A Channel State Information Feedback Method Based on a Complex Convolutional Neural Network

Abstract

The present invention discloses a channel state information feedback method based on a complex convolutional neural network, including: establishing a communication system model for the channel condition in a frequency division duplex (FDD) large-scale multiple-input multiple-output (MIMO) system, and obtaining its final optimization objective; determining a complex network structure; constructing a CSI feedback network based on the complex network structure to obtain an original CSI matrix. The present invention realizes the compression, feedback, and reconstruction processes of the CSI matrix by constructing a complex encoder-decoder structure. In the encoder, complex convolutional downsampling is used to compress the original CSI matrix to reduce the feedback amount; in the decoder, a CDBlock structure is constructed, and by utilizing its characteristic of feature reuse, the codeword is reconstructed with high precision, so that the base station (BS) can obtain the channel state information of the network downlink.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technologies, and particularly to a method for reducing the feedback overhead of channel state information (CSI) in a large-scale multiple input multiple output (MIMO) system with frequency division duplexing (FDD). Background Art

[0002] In recent years, with the rapid development of information technology, the network user group has gradually increased, and the types and quantities of access terminals have reached an unprecedented scale. The communication traffic has grown explosively. Research shows that the number of global mobile users will reach 5.7 billion by the end of 2025, and future wireless communication systems have put forward higher requirements for mobile communication. At the same time, the rapid development of the mobile Internet has also promoted the rapid popularization of mobile communication. As an important entrance to the mobile Internet, the penetration rate of intelligent terminals has increased exponentially, greatly stimulating the development of the mobile Internet and accelerating the global users' need for wireless data services. In recent years, 5G has begun commercial deployment. As a key technology of 5G, the large-scale MIMO system meets the higher requirements of mobile communication systems for capacity and connection density, etc.

[0003] Currently, wireless communication systems have been able to deploy multiple antennas at the transmitter and receiver, and the large-scale MIMO technology has been successfully applied. The diversity and multiplexing functions of this technology can fully improve the system capacity and reliability, give play to the spatial degrees of freedom, and reduce the fading occurring in the wireless channel. The MIMO technology has been widely studied in the past few decades, and the recent research trend is developing towards large-scale MIMO systems. Large-scale MIMO is mostly distributed in the form of an array, and is also called a large antenna system. In large-scale MIMO, a large number of antennas are usually deployed at the base station. Some of the antennas can adaptively provide services for multiple users, and the additional antennas can concentrate the received signal energy for transmission, thereby significantly improving the energy efficiency and throughput. The base station of large-scale MIMO has a large number of antenna elements. When the total transmitted power is constant, each group of antennas only needs to transmit a small amount of power to meet the user requirements. As the number of antennas increases, the CSI matrices at the user side and the base station will gradually tend to be orthogonal during the transmission process, greatly reducing the interference between multiple users and effectively ensuring the communication quality between users. Compared with the traditional point-to-point MIMO technology, the large-scale MIMO technology has great advantages in terms of robustness, spectral efficiency and reliability.

[0004] These advantages of massive MIMO technology are mainly obtained by analyzing the network conditions using CSI at the base station (BS). In wireless communication, CSI describes how a signal propagates from a transmitter to a receiver through a channel, and it characterizes the comprehensive information of a series of effects, such as scattering, fading, and the attenuation of energy with distance. Channel state information enables us to adjust the transmission according to the current channel conditions, which is crucial for high-bit-rate reliable communication in multi-antenna systems. Therefore, it is very important to accurately obtain CSI in a massive MIMO system. However, since the data scale of CSI is proportional to the number of antennas at the BS, and there are hundreds of antenna elements on the BS side of a massive MIMO system, the overhead of downlink channel estimation and CSI link feedback is very large. This excessive feedback overhead makes it difficult to implement massive MIMO technology in wireless communication systems. How to ensure that the complete CSI is fed back to the base station while reducing the amount of CSI feedback is the main problem that needs to be solved by massive MIMO in wireless communication systems. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for channel state information feedback based on a complex convolutional neural network, so as to obtain CSI with a lower system overhead in a massive MIMO system, thereby enabling a wireless communication system to obtain a gain.

