阅读说明：本技术 基于复数卷积网络的混合波束成形方法 (Hybrid beam forming method based on complex convolution network ) 是由 罗杨 刘子健 骆春波 许燕 杨秉鸿 张富鑫 于 2021-07-20 设计创作，主要内容包括：本发明公开了一种基于复数卷积网络的毫米波波束成形方法,所述毫米波波束成形方法包括：S1：获取发射天线和接收天线间的信道状态信息；S2：根据所述信道状态信息,得到信道矩阵；S3：根据所述信道矩阵,利用复数卷积网络,得到基带预编码矩阵和射频预编码矩阵；S4：将所述基带预编码矩阵和所述射频预编码矩阵输入混合波束成型系统中,生成毫米波波束。本发明所提供的基于复数卷积网络的毫米波波束成形方法,能够解决一般深度学习方法无法有效利用通信信号处理中的复数相位信息的问题。(The invention discloses a millimeter wave beam forming method based on a complex convolution network, which comprises the following steps: s1: acquiring channel state information between a transmitting antenna and a receiving antenna; s2: obtaining a channel matrix according to the channel state information; s3: obtaining a baseband precoding matrix and a radio frequency precoding matrix by utilizing a complex convolution network according to the channel matrix; s4: and inputting the baseband pre-coding matrix and the radio frequency pre-coding matrix into a hybrid beam forming system to generate millimeter wave beams. The millimeter wave beam forming method based on the complex convolution network can solve the problem that the common deep learning method cannot effectively utilize the complex phase information in the communication signal processing.)

1. The millimeter wave beam forming method based on the complex convolution network is characterized by comprising the following steps:

s1: acquiring channel state information between a transmitting antenna and a receiving antenna;

s2: obtaining a channel matrix according to the channel state information;

s3: obtaining a baseband precoding matrix and a radio frequency precoding matrix by utilizing a complex convolution network according to the channel matrix;

s4: and inputting the baseband pre-coding matrix and the radio frequency pre-coding matrix into a hybrid beam forming system to generate millimeter wave beams.

2. The complex convolutional network-based millimeter wave beamforming method of claim 1, wherein in step S4, the hybrid beamforming system comprises multiple transmit antennas, multiple transmit chains, multiple data streams and receive antennas, and step S4 comprises the following sub-steps:

s41: converting the multi-path data code stream to a transmitting link to obtain a plurality of transmitting links;

s42: converting the plurality of transmitting chains to a plurality of transmitting antennas to generate transmitting signals;

s43: transmitting the transmission signal to the receiving antenna, and/or receiving the transmission signal to generate a receiving signal.

3. The complex convolutional network-based millimeter wave beamforming method of claim 2, wherein in step S43, the transmission signal is:

X＝F_RFF_BBS

wherein X represents a transmit signal vector and S is N_sX 1-dimensional data code stream vector, N_sAs a number of data streams, F_BBRepresenting the baseband precoding matrix, F_RFRepresenting the radio frequency precoding matrix.

4. The complex convolutional network-based millimeter wave beamforming method of claim 2, wherein in step S44, the received signals are:

wherein Y represents a received signal vector, and H is N_r×N_tAnd a channel matrix ofP represents the average received power and n is subject to independent distributionNoise vector, F_RFF_BBS is a transmit signal and S is N_sX 1-dimensional data code stream vector, N_sAs a number of data streams, F_BBRepresenting the baseband precoding matrix, F_RFRepresenting the radio frequency precoding matrix, N_tIs the number of transmit antennas; n is a radical of_rIs the number of receive antennas.

5. The complex convolutional network based millimeter wave beamforming method according to any one of claims 2 to 4, wherein in step S3, the complex convolutional network comprises an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fully-connected layer and an output layer which are connected in sequence.

6. The complex convolutional network-based millimeter wave beamforming method of claim 5, wherein the length of the input layer is 2N_tN_rIts width is 1, its height is 1;

the first coiled layer has a length of 2N_tWidth 1 and height 46;

the length of the third convolution layer isWidth 1, height 128;

the length of the full connecting layer isWidth 1, height 1;

the length of the output layer isWidth 1, height 1;

wherein N is_tIs the number of transmit antennas; n is a radical of_rIs the number of receiving antennas;is the number of transmit chains; n is a radical of_sIs the number of data code streams.

7. The complex convolutional network-based millimeter wave beam forming method of claim 6, wherein the length, width and height of the second convolutional layer are equal to the length, width and height of the first convolutional layer, and/or

The length, width and height of the fourth convolution layer are equal to the length, width and height of the third convolution layer.

