A RIS-assisted wireless communication system channel estimation method based on attention mechanism

Abstract

The present invention belongs to the field of information and communication technology and relates to a channel estimation method for a RIS-assisted wireless communication system based on an attention mechanism. The method comprises the following steps: S1: constructing a channel model and a signal model for the RIS-assisted wireless communication system; S2: generating an experimental data set through MATLAB simulation; S3: pre-estimating channel state information using the least squares method; S4: constructing an attention mechanism network model based on the system model, inputting the experimental data set into the attention mechanism network model for training until the error loss function of the attention mechanism network model converges to obtain a trained attention mechanism network model; and S5: inputting an actual pilot signal and a corresponding received signal into the trained attention mechanism network model to estimate the actual RIS-assisted wireless communication system channel. The method has a fast convergence speed, and the average mean squared error (MSE) loss after convergence is lower than that of other existing solutions. In other words, the estimation result is more accurate than existing estimation methods.

Description

Technical Field

[0001] The present invention relates to the field of communications, and in particular to a RIS-assisted wireless communication system channel estimation method based on an attention mechanism. Background Art

[0002] Achieving excellent communication performance in RIS-assisted wireless communication systems requires deploying a sufficient number of reflector units. Research has shown that when the number of reflector units is small, RIS-assisted wireless communication does not outperform traditional relay communication. As the number of reflector units increases, the CSI dimension also increases, significantly increasing the pilot overhead required for channel estimation. Channel estimation in RIS-assisted wireless communication systems presents greater challenges than channel estimation in traditional communications. However, the downlink CSI determines whether a RIS-assisted wireless communication system can achieve potential gains. Therefore, designing a high-dimensional downlink CSI channel estimation scheme is a pressing issue in RIS-assisted wireless communication systems.

[0003] Glossary:

[0004] FDD, or frequency division duplex, refers to the use of two separate frequencies (with a certain frequency separation requirement) for the uplink (mobile station to base station) and downlink (base station to mobile station) of wireless communications. This mode operates across symmetrical frequency bands. FDD is suitable for wireless communication systems that provide a single radio frequency channel for each user.

[0005] DenseLayer layer: dense layer.

[0006] conv layer: convolutional layer.

[0007] BN (BatchNormalization, BN) layer: batch normalization layer.

[0008] UE: User Equipment.

[0009] BS: Base station.

[0010] RIS: Smart Reflective Metasurface.

[0011] CSI: Channel State Information.

[0012] RIS-UE link channel: intelligent reflective metasurface-user equipment link channel.

[0013] BS-UE direct link channel: base station-user equipment link channel.

[0014] BS-RIS link channel: base station-intelligent reflective metasurface link channel.

[0015] Transpose layer: a network layer of the pytorch neural network framework. Summary of the Invention

[0016] In order to solve the above technical problems, the present invention discloses a channel estimation method for an auxiliary wireless communication system.

[0017] The technical solution of the present invention is as follows:

[0018] A RIS-assisted wireless communication system channel estimation method based on an attention mechanism comprises the following steps:

[0019] S1. Construct the channel model and signal model of RIS-assisted wireless communication system;

[0020] S2, generate experimental data sets through MATLAB simulation;

[0021] S3. Pre-estimation of channel state information using the least squares method;

[0022] S4. Build an attention mechanism network model and input the experimental data set into the attention mechanism network model for training until the error loss function of the attention mechanism network model converges to obtain a trained attention mechanism network model;

[0023] S5. Input the actual pilot signal and the corresponding received signal into the trained attention mechanism network model to obtain an estimate of the actual RIS-assisted wireless communication system channel.

[0024] As a further improvement, the channel model is constructed as follows:

[0025] In the RIS-assisted wireless communication system in FDD mode, it is assumed that the BS is equipped with M antennas and the RIS is equipped with N reflection units to serve K single-antenna user equipments. The downlink signal y received by the kth user equipment is k Expressed as

[0026]

[0027] in, is the precoded transmit signal at the BS; and They represent the RIS-UE link channel f after matrix transposition. k ; BS-UE direct link channel and BS-RIS link channel; T represents matrix transpose, f k Represents the RIS-UE link channel; z k represents the additive white Gaussian noise of the k-th user equipment; The adjustable phase shift diagonal matrix of RIS is written as

[0028]

