Large-scale MIMO non-sparse channel state information feedback and reconstruction method

Abstract

The invention discloses a large-scale MIMO non-sparse channel state information feedback and reconstruction method, and belongs to the technical field of communication. Based on a deep learning technology and a traditional large-scale MIMO communication system channel state information feedback thought, the invention provides a neural network model SCF-Net for realizing large-scale MIMO non-sparse channel state information feedback, and the large-scale MIMO communication system channel state information is regarded as a grayscale image. A large-scale MIMO communication system channel state information feedback and reconstruction problem is converted into a lossy image restoration problem, and channel state information feedback and reconstruction performance exceeding a traditional compressed sensing algorithm is obtained under the support of a neural network model, data and computing power. According to the method, the problem that the existing channel state information feedback technology excessively depends on channel sparsity can be effectively solved, and the large-scale MIMO non-sparse channel state information feedback and reconstruction capability can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology and relates to a channel state information feedback method for a wireless communication system, in particular to a large-scale MIMO non-sparse channel state information feedback and reconstruction method. Background Art

[0002] With the rapid development of wireless communication technology, demands for data transmission speeds, system capacity, and communication quality are increasing. Against this backdrop, massive multiple-input, multiple-output (MIMO) technology has attracted considerable attention due to its significant performance improvements. By deploying a large number of antennas (typically dozens or even hundreds) at both the transmitter and receiver ends, massive MIMO significantly improves data transmission speeds and spectral efficiency, making it a key enabling technology for the next generation of wireless communications.

[0003] However, as the number of antennas increases, massive MIMO communication systems also face many challenges. Among them, the acquisition and feedback of Channel State Information (CSI) is particularly critical. CSI describes the attenuation, phase offset and other characteristics of the signal during transmission, and is the basis for base stations to perform signal preprocessing and achieve efficient transmission. However, in massive MIMO systems, due to the large number of antennas, the number of channel parameters that need to be estimated and fed back increases dramatically. Traditional CSI feedback algorithms based on codebooks, vector quantization, and compressed sensing face problems such as high computational complexity, huge system feedback overhead, and strict requirements on channel sparsity, making it difficult to meet the needs of massive MIMO systems.

[0004] Research on CSI feedback methods for massive MIMO communication systems has important scientific research and engineering practical significance, specifically in the following aspects:

[0005] 1. Improving Communication System Performance: Accurate CSI feedback is crucial to the performance of massive MIMO communication systems. Precise CSI information enables the transmitter to implement more precise signal preprocessing measures such as beamforming and power allocation, thereby increasing data rates, reducing interference, and enhancing system capacity. Therefore, researching efficient CSI feedback techniques is crucial for improving the overall performance of massive MIMO systems.

[0006] 2. Reducing Feedback Overhead: In massive MIMO systems, due to the large number of antennas, directly feeding back complete CSI information incurs significant overhead. This not only increases system complexity but also potentially impacts the real-time nature of data transmission. Therefore, research on how to reduce feedback overhead while ensuring CSI accuracy has become an urgent issue.

[0007] 3. Promoting technological innovation: The development of massive MIMO technology is inseparable from the continuous research and innovation of CSI feedback technology. With the continuous advancement of wireless communication technology, new application scenarios and communication needs continue to emerge. To meet these challenges, it is necessary to continuously explore new CSI feedback technologies and methods.

[0008] 4. Promote industry upgrades: As a key technology for next-generation wireless communication systems, the development of massive MIMO technology will drive the upgrade and transformation of the entire communications industry chain. In-depth research on CSI feedback technology can promote the development of related hardware equipment, optimize software algorithms, and transform network architectures, thereby accelerating the innovative development of the communications industry.

[0009] In recent years, breakthroughs in computing power and hardware have enabled rapid advancements in artificial intelligence (AI), particularly deep learning, and have successfully enabled technological breakthroughs in multiple scientific fields. Consequently, some researchers have gradually attempted to introduce deep learning into the field of CSI feedback for massive MIMO communication systems. By first compressing feedback and then decompressing and reconstructing it, they are implementing intelligent CSI feedback and reconstruction for massive MIMO communication systems. Based on this foundation, further research has led to the development of numerous intelligent CSI feedback and reconstruction methods for massive MIMO communication systems.

[0010] Chinese patent application number CN110350958A discloses a neural network-based CSI multi-rate compression feedback method for massive MIMO. The method proposes two network design principles for CSI feedback, introduces a new network architecture, CsiNet+, based on these principles, and introduces two different variable-rate compression frameworks: SM-CsiNet+, a serial multi-rate compression framework, and PM-CsiNet+, a parallel multi-rate compression framework. This method requires that the massive MIMO channel must be strictly sparse, which is almost impossible to achieve in practical engineering, and the CSI feedback accuracy of this method is relatively low.

