WO2023035736A1 - Antenna system precoding method and apparatus based on two time scales and deep learning - Google Patents

Abstract

An antenna system precoding method and apparatus based on two time scales and deep learning. The method comprises: constructing a long-time-scale deep neural network (DNN) and a short-time-scale DNN, wherein the constructed DNNs each comprise a plurality of sub-networks corresponding to a transceiver of a large-scale millimeter wave multiple-input-multiple-output system; training the constructed DNNs by means of training data; acquiring a signal to be transmitted, performing high-dimensional pilot frequency estimation and high-dimensional channel feedback by means of a trained long-time-scale DNN, and performing low-dimensional pilot frequency estimation and low-dimensional channel feedback by means of the short-time-scale DNN; and performing analog precoding and digital precoding according to a high-dimensional original channel matrix and by means of the long-time-scale DNN, and performing digital precoding according to a low-dimensional equivalent channel matrix and by means of the short-time-scale DNN. By means of the method, the signaling overheads can be reduced, the robustness of a system can be improved, and modules in the system are jointly designed, thereby reducing the calculation complexity.

Description

Antenna system precoding method and device based on dual time scales and deep learning

Cross References to Related Applications

This application is based on a Chinese patent application with application number 202111056767.2 and a filing date of September 9, 2021, and claims the priority of this Chinese patent application. The entire content of this Chinese patent application is hereby incorporated by reference into this application.

technical field

The present application relates to the technical field of wireless communication, and in particular to a method and device for antenna system precoding based on dual time scales and deep learning.

Background technique

With the development of wireless networks, wireless data services grow explosively. In order to meet the ensuing challenges, the new generation of 5G communication network needs to provide larger bandwidth, higher spectral efficiency, and accommodate more users. The rise of 5G networks is accompanied by a significant increase in the number of users and the amount of data they transmit. Among them, mmWave communication is considered as one of the key technologies to meet the high data rate transmission requirements in 5G wireless network due to its huge bandwidth. The shorter wavelength of the millimeter wave allows the system to deploy a sufficient number of array antennas, among which the massive MIMO system can provide a large enough array gain for spatial multiplexing, thereby increasing system capacity and alleviating the shortage of radio spectrum. However, the massive MIMO system needs to be precoded in the application. However, traditional all-digital precoding needs to configure a radio frequency link for each antenna, which is costly and consumes a lot of energy.

In related technologies, in order to solve this problem, hybrid analog-digital precoding is usually used, that is, a large number of antennas are connected to fewer radio frequency links through a phase shifter. In addition, channel estimation and channel feedback are two important issues in hybrid precoding design. There are mainly two types of channel estimation methods: (1) directly estimate the channel itself, such as the least squares method; (2) estimate the channel parameters by compressive sensing method, and then restore the channel according to these parameters. Channel feedback schemes are mainly divided into two categories: (1) utilizing the space-time correlation of channel state information to reduce feedback overhead; (2) feedback schemes based on codebooks. For the above-mentioned hybrid precoding system, analog precoding and digital precoding need to be carefully designed to approach the performance of an all-digital precoding system.

However, the applicant found that in the above-mentioned technologies, most of the hybrid precoding systems design each module separately, and each module has a high complexity, such as high computational complexity, cannot be applied in real time, and requires accurate Issues such as mathematical modeling and poor robustness against environmental changes. Moreover, most of the existing hybrid precoding algorithms are based on high-dimensional instantaneous channels. In large-scale antenna scenarios, obtaining high-dimensional channel matrices will lead to huge signaling overhead, resulting in serious transmission delays and channel delays. lost pair.

Contents of the invention

This application aims to solve one of the technical problems in the related art at least to a certain extent.

To this end, the present application proposes an antenna system precoding method, device, electronic device and storage medium based on dual time scales and deep learning.

The embodiment of the first aspect of the present application proposes an antenna system precoding method based on dual time scales and deep learning, including the following steps:

A long-time-scale deep neural network DNN and a short-time-scale deep neural network DNN are constructed, wherein the long-time-scale DNN and the short-time-scale DNN respectively include multiple sub-networks corresponding to the transceivers of the massive millimeter-wave MIMO system. The multiple sub-networks include: a channel estimation sub-network and a channel feedback sub-network at the receiving end, and a pilot design sub-network and a hybrid precoding sub-network at the sending end;

Obtain training data with different signal-to-noise ratios, and use the training data to train the long-time scale DNN and the short-time scale DNN to optimize network parameters;

Obtain the signal to be transmitted, perform high-dimensional pilot estimation and high-dimensional channel feedback through the trained long-term scale DNN to restore the high-dimensional original channel matrix, and perform low-dimensional through the trained short-time scale DNN Pilot estimation and low-dimensional channel feedback to obtain a low-dimensional equivalent channel matrix;

Perform analog precoding and digital precoding according to the high-dimensional original channel matrix through the long-term scale DNN, and perform digital precoding through the short-time scale DNN according to the low-dimensional equivalent channel matrix to complete signal transmission .

Optionally, in an embodiment of the present application, the method further includes: dividing the time axis into multiple superframes according to channel statistical characteristics, and dividing each superframe into a first preset number of frames, each Each of the frames includes a second preset number of time slots, the long time scale is determined according to the superframe, and the short time scale is determined according to the time slots.

Optionally, in an embodiment of the present application, training the long-time-scale DNN and the short-time-scale DNN includes: alternately training the long-time-scale DNN and the short-time-scale DNN according to a dual-time-scale frame structure. Time-scale DNN, wherein the digital precoding matrix of the short-time-scale DNN is updated based on the low-dimensional equivalent channel in every time slot except the last time slot of each frame, and the long-time-scale DNN The analog precoding matrix and the digital precoding matrix of are updated based on the high-dimensional equivalent channel at the last time slot of each frame.

Optionally, in one embodiment of the present application, constructing the channel feedback sub-network according to the output of binary neurons in the deep neural network, the training of the long-term scale DNN and the short-time scale DNN also includes : the pilot information is set as the training parameters of the pilot design sub-network, and the target training parameters of the pilot design sub-network are learned by stochastic gradient descent; the gradient of the binary neuron is approximated by the estimator of the sigmoid function, and passed Stochastic gradient descent trains the channel feedback subnetwork.

Optionally, in one embodiment of the present application, the high-dimensional pilot is estimated by the following formula:

是接收端接收到的导频信号矩阵，

是发送端发送的训练导频，

是模拟预编码矩阵，

H是待估计的高维原始信道，

是模拟接收矩阵

的共轭转置矩阵，

N是高斯噪声矩阵，

选取所述训练导频、所述模拟预编码矩阵和所述模拟接收矩阵为所述信道估计子网络的训练参数。 in,

is the pilot signal matrix received by the receiver,

is the training pilot sent by the sender,

is the simulated precoding matrix,

H is the high-dimensional original channel to be estimated,

is the simulated receiving matrix

The conjugate transpose matrix of ,

N is the Gaussian noise matrix,

The training pilot, the simulated precoding matrix and the simulated receiving matrix are selected as training parameters of the channel estimation sub-network.

Optionally, in one embodiment of the present application, high-dimensional channel feedback is performed by the following formula:

表示导频信号矩阵

的向量化结果，

是向量

实部、虚部分开的表示，

是长时间尺度DNN的训练参数，σ _r是长时间尺度DNN第r层的非线性激活函数，sgn(·)是长时间尺度DNN二值层的激活函数。 where q is the feedback bit,

Represents the pilot signal matrix

The vectorized result of

is a vector

The real part and the imaginary part are separated,

is the training parameter of the long-time scale DNN, σ _r is the nonlinear activation function of the rth layer of the long-time scale DNN, and sgn( ) is the activation function of the binary layer of the long-time scale DNN.

Optionally, in one embodiment of the present application, the hybrid precoding subnetwork includes an analog transmitter precoding module, a digital transmitter precoding module, an analog receiver precode module, a digital receiver precode module, and a demodulation module , the analog precoding is performed according to the high-dimensional original channel, including:

The real part and the imaginary part of the high-dimensional original channel matrix are respectively input to the analog transmitter precoding module and the analog receiver precoder module to output the analog encoder phases of the transmitter and the receiver; Constrained complex vectors; the complex vectors satisfying the constant modulus constraints are converted by the following formula to generate an analog precoding matrix:

in,

表示将向量转换成矩阵的操作，

是发送端的模拟编码器相位，

是接收端的模拟编码器相位，N _t是发送端天线的数目，N _r是接收端天线的数目。 Among them, F _RF is the analog precoding matrix of the transmitting end, W _RF is the analog precoding matrix of the receiving end,

Represents the operation of converting a vector into a matrix,

is the phase of the analog encoder at the transmitter,

is the analog encoder phase at the receiver, N _t is the number of antennas at the transmitter, and N _r is the number of antennas at the receiver.

Optionally, in one embodiment of the present application, the training data includes high-dimensional original channel matrix samples, Gaussian noise and data labels to be sent, and the training of the long-term DNN further includes: moving average Or the way of sliding window to obtain high-dimensional original channel matrix samples, the moving average is the current output of the analog transmitter precoding module and the analog receiver precoding module of the long-term scale DNN and the last moment closest to the current The output is weighted average.