[0029] in, is the phase shift of the nth reflection unit of RIS, w∈[0,1] is the amplitude coefficient of RIS; j represents the RIS unit amplitude, e represents the natural constant, θ represents the intermediate variable, θ T represents the transpose of θ; diag() represents a diagonal matrix; and Respectively represent the complex number set of N×M variables, the complex number set of 1×M variables and the complex number set of 1×N variables;

[0030] According to the properties of diagonal matrices, we have The BS-RIS-UE cascade link is represented as

[0031] Channel estimation target H in RIS-assisted wireless communication system k for:

[0032]

[0033] As a further improvement, the channel model is constructed as follows:

[0034] The RIS-UE link channel, BS-UE direct link channel and BS-RIS link channel all adopt the Rice fading model. The BS-RIS link channel is expressed as

[0035]

[0036] Among them, β BR is the Ricean factor of the BS-RIS channel, G LOS and G NLOS denotes the line-of-sight component and the non-line-of-sight component respectively. BR →∞, it represents the line-of-sight channel model; when β BR = 0, it obeys the Rayleigh fading channel, α BR represents the path loss of the BS-RIS channel;

[0037] The RIS-UE link channel, BS-UE direct link channel and BS-RIS link channel are represented as

[0038]

[0039] where γ i ,i∈{1,2,3} is the path loss index, λ0 is the reference distance, α0 is the path loss of the reference distance, represents the distance from the kth UE to the BS, λ BR Indicates the distance between BS and RIS, represents the distance between RIS and the kth UE; α BR and represent the path losses of BS-UE channel, BS-RIS channel and RIS-UE channel respectively.

[0040] As a further improvement, in step S2, the least squares method is used to pre-estimate the channel state information to generate the corresponding received signal:

[0041] In order to estimate the BS-UE direct link channel and the BS-RIS-UE cascade link, it is assumed that T pilot symbols are transmitted for channel estimation and the received signal Y k for:

[0042] Y k =H k S k +Z k (6)

[0043] Where S=[s1,s2,…,s T ] represents the pilot symbol matrix, Z k =[z 1,k ,z 2,k ,…,z T,k ] represents the additive white Gaussian noise matrix; T represents the Tth pilot symbol;

[0044] The result of the least squares estimation is written as:

[0045]

[0046] in, represents the pseudo-inverse of S, Represents the result of least squares estimation, S H represents the conjugate transpose of S.

[0047] As a further improvement, the attention mechanism network model is as follows:

[0048] The input data first passes through a two-dimensional convolution (Conv2d), batch normalization (BatchNorm2d), LeakyReLU activation function, then passes through two DenseLayer network structures, and finally passes through Conv2d and BatchNorm2d to obtain the denoising result;

[0049] The input data in the DenseLayer network structure first passes through a two-dimensional convolution Conv2d, a batch normalization layer BatchNorm2d, a LeakyReLU activation function, and then passes through the spatial attention module. The input data in the DenseLayer network structure and the output data of the spatial attention module are then spliced ​​through the Cat layer. Finally, the spliced ​​output data passes through a two-dimensional convolution Conv2d, a batch normalization layer BatchNorm2d, and a LeakyReLU activation function for output.

[0050] The spatial attention module is divided into two channels. The first channel outputs the input data of the spatial attention mechanism through a linear layer and an activation function ReLU; the second channel first transforms the input data of the spatial attention mechanism through a transpose layer for data dimension transformation, then through a linear layer and an activation function ReLU, and finally through a transpose layer for data dimension transformation and output; the output of the first channel and the output of the second channel are arithmetically added, and then subjected to two-dimensional convolution Conv2d and batch normalization BatchNorm2d to form a two-channel output. Finally, the two-channel output and the input data of the spatial attention module are arithmetically added as the output data of the spatial attention module.

[0051] Further improvement, the attention mechanism network model further improves the channel estimation accuracy by inputting the results of least squares estimation, and the output result value H est Expressed as

[0052]

[0053] Among them, f est (·) represents the attention mechanism network model;

[0054] The error loss function of the attention mechanism network model is as follows:

[0055]

[0056] in, Indicates H est , H k The model parameters when the mean square error of the two is the smallest. Advantages of the present invention:

[0057] The present invention has a fast convergence speed, and the average MSE loss after convergence is lower than the average MSE loss of other existing solutions, that is, compared with the existing estimation method, the estimation result is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a schematic diagram of the RIS-assisted wireless communication system;

[0059] Figure 2 Schematic diagram of the neural network model structure based on the attention mechanism;

[0060] Figure 3 Schematic diagram of the spatial attention mechanism network structure;

[0061] Figure 4 NMSE performance plot of the proposed channel estimation scheme under different signal-to-noise ratio training data;

[0062] Figure 5 The average MSE loss curves of different channel estimation schemes when using [-5,10]dB signal-to-noise ratio training data;

[0063] Figure 6 The NMSE performance of different channel estimation schemes when using [-5,10]dB signal-to-noise ratio training data;

[0064] Figure 7 The NMSE performance of the channel estimation scheme is affected by the number of BS antennas N.

