A hierarchical aerial federated learning control method and device

Abstract

The present application discloses a hierarchical aerial federated learning control method and device. The scheme of the present application is based on a network architecture of hierarchical aerial federated learning from multiple edge devices in a cell to their base stations, and from base stations in each cell to an aggregation center. The aggregation center first obtains the data information of the edge devices, the channel information from the edge devices to the base stations, and from the base stations to the aggregation center, respectively, and constructs a communication parameter optimization model. The transmission power parameters of each edge device and base station and the noise reduction factor of the aggregation center are determined by an alternating optimization algorithm, and then the transmission power parameters are fed back to the corresponding edge devices and base stations respectively, so that they can transmit model learning signals under the condition of simultaneous transmission of the same resource block. The scheme performs interference coordination by optimizing the transmission power and noise reduction factor of the edge devices and base stations, thereby reducing the impact of interference signals, allowing the edge devices and base stations to exchange information under interference conditions, reducing communication delays, and improving training speed.

Description

Technical Field

[0001] The present application relates to the field of computer and communication technology, and in particular to a hierarchical aerial federated learning control method and device. Background Art

[0002] Federated learning is a distributed machine learning method that distributes model training tasks to multiple devices. Each device uses its own local data to train a shared machine learning model under the coordination of the server. Although federated learning can fully utilize the distributed computing and storage capabilities of the edge network, the frequent transmission of high-dimensional gradient parameters makes the wireless communication between devices and servers a bottleneck, resulting in challenges in implementing federated edge learning in wireless communication networks. Air computing technology uses the signal superposition characteristics of wireless link multiple access channels to complete function calculations in the air, allowing all devices to share the same spectrum and resource blocks for simultaneous transmission, overcoming the high latency problem caused by large-scale access and reducing wireless communication overhead. The above characteristics make air computing very suitable as a new air access solution, which is applied to federated edge learning and forms the basis for air federated learning.

[0003] The existing large-scale edge network structure manages devices by dividing them into cells. Each cell has a base station to aggregate signals for multiple devices. In order to overcome the high latency problem caused by large-scale access, the device aggregates the gradient information to the base station through air computing. After receiving the signal, the base station forwards it to the aggregation center through air computing to realize hierarchical air federated learning in large edge networks. However, this model is prone to interference problems between multiple cells in actual application. It is necessary to schedule each cell to perform air computing in turn to eliminate the interference signals generated in this process. Therefore, it affects the training speed of large-scale federated learning networks and limits the application of air federated learning in large-scale network scenarios. Summary of the invention

[0004] The present application provides a hierarchical aerial federated learning control method and device for solving the technical problem of low training efficiency of existing large-scale aerial federated learning networks.

[0005] In order to solve the above technical problems, the first aspect of the present application provides a hierarchical aerial federated learning control method, including:

[0006] According to a network architecture of hierarchical air federated learning from multiple edge devices in a cell to their base stations, and from base stations of each cell to an aggregation center, data information, first channel information, and second channel information of edge devices in the network architecture are obtained, wherein the data information of the edge device is the local gradient mean and standard deviation of the edge device, the first channel information is the channel information from each edge device to the base station in the network architecture, and the second channel information is the channel information from each base station to the aggregation center;

[0007] A communication parameter optimization model is constructed according to the edge device information, the first channel information, and the second channel information, so as to determine the transmission power parameters of each edge device and each base station and the noise reduction factor of the aggregation center by using the communication parameter optimization model and an alternating optimization algorithm, wherein the noise reduction factor is used by the aggregation center to perform noise reduction processing on the received model learning signal;

[0008] The calculated transmission power parameters are fed back to the corresponding edge devices and base stations respectively, so that each edge device and each base station transmits the model learning signal according to the currently received transmission power parameters under the condition of simultaneous transmission of the same resource block.

[0009] Preferably, each edge device and each base station transmits the model learning signal according to the currently received transmit power parameter under the condition of simultaneous transmission of the same resource block, and then performs noise reduction processing on the model learning signal according to the noise reduction factor at the aggregation center end to obtain the final model learning information, which specifically includes:

[0010] Based on the received transmission power parameters, the edge device sends the device gradient signal obtained from the local device model to the corresponding target base station through over-the-air calculation, so that the target base station updates the local cell model based on the received device gradient signal or aggregates the device gradient signal into base station gradient data and then uploads it to the aggregation center. The aggregation center uses the noise reduction factor to perform noise reduction on the aggregated base station gradient data to obtain global gradient data, which is used to update the global model of the aggregation center.

