Channel information compression feedback method using frequency division duplex channel reciprocity

Abstract

The application discloses a channel information compression feedback method using frequency division duplex channel implicit reciprocity. For the frequency division duplex channel information data with implicit reciprocity, the application firstly converts the spatial frequency domain channel matrix to the angle delay domain and performs clipping, and then splits and splices the real part and the imaginary part of the channel matrix into two real number matrices; three encoder networks are used for decoupling characterization learning of the downlink channel, and the downlink channel unique information, the uplink and downlink channel common information and the virtual uplink and downlink channel common information are extracted respectively; the decoder network recovers the original channel information by using the downlink channel unique information and the uplink and downlink channel common information during training, and recovers the original channel information by using the downlink channel unique information and the virtual uplink and downlink channel common information during testing; in the application, the user only needs to feed back the downlink channel unique information, the channel feedback load is effectively reduced, the channel information recovery performance is improved, and the application has a significant performance advantage.

Description

Technical Field

[0001] This invention belongs to the field of compressed sensing and channel information recovery technology, and particularly relates to a channel information compression feedback method that utilizes the implicit reciprocity of frequency division duplex channels. Background Technology

[0002] Massive MIMO has become a key technology in modern wireless communication networks. By deploying massive antenna arrays at base stations, massive MIMO can effectively utilize spatial resources to mitigate the impact of multi-user interference, thereby improving spectral and energy efficiency. However, to achieve performance gains, base stations need accurate channel state information. In time-division duplex systems, base stations can directly obtain downlink channel state information from uplink channel state information using channel reciprocity. Conversely, in frequency-division duplex systems, downlink channel state information must be fed back to the base station by users through bandwidth-constrained feedback links. As the number of antennas deployed at the base station increases in frequency-division duplex massive MIMO systems, obtaining downlink channel information requires significant overhead and resources. This necessitates compression of the downlink channel state information fed back by users. Traditional compression methods include codebook-based and compressed sensing-based methods, but the channel state information feedback overhead of these methods still exhibits a linear relationship with the number of base station antennas. This leads to a significant increase in channel state information feedback overhead as the number of antennas increases.

[0003] Due to the powerful learning and parallel computing capabilities of deep learning technology, using deep learning to compress downlink channel state information fed back by users is a promising direction, and there are already many effective methods in this direction. However, these methods usually assume that there is no reciprocity between the uplink and downlink channels in frequency division duplex massive MIMO systems, thus ignoring the potential role of uplink channel information in acquiring downlink channel information. However, implicit reciprocity may exist in the bidirectional channels of frequency division duplex systems because they actually share the same physically identical propagation path. Therefore, we propose to utilize this implicit reciprocity in frequency division duplex massive MIMO systems to minimize feedback overhead and enhance the recovery of downlink channel information through uplink channel information at the base station. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a channel information compression feedback method that utilizes the implicit reciprocity of frequency division duplex channels. This method can fully leverage the implicit reciprocity of uplink and downlink channels and effectively extract features, thereby solving the technical problems of low compression efficiency and recovery performance in networks.

[0005] Technical Solution: To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0006] A channel information compression feedback method utilizing the implicit reciprocity of frequency division duplex channels includes the following steps:

[0007] The uplink and downlink spatial frequency domain channel matrices are transformed to the angular time delay domain using a two-dimensional discrete Fourier transform, and then subjected to time-domain clipping, complex transformation, and normalization to obtain the preprocessed downlink channel information matrix H. d and uplink channel information matrix H u ;

[0008] The downlink channel-specific information extraction module is used to extract information from the input H. d The unique information representation of the downlink channel is extracted from it, denoted as z. d The uplink and downlink channel shared information extraction module is used to extract information from the input H. d and H u The shared information representation of the uplink and downlink channels is extracted and denoted as z. S The virtual uplink and downlink channel shared information extraction module is used to extract information from the input H. u Extracting z S The approximate distribution representation is denoted as the virtual uplink / downlink shared representation r. S The downlink channel-specific information extraction module, deployed on the user side, compresses the extracted feature map containing downlink-specific information to a codeword with a given compression ratio through a fully connected network, and extracts the downlink-specific representation z through another fully connected network deployed on the base station side. d The virtual uplink and downlink channel shared information extraction module is deployed on the base station side.

[0009] z during training d With z S The downlink channel matrix H is recovered through the decoding network. d During testing, z will be... d With r S The downlink channel matrix H is recovered through the decoding network. d .

