Unmanned aerial vehicle group investigation force scheduling method based on cooperative computing

Abstract

The invention discloses an unmanned aerial vehicle group detection capability scheduling method based on cooperative computing, and aims to solve the problem of how to realize information intercommunication and cluster resource comprehensive cooperative distribution in a cooperative detection process without depending on a ground mobile base station and a cloud computing server. A cluster information collaboration model is established, a cluster collaboration computing alliance is established, cluster detection and computing resources are reasonably coordinated through mutual communication between unmanned aerial vehicles, and the cluster detection capability optimization method is realized. The method comprises the steps that a system optimizes an unmanned aerial vehicle detection position based on a task search vector in a detection topology optimization layer, a system alternating task unloading proportion and a computing resource allocation proportion in a resource optimization layer give a detection force under a current detection topology, and the system performs resource optimization according to the feedback of the detection force. And continuously optimizing the detection topology optimization layer and the resource allocation optimization layer, and finally obtaining an optimal detection topology and resource allocation scheme of the cluster to realize efficient detection.

Description

Technical Field

[0001] The present invention relates to swarm intelligence technology, unmanned aerial vehicle (UAV) cluster communication, unmanned aerial vehicle (UAV) collaborative computing, unmanned aerial vehicle (UAV) collaborative perception, and unmanned aerial vehicle (UAV) resource allocation, and in particular to a method for scheduling detection and computational power in an unmanned aerial vehicle (UAV) swarm collaborative computing and computational detection system. Background Art

[0002] The use of drones has become the norm in low-altitude disaster detection. The use of drones enables rescue workers to quickly understand the damage in the disaster area and quickly launch search and rescue work through a human-machine combined collection mode. However, the current drone detection mode still mainly relies on a single drone for non-intelligent detection operations. The detection data is usually transmitted back to the command center through a wireless network or optical fiber connection for processing, and then the results are transmitted to the search and rescue personnel terminal. This mode limits the flexibility of drones to a certain extent, and lacks intelligent processing capabilities, making it difficult to fully meet the diverse needs in complex environments.

[0003] With the rapid development and innovation of drone technology, the importance of drones in the field of disaster detection has become increasingly prominent. In this context, due to its limited detection and computing resources, a single drone is often unable to perform tasks such as efficient detection of large areas and large-scale data processing alone. For this reason, the detection style of drones is changing, and drone cluster collaborative detection based on collaborative computing has become an important trend in the future of drone intelligent detection. How to determine the detection position of each drone and allocate the computing resources of each drone is of irreplaceable importance for drone clustering and intelligent detection. Summary of the invention

[0004] The purpose of the present invention is to address the defects and shortcomings of the above-mentioned prior art and propose a method for scheduling computing power for drone cluster detection based on collaborative computing. The method is applied to future intelligent drone cluster detection scenarios.

[0005] The technical solution adopted by the present invention to solve the technical problem is: a method for scheduling the detection and computing power of a group of drones based on collaborative computing, which includes five overall steps from S1 to S5:

[0006] S1. Establish an adaptive grid map based on initial prior information and obtain the task search vector

[0007] S2. Analyze the composition of drone detection vectors, including drone detection range model, drone probability detection model, task search vector update model, and establish a drone cluster information coordination model.

[0008] S3. Analyze the composition of drone computing vectors, including drone collaborative computing attributes, cluster collaborative computing alliance, task offloading ratio matrix and computing resource allocation matrix.

[0009] S4. Analyze the composition of the group detection computing power, including the cluster detection power and the cluster task completion delay, and obtain the calculation method of the group detection computing power.

[0010] S5. With the goal of maximizing the swarm detection computing power, a system optimization model is established to jointly optimize the drone cluster detection position and cluster resource allocation through the detection swarm intelligent optimization algorithm.

[0011] The following is a detailed description of the above five steps one by one:

[0012] S1 builds an adaptive grid map, that is, the steps of building an adaptive grid map based on prior information are as follows:

[0013] S1.1. Based on the detection data of low-orbit satellites, obtain the geographical environment information of the target area M = {x M ,y M ,z M}, the number of target points to be detected n target and the target location information R = {x R ,y R ,z R}.

[0014] S1.2, Adaptive rasterized map G = {g 1 ,g 2 ,…,g i}: Divide the task area into grid maps of different grid sizes according to the detection weight. Divide the target area into four grids and calculate the grid detection weight:

[0015]

[0016] in, Indicates the importance of the target to be detected in the grid. represents the complexity of the terrain, α w ,β w is the corresponding coefficient. th The grid is further divided into four sub-grids, and the above process is repeated until no further division is possible.

[0017] S1.3. Task Search Vector The task search vector consists of the grid target existence probability and grid uncertainty. Indicates that drone j detects grid g i The probability of the target existing, grid uncertainty Indicates that drone j detects grid g iIn view of the uncertainty of the grid target, the calculation methods of grid target existence probability and grid uncertainty are introduced below.

[0018] (1) Grid target existence probability: Initial grid target existence probability The grid detection weight W i Normalized to get:

[0019]

[0020] in, η represents the accuracy of prior information, η∈[0,1].

[0021] (2) The grid uncertainty is defined as the information entropy of the probability of the target existing in the grid:

[0022]

[0023] S2 UAV detection vector calculation method

[0024] Drone detection vector ZC j ={vg j ,(PD j ,PF j ),Y j}, where vg j represents the detectable grid set of UAV j, PD j Indicates the probability of successful detection, PF j represents the error detection probability. The success detection probability and the error detection probability are collectively referred to as the drone probability detection model; Y j represents the task search vector of UAV j. The calculation method of each part is introduced in turn below.

[0025] S2.1. UAV detection range model: Set the center point of the grid as the information point of the grid. If the grid information point is within the detection range of the drone and is not blocked by obstacles such as mountains, then the grid is in the detectable grid set vg of the drone. j Inside.

