Channel estimation method for intelligent reflecting surface assisted communication system based on noise suppression

Abstract

The application discloses a channel estimation method for an intelligent reflecting surface assisted communication system based on noise suppression. The method comprises the following steps: constructing an uplink communication system of a mobile user (MU) and a base station (BS) assisted by an intelligent reflecting surface (IRS); constructing a channel estimation neural network based on latent feature space noise suppression, which comprises a variational autoencoder network for extracting latent features and a UNet network for estimating channel parameters; training the variational autoencoder network; fixing the parameters of the variational autoencoder network, training the channel estimation neural network based on latent feature space noise suppression, and obtaining the trained UNet network; inputting a base station received signal into the trained UNet network, estimating channel parameters based on the UNet network, and outputting an estimated channel matrix. The application greatly improves the accuracy of channel estimation based on intelligent reflecting surface assistance by using a network model of feature space noise suppression, and reduces the time complexity of channel estimation.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology using smart reflectors, and more particularly to a channel estimation method for smart reflector-assisted communication systems based on noise suppression. Background Technology

[0002] Intelligent Reflective Surfaces (IRS) have been widely used in wireless communication, radar detection, and signal processing due to their low cost, high efficiency, and flexible configuration. An IRS consists of numerous low-cost passive reflective elements. By controlling the parameters of each reflective element, it can intelligently reconstruct the wireless propagation environment, significantly improving the communication performance of wireless transmission and achieving multiple benefits such as enhanced signal coverage, increased spectral efficiency, reduced energy consumption, and extended communication distance. To adjust the phase / amplitude of each reflective element in the IRS to effectively suppress interference and achieve optimal transmission performance, accurate Channel State Information (CSI) is required. High-quality channel estimation allows the IRS to more accurately adjust the state of its reflective elements to adapt to dynamic changes in wireless propagation. Due to cost and power consumption limitations, traditional IRSs are typically passive and cannot separately acquire channel state information from mobile devices to the IRS and from the IRS to the base station. Therefore, it is necessary to estimate the cascaded channel from the mobile device to the base station. To address these issues, recent research has focused on new channel estimation techniques and algorithms, such as the optimal minimum mean square error estimator, but these involve multidimensional integration, which imposes a significant computational burden. There is also research dedicated to developing semi-active or active IRS components, but these components also greatly increase the complexity and cost of the system.

[0003] Recently, deep learning technology has attracted much attention in IRS-assisted communication systems due to its high efficiency in channel estimation. Using deep learning methods can reduce the dependence on expensive and complex hardware, but how to reduce the time complexity of neural networks while obtaining more accurate channel estimation is also an urgent problem to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a channel estimation method for a noise-suppressed intelligent reflector-assisted communication system. The network model comprises two parts: a variational autoencoder network and a UNet network. First, the variational autoencoder network is trained. Then, with the variational autoencoder network parameters fixed, the entire network is trained, while the estimation phase utilizes only the UNet network. This achieves accurate channel estimation while reducing time complexity.

[0005] Specifically, this invention utilizes the encoder output of a noiseless latent feature vector in a variational autoencoder network, enabling the encoder of the UNet network to learn this noiseless feature. The variational autoencoder network minimizes the estimated noiseless BS received signal. The goal is to achieve a normalized mean square error between the input and the true value, primarily to ensure that the encoder outputs a noise-free latent feature vector. Then, the variational autoencoder is fixed, and the entire network is trained, allowing the UNet network's encoder to learn this noise-free latent feature vector. Since the UNet network's input is noisy received data, the variational autoencoder forces the UNet network's encoder to learn a noise-free latent feature vector, thereby suppressing noise and improving the accuracy of channel estimation in the subsequent decoder. Furthermore, in the estimation phase, only the UNet network is needed for channel estimation, effectively reducing the time cost of channel estimation.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] Firstly, a channel estimation method for a noise-suppressed intelligent reflector-assisted communication system includes the following:

[0008] Step 1: Construct an uplink communication system for mobile subscribers (MU) and base stations (BS) assisted by an intelligent reflector (IRS);

[0009] Step 2: Construct a channel estimation neural network based on latent feature space noise suppression. This neural network includes a variational autoencoder network for extracting latent features and a UNet network for estimating channel parameters.

[0010] Step 3: Train the variational autoencoder network;

[0011] Step 4: Fix the parameters of the variational autoencoder network, train the channel estimation neural network based on latent feature space noise suppression, and obtain the trained UNet network.