[0010] As a preferred option, the downlink spatial frequency domain channel matrix and uplink spatial frequency domain channel matrix A series of preprocessing operations are performed to obtain the preprocessed downlink channel information matrix H. d and uplink channel information matrix H u The specific operations are as follows:

[0011] Two-dimensional Discrete Fourier Transform: Where F a For N t ×N t The Fourier matrix, F b For N f ×Nf The Fourier matrix, where the superscript * denotes the conjugate transpose of the matrix, H′ d and H′ u Let N represent the downlink channel matrix and uplink channel matrix after Fourier transform, respectively. f N represents the number of subcarriers. t This refers to the number of base station antennas.

[0012] Temporal clipping: for H′ d and H′ u Perform cropping, keeping only H′ d and H′ u The first N containing non-zero elements c The columns, and the clipped matrices, are represented as H″ respectively. d and H″ u ;

[0013] Complex number conversion: convert H″ to H″ respectively d and H″ u The real and imaginary parts are extracted and concatenated along the third dimension to obtain two parts of size N. t ×N c A three-dimensional matrix H″′ of size ×2 d and H″′ u ;

[0014] Normalization processing: for H″′ respectively d and H″′ u Normalization is performed by subtracting the minimum value and dividing by the difference between the maximum and minimum values ​​to obtain the normalized downlink channel information matrix H. d and uplink channel information matrix H u .

[0015] Preferably, the downlink-specific information extraction module includes a first nonlinear convolutional layer and at least two representation extraction modules composed of nonlinear convolutional layers. The first nonlinear convolutional layer is used to perform preliminary extraction of features of the downlink channel information matrix. The representation extraction module consists of two parallel nonlinear convolutional branches and a second nonlinear convolutional layer. Each nonlinear convolutional branch consists of multiple cascaded nonlinear convolutional layers. The multiple nonlinear convolutional layers of the first convolutional branch are used to extract dense features of the channel information matrix. The multiple nonlinear convolutional layers of the second convolutional branch are used to extract sparse features of the channel information matrix. The feature maps output by the two convolutional branches are concatenated in the third dimension and finally passed through the second nonlinear convolutional layer to obtain the output of the representation extraction module. The final output of the representation extraction module is the output of the downlink-specific information extraction module.

[0016] Preferably, the uplink and downlink shared information extraction module consists of two convolutional paths identical to those in the downlink-specific information extraction module and a third nonlinear convolutional layer. The first convolutional path is for H...d Extracting common information features, the second convolutional path pairs H u Extract shared information features; concatenate the output features of the two convolutional paths and output z through a third nonlinear convolutional layer. S mean u S Sum of standard deviation N S and N S Multiply by a sample from a standard Gaussian distribution and add u S Finally, z was obtained S .

[0017] Preferably, the virtual uplink / downlink shared information extraction module is reused in the uplink / downlink shared information module for H. u The convolutional path that extracts shared information features is then passed through a fourth nonlinear convolutional layer to output a virtual shared representation r. S mean u r and standard deviation σ r and σ r Multiply by a sample from a standard Gaussian distribution and add u r Finally, we get r S .

[0018] Preferably, the decoding network consists of a fifth nonlinear convolutional layer, at least two information recovery modules, and a nonlinear activation function sigmoid layer connected in sequence. The fifth nonlinear convolutional layer is used to perform preliminary filtering on the input feature map and transform it to a form similar to H. d Same size; the Sigmoid layer is used to map the output to recover the channel matrix;

[0019] The information recovery module includes multiple parallel two-layer nonlinear convolutional layers. Each two-layer nonlinear convolutional layer is used to refine the input channel information feature map, extract effective features and remove invalid features. The multiple output feature maps are concatenated and then input into the sixth nonlinear convolutional layer to compress the channel dimension of the feature map to be consistent with the original input to obtain the output feature map. The original input and output are added together and finally passed through the activation layer to obtain the final output.

[0020] Preferably, the training objective of the entire network framework is to minimize the following expression:

[0021]

[0022] Where I(z) d ;z S ) represents z d With z S Mutual information between them, I(z) S H d H u ) represents z S Hd and H u Mutual information between them, p a (H d ) represents H d The actual probability distribution, p(H) d ) represents H d The reconstruction probability, q(z) S |H d H u ) and q(r S |H u ) represent the z values ​​estimated by the uplink and downlink shared information extraction modules, respectively. S The probability distribution and the estimated r by the virtual uplink and downlink shared information extraction module S The probability distributions are given by β, α, and γ, which are positive weighting coefficients, and K represents the KL divergence between the two distributions.

