Channel Estimation Method for RIS Communication System Based on Improved Residual Shrinkage Network

Abstract

The RIS communication system channel estimation method based on the improved residual shrinkage network belongs to the field of wireless communication technology. The implementation process of the method mainly includes three steps: first, the RIS-assisted wireless communication system generates a data set, and uses the scaled least squares (SLS) algorithm to perform channel pre-estimation to obtain a noisy channel estimation matrix; secondly, build an improved residual shrinkage network, input the noisy channel estimation matrix and the actual channel matrix into the network, and use the supervised learning method for offline training; finally, under different signal-to-noise ratios, input the test data into the trained channel estimation network, and output the estimated channel matrix. The present invention can combine the SLS estimation algorithm with the improved residual shrinkage network when the channel statistical characteristics are unknown, model the channel estimation as a denoising problem, and use the powerful denoising ability of the improved residual shrinkage network to estimate the uplink channel of the RIS-assisted wireless communication system, significantly improving the accuracy of channel estimation.

Description

Technical Field

[0001] The invention relates to a RIS communication system channel estimation method based on an improved residual shrinkage network, belonging to the technical field of wireless channel estimation. Background Art

[0002] Reconfigurable smart surface (RIS) is composed of a large number of passive reflective units. Passive reflective elements change the phase and amplitude of the incident signal, thereby reconfiguring the wireless channel environment to improve the reliability and effectiveness of the system. In the RIS-assisted wireless communication system, the received signal includes the direct channel transmission signal and the RIS reflection signal. The system reliability can be improved by adjusting the RIS phase. RIS is a passive component and cannot process the received signal, so it can only estimate its cascade channel. For the uplink channel estimation problem of this system, traditional pilot-based channel estimation algorithms such as least squares (LS) and linear minimum mean square error (LMMSE) algorithms are generally used. However, traditional channel estimation requires known channel statistical characteristics, and the complexity increases sharply with the increase in the number of reflective units. Deep learning, as an emerging technology in the field of artificial intelligence, has been widely used in various aspects of wireless communication in recent years, such as channel estimation and signal detection. Residual shrinkage network (DRSN) introduces soft threshold processing on the basis of residual network (ResNet), which can significantly reduce the influence of noise. Compared with traditional neural network methods, it solves the problems of network degradation and gradient disappearance well, and shows good application prospects in the field of wireless communication. For a given intermediate feature map, the convolutional attention mechanism module (CBAM) can infer the attention map along two independent dimensions (channel and space) in sequence, multiply the attention map with the input feature map to perform feature optimization, and extract features more accurately.

[0003] Chang Liu et al. (see C. Liu, X. Liu, DW Kwan Ng and J. Yuan, "Deep Residual Network Empowered Channel Estimation for IRS-Assisted Multi-User Communication Systems," ICC 2021-IEEE International Conference on Communications, 2021.) performed channel estimation for RIS-assisted multi-user communication systems based on convolutional neural networks, and proposed a CNN-based denoising network called CDRN, which models the channel estimation problem as a noise elimination problem and uses convolutional neural networks to build a noise reduction module to complete channel estimation. Simulation results show that this method can achieve optimal channel estimation with lower complexity. The above detection scheme uses deep learning to estimate the channel, but the improvement in channel estimation accuracy is still lacking, and the model converges slowly. Summary of the invention

[0004] The traditional LMMSE channel estimation algorithm requires known channel statistics, and the complexity increases sharply with the increase of the number of reflection units. In order to overcome this problem, the present invention proposes an uplink channel estimation method for RIS-assisted wireless communication system based on an improved residual shrinkage network, which can significantly improve the accuracy of channel estimation when the channel statistics are unknown.

