CN111049559A - Deep learning precoding method using position information - Google Patents

Deep learning precoding method using position information Download PDF

Info

Publication number
CN111049559A
CN111049559A CN201911106524.8A CN201911106524A CN111049559A CN 111049559 A CN111049559 A CN 111049559A CN 201911106524 A CN201911106524 A CN 201911106524A CN 111049559 A CN111049559 A CN 111049559A
Authority
CN
China
Prior art keywords
vector
user
log
horizontal
vertical
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.)
Granted
Application number
CN201911106524.8A
Other languages
Chinese (zh)
Other versions
CN111049559B (en
Inventor
李潇
孟声国
杨浩
熊龙辉
何潜翔
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Southeast University
Original Assignee
Southeast University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Southeast University filed Critical Southeast University
Priority to CN201911106524.8A priority Critical patent/CN111049559B/en
Publication of CN111049559A publication Critical patent/CN111049559A/en
Application granted granted Critical
Publication of CN111049559B publication Critical patent/CN111049559B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • H04B7/0413—MIMO systems
    • H04B7/0456—Selection of precoding matrices or codebooks, e.g. using matrices antenna weighting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Abstract

The invention discloses a deep learning precoding method by using position information, which comprises the following steps: acquiring position information and statistical channel information of each user, and calculating the horizontal and vertical optimal direction indexes of each user; constructing a deep neural network model for judging horizontal and vertical indexes of the user, inputting position information of each user into the model, and outputting corresponding horizontal and vertical indexes; training the model to make the direction index output by the model gradually approach to the optimal direction index to obtain model parameters, and determining the precoding vector of the model according to the horizontal and vertical direction indexes output by the model. The invention can reduce the calculation complexity of the precoding vector, can predict the optimal precoding vector only by acquiring the position information of the user, has higher accuracy, and can efficiently realize the downlink precoding design particularly when the number of users and the number of antennas are larger.

