A beam spatial channel estimation method based on DeRePixNet

Abstract

The application discloses a beam space channel estimation method based on DeRePixNet and relates to the technical field of millimeter waves.Compared with previous reconstruction algorithms, the application improves the problems of the slow convergence speed, the low estimation accuracy, the excessive dependence on prior information of channel sparsity and a measurement matrix of the existing reconstruction algorithm, the single convolutional neural network structure and the incapability of the existing convolutional neural network structure to well extract a nonlinear mapping relationship from a measurement signal to an original channel, solves the problems of the excessive dependence on prior knowledge of channel sparsity and an observation matrix in the solving process of the beam space channel estimation based on DeRePixNet, models a channel matrix virtual-real separation into a two-dimensional image, regards a measurement signal as a low-resolution image and an original channel as a high-resolution image, and accurately reconstructs the measurement signal and the original channel by using DeRePixNet.The network structure of the application is relatively simple, and the application range of the algorithm in the field of channel estimation can be further expanded.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of millimeter wave, and particularly relates to a beam space channel estimation method based on DeRePixNet. BACKGROUND

[0002] Millimeter wave massive multiple-input multiple-output (MIMO) technology combines the rich spectrum resources of the millimeter wave frequency band with the array gain of large-scale antenna arrays, fully utilizes the performance advantages of both, and makes up for the path loss and transmission environment sensitivity of the millimeter wave frequency band, becoming an ideal choice for future wireless communication systems to solve the increasing demand for bandwidth. In the millimeter wave massive MIMO system, due to the large number of antennas, channel estimation becomes particularly important. Accurate channel estimation can help the system determine the best beamforming direction to maximize the received power of the signal. In addition, channel estimation can also be used for multi-user interference cancellation. By estimating the channel state information of other users, the system can take appropriate interference suppression measures to improve the capacity and performance of the system.

[0003] In the millimeter wave massive MIMO system, the traditional spatial domain channel can be converted into a beam space channel through a lens antenna array. By utilizing the sparsity of the beam space channel, the channel estimation problem can be further converted into a sparse signal reconstruction problem in the compressive sensing (CS) framework. That is:

[0004]

[0005] Wherein, represents a measurement signal, represents a beam selection network, which can be regarded as an observation matrix here, is a beam space channel matrix in the millimeter wave massive MIMO system, is a Gaussian noise vector. How to recover the original high-dimensional channel from the low-dimensional measurement signal is a difficulty of the existing research technology, and some classic greedy algorithms (such as OMP), approximate message passing algorithms (such as AMP) based on the Bayesian framework and some deep learning-based methods (ReConNet, ConCSNet) have been proposed to solve such problems.

[0006] Traditional sparse signal reconstruction algorithms mostly use structured sparse characteristics to estimate wireless channels, and use iterative optimization strategies to solve underdetermined optimization problems. The intensive calculation of this iterative optimization and the problem of not guaranteeing global optimization have become the bottleneck of CS in the application of sparse channel estimation of large-scale MIMO systems. The weaknesses of these reconstruction algorithms are slow convergence speed, low estimation accuracy and excessive dependence on prior information of channel sparsity and measurement matrix. Some deep learning-based algorithms mostly use standard convolutional neural network (CNN) structure, which is relatively single and cannot well extract the nonlinear mapping relationship from the measured signal to the original channel.

[0007] In order to solve the above problems, the application proposes a channel estimation algorithm based on a deep convolutional reconstruction network (DeRePixNet), which aims to improve the channel estimation accuracy without prior information of the channel. In the network, an MSFM (multi-scale information fusion module) is designed, and the MSFM and the RB (residual block) are organically combined to prevent gradient disappearance while increasing the depth of the network, find the best local sparse structure in the convolutional visual network, and repeat in space. Without the need for prior information of the channel, the DeRePixNet is used to directly solve the inverse transformation process from the measured signal to the original channel in a data-driven manner. SUMMARY

[0008] The application aims to propose a DeRePixNet-based beam space channel estimation method to solve the problems proposed in the background art:

[0009] The existing reconstruction algorithm has slow convergence speed, low estimation accuracy and excessive dependence on prior information of channel sparsity and measurement matrix, and the existing convolutional neural network structure is single and cannot well extract the nonlinear mapping relationship from the measured signal to the original channel.

