RIS auxiliary communication system beam forming optimization method based on depth expansion
By combining deep unfolded networks and multilayer perceptrons, the computational complexity and inference latency of the RIS-assisted communication system are reduced, achieving efficient beamforming optimization, meeting real-time communication requirements, and improving generalization capabilities.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-08
- Publication Date
- 2026-03-13
AI Technical Summary
Existing technologies suffer from high computational complexity, prolonged inference time, large training data requirements, and insufficient generalization ability in RIS-assisted communication systems.
A deep unfolded network is used for channel matrix analysis. Auxiliary variables and RIS phase vectors are iteratively updated through cascaded update modules to optimize the beamforming matrix by reducing dimensions. The beamforming matrix is then trained using a multilayer perceptron and gradient descent algorithm to achieve end-to-end optimization.
It significantly reduces computational complexity and inference latency, improves optimization performance, meets real-time communication requirements, and has good generalization ability under complex channel conditions.
Smart Images
Figure CN121664241A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of wireless communication technology, and in particular to a beamforming optimization method for RIS-assisted communication systems based on depth unfolding. Background Technology
[0002] Reconfigurable Intelligent Surfaces (RIS), a crucial technology for future 6G wireless communication systems, reconstruct the propagation path of electromagnetic waves through numerous controllable passive reflective elements. This effectively improves the coverage performance and spectral efficiency of Multiple-Input Single-Output (MISO) systems. In such systems, it is necessary to jointly optimize the precoding matrix on the base station side and the phase control of the RIS to enhance the desired signal and suppress co-channel interference. However, this joint design problem is a high-dimensional non-convex optimization problem, subject to power constraints and phase limitations. Its solution complexity increases significantly with system size, making real-time optimization difficult.
[0003] Existing technologies primarily employ traditional iterative optimization methods, deep learning methods, and meta-learning methods to address this issue. Traditional iterative optimization methods, such as alternating optimization and weighted least mean square error, possess a certain degree of convergence, but each solution requires numerous matrix operations and iterative steps, with computation times for a single channel instance typically reaching the second level, making it difficult to meet the millisecond-level processing latency requirements of future wireless systems. Deep learning methods achieve a direct mapping from channel state information to precoding and phase configuration by constructing neural networks, resulting in fast inference speeds. However, they heavily rely on large amounts of labeled data, and their performance often degrades significantly outside the training distribution. Furthermore, the network output is prone to violating physical constraints, limiting their engineering usability. Meta-learning methods reduce the need for labeled data by learning transferable update rules, but in practical applications, multiple rounds of online gradient updates are still required during the testing phase, resulting in inference latency still in the second range. Moreover, the optimization results are highly sensitive to initialization, leading to insufficient performance stability. Summary of the Invention
[0004] This application provides a beamforming optimization method for RIS-assisted communication systems based on deep deployment, to address the problems of high computational complexity, long inference time, large training data requirements, and insufficient generalization ability of existing methods. The method includes:
[0005] The channel matrix from the base station to the RIS and the channel matrix from the RIS to the user are obtained. The channel matrix from the base station to the RIS and the channel matrix from the RIS to the user are input into the trained deep unfolding network for analysis and processing to obtain the optimized beamforming matrix. The deep unfolding network includes a cascaded update module composed of multiple update units.
[0006] The step of inputting the channel matrix from the base station to the RIS and the channel matrix from the RIS to the user into the trained deep unfolded network for analysis and processing includes:
[0007] Initialize auxiliary variables and the RIS phase vector;
[0008] The initialized auxiliary variables, the initialized RIS phase vector, the channel matrix from the base station to the RIS, and the channel matrix from the RIS to the user are input into the cascaded update module to iteratively update the auxiliary variables and the RIS phase vector, so as to obtain the updated auxiliary variables and the RIS phase vector.
[0009] The concatenated channel matrix is determined based on the updated RIS phase vector, the channel matrix from the base station to the RIS, and the channel matrix from the RIS to the user.
[0010] The dimensionality of the cascaded channel matrix is reduced based on the updated auxiliary variables to obtain the optimized beamforming matrix.