Description

Deep learning precoding method using position information
Technical Field
The invention relates to a deep learning precoding technology utilizing position information, and belongs to the technical field of communication.
Background
A large-scale multiple-input multiple-output (MIMO) transmission technology is one of the key technologies in the 5G communication system. The technology replaces multi-antenna arrays with large-scale antenna arrays to achieve higher spectral efficiency and transmission reliability. When a base station side obtains downlink Channel State Information (CSI), interference between users can be eliminated through channel adaptive techniques such as precoding, beamforming, and the like, and information is transmitted by using three dimensions of space, time, and frequency, thereby greatly improving system capacity. However, it is difficult to obtain complete channel state information in a real wireless communication system in a timely manner. For a Time Division Duplex (TDD) system, due to reciprocity between an uplink channel and a downlink channel, corresponding downlink CSI can be obtained through uplink channel estimation, but pilot pollution exists in the obtained CSI. However, in a frequency division multiplexing (FDD) system, channels do not have reciprocity and need to be fed back through an uplink, and particularly when the number of users and the number of antennas are large, it is impractical for a ue to obtain downlink CSI that is sent to a base station through a feedback link. One effective way to overcome this difficulty is to use statistical CSI of the channel, such as transmit-receive correlation matrix, mean information, etc. Compared with the instantaneous CSI, the statistical CSI of the channel is approximately constant for a long time, the accuracy is relatively high, and the required feedback overhead is small.
In addition, with the increase of the number of users and the number of antennas, the calculation complexity of designing the user precoding vector by using the statistical CSI increases exponentially, and a large amount of calculation resources and a long calculation time are required by the conventional calculation method. The statistical CSI of a user is closely related to the environment, i.e. the location information of the user. In recent years, the deep learning method has attracted wide attention in the field of wireless communication due to the strong expression capability and the parallel computing structure of the network. In principle, as long as the parameters of the deep neural network are enough, the deep neural network can represent any function, and the trained neural network can classify and deduce input data only by performing some simple basic operations. Therefore, the deep neural network can be used for learning the relation between the user position and the statistical CSI thereof, and further obtaining the mapping relation between the user position and the precoding vector, thereby obtaining the precoding method with lower complexity.
Disclosure of Invention
The invention aims to overcome the defects in the prior art and provides a base station with a downlink of a uniform planar antenna array
The transmission system provides a deep learning precoding method by using position information, solves the problem of higher calculation complexity of precoding vectors, can design the precoding vectors of the users according to the position information of the users, and obtains higher precision by using fewer calculation resources.
In order to achieve the above object, the present invention provides a deep learning precoding design method using position information, which comprises the following steps:
step 1, a base station configures a uniform planar antenna array, wherein the antenna array comprises M rows of antenna array elements in the vertical direction, each row in the horizontal direction is provided with N antenna array elements, the distance between adjacent antenna array elements is half of the carrier wavelength in the horizontal direction and the vertical direction, and a user configures a single receiving antenna; obtaining statistical channel information and position information of U users, respectively calculating statistical parameters of users i, i being 1, …, U, and further calculating each user log represented by binary vector2N x 1 dimensional horizontal best direction index
Figure BDA0002271489510000021
And log2Vertical optimal direction index of mx 1 dimension
Figure BDA0002271489510000022
And combines the two vectors into one (log)2N+log2Vector of M) × 1 dimension
Figure BDA0002271489510000023
Wherein the vector front log2Vector formed by N elements corresponds to user horizontal optimal direction index
Figure BDA0002271489510000024
Rear log2The vector formed by M elements corresponds to the vertical optimal direction index
Figure BDA0002271489510000025
Forming U training samples;
the location information includes: three-dimensional coordinate (x) of user i by using base station as coordinate origini,yi,zi)T。
The statistical channel information includes: rice factor for user i-channel
Figure BDA0002271489510000026
Horizontal line of sight component
Figure BDA0002271489510000027
Vertical line of sight component
Figure BDA0002271489510000028
Horizontal correlation matrix
Figure BDA0002271489510000029
And vertical correlation array
Figure BDA00022714895100000210
Wherein, the matrix HiIs a channel matrix between the base station and user i, HiRow m and column n of [ H ]i]m,nIs the channel coefficient between the antenna unit of the mth row and nth column of the base station and the user i,
Figure BDA00022714895100000211
Figure BDA00022714895100000212
and
