Feature channel and spatial domain double-attention driven super-large scale MIMO channel estimation method

Abstract

The invention discloses a characteristic channel and spatial domain double-attention driven super-large-scale MIMO channel estimation method, which comprises the following steps: modeling channel characteristics in a super-large-scale MIMO system, constructing a channel matrix, and preprocessing signals received by a base station; a deep convolutional neural network model is constructed based on a convolutional network and an encoder, the convolutional network extracts local features of a channel matrix through a plurality of convolutional layers, convergence of the network is accelerated by using a batch normalization layer, and the stability of the network is improved through residual connection and layer normalization; the encoder is integrated with a feature channel and space channel double-attention mechanism, attention weights of different information are obtained, and extraction and utilization of important features in the attention weights are enhanced; training and testing the deep convolutional neural network model; and inputting the noisy channel matrix into the deep convolutional neural network model, and outputting an estimated channel matrix. The method does not depend on priori knowledge, the feature extraction and representation capability is enhanced, and the accuracy and stability of channel estimation are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to a very large-scale MIMO channel estimation method driven by dual attention of feature channels and spatial domains. Background Art

[0002] Extra-Large Multiple Input Multiple Output (XL-MIMO) technology is an innovative technology for the sixth generation of wireless communications. XL-MIMO significantly improves communication capacity and reliability by deploying a large number of antennas at base stations. Thanks to the wide range of antennas, it performs well in spectrum efficiency, energy efficiency, and interference mitigation. In XL-MIMO, the base station needs to obtain channel state information from a large number of antennas, resulting in a high-dimensional channel matrix. Compared with traditional massive MIMO systems, the channel matrix in XL-MIMO not only has a higher dimension, but also has more tricky characteristics. When dealing with large-scale antenna systems, the computational load increases dramatically, requiring efficient algorithms and powerful computing resources for channel estimation. In the XL-MIMO system, due to the large antenna array, the near-field effect is more obvious, which affects the estimation accuracy.

[0003] Traditional channel estimation methods usually rely on linear models and statistical assumptions. The least squares (LS) based scheme can only achieve low received signal-to-noise ratio and lead to considerable pilot overhead. In order to reduce the high pilot overhead in channel estimation, in current massive MIMO systems, some compressed sensing based methods, such as orthogonal matching pursuit and sparse Bayesian learning, estimate high-dimensional channels with low pilot overhead by exploiting the channel sparsity in the angle domain. However, these methods require prior knowledge and usually have higher complexity when dealing with high-dimensional channel matrices. Summary of the invention

[0004] In view of the deficiencies of the above-mentioned prior art, the present invention provides a very large-scale MIMO channel estimation method driven by dual attention of feature channels and spatial domains, and proposes a neural network (MAT-CENet) for very large-scale MIMO channel estimation driven by dual attention of feature channels and spatial domains. The method is suitable for mixed field users with different channels and does not rely on prior knowledge. The method combines convolutional neural networks and Transformer encoder structures to enhance feature extraction and representation capabilities, significantly improve the accuracy and stability of channel estimation, and can solve the complexity of high-dimensional channel estimation and the problems of mixed fields in existing XL-MIMO systems.

[0005] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0006] A method for ultra-large-scale MIMO channel estimation driven by dual attention in feature channels and spatial domains, characterized by comprising:

[0007] Step 1: Model the channel characteristics in the ultra-large-scale MIMO system, construct a channel matrix, and pre-process the signals received by the base station to obtain a training data set and a test data set;

[0008] Step 2: A deep convolutional neural network model is constructed based on the convolutional network and the encoder. The convolutional network extracts local features of the channel matrix through multiple convolutional layers, uses batch normalization layers to accelerate network convergence, and improves network stability through residual connections and layer normalization. The encoder incorporates a dual attention mechanism of feature channels and spatial channels to obtain attention weights for different information, thereby enhancing the extraction and utilization of important features.

[0009] Step 3: Use the training data set and the test data set to train and test the deep convolutional neural network model;

[0010] Step 4: Input the noisy channel matrix into the deep convolutional neural network model obtained in step 3, and output the estimated channel matrix.

