Channel information compression feedback method and system based on mode phase separation and spatial-temporal feature extraction

Abstract

The invention discloses a channel information compression feedback method and system based on mode-phase separation and spatial-temporal feature extraction, and the method comprises the steps: carrying out the preprocessing of uplink and downlink spatial frequency domain channel matrixes, and calculating a mode value matrix and a phase matrix of the channel matrixes; decoupling the uplink and downlink module value matrixes, extracting common information of uplink and downlink channel module values and unique information of downlink channel module values, and extracting space and time characteristics for feedback; the base station combines the module value common information and the downlink channel module value unique information to recover the downlink channel module value; module value screening is carried out, and corresponding position indexes and phases are fed back; and the base station reconstructs a downlink channel matrix by combining the downlink channel screening phase and the recovered downlink channel module value. According to the method, implicit reciprocity and time correlation of the frequency division duplex channel are fully utilized, so that a user only needs to feed back unique information of a downlink module value and a downlink main path phase, the channel feedback load is reduced, the recovery performance of channel information is improved, and the method has remarkable performance advantages.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication, and relates to compressive sensing and channel information recovery technologies, and particularly relates to a channel information compression feedback method and system based on modulus-phase separation and spatio-temporal feature extraction. Background Art

[0002] Massive MIMO has become a key technology in modern wireless communication networks. By deploying a large-scale antenna array at the base station, massive MIMO can effectively utilize spatial resources to mitigate the impact of multi-user interference, thereby improving spectral and energy efficiency; however, in order to obtain performance gains, the base station needs to obtain accurate channel state information. In a frequency division duplex (FDD) system, the uplink and downlink channels do not have reciprocity, and the channel state information of the downlink must be fed back from the user to the base station through a bandwidth-limited feedback link; with the increase in the number of antennas deployed at the base station end in a massive MIMO system of frequency division duplex, obtaining downlink channel information consumes a large amount of overhead and resources, which makes it necessary to study a channel state information (CSI) compression feedback method with small feedback overhead and high recovery accuracy; traditional compression methods include codebook-based methods and compressive sensing-based methods, but there is still a linear relationship between the channel state information feedback overhead of these compression methods and the number of antennas at the base station, which results in a significant increase in the 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 the downlink channel state information fed back by users is a promising direction, and there are currently many effective methods in this direction; recently, a compression feedback method using the implicit reciprocity between the uplink and downlink channels has been invented. Because the two-way channels in a frequency-division duplexing system actually share the same physical propagation path, there may be implicit reciprocity in the two-way channels. Using this implicit reciprocity in a large-scale MIMO system with frequency-division duplexing for MIMO compression feedback can minimize the feedback overhead and enhance the recovery of downlink channel information through the uplink channel information at the base station, improving the recovery accuracy. However, this method directly concatenates the real and imaginary parts of the uplink and downlink channels into a real channel matrix and decouples the uplink and downlink real matrices. Through visual observation, it is found that the similarity of the modulus values of the uplink and downlink channel matrices is higher than that of the real and imaginary part concatenated matrices, while the similarity between the phase matrices is not high, indicating that the common information in the modulus values accounts for a relatively high proportion and the unique information in the phases dominates. Therefore, it is inferred that decoupling and feedback of the modulus matrix and separately feeding back the phase information should be more effective; we propose to separately process and feedback the modulus and phase of the uplink and downlink channel matrices to obtain better reconstruction accuracy. And since only the modulus matrix is decoupled, the volume of the neural network can be reduced, the inference efficiency can be improved, and the computing resources and storage space at the user end can be saved. In addition, most of the existing CSI compression feedback methods based on Artificial Intelligence (AI) only extract the spatial features of CSI and ignore the temporal correlation of CSI. Summary of the Invention

[0004] Object of the Invention: In order to overcome the deficiencies in the prior art, a channel information compression feedback method and system based on modulus-phase separation and spatio-temporal feature extraction are provided, which can make full use of the implicit reciprocity of the modulus values of the uplink and downlink channels and effectively extract spatial and temporal features to solve the technical problems of low compression efficiency and recovery performance of the network.

[0005] Technical Solution: To achieve the above object, the present invention provides a channel information compression feedback method based on modulus-phase separation and spatio-temporal feature extraction, including the following steps:

[0006] S1: Concatenate the channel response vectors received on each subcarrier of the downlink and uplink into a downlink channel information matrix and an uplink channel information matrix The subscripts d and u represent the downlink and uplink respectively; perform preprocessing on and respectively to obtain an uplink channel information modulus matrix |H u |, a downlink channel information modulus matrix |H d | and a downlink channel information phase φ d ;

[0007] S2: Use the downlink channel modulus unique information extraction module to extract the unique information feature map of the downlink channel modulus from the input |H d |, denoted as W. Use the feature fusion coding module to extract the spatial feature and the temporal feature from the downlink modulus unique information representation W and add them to obtain the downlink modulus unique information compressed codeword m and feed it back. Use the feature de-fusion module to decompress the downlink modulus unique information compressed codeword m fed back to obtain the downlink modulus unique information representation, denoted as z d ;

[0008] Use the uplink and downlink channel modulus common information extraction module to extract the common information representation of the uplink and downlink channel moduli from the input |H d | and |H u |, denoted as z S , use the virtual uplink and downlink channel modulus common information extraction module to extract the representation approximately distributed with z u |, and this representation is denoted as the virtual uplink and downlink modulus common representation r S . During training, z S and z d are restored to the downlink channel modulus matrix |H S through the downlink channel modulus restoration module. During testing, z d and r d are restored to the downlink channel modulus matrix |H S through the downlink channel modulus restoration module; d |;

[0009] S3: Through the phase screening module, screen out some elements with large modulus values from the downlink channel matrix H d in descending order of element modulus value, and extract the phase φ d corresponding to the screened elements and the position index P main from the downlink channel phase matrix φ main , and feed them back to the base station;

[0010] S4: Input the restored downlink channel modulus matrix |H d |, the fed-back screened point phase φ main and the screened point position index P main into the reconstruction module to restore the downlink channel information matrix H d .

