CN115021787A - Channel state information feedback method based on complex convolutional neural network - Google Patents
Channel state information feedback method based on complex convolutional neural network Download PDFInfo
- Publication number
- CN115021787A CN115021787A CN202210598753.1A CN202210598753A CN115021787A CN 115021787 A CN115021787 A CN 115021787A CN 202210598753 A CN202210598753 A CN 202210598753A CN 115021787 A CN115021787 A CN 115021787A
- Authority
- CN
- China
- Prior art keywords
- complex
- csi
- csi matrix
- network
- layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B7/00—Radio transmission systems, i.e. using radiation field
- H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
- H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
- H04B7/06—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station
- H04B7/0613—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission
- H04B7/0615—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission of weighted versions of same signal
- H04B7/0619—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission of weighted versions of same signal using feedback from receiving side
- H04B7/0621—Feedback content
- H04B7/0626—Channel coefficients, e.g. channel state information [CSI]
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L25/00—Baseband systems
- H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
- H04L25/0202—Channel estimation
- H04L25/024—Channel estimation channel estimation algorithms
- H04L25/0254—Channel estimation channel estimation algorithms using neural network algorithms
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Abstract
The invention discloses a channel state information feedback method based on a complex convolutional neural network, which comprises the following steps: aiming at the channel condition under a frequency division duplex FDD large-scale MIMO system, establishing a communication system model and obtaining a final optimization target of the communication system model; determining a plurality of network structures; and constructing a CSI feedback network based on a complex network structure to obtain an original CSI matrix. The invention realizes the processes of compression, feedback and reconstruction of the CSI matrix by constructing a complex encoder-decoder structure. Compressing the original CSI matrix using complex convolution downsampling in an encoder to reduce the amount of feedback; a CDBlock structure is constructed in a decoder, and a code word is reconstructed with high precision by using the characteristic of characteristic reuse, so that the BS can obtain downlink channel state information of a network.
Description
Technical Field
The present invention relates to the field of wireless communication technologies, and in particular, to a method for reducing Channel State Information (CSI) feedback overhead in a large-scale Multiple Input Multiple Output (MIMO) system with Frequency Division Duplex (FDD).
Background
In recent years, with the rapid development of information technology, the number of network user groups is gradually increased, the types and the number of access terminals reach unprecedented scales, and the communication traffic is explosively increased, researches show that the number of global mobile users reaches 57 hundred million by the end of 2025, and a wireless communication system puts higher demands on mobile communication in the future. Meanwhile, the rapid development of the mobile internet also promotes the rapid popularization of mobile communication, and the popularization rate of the intelligent terminal serving as an important entrance of the mobile internet exponentially increases, thereby greatly stimulating the development of the mobile internet and accelerating the requirement of global users on wireless data services. In recent years, 5G is beginning to be commercially deployed, and a large-scale MIMO system is taken as a key technology of 5G, so that higher requirements of a mobile communication system on capacity, connection density and the like are met.
Currently, wireless communication systems have been able to deploy multiple antennas at the transmitter and receiver, successfully applying massive MIMO techniques. The diversity and multiplexing functions of the technology can fully improve the system capacity and reliability, exert the spatial degree of freedom and reduce the fading in the wireless channel. MIMO technology has been extensively studied in the past decades, and recent research trends are moving towards massive MIMO systems. Massive MIMO is distributed in the form of arrays, also referred to as large antenna systems. In massive MIMO, a base station is usually deployed with a large number of antennas, some of which can adaptively provide services for multiple users, and additional antennas can transmit received signal energy intensively, thereby significantly improving energy efficiency and throughput. The base station of the large-scale MIMO has a large number of antenna elements, and when the total power transmitted is constant, each group of antennas only needs to transmit smaller power to meet the requirements of users. With the increase of the number of the antennas, the CSI matrixes at the user side and the base station tend to be orthogonal gradually in the sending process, so that the interference among multiple users is greatly reduced, and the communication quality among the users can be effectively ensured. Compared with the traditional point-to-point MIMO technology, the massive MIMO technology has great advantages in robustness, spectrum efficiency and reliability.