The embodiment of the second aspect of the present application proposes an antenna system precoding device based on dual time scales and deep learning, including the following modules:

A building block for constructing a long-time-scale deep neural network DNN and a short-time-scale deep neural network DNN, wherein the long-time-scale DNN and the short-time-scale DNN respectively include transceivers corresponding to massive millimeter-wave multiple-input multiple-output MIMO systems A plurality of subnetworks, the plurality of subnetworks include: a channel estimation subnetwork and a channel feedback subnetwork at the receiving end, and a pilot design subnetwork and a hybrid precoding subnetwork at the sending end;

A training module, configured to obtain training data with different signal-to-noise ratios, and train the long-time scale DNN and the short-time scale DNN through the training data to optimize network parameters;

The acquisition module is used to acquire the signal to be transmitted, perform high-dimensional pilot estimation and high-dimensional channel feedback through the long-term scale DNN completed through training, so as to restore the high-dimensional original channel matrix, and complete the short-term DNN through training Scale DNN performs low-dimensional pilot estimation and low-dimensional channel feedback to obtain low-dimensional equivalent channel matrix;

A coding module, configured to perform analog precoding and digital precoding according to the high-dimensional original channel matrix through the long-time scale DNN, and perform digital precoding through the short-time scale DNN according to the low-dimensional equivalent channel matrix , to complete the signal transmission.

The embodiment of the third aspect of the present application proposes an electronic device, including: a memory, a processor, and a computer program stored on the memory and operable on the processor. When the processor executes the program, the above first The steps in the antenna system precoding method based on dual time scales and deep learning shown in the aspect.

The embodiment of the fourth aspect of the present application proposes a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores a computer program; when the computer program is executed by a processor, the above-mentioned An antenna system precoding method based on dual time scales and deep learning is shown in one aspect.

The technical solutions provided by the embodiments of the present application bring at least the following beneficial effects: the present application performs hybrid precoding based on dual time scales, wherein the analog precoding of the long time scale is obtained based on the channel statistical characteristics, and the digital precoding of the short time scale is obtained according to the low Dimensional real-time equivalent channel matrix is optimized, which can reduce signaling overhead and improve the robustness to channel mismatch caused by transmission delay. Moreover, the method jointly designs each module in the communication system through a deep learning framework to achieve end-to-end performance optimization. Statistical properties do not require precise channel mathematical models, and the calculation of deep neural networks can be parallelized, which improves the communication performance of massive MIMO systems and reduces the computational complexity of hybrid precoding. In addition, based on the dual-time-scale method of the present application, the present application only needs to estimate the low-dimensional equivalent channel matrix most of the time in practical applications, so that the channel error caused by the transmission delay can be greatly reduced.

Additional aspects and advantages of the disclosure will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the disclosure.

Description of drawings

The above and/or additional aspects and advantages of the present application will become apparent and easy to understand from the following description of the embodiments in conjunction with the accompanying drawings, wherein:

FIG. 1 is a schematic flowchart of an antenna system precoding method based on dual time scales and deep learning proposed in an embodiment of the present application;

FIG. 2 is a schematic structural diagram of a hybrid precoding millimeter-wave MIMO system proposed in an embodiment of the present application;

FIG. 3 is a communication schematic diagram of a hybrid precoding millimeter-wave MIMO system proposed in an embodiment of the present application;

FIG. 4 is a schematic structural diagram of a dual-time-scale frame proposed in an embodiment of the present application;

FIG. 5 is a schematic structural diagram of an end-to-end dual-time-scale DNN for a millimeter-wave MIMO system proposed in an embodiment of the present application;

FIG. 6 is a schematic diagram of a hybrid precoding design framework and its data transmission proposed by the embodiment of the present application;

FIG. 7 is a schematic diagram of a sliding window with a cache capacity of 3 proposed by an embodiment of the present application.

FIG. 8 is a schematic structural diagram of a frame structure of a centralized training DNN proposed in an embodiment of the present application;

FIG. 9 is a schematic structural diagram of a frame structure of a distributed training DNN proposed in an embodiment of the present application;

FIG. 10 is a schematic diagram of a comparison of the bit error rates of a dual-time-scale DNN proposed in the embodiment of the present application and a traditional solution under different signal-to-noise ratios;

FIG. 11 is a schematic diagram of a comparison of the bit error rate of a dual-time-scale DNN proposed in the embodiment of the present application and a traditional solution under different feedback bit numbers;

FIG. 12 is a schematic diagram of a comparison of bit error rates of a dual-time-scale DNN proposed in the embodiment of the present application and a traditional solution under different pilot lengths;

FIG. 13 is a schematic diagram of a communication process of an end-to-end dual-scale millimeter-wave MIMO system proposed in an embodiment of the present application;

FIG. 14 is a schematic structural diagram of an antenna system precoding device based on dual time scales and deep learning proposed by an embodiment of the present application.

Detailed ways

Embodiments of the present disclosure are described in detail below, examples of which are illustrated in the drawings, in which the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the figures are exemplary and are intended to explain the present disclosure and should not be construed as limiting the present disclosure.

At present, the existing dual-time-scale algorithms have high computational complexity, and cannot jointly design each module in the communication system. Although optimization algorithms can be used in related technologies, there are still problems such as high computational complexity, impossibility of real-time application, precise mathematical modeling of problems, poor robustness against environmental changes, and differences between communication modules. design issues. However, this application uses deep learning technology to carry out joint design of each module in a massive MIMO system. Compared with optimization algorithms, deep learning has lower computational complexity, does not require precise mathematical modeling of problems, and is robust against errors. Preferably, the communication modules can be jointly designed, and the end-to-end deep learning framework is suitable for the joint design of each module in this system. First of all, unlike the separate design of each module of the traditional communication system, the deep learning framework can jointly design all modules to achieve end-to-end performance optimization. Second, the deep learning framework implicitly learns the statistical properties of the channel in a data-driven manner during the process of optimizing an end-to-end communication system, without requiring an accurate mathematical model of the channel. Furthermore, the calculation of DNN can be parallelized, and the computational complexity is much lower than that of traditional optimization algorithms.

The antenna system precoding method and device based on dual time scales and deep learning proposed by the embodiments of the present disclosure are described below with reference to the accompanying drawings.

Figure 1 is a schematic flowchart of a method for precoding an antenna system based on dual time scales and deep learning proposed in the embodiment of the present application. As shown in Figure 1, the method includes the following steps:

Step 101, constructing a long-time-scale deep neural network DNN and a short-time-scale deep neural network DNN, wherein the long-time-scale DNN and the short-time-scale DNN respectively include a plurality of sub-transceivers corresponding to a large-scale millimeter-wave multiple-input multiple-output MIMO system The multiple subnetworks include: a channel estimation subnetwork and a channel feedback subnetwork at the receiving end, and a pilot design subnetwork and a hybrid precoding subnetwork at the sending end.

Specifically, the application first constructs a deep neural network (DNN for short), and divides the DNN into two parts based on a dual time scale, that is, constructing a long-term scale deep neural network DNN and a short-time scale deep neural network DNN. Among them, the long-term DNN is used for high-dimensional pilot estimation and high-dimensional original channel feedback, and the analog precoding and digital precoding are optimized according to the recovered high-dimensional channel, and the short-time-scale DNN is used for low-dimensional pilot estimation and The low-dimensional equivalent channel is fed back, and the digital precoding is optimized according to the recovered low-dimensional channel.

Wherein, each of the long-time-scale DNN and the short-time-scale DNN encapsulates all modules of a transceiver of a large-scale millimeter wave multiple-in multiple-out (MIMO for short) system, that is, the long-time scale DNN And the short-time-scale DNN includes sub-networks equivalent to channel estimation and channel feedback at the receiving end of MIMO systems, as well as pilot design and hybrid precoding sub-networks at the sending end. That is to say, the hybrid precoding millimeter-wave MIMO system after deploying the DNN of this application performs channel estimation, channel quantization, and channel feedback based on DNN at the receiving end, and performs pilot design and hybrid precoding optimization at the transmitting end. In other words, the received pilot signal is mapped to feedback bits at the receiving end, and then the received feedback bits are mapped to a hybrid precoding matrix at the sending end.

In order to more clearly illustrate the structure of the deep neural network constructed by the present application, the hybrid precoding millimeter-wave MIMO system of the present application will be described in detail below with reference to FIG. 2 .

条射频链路，发送N _s个数据流至接收端，其中，

接收端(可以是用户端)装配有N _r根接收天线和

条射频链路，其中，

在发送端，射频链路连接着一个移相器网络，将

个数字输出信号变为N _t个经过编码的模拟信号。类似地，在接收端，N _r根接收天线连接着一个移相器网络以及

条射频链路。 As shown in Figure 2, the transmitting end of the system is equipped with N _t transmitting antennas and

RF links, sending N _s data streams to the receiving end, where,

The receiving end (could be the user end) is equipped with N _r receiving antennas and

RF links, where,

At the transmit end, the RF link is connected to a network of phase shifters that will

digital output signals into N _t encoded analog signals. Similarly, at the receiving end, N _r receiving antennas are connected to a network of phase shifters and

radio frequency link.