[0026] S2.2, UAV probabilistic detection model:

[0027] The probability of successful detection indicates the probability that the UAV system successfully determines the existence of the target when the target exists, and is defined as an exponential function of the square of the height difference between the UAV and the detectable grid:

[0028]

[0029] In the formula, ω s is the sensitivity of successful detection of UAV; s is the successful detection coefficient of the drone; Indicates the grid g i Whether there is actually a target in Indicates that there is no target in the grid gi. Represents grid g i There is actually a goal in Indicates the specific results of drone detection. Represents that at time t, drone j detects grid g i There is a goal, when Indicates that the drone detects grid g i There is no goal; Represents drone j and grid g i The height difference between .

[0030] The error detection probability indicates the probability that the UAV system incorrectly judges the existence of a target when the target does not exist:

[0031]

[0032] In the formula, ω f is the error detection sensitivity of the UAV, κ f is the error detection coefficient of the UAV.

[0033] S2.3. Task Search Vector Update Model

[0034] Based on the information detected by the drone (i.e. PD and PF), the Bayesian criterion is used to dynamically update the target existence probability to reflect the current distribution of targets.

[0035] When drone j detects grid g i When there is a target in At this time, the probability of the grid target existing is updated as:

[0036]

[0037] When drone j detects grid g i When there is no target in At this time, the probability of the grid target existing is updated as:

[0038]

[0039] At the same time, based on formula (3), the grid uncertainty is updated according to the new target existence probability. Set the grid g i The conditions for detection completion are:

[0040]

[0041] Therefore, combined with formula (3), the grid uncertainty threshold o is obtainedth = 0.0808, when the uncertainty of the grid is lower than o th , it indicates that the grid detection is completed.

[0042] S2.4 Cluster Information Collaboration Model

[0043] During the cluster collaborative detection process, each drone independently completes its own task search vector In order to quickly and accurately complete the detection of the cluster and avoid the drones from repeatedly detecting the grids that have been detected, the drones in the cluster need to obtain the global task search vector information based on the task search vectors of other drones. Therefore, a cluster information coordination model is established.

[0044] From formulas (6) and (7), we can see that the simultaneous detection of k drones can be expressed as:

[0045]

[0046] In order to simplify the calculation, formula (9) is transformed into:

[0047]

[0048] Then, the logarithmic transformation of formula (10) is:

[0049]

[0050] make Therefore, formula (11) can be transformed into:

[0051]

[0052] in,

[0053] From formula (12), we can see that through the cluster information coordination model, the drone only needs to receive the probability offset transmitted by other drones The global search vector information can be obtained. Since the communication only involves the transmission probability offset, the communication delay can be ignored.

[0054] S3, drone calculation vector

[0055] The drone calculates the vector JS = {(xt,AC),λ,f r}, where xt represents the collaborative computing attributes of the UAV; AC represents the collaborative computing alliance matrix; λ represents the task offloading ratio matrix; f r The table calculates the resource allocation ratio matrix. The calculation method of each part is introduced in turn below.

[0056] S3.1. Collaborative calculation of attributes xt

[0057] Non-Cooperative Unmanned Aerial Vehicle (NC-UAV): When the load is high, it needs to offload part of its computing tasks to cooperative UAVs, and can offload to multiple cooperative UAVs in a certain proportion, so that multiple cooperative UAVs can coordinate computing to improve the speed of task completion. Set the cooperative computing attribute xt=0 of non-cooperative UAVs.

[0058] Cooperative Unmanned Aerial Vehicle (C-UAV): When its own task load is low, it can receive task data from non-cooperative UAVs and help them perform collaborative computing. Set the collaborative computing attribute xt=1 for cooperative UAVs.

[0059] S3.2, Cluster Collaborative Computing Alliance AC: Considering that there may be multiple non-cooperative drones in the cluster, all cooperative drones can simultaneously join the computing alliance of multiple non-cooperative drones to participate in task computing. Therefore, the formation method of the collaborative computing alliance AC is as follows:

[0060]

[0061] Among them, AC ij =1 represents that drone j joins the collaborative computing alliance of task i.

[0062] S3.3, Task offloading ratio matrix: Define the task offloading ratio matrix λ = [λ ij ] N×N (λ ij ∈[0,1]), λ ij represents the proportion of tasks that drone i will offload from its own task i to drone j, if drone j is not in the collaborative computing alliance AC of task i i In the ij =0, and it is necessary to ensure Therefore, the task offloading ratio matrix initialization method is as follows:

[0063]

[0064] In the formula, represents the number of drones in the alliance for task i.

[0065] S3.4, computing resource allocation matrix: define the remaining computing resources of drone i as f i , the computing resource allocation ratio matrix is in It means that drone i will use its remaining computing resources f i The proportion assigned to task j must satisfy To ensure that computing resources are not over-allocated, the computing resource allocation ratio matrix is ​​initialized as follows:

[0066]

[0067] In the formula, Represents the number of collaborative computing alliances that UAV j participates in.

[0068] S4. Group detection computing power calculation method

[0069] The group detection computing power is mainly composed of the cluster detection capability and the cluster task completion delay. The calculation methods of each part are described in turn below.

[0070] S4.1. Cluster detection capability

[0071] The cluster detection capability is related to the change of uncertainty in the mission area and the number of detected targets, which can be expressed as:

[0072]

[0073]

[0074] Where μ represents the weight factor of uncertainty change, n F Indicates the number of targets detected by the cluster, n g represents the number of grids in the task area, Ο represents the initial uncertainty of the task area, represents the initial uncertainty of each grid in the mission area, Ο' represents the uncertainty of the mission area after detection is completed, Represents the uncertainty of each grid in the mission area after detection.

