A CSI feedback method for massive MIMO systems based on deep learning

Abstract

The present invention belongs to the technical field of intelligent mobile communication, and particularly relates to a CSI feedback method of a large-scale MIMO system based on deep learning, comprising: obtaining a CSI channel matrix of a downlink in a large-scale MIMO system under an FDD standard; performing a two-dimensional DFT transformation on the obtained CSI channel matrix of the downlink to obtain a CSI matrix that is sparse in an angle delay domain, and obtaining a truncated matrix according to the matrix; constructing an encoder of a CSI feedback network at a user end, and converting the truncated matrix into a codeword vector through the encoder; constructing a decoder of the CSI feedback network at a base station end, and passing the codeword vector obtained by the encoder through the decoder to obtain an estimated value of the truncated matrix; the present invention avoids the feedback network from being too deep, and adopts a horizontal expansion method, that is, increasing the generalization of the feedback network through multi-resolution convolution, and increasing the feature extraction capability of each convolution kernel through asymmetric convolution.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent mobile communication technology, and in particular to a channel state information (CSI) feedback method of a large-scale multiple-input multiple-output (MIMO) system based on deep learning. Background Art

[0002] With the rapid growth of global mobile data traffic, people's demand for data transmission rate and reliability has further increased. Therefore, in order to cope with this development trend, 5G communication systems have proposed three core key technologies: ultra-dense heterogeneous cellular networks, millimeter wave communications, and massive multiple-input multiple-output. Among them, massive MIMO refers to a communication system in which hundreds of antennas are used at the base station to simultaneously serve multiple user cells. Compared with traditional MIMO systems, massive MIMO systems can increase the capacity of communication systems by several times, improve energy transmission efficiency, and reduce air interface delay. The many advantages of the above-mentioned massive MIMO systems are based on the fact that the base station can accurately obtain channel state information and use precoding to eliminate interference between multiple users. However, for massive MIMO systems under the frequency division duplexing (FDD) standard, the uplink and downlink operate at different frequencies, so the downlink CSI is obtained by the user end and transmitted back to the base station through the feedback link. Considering that the base station uses a large number of antennas, feedback of complete CSI will result in huge resource overhead. Therefore, in order to ensure the timeliness of CSI, the feedback delay must be reduced as much as possible.

[0003] On the one hand, the traditional CSI feedback method is based on the codebook, but the number of antennas in the large-scale MIMO system is large, which increases the size and design complexity of the codebook. On the other hand, there is a CSI feedback method based on Compressed Sensing (CS), but the disadvantage of the CS feedback method is that it requires a priori assumptions about the channel, that is, it is very dependent on the sparse characteristics of the CSI matrix, but in actual scenarios, it is impossible to achieve theoretical complete sparsity; the CS recovery algorithm involves a large number of matrix iterative calculations. As the number of antennas increases, the computational complexity increases, affecting the real-time performance of the actual communication system. Therefore, the traditional CSI feedback method cannot be well applied to large-scale MIMO systems. Summary of the invention

[0004] In order to make the feedback network have higher accuracy while reducing its calculation amount and parameter amount, the present invention proposes a CSI feedback method for a large-scale MIMO system based on deep learning, comprising the following steps:

[0005] S1. In a massive MIMO system under FDD standard, obtain the CSI channel matrix of the downlink when the base station antenna adopts a uniform linear array with a half-wavelength spacing and the user end is a single user;

[0006] S2. Perform a two-dimensional DFT transform on the acquired downlink CSI channel matrix to obtain a sparse CSI matrix in the angle delay domain;

[0007] S3, the first N of the CSI matrix obtained in the angle delay domain sparse a The rows are truncated to obtain a truncated matrix;

[0008] S4, constructing an encoder of the CSI feedback network at the user end, and converting the truncation matrix into a codeword vector through the encoder;

[0009] S5. Construct a decoder of the CSI feedback network at the base station end, and pass the codeword vector obtained by the encoder through the decoder to obtain an estimated value of the truncation matrix;

[0010] S6, the encoder and decoder are jointly trained in an end-to-end manner so that the output of the feedback network With the original matrix H a The difference is as small as possible, and the trained feedback network parameters are saved;

[0011] S7. The trained encoder and decoder are used for CSI compression and reconstruction. When the decoder is reconstructing, the estimated value of the truncated matrix is ​​first padded with zeros to restore it to the original CSI matrix size, and then a two-dimensional inverse DFT transform is performed to obtain the reconstructed value of the downlink CSI channel matrix.