These advantages of massive MIMO technology are mainly obtained by analyzing network conditions at the Base Station (BS) using CSI, which describes how a signal propagates from a transmitter to a receiver through a channel in wireless communications, and characterizes a comprehensive set of information about the effects, such as scattering, fading, and energy attenuation with distance. The channel state information enables us to adapt the transmission according to the current channel conditions, which is crucial for high bit rate reliable communication of multi-antenna systems. Therefore, it is very important to accurately obtain CSI in a massive MIMO system, but since the data size of CSI is proportional to the number of antennas at the BS, while there are several hundred antenna elements on the BS side of the massive MIMO system, the overhead of downlink channel estimation and CSI link feedback is very large. The too high feedback overhead causes that the massive MIMO technology is difficult to realize in the wireless communication system, and how to ensure that the CSI feedback quantity is reduced while the complete CSI is fed back to the base station is the main problem that the massive MIMO needs to solve in the wireless communication system.
Disclosure of Invention
The invention aims to provide a channel state information feedback method based on a complex convolutional neural network, so that CSI can be obtained in a large-scale MIMO system at low system overhead, and a wireless communication system can obtain gain.
In order to achieve the above object, the present application provides a channel state information feedback method based on a complex convolutional neural network, including:
aiming at the channel condition under a frequency division duplex FDD large-scale MIMO system, establishing a communication system model and obtaining a final optimization target of the communication system model;
determining a plurality of network structures;
and constructing a CSI feedback network based on a complex network structure to obtain an original CSI matrix.
Further, a communication system model is established for channel conditions under a frequency division duplex FDD large-scale MIMO system, and a final optimization target is obtained, specifically:
a single cell FDD uses massive MIMO system, at base station BS sideWith N t An antenna, wherein N t > 1, there is N at the UE side of the user equipment r An antenna (for convenience of explanation, let N r Equal to 1), received signalExpressed as:
y=Αx+z (1)
wherein N is c Is the number of sub-carriers,a symbol representing the transmission of the data is transmitted,additive Gaussian noise; the original output signal A is represented asWherein i∈{1,...,N c Represents the downlink channel coefficient and beamforming precoding vector of the subcarrier i, respectively (·) H Representing a conjugate transpose.
Further, to derive a beamforming precoding vector p i The base station BS needs the user equipment UE to feed back the corresponding channel coefficient h i Therefore, assume the downlink CSI matrix isWhich comprises N c N t The number of the parameters needing to be fed back is 2N c N t Which is proportional to the number of antennas.
Because the CSI matrix H is sparse in the angular delay domain, the original form of the space-frequency domain CSI is converted into the angular delay domain through two-dimensional discrete Fourier transform, so that:
H′=F c HF t H (2)
wherein F c And F t Respectively is dimension N c ×N c ,N t ×N t For the angular delay domain CSI matrix H', each element corresponds to a particular path delay with an angle of arrival; in H', only the first row contains useful information, the remaining rows represent paths with large propagation delays and consist of values close to zero, which can be omitted without losing too much information; let H a The valid information line representing H'.