[0010] In order to achieve the above purpose, the application adopts the following technical solutions:

[0011] A DeRePixNet-based beam space channel estimation method, comprising the following steps:

[0012] S1: Model the real part and the imaginary part of the training data and the label as a two-dimensional image of a double channel;

[0013] S2: Build an initialization module and introduce a trainable matrix to perform dimensionality lifting to obtain an initialized channel as the input of the reconstruction module;

[0014] S3: Construct a reconstruction module composed of a fully connected layer, a convolutional layer, an MSFM and an RB, and input the output of the initialization module into the reconstruction module as the input of the reconstruction module;

[0015] S4: training and estimation of the real part and the imaginary part of the training data and the label based on the initialization module and the reconstruction module built by DeRePixNet, and finally the real part and the imaginary part are spliced and converted into the final complex channel matrix .

[0016] Preferably, in the initialization module of S2, the low-dimensional signal is obtained by an observation matrix The high-dimensional channel is obtained after sampling and superimposing noise, and a trainable matrix is introduced, wherein , is the transpose of the matrix, and the initialized channel is obtained.

[0017] Preferably, in the reconstruction module of S3, the first layer and the third layer use a 5x5 convolution kernel, the fifth layer and the seventh layer use a 7x7 convolution kernel, the sixth layer and the ninth layer use a 1x1 convolution kernel, the second layer and the fourth layer use an organic combination of MSFM and RB, and the eighth layer uses RB; in addition to the first layer and the third layer using a Relu activation function, the remaining single convolution layers do not use a Relu activation function; and different padding is performed on the different sizes of the convolution kernel in the reconstruction module to ensure that the size of the output feature map in all layers remains unchanged.

[0018] Preferably, the MSFM of the reconstruction module is provided with four branches, the feature matrix height and width of each branch are the same, each branch undergoes different convolution operations, different convolution sizes provide different receptive fields, RB is embedded into two middle branches of the MSFM, and finally the channel dimensions of the output matrix of each branch are spliced to realize information fusion.

[0019] Preferably, the RB of the eighth layer is composed of two 3x3 convolution layers, and the output feature matrix is added to the input feature matrix and then output after a Relu activation function.

[0020] Preferably, in the reconstruction module, a linear mapping network is first used to obtain an initial reconstruction channel, and the channel state information obtained by the linear mapping process is an approximate solution, and the corresponding linear mapping matrix is , and the error of is minimized, and the training set contains training samples, that is, , wherein is the measurement value of the received signal, is a millimeter wave beam space channel state information matrix, and the approximate solution obtained by linear mapping is represented as:

[0021]

[0022] wherein, , denotes the transpose of a matrix; denotes a linear mapping;

[0023] The deep neural network composed of convolutional layer, MSFM, RB is further processed to obtain high-precision channel estimation value; the parameters of the convolutional network in the first layer to the ninth layer are updated using the Adam optimizer , and the mean square error is used as the loss function during the training process :

[0024]

[0025] wherein, denotes a linear mapping based on ; denotes a linear mapping based on ;

[0026] The accurate solution obtained is:

[0027]

[0028] wherein, is the accurate solution obtained by the reconstruction module.

[0029] Preferably, the S4 comprises an offline training stage and an online estimation stage; in the offline training stage, the network is trained by supervised learning, and the training data is represented as , wherein is the input of DeRePixNet, is the corresponding label, is the sample number of the training data, and in the training process is first passed through the initialization module for dimension increasing, matched with dimension, and then enters the reconstruction module, optimizes all trainable parameters by minimizing the loss function, and saves the network model with the lowest loss function value on the validation set for testing in the online estimation stage, and stops training when the preset Epoch number is reached; in the online evaluation stage, the measurement signal of the test set is input into the trained DeRePixNet to obtain the beam space channel estimation result ; wherein the measurement measurement signal and the beam space channel state information are complex numbers, and finally the corresponding real part and imaginary part are spliced and converted into the final complex channel matrix .