[0023] Furthermore, by relaxing the first and second terms using the nonnegativity of the KL divergence and Jensen's inequality, the relaxed result of the above equation is as follows:

[0024]

[0025] Where q(z) d |H d ) represents the z estimated by the downlink-specific information extraction module. d The probability distribution, p(z) d ) and p(z S ) for z d and z S The actual probability distribution.

[0026] Based on the same inventive concept, this invention provides a channel information compression feedback system utilizing the implicit reciprocity of frequency division duplex channels, comprising:

[0027] The preprocessing module transforms the uplink and downlink spatial frequency domain channel matrices into the angular time delay domain using a two-dimensional discrete Fourier transform, and then performs time-domain clipping, complex transformation, and normalization to obtain the preprocessed downlink channel information matrix H. d and uplink channel information matrix H u ;

[0028] The encoding module is used to extract information from the input H using the downlink channel-specific information extraction module. d The unique information representation of the downlink channel is extracted from it, denoted as z. d The uplink and downlink channel shared information extraction module is used to extract information from the input H. d and H u The shared information representation of the uplink and downlink channels is extracted and denoted as z. SThe virtual uplink and downlink channel shared information extraction module is used to extract information from the input H. u Extracting z S The approximate distribution representation is denoted as the virtual uplink / downlink shared representation r. S The downlink channel-specific information extraction module, deployed on the user side, compresses the extracted feature map containing downlink-specific information to a codeword with a given compression ratio through a fully connected network, and extracts the downlink-specific representation z through another fully connected network deployed on the base station side. d The virtual uplink and downlink channel shared information extraction module is deployed on the base station side.

[0029] The decoding module is used to decode z during training. d With z S The downlink channel matrix H is recovered through the decoding network. d During testing, z will be... d With r S The downlink channel matrix H is recovered through the decoding network. d .

[0030] Based on the same inventive concept, the present invention provides a computer system including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the steps of the channel information compression feedback method utilizing the implicit reciprocity of frequency division duplex channels.

[0031] Beneficial effects: This invention first transforms the uplink and downlink spatial frequency domain channel matrices to the angular time delay domain using a two-dimensional Fourier transform and then prunes them. Next, the real and imaginary parts of the channel matrix are split into two real matrices and combined. Three encoder networks are used to decouple and represent the downlink channel, extracting downlink-specific information, uplink-downlink shared information, and virtual uplink-downlink shared information. During training, the decoder network recovers the original channel information using downlink-specific information and uplink-downlink shared information; during testing, it recovers the original channel information using downlink-specific information and virtual uplink-downlink shared information. Compared with existing technologies, this invention utilizes the implicit reciprocity of frequency division duplex channels to extract shared information from uplink and downlink channels, allowing users to only need to feed back downlink-specific information, effectively reducing channel feedback load and improving channel information recovery performance, demonstrating significant performance advantages. Attached Figure Description

[0032] Figure 1 This is a block diagram of the neural network structure of the method of the present invention;

[0033] Figure 2 This is a block diagram of the representation extraction module of the present invention;

[0034] Figure 3This is a block diagram of the information recovery module of the present invention;

[0035] Figure 4 This diagram illustrates the performance of other neural networks and the present invention under the Quadriga channel model. Detailed Implementation

[0036] To better understand the purpose, structure, and function of this invention, the technical solution of this invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

[0037] The method of the present invention is based on, for example Figure 1 The neural network architecture shown consists of an encoder network and a decoder network. The encoder network includes a downlink channel-specific information extraction module, an uplink / downlink channel shared information extraction module, and a virtual uplink / downlink channel shared information extraction module. The downlink-specific module consists of a single nonlinear convolutional layer and at least two representation extraction modules, with the input downlink channel H... d The output feature map W is compressed to a codeword m with a given compression ratio through a fully connected network, and the downlink-specific representation z is extracted through another fully connected network deployed on the base station side. d The uplink and downlink shared information extraction module consists of two convolutional paths and one non-linear convolutional layer. The composition of one convolutional path is the same as that of the downlink-specific information extraction module, and the input is the downlink channel H. d and uplink channel H u The total number of outputs for both uplink and downlink is represented by z. S The virtual uplink / downlink shared information extraction module reuses the H data from the uplink / downlink shared information module. u The convolutional path extracts common information features and is additionally cascaded with a non-linear convolutional layer, input to the uplink channel H. u Output virtual uplink and downlink total representation r S ;z S Used only during training, r S Then use it in the test; such as Figure 2 As shown, the extraction module consists of two parallel nonlinear convolutional branches and one nonlinear convolutional layer. Each nonlinear convolutional branch comprises multiple cascaded nonlinear convolutional layers. The multiple nonlinear convolutional layers in the first branch use smaller kernels to extract dense features of the channel information matrix; the multiple nonlinear convolutional layers in the second branch use larger kernels to extract sparse features of the channel information matrix. During training, the decoding network is used to... d and z S The concatenated feature map is used to restore the original channel matrix; during testing, the decoding network is used to convert z... d and r SThe concatenated feature map is restored to the original channel matrix; the decoding network consists of a nonlinear convolutional layer, at least two information recovery modules, and a nonlinear activation function sigmoid layer connected sequentially. Figure 3 As shown, the information recovery module includes multiple parallel two-layer nonlinear convolutional layers with different kernel sizes. Each two-layer nonlinear convolutional layer is used to refine the input channel information feature map, extract effective features, and remove invalid features.