Figure BDA00022714895100000213
respectively represent matrices
Figure BDA00022714895100000214
And
Figure BDA00022714895100000215
the first column of (a) is,
Figure BDA00022714895100000216
superscript (·)HRepresenting a conjugate transpose, superscript (. cndot.)TRepresenting transposition, and E {. cndot } represents averaging;
the process of calculating the statistical parameters of the user i comprises the following sub-steps:
a1) respectively calculate
Figure BDA00022714895100000217
And
Figure RE-GDA0002397588400000031
wherein FMAnd FNDFT matrices, F, of M and N, respectivelyMAnd FNThe elements of the m-th row and the n-th column are respectively
Figure RE-GDA0002397588400000032
And
Figure RE-GDA0002397588400000033
a2) separately extracting AH,i、AV,i、ΛH,iAnd ΛV,iA diagonal element of (a) to obtainH,i、aV,i、λH,iAnd λV,i(ii) a Wherein, aH,iAnd λH,iIs a column vector of dimension Nx 1, the nth element of which is AH,iAnd ΛH,iThe nth diagonal element of (a); a isV,iAnd λV,iIs a column vector of dimension M × 1, the M-th elements of which are AV,iAnd ΛV,iThe mth diagonal element of (1);
the process of calculating the horizontal and vertical best direction indices of user i in binary vector representation comprises the following sub-steps:
b1) computing
Figure BDA0002271489510000034
b2) To obtain omegaiThe row and column of the largest element in the set, assume ΩiThe largest element in the group
Figure BDA0002271489510000035
Go to the first
Figure BDA0002271489510000036
Elements of the column, then
Figure BDA0002271489510000037
For the decimal horizontal best direction index,
Figure BDA0002271489510000038
indexing for a vertical best direction of the decimal;
b3) indexing the decimal
Figure BDA0002271489510000039
Separately undergo conversion into log2Binary vector of Nx 1 dimension
Figure BDA00022714895100000310
And log2Binary vector of Nx 1 dimension
Figure BDA00022714895100000311
Step 2, constructing and judging a depth neural network model which is expressed by binary vectors and indexed in the horizontal direction and the vertical direction, and setting the three-dimensional coordinates (x) of the user i, i as 1, …, Ui,yi,zi)TInput into neural network model, output (log)2N+log2M) × 1-dimensional binary vector
Figure BDA00022714895100000312
Front log of the vector2Vector formed by N elements corresponds to user horizontal direction index
Figure BDA00022714895100000313
Rear log2The vector formed by M elements corresponds to the vertical direction index
Figure BDA00022714895100000314
Step 3, training the models respectively by using the training samples formed in the step 1 to enable prediction output
Figure BDA00022714895100000315
Gradually approach to
Figure BDA00022714895100000316
Prediction output
Figure BDA00022714895100000317
Gradually approach to
Figure BDA00022714895100000318
To obtain parameters of the model;
step 4, obtaining the three-dimensional coordinate (x) of the user l of the vector to be precodedl,yl,zl)TInput into a trained model, and output by the model a horizontal direction index represented by a binary vector
Figure BDA00022714895100000319
And vertical direction index
Figure BDA00022714895100000320
Step 5, indexing the obtained horizontal direction represented by the binary vector
Figure BDA00022714895100000321
And vertical direction index
Figure BDA00022714895100000322
Converted into decimal representation
Figure BDA00022714895100000323
And
Figure BDA00022714895100000324
step 6, determining the precoding vector of the user l to obtain the precoding vector as
Figure RE-GDA0002397588400000041
Wherein
Figure RE-GDA0002397588400000042
Is a matrix FNTo (1) a
Figure RE-GDA0002397588400000043
The columns of the image data are,
Figure RE-GDA0002397588400000044
is a matrix FMTo (1) a
Figure RE-GDA0002397588400000045
And (4) columns.
As a preferred aspect of the present invention, the deep neural network model that determines the horizontal and vertical indexes represented by the binary vectors in step 2 includes an input layer, four hidden layers, a Dropout layer, and an output layer.
In a preferred embodiment of the present invention, the step 3 uses a cross entropy loss function to make the prediction output
Figure BDA0002271489510000046
And
Figure BDA0002271489510000047
gradually approach to
Figure BDA0002271489510000048
And
Figure BDA0002271489510000049
the cross entropy loss function adopted by the depth neural network model for judging the horizontal direction index and the vertical direction index represented by the binary vector is specifically as follows:
Figure BDA00022714895100000410
wherein U is the number of all samples in the training set,
Figure BDA00022714895100000411
horizontal best direction indexing for binary vector representation
Figure BDA00022714895100000412
And vertical best direction index
Figure BDA00022714895100000413
Combined (log)2N+log2Vector of M) × 1 dimension
Figure BDA00022714895100000414
The kth element of (1);
Figure BDA00022714895100000415
is output (log)2N+log2M) × 1-dimensional binary vector
Figure BDA00022714895100000416
The k-th element of (2), the vector front log2Vector formed by N elements corresponds to user horizontal direction index
Figure BDA00022714895100000417
Rear log2The vector formed by M elements corresponds to the vertical direction index
Figure BDA00022714895100000418
Has the advantages that:
(1) the invention needs small information amount, only needs the position information of the user, and is suitable for various typical wireless communication environments;
(2) the deep neural network designed by the invention is simple and easy to train, and has higher prediction accuracy;
(3) the method has high calculation efficiency, and particularly when the number of users and the number of antennas are large, the calculation efficiency of the method is improved by dozens of times compared with that of the traditional calculation method.
