Single-bit MIMO system channel estimation method based on PDU-RA-GAN

Abstract

The invention provides a single-bit MIMO system channel estimation method based on PDU-RA-GAN, and relates to the technical field of wireless communication. The method comprises the following specific steps: firstly, making a data set and dividing the data set into a training set and a test set; then, a PDU-RA-GAN is designed; secondly, designing a composite loss function; and finally, the PDU-RA-GAN network is trained and tested. Experiments verify that the PDU-RA-GAN designed by the invention has better channel estimation accuracy for a single-bit MIMO (Multiple Input Multiple Output) system.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a single-bit MIMO system channel estimation method based on PDU-RA-GAN. Background Art

[0002] MIMO systems are of great significance in the fifth and sixth generation communication systems. Deploying large-scale antenna arrays at the base station can not only significantly improve the spectrum reuse efficiency of multiple users, but also greatly increase the data transmission rate. However, MIMO systems using high-resolution analog-to-digital converters (ADCs) have problems such as complex hardware design and high power consumption. In contrast, MIMO systems using single-bit analog-to-digital converters (i.e., single-bit MIMO systems) have gradually become a research hotspot due to their simple hardware structure and low power consumption.

[0003] Channel estimation is a key technology in the field of wireless communications, mainly used to evaluate the characteristics of wireless channels, such as fading, multipath effects, and noise. These characteristics will cause attenuation and distortion of wireless signals, thus affecting the communication quality. Only by accurately estimating the channel characteristics can channel compensation be effectively implemented, thereby improving communication performance. Therefore, channel estimation plays a vital role in ensuring the quality of wireless communications. However, in single-bit MIMO systems, the accuracy of channel estimation is seriously affected by the large quantization error introduced by the single-bit analog-to-digital converter.

[0004] With the rapid development of artificial intelligence theory, deep learning has brought a new perspective to solving the channel estimation problem. Using deep neural networks can mine richer feature information from signals, which helps to more accurately describe complex, variable and nonlinear channel characteristics, thereby significantly improving the effect of channel estimation. Summary of the invention

[0005] The present invention proposes a single-bit MIMO system channel estimation method based on PDU-RA-GAN, which relates to the field of wireless communication technology. The specific steps include:

[0006] First, create a dataset and divide it into training and testing sets; then, design PDU-RA-GAN; secondly, design a composite loss function; finally, train and test the PDU-RA-GAN network.

[0007] The scheme of the present invention is as follows:

[0008] A: Create a data set and divide it into training set and test set;

[0009] A1: Create a dataset;

[0010] A2: Divide into training set and test set;

[0011] B: Design a parallel double U-shaped residual attention generative adversarial network;

[0012] B1: Design generator;

[0013] B2: Design the discriminator;

[0014] C: Design a composite loss function;

[0015] C1: Calculate GAN loss L GAN ;

[0016] C2: Calculate the log-cosh loss L logc ;

[0017] C3: Calculate the compound loss L

[0018] D: Training and testing the PDU-RA-GAN network;

[0019] D1: training the PDU-RA-GAN network;

[0020] D2: Test the PDU-RA-GAN network;

[0021] The beneficial effects brought by the method proposed by the present invention include at least:

[0022] (1) A single-bit MIMO system channel estimation method based on PDU-RA-GAN, the present invention uses PDU-RA-GAN to learn the mapping from the quantized received signal matrix Y to the channel matrix H. In the parallel double U-shaped network contained in the generator of PDU-RA-GAN, a dilated group convolution is designed to achieve efficient network calculation while capturing multi-scale context information; in the discriminator of PDU-RA-GAN, a residual attention module is designed to enhance the important spatial position information in the input feature map.

[0023] (2) A single-bit MIMO system channel estimation method based on PDU-RA-GAN. The present invention designs a composite loss function, combining the advantages of GAN loss and log-cosh loss. The global distribution characteristics of the generated samples are optimized through GAN loss to ensure the generation quality. The log-cosh loss is used to focus on numerical errors and improve the robustness to outliers and local accuracy. The combination of the two balances the generation quality and numerical fitting performance while ensuring training stability.