[0012] Furthermore, in the process of acquiring the sparse CSI matrix in the angle delay domain, that is, performing an angle delay domain transformation on the downlink CSI channel matrix, the transformation process includes:

[0013] H′=F c HF t H ;

[0014] Among them, F c is a DFT matrix of size 1024×1024, F t is a 32×32 DFT matrix; H is the CSI channel matrix of the downlink; H′ is a sparse CSI matrix in the angle delay domain.

[0015] Furthermore, the real and imaginary parts of the complex number in the truncated matrix are split, and then the real and imaginary parts are concatenated in the channel dimension. The concatenated truncated matrix containing two channels is used as the input of the encoder. The encoder on the user side includes a cascaded multi-resolution convolution module, a convolution layer with a convolution kernel of 1×1, and a fully connected layer containing N neurons.

[0016] Furthermore, the number of neurons in the fully connected layer N is proportional to the compression ratio, that is, N = 2048 × CR; where CR is the compression ratio.

[0017] Furthermore, the decoder at the base station includes a fully connected layer containing 2048 neurons, a convolutional layer, and a multi-resolution convolutional module. The input codeword vector is reorganized using a fully connected layer containing 2048 neurons, and the reorganized matrix is ​​input into two cascaded convolutional layers with a 3×3 convolution kernel for rough feature extraction. The roughly extracted features are input into a dense network composed of three densely connected multi-resolution convolutional modules. The input of each multi-resolution convolutional module is the concatenation of the output of the previous multi-resolution convolutional module and the rough features. The output of the dense network is normalized by a sigmoid activation function to complete the decoding.

[0018] Furthermore, the multi-resolution convolution module includes a small convolution kernel layer composed of a convolution module and a large convolution kernel layer composed of three convolution modules cascaded. The small convolution kernel layer and the large convolution kernel layer are used to extract features of the input encoder matrix respectively, and the extracted features are concatenated as the output of the multi-resolution convolution module.

[0019] Furthermore, the convolution module includes three convolution layers with convolution kernel sizes of 1×3, 3×1 and 3×3. After the training is completed, the convolution kernel parameters of the convolution kernel sizes of 1×3 and 3×1 are added to the convolution kernel parameters of the convolution kernel size of 3×3, and the addition position is the cross position at the center of the 3×3 convolution layer; in the real-time data processing stage, only the fused convolution kernel size of 3×3 is used for feature extraction as the output of the convolution module. The data input into this convolution module and the convolution kernel parameters are essentially matrices. The convolution is calculated as the matrix of the convolution kernel parameters starting from the upper left corner of the input data and multiplied and accumulated with the corresponding position of the input data. The result obtained is the output of the convolution operation, and the value in the convolution kernel matrix is ​​the convolution kernel parameter.

[0020] Furthermore, the number of convolution kernels in the convolution layer of the encoder's resolution convolution module is 4.

[0021] Furthermore, the number of convolution kernels in the small convolution kernel layer in the resolution convolution module of the decoder is 4, and the number of convolution kernels in the three cascaded convolution kernel layers in the large convolution kernel layer are 8, 16, and 4, respectively.

[0022] Furthermore, the loss function for the joint training of the feedback network composed of the encoder and the decoder adopts the mean square error function, the feedback network is initialized using the Xavier method, and Adam is used to optimize the feedback network according to the loss function. During the optimization process, the cosine annealing algorithm is used to dynamically adjust the learning rate of the feedback network.

