Federal learning network model optimization method
By optimizing the RIS reflection coefficient and signal phase and combining it with air computing technology, the communication bottleneck problem of the federated learning system in large-scale distributed scenarios is solved, efficient model parameter aggregation and rapid convergence are achieved, and the communication efficiency and stability of the system are improved.
Patent Information
- Application Number
- CN202511035352.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-25
- Publication Date
- 2025-10-10
AI Technical Summary
Existing federated learning systems have problems such as high communication overhead, high latency, and low bandwidth utilization in large-scale distributed scenarios. In addition, RIS technology is difficult to adapt to the needs of multi-client collaboration, and the accuracy and efficiency of OAC model parameter aggregation are easily affected by multipath effects and noise interference.
By optimizing the reflection coefficient matrix and signal phase of the intelligent reflecting surface (RIS) and combining it with over-the-air computing (OAC) technology, signal phase alignment and power control are achieved, channel conditions are optimized, a convex optimization algorithm is used to solve the optimal phase shift matrix, the RIS reflection coefficient is dynamically adjusted, phase compensation and power adjustment are performed based on channel state information, and OAC is used for signal superposition and aggregation.
It significantly improves the communication efficiency and model aggregation accuracy of the federated learning system, reduces communication latency, and enhances the scalability and stability of the system. It is particularly suitable for large-scale distributed learning scenarios such as the Internet of Things and edge computing.
Smart Images
Figure CN120768489A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of wireless communications and federated learning technologies, and specifically relates to a method for optimizing the phase of transmitted signals in a federated learning system through reconfigurable intelligent surfaces (RIS) technology and aggregating model parameters through over-the-air computing (OAC) technology. This method takes into account the multipath effect and signal processing technology in wireless communications, aiming to improve the communication efficiency of the federated learning system, reduce latency, and enhance the scalability of the system. The present invention is particularly suitable for large-scale distributed learning scenarios, such as the Internet of Things, edge computing, and mobile computing. Background Art
[0002] Amidst the surging demand for data privacy protection, federated learning, a distributed machine learning framework that supports collaborative model training among multiple clients without sharing raw data, has been widely adopted in fields such as healthcare, finance, and the Internet of Things. In particular, in IoT and edge computing scenarios, wireless networks are a critical bridge for communication between clients and servers. However, traditional communication methods suffer from high communication overhead, high latency, and low bandwidth utilization when transmitting large-scale model parameters. The parameter-by-parameter aggregation method used by existing federated learning systems can easily create communication bottlenecks in large-scale distributed scenarios. Optimization schemes such as model compression and gradient quantization struggle to reduce overhead while ensuring model accuracy and convergence speed.
[0003] Multipath in wireless communications can cause signal interference and fading, severely impacting communication quality. RIS technology dynamically adjusts the phase of reflective elements to optimize signal propagation paths, aligning signals from different paths at the receiving end. This effectively enhances signal strength, suppresses interference, and compensates for channel fading in complex environments (such as urban and indoor environments), significantly improving channel stability.
[0004] Over-the-Air Computing (OAC) technology leverages the overlapping nature of wireless channels, allowing multiple clients to send signals in parallel. This allows for parameter aggregation at the receiving end through signal superposition. Compared to traditional, one-by-one transmission methods, OAC can aggregate parameters from multiple clients within a single time slot, significantly reducing communication latency and fully utilizing limited bandwidth resources. This makes it particularly suitable for large-scale federated learning scenarios.
[0005] However, existing technologies still have shortcomings: RIS technology focuses more on single-link performance optimization and is difficult to adapt to the multi-client collaboration needs of federated learning; OAC is easily affected by multipath effects and noise interference, and the accuracy and efficiency of model parameter aggregation are easily affected.
[0006] To this end, this paper combines multipath effect optimization, RIS technology, and OAC to propose an efficient method for model parameter aggregation in federated learning systems. RIS mitigates multipath interference by adjusting signal phase, while OAC reduces communication overhead. This approach aims to improve system scalability and model aggregation accuracy, providing an innovative solution for large-scale distributed learning scenarios such as the Internet of Things and edge computing. Summary of the Invention
[0007] In response to the problems in the existing technology of how to optimize the RIS phase shift matrix to maximize channel gain, and how to optimize the signal transmission strategy to minimize communication delay under different communication loads, the present invention proposes a federated learning optimization method and system based on intelligent reflecting surface (RIS) and over-the-air computing (OAC).