[0038] Specifically, the channel information compression feedback method utilizing the implicit reciprocity of frequency division duplex channels disclosed in this embodiment of the invention comprises the following detailed steps:

[0039] Step 1: Preprocessing: Transform the uplink and downlink spatial frequency domain channel matrices to the angular time delay domain and perform preprocessing to obtain the corresponding downlink channel information matrix. In a large-scale multiple-input multiple-output (MIMO) frequency division duplex communication system with Orthogonal Frequency Division Multiplexing (OFDM) modulation, the receiver concatenates the received channel response vectors on each downlink and uplink subcarrier to form the downlink channel information matrix. and uplink channel information matrix The subscripts d and u represent downlink and uplink respectively; for and The preprocessed downlink channel information matrix H is obtained by performing two-dimensional discrete Fourier transform, time-domain clipping, complex transformation, and normalization respectively. d and uplink channel information matrix H u .

[0040] Step 1.1: For a subcarrier number of N f In an OFDM system, the number of antennas at the base station is N. t The user configures a single antenna and obtains the complete downlink channel matrix, denoted as... The complete uplink channel matrix is ​​obtained at the base station. and The dimension size is N t ×N f ;

[0041] Step 1.2, for and A series of preprocessing operations are performed to obtain the preprocessed downlink channel matrix and uplink channel matrix, wherein the specific operations are as follows:

[0042] Two-dimensional Discrete Fourier Transform: Where F a For Nt ×N t The Fourier matrix, F b For N f ×N f The Fourier matrix, where the superscript * denotes the conjugate transpose of the matrix, H′ d and H′ u These represent the downlink channel matrix and the uplink channel matrix after Fourier transform, respectively;

[0043] Time-domain clipping: Due to the large number of zero values ​​in the channel matrix in the time domain, H′ d and H′ u Perform cropping, keeping only H′ d and H′ u The first N containing non-zero elements c Column, and there are N f >>N c The clipped matrices are represented as H″. d and H″ u ;

[0044] Complex number conversion: convert H″ to H″ respectively d and H″ u The real and imaginary parts are extracted and concatenated along the third dimension to obtain two parts of size N. t ×N c A three-dimensional matrix H″′ of size ×2 d and H″′ u ;

[0045] Normalization processing: for H″′ respectively d and H″′ u Normalization is performed by subtracting the minimum value and dividing by the difference between the maximum and minimum values ​​to obtain the normalized downlink channel matrix H. d and uplink channel matrix H u .

[0046] Step 2, Encoding: Use the downlink channel-specific information extraction module to extract information from the input H... d The unique information representation of the downlink channel is extracted from it, denoted as z. d The uplink and downlink channel shared information extraction module is used to extract information from the input H. d and H u The shared information representation of the uplink and downlink channels is extracted and denoted as z. S The virtual uplink and downlink channel shared information extraction module is used to extract information from the input H. u Extracting z S An approximate representation of the distribution, denoted as the virtual uplink / downlink shared representation r. S .

[0047] Step 2.1: Construct a downlink channel-specific information extraction module and deploy it on the user side. The input is H. d This is used to extract unique information from the downlink channel matrix and feed it back to the base station; a common uplink and downlink information extraction module is constructed for training purposes, but is not deployed in actual applications, and its input is H. d and H u To extract common information between the downlink and uplink channels; construct a virtual uplink / downlink common information extraction module, deploy it on the base station side, and take H as the input. u This is used to extract virtual uplink and downlink shared information. Following the downlink channel-specific information extraction module is a fully connected network that compresses the extracted feature map W containing downlink-specific information to a codeword m with a given compression ratio. Then, another fully connected network deployed on the base station side extracts the downlink-specific representation z. d mean u d and standard deviation σ d and σ d Multiply by a sample from a standard Gaussian distribution and add u d Finally, z was obtained d The downlink channel-specific information extraction module, the uplink-downlink shared information extraction module, and the virtual uplink-downlink shared information extraction module are all composed of nonlinear convolutional layers of different sizes. The nonlinear convolutional layers include convolutional layers, batch normalization layers, and activation layers.

