable range difference. The specific conceptual model can be found in Figure 1a , b, simulation diagram see Figure 3 .

[0032] Step 4: Flexibly group the data samples T of the reconnaissance target by cosine similarity clustering

[0033] Randomly generate Nv points within the data size range as the initial cluster center points {C 1 ,C 2 ,...,C k}, 1<k≤Nv. In the data sample T, there are Nu objects T={T 1 ,T 2 ,...,T Nu}, each object will have m-dimensional features. In order to reduce track overlap and reduce iteration time, K-Means is used to cluster Nu objects into specified Nv clusters based on the cosine similarity between objects. Each object can only exist in one of the specified Nv clusters. Then the cosine value from each object to each cluster center is calculated as follows:

[0034]

[0035] Where, T i The i-th object represented by 1≤i≤Nv, C k It represents the kth cluster center 1≤k≤Nv, T i m represents the mth feature of the i-th object. represents the mth feature of the kth object. Through continuous iteration and redistribution into clusters, k clusters {S 1 ,S 2 ,...,S k}, 1≤k≤Nv. After all data samples are allocated, the clusters are calculated using the cluster center calculation formula, and then the data samples and cluster centers are iteratively allocated again until a certain number of iterations is reached or the change in cluster center is very small, at which time data can be generated.

[0036] The cluster center can be obtained using the following formula:

[0037]

[0038] In the formula, C k represents the cluster center, T i It means that it belongs to cluster S k The i-th object, num(S k ) represents the cluster S k , ΔC represents the difference between the center point of the k-th cluster and the center point of the k-1-th cluster, I represents the number of iterations, I max It represents the maximum number of iterations. The simulation diagram of flexible grouping of target points is shown in Figure 4 .

[0039] Step 5: Design of discrete pigeon flock algorithm based on global coding crossover and mutation

[0040] 1) Discrete Pigeon Swarm Algorithm Applied to Combinatorial Optimization Problems

[0041] The pigeon flock algorithm is an intelligent optimization algorithm designed to simulate the homing behavior of pigeons. It makes it easier for pigeons to return to their nests through three guidance tools: geomagnetic field information, solar altitude information, and landmark information. It can effectively solve a series of problems such as parameter optimization and numerical design, and usually has certain advantages in continuous optimization problems such as function extreme value problems. However, for complex combinatorial optimization problems, its continuity limits its scope of use. The global coding crossover and mutation method is used to introduce the discretization of compass operators and landmark operators into the combinatorial optimization problem of multi-UAV reconnaissance mission allocation.

[0042] 2) Initialization

[0043] The number of pigeons in the flock is P, the number of tasks to be assigned to the reconnaissance mission is N, Q1 represents the number of geomagnetic navigations, Q2 represents the number of landmark navigations, and the position of the i-th pigeon is X i =[x 1 ,x 2 ,...,x n], 1≤i≤P represents a sequence of UAV reconnaissance mission assignments, where 1≤x n ≤N,1≤n≤N and formula (8) holds

[0044]

[0045] The speed of the i-th pigeon is V i =[v 1 ,v 2 ,...,v n ], 1≤i≤P represents a period that makes the pigeon's position X i t Get another position The transfer sequence, where 0≤v n ≤N, 1≤n≤N.

[0046] 3) Design of map and compass operators based on global coding crossover and mutation

[0047] The iterative formula of the continuous map compass operator is as follows:

[0048]

[0049] In the formula, the position of the i-th pigeon is represented by X i , speed is represented by V i , R is the map compass operator, t represents the number of iterations, rand is a random number between [0, 1], X g It represents the global optimal position.

[0050] The update formula of the compass operator using global encoding crossover and mutation is as follows:

[0051]

[0052]

[0053]

[0054] In the formula, t represents the number of iterations, T t represents the matrix of the global optimal sequence, M t It represents the sequence matrix of the individual at this time, C t It means from M t The matrix becomes T t The operator of the matrix, N represents the number of tasks, i.e., the dimension, α and β are both random numbers in [0, 1], R is the geomagnetic influence factor that has a higher global optimization ability at the beginning and a stronger local optimization ability at the end as the number of iterations increases, and the other expressions are consistent with those in formula (8).

[0055] 4) Landmark operator design based on global encoding crossover and mutation

[0056] The iterative formula of the continuous landmark operator is as follows:

[0057]

[0058] Where N p represents the number of individuals in the population, t represents the number of iterations, and X c represents the position of the central pigeon after each iteration, X i represents the position of the i-th pigeon, f(X i ) represents X i At this time, the fitness function, rand is a random number on [0,1].

