A data transmission method, device and terminal equipment based on federated learning

Abstract

The application discloses a data transmission method, device and terminal equipment based on federal learning, when each local equipment of the data transmission method pre-encodes the uploaded first information symbol, can directly obtain the corresponding channel state estimation information according to the state evolution and estimation of the pilot signal sent by the equipment, and generate the channel matrix corresponding to the channel where each equipment is based on the channel state estimation information, so that the corresponding target precoder can be obtained, and the first information symbol is encoded to generate the second information symbol. Therefore, the application does not need to frequently exchange the channel response information of other time-varying channels to obtain the corresponding channel matrix, but directly estimates the corresponding channel response through the pilot signal corresponding to the current channel, so as to quickly encode the first information symbol that needs to be uploaded, without too much channel overhead, improve the performance of the system, and speed up the convergence speed of the corresponding model of federal learning.

Description

Technical Field

[0001] The present invention relates to the field of precoding technology for federated learning, and in particular to a data transmission method, apparatus, and terminal device based on federated learning. Background Art

[0002] First proposed in 2016, the concept of the federated learning (FL) framework involves users distributing training data locally and uploading model parameters to jointly train a global model. This ensures privacy and reduces communication costs. However, due to the communication and training challenges posed by wireless channels, federated learning is often difficult to implement in practice.

[0003] In the actual application scenarios of federated learning, when each device uploads model parameters, it needs to generate precoding corresponding to each information symbol in the channel, thereby encoding the information symbols and then uploading the corresponding data signal. In the application scenario of the Internet of Vehicles, since the devices in the Internet of Vehicles move at high speeds, the wireless channel between the base station and the Internet of Vehicles devices will change dynamically. As a result, in the federated learning of the Internet of Vehicles, the dynamically changing channels require frequent exchange of channel states. When facing dynamically changing channels, that is, when facing time-varying channels, existing precoding technologies need to frequently exchange channel state information of all devices participating in federated learning in order to perform precoding. This will increase the channel overhead of the system, resulting in a lower data transmission rate for federated learning and a slower convergence speed of the corresponding federated learning model. Summary of the Invention

[0004] The embodiments of the present invention provide a data transmission method, apparatus, and terminal device based on federated learning, which can effectively solve the problem in the prior art that channel state information of all devices participating in federated learning needs to be frequently exchanged before precoding can be performed, which will increase the channel overhead of the system and make the convergence speed of the model corresponding to federated learning slower.

[0005] An embodiment of the present invention provides a data transmission method based on federated learning, applicable to each local device participating in federated learning, including:

[0006] Predicting a pilot signal based on a preset state equation and a preset observation equation to obtain corresponding channel state estimation information; wherein the pilot signal is a signal sent by the local device to the server participating in federated learning in the target time frame;

[0007] Generating, based on the pilot signal and the channel state estimation information, a channel matrix corresponding to the channel where the device is located and used to characterize channel characteristics;

[0008] According to the channel matrix, a corresponding target precoder is generated, and a second information symbol is encoded based on the target precoder, to generate a first information symbol;

[0009] The second information symbol of each device is transmitted to the server, so that the server collects corresponding model training data in a target time frame according to each second information symbol.

[0010] Preferably, the pilot signal is predicted according to the preset state equation and the preset observation equation to obtain corresponding channel state estimation information, comprising:

[0011] The pilot signal is taken as the input of the Kalman filter;

[0012] Repeat the following prediction operation until the iteration number of the prediction operation is reached to generate the final channel state estimation information:

[0013] Based on the state equation and the observation equation corresponding to the current time of the Kalman filter, and the observation value of the pilot signal at the current time, the channel state estimation information predicted last time is re-predicted to obtain the current channel state estimation information and the current covariance;

[0014] After predicting a new covariance according to the current covariance, the Kalman gain is obtained based on the new covariance;

[0015] Based on the current channel state estimation information and the Kalman gain, the channel state estimation information of this prediction is generated;

[0016] When it is determined that the iteration number of the prediction operation has not been reached, the state equation and the observation equation of the Kalman filter are updated according to the observation value of the pilot signal and the channel state estimation information of this prediction.

[0017] Preferably, the channel matrix corresponding to the channel in which the device is located is generated according to the pilot signal and the channel state estimation information, which is used to represent the channel characteristics, comprising:

[0018] According to the pilot signal, and the channel impulse response calculated from the pilot signal and the received signal corresponding to the pilot signal;

[0019] According to the channel impulse response and the amplitude, phase and delay in the channel state estimation information, the channel matrix is generated.

