RIS-assisted millimeter wave channel estimation method and system combining LMMSE algorithm and double-attention mechanism neural network

Abstract

The invention provides an RIS-assisted millimeter wave channel estimation method and system combining an LMMSE algorithm and a double-attention mechanism neural network. S1, establishing a multi-user millimeter wave communication model; s2, deploying a base station antenna and an RIS reflection unit in a uniform planar array manner according to the constructed communication model, and modeling a channel by adopting an SV model to obtain a channel model from the RIS to the base station and a channel model from a user to the RIS; s3, based on the channel model from the RIS to the base station and the channel model from the user to the RIS, according to a base station receiving signal, estimating a cascade channel matrix H by using an LMMSE estimator; s4, constructing an L-DA neural network, and performing offline training on the neural network by using a data set composed of the pre-estimated channel matrix and the actual channel matrix to obtain a weight parameter of an L-DA neural network model; and S5, performing online estimation on the pre-estimated channel matrix by using the trained L-DA neural network, and outputting a final channel estimation matrix.

Description

Technical Field

[0001] The present invention relates to the field of millimeter wave communication technology, and in particular to a RIS-assisted millimeter wave channel estimation method and system combining a LMMSE algorithm with a dual-attention mechanism neural network. Background Art

[0002] In the field of wireless communications, millimeter waves, due to their abundant spectrum resources, are a key technology for future communications development. However, millimeter wave signals are highly susceptible to obstruction and path attenuation during propagation, severely limiting their propagation range and communication quality. Reconfigurable Intelligent Surfaces (RIS), an emerging technology, are considered an effective solution to these problems. Reconfigurable intelligent metasurfaces, by their ability to reflect wireless signals, can provide additional propagation paths for wireless signals, effectively mitigating the impact of obstruction and path loss on signal transmission. Furthermore, they can manipulate the phase and amplitude of wireless signals reflected by their surfaces, further improving signal transmission efficiency. To fully leverage the multipath advantages provided by RIS and enhance the communication quality of millimeter wave systems, the design of RIS reflection parameters must be fully tailored to the current wireless environment. Achieving this goal relies heavily on accurate channel estimation. However, with the increasing number of base station antennas and RIS reflector elements, the number of channels is increasing significantly. This increase dramatically increases the complexity of channel estimation and poses greater challenges to its accuracy.

[0003] Traditional channel estimation algorithms are mostly based on concepts such as least squares (LS) or LMMSE. However, both of these algorithms have significant limitations. The performance of the least squares method significantly degrades in scenarios with non-Gaussian noise (such as impulse noise) or high noise intensity. While the linear minimum mean square error method takes noise into account to some extent, it has high computational complexity and limited estimation accuracy. Therefore, these two traditional algorithms are no longer able to meet the channel estimation requirements of the increasingly complex RIS-assisted millimeter wave channels.

[0004] In recent years, deep learning has been widely used in the field of channel estimation due to its advantages in inference time, estimation accuracy, and universality. Among them, the Skip-Connection Attention (SC-Attention) network and the Convolutional Deep Residual Network (CDRN) based on convolutional neural networks are relatively typical deep learning neural networks. The SC-Attention network effectively improves the accuracy of channel estimation by introducing a self-attention layer and skip connection mechanism. The CDRN cleverly models the channel estimation problem as a denoising problem, using a neural network to learn the residual noise, thereby improving estimation accuracy.

[0005] However, although deep learning has made some progress in channel estimation, due to the extremely complex network environment in which millimeter wave channels are located, the current channel estimation methods still have much room for improvement in accuracy. Summary of the Invention

[0006] The present invention is made to solve the above problems, and its purpose is to provide a RIS-assisted millimeter wave channel estimation method and system combining the LMMSE algorithm with a dual attention mechanism neural network.

[0007] The present invention provides a RIS-assisted millimeter wave channel estimation method combining an LMMSE algorithm with a dual-attention mechanism neural network, which is used for channel estimation and has the following characteristics. The method specifically includes the following steps: S1: a communication model construction step, in which a multi-user millimeter wave communication model is constructed, wherein there is a base station, a RIS board and K users, the base station is equipped with M antennas, the RIS board is equipped with N reflection units, and each user is equipped with a single antenna; S2: a channel modeling step, in which channel modeling is performed according to the millimeter wave communication model, the channel adopts the Saleh-Valenzuela (SV) model, and the antenna and the RIS reflection unit adopt a uniform planar array (Uniform Planar Array). The system comprises a first step of deploying a base station (BS) and a second step of deploying a user-interface (RIS) array (UPA), respectively determining the array response vectors of the base station and the RIS in the azimuth and elevation directions, thereby obtaining the channel model from the RIS to the base station and the channel model from the user to the RIS; a second step of pre-estimating the channel matrix, wherein the base station uses an LMMSE estimator to pre-estimate the cascade channel matrix H according to the received uplink signal of the user, thereby obtaining the pre-estimated channel matrix; a third step of constructing an L-DA neural network model, wherein the pre-estimated channel matrix and the true channel matrix are combined into a label pair, and a data set consisting of multiple label pairs obtained through multiple measurements is format-aligned. The L-DA neural network is trained offline using the format-aligned data set, and the L-DA neural network model and the corresponding weight parameters in the network are obtained through multiple rounds of iterations; and a third step of outputting the pre-estimated channel matrix online using the L-DA neural network model and the corresponding weight parameters, thereby outputting the channel estimation matrix.

[0008] The L-DA RIS-assisted millimeter wave channel estimation method provided by the present invention may also have the following features: wherein, in S1, a multi-user millimeter wave communication model is established, specifically:

[0009] Assume that the channel from RIS to the base station is The channel from user to RIS is The reflection coefficient of RIS is The communication protocol adopts a frame structure protocol. A subframe contains T time slots. Each user sends a pilot sequence in the T time slots of each subframe. Within the same subframe, the reflection coefficient of RIS remains unchanged. Between different subframes, the reflection coefficient of RIS is different. The user end uses an orthogonal pilot sequence.

[0010] The received signal expression of the base station in each subframe is:

[0011]

[0012] Where, is noise, with mean 0 and variance σ 2 The complex Gaussian distribution of Where I is the identity matrix, φ b is the RIS reflection coefficient of the b-th subframe, diag(x) represents the diagonal matrix with vector x as the diagonal, since The uplink signals from all users received by the base station in one subframe can be expressed as:

[0013]

[0014] Where, is the cascade channel from the user to the base station,

[0015] The pilot sequence of each user is set to an orthogonal pilot sequence, and the uplink signal from the kth user received by the base station in one subframe is:

[0016]

[0017] Where z b,k The base station receives the signal from the kth user in the bth subframe, combines the signals received in B subframes, and writes the received signal received by the base station in matrix form:

[0018] Z k =HΘ+U

[0019] Where, is the received signal at the base station after B subframes. is the reflection coefficient matrix of RIS after B subframes, It's noise.

[0020] The RIS-assisted millimeter wave channel estimation method of the L-DA neural network provided by the present invention may also have the following characteristics: wherein, in S2, the channel model from RIS to the base station is expressed as:

[0021]

[0022] Where, L g and are the number of effective paths and path gain from RIS to base station channels, and denote the azimuth and elevation angles at the base station, respectively. and denote the azimuth and elevation angles at RIS, respectively. and denote the array response vectors of the base station and RIS respectively.

[0023] The channel model from user to RIS is expressed as:

[0024]

[0025] Where, L f and are the number of effective paths and path gain of the channel from user to RIS, and represent the azimuth and elevation angles of the user respectively.

[0026] The RIS-assisted millimeter wave channel estimation method of the L-DA neural network provided by the present invention may also have the following characteristics: in S3, the cascade channel matrix H is pre-estimated using the LMMSE estimator to obtain the pre-estimated channel matrix, specifically:

[0027] Find a matrix A such that Denotes the pre-estimated channel matrix, let represents the mean square error between the cascade channel matrix and the pre-estimated channel matrix, Substituting into the formula we get Right now:

[0028]

[0029] Where A is the coefficient matrix to be solved, H is the cascade channel from the user to the base station, and Z k is the received signal at the base station, tr represents the trace of the matrix, It means to find the mathematical expectation of the matrix. By finding the partial derivative of A and setting it equal to 0, the expression of matrix A is:

[0030] A=(Θ H R H Θ+R N ) -1 Θ H R H

[0031] Where R H is the autocorrelation matrix of channel H, R N is the autocorrelation matrix of the noise, and the expression of the pre-estimated channel matrix is:

[0032]

[0033] Where, Represents the pre-estimated channel matrix.

[0034] The RIS-assisted millimeter wave channel estimation method using the L-DA neural network provided by the present invention may also have the following features: wherein, in S5, the method for online estimation of the pre-estimated channel matrix using the L-DA neural network model and the corresponding weight parameters is specifically as follows:

[0035] The L-DA neural network model is composed of two branches. Divided into two channels, real and imaginary, we get Will As the input of the two branches, the upper branch consists of 5 network blocks, each of which contains a convolutional layer, a batch normalization layer and a ReLu activation function layer, and a jump link is added between each network block, that is, the input of each block is obtained by adding the output of the previous block to the input of the previous block. The lower branch is also composed of 5 network blocks, each of which contains a convolutional layer, a batch normalization layer, a ReLu activation function layer and a SE-C-CBAM module, and a jump link is added between each network block. The outputs of the two branches are spliced ​​according to the channel dimension, and then pass through two network blocks. Each network block consists of a convolutional layer, a batch normalization layer, and a ReLu activation function layer, and then passes through a multi-head attention mechanism module. Finally, a convolution layer is used to restore the shape of the output noise so that it is consistent with the input Keep the shape consistent.

[0036] The RIS-assisted millimeter wave channel estimation method of the L-DA neural network provided by the present invention may also have the following characteristics: Expressed as:

[0037]

[0038] Where, yes The real part of yes The imaginary part of .

[0039] The RIS-assisted millimeter wave channel estimation method of the L-DA neural network provided by the present invention may also have the following characteristics: a residual connection is added to the L-DA neural network model, that is, the final output channel estimation matrix is ​​the input minus the output noise.

[0040] The present invention provides an L-DA neural network RIS-assisted millimeter wave channel estimation system, which has the following characteristics: a communication model construction module, which builds a multi-user millimeter wave communication model, in which there is a base station, a RIS board and K users, the base station is equipped with M antennas, the RIS board is equipped with N reflection units, and each user is equipped with a single antenna; a channel modeling module, which performs channel modeling according to the millimeter wave communication model, the channel adopts the SV model, the antenna and the RIS reflection unit adopt UPA deployment, and respectively determines the array response vectors of the base station and the RIS in the azimuth and elevation directions, thereby obtaining the channel model from the RIS to the base station and the channel model from the user to the RIS; a pre-estimated channel matrix module, which performs channel modeling based on the RIS to The base station uses the LMMSE algorithm to pre-estimate the cascade channel matrix H based on the received user uplink signal to obtain the pre-estimated channel matrix. The L-DA neural network model construction module forms a label pair with the pre-estimated channel matrix and aligns the data set composed of multiple label pairs obtained through multiple measurements. The L-DA neural network is trained offline using the aligned data set. After multiple rounds of iterations, the L-DA neural network model and the corresponding weight parameters in the network are obtained. The output module uses the L-DA neural network model and the corresponding weight parameters to perform online estimation of the pre-estimated channel matrix and output the channel estimation matrix.

[0041] Functions and effects of the invention

[0042] The RIS-assisted millimeter wave channel estimation method and system of the L-DA neural network involved in the present invention have the following beneficial effects:

[0043] 1. High-precision estimation. This invention organically combines the preliminary estimation advantages of the LMMSE algorithm with the powerful feature learning capabilities of neural networks, significantly improving the accuracy of millimeter-wave channel estimation. Under different signal-to-noise ratio conditions, the L-DA neural network proposed in this invention can more accurately estimate the channel than existing algorithms such as CDRN, DnCNN, and SC-Attention, achieving a lower channel normalized mean square error (NMSE), providing a strong guarantee for achieving reliable millimeter-wave communications.

[0044] 2. Strong universality. The L-DA neural network proposed in this paper demonstrates more accurate channel estimation effects for different numbers of base station antennas or RIS reflection units, demonstrating the universality of the proposed scheme for different system parameters.

[0045] 3. Module Effectiveness: By designing and implementing ablation experiments, we thoroughly validated the effectiveness of key modules within the L-DA neural network model. The experimental results demonstrate that the SE-C-CBAM module and multi-head attention mechanism play a significant role in improving algorithm performance. This further validates the scientific and rational design of our model and provides solid theoretical support for subsequent technical improvements and optimizations. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a schematic diagram of a millimeter wave communication model in an embodiment of the present invention;

[0047] Figure 2 is a schematic diagram of an L-DA neural network model in an embodiment of the present invention;

[0048] Figure 3 is a schematic diagram of a SE-C-CBAM module in an embodiment of the present invention;

[0049] Figure 4 is a performance comparison chart of various models in the embodiments of the present invention;

[0050] Figure 5 This is a performance comparison chart of different models when the number of base station antennas is 8 in an embodiment of the present invention;

[0051] Figure 6 This is a performance comparison chart of different models when the number of base station antennas is 24 in an embodiment of the present invention;

[0052] Figure 7 This is a first result diagram showing the effect of the number of reflection units and the number of effective paths on the performance of the model in an embodiment of the present invention;

[0053] Figure 8 is a second result diagram showing the effect of the number of reflection units and the number of effective paths on the performance of the model in an embodiment of the present invention;

[0054] Figure 9 Graph showing the results of an ablation experiment in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] To make the technical means, creative features, objectives, and effects achieved by the present invention easier to understand, the following embodiments, combined with the accompanying drawings, specifically describe the RIS-assisted millimeter-wave channel estimation method and system of the present invention, which combines the Linear Minimum Mean Square Error (LMMSE) algorithm with the L-DA (LMMSE-Double Attention) neural network.