[0006] To achieve the above purpose, the present application proposes a method for channel state information feedback based on a complex convolutional neural network, including:

[0007] Establish a communication system model for the channel conditions in a frequency division duplex (FDD) massive MIMO system, and obtain its final optimization objective;

[0008] Determine the complex network structure;

[0009] Construct a CSI feedback network based on the complex network structure to obtain the original CSI matrix.

[0010] Further, for the channel conditions in a frequency division duplex (FDD) massive MIMO system, establish a communication system model and obtain its final optimization objective, specifically:

[0011] A single-cell FDD uses a massive MIMO system with N t antennas on the BS side, where N t >> 1, and N r antennas on the user equipment (UE) side (for the sake of convenience, let N r be equal to 1). The received signal is expressed as:

[0012] y = Αx + z (1)

[0013] Among them, N c is the number of subcarriers, represents the transmitted symbol, is the additive Gaussian noise; the original output signal Α is expressed as where i ∈ {1,..., N c} respectively represent the downlink channel coefficient and the beamforming precoding vector of subcarrier i, (·) H represents the conjugate transpose.

[0014] Furthermore, in order to obtain the beamforming precoding vector p i , the base station BS needs the user equipment UE to feedback the corresponding channel coefficient h i , so it is assumed that the downlink CSI matrix is which includes N c N t elements, and the number of parameters to be feedback is 2N c N t , which is proportional to the number of antennas.

[0015] Since the CSI matrix H is sparse in the angular delay domain, through two-dimensional discrete Fourier transform, the original form of the spatio-frequency domain CSI is converted into the angular delay domain, so that:

[0016] H′ = F c HF t H (2)

[0017] where F c and F t are discrete Fourier transform matrices of dimensions N c × N c , N t × N t respectively. For the angular delay domain CSI matrix H′, each element corresponds to a specific path delay with an angle of arrival; in H′, only the first row contains useful information, and the remaining rows represent paths with larger propagation delays and consist of values close to zero, which can be omitted without losing too much information; let H a represent the effective information row of H′.

[0018] Furthermore, input the effective information row H a into the encoder of the user equipment UE to generate a codeword v according to the given compression ratio η, so that:

[0019] v = f ε (H a , Θ ε ) (3)

[0020] where fε Denotes the encoding process, Θ ε Denotes a set of parameters of the encoder;

[0021] Once the base station BS receives the codeword v, the decoder reconstructs the channel:

[0022]

[0023] where Denotes the decoding process, Denotes a set of parameters of the decoder, so the entire feedback process can be expressed as:

[0024]

[0025] Therefore, the final optimization goal is to minimize the difference between the original H a and the reconstruction This difference is expressed as finding a set of encoder and decoder parameter sets that meet the conditions:

[0026]

[0027] Furthermore, determine the complex network structure, specifically:

[0028] When the convolutional neural network works, the convolutional kernel slides from the upper left corner of the CSI matrix, traverses the values at each position in the CSI matrix according to a certain step size, and performs a convolution operation at each position;

[0029] Complex convolution divides the convolutional kernel into two equal parts and operates on the real and imaginary parts of the CSI matrix respectively. The operation process is shown in Equation 7, where K I 、K R 、M I 、M R Denote the imaginary part convolutional kernel, real part convolutional kernel, imaginary part feature map, and real part feature map respectively;

[0030]

[0031] Even further, use complex batch normalization operation to normalize the convolution results obtained by each layer of the convolutional neural network to a normal distribution with a mean of 0 and a standard deviation of 1; The complex batch normalization operation uses two parameters γ and β. The shift parameter β is a complex parameter with two learnable components (real part and imaginary part); The scaling parameter γ is a 2×2 positive semi-definite matrix with only three degrees of freedom, which is expressed as:

[0032]

[0033] where γ rr Is the scaling parameter of the two real parts; γ riis the scaling parameter with the first being the real part and the second being the imaginary part; γ ir is the scaling parameter with the first being the imaginary part and the second being the real part; γ ii is the scaling parameter for two imaginary parts;

