Unmanned aerial vehicle data acquisition optimization method based on federal learning

Abstract

The invention discloses an unmanned aerial vehicle data acquisition optimization method based on federated learning, and belongs to the technical field of wireless communication, and the method specifically comprises the following steps: firstly, achieving the data transaction between an unmanned aerial vehicle and a ground user through a constructed unmanned aerial vehicle auxiliary data acquisition federated learning system model based on a Stackelberg game; optimizing association between subproblem ground users and unmanned aerial vehicles, data uploading ratio and transmitting power of the ground users, unmanned aerial vehicle positions, computing resources and reward distribution ratio to obtain an optimized ground utility function and an optimized unmanned aerial vehicle utility function; and finally, through the ground utility function and the unmanned aerial vehicle utility function after game optimization, obtaining a final optimal matching between the ground user and the unmanned aerial vehicle, a data uploading ratio and transmitting power of the ground user, the position of the unmanned aerial vehicle, and a calculation resource and reward distribution ratio. According to the invention, the data utilization rate and federal learning efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technologies, and particularly to an optimization method for drone data collection based on federated learning. Background Art

[0002] In the existing scenarios of drone-assisted federated learning, some literature focuses on drones aggregating the models locally trained by ground users. Such methods impose relatively high requirements on the computing capabilities of ground users. Some collect ground data through drones and perform local model training on the drones. This type of literature assumes that drones can collect data from all ground users within the coverage area, such as patient medical data in the medical field and camera data in the security field, without considering the participation willingness of ground users. In practical applications, ground users may lack the willingness to participate due to factors such as energy consumption, and the existing methods' relatively high requirements for the computing capabilities and participation enthusiasm of ground users lead to reduced data utilization and affect the efficiency of federated learning. Summary of the Invention

[0003] The objective of the present invention is to provide an optimization method for drone data collection based on federated learning. By optimizing the drone utility function and the ground user utility function, the final optimal best pairing of ground users and drones, the ground user transmission power ratio and the ground user data upload ratio, the drone position, the drone computing resources, and the drone reward allocation ratio are obtained, solving the problem in the prior art that ground users may lack the willingness to participate due to factors such as energy consumption, resulting in reduced data utilization and affecting the efficiency of federated learning. The present invention is implemented through the following technical solutions.

[0004] In a first aspect, the present invention provides an optimization method for drone data collection based on federated learning, including the following:

[0005] When a drone communicates with local users, the drone data collection optimization method includes:

[0006] Obtain the pre-constructed problem of maximizing the drone utility function and the problem of maximizing the ground user utility function; the problem of maximizing the ground user utility function includes the ground user and drone association optimization sub-problem and the ground user data upload ratio and transmission power optimization sub-problem; the optimization problem of the drone utility function is decomposed into the drone position optimization sub-problem, the drone computing resource optimization sub-problem, and the drone reward allocation ratio optimization sub-problem;

[0007] Adopt a game algorithm to solve the problem of maximizing the drone utility function and the problem of maximizing the ground user utility function, and calculate the optimal best pairing of ground users and drones, the ground user transmission power ratio and the ground user data upload ratio, the drone position, the drone computing resources, and the drone reward allocation ratio.

[0008] The Stackelberg game is a leader-follower game model, in which one party (the leader) makes a decision first, and the other party (the follower) makes a response after observing the leader's decision. The present invention jointly considers the interests of both the UAV and the ground user, and constructs the interest competition between the UAV and the ground user as a Stackelberg game model, where the UAV is the leader and the ground user is the follower.

[0009] The UAV has the power to determine the reward allocation ratio, and this ratio will directly affect the amount of data that the ground user chooses to offload to each UAV. As the amount of data uploaded by the ground user to the UAV changes, the UAV will dynamically adjust its reward allocation ratio according to the received amount of data to maximize its own benefits in the cooperation process; conversely, the adjusted reward allocation ratio of the UAV will affect the ground user's decision on the data upload ratio, and both parties will play a game based on the reward allocation ratio and the data upload ratio. Therefore, it is necessary to play a game on the utility functions of the follower and the leader to obtain the final optimal UAV reward allocation ratio and ground user data upload ratio.

[0010] Optionally, the data transaction when the UAV communicates with the ground user includes the following:

[0011] Under a given amount of data, the UAV determines the corresponding reward allocation ratio decision according to the channel transmission environment and its own needs, and sends a data collection request to the ground user. After receiving the data collection request from the UAV, the ground user optimizes the data upload ratio according to the allocated reward, so as to increase its own benefits obtained in the cooperation process.

[0012] Optionally, the optimization problem of the ground user utility function is defined as P1, the optimization sub-problem of the ground user and the UAV association is defined as P1-1, and the optimization sub-problem of the ground user data upload ratio and transmission power is defined as P1-2;

[0013] The expression of P1 is as follows:

[0014] ,

[0015] In the formula, C1 is the range of the ground user's data upload ratio, C2 is the limit of the transmission power of the ground user's transmitted data, C3 indicates that the association between the ground user and the UAV needs to meet a certain distance, C4 indicates the limitation of the time consumed by the ground user to transmit data, C5 indicates that a ground user can only be associated with one UAV, and C6 indicates that the ground user can participate in the contribution only when the income is positive;

[0016] Among them, is the ground user utility function, is the ground user number, and the value range is 1-N. is the UAV number, and its value range is from 1 to M. is the set of all UAVs. is the th set of ground users within the coverage of the UAV. is the set of all ground users. is the transmission power of the ground user . is the data upload ratio of the ground user . is the UAV and the correlation value with the ground user . indicates that the UAV covers the ground user , indicates that the UAV does not cover the ground user , indicates the reward allocation ratio of the UAV to the ground user. is the reward obtained by the UAV . represents the unit cost of energy consumption for the ground user to upload data, reflecting the importance degree of the ground user to the transmission energy consumption. is the ground user 's data volume. is the transmission bandwidth of the ground user . is the distance between the UAV and the ground user . is the reference distance m and the channel gain magnitude in the wireless channel. is the path loss exponent. is the maximum transmission distance under the UAV coverage. is the power of Gaussian white noise. and are respectively the minimum and maximum transmission powers of the ground user . is the maximum data upload time;

[0017] The expression of P1-1 is as follows:

[0018] ,

[0019] The expression of P1-2 is as follows:

[0020] .

[0021] Optionally, define the optimization problem of the UAV utility function as P2, simplify problem P2 using variable substitution to represent it as P3, define the problem of optimizing the UAV position as P3-1, define the problem of optimizing the UAV computing resources as P3-2, and define the problem of optimizing the UAV reward allocation ratio as P3-3;

[0022] The expression of P2 is as follows:

[0023]

[0024] In the formula, C7 represents the CPU frequency that restricts the UAV to participate in federated learning training, C8 represents the UAV position range, C9 represents that the reward given by the UAV to the ground user cannot be greater than the reward received by the UAV, and C10 represents that the total delay of the UAV local training shall not exceed ;

[0025] is the position of the UAV ; is the position of the ground user ; is the CPU frequency of the UAV ; represents the unit cost of the energy consumption of the UAV local model training, represents the unit cost of the energy consumption generated by the UAV hovering, respectively reflecting the importance of the UAV for the energy consumption of the training model and the hovering energy consumption, is the global iteration number of federated learning, is the local iteration number of the UAV ; is the energy coefficient used by the CPU, is the UAV the number of CPU cycles required for each bit of calculation, is the hovering power of the UAV, and are respectively the minimum and maximum CPU frequencies of the UAV ; is the total delay of the UAV for local training;

[0026] Since the objective function in P2 contains , an auxiliary variable is introduced, and the constraint C11

[0027]

[0028] Therefore, problem P2 is simplified to represent as P3, and the expression of P3 is as follows:

[0029] ,

[0030] The P3-1 expression is as follows:

[0031] ,

[0032] The P3-2 expression is as follows:

[0033] ,

[0034] The P3-3 expression is as follows:

[0035] .