[0011] Chinese patent application number CN108847876A discloses a method for compressed feedback and reconstruction of time-varying channel state information for large-scale MIMO. The method includes obtaining a channel matrix sequence, performing a DFT transform on each of T channel matrices to obtain a sparse channel matrix sequence in the angular delay domain; constructing a channel feedback and reconstruction model CsiNet-LSTM, inputting the channel matrix sequence into an encoder to output codewords; and sequentially inputting the codewords into a decoder to output a reconstructed channel matrix sequence; training the channel feedback and reconstruction model so that the reconstructed channel matrix sequence gradually approaches the channel matrix sequence to obtain model parameters; performing a two-dimensional inverse DFT transform on each channel matrix in the output reconstructed channel matrix sequence to recover the reconstructed value of the original space-frequency domain matrix sequence; inputting the channel state information to be fed back and reconstructed into the model to output the reconstructed value. This method also requires that the channel must be sparse, and the model computational complexity is extremely high. Summary of the Invention

[0012] In order to solve the above problems, the present invention proposes a large-scale MIMO non-sparse channel state information feedback and reconstruction method. This method is based on a deep learning neural network and can solve the problems of low CSI feedback and reconstruction accuracy and strict requirements on channel sparsity in large-scale MIMO communication systems, thereby improving the CSI feedback and reconstruction performance in large-scale MIMO communication systems.

[0013] In order to achieve the above object, the present invention provides the following technical solutions:

[0014] A method for massive MIMO non-sparse channel state information feedback and reconstruction, comprising the following steps:

[0015] Step S1: Simulate and construct a massive MIMO non-sparse communication system, obtain the channel state information matrix H, preprocess the channel state information matrix H, and divide it into a training data set, a verification data set, and a test data set according to the ratio;

[0016] Step S2: constructing a deep neural network model SCF-Net for realizing non-sparse channel state information feedback and reconstruction;

[0017] Step S3: The deep neural network model SCF-Net constructed in step S2 is trained on the training data set and the validation data set constructed in step S1, so that the channel state information matrix output by the model is Approach the actual channel state information matrix H, save the model weight parameter file generated in each generation during the training process, observe the decline of the training data loss function and the validation data loss function until the loss function fully converges and the model training is completed;

[0018] Step S4: Load the model weight parameter files of each generation saved in step S3 respectively, and test the model performance on the test dataset;

[0019] Step S5: Select the deep neural network model with the best test performance for channel state information feedback and reconstruction, deploy the deep neural network model at the transmitting end, and compress the original channel state information matrix H at the receiving end according to the compression rate γ, and send the compressed channel state information matrix to the transmitting end through the feedback link. The transmitting end inputs the feedback data into the deep neural network model to reconstruct the original channel state information before compression.

[0020] Furthermore, in step S1, a massive MIMO non-sparse communication system is simulated and constructed, specifically in the following manner:

[0021] Step S111: Using MATLAB simulation software to perform system modeling, the massive MIMO non-sparse communication system adopts a symmetrical structure of the transmitting and receiving ends. In the simulation software, the number of transmitting and receiving antennas is set to 16, arranged in a 4×4 structure, and the antenna spacing is set to 1 meter;

[0022] Step S112: The transmitting end sends an orthogonal pilot sequence that does not change with time slots Where i is the time slot;

[0023] Step S113: The channel is modeled using the Kronecker channel model. The number of channel multipaths is set to 7. The channel is represented as

[0024]

[0025] Among them, H w It is a Rayleigh channel with independent channel coefficients, generated by the sine wave superposition method;

[0026] For the correlation matrix R between antennas Rx and R Tx The modeling adopts the classic exponential correlation model, that is, R Rx and R Tx The entry in row i and column j of is modeled as:

[0027]

[0028] Among them, d i,j 30 represents the distance from the i-th antenna to the j-th antenna, d0>0 represents the spacing between adjacent antennas, c i,j =d i,j / d0 is a constant, and a∈[0,1] are parameters;

[0029] Step S114: The transmission signal generated in step S112 is received by the receiving end after passing through the channel constructed in step S113. The receiving end obtains the channel state information through channel estimation based on the received signal. That is, the channel state information required for channel state information feedback and reconstruction, or, in the simulation process, the ideal perfect channel state information is directly saved to realize channel state information feedback and reconstruction.

[0030] Furthermore, in step S1, the channel state information matrix H is obtained by channel estimation at the receiving end of the massive MIMO non-sparse communication system constructed by simulation, or is an ideal known channel state information matrix directly obtained during the simulation process.

