阅读说明：本技术 一种面向车联网的大规模mimo智能csi反馈方法 (Internet of vehicles-oriented large-scale MIMO intelligent CSI feedback method ) 是由 陈成瑞 李玉杰 程港 廖勇 于 2021-07-21 设计创作，主要内容包括：本发明提出了一种面向车联网的大规模MIMO智能CSI反馈方法。针对车联网信道状态信息(CSI)反馈方法复杂度高、反馈精度低和反馈开销大的问题,本发明通过端到端的方式构建了一种从用户端(UE)的编码器到基站端(BS)的解码器相结合的网络结构。在UE端采用连续的平均池化层、普通卷积层(CNN)和深度可分离卷积层(DSCNN)所构成的块来逐步缩小原始的CSI矩阵得到压缩后的码字矩阵。在BS端则是首先采用连续的上采样层、CNN和DSCNN所构成的块来逐步扩大压缩的码字矩阵,得到初始CSI重构。接着,利用残差学习网络(ResNet)可以逐步微调的功能构建了残差卷积块将初始重构的CSI逐步逼近原始的CSI以达到更好的重建效果。(The invention provides a large-scale MIMO intelligent CSI feedback method for the Internet of vehicles. Aiming at the problems of high complexity, low feedback precision and high feedback overhead of a vehicle networking Channel State Information (CSI) feedback method, the invention constructs a network structure combining an encoder of a User End (UE) and a decoder of a base station end (BS) in an end-to-end mode. And adopting a block consisting of a continuous average pooling layer, a common convolutional layer (CNN) and a depth separable convolutional layer (DSCNN) to gradually reduce the original CSI matrix at the UE end to obtain a compressed code matrix. At the BS end, the compressed codeword matrix is gradually expanded by using the blocks formed by the consecutive upsampling layers, CNNs, and DSCNNs to obtain the initial CSI reconstruction. Then, a residual rolling block is constructed by utilizing the function that a residual learning network (ResNet) can be finely adjusted step by step, so that the initially reconstructed CSI gradually approaches the original CSI to achieve a better reconstruction effect.)

1. The large-scale MIMO intelligent CSI feedback method for the Internet of vehicles is characterized by comprising the following steps:

s1, constructing a communication system model facing the Internet of vehicles and based on large-scale MIMO, and obtaining batch CSI truncation matrix data;

s2, designing a structure of an encoder in a lightweight CSI feedback framework;

s3, designing a structure of a decoder in a lightweight CSI feedback framework;

and S4, designing a reasonable training scheme, and training the encoder and the decoder which are provided by S2 and S3 as the same end-to-end network to obtain a network model.

2. The massive MIMO intelligent CSI feedback method for Internet of vehicles according to claim 1, wherein the S1 comprises:

in a large-scale MIMO-based system facing Internet of vehicles, a base station end is considered to have N_t(N_t> 1) a transmitting antenna, and a single receiving antenna is arranged at a user side; the system employs Orthogonal Frequency Division Multiplexing (OFDM) with N_sA subcarrier; the resulting receiver-side signal y can be described as:

wherein the content of the first and second substances,is represented by N_sThe dimensions of the received vector are then used to determine,is represented by N_sThe dimensions of the transmitted vector are,is represented by N_s×N_tDimensional channel matrix ofRepresented as a channel vector on the ith subcarrier,wherein u is_iA precoding vector representing the ith subcarrier,is represented by N_sAdditive white gaussian noise of the dimension;

in order to better design the precoding vector u_iIn the CSI feedback process, it is required to obtain a sufficient accuracy at the base stationNamely, a higher reconstruction effect is obtained through a lower feedback quantity; the CSI matrix is sparse in the angular domain, so the virtual angular domain matrix can be obtained by two Discrete Fourier Transform (DFT) matrices:

wherein D is_sIs a number N_s×N_sDFT matrix of, D_tIs a number N_t×N_tThe DFT matrix of (1); will be original N_s×N_tMatrix is cut off to N'_s×N_t(N′_s＜N_s) Of the truncated matrix, i.e. CSI truncation matrix