[0065] Figure 8 The NMSE performance comparison curve of the proposed scheme with different RIS reflection unit numbers N. DETAILED DESCRIPTION

[0066] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0067] 1. The system model considers the RIS-assisted wireless communication system in FDD mode, such as Figure 1 As shown in Figure 1, the BS is equipped with M antennas and the RIS is equipped with N reflection units to serve K single-antenna user equipments. The downlink signal received by the kth user equipment is expressed as

[0068]

[0069] in, is the precoded transmit signal at the BS; and Respectively represent the RIS-UE link channel, BS-UE direct link channel and BS-RIS link channel; k represents the additive white Gaussian noise (AWGN) of the k-th user equipment; The adjustable phase shift diagonal matrix representing RIS can be written as

[0070]

[0071] in, is the phase shift of the nth reflector unit of the RIS, w∈[0,1] is the amplitude coefficient of the RIS, and diag represents the diag function, i.e., the diagonal matrix.

[0072] According to the properties of diagonal matrices, we can get The BS-RIS-UE cascade link can be expressed as

[0073] The channel estimation objectives in RIS-assisted wireless communication systems are:

[0074]

[0075] In this invention, all channels adopt the Rice fading model, and the BS-RIS link channel can be expressed as

[0076]

[0077] Among them, β BR is the Ricean factor of the BS-RIS channel, G LOS and G NLOS denotes the line-of-sight component and the non-line-of-sight component respectively. BR →∞, it represents the line-of-sight channel model; when β BR = 0, it obeys the Rayleigh channel model. BR represents the path loss of the BS-RIS channel;

[0078] The path losses of BS-UE channel, BS-RIS channel and RIS-UE channel are expressed as

[0079]

[0080] where γ i ,i∈{1,2,3} is the path loss index, λ0 is the reference distance, α0 is the path loss of the reference distance, represents the distance from the kth UE to the BS, λ BR Indicates the distance between BS and RIS, represents the distance between RIS and the kth UE; α BR and Denote the path loss of BS-UE channel, BS-RIS channel and RIS-UE channel respectively. 2 Channel estimation design based on attention mechanism

[0081] 2.1 Channel Estimation Process

[0082] In order to estimate the direct channel and the cascade channel, assuming that T pilot symbols are transmitted for channel estimation, the received signal can be rewritten as

[0083] Y k =H k S k +Z k (6)

[0084] Where S=[s1,s2,…,s T ] represents the pilot symbol matrix, Z k =[z 1,k ,z 2,k ,…,z T,k ] represents the additive white Gaussian noise matrix.

[0085] If we use the LS channel estimation value as the network input, the channel estimation in the RIS-assisted wireless communication system can be regarded as a denoising problem, and the result of LS estimation can be written as

[0086]

[0087] in, represents the pseudo-inverse of S.

[0088] The network inputs the result of LS estimation, and the designed neural network further improves the channel estimation accuracy, and the output result value H est Expressed as

[0089]

[0090] Among them, f est (·) represents the neural network model of the attention mechanism.

[0091] The network uses MSE as the loss function to train the network model parameters. Combined with the above formula, we can get

[0092]

[0093] in, Indicates H est , H k The model parameters that minimize the mean square error between the two.

