CN118018082A - CSI feedback reconstruction method and system for RIS-assisted large-scale MIMO system - Google Patents

Abstract

The invention discloses a CSI feedback reconstruction method and system of a RIS auxiliary large-scale MIMO system, wherein the method comprises the following steps: acquiring CSI channel state information fed back by single antenna UE from a UE-RIS-BS path; decomposing the feedback of the CSI channel state information into an index of a non-zero column signal vector and feedback of the non-zero column signal vector; processing the index of the decomposed non-zero column signal vector and the feedback of the non-zero column signal vector to obtain incremental CSI; inputting the increment CSI into a trained feedback model, and outputting the reconstruction CSI; the method is suitable for the RIS-assisted multi-user millimeter wave large-scale MIMO system scene, greatly simplifies the CSI feedback process by considering the sparse characteristic of the millimeter wave propagation environment, reduces a large amount of feedback expenditure, and realizes high-precision reconstruction of the CSI.

Description

CSI feedback reconstruction method and system for RIS-assisted large-scale MIMO system

Technical Field

The invention relates to a CSI feedback reconstruction method and system for a RIS auxiliary large-scale MIMO system, and belongs to the technical field of communication.

Background

Large-scale input multiple output (MIMO) technology is one of the key technologies of the fifth generation communication system, and deploying a large number of antennas at a Base Station (BS) can significantly improve the performance of the communication system. With the advancement of technology, reconfigurable intelligent reflection surface (RIS) technology is gradually becoming a promising new technology for future 6G communications. In addition, RIS-assisted massive MIMO communication systems improve Spectral Efficiency (SE) and Energy Efficiency (EE) at lower cost by employing a large number of passive reflective elements.

The performance gain of massive MIMO systems depends on the base station obtaining accurate Channel State Information (CSI). In a Time Division Duplex (TDD) system, since uplink and downlink channels have reciprocity, the downlink channel CSI can be easily obtained through uplink channel estimation. In the research of the existing RIS-assisted communication system, the TDD mode is considered. However, TDD mode requires very accurate time synchronization between the two transmitting and receiving ends, otherwise serious interference is generated, and Frequency Division Duplex (FDD) technology realizes stronger anti-interference by setting a guard band between uplink and downlink, which is a duplex mode widely adopted in the current mobile communication network. However, in FDD mode, uplink and downlink channel reciprocity is not established, and the BS needs to perform feedback by a User Equipment (UE) to obtain CSI.

In an RIS-assisted massive MIMO system in FDD mode, the main obstacle is that the feedback overhead increases with the number of BS antennas and the number of RIS reflection units. Therefore, it is required to reduce feedback overhead while achieving high-precision CSI reconstruction at the BS side.

The compressed sensing-based algorithm and the deep learning-based algorithm proposed by the prior art do not balance feedback overhead and reconstruction accuracy well.

Disclosure of Invention

The invention aims to provide a CSI feedback reconstruction method and system for an RIS auxiliary large-scale MIMO system, which can reduce feedback overhead and realize high-precision CSI reconstruction at a BS end at the same time so as to solve the problem that the feedback overhead in the prior art is increased along with the increase of the number of BS antennas and RIS reflection units.

A CSI feedback reconstruction method for an RIS-aided massive MIMO system, the method comprising:

acquiring CSI channel state information fed back by single antenna UE from a UE-RIS-BS path;

Decomposing the feedback of the CSI channel state information into an index of a non-zero column signal vector and feedback of the non-zero column signal vector;

processing the index of the decomposed non-zero column signal vector and the feedback of the non-zero column signal vector to obtain incremental CSI channel state information;

inputting the incremental CSI channel state information into a trained feedback model, and outputting the reconstructed CSI channel state information;

the feedback model trains a pair of dimension reduction and reconstruction dictionary groups by utilizing pre-stored CSI channel state information at a Base Station (BS), optimizes the dimension reduction and reconstruction dictionary groups in an alternate iteration mode to obtain two groups of optimal dictionaries, and outputs the reconstructed CSI channel state information by utilizing the two groups of optimal dictionaries.

Further, the obtaining CSI channel state information fed back by the single antenna UE from the UE-RIS-BS path includes:

The signal received by the single antenna UE is expressed as:

(1)

Wherein the method comprises the steps of Is/>Received signals of the individual UEs; /(I)Representing from RIS to/>RIS-UE channels of the individual UEs; is a transmission signal precoded at the BS; /(I) A BS-RIS channel matrix representing BS-to-RIS; /(I)For the total number of RIS elements,/>Total number of antennas for BS; /(I)For/>Additive white gaussian noise at individual UEs,/>A diagonal matrix representing the phase shift values of the RIS reflective elements, anWherein/> (/>) Index for RIS element,/>And/>Respectively represent the/>Amplitude and phase coefficients of the individual reflective elements; /(I)Is an imaginary unit;

Due to Is a diagonal matrix and therefore has/>First/>Equivalent downlink channel/>, for individual UEsExpressed as:

(2)

Thus the third party The BS-RIS-UE cascade channels of the individual UEs are denoted as:

(3)

Wherein the method comprises the steps of Depending on downlink CSI only, if direct feedback/>User feedback/>And the channel fading coefficients.