[0011] To optimize the above technical solutions, the specific measures taken also include:

[0012] The above step 1 includes the following sub-steps:

[0013] Step 1.1: Model the ultra-large-scale MIMO system. Assume that the base station of the XL-MIMO system with M antennas serves a single-antenna user, and there are L paths between the base station and the user. Then the channel characteristic model in the ultra-large-scale MIMO system is:

[0014] y=hx+η

[0015] Where y is the uplink received signal received by the base station, h is the channel vector between the base station and the user, x represents the pilot signal transmitted by the user in a time slot, and η represents additive Gaussian noise;

[0016] Step 1.2: Construct the channel matrix:

[0017] Very large-scale antenna arrays can operate in both far-field and near-field conditions, which are determined by the Rayleigh distance D Ray Sure:

[0018]

[0019] Among them, D a is the array aperture, λ is the carrier wavelength;

[0020] For a uniform linear array, if the spacing between adjacent antennas is Rayleigh distance D Ray It is expressed as:

[0021]

[0022] When the scattering distance is greater than D Ray When , the far-field channel model is:

[0023]

[0024] in, is the array response vector of the base station antenna array to the lth path, represents the azimuth of the lth path, g l represents the gain of the lth path;

[0025] When the scattering distance is less than D Ray When , the near-field channel model is:

[0026]

[0027] in represents the array response vector of the base station antenna array to the lth path, φ l represents the azimuth of the lth path, r l represents the distance from the lth scatterer to the center of the base station antenna array, represents the distance between the lth scatterer and the mth antenna of the base station, where

[0028] m=1,...,M;

[0029] Therefore, the channel matrix of the mixed field is:

[0030]

[0031] From the above formula, we can see that the channel matrix of the mixed field contains L0 far-field paths and L-L0 near-field paths;

[0032] Step 1.3: Pre-process the signal received by the base station:

[0033] Knowing that the power of x is P, simplify y to

[0034]

[0035] Based on the above formula, the least squares estimation complex matrix of h is:

[0036]

[0037] Step 1.4: Generate training data sets and test data sets according to the contents of steps 1.2 and 1.3, generate a noise-free real channel matrix according to the number of near- and far-field paths and the mixed-field channel model, add noise to the noise-free real channel matrix to generate a noisy channel matrix, save the noisy channel matrix as input data and the noise-free real channel matrix as output labels, and ensure that the data sets have sufficient diversity to cover different channel conditions. The noisy channel matrix is ​​estimated by the proposed network model to obtain the estimated channel matrix and the noise-free real channel matrix Comparisons were made to verify the estimated effects.

[0038] The input of the above deep convolutional neural network model is the tensor H input , the tensor H input The method of obtaining is: the least squares estimate complex matrix of the channel matrix h Split into real and imaginary parts of equal dimension, which are then concatenated and reshaped into one The tensor H input .

[0039] The above-mentioned deep convolutional neural network model includes four two-dimensional convolutional layers with a convolution kernel size of 3×3 and 64 convolution kernels, two encoders, one two-dimensional convolutional layer with a convolution kernel size of 3×3 and 2 convolution kernels, layer normalization and residual connection modules.

[0040] For the first four two-dimensional convolutional layers with a kernel size of 3×3 and 64 kernels, a batch normalization layer is added after each convolutional layer to accelerate the network training process. By normalizing small batches of data, the input of each layer is kept stable to prevent gradient vanishing and gradient exploding problems. A rectified linear unit activation function is used on the output of the convolutional layer to introduce nonlinear features and make the network more expressive. A Dropout layer is introduced between convolutional layers to randomly inactivate a certain proportion of neurons to prevent network overfitting.

[0041] The above four convolution layers with a convolution kernel size of 3×3 and a convolution kernel number of 64 are applied to the tensor H input After extracting the features, reshape them into a tensor H1 suitable for encoder input, with a size of M×64, where M is the number of antennas of the base station in the XL-MIMO system;

[0042] The H1 is input into two encoders in sequence, and finally the encoder output tensor H5 is reshaped into The tensor H6 is passed through a two-dimensional convolution layer with a convolution kernel size of 3×3 and a convolution kernel number of 2 to obtain H7;

[0043] Input H7 to the residual connection module, and use the residual connection to input Hinput Added to H7, the output is H output , improve training stability.

[0044] The above encoder includes a feature channel attention mechanism module, a spatial channel attention mechanism module, two fully connected layers, a Dropout layer and a normalization layer in sequence;

[0045] The feature channel attention mechanism module directly applies the multi-head attention mechanism to generate a reconstruction matrix with the same dimension as the input tensor H1. Finally, the reconstruction matrix obtained by the attention mechanism is summed element by element with H1 to obtain the final output tensor H2 with a size of M×64. H2 represents the weighted sum of all high-dimensional feature maps.