[0011] Preferably, the communication parameter optimization model is specifically:

[0012]

[0013] In the formula, Representing edge devices The transmission power, Representative base station The transmission power, represents the noise reduction factor at the aggregation center, represents the learning rate, B represents the gradient variance, represents the smoothness of the loss function, is the aggregation error representation, represents the dimension of the gradient, n is the number of global iterations, It is the difference between the first round of global iterative loss function and the global optimal loss function, representing the initialization optimal gap.

[0014] Preferably, determining the transmission power parameters of each edge device and each base station and the noise reduction factor of the aggregation center by using the communication parameter optimization model and the alternating optimization algorithm specifically includes:

[0015] Calculate a first aggregation error representation according to the current transmit power parameter and the noise reduction factor in combination with an aggregation error representation calculation formula;

[0016] Calculating a relative difference between the first aggregation error representation and the second aggregation error representation, wherein the second aggregation error representation is obtained based on a transmit power parameter and a noise reduction factor of a previous alternating optimization round in combination with the aggregation error representation calculation formula;

[0017] If the relative difference is less than a preset tolerance threshold or the number of iterations reaches a preset upper limit, the transmission power parameters and noise reduction factors of each edge device and each base station are determined based on the current output result.

[0018] Preferably, the aggregation error characterization calculation formula is specifically:

[0019]

[0020] In the formula, Representing edge devices The transmission power, Representative base station The transmission power, represents the noise reduction factor at the aggregation center, is the aggregation error representation, n is the number of iterations, Representative equipment To base station The channel coefficients, Represents other cells' equipment To base station The channel coefficients, Represents other cells' equipment To its associated base station The channel coefficients of stands for conjugate transpose, represents the standard deviation of the gradient, Representative base station The channel coefficient to the aggregation center.

[0021] A second aspect of the present application provides a hierarchical aerial federated learning control device, comprising:

[0022] A channel information acquisition unit is used to acquire data information, first channel information, and second channel information of edge devices in the network architecture according to a network architecture of hierarchical air federated learning from multiple edge devices in a cell to their base stations and from base stations of each cell to an aggregation center, wherein the data information of the edge device is a local gradient mean and standard deviation of the edge device, the first channel information is channel information from each edge device to a base station in the network architecture, and the second channel information is channel information from each base station to the aggregation center;

[0023] A communication parameter optimization unit, configured to construct a communication parameter optimization model according to the data information of the edge device, the first channel information, and the second channel information, so as to determine the transmission power parameters of each edge device and each base station and the noise reduction factor of the aggregation center by using the communication parameter optimization model and an alternating optimization algorithm, wherein the noise reduction factor is used by the aggregation center to perform noise reduction processing on the received model learning signal;

[0024] The communication control unit is used to feed back the calculated transmission power parameters to the corresponding edge devices and base stations respectively, so that each edge device and each base station transmits the model learning signal according to the currently received transmission power parameters under the condition of simultaneous transmission of the same resource block.

[0025] Preferably, the communication control unit is specifically used for:

[0026] The calculated transmission power parameters are fed back to the corresponding edge devices and base stations respectively, so that the edge devices can send the device gradient signal obtained by the local device model to the corresponding target base station through over-the-air calculation according to the received transmission power parameters, so that the target base station can update the local cell model according to the received device gradient signal or aggregate the device gradient signal into base station gradient data and then upload it to the aggregation center. The aggregation center uses the noise reduction factor to perform noise reduction on the aggregated base station gradient data to obtain global gradient data, which is used to update the global model of the aggregation center.

[0027] Preferably, the communication parameter optimization model is specifically:

[0028]

[0029] In the formula, Representing edge devices The transmission power, Representative base station The transmission power, represents the noise reduction factor at the aggregation center, represents the learning rate, B represents the gradient variance, represents the smoothness of the loss function, is the aggregation error characterization, represents the dimension of the gradient, n is the number of global iterations, It is the difference between the first round of global iterative loss function and the global optimal loss function, representing the initialization optimal gap.

[0030] Preferably, the communication parameter optimization unit is specifically used for:

[0031] According to the data information of the edge device, the first channel information and the second channel information, a communication parameter optimization model is constructed, and the transmission power parameters of each edge device and each base station and the noise reduction factor of the aggregation center are determined by an alternating optimization algorithm, specifically including:

[0032] Calculate a first aggregation error representation according to the current transmit power parameter and the noise reduction factor in combination with an aggregation error representation calculation formula;

[0033] Calculating a relative difference between the first aggregation error representation and the second aggregation error representation, wherein the second aggregation error representation is obtained based on a transmit power parameter and a noise reduction factor of a previous alternating optimization round in combination with the aggregation error representation calculation formula;

[0034] If the relative difference is less than a preset tolerance threshold or the number of iterations reaches a preset upper limit, the transmission power parameters and noise reduction factors of each edge device and each base station are determined based on the current output result.