[0011] Furthermore, the preprocessing in the step S1 includes two-dimensional discrete Fourier transform, time domain cropping, modulus matrix calculation, phase matrix calculation and normalization processing. Specifically:

[0012] Two-dimensional discrete Fourier transform: Perform two-dimensional discrete Fourier transform on each instantaneous channel matrix where F a is the Fourier matrix of T×N t ×N t , F b is the Fourier matrix of T×N f ×N f . The superscript * represents the conjugate transpose of the second and third dimensions of the matrix. H' d and H' u represent the downlink channel matrix and the uplink channel matrix after Fourier transform respectively;

[0013] Time domain clipping: Since there are a large number of zero values in the channel matrix in the time domain, the third dimension of H' d and H' u is clipped, and only the first N d columns containing non-zero elements in H' u and H' c are retained, and N f >>N c . The clipped matrices are represented as H″ d and H″ u respectively;

[0014] Modulus matrix calculation: Calculate the modulus matrices of H″ d and H″ u respectively, to obtain two modulus matrices of size T×N t ×N c ×1, |H' d | and |H' u |;

[0015] Phase matrix calculation: Calculate the phase matrix of H″ d , to obtain a phase matrix of size T×N t ×N c ×1, φ d ;

[0016] Normalization: Normalize |H' d | and |H' u | respectively. By subtracting the minimum value and dividing by the difference between the maximum value and the minimum value, the normalized downlink channel modulus matrix |H d | and the uplink channel modulus matrix |H u | are obtained.

[0017] Further, the downlink channel modulus unique information extraction module in step S2 consists of two parallel representation extraction modules and a non-linear convolutional layer. The first representation extraction module consists of three non-linear convolutional layers with convolutional layer sizes of 1×3×3×c×c, 1×1×9×c×c, and 1×9×1×c×c respectively. The second representation extraction module consists of one non-linear convolutional layer with a convolutional layer size of 1×3×3×c×c. The outputs of the two representation extraction modules are concatenated in the second dimension and input into a non-linear convolutional layer of size 1×1×1×2c×c. After each non-linear convolutional layer, a batch normalization layer and a non-linear activation layer LeakyReLU(·) function are used as follows:

[0018]

[0019] where x represents the input feature map or vector of the non-linear layer;

[0020] Finally, the downlink channel modulus unique information representation W is output, with a dimension of T×N t ×N c ×1.

[0021] Further, the feature fusion and encoding module in step S2 consists of two parallel feature extraction modules, namely a spatial feature extraction module and a temporal feature extraction module. The spatial feature extraction module consists of a one-dimensional non-linear convolutional layer, a batch normalization layer, and an activation layer, which extracts the spatial features in the downlink modulus unique information and compresses them to a given compression ratio size of T×M. The temporal feature extraction module consists of a long short-term memory network, which extracts the temporal features in the downlink modulus unique information and compresses them to a given compression ratio size of T×M. The downlink channel modulus unique information feature map W is reshaped into a tensor of T×N t N c and is respectively input into the spatial feature extraction module and the temporal feature extraction module. The outputs of the two paths are added together to obtain the compressed codeword m and fed back;

[0022] The feature de-fusion module consists of a one-dimensional non-linear convolutional layer, a batch normalization layer, and an activation layer, which extracts the downlink channel modulus unique information representation z from the input compressed codeword m d .

[0023] Further, the uplink and downlink channel modulus common information extraction module in step S2 consists of two non-linear convolutional paths and a non-linear convolutional layer. The first convolutional path extracts the common information features of |H d |, which is the same as the downlink modulus unique information extraction module and consists of two representation extraction modules and a non-linear convolutional layer. The second convolutional path extracts the common information features of |H uExtract the common information features, with the same structure as above; the output features of the two paths are concatenated and then output the common information representation z through a non-linear convolutional layer of size 1×3×3×2c×2c S The mean u S And the standard deviation σ S , and multiply σ S By the sampling of a standard Gaussian distribution and add u S Finally, obtain z S ;

[0024] The virtual uplink and downlink modulus common information extraction module reuses the path for extracting the common information features of |H u | in the uplink and downlink modulus common information extraction module, and outputs the virtual common information representation r through a non-linear convolutional layer of size 1×3×3×c×2c for the output features of this path S The mean u r And the standard deviation σ r , and multiply σ r By the sampling of a standard Gaussian distribution and add u r Finally, obtain r S .

[0025] Furthermore, the downlink channel modulus recovery module in step S2 is composed of a non-linear convolutional layer, two cascaded information recovery modules, and a non-linear activation function Sigmoid(·) layer connected in sequence; the size of the non-linear convolutional layer is 1×5×5×2c×c, which is used to perform preliminary filtering on the input feature map and convert it to the same size as |H d |; the information recovery module includes 2 parallel non-linear convolutional paths. The first path consists of 3 non-linear convolutional layers, and the sizes of the convolutional layers are 1×3×3×c×7c, 1×1×9×7c×7c, and 1×9×1×7c×7c respectively. The second path consists of 2 non-linear convolutional layers, and the sizes of the convolutional layers are 1×1×5×c×7c and 1×5×1×7c×7c respectively; the feature maps output by the two paths are concatenated and then input into a non-linear convolutional layer of size 1×1×1×14c×c to compress the channel dimension of the feature map to be the same as the original input; to prevent gradient explosion, the original input and the output are added and then passed through the non-linear activation layer LeakyReLU(·) for output, and finally the output is mapped to the interval [0,1] through the Sigmoid(·) layer to obtain the finally recovered channel modulus matrix