[0031] Furthermore, in step S1, the channel state information matrix H is preprocessed and divided into a training data set, a verification data set, and a test data set according to a ratio, specifically in the following manner:

[0032] Step S121: Converting the one-dimensional channel state information obtained by channel estimation at the receiving end of the simulated massive MIMO non-sparse communication system or the ideal known channel state information directly obtained into a channel state information matrix H of size N×N, where N is the number of antennas equipped at the transceiver end, and each matrix element in the channel state information matrix H is complex data;

[0033] Step S122: Convert the channel state information matrix H into In this form, the complex matrix of size N×N is converted into a real matrix of size 2N×2N, and then the matrix of size N×N is restored through inverse transformation after being output from the deep neural network model.

[0034] Step S123: compressing the real-number channel state information matrix H′ obtained in step S122 according to the compression ratio γ to generate a low-dimensional channel state information matrix H″ of size 2N×2N×γ, where γ<1;

[0035] Step S124: Repeat steps S121 to S123 to generate multiple sets of data under simulation conditions, each set of data consists of an H and an H″, where H″ is the data to be reconstructed by the deep neural network model, and H is the original label data, which is also the ideal target output of the deep neural network model;

[0036] Step S125: Divide the data generated in step S124 into a training data set, a validation data set, and a test data set in a ratio of 7:1:2.

[0037] Furthermore, the deep neural network model SCF-Net includes an input layer, an intermediate layer and an output layer:

[0038] The input layer is a convolutional layer that uses 64 3*3 convolution kernels to extract important features contained in the channel state information matrix. A nonlinear activation function is added after the convolution layer, and the LeakyReLU activation function is used as the nonlinear activation function.

[0039] The middle layer contains two attention-assisted feature extraction subnetworks with the same structure connected in series. The attention-assisted feature extraction subnetwork contains two convolutional layers. Both convolutional layers use 64 3*3 convolution kernels for feature extraction. A LeakyReLU activation function and random dropout operation are added after the first convolutional layer, and a LeakyReLU activation function is added after the second convolutional layer. Then it is connected to the attention mechanism module for feature refinement. A random dropout operation is added after the attention mechanism module. The input of the attention-assisted feature extraction subnetwork and the output of the attention mechanism module are added through a residual jump connection structure as the output of the attention-assisted feature extraction subnetwork; the output of the second attention-assisted feature extraction subnetwork passes through a 3*3 convolutional layer, a LeakyReLU activation function and a random dropout operation, and then added to the output value of the input layer in the form of a residual jump connection structure as the output of the middle layer;

[0040] The output layer is a 3*3 convolutional layer.

[0041] Furthermore, the attention mechanism module is a CBAM attention module, which includes two parts: a channel attention module and a spatial attention module. Given an intermediate feature map, the CBAM attention module infers the attention map in sequence according to the two independent dimensions of channel and space, and then multiplies the attention map into the input feature map for adaptive feature refinement; wherein, the channel attention module uses the feature relationship between channels to generate a channel attention feature map, which is used to find the key input features that the network needs to pay attention to; the spatial attention module generates a spatial attention feature map through the spatial relationship of the feature space, which is used to find the location of important information features that need to be extracted. The channel attention module and the spatial attention module are arranged in series to form the CBAM attention module.

[0042] Furthermore, in step S3, when training the deep neural network model SCF-Net, the Adam gradient optimization algorithm and the end-to-end learning method are used to jointly train the neural network parameters so as to minimize the loss function;

[0043] The loss function is the mean square error loss, specifically the original channel matrix estimate reconstructed by the deep neural network model The mean square error between H and the true label, the loss function expression is:

[0044]

[0045] Where T is the total number of samples in the training set, ||·||2 is the Euclidean norm, and the subscript i represents the number of the reconstructed channel matrix and the original channel matrix.

[0046] Furthermore, in step S4, the performance of the test model is measured using the normalized mean square error (NMSE), which is calculated as follows:

[0047]

[0048] Among them, E represents the mathematical expectation, and H represent the channel matrix recovered by the neural network and the true channel matrix respectively.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] (1) The present invention converts the channel state information feedback and reconstruction problem into a lossy image restoration task, and realizes channel state information feedback and reconstruction with the help of a neural network model. It does not rely on channel sparsity and can solve the channel state information feedback problem in non-sparse channels.

[0051] (2) The deep neural network model SCF-Net used in the present invention to realize large-scale MIMO non-sparse channel state information feedback improves the performance of the neural network by introducing the LeakyReLU nonlinear activation function and the attention mechanism module. By designing the residual jump connection structure and random drop operation, the overfitting and shallow feature loss of the model are avoided.

[0052] (3) The present invention splits the complex channel matrix into a real channel matrix through a channel state information preprocessing method, which effectively solves the problem that neural networks are not good at processing complex data and improves the algorithm effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of a method for feedback and reconstruction of state information of a massive MIMO non-sparse channel.

[0054] Figure 2 It is a structural diagram of the SCF-Net neural network model proposed in this invention.