[0020] Preferably, before generating the corresponding target precoder, it comprises:

[0021] Obtain the channel parameters of each device recorded in real time by the Kalman filter;

[0022] When a change in the channel parameters of the device is detected, the generated channel matrix is ​​updated according to the changed channel parameters, and the updated channel matrix is ​​used as the final channel matrix; wherein the final channel matrix is ​​used to generate the corresponding target precoder.

[0023] Preferably, encoding the first information symbol based on the target precoder to generate the second information symbol includes:

[0024] The first information symbol is encoded according to the following formula to generate the second information symbol:

[0025]

[0026]

[0027]

[0028] Among them, m dd is the second information symbol of the nth device, B dd,n is the channel matrix of the nth device, ρ n is the preset scalar value for the device to meet the transmit power constraint, O n is the target encoder, O x,n is the sub-matrix of the target encoder, g dd,n It is the first information symbol of the nth device.

[0029] Based on the above method embodiments, the present invention provides corresponding device embodiments.

[0030] An embodiment of the present invention provides a data transmission device based on federated learning, comprising: a channel state estimation information generation module, a channel matrix generation module, an encoding module, and an information symbol transmission module;

[0031] The channel state estimation information generation module is configured to predict the pilot signal based on a preset state equation and a preset observation equation to obtain corresponding channel state estimation information; wherein the pilot signal is a signal sent by the local device to the server participating in federated learning in the target time frame;

[0032] The channel matrix generation module is configured to generate a channel matrix corresponding to the channel where the device is located and used to characterize channel characteristics based on the pilot signal and the channel state estimation information;

[0033] The encoding module is configured to generate a corresponding target precoder according to the channel matrix, and encode the first information symbol based on the target precoder to generate a second information symbol;

[0034] The information symbol transmission module is configured to transmit the second information symbol of each device to the server, so that the server collects corresponding model training data in a target time frame according to each second information symbol.

[0035] Preferably, the channel state estimation information generation module is configured to predict the pilot signal according to a preset state equation and a preset observation equation to obtain corresponding channel state estimation information, including:

[0036] The channel state estimation information generation module is configured to take the pilot signal as an input of the Kalman filter.

[0037] The prediction operation is repeated until the iteration number of the prediction operation is reached, and the final channel state estimation information is generated.

[0038] Based on the state equation and the observation equation of the Kalman filter corresponding to the current time and the observation value of the pilot signal at the current time, the channel state estimation information predicted last time is re-predicted to obtain the current channel state estimation information and the current covariance.

[0039] After the new covariance is obtained by prediction according to the current covariance, the Kalman gain is obtained based on the new covariance.

[0040] Based on the current channel state estimation information and the Kalman gain, the channel state estimation information of this prediction is generated.

[0041] When it is determined that the iteration number of the prediction operation is not reached, the state equation and the observation equation of the Kalman filter are updated according to the observation value of the pilot signal and the channel state estimation information of this prediction.

[0042] Preferably, the channel matrix generation module is configured to generate a channel matrix corresponding to a channel in which the device is located and used to represent channel characteristics according to the pilot signal and the channel state estimation information, including:

[0043] The channel matrix generation module is configured to calculate a channel impulse response according to the pilot signal and a received signal corresponding to the pilot signal.

[0044] The channel matrix is generated according to the channel impulse response and the amplitude, phase and delay in the channel state estimation information.

[0045] On the basis of the method embodiments described above, the application further provides terminal device embodiments.

[0046] Another embodiment of the present application provides a terminal device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the data transmission method based on federated learning according to the above-mentioned embodiment of the present application when executing the computer program.

[0047] Based on the above-mentioned method embodiment, the present application correspondingly provides a storage medium embodiment.

[0048] Another embodiment of the present application provides a storage medium, wherein the computer readable storage medium comprises a stored computer program, wherein the computer readable storage medium controls the device where the computer readable storage medium is located to execute the data transmission method based on federated learning according to the above-mentioned embodiment of the present application when the computer program runs.