Drawings
Fig. 1 is a flowchart of the method of the present invention for calculating a statistical precoding vector for a user l to obtain a precoding vector by using a deep neural network obtained by training.
Detailed Description
The following describes embodiments of the present invention with reference to the drawings.
As shown in fig. 1, the present invention discloses a deep learning precoding method using location information, which specifically comprises the following steps:
step 1, a base station configures a uniform planar antenna array, wherein the antenna array comprises M rows of antenna array elements in the vertical direction, each row in the horizontal direction is provided with N antenna array elements, the distance between adjacent antenna array elements is half of the carrier wavelength in the horizontal direction and the vertical direction, and a user configures a single receiving antenna; obtaining statistical channel information and position information of U users, preprocessing the information, respectively calculating statistical parameters of users i, i is 1, …, U, and further calculating log of each user represented by binary vector2N x 1 dimensional horizontal best direction index
Figure BDA0002271489510000051
And log2Vertical optimal direction index of mx 1 dimension
Figure BDA0002271489510000052
And combines the two vectors into one (log)2N+log2Vector of M) × 1 dimension
Figure BDA0002271489510000053
Wherein the vector front log2Vector formed by N elements corresponds to user horizontal optimal direction index
Figure BDA0002271489510000054
Rear log2Vector formed by M elements corresponds to vertical optimal direction cableGuiding device
Figure BDA0002271489510000055
Forming U training samples;
the location information includes: three-dimensional coordinate (x) of user i by using base station as coordinate origini,yi,zi)T。
The statistical channel information includes: rice factor for user i-channel
Figure BDA0002271489510000056
Horizontal line of sight component
Figure BDA0002271489510000057
Vertical line of sight component
Figure BDA0002271489510000058
Horizontal correlation matrix
Figure BDA0002271489510000059
And vertical correlation array
Figure BDA00022714895100000510
Wherein, the matrix HiIs a channel matrix between the base station and user i, HiRow m and column n of [ H ]i]m,nIs the channel coefficient between the antenna unit of the mth row and nth column of the base station and the user i,
Figure BDA00022714895100000511
Figure BDA00022714895100000512
and
Figure BDA00022714895100000513
respectively represent matrices
Figure BDA00022714895100000514
And
Figure BDA00022714895100000515
the first column of (a) is,
Figure BDA00022714895100000516
superscript (·)HRepresenting a conjugate transpose, superscript (. cndot.)TRepresenting transposition, and E {. cndot } represents averaging;
the process of calculating the statistical parameters of the user i comprises the following sub-steps:
a1) respectively calculate
Figure BDA00022714895100000517
And
Figure BDA00022714895100000518
wherein FMAnd FNDFT matrices, F, of M and N, respectivelyMAnd FNThe elements of the m-th row and the n-th column are respectively
Figure BDA00022714895100000519
And
Figure BDA00022714895100000520
a2) separately extracting AH,i、AV,i、ΛH,iAnd ΛV,iA diagonal element of (a) to obtainH,i、aV,i、λH,iAnd λV,i(ii) a Wherein, aH,iAnd λH,iIs a column vector of dimension Nx 1, the nth element of which is AH,iAnd ΛH,iThe nth diagonal element of (a); a isV,iAnd λV,iIs a column vector of dimension M × 1, the M-th elements of which are AV,iAnd ΛV,iThe mth diagonal element of (1);
the process of calculating the horizontal and vertical best direction indices of user i in binary vector representation comprises the following sub-steps:
b1) computing
Figure BDA0002271489510000061
b2) To obtain omegaiThe row and column of the largest element in the set, assume ΩiMiddle maximumThe element is the first
Figure BDA0002271489510000062
Go to the first
Figure BDA0002271489510000063
Elements of the column, then
Figure BDA0002271489510000064
For the decimal horizontal best direction index,
Figure BDA0002271489510000065
indexing for a vertical best direction of the decimal;
b3) indexing the decimal
Figure BDA0002271489510000066
Separately undergo conversion into log2Binary vector of Nx 1 dimension
Figure BDA0002271489510000067
And log2Binary vector of Nx 1 dimension
Figure BDA0002271489510000068
Step 2, constructing and judging a depth neural network model which is expressed by binary vectors and indexed in the horizontal direction and the vertical direction, and setting the three-dimensional coordinates (x) of the user i, i as 1, …, Ui,yi,zi)TInput into neural network model, output (log)2N+log2M) × 1-dimensional binary vector
Figure BDA0002271489510000069
Front log of the vector2Vector formed by N elements corresponds to user horizontal direction index
Figure BDA00022714895100000610
Rear log2The vector formed by M elements corresponds to the vertical direction index
Figure BDA00022714895100000611
The deep neural network model comprises an input layer, four hidden layers, a Dropout layer and an output layer. The input layer is provided with 3 nodes, the four hidden layers and the Dropout layer adopt ReLU activation functions, the Dropout layer can avoid model overfitting by setting loss proportion, and the output layer adopts a Sigmoid activation function;
step 3, training the models respectively by using the training samples formed in the step 1 to enable prediction output
Figure BDA00022714895100000612
And
Figure BDA00022714895100000613
gradually approach to