[0008] In order to solve the above technical problems, the present invention provides the following technical solutions:
[0009] On the one hand, a method for optimizing a federated learning network model is provided, comprising:
[0010] S1: Establish a federated learning network model with smart reflective surfaces, which includes multiple local clients, RIS-assisted wireless channels, and a central server;
[0011] S2: Optimize the reflection coefficient matrix of the smart reflective surface in the federated learning network model;
[0012] S3: Optimize the phase alignment and power control of client signals in the federated learning network model;
[0013] Optionally, the federated learning communication network model in step S1 further includes: a central server for receiving and aggregating model update signals from clients; multiple local clients responsible for local training and generating model update signals; and a RIS for optimizing wireless channels and improving the reliability of signal transmission.
[0014] Optionally, in step S2, optimizing the reflection coefficient matrix of RIS includes:
[0015] S21: Calculate the optimal phase shift matrix of RIS through optimization algorithms to maximize channel gain and improve signal transmission quality. Channel modeling: Establish cascade channel models from client to RIS and RIS to server, taking into account multipath effects and noise interference. Optimization goal: Maximize received signal strength. The objective function is:
[0016] max|H direct +H ris ·θ| 2
[0017] Among them, H direct is a direct channel, Hris is the RIS-assisted cascade channel, and θ is the phase shift matrix of RIS.
[0018] Solve the optimization problem: Use a convex optimization algorithm to solve the optimal phase shift matrix to ensure that the signal strength at the receiving end is maximized.
[0019] Dynamic adjustment: Dynamically adjust the RIS reflection coefficient based on the channel status to enhance the reliability of signal transmission and improve the efficiency and stability of federated learning.
[0020] S22: Dynamically adjust the reflection coefficient of RIS according to the channel status to enhance the reliability of signal transmission and improve the efficiency and stability of federated learning.
[0021] Optionally, in step S3, optimizing the phase alignment and power control of the client signal includes:
[0022] S31: Obtain the channel state information (CSI) of each client through channel estimation, including channel gain and phase offset;
[0023] S32: Perform phase compensation on the signal of each client to eliminate phase offset and ensure that the signals are superimposed in phase at the receiving end;
[0024] S33: Adjust the transmit power of each client based on the channel state information to ensure that all signals have consistent strength at the receiving end;
[0025] S34: Perform weighted superposition on the signals, utilizing the superposition characteristics of the wireless channel to directly aggregate data in the signal domain;
[0026] S35: Recover the aggregate signal through the demodulator, update the global model, and complete one round of federated learning iteration. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. 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 creative work.
[0028] Figure 1 This is a flow chart of the federated learning network model optimization method provided by an embodiment of the present invention;
[0029] Figure 2 This is a flow chart of the federated learning network model optimization method provided by an embodiment of the present invention;
[0030] Figure 3A diagram of the federated learning model framework using Over-the-Air Computing (OAC) technology and optimized Intelligent Super Surface (RIS) provided in an embodiment of the present invention;
[0031] Figure 4 Schematic diagram of the impact of the intelligent hypersurface (RIS) provided in an embodiment of the present invention on the training accuracy and convergence speed of the federated learning model;
[0032] Figure 5 A schematic diagram illustrating the impact of the intelligent hypersurface (RIS) on the training accuracy of the federated learning model under different client scales provided by an embodiment of the present invention;
[0033] Figure 6 A comparison chart of the signal aggregation delay parameters of the traditional data processing solution and the Over-the-Air Computing (OAC) technology solution under the same model training conditions provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0034] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.
[0035] like Figure 1 As shown, an embodiment of the present invention provides a method for optimizing a federated learning network model, including:
[0036] S101: Establish a federated learning network model with smart reflective surfaces, which includes multiple local clients, RIS-assisted wireless channels, and a central server:
[0037] S102: Optimize the reflection coefficient matrix of the smart reflective surface in the federated learning network model.
[0038] S103: Optimize the phase alignment and power control of client signals in the federated learning network model;
[0039] Optionally, the federated learning network model in step S1 further includes: a plurality of intelligent reflective surfaces for reflecting signals sent by different clients; a Rice channel module for signal processing; a signal modulation and demodulation module; and an air computing module.