[0049] By implementing the present application, the following beneficial effects are achieved:

[0050] The embodiments of the present application provide a data transmission method, device and terminal equipment based on federated learning, in which, when each local device pre-encodes the first information symbol to be uploaded, the corresponding channel state estimation information can be obtained by directly performing state evolution and estimation according to the pilot signal transmitted by the device, and the channel matrix corresponding to the channel where each device is located is generated based on the channel state estimation information, so that the corresponding target precoder can be obtained, the first information symbol is encoded to generate the second information symbol, so that the data transmission method of the present application does not need to frequently exchange the channel response information of other time-varying channels to obtain the corresponding channel matrix, but directly estimates the corresponding channel response through the pilot signal corresponding to the current channel, so as to quickly encode the first information symbol to be uploaded, without excessive channel overhead, thereby improving the performance of the system, and the generated second information symbol can be quickly uploaded to the server, so that the server can quickly collect the corresponding model training data in the target time frame according to each second information symbol, thereby accelerating the convergence speed of the corresponding model of federated learning. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 is a flowchart of a data transmission method based on federated learning provided by an embodiment of the present application.

[0052] Figure 2 is an air-federated learning framework assisted by orthogonal time-frequency modulation provided by an embodiment of the present application.

[0053] Figure 3 is a data transmission flowchart of an air-federated learning framework provided by an embodiment of the present application.

[0054] Figure 4is a structural schematic diagram of a data transmission device based on federated learning provided by an embodiment of the present application. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0056] As Figure 1 shown, is a flowchart of a data transmission method based on federated learning provided by an embodiment of the present application. The data transmission method based on federated learning comprises the following steps.

[0057] Step S1: predicting a pilot signal according to a preset state equation and a preset observation equation to obtain corresponding channel state estimation information; wherein the pilot signal is a signal sent by a local device to a server participating in federated learning in a target time frame;

[0058] Step S2: generating a channel matrix corresponding to a channel where the device is located and used to represent channel characteristics according to the pilot signal and the channel state estimation information;

[0059] Step S3: generating a corresponding target precoder according to the channel matrix, and encoding a first information symbol based on the target precoder to generate a second information symbol;

[0060] Step S4: transmitting the second information symbol of each device to the server, so that the server collects corresponding model training data in the target time frame according to each second information symbol.

[0061] For step S1, in a preferred embodiment, as Figure 2 shown, the method of the present application can be applied in Internet of Vehicles, wherein each vehicle as a local device uploads a local data set and a network parameter to a central server, and then the local device participating in federated learning performs local training by using the local data, and the training process is as follows:

[0062]

[0063] wherein F n (w ω ) represents an empirical loss function of a network model of n devices in the ωth iteration, D n represents a local data set of the nth device, d represents a sample set, and w n,ω represents a network model of n devices in the ωth iteration, i.e.:

[0064]

[0065] wherein, denotes the gradient of the empirical loss function. In order to facilitate the design, the embodiment of the application uses g n,y [i] = w n,ω [i] denotes the i-th element in the network model w n,ω , which is also the transmission information to be uploaded to the server participating in federated learning.

[0066] The transmission information represents the actual data transmitted in the network, which can be various types of information such as text, image, audio, video, etc. The information symbol in the embodiment of the application is a representation of the transmission information, which can be a binary, hexadecimal or other encoding method corresponding information symbol. The information symbol can be used to represent the actual transmitted data in the network model, which is processed by encoding, modulation, etc. and transmitted through each element of the network to the central server.

[0067] The connection between the information symbol and the transmission information is that the information symbol is a specific representation form of the transmission information, and the transmission information is the actual content represented by the information symbol. In the network model, the information symbol is transmitted through each element of the network, and finally reaches the destination and is decoded into the original transmission information.

[0068] As Figure 3 shown, in the uplink of federated learning, the application adopts OTFS modulation technology for model parameter transmission. In OTFS, the information symbol can be mapped to a two-dimensional time delay Doppler grid with size XxY, where X and Y represent the number of subcarriers and time slots, respectively. Γ represents the length of a time slot, and the subcarrier spacing is represented as Δf = 1 / Γ Hz.

[0069] For the nth device, the application can use G dd,k to represent the first information symbol in the DD grid, where the (x, y)th element is represented as g n,y [yX+x]. First, G dd,k will be converted to a time-frequency domain signal G tf,n through a symplectic inverse Fourier transform:

[0070]

[0071] Then, the time-frequency domain signal G tf,n continues to be converted to a time domain signal G t,n using the Heisenberg transform.

[0072] At the center server side, it receives signals from different devices simultaneously, and converts the corresponding time domain signals back to time-frequency domain signals M tf by Wigner transform. tf Finally, the second information symbol M dd corresponding to the first information symbol received in the DD domain is obtained by performing the symplectic finite Fourier transform on M dd .