[0075] S4.2. Cluster task completion delay model:

[0076] The cluster task completion delay is mainly composed of the local computing delay and collaborative computing delay of each subtask. The calculation method of each delay is introduced in turn below.

[0077] Local computation latency of task i:

[0078]

[0079] In the formula, λ i,i It means that UAV i will calculate the amount C of task i i The ratio of uninstalls to itself, It means that drone i will use its remaining computing power f i The proportion of the allocation allocated to itself.

[0080]

[0081] In the formula, φi represents the CPU cycles required to calculate 1 bit of data of task i, I i represents the amount of data for task i, Represents the grid g k Area, g k ∈vg i ; Represents the visible grid set vg of drone i i Total area; Indicates vg i Number of grids in data s Indicates the amount of data generated by a drone detection; Represents the distance between drone i and grid g k The number of detections.

[0082] When drone i cannot complete the computation of task i alone, it is necessary to split task i into multiple sub-computing tasks and offload them to the collaborative drones in the collaborative computing alliance for computation. The collaborative computing delay of task i includes the transmission delay of the sub-tasks and the computation delay of the sub-tasks:

[0083]

[0084] Communication rate i,j Calculated by formula (21):

[0085]

[0086] in, It means that drone i distributes the bandwidth equally to all cooperative drones in the task i alliance. represents the transmission power of UAV i, N 0 is the noise power spectral density, g 0 is the channel power gain at the reference distance, d i,j represents the distance between drone i and the coordinated drone j in the alliance. i,j →0: Rate i,j →∞, however, this situation is unlikely to happen in reality. Considering that the infinite approach of drones may cause collisions and other safety accidents, the hovering distance of all drones should be greater than a certain threshold d safe Therefore, the completion delay of task i is:

[0087] t i =max(t loc,i ,t col,i ) (twenty two)

[0088] All tasks are calculated in parallel, so the total task completion delay T is:

[0089] T=max(ti ) (twenty three)

[0090] S5. The steps for establishing the system optimization model are as follows:

[0091] The group detection computing power U can be expressed as:

[0092]

[0093] The optimization goal is to maximize the group detection computing power:

[0094] max U(25) combined with the constraints yields:

[0095]

[0096] In the formula, constraints 1, 2, and 3 represent the UAV detection position constraints, which means that the UAV must detect within the mission area; constraint 4 indicates that the hovering distance duav of the UAV in the cluster should be greater than a certain threshold dsafe; constraints 5 and 6 represent the unloading ratio constraints of the UAV collaborative computing alliance; constraints 7 and 8 represent the computing resource allocation ratio constraints in the UAV collaborative computing alliance.

[0097] The key to the optimization of this invention is to determine a drone cluster detection scheme to maximize the group detection computing power. The scheme includes three main parts: detection topology, i.e., cluster drone detection location, task offloading, and computing resource allocation. The following is the intelligent optimization algorithm of the detection machine:

[0098] S5.1. Initialize algorithm parameters, including the drone population size pop, the maximum number of iterations max_iter, the convergence factor a, and the balance factor l 1 , parameter A.

[0099] S5.2, detect the topology optimization layer and use the good point set to initialize the position of the drone population.

[0100] According to formula (27), the optimal point r = (r 1 ,r 2 ,…,r j ):

[0101]

[0102] Where k is the smallest prime number that satisfies (k-3) / 2≥dim, and dim represents the dimension of the optimization problem.

[0103] Good Points Collection P n (i), can be expressed as:

[0104] P n (i)={({r 1 i},{r 2i},…,{r n i})},1≤i≤pop (28)

[0105] Among them, {r n i} means taking the decimal part.

[0106] Map the good point set into the search space of the population:

[0107] x i (j) = (ub j -lb j )·P n (i)+lb j (29)

[0108] Among them, x i represents the position of all drones in the cluster i, ub j and lb j Represents the upper and lower bounds of the j-th dimension.

[0109] S5.3. Based on the current detection topology, detect all detectable grids in the cluster, use the cluster information interaction model to make the grid uncertainty of all detectable grids lower than the threshold, output the cluster detection force F and transfer the load calculation amount C of each drone i To the resource optimization layer.

[0110] S5.4, resource optimization layer. From the system optimization model (26), it can be seen that when the detection power F is determined, the group detection computing power depends on the task completion delay T. The system optimization model is simplified into an optimization sub-model (30):

[0111]

[0112] Based on the block coordinate descent method, the sub-model (30) is decomposed into a task offloading optimization sub-problem (31) and a computing resource allocation optimization sub-problem (32).

[0113]

[0114] Solution to subproblem P2.1: The calculation factor cf can be expressed as:

[0115]

[0116] The following formula can be used to find the optimal task offloading solution when the computing resource allocation solution is fixed.

[0117]

[0118] Solution to sub-problem P2.2:

[0119] The local calculation factor b is expressed as:

[0120]

[0121] The collaborative calculation factor c is expressed as:

[0122]

[0123] The following formula can be used to find the optimal computing resource allocation plan when the task offloading plan is fixed.

[0124]

[0125] Form a collaborative computing alliance, alternately optimize subproblems P2.1 and P2.2 until the task completion time T converges to the optimal, and output the optimal task completion delay T under the current detection topology, as well as the task offloading ratio matrix λ and the computing resource allocation matrix f r .

[0126] S5.5. Calculate and update the fitness of all populations according to formula (24), and update the α, β and δ drones to the top three populations with better fitness.

[0127] S5.6. Update all population positions and parameters a, A, C according to formula (43).

[0128]

[0129] In the formula, is a random vector in [0,2], is a random integer in [1,2], Indicates the current position of α, β and δ drones, Indicates the current position of the drone.