[0056] Figure 1 Schematic diagram of a millimeter wave communication model in an embodiment of the present invention.

[0057] like Figure 1 As shown, the RIS-assisted millimeter wave channel estimation method of the L-DA neural network of this embodiment is used to estimate the channel, and specifically includes the following steps:

[0058] Step S1 constructs the communication model. A multi-user millimeter-wave communication model is constructed, specifically an uplink multi-user MU-MIMO millimeter-wave wireless communication model assisted by RIS. This model involves a base station, a RIS board, and K users. The base station is equipped with M antennas, the RIS board is equipped with N reflectors, and each user is equipped with a single antenna. In this model, the link from the user directly to the base station is called a direct link, and the link from the user to the RIS and then to the base station is called a cascade link.

[0059] The specific method for building a multi-user millimeter wave communication model is as follows:

[0060] Assume that the channel from RIS to the base station is The channel from user to RIS is The reflection coefficient of RIS is

[0061] The uplink channel estimation protocol uses a frame structure. A subframe consists of T time slots, and each user transmits a pilot sequence in each of the T time slots of each subframe. Within a subframe, the RIS reflection coefficient remains constant. However, the RIS reflection coefficient varies between subframes.

[0062] The user end uses an orthogonal pilot sequence

[0063] The received signal expression of the base station in each subframe is:

[0064]

[0065] Where, is noise, with mean 0 and variance σ 2 The complex Gaussian distribution of Where I is the identity matrix, Φ b is the RIS reflection coefficient of the b-th subframe, and diag(·) represents a diagonal matrix whose diagonal is the vector ·.

[0066] because The uplink signals from all users received by the base station in one subframe can be expressed as:

[0067]

[0068] Where, It is the cascade channel from the user to the base station.

[0069] The pilot sequence of each user is set to an orthogonal pilot sequence, and the uplink signal from the kth user received by the base station in one subframe is:

[0070]

[0071] Where z b,k The base station receives the signal from the kth user in the bth subframe, combines the signals received in B subframes, and writes the received signal received by the base station in matrix form:

[0072] Z k =HΘ+U

[0073] Where, is the received signal at the base station after B subframes. is the reflection coefficient matrix of RIS after B subframes, It's noise.

[0074] S2 is the channel modeling step. Based on the millimeter wave communication model, channel modeling is performed for millimeter wave channels G and F, using the SV model. The antenna and RIS reflector units are deployed using UPAs. The array response vectors of the base station and RIS in azimuth and elevation are determined, respectively. This results in the RIS-to-base station channel model and the user-to-RIS channel model. The specific steps are as follows:

[0075] For a base station, when the azimuth and elevation angles of the received signal are θ and φ respectively, its array response vector is:

[0076]

[0077] in, is the array response vector of the base station in the horizontal direction, is the array response vector of the base station in the vertical direction, The array response vector of RIS is modeled in the same way.

[0078] According to the above formula, the channel model from RIS to base station is obtained, which is expressed as:

[0079]

[0080] Where, L g and are the number of effective paths and path gain from RIS to base station channels, and denote the azimuth and elevation angles at the base station, respectively. and denote the azimuth and elevation angles at RIS, respectively. and denote the array response vectors of the base station and RIS respectively.

[0081] The channel model from user to RIS is expressed as:

[0082]

[0083] Where, L f and are the number of effective paths and path gain of the channel from user to RIS, and represent the azimuth and elevation angles of the user respectively.

[0084] S3 is the pre-estimation channel matrix step. Based on the channel model from RIS to base station and the channel model from user to RIS, the base station uses the LMMSE estimator to pre-estimate the cascade channel matrix H according to the received user uplink signal to obtain the pre-estimated channel matrix. The specific method is:

[0085] The key to the LMMSE algorithm to estimate the channel is to find a matrix A such that Represents the pre-estimated channel matrix.

[0086] make represents the mean square error between the cascade channel matrix and the pre-estimated channel matrix, Substituting into the formula we get get:

[0087]

[0088] Where A is the coefficient matrix to be solved, H is the cascade channel from the user to the base station, and Z k is the received signal at the base station, tr represents the trace of the matrix, It means to find the mathematical expectation of a matrix.

[0089] The goal of the LMMSE algorithm is to find a matrix A to minimize the mean square error J. By taking the partial derivative of A and setting it equal to 0, the expression of the matrix A is:

[0090] A=(Θ H R H Θ+R N ) -1 Θ H R H

[0091] Where R H is the autocorrelation matrix of channel H, R N is the autocorrelation matrix of the noise, and the expression of the pre-estimated channel matrix is:

[0092]

[0093] Where, Represents the pre-estimated channel matrix.