由N _s×log ₂M维的0-1比特组成。然后，这些数据根据M维调制方式被映射为符号

符号向量s满足

这些符号先经过数字发送预编码

的处理，再经过模拟预编码

经过编码的信号可以写成x＝F _RFF _BBs。其中F _RF表示仅能调整相位，并通过移相器网络实现的模拟预编码矩阵，因此需要符合恒模约束

数字预编码矩阵F _BB需要经过功率归一化

使得发送端的功率约束满足，其中，P _T表示最大的传输功率。经过预编码的信号x经过一个窄带快衰落信道，接收端接收到的符号矢量

可以表示成z＝HF _RFF _BBs+n，其中

表示信道矩阵，

为加性高斯噪声。类似地，在接收端，接收信号需要经过接收端模拟编码

以及数字编码

的处理。检测到的信号可以写成

其中W _RF需要满足恒模约束

最终，检测到的信号r被解调用于恢复N _s个数据流，产生恢复比特

Among them, the sender transmits N _s parallel data

It consists of 0-1 bits of N _s ×log ₂ M dimensions. Then, these data are mapped into symbols according to the M-dimensional modulation scheme

The symbolic vector s satisfies

These symbols are first digitally sent pre-encoded

processing, and then simulated pre-encoding

The encoded signal can be written as x=F _RF F _BB s. Among them, F _RF represents the analog precoding matrix that can only adjust the phase and is realized through the phase shifter network, so it needs to meet the constant modulus constraint

The digital precoding matrix F _BB needs to be normalized by power

So that the power constraint of the sending end is satisfied, where _PT represents the maximum transmission power. The precoded signal x passes through a narrowband fast fading channel, and the symbol vector received by the receiving end

Can be expressed as z=HF _RF F _BB s+n, where

represents the channel matrix,

is additive Gaussian noise. Similarly, at the receiving end, the received signal needs to be analog encoded at the receiving end

and digital code

processing. The detected signal can be written as

where W _RF needs to satisfy the constant modulus constraint

Finally, the detected signal r is demodulated to recover the N _s data streams, yielding recovered bits

Further, the communication process of the system will be described below with reference to FIG. 3 . As shown in Figure 3, the communication process of the system includes channel estimation, channel feedback and hybrid precoding. It should be noted that, in the embodiment of the present application, the base station is used as the sending end, and the sending end needs to obtain the channel matrix H to perform hybrid precoding. Therefore, before transmitting data, it is necessary to send pilots to estimate the channel.

然后接收端接收到的导频信号

其中，

和

分别表示模拟发送导频和模拟接收导频，它们的列从离散傅里叶变换(DFT)矩阵选出以满足恒模约束，

表示高斯白噪声矩阵，

在实际应用中，发送端根据时间顺序依次发送导频矩阵中的导频，且发送导频矩阵的第l次传输

(

的第l列，在本申请实施例中传输次数与导频矩阵的列数一一对应)需要满足功率约束，

接收端从接收到的信号

当中估计出信道H，并从中提取出有用的信息，比如，到达角度、信道增益等，然后将这些信息压缩为B比特反馈给发送端

其中，B是根据实际需要预设的比特量，映射

表示反馈方案。 Specifically, the sending end first sends a pilot matrix of length L

Then the pilot signal received by the receiver

in,

and

denote the analog transmit pilot and the analog receive pilot, respectively, whose columns are selected from the discrete Fourier transform (DFT) matrix to satisfy the constant modulus constraint,

Represents the Gaussian white noise matrix,

In practical applications, the sending end sends the pilots in the pilot matrix sequentially according to the time sequence, and sends the lth transmission of the pilot matrix

(

The lth column, in the embodiment of the present application, the number of transmissions corresponds to the number of columns of the pilot matrix) needs to meet the power constraint,

The signal received by the receiver from

The channel H is estimated, and useful information is extracted from it, such as angle of arrival, channel gain, etc., and then the information is compressed into B bits and fed back to the sending end

Among them, B is the amount of bits preset according to actual needs, and the mapping

Indicates a feedback scheme.

It should be noted that in the above system, the joint design of digital-analog hybrid precoding for each high-dimensional original channel requires a lot of overhead, and the computational complexity and hardware cost are relatively high. To this end, this application proposes a dual-time-scale scheme to simultaneously consider the high-dimensional original channel and channel statistical properties.

In one embodiment of the present application, the time axis is divided into a plurality of superframes according to channel statistical characteristics, and each superframe is divided into a first preset number of frames, each frame includes a second preset number of time slots, The long time scale is determined in terms of superframes, and the short time scale is determined in terms of time slots. The following will be described in detail in conjunction with Figure 4

As shown in FIG. 4 , in the embodiment of the present application, for a specific superframe, it is assumed that the channel statistical characteristics are constant during this period. The superframe is composed of T _f frames, and each frame is divided into T _s time slots. In this example, the first preset number is T _f , and the second preset number is T _s . High-dimensional The original channel remains unchanged within each slot. Based on this division, the long-term scale in the embodiment of the present application is that the statistical properties of the channel are fixed in each superframe, a superframe contains T _f frames, and the short-time scale high-dimensional original channel H It is assumed to be fixed at each time slot.

的维度远低于高维原始信道矩阵H，因此，在本申请实施例中通过发送导频，在每个时隙获得低维等效信道矩阵H _eq，并且，由于在大规模MIMO场景中，在每个时隙获取实时的高维原始信道矩阵H会导致大量的信令开销，所以本申请在每帧中仅获取一个高维原始信道矩阵样本H。因此，本申请实施例中假设接收端能够在每一帧获取一个完整的高维原始信道矩阵样本，并且能够在每个时隙获取实时的低维等效信道矩阵H _eq，模拟和数字预编码矩阵需要分别基于H和H _eq，在不同的时间尺度进行优化。如图3所示，长时间尺度的模拟预编码矩阵{F _RF,W _RF}在每帧的最后一个时刻，基于估计得到的高维原始信道矩阵H进行更新，以实现多天线阵列增益。而短时间尺度的数字预编码矩阵{F _BB,W _BB}在每个时隙中，基于估计得到的低维等效信道矩阵H _eq进行更新，以实现空间复用增益，而短时间尺度的更新数字预编码矩阵时，{F _RF,W _RF}固定不变。由此，通过该双时间尺度的方法对混合预编码矩阵进行更新时，在大多数时间只需要估计低维等效信道矩阵，从而可以大大降低由于传输时延导致的信道误差。 It should be noted that, due to the equivalent channel matrix

The dimension of is much lower than the high-dimensional original channel matrix H, therefore, in the embodiment of the present application, by sending pilots, the low-dimensional equivalent channel matrix _Heq is obtained in each time slot, and, because in the massive MIMO scenario, Obtaining a real-time high-dimensional original channel matrix H in each time slot will cause a large amount of signaling overhead, so this application only obtains one high-dimensional original channel matrix sample H in each frame. Therefore, in the embodiment of this application, it is assumed that the receiving end can obtain a complete high-dimensional original channel matrix sample in each frame, and can obtain a real-time low-dimensional equivalent channel matrix _Heq in each time slot, analog and digital precoding The matrices need to be optimized at different time scales based on H and _Heq respectively. As shown in Figure 3, the long-term scale analog precoding matrix {F _RF , W _RF } is updated at the last moment of each frame based on the estimated high-dimensional original channel matrix H to achieve multi-antenna array gain. The short-time-scale digital precoding matrix {F _BB , W _BB } is updated based on the estimated low-dimensional equivalent channel matrix _Heq in each time slot to achieve spatial multiplexing gain, while the short-time-scale digital precoding matrix When updating the digital precoding matrix, {F _RF , W _RF } is fixed. Therefore, when the hybrid precoding matrix is updated by the dual-time-scale method, it is only necessary to estimate the low-dimensional equivalent channel matrix most of the time, so that the channel error caused by the transmission delay can be greatly reduced.

Step 102, acquiring training data with different signal-to-noise ratios, and using the training data to train the long-time-scale DNN and the short-time-scale DNN, so as to optimize network parameters.

Among them, in the training phase, the purpose of offline training of the long-time scale DNN and the short-time scale DNN is to obtain the training parameters of each sub-network in the DNN through training data with different signal-to-noise ratios, and optimize the performance of the DNN through the obtained training parameters. Network parameters, which are convenient for subsequent predictions by fixing and optimizing network parameters in practical applications.

In an embodiment of the present application, the training data with different signal-to-noise ratios includes training samples {H,n,S _b }, that is, high-dimensional original channel matrix samples, Gaussian noise, and data labels to be sent. In practical applications, training data can be obtained in different ways.

As a possible implementation, the historical data can be obtained according to the actual application scenario of the system, and the historical data generated during the communication of the massive millimeter-wave MIMO system pre-stored in the database can be used as the training data.

As another possible implementation, the channel of the scale millimeter-wave MIMO system can be modeled, and training data can be obtained in real time according to the established channel model. For example, a narrowband millimeter wave channel model is established, the narrowband millimeter wave channel model includes N _cl clusters, and each cluster includes N _ray propagation paths. Each path includes the sending and receiving directions of the channel (sending angle, arrival angle), and path complex gain. The channel matrix is expressed as:

为第i簇中第l条路径的复增益，

和

分别表示接收端和发送端的到达角和发送角。

和

分别表示接收和发送导向矢量。对于一个包含N根天线的线性阵列和角度φ， 导向矢量可以写作： in,

is the complex gain of the l-th path in the i-th cluster,

and

are the angles of arrival and angles of arrival at the receiver and transmitter, respectively.

and

denote the receive and transmit steering vectors, respectively. For a linear array of N antennas and angle φ, the steering vector can be written as:

Among them, d and λ represent the distance between adjacent antennas and the wavelength of the carrier, respectively. After the channel model is established, it is assumed that the channel and noise have a certain distribution, correspondingly generate channel and noise samples according to statistical characteristics, and collect the generated channel and noise samples. Thus, training data can be generated in real time through the established channel model.

Further, after obtaining the training data, in the embodiment of the present application, the training parameters can be iteratively updated through stochastic gradient descent (SGD) with binary cross-entropy (BCE) as the target loss function. When using BCE as the loss function of the training phase of the embodiment of the present application, the binary cross entropy can be expressed by the following formula:

表示训练数据集，S _b表示传输符号矩阵，由维度为N _s×log ₂M的0-1比特组成。

代表恢复出的符号矩阵，表示传输的比特为1的概率。由于

是DNN的输出，可以表示为DNN训练参数的函数，本申请中最大化BCE相当于最大化可达速率。 in,

Represents the training data set, S _b represents the transmission symbol matrix, which consists of 0-1 bits with a dimension of N _s ×log ₂ M.

Represents the recovered symbol matrix, indicating the probability that the transmitted bit is 1. because

is the output of DNN, which can be expressed as a function of DNN training parameters. In this application, maximizing BCE is equivalent to maximizing the attainable rate.