Further, the feedback model is modeled using a Saleh Valenzuela geometric channel model, in which the BS-RIS channel is expressed as:

(4)

Wherein the method comprises the steps of Is normalized by factor one,/>For the number of paths between BS and RIS,/>，Is/>Complex gain of strip path,/>、/>Representing array response vectors associated with RIS and base station BS, respectively,/>、/>、/>The angle of arrival is the azimuth angle of the RIS, the angle of arrival is the elevation angle of the RIS's pitch angle, and the azimuth angle of the BS's departure angle, respectively.

Further, the array response vector expression related to the base station BS is:

(5)

(6)

In the middle of 、/>Respectively the spacing of the RIS element from the antennas of the base station BS,/>Is the signal wavelength,/>、/>The number of RIS elements in the vertical and horizontal directions,/>, respectively、/>(/>，/>) Representing the horizontal and vertical indices, respectively,/>, of RIS elements，/>（/>) Is the antenna index of the BS.

Further, the firstRIS and/>, in BS-RIS-UE cascade channels of individual UEsChannel vector between individual UEs/>Expressed as:

(7)

Wherein the method comprises the steps of For RIS and/>Number of paths between UEs,/>For the normalization factor of two,For/>Complex gain of individual paths,/>、/>The azimuth angle of the RIS and the elevation angle of the RIS, respectively; at/>The array response vector at (a) is represented by formula (5);

according to the formulas (3), (4) and (7), in the first place Concatenated channel matrix/>, for individual UEsCan be expressed as:

(8)

Wherein the method comprises the steps of For/>The/>, received by the individual userComplex gains for the individual paths;

Will be Is marked as/>And let/>Then equation (8) reduces to:

(9)。

Further, the decomposing the feedback of the CSI channel state information into the index of the non-zero column signal vector and the feedback of the non-zero column signal vector includes:

Acquiring complete downlink CSI channel state information of each UE through downlink channel estimation;

cascading channel matrices using angular domain limited channel characteristics in millimeter wave propagation environments Conversion to a mixed domain concatenated channel matrix/>The conversion is expressed as:

(10)

Wherein the method comprises the steps of For the angular resolution size of AoD at the base station BS,/>Is a dictionary matrix of angular resolution G of AoD at base station BS, and the AoD at base station BS is quantized to/>, using the dictionary matrixThe grid, dictionary matrix, is shown below:

(11)

Wherein the method comprises the steps of Represents the angle of quantified AoD, and/>; Wherein/>The discrete angle values of the grids are dictionary matrix/>(1 /)The angle of the array response vector in a column is denoted/>Thus/>Is generated according to formula (6);

Is known to be Is a BS-RIS channel matrix shared by a plurality of UEs from a base station BS to a RIS, and is obtained by angular domain transformationNon-zero columns with a small number of dominant channel gains, and the non-zero columns are the path number/>，/>/>The non-zero columns are represented as:

(12)

Wherein the method comprises the steps of ，/>，/>And/>For/>An index of a non-zero column in (b).

Further, the training method of the feedback model comprises the following steps:

Acquiring a sample data set, and performing sampling processing through a non-zero column signal vector of high-dimensional data to obtain a training set;

Constructing an initial model;

training an initial model by adopting a training set, wherein the initial model comprises an initial dimension reduction dictionary set and a reconstruction dictionary set;

The initial dimension reduction dictionary set and the reconstruction dictionary set are respectively subjected to alternating optimization processing by keeping the neighbor relation and the coding relation between sample data sets unchanged, so that an optimal dictionary and an optimal coding vector are obtained, and the optimal dimension reduction dictionary set and the reconstruction dictionary set are obtained;

The optimal reconstruction dictionary set is reserved at the base station end, and the optimal dimension reduction dictionary set is shared to the user end, so that a trained feedback model is obtained.

Further, the training set includes pre-stored CSI channel state information that is first partially transmitted to the BS side of the base station by the decomposed CSI channel state information.

Further, the inputting the incremental CSI channel state information into the trained feedback model, outputting the reconstructed CSI channel state information, includes:

Inputting the increment non-zero column signal vector into a trained feedback model, performing dimension reduction processing on a user side through an optimal dimension reduction dictionary set, obtaining increment low-dimension embedding, and feeding back to a base station side;

After receiving the low-dimensional embedding, the base station reconstructs the low-dimensional embedding through a prestored reconstruction dictionary set to obtain a reconstructed increment non-zero column signal vector;

and outputting the reconstructed downlink CSI from the feedback model by combining the known reconstructed incremental non-zero column vector and the index of the incremental non-zero column vector.