3. The massive MIMO intelligent CSI feedback method for Internet of vehicles according to claim 1, wherein the S2 comprises:

truncating the matrix for CSIThe matrix belongs to a complex matrix, wherein the matrix comprises data of a real part and an imaginary part, and the data are converted into N 'for convenience of processing and training'_s×N_tX 2; rotating shaftThe transformed matrix is used as the input of the encoder, and firstly passes through a ConvBN block (a two-dimensional common convolution layer Conv2D with a 3 × 3 convolution kernel, a Batch Normalization layer (BN) and an activation function (LeakyReLU) to output a feature map); then, obtaining a compressed code word matrix S through four continuous ACN blocks_en(ii) a Each SACN block consists of one average pooling layer (AveragePooling2D) with a pooling window size of 2 × 2 and one secnvbn block; and each SEConvBN block is composed of a depth separable convolution (SeparableConv2D), BN, and leakyreu.

4. The massive MIMO intelligent CSI feedback method for Internet of vehicles according to claim 1, wherein the S3 comprises:

code word matrix S obtained from encoder output_enThen inputting the USCN block into a decoder, firstly passing through 4 continuous USCN blocks, wherein each USCN block consists of a two-dimensional UpSampling layer (UpSamplling 2D) with the data interpolation window size of 2 multiplied by 2 and a ConvBN block; the up-sampling layer can enable the dimension of input data to be increased, and the principle is that the dimension increase is completed through a repeated interpolation process of rows and columns of the data; subsequently, N 'was obtained through a ConvBN layer'_s×N_tThe multiplied by 2 output is input into two continuous ConvBlock blocks, each ConvBlock block is a residual volume block constructed according to the form of a residual network and consists of three ConvBN blocks; taking the sum of the output after passing through one ConvBN layer and the input of the whole ConvBlock as the final output of the ConvBlock; finally, N'_s×N_tThe x 2 output is passed through a common convolution layer Conv2D (activation function Sigmoid, which may adjust the data to [0,1 ] with a 3 x 3 convolution kernel]) Obtaining the final reconstruction result

5. The massive MIMO intelligent CSI feedback method for Internet of vehicles according to claim 1, wherein the S4 comprises:

s1, the training configuration of the vehicle networking MIMO communication system parameters and the network is as follows: the base station antenna adopts a Uniform Linear Array antenna arrangement mode (ULA), and the antenna spacing is half wavelength; the number of base station end antennas is N_t32, the number of subcarriers of the OFDM system is N_s1024; considering sparse properties of massive MIMO we take only front N'_s32 rows; i.e. H is a 32 x 32 complex matrix; the training, validation and test data set sizes were: 100000, 30000 and 20000; the training batch data size (batch size) was 100, the training round (epochs) was 1500 rounds, and the web learning rate (learning rate) was 0.001.

The technical field is as follows:

the invention relates to a large-scale MIMO intelligent CSI (channel State information) feedback method for Internet of vehicles, in particular to a CSI feedback method based on light weight and low complexity.

Background art:

in the car networking environment, the change of the wireless propagation channel is complicated due to the high-speed movement of the vehicle. The research on the characteristics of the high-speed wireless channel is the basis of the research on the communication technology in the high-speed environment, and the characteristics of the wireless channel generally mainly comprise time-frequency domain channel response and time-frequency domain channel correlation coefficients. On one hand, in a high-speed mobile environment, due to the influence of multipath and doppler, the wireless channel at this time will exhibit time domain and frequency domain selective fading characteristics at the same time. On the other hand, when the vehicle moves at a high speed, the spread wireless channel has the characteristics of fast fading, severe doppler effect, high data transmission rate, complex communication environment and the like, which causes the channel to be not a generalized and stable random process at this time, i.e. the wireless channel in the high-speed moving environment is a typical non-stable channel. Due to these characteristics of the wireless channel in the high-speed mobile environment, it also brings more challenges to obtain reliable, stable and fast mobile communication.