[0024] (3) A single-bit MIMO system channel estimation method based on PDU-RA-GAN. Under the mixed noise conditions of Gaussian noise and impulse noise, the PDU-RA-GAN proposed in the present invention shows better channel estimation accuracy than the existing network structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is an overall flow chart of a single-bit MIMO system channel estimation method based on PDU-RA-GAN involved in the present invention;

[0026] Figure 2 It is an overall framework diagram of a single-bit MIMO system channel estimation method based on PDU-RA-GAN involved in the present invention;

[0027] Figure 3 is a structural diagram of the PDU-RA-GAN involved in the present invention;

[0028] Figure 4 It is a specific structural diagram of the dilated group convolution involved in the PDU-RA-GAN involved in the present invention;

[0029] Figure 5 It is a specific structural diagram of the residual attention module involved in the PDU-RA-GAN involved in the present invention;

[0030] Figure 6 It is an overall framework diagram of the PDU-RA-GAN involved in the present invention combined with the composite loss function for training;

[0031] Figure 7 This is a comparison chart of experimental results between a single-bit MIMO system channel estimation method based on PDU-RA-GAN involved in the present invention and some classic methods in the field of channel estimation. DETAILED DESCRIPTION

[0032] To facilitate understanding, the implementation process of the present invention is further described in detail in combination with the technical solution and the accompanying drawings.

[0033] A single-bit MIMO system channel estimation method based on PDU-RA-GAN, the flow chart of its main steps is as shown in the attached Figure 1 Specifically, it includes the following steps:

[0034] A: Create a data set and divide it into training set and test set; it includes two main processes, namely, creating a data set and dividing it into training set and test set;

[0035] A1: Create a data set; this includes two main processes: constructing a channel matrix and calculating a quantized received signal matrix;

[0036] (1) Construct the channel matrix; as shown in the following figure Figure 2 As shown in the figure, N represents the total number of users, each user sends K subcarriers, and the channel matrix corresponding to a user and the base station is The calculation formula is

[0037] H=[h1,…,h k ,…,h K ],(k=1,…,K)

[0038] In the formula, M represents the total number of antennas at the base station, K represents the total number of subcarriers, and k represents the ordinal number of the subcarrier. represents the channel vector in the kth subcarrier, and the calculation formula is

[0039]

[0040] In the formula, L represents the number of paths, P l represents the received power of path l, φ l represents the phase of path l, λ l represents the propagation delay of path l, B represents the system bandwidth, and a(α l,k ,β l,k ) represents the steering vector, a(α l,k ,β l,k ) is calculated as

[0041]

[0042] In the formula, the superscript (·) T represents the transposition operator, m represents the ordinal number of the base station antenna, The calculation formula is

[0043]

[0044] Where, d AS represents the antenna spacing, α l,k and β l,k Respectively represent the azimuth and elevation angle corresponding to the lth path and the kth subcarrier;

[0045] (2) Calculate the quantized received signal matrix; Quantized received signal matrix The calculation formula is

[0046] Y=sgn(Y f )

[0047] Where sgn(·) represents the sign function, Y f represents the received signal matrix, Y f The calculation formula is

[0048] Yf =HΦ+N

[0049] In the formula, represents the pilot signal matrix, represents the mixed noise matrix, τ represents the pilot length, and the calculation formula of the mixed noise matrix is:

[0050] N=Z+V

[0051] In the formula, represents Gaussian noise that follows a Gaussian distribution, It represents the impulse noise that obeys the Alpha stable distribution. This paper uses the generalized signal-to-noise ratio (GSNR) to measure the ratio of the signal intensity to the Alpha stable distribution noise. The calculation formula of the generalized signal-to-noise ratio GSNR is:

[0052]

[0053] Where P s represents the power of the signal, and γ represents the dispersion coefficient of the Alpha stable distribution noise;

[0054] A2: Divide into training set and test set; use normalization to preprocess the data set and divide the data set D into training set and test set;

[0055] B: Design PDU-RA-GAN; as shown in the attached Figure 3 As shown in the figure, the design of Parallel Dual U-Net Residual Attention Generative Adversarial Network (PDU-RA-GAN) includes two main processes, namely, designing the generator and designing the discriminator;

[0056] B1: Design generator; including two main processes, namely designing the pre-processing module (Pre-processingModule) and designing the parallel dual U-net module (Parallel Dual U-Net);

[0057] (1) Design a preprocessing module. The preprocessing module consists of a convolution layer, a batch normalization layer, and a linear rectification function layer. It is responsible for dimensional preprocessing of the quantized received signal matrix Y.

[0058] (2) Design a parallel dual U-shaped network module; the parallel dual U-shaped network module contains two U-shaped network modules in the upper and lower layers. Each U-shaped network module consists of an encoder, a decoder, and a convolutional layer. The encoder consists of four encoding sub-modules, and the decoder consists of five decoding sub-modules, which are responsible for receiving the output of the preprocessing module. All encoding sub-modules in the upper U-shaped network module consist of two convolutional layers and one maximum pooling layer. The first decoding sub-module consists of two convolutional layers and one polynomial interpolation layer, and the following four decoding sub-modules consist of one spatial attention mechanism (SAM), two convolutional layers, and one polynomial interpolation layer. All encoding sub-modules in the lower U-shaped network module consist of two dilated group convolution layers and one maximum pooling layer. The first decoding submodule consists of two convolutional layers and one nearest neighbor interpolation layer. The following four decoding submodules are composed of one channel attention mechanism (CAM), two convolutional layers and one nearest neighbor interpolation layer. Through multi-scale feature extraction, spatial and channel attention mechanisms and flexible interpolation methods, the network's information extraction ability is enhanced.