Figure BDA00022714895100000614
And
Figure BDA00022714895100000615
to obtain parameters of the model; the method adopts an Adam optimization algorithm to minimize a cross entropy loss function, and the cross entropy loss function adopted by the depth neural network model for judging indexes in the horizontal direction and the vertical direction expressed by binary vectors specifically comprises the following steps:
Figure BDA00022714895100000616
wherein U is the number of all samples in the training set,
Figure RE-GDA0002397588400000071
horizontal best direction indexing for binary vector representation
Figure RE-GDA0002397588400000072
And vertical best direction index
Figure RE-GDA0002397588400000073
Combined (log)2N+log2M)×Vector of 1 dimension
Figure RE-GDA0002397588400000074
The kth element of (1);
Figure RE-GDA0002397588400000075
is output (log)2N+log2M) × 1-dimensional binary vector
Figure RE-GDA0002397588400000076
The k-th element of (2), the vector front log2Vector formed by N elements corresponds to user horizontal direction index
Figure RE-GDA0002397588400000077
Rear log2The vector formed by M elements corresponds to the vertical direction index
Figure RE-GDA0002397588400000078
Step 4, obtaining the three-dimensional coordinate (x) of the user l of the vector to be precodedl,yl,zl)TInput into a trained model, and output by the model a horizontal direction index represented by a binary vector
Figure BDA0002271489510000079
And vertical direction index
Figure BDA00022714895100000710
Step 5, indexing the obtained horizontal direction represented by the binary vector
Figure BDA00022714895100000711
And vertical direction index
Figure BDA00022714895100000712
Converted into decimal representation
Figure BDA00022714895100000713
And
Figure BDA00022714895100000714
step 6, determining the precoding vector of the user l to obtain the precoding vector as
Figure BDA00022714895100000715
Wherein
Figure BDA00022714895100000716
Is a matrix FNTo (1) a
Figure BDA00022714895100000717
The columns of the image data are,
Figure BDA00022714895100000718
is a matrix FMTo (1) a
Figure BDA00022714895100000719
And (4) columns.
In order to verify that the method of the present invention can reduce the computation complexity of the precoding vector and has a very high prediction accuracy, a verification example is specifically mentioned for description.
The verification example is a deep learning precoding method using position information, three-dimensional coordinates of a user are input into a deep neural network model, and a horizontal direction index and a vertical direction index which are output by the model and are represented by binary vectors are used for determining a precoding vector of the model. The method reduces the required information and achieves higher calculation precision with less calculation resources, and specifically comprises the following steps:
step 1, considering a downlink transmission system, a base station end is configured with a uniform planar antenna array, the antenna array comprises 16 rows of 32 antenna elements, the spacing between the adjacent antenna elements in the horizontal direction and the vertical direction is half wavelength of carrier, and a single receiving antenna is used by a user end. The statistical channel information and position information of 435000 users are obtained under the above conditions, and the 5 × 1-dimensional horizontal optimal direction index of each user represented by binary vector is calculated
Figure BDA00022714895100000720
And 4 x 1 dimensional vertical best direction index
Figure BDA00022714895100000721
And combines the two vectors into a 9 x 1 dimensional vector
Figure BDA00022714895100000722
Wherein the vector formed by the first 5 elements of the vector corresponds to the user's horizontal best direction index
Figure BDA00022714895100000723
The vector formed by the last 4 elements corresponds to the vertical best direction index
Figure BDA00022714895100000724
Training data is generated and divided into a training set of 400000 samples, a validation set of 30000 samples, and a test set of 5000 samples.
The location information includes: three-dimensional coordinate (x) of user i by using base station as coordinate origini,yi,zi)T。
The statistical channel information includes: rice factor for user i-channel
Figure RE-GDA0002397588400000081
Horizontal line of sight component
Figure RE-GDA0002397588400000082
Vertical line of sight component
Figure RE-GDA0002397588400000083
Horizontal correlation matrix
Figure RE-GDA0002397588400000084
And vertical correlation array
Figure RE-GDA0002397588400000085
Wherein, the matrix HiIs a channel matrix between the base station and user i, HiRow m and column n of [ H ]i]m,nIs the channel coefficient between the antenna unit of the mth row and nth column of the base station and the user i,
Figure RE-GDA0002397588400000086
Figure RE-GDA0002397588400000087
and
Figure RE-GDA0002397588400000088
respectively represent matrices
Figure RE-GDA0002397588400000089
And
Figure RE-GDA00023975884000000810
the first column of (a) is,
Figure RE-GDA00023975884000000811
superscript (·)HRepresenting a conjugate transpose, superscript (. cndot.)TRepresenting transposition, and E {. cndot } represents averaging;
the process of calculating the statistical parameters of the user i comprises the following sub-steps:
a1) respectively calculate
Figure BDA00022714895100000812
And
Figure BDA00022714895100000813
wherein F16And F32DFT matrices, F, of 16 × 16 and 32 × 32, respectively16And F32The elements of the m-th row and the n-th column are respectively
Figure BDA00022714895100000814
And
Figure BDA00022714895100000822