Further, the effective information row H a Input into an encoder of a user equipment UE, a codeword v is generated according to a given compression ratio η such that:
v=f ε (H a ,Θ ε ) (3)
wherein f is ε Denotes the encoding procedure, [ theta ] ε A set of parameters representing an encoder;
once the base station BS receives the codeword v, the decoder reconstructs the channel:
whereinWhich is indicative of the decoding process, is,represents a set of parameters of the decoder, so the whole feedback process can be expressed as:
so the final optimization objective is to minimize the original H a And reconstructionDifference between them, such difference being expressed as finding a satisfaction barThe encoder and decoder parameter sets of piece:
further, determining a plurality of network structures specifically includes:
when the convolutional neural network works, a convolutional kernel starts to slide from the upper left corner of the CSI matrix, the value of each position in the CSI matrix is traversed according to a certain step length, and convolution operation is carried out at each position;
the complex convolution equally divides the convolution kernel into two parts, and respectively operates the real part and the imaginary part of the CSI matrix, and the operation process is shown as formula 7, wherein K is I 、K R 、M I 、M R Respectively representing an imaginary part convolution kernel, a real part convolution kernel, an imaginary part characteristic diagram and a real part characteristic diagram;
furthermore, a plurality of batches of normalization operations are adopted to normalize the convolution results obtained by each layer of the convolutional neural network to normal distribution with the average value of 0 and the standard deviation of 1; the complex batch normalization operation uses two parameters γ and β, the shift parameter β being a complex parameter having two learnable components (real and imaginary parts); the scaling parameter γ is a 2 × 2 semi-positive definite matrix with only three degrees of freedom, which is expressed as:
wherein gamma is rr Scaling parameters that are two real parts; gamma ray ri The scaling parameter is the first real part and the second imaginary part; gamma ray ir Is the scaling parameter with the first as the imaginary part and the second as the real part; gamma ray ii Scaling parameters that are two imaginary parts;
due to normalized values of the CSI matrixWith a real variance and a virtual variance of 1, will be rr And gamma ii Is initialized toTo obtain a normalized value of the modulus of variance 1, gamma ri And gamma ir With an initialization of 0, the plural batch normalization is defined as:
further, a CSI feedback network based on a complex network structure is constructed to obtain an original CSI matrix, which specifically comprises: introducing a plurality of network structures into a CSI feedback network, and reinforcing local connection between neurons of adjacent layers by utilizing spatial local correlation; constructing a complex encoder-decoder structure, extracting CSI characteristics in the encoder by using a complex convolution downsampling mode, and compressing a CSI matrix by using a complex full-connection layer; restoring the original size of the CSI matrix by using a complex full connection layer in the decoder, and reconstructing a compressed code word into the original CSI matrix through a complex dense connection module CDBlock;
furthermore, each layer of the plural dense connection module CDBlock obtains additional feature maps from all previous layers and transfers the feature map of the current layer to all subsequent layers, and the l-th layer receives the feature maps of all previous layers as:
x l =C l ([x 0 ,x 1 ,...,x l-1 ]) (8)
wherein [ x ] 0 ,x 1 ,...,x l-1 ]Showing the concatenation of signatures generated at level 0, …, l-1, C l For complex batch normalization, an exponential linear unit and a complex convolution three continuously operating complex functions.
As a further step, the encoder comprises a feature extraction module and a compression module; the first layer of the feature extraction module is an input layer, and the input data format is 3232 × 2, and the specification of the feature map obtained after 4 times of complex convolution downsampling is 2 × 2 × 512 when the compression rate is 4; the compression module compresses the CSI matrix by adopting a plurality of full connection layers, and compresses the 2 multiplied by 512 structure output by the feature extraction module into oneObtaining a compressed code word v; the codeword v is sent to the base station BS over a feedback link.
As a further step, 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 the final output of the current feature map of each layer is the concatenation of the feature maps of the previous layer, specifically:
after the base station BS obtains the code word v, firstly, the compressed feature vector is directly restored to N through a complex full-connection layer c ×N t The original size of x 2, constituting a rough reconstruction of the CSI matrix; then, the CSI matrix is further refined by a complex dense connection module, and the size of the input characteristic diagram is kept unchanged by zero padding;
the complex dense connection module adopts three convolution layers with 2, 4 and 1 channels as basic reconstruction modules, 5 reconstruction modules continuously refine the reconstruction of the CSI matrix through dense connection and output a characteristic diagram with the shape of 32 multiplied by 12, and the characteristic diagram of 5 reconstruction modules is connected; the reconstructed original CSI matrix is then obtained by convolutional layers of 3 × 3 convolutional kernels with 2 channels.
Compared with the prior art, the technical scheme adopted by the invention has the advantages that: the invention realizes the processes of compression, feedback and reconstruction of the CSI matrix by constructing a complex encoder-decoder structure. Compressing an original CSI matrix in an encoder by using complex convolution downsampling to reduce the feedback quantity; and constructing a CDBlock structure in a decoder, and reconstructing a code word with high precision by using the characteristic of characteristic reuse, so that the BS can obtain the downlink channel state information of the network. The method solves the problems of high feedback overhead and low reconstruction precision of the existing method by using a deep learning method under a large-scale MIMO system in an FDD mode.