[0036] Optionally, the game algorithm is adopted to solve the problems of maximizing the utility function of the UAV and the utility function of the ground user, and calculate the optimal best pairing of the ground user and the UAV, the ground user transmission power ratio and the ground user data upload ratio, the UAV position, the UAV computing resources and the UAV reward allocation ratio, including performing game solutions on P1-1, P1-2, P3-1, P3-2 and P3-3 to obtain the finally optimized UAV utility function and ground user utility function, and obtaining the finally optimal best pairing of the ground user and the UAV, the ground user transmission power ratio and the ground user data upload ratio, the UAV position, the UAV computing resources and the UAV reward allocation ratio based on the finally optimized UAV utility function and ground user utility function. Specifically, it includes the following steps:

[0037] Step 1: Initialize the data reward allocation ratio of all UAV sets and the data upload ratio and transmission power of all ground users on the ground , and enter Step 2;

[0038] Step 2: Use an improved method based on density-based clustering to solve the UAV position set , the clustering result and the number of UAVs , and enter Step 3;

[0039] As the method iterates, when the utility of each UAV no longer changes at a given accuracy , the method runs to completion.

[0040] Step 3: Each UAV uses the particle swarm method, theoretical derivation and golden section method to optimize the UAV position , computing resources and reward allocation ratio in turn, and enter Step 4; where represents the number of iterations;

[0041] Step 4, when a ground user is within the coverage of two or more drones, traverse all drones that meet the constraint conditions, calculate the utility function of the ground user , and take the maximum ground user utility function The corresponding association method to determine the ground user and the drone The best pairing , and enter Step 5;

[0042] Step 5, for each ground user covered by each drone, use the simulated annealing algorithm to optimize the ground user transmission power and the ground user data upload ratio , and enter Step 6;

[0043] Step 6, calculate the utility of each drone. When the utility of each drone does not meet the given accuracy, return to Step 3 until the utility of each drone reaches the accuracy and then stop, so as to obtain the finally optimized drone utility function of each drone and the finally optimized utility function of each ground user , and further obtain the final optimal best pairing of ground users and drones, ground user upload data ratio and ground user transmission power ratio, drone positions, drone computing resources and drone reward allocation ratio;

[0044] Among them, the best pairing of ground users and drones satisfies ; the ground user upload data ratio satisfies ; the ground user transmission power ratio satisfies ; the drone positions satisfy ; the drone computing resources satisfy ; the drone reward allocation ratio satisfies .

[0045] Optionally, solving the P1-1 includes the following steps:

[0046] Step 1, input the ground user set , ground user transmission power , total bandwidth , the moving range of the drone , the maximum number of ground users that can be carried and the drone coverage radius , the parameter clustering radius in the clustering method and the minimum number of samples , and enter Step 2;

[0047] Step 2, use the original density-based clustering method for Cluster the ground users and proceed to step 3;

[0048] Step 3: Calculate the cluster diameter after clustering , if is greater than the UAV coverage diameter , reduce the clustering radius and return to step 2; if is less than or equal to the UAV coverage diameter , the clustering is successful, proceed to step 4;

[0049] Step 4: When a ground user is within the coverage of two or more UAVs, to optimize the benefit of the ground user, based on the determined transmit power and data upload ratio , traverse all UAVs that meet the constraint conditions and calculate the ground user utility function , select the maximum ground user utility function and the corresponding association method to determine the best pairing between the ground user and the UAV . .

[0050] After clustering is completed, the UAVs collect data of ground users within the coverage at a fixed altitude. The amount of data collected is determined by the ground users according to the rewards obtained. The higher the reward assigned by the UAV to the ground user, the more data the ground user contributes. After the UAVs collect the data, the rewards are allocated from the total rewards based on the training quality of the local model and the amount of data participating in the training.

[0051] Optionally, solving the P1-2 includes the following steps:

[0052] Step 1: Input the UAV reward allocation ratio , the data upload ratio of other ground users in the cluster except the currently optimized user object, the transmit power range , the maximum upload time , the maximum upload time , and proceed to step 2;

[0053] Step 2: Randomly initialize the transmit power of the ground user and the data upload ratio of the ground user , calculate the current ground user utility function , and set the initial temperature , and proceed to step 3;

[0054] Step 3: Generate a new solution by introducing small perturbations based on the current solution , calculate the ground user utility function value of the new solution If the new solution is better than the current solution, accept the new solution; if the new solution is worse than the current solution, accept the new solution with probability to avoid local optimality and enter step 4, where is the current temperature;

[0055] Step 4: Gradually reduce the temperature using the exponential annealing strategy and return to step 3; when the temperature drops to the set threshold or there is no improvement after several consecutive iterations, the method terminates and enters step 5;

[0056] Step 5: Output the optimal solution .

[0057] By randomly perturbing the solution within the search space and accepting the worse solution with a certain probability, it is possible to avoid falling into local optimality and finally converge to the global optimal solution.

[0058] Optionally, the P3-1 is solved by the particle swarm method, including the following steps:

[0059] Step 1: Input the set of ground users , the data upload ratio of ground users , the transmission power of ground users , the total bandwidth , the noise power , the signal attenuation factor and the maximum communication distance between the UAV and ground users , and enter step 2;

[0060] Step 2: Each particle consists of the UAV position and the maximum data acquisition delay . Initialize the number of particle swarms , the maximum number of iterations , the inertia weight , the cognitive coefficient , the social coefficient , the th particle position and velocity and enter step 3, where is the particle serial number, and its value range is , , , where and are the horizontal abscissa and horizontal ordinate of the th particle respectively, is the maximum data acquisition delay of the th particle, is the velocity of, The speed of is The speed of is;

[0061] Step 3: Set the individual optimal solution and the global optimal solution of each particle and enter Step 4;

[0062] Step 4: Calculate the objective function value of each particle and enter Step 5;

[0063] Step 5: Update the individual optimal solution and the global optimal solution and enter Step 6;

[0064] Step 6: Update the particle speed with the formula and update the particle position with the formula and enter Step 7; where is the inertia weight, which controls the search ability of the particle; and are the learning factors, which determine the degree to which the particle is affected by its own and the global optimal solutions; and are random numbers, which are used to increase randomness; is the historical optimal solution of the current particle; is the historical optimal solution of the entire population.

[0065] Step 7: Check whether the updated particle position satisfies the constraint conditions. If not, use the penalty function for adjustment and then enter Step 8. For particles that satisfy the constraint conditions, enter Step 8;

[0066] Step 8: Determine whether the maximum number of iterations or the optimal solution has converged. If not, return to Step 4. If so, enter Step 9;

[0067] Step 9: Take of as the optimal solution of the UAV position .

[0068] The present invention uses the particle swarm optimization algorithm to minimize the hovering time on the basis of the UAV collecting a fixed amount of data, thereby minimizing the hovering energy consumption during data collection, and further optimizing the UAV position.

[0069] Optionally, the P3-2 is solved by a direct solution method, including the following steps:

[0070] Step 1: Calculate the minimum value of under the constraint C10 condition​ ;

[0071] Step 2: Obtain the value range of in combination with constraint C7 ;

[0072] Step 3: , obviously when , this formula is monotonically increasing with respect to . Therefore, when takes the minimum value that satisfies the constraint, the objective function value of sub-problem P3-2 is the smallest. Therefore, the CPU frequency of the UAV has the optimal solution of .

[0073] Optionally, the P3-3 is solved by the golden section method, which includes the following steps:

[0074] Step 1: Input the association coefficient between the ground user and the UAV and the UAV position

[0075] Step 2: Initialize the upper bound of the reward allocation ratio of the UAV and the lower bound , and assign values to the golden section ratio . Then enter Step 3. Among them, the golden section ratio is a fixed value , and the initial values of and are 0 and 1;

[0076] Step 3: When the difference between the upper bound and the lower bound of is greater than 0.01, find two golden section points and , and then enter Step 4;

[0077] Step 4: For the golden section points and respectively, obtain the ratio of the amount of data that the ground users covered by the UAV can upload to the UAV under the current reward allocation ratio, and then enter Step 5; Step 5: Obtain the optimal position

[0078] of the UAV according to the ratio , and then enter Step 6; Step 6:

[0079] Step 6: According to the ratio , Drone location and drones The optimal solution for CPU frequency Computing drones The corresponding utility function and , go to step 7;

[0080] Step 7: Determine the utility function and The size of The upper bound of and lower bound Go to step 8;

[0081] Step 8, Calculate The upper bound of and lower bound If the absolute value of the difference is greater than 0.01, return to step 3. The upper bound of and the lower bound If the absolute value of the difference is less than or equal to 0.01, proceed to step 9;

[0082] Step 9, take and The larger value is the drone The optimal value of The optimal solution for reward distribution ratio . Beneficial Effects

[0083] (1) The present invention can optimize the rewards obtained by the drone and the ground user by playing a game between the drone utility function and the ground user utility function, and obtain the final optimal drone reward distribution ratio and ground user upload data ratio.

[0084] (2) The incentive mechanism of the Stackelberg game introduced in this invention, in which the UAV as the leader formulates the reward strategy, and the ground user as the follower determines the optimal data upload plan. The utility function of the follower is defined by comprehensively considering the rewards obtained by the ground user and the energy consumption of transmitting data. The leader utility function is composed of the rewards obtained by the local training model, the rewards allocated to the ground user, the hovering energy consumption and the training energy consumption, which effectively improves the efficiency of UAV data collection and federated learning.

[0085] (3) The present invention optimizes the follower's utility function using the simulated annealing method to obtain the optimal solution of the transmission power, decomposes the problem of maximizing the leader's utility function into three sub-problems: UAV position optimization, UAV computing resource optimization, and UAV reward allocation ratio optimization, and uses the particle swarm method, theoretical derivation method, and golden section method to solve them respectively, obtaining the optimal UAV position, local training CPU frequency, and reward allocation ratio, and further improving the accuracy of the calculation results. Description of the Drawings

[0086] Figure 1 The figure shows a schematic flow chart of the UAV data collection optimization method based on federated learning of the present invention;

[0087] Figure 2 The figure shows a schematic system model diagram corresponding to the UAV data collection optimization method based on federated learning of the present invention;

[0088] Figure 3 The figure shows a schematic diagram of the clustering result corresponding to the UAV data collection optimization method based on federated learning of the present invention;

[0089] Figure 4 The figure shows a schematic diagram of the average UAV utility simulation of the UAV data collection optimization method based on federated learning of the present invention under different comparison methods;

[0090] Figure 5 The figure shows a schematic diagram of the average ground user utility simulation of the UAV data collection optimization method based on federated learning of the present invention under different comparison methods;

[0091] Figure 6 The figure shows a schematic diagram of the average UAV utility simulation of the UAV data collection optimization method based on federated learning of the present invention at different UAV altitudes;

[0092] Figure 7 The figure shows a schematic diagram of the average ground user utility simulation of the UAV data collection optimization method based on federated learning of the present invention at different UAV altitudes;

[0093] Figure 8 The figure shows a schematic diagram of the average UAV utility simulation of the UAV data collection optimization method based on federated learning of the present invention at different data collection times;

[0094] Figure 9 The figure shows a schematic diagram of the average ground user utility simulation of the UAV data collection optimization method based on federated learning of the present invention at different data collection times;

[0095] Figure 10 The figure shows a schematic diagram of the average UAV utility simulation of the UAV data collection optimization method based on federated learning of the present invention at different numbers of ground users;

[0096] Figure 11 The following figure shows a schematic diagram of the average utility of ground users of the UAV data collection optimization method based on federated learning according to the present invention under different numbers of ground users. Detailed implementation manners

[0097] The following will be further described in conjunction with the accompanying drawings and specific embodiments. In the description of the present invention, it should be understood that terms such as "first", "second", etc. are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first", "second", etc. may explicitly or implicitly include one or more of such features.

[0098] Embodiment 1

[0099] This embodiment introduces a UAV data collection optimization method based on federated learning, which specifically includes the following contents:

[0100] When the UAV communicates with local users, the UAV data collection optimization method includes:

[0101] Obtain the pre-constructed problem of maximizing the UAV utility function and the problem of maximizing the ground user utility function; the problem of maximizing the ground user utility function includes the sub-problem of optimizing the association between the ground user and the UAV and the sub-problem of optimizing the ground user's data upload ratio and transmission power; the optimization problem of the UAV utility function is decomposed into the UAV position optimization sub-problem, the UAV computing resource optimization sub-problem, and the UAV reward allocation ratio optimization sub-problem;

[0102] Adopt a game algorithm to solve the problem of maximizing the UAV utility function and the problem of maximizing the ground user utility function, and calculate the optimal best pairing between the ground user and the UAV, the ground user's transmission power ratio and data upload ratio, the UAV position, the UAV computing resources, and the UAV reward allocation ratio.

[0103] Stackelberg game is a leader-follower game model, in which one party (the leader) makes a decision first, and the other party (the follower) makes a response after observing the leader's decision. The present invention jointly considers the interests of both the UAV and the ground users, and constructs the interest competition between the UAV and the ground users as a Stackelberg game model, where the UAV is the leader and the ground users are the followers.

[0104] The UAV has the power to determine the reward distribution ratio, which will directly affect the amount of data that ground users choose to offload to each UAV. As the amount of data uploaded by ground users to the UAV changes, the UAV will dynamically adjust its reward distribution ratio according to the received data volume to maximize its own benefits during the collaboration process; conversely, the adjusted reward distribution ratio of the UAV will affect the ground users' decision on the data upload ratio, and both sides will conduct a game based on the reward distribution ratio and the data upload ratio. Therefore, it is necessary to optimize the utility functions of the followers and the leader to obtain the final optimal UAV reward distribution ratio and the ground user data upload ratio.

[0105] Embodiment 2

[0106] Based on Embodiment 1, this embodiment introduces the specific implementation process of an optimization method for UAV data collection based on federated learning, as Figure 1 shown, which specifically includes the following content:

[0107] Define the data transaction for the UAV and ground users to communicate as follows:

[0108] Under a given data volume, the UAV determines the corresponding reward distribution ratio decision according to the channel transmission environment and its own needs, and sends a data collection request to the ground users. After receiving the UAV's data collection request, the ground users optimize the data upload ratio according to the allocated rewards, so as to improve their own benefits obtained during the collaboration process.

[0109] The UAV data collection optimization method further includes:

[0110] Define the optimization problem of the ground user utility function as P1, define the ground user and UAV association optimization sub-problem as P1-1, and define the ground user data upload ratio and transmit power optimization sub-problem as P1-2;

[0111] The expression of P1 is as follows:

[0112] ,

[0113] In the formula, C1 is the range of the ground user's data upload ratio, C2 is the limit of the transmit power for the ground user to transmit data, C3 indicates that the association between the ground user and the UAV needs to meet a certain distance, C4 indicates that the time consumed for the ground user to transmit data is limited, C5 indicates that a ground user can only be associated with one UAV, and C6 indicates that only when the income is positive can the ground user participate in the contribution;

[0114] Among them, is the ground user utility function, is the ground user number, and the value range is 1-N, is the UAV number, and its value range is 1-M, is the set of all UAVs, is the th set of ground users within the coverage of the UAV, is the set of all ground users, is the ground user 's transmission power, is the ground user 's upload data ratio, is the UAV 's association value with the ground user , indicates that the UAV covers the ground user , indicates that the UAV does not cover the ground user , indicates the reward allocation ratio of the UAV to the ground user, is the reward obtained by the UAV , represents the unit cost of the energy consumption for the ground user to upload data, reflecting the importance of the ground user to the transmission energy consumption, is the ground user 's data volume, is the ground user 's transmission bandwidth, is the distance between the UAV and the ground user , is the reference distance at which the channel gain in the wireless channel is, is the path loss exponent, is the maximum transmission distance under the UAV coverage, is the power of the Gaussian white noise, and are respectively the minimum and maximum transmission powers of the ground user , is the maximum data upload time;

[0115] The expression of P1-1 is as follows:

[0116] ,

[0117] The expression of P1-2 is as follows:

[0118] .

[0119] Solving the P1-1 includes the following steps:

[0120] Step 1: Input the set of ground users , the transmission power of ground users , the total bandwidth , the moving range of the UAV , the maximum number of ground users that can be carried and the coverage radius of the UAV , the clustering radius, a parameter in the clustering method and the minimum number of samples , and then proceed to Step 2;

[0121] Step 2: Cluster the ground users using the original density-based clustering method, and then proceed to Step 3;

[0122] Step 3: Calculate the diameter of the clusters after clustering . If is greater than the coverage diameter of the UAV , reduce the clustering radius and return to Step 2; if is less than or equal to the coverage diameter of the UAV , the clustering is successful, and proceed to Step 4;

[0123] Step 4: When a ground user is within the coverage ranges of two or more UAVs, based on the determined transmission power and data upload ratio , traverse all UAVs that meet the constraint conditions, calculate the ground user utility function , and select the association method corresponding to the maximum ground user utility function to determine the best pairing between the ground user and the UAV .

[0124] The solution to the P1-2 includes the following steps:

[0125] Step 1: Input the UAV reward allocation ratio , the data upload ratios of other ground users in the cluster except the currently optimized user object, the transmission power range , the maximum upload time , and the maximum upload time , and then proceed to Step 2;

[0126] Step 2: Randomly initialize the transmission power of the ground users and the data upload ratio of the ground users, calculate the current ground user utility function , and set the initial temperature , proceed to Step 3;

[0127] Step 3: Generate a new solution by introducing a small perturbation based on the current solution , calculate the value of the ground user utility function for the new solution , if the new solution is better than the current solution, accept the new solution; if the new solution is worse than the current solution, then with probability accept the new solution and proceed to Step 4, where is the current temperature;

[0128] Step 4: Gradually reduce the temperature using the exponential annealing strategy and return to Step 3; when the temperature drops to the set threshold or there is no improvement after several consecutive iterations, the method terminates and proceeds to Step 5;

[0129] Step 5: Output the optimal solution .

[0130] In a specific implementation manner of the embodiment of the present invention, the UAV data acquisition optimization method further includes:

[0131] Define the optimization problem of the UAV utility function as P2, simplify the problem P2 using variable substitution to be represented as P3, define the UAV position optimization sub-problem as P3-1, define the UAV computing resource optimization sub-problem as P3-2, and define the UAV reward allocation ratio optimization sub-problem as P3-3;

[0132] The expression of P2 is as follows:

[0133]

[0134] In the formula, C7 represents the CPU frequency that restricts the UAV's participation in federated learning training, C8 represents the UAV's position range, C9 represents that the reward given by the UAV to the ground user cannot be greater than the reward received by the UAV, and C10 represents that the total time delay of the UAV's local training shall not exceed ;

[0135] is the position of the UAV , is the position of the ground user , is the UAV 's CPU frequency, represents the unit cost of the energy consumption of the UAV's local model training, represents the unit cost of the energy consumption generated by the UAV's hovering, respectively reflecting the importance of the UAV's energy consumption for training the model and the hovering energy consumption, is the global iteration number of federated learning, is the UAV 's local iteration number, is the energy coefficient for the CPU is the number of CPU cycles required for the UAV to perform calculations per bit is the hovering power of the UAV and are the minimum and maximum CPU frequencies of the UAV respectively is the total delay for the UAV to perform local training;

[0136] Since the objective function in P2 contains , an auxiliary variable is introduced, and the constraint C11

[0137]

[0138] is added, so that the problem P2 is simplified and expressed as P3. The expression of P3 is as follows:

[0139] ,

[0140] The expression of P3-1 is as follows:

[0141] ,

[0142] The expression of P3-2 is as follows:

[0143] ,

[0144] The expression of P3-3 is as follows:

[0145] .

[0146] The P3-1 is solved by the particle swarm method, including the following steps:

[0147] Step 1, input the set of ground users , the data upload ratio of ground users , the transmit power of ground users , the total bandwidth , the noise power , the signal attenuation factor and the maximum communication distance between the UAV and ground users , and enter Step 2;

[0148] Step 2, each particle consists of the UAV position and the maximum data acquisition delay . Initialize the number of particle swarms , the maximum number of iterations , Inertia weight , Cognitive coefficient , Social coefficient , The th particle's position , and velocity Enter step 3, where is the particle serial number, and its value range is , , , where and are respectively the horizontal abscissa and the horizontal ordinate of the th particle, is the maximum data acquisition delay of the th particle, is 's velocity, is 's velocity, is 's velocity;

[0149] Step 3, Set the individual optimal solution and the global optimal solution of each particle, and enter step 4;

[0150] Step 4, Calculate the objective function value of each particle, and enter step 5;

[0151] Step 5, Update the individual optimal solution and the global optimal solution of each particle, and enter step 6;

[0152] Step 6, Update the particle velocity with the formula , and update the particle position with the formula , and enter step 7; where is the inertia weight, which controls the search ability of the particle; and are learning factors, which determine the degree to which the particle is affected by its own and the global optimal solutions; and are random numbers, which are used to increase randomness; is the historical optimal solution of the current particle; is the historical optimal solution of the entire population.

[0153] Step 7, Check whether the updated particle position satisfies the constraint conditions. If not, use the penalty function for adjustment and then enter step 8. For particles that satisfy the constraint conditions, enter step 8;

[0154] Step 8: Determine whether the maximum number of iterations or the optimal solution convergence has been reached. If not, return to Step 4; if so, proceed to Step 9.

[0155] Step 9: Take of as the optimal solution for the UAV position .

[0156] Solve the P3-2 through a direct solution method, including the following steps:

[0157] Step 1: Calculate the minimum value of under the constraint C10 ;

[0158] Step 2: Combine the constraint C7 to obtain the value range of ;

[0159] Step 3: , obviously when , this formula is monotonically increasing with respect to . Therefore, when taking the minimum value that satisfies the constraint, the objective function value of the sub-problem P3-2 is the smallest. Therefore, the optimal solution for the CPU frequency of the UAV is .

[0160] Solve the P3-3 through the golden section method, including the following steps:

[0161] Step 1: Input the correlation coefficient between the ground user and the UAV and the UAV position , and proceed to Step 2;

[0162] Step 2: Initialize the upper bound of the reward allocation ratio of the UAV and the lower bound , and assign values to the golden section ratio . Then, proceed to Step 3; among them, the golden section ratio is a fixed value , and are initialized to 0 and 1;

[0163] Step 3: When the difference between the upper bound and the lower bound of is greater than 0.01, find two golden section points and , and proceed to Step 4;

[0164] Step 4: For the golden section points and , obtain the proportion of the data volume that the ground users can upload to the drone under the current reward allocation ratio for the ground users covered by the drone , and proceed to Step 5; Proportion , enter Step 5;

[0165] Step 5: Obtain the optimal position of the drone according to the proportion , and proceed to Step 6;

[0166] Step 6: Calculate the utility function corresponding to the drone according to the proportion , the position of the drone and the optimal solution of the CPU frequency of the drone , and proceed to Step 7; Utility function and , enter Step 7;

[0167] Step 7: Judge the magnitudes of the utility functions and , so as to update the upper bound and the lower bound and enter Step 8; Enter Step 8;

[0168] Step 8: Calculate the difference between the upper bound and the lower bound of . If the absolute value of the difference is greater than 0.01, return to Step 3. If the upper bound and the lower bound of

[0169] the absolute value of the difference is less than or equal to 0.01, proceed to Step 9; and Step 9: Take the larger value of as the optimal value of the drone , and the corresponding solution as the optimal solution of the reward allocation ratio of the drone .

[0170] In a specific implementation manner of the embodiment of the present invention, the game algorithm is adopted to solve the problem of maximizing the utility function of the UAV and the problem of maximizing the utility function of the ground users, and the optimal best pairing of the ground users and the UAV, the ground user transmission power ratio, the ground user data upload ratio, the UAV position, the UAV computing resources, and the UAV reward allocation ratio are calculated, including obtaining the finally optimized UAV utility function and ground user utility function through game solution of P1-1, P1-2, P3-1, P3-2, and P3-3, and obtaining the finally optimal best pairing of the ground users and the UAV, the ground user transmission power ratio, the ground user data upload ratio, the UAV position, the UAV computing resources, and the UAV reward allocation ratio based on the finally optimized UAV utility function and ground user utility function, specifically including the following steps:

[0171] Step 1, initialize all UAV sets 's data reward allocation ratio, all ground users on the ground 's data upload ratio and transmission power, and enter Step 2;

[0172] Step 2, use an improved method based on density-based clustering to solve the UAV position set , clustering results and the number of UAVs , and enter Step 3;

[0173] Step 3, each UAV uses the particle swarm method, theoretical derivation, and golden section method to optimize the UAV position , computing resources and reward allocation ratio in turn, and enter Step 4; where represents the number of iterations;

[0174] Step 4, when a ground user is within the coverage of two or more UAVs, traverse all UAVs that meet the constraint conditions, calculate the ground user utility function , and take the maximum ground user utility function corresponding association method to determine the best pairing between the ground user and the UAV , and enter Step 5;

[0175] Step 5, for each ground user covered by each UAV respectively, use the simulated annealing algorithm to optimize the ground user transmission power and the ground user data upload ratio , and enter Step 6;

[0176] Step 6: Calculate the utility of each UAV. When the utility of each UAV does not meet the given accuracy, return to Step 3 until the utility of each UAV reaches the accuracy, and then stop, so as to obtain the finally optimized UAV utility function for each UAV. and the finally optimized utility function for each ground user. Furthermore, obtain the final optimal best pairing between ground users and UAVs, the data upload ratio of ground users, the transmission power ratio of ground users, the UAV positions, the UAV computing resources, and the UAV reward allocation ratio.

[0177] Among them, the best pairing between ground users and UAVs satisfies ; the data upload ratio of ground users satisfies ; the transmission power ratio of ground users satisfies ; the UAV positions satisfy ; the UAV computing resources satisfy ; the UAV reward allocation ratio satisfies .

[0178] The UAV-assisted data collection federated learning system model proposed by the present invention is as Figure 2 shown. This system model consists of randomly distributed ground users and UAVs. The horizontal position of ground user is represented by , and the horizontal position of UAV is represented by , both are evenly distributed in the area . The UAVs are at a fixed safe altitude above the target area, and collect data from ground users in this area and perform federated learning.

[0179] To reduce the flight energy consumption during the data collection process of UAVs, first cluster the ground users, and deploy one UAV in each cluster for data collection, as Figure 3 shown. After the collection is completed, the UAVs return to the base, and use the data of the users in the cluster collected to perform local model training, and then the elected UAVs complete the global model aggregation. To ensure the security of data transactions, blockchain technology is introduced, and a tightly coupled architecture is adopted. The UAVs not only undertake the local training tasks of federated learning, but also participate in consensus verification and data security guarantee as blockchain nodes.

[0180] The various model mechanisms and specific formulas involved in this embodiment are as follows:

[0181] ​Federated Learning: After the drones complete data collection, they start the local training task. The elected drones act as the model aggregation server. First, the global model is initialized and broadcast. After receiving the global model, the drones train the received model according to the locally collected dataset, and then upload the updated model after training back to the aggregation drone for weighted aggregation of the model.

[0182] Parameterize the local model of the drone as . The model for local training of the drones participating in federated learning is a loss function that measures the difference between the output of the local model and the ground truth labels on the local dataset . The expression of the loss function is as follows: In the formula,

[0183]

[0184] where , represents the sum of the data uploaded by the ground users in the th cluster, represents the loss function associated with the data point , and m is the drone serial number, with a value range of 1 - M.

[0185] To minimize the loss function, in this embodiment, the Adam optimizer is used to solve it in parallel by each drone. At each global iteration , each drone initializes and updates the local model through the following formula:

[0186] ,

[0187] In the formula, is the learning rate.

[0188] After local training is completed, a drone is selected from the drones participating in training as the aggregation drone in turn. Each drone generates the corresponding model and sends it to the aggregation drone for aggregation of the global model. The loss of the global model is defined as

[0189]

[0190] where . Therefore, the FL process can be represented as the following optimization problem:

[0191] .

[0192] ​To address the issue that the traditional federated averaging method (FedAvg) is strongly influenced by the quality of local models and the authenticity of the amount of contributed data when aggregating local models, this paper will draw on the aggregation method and introduce the local model accuracy and data volume to assist in aggregation. Therefore, the expression of the optimized model aggregation method is as follows:

[0193] ,

[0194] In the formula, is the local model quality of the drone , is the local model quality of the drone .

[0195] In practical applications, a threshold can be set to indicate that the training result meets the expectation. When , the model of global federated learning has reached a good performance level and the task is completed. is the parameter of the federated learning model for the th global iteration. The global iteration number can also be set, and the training ends after reaching the global iteration number. In this paper, the second method is adopted for federated learning.

[0196] Communication model: Denote the cluster covered by the drone as cluster, denote the number of ground users in the cluster as , denote the data volume of each user as . For the ground user

[0197] ,

[0198] In the formula, is the data upload ratio of the ground user , . When , it means not uploading any data. When , it means that all the local data of the ground user is uploaded. When , data is selected for upload by means of random sampling.

[0199] The ground users and the drone communicate in a frequency-division multiple access manner. The total communication bandwidth of each drone is , which is evenly distributed to the covered ground users. Therefore, the bandwidth of the ground user is expressed as follows:

[0200] ,

[0201] Unmanned aerial vehicle and the ground user The distance is

[0202] ,

[0203] Wherein, and are the horizontal abscissas of the ground user and the unmanned aerial vehicle respectively, and are the horizontal ordinates of the ground user and the unmanned aerial vehicle respectively.

[0204] Assume that the maximum transmission distance between the unmanned aerial vehicle and the ground user is , then

[0205]

[0206] Unmanned aerial vehicle and the ground user The channel gain The expression is as follows:

[0207]

[0208] Wherein, Is the channel gain magnitude in the wireless channel at the reference distance Meters, Is the path loss exponent. Therefore, the unmanned aerial vehicle and the ground user The data transmission rate (in bits per second) between them can be further expressed by the following expression Expressed as:

[0209]

[0210] Wherein, Is the unmanned aerial vehicle Is the bandwidth allocated to the users within the coverage Is the transmitting power of the ground user

[0211] Is the power of Gaussian white noise.

[0212] Delay and energy consumption calculation:

[0213] (1) Delay

[0214] During the data acquisition and model training process, it always includes user delay and UAV delay. User delay refers to the time for each user to upload data, and UAV delay includes data acquisition delay and local training delay.

[0214] User delay is the time when the user uploads data, and the expression is as follows:

[0215] ,

[0216] The maximum data acquisition time is limited to , so .

[0217] The UAV has a time delay for data acquisition. After the data acquisition is completed, it can return for local training. The data acquisition time delay is the maximum time delay for users within the coverage area to upload data, and the expression is as follows:

[0218] ,

[0219] The UAV The amount of data collected is the sum of the data uploaded by the ground users within the coverage area, and the expression is as follows:

[0220] ,

[0221] The UAV The total time delay for local training The expression is as follows:

[0222] ,

[0223] In the formula, is the global iteration number of UAV m, is the local iteration number of UAV m, represents the number of CPU cycles required for UAV to calculate each sample, represents the CPU frequency of UAV for local training.

[0224] (2) Energy consumption

[0225] The ground user The energy consumption during data upload The expression is as follows:

[0226] ,

[0227] The UAV For The total energy consumption of local training for the global iteration is calculated by the following formula:

[0228] ,

[0229] wherein, is the local iteration number of the UAV m, represents the UAV the number of CPU cycles required for each bit of calculation, represents the UAV the CPU frequency for local training, represents the energy coefficient used by the CPU.

[0230] When the UAV hovers in the air during data collection, hover energy consumption is generated, and represents the hover power of the UAV, that is, the energy consumption generated by hovering per unit time. Then the hover energy consumption during data collection is expressed as follows:

[0231] .

[0232] To verify the performance of the method proposed in the present invention, the present invention considers comparing with the following four methods:

[0233] Comparison method 1: This method does not consider the interest competition between the UAV and the ground users, and the UAV directly collects all the data of the ground users.

[0234] Comparison method 2: The position of the UAV is fixed, and the horizontal position of the UAV is the centroid of the ground users in the current cluster.

[0235] Comparison method 3: The CPU frequency of the UAV for local training is random, that is, the CPU frequency of the UAV for local training is not optimized, and it randomly takes values between .

[0236] Comparison method 4: The transmission power of the ground users is random, that is, the transmission power of the ground users is not optimized, and it randomly takes values between .

[0237] Figure 4 and Figure 5 give the performance comparison between the method proposed in the present invention and the other four methods under different total bandwidths. Among them, Figure 4 compares the method proposed in the present invention with comparison methods 1-3, and compares the utility of the UAV under different scenarios. Figure 5 The method proposed in the present invention is compared with comparison methods 1 and 4, and the total utility of the ground users under different scenarios is compared. Figure 4 and Figure 5 In, Algorithm 1 - Algorithm 4 refer to comparison methods 1 - 4 introduced in this embodiment, and the algorithm in this paper refers to the UAV data collection optimization method based on federated learning introduced in the present invention. As Figure 4 shown, as the total bandwidth increases, the average revenue of the UAVs of the four methods all shows an upward trend, asFigure 5 As shown, the average benefits of ground users under the three methods all show a downward trend. This is because when the total bandwidth increases, the bandwidth allocated to each user increases, the transmission rate of the user increases, and more data is transmitted. In order to maximize the utility, the drone will reduce the reward allocation ratio. Therefore, the average utility of the drone increases with the increase of the bandwidth, while the average utility of the ground user decreases with the increase of the bandwidth.

[0238] As can be seen from Figure 4 and Figure 5 , the method proposed by the present invention brings the highest utility to the drone and the ground user. At the same time, it can be seen that the situation where the CPU frequency of the drone's local training is random brings the lowest utility to the drone, and the situation where the transmission power of the ground user is random also brings the lowest utility to the user. This is because random values cannot effectively utilize computing and power resources. Compared with Comparative Method 1, the solution proposed by the present invention can effectively improve the utility of the ground user and the drone. This is because in Comparative Method 1, there is no game relationship between the drone and the ground user, the ratio of the data uploaded by the ground user is fixed at 1, the energy consumption of the drone for local training and the energy consumption of the ground user for data transmission both increase, and the average utility of the drone and the ground user both decrease. Therefore, the present invention can bring higher utility to the drone. Compared with the case where the drone's position is fixed, the present invention optimizes the drone's position so that the drone can collect more data at a lower cost under the same conditions, and the average utility of the drone is higher.

[0239] Figure 6 and Figure 7 show the influence of different hovering heights of the drone on the utility of the drone and the ground user. As can be seen from Figure 6 and Figure 7 , the average utility of the drone decreases with the increase of the drone's hovering height, and the average utility of the ground user increases with the increase of the drone's hovering height. This is because the increase in the drone's height reduces the channel gain between the drone and the ground user, increases the cost of the user uploading data, reduces the amount of data transmitted, and the drone will increase the reward allocation ratio given to the user in order to obtain more data. Therefore, the average benefit of the drone decreases and the average benefit of the user increases.

[0240] Figure 8 and Figure 9 show the comparison chart of the average utility of the drone and the average utility of the user under different maximum data acquisition time delays. Figure 8 and Figure 9 The T_max in Figure 8It can be seen that the change in the maximum acquisition delay has little impact on the average utility of the UAV. This is because when the maximum acquisition time becomes larger, more data is uploaded by the ground users. To maximize its utility, the UAV will reduce the reward allocation ratio, resulting in a decrease in the expenditure for users. However, the hovering energy consumption increases, and the local training energy consumption also increases. Therefore, the average utility of the UAV changes little. From Figure 9 It can be seen that the average utility of users decreases as the maximum acquisition time increases. This is because the UAV motivates users to contribute data with a smaller reward allocation ratio, leading to a decrease in the average income of ground users. At the same time, the upload time of ground users increases, and the transmission energy consumption increases. Therefore, the average utility of the UAV decreases.

[0241] Figure 10 and Figure 11 give the comparison charts of the average utility of the UAV and the average utility of users with different numbers of ground users. From Figure 10 and Figure 11 It can be seen that the average utilities of the UAV and ground users decrease as the number of ground users increases. This is because with the increase in the number of ground users, the UAV collects more data in order to obtain more benefits during the competition process, resulting in an increase in their respective training energy consumption and reward allocation ratio, thus reducing the average utility. The increase in the number of ground users participating in data upload also reduces the average utility of ground users.

[0242] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit and scope protected by the present invention and the claims. These all fall within the protection scope of the present invention.

Claims

1. An optimization method for UAV data collection based on federated learning, characterized in that When the drone communicates with local users, the drone data collection optimization method includes: Obtain the pre-constructed problems of maximizing the drone utility function and the ground user utility function; the problem of maximizing the ground user utility function includes the sub-problem of optimizing the association between the ground user and the drone and the sub-problem of optimizing the data upload ratio and transmission power of the ground user; the optimization problem of the drone utility function is decomposed into the sub-problem of optimizing the drone position, the sub-problem of optimizing the drone computing resources, and the sub-problem of optimizing the drone reward allocation ratio; Adopt a game algorithm to solve the problems of maximizing the drone utility function and the ground user utility function, and calculate the optimal best pairing between the ground user and the drone, the ground user transmission power ratio and the ground user data upload ratio, the drone position, the drone computing resources, and the drone reward allocation ratio.

2. The method for optimizing drone data collection according to claim 1, characterized in that, The data transaction when the drone communicates with the ground user includes the following: Under a given amount of data, the drone determines the corresponding reward allocation ratio decision according to the channel transmission environment and its own needs, and sends a data collection request to the ground user. After receiving the data collection request from the drone, the ground user optimizes the data upload ratio according to the allocated reward, so as to increase the income obtained during the cooperation process.

3. The method for optimizing drone data collection according to claim 1, wherein The drone data collection optimization method further includes: Define the optimization problem of the ground user utility function as P1, define the sub-problem of optimizing the association between the ground user and the drone as P1-1, and define the sub-problem of optimizing the data upload ratio and transmission power of the ground user as P1-2; The expression of P1 is as follows: ， In the formula, C1 is the range of the data upload ratio of the ground user, C2 is the limit of the transmission power of the ground user's transmitted data, C3 indicates that the association between the ground user and the drone needs to meet a certain distance, C4 indicates the limit of the time consumed by the ground user to transmit data, C5 indicates that a ground user can only be associated with one drone, and C6 indicates that the ground user can participate in the contribution only when the income is positive; Among them, is the ground user utility function, is the ground user number, and its value range is from 1 to N, is the UAV number, and its value range is from 1 to M, is the set of all UAVs, is the set of ground users within the coverage of the th UAV, is the set of all ground users, is the transmission power of ground user , is the upload data ratio of ground user , is the association value between UAV and ground user , indicates that UAV covers ground user , indicates that UAV does not cover ground user , indicates the reward allocation ratio of UAV to the ground user, is the reward obtained by UAV , represents the unit cost of energy consumption for ground users to upload data, reflecting the importance that ground users attach to transmission energy consumption, is the data volume owned by ground user , is the transmission bandwidth of ground user , is the distance between UAV and ground user , is the reference distance meters, and it is the channel gain magnitude in the wireless channel, is the path loss exponent, is the maximum transmission distance under the UAV coverage, is the power of Gaussian white noise, and are respectively the minimum and maximum transmission powers of ground user , is the maximum data upload time; The expression of P1-1 is as follows: ， The expression of P1-2 is as follows: 。 4. The method for optimizing drone data collection according to claim 3, wherein The drone data collection optimization method further includes: Define the optimization problem of the drone utility function as P2, simplify the problem P2 using variable substitution and represent it as P3, define the sub-problem of optimizing the drone position as P3-1, define the sub-problem of optimizing the drone computing resources as P3-2, and define the sub-problem of optimizing the drone reward allocation ratio as P3-3; The expression of P2 is as follows: ， Wherein, C7 represents the CPU frequency that restricts the participation of the UAV in federated learning training, C8 represents the position range of the UAV, C9 represents that the reward given by the UAV to the ground user cannot be greater than the reward received by the UAV, and C10 represents that the total delay of local training of the UAV shall not exceed ; is the position of the drone ; is the position of the ground user ; is the CPU frequency of the drone ; represents the unit cost of the energy consumption for local model training of the drone and represents the unit cost of the energy consumption generated by the drone hovering, respectively reflecting the importance degrees of the drone on the energy consumption for training the model and the hovering energy consumption is the global iteration number of federated learning is the local iteration number of the drone ; is the energy coefficient used by the CPU is the drone The number of CPU cycles required for each bit of calculation is the hovering power of the drone and are respectively the minimum and maximum CPU frequencies of the drone ; is the total time delay for the drone to perform local training; Since the objective function in P2 contains , an auxiliary variable is introduced and the constraint C11 is added ， Thus, simplify the problem P2 and represent it as P3. The expression of P3 is as follows: ， The expression of P3-1 is as follows: ， The expression of P3-2 is as follows: ， The expression of P3-3 is as follows: 。 5. The method for optimizing UAV data collection according to claim 4, wherein The game algorithm is adopted to solve the problems of maximizing the utility function of the UAV and the utility function of the ground user, and calculate the optimal best pairing of the ground user and the UAV, the ground user transmission power ratio, the ground user data upload ratio, the UAV position, the UAV computing resources, and the UAV reward allocation ratio. It includes performing game solutions on P1-1, P1-2, P3-1, P3-2, and P3-3 to obtain the finally optimized UAV utility function and ground user utility function, and obtaining the finally optimal best pairing of the ground user and the UAV, the ground user transmission power ratio, the ground user data upload ratio, the UAV position, the UAV computing resources, and the UAV reward allocation ratio based on the finally optimized UAV utility function and ground user utility function. Specifically, it includes the following steps: Step 1, initialize all drone sets The data reward allocation ratio, all ground users on the ground The data upload ratio and transmission power, and enter Step 2; Step 2: Use an improved method based on density-based clustering to solve the UAV position set , the clustering result and the number of UAVs , and enter Step 3; Step 3: Each drone optimizes its position in turn using the particle swarm method, theoretical derivation, and the golden section method , computing resources , and the reward allocation ratio , and then proceeds to Step 4; where represents the number of iterations; Step 4: When a ground user is within the coverage of two or more drones, traverse all drones that meet the constraint conditions, calculate the utility function of the ground user , and take the maximum utility function of the ground user . The corresponding association method is used to determine the ground user and the drone 's optimal pairing , and proceed to Step 5; Step 5: For each ground user covered by each drone, use the simulated annealing algorithm to optimize the transmission power of the ground user and the ground user data upload ratio , and proceed to Step 6; Step 6: Calculate the utility of each UAV. When the utility of each UAV does not meet the given accuracy, return to Step 3 until the utility of each UAV reaches the accuracy and then stop, so as to obtain the finally optimized UAV utility function of each UAV and the finally optimized utility function of each ground user , and further obtain the final optimal best pairing of ground users and UAVs, the data upload ratio of ground users and the transmit power ratio of ground users, the UAV positions, the UAV computing resources, and the UAV reward allocation ratio; Among them, the optimal pairing between the ground user and the UAV satisfies ; the data upload ratio of the ground user satisfies ; the transmission power ratio of the ground user satisfies ; the UAV position satisfies ; the computing resources of the UAV satisfy ; the reward allocation ratio of the UAV satisfies .

6. The method for optimizing drone data collection according to claim 5, wherein, Solving the P1-1 includes the following steps: Step 1, input the set of ground users , the transmission power of ground users , the total bandwidth , the moving range of the UAV , the maximum number of ground users that can be carried and the coverage radius of the UAV , the parameter clustering radius in the clustering method and the minimum number of samples , enter Step 2; Step 2: Cluster the ground users using the original density-based clustering method, and proceed to Step 3; ​ Step 3, calculate the cluster diameter after clustering , if is greater than the UAV coverage diameter , reduce the clustering radius and return to Step 2; if is less than or equal to the UAV coverage diameter , the clustering is successful and proceed to Step 4; Step 4, when a ground user is within the coverage of two or more drones, based on the transmit power and the data upload ratio determined, traverse all drones that meet the constraint conditions, calculate the ground user utility function , and take the maximum ground user utility function corresponding association method to determine the best pairing between the ground user and the drone .

7. The method for optimizing UAV data acquisition according to claim 5, characterized in that Solving the P1-2 includes the following steps: Step 1, input the drone Reward allocation ratio , cluster Except for the currently optimized user object within the cluster, the upload data ratio of other ground users , transmission power range , maximum upload time , proceed to Step 2; Step 2, randomly initialize the transmission power of the ground user and the ratio of the data uploaded by the ground user , calculate the current utility function of the ground user , and set the initial temperature , enter Step 3; Step 3: Introduce small perturbations to the current solution to generate a new solution , and calculate the value of the ground user utility function for the new solution . If the new solution is better than the current solution, accept the new solution; if the new solution is worse than the current solution, accept the new solution with probability , and enter Step 4, where is the current temperature Step 4, gradually reduce the temperature using the exponential annealing strategy, and return to step 3; when the temperature drops to the set threshold or there is no improvement after several consecutive iterations, the method terminates and enters step 5; Step 5, output the optimal solution .

8. The method for optimizing UAV data collection according to claim 5, characterized in that Solving the P3-1 using the particle swarm method includes the following steps: Step 1, input the set of ground users , the data upload ratio of ground users , the transmission power of ground users , the total bandwidth , the noise power , the signal attenuation factor and the maximum communication distance between the UAV and ground users , proceed to Step 2; Step 2, each particle is composed of the UAV position and the maximum data acquisition time delay to initialize the number of particle swarms , the maximum number of iterations , the inertia weight , the cognitive coefficient , the social coefficient , the th particle position , and velocity to enter Step 3, where is the particle sequence number, and the value range is , , , where and are respectively the horizontal abscissa and horizontal ordinate of the th particle, is the maximum data acquisition time delay of the th particle, is 's velocity, is 's velocity, is 's velocity; Step 3, set the individual optimal solution for each particle and the global optimal solution , and proceed to Step 4; Step 4, calculate the objective function value of each particle , and proceed to Step 5; Step 5, update the individual optimal solution and the global optimal solution , and proceed to Step 6; Step 6, update the particle velocity with the formula and update the particle position with the formula , then enter Step 7; where is the inertia weight, which controls the search ability of the particle; and are the learning factors, which determine the degree to which the particle is affected by its own and the global optimal solutions; and are random numbers, which are used to increase randomness; is the historical optimal solution of the current particle; is the historical optimal solution of the entire population; Step 7, check the updated particle positions to see if the constraint conditions are met. If not, use the penalty function for adjustment and then proceed to Step 8. For particles that meet the constraint conditions, proceed to Step 8; Step 8, determine whether the maximum number of iterations or the optimal solution convergence is reached. If not, return to step 4. If so, enter step 9; Step 9, take of as the optimal solution for the UAV position .

9. The method for optimizing drone data collection according to claim 5, wherein Solving the P3-2 using the direct solution method includes the following steps: Step 1, calculate the minimum value under the constraint C10 ; ; Step 2, obtain the value range of ; Step 3, , obviously when , this formula is monotonically increasing with respect to , so when taking the minimum value that satisfies the constraint, the objective function value of sub-problem P3-2 is the smallest. Therefore, when the CPU frequency of the UAV takes the optimal solution of is .

10. The method for optimizing drone data collection according to claim 5, wherein, Solving the P3-3 using the golden section method includes the following steps: Step 1, input the correlation coefficient between the ground user and the drone and the drone position , and proceed to Step 2; Step 2, initialize the drone Reward allocation ratio of Upper bound of And lower bound And assign values to the golden ratio Then enter Step 3; among them, the golden ratio Is a fixed value , And The initial values of are 0 and 1; Step 3, when the upper bound of and the lower bound of differ by more than 0.01, find two golden section points and , and proceed to Step 4; Step 4, respectively for the golden section points and , obtain the proportion of the data volume that the ground users under the coverage of the drone can upload to the drone at the current reward allocation ratio for the ground users, and enter Step 5; Proportion of the data volume ​ Step 5, according to the ratio obtain the optimal position of the drone , and proceed to Step 6; Step 6, according to the ratio and the UAV position and the CPU frequency of the UAV optimal solution calculate the UAV corresponding utility function and , go to Step 7; Step 7, determine the utility function and to update the upper bound and the lower bound and enter Step 8; Step 8, calculate the upper bound of and the lower bound the difference between them. If the absolute value of the difference is greater than 0.01, return to Step 3. If the upper bound of and the lower bound the absolute value of the difference between them is less than or equal to 0.01, proceed to Step 9; Step 9, take and the larger value as the optimal value of the drone and the corresponding solution as the optimal solution of the reward allocation ratio of the drone .​

Images