阅读说明：本技术 基于深度学习的大规模mimo信道联合估计和反馈方法 (Large-scale MIMO channel joint estimation and feedback method based on deep learning ) 是由 金石 陈彤 郭佳佳 陈慕涵 于 2020-12-25 设计创作，主要内容包括：本发明公开了一种基于深度学习的大规模MIMO信道联合估计和反馈方法。首先,在用户端进行初始信道估计。然后,构建信道估计子网CEnet,通过训练使估计误差最小。其次,构建信道反馈子网CFnet。在用户端,将最优化后的信道估计值输入,输出压缩后的码字；在基站端,将码字输入,输出重建的信道矩阵。两个子网共同构成信道估计和反馈联合网络CEFnet。以往的CSI反馈网络均假设已得到完美的信道状态信息,未考虑实际中信道是估计得到的,存在误差和噪声。本发明通过构建信道估计和反馈联合网络CEFnet,实现了完整的下行信道估计和反馈过程,并通过使用全新的网络架构,达到了消除误差和噪声的目的,在减少反馈开销的同时提高了重建精度。(The invention discloses a large-scale MIMO channel joint estimation and feedback method based on deep learning. First, initial channel estimation is performed at the ue. Then, a channel estimation subnet CEnet is constructed, which minimizes estimation errors through training. Secondly, a channel feedback sub-network CFnet is constructed. Inputting the optimized channel estimation value at a user side, and outputting a compressed code word; and at the base station end, inputting the code words and outputting the reconstructed channel matrix. The two subnetworks together form a channel estimation and feedback joint network CEFnet. The prior CSI feedback networks all assume that perfect channel state information is obtained, and the fact that the channel is obtained by estimation in practice is not considered, so that errors and noises exist. The invention realizes the complete downlink channel estimation and feedback process by constructing the channel estimation and feedback joint network CEFnet, achieves the aim of eliminating errors and noises by using a brand-new network architecture, and improves the reconstruction precision while reducing the feedback overhead.)

1. A large-scale MIMO channel joint estimation and feedback method based on deep learning is characterized by comprising the following steps:

the method comprises the following steps: transmitting a data signal containing a pilot frequency symbol at a base station end of a large-scale MIMO system, extracting the pilot frequency symbol at a pilot frequency subcarrier position from a signal received at a user end after channel transmission, calculating a channel frequency domain response of the pilot frequency subcarrier position by a least square method, then obtaining the channel frequency domain responses of other subcarrier positions by interpolation, and jointly forming an initial channel estimation value containing noise;

step two: obtaining the initial noisy channel estimation value, namely the channel matrix of the channel state information in the space-frequency domainThen, it is converted into a transformation domain which can make it sparse, and a new channel matrix under the transformation domain is obtained

Step three: constructing a super-resolution channel estimation subnet model CEnet; the model consists of a three-layer convolutional neural network with super-resolution function, and is trained, so that estimation errors are gradually reduced, and an estimated channel matrix is gradually close to an ideal channel matrix H to obtain model parameters of CEnet;

step four: constructing a channel feedback sub-network model CFnet, wherein the network comprises an encoder and a decoder, the encoder belongs to a user side, and the decoder belongs to a base station side; at a user terminal, an encoder compresses an input channel matrix H into a low-dimensional code word; at the base station, the decoder reconstructs the channel matrix values from the compressed codewords

Step five: and placing the channel feedback sub-network CFnet behind the channel estimation sub-network CEnet, forming a channel estimation and feedback joint network model CEFnet together with the channel estimation sub-network CFnet, and performing joint training. When in combined training, model parameters obtained by CEnet training in the third step are imported and fixed, the model parameters of the CFnet are updated, so that the training error is gradually reduced, and meanwhile, a channel matrix is reconstructedGradually approaching the channel matrix H to obtain CFnet model parameters, and forming the model parameters of the CEFnet together with the model parameters of the CEnet obtained in the third step;

step six: reconstruction matrix output by joint network model CEFnet of channel estimation and feedbackTurning to the initial transform domain to recover the channel matrix of the space-frequency domainA reconstructed value of (a);

step seven: and the trained channel estimation and feedback combined network model CEFnet is used for estimating and feeding back channel state information under various scenes, downlink channel estimation values are obtained through the model, code words are obtained after network compression, and the code words are fed back to the base station end for reconstruction, so that the reconstruction values of the channel matrix of the original space-frequency domain are recovered.