[0026] Furthermore, the training objective of the entire network framework in step S2 is to minimize the following formula:

[0027]

[0028] Among them, I(z d ; zS ) represents z d and z S The mutual information between, I(z S ; |H d |; |H u |) represents z S , |H d | and |H u | The mutual information between, pa(|H d |) represents |H d | The actual probability distribution of, p(|H d |) represents |H d | The reconstruction probability of, q(z S ||H d |, |H u |) and q(r S ||H u |) respectively represent the probability distribution of z estimated by the uplink and downlink modulus common information extraction module S and the probability distribution of r estimated by the virtual uplink and downlink modulus common information extraction module S The relative entropy between the two distributions, K, and β, α, and γ are positive weighting coefficients respectively;

[0029] The first and second terms of the above formula are relaxed, and transformed into minimizing the upper bounds of the first and second terms of the above formula, and the relaxed result of the above formula is obtained as follows:

[0030]

[0031] Among them, q(z d ||H d |) represents the probability distribution of z estimated by the downlink unique information extraction module d The probability distributions of p(z d ) and p(z S ) are the actual probability distributions of z d and z S The third term of the above formula is replaced by the mean square error between the true downlink channel modulus matrix and the reconstructed downlink channel modulus matrix multiplied by the weighting coefficient; the network parameters are initialized before training, and optimized using the gradient descent method during the training process.

[0032] Furthermore, in step S3, the phase screening module sorts the elements in the downlink channel information matrix in descending order of modulus, screens the first D elements, and takes the phase φ main and the position index P main fed back to the base station; φ main is a D×1 vector, and each element is a real number from -π to π, representing the radian phase of the screening point; P mainis a D×1 integer vector, and each element is an integer in the range of 1 to N t N c The integers within the range represent the positions of the screening points in the N t ×N c matrix.

[0033] Further, in step S4, the reconstruction module uses |H d | as the modulus value of the reconstructed channel matrix, and uses the phase φ main to fill the phase of the elements at the position P main of the reconstruction matrix, and uses the constant phase to fill the phases of the elements at the remaining positions, and finally restores the downlink channel matrix H d .

[0034] Based on the method of the present invention, the present invention also provides a channel information compression feedback system based on modulus-phase separation and spatio-temporal feature extraction, including:

[0035] A preprocessing module for transforming the uplink and downlink spatial-frequency domain channel matrices into the angular-delay domain through two-dimensional discrete Fourier transform, and performing preprocessing to obtain the uplink and downlink channel information modulus matrices |H u | and |H d | and the downlink channel information phase φ d ;

[0036] A modulus decoupling feedback module for using the downlink channel modulus unique information extraction module to extract the unique information feature map of the downlink channel modulus from the input |H d |, denoted as W, using the feature fusion coding module to extract the spatial and temporal features from W and add them to obtain the compressed codeword m and feedback, using the feature de-fusion module to decompress the downlink modulus unique information representation from m, denoted as z d , using the uplink and downlink channel modulus common information extraction module to extract the common information representation of the uplink and downlink channel moduli from the input |H d | and |H u |, denoted as z S , using the virtual uplink and downlink channel modulus common information extraction module to extract the representation approximately distributed with z u from the input |H S |, and this representation is denoted as the virtual uplink and downlink modulus common representation r S . During training, z d and z S are used to restore the downlink channel modulus matrix |H d | through the downlink channel modulus recovery module. During testing, z d and r S are used to restore |H d | through the downlink channel modulus recovery module;

[0037] Phase screening feedback module, used to use the phase screening module to d | Screen out some elements with larger modulus values in descending order of modulus value, and extract the phases φ d corresponding to the screened elements and the position indexes P main at the corresponding positions from the downlink channel phase matrix φ main and feedback them to the base station; main

[0038] Reconstruction module, used to reconstruct the downlink channel information matrix H from the recovered |H d |, the feedback screened point phases φ main and the screened point position indexes P main d .

[0039] For the frequency division duplex channel information data with implicit reciprocity, the present invention first transforms the uplink and downlink spatial frequency domain channel matrices to the angle-delay domain through two-dimensional Fourier transform and performs cropping, and then calculates the modulus matrix and phase matrix of the channel matrix; uses the method of decoupled representation learning to decouple the uplink and downlink modulus matrices, extracts the common information of the uplink and downlink channel moduli and the unique information of the downlink channel modulus, and extracts the spatial and temporal features of the unique information of the downlink modulus for feedback; the base station combines the common information of the modulus and the unique information of the downlink channel modulus to recover the downlink channel modulus; uses the screening module to screen out the points with larger moduli, and feedbacks the corresponding position indexes and phases, and fills the phases at other positions with constants; the base station combines the screened phases of the downlink channel and the recovered downlink channel modulus to reconstruct the downlink channel matrix.

[0040] Advantageous effects: Compared with the prior art, the present invention makes full use of the implicit reciprocity and time correlation of the frequency division duplex channel, enabling the user to only feedback the unique information of the downlink modulus and the phase of the downlink main path, reducing the channel feedback load and improving the recovery performance of the channel information, improving the accuracy of compressed feedback, and having significant performance advantages. Description of the Drawings

[0041] Figure 1 is the neural network structure block diagram of the method of the present invention;

[0042] Figure 2 is the structure block diagram of the feature coupling coding module;

[0043] Figure 3 is the structure block diagram of the phase screening module;

[0044] Figure 4 is the schematic diagram of the performance of other neural networks and the present invention under the Quadriga channel model conditions. Detailed Embodiments

[0045] The present invention will be further illustrated below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, various equivalent forms of modification by those skilled in the art fall within the scope defined by the appended claims of this application.