[0094] 2.2 Channel Estimation Algorithm

[0095] The input data first passes through a two-dimensional convolution (Conv2d), batch normalization (BatchNorm2d), LeakyReLU activation function, then passes through two DenseLayer network structures, and finally passes through Conv2d and BatchNorm2d to obtain the denoising result;

[0096] The input data in DenseLayer first passes through three layers (Conv2d, BatchNorm2d, LeakyReLU), then passes through the spatial attention module, and then the input data in DenseLayer and the output data of the spatial attention module are spliced ​​through the Cat layer, and finally the spliced ​​output data passes through (Conv2d, BatchNorm2d, LeakyReLU) for output;

[0097] The spatial attention mechanism is divided into two channels. The first channel passes the input data of the spatial attention mechanism through the linear layer (Linear) and the activation function ReLU and outputs it; the second channel first passes the input data of the spatial attention mechanism through the Transpose layer for data dimension transformation, then passes it through the linear layer and the activation function ReLU, and finally passes it through the Transpose layer for data dimension transformation and output; the spatial attention mechanism performs arithmetic addition on the outputs of the two channels, then passes it through Conv2d, BatchNorm2d, and finally performs arithmetic addition on the output and the input data of the spatial attention mechanism and outputs it.

[0098] The Batch Normalization (BN) layer not only solves the problem of inconsistent data distribution during training, but also avoids the problems of vanishing and exploding gradients. The specific implementation is as follows:

[0099] First solve the data The mean and variance of each channel

[0100]

[0101]

[0102] Among them, B, C, W, and H represent the size, number of channels, width, and height of the input data respectively; μ c and represent the mean and variance of the c-th channel respectively.

[0103] The data is then normalized, scaled and translated, and the output of BN can be expressed as

[0104]

[0105] Among them, γ and β are learnable affine variation parameters, and eps defaults to 1e-5 to ensure that the denominator is not 0.

[0106] The network model uses two activation functions, Relu and LeakyReLU. The mathematical expression of Relu is:

[0107] Relu(x)=max(0,x) (13)

[0108] Among them, x represents the input data, and the ReLU activation function has a faster convergence speed and lower computational complexity.

[0109] The mathematical expression of LeakyReLU is

[0110]

[0111] Wherein, a is a fixed parameter, and in the present invention, its value is 0.3.

[0112] The spatial attention network structure is as follows Figure 3 As shown in the figure, the input and output are both (b, 32, 16, 64); the upper channel of the attention structure mainly extracts linear features from the reflection unit of the RIS, and the lower channel mainly extracts linear features from the BS antenna. This operation is used to improve the global receptive field of the network.

[0113] Table 1 Implementation process of channel estimation algorithm based on deep learning

[0114]

[0115]

[0116] The specific implementation process of the channel estimation algorithm based on deep learning is shown in Table 1. It mainly includes the offline training stage and the online testing stage. The offline training stage mainly uses the training data set to train the network parameters and obtain the optimal network parameter model. The online testing stage mainly uses the trained network parameter model to predict the LS of the test data set. Perform denoising and calculate the NMSE performance of the network.

[0117] 2.3 Simulation Results

[0118] The experimental simulation parameters and simulation results are given using NMSE as the downlink CSI estimation metric for the RIS-assisted wireless communication system. In the RIS-assisted wireless communication system, the number of BS antennas is M = 16, the number of RIS reflection units is N = 64, and the number of user devices is K = 6. The channel parameters are set as follows: λ0 = 10m reference distance, α0 = -15dB is the path loss at the reference distance, λ BR=90m, γ1=3.6dB, γ2=2.3dB, γ3=2dB, β BR =10,β BU =0,β RU = 0. The network training parameters are set as follows: the training set, test set, and validation set are 60,000, 10,000, and 10,000 respectively; the learning rate, the size of each training batch, and the number of training cycles are 1e-3, 64, and 100 respectively.

[0119] The signal-to-noise ratio SNR can be expressed as

[0120]

[0121] Where P is power, represents the noise variance.

[0122] The normalized mean square error NMSE expression is:

[0123]

[0124] Among them, H and Represent the real CSI and the estimated CSI respectively, represents the second norm, and E represents the statistical mathematical expectation.

[0125] Since the convergence speed of each neural network algorithm model is inconsistent, the learning rate adopts the cosine annealing strategy, and the learning rate can be expressed as

[0126]

[0127] Among them, lr max and lr min Represent the maximum and minimum values ​​of the learning rate respectively; e∈[E warmup ,E decay ] represents the e-th training cycle, E warmup and E decay Respectively represent the E warmup The learning rate starts to increase in the first training cycle and the decay The learning rate starts to decrease after one training cycle.