[0073]

[0074] For the nth device, the present embodiment sets a wireless uplink channel with R n different propagation paths, and the channel is assumed to be lowly spread and the maximum delay is , then the DD channel response of the nth device is expressed as:

[0075]

[0076] where h r,n is the path gain of each path r, is the delay exponent, is the Doppler shift.

[0077] Without considering noise, the present embodiment places the null symbol in the last κ dd,n rows of S max,n , and the input-output relationship when the rectangular pulse is used as the pulse shaping waveform is expressed as:

[0078]

[0079] where m x = [M dd (x,0), M dd (x,1), …, M dd (x,Y-1)] Γ , denotes the delay set, and N x,κ,n denotes the time-varying Doppler spread matrix corresponding to the channel between the center server and the nth device, and is expressed as:

[0080]

[0081] where φ x,κ,n is expressed as:

[0082]

[0083] Therefore, the input-output relationship of the signal of the present embodiment can be written as:

[0084] ​

[0085] wherein, and respectively represent information symbols transmitted and received in a vectorized form in the DD domain. In addition, B dd,n is represented as:

[0086]

[0087] It can be understood that B dd,n is a channel matrix of the uplink DD channel;

[0088] wherein, the Doppler spread matrix N x,κ,n causes various information symbol couplings to cause interference, and for over-the-air computation, the coupling between DD plane grid points must be eliminated. To this end, the embodiments of the present application can also generate a precoder for each device's information symbol to ensure that the interference in the DD grid point is minimized.

[0089] Illustratively, the precoder is a device or algorithm used to preprocess data, which can transform or encode the data so that it can be correctly decoded after transmission through the channel. The design of the precoder needs to consider the characteristics of the channel, including the amplitude, phase, frequency response, etc. of the channel. Then, the corresponding encoder can be obtained after generating a channel matrix representing the characteristics of the channel.

[0090] For steps S2 and S3, in a preferred embodiment, the present application first predicts the pilot signal to obtain the corresponding channel state estimation information, so that the corresponding encoder can be obtained.

[0091] Specifically, the pilot signal is predicted according to a predetermined state equation and a predetermined observation equation to obtain the corresponding channel state estimation information, including:

[0092] The pilot signal is taken as an input of the Kalman filter;

[0093] The following prediction operation is repeated until the iteration number of the prediction operation is reached, and the final channel state estimation information is generated:

[0094] Based on the state equation and the observation equation corresponding to the current time of the Kalman filter and the observation value of the pilot signal at the current time, the channel state estimation information predicted last time is re-predicted to obtain the current channel state estimation information and the current covariance;

[0095] After the new covariance is obtained by predicting according to the current covariance, the Kalman gain is obtained based on the new covariance;

[0096] Based on the current channel state estimation information and the Kalman gain, the channel state estimation information of this prediction is generated;

[0097] When it is determined that the number of iterations of the prediction operation is not reached, the state equation and the observation equation of the Kalman filter are updated according to the observation value of the pilot signal and the channel state estimation information of this prediction.

[0098] It can be understood that the Kalman filter algorithm uses the observation value at the current time and the prediction value at the previous time to make optimal estimation of the state by means of continuous loop iteration and combination of the Kalman filter formula. It is a recursive filter, and its core idea is to estimate the state of the system through the two steps of prediction and update. In each recursive step (i.e., repeated prediction), the Kalman filter updates the state equation and the observation equation.

[0099] That is, after the observation value of the pilot signal at the current time is given, the channel state estimation information of the last prediction can be re-predicted using the state equation and the observation equation of the Kalman filter. The specific steps are as follows:

[0100] Calculate the Kalman gain: calculate the Kalman gain according to the observation value of the pilot signal at the current time and the channel state estimation information of the last prediction. The Kalman gain is used to balance the uncertainty between the state equation and the observation equation.

[0101] Update the channel state estimation: update the channel state estimation information of the last prediction using the Kalman gain. The updated channel state estimation information will consider the observation value of the pilot signal at the current time.

[0102] Calculate the error covariance: calculate the error covariance according to the updated channel state estimation information. The error covariance reflects the uncertainty of the channel state estimation.

[0103] Recursive update: pass the updated channel state estimation information and the error covariance to the next iteration to continue the recursive update of the channel state estimation.

[0104] Through the above steps, the channel state estimation information of the last prediction can be re-predicted based on the state equation and the observation equation of the Kalman filter and the observation value of the pilot signal at the current time. This re-prediction method can reduce the estimation error and improve the accuracy of the channel state.

[0105] The channel state estimation information is the state information (CSI, Channel State Information), which describes the characteristics of the channel at a specific frequency and time point. It includes amplitude, phase and delay information, and these information can be estimated by comparing the characteristics of the received pilot signal with the original signal.

[0106] In order to reduce the dd,n The control overhead and radio resource waste caused by the estimation and frequent switching of the B embodiment of the present invention dd,n The calculation method of (channel matrix) can be shown as follows:

[0107] First, in each time frame, the device sends the pilot signal embedded in the DD domain to the central server and uses the shifted 2D Kronecker function As the pilot signal for each device, and It can be expressed as:

[0108]

[0109] Among them, x n and y n is a device-specific constant. The pilot signals for these devices are OTFS modulated without precoding.

[0110] On the central server, OTFS demodulation is used to convert the combined received pilot signal from the device back to the DD domain. Considering the actual wireless channel, the embodiment of the present invention assumes that the DD domain channel changes with time, that is, B dd,n It also changes with time.

[0111] Based on the above input-output relationship, the received signal in It's B dd,n The xth n Y+y n Therefore, use the appropriate x n and y n You can directly dd Determine B dd,n The given column of .

[0112] The present invention uses m dd,γ and B dd,n,γ In addition, a small number of subcarriers are usually allocated to one IoT device, and the remaining spectrum can be allocated to other users in the network. Therefore, it is necessary to l,n,γ and η l,n,γ Tracked individually.

[0113] The present invention uses m dd,t As the input of the Kalman filter, for the nth device, the central server does the following calculation:

[0114]

[0115] where mdd,n,γ is the received signal vector corresponding to the pilot of the nth device. Furthermore, the state vector p n,γ of the nth device in the γth time frame is given by:

[0116]

[0117] where h L,r,n,γ ,h Q,r,n,γ denote the real and imaginary parts of the path gain h r,n,γ , respectively, which vary according to a first-order Gauss-Markov model, i.e. and where ξ is the correlation coefficient and is a Gaussian random variable. Then, a model for η l,n,γ controlled by a Gaussian noise process is adopted, whose state evolution can be expressed as:

[0118] p n,γ+1 = Zp n,γ + v n,γ ,

[0119] where and where Furthermore, a model for m n,γ+1 based on the state vector p dd,n,γ can be expressed as:

[0120]

[0121] For each device, the corresponding Kalman filter algorithm should be run in each frame γ∈{0,1,2,…}, whose process can be expressed as:

[0122]

[0123] where denotes the gradient of the measurement function. It is assumed that O n,0 =∑v and p n,0 are estimated by conventional methods. It can be understood that in the time frame γ+1, a new channel matrix can be constructed according to the predicted p n,γ+1 and used for precoding operation on the data of the nth device.

[0124] Further, the device of the present application can obtain the channel parameters of each device in real time recorded by the Kalman filter in real time;

[0125] When detecting that the channel parameters of the device change, the generated channel matrix is updated according to the changed channel parameters, and the updated channel matrix is taken as a final channel matrix; wherein the final channel matrix is used to generate a corresponding target precoder.

[0126] Because the wireless channel between the central server and the mobile device is correlated in time, for the nth device, the path gains h of the multipath components r in different time frames are correlated, the delay indices t are correlated, and the Doppler indices η are correlated, which change slowly due to changes in the propagation environment. r,n r,n r,n Therefore, the application uses a Kalman filter framework to track changes in channel parameters on the central server. Once it is found that the channel parameters of all multipath components and devices change, the central server informs the devices of the corresponding changes through downlink transmission. Then, the devices construct the corresponding channel matrix B with the help of the change in channel parameters to obtain a more accurate channel matrix. dd,n

[0127] By updating the channel matrix, the system can better track the channel state and adjust the transmission strategy accordingly to maintain the stability and reliability of communication and improve the accuracy of data transmission.

[0128] Real-time updating of the channel matrix helps to improve the robustness of the system. Even in the case of drastic changes in channel parameters, the system can quickly adapt and maintain stable communication performance. The updated channel matrix can better adapt to the changing wireless communication environment. This helps the system to maintain stable performance in various complex environments and meet the communication needs in different scenarios.

[0129] For step S3, in a preferred embodiment, after obtaining the channel matrix, a corresponding target precoder can be generated, and the first information symbol is encoded based on the target precoder to generate a second information symbol.