a2) separately extracting AH,i、AV,i、ΛH,iAnd ΛV,iA diagonal element of (a) to obtainH,i、aV,i、λH,iAnd λV,i(ii) a Wherein, aH,iAnd λH,iA column vector of 32 × 1 dimension, whose n-th elements are AH,iAnd ΛH,iThe nth diagonal element of (a); a isV,iAnd λV,iA 16 × 1 column vector with m-th elements of AV,iAnd ΛV,iThe mth diagonal element of (1);
the process of calculating the horizontal and vertical best direction indices of user i in binary vector representation comprises the following sub-steps:
b1) computing
Figure BDA00022714895100000815
b2) To obtain omegaiThe row and column of the largest element in the set, assume ΩiThe largest element in the group
Figure BDA00022714895100000816
Go to the first
Figure BDA00022714895100000817
Elements of the column, then
Figure BDA00022714895100000818
For the decimal horizontal best direction index,
Figure BDA00022714895100000819
indexing for a vertical best direction of the decimal;
b3) indexing the decimal
Figure BDA00022714895100000820
Respectively converted into 5 x 1 dimensional binary vectors
Figure BDA00022714895100000821
And 4 x 1 dimensional binary vector
Figure BDA0002271489510000091
Step 2, constructing judgment expressed by binary vectorThe deep neural network model is indexed in the horizontal direction and the vertical direction, and specific parameters of the deep neural network model are as follows: the input layer is provided with 3 nodes, the number of neurons of the first hidden layer is 150, and the activation function is a ReLU function; the number of the neurons of the second hidden layer is 180, the number of the activation functions is a ReLU function, the number of the neurons of the third hidden layer is 230, the activation functions is a ReLU function, the number of the neurons of the fourth hidden layer is 250, the number of the activation functions is a ReLU function, the number of the neurons of the Dropout layer is 300, the activation functions are ReLU functions, and the probability of inactivation is 0.5 in each training process; the number of the neurons of the output layer is 9, and a 9 multiplied by 1 dimensional binary vector is output
Figure BDA0002271489510000092
The vector formed by the first 5 elements of the vector corresponds to the horizontal index of the user
Figure BDA0002271489510000093
The vector formed by the last 4 elements corresponds to the user vertical direction index
Figure BDA0002271489510000094
The activation function is a Sigmoid function.
Step 3, designing and judging the loss function of the deep neural network model which is expressed by the binary vector and indexed in the horizontal direction and the vertical direction as the prediction output of the network
Figure BDA0002271489510000095
And a horizontal best direction index and a vertical best direction index vector represented by a binary vector by a user
Figure BDA0002271489510000096
The cross entropy loss function of (1) is specifically:
Figure BDA0002271489510000097
and U is all sample numbers of the training set, then the training set generated in the step 1 is used, the Adam optimization algorithm is adopted to train the model, and the weight value and the offset value of the model are changed through iterative training, so that the loss function is minimum. And (3) calculating the gradient by using all samples in the training set in each iteration, updating parameters according to an Adam optimization algorithm, traversing the training set 10000 times by the method, and using the learning rate of 0.001. And in the training process, the training iteration number of the model is adjusted by using the verification set, and when the training iteration number is determined to be 8950, the loss function of the model on the verification set is minimum.
And 4, the trained model can be used for user precoding design of a downlink transmission system of a base station configured with a 16 × 32 uniform planar antenna array. Three-dimensional coordinates (x) of user l of the vector to be precoded to be obtainedl,yl,zl)TInput into a trained model, and output by the model a horizontal direction index represented by a binary vector
Figure BDA0002271489510000098
And vertical direction index
Figure BDA0002271489510000099
Step 5, indexing the obtained horizontal direction represented by the binary vector
Figure BDA00022714895100000910
And vertical direction index
Figure BDA00022714895100000911
Converted into decimal representation
Figure BDA00022714895100000912
And
Figure BDA00022714895100000913
step 6, determining the precoding vector of the user l to obtain the precoding vector as
Figure BDA00022714895100000914
Wherein
Figure BDA00022714895100000915
Is a matrix F32To (1) a
Figure BDA00022714895100000916
The columns of the image data are,
Figure BDA00022714895100000917
is a matrix F16To (1) a
Figure BDA00022714895100000918
And (4) columns. The final performance of the test set test model is used, the final prediction accuracy is 92.36%, the time required for calculating the optimal transmission direction vector of the user in the test set by using a deep learning method is 0.0189 second, and the time required for directly calculating the optimal transmission direction vector is 1.0098 second.
In conclusion, the invention can reduce the calculation complexity of the precoding vector, has higher accuracy for determining the optimal transmission direction, and can efficiently realize the downlink transmission statistical precoding design especially when the number of users and the number of antennas are larger.
The embodiments of the present invention have been described in detail with reference to the drawings, but the present invention is not limited to the embodiments described above, and various changes can be made within the knowledge of those skilled in the art without departing from the gist of the present invention.