[0048] Step 2.2: In this example, the downlink-specific information extraction module includes a first nonlinear convolutional layer and two convolutional modules composed of nonlinear convolutional layers. This module is called the representation extraction module. The size of the first nonlinear convolutional layer is 3×3×c×c, which is used to perform preliminary coarse extraction of the features of the downlink channel information matrix. After the convolutional layer, a batch normalization layer and a nonlinear activation layer are used. The LeakyReLU(·) function of the nonlinear activation layer is defined as follows:

[0049]

[0050] Where x represents the input feature map or vector of the nonlinear layer; the extraction module consists of two parallel nonlinear convolutional branches and a second nonlinear convolutional layer. Each nonlinear convolutional branch consists of three cascaded nonlinear convolutional layers. The sizes of the three nonlinear convolutional layers in the first branch are 3×3×c×c, 1×7×c×c, and 7×1×c×c, respectively; the sizes of the three nonlinear convolutional layers in the second branch are 5×5×c×c, 1×9×c×c, and 9×1×c×c, respectively. The feature maps output from the two convolutional branches are concatenated in the third dimension and then passed through a second nonlinear convolutional layer of size 1×1×2c×c to obtain the output of the representation extraction module. The output of the second representation extraction module is the output of the downlink unique information extraction module. It is understood that the size of the convolutional kernel and the number of modules in this embodiment are exemplary and can be adjusted and optimized according to the scale of the scenario and the effect of the algorithm in actual applications.

[0051] Step 2.3: The uplink and downlink shared information extraction module consists of two convolutional paths and a third non-linear convolutional layer. The first convolutional path is for H... d The common information features are extracted, similar to the downlink-specific information extraction module, consisting of a non-linear convolutional layer and two representation extraction modules. The second convolutional path pairs with H. u The extraction of common information features is similar to that of the downlink-specific information extraction module, consisting of a non-linear convolutional layer and two representation extraction modules. The output features of the two convolutional paths are concatenated and then output as z through a third non-linear convolutional layer of size 3×3×2c×2c. S mean u S and standard deviation σ S and σ S Multiply by a sample from a standard Gaussian distribution and add u S Finally, z was obtained S The virtual uplink / downlink shared information extraction module reuses the H data from the uplink / downlink shared information module. u The convolutional path that extracts shared information features is then passed through a fourth nonlinear convolutional layer of size 3×3×c×2c to output a virtual shared representation r. S mean u r and standard deviation σ r r is obtained in the same way S .

[0052] Step 2.4, Matrix H d and H u The dimension is N t ×N c ×2, the channel dimension of the input data is 2, so c=2 is present in the downlink-specific information extraction module, the uplink-downlink shared information extraction module, and the virtual uplink-downlink shared information extraction module.

[0053] Step 2.5: Convert matrix H d The input is the downlink unique information extraction module, which outputs a feature map W and compresses it to a codeword m with a given compression ratio through a fully connected network. Then, the downlink unique representation z is extracted through another fully connected network deployed on the base station side. d ; Matrix H d and H u Input the uplink and downlink total information extraction module, output the uplink and downlink total information representation z s ; Matrix H u Input the virtual uplink and downlink total information extraction module, and output the virtual uplink and downlink total information representation r. s ;z s Used only during training, r s Then it will be used in the test.

[0054] Step 3, Decoding: During training, z d With z s The downlink channel matrix H is recovered through the decoding network. d During testing, z will be... d With r S The downlink channel matrix H is recovered through the decoding network. d .

[0055] Step 3.1: Construct a decoding network on the base station side. During training, this network is used to decode z. d and z S The size after splicing is N t ×N c The ×4 feature map is used to recover the original channel matrix; during testing, it is used to convert z... d and r S The size after splicing is N t ×N c The 3×4 feature map is restored to the original channel matrix; the decoding network consists of a fifth nonlinear convolutional layer, two nonlinear convolutional modules, and a nonlinear activation function sigmoid layer connected in sequence. This convolutional module is called the information recovery module. The fifth nonlinear convolutional layer, with a size of 3×3×2×4, is used to perform preliminary filtering on the input feature map and transform it to a form similar to H. d The same size; finally, the output is mapped to the interval [0,1] by the Sigmoid layer to obtain the final recovered channel matrix;