[0023] In order to achieve higher feedback accuracy while reducing the number of method parameters and the amount of calculation, the present invention, on the one hand, replaces the large convolution kernel by using a cascade of small convolution kernels, and on the other hand, avoids the feedback network from being too deep, and adopts a horizontal expansion method, that is, multi-resolution convolution is used to increase the generalization of the feedback network, and asymmetric convolution is used to increase the feature extraction capability of each convolution kernel; in addition, the feedback network of the present invention adopts a local dense connection method, and performs feature learning by repeatedly utilizing multiple low-level features, thereby further improving the learning ability of the feedback network without increasing parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of a CSI feedback method based on deep learning large-scale MIMO system of the present invention;

[0025] Figure 2 AMRNet feedback network structure diagram of the present invention;

[0026] Figure 3 It is a schematic diagram of NMSE comparison under two different compression ratios;

[0027] Figure 4 Schematic diagram for comparing pseudo grayscale images with different compression ratios under two environments. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] The present invention proposes a CSI feedback method for a deep learning large-scale MIMO system, comprising the following steps:

[0030] S1. In a massive MIMO system under FDD standard, obtain the CSI channel matrix of the downlink when the base station antenna adopts a uniform linear array with a half-wavelength spacing and the user end is a single user;

[0031] S2. Perform a two-dimensional DFT transform on the acquired downlink CSI channel matrix to obtain a sparse CSI matrix in the angle delay domain;

[0032] S3, the first N of the CSI matrix obtained in the angle delay domain sparse a The truncated matrix is ​​obtained by truncating the rows; the CSI matrix in the angle delay domain is obtained after the CSI matrix is ​​transformed by DFT. At this time, due to the existence of multipath effect, the matrix has a sparse characteristic, that is, only the first N a The row has non-zero elements, and the elements of other rows are all 0 or close to 0. The valid information contained is very scarce, so it can be truncated directly to keep only the first N a Just go ahead;

[0033] S4, constructing an encoder of the CSI feedback network at the user end, and converting the truncation matrix into a codeword vector through the encoder;

[0034] S5. Construct a decoder of the CSI feedback network at the base station end, and pass the codeword vector obtained by the encoder through the decoder to obtain an estimated value of the truncation matrix;

[0035] S6, the encoder and decoder are jointly trained in an end-to-end manner so that the output of the feedback network With the original matrix H a The difference is as small as possible, and the trained feedback network parameters are saved;

[0036] S7. The trained encoder and decoder are used for CSI compression and reconstruction. When the decoder is reconstructing, the estimated value of the truncated matrix is ​​first padded with zeros to restore it to the original CSI matrix size, and then a two-dimensional inverse DFT transform is performed to obtain the reconstructed value of the downlink CSI channel matrix.

[0037] In this embodiment, the feedback network composed of the encoder and the decoder is named as an asymmetric multi-resolution feedback network (Asymmetric Multi-Resolution Network, AMRNet), referred to as a feedback network. The encoder is used at the user end to compress and encode the channel state information into a low-dimensional codeword, which is transmitted to the base station decoder via the feedback link and reconstructs the channel state information, thereby reducing the channel state information feedback overhead. The method of the present invention is specifically described below in conjunction with a specific application scenario.

[0038] In the downlink of the massive MIMO system, the COST 2100 model is used to set up two experimental environments, an indoor scene at 5.3 GHz and an outdoor scene at 300 MHz. In the indoor and outdoor scenes, the base station is located at the center of a square area with a length of 20 and 400 meters respectively, and the user end is randomly placed in the square area of ​​each sample. The base station uses Nt = 32 transmitting antennas, and the user end uses a single receiving antenna. The massive MIMO system uses OFDM carrier modulation, using N c =1024 subcarriers. According to the above conditions, 120,000 data sets are generated, of which 100,000 data sets are randomly selected for training. The cross-validation method is used to divide the training set and the validation set of the model in a ratio of 8:2. The remaining 20,000 data sets are used as test sets to evaluate the CSI reconstruction accuracy of the model.