[0094] Step S4 constructs the L-DA neural network model, further improving millimeter-wave channel estimation accuracy. The estimated channel matrix and the true channel matrix are combined into label pairs, and the data set consisting of multiple label pairs obtained through multiple measurements is format-aligned. The L-DA neural network is trained offline using the aligned data set. Through multiple rounds of iteration, the L-DA neural network model and its corresponding network weight parameters are obtained.

[0095] Figure 2 Schematic diagram of the L-DA neural network model in an embodiment of the present invention.

[0096] Figure 3 Schematic diagram of the SE-C-CBAM module in an embodiment of the present invention.

[0097] Specifically, if Figure 2 and Figure 3 As shown in Figure 2, a core attention mechanism of the L-DA neural network is the SE-C-CBAM module, which consists of the SE module and the C-CBAM module.

[0098] For the SE module: its core operation is a combination of global average pooling and fully connected layers. Assume that the input feature map is B, C, H, and W represent the batch number, channel number, height, and width of the input feature map, respectively. First, the feature map is processed using global average pooling to compress the spatial dimension to 1×1, so that the features of each channel are compressed into a single value, thereby obtaining the global information of each channel, which is expressed as:

[0099]

[0100] Where Y avg It is the result of average pooling of the input feature map X in the spatial dimension. The result is then passed through two fully connected layers in sequence, and the channel attention weight is generated by the Sigmoid function. The input is multiplied by the channel attention weight to obtain the output.

[0101] For the C-CBAM module: it contains two parts: channel attention and spatial attention. For channel attention, the input feature map is subjected to global average pooling and global maximum pooling to obtain two features. Global average pooling is similar to the SE module, and global maximum pooling is expressed as:

[0102]

[0103] Where Y max It is the result of taking the global maximum value of the input feature map in the spatial dimension. Different from the CBAM module which uses the fully connected layer to calculate the channel attention weight, in the channel attention mechanism of the C-CBAM module, the two features obtained by global average pooling and global maximum pooling are processed by the convolution layer with the same convolution kernel size of 1×1 respectively. Then the two outputs are added and the channel attention weight is generated by the Sigmoid activation function. The input is multiplied by the attention weight to obtain the output Y1 as the input of the spatial attention mechanism.

[0104] For the spatial attention mechanism, we first perform average pooling and maximum pooling on the input in the channel dimension to obtain two feature maps, which are expressed as:

[0105]

[0106] Where Y s-avg It is the result of averaging the input feature map in the channel dimension and retaining the spatial dimension, Y s-max It is the result of taking the maximum value of the input feature map in the channel dimension and retaining the spatial dimension. The two features are then concatenated in the channel dimension and processed through a convolution layer with a convolution kernel of 3×3 or 7×7. The processed result is passed through the Sigmoid activation function to generate the spatial attention weight. Finally, Y1 is multiplied by the attention weight to obtain the output Y2.

[0107] Step S5 is an output step in which the pre-estimated channel matrix is ​​estimated online using the L-DA neural network model and corresponding weight parameters, and the channel estimation matrix is ​​output.

[0108] Specifically, the received signal Z is input into the LMMSE estimator for pre-estimation to obtain the pre-estimated channel matrix Will Divided into two channels, real and imaginary, we get Expressed as:

[0109]

[0110] Where, yes The real part of yes The imaginary part of . is The dual-channel channel matrix composed of the real and imaginary parts of The branch above consists of five network blocks, each of which includes a convolutional layer, a batch normalization layer, and a Relu activation function layer. A skip link is added between each network block, meaning that the input of each block is the sum of the output and input of the previous block. This layer serves as a baseline branch, preserving the integrity of the original features and preventing information loss that may be caused by the attention mechanism.

[0111] The next branch also consists of five network blocks, each of which includes a convolutional layer, a batch normalization layer, a ReLu activation function layer, and a SE-C-CBAM module. A similar skip link is added between each network block. The main function of this layer is to enhance important features and suppress irrelevant features through channel and spatial attention mechanisms.

[0112] The outputs of the two branches are concatenated according to the channel dimension. The purpose of concatenation is to combine the features enhanced by the attention mechanism with the original features to enhance the diversity of features. Then, the network passes through two network blocks, each consisting of a convolutional layer, a batch normalization layer, and a ReLu activation function layer. Then, it passes through a multi-head attention mechanism module to enhance the ability to extract global information. Finally, it passes through a convolutional layer to restore the shape of the output noise and make it consistent with the input. In addition, a residual connection is added to the L-DA neural network model, that is, the final output channel estimation matrix is ​​the input minus the output noise.

[0113] The number of feature channels of the convolution layer of each block is set to 64, and the size of the convolution kernel is set to 3 × 3. In order to better learn the residual noise, the loss function of the neural network is set to the following formula:

[0114]

[0115] Where N is the output noise, is the input of the L-DA neural network, and H is the cascade channel matrix.