8. The complex convolutional network-based millimeter wave beamforming method of claim 6, wherein the length, width and height of the input layer are obtained by:

a1: respectively vectorizing a real part and an imaginary part of the channel matrix to obtain a real part vector and an imaginary part vector;

a2: connecting the real part vector and the imaginary part vector end to generate a new vector of the input layer;

a3: and outputting the new vector as the length, width and height of the input layer.

Technical Field

The invention relates to the technical field of millimeter wave beam forming, in particular to a hybrid beam forming method based on a complex convolution network.

Background

Millimeter wave beam forming mainly adopts a digital and analog mixed structure, realizes efficient receiving and transmitting beam alignment with lower hardware cost, and is a core technology in 5G and future communication systems. Taking transmit beamforming as an example, the essential task of millimeter wave hybrid beamforming is to find the approximate decomposition of the optimal all-digital precoding matrix by designing a reasonable algorithm, that is, the all-digital precoding matrix is approximately decomposed into a digital precoding matrix and an analog precoding matrix under a certain constraint condition. In the traditional method, solution is performed by optimization methods such as compressed sensing or manifold optimization, but the algorithm with good performance (spectral efficiency) is slow in operation speed, and the algorithm with high operation speed is poor in performance. The method based on deep learning becomes a research hotspot in the field of beam forming in recent years, and the trained neural network is the most possible method for breaking through the bottleneck due to the high operation speed.

All-digital pre-coding matrixes and digital and analog mixed pre-coding matrixes are complex matrixes, and the situation of processing the complex matrixes is inevitable by using a deep learning method. In the existing method, a complex number is split into a real part and an imaginary part and is divided into two networks for processing, or the real part and the imaginary part are split and then are superposed into two channels for processing, and the phase information of the complex number is not completely utilized.

Disclosure of Invention

The invention aims to provide a hybrid beam forming method based on a complex convolution network, which solves the problem that the common deep learning method cannot effectively utilize complex phase information in communication signal processing.

The invention provides a millimeter wave beam forming method based on a complex convolution network, which comprises the following steps:

s1: acquiring channel state information between a transmitting antenna and a receiving antenna;

s2: obtaining a channel matrix according to the channel state information;

s3: obtaining a baseband precoding matrix and a radio frequency precoding matrix by utilizing a complex convolution network according to the channel matrix;

s4: and inputting the baseband pre-coding matrix and the radio frequency pre-coding matrix into a hybrid beam forming system to generate millimeter wave beams.

Optionally, in the step S4, the hybrid beamforming system includes multiple transmitting antennas, multiple transmitting links, multiple data streams, and receiving antennas, and the step S4 includes the following sub-steps:

s41: converting the multi-path data code stream to a transmitting link to obtain a plurality of transmitting links;

s42: converting the plurality of transmitting chains to a plurality of transmitting antennas to generate transmitting signals;

s43: transmitting the transmission signal to the receiving antenna, and/or receiving the transmission signal to generate a receiving signal.

Optionally, in step S43, the transmission signal is:

X＝F_RFF_BBS

wherein X represents a transmit signal vector and S is N_sX 1-dimensional data code stream vector, N_sAs a number of data streams, F_BBRepresenting the baseband precoding matrix, F_RFRepresenting the radio frequency precoding matrix.

Optionally, in step S44, the received signal is:

wherein Y represents a received signal vector, and H is N_r×N_tAnd a channel matrix ofP represents the average received power and n is subject to independent distributionNoise vector, F_RFF_BBS is a transmit signal and S is N_sX 1-dimensional data code stream vector, N_sAs a number of data streams, F_BBRepresenting the baseband precoding matrix, F_RFRepresenting the radio frequency precoding matrix, N_tIs the number of transmit antennas; n is a radical of_rIs the number of receive antennas.

Optionally, in step S3, the complex convolutional network includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fully-connected layer, and an output layer, which are connected in sequence.

Optionally, the input layer is 2N long_tN_rIts width is 1, its height is 1;

the first coiled layer has a length of 2N_tWidth 1 and height 46;

the length of the third convolution layer isWidth 1, height 128;

the length of the full connecting layer isWidth 1, height 1;

the length of the output layer isWidth 1, height 1;

wherein N is_tIs the number of transmit antennas; n is a radical of_rIs the number of receiving antennas;is the number of transmit chains; n is a radical of_sIs the number of data code streams.

Optionally, the second convolution layer has a length, width and height equal to the length, width and height of the first convolution layer, and/or

The length, width and height of the fourth convolution layer are equal to the length, width and height of the third convolution layer.