[0130]

[0131] Where r is a random number in [0,1], and the coefficient A 1 and A 3 The calculation formula is as follows:

[0132]

[0133] In the formula, r 1 is a random number in [0,1], convergence factor a and balance factor l 1 The calculation formula is as follows:

[0134]

[0135] S5.7. Determine whether the number of loops reaches the maximum number of iterations. If not, increase the number of iterations by one and return to step 3 to continue execution. If reached, exit the iteration process and output the calculated optimal drone cluster topology, task offloading plan and computing resource allocation plan.

[0136] Beneficial Effects

[0137] 1. The present invention is suitable for various production applications, including daily inspection and detection of landforms such as forest farms, grasslands, pastures, wetlands, etc., as well as multi-UAV swarms for emergency response after unexpected disasters, and rapid detection and calculation applications in key areas.

[0138] 2. Because the present invention adopts an intelligent drone cluster composed of multiple drones to perform collaborative detection on the mission area, it can complete the task quickly compared with the traditional single-machine detection solution.

[0139] 3. In the traditional drone cluster detection mode, drones rely on ground base stations to interact with the command center. However, in the harsh environment of the disaster area, the ground base stations may fail due to damage. Therefore, the present invention introduces collaborative computing technology in the cluster collaborative detection scenario in the absence of base stations. By reasonably optimizing cluster resources, it can achieve rapid and comprehensive integration of key information in the disaster area compared to traditional detection methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0140] Figure 1 This is a schematic diagram of the grid division of the patrol mission area;

[0141] Figure 2 This is a schematic diagram of the establishment of the UAV Collaborative Computing Alliance;

[0142] Figure 3 It is the diagram of the effects of the system influencing factors;

[0143] Figure 4 It is the flow chart of the intelligent optimization algorithm of the detection fleet;

[0144] Figure 5 It is a diagram of the iterative process of the intelligent optimization algorithm of the detection cluster. DETAILED DESCRIPTION

[0145] The invention will be further described in detail below in conjunction with the accompanying drawings.

[0146] like Figure 1 to Figure 5 The present invention provides a method for scheduling the detection and computing power of a drone group based on collaborative computing, the method comprising the following steps:

[0147] (1) The steps to establish an adaptive grid map based on prior information are as follows:

[0148] (1.1) Based on the detection data of low-orbit satellites, the geographical environment information M of the target area is obtained. M ,y M ,z M}, the number of target points to be detected n target and the target location information R = {x R ,y R ,z R}.

[0149] (1.2) Adaptive rasterized map G = {g 1 ,g 2 ,…,g i}:like Figure 1 As shown in the figure, the task area is divided into grid maps of different grid sizes according to the detection weight. The target area is divided into four grids, and the grid detection weight is expressed as:

[0150]

[0151] in, Indicates the importance of the target to be detected in the grid. represents the complexity of the terrain, α w ,β w is the corresponding coefficient. th The grid is further divided into four sub-grids, and the above process is repeated until no further division is possible.

[0152] (1.3) Task Search Vector The task search vector consists of the grid target existence probability and grid uncertainty. Represents the grid g i The probability of the target existing, grid uncertainty Represents the grid g i In view of the uncertainty of the grid target, the calculation methods of grid target existence probability and grid uncertainty are introduced below.

[0153] Grid target existence probability: initial grid target existence probability The grid detection weight W i Normalized, it can be expressed as:

[0154]

[0155] in, η represents the accuracy of prior information, η∈[0,1].

[0156] The grid uncertainty is defined as the information entropy of the probability of the target existing in the grid:

[0157]

[0158] (2) Drone detection vector

[0159] (2.1) UAV detection range model: The center point of the grid is defined as the information point of the grid. If the grid information point is within the detection range of the UAV and is not blocked by obstacles such as mountains, then the grid is within the visible grid set of the UAV. According to formula (4), the visible grid set vg of each UAV is determined j :

[0160]

[0161] in, Indicates the detection position of the drone j, Represents the grid g i Information point position, Δh j,gi Represents drone j and grid g i The vertical distance between Denotes that UAV j is in Δh j,gi Altitude detection range radius, α u is the detection angle, z k Represents drone j and grid g i The sampling points (x k ,y k ,z k )’s height, z ks It means (x k ,y k ) at the ground level.

[0162] (2.2) Drone Probabilistic Detection Model:

[0163] The probability of successful detection indicates the probability that the UAV system successfully determines the existence of the target when the target exists, and is defined as an exponential function of the square of the height difference between the UAV and the detectable grid:

[0164]

[0165] In the formula, ω s is the sensitivity of successful detection of UAV; s is the successful detection coefficient of the UAV; Indicates the grid g i Whether there is actually a target in Indicates that there is no target in the grid gi. Represents the grid g i There is actually a goal in Indicates the specific results of drone detection. j,gi(t) = 1 means that at time t, UAV j detects grid g i There is a goal, when Indicates that the drone detects grid g i There is no goal; Represents drone j and grid g i The height difference between .

[0166] The error detection probability indicates the probability that the UAV system incorrectly judges the existence of a target when the target does not exist:

[0167]

[0168] In the formula, ω f is the error detection sensitivity of the UAV, κ f is the error detection coefficient of the UAV.

[0169] (2.3) Task Search Vector Update Model

[0170] Based on the information detected by the drone (i.e., PD and PF), the existence probability of the grid target is dynamically updated using the Bayesian criterion. to reflect the current distribution of targets.

[0171] When drone j detects grid g i When there is a target in At this time, the probability of the grid target existing is updated as:

[0172]

[0173] When drone j detects grid g i When there is no target in At this time, the probability of the grid target existing is updated as:

[0174]

[0175] At the same time, based on formula (3), the grid uncertainty is updated according to the new target existence probability. Set the grid g i The conditions for determining whether a target exists are:

[0176]

[0177] Therefore, combined with formula (3), the grid uncertainty threshold o is obtained th = 0.0808, when the uncertainty of the grid is lower than o th , it indicates that the grid detection is completed.