A CSI feedback reconstruction system of a RIS-assisted massive MIMO system, the system comprising:

The acquisition data module is used for acquiring the CSI channel state information fed back by the single-antenna UE from the UE-RIS-BS path;

The decomposition module is used for decomposing the feedback of the CSI channel state information into indexes of non-zero column signal vectors and feedback of the non-zero column signal vectors;

the preprocessing module is used for processing the index of the decomposed non-zero column signal vector and the feedback of the non-zero column signal vector to obtain incremental CSI channel state information;

The data processing module is used for inputting the incremental CSI channel state information into the trained feedback model and outputting the reconstructed CSI channel state information;

the feedback model trains a pair of dimension reduction and reconstruction dictionary groups by utilizing pre-stored CSI channel state information at a Base Station (BS), optimizes the dimension reduction and reconstruction dictionary groups in an alternate iteration mode to obtain two groups of optimal dictionaries, and outputs the reconstructed CSI channel state information by utilizing the two groups of optimal dictionaries.

Compared with the prior art, the invention has the beneficial effects that:

1. The invention is suitable for the RIS-assisted multi-user millimeter wave large-scale MIMO system scene, greatly simplifies the feedback process of the CSI channel state information by considering the sparse characteristic of the millimeter wave propagation environment, and reduces a large amount of feedback expenditure;

2. The invention utilizes manifold learning thought to build CSI feedback model based on manifold learning, trains a pair of dimension-reducing and reconstruction dictionary groups, further reduces feedback expenditure, and ensures reconstruction performance when reducing feedback dimension.

Drawings

FIG. 1 is a flow chart of an implementation of the present invention;

FIG. 2 is a diagram of a RIS assisted multi-user MIMO system of the present invention;

Fig. 3 is a diagram of a CSI feedback framework of the present invention;

FIG. 4 is a diagram illustrating the relationship between normalized mean square error and data set size according to an embodiment of the present invention;

FIG. 5 is a graph of achievable sum rate versus signal-to-noise ratio for different compression ratios according to one embodiment of the invention;

FIG. 6 is a schematic diagram of the comparison of the achievable sum rates of different schemes of one embodiment of the invention;

Fig. 7 is a schematic diagram of the comparison of the achievable sum rates of different schemes of another embodiment of the invention.

Detailed Description

The invention is further described in connection with the following detailed description, in order to make the technical means, the creation characteristics, the achievement of the purpose and the effect of the invention easy to understand.

As shown in fig. 1 to 7, the present invention discloses a CSI feedback reconstruction method for an RIS-aided massive MIMO system, and applies the idea of manifold learning to CSI channel state information feedback in an attempt to recover a low-dimensional manifold from a high-dimensional space while maintaining a data inherent manifold structure, thereby reducing the dimension of high-dimensional data. Considering that in a RIS-assisted large-scale MIMO system in an FDD mode, CSI channel state information feedback is taken as a research object, and a dimension-reduction CSI feedback scheme based on a manifold learning framework is provided aiming at the design target of reducing feedback overhead and guaranteeing feedback precision. Firstly, considering the sparse characteristic of the millimeter wave propagation environment, converting downlink incremental CSI into a mixed domain (angular domain-space domain) representation, so that the feedback of the CSI channel state information is decomposed into the feedback of a non-zero column index and a non-zero column channel vector, and the preliminary reduction of feedback overhead is realized; then, a manifold learning idea is utilized to establish a feedback frame (Manifold Learning-based CSI Feedback Framework, MLF) based on manifold learning, a pair of dimension reduction and reconstruction dictionary groups are trained, the dimension reduction is carried out on the incremental CSI channel state information by utilizing a dimension reduction dictionary at a user end in a manner of reserving a local manifold structure, and the low-dimension to high-dimension data reconstruction is carried out at a base station end by utilizing a reconstruction dictionary, so that the further reduction of feedback expenditure is realized; and finally, optimizing the dimension reduction and reconstruction dictionary groups in an alternate iteration mode to obtain two groups of optimal dictionaries, thereby realizing the reconstruction of the CSI channel state information with higher precision.

As shown in fig. 1, a CSI feedback reconstruction method of an RIS-assisted massive MIMO system is disclosed, and a system involved in the method is shown in fig. 2, where the method includes:

step 1, obtaining CSI channel state information fed back by single antenna UE from a UE-RIS-BS path;

Step 2, decomposing the feedback of the CSI channel state information into indexes of non-zero column signal vectors and feedback of the non-zero column signal vectors;

Step 3, processing the index of the decomposed non-zero column signal vector and the feedback of the non-zero column signal vector to obtain incremental CSI channel state information;

Step 4, inputting the incremental CSI channel state information into a trained feedback model, and outputting the reconstructed CSI channel state information;

The feedback model trains a pair of dimension reduction and reconstruction dictionary groups by utilizing pre-stored CSI channel state information at a Base Station (BS), optimizes the dimension reduction and reconstruction dictionary groups in an alternate iteration mode to obtain two groups of optimal dictionaries, and outputs the reconstructed CSI channel state information by utilizing the two groups of optimal dictionaries.