[0012] Step 5: Input the base station received signal into the UNet network trained in Step 4, estimate the channel parameters based on the UNet network, and output the estimated channel matrix.

[0013] Furthermore, in the channel estimation method for the above-mentioned intelligent reflector-assisted communication system based on noise suppression, the uplink communication system includes a BS, an IRS, and a MU, wherein the BS and MU have M antennas and 1 antenna, respectively, and the IRS has N reflection elements; the equivalent phase shift matrix of the IRS reflection element is denoted as Ω = diag(θ). Where diag(·) denotes the diagonalization operation of a vector, and the phase-shifted vector. β n∈[0,1] and These represent the reflection amplitude and reflection phase of the nth reflecting unit, respectively. T For the transpose operation, let β n When (n=1,2,…,N) is 1, and the channel follows quasi-static flat fading, the communication link from MU to BS is denoted as The direct channel is blocked, and several scattered channels exist; the communication link from IRS to BS is denoted as... The communication link from MU to IRS is denoted as Each has one direct channel and several scattered channels, where ∈ indicates belonging to the symbol. The communication link from MU to BS via IRS, representing the set of complex numbers, is written as:

[0014] h=H1Ωh2=H1diag(θ)h2=H1diag(h2)θ

[0015] Let the channel be C = H1diag(h2), For a cascaded channel, the channel matrix H to be estimated is expressed as:

[0016]

[0017] In the k-th time slot, the received signal y(k) of the BS is expressed as:

[0018]

[0019] in, x(k) represents the pilot signal transmitted by the MU, w(k) represents the noise vector received by the BS, θ(k) is the phase shift vector of the k-th time slot, x(k) is set to 1, and all elements in w(k) follow a function with a mean of 0 and a variance of 0. The complex Gaussian distribution, The noise power is represented as follows: After N+1 time slots, the received signal of the BS is expressed as:

[0020]

[0021] Among them, Y=[y(1),y(2),…,y(N+1)], It is designed in the form of a discrete Fourier transform matrix, W = [w(1), w(2), ..., w(N+1)], where Z represents the noise-free BS received signal.

[0022] Furthermore, in the aforementioned channel estimation method for a noise-suppressed intelligent reflector-assisted communication system, the construction of a channel estimation neural network based on latent feature space noise suppression includes a variational autoencoder network for extracting latent features and a UNet network for estimating channel parameters. The input to the variational autoencoder network is the noise-free BS received signal Z, which outputs a latent feature vector P1 after passing through encoder 1, and then outputs the estimated noise-free BS received signal after passing through decoder 1. The input to the UNet network is the received signal Y from the BS after passing through N+1 time slots. After passing through encoder 2, it outputs a feature vector P2, and after passing through decoder 2, it outputs an estimated channel matrix.

[0023] Furthermore, in the aforementioned channel estimation method for an intelligent reflector-assisted communication system based on noise suppression, the channel estimation neural network based on latent feature space noise suppression comprises two parts: a variational autoencoder network and a UNet network. The variational autoencoder network includes encoder 1 and decoder 1. The encoder 1 consists of four two-dimensional convolutional blocks, fully connected block 1, and fully connected block 2. After mapping the input data to the mean μ and standard deviation σ in the latent space, it performs reparameterization to output the latent feature vector P1. The two-dimensional convolutional block includes a two-dimensional convolutional layer with a kernel size of 5×2 and a stride of (2,1), a batch normalization layer, and a ReLU activation function. The channel dimensions of the four two-dimensional convolutional blocks are 32, 64, 64, and 64, respectively. The fully connected block 1 consists of two... The neural network structure consists of alternating Dropout layers, fully connected layers, batch normalization layers, and ReLU activation functions; the fully connected block 2 has the same structure as fully connected block 1; the decoder 1 structure includes fully connected block 3 and four two-dimensional deconvolution blocks, which decode the encoding in the latent space into the original data; the fully connected block 3 includes a fully connected layer, a batch normalization layer, a ReLU activation function, and a fully connected layer, which is then processed by a batch normalization layer before undergoing a reshape operation; the two-dimensional deconvolution block includes a two-dimensional deconvolution layer with a kernel size of 5×2 and a stride of (2,1), a batch normalization layer, and a ReLU activation function, and the input of each two-dimensional deconvolution layer is obtained by concatenating the output of the previous layer and the output of the corresponding two-dimensional convolution layer; the output latent feature vector P1 is expressed as:

[0024] P1=γ⊙exp(σ)+μ

[0025] Where γ is a random number drawn from a standard normal distribution, ⊙ represents the element-wise multiplication, and finally, after passing through decoder 1, the estimated noise-free BS received signal is output.

[0026] The UNet network includes encoder 2 and decoder 2. Encoder 2 consists of four two-dimensional convolutional blocks and one fully connected block 4, and outputs vector P2 after passing through encoder 2. Decoder 2 consists of one fully connected block 5 and four two-dimensional deconvolutional blocks, and finally outputs the estimated channel matrix after passing through decoder 2.

[0027] Furthermore, in the channel estimation method of the above-mentioned intelligent reflector-assisted communication system based on noise suppression, the training of the variational autoencoder network aims to minimize the estimated noise-free BS received signal. The objective is the normalized mean square error between the true value and the actual value, i.e.:

[0028]

[0029] Where MSE represents the mean square error calculation, σ i and μ i The vector elements of σ and μ are used to train the variational autoencoder network until a set of well-trained variational autoencoder network parameters is obtained.

[0030] Furthermore, in the channel estimation method of the aforementioned intelligent reflector-assisted communication system based on noise suppression, the parameters of the fixed variational autoencoder network are used to train the UNet network in the channel estimation neural network based on latent feature space noise suppression to obtain a well-trained neural network; the network parameters of the fixed variational autoencoder are used so that the output vector P2 of encoder 2 of the UNet network can learn the features of the noise-free latent feature vector P1, and the estimated channel matrix output by decoder 2. The target is a channel matrix H that approximates the true value to be estimated, i.e.:

[0031]

[0032] NMSE represents the normalized mean square error calculation, which trains the channel estimation neural network based on latent feature space noise suppression until a fully trained UNet network is obtained.

[0033] Furthermore, in the channel estimation method of the above-mentioned intelligent reflector-assisted communication system based on noise suppression, the base station receives the signal and inputs it into the trained neural network to estimate the channel parameters and output the estimated channel matrix based on the UNet network; in the actual estimation stage, the base station receives the signal and inputs it into the trained UNet network to obtain the desired estimated channel matrix.

[0034] In a second aspect, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the aforementioned channel estimation method for an intelligent reflector-assisted millimeter-wave communication system based on an autoencoder.

[0035] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the aforementioned channel estimation method for an intelligent reflector-assisted millimeter-wave communication system based on an autoencoder.

[0036] The beneficial effects of this invention are as follows:

[0037] In situations where obstacles obstruct communication between mobile users and base stations, communication can proceed smoothly with the assistance of intelligent reflective surfaces. However, since intelligent reflective surfaces consist of numerous passive reflective elements, obtaining accurate channel estimation is extremely challenging. This invention utilizes the encoder of a variational coding network to acquire noise-free latent feature vectors, which helps the encoder of the UNet network learn noise-free feature vectors, achieving noise suppression and significantly improving the accuracy of channel estimation. Finally, after training, channel estimation is performed using only the UNet network portion, reducing time complexity. Attached Figure Description

[0038] Figure 1 This is the overall flowchart of the channel estimation method for the intelligent reflector-assisted communication system based on noise suppression according to the present invention;

[0039] Figure 2 A schematic diagram of an uplink communication system for mobile users and base stations assisted by an intelligent reflective surface, provided by the present invention;

[0040] Figure 3 A neural network structure diagram based on latent feature space noise suppression provided by the present invention;

[0041] Figure 4 The specific structure of the encoder-decoder in the neural network based on latent feature space noise suppression provided by the present invention;

[0042] Figure 5 A schematic diagram of the neural network training and estimation stages provided by this invention;

[0043] Figure 6 The performance curve of the estimation accuracy as a function of the input signal-to-noise ratio provided by this invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0045] This invention provides a channel estimation method for an intelligent reflector-assisted communication system based on latent feature space noise suppression, such as... Figure 1 As shown, the method mainly includes the following steps:

[0046] Step 1: Construct an uplink communication system for mobile subscribers (MU) and base stations (BS) assisted by an intelligent reflector (IRS);

[0047] Step 2: Construct a channel estimation neural network based on latent feature space noise suppression. This neural network includes a variational autoencoder network for extracting latent features and a UNet network for estimating channel parameters.

[0048] Step 3: Train the variational autoencoder network;

[0049] Step 4: Fix the parameters of the variational autoencoder network, train the channel estimation neural network based on latent feature space noise suppression, and obtain the trained UNet network.

[0050] Step 5: Input the base station received signal into the UNet network trained in Step 4, estimate the channel parameters based on the UNet network, and output the estimated channel matrix.

[0051] In step 1 above, if Figure 2 As shown, the uplink communication system includes a BS, an IRS, and a MU, wherein the BS and MU have M antennas and 1 antenna, respectively, and the IRS has N reflection elements; the equivalent phase shift matrix of the IRS reflection element is denoted as Ω = diag(θ). Where diag(·) denotes the diagonalization operation of a vector, and the phase-shifted vector. β n ∈[0,1] and These represent the reflection amplitude and reflection phase of the nth reflecting unit, respectively. T For the transpose operation, let β n When (n=1,2,…,W) is 1, and the channel follows quasi-static flat fading, the communication link from MU to BS is denoted as The direct channel is blocked, and several scattered channels exist; the communication link from IRS to BS is denoted as... The communication link from MU to IRS is denoted as Each has one direct channel and several scattered channels, where ∈ indicates belonging to the symbol. The communication link from MU to BS via IRS, representing the set of complex numbers, is written as:

[0052] h=H1Ωh2=H1diag(θ)h2=H1diag(h2)θ

[0053] Let the channel be C = H1diag(h2), For a cascaded channel, the channel matrix H to be estimated is expressed as:

[0054]

[0055] In the k-th time slot, the received signal y(k) of the BS is expressed as:

[0056]

[0057] in, x(k) represents the pilot signal transmitted by the MU, w(k) represents the noise vector received by the BS, θ(k) is the phase shift vector of the k-th time slot, x(k) is set to 1, and all elements in w(k) follow a function with a mean of 0 and a variance of 0. The complex Gaussian distribution, The noise power is represented as follows: After N+1 time slots, the received signal of the BS is expressed as:

[0058]

[0059] Among them, Y=[y(1),y(2),…,y(N+1)], It is designed in the form of a discrete Fourier transform matrix, W = [w(1), w(2), ..., w(N+1)], where Z represents the noise-free BS received signal.

[0060] In step 2 above, such as Figure 3 As shown, a channel estimation neural network based on latent feature space noise suppression is constructed. This neural network includes a variational autoencoder network for extracting latent features and a UNet network for estimating channel parameters. The input of the variational autoencoder network is the noise-free BS received signal Z. After passing through encoder 1, it outputs a latent feature vector P1. After passing through decoder 1, it outputs the estimated noise-free BS received signal. The input to the UNet network is the received signal Y from the BS after passing through N+1 time slots. After passing through encoder 2, it outputs a feature vector P2, and after passing through decoder 2, it outputs an estimated channel matrix.

[0061] like Figure 3 and Figure 4As shown, the variational autoencoder network includes encoder 1 and decoder 1. The encoder 1 consists of four two-dimensional convolutional blocks, fully connected block 1, and fully connected block 2. After mapping the input data to the mean μ and standard deviation σ in the latent space, it is reparameterized to output the latent feature vector P1. The two-dimensional convolutional block includes a two-dimensional convolutional layer with a kernel size of 5×2 and a stride of (2,1), a batch normalization layer, and a ReLU activation function. The channel dimensions of the four two-dimensional convolutional blocks are 32, 64, 64, and 64, respectively. The fully connected block 1 is a neural network composed of two Dropout layers, a fully connected layer, a batch normalization layer, and a ReLU activation function, alternating between them. Structure; The fully connected block 2 has the same structure as the fully connected block 1; The decoder 1 includes a fully connected block 3 and four two-dimensional deconvolution blocks, which decode the encoding in the latent space into the original data; The fully connected block 3 includes a fully connected layer, a batch normalization layer, a ReLU activation function, and a fully connected layer, which is then processed by a batch normalization layer before undergoing a reshape operation; The two-dimensional deconvolution block includes a two-dimensional deconvolution layer with a kernel size of 5×2 and a stride of (2,1), a batch normalization layer, and a ReLU activation function. The input of each two-dimensional deconvolution layer is obtained by concatenating the output of the previous layer and the output of the corresponding two-dimensional convolution layer; The output latent feature vector P1 is expressed as:

[0062] P1=γ⊙exp(σ)+μ

[0063] Where γ is a random number drawn from a standard normal distribution, ⊙ represents the element-wise multiplication, and finally, after passing through decoder 1, the estimated noise-free BS received signal is output.

[0064] The UNet network consists of encoder 2 and decoder 2. Encoder 2 comprises four 2D convolutional blocks and one fully connected block 4, outputting vector P2. Decoder 2 comprises one fully connected block 5 and four 2D deconvolutional blocks, ultimately outputting the estimated channel matrix.

[0065] In step 3 above, such as Figure 5 As shown, in the first training phase, the goal is to minimize the estimated noise-free BS received signal. The objective is the normalized mean square error between the true value and the actual value, i.e.:

[0066]

[0067] Where MSE represents the mean square error calculation, σ i and μ i The vector elements of σ and μ are used to train the variational autoencoder network until a set of well-trained variational autoencoder network parameters is obtained.

[0068] In step 4 above, such as Figure 5 As shown, in the second training phase, the network parameters of the variational autoencoder are fixed so that the output vector P2 of encoder 2 of the UNet network can learn the features of the noise-free latent feature vector P1 and the estimated channel matrix output by decoder 2. The target is a channel matrix H that approximates the true value to be estimated, i.e.:

[0069]

[0070] NMSE represents the normalized mean square error calculation, which trains the channel estimation neural network based on latent feature space noise suppression until a fully trained UNet network is obtained.

[0071] In step 5 above, if Figure 5 As shown, in the actual estimation stage, the BS received signal is input into the trained UNet network to obtain the desired estimated channel matrix.

[0072] In one embodiment, a simulation experiment was conducted on the method proposed in this invention. The simulation conditions included: the number of base station antennas M = 16, the number of IRS end reflective elements N = 32, the direct path in the communication link from MU to BS was blocked, and two scattered paths existed; the communication links from IRS to BS and from MU to IRS each had one direct path and two scattered paths, all following a complex Gaussian distribution, with the power of the direct path being 1 and the power of the scattered paths being 0.1. Various channel estimation methods adopted the above-mentioned channel conditions. The image showing the change of the normalized mean square error of channel estimation with the signal-to-noise ratio (SNR) under different transmitting end signal-to-noise ratios is shown below. Figure 6 It can be seen that the channel estimation accuracy of the channel estimation method proposed in this invention is much higher than that of the traditional least squares estimator.

[0073] Based on the same inventive concept, another embodiment of the present invention provides an electronic device (computer, server, smartphone, etc.) including a memory and a processor, wherein the memory stores a computer program configured to be executed by the processor, and the computer program includes instructions for performing the steps of the method of the present invention.

[0074] Based on the same inventive concept, another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, optical disk), which stores a computer program that, when executed by a computer, implements the various steps of the method of the present invention.

[0075] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be defined by the scope of the claims.