[0178] (2.4) Cluster Information Collaboration Model

[0179] During the cluster collaborative detection process, each drone independently completes its own task search vector In order to quickly and accurately complete the detection of the cluster and avoid the drones from repeatedly detecting the grids that have been detected, the drones in the cluster need to obtain the global task search vector information based on the task search vectors of other drones. Therefore, a cluster information coordination model is established.

[0180] From formulas (7) and (8), we can see that the simultaneous detection of k drones can be expressed as:

[0181]

[0182] In order to simplify the calculation, formula (10) is transformed into:

[0183]

[0184] Then, the formula (11) is logarithmically transformed to obtain:

[0185]

[0186] make Therefore, formula (12) can be transformed into:

[0187]

[0188] in,

[0189] From formula (13), we can see that through the cluster information coordination model, the drone only needs to receive the probability offset transmitted by other drones The global search vector information can be obtained. Since the communication only involves the transmission probability offset, the communication delay can be ignored.

[0190] (3) UAV calculation vector

[0191] The drone calculates the vector JS = {(xt,AC),λ,f r}, where xt represents the collaborative computing attributes of the UAV; AC represents the collaborative computing alliance matrix; λ represents the task offloading ratio matrix; f r The table calculates the resource allocation ratio matrix. The calculation method of each part is introduced in turn below.

[0192] (3.1) Collaborative calculation of attribute xt

[0193] Non-Cooperative Unmanned Aerial Vehicle (NC-UAV): When the load is high, it needs to offload part of its computing tasks to cooperative UAVs, and can offload to multiple cooperative UAVs in a certain proportion, so that multiple cooperative UAVs can coordinate computing to improve the speed of task completion. Set the cooperative computing attribute xt=0 of non-cooperative UAVs.

[0194] Cooperative Unmanned Aerial Vehicle (C-UAV): When its own task load is low, it can receive task data from non-cooperative UAVs and help them perform collaborative computing. Set the collaborative computing attribute xt=1 for cooperative UAVs.

[0195] (3.2) Cluster Collaborative Computing Alliance AC: Figure 2 As shown in the figure, considering that there may be multiple non-cooperative drones in the cluster, all cooperative drones can simultaneously join the computing alliance of multiple non-cooperative drones to participate in task computing. Therefore, the formation method of the collaborative computing alliance AC is as follows:

[0196]

[0197] Among them, AC ij =1 represents that drone j joins the collaborative computing alliance of task i.

[0198] (3.3) Task offloading ratio matrix: Define the task offloading ratio matrix λ = [λ ij ] N×N (λ ij ∈[0,1]), λ ij represents the proportion of tasks that drone i will offload from its own task i to drone j, if drone j is not in the collaborative computing alliance AC of task i i In the ij =0, and it is necessary to ensure Therefore, the task offloading ratio matrix initialization method is as follows:

[0199]

[0200] In the formula, represents the number of drones in the alliance for task i.

[0201] (3.4) Computational resource allocation matrix: The remaining computing resources of drone i are defined as f i , the computing resource allocation ratio matrix is in It means that drone i will use its remaining computing resources f i The proportion assigned to task j must satisfy To ensure that computing resources are not over-allocated, the computing resource allocation ratio matrix is ​​initialized as follows:

[0202]

[0203] In the formula, Represents the number of collaborative computing alliances that UAV j participates in.

[0204] (4) Calculation method of group detection computing power

[0205] The group detection computing power is mainly composed of the cluster detection capability and the cluster task completion delay. The calculation methods of each part are described in turn below.

[0206] (4.1) Cluster detection capability

[0207] The detection capability of the cluster is related to the change of uncertainty in the mission area and the number of detected targets, which can be expressed as:

[0208]

[0209]

[0210] Where μ represents the weight factor of uncertainty change, n F Indicates the number of targets detected by the cluster, n g represents the number of grids in the task area, Ο represents the initial uncertainty of the task area, represents the initial uncertainty of each grid in the mission area, Ο' represents the uncertainty of the mission area after detection is completed, Represents the uncertainty of each grid in the mission area after detection.

[0211] (4.2) Cluster task completion delay model:

[0212] The cluster task completion delay is mainly composed of the local computing delay and collaborative computing delay of each subtask. The calculation method of each delay is introduced in turn below.

[0213] (4.3) Local computation delay of task i:

[0214]

[0215] In the formula, λ i,i It means that UAV i will calculate the amount C of task i i The ratio of uninstalls to itself, It means that drone i will use its remaining computing power f i The proportion of the allocation allocated to itself.

[0216]

[0217] In the formula, φ i represents the CPU cycles required to calculate 1 bit of data of task i, I i represents the amount of data for task i, Represents the grid g k Area, g k ∈vg i ; Represents the visible grid set vg of drone i i Total area; Indicates vg i Number of grids in data s Indicates the amount of data generated by a drone detection; Represents the distance between drone i and grid g k The number of detections.