[0034] Due to the values after normalizing the CSI matrix having a real variance and an imaginary variance of 1, γ rr and γ ii are initialized to to obtain a variance modulus of 1 for the normalized values, γ ri and γ ir are initialized to 0, and the complex batch normalization is defined as:

[0035]

[0036] Furthermore, a CSI feedback network based on a complex network structure is constructed to obtain the original CSI matrix. Specifically: The complex network structure is introduced into the CSI feedback network, and the spatial local correlation is utilized to strengthen the local connection between adjacent layer neurons; A complex encoder-decoder structure is constructed. In the encoder, complex convolutional downsampling is used to extract CSI features, and a complex fully connected layer is used to compress the CSI matrix; In the decoder, a complex fully connected layer is used to restore the original size of the CSI matrix, and the compressed codewords are reconstructed into the original CSI matrix through the complex dense connection module CDBlock;

[0037] Furthermore, each layer of the complex dense connection module CDBlock obtains additional feature maps from all previous layers and passes the feature maps of the current layer to all subsequent layers. Then the feature mapping received by the l-th layer from all previous layers is expressed as:

[0038] x l =C l ([x 0 ,x 1 ,...,x l-1 ) (8)

[0039] where [x 0 ,x 1 ,...,x l-1 represents the concatenation of the feature maps generated in layers 0,…,l-1, and C l is a composite function of three consecutive operations of complex batch normalization, exponential linear unit, and complex convolution.

[0040] As a further aspect, the encoder includes a feature extraction module and a compression module; the first layer of the feature extraction module is the input layer, and the input data format is 32×32×2. When the compression ratio is 4, after 4 times of complex convolutional downsampling, the specification of the feature map obtained is 2×2×512; the compression module uses a complex fully connected layer to compress the CSI matrix, taking the 2×2×512 structure output by the feature extraction module as the input, and compressing it into a vector, that is, the compressed codeword v is obtained; the codeword v is sent to the base station BS through the feedback link.

[0041] As a further aspect, the decoder directly connects the features of each layer by matching the feature maps. The current feature map of each layer passes through all subsequent layers, and its final output is the concatenation of the previous layer's feature maps. Specifically:

[0042] After the base station BS obtains the codeword v, first, the compressed feature vector is directly restored to the original size of N c ×N t ×2 through the complex fully connected layer, forming a rough reconstruction of the CSI matrix; subsequently, the complex densely connected module further refines the CSI matrix and uses zero-padding to keep the size of the input feature map unchanged;

[0043] The complex densely connected module uses three convolutional layers with 2, 4, and 1 channels as the basic reconstruction module. 5 reconstruction modules continuously refine the reconstruction of the CSI matrix through dense connection and output a feature map with a shape of 32×32×12. The features of this Figure 5 reconstruction module are connected; then, the original CSI matrix is reconstructed through a convolutional layer with a 3×3 convolutional kernel having 2 channels.

[0044] The above technical solution adopted by the present invention, compared with the prior art, has the following advantages: The present invention realizes the compression, feedback, and reconstruction process of the CSI matrix by constructing a complex encoder-decoder structure. In the encoder, complex convolutional downsampling is used to compress the original CSI matrix to reduce the feedback amount; in the decoder, a CDBlock structure is constructed, and by utilizing its feature reuse characteristic, the codeword is reconstructed with high precision, so that the BS can obtain the channel state information of the network downlink. This method solves the problems of large feedback overhead and low reconstruction accuracy of the existing methods in the large-scale MIMO system in the FDD mode by using deep learning methods. Description of the Drawings

[0045] Figure 1 is the flowchart of the implementation process of the present invention;

[0046] Figure 2 is the implementation diagram of the complex convolution of the present invention;

[0047] Figure 3 CDBlock structure diagram of the present invention;

[0048] Figure 4 CVCsiNet structure and feedback model diagram proposed by the present invention;

[0049] Figure 5 Encoder structure diagram of the present invention when the compression ratio is 1 / 4;

[0050] Figure 6 Decoder structure diagram of the present invention when the compression ratio is 1 / 4. Detailed implementation manners

[0051] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application, that is, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments.