Drawings
FIG. 1 is a flow chart of an implementation of the present invention;
FIG. 2 is a diagram of a complex convolution implementation of the present invention;
FIG. 3 is a diagram of the CDBlock architecture of the present invention;
FIG. 4 is a diagram of a CVCsiNet structure and a feedback model according to the present invention;
FIG. 5 is a block diagram of an encoder of the present invention having a compression ratio of 1/4;
fig. 6 is a block diagram of a decoder having a compression ratio of 1/4 according to the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the application, i.e., the embodiments described are only a subset of, and not all embodiments of the application.
Example 1
As shown in fig. 1, the present application provides a channel state information feedback method based on a complex convolutional neural network, which specifically includes:
s1, aiming at the channel condition under a frequency division duplex FDD large-scale MIMO system, establishing a communication system model and obtaining a final optimization target of the communication system model;
specifically, consider a single cell FDD using massive MIMO system with N on the BS side t An antenna, wherein N t > 1, with N at the UE side r An antenna (for convenience of explanation, let N r Equal to 1), received signalCan be expressed as:
y=Αx+z (1)
wherein N is c Is the number of sub-carriers,presentation renderingThe sign of the input is displayed on the display,additive Gaussian noise; the original output signal A is represented asWherein i∈{1,...,N c Represents the downlink channel coefficient and beamforming precoding vector of the subcarrier i, respectively (·) H Representing a conjugate transpose.
To derive a beamforming precoding vector p i The base station BS needs the user equipment UE to feed back the corresponding channel coefficient h i Therefore, assume the downlink CSI matrix isWhich comprises N c N t The number of the parameters needing to be fed back is 2N c N t Which is proportional to the number of antennas;
because the CSI matrix H is sparse in the angular delay domain, the original form of the space-frequency domain CSI is converted into the angular delay domain by two-dimensional discrete fourier transform, so that:
H′=F c HF t H (2)
wherein F c And F t Respectively is dimension N c ×N c ,N t ×N t For the angular delay domain CSI matrix H', each element corresponds to a particular path delay with an angle of arrival; in H', only the first row contains useful information, the remaining rows represent paths with large propagation delays and consist of values close to zero, which can be omitted without losing too much information. Let H a The valid information line representing H'.
Will be valid information row H a Input to user equipment UEIn the encoder, a codeword v is generated according to a given compression ratio η such that:
v=f ε (H a ,Θ ε ) (3)
wherein f is ε Denotes the encoding procedure, [ theta ] ε A set of parameters representing an encoder;
once the base station BS receives the codeword v, the decoder reconstructs the channel:
whereinWhich is indicative of the decoding process, is,represents a set of parameters of the decoder, so the whole feedback process can be expressed as:
so the final optimization objective is to minimize the original H a And reconstructionThe difference between, expressed as finding the encoder and decoder parameter sets that satisfy the condition:
s2, determining a plurality of network structures;
specifically, according to the operation principle of complex numbers, the complex numbers are regarded as two-dimensional real number pairs, the addition process is the same as real number operation, the complex numbers are expressed as z ═ A + ib, a real component A and an imaginary component b are provided, and the input and the weight of the network are divided into two parts;
when the convolutional neural network works, a convolutional kernel slides from the upper left corner of the CSI matrix, the value of each position in the CSI matrix is traversed according to a certain step length, and convolution operation is carried out at each position. The structure provided by the invention is realized based on a deep complex network, wherein the convolution operation and the real convolution operation are different in that the complex convolution averagely divides a convolution kernel into two parts, the real part and the imaginary part of the input are respectively operated, and the operation flow is as shown in the formula 7. Wherein K I 、K R 、M I 、M R Respectively representing an imaginary part convolution kernel, a real part convolution kernel, an imaginary part characteristic diagram and a real part characteristic diagram;
in the convolutional neural network, the updating of each layer of parameters can cause the data distribution of each layer of input to change, and after multi-layer operation, the input distribution of the high layer can be obviously different from the initial distribution, so that the data distribution required to be adapted by the network in each iteration is always changed, and the difficulty of network training is high. Meanwhile, the change of data distribution can cause the data to fall into the saturation region of the activation function, and the parameter updating of the network is influenced.