[0130] The application develops a precoder for the information symbol of each device to ensure that the interference in the DD grid point is minimized. Specifically, the input-output relationship of the application can be expressed as follows after combining the proposed precoder:

[0131]

[0132] wherein,

[0133]

[0134] and O x,n is a precoder submatrix, and ρ n ​​​is a relevant scalar that guarantees the device to meet the transmit power constraint. Considering the feasibility of precoder design, the present application selects a diagonal matrix as the precoder submatrix O x,n Illustratively, the present application obtains the corresponding precoder based on the precoding technology of the zero-forcing (ZF) standard and the optimal transmit power allocation standard to reduce errors in the air computing process and ensure accurate transmission of learning parameters.

[0135] Specifically, the ZF precoder of the nth device can be characterized by the precoder submatrix:

[0136]

[0137] wherein, and and

[0138] When the transmitted information symbols are precoded, the received information symbols will not interfere with each other in the DD domain. Therefore, there is wherein w dd is a non-additive white Gaussian noise (AWGN) sample with variance σ 2 .

[0139] It can be understood that the main goal of the central server is to estimate the i-th element dd,n [i] from the received symbols based on g n,y [i]. The estimated value of p[i] can be expressed as:

[0140]

[0141] Therefore, the present application performs OTFS modulation on the model, which is different from OFDM in the frequency domain. OTFS transmits information in the delay-Doppler domain, which can adapt to high-speed mobile environments and protect data from time-varying channels using orthogonal time-frequency modulation (OTFS) technology to ensure the integrity and stability of data transmission, thereby solving the training bottleneck under the traditional framework.

[0142] Meanwhile, the channel state information estimation scheme based on the extended Kalman filter of the present application can reduce the overhead of frequent exchange of channel state information caused by existing precoding technologies. The present application directly estimates the corresponding channel response from the pilot signal corresponding to the current channel, thereby realizing fast encoding of the first information symbol to be uploaded without excessive channel overhead, improving the performance of the system.

[0143] For step S4, in a preferred embodiment, each device in the present application can transmit the pre-encoded second information symbol to the server, so that the server collects the corresponding model training data in the target time frame according to each second information symbol. Since the present application can quickly encode the first information symbol to be uploaded, the present application can quickly upload the generated second information symbol to the server, so that the server can quickly collect the corresponding model training data in the target time frame according to each second information symbol, thereby accelerating the convergence speed of the corresponding model of the federated learning.

[0144] The schematic central server aggregates the models of all devices into a global model after collecting the models, and uses a federated averaging algorithm to aggregate the models:

[0145]

[0146] Then the central server distributes the global model to all users participating in the federated learning. In order to accelerate the aggregation calculation of the central server, the present application uses OTA calculation technology, in which the central server sums up the scaled versions of the received local weight vectors from different devices due to the signal superposition characteristics of the wireless channel. In this way, the empirical loss function converges to the minimum solution.

[0147] In a simulation experiment, the parameters X = 32, Y = 64, ξ = 0.995, σ η = 0.01, the maximum power is 32x64W, the iteration number ranges from 0 to 150, and the data set is a set of 10000 handwritten images containing numbers from 0 to 5. By comparing the model trained based on the OFDM-based federated learning framework, the model trained based on the OTFS-based federated learning framework proposed in the present application has higher test accuracy and faster convergence speed, because the over-the-air calculation accelerates the aggregation behavior of the federated learning, and the OTFS ensures the reliability of information transmission in high-speed environment. Therefore, this phenomenon verifies the effectiveness of the present method.

[0148] As Figure 4 shown, based on the above various embodiments of the federated learning-based data transmission method, the present application correspondingly provides device embodiment;

[0149] An embodiment of the present application provides a federated learning-based data transmission device, which comprises a channel state estimation information generation module, a channel matrix generation module, an encoding module and an information symbol transmission module.

[0150] The channel state estimation information generation module is configured to predict a pilot signal according to a preset state equation and a preset observation equation to obtain corresponding channel state estimation information; wherein the pilot signal is a signal sent by a local device to a server participating in federated learning in a target time frame;

[0151] The channel matrix generation module is configured to generate a channel matrix corresponding to a channel in which the device is located and used to represent channel characteristics according to the pilot signal and the channel state estimation information.

[0152] The encoding module is configured to generate a corresponding target precoder according to the channel matrix, and encode a first information symbol based on the target precoder to generate a second information symbol.