[0116] In this embodiment, the performance of the present invention is illustrated by simulation experiments based on MATLAB and Pytorch. The LMMSE algorithm pre-estimation is simulated by MATLAB, and then the generated data set is input into the L-DA neural network model.

[0117] The simulation of the L-DA neural network model in this embodiment is implemented based on Pytorch.

[0118] Table 1 Experimental parameter settings

[0119]

[0120] Table 1 shows the specific experimental parameter settings.

[0121] Table 2 Comparison of inference time of different algorithms

[0122]

[0123] Table 2 shows a comparison of the inference time of the L-DA neural network model proposed in this paper and other neural network models. The proposed L-DA neural network model has a shorter time complexity than the SC-Attention neural network model and higher than the CDRN and DnCNN neural network models. However, the L-DA neural network model offers the best estimation performance among these models, as demonstrated in the following simulation experiments.

[0124] Figure 4 2 is a performance comparison chart of various models in the embodiments of the present invention.

[0125] like Figure 4 As shown, M=16, N=32, L g =5 and L f When ∑ = 3, the channel estimation accuracy differences between the L-DA neural network model and models such as CDRN, DnCNN, and SC-Attention at different signal-to-noise ratios are shown. The figure clearly shows that the L-DA model achieves the highest estimation accuracy under different signal-to-noise ratio conditions. For example, at a low signal-to-noise ratio of 0 dB, the L-DA model improves estimation accuracy by 1.08 dB, 1.44 dB, and 1.69 dB compared to CDRN, DnCNN, and SC-Attention, respectively.

[0126] Figure 5 3 is a performance comparison chart of different models when the number of base station antennas is 8 in an embodiment of the present invention.

[0127] Figure 6 3 is a performance comparison chart of different models when the number of base station antennas is 24 in an embodiment of the present invention.

[0128] like Figure 5 and Figure 6 Figure 2 shows the estimation accuracy trends of different models as the signal-to-noise ratio changes when the number of base station antennas M is 8 and 24, respectively. The results show that the L-DA neural network model still has the highest channel estimation accuracy.

[0129] like Figure 4-Figure 6 The experimental results shown jointly verify that the L-DA neural network model can achieve the highest accuracy channel estimation under different numbers of antennas, highlighting its universality for the number of antennas.

[0130] Figure 7 This is a first result diagram of the impact of the number of reflection units and the number of effective paths on the performance of the model in an embodiment of the present invention.

[0131] Figure 8 This is a second result diagram of the impact of the number of reflection units and the number of effective paths on the performance of the model in an embodiment of the present invention.

[0132] like Figure 7 and Figure 8 As shown in Figure 2, the relationship between the channel estimation accuracy and signal-to-noise ratio of different models is shown under different numbers of RIS reflection units and the number of effective millimeter wave propagation paths. Figure 7 Middle: N=16, L g =3, L f =2 or Figure 8 Medium: N=24, L g =4, L f = 2, the L-DA neural network model can achieve the best channel estimation accuracy, which strongly proves the effectiveness and adaptability of the L-DA neural network model in this embodiment under different reflection unit configurations and different effective paths.

[0133] Figure 9 Graph showing the results of an ablation experiment in an embodiment of the present invention.

[0134] like Figure 9 The figure shows the performance changes of the L-DA neural network model when the SE-C-CBAM module, the multi-head attention mechanism, and all attention mechanisms are removed. It can be seen that eliminating any module reduces the channel estimation accuracy, thus demonstrating the significant impact of these modules on the performance of the L-DA neural network model and fully verifying the key role and effectiveness of these modules in the L-DA neural network model.

[0135] This embodiment also provides a RIS-assisted millimeter wave channel estimation system using an L-DA neural network, including:

[0136] The communication model construction module is used to implement step S1, namely: building a multi-user millimeter wave communication model, in which there is a base station, a RIS board and K users. The base station is equipped with M antennas, the RIS board is equipped with N reflection units, and each user is equipped with a single antenna.

[0137] The channel modeling module is used to implement step S2, namely, channel modeling based on the millimeter wave communication model, using the SV model. The antenna and RIS reflector units are deployed using UPAs. The array response vectors of the base station and RIS in azimuth and elevation are determined, respectively. This results in channel models from the RIS to the base station and from the user to the RIS.

[0138] The pre-estimated channel matrix module is used to implement step S3, that is, based on the channel model from RIS to base station and the channel model from user to RIS, the base station uses the LMMSE algorithm to pre-estimate the cascade channel matrix H according to the received user uplink signal to obtain the pre-estimated channel matrix.