[0011] Optionally, the update unit includes two parallel multilayer perceptrons. The step of inputting the initialized auxiliary variables, the initialized RIS phase vector, the base station-to-RIS channel matrix, and the RIS-to-user channel matrix into the cascaded update module to iteratively update the auxiliary variables and the RIS phase vector includes:
[0012] In each update unit, the concatenated channel matrix is calculated based on the channel matrix from the base station to the RIS, the channel matrix from the RIS to the user, and the RIS phase vector.
[0013] Feature extraction is performed on the concatenated channel matrix to obtain channel features;
[0014] The channel characteristics and auxiliary variables are input into the first multilayer perceptron for update calculation, the update result of the auxiliary variables is obtained, and the power normalization is performed on the update result of the auxiliary variables;
[0015] The channel features and the RIS phase vector are input into the second multilayer perceptron for update calculation to obtain the update result of the RIS phase vector, and the update result of the RIS phase vector is then phase-projected.
[0016] Optionally, the power normalization of the update results of the auxiliary variables includes:
[0017] Calculate the equivalent beamforming matrix based on the cascaded channel matrix and auxiliary variables;
[0018] Calculate the Frobenius norm of the equivalent beamforming matrix, and perform power scaling based on the Frobenius norm and the maximum transmit power to obtain an auxiliary vector that satisfies the power constraint.
[0019] Optionally, the concatenated channel matrix is calculated based on the channel matrix from the base station to the RIS, the channel matrix from the RIS to the user, and the RIS phase vector, satisfying the following formula:
[0020] ;
[0021] in, For the first The concatenated channel matrix output by each update unit. This is the conjugate transpose of the channel matrix from RIS to the user. For the first The RIS phase vector output by each update unit, This is the channel matrix from the base station to the RIS.
[0022] Optionally, the training process of the deep unfolded network includes:
[0023] A sample set for training is constructed, which includes: the channel matrix from RIS to the user, the channel matrix from the base station to RIS, and the user weight vector, and the sample set is divided into a training set and a validation set according to a preset ratio;
[0024] The channel matrix from RIS to user, the channel matrix from base station to RIS and the user weight vector in the training set are input into the depth-unfolded network for forward propagation to obtain the beamforming matrix and RIS phase vector under the current network parameters.
[0025] The parameters of the network are iteratively updated using the gradient descent algorithm based on the preset loss function until the performance on the validation set no longer improves or the preset maximum number of training epochs is reached, resulting in a trained deep unfolded network.
[0026] Optionally, the preset loss function satisfies the following formula:
[0027] ;
[0028] in, Let B be the set of all learnable parameters of the network, and let B be the training batch. The channel matrix from RIS to the user, The channel matrix from the base station to the RIS. For the set of user weight vectors, For the first The weight vector of each user For weighted sum rate, For the network based on the channel The output beamforming matrix, For the network based on the channel The output RIS phase control matrix is constructed based on the RIS phase vector.
[0029] Optionally, the weighted sum rate satisfies the following formula:
[0030] ;
[0031] in, Total number of users For the first The weight of each user For the first Signal-to-interference-to-noise ratio for each user;
[0032] ;
[0033] in, For the first cascaded channel vectors for each user This is the conjugate transpose of the concatenated channel matrix formed by the concatenated channel vectors of all users. For the auxiliary matrix X, the first... List, The first auxiliary matrix X List, For noise variance, Represents the modulo square operation for complex numbers.