[0030] Preferably, the DeRePixNet also performs performance evaluation of the beam space channel estimation method based on normalized mean square error as a metric; specifically as follows:

[0031]

[0032] wherein, is the normalized mean square error; represents an expectation operation; represents the true value of the channel parameter; represents the output value of the network model.

[0033] Compared with the prior art, the present application provides a beam space channel estimation method based on DeRePixNet, which has the following beneficial effects:

[0034] The beam space channel estimation based on DeRePixNet proposed by the present application solves the problems of excessive dependence on prior knowledge of channel sparsity and observation matrix in the solving process, and by modeling the real and imaginary parts of the channel matrix into a two-dimensional image, the channel estimation problem can be expressed as a super-resolution reconstruction problem in the field of image processing. The measured signal is regarded as a low-resolution image and the original channel is regarded as a high-resolution image, and DeRePixNet is used to accurately reconstruct it. The network structure of the present application is relatively simple, and can further expand the application range of the algorithm in the field of channel estimation. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a network structure diagram of the reconstruction module mentioned in embodiment 1 of the present application;

[0036] Figure 2 is a structure diagram of MSFM+RB mentioned in embodiment 1 of the present application;

[0037] Figure 3 is a structure diagram of RB mentioned in embodiment 1 of the present application;

[0038] Figure 4 is a performance comparison diagram of NMSE of DeRePixNet and other algorithms mentioned in embodiment 2 of the present application. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments.

[0040] The beam space channel estimation based on the DeRePixNet solves the problems of excessive dependence on prior knowledge of channel sparsity and an observation matrix in a solving process, and by modeling a real and imaginary part of a channel matrix into a two-dimensional image, a channel estimation problem can be expressed as a super-resolution reconstruction problem in an image processing field. The measured signal is regarded as a low-resolution image and the original channel is regarded as a high-resolution image, and the DeRePixNet is used for accurate reconstruction. The network structure is relatively simple, and the application range of the algorithm in the channel estimation field can be further expanded. Specifically, the following contents are included.

[0041] Embodiment 1

[0042] Please refer to Figures 1-3 The beam space channel estimation method based on the DeRePixNet comprises the following steps.

[0043] S1: modeling the real part and the imaginary part of the training data and the label into a two-dimensional image of a double channel; specifically as follows:

[0044] Since the neural network usually takes real numbers as input, the data needs to be preprocessed, that is, the real part and the imaginary part of the training data and the label are modeled into a two-dimensional image of a double channel.

[0045] S2: building an initialization module, and introducing a trainable matrix to perform dimensionality elevation to obtain an initialized channel as an input of a reconstruction module; specifically as follows:

[0046] In the initialization module, since the final measured signal is obtained by an observation matrix , the high-dimensional channel is obtained after sampling and superimposing noise, in order to solve the dimension mismatch problem between and , a trainable matrix is introduced, wherein , to obtain the initialized channel as an input of the reconstruction.

[0047] S3: constructing a reconstruction module composed of a full connection layer, a convolution layer, an MSFM and an RB respectively, and taking the output of the initialization module as an input of the reconstruction module; specifically as follows:

[0048] In the reconstruction module, the first layer and the third layer use a 5x5 convolution kernel, the fifth layer and the seventh layer use a 7x7 convolution kernel, the sixth layer and the ninth layer use a 1x1 convolution kernel, the second layer and the fourth layer use the organic combination of MSFM and RB, and the eighth layer uses RB; except that the first layer and the third layer use the Relu activation function, the remaining single convolution layer does not use the Relu activation function, and different padding is performed on the different sizes of the convolution kernel in the reconstruction module to ensure that the size of the output feature map in all layers remains unchanged.