[0046] Embodiment 1:

[0047] This embodiment provides a channel information compression feedback method based on modulus-phase separation and spatio-temporal feature extraction. This method is implemented based on the neural network architecture as shown in Figure 1 . This network consists of a modulus feedback network and a phase feedback network. The modulus feedback network includes a downlink channel modulus unique information extraction module, an uplink-downlink channel modulus common information extraction module, a virtual uplink-downlink channel modulus common information extraction module, a feature fusion coding module, a feature de-fusion module, and a downlink channel modulus recovery module. The downlink modulus unique module consists of a non-linear convolutional layer and two representation extraction modules. The input is the downlink channel modulus |H d | and the output is the feature map W. The uplink-downlink modulus common information extraction module consists of two parallel convolutional paths and a non-linear convolutional layer. The composition of each convolutional path is the same as that of the downlink modulus unique information extraction module. The downlink channel modulus |H d | and the uplink channel modulus |H u | are respectively input into the two paths. The results of the two paths are added and finally output the uplink-downlink modulus common representation z S after passing through the non-linear convolutional layer. The virtual uplink-downlink modulus common information extraction module reuses the convolutional path that extracts the common information features of |H u | in the uplink-downlink modulus common information module and additionally cascades a non-linear convolutional layer. The input is the uplink channel modulus |H u | and the output is the virtual uplink-downlink modulus common representation r S ; z S is only used during training, and r S is used during testing. The feature coupling coding module consists of a non-linear convolutional layer and a long short-term memory network in parallel. The input is the downlink channel modulus unique information W, and the output is the compressed codeword m after feature coupling coding. The feature de-coupling module consists of a non-linear convolutional layer. The input is the feedback codeword m, and the output is the decompressed and extracted downlink unique representation z d d .

[0048] Based on the above neural network architecture, the specific process of a channel information compression feedback method based on modulus-phase separation and spatio-temporal feature extraction includes:

[0049] Step 1. Preprocessing: In a large-scale MIMO frequency-division duplex communication system with OFDM modulation as the waveform modulation method, at the receiving end, the received channel response vectors on each subcarrier of the downlink and uplink are respectively concatenated into the downlink channel information matrix and the uplink channel information matrix The subscripts d and u represent the downlink and uplink respectively; for and Two-dimensional discrete Fourier transform, time-domain cropping, modulus value matrix calculation, phase matrix calculation, and normalization processing are respectively performed to obtain the preprocessed downlink and uplink channel information modulus matrices |H u | and |H d | and the downlink channel information phase φ d ;

[0050] Step 1.1: For an OFDM system with the number of subcarriers being N f and the number of antennas at the base station being N t , and a single antenna is configured at the user end. There are T instantaneous channel state information within the coherence time. Obtain the complete downlink channel matrix within the coherence time, denoted as The base station obtains the complete uplink channel matrix and Both have a dimension size of T×N t ×N f ;

[0051] Step 1.2: Perform a series of preprocessing operations on and to obtain the preprocessed downlink channel matrix and uplink channel matrix. The specific operations are as follows:

[0052] Two-dimensional discrete Fourier transform: Perform a two-dimensional discrete Fourier transform on each instantaneous channel matrix where F a is the Fourier matrix of T×N t ×N t , F b is the Fourier matrix of T×N f ×N f , the superscript * represents the conjugate transpose of the second and third dimensions of the matrix, and H' d and H' u represent the downlink channel matrix and uplink channel matrix after Fourier transform respectively;

[0053] Time-domain cropping: Since there are a large number of zero values in the channel matrix in the time domain, crop the third dimension of H' d and H' u , and only retain the first N containing non-zero elements in H' d and H' u inc columns, and there are N f >> N c , the cropped matrices are respectively denoted as H″ d and H″ u ;

[0054] Find the modulus value: Find the modulus value matrices of H″ d and H″ u respectively, and obtain two modulus value matrices of size T×N t ×N c ×1, namely |H' d | and |H' u |;

[0055] Find the phase: Find the phase matrix of H″ d , and obtain a phase matrix of size T×N t ×N c ×1, namely φ' d ;

[0056] Normalization processing: Perform normalization processing on |H' d | and |H' u | respectively. By subtracting the minimum value and dividing by the difference between the maximum value and the minimum value, obtain the normalized downlink channel modulus value matrix |H d | and the uplink channel modulus value matrix |H u |.

[0057] Step 2, Modulus decoupling feedback: Use the downlink channel modulus value unique information extraction module to extract the unique information feature map of the downlink channel modulus value from the input |H d |, denoted as W. Use the feature fusion coding module to extract the spatial feature and the temporal feature from the downlink modulus value unique information representation W and add them to obtain the downlink modulus value unique information compression codeword m and feedback. Use the feature de-fusion module to decompress the downlink modulus value unique information compression codeword m fed back to obtain the downlink modulus value unique information representation, denoted as z d . Use the uplink and downlink channel modulus value common information extraction module to extract the common information representation of the uplink and downlink channel modulus values from the input |H d | and |H u |, denoted as z S . Use the virtual uplink and downlink channel modulus value common information extraction module to extract the representation approximately distributed with z u from the input |H S . This representation is denoted as the virtual uplink and downlink modulus value common representation r S . During training, z d and z S are restored to the downlink channel modulus value matrix |H d | through the downlink channel modulus value recovery module. During testing, z d and rS Restore it to the downlink channel modulus matrix |H through the downlink channel modulus restoration module d ;

[0058] Step 2.1: Construct a downlink channel modulus unique information extraction module, deployed on the user side, with the input being |H d | for extracting the unique information feature map W of the downlink channel modulus matrix; construct a feature fusion and coding module, deployed on the user side, inputting W to extract the spatial and temporal dimension features of the downlink modulus unique information and fuse them, and compress the fused features to the codeword m of the given compression ratio size; construct a feature de-fusion module, deployed on the base station side, inputting m to decompress and output the downlink channel modulus unique information representation z d ; construct a downlink and uplink channel modulus common information extraction module for training, not deployed in actual applications, inputting |H d | and |H u | to extract the common information z of the downlink channel modulus and the uplink channel modulus S ; construct a virtual downlink and uplink channel modulus common information extraction module, deployed on the base station side, inputting |H u | for extracting the virtual downlink and uplink modulus common information r S ; construct a downlink channel modulus restoration module, deployed on the base station side, and restore the concatenated feature map of z d and z S to the original downlink channel modulus during training; restore the concatenated feature map of z d and r S to the original downlink channel modulus during testing.