[0039] Each spatial-frequency domain CSI matrix H in the data set is transformed into the angle delay domain. The transformation process is as follows:

[0040] H′=F c HF t H

[0041] Among them, F c and F t The DFT matrices are 1024×1024 and 32×32 in size respectively. is the angle delay domain matrix obtained after two-dimensional DFT transformation. Because the delay between multipath arrival times is within a limited time range, in the delay domain, the CSI matrix is ​​only a (N a <N c ) row has a value, and the values ​​of the other rows are 0 or close to 0. In this case, set N a =32, that is, the first 32 rows of H′ are truncated, and the truncated matrix is ​​represented by H a .

[0042] AMRNet encoder design reference Figure 2In the Encoder part, since the current deep learning framework does not support complex number calculations, it is necessary to split the real and imaginary parts of the complex numbers in the input CSI matrix, and then splice them in the channel dimension. After splicing, the dimension of the CSI matrix is ​​32×32×2. The encoder includes a cascaded multi-resolution convolution module, a convolution layer with a convolution kernel of 1×1, and a fully connected layer containing N neurons. The multi-resolution convolution module includes a small convolution kernel convolution layer and a large convolution kernel convolution layer. The small convolution kernel convolution layer uses a convolution kernel size of 3×33×3 as the convolution kernel. The large convolution kernel convolution layer uses a 7×7 convolution kernel cascaded with three convolution kernels of 3×3. In each convolution layer, there are 4 convolution kernels. After the convolution operation, each convolution layer is zero-filled, LeakyReLU activation function and batch normalization; the multi-resolution convolution module outputs a 32×32×8 feature map, which is fused through a convolution layer with a convolution kernel of 1×1. At this time, the output feature map size is 32×32×2, and the feature map is reorganized into a 2048×1 vector. A fully connected layer containing N neurons is input, and N is calculated according to the compression ratio CR, that is, N=2048×CR, and an N×1 vector v is output, which is the compressed CSI codeword that the user end needs to feed back to the base station.

[0043] AMRNet decoder design reference Figure 2 In the Decoder part of step 3, the codeword vector v is used as the input of the decoder. First, a fully connected layer containing 2048 (32×32×2) neurons is input, and a 2048×1 vector is output. This vector is reorganized into a 32×32×2 matrix, and then input into a convolution layer with a convolution kernel of 5×5 for rough feature extraction, as shown in Figure 2As shown, the convolution layer with a convolution kernel of 5×5 is implemented by cascading two convolution layers with a convolution kernel of 3×3. DenseNet is densely connected by three multi-resolution convolution modules. The multi-resolution module in the decoder uses a convolution layer with a convolution kernel of 3×3 as a small convolution kernel convolution layer. The number of convolution kernels in this layer is 4. Three convolution layers with a convolution kernel of 3×3 are cascaded to form a convolution layer with a convolution kernel of 7×7 as a large convolution kernel convolution layer. In this large convolution kernel convolution layer, the number of convolution kernels of the three 3×3 convolution layers are 8, 16, and 4 respectively. Each multi-resolution block is followed by a convolution layer with a convolution kernel of 1×1 to prevent the number of channels of the feature map from being too large. The output of the convolution layer with a convolution kernel of 1×1 is used as the output of the multi-resolution module in the decoder. After the convolution operation, the convolution layer is zero-filled, LeakyReLU activation function and batch normalization are performed in sequence, so that the size of the feature map obtained after each convolution remains 32×32. After dense connection, feature fusion is performed again through a convolution layer with a convolution kernel size of 1×1 and a 32×32×2 feature map is output. Finally, the sigmoid activation function is used for normalization. The final output of the decoder, a 32×32×2 real number matrix, is used as the real and imaginary parts of the final reconstructed channel matrix.

[0044] The feedback network model parameters are initialized using the Xavier method, the model training optimizer uses the Adam algorithm, and the model learning rate is dynamically adjusted using the cosine annealing algorithm. The loss function of AMRNet is the mean square error, expressed as:

[0045]

[0046] Among them, T represents the number of samples in one round of iterative training, represents the Euclidean norm; H i represents the CSI matrix of the input encoder, Represents the CSI matrix output by the decoder.