[0046] The spatial channel attention mechanism module adopts a multi-head attention mechanism to transpose the input tensor H2 to exchange dimensions, and finally transposes the attention output matrix again and sums it with H2 one by one to obtain the output H3 with a size of M×64. H3 is the weighted sum of all position features and the original features of each position.

[0047] The above multi-head attention mechanism calculates the dependencies between different channels. If X = [x1, x2, L, x n ] represents n input information, Q, K, V are the matrices of query, key and value respectively, then the attention weight formula is

[0048] Q=W q X, K = W k X, V = W v X,

[0049]

[0050] Among them, W q , W k , W v represents the corresponding weight matrix of Q, K, V, d k is the dimension of vector V;

[0051] The subspace in the multi-head attention mechanism is generated as follows:

[0052] Q i =QW i Q , K i =KW i K , V i =VW i V ,

[0053] The above formulas Q, K, and V are transformed by their respective linear transformation matrices W iQ , W i K , W i V Represents the query Q i , key K i Sum value V i , where W i Q , W i K , W i V Represents the weight matrix group corresponding to different "heads" in the multi-head attention mechanism;

[0054] Then the formula of the multi-head attention mechanism is:

[0055] head i =Attention(Q i ,K i ,V i ),i=1,...,h,

[0056] Among them, h represents the number of "heads", head i represents the attention matrix obtained by the i-th "head", Attention(Q i ,K i ,V i ) is the attention mechanism formula corresponding to the i-th “head”.

[0057] The loss function used in step 3 above is:

[0058]

[0059] in, is the estimated channel matrix, the real channel matrix, N tr Represents the number of training samples in the training data set, and the superscript (n) represents the nth sample. It means taking the Frobenius norm of the matrix and then squaring it.

[0060] The above step 3 uses the Adam optimizer to optimize the network. The Adam optimizer combines momentum and adaptive learning rate technology to accelerate convergence. Its update rules are as follows:

[0061]

[0062]

[0063]

[0064]

[0065]

[0066] Among them, m t is the momentum vector at the tth iteration, representing the weighted average of the gradient; v t is the squared gradient momentum at the tth iteration, which represents the weighted average of the squared gradient and is the attenuation coefficient of the momentum term; since in the initial stage of training, m t and v t The initial values ​​are all 0, and deviation correction is required. and is the corrected m t and v t ; β1 is the attenuation coefficient of the momentum term; β2 is the attenuation coefficient of the squared gradient momentum; is the gradient of the loss function L(θ) with respect to the model parameter θ; α is the learning rate; θ is the parameter vector of the model; ∈ is a constant.

[0067] In the test phase of step 3 above, the noisy channel matrix is ​​input into the trained deep convolutional neural network model, and the model outputs the estimated channel matrix through forward propagation. Calculate the estimated channel matrix With the real channel matrix The normalized mean square error between:

[0068]

[0069] For expectation.

[0070] The accuracy of the model in channel recovery is measured by the normalized mean square error. The smaller the normalized mean square error, the more accurate the channel estimation result.

[0071] The present invention has the following beneficial effects:

[0072] (1) The present invention utilizes the feature channel and spatial domain dual attention mechanism to effectively extract local features in the channel matrix, and optimizes the model through convolutional layers and residual connections, so that the channel matrix can be restored more accurately. Even under complex channel conditions, the estimation error can be kept low, which significantly improves the reliability of channel estimation.

[0073] (2) The modeling method of the present invention can adapt to different channel environments and can handle mixed fields combining far-field and near-field paths. Through this flexible modeling method, high estimation performance is maintained under different channel conditions without frequent adjustment of the model or algorithm. This adaptability enables the present invention to be widely used in various practical communication scenarios, improving the versatility and flexibility of the system.

[0074] (3) The present invention effectively reduces computational complexity. The deep convolutional network significantly improves computational efficiency through its parallel computing capability and efficient feature extraction mechanism. Compared with traditional methods, the present invention uses batch normalization, residual connection, and dropout techniques in the training process, which not only improve the training speed but also optimize the computational overhead of the model. These optimization measures combined enable the present invention to achieve fast and efficient computational processing while ensuring high accuracy, which is suitable for efficient deployment in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is a neural network model diagram of the ultra-large-scale MIMO channel estimation method based on feature channel and spatial domain dual attention drive of the present invention;

[0076] Figure 2 is a curve showing the change of the normalized mean square error value performance with the signal-to-noise ratio under different near-field conditions in the present invention;

