An online federated learning method for air-ground collaboration under unreliable communication

Abstract

This invention provides an air-ground collaborative online federated learning method for unreliable communications. The method involves: drones collect real-time image data based on pre-set flight paths and select samples based on dynamic data distribution to update local models. During local model upload, model parameter errors may occur due to unstable channels such as path loss and shadow fading. A central server then selects drones for global aggregation based on these errors. By jointly optimizing client scheduling and sample selection, the method achieves the technical goal of minimizing training loss while satisfying latency requirements, energy consumption constraints, and basic decision variable restrictions. This method is highly efficient and intelligent, significantly improving the accuracy of drone collaborative target detection.

Description

Technical Field

[0001] The present invention relates to the field of air-ground collaborative online federated learning, and specifically to an air-ground collaborative online federated learning method for unreliable communication. Background Art

[0002] Recent advances in drone technology have shown tremendous potential across multiple sectors due to their high flexibility, three-dimensional mobility, and ease of deployment. Beyond data collection, drones can also train machine learning models onboard, which is crucial for tasks like target recognition. Equipped with sensors and intelligent computing modules, drones can cover larger areas in less time, perform real-time intelligent data processing, and advance the expansion of intelligent network services from the ground to low-altitude airspace.

[0003] Despite their enormous potential, drones face inherent challenges stemming from limited onboard resources. IoT applications, such as autonomous driving and object tracking, involve highly distributed datasets, making centralized data processing challenging. This poses significant challenges for drones, necessitating a distributed collaborative learning paradigm to address these issues. Federated learning, with its superior performance in collaboration, efficiency, and edge processing, is an ideal candidate for collaborative object detection in drones. It works by collaboratively training the same machine learning model across a group of drones using local datasets.

[0004] The drone acts as a client to train a local model and uploads its local model to the base station for global aggregation, forming air-ground collaborative federated learning. However, current work mainly focuses on offline training on existing datasets, which cannot be applied to real-time target detection tasks enabled by drones. Therefore, online federated learning takes precedence over adapting to dynamic environments, including changes in participating clients and changes in data distribution. In practice, incorporating all real-time new sample data into training will impose a huge computational cost on resource-limited drones. In addition, the unreliability of air-ground wireless channel transmission, including path loss and small-scale fading, increases the model parameter error rate. That is, during the uplink transmission of the local model, each layer of the neural network faces the problem of parameter error, which ultimately leads to a decrease in the overall accuracy of the model. Summary of the Invention

[0005] This paper provides an online federated learning method for air-ground collaboration in unreliable communications. By jointly optimizing client scheduling and sample selection, it achieves the technical goal of minimizing training loss while satisfying latency requirements, energy consumption constraints, and basic restrictions on decision variables. The proposed method is efficient and intelligent, significantly improving the accuracy of collaborative target detection by drones.

[0006] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0007] In a first aspect, the present invention discloses an air-ground collaborative online federated learning method for unreliable communication, comprising the following steps:

[0008] Step 1: The drone filters sample images collected in real time in a dynamic environment and generates a local personalized model using meta-gradient update training. The Bernoulli distribution is introduced to describe the probability that errors in model parameter uploads by the drone in unstable channels will affect the accuracy of the global model. The convergence of air-ground collaborative online federated learning with parameter errors is analyzed, and the expected convergence bound of air-ground collaborative online federated learning with parameter errors is used to represent the global training loss.

[0009] Step 2: Calculate the uplink and downlink communication delays of the UAV in one round of global training based on the UAV's uplink / downlink rates and the number of model parameters. Calculate the computational delay of the UAV in one round of global training based on the computational frequency of the UAV's onboard embedded device and the size of the training dataset.

[0010] Step 3: Calculate the user's communication energy consumption in one round of global training based on the drone's transmit power and communication delay. Calculate the drone's computing energy consumption in one round of global training based on the computing frequency of the drone's onboard embedded device and the size of the training dataset.