The invention normalizes the convolution result obtained by each layer of the convolution neural network to normal distribution with the mean value of 0 and the standard deviation of 1 by adopting a plurality of batches of normalization operations, thereby ensuring the distribution consistency of data in network training. However, the real batch normalization operation is not suitable for complex numbers, and to normalize a group 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 complex batch normalization operation uses two parameters γ and β, the shift parameter β being a complex parameter having two learnable components (real and imaginary parts); the scaling parameter y is a 2 x 2 semi-positive definite matrix with only three degrees of freedom, and therefore only three learnable components. The scaling parameter γ is expressed as:
wherein gamma is rr Scaling parameters that are two real parts; gamma ray ri Is a scaling parameter with the first being a real part and the second being an imaginary part; gamma ray ir Is the scaling parameter with the first as the imaginary part and the second as the real part; gamma ray ii Scaling parameters that are two imaginary parts;
due to normalized values of the CSI matrixWith a real variance and a virtual variance of 1, will be rr And gamma ii Is initialized toTo obtain a normalized value of the modulus of variance 1, gamma ri And gamma ir With an initialization of 0, the plural batch normalization is defined as:
s3, constructing a CSI feedback network based on a complex network structure to obtain an original CSI matrix;
specifically, a complex network structure is introduced into a CSI feedback network, and local connection between neurons in adjacent layers is enhanced by utilizing spatial local correlation; constructing a complex encoder-decoder structure, extracting CSI characteristics in the encoder by using a complex convolution downsampling mode, and compressing a CSI matrix by using a complex full-connection layer; and recovering the original size of the CSI matrix by using a complex full-connection layer in the decoder, and reconstructing the compressed code words into the original CSI matrix through a complex dense connection module CDBlock.
The method considers the characteristics of correlation, sparsity and the like of wireless channels in a large-scale MIMO system, and when a complex dense connection module CDBlock is used, firstly the correlation of a CSI matrix is utilized, and secondly the denseblock is considered to promote the characteristic reuse of the network. As shown in fig. 2, wherein CConv represents a complex convolution structure, in order to ensure maximization of information flow between layers of the network, each layer of CDBlock obtains an additional input feature map from all previous layers and passes the current feature map to all subsequent layers, and the l-th layer receives feature maps of all previous layers:
x l =C l ([x 0 ,x 1 ,...,x l-1 ]) (8)
wherein [ x ] 0 ,x 1 ,...,x l-1 ]Showing the concatenation of signatures generated at level 0, …, l-1, C l For complex batch normalization, an exponential linear unit and a complex convolution three continuously operating complex functions. CDBlock helps train deeper network architectures, and dense connections have a regularization effect, which can reduce overfitting of smaller tasks of the training set.
As shown in fig. 3, the present invention provides 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;
the encoder as shown in fig. 2, wherein S1 × S2 × S3 represents the length, width and number of feature maps, respectively; the encoder comprises a feature extraction module and a compression module; the feature extraction module extracts the CSI matrix features by adopting a complex convolution downsampling mode, and meanwhile, the convergence rate of the model is improved by using complex batch normalization, so that the problem of gradient dispersion in the network is relieved, and the network model is easier to stabilize. The complex convolution downsampling process controls the stepping size, information fusion is good, and the traditional method adopts a pooling layer sampling mode, so that the possibility of filtering useful information exists, and the structure used by the method can better extract data characteristics. The specific setting parameters of the network structure of the feature extraction module may be "threads ═ 2", "padding ═ same", and "kernel _ size ═ 3". The first layer is an input layer, the input data format is 32 × 32 × 2, and when the compression rate is 4, the specification of obtaining a feature map is 2 × 2 × 512 after 4 times of complex convolution downsampling; the compression module compresses the CSI matrix by adopting a plurality of full connection layers, and compresses the 2 multiplied by 512 structure output by the feature extraction module into oneObtaining a compressed code word v; the codeword v is sent to the base station BS over a feedback link.