Claims (3)

1. A deep learning precoding method using position information, comprising the steps of:
step 1, a base station configures a uniform planar antenna array, wherein the antenna array comprises M rows of antenna array elements in the vertical direction, each row in the horizontal direction is provided with N antenna array elements, the distance between adjacent antenna array elements is half of the carrier wavelength in the horizontal direction and the vertical direction, and a user configures a single receiving antenna; obtaining statistical channel information and position information of U users, respectively calculating statistical parameters of users i, i being 1, …, U, and further calculating each user log represented by binary vector2N x 1 dimensional horizontal best direction index
Figure FDA0002271489500000011
And log2Vertical optimal direction index of mx 1 dimension
Figure FDA0002271489500000012
And combines the two vectors into one (log)2N+log2Vector of M) × 1 dimension
Figure FDA0002271489500000013
Wherein the vector front log2Vector formed by N elements corresponds to user horizontal optimal direction index
Figure FDA0002271489500000014
Rear log2The vector formed by M elements corresponds to the vertical optimal direction index
Figure FDA0002271489500000015
Forming U training samples;
the location information includes: three-dimensional coordinate (x) of user i by using base station as coordinate origini,yi,zi)T。
The statistical channel information includes: rice factor for user i-channel
Figure FDA0002271489500000016
Horizontal line of sight component
Figure FDA0002271489500000017
Vertical line of sight component
Figure FDA0002271489500000018
Horizontal correlation matrix
Figure FDA0002271489500000019
And vertical correlation array
Figure FDA00022714895000000110
Wherein, the matrix HiIs a channel matrix between the base station and user i, HiRow m and column n of [ H ]i]m,nIs the channel coefficient between the antenna element of the mth row and nth column of the base station and the user i,
Figure FDA00022714895000000111
Figure FDA00022714895000000112
and
Figure FDA00022714895000000113
respectively represent matrices
Figure FDA00022714895000000114
And
Figure FDA00022714895000000115
the first column of (a) is,
Figure FDA00022714895000000116
superscript (·)HRepresenting a conjugate transpose, superscript (. cndot.)TRepresenting transposition, E {. cndot } represents averaging;
the process of calculating the statistical parameters of the user i comprises the following sub-steps:
a1) respectively calculate
Figure FDA00022714895000000117
And
Figure FDA00022714895000000118
wherein FMAnd FNDFT matrices, F, of M and N, respectivelyMAnd FNThe elements of the m-th row and the n-th column are respectively
Figure FDA00022714895000000119
And
Figure FDA00022714895000000120
a2) separately extracting AH,i、AV,i、ΛH,iAnd ΛV,iA diagonal element of (a) to obtainH,i、aV,i、λH,iAnd λV,i(ii) a Wherein, aH,iAnd λH,iIs a column vector of dimension Nx 1, the nth element of which is AH,iAnd ΛH,iThe nth diagonal element of (a); a isV,iAnd λV,iIs a column vector of dimension M × 1, the M-th elements of which are AV,iAnd ΛV,iThe mth diagonal element of (1);
the process of calculating the horizontal and vertical best direction indices of user i in binary vector representation comprises the following sub-steps:
b1) computing
Figure FDA0002271489500000021
b2) To obtain omegaiThe row and column of the largest element in the set, assume ΩiThe largest element in the group
Figure FDA0002271489500000022
Go to the first
Figure FDA0002271489500000023
Elements of the column, then
Figure FDA0002271489500000024
For the decimal horizontal best direction index,
Figure FDA0002271489500000025
indexing for a vertical best direction of the decimal;
b3) indexing the decimal
Figure FDA0002271489500000026
Separately undergo conversion into log2Binary vector of Nx 1 dimension
Figure FDA0002271489500000027
And log2Binary vector of Nx 1 dimension
Figure FDA0002271489500000028
Step 2, constructing a depth neural network model for judging indexes in the horizontal direction and the vertical direction represented by binary vectors, and setting the three-dimensional coordinates (x) of the user i, i as 1, …, Ui,yi,zi)TInput into neural network model, output (log)2N+log2M) × 1-dimensional binary vector
Figure FDA0002271489500000029
Front log of the vector2Vector formed by N elements corresponds to user horizontal direction index
Figure FDA00022714895000000210
Rear log2The vector formed by M elements corresponds to the vertical direction index
Figure FDA00022714895000000211
Step 3, training the model by using the training sample formed in the step 1 to enable prediction output
Figure FDA00022714895000000212
Gradually approach to
Figure FDA00022714895000000213
Prediction output
Figure FDA00022714895000000214
Gradually approach to
Figure FDA00022714895000000215
To obtain parameters of the model;
step 4, the obtained vector to be precodedThree-dimensional coordinates (x) of user ll,yl,zl)TInput into a trained model, and output by the model a horizontal direction index represented by a binary vector
Figure FDA00022714895000000216
And vertical direction index
Figure FDA00022714895000000217
Step 5, indexing the obtained horizontal direction represented by the binary vector
Figure FDA00022714895000000218
And vertical direction index
Figure FDA00022714895000000219
Converted into decimal representation
Figure FDA00022714895000000220
And
Figure FDA00022714895000000221
step 6, determining the precoding vector of the user l to obtain the precoding vector as
Figure FDA00022714895000000222
Wherein
Figure FDA00022714895000000223
Is a matrix FNTo (1) a
Figure FDA00022714895000000224
The columns of the image data are,
Figure FDA00022714895000000225
is a matrix FMTo (1) a
Figure FDA00022714895000000226
And (4) columns.
2. The deep learning precoding method using the position information as claimed in claim 1, wherein: the depth neural network model which is represented by binary vectors and is indexed in the horizontal direction and the vertical direction is judged in the step 2 to comprise an input layer, four hidden layers, a Dropout layer and an output layer.
3. The deep learning precoding method using the position information as claimed in claim 1, wherein: in the step 3, a cross entropy loss function is adopted to enable prediction output
Figure FDA00022714895000000227
And
Figure FDA00022714895000000228
gradually approach to
Figure FDA00022714895000000229
And
Figure FDA00022714895000000230
the cross entropy loss function adopted by the depth neural network model for judging the horizontal direction index and the vertical direction index represented by the binary vector is specifically as follows:
Figure FDA0002271489500000031
wherein U is the number of all samples in the training set,
Figure FDA0002271489500000032
horizontal best direction indexing for binary vector representation
Figure FDA0002271489500000033
And vertical best direction index
Figure FDA0002271489500000034
Combined (log)2N+log2Vector of M) × 1 dimension
Figure FDA0002271489500000035
The kth element of (1);
Figure FDA0002271489500000036
is output (log)2N+log2M) × 1-dimensional binary vector
Figure FDA0002271489500000037
The k-th element of (2), the vector front log2Vector formed by N elements corresponds to user horizontal direction index
Figure FDA0002271489500000038
Rear log2The vector formed by M elements corresponds to the vertical direction index
Figure FDA0002271489500000039
CN201911106524.8A 2019-11-13 2019-11-13 Deep learning precoding method using position information Active CN111049559B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911106524.8A CN111049559B (en) 2019-11-13 2019-11-13 Deep learning precoding method using position information

