Asynchronous federal learning acceleration method for heterogeneous unmanned aerial vehicle electric power inspection

Abstract

The invention discloses an asynchronous federated learning acceleration method for heterogeneous unmanned aerial vehicle electric power inspection. The method comprises the following steps: forming a federated learning system by an unmanned aerial vehicle executing an inspection task and a nearby base station; the unmanned aerial vehicle calculates the time consumption of three processes of local data loading, local model training and model parameter uploading, and the total time consumption obtained through calculation is added to the time consumption of the global model parameter distribution process to determine the total time consumption required by each round of federated learning training of each unmanned aerial vehicle; the base station calculates and obtains an optimal load forwarding scheme according to the time consumption required by each round of training of each unmanned aerial vehicle, and notifies the unmanned aerial vehicle needing load forwarding to execute load forwarding; after the unmanned aerial vehicles execute the load forwarding scheme, the base station initializes global model parameters and broadcasts the global model parameters to the unmanned aerial vehicles participating in training to start asynchronous federated learning; and the unmanned aerial vehicle adopts a gradient descent algorithm to update the local model, and uploads the updated local model parameters to the base station for model cooperation. According to the method, model training acceleration can be realized under the condition of ensuring the accuracy of the global model.

Description

Technical Field

[0001] The present invention relates to the fields of machine learning and image detection, and specifically to an asynchronous federated learning acceleration method for heterogeneous UAV power line inspection. Background Art

[0002] As an indispensable part of the power grid, the safety and stability of the operation of transmission lines are crucial for ensuring the continuity and reliability of power supply. With the continuous expansion and development of the power grid, the inspection and maintenance of transmission lines are also facing more and more challenges. The traditional manual inspection method has been difficult to meet the needs of large-scale transmission line maintenance, and there are significant potential safety hazards to human life when performing high-altitude operations in complex and harsh geographical environments. As an emerging technology, UAVs can carry sensor devices such as high-definition cameras and infrared thermal imagers to take multi-directional and high-definition pictures of transmission lines. Through image detection technology, line defects can be automatically detected and classified, effectively improving the accuracy and reliability of inspection. In addition, UAVs can perform inspection tasks in complex geographical environments and can significantly shorten the inspection cycle. Currently, they have gradually become the main means of transmission line inspection.

[0003] In power line inspection, as a key technical means of UAV inspection, the effectiveness of the image detection model directly determines the efficiency and accuracy of inspection. The model training methods can be divided into centralized training and distributed training. Centralized training requires UAVs to upload the high-definition images collected during the inspection process to the cloud server for unified processing and model training. However, due to the large amount of high-definition image data collected by UAVs, a high network bandwidth resource is required during the image upload process. When multiple UAVs perform inspection tasks and upload data simultaneously, the network link may become congested due to the sudden increase in data traffic, which not only affects the efficiency of data transmission but also may cause delays in the overall detection process, thereby reducing the overall efficiency and timeliness of the inspection work. In addition, the image data collected by UAVs may contain sensitive information such as geographical coordinates. If uploaded to the cloud without encryption protection, there is a risk of privacy leakage. Facing these problems, the distributed training method shows obvious advantages. Taking the federated learning, a distributed learning framework, as an example, UAVs can process data locally and only upload model updates to the central server, reducing the communication resource overhead while strengthening data privacy protection. This training method not only reduces the pressure on the network bandwidth but also enhances the data security, improving the overall feasibility and security of the power line inspection task.

[0004] Although using federated learning for power inspection by drones can effectively protect privacy data and reduce communication overhead, due to the differences in the hardware performance of each drone and the non-independent and identically distributed (non-IID) characteristics of the collected data, there are significant differences in the time taken for each drone to perform local model calculations during the federated learning process. In such an environment of heterogeneous devices and data, the traditional asynchronous federated learning update mechanism will encounter the Staleness problem - that is, the gradients or parameters used for global model updates lag behind the current state of the global model. This not only weakens the convergence performance of the model but also leads to an extended training time. The emergence of the Staleness problem not only increases the consumption of the drones' computing resources but also affects the ability of the drones to perform other tasks. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides an asynchronous federated learning acceleration method for heterogeneous drone power inspection. Before the start of asynchronous federated learning training, a load forwarding scheme is adopted to balance the time consumption of each drone in each round of training, reduce the Staleness problem in the asynchronous federated learning process, and accelerate model training while ensuring the accuracy of the global model.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] An asynchronous federated learning acceleration method for heterogeneous drone power inspection, comprising the following steps:

[0008] S1: The drones performing the inspection tasks and the nearby base stations jointly form a federated learning system. The drones are responsible for performing local model calculations using the data samples and labels collected during the power inspection process. The base station, as the server of the federated learning, is responsible for aggregating the model parameters from each drone and sending the updated model parameters to each drone;

[0009] S2: Each drone calculates the time consumption of the three processes of local data loading, local model training, and model parameter uploading using the parameters related to computing and communication time consumption, and sends the calculated total time consumption to the base station. The base station adds the time consumption of the global model parameter distribution process to the total time consumption to finally determine the total time consumption required for each drone in each round of federated learning training;

[0010] S3. The base station calculates the optimal load forwarding scheme based on the time consumption required for each drone in each round of training, and notifies the drones that need to perform load forwarding to execute load forwarding;

[0011] S4. After the drones execute the load forwarding scheme, the base station initializes the global model parameter w 0, and broadcast to each participating drone to start asynchronous federated learning;

[0012] S5. After receiving the global model parameters broadcast by the base station, the drone 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 parameters to the base station for model collaboration;

[0013] S6. The base station determines whether the preset global training rounds have been reached. If the preset training rounds have been reached, it broadcasts a notice to the drones to end the training and saves the current global model parameters of the base station; if the preset global training rounds have not been reached, it immediately updates the global model parameters and sends the updated model parameters to the drones participating in the aggregation to continue executing S5.

[0014] Furthermore, the parameters related to the computing and communication time consumption include the number of gradient descent iterations k required for one round of local model training, the size D of the data samples used in each iteration s , CPU frequency f n , CPU cycles μ required to train one sample n , signal transmission power p n .

[0015] Furthermore, step S2 specifically includes:

[0016] S21: Calculate the time consumption of the data loading process before the local model training of the drone

[0017] In each round of federated learning process, drone n loads data samples of size D in each gradient descent iteration, and then performs k gradient descent iterations to update the local model parameters. The number of data samples loaded in each round of federated learning process is kD s ; if l s is the speed at which the drone loads each data sample, then the time consumption required for drone n to load data in each round is: n q

[0018] q n = l n kD s

[0019] S22: Calculate the time consumption of the local model training process of the drone

[0020] When the training samples required for this round are ready, the participating drone node n will perform k gradient descent algorithms to update the local model parameters. If f n is the CPU frequency parameter of the drone, and μ n is the CPU cycles required for it to process one data sample, then the time consumption of the local model training process is:

[0021]

[0022] S23: Calculation of the time taken for the UAV to upload the trained local model parameters to the base station

[0023] When the UAVs participating in the training complete local training, they upload the local model parameters to the base station. According to the basic principle of wireless channel transmission, if B is the wireless link bandwidth, h is the channel gain, σ 2 is the noise power, and b n is the link bandwidth allocated to UAV n, p n is the signal transmission power of UAV node n, then the uplink rate of UAV node n is: -

[0024]

[0025] If the size of the model parameters is represented as s w , then the time taken to upload the local model parameters to the base station is:

[0026] t n = s w / r n

[0027] S24: The total time consumption c' of the three processes of UAV local data loading, local model training, and model parameter upload n = q n + v n + t n is sent to the base station.

[0028] S25: The base station adds the time consumption required for global model parameter distribution to the time consumption c' of the previous three processes to obtain the total time consumption required for each UAV for each round of federated learning training, specifically including: n

[0029]

[0029] S251: The base station calculates the time consumption required for global model parameter distribution. According to the basic principle of wireless channel transmission, if p c is the base station transmission power and σ 2 is the noise power, then the downlink rate of the base station is:

[0030]

[0031] The time consumption during the model parameter distribution process is:

[0032] t c = s w / r c

[0033] S252: Accumulate the time consumption of the four processes of local data loading, local model training, model parameter uploading, and global model parameter distribution to obtain the time consumption c required for each round of federated learning training of each UAV. n = t c + c' n 。

[0034] Further, step S3 includes the following steps:

[0035] S31: The base station sorts the UAVs according to the time consumption required for each round of training of each UAV from smallest to largest to obtain the sorted UAV number sequence. The UAVs in the first half after sorting are defined as high-performance UAVs, and the UAVs in the second half are classified as low-performance UAVs.