[0055] Figure 3 yes Figure 2 Schematic diagram of the structure of the attention-assisted feature extraction subnetwork.

[0056] Figure 4 This is a diagram showing the simulation results of some algorithms performance of the present invention. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0058] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the described features. In the description of this application, "plurality" means two or more, unless otherwise expressly specified.

[0059] A method for feedback and reconstruction of large-scale MIMO non-sparse channel state information, which is implemented based on a deep neural network model. The channel state information of a large-scale MIMO communication system is regarded as a grayscale image, and the problem of feedback and reconstruction of the channel state information of a large-scale MIMO communication system is converted into a lossy image restoration problem. With the support of data and computing power, the channel state information feedback and reconstruction performance that exceeds that of traditional compressed sensing algorithms is obtained. The method includes acquisition and preprocessing of channel state information data of a large-scale MIMO non-sparse communication system, construction of a neural network model, and training and testing of the neural network model. Figure 1 As shown, the method specifically includes the following steps:

[0060] Step S1: Dataset acquisition: simulate and construct a large-scale MIMO non-sparse communication system, obtain the channel state information matrix H, preprocess the channel state information matrix H, and divide it into training data set, verification data set, and test data set according to the proportion;

[0061] Step S2: constructing a deep neural network model SCF-Net for realizing non-sparse channel state information feedback and reconstruction;

[0062] Step S3: The deep neural network model SCF-Net constructed in step S2 is trained on the training data set and the validation data set constructed in step S1, so that the channel state information matrix output by the model is Keep the model weight parameter file generated during the training process as close as possible to the actual channel state information matrix H, and observe the decrease of the training data loss function and the validation data loss function until the loss function fully converges and the model training is completed;

[0063] Step S4: Load the model weight parameter file saved in step S3 and test the model performance on the test dataset;

[0064] Step S5: Select the neural network model with the best test performance for channel state information feedback and reconstruction, deploy the neural network model at the transmitting end, and compress the original channel state information matrix H at the receiving end according to the compression rate γ, and send the compressed channel state information matrix to the transmitting end through the feedback link. The transmitting end inputs the feedback data into the neural network model, and finally reconstructs the original channel state information before compression.

[0065] The massive MIMO communication system is constructed using software simulation methods, and the transmitting and receiving ends have a symmetrical structure. The specific method is as follows:

[0066] Step S111: Using MATLAB simulation software to perform system modeling, the massive MIMO non-sparse communication system adopts a symmetrical structure of the transmitting and receiving ends. In the simulation software, the number of transmitting and receiving antennas is set to 16, arranged in a 4×4 structure, and the antenna spacing is set to 1 meter;

[0067] Step S112: The transmitting end sends an orthogonal pilot sequence that does not change with time slots Where i is the time slot;

[0068] Step S113: Modeling the channel. Since antenna correlation may exist between large-scale antenna arrays, this embodiment uses the Kronecker channel model to model the channel. The number of channel multipaths is set to 7. At this time, the channel can be expressed as

[0069]

[0070] Among them, H w The channel coefficients are independent of each other in the Rayleigh channel, which is generated by the sine wave superposition method; for the correlation matrix R between antennas Rx and R Tx The modeling adopts the classic exponential correlation model, that is, R Rx and R Tx The entry in row i and column j of

[0071]

[0072] where d i,j 30 represents the distance from the i-th antenna to the j-th antenna, d0>0 represents the spacing between adjacent antennas, c i,j =d i,j / d0 is a constant, and a∈[0,1] are parameters;

[0073] Step S114: The transmission signal generated in step S112 is received by the receiving end after passing through the channel constructed in step S113. The receiving end obtains the channel state information through channel estimation based on the received signal. That is, the channel state information required for channel state information feedback and reconstruction. The ideal perfect channel state information can also be directly saved during the simulation process to realize channel state information feedback and reconstruction.

[0074] The channel state information matrix H in this embodiment is obtained by channel estimation at the receiving end, and may also be an ideal known channel state information matrix directly obtained during the simulation modeling process.

[0075] The preprocessing of the channel state information matrix H includes transforming the channel state information data to be fed back into a data format suitable for direct processing by the neural network, and constructing a data set based on this data in accordance with the data ratio. The specific steps include:

[0076] Step S121: The receiving end converts the one-dimensional channel state information obtained by channel estimation or the directly obtained ideal known channel state information into matrix data of size N×N to facilitate feature extraction and learning by the convolutional layer of the neural network, where N is the number of antennas equipped at the transmitting and receiving end;

[0077] Step S122: Each matrix element in the channel state information matrix obtained in step S1 is complex data. Since the neural network model is not good at processing complex data, the channel matrix H is converted into In other words, the complex matrix of size N×N is converted into a real matrix of size 2N×2N, and after being output from the neural network, the channel matrix of size N×N is restored according to the inverse transformation of the above transformation;

[0078] Step S123: compressing the real-number channel state information matrix H′ obtained in step S122 according to the compression ratio to generate a low-dimensional channel state information matrix H″. Assuming the compression ratio is γ, the size of the compressed low-dimensional channel state information matrix H″ is 2N×2N×γ, where γ<1;

[0079] Step S124: Repeat steps S121 to S123 to generate a data set under simulation conditions. The data in the data set are all in pairs. The data compressed in step S123 is the data to be reconstructed by the neural network, and the data not compressed in step S123 is the original label data, which is also the ideal target output of the neural network.

