A channel information compression feedback method of asymmetric network structure

Abstract

The application discloses a channel information compression feedback method of an asymmetric network structure. First, a two-dimensional Fourier transform is used to convert a channel state information (CSI) matrix in a space-frequency domain to an angle-delay domain, then a complex matrix is split into two real number sub-matrices and spliced, and the value range of the CSI matrix is adjusted; a coding network extracts a feature map and a position map of the CSI through an amplitude pooling module, and fuses the feature map and the position map into a feedback code word through a position embedding module; a decoding network extracts a feature map and a position map from the received code word, restores a rough CSI through an inverse pooling layer, and finally further improves the accuracy of CSI reconstruction through a feature fine-tuning network. The application introduces a pooling structure at an encoder end, introduces trainable network parameters only at a decoder end, removes the network trainable parameters at the encoder end through the asymmetric network structure, and introduces a residual compression excitation network, so that the CSI recovery accuracy can be effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of compressed sensing and channel information recovery technology, and relates to a channel information compression feedback network method, specifically a channel information compression feedback method with an asymmetric network structure. Background Technology

[0002] Multiple-input multiple-output (MIMO) antenna arrays and orthogonal frequency division multiplexing (OFDM) are two key technologies in fifth-generation (5G) and above (B5G) communication systems. In MIMO-OFDM systems, to achieve sufficient performance gain, the base station needs accurate downlink channel state information for beamforming, subcarrier allocation, and power control. In frequency division duplex (FDD) mode, downlink channel state information needs to be estimated at the user equipment side and fed back to the base station. However, in MIMO-OFDM systems, the number of antennas and subcarriers is typically large, leading to high channel state information feedback costs.

[0003] Typically, the channel information matrix of massive MIMO is very sparse, and efficient compression can significantly reduce the cost of feedback. However, traditional codebook-based methods are unusable when the channel state information (CSI) matrix is ​​very large because their computational and storage costs are exponentially related to the number of feedback bits. Compressed sensing methods struggle to recover good channel information when the number of feedback bits is small. While neural network-based compressed feedback methods offer advantages in reconstruction accuracy, their deployment is limited by the additional computational and storage overhead required for deployment, especially in scenarios where user-side computational and storage resources are extremely limited. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a channel information compression feedback method using an asymmetric network structure. At the user end, an amplitude pooling method is employed, effectively extracting channel information while eliminating the need for learnable parameters in traditional neural network methods. This solves the problem of deployment difficulties in scenarios with extremely limited user-end computing and storage resources. Furthermore, the residual compression excitation network used in this invention can effectively improve the accuracy of channel information compression feedback.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A channel information compression feedback method for an asymmetric network structure includes:

[0007] Step 1, Preprocessing steps:

[0008] Step 1.1 In an orthogonal frequency division multiplexing (OFDM) communication system, the user concatenates the channel response vectors on different subcarriers into a CSI matrix H;

[0009] Step 1.2: Perform two-dimensional discrete Fourier transform, zero-value clipping, complex number splitting, and normalization operations on H in sequence to obtain the preprocessed matrix H. a ;

[0010] Step 2, Encoding Steps:

[0011] On the user side, an amplitude pooling module with no learnable parameters is used to extract feature map F and location map P; F and P are combined into a feature location map through a location embedding module, and then dimensional reconstruction is performed to obtain the transmission codeword s;

[0012] Step 3, Decoding Steps:

[0013] On the base station side, a feature extraction module is used to extract the recovered feature map. and location map Coarse channel state information recovered by the depooling module The accurate CSI matrix is ​​then obtained through a feature fine-tuning network.

[0014] Furthermore, step 1.1 specifically includes the following steps:

[0015] For a subcarrier number of OFDM system, base station configured with N t With a single antenna configured by the user, the user side obtains a complete channel information matrix, denoted as H, with a dimension of [missing information].

[0016] Furthermore, step 1.2 specifically includes the following steps:

[0017] Two-dimensional Discrete Fourier Transform: Where F c and F d They are respectively and N t ×N t The Fourier matrix of , with superscript (·) H H represents the conjugate transpose of a matrix. d This represents the processed CSI matrix;

[0018] Zero-value clipping: for H a Perform cropping, retaining H. d The first N containing non-zero elements c Okay, we get the clipped matrix H. c ;

[0019] Complex number splitting operation: split H c The matrix is ​​split into two parts: a real part matrix and an imaginary part matrix. These two parts are then concatenated to obtain a matrix of size 2×N. c ×N t The three-dimensional matrix H b ;

[0020] Normalization operation: for H b Subtract 0.5 from each element in the matrix, so that the minimum and maximum values ​​in the CSI matrix are -0.5 and 0.5 respectively, resulting in the preprocessed matrix H. a .