[0034] This application provides a beamforming optimization method for a RIS-assisted communication system based on deep unfolding. This method obtains the channel matrices from the base station to the RIS and from the RIS to the user, and inputs these matrices into a trained deep unfolding network for analysis and processing to obtain an optimized beamforming matrix. The deep unfolding network includes a cascaded update module composed of multiple update units. The analysis and processing of the channel matrices from the base station to the RIS and from the RIS to the user involves: initializing auxiliary variables and the RIS phase vector; iteratively updating the auxiliary variables and the RIS phase vector by inputting the initialized auxiliary variables, the initialized RIS phase vector, the channel matrix from the base station to the RIS, and the channel matrix from the RIS to the user into the cascaded update module to obtain updated auxiliary variables and the RIS phase vector; determining the cascaded channel matrix based on the updated RIS phase vector, the channel matrix from the base station to the RIS, and the channel matrix from the RIS to the user; and reducing the dimensionality of the cascaded channel matrix based on the updated auxiliary variables to obtain the optimized beamforming matrix. This method reduces computational complexity and inference latency while ensuring optimization performance. Attached Figure Description
[0035] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0036] Figure 1 A schematic diagram of a RIS-assisted communication system model provided in an embodiment of this application;
[0037] Figure 2 A schematic diagram illustrating the processing flow of the deep unfolded network provided in the embodiments of this application;
[0038] Figure 3 A schematic diagram of the processing flow of the update unit provided in an embodiment of this application;
[0039] Figure 4 This is a schematic diagram illustrating how the sum and rate change with the base station's transmit power, as provided in an embodiment of this application.
[0040] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation
[0041] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0042] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein.
[0043] In this application, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.
[0044] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.
[0045] Figure 1 This is a schematic diagram of a RIS-assisted communication system model provided in an embodiment of this application. Figure 1 As shown, the RIS-assisted communication system includes: a base station equipped with M antennas, a RIS configured with N reflector units, and K single-antenna users.
[0046] The system's channel components include the RIS-to-user channel matrix. Channel matrix from base station to RIS and RIS phase control matrix ,in This represents the phase adjustment amount of the nth RIS unit.
[0047] Define the cascaded channel vector of the k-th user as follows: , Given the conjugate transpose of the channel matrix from RIS to the k-th user, the concatenated channel matrix for all users is: ,matrix The k-th row is the concatenated channel vector of the k-th user. .
[0048] To reduce the dimensionality of the optimization variables, low-dimensional auxiliary variables are introduced. The complete beamforming matrix is recovered through cascaded channel mapping, satisfying the following formula:
[0049]
[0050] in, For beamforming matrix, This indicates the conjugate transpose.
[0051] This strategy will optimize the dimension from Reduce to This reduces the optimization dimension from the original M×K complex variables to This method significantly reduces the size of the optimization space by using multiple complex variables. It leverages the fact that the effective rank of a cascaded channel is typically limited by the number of users K, thus achieving efficient variable representation while maintaining optimization capabilities.
[0052] The base station transmits a signal vector as follows:
[0053] ,
[0054] in, The signal transmitted by the base station to the k-th user is T, where T represents the matrix transpose.
[0055] The transmitted signal vector satisfies the transmitted energy normalization condition:
[0056] ,
[0057] in, It is the expected operation operator. Represents signal vector The square of the Euclidean norm.
[0058] The received signal for the kth user is:
[0059] ,
[0060] When expanded, it appears as follows:
[0061] ,
[0062] in, Let W be the k-th column vector of the base station beamforming matrix. Let J be the j-th column vector of the base station beamforming matrix W. The signal transmitted by the base station to the k-th user. The signal transmitted by the base station to the j-th user. For the signal that user k expects to receive, This represents the interference signal caused by other users to user k. For the k-th user, the additive white Gaussian noise satisfies... , This represents the noise variance.
[0063] Using dimensionality reduction mapping, we have
[0064] ,
[0065] in, is the k-th column of the auxiliary matrix X.
[0066] Based on the above received signal model, the signal-to-interference-plus-noise ratio (SIR) for the k-th user is:
[0067] ,
[0068] System performance is measured by spectral efficiency, which in this multi-user system is specifically reflected as the weighted sum rate, defined as:
[0069] ,
[0070] in, It is the weight of the k-th user. .
[0071] To apply the dimensionality reduction strategy, Substitute the above Expressions, and combinations .because It is determined by the RIS phase adjustment amount Constructed using RIS phase vector For optimization purposes, the signal-to-interference-plus-noise ratio (SINR) can be simplified into a low-dimensional auxiliary optimization variable. Phase vector of RIS Functions:
[0072] ,
[0073] Therefore, the weighted sum rate of the system can ultimately be expressed as a low-dimensional auxiliary optimization variable. Phase vector of RIS function .
