A CSI Compression and Feedback Method and System for Large-Scale MIMO Based on Neural Networks

Abstract

The present invention discloses a CSI compression and feedback method and system for large-scale MIMO based on neural networks. The present invention uses a new deep neural network structure to complete the feedback of CSI, and this network is called Dilated Attention Inception Net (DAINet). The CSI data to be compressed and feedback is sequentially input into the pre-denoising module and encoder of the trained DAINet after 2D DFT transformation. After compressing and outputting a one-dimensional codeword for feedback, it is received and decoded by the trained decoder at the feedback receiving end to obtain the CSI matrix in the angular delay domain, and finally the restored value of the CSI matrix in the spatial frequency domain is obtained through 2D IDFT transformation; in the case of ideal channel estimation and non-ideal channel estimation, considering the accuracy rate and training cost comprehensively, the proposed DAINet is superior to the existing methods. Specifically, in the case of ideal channel estimation, the NMSE (Normalized Mean Square Error) of DAINet is about 1 / 2 of AnciNet and 1 / 12 of CsiNet; in the non-ideal case, the NMSE of DAINet is about 9 / 10 of AnciNet and 1 / 2 of CsiNet. The number of network parameters of DAINet is about 80% of AnciNet and 130% of CsiNet.

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and in particular to a CSI compression feedback method and system for large-scale MIMO based on a neural network considering non-ideal channel estimation. Background Art

[0002] Large-scale multiple-input multiple-output (m-MIMO) has become an essential component of 5G wireless networks. However, to fully utilize this technology and effectively eliminate multi-user interference in the channel and improve the transmission efficiency of the communication system, accurate channel state information (CSI) must be obtained at the transmitter. In a frequency-division duplex (FDD) system, the downlink CSI is usually estimated by a channel at a user equipment (UE) and then fed back to a base station (BS).

[0003] However, due to the large number of antennas and the large channel matrix in the m-MIMO system, directly feeding back the CSI without compression makes CSI estimation and feedback very challenging, especially through a feedback channel with limited bandwidth. The urgent need to reduce the cost of CSI feedback has inspired various techniques, including traditional methods and neural network-based methods.

[0004] Traditional methods such as codebook-based and compressive sensing (CS)-based methods. The implementation cost of the codebook-based method is linearly related to the number of transmit antennas and is not applicable in the m-MIMO system. The practical operation of the compressive sensing (CS)-based method is more difficult.

[0005] In recent years, the successful application of deep learning (DL)-based methods in channel estimation and signal detection has inspired a series of research results, which has also attracted more and more attention to m-MIMO CSI feedback. Therefore, DL-based methods have been introduced into the CSI feedback task and have shown great potential in CSI recovery.

[0006] With the further research on CSI feedback, various neural network-based methods have been continuously proposed, including CsiNet, CsiNet+, CsiNet-LSTM, CRNet, DS-RefineNet, and CsiNet Pro. These networks have achieved better results than traditional methods and have continuously broken through in terms of performance. However, these neural network-based methods design the CSI feedback system under the assumption of ideal channel estimation. Therefore, researchers have also discussed the design of CSI feedback neural networks in the case of non-ideal channel estimation. Considering the channel estimation error, AnciNet was proposed. Through reasonable design, AnciNet has shown excellent results under different signal-to-noise ratios. However, the training cost of AnciNet is much higher than that of previous networks, and there is still a large room for improvement in its accuracy. Moreover, due to the overfitting problem, the accuracy is extremely unstable under end-to-end training. Summary of the Invention

[0007] In view of the deficiencies in the above-mentioned background technology, the present invention provides a method and system for CSI compression and feedback of large-scale MIMO based on neural networks. The neural network model DAINet adopted by the present invention applies the attention mechanism, dilated convolution, and residual learning to solve the problems of high training cost, serious overfitting problem, and low accuracy of existing neural networks.