[0047] The encoder and decoder are jointly trained using the training data, including weights, biases, and convolution kernel parameters, to minimize the cost function. 200 samples in the training set are used to calculate the gradient in each iteration, and the parameters are updated according to the formula of the Adam algorithm. The entire training set is trained 1000 times in this way, and the model parameters are debugged according to the results of the validation set. The performance of the final model is verified by the results of the test set.

[0048] After the network training phase, the asymmetric module ( Figure 2In the ACBlock, the two one-dimensional convolution kernel parameters in the convolution module described in the basic invention are added to the symmetric convolution kernel parameters, and the addition position is at the central cross position of the symmetric convolution kernel. In the online derivation stage, there is no need to use asymmetric convolution kernels for calculations, but to use symmetric convolution kernels after feature fusion for calculations. The convolution kernel fusion process is as follows: Assume that the convolution kernel size of a convolution layer is H×W×C, where C represents the number of channels of the convolution kernel and H×W represents the convolution kernel size, so To represent the size of a convolution kernel. As the size of the input data of the convolution layer, J×K is the dimension of the input data. Let the number of convolution kernels of the convolution layer be D, so the output of the convolution layer can be expressed as For each convolution kernel, the calculation process can be expressed as follows:

[0049]

[0050] Among them, * represents 2D convolution operation, M :,:,k is the kth channel of the input data, a matrix of size J×K, is the convolution matrix of the kth channel, size H×W; O :,:,j is the jth channel of the output data.

[0051] Batch normalization is currently proven to be an effective means to reduce overfitting and accelerate network model training. Therefore, after adding batch normalization to the convolutional layer, its calculation process can be expressed as follows:

[0052]

[0053] Among them, μ j and σ j Represent the mean and standard deviation of the calculation results of the jth convolution kernel, γ j and β j They represent the scaling factor and bias term of the j-th convolution kernel calculation result respectively.

[0054] The theoretical basis of asymmetric convolution is based on the additivity of the convolution operation. The process can be expressed as:

[0055]

[0056] Among them, I represents the feature matrix, K represents the convolution kernel with compatible size, that is, when the number of channels is the same, the size of the asymmetric convolution kernel cannot exceed the size of the symmetric convolution kernel, and the superscript only indicates that the two convolution kernels are different convolution kernels; Indicates the addition of elements at corresponding positions in the matrix. Therefore, a convolution kernel can be split into H×1, 1×W and H×W to form three-way parallel convolution operations. At this time, the process of batch normalization of convolution kernel calculation is:

[0057]

[0058] Among them, F′ j represents the fused convolution kernel, F j , and They represent the symmetric convolution kernel before convolution fusion, and the asymmetric convolution kernels of H×1 and 1×W respectively. and Represents the scaling factor when the asymmetric convolution kernel is calculated for batch normalization, and Represents the bias term when calculating batch normalization for asymmetric convolution kernels.

[0059] Assume b j It represents the sum of the mean term and the bias term in the batch normalization of the three convolution kernels and is expressed as:

[0060]

[0061] in and They represent the scale factor and bias term of the asymmetric convolution kernel calculation results respectively, so the complete asymmetric convolution fusion calculation process can be deduced as follows:

[0062]

[0063] The trained AMRNet feedback network with asymmetric convolution kernel parameter fusion is used for CSI compression and reconstruction. The reconstructed CSI matrix First, a zero padding operation is performed to restore the original CSI matrix H to the same size, and then a two-dimensional inverse DFT transform is performed to obtain the reconstructed value of the CSI matrix in the original space-frequency domain.

[0064] The present invention is verified according to the above steps, and the CsiNet feedback network based on deep learning that has been proposed is trained with the same data set. The effectiveness of the present invention is demonstrated by comparing it with CsiNet.