[0056] Step 3.2: The information recovery module includes four parallel two-layer nonlinear convolutional layers. The convolutional layer sizes of the four parallel two-layer nonlinear convolutional layers are 3×3×2×8 and 5×5×8×8, 5×5×2×8 and 3×3×8×8, 1×7×2×8 and 7×1×8×8 and 1×9×2×8 and 9×1×8×8, respectively. The feature maps of the four outputs are concatenated and then input to a nonlinear convolutional layer with a size of 1×1×32×2. This nonlinear convolutional layer only includes convolutional layers and batch normalization layers to compress the feature map channel dimensions to be consistent with the original output to obtain the output feature map. To prevent gradient explosion, the original input and output are added together and then passed through an activation layer to obtain the final output.

[0057] Step 3.3: Minimize z d With z S Mutual information between I(z) d ;z S ) and maximizing z S H d and H u Mutual information between I(z) S H d H u This achieves decoupling of downlink-specific information and uplink / downlink shared information, reducing z d With z S This minimizes information redundancy and maximizes the shared information that can be extracted at the base station, thereby reducing the amount of information that users need to provide. Simultaneously, it minimizes z. S With r S The Kullback-Leibler (KL) divergence between the distributions makes r S Approaching z S This allows us to achieve channel reconstruction accuracy during testing that is similar to that during training; finally, we minimize the channel information reconstruction error, which can be modeled as minimizing H. d The negative value of the reconstruction probability leads to the training objective of the entire network framework being to minimize the following equation:

[0058]

[0059] Where p a (H d ) represents H d The actual probability distribution, p(H) d ) represents H d The reconstruction probability, q(z) S |H d H u ) and q(r S |H u ) represent the z values ​​estimated by the uplink and downlink shared information extraction modules, respectively. SThe probability distribution and the estimated r by the virtual uplink and downlink shared information extraction module S The probability distributions are given by β, α, and γ, which are positive weighting coefficients, and K represents the KL divergence between the two distributions. However, directly calculating the first and second terms of the above equation is difficult in neural network training. Therefore, we relax the first and second terms of the above equation, transforming it into minimizing the upper bounds of the first and second terms. We relax the first and second terms using the non-negativity of the KL divergence and Jensen's inequality, respectively, and obtain the relaxed result of the above equation as follows:

[0060]

[0061] Where q(z) d |H d ) represents the z estimated by the downlink-specific information extraction module. d The probability distribution, p(z) d ) and p(z S ) for z d and z S In the actual probability distribution, α, β, and γ are generally taken as 5 × 10⁻⁶. -7 5×10 -7 2×10 4 The third term in the above equation is obtained by multiplying the mean square error between the actual downlink channel and the downlink channel reconstructed by the decoding network by a weighting coefficient, which is typically taken as 5 × 10. 10 ;

[0062] Step 3.4: Initialize network parameters before training begins. During training, optimize using gradient descent, and use the loss obtained in step 3.3 as the loss term for neural network training.

[0063] Step 3.5: Deploy the trained downlink-specific information extraction module on the user side, deploy the virtual uplink-downlink shared information extraction module and decoding network on the base station side, and the uplink-downlink shared module does not participate in the actual deployment.

[0064] Figure 4 To compare the performance of other neural network methods with that of this invention under the channel model generated by Quadriga, Quadriga is a tool for generating highly accurate and controllable radio frequency channels. Its full name is Quasi-Deterministic Radio channel Generator. The comparison method is CRNet network structure, the comparison method is ACRNet 1x network structure, and the comparison method is ACRNet 10x network structure. As can be seen from the figure, the method proposed in this invention has a performance gain compared with other neural network methods at different compression ratios, and the gain is even higher at high compression ratios.

[0065] Based on the same inventive concept, this invention discloses a channel information compression feedback system utilizing the implicit reciprocity of frequency division duplex channels, comprising: a preprocessing module, used to transform the uplink and downlink spatial frequency domain channel matrices to the angular time delay domain through a two-dimensional discrete Fourier transform, and perform time-domain clipping, complex transformation, and normalization to obtain the preprocessed downlink channel information matrix H. d and uplink channel information matrix H u The encoding module is used to extract information from the input H using the downlink channel-specific information extraction module. d The unique information representation of the downlink channel is extracted from it, denoted as z. d The uplink and downlink channel shared information extraction module is used to extract information from the input H. d and H u The shared information representation of the uplink and downlink channels is extracted and denoted as z. S The virtual uplink and downlink channel shared information extraction module is used to extract information from the input H. u Extracting z S The approximate distribution representation is denoted as the virtual uplink / downlink shared representation r. S The downlink channel-specific information extraction module, deployed on the user side, compresses the extracted feature map containing downlink-specific information to a codeword with a given compression ratio through a fully connected network, and extracts the downlink-specific representation z through another fully connected network deployed on the base station side. d The virtual uplink and downlink channel shared information extraction module is deployed on the base station side; the decoding module is used to extract z during training. d With z S The downlink channel matrix H is recovered through the decoding network. d During testing, z will be... d With r S The downlink channel matrix H is recovered through the decoding network. d .