[0011] Step 4: Based on the idea of ​​alternating optimization, the problem of minimizing the global training loss is decomposed into the client scheduling problem and the sample screening problem. Based on the energy consumption constraints, delay constraints, and basic restrictions on optimization variables, the optimal solution for drone scheduling and sample screening is obtained through iterative optimization.

[0012] As a preferred example, in step 1, the expected convergence bound G of air-ground collaborative online federated learning with parameter errors is loss The calculation formula is:

[0013]

[0014] in,

[0015]

[0016] Where η1 is the learning rate, L is the Lipschitz constant, and D i is the size of the local dataset of drone i, D is the size of the training set after filtering samples, N is the number of drones, α i ∈(0,1) is the sample screening variable, β i ∈{0,1} is the client screening variable, e i is the parameter error rate during the transmission process of UAV i, ε is a positive parameter, η2 is the meta-learning rate, t is the training round, θ is the waterfall threshold, E|h i | 2is the channel gain, B U is the uplink bandwidth, p i is the transmission power of UAV i, N0 is the power spectrum density of Gaussian noise, φ is the intermediate variable expressed in the simplified formula, t is the result of recursive summation of the intermediate variable φ in the global round.

[0017] As a preferred example, in step 2, the uplink communication delay T i U , downlink communication delay T i D and calculation delay T i C The calculation formula is:

[0018]

[0019] in,

[0020]

[0021] Where S i (ω) is the number of model parameter bits uploaded by UAV i, S(Ω) is the number of global model parameter bits broadcast, B U 、B D are uplink / downlink bandwidth, p i 、p B are the transmission powers of UAV i and base station respectively, N0 is the power spectrum density of Gaussian noise, c i The number of CPU cycles required to process a single data sample for drone i, g i is the CPU frequency of drone i, s is the speed of light, f c is the carrier frequency, v i is the flight speed of UAV i, d i is the distance from UAV i to the base station, are the line-of-sight link / non-line-of-sight link losses, ζ i is the probability of line-of-sight link between UAV i and the base station, α i ∈(0,1) is the sample screening variable.

[0022] As a preferred example, in step 3, the communication energy consumption E i and calculate energy consumption The calculation formula is:

[0023]

[0024] Where, Calculate the chip effective capacitance coefficient, S, for drone i i (ω) is the number of bits of model parameters uploaded by UAV i, pi is the transmission power of UAV i, B U is the uplink bandwidth, E|h i | 2 is the channel gain, α i ∈(0,1) is the sample screening variable, g i is the CPU frequency of drone i, β i ∈{0,1} is the client screening variable, and N0 is the power spectral density of Gaussian noise.

[0025] As a preferred example, step 4 specifically includes the following sub-steps:

[0026] Step 4.1: Based on the first delay constraint, the first energy consumption constraint, the basic restrictions on the client screening variables, and the initial feasible solution of the sample screening variables, optimize the client scheduling using the branch and bound method to obtain the optimal solution for the current client screening variables;

[0027] Step 4.2: Based on the optimal solution of the client screening variable obtained in step 4.1, combined with the second delay constraint, the second energy consumption constraint, the gradient error constraint, and the basic restrictions of the sample screening variable, the sample screening variable is optimized using the Dinkelbach algorithm to obtain the optimal solution for the current sample screening variable;

[0028] Step 4.3: Repeat steps 4.1 to 4.2 alternately until the relative error of the optimization objective of the two iterations is less than the predefined threshold, the global training loss is minimized, and the optimal solution for client scheduling and sample screening is obtained.

[0029] As a preferred example, in step 4.1, the first delay constraint is calculated as follows:

[0030]

[0031] Where, T th Indicates the upper limit of the delay of a round of training, Screening the initial feasible solutions of variables for the sample;

[0032] The first energy consumption constraint is calculated as follows:

[0033]

[0034] Where, E th Indicates the upper limit of energy consumption for one round of training;

[0035] The basic limitations of client-side filter variables are as follows:

[0036] β i ∈{0,1}.