[0077] Figure 3 is a curve of normalized mean square error performance changing with signal-to-noise ratio under different far-field conditions in the present invention;

[0078] Figure 4 It is a curve showing the change of the normalized mean square error performance of various schemes with the number of far-field paths in the mixed field under different conditions in the present invention. DETAILED DESCRIPTION

[0079] The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0080] like Figure 1 This is a neural network model diagram of a very large-scale MIMO channel estimation method based on feature channel and spatial domain dual attention drive of the present invention. The channel estimation method specifically includes the following steps:

[0081] Step 1: Model the channel characteristics in the ultra-large-scale MIMO system, construct a channel matrix, and pre-process the signals received by the base station to obtain a training data set and a test data set;

[0082] The specific steps include the following:

[0083] Step 1.1: Model the ultra-large-scale MIMO system. Assume that the base station of the XL-MIMO system with M antennas serves a single-antenna user and there are L paths between the base station and the user. The uplink received signal y received by the base station can be expressed as

[0084] y=hx+η

[0085] Where h is the channel vector between the base station and the user, x represents the pilot signal transmitted by the user in a time slot, and η represents additive Gaussian noise.

[0086] Step 1.2: The ultra-large-scale antenna array can operate in far-field and near-field conditions, which are determined by the Rayleigh distance D. Ray Sure

[0087]

[0088] Among them, D a is the array aperture, λ is the carrier wavelength, for a uniform linear array, if the spacing between adjacent antennas is Rayleigh distance D Ray It can be expressed as

[0089]

[0090] When the scattering distance is greater than D Ray When , the channel model is expressed using the far-field formula as

[0091]

[0092] in, is the array response vector of the base station antenna array to the lth path, represents the azimuth of the lth path, g l represents the gain of the lth path.

[0093] When the scattering distance is less than D Ray The channel is modeled using the near-field formulation when

[0094]

[0095] in represents the array response vector of the base station antenna array to the lth path, g l represents the gain of the lth path, φ l represents the azimuth of the lth path, r l represents the distance from the lth scatterer to the center of the base station antenna array, represents the distance between the lth scatterer and the mth antenna of the base station, where m=1,…,M.

[0096] Therefore, the mixed field channel model can be expressed as

[0097]

[0098] It can be seen from the above formula that the mixed field model contains L0 far-field paths and L-L0 near-field paths.

[0099] Step 1.3: Preprocess the data. x is a pilot signal with known power. Assume its power is P. Simplify y to

[0100]

[0101] Based on the above formula, the least squares estimate of h is expressed as

[0102]

[0103] Step 1.4: Generate training data sets and test data sets according to the contents of steps 1.2 and 1.3, generate a noise-free real channel matrix according to the number of near- and far-field paths and the mixed-field channel model, add noise to the noise-free real channel matrix to generate a noisy channel matrix, save the noisy channel matrix as input data and the noise-free real channel matrix as output labels, and ensure that the data sets have sufficient diversity to cover different channel conditions. The noisy channel matrix is ​​estimated by the proposed network model to obtain the estimated channel matrix and the noise-free real channel matrix Comparisons were made to verify the estimated effects.

[0104] Step 2: A deep convolutional neural network model is constructed based on the convolutional network and the encoder. The convolutional network extracts local features of the channel matrix through multiple convolutional layers, uses batch normalization layers to accelerate network convergence, and improves network stability through residual connections and layer normalization. The encoder incorporates a dual attention mechanism of feature channels and spatial channels to obtain attention weights for different information, thereby enhancing the extraction and utilization of important features.

[0105] (1) Construct a convolutional network structure, extract local features of the channel matrix through multiple convolutional layers, use batch normalization layers to accelerate network convergence, and improve network stability through residual connections and layer normalization;

[0106] The details are as follows:

[0107] The deep convolutional neural network model is MAT-CENet, which converts the complex matrix Split into real and imaginary parts of equal dimension, which are then concatenated and reshaped into one The tensor H input MAT-CENet uses four two-dimensional convolutional layers with a convolution kernel size of 3×3 and 64 convolution kernels to perform a multi-dimensional convolution on the tensor H. inputAfter extracting the features, reshape it into a tensor H1 suitable for encoder input, with a size of M×64. Add a batch normalization layer after each convolution layer to accelerate the network training process. By normalizing small batches of data, the input of each layer remains stable to prevent gradient vanishing and gradient explosion problems. Use the rectified linear unit (Relu) activation function on the output of the convolution layer to introduce nonlinear features and make the model more expressive. Introduce a Dropout layer between convolutional layers to randomly inactivate a certain proportion of neurons to prevent overfitting of the model. H1 is input to the two encoders in turn, and finally the tensor H5 output from the encoder is reshaped into The tensor H6 is passed through a 2D convolution layer with a convolution kernel size of 3×3 and 2 convolution kernels, and then layer normalization and residual connection are used to improve the stability of the network. In order to avoid the gradient disappearance problem, the residual connection is used to convert the input H input Directly added to the output. At the same time, use layer normalization to normalize the output of each layer to improve training stability.

[0108] (2) Design the encoder structure, integrate the dual attention mechanism of feature channel and spatial domain, obtain the importance score matrix of different information, and enhance the extraction and utilization of important features;

[0109] The details are as follows:

[0110] The encoder structure of the Transformer in the attention mechanism is applied in the system model network. The multi-head attention mechanism is used in the encoder structure. For different channels, the attention mechanism is used to calculate the dependencies between different channels. n ] represents n input information, Q, K, V are matrices of query, key and value respectively, and the attention weight formula can be expressed as

[0111] Q=W q X, K = W k X, V = W v X,

[0112]

[0113] Among them, W q , W k , W v represents the corresponding weight matrix of Q, K, V, d k is the dimension of the vector V.

[0114] In the feature channel attention module and spatial domain attention module of the system, a multi-head attention mechanism is adopted. Instead of executing a single attention mechanism, multi-head attention runs multiple attention mechanisms in parallel, called "heads". The generation of subspaces can be expressed as

[0115] Q i =QW i Q , K i =KW i K , V i =VW i V ,

[0116] The above formulas Q, K, and V are transformed by their respective linear transformation matrices W i Q , W i K , W i V Represents the query Q i , key K i Sum value V i ,

[0117] Among them, W i Q , W i K , W i V represents multiple weight matrix groups corresponding to different heads in the multi-head attention mechanism; furthermore,

[0118] The formula of the multi-head attention mechanism is:

[0119] head i =Attention(Q i ,K i ,V i ),i=1,...,h,

[0120] Among them, h represents the number of "heads", head i represents the attention matrix obtained by the i-th "head", Attention(Q i ,K i ,V i ) is the attention mechanism formula corresponding to the i-th “head”.

[0121] The multi-head attention mechanism is applied in the feature channel and the spatial domain in turn. In the feature channel, the multi-head attention mechanism is directly applied to generate a matrix with the same dimension as H1. Finally, the reconstructed matrix obtained by the attention mechanism is summed element-wise with H1 to obtain the final output tensor H2 of size M×64. H2 represents the weighted sum of all these high-dimensional feature maps. After using the multi-feature channel attention mechanism, the feature map is processed in the spatial domain to capture the spatial correlation of the channel matrix. Unlike the feature map attention mechanism, the spatial domain attention module cannot directly use the input matrix. In order to extract the spatial feature weight matrix, the input tensor H2 needs to be transposed to exchange the dimensions. Finally, the attention output matrix is ​​transposed again and summed with H2 to obtain the output H3 of size M×64. H3 is the weighted sum of all position features and the original features at each position.

[0122] Step 3: Use the training data set and the test data set to train and test the deep convolutional neural network model;

[0123] (1) Train the entire network including the convolutional layer and the dual attention encoder until the model converges;

[0124] The specific steps include the following:

[0125] Step 3.1, use mean squared error (MSE) as the loss function. The estimated channel matrix is The real channel matrix is The mean square error loss function is defined as

[0126]

[0127] Among them, N tr represents the number of training samples, the superscript (n) represents the nth sample, It means taking the Frobenius norm of the matrix and then squaring it.

[0128] Step 3.2: Use the Adam optimizer to optimize the network. The Adam optimizer combines momentum and adaptive learning rate techniques to converge quickly within a small number of iterations. The update rules are as follows:

[0129] m t =β1m t-1 +(1-β1)▽ θ L(θ),

[0130]

[0131]

[0132]

[0133]

[0134] Among them, m t is the momentum vector at the tth iteration, representing the weighted average of the gradient, v t is the squared gradient momentum at the tth iteration, which represents the weighted average of the squared gradient and is the attenuation coefficient of the momentum term. Since in the initial stage of training, m t and v t The initial values ​​are all 0, and deviation correction is required. and is the corrected m t and v t ; β1 is the attenuation coefficient of the momentum term, β2 is the attenuation coefficient of the squared gradient momentum, is the gradient of the loss function L(θ) with respect to the model parameters θ, α is the learning rate, θ is the parameter vector of the model, and ∈ is a small constant used to avoid division by zero.