[0066] Based on the same inventive concept, an embodiment of the present invention discloses a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the steps of the channel information compression feedback method utilizing the implicit reciprocity of frequency division duplex channels.

[0067] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A channel information compression feedback method utilizing the implicit reciprocity of frequency division duplex channels, characterized in that, Includes the following steps: The uplink and downlink spatial frequency domain channel matrices are transformed to the angular time delay domain using a two-dimensional discrete Fourier transform, and then subjected to time-domain clipping, complex transformation, and normalization to obtain the preprocessed downlink channel information matrix H. d and uplink channel information matrix H u ; The downlink channel-specific information extraction module is used to extract information from the input H. d The unique information representation of the downlink channel is extracted from it, denoted as z. d The uplink and downlink channel shared information extraction module is used to extract information from the input H. d and H u The shared information representation of the uplink and downlink channels is extracted and denoted as z. S The virtual uplink and downlink channel shared information extraction module is used to extract information from the input H. u Extracting z S The approximate distribution representation is denoted as the virtual uplink / downlink shared representation r. S The downlink channel-specific information extraction module, deployed on the user side, compresses the extracted feature map containing downlink-specific information to a codeword with a given compression ratio through a fully connected network, and extracts the downlink-specific representation z through another fully connected network deployed on the base station side. d The virtual uplink and downlink channel shared information extraction module is deployed on the base station side. z during training d With z S The downlink channel matrix H is recovered through the decoding network. d During testing, z will be... d With r S The downlink channel matrix H is recovered through the decoding network. d .

2. The channel information compression feedback method utilizing the implicit reciprocity of frequency division duplex channels according to claim 1, characterized in that, For the downlink spatial frequency domain channel matrix and uplink spatial frequency domain channel matrix A series of preprocessing operations are performed to obtain the preprocessed downlink channel information matrix H. d and uplink channel information matrix H u The specific operations are as follows: Two-dimensional Discrete Fourier Transform: Where F a For N t ×N t The Fourier matrix, F b For N f ×N f The Fourier matrix, where the superscript * denotes the conjugate transpose of the matrix, H′ d and H′ u Let N represent the downlink channel matrix and uplink channel matrix after Fourier transform, respectively. f N is the number of subcarriers. t This refers to the number of base station antennas. Temporal clipping: for H′ d and H′ u Perform cropping, keeping only H′ d and H′ u The first N containing non-zero elements c The columns, and the clipped matrices, are represented as H″ respectively. d and H″ u ; Complex number conversion: convert H″ to H″ respectively d and H″ u The real and imaginary parts are extracted and concatenated along the third dimension to obtain two parts of size N. t ×N c A three-dimensional matrix H″′ of size ×2 d and H″′ u ; Normalization processing: for H″′ respectively d and H″′ u Normalization is performed by subtracting the minimum value and dividing by the difference between the maximum and minimum values ​​to obtain the normalized downlink channel information matrix H. d and uplink channel information matrix H u .

3. The channel information compression feedback method utilizing the implicit reciprocity of frequency division duplex channels according to claim 1, characterized in that, The downlink-specific information extraction module includes a first nonlinear convolutional layer and at least two representation extraction modules composed of nonlinear convolutional layers. The first nonlinear convolutional layer is used to perform preliminary extraction of features of the downlink channel information matrix. The representation extraction module consists of two parallel nonlinear convolutional branches and a second nonlinear convolutional layer. Each nonlinear convolutional branch consists of multiple cascaded nonlinear convolutional layers. The multiple nonlinear convolutional layers of the first convolutional branch are used to extract dense features of the channel information matrix. The second convolutional branch has multiple non-linear convolutional layers used to extract sparse features of the channel information matrix; the feature maps output by the two convolutional branches are concatenated in the third dimension and finally passed through the second non-linear convolutional layer to obtain the output of the representation extraction module. The final output of the representation extraction module is the output of the downlink-specific information extraction module.