[0080] Step S125: Divide the data set generated in step S124 into a ratio of 7:1:2 to obtain training, verification, and test data sets.

[0081] like Figure 2 and 3 As shown in the figure, the SCF-Net neural network model contains a total of 7 convolutional layers, including:

[0082] The input layer is a convolutional layer that uses 64 3*3 convolution kernels to extract important features contained in the channel information matrix. A nonlinear activation function is added after the convolution layer. Since the elements in the channel information matrix contain both positive and negative numbers, the LeakyReLU activation function is used to prevent the "dead neurons" phenomenon. The output of the input layer is added to the output of the intermediate layer through a residual skip connection structure.

[0083] The middle layer contains two identical attention-assisted feature extraction subnetworks. The attention-assisted feature extraction subnetwork contains two convolutional layers, both of which use 64 3*3 convolution kernels for feature extraction. A LeakyReLU activation function and random dropout operation are added after each convolutional layer. The output after the second random dropout operation is connected to the attention mechanism module for feature refinement. The input of the attention-assisted feature extraction subnetwork and the output of the attention mechanism module are added through a residual jump connection structure as the output of the attention-assisted feature extraction subnetwork; the output of the second attention-assisted feature extraction subnetwork passes through a 3*3 convolutional layer, a LeakyReLU activation function and a random dropout operation, and then is added to the output value of the model input layer in the form of a residual jump connection structure; the output layer is a 3*3 convolutional layer.

[0084] The attention mechanism module is the CBAM attention module, which consists of two parts: the channel attention module and the spatial attention module. Given an intermediate feature map, the CBAM attention module infers the attention map in sequence according to the two independent dimensions of channel and space, and then multiplies the attention map into the input feature map for adaptive feature refinement; among them, the channel attention module uses the feature relationship between channels to generate a channel attention feature map, which is used to find the key input features that the network needs to pay attention to; the spatial attention module generates a spatial attention feature map through the spatial relationship of the feature space, which is used to find the location of important information features that need to be extracted. The channel attention module and the spatial attention module are arranged in series to form the CBAM attention module.

[0085] The training and testing of the above neural network model uses the Adam gradient optimization algorithm and end-to-end learning method to jointly train the neural network parameters to minimize the loss function;

[0086] The loss function uses the mean square error loss, which is the original channel matrix estimate reconstructed by the neural network model The mean square error between H and the true label, the loss function expression is:

[0087]

[0088] Where T is the total number of samples in the training set, ||·||2 is the Euclidean norm, and the subscript i represents the number of the reconstructed channel matrix and the original channel matrix;

[0089] The constructed neural network SCF-Net is placed on the training and validation data sets to complete model training and validation. If the de novo training mode is used, the weight model parameter file is not loaded; if the continuous training mode is used, the original model parameter file needs to be loaded. The loss function decreases during the training and validation process to ensure that the neural network model converges and there is no overfitting phenomenon. The model weights during the training process are saved.

[0090] Finally, load the saved model weights on the test dataset, read the data in the test dataset, test the model performance, and select the neural network model with the best performance for deployment.

[0091] Among them, the model performance test uses normalized mean square error as a measurement indicator, and the calculation formula is:

[0092]

[0093] in, and H represent the channel matrix recovered by the neural network and the true channel matrix respectively.

[0094] This method is based on deep learning technology and the idea of ​​channel state information feedback in traditional massive MIMO communication systems. It regards the channel state information of massive MIMO communication systems as a grayscale image, and converts the channel state information feedback and reconstruction problem of massive MIMO communication systems into a lossy image restoration problem. With the support of data and computing power, it obtains channel state information feedback and reconstruction performance that surpasses traditional compressed sensing algorithms.

[0095] Here's a more specific example:

[0096] A massive MIMO non-sparse communication system uses a point-to-point single-user communication mode. The transmit and receive ends are symmetrically structured, using 16 identical antennas arranged in a 4×4 square array with equal spacing. The channel fading follows Rayleigh fading. Based on these conditions, channel state information data is obtained using two methods:

[0097] In the first method, the transmitter sends a known pilot signal, which is then received by the receiver after passing through the channel. The receiver then uses a channel estimation algorithm to obtain the original channel state information matrix H.