[0074] The system optimization objective is to maximize the weighted sum-rate (WSR), and the optimization problem is expressed as:
[0075]
[0076] in, Represents constraints. For low-dimensional auxiliary optimization variables, For the RIS phase vector, Let be the weight vector of the k-th user. For base station beamforming matrix, The power constraint is the conjugate transpose of the cascaded channel matrix. This means that the total transmit signal power of the base station must not exceed the maximum allowable transmit power.
[0077] The beamforming optimization method for RIS-assisted communication systems based on deep unfolding provided in this embodiment includes: obtaining the channel matrix from the base station to the RIS and the channel matrix from the RIS to the user; inputting the channel matrix from the base station to the RIS and the channel matrix from the RIS to the user into the trained deep unfolding network for analysis and processing to obtain the optimized beamforming matrix.
[0078] Figure 2 This is a schematic diagram of the processing flow of the deep unfolded network provided in an embodiment of this application. The deep unfolded network includes a cascaded update module consisting of L update units (L=50 in this embodiment).
[0079] Specifically, the channel matrix from the base station to the RIS and the channel matrix from the RIS to the user are input into the trained deep unfolded network for analysis and processing, including:
[0080] S1: Initialize auxiliary variables and RIS phase vector.
[0081] For example, for auxiliary variables Phase vector of RIS Initialization is performed, where Each element is randomly sampled from a complex Gaussian distribution with a mean of 0 and a small variance; Each element from Random sampling is performed from a uniform distribution over the interval.
[0082] S2: Input the initialized auxiliary variables, the initialized RIS phase vector, the channel matrix from the base station to the RIS, and the channel matrix from the RIS to the user into the cascaded update module to iteratively update the auxiliary variables and the RIS phase vector, and obtain the updated auxiliary variables and the RIS phase vector.
[0083] Each update unit of the cascaded update module includes two parallel multilayer perceptrons, used to simultaneously update auxiliary variables and RIS phase vectors.
[0084] Figure 3 This is a schematic diagram of the processing flow of the update unit provided in the embodiments of this application. Specifically, in each update unit, a concatenated channel matrix is calculated based on the channel matrix from the base station to the RIS, the channel matrix from the RIS to the user, and the RIS phase vector; features are extracted from the concatenated channel matrix to obtain channel features; the channel features and auxiliary variables are input into the first multilayer perceptron for update calculation to obtain the update result of the auxiliary variables, and the update result of the auxiliary variables is power normalized; the channel features and the RIS phase vector are input into the second multilayer perceptron for update calculation to obtain the update result of the RIS phase vector, and the update result of the RIS phase vector is phase projected.
[0085] More specifically, the first step in deep unfolding networks Layer update unit Variables that receive output from the previous layer and channel matrix As input,
[0086] Based on the RIS reflection phase vector The concatenated channel matrix is calculated to satisfy the following formula:
[0087]
[0088] in, For the first The concatenated channel matrix output by each update unit. This is the conjugate transpose of the channel matrix from RIS to the user. For the first The RIS phase vector output by each update unit, This is the channel matrix from the base station to the RIS.
[0089] Furthermore, a unified input feature vector for the multilayer perceptron is constructed. Specifically, this is based on the phase vector of the previous layer. The calculated cascaded channel matrix The real and imaginary parts are separated and expanded, along with the auxiliary variables of the previous level. and phase vector The input vector is concatenated and then fed into two multilayer perceptrons with identical structures but independent parameters. and In this process, the update increments of auxiliary variables and phase vectors are generated.
[0090] A multilayer perceptron consists of several cascaded hidden layers. Each hidden layer sequentially performs linear mapping, normalization, and ReLU activation to extract and transform input features. The feature vector processed by the hidden layers is then transformed into an update increment that matches the dimension of the variable to be updated through linear mapping. and .
[0091] Based on the update increment, the specific update calculation formulas for the auxiliary variable and the RIS phase vector are as follows:
[0092] , ,
[0093] in, , For learnable step size parameters, Multilayer perceptron The update increment of the output auxiliary variable. Multilayer perceptron The update increment of the output RIS phase vector.