[0218] (4.4) When drone i cannot complete the computation of task i alone, task i needs to be split into multiple sub-computing tasks and offloaded to the collaborative drones in the collaborative computing alliance for computation. The collaborative computing delay of task i includes the transmission delay of the sub-tasks and the computation delay of the sub-tasks:

[0219]

[0220] Communication rate i,j Calculated by formula (22):

[0221]

[0222] in, It means that drone i distributes the bandwidth equally to all cooperative drones in the task i alliance. represents the transmission power of UAV i, N 0 is the noise power spectral density, g 0 is the channel power gain at the reference distance, d i,j represents the distance between drone i and the coordinated drone j in the alliance. i,j →0: Rate i,j →∞, however, this situation is unlikely to happen in reality. Considering that the infinite approach of drones may cause collisions and other safety accidents, the hovering distance of all drones should be greater than a certain threshold d safe Therefore, the completion delay of task i is:

[0223] t i =max(t loc,i ,t col,i ) (twenty three)

[0224] All tasks are calculated in parallel, so the total task completion delay T is:

[0225] T=max(t i ) (twenty four)

[0226] (5) The steps for establishing the system optimization model are as follows:

[0227] The system group detection power U can be expressed as:

[0228]

[0229] The optimization goal is to maximize the system group detection computing power:

[0230] max U(26) combined with the constraints yields:

[0231]

[0232] Wherein, constraints 1, 2, and 3 represent the detection position constraints of UAVs, which means that UAVs need to detect within the mission area; constraint 4 indicates that the hovering distance duav of UAVs in the cluster should be greater than a certain threshold dsafe; constraints 5 and 6 represent the unloading ratio constraints of the UAV collaborative computing alliance; constraints 7 and 8 represent the computing resource allocation ratio constraints in the UAV collaborative computing alliance.

[0233] From formula (25), we can see that the group detection computing power depends on the detection power and the task completion delay. Figure 3 As shown in the figure, the number of targets found and the uncertain changes in the mission area are different under different detection topologies. The resulting cluster detection power and the amount of task computing carried by each drone are different. When the detection topology is determined, the task completion delay is related to the task unloading in the cluster and the allocation of computing resources in the cluster. Therefore, the key to optimizing this system is to determine a drone cluster detection scheme that includes the detection topology, i.e., the detection location of the cluster drone, task unloading, and computing resource allocation to maximize the cluster detection computing power. The following is the intelligent algorithm for the detection swarm:

[0234] The working theory of the Reconnaissance Uav Swarm Intelligent Optimization (RUSIO) algorithm is as follows: in the process of executing the task, different detection swarms have different group detection computing powers. The top three detection swarms with the highest group detection computing powers are α-uavs, β-uavs and δ-uavs, and the remaining detection swarms are ω-uavs. α-uavs, β-uavs and δ-uavs are used to guide the remaining detection swarms to explore the mission area, and the positions of α-uavs, β-uavs and δ-uavs are continuously iterated to obtain the optimal solution.

[0235] like Figure 4 As shown in the figure, the specific steps of the intelligent optimization algorithm of the detection cluster are as follows:

[0236] (5.1) Initialize algorithm parameters, including drone population size pop, maximum number of iterations max_iter, convergence factor a, balance factor l 1 .

[0237] (5.2) Use the good point set to initialize the position of the drone population.

[0238] According to formula (28), the optimal point r = (r 1 ,r 2 ,…,r j ):

[0239]

[0240] Where k is the smallest prime number that satisfies (k-3) / 2≥dim, and dim represents the dimension of the optimization problem.

[0241] Good Points Collection R n (i), can be expressed as:

[0242] R n (i)={({r 1 i},{r 2 i},…,{r n i})},1≤i≤pop (29)

[0243] Among them, {r n i} means taking the decimal part.

[0244] Map the good point set to the search space of the population:

[0245] x i (j) = (ub j -lb j )·R n (i)+lb j (30)

[0246] Among them, x i represents the position of all drones in the cluster i, ub j and lb j Represents the upper and lower bounds of the j-th dimension.

[0247] (5.3) Based on the current detection topology, detect all detectable grids in the cluster, use the cluster information interaction model to make the grid uncertainty of all detectable grids lower than the threshold, output the cluster detection force F and transfer the load calculation amount C of each drone i To the resource optimization layer.

[0248] (5.4) The optimization steps of the resource optimization layer are as follows:

[0249] From the system optimization model (27), it can be seen that when the detection force F is determined, the detection power of the system group depends on the task completion delay T. The system optimization model is simplified into the optimization sub-model (31):

[0250]

[0251] Based on the block coordinate descent method, the sub-model (31) is decomposed into a task offloading optimization sub-problem (32) and a computing resource allocation optimization sub-problem (33).

[0252]

[0253] Subproblem P2.1 Optimization steps: The calculation factor cf can be expressed as:

[0254]

[0255] The following formula can be used to find the optimal task offloading solution when the computing resource allocation solution is fixed.

[0256]

[0257] Optimization steps for subproblem P2.2:

[0258] The local calculation factor b is expressed as:

[0259]

[0260] The collaborative calculation factor c is expressed as:

[0261]

[0262] The following formula can be used to find the optimal computing resource allocation plan when the task offloading plan is fixed.

[0263]

[0264] Form a collaborative computing alliance, alternately optimize subproblems P2.1 and P2.2 until the task completion time T converges to the optimal, and output the optimal task completion delay T under the current detection topology, as well as the task offloading ratio matrix λ and the computing resource allocation matrix f r .

[0265] (5.5) According to formula (25), the fitness of all populations is calculated and updated, and the α, β and δ drones are updated to be the top three populations with better fitness.

[0266] (5.6) Update all population positions and parameters a, A, C according to the following formula.

[0267]

[0268] In the formula, is a random vector in [0,2], is a random integer in [1,2], Indicates the current position of α, β and δ drones, Indicates the current position of the drone.

[0269]

[0270] Where r is a random number in [0,1], and the coefficient A 1 and A 3 The calculation formula is as follows:

[0271]

[0272] In the formula, r 1 is a random number in [0,1], convergence factor a and balance factor l 1 The calculation formula is as follows:

[0273]

[0274] (5.7) Determine whether the number of loops reaches the maximum number of iterations. If not, increase the number of iterations by one and return to step 3 to continue execution. If it reaches the maximum number of iterations, exit the iteration process and output the calculated optimal drone cluster topology, task offloading plan and computing resource allocation plan.