In order to improve system performance, the downlink CSI needs to be accurately acquired at the transmitting end, and if all CSI is fed back, a serious burden is imposed on a feedback link. Therefore, in order to reduce excessive feedback overhead, the receiving end only needs to feed back part of CSI to the transmitting end. In recent years, the deep learning method is gradually applied to CSI feedback, and many researchers have proposed some CSI feedback frameworks based on deep learning. Deep learning based real-time channel recovery schemes utilize deep neural networks in training and prediction to reduce feedback overhead. Existing schemes, such as the channel recovery framework CsiNet based on deep learning, use an encoder at the user end to convert the channel matrix into codewords and a decoder at the base end to reconstruct the CSI from the codewords. Because the code word is small and can not identify the information of the channel matrix, the overfitting phenomenon is serious and the recovery effect is common. There is also an improved network architecture csiintplus, which uses a larger convolution kernel and more reinenet blocks to improve the reconstruction effect of the network on the CSI. Although the CSI feedback frameworks can obtain better reconstruction effects than methods based on compressed sensing and codebooks, their complex deployment at the user end and high requirements for equipment computation are still problems to be solved.

The invention content is as follows:

aiming at the problem of high complexity feedback overhead, the invention provides a large-scale MIMO intelligent CSI feedback method for the Internet of vehicles, which is characterized by comprising the following steps:

s1, constructing a communication system model facing the Internet of vehicles and based on large-scale MIMO, and obtaining batch CSI truncation matrix data;

s2, designing a structure of an encoder in a lightweight CSI feedback framework;

s3, designing a structure of a decoder in a lightweight CSI feedback framework;

and S4, designing a reasonable training scheme, and training the encoder and the decoder which are provided by S2 and S3 as the same end-to-end network to obtain a network model.

The intelligent CSI feedback method based on the massive MIMO system and oriented to the Internet of vehicles is characterized in that the S1 comprises the following steps:

in a large-scale MIMO-based system facing Internet of vehicles, a base station end is considered to have N_t(N_t> 1) a transmitting antenna, and a user side has a single receiving antenna. The system employs Orthogonal Frequency Division Multiplexing (OFDM) and has N_sAnd (4) sub-carriers. The resulting receiver-side signal y can be described as:

wherein the content of the first and second substances,is represented by N_sThe dimensions of the received vector are then used to determine,is represented by N_sThe dimensions of the transmitted vector are,is represented by N_s×N_tDimensional channel matrix ofRepresented as a channel vector on the ith subcarrier,wherein u is_iA precoding vector representing the ith subcarrier,is represented by N_sAdditive white gaussian noise of the dimension.

In order to better design the precoding vector u_iIn the CSI feedback process, it is required to obtain a sufficient accuracy at the base stationI.e. a higher reconstruction result is obtained with a lower amount of feedback. The CSI matrix is sparse in the angular domain, so the virtual angular domain matrix can be obtained by two Discrete Fourier Transform (DFT) matrices:

wherein D is_sIs a number N_s×N_sDFT matrix of, D_tIs a number N_t×N_tThe DFT matrix of (a). Will be original N_s×N_tMatrix is cut off to N'_s×N_t(N′_s＜N_s) Of the truncated matrix, i.e. CSI truncation matrix

The intelligent CSI feedback method based on the massive MIMO system and oriented to the Internet of vehicles is characterized in that the S2 comprises the following steps:

truncating the matrix for CSIThe matrix belongs to a complex matrix, wherein the matrix comprises data of a real part and an imaginary part, and the data are converted into N 'for convenience of processing and training'_s×N_tX 2. The converted matrix is used as input of the coderFirst, a ConvBN block (the feature map is output by a two-dimensional normal convolution layer Conv2D, batch normalization layer (BN) and activation function (LeakyReLU) of a 3 × 3 convolution kernel). Then, obtaining a compressed code word matrix S through four SACN blocks_en. Each SACN block consists of one average pooling layer (AveragePooling2D) with a pooling window size of 2 x 2 and one secnvbn block. And each SEConvBN block is composed of a depth separable convolution (SeparableConv2D), BN, and leakyreu.