[0035] 10. A hierarchical aerial federated learning control device according to claim 4, characterized in that the aggregate error characterization calculation formula is specifically:

[0036]

[0037] In the formula, Representing edge devices The transmission power, Representative base station The transmission power, represents the noise reduction factor at the aggregation center, is the aggregation error representation, n is the number of iterations, Representative equipment To base station The channel coefficients, Represents the equipment of other cells To base station The channel coefficients, Represents other cells' equipment To its associated base station The channel coefficients of stands for conjugate transpose, represents the standard deviation of the gradient, Representative base station The channel coefficient to the aggregation center.

[0038] It can be seen from the above technical solutions that this application has the following advantages:

[0039] The scheme of the present application first obtains the data information of the edge devices in the network architecture, the channel information from each edge device to the base station, and the channel information from each base station to the aggregation center according to the network architecture of hierarchical air federated learning from multiple edge devices in a cell to their base stations, and from each cell base station to the aggregation center, and then constructs a communication parameter optimization model based on the obtained information, so as to determine the transmission power parameters of each edge device and each base station and the noise reduction factor of the aggregation center through the communication parameter optimization model using an alternating optimization algorithm, and then feeds back the calculated transmission power parameters to the corresponding edge devices and base stations respectively, so that each edge device and each base station transmits the model learning signal according to the currently received transmission power parameters under the condition of simultaneous transmission of the same resource block, and then obtains the final model learning information after the model learning signal is denoised according to the noise reduction factor at the aggregation center end. This scheme performs interference coordination by optimizing the transmission power and noise reduction factor of the edge devices and base stations, thereby reducing the impact of interference signals, so that the edge devices and base stations can maintain communication quality while also being able to exchange information under the condition of simultaneous transmission of the same resource block, thereby reducing the communication delay between devices and improving the training speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0041] Figure 1 Schematic diagram of the hierarchical aerial federated learning system architecture provided for this application.

[0042] Figure 2 A flowchart of an embodiment of a hierarchical over-the-air federated learning control method provided in this application.

[0043] Figure 3 A comparison chart of the optimal gap between a hierarchical aerial federated learning control method provided in this application and existing methods.

[0044] Figure 4A comparison chart of the prediction errors of a hierarchical aerial federated learning control method provided in this application and existing methods.

[0045] Figure 5 A schematic diagram of the structure of a hierarchical aerial federated learning control device provided for this application. DETAILED DESCRIPTION

[0046] In response to the problems existing in large-scale airborne federated learning, research has found that since airborne computing utilizes the same resource block characteristics of the wireless link multiple access channel, if airborne computing is performed in different cells at the same time, interference between cells will occur. Although the existing approach uses time division multiple access, each cell performs airborne computing in turn to eliminate the interference signals that occur in this process. However, this strategy will increase the delay as the network scale increases, attenuating the training speed of federated learning and cannot be applied in large networks. As a result, the current airborne federated learning technology is limited to a small-scale network structure of a single cell and is difficult to apply in a large-scale network structure involving multiple cells.

[0047] The present application provides a hierarchical aerial federated learning control method and device for solving the technical problem of low training efficiency of existing large-scale aerial federated learning networks.

[0048] In order to make the purpose, features, and advantages of the invention of this application more obvious and easy to understand, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the embodiments described below are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0049] First, a detailed description of an embodiment of a hierarchical aerial federated learning control method provided by the present application is as follows:

[0050] This application is a hierarchical air federated learning control method and device for large edge networks. This solution consists of an aggregation center, multiple base stations, and multiple edge devices. The aggregation center is configured with a global model, each base station is configured with a cell model in the middle layer, and each edge device is configured with a bottom-level device model. The overall architecture is as follows: Figure 1As shown. Define that a cell iteration is when a device exchanges a device gradient signal / cell model with a base station. Define that after multiple rounds of cell iterations, the base station sends a base station gradient signal to the aggregation center, the aggregation center updates the global model, and then broadcasts the new global model to all base stations and devices as a global iteration. This solution optimizes the transmission power and noise reduction factor of devices and base stations, and coordinates interference between cells, so that hierarchical air federated learning with multiple cell iterations can reach convergence faster and improve the performance of model training.