[0139] The L-DA neural network model construction module is used to implement step S4, namely, forming label pairs from the estimated channel matrix and the true channel matrix, and aligning the data set consisting of multiple label pairs obtained through multiple measurements. The L-DA neural network is trained offline using the aligned data set. Through multiple rounds of iterations, the L-DA neural network model and the corresponding network weight parameters are obtained.

[0140] The output module is used to implement step S5, namely: use the L-DA neural network model and the corresponding weight parameters to perform online estimation on the pre-estimated channel matrix and output the channel estimation matrix.

[0141] Functions and Effects of the Embodiments

[0142] The RIS-assisted millimeter wave channel estimation method and system of the L-DA neural network involved in the present invention have the following beneficial effects:

[0143] 1. High-precision estimation. This invention organically combines the preliminary estimation advantages of the LMMSE algorithm with the powerful feature learning capabilities of neural networks, significantly improving the accuracy of millimeter-wave channel estimation. Under different signal-to-noise ratio conditions, the L-DA neural network proposed in this invention can more accurately estimate the channel than existing algorithms such as CDRN, DnCNN, and SC-Attention, achieving a lower channel normalized mean square error (NMSE), providing a strong guarantee for achieving reliable millimeter-wave communications.

[0144] 2. Strong universality. The L-DA neural network proposed in this paper demonstrates more accurate channel estimation effects for different numbers of base station antennas or RIS reflection units, demonstrating the universality of the proposed scheme for different system parameters.

[0145] 3. Module Effectiveness: By designing and implementing ablation experiments, we thoroughly validated the effectiveness of key modules within the L-DA neural network model. The experimental results demonstrate that the SE-C-CBAM module and multi-head attention mechanism play a significant role in improving algorithm performance. This further validates the scientific and rational design of our model and provides solid theoretical support for subsequent technical improvements and optimizations.

[0146] Those skilled in the art will appreciate that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A RIS-assisted millimeter wave channel estimation method combining the LMMSE algorithm with a dual attention mechanism neural network is used to estimate the channel, characterized by: The specific steps include: S1: Communication model construction step, building a multi-user millimeter wave communication model, which includes a base station, a RIS board and K users. The base station is equipped with M antennas, the RIS board is equipped with N reflection units, and each user is equipped with a single antenna; S2: a channel modeling step, performing channel modeling according to the millimeter wave communication model, wherein the channel adopts the Saleh-Valenzuela (SV) model, and the reflection units of the antenna and the RIS are deployed using a uniform planar array (UPA). The array response vectors of the base station and the RIS in the azimuth and elevation directions are determined respectively, thereby obtaining a channel model from the RIS to the base station and a channel model from the user to the RIS; S3: Pre-estimating the channel matrix step, based on the channel model from the RIS to the base station and the channel model from the user to the RIS, the base station uses an LMMSE estimator to pre-estimate the cascade channel matrix H according to the received uplink signal of the user to obtain a pre-estimated channel matrix; S4: an L-DA neural network model construction step, wherein the pre-estimated channel matrix and the true channel matrix form a label pair, and a data set consisting of a plurality of label pairs obtained through multiple measurements is format-aligned, and the L-DA neural network is trained offline using the data set after the format alignment, and the L-DA neural network model and the corresponding weight parameters in the network are obtained through multiple rounds of iterations; S5: Output step, using the L-DA neural network model and the corresponding weight parameters to perform online estimation on the pre-estimated channel matrix, and output the channel estimation matrix.

2. The RIS-assisted millimeter wave channel estimation method combining the LMMSE algorithm and the dual attention mechanism neural network according to claim 1 is characterized by: in, In S1, a multi-user millimeter wave communication model is built, specifically: Assume that the channel from RIS to the base station is The channel from user to RIS is The reflection coefficient of RIS is The communication protocol adopts a frame structure protocol. A subframe contains T time slots. Each user sends a pilot sequence in the T time slots of each subframe. Within the same subframe, the reflection coefficient of RIS remains unchanged. Between different subframes, the reflection coefficient of RIS is different. The user end uses an orthogonal pilot sequence. T ≥ K, The received signal expression of the base station in each subframe is: Where, is noise, with mean 0 and variance σ 2 The complex Gaussian distribution of Where I is the identity matrix, φ b is the RIS reflection coefficient of the b-th subframe, diag(x) represents the diagonal matrix with vector x as the diagonal, since The uplink signals from all users received by the base station in one subframe can be expressed as: Where, is the cascade channel from the user to the base station, The pilot sequence of each user is set to an orthogonal pilot sequence, and the uplink signal from the kth user received by the base station in one subframe is: Where z b,k The base station receives the signal from the kth user in the bth subframe, combines the signals received in B subframes, and writes the received signal received by the base station in matrix form: Z k =HΘ+U Where, is the received signal at the base station after B subframes. is the reflection coefficient matrix of RIS after B subframes, It's noise.

