Joint beamforming method, system, medium, terminal of heterogeneous intelligent reflecting surface system

Abstract

This invention provides a joint beamforming method, system, medium, and terminal for a heterogeneous intelligent reflector system, comprising the following steps: obtaining the number of users participating in federated training and the number of reflective units contained in each intelligent reflector; initializing a global federated model based on the number of users and the number of reflective units; providing the global federated model to each user in each training cycle, enabling each user to update their local federated model using the global federated model and train their local federated model using a local dataset; aggregating the local federated models of all users to obtain the aggregated global federated model; and predicting the joint beamforming of the heterogeneous intelligent reflector system based on the aggregated global federated model. The joint beamforming method, system, medium, and terminal for the heterogeneous intelligent reflector system of this invention solves the joint active and passive beamforming problem of heterogeneous IRS-assisted downlink systems based on the SFL algorithm, effectively reducing computational complexity and training costs.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a joint beamforming method, system, medium, and terminal for a heterogeneous intelligent reflecting surface (IRS) system. Background Technology

[0002] The goal of next-generation wireless communication systems is to provide high-speed, energy-efficient, highly reliable, and ubiquitous communication. Intelligent reflective surfaces (IRS) have attracted significant attention due to their potential to improve spectral and energy efficiency. IRS consists of numerous low-cost passive reflective elements developed using metamaterials technology, possessing programmable electromagnetic properties. By intelligently adjusting the phase shift of each reflective element, passively reflected signals can be constructively added to the active incident signal to enhance the desired signal power.

[0003] New challenges have emerged in the design and implementation of IRS-assisted communication systems. A key issue is how to jointly optimize the active beamforming of the base station and the passive beamforming of the IRS at both the base station and the IRS. However, the highly non-convex unity modulus constraint in the optimization equations makes it difficult to find the optimal solution. Existing optimization algorithms mainly include semidefinite relaxation (SDR) algorithms, alternating optimization algorithms, manifold optimization algorithms, gradient projection-based algorithms, and branch-and-bound algorithms. However, these algorithms either have high computational complexity or require approximations to simplify the original objective function.

[0004] In recent years, research has focused on incorporating machine learning (ML) techniques into IRS-assisted communication networks to address non-convex problems and reduce computational complexity. Supervised learning, unsupervised learning, deep learning, and reinforcement learning algorithms are widely used. However, in traditional machine learning algorithms, the channel matrix is ​​used to form input features. When the channel distribution changes, the dimensionality of the input data changes accordingly, requiring model retraining. Therefore, for different communication links, training data and the training process are separate, leading to significant dataset overhead and training costs. Training different IRS-assisted systems separately without prior knowledge is often inefficient. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a joint beamforming method, system, medium, and terminal for heterogeneous intelligent reflector systems. Based on the unsupervised hierarchical federated learning (SFL) algorithm, it solves the joint active and passive beamforming problem of heterogeneous IRS-assisted downlink systems, effectively reducing computational complexity and training costs.

[0006] To achieve the above and other related objectives, this invention provides a joint beamforming method for a heterogeneous intelligent reflector system. The heterogeneous intelligent reflector system includes a base station and several intelligent reflectors, each serving one or more users. The joint beamforming method for the heterogeneous intelligent reflector system includes the following steps: obtaining the number of users participating in federated training and the number of reflector units contained in each intelligent reflector; initializing a global federated model based on the number of users and the number of reflector units; providing the global federated model to each user in each training cycle, enabling each user to update their local federated model using the global federated model and train their local federated model using a local dataset; aggregating the local federated models of all users to obtain an aggregated global federated model; and predicting the joint beamforming of the heterogeneous intelligent reflector system based on the aggregated global federated model.

[0007] In one embodiment of the present invention, each user performs multiple rounds of training on the local federated model within a training cycle; the training objective of the local federated model is to maximize the user's speed.

[0008] In one embodiment of the present invention, the global loss function used in the aggregated global federated model is: in This represents the training model stored locally by the k-th user, where the training model is a locally federated model. and local non-federated models It is composed of splicing, where m and n are the number of layers in the local federated model and the local non-federated model, respectively, and F k Let represent the loss function used by the k-th user when training the locally stored training model. This refers to the local dataset.