[0051] See also Figure 2 , this embodiment provides a hierarchical aerial federated learning control method, including:

[0052] Step 101, according to the network architecture of hierarchical air federated learning from multiple edge devices in a cell to their base stations, and from each cell base station to an aggregation center, data information, first channel information, and second channel information of edge devices in the network architecture are obtained;

[0053] The data information of the edge device is the local gradient mean and standard deviation of the edge device, the first channel information is the channel information from each edge device to the base station in the network architecture, and the second channel information is the channel information from each base station to the aggregation center;

[0054] It should be noted that, first of all, according to the network architecture of hierarchical aerial federated learning from multiple edge devices in the cell to their base stations, and from each cell base station to the aggregation center, the aggregation center broadcasts the global model to all base stations and devices under its jurisdiction, obtains the edge device data information in the network architecture, and the first channel information from each edge device to the base station and the second channel information from each base station to the aggregation center, which can be summarized into a unified central control module for processing. The central control module mentioned here is generally integrated in the aggregation center.

[0055] Step 102: construct a communication parameter optimization model according to the data information, the first channel information, and the second channel information of the edge device. The central control module uses an alternating optimization algorithm to optimize the communication parameter optimization model, first fix the base station transmit power coefficient and the noise reduction factor to optimize the device transmit power coefficient, then fix the device transmit power coefficient to optimize the base station transmit power coefficient and the noise reduction factor, and perform iterative optimization to determine the transmit power parameters and noise reduction factors of each edge device and each base station.

[0056] Among them, the noise reduction factor is used by the aggregation center to perform noise reduction processing on the received model learning signal;

[0057] It should be noted that, according to the information obtained in step 101, a communication parameter optimization model is constructed to perform interference coordination operations through the communication parameter optimization model to determine the transmission power parameters and noise reduction factors of each edge device and each base station.

[0058] Step 103: Feedback the calculated transmission power parameters to the corresponding edge devices and base stations respectively, so that each edge device and each base station transmits the model learning signal according to the currently received transmission power parameters under the condition of simultaneous transmission of the same resource block.

[0059] It should be noted that, according to the transmit power parameters and noise reduction factors of each edge device and each base station calculated in step 102, the noise reduction factor is left to the aggregation center, and when the aggregation center receives the model learning signal of the subordinate base station, the received model learning signal is subjected to noise reduction processing using the corresponding noise reduction factor, so as to suppress noise and reduce errors, which is more conducive to the update training of the global model, and the transmission power parameters are fed back to the corresponding edge devices and base stations respectively, so that each edge device can communicate with the base station to which it belongs according to the currently received transmit power parameters under the condition of simultaneous transmission of the same resource block, so as to calculate the gradient signal of the base station aggregation device through the air;

[0060] Similarly, the base station can also aggregate the base station gradient signal to the aggregation center through air calculation according to the received transmission power parameter. More specifically, all base stations obtain the base station gradient signal through the aggregation device gradient signal to determine whether it is the last round of cell iteration; if not, the base station gradient signal is used to update the cell model, and the cell model is broadcast to the associated edge device to perform the next round of cell iteration; if so, the communication controllers on all base stations aggregate the base station gradient signal to the aggregation center through air calculation according to the obtained base station transmission power, and finally the aggregation center uses the noise reduction factor to perform noise reduction processing on the aggregated base station gradient signal to obtain the global gradient signal;

[0061] Among them, the device gradient signal is specifically a gradient signal calculated by the edge device based on the local learning model according to the input of the global\cell model. The base station gradient signal is aggregated by the base station according to the received device gradient signals. It can be used for training and updating the local cell model or uploaded to the aggregation center to obtain the global gradient signal for updating the global model.

[0062] by Figure 1 As an example, each cell has a base station and associated devices , In the first round of global iteration, the aggregation center broadcasts the initialized global model to all base stations and devices. The devices use the gradient calculation module to calculate the device gradient signal according to the model. When the same frequency and resource block are transmitted simultaneously and inter-cell interference occurs, the device transmission power optimized by the central control module is used for communication control. The device gradient signal is aggregated to the base station through air calculation. The base station updates the cell model and broadcasts the new cell model to the associated devices for the next cell iteration. After the cell iteration, the base station uses the base station transmission power optimized by the central control module for communication control, and aggregates the base station gradient signal to the aggregation center through air calculation again. The aggregation center performs noise reduction and global model update, and broadcasts the new global model to all base stations and devices for the next global iteration. This global iteration will continue wheel.

[0063] Select a specific round of global iteration The last round of cell iteration process is used as the research. First, all devices use the stochastic gradient descent algorithm to calculate the local loss function gradient of random data , the calculation formula is as follows:

[0064] (1)

[0065] in Represents the size of the local mini-batch training set data, and Respectively represent The data and their correct label values, Represents when the device Model The local loss function gradient is obtained based on the data and its correct label value.