The intelligent CSI feedback method based on the massive MIMO system and oriented to the Internet of vehicles is characterized in that the S3 comprises the following steps:

code word matrix S obtained from encoder output_enThen, the data is input into a decoder, and the data is firstly input into 4 consecutive USCN blocks, wherein each USCN block consists of a two-dimensional UpSampling layer (UpSamplling 2D) with the data interpolation window size of 2 multiplied by 2 and a ConvBN block. The upsampling layer can enable the dimension of input data to be increased, and the principle is that the dimension is increased through repeated interpolation processes of rows and columns of the data. Subsequently, N 'was obtained through a ConvBN layer'_s×N_tThe x 2 output is input to two consecutive ConvBlock blocks, each of which is a residual volume block constructed according to the form of a residual network, and is composed of three ConvBN blocks. The sum of the output after passing through one ConvBN layer and the input of the entire ConvBlock block is taken as the final output of the ConvBlock block. Finally, N'_s×N_tThe x 2 output is passed through a common convolution layer Conv2D (activation function Sigmoid, which may adjust the data to [0,1 ] with a 3 x 3 convolution kernel]) Obtaining the final reconstruction result

In conclusion, the beneficial effects of the invention are as follows:

the intelligent CSI feedback method based on the large-scale MIMO system and oriented to the Internet of vehicles has good reconstruction quality of a channel matrix, and obtains good performance while reducing the complexity of the network. For the equipment with low computing power and low storage of the current user terminal, the CSI feedback method provided by the invention has more excellent deployability and performance.

Drawings

The above and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:

FIG. 1 is a general flow diagram of the present invention;

fig. 2 is a diagram of a structure of client side coding;

fig. 3 is a structural design diagram of a decoder at a base station end.

Detailed Description

Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the accompanying drawings are illustrative only for the purpose of explaining the present invention, and are not to be construed as limiting the present invention.

In the present invention, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.

The invention provides a large-scale MIMO intelligent CSI feedback method for the Internet of vehicles, which has better performance in complexity, reconstruction quality and running time.

The invention is explained in detail with reference to the attached figure 1, which mainly comprises the following steps:

step 1: and starting.

Step 2: and constructing a communication system model facing the Internet of vehicles and based on large-scale MIMO, and establishing a CSI truncation matrix.

In a large-scale MIMO-based system facing Internet of vehicles, a base station end is considered to have N_t(N_t> 1) a transmitting antenna, and a user side has a single receiving antenna. The system employs Orthogonal Frequency Division Multiplexing (OFDM) with N_sAnd (4) sub-carriers. The resulting receiver-side signal y can be described as:

wherein the content of the first and second substances,is represented by N_sThe dimensions of the received vector are then used to determine,is represented by N_sThe dimensions of the transmitted vector are,is represented by N_s×N_tDimensional channel matrix ofRepresented as a channel vector on the ith subcarrier,wherein u is_iA precoding vector representing the ith subcarrier,is represented by N_sAdditive white gaussian noise of the dimension.

In order to better design the precoding vector u_iIn the CSI feedback process, it is required to obtain a sufficient accuracy at the base stationI.e. by lowerThe feedback amount achieves a higher reconstruction effect. The CSI matrix is sparse in the angular domain, so the virtual angular domain matrix can be obtained by two Discrete Fourier Transform (DFT) matrices:

wherein D is_sIs a number N_s×N_sDFT matrix of, D_tIs a number N_t×N_tThe DFT matrix of (a). Will be original N_s×N_tMatrix is cut off to N'_s×N_t(N′_s＜N_s) Of the truncated matrix, i.e. CSI truncation matrixTruncating the matrix for CSIThe matrix belongs to a complex matrix, wherein the matrix comprises data of a real part and an imaginary part, and the data are converted into N 'for convenience of processing and training'_s×N_tX 2. The transformed matrix serves as the input to the encoder.