[0036] S32: According to the load forwarding rule, the base station calculates the load forwarding ratio of the low-performance UAVs and notifies the low-performance UAVs to execute the load forwarding scheme according to the calculated load forwarding ratio, forwarding the load to the specified high-performance UAVs in the rule, and using the computing power of the high-performance UAVs to help them complete part of the computing tasks.

[0037] Further, the load forwarding rule is:

[0038] (1) Each low-performance UAV that needs to forward the load can only forward the load to one high-performance UAV, denoted as n and n' respectively.

[0039] (2) During the load forwarding process, the high-performance UAV n' needs to obtain data and model parameters from the low-performance UAV n. To reduce communication overhead, the high-performance UAV n' directly obtains the model parameters of the UAV n from the base station and uploads the model parameters after training and updating.

[0040] (3) In an ideal situation, using the load forwarding scheme, the time required for the high-performance UAV n' and the low-performance UAV n to execute one round of federated learning should be equal. Based on this, the base station calculates the optimal load forwarding ratio of each low-performance UAV and broadcasts and notifies each low-performance UAV to execute the load forwarding scheme.

[0041] Further, in step S5:

[0042] For UAV n, the local model update formula is:

[0043] is the local model parameter for the (k + 1)-th iteration, is the local model parameter for the k-th iteration, η n is the learning rate of the model, is the gradient.

[0044] Further, the global model parameter update formula of the base station under the asynchronous federated mechanism is w i+1 is the global model parameter of the (i + 1)-th round, w i is the global model parameter of the i-th round, χ n is the weight parameter of the drone n participating in the collaboration, and is its gradient.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. Each drone in the present invention adopts a distributed asynchronous federated learning framework to collaboratively learn and obtain an image detection model in power inspection. Compared with the traditional centralized model training method, that is, drones centrally upload original data to the cloud server for training, it avoids the consumption of communication resources for uploading a large amount of high-definition image data and the leakage of privacy data.

[0047] 2. According to the data sets and hardware computing capabilities of the drones participating in the collaboration, the present invention adopts a load forwarding method to forward part of the data of low-performance drones to high-performance drones for model training, balances the aggregation times of each drone, and thus effectively controls the Staleness problem in asynchronous federated learning. Without affecting the accuracy, it reduces the model training duration. For resource-constrained drones, it can reduce unnecessary computing waste, thereby saving energy consumption and extending the working life of the drones. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a schematic flowchart of an asynchronous federated learning acceleration method for heterogeneous drone power inspection according to an embodiment of the present invention;

[0049] Figure 2 is a schematic diagram of the federated learning process of drone power inspection according to an embodiment of the present invention;

[0050] Figure 3 are the training time-consuming results of different federated learning methods in different scenarios. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] 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.

[0052] Please refer to Figure 1, an embodiment of the present invention provides an asynchronous federated learning acceleration method for heterogeneous UAV power inspection, including the following steps:

[0053] S1: The UAVs performing inspection tasks and nearby base stations jointly form a federated learning system, as Figure 2 shown. The UAVs are responsible for performing local model calculations using the data samples and labels collected during the power inspection process. The base station, as the server of the federated learning, is responsible for aggregating the model parameters from each UAV and sending the updated model parameters to each UAV.

[0054] Specifically, consider a wireless federated learning system consisting of one base station and N UAVs (denoted as the set ). Each UAV has its own local dataset, that is, the image data (images of power facilities such as power towers, insulators, and conductors) and corresponding labels (types of power facilities, such as iron towers, concrete towers, insulator strings, and states, such as normal, cracked, rusted, damaged, etc.) collected during the inspection process, denoted as D n , n ∈ N.

[0055] S2: Each UAV uses parameters related to computing and communication time consumption, including the number of gradient descent iterations k required for one round of local model training, the size of the data samples D s used each time, the CPU frequency f n , the CPU cycles μ required to train one sample n , the signal transmission power p n , calculates the time consumption of the three processes of local data loading, local model training, and model parameter uploading, and sends the calculated total time consumption to the base station. Based on this, the base station adds the time consumption of the global model parameter distribution process and finally determines the time consumption required for each UAV for each round of federated learning training.