4. The channel information compression feedback method utilizing the implicit reciprocity of frequency division duplex channels according to claim 1, characterized in that, The shared uplink and downlink information extraction module consists of two convolutional paths identical to the downlink-specific information extraction module and a third nonlinear convolutional layer. The first convolutional path is for H... d Extracting common information features, the second convolutional path pairs H u Extract shared information features; concatenate the output features of the two convolutional paths and output z through a third nonlinear convolutional layer. S mean u S and standard deviation σ S and σ S Multiply by a sample from a standard Gaussian distribution and add u S Finally, z was obtained S .

5. The channel information compression feedback method utilizing the implicit reciprocity of frequency division duplex channels according to claim 1, characterized in that, The virtual uplink / downlink shared information extraction module is reused in the uplink / downlink shared information module for H. u The convolutional path that extracts shared information features is then passed through a fourth nonlinear convolutional layer to output a virtual shared representation r. S mean u r and standard deviation σ r and σ r Multiply by a sample from a standard Gaussian distribution and add u r Finally, we get r S .

6. The channel information compression feedback method utilizing the implicit reciprocity of frequency division duplex channels according to claim 1, characterized in that, The decoding network consists of a fifth nonlinear convolutional layer, at least two information recovery modules, and a sigmoid layer connected in sequence. The fifth nonlinear convolutional layer is used to perform preliminary filtering on the input feature map and transform it to a form similar to H. d Same size; the Sigmoid layer is used to map the output to recover the channel matrix; The information recovery module includes multiple parallel two-layer nonlinear convolutional layers. Each two-layer nonlinear convolutional layer is used to refine the input channel information feature map, extract effective features and remove invalid features. The multiple output feature maps are concatenated and then input into the sixth nonlinear convolutional layer to compress the channel dimension of the feature map to be consistent with the original input to obtain the output feature map. The original input and output are added together and finally passed through the activation layer to obtain the final output.

7. The channel information compression feedback method utilizing the implicit reciprocity of frequency division duplex channels according to claim 1, characterized in that, The training objective of the entire network framework is to minimize the following expression: Where I(z) d ;z S ) represents z d With z S Mutual information between them, I(z) S H d H u ) represents z S H d and H u Mutual information between them, p a (H d ) represents H d The actual probability distribution, p(H) d ) represents H d The reconstruction probability, q(z) S |H d H u ) and q(r S |t u ) represent the z values ​​estimated by the uplink and downlink shared information extraction modules, respectively. S The probability distribution and the estimated r by the virtual uplink and downlink shared information extraction module S The probability distributions are given by β, α, and γ, which are positive weighting coefficients, and K represents the KL divergence between the two distributions.

8. The channel information compression feedback method utilizing the implicit reciprocity of frequency division duplex channels according to claim 7, characterized in that, By relaxing the first and second terms using the non-negativity of the KL divergence and Jensen's inequality, the relaxed result of the above equation is as follows: Where q(z) d |H d ) represents the z estimated by the downlink-specific information extraction module. d The probability distribution, p(z) d ) and p(z S ) for z d and z S The actual probability distribution.

9. A channel information compression feedback system utilizing the implicit reciprocity of frequency division duplex channels, characterized in that, include: The preprocessing module transforms the uplink and downlink spatial frequency domain channel matrices into the angular time delay domain using a two-dimensional discrete Fourier transform, and then performs time-domain clipping, complex transformation, and normalization to obtain the preprocessed downlink channel information matrix H. d and uplink channel information matrix H u ; The encoding module is used to extract information from the input H using the downlink channel-specific information extraction module. d The unique information representation of the downlink channel is extracted from it, denoted as z. d The uplink and downlink channel shared information extraction module is used to extract information from the input H. d and H u The shared information representation of the uplink and downlink channels is extracted and denoted as z. S The virtual uplink and downlink channel shared information extraction module is used to extract information from the input H. u Extracting z S The approximate distribution representation is denoted as the virtual uplink / downlink shared representation r. S The downlink channel-specific information extraction module, deployed on the user side, compresses the extracted feature map containing downlink-specific information to a codeword with a given compression ratio through a fully connected network, and extracts the downlink-specific representation z through another fully connected network deployed on the base station side. d The virtual uplink and downlink channel shared information extraction module is deployed on the base station side. The decoding module is used to decode z during training. d With z S The downlink channel matrix H is recovered through the decoding network. d During testing, z will be... d With r S The downlink channel matrix H is recovered through the decoding network. d .

10. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the steps of the channel information compression feedback method utilizing the implicit reciprocity of frequency division duplex channels according to any one of claims 1-8.

Images