2. The deep learning-based joint estimation and feedback method for large-scale MIMO channels according to claim 1, wherein the transmission pilot at the base station in the first step is an orthogonal comb-shaped pilot placed at equal intervals in the frequency domain, and after BPSK modulation, the transmission pilot and the data symbols at other subcarrier positions form a transmission signal together. And correspondingly, a channel estimation method based on pilot symbols is adopted at the user terminal to obtain an initial channel estimation value, an LS method with the lowest operation complexity is selected to obtain channel frequency domain response of the pilot subcarrier position, and then channel frequency domain response of other subcarrier positions is obtained through simple operations such as interpolation or DFT conversion and the like, so that the channel estimation values of all the subcarrier positions are obtained.

3. The deep learning-based large-scale MIMO channel joint estimation and feedback method according to claim 1, wherein the super-resolution channel estimation subnet model CEnet in step three is a lightweight convolutional neural network, and is composed of three convolutional layers, the lightweight structure of the three layers fully considers the computation capability of the user side, does not occupy too much storage space of the user side, initializes the parameters of the layers randomly, inputs the channel matrix estimated by LS method and interpolation into the subnet, and outputs the channel matrix with lower estimation error

4. The deep learning-based joint estimation and feedback method for large-scale MIMO channels according to claim 1, wherein the CFnet in the feedback sub-network model in step four is composed of two modules, i.e. an encoder and a decoder, wherein the encoder at the user end only includes a full connection layer, randomly initializes its parameters, and estimates the complex channel matrix with sparse transform domainAfter the real part and the imaginary part are separated, the components are flattened and spliced into a one-dimensional high-dimensional vector which is used as the input of the encoder, and the output is the ratioA one-dimensional low-dimensional vector with low dimensionality, namely a code word s after compression coding, is sent to a base station end through an uplink; the decoder at the base station end consists of a full connection layer, a Reshape layer and a deep de-noising residual neural network module Refinement Block, randomly initializes parameters of each layer, and inputs the received compressed code wordOutputting a reconstructed channel matrix with the same dimension as the channel matrix H

5. The deep learning based joint estimation and feedback method for massive MIMO channels according to claim 1, wherein: the model parameters in the third step and the fifth step mainly comprise the weight and the bias of the full connection layer and the convolution kernel and the bias of the convolution layer.

6. The deep learning based joint estimation and feedback method for massive MIMO channels according to claim 1, wherein: and step three, training parameters of a channel estimation subnet CEnet by adopting an Adam optimization algorithm to minimize a cost function, wherein the cost function is described as follows:

wherein M is the number of all samples in the training set, | | · | | non |)₂Is the Euclidean norm, H_iIs an ideal channel matrix;and the channel matrix after super-resolution estimation is obtained.

7. The deep learning based joint estimation and feedback method for massive MIMO channels according to claim 1, wherein: and fifthly, jointly training parameters of a combined channel estimation and feedback network CEFnet by adopting an Adam optimization algorithm and an end-to-end learning mode, wherein the parameters of the CEnet are obtained by training in the third step, are imported and fixed, and only the parameters of a channel feedback subnet CFnet are trained and updated, so that a cost function is minimum, and the cost function is described as follows:

wherein M is the number of all samples in the training set, | | · | | non |)₂Is the Euclidean norm, H_iIs an ideal channel matrix;to reconstruct the channel matrix.

8. The deep learning based joint estimation and feedback method for massive MIMO channels according to claim 3, wherein: the three convolutional layers of the super-resolution channel estimation subnet model CEnet are respectively as follows: the first layer is to extract the characteristics of the noisy channel, the second layer is to perform nonlinear mapping, the characteristics of the noisy channel are mapped to the characteristics of an ideal channel, and the third layer is to perform weighted combination on the mapped characteristics to recover the original ideal channel, so as to realize super-resolution estimation of the noisy channel matrix.