[0049] With reference to Figure 1 The full connection layer adopts a 256x256 linear mapping matrix to obtain a channel approximate solution. The first layer and the third layer use a 5x5 convolution kernel, and the output channel numbers are 10 and 20 respectively; the fifth layer and the seventh layer use a 7x7 convolution kernel, and the output channel numbers are 64 and 16 respectively; the sixth layer and the ninth layer use a 1x1 convolution kernel, and the output channel numbers are 32 and 2 respectively. Except that the first layer and the third layer need the activation function Relu, the remaining layers do not use the activation function. In addition, different padding is performed on different sizes of the convolution kernel to ensure that the size of the output feature map in all layers remains unchanged. The second layer and the fourth layer use the organic combination of MSFM and RB:

[0050] With reference to Figure 2 The MSFM of the reconstruction module is provided with four branches, wherein the first branch uses a 1x1 convolution layer, the second branch uses a 1x1 convolution layer stacked with a 5x5 convolution layer, the third branch uses a 1x1 convolution layer stacked with two 3x3 convolution layers, and the fourth branch uses an average pooling layer stacked with a 1x1 convolution layer; the RB is embedded into the two middle branches of the MSFM, i.e. the second branch and the third branch, and finally the channel dimensions of the output matrices of the branches are spliced to realize information fusion. The output feature matrix height and width of each branch are the same, each branch is subjected to different convolution operations, different convolution sizes provide different receptive fields, and different levels of feature extraction can be performed. The pooling operation itself has the function of extracting features, and because there is no parameter, overfitting will not occur, so the pooling operation is also a branch of this module. The input feature map is obtained through the four branches in parallel to obtain four outputs. Since the middle two branches of the MSFM are deep, embedding the residual block into the middle branches of the module effectively alleviates the gradient vanishing problem caused by increasing the depth of the branch network. In order to enable the outputs of the four branches to be spliced in the depth direction, it is necessary to ensure that the feature matrix height and width of the outputs of the four branches are the same, so different padding operations are performed on the different sizes of the convolution kernel in the MSFM to ensure that the size of the output feature map of all branches remains unchanged.

[0051] The eighth layer uses a single RB:

[0052] With reference toFigure 3 The feature matrix is outputted by two 3x3 convolution layers, and then added with the input and outputted by the activation function Relu. The shapes of the output and input feature matrices must be one-to-one corresponding. Finally, the output matrices of each branch are spliced in the depth dimension (channel dimension) to realize efficient fusion of information.

[0053] First, a linear mapping network is used to obtain the initial reconstructed channel, and the channel state information obtained by the linear mapping process is an approximate solution, and the corresponding linear mapping matrix is , and the error of is minimized. The training set contains training samples, i.e. , wherein is the measurement value of the received signal, is the millimeter wave beam space channel state information matrix, and the approximate solution obtained by linear mapping can be expressed as:

[0054]

[0055] wherein, , represents the transpose of the matrix; Table linear mapping. In order to obtain a high-precision channel estimation value, further processing is required, i.e. adding a deep neural network composed of convolution layers, MSFM and RB. The Adam optimizer is used to update the parameters of the convolution network in the first to ninth layers , and the mean square error is used as the loss function during the training process:

[0056]

[0057] wherein, represents linear mapping based on ; represents linear mapping based on ;

[0058] The accurate solution obtained is:

[0059]

[0060] wherein, is the accurate solution obtained by the reconstruction module.