[0052] Embodiment 1

[0053] As Figure 1 shown, the present application provides a channel state information feedback method based on a complex convolutional neural network, specifically including:

[0054] S1. For the channel conditions in a frequency division duplexing (FDD) large-scale MIMO system, establish a communication system model and obtain its final optimization objective;

[0055] Specifically, consider a single-cell FDD large-scale MIMO system, with N t antennas on the BS side, where N t >> 1, and N r antennas on the UE side (for the sake of convenience of explanation, let N r be equal to 1). The received signal can be expressed as:

[0056] y = Αx + z (1)

[0057] where N c is the number of subcarriers, represents the transmitted symbol, is the additive Gaussian noise; the original output signal Α is expressed as where i ∈ {1,..., N c} respectively represent the downlink channel coefficient and beamforming precoding vector of subcarrier i, and (·) H represents the conjugate transpose.

[0058] To obtain the beamforming precoding vector p i , the base station BS needs the user equipment UE to feedback the corresponding channel coefficient h i . Therefore, it is assumed that the downlink CSI matrix is which includes N c N t elements, and the number of parameters to be fed back is 2N c N t , which is proportional to the number of antennas;

[0059] Since the CSI matrix H is sparse in the angular delay domain, through two-dimensional discrete Fourier transform, the original form of the CSI in the spatial-frequency domain is converted into the angular delay domain, such that:

[0060] H′ = F c HF t H (2)

[0061] where F c and F t are the discrete Fourier transform matrices of dimension N c ×N c , N t ×N t respectively. For the angular delay domain CSI matrix H′, each element corresponds to a specific path delay with an angle of arrival; in H′, only the first row contains useful information, and the remaining rows represent paths with larger propagation delays and are composed of values close to zero, which can be omitted without losing too much information. Let H a represent the effective information row of H′.

[0062] Input the effective information row H a into the encoder of the user equipment UE, and generate a codeword v according to the given compression ratio η, such that:

[0063] v = f ε (H a , Θ ε ) (3)

[0064] where f ε represents the encoding process, and Θ ε represents a set of parameters of the encoder;

[0065] Once the base station BS receives the codeword v, the decoder reconstructs the channel:

[0066]

[0067] where represents the decoding process, represents a set of parameters of the decoder. Therefore, the entire feedback process can be expressed as:

[0068]

[0069] Therefore, the final optimization goal is to minimize the difference between the original H a and the reconstruction This difference is expressed as finding a set of encoder and decoder parameter sets that meet the conditions:

[0070]

[0071] S2. Determine the complex network structure;

[0072] Specifically, according to the operation principle of complex numbers, the present invention regards complex numbers as two-dimensional real number pairs. The addition process is the same as that of real number operations. In the present invention, a complex number is represented as z = A + ib, with a real component A and an imaginary component b. The input and weights of the network are divided into two parts;

[0073] When the convolutional neural network works, the convolutional kernel slides from the upper left corner of the CSI matrix, traverses the values at each position in the CSI matrix according to a certain step size, and performs a convolution operation at each position. The structure proposed by the present invention is based on the deep complex network. The difference between the complex convolution operation and the real number convolution operation is that the complex convolution divides the convolutional kernel into two parts on average, and operates on the real part and the imaginary part of the input respectively. The operation process is as shown in Equation 7. Where K I 、K R 、M I 、M R respectively represent the imaginary part convolutional kernel, the real part convolutional kernel, the imaginary part feature map, and the real part feature map;

[0074]

[0075] In the convolutional neural network, the update of the parameters of each layer will cause the data distribution of the input of each layer to change. After multiple layers of operations, the input distribution of the high layer will be significantly different from the initial distribution. This causes the data distribution that the network needs to adapt to change continuously in each iteration, making the network training difficult. At the same time, the change of the data distribution may cause the data to fall into the saturation area of the activation function, affecting the parameter update of the network.