9. The deep learning based joint estimation and feedback method for massive MIMO channels according to claim 4, wherein: in the decoder, a deep denoising residual neural network module Refinement Block of L layers comprises an input convolutional layer, a residual convolutional network of (L-2) layers and an output convolutional layer, wherein output data of the input layer is added with output data of the last convolutional layer of the residual network, and the output data is output after passing through the last convolutional layer.

Technical Field

The invention relates to a large-scale MIMO channel joint estimation and feedback method based on deep learning, and belongs to the technical field of communication.

Background

A Massive Multiple-Input Multiple-Output (Massive MIMO) system is widely considered as a main technology of a 5G wireless communication system. Such a system can greatly reduce multi-user interference by configuring hundreds or even thousands of antennas to the base station to form an antenna array, thereby serving multiple users simultaneously on the same time-frequency resource block and providing a multiplied increase in cell throughput. However, the potential benefits described above are primarily obtained by utilizing CSI in the base station. Time-Division Duplexing (TDD) technology can obtain CSI from the uplink, but requires a complex calibration process, while Frequency-Division Duplexing (FDD) technology completely requires CSI to be obtained through feedback. In the current FDD Massive MIMO system, during training, a User Equipment (UE) is used as a receiving end to acquire CSI of a downlink, and the CSI is returned to a Base Station (BS) through a feedback link. Because the number of antennas in the massive MIMO system is greatly increased, and feedback of complete CSI brings huge overhead, vector quantization or a codebook-based limited feedback algorithm is usually adopted, but problems of increased quantization error, complex codebook design, linear increase of feedback overhead, and the like occur, so that the method is not suitable for channel feedback of the massive MIMO system.

The large-scale MIMO channel state information feedback and reconstruction model CsiNet based on deep learning and other models based on the large-scale MIMO channel state information feedback and reconstruction model CsiNet which are proposed at present reconstruct CSI by utilizing the theory of space-time correlation and compressed sensing of the channel state information, and reduce feedback overhead. However, in the prior art, all models are performed on the premise that the receiving end is assumed to have obtained an ideal channel matrix, and in an actual communication system, a complete CSI feedback process should include two parts, namely, channel estimation and channel feedback, that is, a downlink channel is estimated first, and then an undesired channel matrix obtained through estimation is fed back to the base station after being processed by compression and the like. Because the feedback is the estimated channel matrix containing noise and estimation error, higher requirements are put on the decompression reconstruction network at the base station end. Therefore, it is considered to establish a complete joint channel estimation and feedback model with better reconstruction capability to achieve true CSI feedback.

Disclosure of Invention

The technical problem is as follows: the technical problem to be solved by the invention is to overcome the defects of the existing CsiNet and other models based on the CsiNet for structure optimization, provide a large-scale MIMO downlink channel joint estimation and feedback method by transmitting pilot frequency by a base station and receiving the pilot frequency by a user terminal for channel estimation and feedback, make up the defect that the channel estimation is not considered in the traditional models, and improve the reconstruction performance of the network by adopting a brand-new decoder network structure at the base station end because the feedback is the estimated channel matrix containing noise and estimation errors.

The technical scheme is as follows: the invention relates to a large-scale MIMO channel joint estimation and feedback method based on deep learning, which solves the technical problems by adopting the following technical scheme:

the method comprises the following steps:

the method comprises the following steps: transmitting a data signal containing a pilot frequency symbol at a base station end of a large-scale MIMO system, extracting the pilot frequency symbol at a pilot frequency subcarrier position from a signal received at a user end after channel transmission, calculating a channel frequency domain response of the pilot frequency subcarrier position by a least square method, then obtaining the channel frequency domain responses of other subcarrier positions by interpolation, and jointly forming an initial channel estimation value containing noise;

step two: obtaining the initial noisy channel estimation value, namely the channel matrix of the channel state information in the space-frequency domainThen, it is converted into a transformation domain which can make it sparse, and a new channel matrix under the transformation domain is obtained

Step three: constructing a super-resolution channel estimation subnet model CEnet; the model consists of a three-layer convolutional neural network with super-resolution function, and is trained, so that estimation errors are gradually reduced, and an estimated channel matrix is gradually close to an ideal channel matrix H to obtain model parameters of CEnet;

step four: constructing a channel feedback sub-network model CFnet, wherein the network comprises an encoder and a decoder, the encoder belongs to a user side, and the decoder belongs to a base station side; at a user terminal, an encoder compresses an input channel matrix H into a low-dimensional code word; at the base station, the decoder reconstructs the channel matrix values from the compressed codewords

Step five: and placing the channel feedback sub-network CFnet behind the channel estimation sub-network CEnet, forming a channel estimation and feedback joint network model CEFnet together with the channel estimation sub-network CFnet, and performing joint training. When in combined training, model parameters obtained by CEnet training in the third step are imported and fixed, the model parameters of the CFnet are updated, so that the training error is gradually reduced, and meanwhile, a channel matrix is reconstructedGradually approaching the channel matrix H to obtain CFnet model parameters, and forming the model parameters of the CEFnet together with the model parameters of the CEnet obtained in the third step;

step six: reconstruction matrix output by joint network model CEFnet of channel estimation and feedbackTurning to the initial transform domain to recover the channel matrix of the space-frequency domainA reconstructed value of (a);

step seven: and the trained channel estimation and feedback combined network model CEFnet is used for estimating and feeding back channel state information under various scenes, downlink channel estimation values are obtained through the model, code words are obtained after network compression, and the code words are fed back to the base station end for reconstruction, so that the reconstruction values of the channel matrix of the original space-frequency domain are recovered.

Wherein the content of the first and second substances,

and in the first step, the transmitting pilot frequency at the base station end is an orthogonal comb-shaped pilot frequency which is placed at equal intervals in a frequency domain, and after BPSK modulation, the transmitting pilot frequency and data symbols at other subcarrier positions jointly form a transmitting signal. And correspondingly, a channel estimation method based on pilot symbols is adopted at the user terminal to obtain an initial channel estimation value, an LS method with the lowest operation complexity is selected to obtain channel frequency domain response of the pilot subcarrier position, and then channel frequency domain response of other subcarrier positions is obtained through simple operations such as interpolation or DFT conversion and the like, so that the channel estimation values of all the subcarrier positions are obtained.

The super-resolution channel estimation subnet model CEnet in the third step is a lightweight convolutional neural network and is composed of three layers of convolutional layers, the lightweight structure of the three layers fully considers the computing capability of the user end, does not occupy too much storage space of the user end, randomly initializes parameters of each layer, inputs the channel matrix estimated by the LS method and interpolation into the subnet, and outputs the channel matrix with lower estimation error

The channel feedback sub-network model CFnet in the fourth step consists of two modules, namely an encoder and a decoder, wherein the encoder at the user end only comprises a full connection layer, the parameters of the full connection layer are initialized randomly, and the estimated complex channel matrix with sparse transform domain is usedAfter the real part and the imaginary part are separated, the components are flattened and spliced into a one-dimensional high-dimensional vector which is used as the input of the encoder, and the output is the ratioA one-dimensional low-dimensional vector with low dimensionality, namely a code word s after compression coding, is sent to a base station end through an uplink; the decoder at the base station end consists of a full connection layer, a Reshape layer and a deep de-noising residual neural network module Refinement Block, randomly initializes parameters of each layer, and inputs the received compressed code wordOutputting a reconstructed channel matrix with the same dimension as the channel matrix H

The model parameters in the third step and the fifth step mainly comprise the weight and the bias of the full connection layer and the convolution kernel and the bias of the convolution layer.

And step three, training parameters of a channel estimation subnet CEnet by adopting an Adam optimization algorithm to minimize a cost function, wherein the cost function is described as follows:

wherein M is the number of all samples in the training set, | | · | | non |)₂Is the Euclidean norm, H_iIs an ideal channel matrix;and the channel matrix after super-resolution estimation is obtained.

And fifthly, jointly training parameters of a combined channel estimation and feedback network CEFnet by adopting an Adam optimization algorithm and an end-to-end learning mode, wherein the parameters of the CEnet are obtained by training in the third step, are imported and fixed, and only the parameters of a channel feedback subnet CFnet are trained and updated, so that a cost function is minimum, and the cost function is described as follows:

wherein the content of the first and second substances,m is all the sample numbers of the training set, | · | | non-woven phosphor₂Is the Euclidean norm, H_iIs an ideal channel matrix;to reconstruct the channel matrix.

The three convolutional layers of the super-resolution channel estimation subnet model CEnet are respectively as follows: the first layer is to extract the characteristics of the noisy channel, the second layer is to perform nonlinear mapping, the characteristics of the noisy channel are mapped to the characteristics of an ideal channel, and the third layer is to perform weighted combination on the mapped characteristics to recover the original ideal channel, so as to realize super-resolution estimation of the noisy channel matrix.

In the decoder, the deep denoising residual neural network module RefinementBlock of the L layers comprises an input convolutional layer, a (L-2) layer of residual convolutional network and an output convolutional layer, wherein output data of the input layer is added with output data of the last convolutional layer of the residual network, and the output data is output after passing through the last convolutional layer.

Has the advantages that: by adopting the technical scheme, the invention can produce the following technical effects:

the invention considers the acquisition mode of the CSI data of the downlink channel in the actual communication system, makes up the defects of CsiNet and other models which carry out structure optimization based on CsiNet, and forms a complete downlink channel estimation and feedback process. Considering the computing power and the storage space of the user terminal, the channel estimation module of the user terminal adopts a lightweight network with only three layers of convolution layers. In addition, because the estimated downlink CSI contains noise and estimation errors, the original structure of the automatic encoder is converted into a noise reduction structure of the automatic encoder, which puts higher requirements on the reconstruction performance of the decoder at the base station. Therefore, the decoder module adopts a brand-new residual error network structure, obtains a larger receptive field by stacking small convolution kernels and deepening the network layer number, and ensures that the reconstruction performance can be improved by fully utilizing the original channel information.

The invention relates to a channel estimation and feedback combined network CEFnet, which mainly comprises a channel estimation sub-network and a channel feedback sub-network, wherein the channel estimation sub-network CEnet is a full convolutional layer network, the channel feedback sub-network CFnet is composed of a convolutional layer and a full connection layer, a channel matrix is estimated from a received pilot frequency through a two-step, end-to-end and data-driven training scheme, an effective compressed code word is obtained through learning a channel structure through the full connection network, the full connection layer is utilized to recover to an initial channel dimension at a base station end, and then a deep residual error network composed of the convolutional layer recovers to the original channel matrix. The scheme overcomes the defects of the existing CsiNet and the derivative model thereof in the problem of actual channel acquisition, uses the lightweight network for channel estimation, controls the increase of network parameters to a greater extent, promotes the improvement of reconstruction precision, and further improves the reconstruction performance of the network by using a brand-new deep residual error network at the base station end.

Therefore, the invention can realize the complete downlink channel estimation and feedback process based on deep learning, has higher practical significance, greatly reduces the generated estimation error of the channel estimation subnet, promotes the improvement of reconstruction precision, and further improves the reconstruction precision while reducing the feedback overhead by a brand-new noise reduction automatic encoder structure.

Drawings

FIG. 1 is a diagram of the CEFnet network architecture employed in the method of the present invention;

FIG. 2 is a three-layer lightweight convolutional super-resolution network in CEnet employed in the method of the present invention;

FIG. 3 is a Block diagram of a denoised residual neural network module Refinement Block in CFnet according to the present invention.

Detailed Description

The following describes embodiments of the present invention with reference to the drawings.

As shown in fig. 1, the present invention designs a deep learning-based large-scale MIMO channel joint estimation and feedback method. In order to verify that the method of the present invention can obtain the downlink channel estimation with a small error at the user end, and feed back the estimated noisy channel to the base station for reconstruction, and ensure a very high reconstruction accuracy, a verification example is specifically mentioned for description.

The verification example is a large-scale MIMO channel joint estimation and feedback method based on deep learning, super-resolution estimation of a channel matrix is realized at a user side through a lightweight convolutional layer structure, then, a noise reduction automatic encoder framework driven by data is adopted, the channel estimation value to be fed back is compressed and encoded into low-dimensional code words by encoders with different compression ratios at the user side, the low-dimensional code words are transmitted to a base station side through a feedback link, and an original channel matrix is reconstructed by a brand-new decoder with a 16-layer deep residual error network structure. By adopting a lightweight channel estimation subnet and a brand-new decoder structure, the method not only controls the increase of user side network parameters, but also reduces estimation errors, simultaneously reduces the feedback overhead of channel state information, and improves the channel reconstruction quality, and specifically comprises the following steps:

the method comprises the following steps: in a large-scale MIMO FDD system, a base station end is configured with 32 transmitting antennas, a user end uses a single receiving antenna, adopts an OFDM carrier modulation mode and uses 256 subcarriers. 15000 samples of space-frequency domain channel matrixes are generated in an indoor micro-cellular scene of 5.3GHz and an outdoor country scene of 300MHz by using a COST 2100 model according to the conditions, and a sample set is divided into a training set, a verification set and a test set, wherein the training set, the verification set and the test set respectively comprise 100000 samples, 30000 samples and 20000 samples. For each space-frequency domain channel matrix in the samples, using DFT matrixes F with the dimensions of 1024 × 1024 and 32 × 32 respectively_dAnd F_aChannel matrix of air-to-space frequency domainPerforming two-dimensional DFT transformation to obtain sparse channel matrix H in angular delay domain, i.e.Because the delay between the multipath arrival times is in a limited time range, the channel matrix H only has values in the first 32 rows in the time delay domain, so that the element values in the first 32 rows are retained and modified into the channel matrix H of 32 × 32, and finally the output target sample set is obtained.

Step two: constructing a communication system meeting the above transmitting and receiving conditions, and generating at the base stationAnd generating a transmitting signal containing pilot symbols, wherein the transmitting pilot is an orthogonal comb-shaped pilot frequency which is placed at equal intervals in a frequency domain, and the transmitting signal is formed together with the data symbols at other subcarrier positions after BPSK modulation. After the transmission of the sample channel generated by the COST 2100 model, a receiving signal containing a receiving pilot frequency is obtained. A channel estimation method based on pilot frequency symbols is adopted to extract the pilot frequency symbols at the positions of pilot frequency sub-carriers from signals received at a user terminal, the channel frequency domain response at the positions of the pilot frequency sub-carriers is calculated by selecting a Least Square (LS) method with the lowest operation complexity, and then the channel frequency domain response at the positions of other sub-carriers is obtained through interpolation, so that the channel estimation values of all the sub-carrier positions are obtained, and the initial channel estimation containing noise is formed. And performing two-dimensional DFT conversion on the matrix same as the output target sample matrix to obtain channel estimation in an angular delay domain

Step three: the channel estimation subnet of the user side is designed as shown in the CEnet part of the CEFnet architecture shown in fig. 1. Channel estimation matrix of complex fieldAnd the real part and the imaginary part of the ideal channel matrix H are split into two real matrices of 32 × 32 size, respectively, as input and output target samples of the two channels. Inputting initial channel estimation into a convolutional layer with a convolution kernel of 64 multiplied by 9 of CEnet in a first layer, extracting the characteristics of a noisy channel matrix, forming nonlinear mapping between the characteristics of the noisy channel and an ideal channel by a convolutional layer with a convolution kernel of 32 multiplied by 1 of a second layer, and obtaining a channel matrix estimation value with minimum estimation error after super resolution by a convolutional layer with a convolution kernel of 2 multiplied by 5 of a last layer

Step four: designing cost function of CEnet as super-resolved channel matrix output by networkThe mean square error with the true channel matrix H isWherein M is the number of all samples in the training set, | | · | | non |)₂Is the Euclidean norm, H_iIs an ideal channel matrix;and the channel matrix after super-resolution estimation is obtained. And (3) training parameters of the CEnet by using input and output training samples generated in the first step and the second step and adopting an Adam optimization algorithm, wherein the parameters comprise convolution kernels and bias of all convolution layers, so that the cost function is minimum. Each iteration calculates the gradient by dividing the training set into a batch of 200 samples, and parameters are updated according to an Adam optimization algorithm, so that the training set is traversed for 100 rounds. The learning rate is halved when the cost function value tends to be stable by adopting a variable dynamic learning rate, namely setting the initial learning rate to be 0.01. And in the training process, the verification set is used for adjusting the hyper-parameters of the model, and the test set is used for testing the final performance of the model.

Step five: the encoder of the channel feedback sub-network at the user end is designed as shown in the CFnet part of the CEFnet architecture shown in fig. 1. In the encoder, the input matrix is compressed after being flattened into a 2048 × 1 vector. Wherein the 4 times, 8 times, 16 times and 32 times of compression correspond to outputting compressed codewords of 512 dimensions, 256 dimensions, 128 dimensions and 64 dimensions, respectively.

Step six: the decoder of the channel feedback sub-network at the base station side is designed as shown in the CFnet part in the CEFnet architecture shown in fig. 1. The fed-back compressed code words are used as input of a decoder, the original 2048 × 1 vector is restored through a full connection layer, and then the vector is recombined into two matrixes with the size of 32 × 32. Inputting the two matrixes as characteristic graphs of two channels, continuously extracting characteristics by a deep residual error network module Refinement block of 16 layers, refining, and obtaining a finally reconstructed channel matrixReal and imaginary parts of (c).

Step six: designing cost function of whole CEFnet as channel matrix of network outputThe mean square error with the true channel matrix H isWherein M is the number of all samples in the training set, | | · | | non |)₂Is the Euclidean norm, H_iIn order to be the original ideal channel matrix,to reconstruct the channel matrix. And (3) carrying out joint training by using input and output training samples generated in the first step and the second step and adopting an Adam optimization algorithm and an end-to-end learning mode, importing and fixing model parameters obtained by CEnet training in the fourth step during training, and updating the model parameters of the CFnet, including convolution kernels of all convolution layers of the CFnet and weights and offsets of all connection layers, so that a cost function is minimum. Each iteration calculates the gradient by dividing the training set into 200 samples, and parameters are updated according to an Adam optimization algorithm, so that 200 rounds of the training set are traversed. The learning rate is halved when the cost function value tends to be stable by adopting a variable dynamic learning rate, namely setting the initial learning rate to be 0.001. And in the training process, the verification set is used for adjusting the hyper-parameters of the model, and the test set is used for testing the final performance of the model.

Step seven: and the trained CEFnet model is used for downlink channel estimation and feedback of the FDD large-scale MIMO system. According to the second step, after channel estimation is carried out according to the received pilot frequency, the estimated channel matrix is transformed into an angle-time delay domain channel matrixAfter the CEFnet model is input, the reconstructed angular delay domain channel sequence can be output

In summary, the invention can realize a channel estimation and feedback joint network CEFnet based on deep learning, so that compared with the prior model, the channel is obtained by estimation before feedback is considered, a complete channel estimation and feedback process is formed, the estimation error is greatly reduced through a super-resolution channel estimation subnet, the reconstruction precision is promoted to be improved, meanwhile, the feedback expense of large-scale MIMO channel information is reduced through a brand-new noise reduction automatic encoder structure, the reconstruction precision is further improved, and the efficient and practical feedback of channel state information is realized under the limited resource expense.

The above embodiments are only for illustrating the technical idea of the present invention, and the protection scope of the present invention is not limited thereby, and any modifications made on the basis of the technical solution according to the technical idea of the present invention fall within the protection scope of the present invention.