The bit-error rate (BER) of the training set can be expressed as:

时，

否则，

Among them, when

hour,

otherwise,

Furthermore, since this application adopts a dual-time-scale precoding method, when performing the above stochastic gradient descent training in the embodiment of this application, the long-time-scale DNN and the short-time-scale DNN can be alternately trained according to the frame structure of the dual-time scale , where the digital precoding matrix of the short-time scale DNN is updated based on the low-dimensional equivalent channel in each frame except the last slot, and the analog precoding matrix and digital precoding matrix of the long-time scale DNN The matrix is updated based on the high-dimensional equivalent channel at the last slot of each frame.

短时间尺度DNN输出数字预编码矩阵{F _BB,W _BB}。在每一帧的最后一个时隙，发送的是导频

长时间尺度DNN输出混合预编码(包括数字、模拟)矩阵{F _BB,F _RF,W _BB,W _RF}。 Specifically, a short-time-scale DNN is trained for the first T _s −1 slots of each frame, with the input {H,F _RF ,W _RF ,n,S _b }, where the simulated precoding matrix {F _RF ,W _RF } is computed by a long-time scale DNN. In the last time slot of each frame, the long-term scale DNN is trained, and the input is {H,n,S _b }, that is, the long-time scale DNN is trained once in each frame, and the short-time scale DNN is trained once in each time slot. The two are trained alternately until convergence. In the prediction stage of subsequent practical applications, the principle of DNN output hybrid precoding is the same as above. Specifically, in the first T _s -1 time slots of each frame, the pilot

The short-time scale DNN outputs a digital precoding matrix {F _BB , W _BB }. In the last slot of each frame, the pilot is sent

The long-term scale DNN outputs a mixed precoding (including digital and analog) matrix {F _BB , F _RF , W _BB , W _RF }.

In order to more clearly describe the training scheme of the long-term DNN and the short-time DNN in this application, the following is a detailed description of each training phase in DNN in combination with Figure 5 and Figure 6, including: pilot training phase, channel feedback training stage and hybrid precoding design training stage. Among them, the training carried out in each training stage can be regarded as the training of the corresponding sub-network. For example, in the pilot training stage, the pilot design sub-network is trained, and the channel feedback training stage is the channel estimation sub-network. Network and channel feedback sub-network for training and so on.

Specifically, as shown in Figure 5, the end-to-end dual-time-scale DNN for mmWave MIMO systems includes a long-time scale DNN10 and a short-time scale DNN20. In the hybrid precoding design stage, DC-NN and DP-NN can Shared learning parameters. First, in the pilot training phase, the receiver needs to estimate the low-dimensional equivalent channel matrix H _eq in the first T _s -1 slots of each frame, and estimate the high-dimensional original channel matrix H in the last slot of each frame .

以及模拟预编码矩阵(在信道估计阶段称作模拟导频矩阵)

其中L表示导频长度。然后，接收到的导频信号矩阵经过模拟接收矩阵

表示为

其中，

表示高斯噪声矩阵。即在本申请实施例中对长时间尺度DNN进行导频训练时，长时间尺度DNN的输入和输出分别为H和

其中，H为获取的训练数据。 Among them, when performing pilot training on long-term scale DNN, in order to estimate the high-dimensional original channel matrix H, the sending end sends training pilot

And the analog precoding matrix (called the analog pilot matrix in the channel estimation stage)

Where L represents the pilot length. Then, the received pilot signal matrix goes through the simulated receiving matrix

Expressed as

in,

Represents a Gaussian noise matrix. That is, when the pilot training is performed on the long-term scale DNN in the embodiment of the present application, the input and output of the long-time scale DNN are H and

Among them, H is the acquired training data.

代替DNN网络中的参数作为训练参数，即在本申请中将上述导频信息设为导频设计子网络训练参数，相比较采用DNN网络中的参数作为训练参数，本申请采用的高斯导频以及

等通过DFT矩阵中选出来的导频，训练得到的导频参数

可以实现更好的信道估计性能，更适配信道统计特性。并且，为了保证模拟导频

满足恒模约束，在本申请实施例中将这两个矩阵的每一个元素设为训练参数的同时，在每次训练完成后除以自身的模

此外，为了确保导频矩阵满足功率约束，本申请实施例还通过放缩

使得

其中

为第l次传输的导频，即矩阵

的第l列。 It should be noted that, in order to design the best pilot that adapts to the statistical characteristics of the current channel, so as to make the estimation of the high-dimensional original channel matrix H more accurate, the parameter set is selected in the embodiment of this application

Instead of the parameters in the DNN network as training parameters, that is, in this application, the above-mentioned pilot information is set as the pilot design subnetwork training parameters. Compared with using the parameters in the DNN network as training parameters, the Gaussian pilot used in this application and

Waiting for the pilot selected from the DFT matrix, the pilot parameters obtained by training

It can achieve better channel estimation performance and better adapt to channel statistical characteristics. And, to ensure that the simulated pilot

To satisfy the constant modulus constraint, in the embodiment of this application, each element of the two matrices is set as a training parameter, and after each training is completed, it is divided by its own modulus

In addition, in order to ensure that the pilot matrix satisfies the power constraint, the embodiment of the present application also scales

make

in

is the pilot frequency of the lth transmission, that is, the matrix

column l of .

接收端接收到

其中

为高斯噪声矩阵。本申请的短时间尺度DNN的输入和输出分别为H _eq和

其中，H _eq为获取的训练数据。 When performing pilot training on a short-time scale DNN, in order to estimate the low-dimensional equivalent channel matrix _Heq , the sender sends the training pilot matrix

Receiver receives

in

is a Gaussian noise matrix. The input and output of the short-time scale DNN in this application are _Heq and

Among them, _Heq is the acquired training data.

在信道估计阶段，和长时间尺度DNN不同，短时间尺度DNN的模拟编码器{F _RF,W _RF}并非通过训练得到的，而是和上一帧在数据传输阶段使用的模拟编码器相同。即短时间尺度DNN估计的是等效信道

模拟编码器{F _RF,W _RF}为等效信道的一部分，并且，本申请还通过缩放

使得导频满足功率约束。 It should be noted that, in order to design the best pilot suitable for the current channel statistical characteristics, so as to make the estimation of the equivalent channel matrix H _eq more accurate, the application sets the training parameters as

In the channel estimation stage, unlike the long-time-scale DNN, the analog encoder {F _RF , W _RF } of the short-time-scale DNN is not obtained through training, but is the same as the analog encoder used in the data transmission stage of the previous frame. That is, the short time scale DNN estimates the equivalent channel

The analog encoder {F _RF ,W _RF } is part of the equivalent channel, and this application also scales

Such that the pilot satisfies the power constraint.

Secondly, in the channel feedback training phase, in the first T _s -1 time slots of each frame, the receiving end feeds back the low-dimensional equivalent channel matrix _Heq after quantization, and in the last time slot of each frame, the receiving end feeds back The high-dimensional original channel matrix H after quantization.

估计高维原始信道矩阵H。然后，接收端从中提取出有用的信息并将其量化为B比特反馈给发送端，用于后续混合预编码设计。这两步可以用一个R层的全连接DNN实现，即接收端的反馈比特可以通过以下公式进行表示： Among them, when performing channel feedback training on the long-term scale DNN, in the last time slot of each frame, first, the receiving end based on the received pilot signal matrix

Estimate the high-dimensional original channel matrix H. Then, the receiving end extracts useful information from it and quantizes it into B bits to feed back to the sending end for subsequent hybrid precoding design. These two steps can be implemented with an R-layer fully connected DNN, that is, the feedback bits at the receiving end can be expressed by the following formula:

表示导频信号矩阵

的向量化结果，DNN的输入为向量

实部、虚部分开的表示

表示训练参数，σ _r代表第r层的非线性激活函数。符号函数sgn(·)为最后一层(二值层)的激活函数，用于产生反馈比特向量q(q的每一个元素取值都为0或1)。 in,

Represents the pilot signal matrix

The vectorization result of DNN, the input of DNN is a vector

Representation with real and imaginary parts separated

Indicates the training parameters, and _σr represents the nonlinear activation function of the rth layer. The sign function sgn(·) is the activation function of the last layer (binary layer), which is used to generate the feedback bit vector q (each element of q takes the value of 0 or 1).

估计出低维等效信道矩阵H _eq，并从中提取出有用的信息，并将其量化为B _eq比特反馈给发送端，用于后续数字预编码设计。这两步可以用一个R _eq层的全连接DNN实现，即接收端的反馈比特可以通过以下公式进行表示： When performing channel feedback training on a short-time-scale DNN, the feedback process of the low-dimensional equivalent channel matrix _Heq is similar to the feedback process of the above-mentioned long-time-scale DNN. Specifically, in the first T _s -1 time slots of each frame, the receiving end first bases on the received pilot signal matrix

Estimate the low-dimensional equivalent channel matrix _Heq , extract useful information from it, quantize it into _Beq bits and feed it back to the transmitter for subsequent digital precoding design. These two steps can be implemented with a fully connected DNN at the _Req layer, that is, the feedback bits at the receiving end can be expressed by the following formula:

表示矩阵

的向量化结果，DNN的输入为向量

实部、虚部分开的表示

表示DNN的训练参数，符号函数sgn(·)为最后一层(二值层)的激活函数，用于产生反馈比特向量

(q _eq的每一个元素取值都为0或1)，由此，针对二进制进制神经元的离散输出，通过sigmoid函数的估计器近似二进制神经元的梯度，便于后续通过随机梯度下降训练信道反馈子网络，使基于梯度的训练成为可能。而由于反馈向量q _eq的维度要远小于q的维度，因为H _eq的维度远小于H的维度，因此，在本申请实施例中可以使用更少层数和参数数量的DNN来获得q _eq。 in,

representation matrix

The vectorization result of DNN, the input of DNN is a vector

Representation with real and imaginary parts separated

Represents the training parameters of DNN, and the sign function sgn( ) is the activation function of the last layer (binary layer), which is used to generate the feedback bit vector

(Each element of q _eq has a value of 0 or 1). Therefore, for the discrete output of the binary neuron, the gradient of the binary neuron is approximated by the estimator of the sigmoid function, which is convenient for the subsequent stochastic gradient descent training channel Feedback sub-networks, enabling gradient-based training. Since the dimension of the feedback vector q _eq is much smaller than the dimension of q, because the dimension of He _eq is much smaller than the dimension of H, therefore, in the embodiment of the present application, DNN with fewer layers and parameters can be used to obtain q _eq .