[0065] from Figure 3It can be seen that the proposed AMRNet outperforms CsiNet in different CRs in both indoor and outdoor scenarios. At the same time, using only small convolution kernels does not reduce the learning ability of the feedback network. Under the combination of asymmetric convolution and dense connection, it can better complete the CSI feedback task than CsiNet in different environments. With the increase of compression ratio, the reconstruction accuracy gradually decreases, which shows that with the increase of compression ratio, the hidden information lost by the compressed CSI matrix is ​​also increasing. The outdoor environment contains more noise information than the indoor environment, which makes the reconstruction accuracy decrease more obviously. On the other hand, by comparing asymmetric convolution and symmetric convolution networks, it is found that in different environments and under different CRs, the performance of the network with asymmetric convolution structure is slightly better than that of the symmetric convolution network MRNet. The NMSE in indoor environments is increased by about 9% on average, and the NMSE in outdoor environments is increased by about 5% on average. This proves that asymmetric convolution enhances the feature extraction ability of the convolution kernel.

[0066] Figure 4 The pseudo grayscale images of the original CSI matrix under 4 different CRs and the reconstructed CSI matrix by CsiNet and AMRNet in indoor and outdoor environments are plotted. As can be seen from the figure, in indoor environments, CsiNet can reconstruct the main information of the CSI matrix under different CRs, and only loses the secondary information in the right half. In outdoor environments, starting from CR=1 / 16, CsiNet can no longer reconstruct the CSI matrix more completely, which proves that CsiNet can only achieve good results when CSI contains more useful information. In outdoor environments or at large compression ratios, as the useful information in the CSI matrix decreases, the reconstruction accuracy decreases. In both environments, the AMRNet proposed in this paper under 4 different CRs benefits from asymmetric convolution and multi-resolution convolution, so that the proposed network can reconstruct the CSI matrix more completely while reducing the number of network model parameters and the amount of calculation compared to CsiNet.

[0067] Table 1 Comparison of complexity of different models

[0068]

[0069] As can be seen from Table 1, the difference between the AMRNet proposed in this paper and CsiNet in terms of parameter quantity is not much, because the parameter quantity of both networks mainly comes from the fully connected layers at the encoding and decoding ends. However, the amount of computation is greatly reduced compared to CsiNet. When the CR is 1 / 4, the amount of computation is reduced by about 30%, and when the CR is 1 / 32, the amount of computation is reduced by about 46%. This difference also comes from the fully connected layer. Compared with CR 1 / 4, the number of neuron nodes at the output end of the fully connected layer is 512, and when the CR is 1 / 32, the number is only 64. Therefore, the larger the CR, the more obvious the reduction in computational complexity. Therefore, the AMRNet proposed in this paper reduces the number of model parameters and computational complexity by using a 3×3 convolution kernel cascade to realize a large convolution kernel. At the same time, the asymmetric convolution in the network, the multi-resolution convolution and the densely connected network structure at the decoding end enhance the convolution kernel feature reuse capability, improve the network's feature reconstruction performance for the CSI matrix, and can further reduce the amount of computation involved in the decoding process while ensuring the accuracy of CSI matrix reconstruction.