As shown in fig. 2 for a complex densnet decoder, to maximize the inter-layer information fusion, the densnet directly connects the layers in the network by matching the profiles. The current feature map of each layer passes through all the subsequent layers, and the final output is the concatenation of the feature maps of the previous layer. DenseNet exploits the potential of the network by feature reuse, resulting in an enriched model that is easy to train and parameter efficient. The feature map connecting the different layers increases the variation of the input of the subsequent layers and improves the efficiency, which is simpler and more efficient than before. On the basis of the DenseNet idea, the structure not only utilizes the powerful characteristic reuse capacity of DenseNet, but also combines the characteristics of a complex convolution neural network, and can improve the CSI recovery precision.
After the base station BS obtains the code word v, firstly, the compressed feature vector is directly restored to N through a complex full-connection layer c ×N t The original size of x 2, constituting a rough reconstruction of the CSI matrix; then, the CSI matrix is further refined by a complex dense connection module, and the size of the input characteristic diagram is kept unchanged by zero padding; by using a globally dense connection, information from the first layer can be detected at the last layer, which improves the efficiency of the information flow with a slightly increased computational cost.
The complex dense connection module adopts three convolution layers with 2, 4 and 1 channels as basic reconstruction modules, 5 reconstruction modules continuously refine the reconstruction of the CSI matrix through dense connection and output a characteristic diagram with the shape of 32 multiplied by 12, and the characteristic diagram of 5 reconstruction modules is connected; the reconstructed original CSI matrix is then obtained by convolutional layers of 3 × 3 convolutional kernels with 2 channels.
The foregoing descriptions of specific exemplary embodiments of the present invention have been presented for purposes of illustration and description. It is not intended to limit the invention to the precise form disclosed, and obviously many modifications and variations are possible in light of the above teaching. The exemplary embodiments were chosen and described in order to explain certain principles of the invention and its practical application to enable one skilled in the art to make and use various exemplary embodiments of the invention and various alternatives and modifications as are suited to the particular use contemplated. It is intended that the scope of the invention be defined by the claims and their equivalents.
Claims (10)
1. A channel state information feedback method based on a complex convolutional neural network is characterized by comprising the following steps:
aiming at the channel condition under a frequency division duplex FDD large-scale MIMO system, establishing a communication system model and obtaining a final optimization target of the communication system model;
determining a plurality of network structures;
and constructing a CSI feedback network based on a complex network structure to obtain an original CSI matrix.
2. The method according to claim 1, wherein a communication system model is established for a channel condition under a frequency division duplex, FDD, massive MIMO system, and a final optimization objective is obtained, specifically:
one single cell FDD uses massive MIMO system, there is N at base station BS side t An antenna, wherein N t > 1, there is N at the UE side of the user equipment r An antenna for receiving the signalExpressed as:
y=Αx+z (1)
wherein N is c Is the number of sub-carriers,a symbol representing the transmission is transmitted,additive Gaussian noise; the original output signal A is represented asWherein Downlink channel coefficients and beamforming precoding vectors, each representing a subcarrier i (·) H Representing a conjugate transpose.
3. The method of claim 2, wherein the beamforming precoding vector p is derived from a complex convolutional neural network-based channel state information feedback method i The base station BS needs the user equipment UE to feed back the corresponding channel coefficient h i Therefore, assume the downlink CSI matrix isWhich comprises N c N t The number of the parameters needing to be fed back is 2N c N t Which is proportional to the number of antennas;
because the CSI matrix H is sparse in the angular delay domain, the original form of the space-frequency domain CSI is converted into the angular delay domain by two-dimensional discrete fourier transform, so that:
H′=F c HF t H (2)
wherein F c And F t Respectively is dimension N c ×N c ,N t ×N t For the angular delay domain CSI matrix H', each element corresponds to a particular path delay with an angle of arrival; in H', only the first row contains useful information, the remaining rows represent paths with large propagation delays and consist of values close to zero; let H a The valid information line representing H'.