Furthermore, in the hybrid precoding design training phase, in the first T _s -1 time slots of each frame, this application uses a short time scale DNN based on q _eq to update the digital precoding matrix {F _BB , W _BB }, in At the last slot of each frame, the digital and analog precoding matrices {F _RF , F _BB , W _RF , W _BB } are updated using a long-term scale DNN based on q.

然后，发送端基于恢复出的

用DNN设计混合预编码矩阵{F _RF,F _BB,W _RF,W _BB}。如图5所示，该DNN包括了5个子网络，模拟发送端预编码网络(analog precoder NN，AP-NN)，数字发送端预编码网络(digital precoder NN，DP-NN)，模拟接收端预编码网络(analog combiner NN，AC-NN)，数字接收端预编码网络(digital combiner NN，DC-NN)，以及解调网络。具体而言，在本申请一个实施例中，先将恢复出的信道矩阵

的实部、虚部分开存储，成为一个实数矩阵，然后，分别将实部和虚部输入AP-NN和AC-NN，通过AP-NN和AC-NN分别输出发送端和接收端的模拟编码器相位

和

由此可以通过以下公式计算出满足恒模约束的复数向量： Among them, when performing hybrid precoding design training on long-term scale DNN, in the last time slot of each frame, the sender receives the feedback bit q to restore the high-dimensional original channel matrix

Then, based on the recovered

Design hybrid precoding matrix {F _RF , F _BB , W _RF , W _BB } with DNN. As shown in Figure 5, the DNN includes 5 sub-networks, the analog precoder network (analog precoder NN, AP-NN), the digital transmitter precoder network (digital precoder NN, DP-NN), and the analog receiver precoder network. Coding network (analog combiner NN, AC-NN), digital receiver precoding network (digital combiner NN, DC-NN), and demodulation network. Specifically, in one embodiment of the present application, the recovered channel matrix is first

The real part and imaginary part are stored separately to become a real number matrix. Then, the real part and imaginary part are input into AP-NN and AC-NN respectively, and the analog encoders at the sending end and receiving end are respectively output through AP-NN and AC-NN. phase

and

From this, the complex vector that satisfies the constant modulus constraint can be calculated by the following formula:

Then, the simulated precoding matrix is generated by the following formula:

表示将向量转换成矩阵的操作，N _t是发送端天线的数目，N _r是接收端天线的数目。进而，根据原始信道矩阵

和得到的模拟预编码矩阵{F _RF,W _RF}计算出低维等效信道矩阵

in,

Indicates the operation of converting a vector into a matrix, N _t is the number of antennas at the transmitting end, and N _r is the number of antennas at the receiving end. Furthermore, according to the original channel matrix

and the obtained simulated precoding matrix {F _RF , W _RF } to calculate the low-dimensional equivalent channel matrix

实部、虚部分开存储，成为一个实数矩阵，输入DP-NN和DC-NN，分别输出发送端和接收端的数字编码器

(实部虚部分开存储)。则可以通过以下公式生成数字预编码矩阵：

再通过下述公式进行功率归一化，以得到最终的数字预编码矩阵，同时保证功率约束满足： Further, referring to the above process, the equivalent channel matrix

The real part and imaginary part are stored separately and become a real matrix, which is input into DP-NN and DC-NN, and output to the digital encoders at the sending end and receiving end respectively

(Real and imaginary parts are stored separately). Then the digital precoding matrix can be generated by the following formula:

Then perform power normalization by the following formula to obtain the final digital precoding matrix, while ensuring that the power constraints are satisfied:

然后，基于恢复出来的信道矩阵

发送端设计数字预编码矩阵{F _BB,W _BB}，此时模拟预编码矩阵{F _RF,W _RF}固定不变(直接采用上一帧最后一个时隙，长时间尺度DNN计算得到的模拟预编码矩阵)。如图5所示，短时间尺度DNN包含DP-NN和DC-NN，分别 用于产生发送端和接收端数字预编码矩阵F _BB和W _BB，其产生矩阵的原理可参照上述对长时间尺度DNN进行混合预编码设计训练时的步骤，此处不再赘述。 Among them, when performing hybrid precoding design training on short-time-scale DNN, in the first T _s -1 time slots of each frame, the sender receives feedback bits q _eq , which are used to restore the low-dimensional equivalent channel matrix

Then, based on the recovered channel matrix

The sending end designs the digital precoding matrix {F _BB , W _BB }, and the analog precoding matrix {F _RF , W _RF } is fixed at this time (directly use the last time slot of the previous frame, the simulation obtained by the long-term scale DNN precoding matrix). As shown in Figure 5, the short-time-scale DNN includes DP-NN and DC-NN, which are used to generate digital precoding matrices F _BB and W _BB at the transmitter and receiver respectively. The principle of generating the matrix can refer to the above-mentioned long-term scale The steps of DNN performing hybrid precoding design training will not be repeated here.

进而，根据S _b和

通过上述的最小化端到端二元交叉熵，并通过随机梯度下降迭代更新DNN的训练参数的方式完成训练。 During the training process, the DNN signal flow is shown in the signal flow chart in Figure 5. The whole process simulates the sending signal S _b through the mixed precoding {F _RF , F _BB } at the sending end, channel fading H, noise n, receiving end Hybrid precoding {W _RF , W _BB }, the process of recovering the signal at the receiving end. In this application, the training sample {H,n,S _b } is input into DNN to generate a mixed precoding matrix, and finally the received signal r is obtained. The real part and imaginary part of the received signal r are separated, r is converted into a real-valued vector, and input to the demodulation network to generate a restored signal

Furthermore, according to S _b and

The training is completed by minimizing the end-to-end binary cross entropy as described above, and iteratively updating the training parameters of the DNN through stochastic gradient descent.

Therefore, in the embodiment of the present application, after off-line training of the long-term scale DNN and the short-time scale DNN through the obtained training data, the training parameters of each sub-network in the DNN are obtained, and the optimized training parameters are fixed to facilitate subsequent prediction The stage makes predictions based on determined network parameters.

Step 103, obtain the signal to be transmitted, perform high-dimensional pilot estimation and high-dimensional channel feedback through the trained long-term DNN to restore the high-dimensional original channel matrix, and perform low-dimensional pilot through the trained short-time scale DNN Frequency estimation and low-dimensional channel feedback to obtain low-dimensional equivalent channel matrix.

Step 104: Perform analog precoding and digital precoding according to the high-dimensional original channel matrix through the long-time scale DNN, and perform digital precoding according to the low-dimensional equivalent channel matrix through the short-time scale DNN to complete signal transmission.

Wherein, the signal to be transmitted is the parallel data actually to be sent by the sending end.

In the specific implementation, in the prediction stage, the implementation method in the above prediction stage can be referred to, and the high-dimensional pilot estimation and high-dimensional channel feedback can be performed through the trained long-term scale DNN to restore the high-dimensional original channel matrix, and the training can be completed The short-time scale DNN of , performs low-dimensional pilot estimation and low-dimensional channel feedback to obtain a low-dimensional equivalent channel matrix. Then analog precoding and digital precoding are performed according to the high-dimensional original channel matrix by a long-time scale DNN, and digital precoding is performed according to a low-dimensional equivalent channel matrix by a short-time scale DNN. That is, based on the determined training parameters, the received pilot signal is mapped to feedback bits at the receiving end, and then the received feedback bits are mapped to a hybrid precoding matrix at the sending end. The specific implementation principle and implementation process can refer to the above prediction stage The scheme in the above-mentioned training phase is not repeated here, that is, after the fixed and optimized training parameters and the design of the hybrid precoder matrix are obtained through the above training phase scheme, in the prediction phase, each module in the signal flow shown in Figure 6 is For the actual data to be transmitted, the optimized parameters are substituted into the formulas in the above embodiments for channel estimation, channel feedback and hybrid precoding design, so that the trained After the network performs channel estimation, channel feedback and hybrid precoding design, the transmitted signal can be accurately recovered at the receiving end.

It should be noted that, in the channel feedback training phase of this application, the DNN with discrete output 0-1 variables is trained. Since the binary layer (the output is 0 or 1, the sign function sgn(x) is used as the activation function The derivative of ) is almost 0 everywhere, and it is not derivable at the origin. The existing backpropagation method cannot be directly used to train the layer before the binary layer. Therefore, in one embodiment of the present application, in the process of gradient backpropagation, a smooth everywhere-differentiable function is used to approximate the sign function. Specifically, the sign function sgn(x) is replaced by a 2sigm(x)-1 function, where sigm(x)=1/(1+exp(-x)) represents a sigmoid function. Moreover, in order to obtain a better training effect, a stepwise approximation method is also adopted in the embodiment of the present application, that is, as the training process progresses, the slope of the substitution function is slowly increased, so that the substitution function gradually approaches the sign function. Therefore, in the initial stage of training, the DNN training can achieve relevant effects relatively easily and quickly, just like ordinary DNNs. After the network training reaches the preset effect, the slope of the substitution function is increased to avoid numerical instability and make the training more stable and converge faster. Among them, the expression of the replacement function is:

Among them, α ⁽ⁱ⁾ is the parameter of the i-th epoch, which needs to satisfy α ⁽ⁱ⁾ ≥ α ^(i-1) .