[0135] Step 3.3: Update the network weights and bias terms by continuous iteration until the loss function converges and the training process terminates.

[0136] (2) In the testing phase, the noisy channel matrix is ​​input into the trained deep neural network model, and the estimated channel matrix is ​​output to evaluate the channel recovery effect of the model;

[0137] The details are as follows:

[0138] In the test phase, the noisy channel matrix is ​​input into the trained deep neural network. The model estimates the channel matrix through forward propagation. Calculate the estimated channel matrix With the real channel matrix Normalized Mean Squared Error (NMSE) between

[0139]

[0140] This indicator is used to measure the accuracy of the model in channel recovery. The smaller the normalized mean square error, the more accurate the channel estimation result.

[0141] Step 4: Input the noisy channel matrix into the deep convolutional neural network model obtained in step 3, and output the estimated channel matrix.

[0142] The present invention performs channel estimation for ultra-large-scale MIMO systems based on dual attention drive of feature channels and spatial domains, and adopts a neural network including convolutional layers and dual attention encoders to achieve high-precision channel estimation under various channel conditions. The local features in the channel matrix are extracted through a deep convolutional network, and the dual attention mechanism of feature channels and spatial domains is combined to effectively improve the accuracy and robustness of channel estimation, overcoming the limitations of traditional methods under complex channel conditions. The present invention not only performs well in accuracy, flexibility and efficiency, but is also suitable for dynamically changing channel environments, and has broad application prospects and practical value.

[0143] Example 1

[0144] Set the simulation scenario parameters: number of base station antennas M = 256, wavelength λ = 0.01 m, average path gain σ 2 =1, r l ~U(10,80) meters. In the simulation, the training set has 9000 samples and the validation set has 1000 samples, all of which are data generated by a mixed field channel model, where the mixed field contains 1 far-field path and 5 near-field paths. The test set has 2000 samples, corresponding to near-field conditions and far-field conditions, respectively. Adam is used as the optimizer. Specifically, the method proposed in the present invention is compared with the traditional LS estimation method, LMMSE estimation method, HY-OMP and XLCNet.

[0145] Figure 2 The normalized mean square error performance of near-field users versus signal-to-noise ratio is depicted, with two sub-graphs corresponding to 3 and 6 near-field paths, respectively. Compared with other methods, MAT-CENet shows excellent performance in near-field conditions, significantly reducing the estimation error, especially in low signal-to-noise ratio conditions, showing higher accuracy. Despite the increase in the number of paths in Figure (b), the estimation accuracy is still high.

[0146] Figure 3 The normalized mean square error performance versus signal-to-noise ratio for far-field users is shown, with the two sub-graphs corresponding to three and six far-field paths, respectively. Figure 3 This shows its effectiveness in extracting channel features in multi-far-field path scenarios. Although the overall estimation error increases with the number of paths, MAT-CENet still maintains its advantage.

[0147] Example 2

[0148] Set the simulation scenario parameters: number of base station antennas M = 256, wavelength λ = 0.01 m, average path gain σ 2 =1, r l~U(10,80) meters. In the simulation, the training set has 9000 samples and the validation set has 1000 samples, all of which are data generated by a mixed field channel model. The mixed field contains 1 far-field path and 5 near-field paths. The test set has 2000 samples, corresponding to different mixed field conditions. Adam is used as the optimizer. Specifically, the method proposed in the present invention is compared with the traditional LS estimation method, LMMSE estimation method, HY-OMP and XLCNet.

[0149] In order to better demonstrate the performance of the method proposed in the present invention, Figure 4 The normalized mean square error performance of various schemes is shown as the number of far-field paths increases, with a signal-to-noise ratio of 10dB and a total of 6 paths in the mixed field. As shown in the figure, regardless of the number of far-field paths, MAT-CENet always outperforms other schemes in the mixed-field scenario. In addition, since the training dataset was generated using a channel model with 1 far-field path and 5 near-field paths, the network achieved the best performance in this specific channel estimation task. However, MAT-CENet also showed stable performance in other mixed-field channel estimations, highlighting the strong generalization ability of the invention.