[0070] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A CSI feedback method for a large-scale MIMO system based on deep learning, characterized in that: The following steps are involved: S1. In a massive MIMO system under FDD standard, obtain the CSI channel matrix of the downlink when the base station antenna adopts a uniform linear array with a half-wavelength spacing and the user end is a single user; S2. Perform a two-dimensional DFT transform on the acquired downlink CSI channel matrix to obtain a sparse CSI matrix in the angle delay domain; S3, the first N of the CSI matrix obtained in the angle delay domain sparse a The non-zero elements of the row are truncated to obtain a truncated matrix; S4, constructing an encoder of the CSI feedback network at the user end, and converting the truncation matrix into a codeword vector through the encoder; S5. Construct a decoder of the CSI feedback network at the base station end, and pass the codeword vector obtained by the encoder through the decoder to obtain an estimated value of the truncation matrix; S6. The encoder and decoder are jointly trained in an end-to-end manner so that the output H of the feedback network a With the original matrix H a The difference is as small as possible, and the trained feedback network parameters are saved; wherein, the encoder at the user end includes a cascaded multi-resolution convolution module, a convolution layer with a convolution kernel of 1×1, and a fully connected layer containing N neurons, and the multi-resolution convolution module includes a small convolution kernel layer composed of a convolution module and a large convolution kernel layer composed of three convolution modules cascaded, and the small convolution kernel layer and the large convolution kernel layer are used to extract the features of the input encoder matrix respectively, and the extracted features are spliced ​​as the output of the multi-resolution convolution module; The convolution module includes three convolution layers with convolution kernel sizes of 1×3, 3×1, and 3×3. After training, the convolution kernel parameters of the convolution kernels of 1×3 and 3×1 are added to the convolution kernel parameters of the convolution kernel of 3×3, and the addition position is the cross position at the center of the 3×3 convolution layer. In the real-time data processing stage, only the fused convolution kernel of 3×3 is used for feature extraction as the output of the convolution module. S7. The trained encoder and decoder are used for CSI compression and reconstruction. When the decoder is reconstructing, the estimated value of the truncated matrix is ​​first padded with zeros to restore it to the original CSI matrix size, and then a two-dimensional inverse DFT transform is performed to obtain the reconstructed value of the downlink CSI channel matrix.

2. According to claim 1, a CSI feedback method based on deep learning massive MIMO system is characterized in that: The process of obtaining the sparse CSI matrix in the angle delay domain is to transform the downlink CSI channel matrix in the angle delay domain. The transformation process includes: H′=F c HF t H ; Among them, F c is a DFT matrix of size 1024×1024, F t is a 32×32 DFT matrix; H is the CSI channel matrix of the downlink; H′ is a sparse CSI matrix in the angle delay domain.

3. The CSI feedback method based on deep learning massive MIMO system according to claim 1, characterized in that: The real and imaginary parts of the complex number in the truncated matrix are split, and then the real and imaginary parts are concatenated in the dimension of the channel, and the concatenated truncated matrix containing two channels is used as the input of the encoder.

4. The CSI feedback method based on deep learning massive MIMO system according to claim 3, characterized in that: The number of neurons in the fully connected layer N is proportional to the compression ratio, that is, N = 2048 × CR; where CR is the compression ratio.

5. The CSI feedback method based on deep learning massive MIMO system according to claim 1, characterized in that: The decoder at the base station end includes a fully connected layer containing 2048 neurons, a convolutional layer, and a multi-resolution convolutional module. The input codeword vector is reorganized using a fully connected layer containing 2048 neurons, and the reorganized matrix is ​​input into two cascaded convolutional layers with a 3×3 convolution kernel for rough feature extraction. The roughly extracted features are input into a dense network composed of three densely connected multi-resolution convolutional modules. The input of each multi-resolution convolutional module is the concatenation of the output of the previous multi-resolution convolutional module and the rough features. The output of the dense network is normalized by a sigmoid activation function to complete the decoding.

6. The CSI feedback method based on deep learning massive MIMO system according to claim 1, characterized in that: The number of convolution kernels in the convolution layer of the encoder's resolution convolution module is 4.

7. The CSI feedback method based on deep learning massive MIMO system according to claim 1, characterized in that: The number of convolution kernels in the small convolution kernel layer in the resolution convolution module of the decoder is 4, and the number of convolution kernels in the three cascaded convolution kernel layers in the large convolution kernel layer are 8, 16, and 4 respectively.

8. The CSI feedback method based on deep learning massive MIMO system according to claim 1, characterized in that: The loss function of the feedback network composed of the encoder and the decoder for joint training adopts the mean square error function. The feedback network is initialized using the Xavier method, and Adam is used to optimize the feedback network according to the loss function. During the optimization process, the cosine annealing algorithm is used to dynamically adjust the learning rate of the feedback network.

Images