In one embodiment of the present application, in order to improve the performance of dual-time-scale hybrid precoding, for the simulated precoding matrix, high-dimensional original channel matrix samples can be obtained by means of a sliding average or a sliding window. The sliding average is a long-term scale The current output of the analog transmitter precoding module and the analog receiver precoding module of the DNN are weighted averaged with the output at the last moment closest to the current one.

Specifically, when acquiring high-dimensional original channel matrix samples by means of moving average, the long-term scale variable (analog precoder) {F _RF , W _RF } needs to adapt to the statistical characteristics of the channel. Therefore, the optimization of long-term scale variables needs to be based on enough high-dimensional original channel samples H, however, since only one sample H can be obtained for each frame. For this reason, in the embodiment of this application, the channel sample H is fully utilized by using the sliding average method, that is, a weighted average is made between the current output result of the DNN and the result at the previous moment, and the weighted average is performed by the following formula:

和

分别表示当前时刻的发送端模拟编码器相位和上一帧AP-NN的输出，

和

分别表示当前时刻的接收端模拟编码器相位和上一帧AC-NN的输出。{γ _t,t＝1,2,…,T _f}表示滑动平均的步长序列，需要满足下述条件：

in,

and

Respectively represent the phase of the analog encoder at the sending end and the output of the previous frame AP-NN at the current moment,

and

Respectively represent the phase of the analog encoder at the receiving end at the current moment and the output of the AC-NN of the previous frame. {γ _t ,t=1,2,…,T _f } represents the step sequence of the moving average, which needs to meet the following conditions:

来存储之前若干帧恢复出来的高维原始信道样本H，如图7所示，在第t帧输入AC-NN和AP-NN网络的矩阵，为从第t-D+1帧到第t帧所有的信道样本，即

图7所示的滑动窗缓存容量为3，在本申请的一些实施例中，还可以根据实际需要确定缓存容量。 When obtaining high-dimensional original channel matrix samples by means of sliding windows, in order to make the long-term scale variables better adapt to the statistical characteristics of the channel and make full use of the channel samples H, the embodiment of this application proposes to use a sliding window with a certain buffer capacity D

To store the high-dimensional original channel samples H recovered from several previous frames, as shown in Figure 7, the matrix input to the AC-NN and AP-NN networks in the tth frame is from the t-D+1th frame to the tth frame All channel samples, i.e.

The sliding window buffer capacity shown in FIG. 7 is 3, and in some embodiments of the present application, the buffer capacity may also be determined according to actual needs.

It should also be noted that the dual-time-scale DNN training methods proposed in the embodiment of the present application can be divided into two types: centralized and distributed, and any of them can be used for training. In order to adapt to different training methods, this application also designs corresponding frame structures for centralized and distributed training to complete the training.

Specifically, Figure 8 shows the frame structure design of centralized training. Before DNN can be officially deployed, it needs to be trained offline. After the training is completed, the DNN corresponding to the receiving end (including structure and parameters) needs to be distributed to the user end. During use, a frame includes several time slots, and the structure of a time slot consists of the following four parts: indication bits, pilot symbols, feedback bits, and transmission data. Wherein, the indication bit represents whether the current time slot uses a long-time-scale DNN or a short-time-scale DNN, whether the current channel statistical characteristics have changed, and whether the channel change speed has changed. When the statistical characteristics of the channel change, because generally there will be no major changes in a short period of time, you can use new channel samples for fine-tuning (online training) on the basis of previous training, and after training for several time slots, you can restore use. If the channel change speed changes, the frame and time slot lengths need to be adaptively adjusted. When the channel changes faster, the frame and time slot lengths need to be shortened to obtain more high-dimensional original channel samples to track channel changes.

Figure 9 shows the frame structure design of distributed training. The difference from the frame structure of centralized training is that before DNN is officially used and deployed, it needs to be distributed offline training. This process includes the interaction of DNN input and output, gradient information, etc. between the base station (transmitter) and the user (receiver). . After the training is completed, it can be directly deployed and used without the process of distributing DNN (including structure and parameters) to the client.

Therefore, this application performs hybrid precoding based on dual time scales, and jointly designs each module in the communication system through a deep learning framework to achieve end-to-end performance optimization. In order to more clearly illustrate the beneficial effect of the precoding method of the embodiment of the present application, the following is combined with the actual application of the antenna system precoding method based on dual time scales and deep learning of the present application and the precoding scheme in the prior art The test results are compared, among which, Figure 10 compares the bit error rate of the dual-time-scale DNN and the traditional scheme under different signal-to-noise ratios, and Figure 11 compares the bit-error rate of the dual-time-scale DNN and the traditional scheme under different feedback bit numbers, Fig. 12 compares the BER of dual-time-scale DNN and traditional schemes under different pilot lengths. Therefore, it can be seen that compared with the traditional solution, the dual-time-scale DNN of the present application can significantly reduce the channel feedback overhead and the pilot length, while maintaining a good bit error rate performance.

In summary, the antenna system precoding method based on dual time scales and deep learning in the embodiment of the present application performs hybrid precoding based on dual time scales, in which the long-term analog precoding is obtained based on channel statistical characteristics, and the short-time scale The digital precoding of is optimized according to the low-dimensional real-time equivalent channel matrix, which can reduce signaling overhead and improve the robustness to channel mismatch caused by transmission delay. Moreover, the method jointly designs each module in the communication system through a deep learning framework to achieve end-to-end performance optimization. Statistical characteristics, no precise channel mathematical model is required, and the calculation of deep neural network can be parallelized, which improves the communication performance of massive MIMO system, reduces the computational complexity of hybrid precoding, and improves the bit error rate performance.

In order to more clearly illustrate the communication process of the end-to-end dual-scale millimeter-wave MIMO system in this application, a specific embodiment will be described in detail below in conjunction with FIG. 13:

接收端根据接收到的导频信号估计出低维等效信道矩阵H _eq，并对其进行量化，并将量化之后的信道信息q _eq反馈给发送端。随后，发送端根据反馈的结果恢复出低维等效信道矩阵

并设计数字预编码器{F _BB,W _BB}，同时保持模拟预编码器{F _RF,W _RF}不变。最后，按照如图6所示的信号流进行数据传输。在每一帧的最后一个时隙，发送端首先发送训练导频和模拟导频

接收端根据接收到的导频信号恢复出高维原始信道H，并对其进行量化，并将量化后的结果q位反馈给发送端。接着，发送端根据反馈的结果恢复出

并设计混合预编码器{F _BB,F _RF,W _BB,W _RF}。最后传输实际要发送的数据s。由于H _eq的维度远小于H，反馈信息q _eq的维度比q的维度小很多。 As shown in Figure 13, in the communication process of the MIMO system based on dual-time-scale hybrid precoding in the embodiment of the present application, in the first T _s -1 time slots of each frame, the transmitting end sends the pilot matrix

The receiving end estimates the low-dimensional equivalent channel matrix _Heq according to the received pilot signal, quantizes it, and feeds back the quantized channel information q _eq to the sending end. Subsequently, the sender restores the low-dimensional equivalent channel matrix according to the feedback result

And design the digital precoder {F _BB , W _BB } while keeping the analog precoder {F _RF , W _RF } unchanged. Finally, data transmission is performed according to the signal flow shown in FIG. 6 . In the last slot of each frame, the sender first sends training pilots and simulated pilots

The receiving end restores the high-dimensional original channel H according to the received pilot signal, quantizes it, and feeds back the quantized result q bits to the sending end. Then, the sender restores the output according to the feedback result

And design a hybrid precoder {F _BB , F _RF , W _BB , W _RF }. Finally transmit the actual data s to be sent. Since the dimension of _Heq is much smaller than H, the dimension of feedback information _qeq is much smaller than that of q.

In order to improve the applicability of the antenna system precoding method based on dual time scale and deep learning in this application, the MIMO system based on dual time scale hybrid precoding in this application can also be extended to other different types of systems to adapt to different application scenarios needs.

Specifically, in one embodiment of the present application, the network in the present application is extended to the TDD system through fewer steps, wherein only the channel feedback part in the MIMO system in the above embodiment needs to be omitted, and the channel can be fully utilized Reciprocity does. In this embodiment, due to the original MIMO system based on dual-time scale hybrid precoding, the base station sends the pilot, the user receives the pilot and estimates the downlink channel, quantizes the channel and feeds it back to the base station, and the base station restores The downlink channel is used for precoding, but in TDD, only the following modifications are required: remove the channel feedback part, change the base station transmission pilot to the user transmission pilot, and the other network structures remain unchanged. Then the corresponding specific process is. The user sends the pilot frequency, the base station estimates the uplink channel, and according to the channel reciprocity, can directly obtain the downlink channel, and directly perform precoding according to the downlink channel.

In another embodiment of the present application, the dual-time scale hybrid precoding-based MIMO system of the present application can also be simply extended to the OFDM system. For example, taking an OFDM system with 1024 subcarriers as an example, the channels of the extended OFDM system are equivalent to the 1024 channels (channels that each subcarrier has) in the system described in the above-mentioned embodiments, then these The channel is input as sample data into the long-time scale deep neural network DNN and short-time scale deep neural network DNN of this application for training, that is, if the OFDM to be extended contains 1024 subcarriers, then repeat the above steps 101 to 104, The extended OFDM system can be obtained by repeating 1024 times.

In order to realize the above embodiments, the present application also proposes an antenna system precoding device based on dual time scales and deep learning.

FIG. 14 is a schematic structural diagram of an antenna system precoding device based on dual time scales and deep learning proposed by an embodiment of the present application. As shown in FIG. 14 , the device includes a construction module 100 , a training module 200 , an acquisition module 300 and an encoding module 400 .