[0128] Figure 4This figure demonstrates the NMSE performance of the proposed channel estimation algorithm under training data with different signal-to-noise ratios. The number of RIS reflectors, N, is 64, the number of BS antennas, M, is 16, and the number of pilots, T, is N+1. The training data for the signal-to-noise ratio [-5dB, 10dB] consists of 15,000 data points generated at each of -5dB, 0dB, 5dB, and 10dB, then merged and randomly shuffled. The 5dB training data consists of 60,000 data points sampled at a 5dB signal-to-noise ratio. The figure shows that when the network model is trained using -5dB data, the NMSE performance is relatively good for -10dB and -5dB test data, but as the signal-to-noise ratio increases, the NMSE performance falls far short of the performance of the high-SNR training data. When the network model is trained using 10dB data, the NMSE performance is relatively good for 10dB and 15dB test data, but performance deteriorates at low SNRs. Therefore, this paper combines and shuffles the training data with a signal-to-noise ratio range of -5dB to 10dB to form new training data for model training. Experiments have verified that network performance trained with this combined data is superior to that of network models trained with single signal-to-noise ratio data. Combining data not only improves network performance but also enhances network robustness. Because channel noise is random, selecting a single signal-to-noise ratio as training data is not a good choice in engineering practice. Based on this experimental conclusion, all subsequent experiments used combined data with a signal-to-noise ratio of [-5, 10]dB for network model training.

[0129] In the experiment, the channel estimation scheme proposed in this invention was also compared with other channel estimation network models, including CDRN and DnCNN. When the number of RIS reflection units N = 64, the number of BS antennas M = 16, the number of pilots T = N + 1, and the combined training data interval [-5, 10] dB, Figure 5 The average MSE loss curves of different channel estimation schemes are shown. The proposed scheme converges faster than the other schemes, and the average MSE loss after convergence is lower than the average MSE loss of the other two schemes, which also shows that the channel estimation scheme proposed in this paper is superior to the other two channel estimation schemes.

[0130] Figure 6This figure shows a comparison of the NMSE performance of the proposed channel estimation scheme and existing deep learning channel estimation schemes. The number of RIS reflection units (N) is 64, the number of BS antennas (M) is 16, the number of pilot signals (T) is N+1, and the combined data interval is [-5, 10] dB. The LS line represents the NMSE performance curve at different signal-to-noise ratios after pre-estimation using the least squares method. The figure clearly shows that the deep learning-based channel estimation scheme achieves superior NMSE performance at low signal-to-noise ratios. The proposed channel estimation scheme outperforms the existing deep learning channel estimation schemes at different signal-to-noise ratios.

[0131] Figure 7 The NMSE performance curves of the proposed channel estimation scheme are presented for different BS antenna numbers M. The number of RIS reflection units (RIS) is N = 64, the number of pilot signals (T) is N + 1, and the combined data interval is [-5, 10] dB. LS:M16-N64 represents the NMSE performance curve after LS pre-estimation when the number of BS antennas (M) is 16 and the number of RIS reflection units (N) is 64. The figure shows that as the signal-to-noise ratio (SNR) increases, the impact of the number of BS antennas (M) on the channel estimation network performance decreases. At low SNRs, the performance of the proposed channel estimation network improves as the number of BS antennas (M) increases.

[0132] Figure 8 The NMSE performance curves of the proposed channel estimation scheme are shown for different RIS reflector numbers N. The curves assume that the number of BS antennas M = 16, the number of pilots T = N + 1, and the combined data interval is [-5, 10] dB. As can be seen from the figure, the NMSE performance of the proposed channel estimation network improves with increasing RIS reflector numbers N and pilot numbers T.