And step 3: designing a CSI truncation matrix encoder, and inputting a CSI truncation matrix, wherein the specific implementation process comprises the following steps:

first, 64 signature numbers were output through a ConvBN layer (a two-dimensional normal convolution layer Conv2D with a 3 × 3 convolution kernel, Batch Normalization layer (Batch Normalization) and activation function (LeakyReLU)). Then, obtaining a compressed code word matrix S through four SACN blocks_enConversion of the number of feature maps intoEach SACN block consists of an average pooling layer (AveragePooling2D) with a pooling window size of 2 × 2 and a SEConvBN block (unlike the ConvBN block, in which the ordinary two-dimensional convolutional layer (Conv2D) is replaced with a depth-separable convolutional layer of 3 × 3 convolutional kernels: (the SACN block is a block with a pooling window size of 2 × 2) ((Conv 2D)SeparableConv 2D). Compared with the common convolutional layer, the SeparableConv2D has greatly reduced network parameters. Assuming that the matrix specification of the input two convolutional layers is (w, l, c), the convolutional kernel size (k, k) and no bias terms are calculated, the final output channel number is m. For a normal convolutional block, there are as many convolution kernels as there are input channels, and the parameters of the normal convolutional block are: param is m × k × k × 0 c. The depth separable convolution block performs channel convolution (i.e., the number of feature maps equal to the number of input channels can be obtained), and then performs 1 × 1 point-by-point convolution on each output feature map, and calculates the parameter number to obtain param ═ k × k × c +1 × 1 × c × m ═ k × k + m × c.

And 4, step 4: designing a code matrix decoder, and inputting the code matrix, wherein the specific implementation process comprises the following steps:

obtaining a code matrix S from the output of the encoder_enThe input decoder is firstly 4 consecutive USCNs, each USCN block is composed of a two-dimensional UpSampling layer (Up Sampling 2D) with the data interpolation window size of 2 multiplied by 2 and a ConvBN block, and each USCN block is composed of an UpSampling2D layer and a SEConvBN block. The upsampling layer may enable the dimensionality of the input data to be increased, which is accomplished by performing a repeated interpolation process on the rows and columns of the data. The feature map number conversion process is as follows:subsequently, N 'was obtained through a ConvBN layer'_s×N_tThe x 2 output is input to two consecutive ConvBlock blocks, each of which is a residual volume block constructed according to the form of a residual network. Each ConvBlock contains three ConvBN layers with 8, 16, 2 signature graphs. Finally, N'_s×N_tThe x 2 output is passed through a common convolution layer Conv2D (activation function Sigmoid, which may adjust the data to [0,1 ] with a 3 x 3 convolution kernel]) Obtaining the final reconstruction result

And 5: designing a training scheme, training a network model through a large amount of data, and specifically implementing the following processes:

the training configuration of the large-scale MIMO-based communication system parameters and network facing the Internet of vehicles is as follows: the base station antenna adopts a Uniform Linear Array antenna arrangement mode (ULA), and the antenna spacing is half a wavelength. The number of base station end antennas is N_t32, the number of subcarriers of the OFDM system is N_s1024. Considering sparse properties of massive MIMO we take only front N'_s32 rows. I.e., H is a 32 x 32 complex matrix. The training, validation and test data set sizes were: 100000, 30000 and 20000. The training batch data size (batch size) was 100, the training round (epochs) was 1500 rounds, and the web learning rate (learning rate) was 0.001.

Step 6: and (6) ending.

In the description herein, references to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

While embodiments of the invention have been shown and described, it will be understood by those of ordinary skill in the art that: various changes, modifications, substitutions and alterations can be made to the embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.