[0150] The ultra-large-scale MIMO channel estimation method based on the dual attention driven by feature channels and spatial domains in the present invention applies neural networks to the channel estimation of ultra-large-scale MIMO. It does not rely on channel statistical knowledge and has not only excellent performance, but also high sensitivity and accuracy. It also has strong generalization ability, which is specifically reflected in the robustness to mixed scenarios. In mixed scenarios, regardless of the number of far-field paths, MAT-CENet is always better than other schemes.

[0151] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A very large-scale MIMO channel estimation method driven by dual attention in feature channels and spatial domains, characterized in that: include: Step 1: Model the channel characteristics in the ultra-large-scale MIMO system, construct a channel matrix, and pre-process the signals received by the base station to obtain a training data set and a test data set; Step 2: A deep convolutional neural network model is constructed based on the convolutional network and the encoder. The convolutional network extracts local features of the channel matrix through multiple convolutional layers, uses batch normalization layers to accelerate network convergence, and improves network stability through residual connections and layer normalization. The encoder incorporates a dual attention mechanism of feature channels and spatial channels to obtain attention weights for different information, thereby enhancing the extraction and utilization of important features. Step 3: Use the training data set and the test data set to train and test the deep convolutional neural network model; Step 4: Input the noisy channel matrix into the deep convolutional neural network model obtained in step 3, and output the estimated channel matrix.

2. According to claim 1, a feature channel and spatial domain dual attention driven ultra-large-scale MIMO channel estimation method is characterized in that: The step 1 includes the following sub-steps: Step 1.1: Model the ultra-large-scale MIMO system. Assume that the base station of the XL-MIMO system with M antennas serves a single-antenna user, and there are L paths between the base station and the user. Then the channel characteristic model in the ultra-large-scale MIMO system is: y=hx+η Where y is the uplink received signal received by the base station, h is the channel vector between the base station and the user, x represents the pilot signal transmitted by the user in a time slot, and η represents additive Gaussian noise; Step 1.2: Construct the channel matrix: Very large-scale antenna arrays can operate in both far-field and near-field conditions, which are determined by the Rayleigh distance D Ray Sure: Among them, D a is the array aperture, λ is the carrier wavelength; For a uniform linear array, if the spacing between adjacent antennas is Rayleigh distance D Ray It is expressed as: When the scattering distance is greater than D Ray When , the far-field channel model is: in, is the array response vector of the base station antenna array to the lth path, represents the azimuth of the lth path, g l represents the gain of the lth path; When the scattering distance is less than D Ray When , the near-field channel model is: in represents the array response vector of the base station antenna array to the lth path, φ l represents the azimuth of the lth path, r l represents the distance from the lth scatterer to the center of the base station antenna array, represents the distance between the lth scatterer and the mth antenna of the base station, where Therefore, the channel matrix of the mixed field is: From the above formula, we can see that the channel matrix of the mixed field contains L0 far-field paths and L-L0 near-field paths; Step 1.3: Pre-process the signal received by the base station: Knowing that the power of x is P, simplify y to Based on the above formula, the least squares estimation complex matrix of h is: Step 1.4: Generate training and test data sets according to the contents of step 1.2 and step 1.3, respectively saving the noisy channel matrix as input data and the noise-free true channel matrix as output labels, ensuring that the data sets have sufficient diversity to cover different channel conditions.

3. The ultra-large-scale MIMO channel estimation method driven by dual attention in feature channel and spatial domain according to claim 1, characterized in that: The input of the deep convolutional neural network model is the tensor H input , the tensor H input The method of obtaining is: the least squares estimate complex matrix of the channel matrix h Split into real and imaginary parts of equal dimension, which are then concatenated and reshaped into one The tensor H input .

4. The ultra-large-scale MIMO channel estimation method driven by dual attention in feature channel and spatial domain according to claim 1, characterized in that: The deep convolutional neural network model includes four two-dimensional convolutional layers with a convolution kernel size of 3×3 and 64 convolution kernels, two encoders, a two-dimensional convolutional layer with a convolution kernel size of 3×3 and 2 convolution kernels, and a layer normalization and residual connection module. For the first four two-dimensional convolutional layers with a kernel size of 3×3 and 64 kernels, a batch normalization layer is added after each convolutional layer to accelerate the network training process. By normalizing small batches of data, the input of each layer is kept stable to prevent gradient vanishing and gradient exploding problems. A rectified linear unit activation function is used on the output of the convolutional layer to introduce nonlinear features and make the network more expressive. A Dropout layer is introduced between convolutional layers to randomly inactivate a certain proportion of neurons to prevent network overfitting.