[0059] The landmark operator update formula using global encoding cross-mutation is as follows:

[0060] P(t)=P(t-1) / 2 (14)

[0061]

[0062]

[0063] In the formula, ceil(·) is the rounding function, O(·) is a missing integer complement operator, which means that when When the element in the sequence is 0, use The middle is not equal to the original The data completion sequence in is, γ is a random number in [0,1], and the other expressions are consistent with formulas (11) and (12). The overall algorithm framework diagram is shown in Figure 2 .

[0064] Step 6: Multi-UAV Dubins model uses discrete pigeon flock algorithm to optimize the fitness function

[0065] From step 4, we get the k cluster sequence S for unsupervised learning i ={S 1 ,S 2 ,...,S k Each cluster sequence is considered as a Dubins traveling salesman problem for pre-allocation. The optimization algorithm ensures that when making path selection for each target in each cluster, the one with the smallest path length is selected from the set D = {LSL, RSR, RSL, LSR, RLR, LRL} and finally returns to the starting point.

[0066]

[0067] In the formula, Nu represents the number of target points. It means S i The Dubins distance from the cluster sequence to point j, T i It represents the cluster S i The time taken to complete the reconnaissance.

[0068] Step 7: Output the unsupervised learning discrete pigeon flock optimization multi-UAV reconnaissance task allocation result graph

[0069] Determine whether the number of iterations of unsupervised learning has reached N max Or the cluster center reaches stability, and then determine whether the number of iterations of discrete pigeon group optimization reaches Q1+Q2. If so, the output of the multi-UAV reconnaissance task allocation result is shown in the figure. Figure 6 , Figure 8 , compared with the simple random selection of genetic algorithm, see Figure 5 , Figure 7 ; Otherwise, update the number of iterations and go to step 2.

[0070] The advantages and effects of the multi-UAV reconnaissance task allocation method based on unsupervised learning discrete pigeon group optimization of the present invention are: first, it provides an optimization algorithm for solving the multi-UAV reconnaissance task allocation, reduces track overlap, reduces calculation time and the number of adjustable variables, and improves robustness; second, it proposes a dynamic allocation method that supports task pre-allocation and reallocation and has certain environmental adaptability; third, it has the characteristics of flexible partitioning, which can improve the execution efficiency when the UAV group allocates reconnaissance tasks to the target. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1a , b Schematic diagram of target model and sensor model

[0072] Figure 2 The present invention is based on unsupervised learning discrete pigeon group optimization algorithm flow chart

[0073] Figure 3 Schematic diagram of multi-UAV reconnaissance mission allocation target model simulation

[0074] Figure 4 Simulation diagram of flexible grouping model for multi-UAV reconnaissance mission allocation

[0075] Figure 5 Schematic diagram of genetic algorithm simple random selection simulation (top view)

[0076] Figure 6 Schematic diagram of unsupervised learning discrete pigeon group optimization simulation in an embodiment of the present invention (top view)

[0077] Figure 7 Schematic diagram of genetic algorithm simple random selection simulation (side view)

[0078] Figure 8Schematic diagram of unsupervised learning discrete pigeon group optimization simulation in an embodiment of the present invention (top view)

[0079] Fig. 9 Schematic diagram of multi-UAV fitness function value simulation

[0080] The numbers and symbols in the figure are explained as follows:

[0081] XYZ——The geodetic coordinate system of the drone relative to the ground

[0082] H——The flight altitude of the drone when performing reconnaissance missions

[0083] R——The radius of the drone sensor model’s reconnaissance range

[0084] r——Radius of the reconnaissance target

[0085] T Nu ——The serial number of the reconnaissance target DETAILED DESCRIPTION

[0086] The effectiveness of the method proposed in the present invention is verified by a specific example of multi-UAV reconnaissance task allocation. The experimental computer is configured with an Intel Core i7-10700 processor, 2.90Ghz main frequency, 16G memory, and the software is MATLAB 2020b version. A method for multi-UAV reconnaissance task allocation based on unsupervised learning discrete pigeon group optimization has the following specific steps:

[0087] Step 1: Establish the UAV swarm model U total = {U 1 ,U 2 ,...,U Nv}

[0088] The initial heading angles of the 10 drones are all 0°, the initial horizontal and vertical coordinate positions are [0,0] and the heights are [22:2:40] respectively, the speed is constant at 10m / s, the minimum turning radius is 5m, the simulation step is 0.01s, and the simulation maximum threshold is 100s.