[0037] As a preferred example, in step 4.2, the second delay constraint is calculated as follows:

[0038]

[0039] Where, Screen the optimal solution of variables for the client;

[0040] The second energy consumption constraint is calculated as follows:

[0041]

[0042] The gradient error is calculated as follows:

[0043]

[0044] Where, is the variability of the gradient data, is the variability of the Hessian data, ▽ th is the gradient error threshold;

[0045] The basic restrictions on sample screening variables are as follows:

[0046] 0≤α i =α i,1 +α i,2 +α i,3 ≤1

[0047] Where, α i,1 , α i,2 , α i,3 Represent the screening variables of three independent small batches of samples.

[0048] As a preferred example, in step 4.3, when φ>1, the expected convergence bound of the minimized global training loss t→∞ diverges;

[0049] When φ<1, the t→∞ expected convergence bound converges,

[0050]

[0051] The global training loss to be minimized is calculated as:

[0052]

[0053] Where α represents the sample screening variable set, and β represents the client screening variable set.

[0054] In a second aspect, the present invention further discloses a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program enables a computer to execute the air-ground collaborative online federated learning method for unreliable communication as described above.

[0055] In a third aspect, the present invention further discloses an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the air-ground collaborative online federated learning method for unreliable communication as described above is implemented.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] The present invention's air-ground collaborative online federated learning method for unreliable communications uses drones to collect real-time image data based on pre-set flight paths and select samples based on dynamic data distribution to update local models. During the upload process, unstable channels such as path loss and shadow fading can cause model parameter errors. The central server then selects drones for global aggregation based on these errors. Furthermore, based on the principle of alternating optimization, the minimization problem is decomposed into a client scheduling problem and a sample selection problem. Through iterative optimization, global training loss is minimized. This proposed method is highly efficient and intelligent, significantly improving training effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a scene diagram of air-ground collaborative online federated learning under unreliable communication conditions of the present invention.

[0059] Figure 2 This is a flow chart of the air-ground collaborative online federated learning method for unreliable communication of the present invention.

[0060] Figure 3 It is a schematic diagram of the convergence numerical simulation results of the method of the present invention under different numbers of clients and different parameter error rates.

[0061] Figure 4 This is a schematic diagram showing how the convergence gap between the method of the present invention and the comparison algorithm changes with the number of iterations.

[0062] Figure 5 This is a schematic diagram of the impact of channel bandwidth on model performance. DETAILED DESCRIPTION

[0063] The embodiments of the present invention are described in further detail below with reference to the accompanying drawings.

[0064] This invention provides an air-ground collaborative online federated learning method for unreliable communication. This method considers a complex and dynamic environment in which multiple drones equipped with embedded intelligent computing devices perform collaborative target detection tasks under the coordination of a central server. Each drone acts as an independent client and, without sharing data with others, jointly trains a model. The drones collect real-time image data according to pre-set flight paths and select samples based on the dynamic data distribution to update the local model. During the local model upload process, model parameter errors may occur due to unstable channels such as path loss and shadow fading. The central server then selects drones for global aggregation based on these errors. By jointly optimizing client scheduling and sample selection, the method minimizes training loss while satisfying latency requirements, energy consumption constraints, and basic decision variable constraints. In this invention, the clients are drones equipped with embedded intelligent computing devices in the Internet of Things, and the central server is a ground base station. In practice, real-time data collection by drones leads to a continuous increase in the computing resources required for training. Furthermore, to prevent model parameter errors caused by unstable channels such as path loss and shadow fading during local model upload from affecting the accuracy of the global model, it is necessary to schedule drones to participate in aggregation effectively. Therefore, combining online federated learning with unreliable air-ground channels presents the following challenges: 1) The design of air-ground collaborative online federated learning should consider the computational overhead of real-time samples on resource-constrained drones and screen samples; 2) The design of air-ground collaborative online federated learning should strictly consider the impact of unreliable air-ground channels; 3) The design of air-ground collaborative online federated learning should strictly consider timeliness and energy saving to ensure that computing and communication are completed within the specified time and energy consumption thresholds.