[0059] Step 2.2: The downlink channel modulus unique information extraction module consists of two parallel representation extraction modules and a non-linear convolutional layer; the first representation extraction module consists of 3 non-linear convolutional layers, and the sizes of the convolutional layers are 1×3×3×1×1, 1×1×9×1×1, 1×9×1×1×1 respectively, and the second representation extraction module consists of a non-linear convolutional layer, and the size of the convolutional layer is 1×3×3×1×1; concatenate the outputs of the two representation extraction modules in the second dimension and input them into the non-linear convolutional layer with the size of 1×1×1×2×1; use the batch normalization layer and the non-linear activation layer LeakyReLU(·) function after each non-linear convolutional layer as follows:

[0060]

[0061] where x represents the input feature map or vector of the non-linear layer; finally, output the downlink channel modulus unique information representation W, with the dimension of T×N t ×N c ×1.

[0062] Step 2.3: As Figure 2 shown, the feature coupling encoding module consists of parallel non-linear convolutional branches and long short-term memory network branches. The non-linear convolutional branch consists of a convolutional layer, a batch normalization layer, and an activation layer, which extracts and compresses the spatial correlation features in the unique information of the downlink modulus value. The long short-term memory network branch stacks five layers of long short-term memory networks, which extracts and compresses the time correlation features in the unique information of the downlink modulus value. Input the unique information W of the downlink channel modulus value, pass through two branches and add the results to obtain the compressed codeword m;

[0063] In this embodiment, the feature fusion encoding module consists of two parallel feature extraction modules, namely the spatial feature extraction module and the time feature extraction module. The spatial feature extraction module consists of a one-dimensional non-linear convolutional layer, a batch normalization layer, and an activation layer, which extracts and compresses the spatial features in the unique information of the downlink modulus value to a given compression ratio size of T×M. The time feature extraction module consists of a long short-term memory network, which extracts and compresses the time features in the unique information of the downlink modulus value to a given compression ratio size of T×M. Reshape the feature map W of the unique information of the downlink channel modulus value into a tensor of T×N t N c and input it into the spatial feature extraction module and the time feature extraction module respectively. Add the outputs of the two paths to obtain the compressed codeword m and feedback it. The feature de-fusion module consists of a one-dimensional non-linear convolutional layer, a batch normalization layer, and an activation layer, which extracts the representation z of the unique information of the downlink channel modulus value from the input compressed codeword m d .

[0064] Step 2.4: The common information extraction module for the uplink and downlink channel modulus values consists of two non-linear convolutional paths and a non-linear convolutional layer. The first convolutional path extracts the common information features for |H d |, which is the same as the unique information extraction module for the downlink modulus value, and consists of two representation extraction modules and a non-linear convolutional layer. The second convolutional path extracts the common information features for |H u |, and the structure is the same as above. The output features of the two paths are concatenated and then output the common information representation z S of the mean value u S and the standard deviation σ S through a non-linear convolutional layer of size 1×3×3×2×2. Multiply σ S by a sample of a standard Gaussian distribution and add u S to finally obtain z S . The virtual common information extraction module for the uplink and downlink modulus values reuses the path for extracting the common information features for |H u | in the common information extraction module for the uplink and downlink modulus values, and outputs the virtual common information representation r S of the mean value u rand standard deviation σ r , and get r in the same way S .

[0065] Step 2.5: The downlink channel modulus recovery module consists of a nonlinear convolution layer, two serial information recovery modules, and a nonlinear activation function Sigmoid (·) layer connected in sequence; the size of the nonlinear convolution layer is 1×5×5×2×1, which is used to perform preliminary filtering on the input feature map and convert it into a signal with the same resolution as |H d |same size; the information recovery module includes two parallel nonlinear convolution paths. The first path consists of three nonlinear convolutional layers with sizes of 1×3×3×1×7, 1×1×9×7×7, and 1×9×1×7×7, respectively. The second path consists of two nonlinear convolutional layers with sizes of 1×1×5×1×7 and 1×5×1×7×7, respectively. The feature maps output by the two paths are concatenated and input into a nonlinear convolutional layer of size 1×1×1×14×1, so that the channel dimension of the feature map is compressed to be consistent with the original input. To prevent gradient explosion, the original input and output are added and then output through a nonlinear activation layer LeakyReLU(·). Finally, the output is mapped to the interval [0,1] by a Sigmoid(·) layer to obtain the final recovered channel modulus matrix.

[0066] Step 2.6: By minimizing z d With z S The mutual information I(z d ;z S ) to fully decouple the unique information and shared information of the downstream module value; by maximizing z S 、|H d |和|H u The mutual information I(z S ;|H d |;|H u |) to extract the shared information of uplink and downlink channel modulus values to the greatest extent, thereby reducing the information that users need to feedback; minimize z S With r S The relative entropy between the distributions of r S Approximation z S , so that the shared information of uplink and downlink modulus values similar to those in training can be obtained during testing; the reconstruction accuracy of the downlink channel can be improved by minimizing the reconstruction error of the downlink channel modulus value. Here, the model is to minimize |H d |The negative value of the reconstruction probability, so the training objective of the entire network framework is to minimize the following formula:

[0067]

[0068] Among them, p a (|Hd |) means |H d The actual probability distribution of |, p(|H d |) means |H d The reconstruction probability of |,q(z S ||H d |,|H u |) and q(r S ||H u |) represent the z estimated by the uplink and downlink module common information extraction modules S The probability distribution of the virtual uplink and downlink modulus values is estimated by the information extraction module r S The probability distribution of , β, α, γ are positive weight coefficients, and K represents the relative entropy between the two distributions;

[0069] Relax the first and second terms of the above equation and transform it into minimizing the upper bounds of the first and second terms of the above equation. The result after relaxation is as follows:

[0070]

[0071] The value of the last item represents the reconstructed mean square error of the downlink channel modulus The network parameters are initialized before training begins, and the gradient descent method is used for optimization during training. The relaxation result obtained above is used as the loss function of the neural network training.