Step 1, obtaining CSI channel state information fed back by single antenna UE from UE-RIS-BS path includes:

The signal received by the single antenna UE is expressed as:

(1)

Wherein the method comprises the steps of Is/>Received signals of the individual UEs; /(I)Representing from RIS to/>RIS-UE channels of the individual UEs; is a transmission signal precoded at the BS; /(I) A BS-RIS channel matrix representing BS-to-RIS; /(I)For the total number of RIS elements,/>Total number of antennas for BS; /(I)For/>Additive white gaussian noise at individual UEs,/>A diagonal matrix representing the phase shift values of the RIS reflective elements, anWherein/> (/>) Index for RIS element,/>And/>Respectively represent the/>Amplitude and phase coefficients of the individual reflective elements; /(I)Is an imaginary unit;

Due to Is a diagonal matrix and therefore has/>First/>Equivalent downlink channel/>, for individual UEsExpressed as:

(2)

Thus the third party The BS-RIS-UE cascade channels of the individual UEs are denoted as:

(3)

Wherein the method comprises the steps of Depending on downlink CSI only, if direct feedback/>User feedback/>And the channel fading coefficients.

The feedback model is modeled using a Saleh Valenzuela geometric channel model in which the BS-RIS channelCan be expressed as:

(4)

Wherein the method comprises the steps of Is normalized by factor one,/>For the number of paths between BS and RIS,/>，Is/>Complex gain of strip path,/>、/>Representing array response vectors associated with RIS and base station BS, respectively,/>、/>、/>An arrival angle AoA of azimuth of the RIS, an arrival angle AoA of pitch angle of the RIS, and an departure angle AoD of azimuth of the BS, respectively.

The BS-related array response vector expression is:

(5)

(6)

In the middle of 、/>Respectively the spacing of the RIS element from the antennas of the base station BS,/>Is the signal wavelength,/>、/>The number of RIS elements in the vertical and horizontal directions,/>, respectively、/>(/>，/>) Representing the horizontal and vertical indices, respectively,/>, of RIS elements，/>（/>) Is the antenna index of the BS.

First, theRIS and/>, in BS-RIS-UE cascade channels of individual UEsChannel vector between individual UEs/>Expressed as:

(7)

Wherein the method comprises the steps of For RIS and/>Number of main paths between individual UEs,/>For the normalization factor of two,For/>Complex gain of individual paths,/>、/>An exit angle AoD of azimuth of the RIS and an exit angle AoD of pitch of the RIS, respectively; at/>The array response vector at (a) is represented by formula (5);

according to the formulas (3), (4) and (7), in the first place Concatenated channel matrix/>, for individual UEsCan be expressed as:

(8)

Wherein the method comprises the steps of For/>The/>, received by the individual userComplex gains for the individual paths;

Will be Is marked as/>And let/>Then equation (8) reduces to:

(9)。

step 2, decomposing the feedback of the CSI channel state information into an index of a non-zero column signal vector and a feedback of the non-zero column signal vector, including:

Acquiring complete downlink CSI channel state information of each UE through downlink channel estimation;

cascading channel matrices using angular domain limited channel characteristics in millimeter wave propagation environments Conversion to a mixed domain concatenated channel matrix/>The conversion is expressed as:

(10)

Wherein the method comprises the steps of For the angular resolution size of AoD at the base station BS,/>Is a dictionary matrix of angular resolution G of AoD at base station BS, and the AoD at base station BS is quantized to/>, using the dictionary matrixThe grid, dictionary matrix, is shown below:

(11)

Wherein the method comprises the steps of Represents the angle of quantified AoD, and/>; Wherein/>The discrete angle values of the grids are dictionary matrix/>(1 /)The angle of the array response vector in a column is denoted/>Thus/>Is generated according to formula (6);

Is known to be Is a BS-RIS channel matrix shared by a plurality of UEs from a base station BS to a RIS, and is obtained by angular domain transformationNon-zero columns with a small number of dominant channel gains, and the non-zero columns are the path number/>， />/>The non-zero columns are represented as:

(12)

Wherein the method comprises the steps of ，/>，/>And/>For/>An index of a non-zero column in (b).

The method for training the optimal feedback model comprises the following steps:

Acquiring a sample data set, and performing sampling processing through a non-zero column signal vector of high-dimensional data to obtain a training set;

Constructing an initial model;

training an initial model by adopting a training set, wherein the initial model comprises an initial dimension reduction dictionary set and a reconstruction dictionary set;

The initial dimension reduction dictionary set and the reconstruction dictionary set are respectively subjected to alternating optimization processing by keeping the neighbor relation and the coding relation between sample data sets unchanged, so that an optimal dictionary and an optimal coding vector are obtained, and the optimal dimension reduction dictionary set and the reconstruction dictionary set are obtained;

the optimal reconstruction dictionary set is reserved at the base station end, the optimal dimension reduction dictionary set is shared to the user end, a trained feedback model is obtained, and aiming at the training method, in the embodiment, the following is further expressed:

The method comprises the steps of obtaining a sample data set, firstly, naming a high-dimensional space where an original channel matrix is sampled as an 'input space', naming a low-dimensional space where the original channel matrix is embedded as a 'characteristic space', and obtaining from downlink pilot frequency />Samples, will/>The samples are trained as high-dimensional training data by equation (13):

(13)

Wherein the method comprises the steps of For/>(1 /)Subsampling, from/>The channel matrix of the dimensional input spatial samples contains a dataset/>The manifold structure of the input space is characterized;

local cut space alignment (LTSA) calculation using manifold learning algorithm Is embedded as follows:

(14)

Wherein the method comprises the steps of Is formed by/>Sample composition in dimensional feature space,/>Is/>Embedding of dimension/>Much smaller than/>Compression ratio is defined as/>；

Searching data setsTo/>Mapping relation/>Introducing a manifold structure in which dictionary description high-dimensional non-zero column vector sampling or low-dimensional embedding is located, and setting/>Is a high-dimensional dictionary that characterizes the input spatial manifold structure, where/>Is the size of the dictionary,/>Is/>(1 /)A column;

Will be The approximation is:

(15)

Wherein, Needs to meet/>Let/>Is/>At/>Middle/>A set of nearest neighbors if/>Then/>Dictionary/>Upper/>Is expressed as the code vector of (a)，/>Is defined as the coding matrix of (a)；

Based on the characteristic spaceAnd a low-dimensional dictionary/>The linear approximation between them holds, yielding:

(16)

Wherein, And/>The set of nearest neighbors is the same as in equation (16);

And obtaining the minimized cost function of the coding vector and the dictionary by keeping the neighbor relation and the coding relation of the input space and the feature space unchanged:

(17a)

(17b)

If/> (17c)

Wherein the method comprises the steps ofIs a constant for adjusting the proportion of the last term in the cost function.

Mapping relation is searched based on reconstruction increment in low-dimensional embeddingThe low-dimensional dictionary in feature space is set to/>The corresponding high-dimensional dictionary in the input space isDictionary/>Upper/>Is expressed as the code vector of (a)，/>Is defined as the coding matrix of (a)By keeping the neighbor relation and the coding relation, the objective function is minimized as follows: /(I) (18)

Acquiring an optimal dictionaryAnd optimal coding vector/>。

And combining dimension reduction and reconstruction of the dictionary set, and outputting a trained optimal feedback model.

And 4, solving a feedback model, wherein the specific solving process is as follows:

The first step: assume that Is initialized or updated in the last iteration, i.e. at this point/>Fixing. Measurement of/> using Euclidean distanceAnd/>Similarity between, i.e./>Select/>/>Nearest neighbors, i.e. slave/>Selected/>Personal and/>The index of the corresponding element forms an index vector/>. All neighbor columns form a matrix。

And a second step of: the vector of non-zero entries is noted asWherein/>Equal to/>，/>. In formula (18), use/>Substitution/>Fixing/>Introducing Lagrangian multiplier method, solving the problem in the formula (18) to obtain coding vector/>：

(19)

Wherein the constraint is thatAnd/>，，/>For/>Diagonal elements of (a) are included. Repeating equation (19)/>Next time, useUpdating coding matrix/>(1 /)Columns, the remaining elements are zero.

And a third step of: by fixing the coding matrixBy the formula (18) pair/>When the first derivative of (2) is 0, there is an extremum. Obtaining a high-dimensional dictionary/>First/>Solution of columns/>(20)

Wherein the method comprises the steps ofFor/>(1 /)Line/>Representation/>A matrix of squares of each element of (a) a matrix of elements. Repeating the calculation formula (20)/>Second, update the entire dictionary/>。

Fourth step: by alternately optimizing the coding matrixAnd dictionary/>Gradually converging the cost function to obtain the optimal coding matrix/>After that, the cost function can be calculated by minimizing the following cost function,(21)

Obtaining a low-dimensional dictionaryThe least squares solution of formula (21) is/>. BS is known to acquire a dimensionality-reduced high-dimensional dictionary/>And a low-dimensional dictionary/>And sharing the data to the UE side for calculating the embedding of the increment.

Fifth step: similar to the process of obtaining the dimension-reduced dictionary set, the low-dimension dictionary is obtained through alternate optimizationAnd coding matrix/>Obtaining the solution of the high-dimensional dictionary as/>, by a least square method. To this end, the dictionary set/>, is reconstructedAndPre-stored at the BS side is known for reconstructing high-dimensional increments.

Sixth step: a solution of the feedback model is obtained.

Step 4, inputting the incremental CSI channel state information into a trained feedback model, and outputting the reconstructed CSI channel state information, wherein the method comprises the following steps:

acquiring incremental CSI channel state information, and decomposing feedback of the incremental CSI channel state information into indexes of non-zero column signal vectors and feedback of the non-zero column signal vectors;

Inputting the increment non-zero column signal vector into a trained feedback model, performing dimension reduction processing on a user side through an optimal dimension reduction dictionary set, obtaining increment low-dimension embedding, and feeding back to a base station side;

After receiving the low-dimensional embedding, the base station reconstructs the low-dimensional embedding through a prestored reconstruction dictionary set to obtain a reconstructed increment non-zero column signal vector;

and outputting the reconstructed downlink CSI from the feedback model by combining the known reconstructed incremental non-zero column vector and the index of the incremental non-zero column vector.