Wherein, the building block 100 is used to construct a long-time-scale deep neural network DNN and a short-time-scale deep neural network DNN, wherein the long-time-scale DNN and the short-time-scale DNN respectively include transceivers with a large-scale millimeter-wave multiple-input multiple-output MIMO system Multiple sub-networks corresponding to the machine, the multiple sub-networks include: the channel estimation sub-network and the channel feedback sub-network at the receiving end, and the pilot design sub-network and hybrid precoding sub-network at the sending end.

The training module 200 is used to obtain training data with different signal-to-noise ratios, and use the training data to train the long-time-scale DNN and the short-time-scale DNN, so as to optimize network parameters.

The acquisition module 300 is used to acquire the signal to be transmitted, and perform high-dimensional pilot estimation and high-dimensional channel feedback through the trained long-term DNN to restore the high-dimensional original channel matrix, and perform high-dimensional pilot estimation and high-dimensional channel feedback through the trained short-time scale DNN. Low-dimensional pilot estimation and low-dimensional channel feedback to obtain low-dimensional equivalent channel matrix.

The coding module 400 is used to perform analog precoding and digital precoding according to the high-dimensional original channel matrix through the long-time scale DNN, and perform digital precoding according to the low-dimensional equivalent channel matrix through the short-time scale DNN to complete signal transmission.

In one embodiment of the present application, the construction module 100 is further configured to divide the time axis into multiple superframes according to the channel statistical characteristics, and divide each superframe into a first preset number of frames, and each frame includes a second preset Given a number of time slots, the long time scale is determined in terms of superframes, and the short time scale is determined in terms of time slots.

In one embodiment of the present application, the training module 200 is also used to alternately train the long-time-scale DNN and the short-time-scale DNN according to the dual-time-scale frame structure, wherein the digital precoding matrix of the short-time-scale DNN is divided in each frame Each time slot other than the last time slot is updated based on the low-dimensional equivalent channel, and the analog precoding matrix and digital precoding matrix of the long-term DNN are in the last time slot of each frame, based on the high-dimensional equivalent channel to update.

In one embodiment of the present application, the training module 200 is also used to set the pilot information as the training parameters of the pilot design sub-network, and learn the target training parameters of the pilot design sub-network through stochastic gradient descent; through the estimator of the sigmoid function Approximate gradients of binary neurons and train channel feedback subnetworks via stochastic gradient descent.

In one embodiment of the present application, the training module 200 is also used to perform high-dimensional pilot estimation by the following formula:

是接收端接收到的导频信号矩阵，

是发送端发送的训练导频，

是模拟预编码矩阵，

H是待估计的高维原始信道，

是模拟接收矩阵

的共轭转置矩阵，

N是高斯噪声矩阵，

选取训练导频、模拟预编码矩阵和模拟接收矩阵为信道估计子网络的训练参数。 in,

is the pilot signal matrix received by the receiver,

is the training pilot sent by the sender,

is the simulated precoding matrix,

H is the high-dimensional original channel to be estimated,

is the simulated receiving matrix

The conjugate transpose matrix of ,

N is the Gaussian noise matrix,

The training pilot, simulated precoding matrix and simulated receiving matrix are selected as the training parameters of the channel estimation sub-network.

In one embodiment of the present application, the training module 200 is also used to perform high-dimensional channel feedback through the following formula:

表示导频信号矩阵

的向量化结果，

是向量

实部、虚部分开的表示，

是长时间尺度DNN的训练参数，σ _r是长时间尺度DNN第r层的非线性激活函数，sgn(·)是长时间尺度DNN二值层的激活函数。 where q is the feedback bit,

Represents the pilot signal matrix

The vectorized result of

is a vector

The real part and the imaginary part are separated,

is the training parameter of the long-time scale DNN, σ _r is the nonlinear activation function of the rth layer of the long-time scale DNN, and sgn( ) is the activation function of the binary layer of the long-time scale DNN.

In one embodiment of the present application, the hybrid precoding sub-network includes an analog transmitting end precoding module, a digital transmitting end precoding module, an analog receiving end precoding module, a digital receiving end precoding module and a demodulation module, and the encoding module 400 is also used to: input the real part and the imaginary part of the high-dimensional original channel matrix to the precoding module of the analog sending end and the precoding module of the analog receiving end respectively, so as to output the analog encoder phases of the sending end and the receiving end; The complex number vector of the modulus constraint; the complex number vector satisfying the constant modulus constraint is transformed by the following formula to generate the analog precoding matrix:

in,

表示将向量转换成矩阵的操作，

是发送端的模拟编码器相位，

是接收端的模拟编码器相位，N _t是，N _r是。 Among them, F _RF is the analog precoding matrix of the transmitting end, W _RF is the analog precoding matrix of the receiving end,

Represents the operation of converting a vector into a matrix,

is the phase of the analog encoder at the transmitter,

is the analog encoder phase at the receiving end, N _t is, N _r is.

In one embodiment of the present application, the training module 200 is also used to acquire high-dimensional original channel matrix samples by means of sliding average or sliding window. The current output of the encoding module and the output at the last moment closest to the current are weighted averaged.

In summary, the antenna system precoding device based on dual time scales and deep learning in the embodiment of the present application performs hybrid precoding based on dual time scales, in which the long-term analog precoding is obtained based on channel statistical characteristics, and the short-time scale The digital precoding of is optimized according to the low-dimensional real-time equivalent channel matrix, which can reduce signaling overhead and improve the robustness to channel mismatch caused by transmission delay. Moreover, the device jointly designs each module in the communication system through a deep learning framework to achieve end-to-end performance optimization. The deep learning framework implicitly learns the channel performance in a data-driven manner during the process of optimizing the end-to-end communication system. Statistical characteristics, no precise channel mathematical model is required, and the calculation of deep neural network can be parallelized, which improves the communication performance of massive MIMO system, reduces the computational complexity of hybrid precoding, and improves the bit error rate performance.

In order to realize the above-mentioned embodiments, the present disclosure also proposes an electronic device, including a memory, a processor, and a computer program stored on the memory and operable on the processor. When the processor executes the program, the present disclosure based on Steps in an antenna system precoding method with dual time scales and deep learning.

In order to realize the above-mentioned embodiments, the present disclosure also proposes a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements a An Antenna System Precoding Method Based on Dual Time Scales and Deep Learning.

Provided in the present disclosure may be computing devices such as desktop computers, notebooks, palmtop computers, and cloud servers. The terminal device may include, but not limited to, a processor and a memory. The processor can be a central processing unit (Central Processing Unit, referred to as CPU), and can also be other general-purpose processors, digital signal processors (Digital Signal Processor, referred to as DSP), application specific integrated circuits (Application Specific Integrated Circuit, referred to as ASIC) ), off-the-shelf programmable gate array (Field-Programmable Gate Array, referred to as FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The memory can be used to store the computer programs and/or modules, and the processor implements the terminal by running or executing the computer programs and/or modules stored in the memory and calling the data stored in the memory various functions of the device.

The electronic device of the present disclosure can implement the steps in the foregoing method embodiments when the processor executes the computer program. Alternatively, when the processor executes the computer program, the functions of the modules/units in the above device embodiments are implemented.

Any process or method descriptions in flowcharts or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing custom logical functions or steps of a process , and the scope of preferred embodiments of the present application includes additional implementations in which functions may be performed out of the order shown or discussed, including in substantially simultaneous fashion or in reverse order depending on the functions involved, which shall It should be understood by those skilled in the art to which the embodiments of the present application belong.

The logic and/or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced listing of executable instructions for implementing logical functions, can be embodied in any computer-readable medium, For use with instruction execution systems, devices, or devices (such as computer-based systems, systems including processors, or other systems that can fetch instructions from instruction execution systems, devices, or devices and execute instructions), or in conjunction with these instruction execution systems, devices or equipment used. For the purposes of this specification, a "computer-readable medium" may be any device that can contain, store, communicate, propagate or transmit a program for use in or in conjunction with an instruction execution system, device or device. More specific examples (non-exhaustive list) of computer-readable media include the following: electrical connection with one or more wires (electronic device), portable computer disk case (magnetic device), random access memory (RAM), Read Only Memory (ROM), Erasable and Editable Read Only Memory (EPROM or Flash Memory), Fiber Optic Devices, and Portable Compact Disc Read Only Memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program can be printed, since the program can be read, for example, by optically scanning the paper or other medium, followed by editing, interpretation or other suitable processing if necessary. processing to obtain the program electronically and store it in computer memory.

It should be understood that each part of the present application may be realized by hardware, software, firmware or a combination thereof. In the embodiments described above, various steps or methods may be implemented by software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented by any one or a combination of the following techniques known in the art: a discrete Logic circuits, ASICs with suitable combinational logic gates, Programmable Gate Arrays (PGA), Field Programmable Gate Arrays (FPGA), etc.

Those of ordinary skill in the art can understand that all or part of the steps carried by the methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. During execution, one or a combination of the steps of the method embodiments is included.

In addition, each functional unit in each embodiment of the present application may be integrated into one processing module, each unit may exist separately physically, or two or more units may be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware or in the form of software function modules. If the integrated modules are realized in the form of software function modules and sold or used as independent products, they can also be stored in a computer-readable storage medium.

The storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disk, and the like.

In the description of this specification, descriptions referring to the terms "one embodiment", "some embodiments", "example", "specific examples", or "some examples" mean that specific features described in connection with the embodiment or example , structure, material or characteristic is included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Furthermore, the described specific features, structures, materials or characteristics may be combined in any suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or examples and features of different embodiments or examples described in this specification without conflicting with each other.