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911106524.8A CN111049559B (en) 2019-11-13 2019-11-13 Deep learning precoding method using position information

Publications (2)

Publication Number Publication Date
CN111049559A true CN111049559A (en) 2020-04-21
CN111049559B CN111049559B (en) 2022-03-11

Family

ID=70232623

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911106524.8A Active CN111049559B (en) 2019-11-13 2019-11-13 Deep learning precoding method using position information

Country Status (1)

Country Link
CN (1) CN111049559B (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104320171A (en) * 2014-11-03 2015-01-28 东南大学 Three-dimensional beam forming space division multi-access adaptive transmission method adopting statistic channel state information
CN107332598A (en) * 2017-06-26 2017-11-07 浙江理工大学 A kind of precoding of mimo system joint and antenna selecting method based on deep learning
CN109560846A (en) * 2019-01-04 2019-04-02 东南大学 A kind of three-dimensional method for precoding based on model-driven deep learning
CN109687897A (en) * 2019-02-25 2019-04-26 西华大学 Superposition CSI feedback method based on the extensive mimo system of deep learning
KR20190069332A (en) * 2017-12-11 2019-06-19 한국전자통신연구원 Beamforming Method Performing the Deep-Learning

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104320171A (en) * 2014-11-03 2015-01-28 东南大学 Three-dimensional beam forming space division multi-access adaptive transmission method adopting statistic channel state information
CN107332598A (en) * 2017-06-26 2017-11-07 浙江理工大学 A kind of precoding of mimo system joint and antenna selecting method based on deep learning
KR20190069332A (en) * 2017-12-11 2019-06-19 한국전자통신연구원 Beamforming Method Performing the Deep-Learning
CN109560846A (en) * 2019-01-04 2019-04-02 东南大学 A kind of three-dimensional method for precoding based on model-driven deep learning
CN109687897A (en) * 2019-02-25 2019-04-26 西华大学 Superposition CSI feedback method based on the extensive mimo system of deep learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
XIAOXIANG YU; XI YANG; XIAO LI; SHI JIN: "Deep Learning based Beamforming for FD-MIMO Downlink Transmission", 《2019 IEEE/CIC INTERNATIONAL CONFERENCE ON COMMUNICATIONS IN CHINA (ICCC)》 *
刘斌; 任欢; 李立欣: "基于机器学习的毫米波大规模MIMO混合预编码技术", 《移动通信》 *

Also Published As

Publication number Publication date
CN111049559B (en) 2022-03-11

Similar Documents

Publication Publication Date Title
CN109560846B (en) Three-dimensional precoding method based on model-driven deep learning
CN108964726B (en) Low-complexity large-scale MIMO uplink transmission channel estimation method
Lin et al. Tensor-based algebraic channel estimation for hybrid IRS-assisted MIMO-OFDM
CN107332598A (en) A kind of precoding of mimo system joint and antenna selecting method based on deep learning
CN110557177A (en) DenseNet-based hybrid precoding method in millimeter wave large-scale MIMO system
CN109617584A (en) A kind of mimo system beamforming matrix design method based on deep learning
Huang et al. Beam-blocked channel estimation for FDD massive MIMO with compressed feedback
Lin et al. Deep learning-based antenna selection and CSI extrapolation in massive MIMO systems
CN110430150B (en) Receiver design method of cellular mobile communication system based on neural network
CN105846879A (en) Iterative beam forming method of millimeter wave precoding system
CN108881076A (en) A kind of compressed sensing based MIMO-FBMC/OQAM system channel estimation method
CN111181671A (en) Deep learning-based downlink channel rapid reconstruction method
CN107086886A (en) The double-deck Precoding Design of extensive mimo system fusion ZF and Taylor series expansion
CN114338301B (en) Channel estimation method of RIS auxiliary millimeter wave system based on compressed sensing
Hu et al. Two-timescale end-to-end learning for channel acquisition and hybrid precoding
CN108933619B (en) Large-scale MIMO hybrid precoding codebook design method
CN111010222B (en) Deep learning-based large-scale MIMO downlink user scheduling method
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
CN114884775A (en) Deep learning-based large-scale MIMO system channel estimation method
Li et al. Multi-task deep learning based hybrid precoding for mmWave massive MIMO system
Turan et al. A versatile low-complexity feedback scheme for FDD systems via generative modeling
CN111988069B (en) Large-scale MIMO generalized eigenvector structure precoding solving method and device
CN111049559B (en) Deep learning precoding method using position information
CN112929304A (en) Wireless channel acquisition method assisted by environment knowledge base

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
GR01 Patent grant
GR01 Patent grant