4. The method of claim 3, wherein the channel state information feedback method based on the complex convolutional neural network,will be valid information row H a Input into an encoder of a user equipment UE, a codeword v is generated according to a given compression ratio η such that:
v=f ε (H a ,Θ ε ) (3)
wherein f is ε Denotes the encoding procedure, [ theta ] ε A set of parameters representing an encoder;
once the code word v is received by the base station BS, the decoder reconstructs the channel:
whereinWhich is indicative of the decoding process, is,represents a set of parameters of the decoder, so the whole feedback process is represented as:
so the final optimization objective is to minimize the original H a And reconstructionThe difference between, expressed as finding the encoder and decoder parameter sets that satisfy the condition:
5. the channel state information feedback method based on the complex convolutional neural network as claimed in claim 1, wherein determining the complex network structure specifically comprises:
when the convolutional neural network works, a convolutional kernel starts to slide from the upper left corner of the CSI matrix, the value of each position in the CSI matrix is traversed according to a certain step length, and convolution operation is carried out at each position;
the complex convolution equally divides the convolution kernel into two parts, and respectively operates the real part and the imaginary part of the CSI matrix, and the operation flow is shown as formula 7, wherein K is I 、K R 、M I 、M R Respectively representing an imaginary part convolution kernel, a real part convolution kernel, an imaginary part characteristic diagram and a real part characteristic diagram;
6. the channel state information feedback method based on the complex convolutional neural network as claimed in claim 5, wherein the convolution results obtained from each layer of the convolutional neural network are normalized to normal distribution with a mean value of 0 and a standard deviation of 1 by using a plurality of batch normalization operations; the complex batch normalization operation uses two parameters γ and β, the shift parameter β is a complex parameter having two learnable components; the scaling parameter γ is a 2 × 2 semi-positive definite matrix with only three degrees of freedom, which is expressed as:
wherein gamma is rr Scaling parameters that are two real parts; gamma ray ri The scaling parameter is the first real part and the second imaginary part; gamma ray ir Is the scaling parameter with the first as the imaginary part and the second as the real part; gamma ray ii Scaling parameters that are two imaginary parts;
due to normalized values of the CSI matrixWith a real variance and a virtual variance of 1, will be rr And gamma ii Is initialized toTo obtain a normalized value of the modulus of variance 1, gamma ri And gamma ir With an initialization of 0, the plural batch normalization is defined as:
7. the channel state information feedback method based on the complex convolutional neural network as claimed in claim 1, wherein a CSI feedback network based on a complex network structure is constructed to obtain an original CSI matrix, specifically: introducing a plurality of network structures into a CSI feedback network, and reinforcing local connection between neurons of adjacent layers by utilizing spatial local correlation; constructing a complex encoder-decoder structure, extracting CSI characteristics in the encoder by using a complex convolution downsampling mode, and compressing a CSI matrix by using a complex full-connection layer; and recovering the original size of the CSI matrix by using a complex full-connection layer in the decoder, and reconstructing the compressed code words into the original CSI matrix through a complex dense connection module CDBlock.
8. The method as claimed in claim 7, wherein each layer of the CDBlock acquires additional feature maps from all previous layers and transfers the feature map of the current layer to all subsequent layers, and then the l-th layer receives the feature map representation of all previous layers as:
x l =C l ([x 0 ,x 1 ,...,x l-1 ]) (8)
wherein [ x ] 0 ,x 1 ,...,x l-1 ]Showing the concatenation of signatures generated at level 0, …, l-1, C l For complex batch normalization, an exponential linear unit and a complex convolution three continuously operating complex functions.
9. The method of claim 7, wherein the encoder comprises a feature extraction module and a compression module; the first layer of the characteristic extraction module is an input layer, the input data format is 32 multiplied by 2, and under the condition that the compression rate is 4, the specification of the characteristic diagram obtained after 4 times of complex convolution downsampling is 2 multiplied by 512; the compression module compresses the CSI matrix by adopting a plurality of full connection layers, and compresses the 2 multiplied by 512 structure output by the feature extraction module into oneObtaining a compressed code word v; the codeword v is sent to the base station BS over a feedback link.