[0072] Step 3: Phase screening feedback: Downlink channel matrix H d According to the element modulus from large to small, select the elements with large modulus values from the downlink channel phase matrix φ d Extract the phase φ corresponding to the screening element main and position index P main , and fed back to the base station;

[0073] like Figure 3 As shown, the phase screening module inputs the downlink channel modulus matrix |H d | and the phase matrix φ d , according to the modulus matrix, filter the elements with large modulus values in the downlink channel matrix from large to small, and extract the phase φ at the corresponding position main and position index P main Output and feedback;

[0074] In this embodiment, differences in factors such as the center frequency, initial phase, and path loss factor of the uplink and downlink channels result in the relative independence of the phase of the uplink and downlink channel information matrices. Therefore, the phase is fed back separately. Only the phase of the points representing the main path with relatively large modulus values has a greater impact. Therefore, only the phases of the main path points are selected for feedback, and the phases of the remaining points are approximated by constant phases. The elements in the downlink channel information matrix are sorted in descending order of modulus, and the first D elements are selected. The phases φ main and the position index P main are fed back to the base station; φ main is a D×1 vector, and each element is a real number from -π to π, representing the radian phase of the selected points; P main is a D×1 integer vector, and each element is an integer in the range from 1 to N t N c The integers within the range indicate the positions of the selected points in the N t ×B c matrix;

[0075] Step 4, Reconstruction: Input the restored downlink channel modulus matrix |H d |, the fed-back selected point phase φ main and the selected point position index P main into the reconstruction module to restore the downlink channel information matrix H d .

[0076] In this embodiment, the restored downlink channel modulus matrix |H d |, the fed-back selected point phase φ main and the selected point index P main are input into the reconstruction module to restore the downlink channel information matrix H d ; The reconstruction module uses |H d | as the modulus of the reconstructed channel matrix, fills the phase of the elements at the position P main of the reconstructed matrix with φ main , and fills the phases of the elements at the remaining positions with the constant phase to finally restore the downlink channel matrix H d .

[0077] Embodiment 2:

[0078] Based on the method of Embodiment 1, this embodiment provides a channel information compression feedback system based on modulus-phase separation and spatio-temporal feature extraction, including:

[0079] A preprocessing module, which is used to transform the uplink and downlink spatial frequency domain channel matrices to the angular delay domain through two-dimensional discrete Fourier transform, and perform preprocessing to obtain the uplink and downlink channel information modulus matrices |H u | and |H d| and the phase φ of the downlink channel information d ;

[0080] The modulus decoupling feedback module is used to extract the unique information feature map of the downlink channel modulus from the input |H d |, denoted as W, extract spatial and temporal features from W using the feature fusion coding module, add them to obtain the compressed codeword m and feedback it, and decompress the downlink modulus unique information representation, denoted as z, from m d , use the common information extraction module of the uplink and downlink channel moduli to extract the common information representation of the uplink and downlink channel moduli from the input |H d | and |H u |, denoted as z S , use the virtual common information extraction module of the uplink and downlink channel moduli to extract the representation approximately distributed with z u | from the input |H S , and this representation is denoted as the virtual common uplink and downlink modulus representation r S , during training, z d and z S are used to recover the downlink channel modulus matrix |H d | through the downlink channel modulus recovery module, and during testing, z d and r S are used to recover | d | through the downlink channel modulus recovery module; the downlink channel modulus unique information extraction module and the feature fusion coding module are deployed on the user side, the virtual common information extraction module of the uplink and downlink channel moduli, the feature de-fusion module, and the downlink channel modulus recovery module are deployed on the base station side, and the common information extraction module of the uplink and downlink channel moduli is not actually deployed;

[0081] The phase screening feedback module is used to use the phase screening module to screen out some elements with large moduli from | d | in descending order of modulus, and extract the phase φ d corresponding to the screened elements and the position index P main from the downlink channel phase matrix φ main , and P main and feedback them to the base station; the phase screening module is deployed on the user side;

[0082] The reconstruction module is used to reconstruct the downlink channel information matrix H d | from the recovered |H main , the feedback screened point phase φ main and the screened point position index P d ; the reconstruction module is deployed on the base station side.

[0083] Embodiment 3:

[0084] To verify the effectiveness and effect of the solution of the present invention, this embodiment conducts a comparison through experiments, and the specific data and analysis are as follows:

[0085] Figure 4 For the performance comparison between other neural network methods and the present invention under the channel model generated by Quadriga. Quadriga is a tool for generating highly accurate and controllable radio frequency channels, and its full name is Quasi-Deterministic Radio channel Generator. Among them, the comparison method 1 is the CRNet network structure, the comparison method 2 is the DrCsiNet network structure, and the comparison method 3 is the method of deleting the feature fusion module from the framework proposed by the present invention. It can be seen from Figure 4 that the method of the present invention has performance gains compared with other neural network methods under different compression ratios, and the gains are more obvious under high compression ratios.

[0086] It can be seen that compared with the prior art, the present invention makes full use of the implicit reciprocity and time correlation of the frequency division duplex channel, enabling the user to only feedback the unique information of the downlink modulus value and the downlink main path phase, reducing the channel feedback load and improving the channel information recovery performance, and improving the accuracy of compressed feedback, having significant performance advantages.