[0040] Optionally, in step S12, optimizing the reflection coefficient matrix of the smart reflective surface in the federated learning network model includes:
[0041] S121: Setting the direct channel path loss value from the base station to the data center and the path loss setting value through the RIS channel through code simulation;
[0042] S122: constructing an objective function under the condition of the set path loss value;
[0043] S123: Use the convex optimization method to optimize and solve the objective function to complete the optimization of the reflection coefficient matrix of the intelligent reflective surface in the federated learning network model.
[0044] Optionally, in step S122, under the condition of the set path loss value, constructing an objective function includes:
[0045] Construct the objective function as follows:
[0046] |H direct +H ris ·θ| 2 (1)
[0047] Optionally, in step S123, the direct channel is defined as H direct RIS reflection channel is H ris , the phase matrix of RIS is θ. The convex optimization method is used to optimize the objective function and complete the optimization of the reflection coefficient matrix of the intelligent reflective surface in the federated learning network model, including:
[0048] Maximize received signal power:
[0049]
[0050] Constraints:
[0051]
[0052] This constraint ensures that RIS can only change the phase of the signal, not the amplitude.
[0053] Algorithm conversion:
[0054]
[0055] Using the complex modulus square formula:
[0056] |a+b| 2 =|a| 2 +|b| 2 +2·Re(a*b)
[0057] The objective function can be converted to:
[0058]
[0059] Note |h d | 2 is a constant term, But in practice it is much smaller than the cross term So the optimization problem can be simplified to:
[0060]
[0061] Phase optimization solution:
[0062] When the direct channel h d With cascade channel When the phases are aligned, the received power is maximum:
[0063]
[0064] By optimizing the phase of θ to ensure the above conditions hold, signal enhancement can be achieved. The optimized phase matrix θ is used to adjust the reflection characteristics of the RIS unit. By adjusting the phase, the signals are superimposed in phase at the receiving end, enhancing the received signal strength and completing the optimization of the reflection coefficient matrix of the intelligent reflective surface in the federated learning network model.
[0065] Among them, φ is the optimization parameter. By optimizing the vector φ, the matrix parameters in the system are adjusted to maximize the transmission power. d is the direct channel matrix, h r，c represents the transmission channel matrix from client to RIS, h r,s represents the transmission channel matrix from RIS to the server. The superscript H represents the transposed matrix. The core of the arg() function is to find the independent variable that makes the function reach the extreme value (or meet the conditions). Here, it is to find the variable phase matrix θ that meets the conditions.
[0066] Optionally, in step S13, optimizing the phase alignment and power control of the client signal in the federated learning network model includes:
[0067] S131: Obtain channel state information (CSI) of each client through channel estimation, specifically including channel gain and phase offset;
[0068] S132: Based on the channel estimation result, phase compensation is performed on each client signal, and the transmit power is adjusted according to the channel state information.
[0069] Optionally, phase compensation is performed on the transmission signal, including:
[0070] According to formula (2), the signal after phase compensation is obtained:
[0071]
[0072] Where i represents the signal transmitted in the i-th wireless channel, is the signal after phase compensation, s i is a signal without phase compensation, is the phase compensation factor. The client signal has a phase offset due to different channels (φ i =arg(h i)), compensation is required to achieve in-phase superposition, so each client signal is multiplied by the phase compensation factor So that it can perform better aerial calculations.
[0073] Optionally, perform on-the-fly calculations on different transmitted signals, including:
[0074] According to formula (3), the weights of different signals are obtained:
[0075]
[0076] Among them, N i is the data sample size of client i, M is the number of participating clients, is the weight each signal receives, is the total data sample size of all participating aggregation clients. Here, a weighted approach is used to ensure that clients with large data volumes contribute more to the global model, thereby more accurately reflecting the data distribution.
[0077] like Figure 2 As shown, the embodiment of the present invention provides a method for optimizing a federated learning network model, and the specific steps include:
[0078] S201: Establish a federated learning network model with a smart reflective surface, the model including multiple local clients, a RIS-assisted wireless channel, and a central server;
[0079] Among them, the federated learning network model not only includes multiple local clients, RIS-assisted wireless channels, and a central server, but also includes multiple intelligent reflective surfaces for reflecting signals from different clients; Rice channel modules for signal processing; signal modulation and demodulation modules; and air computing modules.
[0080] In a feasible implementation, the present invention establishes a simulation model of intelligent reflective surface (RIS) and over-the-air computing (OAC) assisted federated learning, and seeks the optimal phase shift matrix of RIS and the lowest model parameter aggregation delay strategy through optimization methods.