[0275] (6) Analysis of simulation results:

[0276] The above algorithm is simulated and analyzed using MATLAB to prove the effectiveness of the Reconnaissance Uav Swarm Intelligent Optimization (RUSIO) algorithm. Consider a 0.5km×0.5km detection area, which contains 6 targets to be detected and 6 detection drones. The drone's channel bandwidth B=10MHz and transmission power Pt=10dB.

[0277] The other simulation parameters are set as follows: safe hovering distance between drones dsafe = 5m, channel gain power at reference distance g0 = -50dB, noise power spectrum density N 0 =-170dBm / Hz, the detection angle of the drone αu = 45°, the parameters of the drone probability detection model: κf = 0.9, κs = 0.9; the drone computing power f = 2GHz.

[0278] like Figure 5As shown in the figure, under the intelligent optimization algorithm of the detection cluster, the cluster detection computing power becomes 2624.72 bit / s, and the calculated cluster detection power is 5.4648×10 4 bit, the task completion delay is 20.8343s. Compared with the particle swarm optimization algorithm (PSO) in the traditional algorithm, the swarm detection computing power is improved by 254.6%.

Claims

1. A method for scheduling computing power of drone swarm based on collaborative computing, characterized in that: include: S1. Establish an adaptive grid map based on initial prior information and obtain the task search vector Among them, j represents the drone number and i represents the grid number; is the probability of grid target existence, indicating that drone j detects grid g i The probability of the target existing, grid uncertainty Indicates that drone j detects grid g i uncertain situation; Adaptive rasterized map G = {g1, g2, ..., g i }: According to the detection weight, the task area is divided into grid maps of different grid sizes. The target area is divided into four grids. The grid detection weight is expressed as: in, Indicates the importance of target detection within the grid, represents the complexity of the terrain, α w ,β w is the corresponding coefficient, and the weight exceeds the threshold W th The grid is further divided into four sub-grids, and the above process is repeated until no further division is possible; Grid target existence probability: Initial grid target existence probability The grid detection weight W i Normalized to get: in, η represents the accuracy of prior information, η∈[0,1]; The grid uncertainty is defined as the information entropy of the probability of the target existing in the grid: S2. Analyze the composition of drone detection vectors, including drone detection range, drone probability detection model, task search vector update model, and establish a drone cluster information coordination model; S3, analyze the composition of drone computing vectors, including drone collaborative computing attributes, cluster collaborative computing alliance, task offloading ratio matrix and computing resource allocation matrix; S4. Analyze the composition of the group detection computing power, including the cluster detection power and the cluster task completion delay, and obtain the calculation method of the group detection computing power; S5. With the goal of maximizing the swarm detection computing power, a system optimization model is established to jointly optimize the drone cluster detection position and cluster resource allocation through the detection swarm intelligent optimization algorithm.

2. The method for scheduling the computing power of a drone group based on collaborative computing according to claim 1 is characterized in that: The specific steps of step S2 are as follows: (a) Drone detection vector ZC j ={vg j ,(PD j ,PF j ),Y j }, where vg j represents the detectable grid set of UAV j, PD j Indicates the probability of successful detection, PF j represents the error detection probability. The success detection probability and the error detection probability are collectively referred to as the drone probability detection model; Y j represents the mission search vector of UAV j; (b) Drone Probabilistic Detection Model: The probability of successful detection indicates the probability that the UAV system successfully determines the existence of the target when the target exists, and is defined as an exponential function of the square of the height difference between the UAV and the detectable grid: In the formula, Defined as the probability of successful detection, Indicates the grid g i Whether there is actually a target in Indicates the specific results of drone detection, ω s is the successful detection sensitivity of the UAV, κ s is the successful detection coefficient of the drone, Represents drone j and grid g i The height difference between j represents the set of detectable grids of UAV j; The error detection probability indicates the probability that the UAV system incorrectly judges the existence of a target when the target does not exist: In the formula, ω f is the error detection sensitivity of the UAV, κ f is the error detection coefficient of the UAV; (c) Cluster Information Collaboration Model During the cluster collaborative detection process, each drone independently completes its own task search vector In order to quickly and accurately complete the detection of the cluster and avoid the drones from repeatedly detecting the grids that have been detected, the drones in the cluster need to obtain the global task search vector information based on the task search vectors of other drones. Therefore, a cluster information coordination model is established: According to the Bayesian criterion, the simultaneous detection of k drones can be expressed as: In order to simplify the calculation, formula (6) is transformed into: make Therefore, formula (7) can be transformed into: From formula (8), we can see that through the cluster information coordination model, the drone only needs to receive the probability offset transmitted by other drones The global search vector information can be obtained. Since the communication only involves the transmission probability offset, the communication delay can be ignored.

3. The method for scheduling the computing power of drone swarm based on collaborative computing according to claim 1, characterized in that: The specific steps of step S3 are as follows: (a) UAV calculation vector The drone calculates the vector JS = {(xt,AC),λ,f r }, where xt represents the collaborative computing attribute of the UAV, AC represents the collaborative computing alliance matrix, λ represents the task offloading ratio matrix, and f r Table calculation resource allocation ratio matrix; (b) Collaborative calculation of attribute xt Non-Cooperative Unmanned Aerial Vehicle (NC-UAV): When the load is high, it needs to offload part of its computing tasks to cooperative UAVs, and can offload to multiple cooperative UAVs in a certain proportion, so that multiple cooperative UAVs can coordinate computing to improve the task completion speed. Set the cooperative computing attribute xt=0 of non-cooperative UAVs; Cooperative Unmanned Aerial Vehicle (C-UAV): When its own task load is low, it can receive task data from non-cooperative UAVs and help non-cooperative UAVs to perform collaborative computing, and set the collaborative computing attribute xt=1 of the cooperative UAV; (c) Cluster collaborative computing alliance formation: Considering that there may be multiple non-cooperative drones in the cluster, in order to improve the cluster resource utilization, all collaborative drones can dynamically join the computing alliance of multiple non-cooperative drones to participate in task calculations at the same time, considering their own load conditions and collaborative benefits.