For step 4, specifically set forth:

The first step: taking a single user as an example, use An incremental channel matrix representing the user, and an angular domain concatenated channel matrix/>, obtained using (10)The set of index and non-zero column vectors is found as matrix/>Is thatWherein/>Is/>Is a non-zero column in (b).

And a second step of: will beInputting into a feedback model, the UE adopts a prestored dimension reduction dictionary group, namely/>, and the method comprises the following steps ofAnd/>Substituting formula (17 a) and setting/>In dictionary/>The above coding vector is/>，/>The coding matrix is defined as/>. To keep/>The input space and the feature space have the same coding and neighbor relation principle, and the dimension reduction is realized by solving the following optimization problems:

(22a)

(22b)

If/> (22c)

Wherein the method comprises the steps ofIs a constant for adjusting the proportion of the last term in the cost function. The optimization problem is similar to equation (17 a), except that at this time the high-dimensional dictionary/>It is known that no further alternating optimization of the two variables is necessary. By means of a known dictionary/>Directly solve for/>, using formula (19)(1 /)Column/>。

And a third step of: acquisition ofThereafter, the/>, is further calculated using the matrix form of formula (16)Is embedded as a low dimension ofAnd then fed back to the BS by the UE.

Fourth step: after the BS receives the low-dimensional embedding, a prestored reconstruction dictionary is utilized、/>The low-dimensional embedding is substituted into formula (18), the following minimization objective function is solved to obtain the encoded vector,

(23)

Wherein the method comprises the steps ofIs the coding matrix/>(1 /)Column/>Is/>(1 /)Columns. Based on dictionary/>Coding matrixBy/>Acquisition of the reconstruction/>。

Fifth step: combining the known non-zero column vector and the non-zero column index, the method is performed by the formulaAnd outputting the reconstructed downlink CSI channel state information from the feedback model.

The invention is suitable for the RIS-assisted multi-user millimeter wave large-scale MIMO system scene, greatly simplifies the CSI feedback process and reduces a large amount of feedback expenditure by considering the sparse characteristic of the millimeter wave propagation environment; the invention utilizes manifold learning thought to build CSI feedback frame based on manifold learning, trains a pair of dimension-reducing and reconstruction dictionary groups, further reduces feedback expenditure, and ensures reconstruction performance when reducing feedback dimension.

As shown in fig. 3, the feedback process of the CSI feedback model in the method of the present invention needs to be performed only once in a longer angular coherence time; step two, only need to be executed once in the whole communication process; step three is executed once in the channel coherence time. Step one needs to be performed again within a different angular coherence time.

As shown in fig. 4, the relationship of mean square error and data set size is normalized at different compression ratios. FIG. 4 shows the sizes of training data sets with compression ratios of 1/32, 1/16 and 1/8Effect on reconstruction Performance, describes/>And reconstructing the relationship between NMSEs. For a certain compression ratio, with/>The value increases and the NMSE becomes smaller and eventually tends to converge. When/>After 4000 is reached, the reconstructed NMSE gradually converges, and particularly, the convergence effect is best when the compression ratio is 1/8.

As shown in fig. 5, the achievable sum rate of the present invention is related to the signal-to-noise ratio at different compression ratios. Fig. 5 shows the achievable sum rate obtained by the CSI feedback scheme proposed by the present invention at different compression ratios, the "perfect CSI" curve being the upper limit for the achievable sum rate obtained by the time-consuming system. It can be observed that the achievable and rate performance approaches ideal when the compression ratio is 1/8, indicating that the error between the reconstructed channel and the original channel is small. However, as the compression ratio decreases, performance decreases.

As shown in fig. 6, RIS of different schemes assists the reachability and rate of MU-MIMO systems. Fig. 6 compares the reachability and rate performance of the present invention, compressed sensing scheme, deep learning scheme, csiNet scheme and codebook scheme at different signal to noise ratios, and it is apparent that the present invention is superior to the rest of the schemes. In a high signal-to-noise ratio scene, the inter-user interference is a main influence relative to the system noise, and in order to obtain good performance, larger expenditure is often caused, but the invention does not need to perform dictionary learning for many times for a long period of time, the expenditure is obviously lower than that of other schemes, and the feedback expenditure and the performance of the system are well balanced.

As shown in fig. 7, the achievable sum rate of different schemes varies with the number of UEs. Fig. 7 shows that the present invention has better performance than the compressed sensing scheme, the deep learning scheme, the CsiNet scheme, and the codebook scheme. As can be seen from fig. 7, in the perfect CSI environment and the imperfect CSI environment, the achievable and rate tends to decrease when the number of UEs is greater than 4. This phenomenon stems from the low rank nature of the channel matrix, which is a common characteristic of massive MIMO channels in a limited scattering environment, particularly in millimeter wave communication systems.