[0094] Furthermore, to ensure that the optimization variables satisfy the system constraints, the update results need to be projected as follows: First, based on the relationship... Calculate the equivalent beamforming matrix Calculate its total power. .like Then for Scaling: This ensures that the transmit power constraint is strictly met. Subsequently, a periodic phase projection is applied to the RIS phase vector, mapping it to... The interval, which is related to This is equivalent in the complex exponential sense, thus ensuring that the reflection coefficient satisfies the unit modulus constraint. After the above projection operation, the feasible solution of the final output of this layer is obtained. and .
[0095] S3: Determine the cascaded channel matrix based on the updated RIS phase vector, the channel matrix from the base station to the RIS, and the channel matrix from the RIS to the user.
[0096] Specifically, it satisfies the following formula:
[0097]
[0098] in, The concatenated channel matrix output by the Lth update unit. This is the conjugate transpose of the channel matrix from RIS to the user. The RIS phase vector output by the Lth update unit. This is the channel matrix from the base station to the RIS.
[0099] S4: Reduce the dimensionality of the cascaded channel matrix based on the updated auxiliary variables to obtain the optimized beamforming matrix.
[0100] Specifically, it satisfies the following formula:
[0101]
[0102] in, Let L be the beamforming matrices corresponding to the L update units. This is the auxiliary variable output by the Lth update unit.
[0103] Furthermore, the obtained beamforming matrix is subjected to power normalization to satisfy the maximum transmit power constraint, thereby obtaining the final beamforming matrix of the base station.
[0104] In this embodiment, the training process of the deep unfolded network includes:
[0105] (1) Construct a sample set for training, which includes: the channel matrix from RIS to the user, the channel matrix from the base station to RIS and the user weight vector, and divide the sample set into a training set and a validation set according to a preset ratio.
[0106] Specifically, the training data consists of a channel sample set. Composition, in which For the first Channel matrix from RIS to user for each sample For the first The channel matrix from the base station to the RIS for each sample. For the first Each sample contains a user weight vector. These channel samples can be generated using a standard channel model or obtained through measurements from an actual system. For example, the sample set is divided into 80% training set and 20% validation set.
[0107] Understandably, this invention eliminates the need to pre-generate optimal beamforming and phase configurations for channel samples as supervisory labels, thereby significantly reducing the cost and complexity of training data preparation.
[0108] (2) Input the channel matrix from RIS to user, the channel matrix from base station to RIS and the user weight vector from the training set into the depth-expanded network for forward propagation to obtain the beamforming matrix and RIS phase vector under the current network parameters.
[0109] (3) The parameters of the network are iteratively updated using the gradient descent algorithm according to the preset loss function until the performance on the validation set no longer improves or the preset maximum number of training rounds is reached, and the trained deep unfolded network is obtained.
[0110] Specifically, to achieve joint optimization of the beamforming matrix and the RIS phase, the loss function of the training process is constructed as the negative of the weighted sum rate within the batch.
[0111] The loss function satisfies the following formula:
[0112]
[0113] in, Let B be the set of all learnable parameters of the network, and let B be the training batch. The channel matrix from RIS to the user, The channel matrix from the base station to the RIS. For the set of user weight vectors, For the first The weight vector of each user For weighted sum rate, For the network based on the channel The output beamforming matrix, For the network based on the channel The output RIS phase control matrix is a diagonal matrix constructed based on the RIS phase vector;
[0114] The weighted sum rate satisfies the following formula:
[0115]
[0116] in, Total number of users For the first The weight of each user For the first Signal-to-interference-to-noise ratio for each user;
[0117]
[0118] in, For the first cascaded channel vectors for each user This is the conjugate transpose of the concatenated channel matrix formed by the concatenated channel vectors of all users. The first auxiliary matrix X List, The first auxiliary matrix X List, For noise variance, Represents the modulo square operation for complex numbers.
[0119] For example, this deep unfolded network is trained end-to-end. The Adam optimization algorithm is used during training, and the learning rate is set to... The batch size is 8. The entire training process includes the following steps:
[0120] At the beginning of each training round, channel samples are read from the training dataset in batches, including user-end channel matrices, RIS reflection channel matrices, and user weights. For each batch of samples, they are first input into the deep unwrap network to obtain the beamforming matrix and RIS phase vector under the current network parameters.