Optionally, the width and length of the input layer are obtained by:

a1: respectively vectorizing a real part and an imaginary part of the channel matrix to obtain a real part vector and an imaginary part vector;

a2: connecting the real part vector and the imaginary part vector end to generate a new vector of the input layer;

a3: and outputting the new vector as the length, width and height of the input layer.

The invention has the following beneficial effects:

compared with the prior art, the invention has the beneficial effects that: in one aspect, a single neural network may be used to process complex data, with the input being complex and the output also being complex. Compared with two networks which separately process a real part and an imaginary part, the method adopts a parameter sharing strategy to ensure that the parameters of the networks are not obviously increased. On the other hand, a complex convolution network structure is proposed for the first time, and forward transmission of parameters is realized according to a complex operation rule on the basis of the structure of the convolution network. Compared with a complex fully-connected network, the performance of the system is further improved.

Drawings

Fig. 1 is a flowchart of a millimeter wave beam forming method based on a complex convolutional network according to the present invention;

FIG. 2 is a flowchart illustrating the substeps of step S4 in FIG. 1;

fig. 3 is a structural diagram of a complex convolutional network of a millimeter wave beam forming method based on the complex convolutional network according to an embodiment of the present invention;

fig. 4 is a flowchart of a method for obtaining the length, width, and height of the input layer of the complex convolutional network based on the millimeter wave beamforming method of the complex convolutional network according to the embodiment of the present invention.

Detailed Description

The principles and features of this invention are described below in conjunction with the following drawings, which are set forth by way of illustration only and are not intended to limit the scope of the invention.

Examples

The technical scheme for solving the technical problems is as follows:

the invention provides a millimeter wave beam forming method based on a complex convolution network, which comprises the following steps:

s1: acquiring channel state information between a transmitting antenna and a receiving antenna;

s2: obtaining a channel matrix according to the channel state information;

s3: obtaining a baseband precoding matrix and a radio frequency precoding matrix by utilizing a complex convolution network according to the channel matrix;

s4: and inputting the baseband pre-coding matrix and the radio frequency pre-coding matrix into a hybrid beam forming system to generate millimeter wave beams.

The invention has the following beneficial effects:

compared with the prior art, the invention has the beneficial effects that: in one aspect, a single neural network may be used to process complex data, with the input being complex and the output also being complex. Compared with two networks which separately process a real part and an imaginary part, the method adopts a parameter sharing strategy to ensure that the parameters of the networks are not obviously increased. On the other hand, a complex convolution network structure is proposed for the first time, and forward transmission of parameters is realized according to a complex operation rule on the basis of the structure of the convolution network. Compared with a complex fully-connected network, the performance of the system is further improved.

Optionally, in the step S4, the hybrid beamforming system includes multiple transmitting antennas, multiple transmitting links, multiple data streams, and receiving antennas, and the step S4 includes the following sub-steps:

s41: converting the multi-path data code stream to a transmitting link to obtain a plurality of transmitting links;

s42: converting the plurality of transmitting chains to a plurality of transmitting antennas to generate transmitting signals;

s43: transmitting the transmission signal to the receiving antenna, and/or receiving the transmission signal to generate a receiving signal.

Specifically, the transmitting end has N_tRoot antenna transmitting N_sA data code stream, a receiving end has N_rThe root receives the antenna. The transmitting end hasA transmission chain, andfor the hardware structure of digital and analog pre-coding, the transmitting end has one in the first placeDimensional baseband (digital) precoding matrix F_BBIs a reaction of N_sConversion of way data code stream toOn a radio frequency link, then by oneDimensional radio frequency precoding matrix F_RFAnd different radio frequency links are converted to corresponding antennas, namely the number of transmission code streams is determined by the digital pre-coding matrix, and the connection mode of the radio frequency links and the antennas is determined by the radio frequency pre-coding matrix. One radio frequency link passes through N_tA phase shifter and N_tAn adder is connected to N_tA root antenna, thenNeed of a radio frequency linkPhase shifters, i.e. precoding matrices F_RFEach element of (a) represents a parameter of a phase shifter.

Optionally, in step S43, the transmission signal is:

X＝F_RFF_BBS

wherein X represents a transmit signal vector and S is N_sX 1-dimensional data code stream vector, N_sAs a number of data streams, F_BBRepresenting the baseband precoding matrix, F_RFRepresenting the radio frequency precoding matrix.