[0065] The air-ground collaborative online federated learning method for unreliable communication provided by the present invention specifically includes the following steps:

[0066] (1) The UAV filters sample images collected in real time in a dynamic environment and uses meta-gradient update training to generate a local personalized model. The Bernoulli distribution is introduced to describe the probability that errors will occur when the UAV uploads model parameters in an unstable channel, thereby affecting the accuracy of the global model. The convergence of air-ground collaborative online federated learning with parameter errors is analyzed, and the expected convergence bound is used to characterize the implicit global training loss.

[0067] The expected convergence bound G of the air-ground collaborative online federated learning with parameter errors in this invention loss The calculation process is as follows:

[0068]

[0069] in,

[0070]

[0071] Where η1 is the learning rate, L is the Lipschitz constant, and D i is the size of the local dataset of drone i, D is the size of the training set after filtering samples, N is the number of drones, α i ∈(0,1) is the sample screening variable, β i ∈{0,1} is the client screening variable, e i is the parameter error rate during the transmission process of UAV i, ε is a positive parameter, η2 is the meta-learning rate, t is the training round, θ is the waterfall threshold, E|h i | 2 is the channel gain, B U is the uplink bandwidth, p i is the transmission power of UAV i, N0 is the power spectrum density of Gaussian noise, φ is the intermediate variable expressed in the simplified formula, t is the result of recursive summation of the intermediate variable φ in the global round

[0072] (2) The communication delay of the UAV in one round of global training is calculated based on the UAV uplink / downlink rate and the number of model parameters. The computational delay of the UAV in one round of global training is calculated based on the computational frequency of the UAV's onboard embedded device and the size of the training dataset.

[0073] The uplink communication delay T in the present invention i U , downlink communication delay T i D and calculation delay T i C The calculation process is as follows:

[0074]

[0075] in,

[0076]

[0077] Where S i (ω) is the number of model parameter bits uploaded by UAV i, S(Ω) is the number of global model parameter bits broadcast, B U 、B D are uplink / downlink bandwidth, p i 、p B are the transmission powers of UAV i and base station respectively, N0 is the power spectrum density of Gaussian noise, c i The number of CPU cycles required to process a single data sample for drone i, g i is the CPU frequency of drone i, s is the speed of light, f c is the carrier frequency, v i is the flight speed of UAV i, d i is the distance from UAV i to the base station, are the line-of-sight link / non-line-of-sight link losses, ζ i is the probability of line-of-sight link between UAV i and the base station.

[0078] (3) The communication energy consumption of the user in one round of global training is calculated based on the drone's transmission power and communication delay, and the computing energy consumption of the drone in one round of global training is calculated based on the computing frequency of the drone's onboard embedded device and the size of the training data set.

[0079] The communication energy consumption E in the present invention i and calculate energy consumption The calculation process is as follows:

[0080]

[0081] Where, Calculate the chip effective capacitance coefficient for drone i.

[0082] (4) Decompose the original problem into two optimization problems: drone scheduling and sample screening, based on the nature of the decision variables. Repeat the iterations until the expected convergence bound is minimized, and the optimal solution for client scheduling and sample screening is obtained. The specific steps include the following:

[0083] In step 4.1, an initial feasible solution for the sample screening variable is selected. The original problem is transformed into a mixed integer linear programming problem to solve the optimal solution for client scheduling, which greatly reduces the difficulty of solving the original problem. Therefore, based on the first delay constraint, the first energy consumption constraint, and the basic restrictions of the client screening variable, the client scheduling is optimized through the branch and bound method to obtain the optimal solution for the current client screening variable.