[0081] S202: Preprocess the local data and train it through a convolutional neural network (CNN) to obtain local model parameters:
[0082] In a feasible implementation, obtaining local model parameters includes:
[0083] In the first convolution layer (conv1), the input data is convolved with the convolution kernel and then nonlinearly transformed using the Relu activation function. The pooling layer (pool1) downsamples the convolution result to reduce the data dimension. The second convolution layer (conv2) repeats the convolution and pooling operations to further extract features. The data then enters the fully connected layer for processing. The first fully connected layer (fc1) flattens the convolution layer output into a one-dimensional vector, performs a linear transformation through a fully connected layer, and then activates the Relu. The second fully connected layer (fc2) outputs the final classification result. The cross-entropy loss function is used to calculate the error between the predicted value and the true label. Backpropagation is performed to calculate the gradient of each layer from the output layer, and an optimizer (such as Adam) is used to update the weights and biases of each layer based on the gradient.
[0084] S203: Quantize and modulate local model parameters:
[0085] In one feasible implementation, quantization maps the floating-point parameter θ to an integer range, as shown in formula (4):
[0086]
[0087] Among them, θ is the model parameter, b is the number of quantization bits (here defined as 8 bits), max(|θ|) is the maximum absolute value of the parameter, and the round function represents the rounding operation on the value x to make it approximate to the nearest integer.
[0088] Optionally, quantize the model parameters. Extract all parameters from the model and flatten them into a one-dimensional array. Calculate a scaling factor based on the number of quantization bits and the maximum absolute value of the parameter. Multiply the floating-point parameter by the scaling factor and round it to an integer. Convert the quantized integer parameters to a bitstream to generate a complex signal through the modulator.
[0089] S204: Optimizing the phase of the transmission signal:
[0090] Optionally, a Rice channel is selected as the wireless channel, where some phase offsets and noise are randomly introduced into the transmission signal to simulate the real propagation path. The intelligent reflecting surface (RIS) dynamically optimizes and compensates the phase of the signal to improve signal transmission. RIS optimizes the phase of each reflecting element by maximizing the received signal strength.
[0091] A feasible implementation method is to optimize the signal phase using RIS, as follows:
[0092] Channel modeling: Establish cascade channel models from client to RIS and from RIS to server, taking into account multipath effects and noise interference.
[0093] Optimization goal: maximize received signal strength
[0094] The objective function is:
[0095] max|H direct +H ris ·θ| 2
[0096] Among them, H direct is a direct channel, H ris is the RIS-assisted cascade channel, and θ is the phase shift matrix of RIS.
[0097] Solve the optimization problem: Use a convex optimization algorithm to solve the optimal phase shift matrix to ensure that the signal strength at the receiving end is maximized.
[0098] Dynamic adjustment: Dynamically adjust the RIS reflection coefficient based on the channel status to enhance the reliability of signal transmission and improve the efficiency and stability of federated learning.
[0099] Maximize received signal power:
[0100]
[0101] Constraints:
[0102]
[0103] This constraint ensures that RIS can only change the phase of the signal, not the amplitude.
[0104] Algorithm conversion:
[0105]
[0106] Using the complex modulus square formula:
[0107] |a+b| 2 =|a| 2 +|b| 2 +2·Re(a*b)
[0108] The objective function can be converted to:
[0109]
[0110] Note |h d | 2 is a constant term, But in practice it is much smaller than the cross term So the optimization problem can be simplified to:
[0111]
[0112] Phase optimization solution:
[0113] When the direct channel h d With cascade channel When the phases are aligned, the received power is maximum:
[0114]
[0115] By optimizing the phase of θ to ensure the above conditions hold, signal enhancement can be achieved. The optimized phase matrix θ is used to adjust the reflection characteristics of the RIS unit. By adjusting the phase, the signals are superimposed in phase at the receiving end, enhancing the received signal strength and completing the optimization of the reflection coefficient matrix of the intelligent reflective surface in the federated learning network model.
[0116] S205: Aggregate the transmission signal using air computing technology:
[0117] In one feasible implementation, the server first uses channel estimation to obtain each client's channel state information (CSI), including channel gain and phase offset. Based on the channel estimation results, phase compensation is performed on each client's signal, and the transmit power is adjusted based on the CSI. Subsequently, the signals are weighted and summed according to the client's weight. Over-the-air computing (OAC) technology is used to directly superimpose the signals at the physical layer of the wireless channel to achieve data aggregation.