[0021] Furthermore, step 2 uses an amplitude pooling module with no learnable parameters to extract the feature map F and the location map P, specifically including the following steps: given a compression ratio of η, for H... a The magnitude matrix |H is obtained by performing an absolute value operation on each element of the matrix. a |, then for |H a |with Given the pooling kernel and pooling step size, we obtain There are three pooling regions. Max pooling is performed within each pooling region. The position information of the maximum element obtained from max pooling within each region is denoted as p. i，j，k p i，j，k ∈{1,2,...,η}, where i=1,2,j=1,2,...,N c , p i，j，k The dimensions that can be formed are Location map P, then H a by Given the pooling kernel and pooling step size, we obtain For each pooling region, index the position of each pooling region according to the elements in P to obtain the feature f pooled from each region. i，j，k =H a [i, j, p] i，j，k ], where i = 1, 2, j = 1, 2, ..., N c , [·] indicates indexing the matrix elements. The feature f obtained from pooling... i，j，k The dimensions that can be formed are The feature map F. The amplitude pooling module can efficiently extract CSI features while removing learnable parameters.

[0022] Furthermore, in step 2, F and P are synthesized into a feature location map F using a location embedding module. pThen, dimensional reconstruction is performed to obtain the transmission codeword s. Specifically, this includes the following steps: Utilizing the property that F is a decimal while P is an integer, the following formula is used to generate the feature location map F after position embedding. p :

[0023]

[0024] Where sgn(·) represents the sign function. F represents element-wise multiplication. p Dimensions For F p Perform dimensional reconstruction, transforming dimensions into That is, the transmission codeword s is obtained.

[0025] Furthermore, in step 3, a feature extraction module is used to extract the recovered feature map. and location map Specifically, the following steps are included:

[0026] On the base station side, using the received feature location map The feature maps can be recovered sequentially using the following formula. and feature map

[0027]

[0028]

[0029] Where |·| represents the absolute value operation. This indicates a round-down operation.

[0030] Furthermore, in step 3, the depooling module first generates a 2×N dimension. c ×N t All-zero matrix Then according to The position information of the middle element will Fill in the feature information

[0031]

[0032] Where i = 1, 2, j = 1, 2, ..., N c , ← indicates an assignment operation.

[0033] Furthermore, in step 3, precise channel state information is obtained through a feature fine-tuning network. Specifically, the following steps are included:

[0034] Through a convolutional layer with dimensions 2×c×7×7 Expand the channel dimension, where 2 represents the number of input channels, c represents the number of output channels, and 7×7 represents the width and height of the convolution kernel; let the input of the convolution operation be X and the output be Y, then the convolution operation is represented as:

[0035]

[0036] Where W and h are the parameters of the convolution kernel, the subscript (e, f, g, h) represents the index of the element at position (e, f, g, h) in the convolution parameter W, the subscript (e, i+g, j+h) represents the index of the element at position (e, i+g, j+h) in the convolution input X, the subscript f represents the index of the f-th element in the convolution parameter h, and the subscript (f, i, j) represents the index of the element at position (f, i, j) in the convolution output Y. After convolution, a batch normalization layer and a LeakyReLU activation function layer are used, where the LeakyReLU function is defined as...

[0037]

[0038] The convolution output has a dimension of c×N. c ×N t The feature maps are fed into two consecutive residual compression excitation networks to fine-tune the channel state information. Finally, a c×2×1×1 convolutional kernel is used to reduce the channel dimension to 2, and a batch normalization layer and a sigmoid activation function layer are used to reconstruct the accurate CSI matrix. The Sigmoid function is defined as follows:

[0039]

[0040] Furthermore, in the residual compression excitation network, let I1 be the input, where the dimension of I1 is c×N. c ×N t The input I1 is fed into three consecutive convolutional layers to extract spatial features of channel state information, with the output denoted as I2. The convolutional dimensions are c×c×3×3, c×c×1×9, and c×c×9×1, respectively. Each convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function layer. I2 is then fed into a global average pooling layer to obtain a vector of dimension c×1, which is then fed into a fully connected layer of dimension c×d to compress the vector dimension to d×1. Let the input and output of the fully connected layer be... The fully connected layer operation can then be represented as:

[0041]

[0042] in and This represents the parameters of the fully connected layer. and Representing vectors and The j-th element, Representing vectors The i-th element, express The element at position (i, j) is selected; then, a d×1 vector is input into a d×c fully connected layer, resulting in a c×1 channel-weighted vector I3. Multiplying I2 and I3 yields I4, represented as...