3. The RIS-assisted millimeter wave channel estimation method combining the LMMSE algorithm and the dual attention mechanism neural network according to claim 1 is characterized by: in, In S2, the channel model from RIS to base station is expressed as: Where, L g and are the number of effective paths and path gain from RIS to base station channels, and denote the azimuth and elevation angles at the base station, respectively. and denote the azimuth and elevation angles at RIS, respectively. and denote the array response vectors of the base station and RIS respectively, The channel model from user to RIS is expressed as: Where, L f and are the number of effective paths and path gain of the channel from user to RIS, and represent the azimuth and elevation angles of the user respectively.

4. The RIS-assisted millimeter wave channel estimation method combining the LMMSE algorithm and the dual attention mechanism neural network according to claim 1 is characterized by: in, In S3, the method of using the LMMSE estimator to pre-estimate the cascade channel matrix H to obtain the pre-estimated channel matrix is ​​specifically as follows: Find a matrix A such that Denotes the pre-estimated channel matrix, let represents the mean square error between the cascade channel matrix and the pre-estimated channel matrix, Substituting into the formula we get Right now: Where A is the coefficient matrix to be solved, H is the cascade channel from the user to the base station, and Z k is the received signal at the base station, tr represents the trace of the matrix, It means to find the mathematical expectation of the matrix. By finding the partial derivative of A and setting it equal to 0, the expression of matrix A is: A=(Θ H R H I+R N ) -1 I H R H Where R H is the autocorrelation matrix of channel H, R N is the autocorrelation matrix of the noise, and the expression of the pre-estimated channel matrix is: Where, Represents the pre-estimated channel matrix.

5. The RIS-assisted millimeter wave channel estimation method combining the LMMSE algorithm and the dual attention mechanism neural network according to claim 4 is characterized in that: in, In S5, the method for online estimating the pre-estimated channel matrix using the L-DA neural network model and the corresponding weight parameters is specifically as follows: The L-DA neural network model is composed of two branches. Divided into two channels, real and imaginary, we get Will As the input of the two branches, the above branch consists of 5 network blocks. Each network block contains a convolution layer, a batch normalization layer and a ReLu activation function layer, and a skip link is added between each network block, that is, the input of each block is obtained by adding the output of the previous block to the input of the previous block. The next branch also consists of 5 network blocks. Each network block contains a convolution layer, a batch normalization layer, a ReLu activation function layer and a SE-C-CBAM module, and a skip link is added between each network block. The outputs of the two branches are spliced ​​according to the channel dimension, and then pass through two network blocks. Each network block consists of a convolution layer, a batch normalization layer, and a ReLu activation function layer. Then it passes through a multi-head attention mechanism module, and finally passes through a convolution layer to restore the shape of the output noise so that it is consistent with the input. Keep the shape consistent.

6. The RIS-assisted millimeter wave channel estimation method combining the LMMSE algorithm and the dual attention mechanism neural network according to claim 5 is characterized by: in, described Expressed as: Where, yes The real part of yes The imaginary part of .

7. The RIS-assisted millimeter wave channel estimation method combining the LMMSE algorithm and the dual attention mechanism neural network according to claim 5 is characterized by: in, A residual connection is also added to the L-DA neural network model, that is, the final output channel estimation matrix is ​​the input minus the output noise.

8. The RIS-assisted millimeter wave channel estimation system combines the LMMSE algorithm with a dual-attention mechanism neural network, characterized by: include: The communication model construction module builds a multi-user millimeter wave communication model, which includes a base station, a RIS board, and K users. The base station is equipped with M antennas, the RIS board is equipped with N reflection units, and each user is equipped with a single antenna. a channel modeling module, performing channel modeling according to the millimeter wave communication model, wherein the channel adopts an SV model, and the reflection units of the antenna and RIS adopt UPA deployment, and determining the array response vectors of the base station and the RIS in the azimuth and elevation directions, respectively, to obtain a channel model from the RIS to the base station and a channel model from the user to the RIS; A pre-estimation channel matrix module, based on the channel model from the RIS to the base station and the channel model from the user to the RIS, wherein the base station pre-estimates the cascade channel matrix H using the LMMSE algorithm according to the received user uplink signal to obtain a pre-estimated channel matrix; An L-DA neural network model construction module, which forms label pairs from the pre-estimated channel matrix and the true channel matrix, aligns the format of a data set consisting of a plurality of label pairs obtained through multiple measurements, and uses the aligned data set to perform offline training on the L-DA neural network, obtaining the L-DA neural network model and corresponding weight parameters in the network through multiple rounds of iterations; An output module uses the L-DA neural network model and the corresponding weight parameters to perform online estimation on the pre-estimated channel matrix and output a channel estimation matrix.

Images