The invention also discloses a CSI feedback reconstruction system of the RIS auxiliary large-scale MIMO system, which comprises:

The acquisition data module is used for acquiring the CSI channel state information fed back by the single-antenna UE from the UE-RIS-BS path;

The decomposition module is used for decomposing the feedback of the CSI channel state information into indexes of non-zero column signal vectors and feedback of the non-zero column signal vectors;

the preprocessing module is used for processing the index of the decomposed non-zero column signal vector and the feedback of the non-zero column signal vector to obtain incremental CSI channel state information;

The data processing module is used for inputting the incremental CSI channel state information into the trained feedback model and outputting the reconstructed CSI channel state information;

The feedback model comprises training a pair of dimension reduction and reconstruction dictionary groups by utilizing pre-stored CSI channel state information at a Base Station (BS), optimizing the dimension reduction and reconstruction dictionary groups in an alternate iteration mode to obtain two groups of optimal dictionaries, and outputting the reconstructed CSI channel state information by utilizing the two groups of optimal dictionaries.

The invention utilizes manifold learning thought to build CSI feedback frame based on manifold learning, trains a pair of dimension-reducing and reconstruction dictionary groups, further reduces feedback expenditure, and ensures reconstruction performance when reducing feedback dimension.

The foregoing is merely a preferred embodiment of the present invention, and it should be noted that modifications and variations could be made by those skilled in the art without departing from the technical principles of the present invention, and such modifications and variations should also be regarded as being within the scope of the invention.

Claims (10)

1. A CSI feedback reconstruction method for an RIS-assisted massive MIMO system, the method comprising:

acquiring CSI channel state information fed back by single antenna UE from a UE-RIS-BS path;

Decomposing the feedback of the CSI channel state information into an index of a non-zero column signal vector and feedback of the non-zero column signal vector;

processing the index of the decomposed non-zero column signal vector and the feedback of the non-zero column signal vector to obtain incremental CSI channel state information;

inputting the incremental CSI channel state information into a trained feedback model, and outputting the reconstructed CSI channel state information;

the feedback model trains a pair of dimension reduction and reconstruction dictionary groups by utilizing pre-stored CSI channel state information at a Base Station (BS), optimizes the dimension reduction and reconstruction dictionary groups in an alternate iteration mode to obtain two groups of optimal dictionaries, and outputs the reconstructed CSI channel state information by utilizing the two groups of optimal dictionaries.

2. The method for reconstructing CSI feedback of an RIS-assisted massive MIMO system according to claim 1, wherein said obtaining CSI channel state information fed back by a single antenna UE from a UE-RIS-BS path comprises:

The signal received by the single antenna UE is expressed as:

(1)

Wherein the method comprises the steps of Is/>Received signals of the individual UEs; /(I)Representing from RIS to/>RIS-UE channels of the individual UEs; /(I)Is a transmission signal precoded at the BS; /(I)A BS-RIS channel matrix representing BS-to-RIS; /(I)For the total number of RIS elements,/>Total number of antennas for BS; /(I)For/>Additive white gaussian noise at individual UEs,A diagonal matrix representing the phase shift values of the RIS reflective elements, anWherein/> (/>) Index for RIS element,/>And/>Respectively represent the/>Amplitude and phase coefficients of the individual reflective elements; /(I)Is an imaginary unit;

Due to Is a diagonal matrix and therefore has/>First/>Equivalent downlink channel of individual UEsExpressed as:

(2)

Thus the third party The BS-RIS-UE cascade channels of the individual UEs are denoted as:

(3)

Wherein the method comprises the steps of Depending on downlink CSI only, if direct feedback/>User feedback/>And the channel fading coefficients.

3. The method of claim 2, wherein the feedback model is modeled using a Saleh Valenzuela geometric channel model, and wherein the BS-RIS channel is expressed as:

(4)

Wherein the method comprises the steps of Is normalized by factor one,/>For the number of paths between BS and RIS,/>，Is/>Complex gain of strip path,/>、/>Representing array response vectors associated with RIS and base station BS, respectively,/>、/>、/>The angle of arrival of the azimuth of the RIS, the angle of arrival of the elevation of the RIS, and the angle of departure of the azimuth of the BS, respectively.

4. The method for reconstructing CSI feedback of an RIS-assisted massive MIMO system according to claim 3, wherein the array response vector expression related to the base station BS is:

(5)

(6)

In the middle of 、/>Respectively the spacing of the RIS element from the antennas of the base station BS,/>Is the signal wavelength,/>、/>The number of RIS elements in the vertical and horizontal directions,/>, respectively、/>(/>，/>) Representing the horizontal and vertical indices, respectively,/>, of RIS elements，/>（/>) Is the antenna index of the BS.