[0084] The first delay constraint involved in this process is specifically:

[0085]

[0086] Where, T th Indicates the upper limit of the delay of a round of training, Screen the initial feasible solutions of the variables for the sample.

[0087] The first energy consumption constraint is:

[0088]

[0089] Where, E th Indicates the upper limit of energy consumption for one round of training.

[0090] The basic restrictions on client-side filter variables are:

[0091] β i ∈{0,1}

[0092] In step 4.2, substitute the optimal solution of the client screening variable obtained in step 4.1 into the original problem and transform it into a fractional linear programming problem. Combined with the second delay constraint, the second energy consumption constraint, the gradient error constraint, and the basic restrictions of the sample screening variable, the sample screening variable is optimized using the Dinkelbach algorithm to obtain the optimal solution for the current sample screening variable.

[0093] The second delay constraint involved in this process is specifically:

[0094]

[0095] Where, Optimal solution for the client screening variables found in step 4.1.

[0096] The second energy consumption constraint is:

[0097]

[0098] The gradient error is:

[0099]

[0100] Where, is the variability of the gradient data, is the variability of the Hessian data, ▽ th is the gradient error threshold.

[0101] The basic restrictions on sample screening variables are:

[0102] 0≤α i =α i,1 +α i,2 +α i,3 ≤1

[0103] Where, α i,1 , α i,2 , α i,3 Represent the screening variables of three independent small batches of samples.

[0104] In step 4.3, repeat steps 4.1 to 4.2 alternately until the relative error of the trade-off objectives of the two iterations is less than the predefined threshold, the global training loss is minimized, and the optimal solution for drone scheduling and sample screening is obtained.

[0105] The global training loss involved in the process of minimization is t→∞, and the expected convergence bound G loss The expression:

[0106] When φ>1, the t→∞ expected convergence bound diverges, so this situation is not considered;

[0107] When φ<1, the t→∞ expected convergence bound converges,

[0108]

[0109] In the formula, the numerator and denominator both contain And the signs are different, so the problem can be directly optimized. Finally, the minimized global training loss is:

[0110]

[0111] Where α represents the sample screening variable set, and β represents the client screening variable set.

[0112] Figure 1 This paper describes an air-ground collaborative online federated learning scenario under unreliable communication conditions. This embodiment considers a total of N drones equipped with embedded intelligent computing devices performing collaborative target detection tasks coordinated by a central server. Each drone acts as an independent client, with the goal of collaboratively training a target detection model and does not share its own dataset. In the considered scenario, drones move to different locations to collect image data in real time. During local model training, whenever new samples are added, the drones filter the samples based on the local dataset distribution and use meta-gradients to update the local model. Subsequently, the drones transmit their trained local models to the base station. Due to the unreliability of the air-ground wireless channel, model transmission may encounter varying degrees of parameter error. We use the Bernoulli distribution to describe the parameter error rate. To prevent the global model accuracy from decreasing due to aggregating models with high parameter error rates, the base station must filter drones to perform aggregation.

[0113] Figure 2 This is a flow chart of an air-ground collaborative online federated learning method for unreliable communications. ① The base station broadcasts the initial model. ② The drone collects image data in real time and adds new samples to the original dataset. ③ The drone filters samples. ④ The drone performs local training using the filtered training set. ⑤ The drone uploads the local model to the base station. ⑥ Clients filter. ⑦ The base station performs global aggregation. ⑧ Repeat the above steps. It is important to note that sample and client selection are joint decisions made by solving an optimization problem, performed separately by the drone and the base station.