[0089] Step 2: Simplify the UAV model based on the Dubins curve hypothesis

[0090] Consider the six-degree-of-freedom model of the drone swarm in step 1, which is simplified to a three-degree-of-freedom drone model through the Dubins curve, and a Dubins distance calculation function is generated. The selection strategies are:

[0091] D={LSL,RSR,RSL,LSR,RLR,LRL}

[0092] Each selection strategy represents the distance between two target points, and the distance between target points under all strategies is calculated using the D formula.

[0093] Step 3: Build reconnaissance target and drone sensor models

[0094] Randomly generate 50 horizontal and vertical target point data within the horizontal axis [-200, 200] and the vertical axis [-200, 200], with a radius of r = 2.5m, and the detection radius of the drone sensor model is R = 5m, l max =2.5m.

[0095] Step 4: Flexibly group the data samples T of the reconnaissance target by cosine similarity clustering

[0096] According to step 1 and step 3, 10 cluster center points are randomly generated for the 50 target points within the range. Each target point and the center point have two dimensions, horizontal and vertical. The cosine similarity of each point to the cluster center point is calculated according to formula (6). The data are divided into 10 groups according to their size. Then, the cluster center of each group of data is recalculated by formula (7) until the cluster center point is stable, that is, ΔC approaches 0 or the number of iterations I reaches 10000. The grouping is completed, and 10 groups {S 1 ,S 2 ,...,S 10}.

[0097] Step 5: Design of discrete pigeon flock algorithm based on global coding crossover and mutation

[0098] 1) Initialization

[0099] The total number of pigeons is 30, the number of tasks to be assigned to the reconnaissance mission is 50, the number of geomagnetic navigation times Q1 = 150, the number of landmark navigation times Q2 = 50, and the position X of the i-th pigeon i =[x 1 ,x 2 ,...,x n ], speed is V i =[v 1 ,v 2 ,...,v n ], the algorithm needs to perform optimization calculations on 10 groups of target points in parallel.

[0100] 2) Design of map and compass operators based on global coding crossover and mutation

[0101] Find the optimal individual X when the number of iterations is t g , expanding it to Then, Instead, find out t (Ct )=T t The transfer matrix Keep C by random probability mutation t The elements in each row and column of .

[0102] Then the geomagnetic influence factor R is designed so that it has a strong global search capability in the early stage of iteration and a strong local search capability in the later stage of iteration. -Rt The probability of continuing mutation and retaining C t The elements in each row and column get V t .

[0103] Finally, the retrieved transfer matrix V t Substituting into formula (12) we get the new M t Individual sequence matrix.

[0104] 3) Landmark operator design based on global encoding crossover and mutation

[0105] Because the product of the fitness function value used to calculate the position of the central pigeon is only applicable to the continuous pigeon group algorithm, it is necessary to find the approximate central pigeon in the discrete algorithm and optimize the landmark operator. First, the number of populations is sorted by fitness using equation (14), and half of the individuals with larger fitness values ​​are subtracted. Then, the average of all the optimal pigeons of each generation that have been calculated is calculated and the integer is obtained. It may not meet the requirements of the target sequence. The solution is to use formula (15) to convert the out-of-range and those with repeated data Set to 0, and get Then fill in the missing positions from the first column with the missing position variables to get the complete target sequence X c .

[0106] At this time, X c Expand to The form of Make a comparison and find out t (C t )=T t The transfer matrix Keep C by random probability mutation t The elements in each row and column of .

[0107] Finally, the retrieved transfer matrix C t Substituting into formula (16) we get the new M t Individual sequence matrix.

[0108] Step 6: Multi-UAV Dubins model uses discrete pigeon flock algorithm to optimize the fitness function

[0109] For the k cluster sequences S that have been unsupervised learning i ={S 1 ,S 2 ,...,S k Each cluster sequence is considered as a DTSP problem for pre-allocation. The discrete pigeon flock algorithm ensures that when making path selection for each target in each cluster, the one with the smallest path length is selected from the set D = {LSL, RSR, RSL, LSR, RLR, LRL} and finally returns to the starting point.

[0110]

[0111] In the formula, Nu represents the number of target points. It means S i The Dubins distance from the cluster sequence to point j, T i It represents the cluster S i The time taken to complete the reconnaissance.