In addition, the terms "first" and "second" are used for descriptive purposes only, and cannot be interpreted as indicating or implying relative importance or implicitly specifying the quantity of indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include at least one of these features. In the description of the present application, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limitations on the present application, and those skilled in the art can make the above-mentioned The embodiments are subject to changes, modifications, substitutions and variations.

Claims (18)

An antenna system precoding method based on dual time scales and deep learning, comprising the following steps: A long-time-scale deep neural network DNN and a short-time-scale deep neural network DNN are constructed, wherein the long-time-scale DNN and the short-time-scale DNN respectively include multiple sub-networks corresponding to the transceivers of the massive millimeter-wave MIMO system. The multiple sub-networks include: a channel estimation sub-network and a channel feedback sub-network at the receiving end, and a pilot design sub-network and a hybrid precoding sub-network at the sending end; Obtaining training data with different signal-to-noise ratios, and using the training data to train the long-time-scale DNN and the short-time-scale DNN to optimize network parameters; Obtain the signal to be transmitted, perform high-dimensional pilot estimation and high-dimensional channel feedback through the trained long-term scale DNN to restore the high-dimensional original channel matrix, and perform low-dimensional through the trained short-time scale DNN Pilot estimation and low-dimensional channel feedback to obtain a low-dimensional equivalent channel matrix; Perform analog precoding and digital precoding according to the high-dimensional original channel matrix through the long-term scale DNN, and perform digital precoding through the short-time scale DNN according to the low-dimensional equivalent channel matrix to complete signal transmission . The method according to claim 1, wherein the time axis is divided into a plurality of superframes according to channel statistical characteristics, and each said superframe is divided into a first preset number of frames, and each said frame includes a second preset Given a number of time slots, the long time scale is determined from the superframe, and the short time scale is determined from the time slots. The method according to claim 1 or 2, wherein said training the long-time scale DNN and the short-time scale DNN comprises: According to the dual-time-scale frame structure, the long-time-scale DNN and the short-time-scale DNN are alternately trained, wherein the digital precoding matrix of the short-time-scale DNN is in each frame except the last time slot The time slot is updated based on the low-dimensional equivalent channel, and the analog precoding matrix and digital precoding matrix of the long-term DNN are updated based on the high-dimensional equivalent channel in the last time slot of each frame . The method according to any one of claims 1 to 3, wherein the channel feedback sub-network is constructed according to the output of binary neurons in the deep neural network, and the long-term scale DNN and the short-time scale DNN are trained ,Also includes: Setting the pilot information as the training parameters of the pilot design sub-network, learning the target training parameters of the pilot design sub-network through stochastic gradient descent; The gradient of the binary neuron is approximated by an estimator of the sigmoid function, and the channel feedback subnetwork is trained by stochastic gradient descent. The method according to any one of claims 1 to 4, wherein the high-dimensional pilot estimation is performed by the following formula:

in,

is the pilot signal matrix received by the receiver,

is the training pilot sent by the sender,

is the simulated precoding matrix,

H is the high-dimensional original channel to be estimated,

is the simulated receiving matrix

The conjugate transpose matrix of ,

N is the Gaussian noise matrix,

The training pilot, the simulated precoding matrix and the simulated receiving matrix are selected as training parameters of the channel estimation sub-network. The method according to any one of claims 1 to 5, wherein high-dimensional channel feedback is performed by the following formula:

where q is the feedback bit,

Represents the pilot signal matrix

The vectorized result of

is a vector

The real part and the imaginary part are separated,

is the training parameter of the long-time scale DNN, σ _r is the nonlinear activation function of the rth layer of the long-time scale DNN, and sgn( ) is the activation function of the binary layer of the long-time scale DNN. The method according to any one of claims 1 to 6, wherein the hybrid precoding sub-network includes an analog transmitting end precoding module, a digital transmitting end precoding module, an analog receiving end precoding module, a digital receiving end precoding module A module and a demodulation module, the analog precoding is performed according to the high-dimensional original channel, including: Inputting the real part and the imaginary part of the high-dimensional original channel matrix to the analog transmitting end precoding module and the analog receiving end precoding module respectively, so as to output the analog encoder phases of the transmitting end and the receiving end; Calculate the complex vector that satisfies the constant modulus constraint; The complex number vector that satisfies the constant modulus constraint is converted by the following formula to generate an analog precoding matrix:

in,

Among them, F _RF is the analog precoding matrix of the transmitting end, W _RF is the analog precoding matrix of the receiving end,

Represents the operation of converting a vector into a matrix,

is the phase of the analog encoder at the transmitter,

is the analog encoder phase at the receiver, N _t is the number of antennas at the transmitter, and N _r is the number of antennas at the receiver. The method according to any one of claims 1 to 7, wherein the training data includes high-dimensional original channel matrix samples, Gaussian noise and data labels to be sent, and the long-term scale DNN is trained, further comprising: The high-dimensional original channel matrix samples are obtained by means of a sliding average or a sliding window, and the sliding average is the current output of the analog transmitting-end precoding module and the analog receiving-end precoding module of the long-term scale DNN and the current closest The output of the previous moment is weighted average. An antenna system precoding device based on dual time scales and deep learning, including: A building block for constructing a long-time-scale deep neural network DNN and a short-time-scale deep neural network DNN, wherein the long-time-scale DNN and the short-time-scale DNN respectively include transceivers corresponding to massive millimeter-wave multiple-input multiple-output MIMO systems A plurality of subnetworks, the plurality of subnetworks include: a channel estimation subnetwork and a channel feedback subnetwork at the receiving end, and a pilot design subnetwork and a hybrid precoding subnetwork at the sending end; A training module, configured to obtain training data with different signal-to-noise ratios, and train the long-time scale DNN and the short-time scale DNN through the training data to optimize network parameters; The acquisition module is used to acquire the signal to be transmitted, perform high-dimensional pilot estimation and high-dimensional channel feedback through the long-term scale DNN completed through training, so as to restore the high-dimensional original channel matrix, and complete the short-term DNN through training Scale DNN performs low-dimensional pilot estimation and low-dimensional channel feedback to obtain low-dimensional equivalent channel matrix; A coding module, configured to perform analog precoding and digital precoding according to the high-dimensional original channel matrix through the long-time scale DNN, and perform digital precoding through the short-time scale DNN according to the low-dimensional equivalent channel matrix , to complete the signal transmission. The device according to claim 9, wherein the building block is further configured to divide the time axis into a plurality of superframes according to channel statistical characteristics, and divide each superframe into a first preset number of frames, each frame comprising For the second preset number of time slots, the long time scale is determined according to the superframe, and the short time scale is determined according to the time slots. The device according to claim 9 or 10, wherein the training module is further used to alternately train a long-time scale DNN and a short-time-scale DNN according to a dual-time-scale frame structure, wherein the digital precoding matrix of the short-time scale DNN In every slot except the last slot of each frame, the update is based on the low-dimensional equivalent channel, and the analog precoding matrix and digital precoding matrix of the long-term DNN are in the last slot of each frame, based on The high-dimensional equivalent channel is updated. Apparatus according to any one of claims 9 to 11, wherein said training module is further used to The pilot information is set as the training parameters of the pilot design sub-network, and the target training parameters of the pilot design sub-network are learned by stochastic gradient descent; The gradient of the binary neuron is approximated by an estimator of the sigmoid function, and the channel feedback subnetwork is trained by stochastic gradient descent. The device according to any one of claims 9 to 12, wherein the training module is also used for high-dimensional pilot estimation by the following formula:

in,

is the pilot signal matrix received by the receiver,

is the training pilot sent by the sender,

is the simulated precoding matrix,

H is the high-dimensional original channel to be estimated,

is the simulated receiving matrix

The conjugate transpose matrix of ,

N is the Gaussian noise matrix,

The training pilot, simulated precoding matrix and simulated receiving matrix are selected as the training parameters of the channel estimation sub-network. The device according to any one of claims 9 to 13, wherein the training module is also used for high-dimensional channel feedback by the following formula:

where q is the feedback bit,

Represents the pilot signal matrix

The vectorized result of

is a vector

The real part and the imaginary part are separated,

is the training parameter of the long-time scale DNN, σ _r is the nonlinear activation function of the rth layer of the long-time scale DNN, and sgn( ) is the activation function of the binary layer of the long-time scale DNN. The device according to any one of claims 9 to 14, wherein the hybrid precoding sub-network includes an analog transmitting end precoding module, a digital transmitting end precoding module, an analog receiving end precoding module, a digital receiving end precoding module, and a digital receiving end precoding module. module and a demodulation module, the encoding module is also used for: Inputting the real part and the imaginary part of the high-dimensional original channel matrix to the precoding module of the analog transmitting end and the precoding module of the analog receiving end respectively, so as to output the analog encoder phases of the transmitting end and the receiving end; Calculate the complex vector that satisfies the constant modulus constraint; The complex number vector that satisfies the constant modulus constraint is transformed by the following formula to generate an analog precoding matrix:

in,

Among them, F _RF is the analog precoding matrix of the transmitting end, W _RF is the analog precoding matrix of the receiving end,

Represents the operation of converting a vector into a matrix,

is the phase of the analog encoder at the transmitter,

is the analog encoder phase at the receiving end, N _t is, N _r is. The device according to any one of claims 9 to 15, wherein the training module is also used to obtain high-dimensional original channel matrix samples by means of a sliding average or a sliding window, and the sliding average is to send the simulation of the long-term scale DNN The current output of the end precoding module and the analog receiving end precoding module are weighted and averaged with the output at the last moment closest to the current one. An electronic device comprising: A memory, a processor, and a computer program stored on the memory and operable on the processor, when the processor executes the program, the dual-time-scale and deep learning based on any one of claims 1 to 8 is realized The steps in the antenna system precoding method. A non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the antenna system based on dual time scales and deep learning according to any one of claims 1 to 8 is realized precoding method.

Images