5. The method for CSI feedback reconstruction of an RIS-aided massive MIMO system of claim 4, wherein the thRIS and/>, in BS-RIS-UE cascade channels of individual UEsChannel vector between individual UEs/>Expressed as:

(7)

Wherein the method comprises the steps of For RIS and/>Number of paths between UEs,/>Is normalized by a factor of two,/>For/>Complex gain of individual paths,/>、/>The azimuth angle of the RIS and the elevation angle of the RIS, respectively; at/>The array response vector at (a) is represented by formula (5);

according to the formulas (3), (4) and (7), in the first place Concatenated channel matrix/>, for individual UEsCan be expressed as:

(8)

Wherein the method comprises the steps of For/>The/>, received by the individual userComplex gains for the individual paths;

Will be Is marked as/>And let/>Then equation (8) reduces to:

(9)。

6. The method for reconstructing CSI feedback of an RIS-aided massive MIMO system according to claim 4, wherein said decomposing the feedback of CSI channel state information into an index of non-zero column signal vectors and a feedback of non-zero column signal vectors comprises:

Acquiring complete downlink CSI channel state information of each UE through downlink channel estimation;

cascading channel matrices using angular domain limited channel characteristics in millimeter wave propagation environments Conversion to a mixed domain concatenated channel matrix/>The conversion is expressed as:

(10)

Wherein the method comprises the steps of For the angular resolution size of AoD at the base station BS,/>Is a dictionary matrix of angular resolution G of AoD at base station BS, and the AoD at base station BS is quantized to/>, using the dictionary matrixThe grid, dictionary matrix, is shown below:

(11)

Wherein the method comprises the steps of Represents the angle of quantified AoD, and/>; Wherein/>The discrete angle values of the grids are dictionary matrix(1 /)The angle of the array response vector in a column is denoted/>Thus/>Is generated according to formula (6);

Is known to be Is a BS-RIS channel matrix shared by a plurality of UEs from a base station BS to a RIS, obtained by angular domain transformation/>Non-zero columns with a small number of dominant channel gains, and the non-zero columns are the path number/>，/>/>The non-zero columns are represented as:

(12)

Wherein the method comprises the steps of ，/>，/>And/>For/>An index of a non-zero column in (b).

7. The method for reconstructing CSI feedback for an RIS-aided massive MIMO system according to claim 1, wherein said training method for the feedback model comprises:

Acquiring a sample data set, and performing sampling processing through a non-zero column signal vector of high-dimensional data to obtain a training set;

Constructing an initial model;

training an initial model by adopting a training set, wherein the initial model comprises an initial dimension reduction dictionary set and a reconstruction dictionary set;

The initial dimension reduction dictionary set and the reconstruction dictionary set are respectively subjected to alternating optimization processing by keeping the neighbor relation and the coding relation between sample data sets unchanged, so that an optimal dictionary and an optimal coding vector are obtained, and the optimal dimension reduction dictionary set and the reconstruction dictionary set are obtained;

The optimal reconstruction dictionary set is reserved at the base station end, and the optimal dimension reduction dictionary set is shared to the user end, so that a trained feedback model is obtained.

8. The method for reconstructing CSI feedback of an RIS-aided massive MIMO system according to claim 7, wherein the training set includes pre-stored CSI channel state information that is first partially transmitted to a BS side of the base station by the decomposed CSI channel state information.

9. The method for reconstructing CSI feedback of an RIS-aided massive MIMO system according to claim 7, wherein said inputting incremental CSI channel state information into a trained feedback model and outputting reconstructed CSI channel state information comprises:

Inputting the increment non-zero column signal vector into a trained feedback model, performing dimension reduction processing on a user side through an optimal dimension reduction dictionary set, obtaining increment low-dimension embedding, and feeding back to a base station side;

After receiving the low-dimensional embedding, the base station reconstructs the low-dimensional embedding through a prestored reconstruction dictionary set to obtain a reconstructed increment non-zero column signal vector;

and outputting the reconstructed downlink CSI from the feedback model by combining the known reconstructed incremental non-zero column vector and the index of the incremental non-zero column vector.

10. A CSI feedback reconstruction system for an RIS-assisted massive MIMO system, the system comprising:

The acquisition data module is used for acquiring the CSI channel state information fed back by the single-antenna UE from the UE-RIS-BS path;

The decomposition module is used for decomposing the feedback of the CSI channel state information into indexes of non-zero column signal vectors and feedback of the non-zero column signal vectors;

the preprocessing module is used for processing the index of the decomposed non-zero column signal vector and the feedback of the non-zero column signal vector to obtain incremental CSI channel state information;

The data processing module is used for inputting the incremental CSI channel state information into the trained feedback model and outputting the reconstructed CSI channel state information;

the feedback model trains a pair of dimension reduction and reconstruction dictionary groups by utilizing pre-stored CSI channel state information at a Base Station (BS), optimizes the dimension reduction and reconstruction dictionary groups in an alternate iteration mode to obtain two groups of optimal dictionaries, and outputs the reconstructed CSI channel state information by utilizing the two groups of optimal dictionaries.