[0056] The step S2 includes the following sub-steps:

[0057] S21: Calculation of the time consumption of the data loading process before the UAV performs local model training

[0058] In each round of federated learning process, UAV n loads data samples of size D s each time during gradient descent iteration, and then performs k gradient descent iterations to update the local model parameters. Therefore, the number of data samples loaded during each round of federated learning process is kD s ; if l n is the speed at which the UAV loads each data sample, then the time consumption required for UAV n to load data for each round is:

[0059] q n = l n kD s

[0060] S22: Calculation of the time consumption during the local model training process of the UAV

[0061] When the training samples required for this round are ready, the UAV node n participating in the collaboration will execute the gradient descent algorithm k times to update the local model parameters. If f n is the CPU frequency parameter of the UAV, and μ n is the CPU cycles required for it to process one data sample, then the time consumption during the local model training process is:

[0062]

[0063] S23: Calculation of the time consumption for the UAV to upload the trained local model parameters to the base station

[0064] When the UAVs participating in the training complete local training, they will upload the local model parameters to the base station. According to the basic principle of wireless channel transmission, if B is the wireless link bandwidth, h is the channel gain, and σ 2 is the noise power, and b n is the link bandwidth allocated to the UAV n, p n is the signal transmission power of the UAV node n, then the uplink rate of the UAV node n is: -

[0065]

[0066] If the size of the model parameters is represented as s w , then the time consumption for uploading the local model parameters to the base station is:

[0067] t n = s w / r n

[0068] S24: The total time consumption c' of the three processes of local data loading, local model training, and model parameter uploading by the UAV n = q n + v n + t n is sent to the base station.

[0069] S25: The base station adds the time consumption required for downloading the global model parameters to the time consumption c' n in the previous three processes to obtain the total time consumption required for each UAV for each round of federated learning training. Specifically, it includes:

[0070] S251: The base station calculates the time consumption required for downloading the global model parameters. According to the basic principle of wireless channel transmission, if p c is the transmission power of the base station and σ 2 is the noise power, then the downlink rate of the base station is:

[0071]

[0072] The time consumption of the model parameter distribution process is:

[0073] t c = s w / r c

[0074] S252: Accumulate the time consumptions of the four processes of local data loading, local model training, model parameter uploading, and global model parameter distribution to obtain the time consumption c required for each round of federated learning training of each UAV. n = t c + c' n .

[0075] S3: The base station calculates the optimal load forwarding scheme based on the time consumption required for each round of training of each UAV, and notifies the UAVs that need to perform load forwarding to execute the load forwarding.

[0076] Specifically, step S3 may include the following steps:

[0077] S31: The base station sorts the time consumptions c required for each round of training of each UAV n from small to large to obtain the sorted UAV number sequence wherein, represents the UAV with the shortest time consumption per round of training, represents the UAV with the longest time consumption per round of training. Then, the UAVs are divided into two groups according to the sorting result: high-performance UAVs, that is, the UAVs in the first half after sorting low-performance UAVs, that is, the UAVs in the second half of the sequence.

[0078] S32: According to the load forwarding rule, the base station calculates the load forwarding ratio of the low-performance UAVs, and notifies the low-performance UAVs to execute the load forwarding scheme according to the calculated load forwarding ratio, forward the load to the specified high-performance UAVs in the rule, and use the computing power of the high-performance UAVs to help them complete part of the computing tasks.

[0079] The specific rule is:

[0080] 1. To protect data security, the present invention stipulates that each low-performance UAV that needs to forward can only forward the load to one high-performance UAV, which are represented by n and n' respectively.

[0081] 2. During the load forwarding process, the high-performance UAV n' needs to obtain data and model parameters from the low-performance UAV n. To reduce communication overhead, the high-performance UAV n' directly obtains the model parameters of the UAV n from the base station and uploads the model parameters updated after training.

[0082] The time for the high-performance UAV n' to obtain data from the low-performance UAV n is:

[0083] o n,n' = ρ n,n' kD s / r n

[0084] where ρ n,n' is the proportion of data forwarding in each round.

[0085] The time overhead for the high-performance UAV n' to help the low-performance UAV n complete a federated learning process is:

[0086] c n,n' = t c + ρ n,n' q n' + o n,n' + ρ n,n' v n' + t n'

[0087] After performing data forwarding, the time required for the high-performance UAV n' to perform one round of federated learning is:

[0088] c n' + c n,n' = 2t c +(1 + ρ n,n' )q n' + o n,n' +(1 + ρ n,n' )v n' + 2t n x

[0089] The time required for the low-performance UAV n to perform one round of federated learning is:

[0090] c n = t c +(1 - ρ n,n' )q n +(1 - ρ n,n' )v n + t n

[0091] 3. In the ideal case, using the load forwarding scheme, the time required for the high-performance UAV n' and the low-performance UAV n to perform one round of federated learning should be equal. The base station calculates the optimal load forwarding ratio for each low-performance UAV based on this and broadcasts a notice to each low-performance UAV to execute the load forwarding scheme.