Claims

1. A channel information compression feedback method based on modulo-phase separation and spatio-temporal feature extraction, characterized in that It includes the following steps: S1: Concatenate the channel response vectors received on each subcarrier for downlink and uplink into a downlink channel information matrix and an uplink channel information matrix The subscripts d and u represent downlink and uplink respectively; for and perform preprocessing respectively to obtain the uplink channel information magnitude matrix |H u |, the downlink channel information magnitude matrix |H d | and the downlink channel information phase φ d ; S2: Use the downlink channel modulus unique information extraction module to extract the unique information feature map of the downlink channel modulus from the input |H d |, denoted as W. Use the feature fusion coding module to extract the spatial feature and the temporal feature from the downlink modulus unique information representation W and add them to obtain the downlink modulus unique information compressed codeword m and feed it back. Use the feature de-fusion module to decompress the downlink modulus unique information compressed codeword m fed back to obtain the downlink modulus unique information representation, denoted as z d ; Use the common information extraction module of the uplink and downlink channel modulus to extract the common information representation of the uplink and downlink channel modulus from the input |H d | and |H u |, denoted as z S ; use the virtual common information extraction module of the uplink and downlink channel modulus to extract the representation approximately distributed with z u | from the input |H S , and this representation is denoted as the virtual common modulus representation r S ; during training, restore z d and z S into the downlink channel modulus matrix |H d | through the downlink channel modulus restoration module; during testing, restore z d and r S into the downlink channel modulus matrix |H d |; S3: Through the phase screening module, for the downlink channel matrix H d screen a part of the elements with larger modulus values from largest to smallest according to the element modulus values, and extract the phase φ d corresponding to the screened elements and the position index P main from the downlink channel phase matrix φ, and feed them back to the base station; main ​ S4: Input the restored downlink channel modulus matrix |H d , the phase φ of the feedback screening point main and the position index P of the screening point main into the reconstruction module to restore the downlink channel information matrix H d .

2. The channel information compression and feedback method based on mode-phase separation and spatio-temporal feature extraction according to claim 1, characterized in that The preprocessing in step S1 includes two-dimensional discrete Fourier transform, time-domain cropping, modulus value matrix calculation, phase matrix calculation, and normalization processing. Specifically: Two-dimensional discrete Fourier transform: Perform a two-dimensional discrete Fourier transform on each instantaneous channel matrix where F a is a Fourier matrix of T×N t ×N t , and F b is a Fourier matrix of T×N f ×N f . The superscript * represents the conjugate transpose of the second and third dimensions of the matrix. H' d and H' u represent the downlink channel matrix and the uplink channel matrix after Fourier transform respectively; Time-domain cropping: Crop the third dimension of H' d and H' u to only retain the first N d and H' u columns that contain non-zero elements, where N c >> N f c The cropped matrices are respectively denoted as H” d and H” u ;​ Modulus matrix: Calculate the modulus matrices of H” d and H” u respectively, obtaining two modulus matrices of size T×N t ×N c ×1, namely |H' d | and |H' u |; Find the phase matrix: Find the phase matrix of H' d ', and obtain a phase matrix φ of size T×N t ×N c ×1 d ; Normalization: Normalize |H' d | and |H' u | respectively. The normalized downlink channel magnitude matrix |H d | and uplink channel magnitude matrix |H u | are obtained by subtracting the minimum value and dividing by the difference between the maximum value and the minimum value.

3. A channel information compression feedback method based on mode-phase separation and spatio-temporal feature extraction according to claim 1, characterized in that, In step S2, the downlink channel modulus unique information extraction module consists of two parallel representation extraction modules and a non-linear convolutional layer. The first representation extraction module consists of 3 non-linear convolutional layers with convolutional layer sizes of 1×3×3×c×c, 1×1×9×c×c, and 1×9×1×c×c respectively. The second representation extraction module consists of a non-linear convolutional layer with a convolutional layer size of 1×3×3×c×c. The outputs of the two representation extraction modules are concatenated in the second dimension and input into a non-linear convolutional layer with a size of 1×1×1×2c×c. After each non-linear convolutional layer, a batch normalization layer and a non-linear activation layer LeakyReLU(·) function are used as follows: where x represents the input feature map or vector of the non-linear layer; The final output of the downlink channel modulus unique information is represented by W, with dimensions of T×N t ×N c ×1.

4. A channel information compression and feedback method based on mode-phase separation and spatio-temporal feature extraction according to claim 1, characterized in that In step S2, the feature fusion and encoding module is composed of two parallel feature extraction modules, namely the spatial feature extraction module and the temporal feature extraction module; The spatial feature extraction module consists of a one-dimensional non-linear convolutional layer, a batch normalization layer, and an activation layer, which extracts the spatial features in the unique information of the downlink modulus value and compresses them to the given compression rate size T×M. The temporal feature extraction module consists of a long short-term memory network, which extracts the temporal features in the unique information of the downlink modulus value and compresses them to the given compression rate size T×M. The feature map W of the unique information of the downlink channel modulus value is reshaped into a tensor of T×N t N c and is respectively input into the spatial feature extraction module and the temporal feature extraction module. The two outputs are added together to obtain the compressed codeword m and fed back The feature solution fusion module consists of a one-dimensional non-linear convolutional layer, a batch normalization layer, and an activation layer, and extracts the unique information representation z of the downlink channel modulus value from the input compressed codeword m d .