[0043] I4[i,j,k]←I2[i,j,k]*I3[i,1],

[0044] Where i = 1, 2, ..., c, j = 1, 2, ..., N c k = 1, 2, ..., N t The dimension of I4 is c×N c ×N t I4 and I1 are element-wise added to obtain I5. Finally, I5 is fed into a c×c×1×1 convolutional layer for feature fusion. After the convolutional layer, a batch normalization layer and a LeakyReLU activation function layer are used.

[0045] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0046] This invention fully considers the deployment problem of neural networks in channel information compression feedback at resource-constrained user terminals. It employs an amplitude pooling method in the encoder at the user terminal for feature extraction of channel information. This method completely eliminates learnable parameters at the user terminal, saving computational and storage overhead. The invention also proposes a location embedding module, which effectively reduces the feedback overhead caused by location information. Finally, the invention uses a residual compression excitation network at the decoder, which can better improve the accuracy of channel information recovery. With an asymmetric network structure, this invention completely eliminates the number of parameters at the encoder, achieving an extremely lightweight deployment advantage while still maintaining a significant performance gain compared to traditional lightweight channel information compression feedback networks. Attached Figure Description

[0047] Figure 1 This is a neural network block diagram of the method of the present invention;

[0048] Figure 2 It is a residual compression excitation network in the user-side feature fine-tuning network.

[0049] Figure 3This is a schematic diagram of the parameter quantities of the user end of the traditional method CsiNet, BCsiNet, and the PoolNet method of this invention under COST2100 channel conditions and different compression rates.

[0050] Figure 4 This diagram illustrates the normalized mean square error performance of channel state information reconstruction using the traditional methods CsiNet, BCsiNet, and the PoolNet method of this invention under different compression rates in a COST2100 indoor channel environment.

[0051] Figure 5 This diagram illustrates the normalized mean square error performance of channel state information reconstruction using the traditional methods CsiNet, BCsiNet, and the PoolNet method of this invention under different compression rates in outdoor channel conditions of COST2100. Detailed Implementation

[0052] The technical solutions provided by the present invention will be described in detail below with reference to specific examples. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0053] The method of the present invention is based on, for example Figure 1 The channel information compression feedback network architecture shown consists of an encoding network and a decoding network. The encoding network extracts features from the CSI matrix using amplitude pooling and generates feedback codewords. The decoding network reconstructs the original CSI matrix based on the feedback codewords. The encoding network includes an amplitude pooling module, a position embedding module, and a dimension reconstruction module. The amplitude pooling module efficiently extracts channel feature information using parameterless pooling parameters (details are explained in step 3 of the construction method). The position embedding module combines the feature map and position map generated by the amplitude pooling module into a feature position map to reduce codeword overhead. The dimension reconstruction module reconstructs the feature position map into a codeword vector. The decoding network consists of a cascaded dimension reconstruction module, a feature extraction layer, an unpooling layer, and a feature fine-tuning network. The dimension reconstruction module reconstructs the codewords from a vector into a feature position map. The feature extraction layer re-extracts the feature map and position map. The unpooling layer regenerates coarse channel state information from the obtained feature map and position map, and the feature fine-tuning network improves the accuracy of channel state information reconstruction. The feature fine-tuning network first expands the dimensions using a convolutional layer, then uses two cascaded residual compression activation networks, and finally restores the original channel dimensions using a convolutional layer. For example... Figure 2As shown, the residual compression excitation network first extracts spatial features through three convolutional layers of different shapes, then extracts channel features through a series of average pooling layers and two fully connected layers, and performs a channel-dimensional product. The result of the product is element-wise added to the input of the residual compression excitation network, and then a convolutional layer is used for feature fusion. Therefore, there are no learnable parameters at the encoder end, and the network with learnable parameters is only deployed at the decoder end. This asymmetric network structure can effectively reduce user-end overhead while efficiently realizing the CSI compression feedback process.