[0153] The information symbol transmission module is configured to transmit the second information symbol of each device to the server, so that the server collects corresponding model training data in the target time frame according to each second information symbol.

[0154] The channel state estimation information generation module is configured to predict a pilot signal according to a preset state equation and a preset observation equation to obtain corresponding channel state estimation information, including:

[0155] The channel state estimation information generation module is configured to take the pilot signal as an input of the Kalman filter.

[0156] The following prediction operation is repeated until the iteration number of the prediction operation is reached to generate the final channel state estimation information:

[0157] Based on the state equation and the observation equation of the Kalman filter corresponding to the current time and the observation value of the pilot signal at the current time, the channel state estimation information predicted last time is re-predicted to obtain the current channel state estimation information and the current covariance.

[0158] After predicting a new covariance based on the current covariance, the Kalman gain is obtained based on the new covariance.

[0159] Based on the current channel state estimation information and the Kalman gain, the channel state estimation information of this prediction is generated.

[0160] When it is determined that the iteration number of the prediction operation has not been reached, the state equation and the observation equation of the Kalman filter are updated according to the observation value of the pilot signal and the channel state estimation information of this prediction.

[0161] The channel matrix generation module is configured to generate a channel matrix corresponding to a channel in which the device is located and used to represent channel characteristics according to the pilot signal and the channel state estimation information, including:

[0162] The channel matrix generation module is configured to calculate a channel impulse response based on the pilot signal and a received signal corresponding to the pilot signal.

[0163] The channel matrix is generated based on the channel impulse response and the amplitude, phase and delay in the channel state estimation information.

[0164] It should be noted that the apparatus embodiments described above are merely illustrative, and the modules described as separate components can or can not be physically separated, and the components shown as modules can or can not be physical modules, and can be located in one place or distributed on multiple network modules. Part or all of the modules can be selected to achieve the purpose of the embodiment according to actual needs. In addition, the connection relationship between the modules in the apparatus embodiments provided by the present application indicates that there is a communication connection between them, which can be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement it without creative labor.

[0165] Those skilled in the art can clearly understand that, for the convenience and brevity of the above-described specific working process of the apparatus, the corresponding process in the foregoing method embodiments can be referred to, and will not be described here.

[0166] On the basis of the above-mentioned various embodiments of the data transmission method based on federated learning, the present application correspondingly provides terminal device embodiments.

[0167] An embodiment of the present application provides a terminal device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements a data transmission method based on federated learning according to any one of the method embodiments of the present application when executing the computer program.

[0168] The terminal device can be a desktop computer, a notebook computer, a palm computer, a cloud server and other computing terminal devices. The terminal device can include, but is not limited to, a processor and a memory.

[0169] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the terminal device, connecting various parts of the entire terminal device using various interfaces and lines.

[0170] The memory can be used to store the computer program, and the processor implements various functions of the terminal device by running or executing the computer program stored in the memory and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function, etc.; the data storage area can store data created based on the use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory and can also include a non-volatile memory, such as a hard disk, internal memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device or other volatile solid-state storage device.

[0171] Based on the above-mentioned various embodiments of the data transmission method based on federated learning, the present invention provides corresponding embodiments of storage medium items.

[0172] An embodiment of the present invention provides a storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute a data transmission method based on federated learning as described in any method embodiment of the present invention.

[0173] The storage medium is a computer readable storage medium, and the computer program is stored in the computer readable storage medium. When the computer program is executed by a processor, steps of each method embodiment described above can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate forms, etc. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer readable medium does not include electrical carrier signals and telecommunication signals.

[0174] The above is the preferred embodiment of the present application. It should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which are also considered within the scope of protection of the present application.

Claims

1. A data transmission method based on federated learning, applicable to each local device participating in federated learning, characterized in that: include: Predicting a pilot signal based on a preset state equation and a preset observation equation to obtain corresponding channel state estimation information; wherein the pilot signal is a signal sent by the local device to the server participating in federated learning in the target time frame; Generating, based on the pilot signal and the channel state estimation information, a channel matrix corresponding to the channel where the device is located and used to characterize channel characteristics; generating a corresponding target precoder according to the channel matrix, and encoding the first information symbol based on the target precoder to generate a second information symbol; Transmitting the second information symbol of each device to the server, so that the server collects corresponding model training data within the target time frame according to each second information symbol; The predicting of the pilot signal according to the preset state equation and the preset observation equation to obtain corresponding channel state estimation information includes: The pilot signal is used as the input of the Kalman filter; Repeat the following prediction operation until the number of prediction operation iterations is reached to generate the final channel state estimation information: Based on the state equation and observation equation corresponding to the Kalman filter at the current moment and the observation value of the pilot signal at the current moment, re-predicting the channel state estimation information predicted last time to obtain current channel state estimation information and current covariance; After predicting the new covariance based on the current covariance, the Kalman gain is obtained based on the new covariance; Generate the predicted channel state estimation information based on the current channel state estimation information and the Kalman gain; When it is determined that the number of iterations of the prediction operation has not been reached, the state equation and the observation equation of the Kalman filter are updated according to the observation value of the pilot signal and the channel state estimation information predicted this time.