Claims

1. A channel estimation method for a smart reflector-assisted communication system based on noise suppression, characterized in that, include: Step 1: Construct an uplink communication system for mobile subscribers (MU) and base stations (BS) assisted by an intelligent reflector (IRS); Step 2: Construct a channel estimation neural network based on latent feature space noise suppression. This neural network includes a variational autoencoder network for extracting latent features and a UNet network for estimating channel parameters. The channel estimation neural network based on latent feature space noise suppression consists of two parts: a variational autoencoder network and a UNet network. The variational autoencoder network includes encoder 1 and decoder 1. The structure of encoder 1 includes four two-dimensional convolutional blocks, fully connected block 1, and fully connected block 2, which map the input data to the mean in the latent space. and standard deviation Then, reparameterization is performed to output the latent feature vector. The two-dimensional convolutional block includes a two-dimensional convolutional layer with a kernel size of 5×2 and a stride of (2,1), a batch normalization layer, and a ReLU activation function. The channel dimensions of the four two-dimensional convolutional blocks are 32, 64, 64, and 64, respectively. The fully connected block 1 is a neural network structure composed of two Dropout layers, a fully connected layer, a batch normalization layer, and a ReLU activation function, alternating between the two layers. The fully connected block 2 has the same structure as the fully connected block 1. The decoder 1 includes a fully connected block 3 and four two-dimensional deconvolutional blocks, which decode the encoding in the latent space into the original data. The fully connected block 3 includes a fully connected layer, a batch normalization layer, and a ReLU activation function. The activation function, the fully connected layer, and then the batch normalization layer are processed before the reshape operation; the two-dimensional deconvolution block includes a two-dimensional deconvolution layer with a kernel size of 5×2 and a stride of (2,1), a batch normalization layer and a ReLU activation function. The input of the two-dimensional deconvolution layer in each two-dimensional deconvolution block is obtained by concatenating the output of the previous layer and the output of the corresponding two-dimensional convolution layer. Output latent feature vector Represented as: in, These are random numbers drawn from a standard normal distribution. This indicates that corresponding elements are multiplied, and the final output after passing through decoder 1 is the estimated noise-free BS received signal. ; The UNet network includes an encoder 2 and a decoder 2; the encoder 2 consists of four two-dimensional convolutional blocks and one fully connected block 4, and outputs a vector after passing through the encoder 2. Decoder 2 consists of a fully connected block 5 and four two-dimensional deconvolution blocks; the final output after passing through decoder 2 is the estimated channel matrix. ; Step 3: Train the variational autoencoder network; including minimizing the estimated noise-free BS received signal. The objective is the normalized mean square error between the true value and the actual value, i.e.: in, This indicates the calculation of the mean square error. and They are and The vector elements are used to train the variational autoencoder network until a set of well-trained variational autoencoder network parameters is obtained. Step 4: Fix the parameters of the variational autoencoder network and train the channel estimation neural network based on latent feature space noise suppression to obtain the trained UNet network; this includes fixing the network parameters of the variational autoencoder and using the encoder 2 output vector of the UNet network. Able to learn noise-free latent feature vectors Features, estimated channel matrix output by decoder 2 Approximately true channel matrix to be estimated For the goal, namely: in, This indicates the calculation of normalized mean square error, which is used to train the channel estimation neural network based on latent feature space noise suppression until a well-trained UNet network is obtained. Step 5: Input the base station received signal into the UNet network trained in Step 4, estimate the channel parameters based on the UNet network, and output the estimated channel matrix.

2. The channel estimation method for a noise-suppressed intelligent reflector-assisted communication system according to claim 1, characterized in that, The uplink communication system includes a BS, an IRS, and a MU, wherein the BS and MU each have One antenna and one antenna, the IRS has One reflection unit; the equivalent phase shift matrix of the IRS reflection unit is denoted as... , ,in This represents a diagonalization operation on a vector, a phase-shifted vector. , and Representing the first The reflection amplitude and reflection phase of each reflecting unit. For the transpose operation, let When the value is 1 and the channel follows quasi-static flat fading, the communication link from MU to BS is denoted as . The direct channel is blocked, and there are several scattering channels; The communication link from IRS to BS is denoted as The communication link from MU to IRS is denoted as Each has one direct channel and several scattering channels, among which Indicates that it belongs to the symbol. The communication link from MU to BS via IRS, representing the set of complex numbers, is written as: Channel For cascaded channels, the channel matrix to be estimated is... Represented as: In the Each time slot, the BS receives the signal. Represented as: in, , This indicates the pilot signal transmitted by MU. This represents the noise vector received by the BS. For the first The phase shift vector of each time slot, Set to 1, The elements in the set all follow a mean of 0 and a variance of . The complex Gaussian distribution, Noise power; after After one time slot, the received signal of the BS is represented as: in, , Designed in the form of a discrete Fourier transform matrix, , This represents a noise-free BS received signal.

3. The channel estimation method for a noise-suppressed intelligent reflector-assisted communication system according to claim 1, characterized in that, Step 2 includes the following: the input to the variational autoencoder network is the noise-free BS received signal. After passing through encoder 1, the output latent feature vector is obtained. After passing through decoder 1, the output is the estimated noise-free BS received signal. ; The input to the UNet network is the one that has passed through After one time slot, the BS receives the signal. After passing through encoder 2, the output feature vector is... After passing through decoder 2, the estimated channel matrix is ​​output. .

4. The channel estimation method for a smart reflector-assisted communication system based on noise suppression as described in claim 1, characterized in that, Step 5 includes, in the actual estimation stage, inputting the BS received signal into the trained UNet network to obtain the desired estimated channel matrix.

5. An electronic device, characterized in that, It includes a memory and a processor, wherein when the processor executes a program stored in the memory, it is able to implement the channel estimation method of the intelligent reflector-assisted communication system based on noise suppression as described in any one of claims 1-4.

6. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, can implement the channel estimation method for the noise-suppressed intelligent reflector-assisted communication system according to any one of claims 1-4.

Images