5. A channel information compression feedback method based on mode-phase separation and spatio-temporal feature extraction according to claim 1, characterized in that In the step S2, the common information extraction module for the uplink and downlink channel modulus values consists of two non-linear convolutional paths and a non-linear convolutional layer. The first convolutional path extracts the common information features from |H d |, which is the same as the exclusive information extraction module for the downlink modulus value and consists of two representation extraction modules and a non-linear convolutional layer. The second convolutional path extracts the common information features from |H u |; the output features of the two paths are concatenated and then output the common information representation z S through a non-linear convolutional layer with a size of 1×3×3×2c×2c. The mean u S and the standard deviation σ S are obtained, and σ S is multiplied by a sample from a standard Gaussian distribution and added to u S to finally obtain z S ; The virtual uplink and downlink modulus common information extraction module reuses the path for extracting common information features in the uplink and downlink modulus common information extraction module for |H u |, and outputs the virtual common information representation r through a non-linear convolutional layer of size 1×3×3×c×2c for the output features of this path S to obtain the mean value u r and the standard deviation σ r . Then multiply σ r by a sample from a standard Gaussian distribution and add u r to finally obtain r S .

6. The channel information compression feedback method based on module-phase separation and spatio-temporal feature extraction according to claim 1, wherein In the step S2, the downlink channel modulus recovery module is composed of a non-linear convolution layer, two cascaded information recovery modules, and a non-linear activation function Sigmoid(·) layer connected in sequence; the size of the non-linear convolution layer is 1×5×5×2c×c, which is used to preliminarily filter the input feature map and convert it to the same size as |H d |; the information recovery module includes two parallel non-linear convolution paths. The first path consists of three non-linear convolution layers with sizes of 1×3×3×c×7c, 1×1×9×7c×7c, and 1×9×1×7c×7c respectively. The second path consists of two non-linear convolution layers with sizes of 1×1×5×c×7c and 1×5×1×7c×7c respectively; the feature maps output by the two paths are concatenated and then input into a non-linear convolution layer with a size of 1×1×1×14c×c to compress the channel dimension of the feature map to be consistent with the original input; after adding the original input and the output and passing through the non-linear activation layer LeakyReLU(·), the output is finally mapped to the interval [0,1] through the Sigmoid(·) layer to obtain the finally recovered channel modulus matrix.

7. A channel information compression and feedback method based on mode-phase separation and spatio-temporal feature extraction according to claim 1, characterized in that In step S2, the training objective of the entire network framework is to minimize the following formula: where, I(z d ; z S ) represents the mutual information between z d and z S , I(z S ; |H d |; |H u |) represents the mutual information among z S , |H d | and |H u |, p a (|H d |) represents the actual probability distribution of |H d |, p(|H d |) represents the reconstructed probability of |H d |, q(z S ||H d |,|H u |) and q(r S ||H u |) respectively represent the probability distributions of z S estimated by the up - down link modulus common information extraction module and the probability distribution of r S estimated by the virtual up - down link modulus common information extraction module, K represents the relative entropy between the two distributions, and β, α, γ are positive weighting coefficients respectively; The first and second terms of the above formula are relaxed and transformed into minimizing the upper bounds of the first and second terms of the above formula, and the relaxed result of the above formula is obtained as follows: where q(z d ||H d |) represents the probability distribution of z estimated by the downlink unique information extraction module, p(z d ) and p(z d ) are the actual probability distributions of z S and z d and z S , and the third term in the above formula is replaced by the mean square error between the true downlink channel modulus matrix and the reconstructed downlink channel modulus matrix multiplied by the weighting coefficient; the network parameters are initialized before the training starts, and the gradient descent method is used for optimization during the training process.

8. A channel information compression feedback method based on modulo-phase separation and spatio-temporal feature extraction according to claim 1, characterized in that In step S3, the phase screening module sorts the elements in the downlink channel information matrix in descending order according to the modulus value, selects the first D elements, and extracts the phase φ main and the position index P main and feeds them back to the base station; φ main is a D×1 vector, and each element is a real number between -π and π, representing the radian phase of the screening point; P main is a D×1 integer vector, and each element is an integer within the range of 1 to N t N c and represents the position of the screening point in the N t ×N c matrix.

9. A channel information compression and feedback method based on mode-phase separation and spatio-temporal feature extraction according to claim 1, characterized in that In the step S4, the reconstruction module uses |H d | as the modulus value of the reconstructed channel matrix, and uses φ main to fill the phases of the elements at the position P main of the reconstruction matrix, and uses a constant phase to fill the phases of the elements at the remaining positions, and finally restores the downlink channel matrix H d .

10. A channel information compression and feedback system based on modulo-phase separation and spatio-temporal feature extraction, characterized in that, It includes: A preprocessing module, configured to transform the uplink and downlink spatial frequency domain channel matrices into the angular delay domain through two-dimensional discrete Fourier transform, and perform preprocessing to obtain the magnitude matrices |H u | and |H d | and the downlink channel information phase φ d ; The modulus decoupling feedback module is used to extract the unique information feature map of the downlink channel modulus from the input |H d | using the downlink channel modulus unique information extraction module, denoted as W. The spatial and temporal features are extracted from W using the feature fusion coding module and added to obtain the compressed codeword m and then fed back. The downlink modulus unique information representation, denoted as z, is decompressed from m using the feature de-fusion module d . The common information representation of the uplink and downlink channel moduli is extracted from the input |H d | and |H u | using the uplink and downlink channel modulus common information extraction module, denoted as z S . The representation approximately distributed with z u is extracted from the input |H S using the virtual uplink and downlink channel modulus common information extraction module, and this representation is denoted as the virtual uplink and downlink modulus common representation r S . During training, z d and z S are used to recover the downlink channel modulus matrix |H d | through the downlink channel modulus recovery module. During testing, z d and r S are used to recover |H d | through the downlink channel modulus recovery module; Phase screening feedback module, which is used to use the phase screening module to screen out some elements with large modulus values from |H d in descending order of modulus value, and extract the phase φ d corresponding to the screened elements and the position index P main from the downlink channel phase matrix φ main , P main and feed them back to the base station; A reconstruction module for reconstructing the |H from the recovered d |, the phase φ of the feedback screening point main and the screening point position index P main to reconstruct the downlink channel information matrix H d .

Images