[0092] Specifically, using the load forwarding scheme, the time required for the high-performance UAV n' and the low-performance UAV n to perform one round of federated learning should be equal, that is:

[0093] c n' +c n,n' =c n

[0094] In this case, the optimal load forwarding ratio is as follows, where the set represents the set of all drones that need to forward.

[0095]

[0096] S4. After the drones execute the load forwarding scheme, the base station determines the initial model parameter w 0 , and broadcasts it to each participating drone to start asynchronous federated learning.

[0097] S5. After receiving the broadcast global model, the drones load the data samples required for this round of training from the memory, update the local model using the gradient descent algorithm, and upload the updated local model to the base station for model collaboration.

[0098] Specifically, for drone n, the local model update formula is:

[0099] is the local model parameter for the (k + 1)-th iteration, is the local model parameter for the k-th iteration, η n is the learning rate of the model, is the gradient.

[0100] S6. The base station determines whether the preset global training round has been reached. If the preset training round has been reached, it broadcasts a notification to the drones to end the training and saves the current global model parameters of the base station; if the preset global training round has not been reached, it immediately updates the global model and sends the updated model to the drones participating in the aggregation to continue executing S5.

[0101] The global model update formula of the base station under the asynchronous federated mechanism is w i+1 is the global model parameter for the (i + 1)-th round, w i is the global model parameter for the i-th round, χ n is the weight parameter of drone n participating in the collaboration, is its gradient.

[0102] To verify the effectiveness of the proposed invention, a federated learning system consisting of 10 drones and a base station was constructed for the detection and identification of insulator defects in transmission lines. YOLOv5 was used as the detection model in the experiment. The experimental dataset contained approximately 2000 images covering 3 categories of labels. The dataset was divided into a training set (80% of the total data volume) and a test set (20% of the total data volume).

[0103] Two different scenarios were designed in the experiment for model training: Scenario 1: Each drone was randomly assigned 20% of the samples in the training set for local model training; Scenario 2: Each drone was randomly assigned 50% of the samples in the training set for local model training;

[0104] During the training process, the test set was retained at the base station for evaluating the performance of the global model. When the detection accuracy of the global model reached 90%, the training process ended, and the training time consumption was recorded and obtained. To prove the effectiveness of the present invention, the present invention was compared with the traditional federated learning scheme FedAvg. The training time consumption results of the FedAvg scheme and the present invention scheme under the above two scenarios are as Figure 3 . From Figure 3 It can be seen that the scheme of the present invention can reduce the training time consumption by about 50% compared with the traditional FedAvg, showing a significant acceleration effect.

[0105] As described above, it 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 within 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. An asynchronous federated learning acceleration method for heterogeneous UAV power inspection, characterized in that: The steps include: S1: The drones that perform inspection tasks and nearby base stations form a federated learning system. The drones are responsible for using the data samples and labels collected during the power inspection process to perform local model calculations. The base stations serve as federated learning servers and are responsible for aggregating model parameters from various drones and sending updated model parameters to each drone. S2: Each UAV uses parameters related to computing and communication time to calculate the time consumption of local data loading, local model training, and model parameter uploading, and sends the calculated total time consumption to the base station. The base station adds the time consumption of the global model parameter distribution process to the total time consumption, and finally determines the total time consumption required for each round of federated learning training for each UAV. S3: The base station calculates the optimal load forwarding solution based on the time required for each round of training for each drone, and notifies the drones that need to forward loads to perform load forwarding; S4. After the drone executes the load forwarding scheme, the base station initializes the global model parameter w0 and broadcasts it to each drone participating in the training to start asynchronous federated learning; S5. After receiving the global model parameters broadcast by the base station, the drone 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 parameters to the base station for model collaboration; S6. The base station determines whether the preset global training rounds have been reached. If so, the base station broadcasts a notification to the drone to end the training and saves the current global model parameters of the base station. If not, the global model parameters are updated immediately and the updated model parameters are sent to the drones participating in the aggregation to continue executing S5.