[0133] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A RIS-assisted wireless communication system channel estimation method based on attention mechanism, characterized in that: The following steps are involved: S1. Construct the channel model and signal model of RIS-assisted wireless communication system; S2, generate experimental data sets through MATLAB simulation; S3. Pre-estimation of channel state information using the least squares method; S4. Build an attention mechanism network model and input the experimental data set into the attention mechanism network model for training until the error loss function of the attention mechanism network model converges to obtain a trained attention mechanism network model; S5. Input the actual pilot signal and the corresponding received signal into the trained attention mechanism network model to obtain an estimate of the actual RIS-assisted wireless communication system channel; The channel model is constructed as follows: In the RIS-assisted wireless communication system in FDD mode, it is assumed that the BS is equipped with M antennas and the RIS is equipped with N reflection units to serve K single-antenna user equipments. The downlink signal y received by the kth user equipment is k Expressed as in, is the precoded transmit signal at the BS; and They represent the RIS-UE link channel f after matrix transposition. k ; BS-UE direct link channel and BS-RIS link channel; T represents matrix transpose, f k Represents the RIS-UE link channel; z k represents the additive white Gaussian noise of the k-th user equipment; The adjustable phase shift diagonal matrix of RIS is written as in, is the phase shift of the nth reflection unit of RIS, w∈[0,1] is the amplitude coefficient of RIS; j represents the RIS unit amplitude, e represents the natural constant, θ represents the intermediate variable, θ T represents the transpose of θ; diag() represents a diagonal matrix; and Respectively represent the complex number set of N×M variables, the complex number set of 1×M variables and the complex number set of 1×N variables; According to the properties of diagonal matrices, we have The BS-RIS-UE cascade link is represented as Channel estimation target H in RIS-assisted wireless communication system k for: The signal model is constructed as follows: The RIS-UE link channel, BS-UE direct link channel and BS-RIS link channel all adopt the Rice fading model. The BS-RIS link channel is expressed as Among them, β BR is the Ricean factor of the BS-RIS channel, G LOS and G NLOS denotes the line-of-sight component and the non-line-of-sight component respectively; when β BR →∞, it represents the line-of-sight channel model; when β BR = 0, it obeys the Rayleigh fading channel, α BR represents the path loss of the BS-RIS channel; The RIS-UE link channel, BS-UE direct link channel and BS-RIS link channel are represented as where γ i ,i∈{1,2,3} is the path loss index, λ0 is the reference distance, α0 is the path loss of the reference distance, represents the distance from the kth UE to the BS, λ BR Indicates the distance between BS and RIS, represents the distance between RIS and the kth UE; and represent the path losses of BS-UE channel, BS-RIS channel and RIS-UE channel respectively.

2. The RIS-assisted wireless communication system channel estimation method based on the attention mechanism according to claim 1, characterized in that: In step S3, the least square method is used to pre-estimate the channel state information to obtain a pre-estimation result: In order to estimate the BS-UE direct link channel and the BS-RIS-UE cascade link, it is assumed that T pilot symbols are transmitted for channel estimation and the received signal Y k for: Y k =H k S k +Z k (6) Where S=[s1,s2,…,s T ] represents the pilot symbol matrix, Z k =[z 1,k ,z 2,k ,…,z T,k ] represents the additive white Gaussian noise matrix; T represents the Tth pilot symbol; The result of the least squares estimation is written as: in, represents the pseudo-inverse of S, Represents the result of least squares estimation, S H represents the conjugate transpose of S.

3. The RIS-assisted wireless communication system channel estimation method based on the attention mechanism according to claim 1, characterized in that: The attention mechanism network model is as follows: The input data first passes through a two-dimensional convolution (Conv2d), batch normalization (BatchNorm2d), LeakyReLU activation function, then passes through two DenseLayer network structures, and finally passes through Conv2d and BatchNorm2d to obtain the denoising result; The input data in the DenseLayer network structure first passes through a two-dimensional convolution Conv2d, a batch normalization layer BatchNorm2d, a LeakyReLU activation function, and then passes through the spatial attention module. The input data in the DenseLayer network structure and the output data of the spatial attention module are then spliced ​​through the Cat layer. Finally, the spliced ​​output data passes through a two-dimensional convolution Conv2d, a batch normalization layer BatchNorm2d, and a LeakyReLU activation function for output. The spatial attention module is divided into two channels. The first channel outputs the input data of the spatial attention mechanism through a linear layer and an activation function ReLU; the second channel first transforms the input data of the spatial attention mechanism through a transpose layer for data dimension transformation, then through a linear layer and an activation function ReLU, and finally through a transpose layer for data dimension transformation and output; the output of the first channel and the output of the second channel are arithmetically added, and then subjected to two-dimensional convolution Conv2d and batch normalization BatchNorm2d to form a two-channel output. Finally, the two-channel output and the input data of the spatial attention module are arithmetically added as the output data of the spatial attention module.

4. The RIS-assisted wireless communication system channel estimation method based on the attention mechanism according to claim 1, characterized in that: The attention mechanism network model further improves the channel estimation accuracy by inputting the results of the least squares estimation, and the output result value H est Expressed as Among them, f est (·) represents the attention mechanism network model; The error loss function of the attention mechanism network model is as follows: in, Indicates H est , H k The model parameters that minimize the mean square error between the two models.

Images