[0098] The second method is that during the simulation modeling process, the modeled Rayleigh fading channel can be directly obtained, and the ideal known channel state information matrix H can be directly captured for channel state information feedback;

[0099] A data set was constructed in the aforementioned communication system scenario. Specifically, 7,000 channel state information matrices H were generated through simulation to form a training data set, 2,000 channel state information matrices H were generated to form a test set, and 1,000 channel state information matrices H were generated as a validation set. Subsequently, data preprocessing was performed on the channel state information matrices H in the aforementioned data set, specifically including the following steps:

[0100] (1) The channel matrix H is transformed into In this form, the complex matrix of size 16×16 is converted into a real matrix of size 32×32, and after being output from the neural network, the complex channel matrix of size 16×16 is restored according to the inverse transformation of the above transformation;

[0101] (2) The real channel state information matrix H′ obtained by the process (1) is compressed according to the compression rate to generate a low-dimensional channel state information matrix H″. Assuming the compression rate is γ, the size of the compressed low-dimensional channel state information matrix H″ is 2N×2N×γ, where γ<1;

[0102] (3) The compressed channel matrix and the uncompressed channel matrix are matched one by one to form a data pair, wherein the data compressed according to step (2) is the data to be reconstructed by the neural network, and the data not compressed in step (2) is the original label data, which is also the ideal target output of the neural network.

[0103] Construct a deep neural network model SCF-Net, whose architecture is as follows Figure 2 and Figure 3As shown in the figure, the SCF-Net neural network model contains a total of 7 layers. The input layer is a convolutional layer, which uses 64 3*3 convolution kernels to extract the important features contained in the channel information matrix. A nonlinear activation function is added after the convolution layer. Since the elements in the channel information matrix contain both positive and negative numbers, in order to prevent the "dead neurons" phenomenon, the nonlinear activation function uses the LeakyReLU activation function. The output of the input layer is added to the output of the middle layer through a residual jump connection structure; the middle layer contains two identical attention-assisted feature extraction sub-networks, and the attention-assisted feature extraction sub-network contains two convolutional layers, both of which use 64 3*3 sizes. The output of the attention-assisted feature extraction subnetwork is added with the output of the attention mechanism module through a residual jump connection structure as the output of the attention-assisted feature extraction subnetwork; the output of the second attention-assisted feature extraction subnetwork passes through a 3*3 convolution layer, a LeakyReLU activation function and a random drop operation, and then is added to the output value of the model input layer in the form of a residual jump connection structure; the output layer is a 3*3 convolution layer.

[0104] The attention mechanism module is the CBAM attention module, which consists of two parts: the channel attention module and the spatial attention module. Given an intermediate feature map As input, the CBAM attention module infers a one-dimensional channel attention map in two independent dimensions, channel and space. and a 2D spatial attention map The attention map is then multiplied into the input feature map for adaptive feature refinement. The entire attention process can be expressed as

[0105]

[0106] in, represents element-wise multiplication. During the multiplication, the channel attention value is propagated along the spatial dimension for replication and vice versa. F″ is the refined result of the final output of the CBAM attention module.

[0107] The channel attention module uses the feature relationship between channels to generate a channel attention feature map, which is used to find the key input features that the network needs to pay attention to. The specific operation process includes:

[0108] First, the spatial information of the feature map is aggregated using mean pooling and maximum pooling operations to generate two different spatial feature maps. and in represents the mean pooling feature, Represents the maximum pooling feature.

[0109] Then, these two features Figure 1 Send them into a shared network to generate a channel attention map The shared network is a two-layer neural network, including a multi-layer perceptron (MLP) and a hidden layer. After the two feature maps are respectively fed into the shared network, the output feature vector is merged using the element-wise summation method.

[0110] The calculation process of channel attention can be expressed as:

[0111]

[0112] Among them, s represents the Sigmoid function, And the weights W0 and W1 of the MLP are shared for both inputs.

[0113] The spatial attention module generates a spatial attention feature map based on the spatial relationship of the feature space, which is used to find the location of important information features to be extracted. The specific implementation steps are as follows:

[0114] By using two pooling operations to aggregate the channel information of the feature map, two two-dimensional maps are generated: and These two-dimensional maps represent the mean pooling feature and the maximum pooling feature in the channel direction respectively;

[0115] Then, the two features are connected through a convolution layer and convolution operation is performed to generate a two-dimensional spatial attention feature map. The specific calculation method of spatial attention is as follows:

[0116]

[0117] Among them, s represents the Sigmoid activation function, f 7×7 Indicates that the convolution operation uses a convolution kernel of size 7×7.

[0118] The channel attention module and the spatial attention module are arranged in series to form the CBAM attention module.