2. A data transmission method based on federated learning according to claim 1, characterized in that: Generating, according to the pilot signal and the channel state estimation information, a channel matrix corresponding to the channel where the device is located and used to characterize channel characteristics, includes: Calculating a channel impulse response based on the pilot signal and a received signal corresponding to the pilot signal; The channel matrix is ​​generated according to the channel impulse response and the amplitude, phase and delay in the channel state estimation information.

3. The data transmission method based on federated learning according to claim 2, characterized in that: Before generating the corresponding target precoder, the following steps are included: Obtain channel parameters of each device recorded in real time by the Kalman filter; When a change in the channel parameters of the device is detected, the generated channel matrix is ​​updated according to the changed channel parameters, and the updated channel matrix is ​​used as the final channel matrix; wherein the final channel matrix is ​​used to generate the corresponding target precoder.

4. The data transmission method based on federated learning according to claim 1, characterized in that: Encoding the first information symbol based on the target precoder to generate a second information symbol includes: The first information symbol is encoded according to the following formula to generate the second information symbol: ； ； ； in, is the second information symbol of the nth device, is the channel matrix of the nth device, is the preset scalar value for the device to meet the transmit power constraint, is the target encoder, is the sub-matrix of the target encoder, It is the first information symbol of the nth device.

5. A data transmission device based on federated learning, characterized in that: include: Channel state estimation information generation module, channel matrix generation module, encoding module and information symbol transmission module; The channel state estimation information generation module is configured to predict the pilot signal based on a preset state equation and a preset observation equation to obtain corresponding channel state estimation information; wherein the pilot signal is a signal sent by the local device to the server participating in federated learning in the target time frame; The channel matrix generation module is configured to generate a channel matrix corresponding to the channel where the device is located and used to characterize channel characteristics based on the pilot signal and the channel state estimation information; The encoding module is configured to generate a corresponding target precoder according to the channel matrix, and encode the first information symbol based on the target precoder to generate a second information symbol; The information symbol transmission module is used to transmit the second information symbol of each device to the server, so that the server collects corresponding model training data within the target time frame according to each second information symbol; The channel state estimation information generation module is used to predict the pilot signal according to a preset state equation and a preset observation equation to obtain corresponding channel state estimation information, including: The channel state estimation information generating module is used to use the pilot signal as the input of the Kalman filter; Repeat the following prediction operation until the number of prediction operation iterations is reached to generate the final channel state estimation information: Based on the state equation and observation equation corresponding to the Kalman filter at the current moment and the observation value of the pilot signal at the current moment, re-predicting the channel state estimation information predicted last time to obtain current channel state estimation information and current covariance; After predicting the new covariance based on the current covariance, the Kalman gain is obtained based on the new covariance; Generate the predicted channel state estimation information based on the current channel state estimation information and the Kalman gain; When it is determined that the number of iterations of the prediction operation has not been reached, the state equation and the observation equation of the Kalman filter are updated according to the observation value of the pilot signal and the channel state estimation information predicted this time.

6. The data transmission device based on federated learning according to claim 5, characterized in that: The channel matrix generation module is configured to generate a channel matrix corresponding to the channel where the device is located and used to characterize channel characteristics based on the pilot signal and the channel state estimation information, including: The channel matrix generation module is configured to calculate a channel impulse response based on the pilot signal and a received signal corresponding to the pilot signal; The channel matrix is ​​generated according to the channel impulse response and the amplitude, phase and delay in the channel state estimation information.

7. A terminal device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the method implements a data transmission method based on federated learning as described in any one of claims 1 to 4.

8. A storage medium, characterized in that: The storage medium includes a stored computer program, wherein when the computer program is running, the device where the storage medium is located is controlled to execute the data transmission method based on federated learning according to any one of claims 1 to 4.

Images