[0066] In order to make the gradient unbiased, the device can normalize the local gradient and obtain the normalized signal to be transmitted. , the calculation formula is as follows:

[0067] (2)

[0068] (3)

[0069] represents the mean of the local gradient, represents the dimension of the gradient, represents the standard deviation of the gradient, represent The vector Before transmitting, the device performs channel pre-compensation. Multiply by the precoding factor ,in Representative equipment Transmit power, Representative equipment To base station The channel coefficients of represents the conjugate transpose. From this we can get time slots, base station Received signal for:

[0070] (4)

[0071] in represents the complex symmetric Gaussian white noise at the base station end, and the last item represents the inter-cell interference signal caused by the simultaneous transmission of the same frequency and resource block.

[0072] After receiving their respective signals, each base station multiplies them by the precoding factor .in Representative base station Transmit power, Representative base station The channel coefficient to the aggregation center. time slots, the signal received by the aggregation center for:

[0073] (5)

[0074] in, Represents the complex symmetric Gaussian white noise at the aggregation center. After receiving the signal, the aggregation center performs an inverse normalization operation:

[0075] (6)

[0076] The local gradient mean With standard deviation Relative to the small dimension of the gradient, it is assumed that it can be transmitted to the aggregation center without loss. represents the unbiased value of the global gradient, represents the error in the gradient estimate, represents the noise reduction factor at the aggregation center. Furthermore, by taking The real part of is used to obtain an estimate of the global gradient, and we can get:

[0077] (7)

[0078] in To obtain the real part, superscript represents the normal transpose, , , . Therefore, the global model can be updated based on this gradient estimate, expressed as:

[0079] (8)

[0080] The difference between the global loss function of the maximum global iteration round and the optimal global loss function, that is, the optimal gap, is derived as the convergence performance indicator of federated learning. Its expression is as follows:

[0081] (9)

[0082] in , , represents the learning rate, represents the gradient variance, Represents the smoothness of the loss function. Aggregate error characterization (Mean Squared Error, MSE) is the most important factor affecting the optimal gap, and its expression is:

[0083] (10)

[0084] The optimization goal of this application is to minimize the optimal gap to improve the convergence performance of federated learning. The optimization variables include the base station transmission power , device transmission power and noise reduction factor Therefore, the following questions can be obtained:

[0085] (11)

[0086] Through theoretical derivation, it can be proved that minimizing the optimal gap is equivalent to minimizing the MSE. Assuming that each round of global iterations can be optimized independently, the superscript of the global iteration round is omitted. , so the problem P0 can be equivalently transformed into the following MSE minimization problem:

[0087] (12)

[0088] in , , For each device's maximum power constraint, Represents the maximum power constraint of each base station.

[0089] More specifically, because the optimization problem (P1) is a non-convex problem that is difficult to solve, the problem is solved by an alternating optimization algorithm, which decomposes (P1) into two sub-problems, alternately fixes the optimization variables, and iterates the solution.

[0090] First, fix the base station transmission power and noise reduction factor , optimize the device transmission power . To facilitate the solution, variable substitution is performed , then sub-problem 1 can be expressed as:

[0091] (13)

[0092] in It is easy to see that the above problem is convex and satisfies the Slater condition, so the Lagrange dual method and KKT condition can be used to solve the problem. Finally, it can be obtained that when the base station transmits power and noise reduction factor Optimal solution for device transmit power for:

[0093] (14)

[0094] in is the Lagrange multiplier, which satisfies the complementary relaxation condition .

[0095] For sub-problem 2, the fixed device transmit power , optimize base station transmission power and noise reduction factor , also for the convenience of solving, variable substitution is performed , , we can get sub-problem 2 as:

[0096] (15)

[0097] in Similarly, it is easy to see that the above problem is a convex problem and satisfies the Slater condition. Therefore, the Lagrange dual method and KKT condition can be used to solve the problem. Finally, the optimal solution of the variables replaced as follows can be obtained:

[0098] (16)

[0099] (17)

[0100] in , is the Lagrange multiplier, which satisfies the complementary relaxation condition , so the transmission power of the fixed optimized device can be obtained The optimal solution for base station transmission power is And the optimal solution of noise reduction factor The are:

[0101] , (18)

[0102] The above-mentioned power control schemes for devices and base stations all follow a regularized channel inversion structure. At the same time, it can be observed from (14) that the aggregation center needs to balance the maximum useful signal power and the minimum interference signal power by adjusting the device transmission power to perform interference coordination.

[0103] From this we can get that The process of the global iterative alternating optimization algorithm is as follows:

[0104] 1. Set the maximum number of alternating optimization algorithms ; Tolerance threshold , initialize the variable , calculate the initial MSE function value (i.e. the objective function value in formula (10)), denoted as ;

[0105] 2. Order , No. The process of round iteration is as follows;

[0106] 3. Fixed According to (12), we can get the optimal solution for the device transmission power: ;

[0107] 4. Fixation According to (14)-(16), we can get the optimal solution for base station transmission power: And the optimal solution of noise reduction factor ;

[0108] 5. Given The following calculation formula (10) ;

[0109] 6. Calculate relative difference , to determine whether it is less than the tolerance threshold , if greater than , then let , repeat steps 3-5 again; if the relative difference is less than the threshold or Greater than Then exit the loop and output This is the desired result.

[0110] In order to more intuitively demonstrate the effect of the technical solution of this embodiment, the present application also provides an explanation of a simulation example based on the above method, which is as follows:

[0111] In the simulation, in addition to the proposed optimization scheme, fixed transmission power and noise reduction factor are also considered as reference schemes to verify the effectiveness of the scheme. The ridge regression loss function, as well as the training set and test set data are constructed in MATLAB; the performance measurement indicators use the optimal gap and prediction error, where the prediction error represents the loss function value of the test set in machine learning.

[0112] In the simulation, the following parameters are used: , , , , , , , , , , , , , the test set data size is 200, and the training set data size on each device is 100. The channel meets , ,in is the channel attenuation value at the reference distance, is the large-scale fading coefficient, where Representative equipment To base station The distance Representative base station Distance to the aggregation center. , .

[0113] Figure 3 The comparison chart of the optimal gap between the proposed optimization scheme and the scheme with fixed transmission power and noise reduction factor is shown. It can be seen that with the increase of global iteration rounds, the optimal gaps of both schemes show a downward trend, indicating that the global loss function is getting closer and closer to the global optimal loss function as the communication rounds of air calculation decrease; because the proposed optimization scheme reduces the interference error between cells through power optimization and performs interference coordination, compared with the comparison scheme, the proposed scheme has a smaller optimal gap, indicating that the proposed power control scheme has better machine learning training effect, is faster and closer to the global optimal loss function.

[0114] Figure 4The comparison chart of the prediction error of the proposed optimization scheme and the fixed transmission power and noise reduction factor scheme is shown. It can be seen that with the increase of global iteration rounds, the prediction errors of both schemes show a downward trend, and finally tend to converge. Because the proposed optimization scheme reduces the inter-cell interference error through power optimization control, it reaches a convergence state after 60 rounds of global iterations compared with the fixed transmission power and noise reduction factor scheme, indicating that the proposed optimization scheme has a faster convergence speed and better machine learning performance on the test set, which shows the effectiveness of the proposed algorithm.

[0115] The above is a detailed description of an embodiment of a hierarchical aerial federated learning control method provided in the present application. The following is a detailed description of an embodiment of a hierarchical aerial federated learning control device provided in the present application.

[0116] See also Figure 5 The embodiment of the present application provides a hierarchical air federated learning control device, corresponding to the central control module that can be integrated in the aggregation center mentioned in the previous embodiment, including:

[0117] The channel information acquisition unit 201 is used to acquire data information, first channel information, and second channel information of edge devices in the network architecture according to the network architecture of hierarchical air federated learning from multiple edge devices in the cell to their base stations and from each cell base station to the aggregation center, wherein the data information of the edge device is the local gradient mean and standard deviation of the edge device, the first channel information is the channel information from each edge device to the base station in the network architecture, and the second channel information is the channel information from each base station to the aggregation center;

[0118] A communication parameter optimization unit 202 is used to construct a communication parameter optimization model according to the data information of the edge device, the first channel information and the second channel information, so as to determine the transmission power parameters of each edge device and each base station and the noise reduction factor of the aggregation center by using the communication parameter optimization model and an alternating optimization algorithm, wherein the noise reduction factor is used by the aggregation center to perform noise reduction processing on the received model learning signal;

[0119] The communication control unit 203 is used to feed back the calculated transmission power parameters to the corresponding edge devices and base stations respectively, so that each edge device and each base station transmits the model learning signal according to the currently received transmission power parameters under the condition of simultaneous transmission of the same resource block.

[0120] Furthermore, the communication control unit 203 is specifically configured to:

[0121] The calculated transmission power parameters are fed back to the corresponding edge devices and base stations respectively, so that the edge devices can send the device gradient signal obtained by the local device model to the corresponding target base station through over-the-air calculation according to the received transmission power parameters, so that the target base station can update the local cell model according to the received device gradient signal or aggregate the device gradient signal into base station gradient data and then upload it to the aggregation center. The aggregation center uses the noise reduction factor to perform noise reduction on the aggregated base station gradient data to obtain global gradient data, which is used to update the global model of the aggregation center.

[0122] Furthermore, the communication parameter optimization model is specifically as follows:

[0123]

[0124] In the formula, Representing edge devices The transmission power, Representative base station The transmission power, represents the noise reduction factor at the aggregation center, represents the learning rate, B represents the gradient variance, represents the smoothness of the loss function, is the aggregation error representation, represents the dimension of the gradient, n is the number of global iterations, It is the difference between the first round of global iterative loss function and the global optimal loss function, representing the initialization optimal gap.

[0125] Furthermore, the communication parameter optimization unit is specifically used for:

[0126] According to the data information of the edge device, the first channel information and the second channel information, a communication parameter optimization model is constructed, and the transmission power parameters of each edge device and each base station and the noise reduction factor of the aggregation center are determined by an alternating optimization algorithm. Specifically, the following are included:

[0127] Calculate a first aggregation error representation according to the current transmit power parameter and the noise reduction factor in combination with an aggregation error representation calculation formula;

[0128] Calculate a relative difference between the first aggregation error representation and the second aggregation error representation, wherein the second aggregation error representation is obtained based on the transmit power parameter and the noise reduction factor of the previous alternating optimization round combined with the aggregation error representation calculation formula;

[0129] If the relative difference is less than the preset tolerance threshold or the number of iterations reaches the preset upper limit, the transmission power parameters and noise reduction factors of each edge device and each base station are determined based on the current output results.

[0130] Furthermore, the aggregation error characterization calculation formula is specifically as follows:

[0131]

[0132] In the formula, Representing edge devices The transmission power, Representative base station The transmission power, represents the noise reduction factor at the aggregation center, is the aggregation error representation, n is the number of iterations, Representative equipment To base station The channel coefficients, Represents other cells' equipment To base station The channel coefficients, Represents other cells' equipment To its associated base station The channel coefficients of stands for conjugate transpose, represents the standard deviation of the gradient, Representative base station The channel coefficient to the aggregation center.

[0133] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the terminals, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0134] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0135] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application described herein can, for example, be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0136] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0137] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0138] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0139] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or part of the contribution to the prior art or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code. As described above, the above embodiments are only used to illustrate the technical solution of the present application, not to limit it; although the present application is described in detail with reference to the aforementioned embodiments, ordinary technicians in the field should understand that: it can still modify the technical solutions recorded in the aforementioned embodiments, or replace some of the technical features therein with equivalents; and these modifications or replacements do not make the essence of the corresponding technical solution deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A hierarchical aerial federated learning control method, characterized in that: include: According to a network architecture of hierarchical air federated learning from multiple edge devices in a cell to their base stations, and from base stations of each cell to an aggregation center, data information, first channel information, and second channel information of edge devices in the network architecture are obtained, wherein the data information of the edge device is the local gradient mean and standard deviation of the edge device, the first channel information is the channel information from each edge device to the base station in the network architecture, and the second channel information is the channel information from each base station to the aggregation center; According to the edge device information, the first channel information and the second channel information, a communication parameter optimization model is constructed to determine the transmission power parameters of each edge device and each base station and the noise reduction factor of the aggregation center by using the communication parameter optimization model and an alternating optimization algorithm, wherein the noise reduction factor is used by the aggregation center to perform noise reduction processing on the received model learning signal. The communication parameter optimization model is specifically: In the formula, represents the transmit power of edge device k, represents the transmission power of base station m, θ (n) represents the noise reduction factor at the aggregation center, η (n) represents the learning rate, B represents the gradient variance, and L represents the smoothness of the loss function. is the aggregation error representation, C represents the dimension of the gradient, It is the difference between the first round of global iterative loss function and the global optimal loss function, representing the initialization optimal gap, n is the number of iterations, represents the channel coefficient from device k to base station m, represents the channel coefficient from device i in other cells to base station m, represents the channel coefficient from device i in other cells to its associated base station j. The superscript H represents the conjugate transpose. represents the standard deviation of the gradient, represents the channel coefficient from base station m to the aggregation center; The calculated transmission power parameters are fed back to the corresponding edge devices and base stations respectively, so that each edge device and each base station transmits the model learning signal according to the currently received transmission power parameters under the condition of simultaneous transmission of the same resource block.

2. A hierarchical aerial federated learning control method according to claim 1, characterized in that: Under the condition of simultaneous transmission of the same resource block, each edge device and each base station transmits the model learning signal according to the currently received transmit power parameter, and then the aggregation center performs noise reduction processing on the model learning signal according to the noise reduction factor to obtain the final model learning information, which specifically includes: Based on the received transmission power parameters, the edge device sends the device gradient signal obtained from the local device model to the corresponding target base station through over-the-air calculation, so that the target base station updates the local cell model based on the received device gradient signal or aggregates the device gradient signal into base station gradient data and then uploads it to the aggregation center. The aggregation center uses the noise reduction factor to perform noise reduction on the aggregated base station gradient data to obtain global gradient data, which is used to update the global model of the aggregation center.

3. A hierarchical aerial federated learning control method according to claim 1, characterized in that: By using the communication parameter optimization model, the transmission power parameters of each edge device and each base station and the noise reduction factor of the aggregation center are determined by using an alternating optimization algorithm, specifically including: Calculate a first aggregation error representation according to the current transmit power parameter and the noise reduction factor in combination with an aggregation error representation calculation formula; Calculating a relative difference between the first aggregation error representation and the second aggregation error representation, wherein the second aggregation error representation is obtained based on a transmit power parameter and a noise reduction factor of a previous alternating optimization round in combination with the aggregation error representation calculation formula; If the relative difference is less than a preset tolerance threshold or the number of iterations reaches a preset upper limit, the transmission power parameters and noise reduction factors of each edge device and each base station are determined based on the current output result.

4. A hierarchical aerial federated learning control device, characterized in that: include: A channel information acquisition unit is used to acquire data information, first channel information, and second channel information of edge devices in the network architecture according to a network architecture of hierarchical air federated learning from multiple edge devices in a cell to their base stations and from base stations of each cell to an aggregation center, wherein the data information of the edge device is a local gradient mean and standard deviation of the edge device, the first channel information is channel information from each edge device to a base station in the network architecture, and the second channel information is channel information from each base station to the aggregation center; A communication parameter optimization unit is used to construct a communication parameter optimization model according to the data information of the edge device, the first channel information and the second channel information, so as to determine the transmission power parameters of each edge device and each base station and the noise reduction factor of the aggregation center by using the communication parameter optimization model and an alternating optimization algorithm, wherein the noise reduction factor is used by the aggregation center to perform noise reduction processing on the received model learning signal, and the communication parameter optimization model is specifically: In the formula, represents the transmit power of edge device k, represents the transmission power of base station m, θ (n) represents the noise reduction factor at the aggregation center, η (n) represents the learning rate, B represents the gradient variance, and L represents the smoothness of the loss function. is the aggregation error representation, C represents the dimension of the gradient, It is the difference between the first round of global iterative loss function and the global optimal loss function, representing the initialization optimal gap, n is the number of iterations, represents the channel coefficient from device k to base station m, represents the channel coefficient from device i in other cells to base station m, represents the channel coefficient from device i in other cells to its associated base station j. The superscript H represents the conjugate transpose. represents the standard deviation of the gradient, represents the channel coefficient from base station m to the aggregation center; The communication control unit is used to feed back the calculated transmission power parameters to the corresponding edge devices and base stations respectively, so that each edge device and each base station transmits the model learning signal according to the currently received transmission power parameters under the condition of simultaneous transmission of the same resource block.

5. A hierarchical aerial federated learning control device according to claim 4, characterized in that: The communication control unit is specifically used for: The calculated transmission power parameters are fed back to the corresponding edge devices and base stations respectively, so that the edge devices can send the device gradient signal obtained by the local device model to the corresponding target base station through over-the-air calculation according to the received transmission power parameters, so that the target base station can update the local cell model according to the received device gradient signal or aggregate the device gradient signal into base station gradient data and then upload it to the aggregation center. The aggregation center uses the noise reduction factor to perform noise reduction on the aggregated base station gradient data to obtain global gradient data, which is used to update the global model of the aggregation center.

6. A hierarchical aerial federated learning control device according to claim 4, characterized in that: The communication parameter optimization unit is specifically used for: According to the data information of the edge device, the first channel information and the second channel information, a communication parameter optimization model is constructed, and the transmission power parameters of each edge device and each base station and the noise reduction factor of the aggregation center are determined by an alternating optimization algorithm, specifically including: Calculate a first aggregation error representation according to the current transmit power parameter and the noise reduction factor in combination with an aggregation error representation calculation formula; Calculating a relative difference between the first aggregation error representation and the second aggregation error representation, wherein the second aggregation error representation is obtained based on a transmit power parameter and a noise reduction factor of a previous alternating optimization round in combination with the aggregation error representation calculation formula; If the relative difference is less than a preset tolerance threshold or the number of iterations reaches a preset upper limit, the transmission power parameters and noise reduction factors of each edge device and each base station are determined based on the current output result.

Images