4. The method for scheduling the computing power of a drone group based on collaborative computing according to claim 1 is characterized in that: The specific steps of step S4 are as follows: (a) Calculation method of group detection computing power: The group detection computing power is mainly composed of cluster detection power and cluster task completion delay. The calculation methods of each part are described in turn below; (b) Cluster detection capability: The cluster detection capability is related to the change of uncertainty in the mission area and the number of detected targets, expressed as: Where μ represents the weight factor of uncertainty change, n F Indicates the number of targets detected by the cluster, n g represents the number of grids in the task area, Ο represents the initial uncertainty of the task area, represents the initial uncertainty of each grid in the mission area, Ο' represents the uncertainty of the mission area after detection is completed, It represents the uncertainty of each grid in the mission area after the detection is completed; (c) Cluster task completion delay model: The cluster task completion delay is mainly composed of the local computing delay and collaborative computing delay of each subtask. The calculation methods of each delay are introduced in turn below: Local computation latency of task i: In the formula, λ i,i It means that UAV i will calculate the amount C of task i i The ratio of uninstalls to itself, It means that drone i will use its remaining computing power f i The proportion of allocation allocated to oneself, In the formula, φ i represents the CPU cycles required to calculate 1 bit of data of task i, I i represents the amount of data for task i, Represents the grid g k The area of ​​​​g k ∈vg i , Represents the visible grid set vg of drone i i The total area, Indicates vg i The number of grids in data s Indicates the amount of data generated by a drone detection. Represents the distance between drone i and grid g k The number of detections; The collaborative computing delay of task i includes the transmission delay of subtasks and the computing delay of subtasks: Communication rate i,j Calculated by formula (14): in, It means that drone i distributes the bandwidth equally to all cooperative drones in the task i alliance. represents the transmission power of drone i, N0 is the noise power spectrum density, g0 is the channel power gain at the reference distance, d i,j represents the distance between drone i and the coordinated drone j in the alliance; when d i,j →0: Rate i,j →∞, however, this situation is unlikely to happen in reality. Considering that the infinite approach of drones may cause collisions and other safety accidents, the hovering distance of all drones should be greater than a certain threshold d safe , therefore, the completion delay of task i is: t i =max(t loc,i ,t col,i )。 (15) All tasks are calculated in parallel, so the total task completion delay T is: T=max(t i ) (16)。 5. The method for scheduling the computing power of drone swarm based on collaborative computing according to claim 1 is characterized in that: S5 proposes an intelligent optimization algorithm for detection swarms, which optimizes the detection and computing power of drone swarms by alternately optimizing the detection topology optimization layer and the resource optimization layer; (a) Initialize the algorithm parameters, including the UAV population size pop, the maximum number of iterations max_iter, the convergence factor a, the balance factor l1, and the parameter A; (b) In the detection topology optimization layer, the detection swarm intelligent optimization algorithm uses the best point set to initialize the position of the drone population to improve the diversity of the population. The best point set initialization method is as follows, and the best point is calculated according to formula (17): Where k is the smallest prime number that satisfies (k-3) / 2≥dim, and dim represents the dimension of the optimization problem; Good Point Collection P n (i), can be expressed as: P n (i)={({r1i},{r2i},…,{r n i})},1≤i≤pop (18) In the formula, {r n i} indicates taking the decimal part, and pop indicates the size of the drone population; Map the good point set into the search space of the population: x i (j)=(ub j -lb j )·P n (i)+lb j (19) In the formula, x i represents the location of all drones in the cluster i, i.e., the detection topology, ub j and lb j represents the upper and lower bounds of the j-th dimension; (c) Based on the current detection topology, detect all detectable grids in the cluster, use the cluster information interaction model to make the grid uncertainty of all detectable grids lower than the threshold, output the cluster detection force F and transfer the load calculation amount C of each drone i To the resource optimization layer; (d) In the resource optimization layer, the intelligent optimization algorithm of the detection cluster adopts the idea of ​​block coordinate descent algorithm to decouple the original optimization problem into task offloading optimization sub-problem (20) and computing resource allocation optimization sub-problem (21). The problem is expressed as follows: The drone swarm forms a collaborative computing alliance and obtains the best resource allocation solution by alternately solving two sub-optimization problems to minimize the task completion delay. The solutions to the sub-optimization problems are as follows: Solution to subproblem P2.1: Define the calculation factor cf as: The optimal task offloading solution can be obtained by the following formula when the computing resource allocation solution is fixed: Solution to subproblem P2.2: Define the local calculation factor b as: Define the collaborative calculation factor c as: The optimal computing resource allocation scheme can be obtained by the following formula when the task offloading scheme is fixed: Output the optimal task completion delay T under the current detection topology, as well as the task offloading ratio matrix λ and the computing resource allocation matrix f r ; (e) According to the formula Calculate and update the fitness of all populations, and update α, β and δ drones to the top three populations with better fitness; (f) In the detection topology optimization layer, based on the guidance of α, β and δ drones, the positions of the remaining drones are updated. The updating steps are as follows: In the formula, is a random number in [0,2], is a random integer in [1,2], Indicates the current position of α, β and δ drones, Indicates the current position of the drone; Where r is a random number between 0 and 1. Coefficients A1 and A3 are used to balance the exploration and development ratios of the algorithm. The calculation formula is as follows: Where r1 is a random number between 0 and 1. The calculation formulas for the convergence factor a and the balance factor l1 are as follows: (g) Determine whether the number of loops reaches the maximum number of iterations. If not, increase the number of iterations by one and return to step 3 to continue execution. If reached, exit the iteration process and output the calculated optimal UAV cluster topology, task offloading plan, and computing resource allocation plan.

Images