[0054] The method for constructing a channel information compression feedback network proposed in this invention includes the following steps:

[0055] 1) For a subcarrier with a number of OFDM system, base station configured with N t With a single antenna configured by the user, the user side obtains the complete CSI matrix, denoted as H, with a dimension of [missing information].

[0056] 2) Perform two-dimensional discrete Fourier transform, zero-value clipping, complex number splitting, and normalization operations on H in sequence to obtain the preprocessed CSI matrix H. a The specific operations for each step are as follows:

[0057] Two-dimensional Discrete Fourier Transform: Where F c and F d They are respectively and N t ×N t The Fourier matrix of , with superscript (·) H H represents the conjugate transpose of a matrix. d This represents the processed CSI matrix;

[0058] Zero-value clipping: for H d Perform cropping, retaining H. d The first N containing non-zero elements c Okay, we get the clipped matrix H. c ;

[0059] Complex number splitting operation: split H c The matrix is ​​split into two parts: a real part matrix and an imaginary part matrix. These two parts are then concatenated to obtain a matrix of size 2×N. c ×N t The three-dimensional matrix H b ;

[0060] Normalization operation: for H b Subtract 0.5 from each element in the matrix, so that the minimum and maximum values ​​in the CSI matrix are -0.5 and 0.5 respectively, resulting in the preprocessed matrix H.a .

[0061] 3) On the user side, an amplitude pooling module with no learnable parameters is used to extract the feature map F and the location map P; the location embedding module then combines F and P into a feature location map F. p Then, dimensional reconstruction is performed to obtain the transmission codeword s, where the specific operations of each step are as follows:

[0062] Using an amplitude pooling module with no learnable parameters, extract the feature map F and the location map P: given a compression ratio of η, for H a The magnitude matrix |H is obtained by performing an absolute value operation on each element of the matrix. a |, then for |H a |with Given the pooling kernel and pooling step size, we obtain There are three pooling regions. Max pooling is performed within each pooling region. The position information of the maximum element obtained from max pooling within each region is denoted as p. i，j，k p i，j，k ∈{1,2,...,η}, where i=1,2,j=1,2,...,N c , p i，j，k The dimensions that can be formed are Location map P, then H a by Given the pooling kernel and pooling step size, we obtain For each pooling region, index the position of each pooling region according to the elements in P to obtain the feature f pooled from each region. i，j，k =H a [i, j, p] i，j，k ], where i = 1, 2, j = 1, 2, ..., N c , [·] indicates indexing the matrix elements. The feature f obtained from pooling... i，j，k The dimensions that can be formed are The feature map F. The amplitude pooling module can efficiently extract CSI features while removing learnable parameters.

[0063] F and P are synthesized into a feature location map F using a location embedding module. p Then, dimensional reconstruction is performed to obtain the transmission codeword s: taking advantage of the property that F is a decimal while P is an integer, the feature location map F after position embedding is generated using the following formula. p :

[0064]

[0065] Where sgn(·) represents the sign function. F represents element-wise multiplication. p Dimensions For F p Perform dimensional reconstruction, transforming dimensions into That is, the transmission codeword s is obtained.

[0066] 4) On the base station side, the feature extraction module is used to extract the recovered feature map. and location map The coarse CSI matrix recovered by the depooling module The accurate CSI matrix is ​​then obtained through a feature fine-tuning network. The specific steps are as follows:

[0067] The feature extraction module is used to extract the recovered feature map. and location map On the base station side, using the received feature location map The feature maps can be recovered sequentially using the following formula. and feature map

[0068]

[0069]

[0070] Where |·| represents the absolute value operation. This indicates a round-down operation.

[0071] The unpooling module first generates a 2×N dimension. c ×N t All-zero matrix Then according to The position information of the middle element will Fill in the feature information

[0072]

[0073] Where i = 1, 2, j = 1, 2, ..., N c , ← indicates an assignment operation.