[0009] In one embodiment of the present invention, aggregating the local federation models of all users includes the following steps:

[0010] Obtain the weights of each user's local federated model;

[0011] A weighted average is applied to the local federated models for all users.

[0012] In one embodiment of the present invention, when aggregating the local federated models of all users, only the parameters of the specified fully connected layer of the local federated model are aggregated.

[0013] In one embodiment of the present invention, the number of intelligent reflective surfaces is greater than one, and different intelligent reflective surfaces contain different numbers of reflective units.

[0014] This invention provides a joint beamforming system for a heterogeneous smart reflector system, wherein the heterogeneous smart reflector system includes a base station and a plurality of smart reflectors, each smart reflector serving one or more users;

[0015] The joint beamforming system of the heterogeneous intelligent reflective surface system includes an acquisition module, an initialization module, a training module, and a prediction module;

[0016] The acquisition module is used to acquire the number of users participating in federated training and the number of reflection units contained in each smart reflective surface;

[0017] The initialization module is used to initialize the global federation model based on the number of users and the number of reflection units;

[0018] The training module is used to provide the global federated model to each user in each training cycle, so that each user can update their local federated model using the global federated model and train the local federated model using the local dataset; and to aggregate the local federated models of all users to obtain the aggregated global federated model.

[0019] The prediction module is used to predict the joint beamforming of the heterogeneous smart reflector system based on the aggregated global federated model.

[0020] The present invention provides a storage medium storing a computer program that, when executed by a processor, implements the aforementioned joint beamforming method for heterogeneous intelligent reflector systems.

[0021] This invention provides a combined beamforming terminal for a heterogeneous intelligent reflector system, characterized in that it includes: a processor and a memory;

[0022] The memory is used to store computer programs;

[0023] The processor is used to execute the computer program stored in the memory, so that the joint beamforming terminal of the heterogeneous smart reflector system performs the above-described joint beamforming method of the heterogeneous smart reflector system.

[0024] The present invention provides a joint beamforming system for a heterogeneous smart reflector system, including a joint beamforming terminal for the heterogeneous smart reflector system and a client corresponding one-to-one with the smart reflectors contained in the heterogeneous smart reflector system.

[0025] The heterogeneous smart reflective surface system includes a base station and several smart reflective surfaces, with each smart reflective surface serving one or more users;

[0026] The client corresponds one-to-one with the user and is used to update the local federated model using the global federated model provided by the joint beamforming terminal of the heterogeneous intelligent reflector system in each training cycle, train the local federated model using the local dataset, and send the trained local federated model to the joint beamforming terminal of the heterogeneous intelligent reflector system.

[0027] As described above, the combined beamforming method, system, medium, and terminal of the heterogeneous intelligent reflective surface system of the present invention have the following beneficial effects:

[0028] (1) Solve the joint problem of active and passive beamforming in heterogeneous IRS-assisted downlink systems based on the SFL algorithm;

[0029] (2) By adopting a specific set of deep neural network structures and performing hierarchical federation, the latency and energy consumption in federation training are effectively reduced, and the problem that the channel data of different IRS auxiliary cascaded channels follow independent and different distributions, and it is impossible to use a unified model for federation averaging for all users is solved.

[0030] (3) Under the condition of satisfying the maximum transmit power constraint of the base station, the received signal power of the user is maximized by jointly optimizing the beamforming of the base station and the phase of the IRS;

[0031] (4) Effectively reduced training costs, reducing the average number of training samples required for each user by more than half;

[0032] (5) Experimental results show that, compared with the classical semidefinite relaxation (SDR) algorithm for convex optimization, the computational complexity is reduced by four orders of magnitude. Reduce to Where K represents the number of users participating in federated training, while ensuring more than 90% performance, surpassing the vast majority of machine learning algorithms; compared with classic centralized machine learning, it can reduce the average amount of training data required per user by 50% while maintaining the same level of complexity and comparable performance. Attached Figure Description

[0033] Figure 1The diagram shows a structural schematic of a heterogeneous smart reflector-assisted downlink wireless communication system in one embodiment of the prior art.