[0061] S4: The DeRePixNet built based on the initialization module and the reconstruction module is used to train and estimate the real part and the imaginary part of the training data and the label respectively, and finally the real part and the imaginary part are spliced to convert into the final complex channel matrix . Specifically as follows:

[0062] The offline training phase trains the network using supervised learning, and the training data is represented as , where is the input of DeRePixNet, is the corresponding label, is the number of samples of the training data, and in the training process is first passed through the initialization module for dimensionality increase, and is matched in dimension to enter the reconstruction module, optimize all trainable parameters by minimizing the loss function, and save the network model with the lowest loss function value on the validation set for testing in the online estimation phase, stop training when the preset number of Epochs is reached; in the online evaluation phase, input the measurement signal of the test set into the trained DeRePixNet to obtain the beam space channel estimation result ; where the measurement signal and the beam space channel state information are complex numbers, so the data is preprocessed. The real and imaginary parts of the training data and labels are modeled as two-dimensional images of two channels, and finally the real and imaginary parts are spliced and converted into the final complex channel matrix .

[0063] Embodiment 2:

[0064] Compared with traditional algorithms and deep learning-based algorithms, the beam space channel evaluation based on normalized mean square error as the measurement standard:

[0065]

[0066] where, denotes the expectation operation; denotes the true value of the channel parameter; denotes the output value of the network model.

[0067] The simulation scenario is selected as the uplink transmission link of the TDD millimeter wave massive MIMO system. It is assumed that the base station is placed with a lens antenna array with N = 256 antennas, and the antenna type is a uniform planar antenna array (UPA), and the channel matrix size is 16 × 16. In order to facilitate the study, the channel matrix of the UPA is vectorized. The number of RF links , the number of users , and both are single-antenna users, and the length of the measured signal is . The signal-to-noise ratio of the uplink estimation is . The channel parameter settings are: the number of effective paths ; complex gain ; azimuth and elevation range Under the above parameter settings, the sparse beam space channel and the measured signal under different signal-to-noise ratios are generated according to the background technology. When the traditional orthogonal matching pursuit algorithm is used for channel estimation, the sparsity of the beam space channel vector is set to 4 When the channel estimation based on the approximate message passing algorithm is used, the number of iterations is 10, and the empirical shrinkage parameter of each iteration is set to 0.1 .

[0068] The generation of the data set is carried out on the Matlab software, and the sizes of the training set, the validation set and the test set are 80000, 2000 and 2000 respectively. The Epoch of training is set to 150, the initial learning rate is set to 0.0001, the attenuation factor is 0.1, and the attenuation period is 100. The batch size is set to 16, and the Adam optimizer is used to optimize the neural network parameters. The embodiment is carried out on the PyTorch deep learning framework and the Matlab simulation software, and the simulation is carried out on a GPU server with RTX3080Ti.

[0069] The simulation results of the traditional OMP algorithm, the AMP algorithm, the model-driven deep learning algorithm LAMP, GA-LAMP, the data-driven deep learning algorithm RePixNet, U-Net and the DeRePixNet algorithm of the present application are compared, and the specific simulation results can be referred to Figure 4 As can be seen from the figure, the performance of the beam space channel estimation is improved with the increase of the signal-to-noise ratio; the NMSE performance of the traditional OMP and AMP algorithms is poor, and LAMP, GM-LAMP and DeRePixNet all have better performance; at the same time, when the signal-to-noise ratio is 10dB, the performance of DeRePixNet is improved by 1.81dB compared with GM-LAMP, by 2.29dB compared with LAMP, by 10.74dB and 8.78dB compared with OMP and AMP respectively, and when the signal-to-noise ratio is 20dB, DeRePixNet is improved by 8.63dB and 7.92dB compared with RePixNet and U-Net respectively. Compared with the existing algorithms, the DeRePixNet network of the present application has a great improvement in channel estimation accuracy in all signal-to-noise ratio ranges. Since the orthogonal pilot transmission strategy is used between multiple users, the system can distinguish the pilot signals of different users without being affected by the interference between other users. Therefore, the NMSE performance of the beam space channel estimation is independent of the number of users.