[0114] Figure 3The convergence of the method of the present invention was numerically simulated. In general, the algorithm converged after about 6 to 8 iterations, and the loss value after convergence was about 70% lower than the initial value. This method takes into account online learning and wireless factors, optimizes client selection and sample screening, reduces the impact of wireless transmission errors on federated learning training, and improves the performance of federated learning. In addition, we also simulated the convergence under different numbers of clients and different model parameter errors. As can be seen from the figure, with lower parameter error rates (that is, when the drone is closer to the base station to perform the training task), the final loss value decreases as the number of drones increases, while the opposite is true for higher error rates. This is because the increase in the number of drones provides more data for federated learning training, thereby improving the approximate accuracy of the gradient of the loss function at low parameter error rates. On the contrary, at high parameter error rates, the number of drones that cannot participate in aggregation increases.

[0115] Figure 4 The vertical axis represents the expected convergence bound, which is the difference between the proposed method and the comparison algorithm. Figure 4 It can be seen that the difference between the simulation results obtained by the method of the present invention and the brute force search is less than 7%, which is better than the other three baselines. This confirms the effectiveness of the present invention. Figure 4 The increasing trend of the two curves also shows that as the number of drones increases, the number of drones that cannot participate in aggregation will also increase.

[0116] Figure 5 All the curves in the figure show a common feature, that is, the target value increases with the increase of bandwidth due to the negative impact of the increase in parameter error rate on model performance. The difference lies in the slope of each curve. When the bandwidth is high, the parameter error rate is also high. As the number of drones increases, the number of drones that do not participate in the aggregation will also increase, resulting in a decrease in the target value. This is consistent with the Figure 3 The results in accord with .

[0117] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal translation scripting language JavaScript, etc.

[0118] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0119] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0120] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions for executing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0121] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0122] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A method for online federated learning of air-ground collaboration for unreliable communication, characterized by: The following steps are involved: Step 1: The drone filters sample images collected in real time in a dynamic environment and generates a local personalized model using meta-gradient update training. The Bernoulli distribution is introduced to describe the probability that errors in model parameter uploads by the drone in unstable channels will affect the accuracy of the global model. The convergence of air-ground collaborative online federated learning with parameter errors is analyzed, and the expected convergence bound of air-ground collaborative online federated learning with parameter errors is used to represent the global training loss. Step 2: Calculate the uplink and downlink communication delays of the UAV in one round of global training based on the UAV's uplink / downlink rates and the number of model parameters. Calculate the computational delay of the UAV in one round of global training based on the computational frequency of the UAV's onboard embedded device and the size of the training dataset. Step 3: Calculate the user's communication energy consumption in one round of global training based on the drone's transmit power and communication delay. Calculate the drone's computing energy consumption in one round of global training based on the computing frequency of the drone's onboard embedded device and the size of the training dataset. Step 4: Based on the idea of ​​alternating optimization, the problem of minimizing the global training loss is decomposed into the client scheduling problem and the sample screening problem. Based on the energy consumption constraints, delay constraints, and basic restrictions on optimization variables, the optimal solution for drone scheduling and sample screening is obtained through iterative optimization.

2. The air-ground collaborative online federated learning method for unreliable communication according to claim 1 is characterized by: In step 1, the expected convergence bound G of air-ground collaborative online federated learning with parameter errors is loss The calculation formula is: in, Where η1 is the learning rate, L is the Lipschitz constant, and D i is the size of the local dataset of drone i, D is the size of the training set after filtering samples, N is the number of drones, α i ∈(0,1) is the sample screening variable, β i ∈{0,1} is the client screening variable, e i is the parameter error rate during the transmission process of UAV i, ε is a positive parameter, η2 is the meta-learning rate, t is the training round, θ is the waterfall threshold, E|h i | 2 is the channel gain, B U is the uplink bandwidth, p i is the transmission power of UAV i, N0 is the power spectrum density of Gaussian noise, φ is the intermediate variable, φ t is the result of recursive summation of the intermediate variable φ in the global round.