[0034] Figure 2 The flowchart shown is a joint beamforming method for a heterogeneous intelligent reflective surface system according to an embodiment of the present invention.

[0035] Figure 3(a) shows a schematic diagram of the SFL training of deep neural networks (DNN) of the present invention in one embodiment;

[0036] Figure 3(b) shows a schematic diagram of the DNN structure of the k-th user in the SFL algorithm of the present invention in one embodiment;

[0037] Figure 4 The diagram shows a comparison of the spectral efficiency performance of SFL, SDR, centralized machine learning, and random phase configuration algorithms in one embodiment.

[0038] Figure 5 The diagram illustrates the relationship between the sum rate and the average number of training samples per user in one embodiment.

[0039] Figure 6 The diagram shown is a structural schematic of the combined beamforming system of the heterogeneous intelligent reflective surface system of the present invention in one embodiment.

[0040] Figure 7 The diagram shown is a structural schematic of a combined beamforming terminal of the heterogeneous intelligent reflective surface system of the present invention in one embodiment.

[0041] Figure 8 The diagram shown is a structural schematic of the combined beamforming system of the heterogeneous intelligent reflective surface system of the present invention in another embodiment.

[0042] Component designation explanation

[0043] 61 Acquisition Module

[0044] 62 Initialization Module

[0045] 63 Training Module

[0046] 64 Prediction Module

[0047] 71 processor

[0048] 72 Memory

[0049] 81 Combined Beamforming Terminal for Heterogeneous Intelligent Reflector Systems

[0050] 82 Client Detailed Implementation

[0051] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0052] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0053] Federated learning (FL) is a distributed machine learning algorithm that shares experiential knowledge while maintaining local data security. Each user trains a local model, which is then sent to a server for aggregation and updating to obtain a global model. The global model is then distributed back to users to update their local models, and this process is repeated continuously, optimizing both the global and local models until a globally optimal model is established. Therefore, this invention applies unsupervised hierarchical federated learning to a heterogeneous intelligent reflector system. Utilizing the federated mechanism, it extracts and shares knowledge related to the propagation characteristics of different IRS-assisted wireless channels, reducing the machine learning's dependence on large amounts of data samples. This allows the federated training of models for each intelligent reflector to effectively predict their respective optimal phase configurations, thereby achieving joint beamforming optimization with enhanced signal transmission, effectively reducing computational complexity and improving training efficiency.

[0054] like Figure 1 As shown, in one embodiment, in a heterogeneous smart reflector system for assisting downlink wireless communication, it is assumed that there is one base station (BS) configured with M antennas and K smart reflectors (IRS1...IRS). k ...IRS K ) respectively serve K users (U1....U k .....U K ), K > 1. Each smart reflector has a different number of reflective units, where the k-th smart reflector has N k N reflective units i ≠N jThe signals, i, j ∈ 1, ..., K, are connected to the base station via a smart reflector controller. Each user connects to a different smart reflector based on their location, receiving both direct signals from the base station and signals reflected by the smart reflector. Due to significant path loss, it is generally assumed that signals reflected two or more times by the smart reflector are ignored; and that interference between multiple smart reflectors is negligible because they are far apart.

[0055] like Figure 2 As shown, in one embodiment, the joint beamforming method of the heterogeneous smart reflector system of the present invention includes the following steps:

[0056] Step S1: Obtain the number of users participating in federated training and the number of reflective units contained in each smart reflective surface.

[0057] Step S2: Initialize the global federation model based on the number of users and the number of reflection units.

[0058] Specifically, based on the number of users K and the number of reflection units N k K = 1, 2, ..., K, initialize the global federation model. The number of neurons N FL Because N FL There is a trade-off between the choice of N size and the latency, energy consumption, and performance of the proposed algorithm. A larger N size will result in a better performance. FL This allows SFL to better aggregate and share experiential knowledge, but it also introduces significant latency and energy consumption. Here, we take...

[0059] Step S3: In each training cycle, the joint beamforming terminal of the heterogeneous intelligent reflector system provides the global federated model to each user, so that each user can update their local federated model using the global federated model and train the local federated model using the local dataset; the local federated models of all users are aggregated to obtain the aggregated global federated model.

[0060] Specifically, for each training cycle, the user downloads the fully federated model. Locally, the global federation model is adopted. Update the local federation model in and These are all parameter tuples. Simultaneously, users use locally stored channel data as their local dataset and employ this local dataset to train the local federated model. The trained local federated model is stored locally and can predict the phase of different smart reflectors, maximizing the rate for all users. When the rate of each user is maximized, the total number of users and the rate are maximized. The mini-batch stochastic gradient descent (SFL) algorithm is used, employing the Adam (Adaptive Momentestimation) optimizer with an initial learning rate of 0.001. The batch size and local update steps are hyperparameters of the SFL algorithm and need to be configured based on the actual parameters of the users participating in the federated training. The training data from all local datasets is traversed to complete one round of training, updating the user's local model parameters. in It is based on the updated local federation model. and local non-federated models It is composed of spliced ​​components. Preferably, within a training cycle, each user performs multiple rounds of training on the local federated model and updates the local model parameters, thereby reducing the number of communications between the joint beamforming terminal of the heterogeneous intelligent reflector system and the user, and improving SFL training efficiency.

[0061] Once the local federated model training is complete, the user will use their own local federated model. The parameters are uploaded to the joint beamforming terminal of the heterogeneous intelligent reflector system. The joint beamforming terminal of the heterogeneous intelligent reflector system processes all local federated models. The parameters are aggregated using the Federated Averaging Algorithm (FedAvg) to obtain the aggregated global federated model. Here, dk represents the weight of the k-th user. Specifically, first, the weights of each user's local federated model are obtained; then, a weighted average is calculated for the local federated models of all users.

[0062] After one training cycle is completed, the next training cycle needs to be started until the expected convergence is achieved. In one embodiment, the training is terminated when the absolute value of the change in the loss function over 10 training cycles is less than 0.001.

[0063] Assume all channels exhibit quasi-static flat Rayleigh fading, and that this fading can be estimated by a smart reflector controller using existing channel estimation techniques. The channel path loss model is C0(d / d0). -λ Where C0 represents the path loss at the reference distance d0, d represents the distance between the base station and the user, and λBR =2.0, λ RU =2.8, λ BU =3.5, representing the path loss indices from base station to smart reflector, from smart reflector to user, and from base station to user, respectively. Assuming the locations of the base station and smart reflector are fixed, d... BR d represents the distance between the base station and the smart reflector. v d h Let represent the vertical and horizontal distances from the user to the base station, respectively. Therefore, the straight-line distance from the user to the base station is... The straight-line distance from the user to the smart reflective surface is Assume there are three smart reflective surfaces of different sizes, with corresponding numbers of reflective units N1, N2, and N3 = 64, 72, and 80, respectively. v and d h If the data is randomly distributed across [1, 6] and [40, 50], then the maximum transmit power at the base station is: Noise power is σ 2 = -80dBm.

[0064] The signal received by user k is:

[0065]

[0066] in, Let Θ represent the channel matrices from the base station to the smart reflector, from the smart reflector to the user, and from the base station to the user, respectively. k Let Θ be the phase shift matrix of the intelligent reflector. For simplicity of analysis, assume that the reflected signal from the intelligent reflector has no amplitude loss, then Θ k =diag(θ) k ),in And θ k，n ∈(0,2π],n=1,...,N k . This indicates the transmitted beamforming and satisfies... in This represents the maximum transmit power of the base station. k It is the data stream sent by the base station to user k, satisfying This represents additive noise. Therefore, the signal power received by user k is: p k =|(G k Θ k h r，k +h d，k ) H w k | 2 .

[0067] To maximize the received signal power of the user while satisfying the maximum transmit power constraint of the base station, by jointly optimizing the beamforming of the base station and the phase of the smart reflector, the optimization equation can be expressed as:

[0068] P1:

[0069]

[0070]

[0071]

[0072] Where w = [w1, w2, ..., w K When Θ is fixed, according to the Maximum-ratio transmission (MRT) rule, the optimal beamforming at the reflecting end of P1 is:

[0073]

[0074] w k，opt Substituting P1, the optimization equation is transformed into maximizing the joint channel power gain:

[0075] P2:

[0076] st|θ n，k |=1, n=1, ...,N,

[0077] k = 1, ..., K.

[0078] Therefore, the global loss function used in the aggregated global federation model is: in This represents the training model stored locally by the k-th user, where the training model is a locally federated model. and local non-federated models It is composed of splicing, where m and n are the number of layers in the local federated model and the local non-federated model, respectively, and F k Let represent the loss function used by the k-th user when training the locally stored training model. (k = 1, 2, ..., K) represents the local dataset. The number of training data samples stored locally by the k-th user is... Therefore, the total number of training data samples is The training objective of the global federated model is to minimize the global loss function.

[0079] Since data labels for heterogeneous intelligent reflective surface systems are often difficult to obtain, the loss function used by users when training the local federated model is... Therefore, convergence can be achieved without data labels. Here, T represents the batch size of the mini-batch gradient descent algorithm. Represents channel information. Θ k This represents the phase shift matrix output by the k-th local network. Since the local model uses a real-number neural network, it is passed through a Lambda layer. The intelligent reflector phase configuration vector p output by the network k θ converted to complex form k .

[0080] In one embodiment of the present invention, when aggregating the local federated models of all users, only the parameters of the specified fully connected layers of the local federated models are aggregated. Preferably, only the parameters of the FC3 and FC4 fully connected layers of the local federated models are aggregated, thereby ensuring good federated training performance while maintaining low latency and power consumption. As shown in Figures 3(a) and 3(b), when training a DNN model based on the SFL algorithm, it can be seen from the optimization equation P2 that using G and h r The inner product is then connected to h. d As input to the local neural network, θ is meaningful as output. Since the neural network is real-valued, while the channel matrix is ​​complex-valued, the real and imaginary parts of the channel matrix can be considered independent features. This takes into account the dimension of the cascaded channel matrix and the number of reflective elements N of the intelligent reflector. k Directly related, this invention sets the number of neurons in the input layer to 32N. k , where N k To ensure the neural network maintains good learning capabilities even when the dimensionality of the input data changes, the number of neurons in the output layer is naturally set to N. k The three hidden layers each have 16N. FL 8N FL and 4N FL Neurons. These layers are designed to extract some common knowledge inherent in the proposed optimization problem and share it among SFL users.

[0081] Step S4: Predict the joint beamforming of the heterogeneous smart reflector system based on the aggregated global federated model.

[0082] Specifically, after the aggregated global federated model has been trained, channel data of the heterogeneous smart reflector system is input to predict the joint beamforming of the heterogeneous smart reflector system.

[0083] like Figure 4As shown, the spectral efficiency (SE) performance of algorithms based on SFL, SDR, Centralized Learning (CL), and stochastic phase configuration is described in detail. Simulations of the SFL and CL algorithms are performed under three different conditions, with the average training data sample size for each user being: 1)d k =1.5e5; 2)d k =2e5; 3)d k =3.5e5, k = 1, ..., K. As the number of reflection units N increases, the spectral efficiency implemented based on these algorithms improves. This is because the effective gain of the reflection path increases with increasing N. Therefore, users based on hierarchical federated learning can achieve over 90% of the performance of the SDR method and significantly outperform the performance of random phase. Simultaneously, this invention can achieve performance very close to CL while saving more than half of the training data samples. Relying on the federated sharing mechanism, SFL can achieve the same performance as CL with less training data.

[0084] Figure 5 The relationship between the summation rate of the present invention and the average number of training samples per federated user is shown. As the average number of training samples per federated user increases, the summation rate of the SFL and CL algorithms also increases. This is because as d... k With the increase in the number of data samples, clients can use more data samples during the offline training phase. Furthermore, since users have sufficient samples to approximate the gradient of the loss function, the training loss on the validation set remains constant as the number of data samples increases. As shown in the figure, the minimum number of data samples required per client based on the SFL algorithm is much smaller than that of the CL-based method, thus the SFL algorithm significantly reduces the amount of training samples required for machine learning training.

[0085] like Figure 6 As shown, in one embodiment, the combined beamforming system of the heterogeneous smart reflector system of the present invention includes an acquisition module 61, an initialization module 62, a training module 63, and a prediction module 64.

[0086] The acquisition module 61 is used to acquire the number of users participating in federated training and the number of reflection units contained in each smart reflector.

[0087] The initialization module 62 is connected to the acquisition module 61 and is used to initialize the global federation model based on the number of users and the number of reflection units.

[0088] The training module 63 is connected to the initialization module 62 and is used to provide the global federated model to each user in each training cycle, so that each user can update their local federated model using the global federated model and train the local federated model using the local dataset; and to aggregate the local federated models of all users to obtain the aggregated global federated model.

[0089] The prediction module 64 is connected to the training module 63 and is used to predict the joint beamforming of the heterogeneous smart reflector system based on the aggregated global federated model.

[0090] The structure and principle of the acquisition module 61, initialization module 62, training module 63 and prediction module 64 correspond one-to-one with the steps in the joint beamforming method of the heterogeneous intelligent reflector system described above, so they will not be repeated here.

[0091] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls, entirely in hardware, or partially in software calls via processing elements and partially in hardware. For example, module x can be a separate processing element or integrated into a chip within the device. Additionally, module x can be stored as program code in the device's memory, invoked and executed by a processing element. The implementation of other modules is similar. These modules can be fully or partially integrated together or implemented independently. The processing element mentioned here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions. These modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), one or more Field Programmable Gate Arrays (FPGAs), etc. When a module is implemented through processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. These modules can be integrated together to implement a System-on-a-Chip (SOC).

[0092] The storage medium of this invention stores a computer program, which, when executed by a processor, implements the aforementioned joint beamforming method for heterogeneous intelligent reflector systems. Preferably, the storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disk, USB flash drive, memory card, or optical disk.

[0093] like Figure 7 As shown, in one embodiment, the combined beamforming terminal of the heterogeneous smart reflector system of the present invention includes a processor 71 and a memory 72.

[0094] The memory 72 is used to store computer programs.

[0095] The memory 72 includes various media capable of storing program code, such as ROM, RAM, magnetic disk, USB flash drive, memory card, or optical disk.

[0096] The processor 71 is connected to the memory 72 and is used to execute the computer program stored in the memory so that the joint beamforming terminal of the heterogeneous smart reflector system performs the above-described joint beamforming method of the heterogeneous smart reflector system.

[0097] Preferably, the processor 71 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0098] like Figure 8 As shown, in one embodiment, the joint beamforming system of the heterogeneous smart reflector system of the present invention includes the joint beamforming terminal 81 of the heterogeneous smart reflector system and the client 82 corresponding one-to-one with the smart reflectors included in the heterogeneous smart reflector system.

[0099] The heterogeneous smart reflective surface system includes a base station and several smart reflective surfaces, with each smart reflective surface serving one or more users;

[0100] The client 82 corresponds one-to-one with the user and is used to update the local federated model using the global federated model provided by the joint beamforming terminal 81 of the heterogeneous intelligent reflector system in each training cycle, train the local federated model using the local dataset, and send the trained local federated model to the joint beamforming terminal 81 of the heterogeneous intelligent reflector system.

[0101] In summary, the joint beamforming method, system, medium, and terminal of the heterogeneous intelligent reflector system of this invention solves the joint active and passive beamforming problem of heterogeneous IRS-assisted downlink systems based on the SFL algorithm. It employs a specific deep neural network structure and performs hierarchical federation, effectively reducing latency and energy consumption in federated training. It addresses the problem that channel data from different IRS-assisted cascaded channels follow independent and different distributions, making it impossible to use a unified model for federated averaging across all users. Under the condition of satisfying the maximum transmit power constraint of the base station, it maximizes the received signal power of users by jointly optimizing the beamforming of the base station and the phase of the IRS. It effectively reduces training costs, reducing the average training sample size required for each user by more than half. Experimental verification shows that compared with the classical semidefinite relaxation (SDR) algorithm for convex optimization, the computational complexity is reduced by four orders of magnitude. Reduce to Where K represents the number of users participating in federated training, it can guarantee more than 90% performance, surpassing the vast majority of machine learning algorithms; compared with classic centralized machine learning, it can reduce the average amount of training data required per user by 50% while maintaining the same level of complexity and comparable performance. Therefore, this invention effectively overcomes the various shortcomings of the prior art and has high industrial application value.

[0102] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A joint beamforming method for a heterogeneous intelligent reflector system, characterized in that, The heterogeneous smart reflective surface system includes a base station and several smart reflective surfaces, with each smart reflective surface serving one or more users; The joint beamforming method for the heterogeneous smart reflector system includes the following steps: Obtain the number of users participating in federated training and the number of reflective units contained in each smart reflective surface; The global federation model is initialized based on the number of users and the number of reflection units; In each training cycle, the global federated model is provided to each user so that each user can update their local federated model using the global federated model and train their local federated model using their local dataset; the local federated models of all users are aggregated to obtain the aggregated global federated model. The combined beamforming of the heterogeneous smart reflector system is predicted based on the aggregated global federated model.

2. The joint beamforming method for a heterogeneous intelligent reflector system according to claim 1, characterized in that, Within a training cycle, each user performs multiple rounds of training on the local federated model; the training objective of the local federated model is to maximize the user's speed.

3. The joint beamforming method for a heterogeneous intelligent reflector system according to claim 1, characterized in that, The global loss function used in the aggregated global federated model is: in This represents the training model stored locally by the k-th user, where the training model is a locally federated model. and local non-federated models It is composed of splicing, where m and n are the number of layers in the local federated model and the local non-federated model, respectively, and F k Let represent the loss function used by the k-th user when training the locally stored training model. This refers to the local dataset.

4. The joint beamforming method for a heterogeneous intelligent reflector system according to claim 1, characterized in that, Aggregating the local federated models for all users includes the following steps: Obtain the weights of each user's local federated model; A weighted average is applied to the local federated models for all users.

5. The joint beamforming method for a heterogeneous intelligent reflector system according to claim 1, characterized in that, When aggregating the local federated models of all users, only the parameters of the specified fully connected layer of the local federated model are aggregated.

6. The joint beamforming method for a heterogeneous intelligent reflector system according to claim 1, characterized in that, The number of intelligent reflective surfaces is greater than one, and different intelligent reflective surfaces contain different numbers of reflective units.

7. A combined beamforming system for a heterogeneous intelligent reflector system, characterized in that, The heterogeneous smart reflective surface system includes a base station and several smart reflective surfaces, with each smart reflective surface serving one or more users; The joint beamforming system of the heterogeneous intelligent reflective surface system includes an acquisition module, an initialization module, a training module, and a prediction module; The acquisition module is used to acquire the number of users participating in federated training and the number of reflection units contained in each smart reflective surface; The initialization module is used to initialize the global federation model based on the number of users and the number of reflection units; The training module is used to provide the global federated model to each user in each training cycle, so that each user can update their local federated model using the global federated model and train the local federated model using the local dataset; and to aggregate the local federated models of all users to obtain the aggregated global federated model. The prediction module is used to predict the joint beamforming of the heterogeneous smart reflector system based on the aggregated global federated model.

8. A storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the joint beamforming method for the heterogeneous smart reflector system as described in any one of claims 1 to 6.

9. A combined beamforming terminal for a heterogeneous intelligent reflector system, characterized in that, include: Processor and memory; The memory is used to store computer programs; The processor is used to execute the computer program stored in the memory, so that the joint beamforming terminal of the heterogeneous smart reflector system performs the joint beamforming method of the heterogeneous smart reflector system according to any one of claims 1 to 6.

10. A combined beamforming system for a heterogeneous intelligent reflector system, characterized in that, It includes a joint beamforming terminal for the heterogeneous smart reflector system as described in claim 9, and a client that corresponds one-to-one with the smart reflectors contained in the heterogeneous smart reflector system. The heterogeneous smart reflective surface system includes a base station and several smart reflective surfaces, with each smart reflective surface serving one or more users; The client corresponds one-to-one with the user and is used to update the local federated model using the global federated model provided by the joint beamforming terminal of the heterogeneous intelligent reflector system in each training cycle, train the local federated model using the local dataset, and send the trained local federated model to the joint beamforming terminal of the heterogeneous intelligent reflector system.

Images