[0118] Optionally, phase compensation is performed on the transmission signal, including:
[0119] The phase-compensated signal is obtained according to the following formula:
[0120]
[0121] Where i represents the signal transmitted in the i-th wireless channel, is the signal after phase compensation, s i is a signal without phase compensation, is the phase compensation factor. The client signal has a phase offset due to different channels (φ i =arg(h i )), compensation is required to achieve in-phase superposition, so each client signal is multiplied by the phase compensation factor So that it can perform better aerial calculations.
[0122] Optionally, perform on-the-fly calculations on different transmitted signals, including:
[0123] According to formula (2), the weights of different signals are obtained:
[0124]
[0125] Where i is the data sample size of the client, M is the number of participating clients, is the weight given to each signal, and N is the number of client training samples. The weighted approach here ensures that clients with large amounts of data contribute more to the global model, thereby more accurately reflecting the data distribution.
[0126] Perform weighted aggregation according to formula (3):
[0127]
[0128] Among them, s agg is the aggregate signal, M is the number of participating clients, is the weight each signal receives, The signal after phase compensation. The superimposed signal contains the parameter update information of all clients and is phase-consistent.
[0129] S206: Performing a recovery operation on the transmission signal:
[0130] In a feasible implementation manner, the aggregated signal is restored to a bit stream via a demodulator and converted into model parameters via an inverse quantization operation.
[0131] Optionally, perform an inverse quantization operation on the signal, including:
[0132] Dequantization is performed according to formula (4):
[0133]
[0134] Among them, k is the scaling factor during quantization, θ is the floating-point parameter after recovery, is the aggregate signal s agg After averaging the signal, the server loads the recovered model parameters into the global model, completing the model update. The updated global model is distributed to all clients via wireless channels. After the client loads the global model, it initiates the next round of local training. This process continues until the model converges or the preset number of iterations is reached.
[0135] like Figure 3The figure shows the system model diagram of the present invention. The present invention provides a federated learning optimization method and system based on over-the-air computing (OAC) and intelligent reflective surface (RIS) assistance. The present invention deploys a RIS composed of a large number of low-cost passive reflective elements, and based on preset channel state information, intelligently controls the propagation path, phase and amplitude of the wireless signal, thereby improving the channel conditions, reducing signal fading and interference, and improving the reliability and stability of the communication link; in the federated learning model parameter aggregation stage, OAC technology is used to convert the model parameters of each client into an analog signal at the physical layer of the wireless channel and perform superimposed transmission to achieve automatic parameter aggregation, effectively avoiding the high communication overhead and long delay problems caused by traditional one-by-one polling and serial transmission methods. The RIS and OAC technologies work together to improve the communication efficiency of the federated learning system, significantly accelerate the model convergence speed, improve the final performance of the model, and enhance the stability of the training process, providing a technical solution for the efficient application of federated learning in wireless communication scenarios.
[0136] like Figure 4 The figure shows the model training results. By comparing the model accuracy with and without RIS optimization over the number of training rounds, we can clearly see that RIS significantly improves the convergence speed of the federated learning model. Specifically, in the early stages of training (approximately 40 rounds), the model optimized with RIS has reached a high level of accuracy, while the model without RIS optimization requires more rounds (approximately 50 rounds) to achieve the same accuracy. This shows that RIS effectively accelerates model convergence by optimizing wireless channel conditions. In the later stages of training (200 rounds), the accuracy of the model with RIS optimization stabilizes at a high level of approximately 90%, while the accuracy of the model without RIS optimization is approximately 80%, indicating that RIS not only improves the convergence speed but also the final performance of the model. In addition, the accuracy curve of the model with RIS optimization is smoother and less volatile, reflecting that RIS enhances the stability of model training by improving channel conditions.
[0137] like Figure 5 The following graph shows the model training results. Comparing the accuracy of models with and without RIS optimization under different client numbers clearly demonstrates the performance advantages of RIS in various scale scenarios. In small-scale client scenarios (10-50 clients), the accuracy of models with RIS optimization far exceeds that without RIS optimization, confirming that RIS significantly improves model performance in small-scale scenarios. In medium- to large-scale client scenarios (100-500 clients), the accuracy of models without RIS optimization decreases significantly as the number of clients increases, while the accuracy of models with RIS optimization decreases more gradually, demonstrating that RIS can maintain a high level of performance even in large-scale scenarios.