[0076] The present invention uses a plural batch normalization operation to normalize the convolution results obtained by each layer of the convolutional neural network to a normal distribution with a mean of 0 and a standard deviation of 1, ensuring the consistency of the data distribution during network training. However, the real number batch normalization operation is not applicable to complex numbers. To normalize a set of complex numbers to a standard normal complex distribution, a normalization algorithm is required to decorrelate the imaginary part and the real part of a unit, thereby reducing the risk of overfitting. The plural batch normalization operation uses two parameters γ and β. The shift parameter β is a complex parameter with two learnable components (real part and imaginary part); the scaling parameter γ is a 2×2 positive semi - definite matrix with only three degrees of freedom, so there are only three learnable components. The scaling parameter γ is expressed as:

[0077]

[0078] where γ rr is the scaling parameter of the two real parts; γ ri is the scaling parameter with the first being the real part and the second being the imaginary part; γ ir is the scaling parameter with the first being the imaginary part and the second being the real part; γ ii is the scaling parameter of the two imaginary parts;

[0079] Since the normalized value of the CSI matrix has a real variance and an imaginary variance of 1, γ rr and γ ii are initialized to to obtain a variance modulus of 1 for the normalized value, and γ ri and γ ir are initialized to 0, then the plural batch normalization is defined as:

[0080]

[0081] S3. Construct a CSI feedback network based on a complex network structure to obtain the original CSI matrix;

[0082] Specifically, introduce a complex network structure into the CSI feedback network, utilize the spatial local correlation to strengthen the local connection between adjacent - layer neurons; construct a complex - form encoder - decoder structure. In the encoder, use the method of complex convolution downsampling to extract CSI features, and use a complex fully - connected layer to compress the CSI matrix; in the decoder, use a complex fully - connected layer to restore the original size of the CSI matrix, and reconstruct the compressed codeword into the original CSI matrix through a complex dense connection module CDBlock.

[0083] The present invention considers the characteristics such as the correlation and sparsity of the wireless channel in a large - scale MIMO system. When using the complex dense connection module CDBlock, first utilize the correlation of the CSI matrix, and secondly consider that the denseblock can promote the feature reuse of the network. As shown in the appendixFigure 2 As shown, where CConv represents the complex convolution structure. To ensure the maximization of information flow between layers of the network, each layer of the CDBlock obtains additional input feature maps from all previous layers and passes the current feature map to all subsequent layers. The l-th layer receives the feature maps from all previous layers:

[0084] x l = C l ([x 0 , x 1 ,..., x l-1 ) (8)

[0085] where [x 0 , x 1 ,..., x l-1 represents the concatenation of the feature maps generated at layers 0,…, l-1, and C l is a composite function of three consecutive operations: complex batch normalization, exponential linear unit, and complex convolution. The CDBlock helps train deeper network architectures. In addition, the dense connection has a regularization effect, which can reduce overfitting for tasks with a small training set.

[0086] As Figure 3 shown, the present invention proposes a CSI feedback network CVCsiNet based on a complex network structure, which uses a complex convolutional neural network to construct an encoder and a decoder;

[0087] As Figure 2 shown, for the encoder, where S1×S2×S3 respectively represent the length, width, and number of the feature maps; the encoder includes a feature extraction module and a compression module; the feature extraction module extracts the CSI matrix features in the way of complex convolution downsampling, and at the same time uses complex batch normalization to improve the convergence speed of the model and alleviate the problem of gradient dispersion in the network, so that the network model is more likely to be stable. The complex convolution downsampling process controls the step size and has good information fusion. While the traditional method uses the sampling method of the pooling layer, there is a possibility of filtering out useful information. Therefore, the structure used in the present invention can better extract data features. The specific setting parameters of the network structure of the feature extraction module can be strides = 2, padding = same, kernel_size = 3. Its first layer is the input layer, and the input data format is 32×32×2. Under the compression ratio of 4, after 4 times of complex convolution downsampling, the specification of the obtained feature map is 2×2×512; the compression module uses a complex fully connected layer to compress the CSI matrix, takes the 2×2×512 structure output by the feature extraction module as the input, and compresses it into a vector, that is, the compressed codeword v is obtained; the codeword v is sent to the base station BS through the feedback link.

[0088] As Figure 2 shown by the plural DenseNet decoders, in order to maximize the inter-layer information fusion, DenseNet directly connects each layer in the network by matching feature maps. The current feature map of each layer passes through all subsequent layers, and the final output is the concatenation of the feature maps of the previous layer. DenseNet utilizes the potential of the network through feature reuse to generate a compact model that is easy to train and has high parameter efficiency. Connecting the feature maps of different layers increases the variation of the input to the subsequent layer and improves the efficiency. This structure is simpler and more efficient than before. Based on the idea of DenseNet, this structure not only utilizes the powerful feature reuse ability of DenseNet but also combines the characteristics of the complex convolutional neural network, which can improve the accuracy of CSI recovery.