[0005] The technical solution of the present invention is as follows:

[0006] A channel estimation method for a RIS communication system based on an improved residual shrinkage network is implemented by a RIS-assisted wireless communication system. The system includes a transmitter, a RIS and a receiver. The transmitter includes K single-antenna users, the RIS includes N passive reflection units, and the receiver includes N r The base station is composed of a base station composed of a root receiving antenna and a channel estimation module based on an improved residual shrinkage network. The links from user to RIS and user to base station experience independent Rayleigh fading, and the link from RIS to base station experiences independent Rice fading. The noise is additive Gaussian white noise. First, the RIS-assisted wireless communication system generates a data set, and uses the SLS algorithm to pre-estimate the channel to obtain a noisy channel estimation matrix. Secondly, an improved residual shrinkage network is built, and the noisy channel estimation matrix and the actual channel matrix are input into the network, and supervised learning methods are used for offline training. Finally, under different signal-to-noise ratios, the test data is input into the trained channel estimation network, and the estimated channel matrix is ​​output to evaluate the channel estimation performance. The specific steps are as follows:

[0007] 1) RIS assists in the simulation of wireless communication system to generate data sets and preprocess the data:

[0008] The RIS-assisted wireless communication system adopts time division multiplexing full-duplex mode and uses the uplink to estimate the channel. The channel matrix from RIS to the base station is: Its elements follow the Rice distribution, N r ×N-dimensional complex vector set, the channel matrices from user to RIS and from user to base station are and Each element is an independent complex Gaussian random variable, obeying a complex Gaussian distribution with a mean of 0 and a variance of 1. Each reflection unit of the RIS can independently adjust the phase of the incident signal. The reflection phase matrix can be expressed as in is the reflection phase at the nth reflection unit, diag represents the construction of a diagonal matrix, [] T represents matrix transpose. RIS is a passive reflective element and cannot process signals. It is impossible to estimate the channel matrices from user to RIS and from RIS to base station separately. When estimating the cascade channel G = Bdiag(h) and direct channel d from user to base station in a RIS-assisted multi-user communication system, it is necessary to perform channel estimation based on pilot signals. The channel matrix to be estimated is expressed as H = [d, G]. The pilot signal contains channel estimation pilot signals of K users. The pilot sequences are orthogonal to each other. The pilot sequence sent by the kth user is x k =[u k,1 ,...,u k,l ...,u k,L ], L represents the pilot sequence length, u k,l represents the lth pilot symbol of the kth user. Each channel estimation data frame contains D different subframes. Let D = N + 1. The dth phase shift matrix can be expressed as in Represents the reflection phase on the nth RIS reflection unit in the dth phase shift matrix. The signal of the dth subframe received by the base station is Among them, H k represents the channel matrix of the kth user, p d =[1,r d ] T is the phase shift matrix, is additive Gaussian white noise, whose element v d,l represents the lth noise vector of the dth subframe. Considering the orthogonality between user pilots, after the kth user sends D subframes, the base station receives the signal: in represents D reflection phase shift matrices, represents the noise matrix, P can be represented by a discrete Fourier matrix, and the received signal of the kth user can be expressed as where w = e j2π / D The receiving end first performs SLS channel pre-estimation on the received signal. The SLS algorithm further considers the impact of noise on channel estimation based on the least squares algorithm, namely the LS algorithm. Its expression is: where σ 2 is the noise power, is the autocorrelation matrix of the channel matrix, tr{} represents the trace of the matrix, represents the matrix pseudo-inverse; P H It means to find the conjugate transpose of P. represents the channel matrix estimated by the LS algorithm;

[0009] The data set is generated by the RIS-assisted wireless communication system and its size is 4.8×10 5 , 75% of the data is used for network training and 25% of the data is used for testing;

[0010] 2) Build a network model based on the residual shrinkage network, i.e., DRSN, input the noisy channel matrix and the actual channel matrix into the network, and use the supervised learning method for offline training:

[0011] Soft thresholding and attention mechanism are introduced into the ResNet network structure, and a network model based on DRSN is built. The network model includes a residual shrinkage module and a convolutional attention mechanism module, namely the CBAM module. The residual shrinkage module consists of three convolutional layers, two batch normalization layers, namely BN, two ReLu activation functions and a soft threshold module. The CBAM module is embedded in the residual shrinkage module, including a channel attention module, namely the CAM module, and a spatial attention module, namely the SAM module. The CAM module includes an average pooling layer, a maximum pooling layer, a two-layer neural network, a ReLU activation function and a Sigmoid activation function. The SAM module includes a maximum pooling layer, an average pooling layer, a convolutional layer and a Sigmoid activation function. The specific training steps are as follows:

[0012] ① Let x represent the input data feature. After x passes through the first convolution layer, BN layer and ReLU activation function of the residual shrinkage module, the output feature map F is obtained. 1 ∈R C×H×W , where C represents the length of the feature image, H represents the height of the feature image, and W represents the number of channels of the feature image;

[0013] ②F 1 Then, after the CBAM module embedded in the residual shrinkage module, the input feature map F 1, the CBAM module infers the channel attention map and the spatial attention map along two independent dimensions, namely the channel and spatial dimensions, and then compares the channel attention map and the spatial attention map with F 1 The CAM module compresses the input feature map in the spatial dimension and performs one-dimensional convolution. Through the parallel average pooling layer and maximum pooling layer, the feature points in the neighborhood are averaged and the maximum feature point is taken to keep the channel dimension unchanged and the spatial dimension compressed. The feature image dimension is changed from C×H×W to C×1×1. Then, the number of channels is compressed to the original 1 / r through a two-layer neural network, where r is the attenuation rate of MLP, and then expanded to the original number of channels. Finally, the results of the ReLU activation function are added, and then the output result is obtained through the Sigmoid activation function and combined with F 1 Multiplying back to the size of C×H×W, the feature map of channel attention is expressed as in and F c 1max They represent the input feature F in the CAM module. 1 The feature vector after average pooling and maximum pooling, W 1 and W 2 They represent the weights of the first and second layers of the two-layer neural network respectively, and σ() represents the Sigmoid activation function;

[0014] The SAM module is connected in series after the CAM module, and the output result F of the CAM module is 1 ′ Two 1×H×W-dimensional feature maps are obtained through the maximum pooling layer and the average pooling layer. The two feature maps are concatenated and converted into a feature map of one channel through a 7*7 convolution layer. Finally, a Sigmoid activation function is used to obtain the feature map of SAM, and the output result is compared with F 1 ′ is multiplied back to C×H×W dimensions, and the feature map of spatial attention is expressed as in and In the SAM module, the input feature F 1 ′ is the feature vector after average pooling and maximum pooling, f 7×7 () represents a 7*7 convolution operation;

[0015] The characteristic graph of the CAM module is M c (F 1 )∈R C×1×1 , and compare it with F 1 Multiply to get the output result of the CAM module F 1 ′ passes through the SAM module and is combined with its feature map M s (F 1 ′) and multiply to get the output result of CBAM module in represents the Kronecker product operation;

[0016] ③ Repeat steps ① and ② to get the output result F of the second CBAM module 2 ″, and then the third convolutional layer of the residual contraction module obtains the output feature F with the same dimension as x out Finally, the soft threshold module is used to process the noise, and its formula is: Where x represents the input feature, y represents the output feature, and τ represents the positive threshold. After soft thresholding, the feature whose absolute value of x is less than τ is set to 0, thereby achieving the effect of reducing data noise. The output result is expressed as F f ;

[0017] ④The input feature x is mapped to the denoised F by identity mapping f Add, and F f If the dimension of y is the same as that of x, then y can be defined as y = σ(F f )+x;

[0018] 3) Generate test data at different signal-to-noise ratios and input them into the trained channel estimation network. Output the estimated channel matrix and further compare it with the actual channel matrix to evaluate the network estimation performance:

[0019] After the channel estimation network training is completed, online deployment is achieved, and the RIS-assisted wireless communication system generates 10 5 For each test data, the SLS channel estimation algorithm is used to estimate the channel. Then the noise is eliminated through the channel estimation network, and finally a reliable channel estimation result is output. And compare it with the actual channel matrix H to evaluate the channel estimation performance.

[0020] The DRSN is the abbreviation of Deep Residual Shrinkage Networks, which means deep residual shrinkage network.

[0021] The SLS is the abbreviation of Scaled Least Squares, which means scaled least squares.

[0022] The present invention proposes an uplink channel estimation method for a reconfigurable intelligent surface (RIS) assisted wireless communication system based on an improved residual shrinkage network. When the statistical characteristics of the channel are unknown, the SLS estimation algorithm is combined with the improved residual shrinkage network to estimate the uplink channel of the RIS assisted wireless communication system, thereby significantly improving the accuracy of channel estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Schematic diagram of the system model of the method of the present invention.

[0024] Figure 2 It is a schematic diagram of the channel estimation network structure of the method of the present invention.

[0025] Figure 3 The transmitting antenna N t =1, receiving antenna N r =8, and N =32 reflection units, the normalized mean square error performance simulation comparison diagram of the method of the present invention and traditional channel estimation methods such as least squares (LS) and scaled least squares (SLS). Figure 3 It can be seen that the normalized mean square error of this method is much smaller than that of the LS and SLS algorithms. DETAILED DESCRIPTION

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments, but is not limited thereto.

[0027] Example:

[0028] A channel estimation method for a RIS communication system based on an improved residual shrinkage network is implemented by a RIS-assisted wireless communication system. The system includes a transmitter, a RIS and a receiver. The transmitter includes K single-antenna users, the RIS includes N passive reflection units, and the receiver includes N r The base station is composed of a base station composed of a root receiving antenna and a channel estimation module based on an improved residual shrinkage network. The links from user to RIS and user to base station experience independent Rayleigh fading, and the link from RIS to base station experiences independent Rice fading. The noise is additive Gaussian white noise. First, the RIS-assisted wireless communication system generates a data set, and uses the SLS algorithm to pre-estimate the channel to obtain a noisy channel estimation matrix. Secondly, an improved residual shrinkage network is built, and the noisy channel estimation matrix and the actual channel matrix are input into the network, and supervised learning methods are used for offline training. Finally, under different signal-to-noise ratios, the test data is input into the trained channel estimation network, and the estimated channel matrix is ​​output to evaluate the channel estimation performance. The specific steps are as follows:

[0029] 1) RIS assists in the simulation of wireless communication system to generate data sets and preprocess the data:

[0030] The RIS-assisted wireless communication system adopts time division multiplexing full-duplex mode and uses the uplink to estimate the channel. The channel matrix from RIS to the base station is: Its elements follow the Rice distribution, N r ×N-dimensional complex vector set, the channel matrices from user to RIS and from user to base station are and Each element is an independent complex Gaussian random variable, obeying a complex Gaussian distribution with a mean of 0 and a variance of 1. Each reflection unit of the RIS can independently adjust the phase of the incident signal. The reflection phase matrix can be expressed as in is the reflection phase at the nth reflection unit, diag represents the construction of a diagonal matrix, [] T represents matrix transpose. RIS is a passive reflective element and cannot process signals. It is impossible to estimate the channel matrices from user to RIS and RIS to base station separately. When estimating the cascade channel G = Bdiag(h) and direct channel d from user to base station in a RIS-assisted multi-user communication system, it is necessary to perform channel estimation based on pilot signals. The channel matrix to be estimated is expressed as H = [d, G]. The pilot signal contains channel estimation pilot signals of K users. The pilot sequences are orthogonal to each other. The pilot sequence sent by the kth user is x k =[u k,1 ,...,u k,l ...,u k,L ], L represents the pilot sequence length, u k,l represents the lth pilot symbol of the kth user. Each channel estimation data frame contains D different subframes. Let D = N + 1. The dth phase shift matrix can be expressed as in Represents the reflection phase on the nth RIS reflection unit in the dth phase shift matrix. The signal of the dth subframe received by the base station is Among them, H k represents the channel matrix of the kth user, p d =[1,r d ] T is the phase shift matrix, is additive Gaussian white noise, whose element v d,l represents the lth noise vector of the dth subframe. Considering the orthogonality between user pilots, after the kth user sends D subframes, the base station receives the signal: in represents D reflection phase shift matrices, represents the noise matrix, P can be represented by a discrete Fourier matrix, and the received signal of the kth user can be expressed as where w = e j2π / D The receiving end first performs SLS channel pre-estimation on the received signal. The SLS algorithm further considers the impact of noise on channel estimation based on the least squares algorithm, namely the LS algorithm. Its expression is: where σ 2 is the noise power, is the autocorrelation matrix of the channel matrix, tr{} represents the trace of the matrix, represents the matrix pseudo-inverse; P HIt means to find the conjugate transpose of P. represents the channel matrix estimated by the LS algorithm;

[0031] The dataset is generated by the RIS-assisted wireless communication system and its size is 4.8×10 5 , 75% of the data is used for network training and 25% of the data is used for testing;

[0032] 2) Build a network model based on the residual shrinkage network, i.e., DRSN, input the noisy channel matrix and the actual channel matrix into the network, and use the supervised learning method for offline training:

[0033] Soft thresholding and attention mechanism are introduced into the ResNet network structure, and a network model based on DRSN is built. The network model has a residual shrinkage module and a convolutional attention mechanism module, namely the CBAM module. The residual shrinkage module consists of three convolutional layers, two batch normalization layers, namely BN, two ReLu activation functions and a soft threshold module. The CBAM module is embedded in the residual shrinkage module, including a channel attention module, namely the CAM module, and a spatial attention module, namely the SAM module. The CAM module includes an average pooling layer, a maximum pooling layer, a two-layer neural network, a ReLU activation function and a Sigmoid activation function. The SAM module includes a maximum pooling layer, an average pooling layer, a convolutional layer and a Sigmoid activation function. The specific training steps are as follows:

[0034] ① Let x represent the input data feature. After x passes through the first convolution layer, BN layer and ReLU activation function of the residual shrinkage module, the output feature map F is obtained. 1 ∈R C×H×W , where C represents the length of the feature image, H represents the height of the feature image, and W represents the number of channels of the feature image;

[0035] ②F 1 Then, after the CBAM module embedded in the residual shrinkage module, the input feature map F 1 , the CBAM module infers the channel attention map and the spatial attention map along two independent dimensions, namely the channel and spatial dimensions, and then compares the channel attention map and the spatial attention map with F 1The CAM module compresses the input feature map in the spatial dimension and performs one-dimensional convolution. Through the parallel average pooling layer and maximum pooling layer, the feature points in the neighborhood are averaged and the maximum feature point is taken to keep the channel dimension unchanged and the spatial dimension compressed. The feature image dimension is changed from C×H×W to C×1×1. Then, the number of channels is compressed to the original 1 / r through a two-layer neural network, where r is the attenuation rate of MLP, and then expanded to the original number of channels. Finally, the results of the ReLU activation function are added, and then the output result is obtained through the Sigmoid activation function and combined with F 1 Multiplying back to the size of C×H×W, the feature map of channel attention is expressed as in and F c 1max They represent the input feature F in the CAM module. 1 The feature vector after average pooling and maximum pooling, W 1 and W 2 They represent the weights of the first and second layers of the two-layer neural network respectively, and σ() represents the Sigmoid activation function;

[0036] The SAM module is connected in series after the CAM module, and the output result F of the CAM module is 1 ′ Two 1×H×W-dimensional feature maps are obtained through the maximum pooling layer and the average pooling layer. The two feature maps are concatenated and converted into a feature map of one channel through a 7*7 convolution layer. Finally, a Sigmoid activation function is used to obtain the feature map of SAM, and the output result is compared with F 1 ′ is multiplied back to C×H×W dimensions, and the feature map of spatial attention is expressed as in and In the SAM module, the input feature F 1 ′ is the feature vector after average pooling and maximum pooling, f 7×7 () represents a 7*7 convolution operation;

[0037] The characteristic graph of the CAM module is M c (F 1 )∈R C×1×1 , and compare it with F 1 Multiply to get the output result of the CAM module F 1 ′ passes through the SAM module and is combined with its feature map M s (F 1 ′) and multiply to get the output result of CBAM module in represents the Kronecker product operation;

[0038] ③ Repeat steps ① and ② to get the output result F of the second CBAM module 2 ″, and then the third convolutional layer of the residual contraction module obtains the output feature F with the same dimension as x out Finally, the soft threshold module is used to process the noise, and its formula is: Where x represents the input feature, y represents the output feature, and τ represents the positive threshold. After soft thresholding, the feature whose absolute value of x is less than τ is set to 0, thereby achieving the effect of reducing data noise. The output result is expressed as F f ;

[0039] ④The input feature x is mapped to the denoised F by identity mapping f Add, and F f If the dimension of y is the same as that of x, then y can be defined as y = σ(F f )+x;

[0040] 3) Generate test data at different signal-to-noise ratios and input them into the trained channel estimation network. Output the estimated channel matrix and further compare it with the actual channel matrix to evaluate the network estimation performance:

[0041] After the channel estimation network training is completed, online deployment is achieved, and the RIS-assisted wireless communication system generates 10 5 For each test data, the SLS channel estimation algorithm is used to estimate the channel. Then the noise is eliminated through the channel estimation network, and finally a reliable channel estimation result is output. And compare it with the actual channel matrix H to evaluate the channel estimation performance.

Claims

1. A channel estimation method for a RIS communication system based on an improved residual shrinkage network is implemented by a RIS-assisted wireless communication system, the system comprising a transmitter, a RIS and a receiver, the transmitter comprising K single-antenna users, the RIS comprising N passive reflection units, and the receiver comprising N r The base station is composed of a base station composed of a root receiving antenna and a channel estimation module based on an improved residual shrinkage network. The links from user to RIS and user to base station experience independent Rayleigh fading, and the link from RIS to base station experiences independent Rice fading. The noise is additive Gaussian white noise. First, the RIS-assisted wireless communication system generates a data set, and uses the SLS algorithm to pre-estimate the channel to obtain a noisy channel estimation matrix. Secondly, an improved residual shrinkage network is built, and the noisy channel estimation matrix and the actual channel matrix are input into the network, and supervised learning methods are used for offline training. Finally, under different signal-to-noise ratios, the test data is input into the trained channel estimation network, and the estimated channel matrix is ​​output to evaluate the channel estimation performance. The specific steps are as follows: 1) RIS assists in the simulation of wireless communication system to generate data sets and preprocess the data: The RIS-assisted wireless communication system adopts time division multiplexing full-duplex mode and uses the uplink to estimate the channel. The channel matrix from RIS to the base station is: Its elements follow the Rice distribution, N r ×N-dimensional complex vector set, the channel matrices from user to RIS and from user to base station are and Each element is an independent complex Gaussian random variable, obeying a complex Gaussian distribution with a mean of 0 and a variance of 1. Each reflection unit of the RIS can independently adjust the phase of the incident signal. The reflection phase matrix is ​​expressed as in is the reflection phase at the nth reflection unit, diag represents the construction of a diagonal matrix, [] T represents matrix transpose. RIS is a passive reflective element and cannot process signals. It is impossible to estimate the channel matrices from user to RIS and RIS to base station separately. When estimating the cascade channel G = Bdiag(h) and direct channel d from user to base station in a RIS-assisted multi-user communication system, it is necessary to perform channel estimation based on pilot signals. The channel matrix to be estimated is expressed as H = [d, G]. The pilot signal contains channel estimation pilot signals of K users. The pilot sequences are orthogonal to each other. The pilot sequence sent by the kth user is x k =[u k,1 ,...,u k,l ...,u k,L ], L represents the pilot sequence length, u k,l represents the lth pilot symbol of the kth user. Each channel estimation data frame contains D different subframes. Let D = N + 1. The dth phase shift matrix is ​​expressed as in Represents the reflection phase on the nth RIS reflection unit in the dth phase shift matrix. The signal of the dth subframe received by the base station is Among them, H k represents the channel matrix of the kth user, p d =[1,r d ] T is the phase shift matrix, is additive Gaussian white noise, whose element v d,l represents the lth noise vector of the dth subframe. Considering the orthogonality between user pilots, after the kth user sends D subframes, the base station receives the signal: in represents D reflection phase shift matrices, represents the noise matrix, P can be represented by a discrete Fourier matrix, and the received signal of the kth user is expressed as where w = e j2π / D The receiving end first performs SLS channel pre-estimation on the received signal. The SLS algorithm further considers the impact of noise on channel estimation based on the least squares algorithm, namely the LS algorithm. Its expression is: where σ 2 is the noise power, is the autocorrelation matrix of the channel matrix, tr{} represents the trace of the matrix, represents the matrix pseudo-inverse; P H It means to find the conjugate transpose of P. represents the channel matrix estimated by the LS algorithm; The dataset is generated by the RIS-assisted wireless communication system and its size is 4.8×10 5 , 75% of the data is used for network training and 25% of the data is used for testing; 2) Build a network model based on the residual shrinkage network, i.e., DRSN, input the noisy channel matrix and the actual channel matrix into the network, and use the supervised learning method for offline training: Soft thresholding and attention mechanism are introduced into the ResNet network structure, and a network model based on DRSN is built. The network model includes a residual shrinkage module and a convolutional attention mechanism module, namely the CBAM module. The residual shrinkage module consists of three convolutional layers, two batch normalization layers, namely BN, two ReLu activation functions and a soft threshold module. The CBAM module is embedded in the residual shrinkage module, including a channel attention module, namely the CAM module, and a spatial attention module, namely the SAM module. The CAM module includes an average pooling layer, a maximum pooling layer, a two-layer neural network, a ReLU activation function and a Sigmoid activation function. The SAM module includes a maximum pooling layer, an average pooling layer, a convolutional layer and a Sigmoid activation function. The specific training steps are as follows: ① Let x represent the input data feature. After x passes through the first convolution layer, BN layer and ReLU activation function of the residual shrinkage module, the output feature map F is obtained. 1 ∈R C×H×W , where C represents the length of the feature image, H represents the height of the feature image, and W represents the number of channels of the feature image; ②F 1 Then, after the CBAM module embedded in the residual shrinkage module, the input feature map F 1 , the CBAM module infers the channel attention map and the spatial attention map along two independent dimensions, namely the channel and spatial dimensions, and then compares the channel attention map and the spatial attention map with F 1 The CAM module compresses the input feature map in the spatial dimension and performs one-dimensional convolution. Through the parallel average pooling layer and maximum pooling layer, the feature points in the neighborhood are averaged and the maximum feature point is taken to keep the channel dimension unchanged and the spatial dimension compressed. The feature image dimension is changed from C×H×W to C×1×1. Then, the number of channels is compressed to the original 1 / r through a two-layer neural network, where r is the attenuation rate of MLP, and then expanded to the original number of channels. Finally, the results of the ReLU activation function are added, and then the output result is obtained through the Sigmoid activation function and combined with F 1 Multiplying back to the size of C×H×W, the feature map of channel attention is expressed as in and They represent the input feature F in the CAM module. 1 The feature vector after average pooling and maximum pooling, W 1 and W 2 They represent the weights of the first and second layers of the two-layer neural network respectively, and σ() represents the Sigmoid activation function; The SAM module is connected in series after the CAM module, and the output result F of the CAM module is 1 ′ Two 1×H×W-dimensional feature maps are obtained through the maximum pooling layer and the average pooling layer. The two feature maps are concatenated and converted into a feature map of one channel through a 7*7 convolution layer. Finally, a Sigmoid activation function is used to obtain the feature map of SAM, and the output result is compared with F 1 ′ is multiplied back to C×H×W dimensions, and the feature map of spatial attention is expressed as in and In the SAM module, the input feature F 1 ′ is the feature vector after average pooling and maximum pooling, f 7×7 () represents a 7*7 convolution operation; The characteristic graph of the CAM module is M c (F 1 )∈R C×1×1 , and compare it with F 1 Multiply to get the output result of the CAM module F 1 ′ passes through the SAM module and is combined with its feature map M s (F 1 ′) and multiply to get the output result of CBAM module in represents the Kronecker product operation; ③ Repeat steps ① and ② to get the output result F of the second CBAM module 2 ″, and then the third convolutional layer of the residual contraction module obtains the output feature F with the same dimension as x out Finally, the soft threshold module is used to process the noise, and its formula is: Where x represents the input feature, y represents the output feature, and τ represents the positive threshold. After soft thresholding, the feature whose absolute value of x is less than τ is set to 0, thereby achieving the effect of reducing data noise. The output result is expressed as F f ; ④The input feature x is mapped to the denoised F by identity mapping f Add, and F f If the dimension of y is the same as x, then y is defined as y = σ(F f )+x; 3) Generate test data at different signal-to-noise ratios and input them into the trained channel estimation network. Output the estimated channel matrix and further compare it with the actual channel matrix to evaluate the network estimation performance: After the channel estimation network training is completed, online deployment is achieved, and the RIS-assisted wireless communication system generates 10 5 For each test data, the SLS channel estimation algorithm is used to estimate the channel. Then the noise is eliminated through the channel estimation network, and finally a reliable channel estimation result is output. And compare it with the actual channel matrix H to evaluate the channel estimation performance.

Images