[0112] Step 7: Output the unsupervised learning discrete pigeon flock optimization multi-UAV reconnaissance task allocation result graph

[0113] Determine whether the number of iterations of unsupervised learning reaches 10,000 or the cluster center reaches stability, and then determine whether the number of iterations of discrete pigeon group optimization reaches 200. If so, output the multi-UAV reconnaissance task allocation result graph; otherwise, update the number of iterations and go to step 2.

[0114] The present invention adopts 10 UAVs and 50 target points. First, the target points are grouped by unsupervised learning method. Then, the UAV Dubins model is considered to fly at a constant speed at different altitudes with a minimum turning radius. The discrete pigeon flock algorithm is used to optimize the reconnaissance task allocation of 50 target points in terms of path and time.

[0115] In order to verify the effectiveness of the method proposed in this invention, this patent also conducted corresponding comparative experiments, comparing it with the simple selection of traditional genetic algorithms, and the changes in the fitness function value are as follows: Fig. 9 shown.

Claims

1. A method for optimizing multi-UAV reconnaissance mission allocation based on unsupervised learning discrete pigeon flocks. Features: The specific steps of this method are as follows: Step 1: Establish the UAV swarm model U total = {U 1 ,U 2 ,...,U Nv } Step 2: Simplify the UAV model based on the Dubins curve hypothesis Step 3: Build reconnaissance target and drone sensor models The target of this task assignment is a stationary target T total ={T 1 ,T 2 ,...,T Nu }, the radius of the target is r, the reconnaissance range of the drone's sensor can be assumed to be a circular range, and the field of view radius is R; when the target point is within the field of view, it means that the drone has performed reconnaissance on it; at this time, the following formula is established: R-r＜l In the formula, R and r are the reconnaissance radius of the UAV sensor and the radius of the target point respectively, and l represents the allowable range difference; Step 4: Flexibly group the data samples T of the reconnaissance target by cosine similarity clustering Step 5: Design of discrete pigeon flock algorithm based on global coding crossover and mutation Step 6: Multi-UAV Dubins model uses discrete pigeon flock algorithm to optimize the fitness function From step 4, we get the k cluster sequence S for unsupervised learning i ={S 1 ,S 2 ,...,S k Each cluster sequence is considered as a Dubins traveling salesman problem for pre-allocation. The optimization algorithm ensures that when selecting a path for each target in each cluster, the path with the smallest length is selected from the set D = {LSL, RSR, RSL, LSR, RLR, LRL}, and finally returns to the starting point; In the formula, Nu represents the number of target points. It means S i The Dubins distance from the cluster sequence to point j, T i It represents the cluster S i The time taken to complete the reconnaissance; Step 7: Output the unsupervised learning discrete pigeon flock optimization multi-UAV reconnaissance task allocation result graph Determine whether the number of iterations of unsupervised learning has reached N max Or the cluster center reaches stability, and then determine whether the number of iterations of discrete pigeon group optimization reaches Q1+Q2; if so, output the multi-UAV reconnaissance task allocation result graph; otherwise, update the number of iterations and go to step 2.

2. According to claim 1, a method for allocating multiple UAV reconnaissance tasks based on unsupervised learning discrete pigeon group optimization, Features: In step 1, it is assumed that the drone U Nv Equipped with speed, heading angle and altitude, the drone model can be simplified to the following 6-state model: Also consider the practical constraints of drones: Where: is the position of UAV i, V t , ψ i and λ i are the horizontal velocity, heading angle and altitude change rate respectively, and are the control inputs on the three loops of the autopilot, τ v , τ ψ and (τ λ ,τ h ) are three time constants, Nv represents the number of drones, V max ,V min ,n max and λ max are all greater than 0, which are the maximum speed, minimum horizontal speed, maximum lateral overload and maximum height change rate, respectively, min is the minimum height change rate, and is less than 0, g is the gravitational acceleration, which is 10m / s 2 .