[0008] The object of the present invention is achieved through the following technical solutions:

[0009] A method for CSI compression and feedback of large-scale MIMO based on neural networks is specifically as follows:

[0010] A method for CSI compression and feedback based on neural networks considering non-ideal channel estimation is specifically as follows:

[0011] The CSI data to be compressed and feedback is sequentially input into the trained pre-denoising module and encoder after 2D DFT transformation. After compressing and outputting a one-dimensional codeword for feedback, the trained decoder at the feedback receiving end receives and decodes it to obtain the CSI matrix in the angular-delay domain, and finally, the restored value of the CSI matrix in the spatial-frequency domain is obtained through 2D IDFT transformation;

[0012] Among them, the pre-denoising module consists of multiple dilated convolution groups, standard convolution groups, multiple DAI Blocks, activation function layers, convolution layers, a network splicing layer that merges and splices the input of the first DAI Block and the output of the convolution layer, activation function Tanh, attention blocks, and reconstruction blocks connected in sequence, and is responsible for pre-denoising the CSI matrix obtained by channel estimation to obtain the denoised CSI matrix The DAI Block is composed of an atrous convolution group, multiple standard convolution groups, a network splicing layer, and a convolution layer in a way of first parallel and then splicing. The atrous convolution group is used to capture more detailed features of the input data of the DAI Block; the standard convolution groups with different convolution kernel sizes extract information from the detailed features captured by the atrous convolution group from different aspects, then merge in the output channel dimension, and then use the convolution layer to adjust the output size, and finally fuse with the input data of the DAI Block as the output of the DAI Block; the reconstruction block is used to reconstruct the pre-denoised channel state information according to the noise information output by the attention block.

[0013] An encoder, used to achieve compression at a specific compression ratio, compresses the denoised CSI matrix into the codeword s by compression quantization;

[0014] A decoder, used to select the corresponding magnification for decompression, and restore and convert the codeword s into the CSI matrix in the angular delay domain.

[0015] The pre-denoising module, encoder, and decoder are connected in sequence to form the DAINet and then trained based on the training dataset.

[0016] Further, the encoder is composed of one or more atrous convolution groups, standard convolution groups, DAI Blocks, a data reorganization layer, a fully connected layer, and a quantization layer.

[0017] Further, the decoder is composed of an inverse quantization layer, a fully connected layer, a data reorganization layer, a standard convolution group, multiple DAI Blocks, a convolution layer, and an activation function.

[0018] Further, each training sample of the training dataset is composed of the CSI matrix under non-ideal channel estimation; the CSI matrix under non-ideal channel estimation of each training sample pair is used as the input of the DAINet, and the trained DAINet is obtained by training with the goal of minimizing the cost function of the output of the DAINet and the true value.

[0019] Further, the cost function is:

[0020]

[0021] where T represents the number of training samples; H d [i] represents the true value of the CSI matrix in the angular delay domain of the i-th sample, ||*||2 is the L2 norm, represents the output of the decoder of the i-th sample.

[0022] A neural network-based CSI compression and feedback system considering non-ideal channel estimation, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the neural network-based CSI compression and feedback method considering non-ideal channel estimation as described above.

[0023] A storage medium containing computer-executable instructions that, when executed by a computer processor, implement the neural network-based CSI compression and feedback method considering non-ideal channel estimation as described above.

[0024] The beneficial effect of the present invention is that, considering the non-ideal situation of channel estimation, the present invention uses numerous neural network solutions to design a new neural network. Compared with the prior art, the present invention has the highest accuracy rate under both ideal channel estimation and non-ideal channel estimation. Under the condition of ideal channel estimation, the NMSE (Normalized Mean Square Error) of DAINet is about 1 / 2 of AnciNet and 1 / 12 of CsiNet; under non-ideal conditions, the NMSE of DAINet is about 9 / 10 of AnciNet and 1 / 2 of CsiNet. The number of network parameters of the present invention is less than that of the prior research on non-ideal channel estimation, reducing the training cost. The number of network parameters of DAINet is about 80% of AnciNet. The present invention solves the problem of extremely unstable accuracy rate in the case of end-to-end training in the prior art. Description of the Drawings

[0025] Figure 1 is the idea diagram of the present invention;

[0026] Figure 2 is the structure diagram of the DAINet network pre-denoising module adopted by the method of the present invention;

[0027] Figure 3 is the structure diagram of the encoder and decoder of the DAINet network adopted by the method of the present invention;

[0028] Figure 4 is the NMSE performance comparison result of the present invention with CsiNet, CsiNet+, and AnciNet under ideal channel estimation and different compression ratios;

[0029] Figure 5 is the NMSE performance comparison result of the present invention with CsiNet, CsiNet+, and AnciNet under non-ideal channel estimation and different compression ratios;

[0030] Figure 6 is the NMSE performance comparison result of the present invention with CsiNet, CsiNet+, and AnciNet under non-ideal channel estimation and different signal-to-noise ratios;

[0031] Figure 7 is the comparison result of the network parameter quantities between the present invention and CsiNet, CsiNet+, and AnciNet;

[0032] Figure 8 is the comparison result of the network training processes between the present invention and CsiNet, CsiNet+, and AnciNet under ideal channel estimation and non-ideal channel estimation; Detailed implementation manners

[0033] To make the technical solutions and advantages of the embodiments of the present invention clearer, the following will further describe the technical solutions in more detail with reference to the accompanying drawings:

[0034] Figure 1 shows the basic idea of the present invention. The present invention selects to use the method based on the neural network DAINet to solve the three parts of CSI compression, feedback link, and CSI recovery in the MIMO CSI feedback system. Specifically, at the user side, the CSI data to be feedback-compressed is sequentially input into the trained pre-denoising module and encoder after 2D DFT transformation, and after compressing and outputting a one-dimensional codeword for feedback, at the feedback receiving end, the trained decoder receives and decodes it to obtain the CSI matrix in the angular-delay domain, and finally, the CSI matrix recovery value in the spatial-frequency domain is obtained through 2D IDFT transformation; among them, the pre-denoising module, encoder, and decoder are connected in sequence to form DAINet and are trained based on the training data set. In the present invention, DAINet uses many neural network solutions, namely dilated convolution, attention mechanism, residual learning, and Inception structure. Specifically, Figure 2 And Figure 3 shows a network structure diagram of DAINet for CSI feedback. Among them, the pre-denoising module is responsible for pre-denoising the CSI matrix obtained by non-ideal channel estimation to obtain the denoised CSI matrix The encoder is used to achieve compression at a specific compression ratio and compress and quantize it into the codeword s; the decoder is used to select the corresponding magnification for decompression and recover the channel matrix from the codeword s.

[0035] The pre-denoising module is composed of multiple sequentially connected dilated convolution groups, standard convolution groups, multiple DAI Blocks, attention blocks, reconstruction blocks, etc. Figure 2The following is a structural diagram of a pre-denoising module, which is specifically composed of a composite dilated convolution group (Composite Dilated Conv), a composite convolution group (Composite Conv), four DAI Blocks, an activation function layer, a convolutional layer, a network concatenation layer (Cat), an activation function Tanh, an attention block, and a reconstruction block (Reconstruction Block). The combination order is the composite dilated convolution group, the composite convolution group, DAI Block Ⅰ, DAI Block Ⅱ, DAI Block Ⅲ, DAI Block Ⅳ, the convolutional layer, the activation function Tanh, the network concatenation layer (Cat) that merges and concatenates the input of the first DAI Block and the output of the convolutional layer, the activation function Tanh, the attention block, and the reconstruction block. Among them, the four DAI Blocks are connected in a residual learning manner. The DAI Block is an Inception-tiny structure, which is specifically composed of a composite dilated convolution group and multiple composite convolution groups, as well as a network concatenation (Cat) and a convolutional layer in a parallel-then-concatenated manner. This composition method is inspired by the Inception structure. The first layer of the DAI Block is a dilated convolutional neural network layer with a larger receptive field, which can capture more feature details. However, the dilated convolutional neural network may cause partial loss of local information, may also affect the local consistency of feature maps, and the larger receptive field can also smooth image details. Therefore, the composite convolution group convolutional layer with a smaller convolutional kernel is placed behind. This structure contains two groups of parallel composite convolution groups, which extract information from different levels through 2 paths, then merge in the output channel dimension, and finally use a convolutional layer to adjust the output size to match the subsequent neural network. This structure can not only extract feature maps at different scales, which is beneficial for subsequent processing, but also enhances the robustness of the neural network where it is located [S. Lazebnik, C. Schmid, and J. Ponce. Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories [C]. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2006: 2169–2178]. At the same time, the DAI Net selects a dilated convolutional neural network. The advantage of dilated convolution is that it can capture more image details by increasing the receptive field and can effectively solve the overfitting problem.

[0036] The Composite Dilated Conv consists of a dilated convolution and a batch normalization (BN) layer, and uses Leaky ReLU as the activation function. The expression of this activation function is

[0037]

[0038] where x represents the input of the activation function.

[0039] The Composite Conv consists of a standard convolution and a batch normalization (BN) layer, and also uses Leaky ReLU as the activation function;

[0040] The attention block consists of convolutional layers and completes its work mainly in two steps. The first step is to use a 1×1 convolutional layer neural network to compress the input multi-dimensional feature matrix to a certain size to adapt to the subsequent network. The second step is to multiply the output of the convolutional layer with a 1×1 convolutional kernel by the output of the convolutional layer with a 3×3 convolutional kernel in the previous DAI Block Ⅳ to extract more prominent noise features. The part within a dashed box after the attention block is called the reconstruction block, whose role is to restore the CSI after denoising, that is, the noise-free CSI. The reconstruction block is used to reconstruct the pre-denoised channel state information. Let the noise features output by the attention block, which is the input of the reconstruction block, be H R , then it can be considered that the reconstruction process can be expressed as where H LC is the latent clean channel state information matrix. [Tian C, Xu Y, Li Z, et al. Attention-guided CNN for image denoising[J]. Neural Networks, 2020, 124: 117-129.]

[0041] The pre-denoising module with the above structure first passes the data through two dilated convolutional groups with convolutional kernel sizes of 3×3 and 1×1 respectively, and a standard convolutional group, generating 64-layer and 16-layer feature maps respectively. The dilated convolutional neural network has a larger receptive field. Placing it in the front position is for two reasons: one is to obtain more comprehensive and profound feature maps, and the other is to adjust the data size to match the subsequent network layers. Then, a residual neural network containing 4 DAI Blocks (Dilated Attention Inception Block) is used. For a neural network with a relatively deep depth, the phenomenon of gradient disappearance will occur, which will cause the performance of the network to decline. Using residual learning can solve this problem. Finally, the pre-denoised CSI data will pass through a convolutional layer with the activation function Tanh, and be concatenated (Cat) with the input network of DAI Block Ⅰ, and then enter the attention block AB (Attention Block) after passing through the activation function. AB is used to guide the previous-stage CNN to learn the relevant information of the noise, and this operation is very effective for dealing with unknown noise. Finally, the CSI data stream output by AB enters the reconstruction block RB (Reconstruction Block). After the reconstruction is completed in RB, it is input to the encoder.

[0042] Figure 3 Figure 4 shows a structural diagram of an encoder and a decoder. Figure 3 The DAIBlock is also applied to the shown encoder and decoder. Specifically, the encoder consists of a composite dilated convolutional group (Composite Dilated Conv), two composite convolutional groups (Composite Conv), a DAI Block, and a quantization layer (Quantization); their combination order is the composite dilated convolutional group, composite convolutional group Ⅰ, DAI Block, composite convolutional group Ⅱ, reshape layer (Reshape), fully connected layer (FC), and quantization layer (Quantization). The quantization layer uses μ-law non-uniform quantization, and this method is optimized using the companding function f(.).

[0043]

[0044] where x ∈ [-1, 1] represents the input of the quantization layer and is a weak signal, and μ is a constant; the functions of the composite dilated convolutional group and composite convolutional group Ⅰ in the encoder are similar to those of the pre-denoising module. Then, a DAI Block is used to generate pure feature maps, and then a combined convolutional network layer (composite convolutional group Ⅱ, reshape layer) is used to compress the feature maps. At the end of the encoder, a fully connected layer and a quantization layer are used to further compress the information and transform it into a one-dimensional codeword. By adjusting the output size of the fully connected layer, the compression ratio can be adjusted. Figure 3Below is the decoder shown. The decoder consists of a dequantization layer, a composite convolution group, and three DAI Blocks, etc. The combination order is the dequantization layer, fully connected layer, reshape layer, composite convolution group I, DAI Block I, DAI Block II, DAI Block III, convolutional layer, activation function Tanh. The three DAI Blocks are connected in a residual learning manner. After receiving the codeword, the decoder first reconstructs the codeword into a higher dimension by a dequantization layer, fully connected layer, and reshape layer, and at the same time completes the preliminary estimation of the CSI matrix. Then, the residual network composed of three DAI Blocks, convolutional layer, and activation function Tanh will complete the further estimation of the CSI matrix, and finally output the reconstructed CSI matrix.

[0045] In summary, the entire neural network f DL The working principle can be expressed by the following formula

[0046]

[0047] f DE (), f EN (), f DN () represent the decoder, encoder, and pre-denoising module respectively, represents the angle-delay domain CSI matrix obtained under the non-ideal channel, and H de represents the angle-delay domain CSI matrix recovered by DAINet.

[0048] The training of the DAINet based on the training dataset is specifically as follows. The present invention can be used for the compression and feedback of CSI data under non-ideal channel estimation. Therefore, each training sample of the training dataset consists of the angle-delay domain CSI matrix under non-ideal channel estimation. The angle-delay domain CSI matrix under non-ideal channel estimation can be obtained by collecting the CSI matrix under non-ideal channel estimation at the user end and then performing two-dimensional Fourier transform to convert the data obtained from non-ideal channel estimation from the spatial frequency domain to the angle-delay domain, obtaining the angle-delay domain data F c , F t represents the DFT matrix; represents taking the conjugate transpose of the F t matrix, Denotes the spatial frequency domain CSI matrix obtained under non-ideal channel estimation. It is also possible to add noise to the ideal channel dataset by following the typical linear model of CSI error [A.S. Housfater, T.J. Lim. Noisy feedback linear precoding: A Bayesian Cram′er-Rao bound [C]. IEEE International Symposium on Information Theory, 2009: 1689 - 1693], and then through Fourier transform, taking the final result as the CSI data under non-ideal channel estimation in the angular delay domain; the spatial frequency domain CSI matrix under non-ideal channel estimation can be expressed as where E represents additive noise, that is, the CSI matrix after non-ideal channel estimation can be represented by adding noise to the matrix after ideal channel estimation.

[0049] Taking the CSI matrix under non-ideal channel estimation of each training sample pair as the input of DAINet, and training to obtain the trained DAINet with the goal of minimizing the cost function between the output of DAINet and the true value. The cost function can adopt conventional loss functions, such as MSE (mean square error), MAE (mean absolute error, etc. As an alternative, the cost function MSE is expressed as follows:

[0050]

[0051] where T represents the number of training samples; H d [i] represents the true value of the i-th angular delay domain CSI matrix, that is, the input of the pre-denoising module, represents the output of DAINet, that is, the decoder, and ||*||2 is the L2 norm.

[0052] The DAINet of the present invention extracts noise-free features from noisy CSI samples obtained from non-ideal channel estimation to achieve effective compression of CSI feedback.

[0053] Figure 4 Represents the performance of different neural networks under different compression ratios, that is in the case of, on the same dataset obtained under ideal channel estimation. As 1 / increases, the NMSE of each neural network increases, and after 1 / = 32, the curve reaches convergence; under ideal channel estimation, the NMSE of DAINet is about 1 / 2 of AnciNet and about 1 / 12 of CsiNet and CsiNet+. Therefore, DAINet has better performance than the other three networks under ideal channel estimation.

[0054] Figure 5 Shows the performance of different neural networks under different compression ratios, that is The performance of different neural networks on the same dataset, which is obtained under non-ideal channel estimation with SNR = 20. Similar to Figure 4 , as 1 / increases, the NMSE of each network increases, and after 1 / = 16, the curve reaches convergence; in the case of non-ideal channel estimation, the NMSE of DAINet is about 9 / 10 of AnciNet and about 1 / 2 of CsiNet and CsiNet+. Therefore, DAINet also has better performance than the other three networks in the case of non-ideal channel estimation.

[0055] Figure 6 Denotes the compression ratio When the channel estimation is non-ideal, the performance comparison of different neural networks under different signal-to-noise ratios, i.e., SNR = 5, 10, 15, 20, 25, 30. It can be seen from the figure that DAINet still performs the best in the case of non-ideal channel estimation, that is, its NMSE is still the smallest.

[0056] Figure 7 Shows the number of training parameters of different neural networks at different compression ratios. The number of training parameters of DAINet is more than that of CsiNet and CsiNet+. The number of network parameters of DAINet is about 130% of CsiNet and CsiNet+, but DAINet has achieved a great improvement in accuracy; compared with AnciNet, the number of network parameters of DAINet is about 80% of AnciNet, and the accuracy of DAINet is higher than that of AnciNet. Therefore, DAINet not only improves the accuracy but also greatly reduces the number of training parameters. In summary, DAINet is the best choice considering both the number of training parameters and accuracy.

[0057] Figure 8 Represents the comparison chart of the training results of different networks under ideal and non-ideal channel estimations. It can be found that DAINet has the best convergence and the lowest loss rate. Although the MSE loss curves of CsiNet and CsiNet+ show a downward trend as the epoch increases, the MSE is larger than that of DAINet after the curves stabilize; the AnciNet curve is affected by the overfitting problem every certain number of epochs, and its MSE loss curve is very unstable, and its MSE is even larger than that of DAINet even when the curve is stable; during the continuous increase of the epoch, DAINet does not have the problem of overfitting, and the curve changes stably.

[0058] Corresponding to the embodiment of the neural network-based CSI compression and feedback method considering non-ideal channel estimation described above, the present invention also provides an embodiment of a neural network-based CSI compression and feedback system considering non-ideal channel estimation.

[0059] A neural network-based CSI compression and feedback system considering non-ideal channel estimation provided by an embodiment of the present invention includes one or more processors for implementing the neural network-based CSI compression and feedback method considering non-ideal channel estimation in the above embodiment.

[0060] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can be referred to the partial description of the method embodiment. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present invention. A person of ordinary skill in the art can understand and implement it without creative work.

[0061] An embodiment of the present invention also provides a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements the neural network-based CSI compression and feedback method considering non-ideal channel estimation in the above embodiment.

[0062] The computer-readable storage medium may be an internal storage unit of any device with data processing capabilities described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium may also be any device with data processing capabilities, such as a plug-in hard disk, a Smart Media Card (SMC), an SD card, a Flash Card, etc. equipped on the device. Further, the computer-readable storage medium may also include both an internal storage unit of any device with data processing capabilities and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and can also be used to temporarily store data that has been output or will be output.

[0063] Obviously, the above embodiments are only examples for clear illustration and not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation manners here. The obvious changes or modifications derived therefrom are still within the protection scope of the present invention.

Claims

1. A CSI compression and feedback method based on neural network considering non-ideal channel estimation, characterized in that, Specifically: The CSI data to be feedback-compressed is successively input into the trained pre-denoising module and encoder after 2D DFT transformation. After compressing and outputting a one-dimensional codeword for feedback, it is received and decoded by the trained decoder at the feedback receiving end to obtain the CSI matrix in the angle-delay domain, and finally the restored value of the CSI matrix in the spatial-frequency domain is obtained through 2D IDFT transformation; Among them, the pre-denoising module is responsible for pre-denoising the CSI matrix obtained by channel estimation by a plurality of dilated convolution groups, standard convolution groups, a plurality of DAI Blocks, activation function layers, convolution layers, network splicing layers that merge and splice the input of the first DAI Block and the output of the convolution layer, activation function Tanh, attention blocks, and reconstruction blocks connected in sequence to obtain a denoised CSI matrix wherein the DAI Block is composed of a dilated convolution group, a plurality of standard convolution groups, a network splicing layer, and a convolution layer in a parallel-then-spliced manner. The dilated convolution group is used to capture the feature details of the input data of more DAI Blocks; the standard convolution groups with different convolution kernel sizes extract information from the feature details captured by the dilated convolution group from different aspects, then splice and merge them in the output channel dimension, then use the convolution layer to adjust the output size, and finally fuse with the input data of the DAI Block as the output of the DAI Block; the reconstruction block is used to reconstruct the pre-denoised channel state information according to the noise information output by the attention block An encoder, which is used to perform compression at a specific compression ratio and compress and quantize the denoised CSI matrix into a codeword s; The decoder is used to select the corresponding magnification for decompression and restore and convert the codeword s into the CSI matrix in the angle-delay domain; The pre-denoising module, encoder, and decoder are connected in sequence to form the DAINet and are obtained by training based on the training dataset.

2. The method according to claim 1, wherein The encoder consists of one or more dilated convolution groups, standard convolution groups, DAI Blocks, data reorganization layers, fully connected layers, and quantization layers.

3. The method according to claim 1, wherein The decoder consists of an inverse quantization layer, a fully connected layer, a data reorganization layer, a standard convolution group, multiple DAI Blocks, a convolution layer, and an activation function.

4. The method according to claim 1, characterized in that, Each training sample of the training dataset consists of the CSI matrix under non-ideal channel estimation; the CSI matrix under non-ideal channel estimation of each training sample pair is used as the input of the DAINet, and the trained DAINet is obtained by training with the goal of minimizing the cost function of the output of the DAINet and the true value.

5. The method according to claim 4, wherein The cost function is: where T represents the number of training samples; H d [i] represents the true value of the angular delay domain CSI matrix of the i-th sample, and ||*||2 is the L2 norm, represents the output of the decoder for the i-th sample.

6. A neural network-based CSI compression and feedback system considering non-ideal channel estimation, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the neural network-based CSI compression and feedback method considering non-ideal channel estimation according to any one of claims 1-5.

7. A storage medium containing computer-executable instructions, where the computer-executable instructions implement the neural network-based CSI compression and feedback method considering non-ideal channel estimation according to any one of claims 1-5 when executed by a computer processor.

Images