[0059] B2: Design the discriminator; the discriminator consists of four residual attention modules and two convolutional layers; as shown in the attached Figure 5 As shown in Figure 2, the upper branch of the residual attention module consists of 1 convolutional layer, 1 batch normalization layer, 1 linear rectifier function layer and 1 convolutional block attention module (CBAM), and the lower branch consists of 1 skip connection. The output results of the two branches are added and divided by Batch processing is performed to balance the feature amplitudes of the residual branch and the main branch, improving the convergence performance of the network; the convolutional block attention module consists of channel attention and spatial attention mechanisms;

[0060] C: Design a composite loss function; as shown in the attached Figure 6 As shown, it includes two main processes, namely calculating the GAN loss L GAN , calculate the log-cosh loss L Log and calculate the composite loss L;

[0061] C1: Calculate GAN loss L GAN ; L GAN The calculation formula is

[0062]

[0063] In the formula, E(·) represents the statistical expectation operator, Indicated by A parameterized discriminator that distinguishes the approximate channel matrix generated by the generator and the true channel matrix H, output probability value, G ψ represents a generator parameterized by ψ, which is used to generate an approximate channel matrix that is as similar as possible to the true channel matrix H

[0064] C2: Calculate log-cosh loss L logc ; L logc The calculation formula is

[0065]

[0066] In the formula, h mk represents the element in the mth row and kth column of the real channel matrix H, represents the approximate channel matrix The element in the mth row and kth column of The calculation formula is

[0067]

[0068] Where cosh(·) represents the hyperbolic cosine function operator;

[0069] C3: Calculate the composite loss L; the calculation formula for L is

[0070]

[0071] In the formula, λ logc represents the log-cosh loss L logc The weight parameter of

[0072] D: Training and testing the PDU-RA-GAN network; including two main processes, namely training the PDU-RA-GAN network and testing the PDU-RA-GAN network;

[0073] D1: Training the PDU-RA-GAN network; input the data samples in the training set into the PDU-RA-GAN network, and use the Adam optimization method to train the PDU-RA-GAN network. The loss function uses the composite loss function L in step D until the loss converges and the training ends;

[0074] D2: Test the PDU-RA-GAN network; take the quantized received signal matrix Y in the test set as input and generate an approximate channel matrix through the generator of DU-RA-GAN Calculate the resulting approximate channel matrix The error between the real channel matrix H in the test set is calculated, and the normalized mean square error (NMSE) is used as the evaluation index to evaluate the performance of the channel estimation method proposed in the present invention. The calculation formula of NMSE is:

[0075]

[0076] In the formula, ||·|| F Represents the Frobenius norm operator, and the unit of the calculation result is decibel (dB);

[0077] As attached Figure 7 As shown, the single-bit MIMO channel estimation method based on PDU-RA-GAN proposed in the present invention has higher channel estimation accuracy.