[0121] Subsequently, the weighted sum rate corresponding to this batch of samples is calculated, and its negative value is used as the loss value. Based on the gradient information of the loss with respect to the network parameters, the Adam optimizer is used to update all learnable parameters, so that the network gradually approaches the optimal mapping relationship.
[0122] The batch-level training process is iterated repeatedly until the performance on the validation set no longer improves or the maximum number of training rounds is reached, thus completing the entire end-to-end training process.
[0123] End-to-end training can approximate the nonlinear mapping from the system channel state to the jointly optimized beamforming matrix and RIS phase configuration, allowing the inference phase to obtain near-optimal results with only one forward propagation. and This significantly reduces the computational latency of traditional iterative methods.
[0124] The beamforming optimization method for RIS-assisted communication systems based on deep unfolding proposed in this embodiment enables the model to achieve effective convergence with only 80 training samples by structured unfolding and embedding physical constraints such as power and phase in the network, which greatly reduces the dependence on large-scale data. During the inference stage, there is no need to perform multiple iterations of traditional alternating optimization (AO) methods, and the processing latency of a single forward propagation can be stabilized within 60 milliseconds, which can meet the real-time communication requirements of sub-100ms.
[0125] To verify the performance of the method in this embodiment, a simulated communication scenario was constructed, including M=64 base station antennas, K=4 users, and N=100 RIS reflection units. Under the condition of a base station transmit power of 10dBm, the proposed method achieves a weighted sum rate of 10.37bps / Hz on an independent unseen channel verification set. Compared with the comparative algorithm that requires online iterative solutions, the output results of this method are more stable, indicating that the proposed deep unfolding structure has good generalization ability and reliability under complex channel conditions.
[0126] Furthermore, by performing mixed training within the 0-10dBm base station transmit power range, the network can adapt to resource allocation strategies within this typical power range. For example... Figure 4 As shown, within this power range, the system weighted sum rate obtained by the proposed method monotonically increases with increasing transmit power. Specifically, when the transmit power exceeds approximately 6 dBm, its performance consistently outperforms the comparative gradient-based manifold meta-learning (GMML) benchmark method, and the performance advantage expands with increasing power. These results validate that the proposed method possesses greater performance improvement potential in scenarios with ample system power.
[0127] This application provides a beamforming optimization method for RIS-assisted communication systems based on deep unfolding. This method expands the iterative optimization process into a deep neural network with a fixed number of layers, achieving a paradigm shift from online optimization during testing to offline learning. By learning the optimization strategy through offline training, optimization can be completed with only a single forward propagation during online inference, significantly reducing inference time. In the deep unfolded network, low-dimensional auxiliary variables are introduced to reduce the dimensionality of the high-dimensional precoding matrix, and the complete beam matrix is recovered using cascaded channel relationships. Each layer of the network uses two parallel multilayer perceptrons to simultaneously update the low-dimensional variables and the RIS phase matrix. Each layer includes channel feature extraction, low-dimensional variable update, phase update, and constraint projection modules. The feature extraction module expands the real and imaginary parts of the complex channel matrix into real vectors; the low-dimensional variable and phase update modules calculate the update amount through independent multilayer perceptrons and learn adjustable step size parameters; the constraint projection module processes the output through power normalization and phase period constraints. During the training phase, an end-to-end strategy is adopted, using weighted sum rate as the loss function, and gradient descent is used to optimize the network parameters, enabling the network to automatically learn the update weights and step sizes for each layer. The parameters of each network layer can be learned independently to enhance the expressive power of the network.
[0128] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.
[0129] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.