[0119] The constructed SCF-Net neural network model was trained on the training and validation datasets constructed above. The Adam optimizer was selected for gradient optimization. An end-to-end learning approach was used to jointly train the neural network model parameters, including weights, biases, step sizes, and convolution kernel sizes, to minimize and stabilize the loss function. The Adam algorithm used a learning rate of 0.001. Each iteration used 512 samples from the training set to calculate the gradient and updated the parameters according to the Adam algorithm's formula. This process was repeated 500 times throughout the training set. The model weight parameter files related to the training process were saved, and the performance of the network model could be tested on the test set.

[0120] The SCF-Net network model with good test performance is selected for the channel state information feedback of the massive MIMO non-sparse communication system. The compressed lossy channel state information matrix H″ is input into the neural network model, and the real channel state information matrix reconstructed by the neural network model is output. Then, the reconstructed complex channel matrix is ​​obtained by combining the split inverse operations

[0121] The results of the test show that the channel state information feedback and reconstruction method based on the SCF-Net neural network model in the massive MIMO non-sparse communication system of this embodiment performs significantly better than traditional algorithms and existing deep learning algorithms. Figure 4 As shown, where NMSE represents the normalized mean square error, the reconstructed channel matrix is ​​calculated The expected value of the square of the difference between the original channel matrix H and r represents the reconstructed channel matrix The cosine similarity between θ and the original channel matrix H.

[0122] In summary, based on deep learning technology and the idea of ​​channel state information feedback in traditional massive MIMO communication systems, this paper proposes a neural network model SCF-Net for realizing non-sparse channel state information feedback in massive MIMO communication systems. The channel state information of massive MIMO communication systems is regarded as a grayscale image, and the channel state information feedback and reconstruction problem of massive MIMO communication systems is converted into a lossy image restoration problem. With the support of neural network models, data and computing power, the channel state information feedback and reconstruction performance that exceeds that of traditional compressed sensing algorithms is obtained.

[0123] The present invention can effectively solve the problem of excessive reliance on channel sparsity in existing channel state information feedback technologies, and can effectively improve the feedback and reconstruction capabilities of large-scale MIMO non-sparse channel state information.

Claims

1. A method for large-scale MIMO non-sparse channel state information feedback and reconstruction, characterized in that: The following steps are involved: Step S1: Simulate and construct a large-scale MIMO non-sparse communication system, obtain a channel state information matrix H, pre-process the channel state information matrix H, and divide it into a training data set, a verification data set, and a test data set according to a ratio; Step S2: constructing a deep neural network model SCF-Net for realizing non-sparse channel state information feedback and reconstruction; Step S3: The deep neural network model SCF-Net constructed in step S2 is trained on the training data set and the validation data set constructed in step S1, so that the channel state information matrix output by the model is Approach the actual channel state information matrix H, save the model weight parameter file generated in each generation during the training process, observe the decrease of the training data loss function and the verification data loss function, until the loss function fully converges and the model training is completed; Step S4: Load the model weight parameter files of each generation saved in step S3 respectively, and test the model performance on the test data set; Step S5: Select the deep neural network model with the best test performance for channel state information feedback and reconstruction, deploy the deep neural network model at the transmitting end, and the receiving end compresses the original channel state information matrix H according to the compression rate γ, and sends the compressed channel state information matrix to the transmitting end through the feedback link. The transmitting end inputs the feedback data into the deep neural network model to reconstruct the original channel state information before compression.

2. A method for massive MIMO non-sparse channel state information feedback and reconstruction according to claim 1, characterized in that: In step S1, a large-scale MIMO non-sparse communication system is simulated and constructed, specifically in the following manner: Step S111: Use MATLAB simulation software to model the system. The massive MIMO non-sparse communication system adopts a symmetrical structure of the transmitting and receiving ends. In the simulation software, the number of transmitting and receiving antennas is set to 16, arranged in a 4×4 structure, and the antenna spacing is set to 1 meter. Step S112: The transmitting end sends an orthogonal pilot sequence that does not change with time slots Where i is the time slot; Step S113: The channel is modeled using the Kronecker channel model. The number of channel multipaths is set to 7. The channel is represented as Among them, H w It is a Rayleigh channel with independent channel coefficients, generated by the sine wave superposition method; For the correlation matrix R between antennas Rx and R Tx The modeling adopts the classic exponential correlation model, that is, R Rx and R Tx The entry in the i-th row and j-th column of is modeled as: Among them, d i,j ≥0 indicates the distance from the ith antenna to the jth antenna, d0>0 indicates the spacing between adjacent antennas, c i,j =d i,j / d0 is a constant, and a∈[0,1] are parameters; Step S114: The transmission signal generated in step S112 is received by the receiving end after passing through the channel constructed in step S113. The receiving end obtains channel state information through channel estimation based on the received signal. That is, the channel state information required for channel state information feedback and reconstruction, or, in the simulation process, the ideal perfect channel state information is directly saved to realize channel state information feedback and reconstruction.