[0070] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A beam spatial channel estimation method based on DeRePixNet, characterized in that, Includes the following steps: S1: Separate the real and imaginary parts of the training data and labels to model a two-channel two-dimensional image; S2: Set up the initialization module and import trainable matrices. Dimensional upscaling is performed to obtain the initial channel. As input to the refactoring module; S3: Construct a reconstruction module consisting of fully connected layers, convolutional layers, MSFM, and RB, and use the output of the initialization module as the input of the reconstruction module; in the reconstruction module, a linear mapping network is first used. The initial reconstructed channel is obtained, and the channel state information is obtained through the linear mapping process. As an approximate solution, its corresponding linear mapping matrix is: and make The error is minimized, and the training set contains training samples, i.e. ,in, The measured value of the received signal. The approximate solution obtained by linear mapping for the millimeter-wave beam space channel state information matrix is ​​as follows: in, , Represents the transpose of a matrix; Represents a linear mapping; A deep neural network consisting of convolutional layers, MSFM, and RB is added for further processing to obtain high-precision channel estimates; the Adam optimizer is used to update the parameters of the convolutional networks from the first to the ninth layer. Mean squared error is used as the loss function during training. : in, Indicates based on A linear mapping; Indicates based on A linear mapping; The exact solution obtained is: in, The exact solution obtained from the reconstructed module; S4: DeRePixNet, built upon the initialization and reconstruction modules, trains and estimates the real and imaginary parts of the training data and labels separately. Finally, it concatenates the real and imaginary parts to convert them into the final complex channel matrix. S4 includes an offline training phase and an online estimation phase; in the offline training phase, supervised learning is used to train the network, and the training data is represented as follows: ,in As input to DeRePixNet, For the corresponding tags, The number of samples in the training data during the training process. First, the dimensionality is increased through the initialization module, and... After dimensionality matching, the network proceeds to the reconstruction module. It optimizes all trainable parameters by minimizing the loss function and saves the network model with the lowest loss function value on the validation set for testing during the online estimation phase. Training stops when a preset number of epochs are reached. During the online evaluation phase, the measurement signals from the test set are input into the trained DeRePixNet to obtain the beam space channel estimation results. Among them, the measurement signal and beam space channel state information All are complex numbers, and finally the corresponding real and imaginary parts are concatenated to convert them into the final complex channel matrix. .

2. The beam spatial channel estimation method based on DeRePixNet according to claim 1, characterized in that, In the initialization module of S2, low-dimensional signals From the observation matrix High-dimensional channels The sample is obtained by adding noise, and a trainable matrix is ​​introduced. ,in , The transpose of the matrix yields the initialized channel. As input to the refactoring module.

3. The beam spatial channel estimation method based on DeRePixNet according to claim 1, characterized in that, In the reconstruction module of S3, the first and third layers both use 5×5 convolution kernels, the fifth and seventh layers both use 7×7 convolution kernels, the sixth and ninth layers both use 1×1 convolution kernels, the second and fourth layers both use an organic combination of MSFM and RB, and the eighth layer uses RB. Except for the first and third layers which use the ReLU activation function, the remaining individual convolutional layers do not use the ReLU activation function; in addition, different padding is applied to the convolutional kernels of different sizes in the reconstruction module to ensure that the size of the output feature map remains unchanged in all layers.

4. The beam spatial channel estimation method based on DeRePixNet according to claim 3, characterized in that, The MSFM in the reconstruction module has four branches, and the output feature matrix of each branch has the same height and width. Each branch undergoes different convolution operations, and different convolution sizes provide different receptive fields. RB is embedded into the two middle branches of MSFM. Finally, the channel dimensions of the output matrices of each branch are spliced ​​to achieve information fusion.

5. The beam spatial channel estimation method based on DeRePixNet according to claim 4, characterized in that, The eighth layer, RB, consists of two 3×3 convolutional layers. The output feature matrix is ​​added to the input feature matrix and then passed through the ReLU activation function before being output.

6. The beam spatial channel estimation method based on DeRePixNet according to claim 5, characterized in that, The DeRePixNet method also performs performance evaluation of the beam space channel estimation method based on the normalized mean square error as a metric; specifically as follows: in, This is the normalized mean square error; This represents the expectation operation; This represents the true value of the channel parameters; This represents the output value of the network model.

Images