Claims
1. A beamforming optimization method for a RIS-assisted communication system based on depth unfolding, characterized in that, The method includes: The channel matrix from the base station to the RIS and the channel matrix from the RIS to the user are obtained. The channel matrix from the base station to the RIS and the channel matrix from the RIS to the user are input into the trained deep unfolding network for analysis and processing to obtain the optimized beamforming matrix. The deep unfolding network includes a cascaded update module composed of multiple update units. The step of inputting the channel matrix from the base station to the RIS and the channel matrix from the RIS to the user into the trained deep unfolded network for analysis and processing includes: Initialize auxiliary variables and the RIS phase vector; The initialized auxiliary variables, the initialized RIS phase vector, the channel matrix from the base station to the RIS, and the channel matrix from the RIS to the user are input into the cascaded update module to iteratively update the auxiliary variables and the RIS phase vector, so as to obtain the updated auxiliary variables and the RIS phase vector. The concatenated channel matrix is determined based on the updated RIS phase vector, the channel matrix from the base station to the RIS, and the channel matrix from the RIS to the user. The dimensionality of the cascaded channel matrix is reduced based on the updated auxiliary variables to obtain the optimized beamforming matrix.
2. The method according to claim 1, characterized in that, The update unit includes two parallel multilayer perceptrons. The process of inputting initialized auxiliary variables, initialized RIS phase vectors, base station-to-RIS channel matrix, and RIS-to-user channel matrix into a cascaded update module to iteratively update the auxiliary variables and RIS phase vectors includes: In each update unit, the concatenated channel matrix is calculated based on the channel matrix from the base station to the RIS, the channel matrix from the RIS to the user, and the RIS phase vector. Feature extraction is performed on the concatenated channel matrix to obtain channel features; The channel characteristics and auxiliary variables are input into the first multilayer perceptron for update calculation, the update result of the auxiliary variables is obtained, and the power normalization is performed on the update result of the auxiliary variables; The channel features and the RIS phase vector are input into the second multilayer perceptron for update calculation to obtain the update result of the RIS phase vector, and the update result of the RIS phase vector is then phase-projected.
3. The method according to claim 2, characterized in that, The power normalization of the update results of the auxiliary variables includes: Calculate the equivalent beamforming matrix based on the cascaded channel matrix and auxiliary variables; Calculate the Frobenius norm of the equivalent beamforming matrix, and perform power scaling based on the Frobenius norm and the maximum transmit power to obtain an auxiliary vector that satisfies the power constraint.
4. The method according to claim 2, characterized in that, The concatenated channel matrix is calculated based on the channel matrix from the base station to the RIS, the channel matrix from the RIS to the user, and the RIS phase vector, satisfying the following formula: ； in, For the first The concatenated channel matrix output by each update unit. This is the conjugate transpose of the channel matrix from RIS to the user. For the first The RIS phase vector output by each update unit, This is the channel matrix from the base station to the RIS.
5. The method according to claim 1, characterized in that, The training process of the deep unfolded network includes: A sample set for training is constructed, which includes: the channel matrix from RIS to the user, the channel matrix from the base station to RIS, and the user weight vector, and the sample set is divided into a training set and a validation set according to a preset ratio; The channel matrix from RIS to user, the channel matrix from base station to RIS and the user weight vector in the training set are input into the depth-unfolded network for forward propagation to obtain the beamforming matrix and RIS phase vector under the current network parameters. The parameters of the network are iteratively updated using the gradient descent algorithm based on the preset loss function until the performance on the validation set no longer improves or the preset maximum number of training epochs is reached, resulting in a trained deep unfolded network.
6. The method according to claim 5, characterized in that, The preset loss function satisfies the following formula: ； in, Let B be the set of all learnable parameters of the network, and let B be the training batch. The channel matrix from RIS to the user, The channel matrix from the base station to the RIS. For the set of user weight vectors, For the first The weight vector of each user For weighted sum rate, For the network based on the channel The output beamforming matrix, For the network based on the channel The output RIS phase control matrix is constructed based on the RIS phase vector.
7. The method according to claim 6, characterized in that, The weighted sum rate satisfies the following formula: ； in, Total number of users For the first The weight of each user For the first Signal-to-interference-to-noise ratio for each user; ； in, For the first cascaded channel vectors for each user This is the conjugate transpose of the concatenated channel matrix formed by the concatenated channel vectors of all users. For the auxiliary matrix X, the first... List, For the auxiliary matrix X, the first... List, For noise variance, Represents the modulo square operation for complex numbers.