3. According to claim 1, a method for allocating multiple UAV reconnaissance tasks based on unsupervised learning discrete pigeon group optimization, Features: The specific process of step 2 is as follows: Assumption: 1) The drone swarm has no combat damage while performing the mission; 2) The altitude and speed of the drone swarm when performing reconnaissance missions are fixed; 3) Each drone will operate at a different altitude; 4) Each drone should have sufficient payload reserves; Based on the above assumptions, the drone in step 1 is simplified into three state quantities Q = (x, y, ψ). For each drone U Nv , its motion model can be expressed as: In the formula, x and y represent the position of the drone, v Nv It indicates the speed of the drone Nv. It represents the minimum turning radius, c represents the control input, |c|≤1, when c>0, it means the drone turns counterclockwise; When c < 0, it means the drone turns clockwise; When c = 0, it means that the drone moves straight along the original direction; Set the Dubins curve to have the following motion: D={LSL,RSR,RSL,LSR,RLR,LRL} (4) In the formula, D represents the set of Dubins paths that can be selected, L represents counterclockwise rotation, R represents clockwise rotation, and S represents going straight in the original direction.

4. According to claim 1, a method for allocating multiple UAV reconnaissance tasks based on unsupervised learning discrete pigeon group optimization, Features: The specific process of step 4 is as follows: Randomly generate Nv points within the data size range as the initial cluster center points {C 1 ,C 2 ,...,C k }, 1<k≤Nv; In the data sample T, there are Nu objects T={T 1 ,T 2 ,...,T Nu }, each of which will have m-dimensional features; in order to reduce track overlap and reduce iteration time, K-Means is used to cluster Nu objects into specified Nv clusters based on the cosine similarity between objects. Each object can only exist in one of the specified Nv clusters; then the cosine value from each object to each cluster center is calculated as follows: Where, T i The i-th object represented by 1≤i≤Nv, C k It represents the kth cluster center 1≤k≤Nv, T i m represents the mth feature of the i-th object. represents the mth feature of the kth object; through continuous iteration and redistribution into clusters, k clusters {S 1 ,S 2 ,...,S k }, 1≤k≤Nv; After all data samples are allocated, the allocated clusters are calculated using the cluster center calculation formula, and then the data samples and cluster centers are iteratively allocated again; The cluster center can be obtained using the following formula: In the formula, C k represents the cluster center, T i It means that it belongs to cluster S k The i-th object, num(S k ) represents the cluster S k , ΔC represents the difference between the center point of the k-th cluster and the center point of the k-1-th cluster, I represents the number of iterations, I max It represents the maximum number of iterations.

5. According to claim 1, a method for allocating multiple UAV reconnaissance tasks based on unsupervised learning discrete pigeon group optimization, Features: The specific process of step five is as follows: 1) Initialization The number of pigeons in the flock is P, the number of tasks to be assigned to the reconnaissance mission is N, Q1 represents the number of geomagnetic navigations, Q2 represents the number of landmark navigations, and the position of the i-th pigeon is X i =[x 1 ,x 2 ,...,x n ], 1≤i≤P represents a sequence of UAV reconnaissance mission assignments, where 1≤x n ≤N,1≤n≤N and formula (8) holds The speed of the i-th pigeon is V i =[v 1 ,v 2 ,...,v n ], 1≤i≤P represents a section that makes the pigeon's position X i t Get another position The transfer sequence, where 0≤v n ≤N, 1≤n≤N; 2) Design of map and compass operators based on global coding crossover and mutation The iterative formula of the continuous map compass operator is as follows: In the formula, the position of the i-th pigeon is represented by X i , speed is represented by V i , R is the map compass operator, t represents the number of iterations, rand is a random number between [0, 1], X g It represents the global optimal position; The update formula of the compass operator using global encoding crossover and mutation is as follows: In the formula, t represents the number of iterations, T t represents the matrix of the global optimal sequence, M t It represents the sequence matrix of the individual at this time, C t It means from M t The matrix becomes T t The operator of the matrix, N represents the number of tasks, i.e., the dimension, α and β are both random numbers in [0,1], R is the geomagnetic influence factor that has a higher global optimization ability at the beginning and a stronger local optimization ability at the end as the number of iterations increases, and the other expressions are consistent with those in formula (8); 3) Landmark operator design based on global encoding crossover and mutation The iterative formula of the continuous landmark operator is as follows: Where N p represents the number of individuals in the population, t represents the number of iterations, and X c represents the position of the central pigeon after each iteration, X i represents the position of the i-th pigeon, f(X i ) represents X i The fitness function at this time, rand is a random number on [0,1]; The landmark operator update formula using global encoding cross-mutation is as follows: P(t)=P(t-1) / 2 (14) In the formula, ceil(·) is the rounding function, O(·) is a missing integer complement operator, which means that when When the element in the sequence is 0, use The middle is not equal to the original The data completion sequence in is , γ is a random number in [0,1], and the other expressions are consistent with formulas (11) and (12).

Images