3. The air-ground collaborative online federated learning method for unreliable communication according to claim 1 is characterized by: In step 2, the uplink communication delay T i U , downlink communication delay T i D and calculation delay T i C The calculation formula is: in, Where S i (ω) is the number of model parameter bits uploaded by UAV i, S(Ω) is the number of global model parameter bits broadcast, B U 、B D are uplink / downlink bandwidth, p i 、p B are the transmission powers of UAV i and base station respectively, N0 is the power spectrum density of Gaussian noise, c i The number of CPU cycles required to process a single data sample for drone i, g i is the CPU frequency of drone i, s is the speed of light, f c is the carrier frequency, v i is the flight speed of UAV i, d i is the distance from UAV i to the base station, are the line-of-sight link / non-line-of-sight link losses, ζ i is the probability of line-of-sight link between UAV i and the base station, α i ∈(0,1) is the sample screening variable.

4. The air-ground collaborative online federated learning method for unreliable communication according to claim 1 is characterized by: In step 3, the communication energy consumption E i and calculate energy consumption The calculation formula is: Where, Calculate the chip effective capacitance coefficient, S, for drone i i (ω) is the number of bits of model parameters uploaded by UAV i, p i is the transmission power of UAV i, B U is the uplink bandwidth, E|h i | 2 is the channel gain, α i ∈(0,1) is the sample screening variable, g i is the CPU frequency of drone i, β i ∈{0,1} is the client screening variable, and N0 is the power spectral density of Gaussian noise.

5. The air-ground collaborative online federated learning method for unreliable communication according to claim 2 is characterized by: Step 4 specifically includes the following sub-steps: Step 4.1: Based on the first delay constraint, the first energy consumption constraint, the basic restrictions on the client screening variables, and the initial feasible solution of the sample screening variables, optimize the client scheduling using the branch and bound method to obtain the optimal solution for the current client screening variables; Step 4.2: Based on the optimal solution of the client screening variable obtained in step 4.1, combined with the second delay constraint, the second energy consumption constraint, the gradient error constraint, and the basic restrictions of the sample screening variable, the sample screening variable is optimized using the Dinkelbach algorithm to obtain the optimal solution for the current sample screening variable; Step 4.3: Repeat steps 4.1 to 4.2 alternately until the relative error of the optimization objective of the two iterations is less than the predefined threshold, the global training loss is minimized, and the optimal solution for client scheduling and sample screening is obtained.

6. The air-ground collaborative online federated learning method for unreliable communication according to claim 5 is characterized by: In step 4.1, the first delay constraint is calculated as follows: Where, T th Indicates the upper limit of the delay of a round of training, Screening the initial feasible solutions of variables for the sample; The first energy consumption constraint is calculated as follows: Where, E th Indicates the upper limit of energy consumption for one round of training; The basic limitations of client-side filter variables are as follows: β i ∈{0,1}。 7. The air-ground collaborative online federated learning method for unreliable communication according to claim 5 is characterized by: In step 4.2, the second delay constraint is calculated as follows: Where, Screen the optimal solution of variables for the client; The second energy consumption constraint is calculated as follows: The gradient error is calculated as follows: Where, is the variability of the gradient data, is the variability of the Hessian data, ▽ th is the gradient error threshold; The basic restrictions on sample screening variables are as follows: 0≤α i ＝α i,1 +α i,2 +α i,3 ≤1 Where, α i,1 , α i,2 , α i,3 Represent the screening variables of three independent small batches of samples.

8. The air-ground collaborative online federated learning method for unreliable communication according to claim 5 is characterized by: In step 4.3, when φ>1, the expected convergence bound of the minimized global training loss t→∞ diverges; When φ<1, the t→∞ expected convergence bound converges, The global training loss to be minimized is calculated as: Where α represents the sample screening variable set, and β represents the client screening variable set.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which enables a computer to execute the air-ground collaborative online federated learning method for unreliable communication as described in any one of claims 1 to 8.

10. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for air-ground collaborative online federated learning for unreliable communication as claimed in any one of claims 1 to 8 is implemented.

Images