Claims

1. A single-bit MIMO system channel estimation method based on PDU-RA-GAN, characterized in that The following steps are involved: A: Create a data set and divide it into training set and test set; It includes two main processes, namely making a data set and dividing it into a training set and a test set; A1: Create a dataset; It includes two main processes, namely, constructing the channel matrix and calculating the quantized received signal matrix; (1) Constructing a channel matrix; channel matrix The calculation formula is H=[h1,…,h k ,…,h K ],(k=1,…,K) In the formula, M represents the total number of antennas at the base station, K represents the total number of subcarriers, and k represents the ordinal number of the subcarrier. represents the channel vector in the kth subcarrier, and the calculation formula is In the formula, L represents the number of paths, P l represents the received power of path l, φ l represents the phase of path l, λ l represents the propagation delay of path l, B represents the system bandwidth, and a(α l,k ,β l,k ) represents the steering vector, a(α l,k ,β l,k ) is calculated as In the formula, the superscript (·) T represents the transposition operator, m represents the ordinal number of the base station antenna, The calculation formula is Where, d AS represents the antenna spacing, α l,k and β l,k Respectively represent the azimuth and elevation angle corresponding to the lth path and the kth subcarrier; (2) Calculate the quantized received signal matrix; Quantized received signal matrix The calculation formula is Y=sgn(Y f ) Where sgn(·) represents the sign function, Y f represents the received signal matrix, Y f The calculation formula is AND f =HΦ+N In the formula, represents the pilot signal matrix, represents the mixed noise matrix, τ represents the pilot length, and the calculation formula of the mixed noise matrix is: N=Z+V In the formula, represents Gaussian noise that follows a Gaussian distribution, It represents the impulse noise that obeys the Alpha stable distribution. This paper uses the generalized signal-to-noise ratio (GSNR) to measure the ratio of the signal intensity to the Alpha stable distribution noise. The calculation formula of the generalized signal-to-noise ratio GSNR is: Where P s represents the power of the signal, and γ represents the dispersion coefficient of the Alpha stable distribution noise; A2: Divide into training set and test set; use normalization to preprocess the data set and divide the data set into training set and test set; B: Design PDU-RA-GAN; Designing a Parallel Dual U-Net Residual Attention Generative Adversarial Network (PDU-RA-GAN) includes two main processes, namely designing the generator and designing the discriminator; B1: Design generator; It includes two main processes, namely designing the pre-processing module and designing the parallel dual U-Net module; (1) Design a preprocessing module; The preprocessing module consists of a convolution layer, a batch normalization layer, and a linear rectification function layer (ReLU Layer), which is responsible for dimensional preprocessing of the quantized received signal matrix Y; (2) Design a parallel dual U-shaped network module; the parallel dual U-shaped network module contains two U-shaped network modules in the upper and lower layers. Each U-shaped network module consists of an encoder, a decoder, and a convolutional layer. The encoder consists of four encoding sub-modules, and the decoder consists of five decoding sub-modules, which are responsible for receiving the output of the preprocessing module. All encoding sub-modules in the upper U-shaped network module consist of two convolutional layers and one maximum pooling layer. The first decoding sub-module consists of two convolutional layers and one polynomial interpolation layer, and the following four decoding sub-modules consist of one spatial attention mechanism (SAM), two convolutional layers, and one polynomial interpolation layer. All encoding sub-modules in the lower U-shaped network module consist of two dilated group convolution layers and one maximum pooling layer. The first decoding submodule consists of two convolutional layers and one nearest neighbor interpolation layer. The following four decoding submodules are composed of one channel attention mechanism (CAM), two convolutional layers and one nearest neighbor interpolation layer. The network information extraction capability is enhanced through multi-scale feature extraction, spatial and channel attention mechanisms and flexible interpolation methods. B2: Design the discriminator; The discriminator consists of 4 residual attention modules and 2 convolutional layers; the upper branch of the residual attention module consists of 1 convolutional layer, 1 batch normalization layer, 1 linear rectifier function layer and 1 convolutional block attention module (CBAM), and the lower branch consists of 1 skip connection. The output results of the two branches are added and divided by Batch processing is performed to balance the feature amplitudes of the residual branch and the main branch, improving the convergence performance of the network; the convolutional block attention module consists of channel attention and spatial attention mechanisms; C: Design a composite loss function; It includes two main processes, namely calculating GAN loss Calculate the log-cosh loss and calculate compound loss C1: Calculating GAN loss The calculation formula is In the formula, E(·) represents the statistical expectation operator, Indicated by A parameterized discriminator that distinguishes the approximate channel matrix generated by the generator and the true channel matrix H, output probability value, G ψ represents a generator parameterized by ψ, which is used to generate an approximate channel matrix that is as similar as possible to the true channel matrix H C2: Calculate the log-cosh loss The calculation formula is In the formula, h mk represents the element in the mth row and kth column of the real channel matrix H, represents the approximate channel matrix The element in the mth row and kth column of The calculation formula is Where cosh(·) represents the hyperbolic cosine function operator; C3: Calculating compound losses The calculation formula is In the formula, λ logc represents log-cosh loss The weight parameter of D: Training and testing the PDU-RA-GAN network; It includes two main processes, namely training the PDU-RA-GAN network and testing the PDU-RA-GAN network; D1: Training the PDU-RA-GAN network; input the data samples in the training set into the PDU-RA-GAN network, and use the Adam optimization method to train the PDU-RA-GAN network. The loss function uses the composite loss function in step D. Until the loss converges, the training ends; D2: Test the PDU-RA-GAN network; take the quantized received signal matrix Y in the test set as input and generate an approximate channel matrix through the generator of DU-RA-GAN Calculate the resulting approximate channel matrix The error between the real channel matrix H in the test set is calculated, and the normalized mean square error (NMSE) is used as the evaluation index to evaluate the performance of the channel estimation method proposed in the present invention. The calculation formula of NMSE is: In the formula, ||·|| F Represents the Frobenius norm operator. The unit of the calculation result is decibel (dB).

Images