[0074] Precise channel state information is obtained through a feature fine-tuning network. The specific steps are as follows:

[0075] Through a convolutional layer with dimensions 2×c×7×7 Expand the channel dimension, where 2 represents the number of input channels, c represents the number of output channels, and 7×7 represents the width and height of the convolution kernel; let the input of the convolution operation be X and the output be Y, then the convolution operation is represented as:

[0076]

[0077] Where W and h are the parameters of the convolution kernel, the subscript (e, f, g, h) represents the index of the element at position (e, f, g, h) in the convolution parameter W, the subscript (e, i+g, j+h) represents the index of the element at position (e, i+g, j+h) in the convolution input X, the subscript f represents the index of the f-th element in the convolution parameter h, and the subscript (f, i, j) represents the index of the element at position (f, i, j) in the convolution output Y. After convolution, a batch normalization layer and a LeakyReLU activation function layer are used, where the LeakyReLU function is defined as...

[0078]

[0079] The convolution output has a dimension of 64×N. c ×N t The feature maps are fed into two consecutive residual compression excitation networks to fine-tune the channel state information. Finally, a 64×2×1×1 convolutional kernel is used to reduce the channel dimension to 2, and a batch normalization layer and a sigmoid activation function layer are used to reconstruct the accurate CSI matrix. The Sigmoid function is defined as follows:

[0080]

[0081] 5) Randomly initialize network parameters before training begins. During training, use the Adam optimizer to minimize the mean square error between the actual CSI matrix of the input encoding network and the CSI matrix recovered by the decoding network.

[0082] 6) Deploy the encoding algorithm and the trained decoding network on the user side and the base station side, respectively.

[0083] After the network is deployed, the channel information compression feedback method for the asymmetric network structure provided by this invention can be implemented, specifically including:

[0084] Preprocessing steps: First, the CSI matrix in the frequency domain is transformed to the angle-time delay domain using a two-dimensional Fourier transform. Then, the complex matrix is ​​split into two real submatrices and concatenated. Specifically, this is the same as steps 1) and 2) in the construction method described above.

[0085] Encoding steps: On the user side, an encoding network is used to extract H data through feature extraction and dimensionality reconstruction operations. a Convert into codeword vectors;

[0086] Decoding steps: On the base station side, a decoding network is used to recover the received codeword vectors into a CSI matrix through dimension recovery and CSI recovery operations;

[0087] The encoding steps include: extracting feature maps and location maps using an amplitude pooling module with no learnable parameters, then combining them into a feature location map using a location embedding module, and finally mapping it into codeword vectors through dimension reconstruction. The specific steps are the same as step 3 in the construction method described above.

[0088] The decoding steps include: first, mapping the received codeword vectors into a feature location map; then, extracting the feature map and location map using a feature extractor; next, recovering a coarse CSI from the feature map and location map using an unpooling layer; and finally, recovering the precise CSI using a feature fine-tuning network. The specific steps are the same as step 4 in the construction method described above.

[0089] Figure 3 For the COST2100 dataset, the number of parameters required on the user side for various traditional lightweight methods and the method of this invention at different compression ratios is clearly different. The method of this invention requires no parameters on the user side at any compression ratio, completely eliminating the storage overhead on the user side. Figure 4 , Figure 5 The figures show the normalized mean square error performance of CSI reconstruction by various traditional lightweight methods and the method of this invention under different compression ratios in indoor and outdoor channel conditions with a COST2100 rating. It is readily apparent that the PoolNet method proposed in this invention exhibits a significant advantage in CSI recovery accuracy across various scenarios and compression ratios.

[0090] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A method for channel information compression feedback of asymmetric network structure, characterized in that, The method comprises the following steps: Step 1, a preprocessing step: Step 1.