2. The asynchronous federated learning acceleration method for heterogeneous UAV power inspection as claimed in claim 1 is characterized by: The parameters related to computation and communication time consumption include the number of gradient descent iterations k required for a round of local model training, the data sample size D used in each iteration, and the number of gradient descent iterations k required for a round of local model training. s 、CPU frequency f n , the CPU cycles required to train a sample μ n , signal transmission power p n .

3. The asynchronous federated learning acceleration method for heterogeneous UAV power inspection as claimed in claim 2 is characterized by: Step S2 specifically includes: S21: Calculation of the time consumption of data loading process before local model training of drones In each round of federated learning, drone n loads D s The number of data samples of size is kD, and then k gradient descent iterations are performed to update the local model parameters. The number of data samples loaded in each round of federated learning process is kD s If l n is the speed at which the drone loads each data sample, then the time required for each round of data loading for drone n is: q n =l n kD s S22: Calculation of the time consumption of the local model training process of the drone When the training samples required for this round are ready, the drone node n participating in the collaboration will execute the gradient descent algorithm k times to update the local model parameters. n is the CPU frequency parameter of the drone, μ n The CPU cycles required to process one data sample are as follows: S23: The time-consuming calculation of the drone uploading the trained local model parameters to the base station When the participating UAV completes local training, it uploads the local model parameters to the base station. According to the basic principle of wireless channel transmission, if B is the wireless link bandwidth, h is the channel gain, σ 2 is the noise power, b n is the link bandwidth allocated to UAV n, p n is the signal transmission power of drone node n, then the uplink rate of drone node n is: If the model parameter size is represented by s w , then the time taken to upload the local model parameters to the base station is: t n =s w / r n S24: The total time consumed by the drone for the three processes of loading local data, training local models, and uploading model parameters c' n =q n +v n +t n Send to the base station. S25: The base station spends c' time in the first three processes n Add the time required for sending global model parameters to get the total time required for each round of federated learning training for each drone, including: S251: The base station calculates the time required to send the global model parameters. According to the basic principle of wireless channel transmission, if p c is the base station transmission power, σ 2 is the noise power, then the downlink rate of the base station is: The time consumption of model parameter delivery process is: t c =s w / r c S252: Accumulate the time consumption of the four processes of local data loading, local model training, model parameter uploading and global model parameter sending to obtain the time consumption c required for each round of federated learning training for each drone n =t c +c' n .

4. The asynchronous federated learning acceleration method for heterogeneous UAV power inspection as claimed in claim 1 is characterized by: The step S3 comprises the following steps: S31: The base station sorts the drones from small to large according to the time required for each round of training, and defines the drones in the first half as high-performance drones, while the drones in the second half are classified as low-performance drones; S32: According to the load forwarding rule, the base station calculates the load forwarding ratio of the low-performance UAV, and notifies the low-performance UAV to execute the load forwarding plan according to the calculated load forwarding ratio, and forwards the load to the high-performance UAV specified in the rule, using the computing power of the high-performance UAV to help it complete part of the computing task.

5. The asynchronous federated learning acceleration method for heterogeneous UAV power inspection as claimed in claim 4 is characterized in that: The load forwarding rules are: (1) Each low-performance UAV that needs to forward load can only forward the load to one high-performance UAV, represented by n and n' respectively; (2) During the load forwarding process, the high-performance UAV n' needs to obtain data and model parameters from the low-performance UAV n. To reduce communication overhead, the high-performance UAV n' directly obtains the model parameters of UAV n from the base station and uploads the trained and updated model parameters; (3) Ideally, using the load forwarding scheme, the time required for high-performance UAV n' and low-performance UAV n to perform a round of federated learning should be equal. Based on this, the base station calculates the optimal load forwarding ratio of each low-performance UAV And broadcast to notify each low-performance drone to execute the load forwarding plan.

6. The asynchronous federated learning acceleration method for heterogeneous UAV power inspection as claimed in claim 1, characterized in that: In step S5: For drone n, the local model update formula is: is the local model parameter of the k+1th iteration, is the local model parameter of the kth iteration, η n is the learning rate of the model, is the gradient.

7. The asynchronous federated learning acceleration method for heterogeneous UAV power inspection as claimed in claim 1, characterized in that: The global model parameter update formula of the base station under the asynchronous federation mechanism is: w i+1 is the global model parameter of the i+1th round, w i is the global model parameter of the i-th round, χ n is the weight parameter of the UAV n participating in the collaboration, for its gradient.

Images