[0089] After the base station BS obtains the codeword v, it first directly restores the compressed feature vector to the original size of N c ×N t ×2 through a complex fully connected layer to form a rough reconstruction of the CSI matrix; then the complex dense connection module further refines the CSI matrix and uses zero padding to keep the size of the input feature map unchanged; by using global dense connection, information from the first layer can be detected in the last layer, and this globality improves the efficiency of the information flow at the cost of a slightly increased computational cost.

[0090] The complex dense connection module uses three convolutional layers with 2, 4, and 1 channels as the basic reconstruction module. Five reconstruction modules continuously refine the reconstruction of the CSI matrix through dense connection and output a feature map with a shape of 32×32×12. The features Figure 5 of these five reconstruction modules are connected; then the original CSI matrix is reconstructed through a convolutional layer with a 3×3 convolutional kernel having 2 channels.

[0091] The foregoing description of the specific exemplary embodiments of the present invention is for purposes of illustration and exemplification. These descriptions are not intended to limit the present invention to the precise forms disclosed, and obviously, many changes and variations are possible in light of the above teachings. The purpose of selecting and describing the exemplary embodiments is to explain the specific principles of the present invention and its practical applications, so that those skilled in the art can implement and utilize various different exemplary embodiments of the present invention as well as various different selections and changes. The scope of the present invention is intended to be defined by the claims and their equivalents.

Claims

1. A method for channel state information feedback based on a complex convolutional neural network, characterized in that, it includes: For the channel condition in a frequency division duplex (FDD) massive multiple-input multiple-output (MIMO) system, a communication system model is established and its final optimization objective is obtained, specifically: A single-cell FDD uses a massive MIMO system with N t antennas at the base station BS side, where N t >> 1, and N r antennas at the user equipment UE side. The received signal is expressed as: y = Αx + z (1) where N c is the number of subcarriers, represents the transmitted symbol, is the additive Gaussian noise; the original output signal Α is expressed as where respectively represent the downlink channel coefficient and the beamforming precoding vector of subcarrier i, (·) H denotes the conjugate transpose; To obtain the beamforming precoding vector p i , the base station BS needs the user equipment UE to feedback the corresponding channel coefficient h i . Therefore, it is assumed that the downlink CSI matrix is which includes N c N t elements, and the number of parameters to be feedback is 2N c N t , which is proportional to the number of antennas; Since the CSI matrix H is sparse in the angular delay domain, through two-dimensional discrete Fourier transform, the original form of the CSI in the spatial-frequency domain is converted into the angular delay domain, so that: where F c and F t are the discrete Fourier transform matrices of dimension N c ×N c , N t ×N t respectively. For the angular delay domain CSI matrix H′, each element corresponds to a specific path delay with an angle of arrival; in H′, only the first row contains useful information, and the remaining rows represent paths with larger propagation delays and consist of values close to zero; let H a represent the valid information row of H′; Input the valid information line H a into the encoder of the user equipment UE, and generate a codeword v according to the given compression ratio η such that: v = f E (H a , Θ E ) (3) where f E represents the encoding process, and Θ E represents a set of parameters of the encoder; Once the base station BS receives the codeword v, the decoder reconstructs the channel: Among them represents the decoding process, represents a set of parameters of the decoder. Therefore, the entire feedback process is expressed as: Therefore, the final optimization goal is to minimize the difference between the original H a and the reconstruction This difference is expressed as finding a set of encoder and decoder parameters that satisfy the conditions: Determine the complex network structure; Construct a CSI feedback network based on the complex network structure to obtain the original CSI matrix.

2. The method for channel state information feedback based on a complex convolutional neural network according to claim 1, characterized in that, Determine the complex network structure, specifically: When the convolutional neural network is working, the convolutional kernel slides from the upper left corner of the CSI matrix, traverses the values at each position in the CSI matrix according to a certain step size, and performs a convolution operation at each position; Complex convolution divides the convolution kernel into two equal parts and operates on the real and imaginary parts of the CSI matrix respectively. The operation process is shown in Equation 7, where K I and K R , M I and M R represent the imaginary part convolution kernel, the real part convolution kernel, the imaginary part feature map, and the real part feature map respectively; 3. The method for channel state information feedback based on a complex convolutional neural network according to claim 2, characterized in that, Adopt complex batch normalization operation to normalize the convolution results obtained by each layer of the convolutional neural network to a normal distribution with a mean of 0 and a standard deviation of 1; the complex batch normalization operation uses two parameters γ and β, and the shift parameter β is a complex parameter with two learnable components; the scaling parameter γ is a 2×2 positive semi-definite matrix with only three degrees of freedom, which is expressed as: where γ rr is the scaling parameter for two real parts; γ ri is the scaling parameter with the first being the real part and the second being the imaginary part; γ ir is the scaling parameter with the first being the imaginary part and the second being the real part; γ ii is the scaling parameter for two imaginary parts; Since the values after CSI matrix normalization have real variance and imaginary variance 1, initialize γ rr and γ ii to obtain the variance modulus 1 of the normalized values, initialize γ ri and γ ir to 0, then the complex batch normalization is defined as:

4. The method for channel state information feedback based on a complex convolutional neural network according to claim 1, characterized in that, Construct a CSI feedback network based on the complex network structure to obtain the original CSI matrix, specifically: introduce the complex network structure into the CSI feedback network, utilize the spatial local correlation to strengthen the local connection between adjacent layer neurons; construct a complex encoder-decoder structure, use the method of complex convolutional downsampling to extract CSI features in the encoder, and use a complex fully connected layer to compress the CSI matrix; in the decoder, use a complex fully connected layer to restore the original size of the CSI matrix, and reconstruct the compressed codeword into the original CSI matrix through a complex dense connection module CDBlock.

5. The method for channel state information feedback based on a complex convolutional neural network according to claim 4, characterized in that, Each layer of the complex dense connection module CDBlock obtains additional feature maps from all previous layers and passes the feature maps of the current layer to all subsequent layers. Then the feature mapping received by the l-th layer from all previous layers is expressed as: x l = C l ([x 0 , x 1 ,..., x l-1 ) (8) where [x 0 , x 1 ,..., x l-1 represents the concatenation of the feature maps generated at layers 0, …, l - 1, and C l is a composite function of three consecutive operations: complex batch normalization, exponential linear unit, and complex convolution.

6. The method for channel state information feedback based on a complex convolutional neural network according to claim 4, characterized in that, The encoder includes a feature extraction module and a compression module; The first layer of the feature extraction module is the input layer, and the input data format is 32×32×2. Under the condition of a compression ratio of 4, after 4 times of complex convolutional downsampling, the specification of the feature map obtained is 2×2×512; the compression module uses a complex fully connected layer to compress the CSI matrix, takes the 2×2×512 structure output by the feature extraction module as the input, and compresses it into a vector, that is, the compressed codeword v is obtained; the codeword v is sent to the base station BS through the feedback link.

7. The method for channel state information feedback based on a complex convolutional neural network according to claim 4, characterized in that, The decoder directly connects the features of each layer by matching the feature maps. The current feature map of each layer passes through all subsequent layers, and its final output is the concatenation of the previous layer's feature maps. Specifically: After the base station BS obtains the codeword v, the compressed feature vector is first directly restored to the original size of N c ×N t ×2 through the complex fully connected layer, forming a rough reconstruction of the CSI matrix; subsequently, the complex densely connected module further refines the CSI matrix and uses zero padding to keep the size of the input feature map unchanged; The complex dense connection module uses three convolutional layers with 2, 4, and 1 channels as the basic reconstruction module. Five reconstruction modules continuously refine the reconstruction of the CSI matrix through dense connection and output a feature map with a shape of 32×32×12. The feature maps of the five reconstruction modules are connected; then, the original CSI matrix is reconstructed through a convolutional layer with a 3×3 convolutional kernel having 2 channels.

Images