5. The ultra-large-scale MIMO channel estimation method driven by dual attention in feature channel and spatial domain according to claim 4, characterized in that: The four convolutional layers with a size of 3×3 and a number of 64 convolutional kernels are used to construct the tensor H. input After extracting the features, reshape them into a tensor H1 suitable for encoder input, with a size of M×64, where M is the number of antennas of the base station in the XL-MIMO system; The H1 is input into two encoders in sequence, and finally the encoder output tensor H5 is reshaped into The tensor H6 is passed through a two-dimensional convolution layer with a convolution kernel size of 3×3 and a convolution kernel number of 2 to obtain H7; Input H7 to the residual connection module, and use the residual connection to input H input Added to H7, the output is H output , improve training stability.

6. The ultra-large-scale MIMO channel estimation method driven by dual attention in feature channel and spatial domain according to claim 1, characterized in that: The encoder includes a feature channel attention mechanism module, a spatial channel attention mechanism module, two fully connected layers, a Dropout layer and a normalization layer in sequence; The feature channel attention mechanism module directly applies the multi-head attention mechanism to generate a reconstruction matrix with the same dimension as the input tensor H1. Finally, the reconstruction matrix obtained by the attention mechanism is summed element by element with H1 to obtain the final output tensor H2 with a size of M×64. H2 represents the weighted sum of all high-dimensional feature maps. The spatial channel attention mechanism module adopts a multi-head attention mechanism to transpose the input tensor H2 to exchange dimensions, and finally transposes the attention output matrix again and sums it with H2 one by one to obtain the output H3 with a size of M×64. H3 is the weighted sum of all position features and the original features of each position.

7. The ultra-large-scale MIMO channel estimation method driven by dual attention in feature channel and spatial domain according to claim 6, characterized in that: The multi-head attention mechanism calculates the dependencies between different channels. If X = [x1, x2, L, x n ] represents n input information, Q, K, V are the matrices of query, key and value respectively, then the attention weight formula is Q=W q X,K=W k X,V=W v X, Among them, W q , W k , W v represents the corresponding weight matrix of Q, K, V, d k is the dimension of vector V; The subspace in the multi-head attention mechanism is generated as follows: The above formulas Q, K, and V are transformed by their respective linear transformation matrices Represents the query Q i , key K i Sum value V i ,in, Represents multiple weight matrix groups corresponding to different heads in the multi-head attention mechanism; Then the formula of the multi-head attention mechanism is: head i =Attention(Q i ,K i ,V i ),i=1,…,h, Among them, h represents the number of "heads", head i represents the attention matrix obtained by the i-th "head", Attention(Q i ,K i ,V i ) is the attention mechanism formula corresponding to the i-th "head".

8. The ultra-large-scale MIMO channel estimation method driven by dual attention in feature channel and spatial domain according to claim 1, characterized in that: The loss function used in step 3 is: in, is the estimated channel matrix, the real channel matrix, N tr Represents the number of training samples in the training data set, and the superscript (n) represents the nth sample. It means taking the Frobenius norm of the matrix and then squaring it.

9. The ultra-large-scale MIMO channel estimation method driven by dual attention in feature channel and spatial domain according to claim 1, characterized in that: The step 3 uses the Adam optimizer to optimize the network. The Adam optimizer combines momentum and adaptive learning rate technology to accelerate convergence. The update rule is as follows: Among them, m t is the momentum vector at the tth iteration, representing the weighted average of the gradient; v t is the squared gradient momentum at the tth iteration, which represents the weighted average of the squared gradient and is the attenuation coefficient of the momentum term; and is the corrected m t and v t ; β1 is the attenuation coefficient of the momentum term; β2 is the attenuation coefficient of the squared gradient momentum; is the gradient of the loss function L(θ) with respect to the model parameter θ; α is the learning rate; θ is the parameter vector of the model; ∈ is a constant.

10. The ultra-large-scale MIMO channel estimation method driven by dual attention in feature channel and spatial domain according to claim 1, characterized in that: In the test phase, step 3 inputs the noisy channel matrix into the trained deep convolutional neural network model, and the model outputs the estimated channel matrix through forward propagation. Calculate the estimated channel matrix With the real channel matrix The normalized mean square error between: For expectations; The accuracy of the model in channel recovery is measured by the normalized mean square error. The smaller the normalized mean square error, the more accurate the channel estimation result.

Images