Optionally, in step S44, the received signal is:

wherein Y represents a received signal vector, and H is N_r×N_tAnd a channel matrix ofP represents the average received power and n is subject to independent distributionNoise vector, F_RFF_BBS is a transmit signal and S is N_sX 1-dimensional data code stream vector, N_sAs a number of data streams, F_BBRepresenting the baseband precoding matrix, F_RFRepresenting the radio frequency precoding matrix, N_tIs the number of transmit antennas; n is a radical of_rIs the number of receive antennas.

Optionally, in step S3, the complex convolutional network includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fully-connected layer, and an output layer, which are connected in sequence.

Optionally, the input layer is 2N long_tN_rIts width is 1, its height is 1;

specifically, H is an N_t×N_rEach element of the complex matrix is corresponding to a channel state factor from a certain transmitting antenna to a certain receiving antenna, and in a millimeter wave massive MIMO scene, adjacent elements in H may have correlation, so that a convolutional network is considered to be used to replace a traditional precoding method. The convolutional network shown in FIG. 3 is composed of 1 input layer, 1 output layer, 4 convolutional layers, and 1 fully-connected layer. For ease of comparison, we designed the convolutional layer to be 4 layers. The number of nodes in each layer is expressed as length x width x height, i.e. the number of nodes isWhereinThe length of the nodes representing the real or imaginary part of the l-th layer, so that the total length isIs the node width of the l-th layer;is the node height of the l-th layer, l ═ 0,1,2 …, 6. The corresponding convolution kernel is also denoted as length x width x height, and for ease of understanding, we plot the different channels of the convolution kernel in the figure, where a complete convolution kernel includes several single-channel convolution kernels, and the number of channels of the convolution kernel determines the height of the next layer of convolution layers. Using four types of convolution kernels, respectively denoted asWhereinRepresenting the q channel of the first convolutional layer. Similar to the parameter sharing of the fully-connected layer, we consider sharing the parameters of the convolution kernel, i.e. for l 1,2,3,4, there is

The first coiled layer has a length of 2N_tWidth 1 and height 46;

(1) input layer to first layer convolutional layer (conv.1layer):

specifically, the input layer of the neural network respectively vectorizes the real part and the imaginary part of H and then carries out head-to-tail phaseThen form a layer of 2N_tN_rOne-dimensional vector of elements, length, width and height of input layer being 2N_tN_rX 1X 1, i.eThe input layer and the first convolutional layer are connected as shown in the right block diagram of FIG. 1, and the size of the single-channel convolution kernel is N_rX 1, the number of convolution channels depending on the height of the next layerI.e. q 1, …, 64. Let stride equal to N_rThat is, for the first layer convolution layer, each type of convolution kernel needs to be shifted for useLet p be 1,2, …, N_t。

At the time of the p-th convolution, willAnd the real part of the first layer [ (p-1) stride +1, (p-1) stride + N_r]The values of the individual nodes are multiplied and summed and thenAnd the imaginary part of the first layer [ (p-1) stride +1, (p-1) stride + N_r]The values of the nodes are dot multiplied and summed, and the two results are subtracted to obtain the value of the position of the first layer convolution layer (p,1, 1). By analogy, for q 1, …,64 use is made ofAndobtaining the value of the first layer convolution layer (p,1, q) nodeAs shown in the following formula:

will be provided withAnd the real part of the first layer [ (p-1) stride +1, (p-1) stride + N_r]The nodes are subjected to dot multiplication and summation, and thenAnd an imaginary part [ (p-1) stride +1, (p-1) stride + N) of the first layer_r]The nodes are dot multiplied and summed, and the two results are added to obtain (N) of the first convolution layer_tValue of + p,1,1) position. By analogy, for q 1, …,64 use is made ofAndobtaining a first layer of a convolution layerValue of a node

The node size of the resulting first layer convolutional layer is 2N_tX 1X 64, i.e

The length, width and height of the second winding layer are equal to the length, width and height of the first winding layer.

(2) First to second layer of convolutional layers (conv.1layer):

specifically, unlike (1), since the height of the first layer convolution layer is not 1, the convolution kernel used is different, that is, it isThe height of the middle q channel convolution kernel is determined by the height of the input first layer convolution layer, and the size of the single channel convolution kernel is 3 multiplied by 1 multiplied by 64. Number of convolution channels ofQ is 1, …,64, and if step stride is 1, p is 1, …, N_t。

At the time of the p-th convolution, willAnd the real number part [ (p-1) stride +1, (p-1) stride + 3) of the first layer convolution layer]Multiplying and summing the values of x 1 x 64 nodes, and then multiplying and summingAnd the imaginary part [ (p-1) stride +1, (p-1) stride + 3) of the first layer convolution layer]The values of x 1 x 64 nodes are dot multiplied and summed, and the two results are subtracted to obtain the value of the (p,1,1) position of the second layer convolution layer. By analogy, for q 1, …,64 use is made ofAndthe value of the second layer convolutional layer (p,1, q) node is obtained. Without loss of generality, there is the following formula:

in this step, l is 2.