[0138] like Figure 6The figure shows the model training results. By comparing the aggregate latency of the traditional solution and the Over-the-Air Computing (OAC) solution at different client numbers, we can clearly see the significant advantage of OAC in reducing communication latency. The aggregate latency of the traditional solution increases nonlinearly with the number of clients, with large fluctuations. When the number of clients is small (10-100), the latency is low at 5-20ms. However, when the number of clients increases to 100-1000, the latency quickly soars to over 100ms, indicating that communication overhead and latency increase dramatically in large-scale scenarios. In contrast, the aggregate latency of the OAC solution increases linearly with the number of clients, with minimal fluctuations. When the number of clients is 10-100, the latency is only 1-7ms. Even in large-scale scenarios with 100-1000 clients, the maximum latency is only about 60ms. When the number of clients reaches 1000, the aggregate latency of the OAC solution is only about 60% of that of the traditional solution, significantly improving the system's real-time performance and scalability. This is due to the fact that OAC effectively reduces communication latency through single-slot wireless aggregation and physical layer processing. It has outstanding advantages in large-scale client scenarios and is of great application value in resource-constrained scenarios, providing key support for the practical application of federated learning systems.
Claims
1. A federated learning model optimization method based on intelligent hypersurface (RIS), characterized by: The following steps are involved: The client quantizes the local model parameters into a bit stream and converts the bit stream into a complex signal through a modulator; The client transmits a complex signal to the server via a Ricean wireless channel, and the signal is phase-optimized using a smart reflecting surface (RIS). The server receives signals from multiple clients and aggregates them directly in the signal domain using over-the-air computing technology. The server restores the aggregated signal to a bit stream through a demodulator and dequantizes it into model parameters. The server loads the dequantized model parameters into the global model to complete the update of the global model. The server distributes the updated global model to all clients through wireless channels. After the client loads the global model, it starts the next round of training.
2. The method according to claim 1, characterized in that The specific steps of the client quantizing the local model parameters into a bit stream include: extracting the local model parameters and flattening them into a one-dimensional array; quantizing the one-dimensional array to convert floating-point numbers into integers; and converting the quantized integers into a bit stream.
3. The method according to claim 1, characterized in that The wireless channel is a Ricean channel, whose channel coefficients include direct components and scattered components, and phase optimization is performed through a smart reflecting surface (RIS) to maximize the received signal strength.
4. The method according to claim 1, wherein The over-the-air computing technology includes the following steps: performing channel estimation on signals from multiple clients to obtain the channel coefficient of each client; performing phase alignment on the signals based on the channel coefficient to ensure that the phases of the signals at the receiving end are consistent; performing weighted superposition on the signals based on the weights of the clients; and demodulating the superposed signals into a bit stream and dequantizing them into model parameters.
5. The method according to claim 4, characterized in that The channel estimation includes the following steps: the client sends a known pilot signal, the server receives the pilot signal and estimates the channel coefficient; the server feeds back the estimated channel coefficient to the client, and the client adjusts the signal phase and signal transmission power based on the fed-back channel coefficient; and the intelligent reflecting surface (RIS) is used to optimize the phase to further compensate for channel fading.
6. The method according to claim 1, characterized in that The server distributes the updated global model to all clients through wireless channels. After the client loads the global model, it starts the next round of training until the federated learning model reaches global optimization.
7. A federated learning system based on air computing, characterized in that: include: Multiple clients for local model training, parameter quantization, signal modulation, and signal transmission over wireless channels; The server is used to receive client signals, aggregate signals through over-the-air computing technology, dequantize model parameters, and update the global model; the intelligent reflecting surface (RIS) is used to optimize the phase of the wireless channel to enhance signal transmission effect; the wireless channel is used for signal transmission between the client and the server.
8. The system according to claim 7, characterized in that The client includes: a local model for performing local training tasks; a modem for converting model parameters into complex signals and demodulating received signals into model parameters; and a wireless transmission module for communicating with a server through a wireless channel.
9. The system according to claim 7, wherein: The server includes: a global model for storing and updating global model parameters; an air calculation module for directly aggregating signals from multiple clients in the signal domain; a demodulator for restoring the aggregated signals into model parameters; and a channel estimation module for estimating the channel coefficients of the wireless transmission channel and feeding them back to the client for phase alignment.