1. In an Orthogonal Frequency Division Multiplexing (OFDM) communication system, a user concatenates channel response vectors on different subcarriers into a CSI matrix ; Step 1.1 specifically comprises the following steps: For an OFDM system with a subcarrier number of , the base station configures root antennas, and the user configures a single antenna, so that the user side obtains a complete channel information matrix, denoted as , with a dimension size of ; Step 1.2: Analyze the CSI matrix. The preprocessed CSI matrix is ​​obtained by sequentially performing two-dimensional discrete Fourier transform, zero-value clipping, complex number splitting, and normalization operations. ; Step 1.2 specifically comprises the following steps: a two-dimensional discrete Fourier transform: wherein is a Fourier matrix, is a Fourier matrix, is a Fourier matrix, is a Fourier matrix, denotes the conjugate transpose of a matrix, denotes the CSI matrix after the two-dimensional discrete Fourier transform. Zero value pruning: pruning the matrix , retaining the rows containing non-zero elements , resulting in a pruned matrix ; Complex splitting operation: split into real and imaginary parts, and concatenate the two parts to get a 3D matrix of size ; Normalization operation: for Subtract 0.5 from each element in the CSI matrix to transform the minimum value to -0.5 and the maximum value to 0.5, thus obtaining the preprocessed matrix. ; Step 2, an encoding step: At the user side, a magnitude pooling module without learnable parameters is used to extract the feature map and the position map ; the feature map and the position map are integrated into a feature position map by a position embedding module, and then dimension reconstruction is performed to obtain a transmission code word ; Step 3, a decoding step: At the base station side, a feature extraction module is used to extract the recovered feature map and a position map The rough CSI matrix recovered by the inverse pooling module , and the accurate CSI matrix obtained by the feature fine-tuning network ; The amplitude pooling module without learnable parameters used in step 2 extracts the feature map and a position map , specifically comprising the following steps: Given a compression ratio ,right The magnitude matrix is ​​obtained by performing an absolute value operation on each element of the matrix. Then to by Given the pooling kernel and pooling step size, we obtain There are three pooling regions. Max pooling is performed within each pooling region. The position information of the maximum element obtained from max pooling within each region is denoted as... , ； The position map has a dimension of , and then is taken as the pooling kernel and the pooling step length, to obtain pooling regions, and each pooling region is indexed according to elements to obtain the features pooled by each pooling region , , wherein the matrix elements are indexed; the pooled features form a feature map with a dimension of ;​​​ In step 2, the location embedding module is used to... and Synthesized into a feature location map Then, dimensional reconstruction is performed to obtain the transmission codewords. Specifically, it includes the following steps: using The value can be a decimal, while Given the property that the value of is an integer, the following formula is used to generate the feature location map after position embedding. : ； in Represents a symbolic function. This indicates element-wise multiplication. Dimensions ;right Perform dimensional reconstruction, transforming dimensions into That is, the transmission codeword is obtained. ; At the base station side, the feature extraction module is used in step 3 to extract the recovered feature map and the position map Specifically, the following steps are included: at the base station side, the received feature position map is used to recover the feature map The feature extraction module is used to recover the feature map in sequence according to the following formula and the feature map : ； wherein denotes an absolute value operation, denotes a floor operation; The anti-pooling module in step 3 first generates a full zero matrix with a dimension of ,then fills the feature information in into according to the position information of the elements in . ； wherein , represents an assignment operation; The step 3 obtains the accurate channel state information through the feature fine-tuning network , specifically comprising the following steps: ​​​​​​​​ ； in and The parameters of the convolution kernel are represented by the subscripts. Represents convolution parameters The middle position is Element index, subscript Represents convolution input The middle position is Element index, subscript Represents convolution parameters The Middle Element index, subscript Represents convolution output The middle position is The element index; after convolution, a batch normalization layer and a LeakyReLU activation function layer are used, where the LeakyReLU function is defined as... ； The dimension of the convolution output is The feature map is input into two consecutive residual compression excitation networks to fine-tune the channel state information; finally, a convolution kernel is used to reduce the channel dimension to 2, and a batch normalization layer and a Sigmoid activation function layer are used to reconstruct the accurate CSI matrix ; wherein the Sigmoid function is defined as: 。 2. The method of claim 1, wherein the asymmetric network structure is a tree structure. In the residual compression excitation network, the input of the residual compression excitation network is denoted as... ,in The dimension is ,Will The input is fed into three consecutive sequential convolutional layers to extract spatial features of channel state information, and the output is denoted as... The dimensions of the convolution are respectively ； A batch normalization layer and a LeakyReLU activation function layer are used after each convolutional layer. is input into the global average pooling layer to obtain a vector with a dimension of is input into a fully connected layer with a dimension of to compress the vector dimension to ; let the input and output of the fully connected layer be , , then the fully connected layer operation is represented as: ； in and This represents the parameters of the fully connected layer. and Representing vectors and The One element, Representing vectors The One element, express The middle position is The elements; then the dimension is The vector input to a dimension of In the fully connected layer, the dimension is thus obtained. Channel-dimensional weighted vector Then and Multiplication yields , represented as: ； in , The dimension is Then and Element-wise addition Finally, Input to Feature fusion is performed in the convolutional layers, followed by batch normalization layers and LeakyReLU activation function layers.

Images