Formula (3)

Will be provided withAnd the real number part [ (p-1) stride +1, (p-1) stride + 3) of the first layer convolution layer]Multiplying and summing the multiplied nodes by 1 × 64And the imaginary part [ (p-1) stride +1, (p-1) stride + N) of the first convolution layer_r]Multiplying x 1 x 64 nodes by dot and summing the two results to obtain the convolution layer of the second layerThe value of the position. By analogy, for q 1, …,64 use is made ofAndobtaining a second layer of the convolution layerThe value of the node. Without loss of generality, there is the following formula:

in this step, l is 2. The node size of the resulting second convolutional layer is 2N_tX 1X 64, i.e

The length of the third convolution layer isWidth 1, height 128;

(3) second to third layer (conv.2layer):

the connection mode is shown in the left block diagram of FIG. 1, and the size of the single-channel convolution kernel is 3 × 1 × 64, and the number of convolution channels isQ is 1, …,128, and stride is N_rPer 4, each type of convolution kernel requires a shift using N_t/(N_r/4)＝4N_t/N_rLet p be 1,2, …,4N_t/N_r。

At the p-th convolution, equation (3) will be usedAnd the real part [ (p-1) stride +1, (p-1) stride + 3) of the second convolution layer]Multiplying and summing the values of x 1 x 64 nodes, and then multiplying and summingAnd the imaginary part [ (p-1) stride +1, (p-1) stride + 3) of the convolution layer of the second layer]The values of x 1 x 64 nodes are dot multiplied and summed, and the two results are subtracted to obtain the value of the (p,1,1) position of the third layer convolution layer. By analogy, for q 1, …,128, use is made ofAndthe value of the third layer convolutional layer (p,1, q) node is obtained.

Using the formula (4) willAnd the real part [ (p-1) stride +1, (p-1) stride + 3) of the second convolution layer]Multiplying and summing the multiplied nodes by 1 × 64And the imaginary part [ (p-1) stride +1, (p-1) stride + N) of the second convolution layer_r]Multiplying x 1 x 64 nodes by dot and summing the two results to obtain the third convolutional layerThe value of the position. By analogy, for q 1, …,128, use is made ofAndobtaining a third layer of the convolution layerThe value of the node.

The node size of the resulting third convolutional layer is 8N_t/N_rX 1X 128, i.e

The length, width and height of the fourth convolution layer are equal to the length, width and height of the third convolution layer.

(4) Third to fourth convolutional layers (conv.3layer):

the convolution is performed in the same manner as the previous step, wherein the size of the single-channel convolution kernel is 3 × 1 × 128. Let the number of convolution channels beq is 1, …,128, and if the step length stride is 1, then p is 1, …,4N_t/N_r. The node size of the fourth convolution layer obtained using equations (3) and (4) is 8N_t/N_rX 1X 128, i.e

The length of the full connecting layer isWidth 1, height 1;

(5) fourth layer (conv.4layer) to full-link layer (Linear 1layer):

spreading all nodes of the fourth layer of convolution layer in a height-width-length mode and arranging the nodes into oneA matrix of dimensions, i.e.

The length of the output layer isWidth 1, height 1;

(6) full connection layer (linear 1layer) to output layer (output layer):

connection mode of reference multiple full-connection network [1]. Due to F_RFAnd F_BBRespectively have N_tN_rAnda non-zero element, so that the size of the output layer is fixed

Wherein N is_tIs the number of transmit antennas; n is a radical of_rIs the number of receiving antennas;is the number of transmit chains; n is a radical of_sAnd transmitting the data code stream for the transmitting antenna.

TABLE 1 Complex convolution network layer parameter value summarization

Alternatively, referring to fig. 4, the width and length of the input layer are obtained by:

a1: respectively vectorizing a real part and an imaginary part of the channel matrix to obtain a real part vector and an imaginary part vector;

a2: connecting the real part vector and the imaginary part vector end to generate a new vector of the input layer; here, the generated new vector of the input layer is obtained by performing an operation in the input layer.

A3: and outputting the new vector as the length, width and height of the input layer.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.