10. The method as claimed in claim 7, wherein 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 the final output of the current feature map of each layer is a concatenation of the feature maps of the previous layer, specifically:
after the base station BS obtains the code word v, firstly, the compressed feature vector is directly restored to N through a complex full-connection layer c ×N t The original size of x 2, constituting a rough reconstruction of the CSI matrix; then, the CSI matrix is further refined by a complex dense connection module, and the size of the input characteristic diagram is kept unchanged by zero padding;
the complex dense connection module adopts three convolution layers with 2, 4 and 1 channels as basic reconstruction modules, 5 reconstruction modules continuously refine the reconstruction of the CSI matrix through dense connection and output a characteristic diagram with the shape of 32 multiplied by 12, and the characteristic diagram of 5 reconstruction modules is connected; the reconstructed original CSI matrix is then obtained by convolutional layers of 3 × 3 convolutional kernels with 2 channels.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210598753.1A CN115021787A (en) | 2022-05-30 | 2022-05-30 | Channel state information feedback method based on complex convolutional neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210598753.1A CN115021787A (en) | 2022-05-30 | 2022-05-30 | Channel state information feedback method based on complex convolutional neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115021787A true CN115021787A (en) | 2022-09-06 |
Family
ID=83070385
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210598753.1A Pending CN115021787A (en) | 2022-05-30 | 2022-05-30 | Channel state information feedback method based on complex convolutional neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115021787A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115664898A (en) * | 2022-10-24 | 2023-01-31 | 四川农业大学 | OFDM system channel estimation method and system based on complex convolution neural network |
-
2022
- 2022-05-30 CN CN202210598753.1A patent/CN115021787A/en active Pending
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115664898A (en) * | 2022-10-24 | 2023-01-31 | 四川农业大学 | OFDM system channel estimation method and system based on complex convolution neural network |
CN115664898B (en) * | 2022-10-24 | 2023-09-08 | 四川农业大学 | OFDM system channel estimation method and system based on complex convolution neural network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112737985B (en) | Large-scale MIMO channel joint estimation and feedback method based on deep learning | |
CN111630787B (en) | MIMO multi-antenna signal transmission and detection technology based on deep learning | |
CN109672464B (en) | FCFNN-based large-scale MIMO channel state information feedback method | |
CN108390706B (en) | Large-scale MIMO channel state information feedback method based on deep learning | |
CN111698182B (en) | Time-frequency blocking sparse channel estimation method based on compressed sensing | |
CN110912598B (en) | Large-scale MIMO system CSI feedback method based on long-time attention mechanism | |
CN111713035B (en) | MIMO multi-antenna signal transmission and detection technology based on artificial intelligence | |
CN113748614A (en) | Channel estimation model training method and device | |
CN111555781B (en) | Large-scale MIMO channel state information compression and reconstruction method based on deep learning attention mechanism | |
CN113872652B (en) | CSI feedback method based on 3D MIMO time-varying system | |
CN112600596B (en) | Millimeter wave system channel feedback method based on tensor parallel compression | |
CN110289898A (en) | A kind of channel feedback method based on the perception of 1 bit compression in extensive mimo system | |
Guo et al. | Deep learning for joint channel estimation and feedback in massive MIMO systems | |
CN115865145A (en) | Large-scale MIMO channel state information feedback method based on Transformer | |
Li et al. | A novel CSI feedback approach for massive MIMO using LSTM-attention CNN | |
CN116192209A (en) | Gradient uploading method for air computing federal learning under MIMO channel | |
CN114157331B (en) | Large-scale MIMO channel state information feedback method based on pseudo complex value input | |
CN114884549A (en) | Large-scale MIMO channel state information feedback method based on deep learning | |
CN115021787A (en) | Channel state information feedback method based on complex convolutional neural network | |
CN114567359A (en) | CSI feedback method based on multi-resolution fusion convolution feedback network in large-scale MIMO system | |
CN115022134B (en) | Millimeter wave large-scale MIMO system channel estimation method and system based on non-iterative reconstruction network | |
CN113556158B (en) | Large-scale MIMO intelligent CSI feedback method for Internet of vehicles | |
An et al. | Channel estimation for one-bit massive MIMO based on improved cGAN | |
CN113556159A (en) | Channel feedback method of large-scale MIMO multi-user system | |
CN113078931A (en) | Millimeter wave large-scale MIMO hybrid precoding method and system based on QGA-OMP algorithm |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |