Resource allocation optimization method for unmanned aerial vehicle power inspection

Abstract

The invention discloses a resource allocation optimization method for unmanned aerial vehicle power inspection. The method comprises the following steps: forming a federated learning system by an unmanned aerial vehicle executing an inspection task, a nearby edge server and a nearby ground base station; the unmanned aerial vehicle uploads edge server parameters to the ground base station; the ground base station calculates the optimal calculation load of each edge server according to the edge server parameters; the ground base station calculates an edge server most suitable for transmission of the unmanned aerial vehicle as an optimal edge server transmission node and broadcasts the optimal edge server transmission node to the unmanned aerial vehicle, and the unmanned aerial vehicle sends data to the edge server according to the optimal edge server transmission node and the optimal calculation load; the ground base station broadcasts the initial model parameters to the edge server to start a new round of federated learning model training; and the ground base station judges whether the model training reaches the optimal model coordination round number or not, and if yes, global model training is stopped. The task allocation of the unmanned aerial vehicle is reasonably scheduled, so that the total communication cost of the system is reduced, and the operation efficiency and inspection quality of the power system are improved.

Description

Technical Field

[0001] The present invention relates to the field of UAV power inspection, and specifically to an optimization method for resource allocation for UAV power inspection. Background Art

[0002] With the continuous growth of the power system's demand for equipment inspection and maintenance, UAVs, as an efficient inspection tool, have been widely used in the monitoring and maintenance of power facilities. UAVs have high flexibility, a wide coverage range, and low costs, and can collect data in real time in complex environments. However, with the increase in the number of UAVs and the improvement of task complexity, how to efficiently allocate the computing tasks of UAVs has become an urgent problem to be solved.

[0003] In the power system, UAVs will pass through multiple ground relay nodes during flight, including ground base stations (abbreviated as BSs) and ground nodes (abbreviated as GNs). UAVs can transfer computing tasks to these ground relay nodes through two methods: licensed band transmission (abbreviated as LBT) or unlicensed bandwidth transmission (abbreviated as UBT). After receiving the computing tasks, the relay nodes will uniformly send them to the cloud or edge server for processing.

[0004] Most traditional computing task allocation methods rely on a centralized management system. This method may be effective when dealing with simple tasks, but it often seems powerless when facing a dynamically changing power system and multi-UAV collaborative work. In addition, the existing technology fails to fully consider the optimization of communication costs during the task allocation process, resulting in low system efficiency and increased unnecessary communication overhead.

[0005] Therefore, the present invention proposes an optimization method for task allocation based on the existing computing tasks of UAVs, aiming to reduce the total communication cost of the system, improve the operation efficiency and inspection quality of the power system by reasonably scheduling the task allocation of UAVs. This method can not only improve the working efficiency of UAVs, but also very effectively reduce the waste of resources, and has important theoretical and practical significance. In addition, this method should also consider the energy limitation of UAVs, the size of information transmission time, and the signal coverage of different regions to ensure the fairness and real-time nature of task allocation. By introducing advanced algorithms, such as machine learning or artificial intelligence technologies, the intelligent level of task allocation can be further improved, and more accurate and dynamic task scheduling can be achieved. This will be very helpful to improve the intelligent level of the entire power system and ensure the continuity and reliability of power supply. Summary of the Invention

[0006] In view of this, the present invention proposes an allocation optimization method for existing computing tasks of unmanned aerial vehicles, aiming to reduce the total communication cost of the system and improve the operation efficiency and inspection quality of the power system by reasonably scheduling the task allocation of unmanned aerial vehicles.

[0007] To achieve the above object, the present invention provides a resource allocation optimization method for unmanned aerial vehicle power inspection, including the following steps:

[0008] S1: The unmanned aerial vehicles performing inspection tasks, nearby edge servers and ground base stations jointly form a federated learning system;

[0009] S2: Each unmanned aerial vehicle uploads the edge server parameters to the ground base station;

[0010] S3: The ground base station calculates the optimal computing load of each edge server according to the edge server parameters uploaded by the unmanned aerial vehicles;

[0011] S4: The ground base station calculates the most suitable edge server for each unmanned aerial vehicle to transmit as the optimal edge server transmission node, and broadcasts it to each unmanned aerial vehicle. The unmanned aerial vehicle sends data to the edge server according to the obtained optimal edge server transmission node and optimal computing load;

[0012] S5: The ground base station broadcasts the initial model parameters to each edge server, and starts a new round of federated learning model training. After receiving the broadcast global model, the unmanned aerial vehicle loads the data samples required for this round of training from the memory, updates the local model using the gradient descent algorithm, and uploads the updated local model to the ground base station for weighted aggregation processing;

[0013] S6: The ground base station determines whether the model training has reached the optimal model coordination round. If it has reached, the global model training is stopped. If it has not reached, the aggregated model parameters are continued to be broadcast to the edge servers participating in the federated learning training, and a new round of model training is started.

[0014] Further, the unmanned aerial vehicle, as a participating client in federated learning, is used to transmit relevant data information obtained during the inspection of the power grid to the edge server for local model training processing; the ground base station, as a central server, is used to broadcast the initial model parameters to each edge server and aggregate the model parameters uploaded by each edge server after completing local model training; the edge server is used to use the relevant data information provided by the unmanned aerial vehicle inspection for local model training, and upload the result of the local model training to the ground base station for model update processing, so as to obtain new model parameters, and send the new model parameters to each edge server.

[0015] Further, the edge server parameters include the data type f of edge server k, the dataset size n k , the number of local model training rounds e k , the size b of the data volume required for each round of training k and the location coordinates.

[0016] Further, step S3 specifically includes:

[0017] Set an empirical mapping function to represent the relationship between edge server parameters and computing load:

[0018]

[0019] where represents the computing load of edge server k, r represents the convergence requirement, and the above formula indicates that when performing federated learning model training, given the data type, convergence requirement, and related parameters n k , e k , b k and the relationship with the computing load, where n k represents the dataset size of edge server k, e k represents the number of iterations of edge server k in each round of model training, and b k represents the dataset size used by edge server k in each model training;

[0020] The problem statement of the resource allocation scheme for the optimal computing load is in the following form:

[0021]

[0022] where u(n) represents the average resource utilization rate of each edge server in the edge computing system, n represents the vector related to the number of data samples, n = <n k > k∈K , represents the maximum computing load supported by each edge server; K represents the set of all edge servers participating in federated learning, and it is assumed here that there is always and where represents the maximum number of sample data that edge server k can receive and process;

[0023] By calculating obtain the optimal result of each n k , denoted as Then check whether each satisfies these two constraint conditions. If not satisfied, adjust to meet the constraints;

[0024] Then use the gradient descent algorithm to update n k :

[0025]

[0026] where i 0 represents the update iteration number, and α represents the step size;

[0027] Stop the iteration until one of the following conditions is met:

[0028] (1) The change in the objective function is less than a preset threshold;

[0029] (2) The maximum number of iterations is reached;

[0030] Thus, we obtain represents the data sample size that achieves the optimal trade-off of the computing resource utilization of edge server k under the given model convergence requirements and the computing resource limitations of edge server k.

[0031] Furthermore, step S4 specifically includes:

[0032] S41: Based on the edge server parameters uploaded by the UAV in step S2, the ground base station calculates and uses a fictitious sequence d ik ={d i1 , d i2 , d i3 ,..., d iK} to represent the distance between the UAV with serial number i and each edge server, where d iK represents the distance from UAV i to edge server k;

[0033] S42: The ground base station obtains the computing power of each edge server and uses a fictitious sequence c k ={c 1 , c 2 , c 3 ..., c K} to represent the computing power of each edge server;

[0034] S43: Calculate the time required for UAV i to transmit and process data to the corresponding edge server. The specific modeling is as follows:

[0035]

[0036] where p k is a manually set weight coefficient, aiming to balance the importance of the time for data transmission of the UAV and data processing analysis of the edge server, and pk The value range is 0≤p k ≤1; N 0 Indicates the power of Gaussian noise, S(d ik ) indicates that the signal changes with the distance d during the transmission process. ik The changed signal power, B is the transmission bandwidth of the drone, T ik It represents the total time required for UAV i to transmit data to edge server k and for edge server k to process the data;

[0037] In calculating T ik The following limiting factors should be considered:

[0038] (a) If drone i selects edge server k, then other drones j, with j≠i, cannot select the same edge server k; and i,j∈I, I represents the set of all drones; specifically, if for drone i, k=arg(min(T ik )), then for drone j, k≠arg(min(T jk ));

[0039] (b) The computational load of each edge server needs to be within its processing capacity to avoid overload and resource waste, as shown below.

[0040]

[0041] Where S k,max represents the maximum data set size that edge server k can process;

[0042] (c) The size of the data set uploaded by the drone may be limited by its own storage and processing capabilities. Therefore, it is necessary to reasonably control the size of the data set transmitted by the drone while ensuring the model training effect. Specifically, it can be expressed as:

[0043]

[0044] Where n i,max represents the maximum data set size that can be collected by UAV i;

[0045] (d) The battery life of the drone limits its computing and communication capabilities when performing tasks, so it is necessary to find a balance between energy consumption and task execution, as shown below:

[0046] E i ≥E i,min

[0047] Where E i It represents the total energy of UAV i, Ei,min It represents the minimum energy required for the UAV to perform one mission.

[0048] The ground base station calculates the minimum T corresponding to UAV i ik where k in ik ) and at this time, the interior point method is used to solve for the minimum value of k. The specific steps are as follows:

[0049] The goal is to minimize the total time required for the UAV to transmit data to the edge server and process it, which is expressed as:

[0050] f(k) = min(max∑ k∈K T ik )

[0051] Next, considering the four restrictive factors (a)-(d) above, a barrier function is used to handle the inequality constraints, in the form:

[0052]

[0053] where x ik is a binary decision variable indicating whether UAV i selects edge server k for transmission, and μ is a barrier parameter that controls the weight of the logarithmic term in the barrier function;

[0054] The above formula is solved through multiple iterative operations, and the value of μ is updated simultaneously. When one of the following conditions is met, the algorithm can terminate:

[0055] (1) The change in the objective function is less than a preset threshold;

[0056] (2) The violation degree of all constraints is less than a preset threshold;

[0057] (3) The maximum number of iterations is reached;

[0058] Once the algorithm terminates, the optimal k * is output as the optimal solution, that is, the edge server k * at this time is the best edge server transmission node;

[0059] S44: The ground base station broadcasts the determined best edge server transmission node to each UAV;

[0060] S45: According to the best edge server transmission node provided by the ground base station, each UAV starts to transmit data to the corresponding best edge server transmission node, and the amount of data transmitted is the best computing load of each edge server.

[0061] The present invention has the following advantages compared with the prior art:

[0062] 1. Optimize resource utilization and improve efficiency: By allocating appropriate computing tasks to each edge server in federated learning, it can ensure that the computing resources of each device are fully utilized, avoiding problems such as resource waste. This method is particularly suitable for resource-constrained environments and can significantly improve the overall efficiency of the federated learning system. In federated learning, the computing power, communication speed, and storage capacity of client devices vary, and this difference poses a great challenge to the overall performance of the model. Therefore, dynamically adjusting computing tasks to adapt to the computing power of different devices can improve the flexibility and efficiency of the system.

[0063] 2. Reduce communication costs and enhance data privacy protection: Since only model updates rather than raw data are uploaded to the central server, this method reduces the amount of data transmitted over the network, thus reducing communication costs. At the same time, it also enhances data privacy protection because the raw data does not leave the local device. This is particularly important when dealing with sensitive data and can increase user trust in the federated learning system. As described in the search results, in federated learning, the data remains on the local client, and only model update information is exchanged with the central server, so privacy data is not directly leaked, providing great assistance to data privacy protection.

[0064] 3. Improve the robustness and adaptability of the system: By reasonably allocating the computing load, it can reduce training interruption problems caused by device failures or network issues, thus improving the robustness of the system. In addition, this method allows the system to adapt to different network conditions and device performances, enabling federated learning to be deployed in more diverse environments. This is also reflected in the search results. Federated learning needs to adapt to the situation where client devices may drop out of the current training due to network failures, computing power limitations, etc.

[0065] 4. Enhance the intelligence level of the power system: By intelligently allocating computing tasks, it not only improves resource utilization efficiency and system robustness but also reduces communication costs and enhances data privacy protection. This is of great significance for the application and development of the federated learning system. Especially in the field of power inspection, the application of this technology can greatly enhance the intelligence level of the power system and ensure the continuity and reliability of power supply.

[0066] 5. Achieve precise and dynamic task scheduling: Introducing advanced algorithms, such as machine learning or artificial intelligence technologies, can further improve the intelligence level of task allocation and achieve more precise and dynamic task scheduling. This will be very helpful in enhancing the intelligence level of the entire power system and ensuring the continuity and reliability of power supply. In this way, drones can flexibly adjust their computing load according to the real-time changing environment and task requirements, thus achieving more efficient power inspection and maintenance work.

[0067] 6. Improve the real-time performance of task execution: By evaluating the computing power of edge servers and the task requirements of drones in real time, this technology can make quick decisions and allocate the most suitable amount of tasks to drones. This real-time performance is crucial for tasks that require quick responses, such as emergency repairs or accident responses. Drones can quickly receive task assignments and immediately start execution, which greatly improves the response speed and processing ability of the power system to emergencies.

[0068] In summary, by intelligently allocating computing tasks, the present invention not only improves the resource utilization efficiency and system robustness, but also reduces communication costs and enhances data privacy protection, which is of great significance for the application and development of federated learning systems. Brief Description of the Drawings

[0069] Figure 1 It is a flowchart of an optimized resource allocation method for drone power inspection according to an embodiment of the present invention;

[0070] Figure 2 It is a flowchart for a drone to select the best edge server according to an embodiment of the present invention. Detailed Embodiments

[0071] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0072] Please refer to Figure 1 , an embodiment of the present invention provides an optimized resource allocation method for drone power inspection, including the following steps:

[0073] S1: A drone performing an inspection task, nearby edge servers, and a ground base station jointly form a federated learning system. Among them, the drone serves as a participating client in the federated learning and transmits relevant data information obtained during the inspection of the power grid to the edge server for local model training. The ground base station serves as the central server of the federated learning, broadcasts the initial model parameters to each edge server, and aggregates the model parameters uploaded by each edge server after completing local model training to complete the model training. The edge server uses the relevant data information provided by the drone inspection for local model training, uploads the results of the local model training to the central server for model update processing to obtain new model parameters, and sends the new model parameters to each edge server.

[0074] S2: Before the start of federated learning training, each drone uploads the edge server parameters to the ground base station, including the data type f, the dataset size n of edge server k k , the number of local model training rounds e k , the size b of the data volume required for each round of training k and related parameters such as location coordinates.

[0075] S3: The ground base station will evaluate and determine the optimal computing load that each edge server should undertake based on the edge server parameters uploaded by each drone. This process involves a comprehensive consideration of parameters such as the data processing capabilities of the edge server, the scale of the dataset, and the number of training rounds to ensure that each server can achieve the optimal allocation and utilization of resources without exceeding its performance limit.

[0076] Specifically, an empirical mapping function is set to represent the relationship between the edge server parameters and the computing load:

[0077]

[0078] where represents the computing load of edge server k, and r represents the convergence requirement. The above formula shows that when training the federated learning model, given the data type, convergence requirement, and related parameters n k , e k , b k and the relationship between the computing load, where n k represents the dataset size of edge server k, e k represents the number of iterations of edge server k in each round of model training, and b k represents the dataset size used by edge server k in each model training.

[0079] The problem of the resource allocation scheme for the optimal computing load can be expressed in the following form:

[0080]

[0081] where u(n) represents the average resource utilization rate of each edge server in the edge computing system, and n represents the vector related to the number of data samples. Specifically, n = <n k > k∈K . represents the maximum computing load supported by each edge server; K represents the set of all edge servers participating in federated learning, and it is assumed here that there is always and where represents the maximum number of sample data that edge server k can receive and process.

[0082] Therefore, by calculating the optimal result for each n k is obtained, denoted as Then, check each to see if it satisfies these two constraint conditions. If not, then adjustments need to be made to meet the constraint conditions.

[0083] Then use the gradient descent algorithm to update n k :

[0084]

[0085] where i 0 represents the update iteration number, and α represents the step size.

[0086] Stop the iteration until one of the following conditions is met:

[0087] (3) The change in the objective function is less than a preset threshold.

[0088] (4) The maximum iteration number is reached.

[0089] From this, we can obtain This value represents the data sample size that achieves the optimal trade-off of the computing resource utilization of edge server k under the given model convergence requirements and the computing resource limitations of edge server k.

[0090] S4: The ground base station calculates the most suitable edge server for each UAV to transmit as the best edge server transmission node and broadcasts it to each UAV. The UAVs send data to the edge server according to the obtained best edge server transmission node and the best computing load. The specific steps are as Figure 2 shown:

[0091] S41: Based on the edge server parameters uploaded by the UAVs in step S2, the ground base station calculates and uses a fictional sequence d ik ={d i1 , d i2 , d i3 ,..., d iK} to represent the distance between the UAV with serial number i and each edge server, where d iK represents the distance from UAV i to edge server k;

[0092] S42: The ground base station obtains the computing power of each edge server and uses a fictional sequence c k ={c 1 , c 2 , c 3..., c K represent the computing power of each edge server;

[0093] S43: The ground base station calculates the edge nodes that are most suitable for data transmission for each UAV based on the uploaded relevant information data.

[0094] Specifically, calculate the time required for UAV i to transmit and process data to the corresponding edge server. The specific modeling is as follows:

[0095]

[0096] where p k is a manually set weight coefficient, aiming to balance the importance of the time for data transmission of the UAV and data processing analysis of the edge server, and p k has a value range of 0 ≤ p k ≤ 1. N 0 represents the power of Gaussian noise, and S(d ik ) represents the signal power that changes with the distance d ik during the signal transmission process. B is the transmission bandwidth of the UAV, and T ik represents the total time required for UAV i to transmit data to edge server k and for edge server k to process this data.

[0097] In addition, when calculating T ik , many restrictive factors need to be considered, as follows:

[0098] (1) The data types collected by the UAV need to meet the requirements of the corresponding edge server for processing, and at the same time, the data quality needs to reach a certain standard to ensure the effectiveness of model training.

[0099] Assume that D ik represents the data type and quality of the dataset collected by UAV i at edge server k. It is necessary to ensure that D ik meets the processing requirements of edge server k:

[0100] D ik ∈ D k

[0101] where D k represents the set of data types and qualities that edge server k can process.

[0102] (2) If UAV i selects edge server k, then for other UAVs j, and j ≠ i, the same edge server k cannot be selected. And i, j ∈ I, where I represents the set of all UAVs. Specifically, if for UAV i, k = arg(min(Tik ), then for the UAV j, k ≠ arg(min(T jk ).

[0103] (3) The computing load of each edge server needs to be within their processing capabilities to avoid overload and resource waste, which is specifically expressed as follows.

[0104]

[0105] where S k,max represents the maximum dataset size that the edge server k can handle.

[0106] (4) The size of the dataset uploaded by the UAV may be limited by its own storage and processing capabilities. Therefore, it is necessary to reasonably control the size of the dataset transmitted by the UAV on the premise of ensuring the model training effect, which can be specifically expressed as:

[0107]

[0108] where n i,max represents the maximum dataset size that the UAV i can collect.

[0109] (5) The battery life of the UAV limits its computing and communication capabilities during mission execution. Therefore, it is necessary to find a balance between energy consumption and mission execution, which is specifically expressed as follows:

[0110] E i ≥E i,min

[0111] where E i represents the total energy size of the UAV i, and E i,min represents the minimum energy size required for the UAV to execute one mission.

[0112] (6) During the data transmission and model update process, it is necessary to ensure data security and privacy protection to avoid the leakage of sensitive information.

[0113] Use an indicator function to express the situation of whether the data is leaked, which is specifically as follows:

[0114] Μ(data leakage) = 0

[0115] where Μ is the indicator function, with a value of 1 when there is data leakage and a value of 0 when there is no data leakage.

[0116] The ground base station calculates the minimum T corresponding to the UAV i ik where k = arg(min(T ik ). At this time, the interior point method is used to solve the minimum value of k, and the specific steps are as follows:

[0117] The goal is to minimize the total time required for the UAV to transmit data to the edge server and process it. This can be expressed as:

[0118] f(k) = min(max∑ k∈K T ik )

[0119] Next, considering the restrictive factors among the above six points, since the inner product method mainly considers inequality constraint conditions, we mainly consider the remaining inequality factors except for the first and sixth points. The barrier function will be used to handle the inequality constraints, in the form as follows:

[0120]

[0121] where x ik is the binary decision variable indicating whether UAV i selects edge server k for transmission. μ is the barrier parameter, which controls the weight of the logarithmic term in the barrier function. As the algorithm progresses, μ will gradually decrease, making it closer and closer to the boundary of the feasible region.

[0122] The above formula is solved through multiple iterative operations, and the value of μ is updated simultaneously. When one of the following conditions is met, the algorithm can terminate:

[0123] (1) The change in the objective function is less than the preset threshold.

[0124] (2) The violation degree of all constraints is less than the preset threshold.

[0125] (3) The maximum number of iterations is reached.

[0126] Once the algorithm terminates, the optimal k * is output as the optimal solution, that is, the edge server k * at this time is the best edge server transmission node.

[0127] S44: Through the above method, the ground base station can select the most suitable edge server for each UAV to transmit data and broadcast the result to each UAV.

[0128] S45: According to the best edge server transmission node provided by the ground base station, each UAV starts to transmit data to the corresponding best edge server transmission node, and the size of the transmitted data volume is the best computing load of each edge server.

[0129] S5: The ground base station determines the initial parameter size w 0 , and broadcasts the result to each edge server. Each edge server uses the gradient descent algorithm for w 0Perform an update and then transmit the updated result to the ground base station for weighted aggregation processing.

[0130] Specifically, for the edge server with serial number k, the local model update formula is:

[0131]

[0132] Among them, is the local model parameter for the (e + 1)-th iteration, is the local model parameter for the e-th iteration, and η k is the learning rate of the model, is the gradient magnitude at the e-th iteration.

[0133] S6: The ground base station determines whether the model training coordination round has been reached. If the ground base station has reached the initially set total model coordination training cycle, it will send a broadcast to the drones and each edge server, instructing them to stop training and save the current global model parameters. If not, it will broadcast the weighted result again to each edge server for the next round of model training.

[0134] The present invention can not only improve the working efficiency of the drones, but also effectively reduce resource waste, having important theoretical and practical significance. In addition, this method should also consider the energy limitation and flight time of the drones, as well as the signal coverage in different regions, to ensure the fairness and real-time nature of task allocation. By introducing advanced algorithms, such as machine learning or artificial intelligence technologies, the intelligent level of task allocation can be further improved, realizing more accurate and dynamic task scheduling. This will help to improve the intelligent level of the entire power system and ensure the continuity and reliability of power supply.

[0135] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A resource allocation optimization method for UAV power inspection, characterized by: The steps include: S1: The drones that perform inspection tasks, the nearby edge servers and ground base stations form a federated learning system; S2: Each drone uploads edge server parameters to the ground base station; S3: The ground base station calculates the optimal computing load of each edge server based on the edge server parameters uploaded by the drone; S4: The ground base station calculates the edge server that is most suitable for each drone to transmit as the best edge server transmission node, and broadcasts it to each drone. The drone sends data to the edge server based on the obtained best edge server transmission node and the best computing load. S5: The ground base station broadcasts the initial model parameters to each edge server and starts a new round of federated learning model training. After receiving the broadcast global model, the drone loads the data samples required for this round of training from the storage, uses the gradient descent algorithm to update the local model, and uploads the updated local model to the ground base station for weighted aggregation processing; S6: The ground base station determines whether the model training has reached the optimal number of model coordination rounds. If so, the global model training is stopped. If not, the aggregated model parameters are broadcasted to the edge servers participating in the federated learning training to start a new round of model training.

2. The resource allocation optimization method for UAV power inspection according to claim 1, characterized in that: The drone serves as a participating client of federated learning, and is used to transmit relevant data information obtained during the inspection of the power grid to the edge server for local model training processing; the ground base station serves as a central server, and is used to broadcast the initial model parameters to each edge server, and aggregate the model parameters uploaded by each edge server after completing the local model training; the edge server is used to use the relevant data information provided by the drone inspection for local model training, and upload the results of the local model training to the ground base station for model update processing, so as to obtain new model parameters, and send the new model parameters to each edge server.

3. The resource allocation optimization method for UAV power inspection according to claim 2, characterized in that: The edge server parameters include the data type f of the edge server k, the data set size n k , number of local model training rounds e k , the amount of data required for each round of training b k and location coordinates.

4. The resource allocation optimization method for UAV power inspection according to claim 3 is characterized by: Step S3 specifically includes: Set an empirical mapping function to express the relationship between edge server parameters and computing load: in represents the computing load of edge server k, r represents the convergence requirement, and the above formula shows that when training the federated learning model, given the data type, convergence requirement and related parameters n k , e k , b k The relationship between the computational load and the k represents the dataset size of edge server k, e k represents the number of iterations of edge server k in each round of model training, b k It represents the size of the data set used by edge server k in each model training; The problem of optimal resource allocation for computing load is formulated as follows: Among them, u(n) represents the average resource utilization of each edge server in the edge computing system, n represents a vector related to the number of data samples, n=<n k > k∈K , represents the maximum computing load supported by each edge server; K represents the set of all edge servers participating in federated learning, and it is assumed that there is always and in Indicates the maximum amount of sample data that edge server k can receive and process; By calculation Get each n k The optimal result is denoted as Then check each Is it satisfied? If these two constraints are not met, adjust To meet the constraints; Then use the gradient descent algorithm to update n k : Where i0 represents the number of update iterations, and α represents the step size; The iteration stops when one of the following conditions is met: (1) The change of the objective function is less than the preset threshold; (2) reaching the maximum number of iterations; From this we get It represents the data sample size that achieves the optimal trade-off in the computing resource utilization of edge server k under the given model convergence requirements and computing resource constraints of edge server k.

5. The resource allocation optimization method for UAV power inspection according to claim 3, characterized in that: Step S4 specifically includes: S41: Based on the edge server parameters uploaded by the drone in step S2, the ground base station calculates and uses a fictitious sequence d ik ={d i1 ,d i2 ,d i3 ,...,d iK } to represent the distance between the drone with serial number i and each edge server, where d iK represents the distance from UAV i to edge server k; S42: The ground base station obtains the computing power of each edge server and uses a fictitious sequence c k ={c1,c2,c3...,c K }Indicate the computing power of each edge server; S43: Calculate the time required for drone i to transmit and process data to the corresponding edge server. The specific modeling is as follows: where p k is a manually set weight coefficient, which aims to weigh the importance of the two parts of time for drone data transmission and edge server data processing and analysis, and p k The value range is 0≤p k ≤1; N0 represents the power of Gaussian noise, S(d ik ) indicates that the signal changes with the distance d during the transmission process. ik The changed signal power, B is the transmission bandwidth of the drone, T ik It represents the total time required for UAV i to transmit data to edge server k and for edge server k to process the data; In calculating T ik The following limiting factors should be considered: (a) If drone i selects edge server k, then other drones j, with j≠i, cannot select the same edge server k; and i,j∈I, I represents the set of all drones; specifically, if for drone i, k=arg(min(T ik )), then for drone j, k≠arg(min(T jk )); (b) The computational load of each edge server needs to be within its processing capacity to avoid overload and resource waste, as shown below. Where S k,max represents the maximum data set size that edge server k can process; (c) The size of the data set uploaded by the drone may be limited by its own storage and processing capabilities. Therefore, it is necessary to reasonably control the size of the data set transmitted by the drone while ensuring the model training effect. Specifically, it can be expressed as: Where n i,max represents the maximum data set size that can be collected by UAV i; (d) The battery life of the drone limits its computing and communication capabilities when performing tasks, so it is necessary to find a balance between energy consumption and task execution, as shown below: AND i ≥E i,min Where E i Indicates the total energy of drone i, E i,min It indicates the minimum amount of energy required by the drone to perform a mission; The ground base station calculates the minimum T corresponding to UAV i ik k=arg(min(T ik )), at this time, the interior point method is used to solve the minimum value of k. The specific steps are as follows: The goal is to minimize the total time required for the drone to transmit data to the edge server and process it, expressed as: f(k)=min(max∑ k∈K T ik ) Next, considering the four restrictive factors (a)-(d) above, the barrier function is used to handle the inequality constraints in the following form: where x ik is the binary decision variable for whether UAV i chooses edge server k for transmission, μ is the barrier parameter, which controls the weight of the logarithmic term in the barrier function; The above formula is solved through multiple iterative operations, and the value of μ is updated at the same time. The algorithm can be terminated when one of the following conditions is met: (1) The change of the objective function is less than the preset threshold; (2) The violation degree of all constraints is less than the preset threshold; (3) reaching the maximum number of iterations; Once the algorithm terminates, the optimal k is output * As the optimal solution, the edge server k at this time * Transmit nodes for optimal edge servers; S44: The ground base station broadcasts the determined optimal edge server transmission node to each drone; S45: According to the best edge server transmission node provided by the ground base station, each drone starts to transmit data to the corresponding best edge server transmission node, and the amount of data transmitted is the best computing load of each edge server.

Images