3. The method for massive MIMO non-sparse channel state information feedback and reconstruction according to claim 1, characterized in that: In step S1, the channel state information matrix H is obtained by channel estimation at the receiving end of the massive MIMO non-sparse communication system constructed by simulation, or is an ideal known channel state information matrix directly obtained during the simulation process.

4. The method for massive MIMO non-sparse channel state information feedback and reconstruction according to claim 1, characterized in that: In step S1, the channel state information matrix H is preprocessed and divided into a training data set, a verification data set and a test data set according to a ratio, and the specific method is as follows: Step S121: converting the one-dimensional channel state information obtained by channel estimation at the receiving end of the simulated massive MIMO non-sparse communication system or the ideal known channel state information directly obtained into a channel state information matrix H of size N×N, where N is the number of antennas equipped at the transceiver end, and each matrix element in the channel state information matrix H is complex data; Step S122: Convert the channel state information matrix H into In this form, a complex matrix of size N×N is converted into a real matrix of size 2N×2N, and then a matrix of size N×N is restored through inverse transformation after being output from the deep neural network model. Step S123: compressing the real channel state information matrix H′ obtained in step S122 according to the compression rate γ to generate a low-dimensional channel state information matrix H″ of size 2N×2N×γ, γ<1; Step S124: repeating steps S121 to S123 to generate multiple sets of data under simulation conditions, each set of data consisting of an H and an H″, wherein H″ is the data to be reconstructed by the deep neural network model, and H is the original label data, which is also the ideal target output of the deep neural network model; Step S125: Divide the data generated in step S124 into a training data set, a validation data set, and a test data set in a ratio of 7:1:

2.

5. The method for massive MIMO non-sparse channel state information feedback and reconstruction according to claim 1, characterized in that: The deep neural network model SCF-Net includes an input layer, an intermediate layer and an output layer: The input layer is a convolution layer, which uses 64 3*3 convolution kernels to extract the important features contained in the channel state information matrix. A nonlinear activation function is added after the convolution layer, and the nonlinear activation function uses the LeakyReLU activation function; The middle layer contains two attention-assisted feature extraction subnetworks with the same structure connected in series. The attention-assisted feature extraction subnetwork contains two convolutional layers. Both convolutional layers use 64 3*3 convolution kernels for feature extraction. A LeakyReLU activation function and a random drop operation are added after the first convolutional layer. A LeakyReLU activation function is added after the second convolutional layer, and then connected to the attention mechanism module for feature refinement. A random drop operation is added after the attention mechanism module. The input of the attention-assisted feature extraction subnetwork and the output of the attention mechanism module are added through a residual jump connection structure as the output of the attention-assisted feature extraction subnetwork; the output of the second attention-assisted feature extraction subnetwork passes through a 3*3 convolutional layer, a LeakyReLU activation function and a random drop operation, and then added to the output value of the input layer in the form of a residual jump connection structure as the output of the middle layer; The output layer is a 3*3 convolutional layer.

6. A method for massive MIMO non-sparse channel state information feedback and reconstruction according to claim 5, characterized in that: The attention mechanism module is a CBAM attention module, which includes two parts: a channel attention module and a spatial attention module. Given an intermediate feature map, the CBAM attention module infers the attention map in turn according to the two independent dimensions of channel and space, and then multiplies the attention map into the input feature map for adaptive feature refinement; wherein, the channel attention module uses the feature relationship between channels to generate a channel attention feature map, which is used to find the key input features that the network needs to pay attention to; the spatial attention module generates a spatial attention feature map through the spatial relationship of the feature space, which is used to find the location of the important information features that need to be extracted. The channel attention module and the spatial attention module are arranged in series to form the CBAM attention module.

7. The method for massive MIMO non-sparse channel state information feedback and reconstruction according to claim 1, characterized in that: In step S3, when training the deep neural network model SCF-Net, the Adam gradient optimization algorithm and the end-to-end learning method are used to jointly train the neural network parameters so as to minimize the loss function; The loss function is the mean square error loss, specifically the original channel matrix estimate reconstructed by the deep neural network model The mean square error between H and the true label, the loss function expression is: Where T is the number of all samples in the training set, ||·||2 is the Euclidean norm, and the subscript i represents the number of the reconstructed channel matrix and the original channel matrix.

8. The method for massive MIMO non-sparse channel state information feedback and reconstruction according to claim 1, characterized in that: In step S4, the normalized mean square error is used as the measurement indicator for the test model performance, and the calculation formula is: Where E represents the mathematical expectation, and H represent the channel matrix restored by the neural network and the true channel matrix respectively.

Images