MIMO channel state information feedback method based on deep learning and attention mechanism

Abstract

The present application relates to the technical field of communication, especially provide a MIMO channel state information feedback method based on deep learning and attention mechanism, the method first introduces the coordinate attention mechanism into the channel state information (CSI) feedback task, the spatial separable self-attention mechanism is introduced in the decoder, improve the quality of channel reconstruction, CANet model reduces the network complexity, improves the performance, CANet especially performs well in indoor scene, and has the lowest calculation complexity, has better performance in complexity, reconstruction quality and running time, the method improves the precision of the feedback network, reduces the calculation amount and parameter quantity, to reduce the complexity.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to a MIMO channel state information feedback method based on deep learning and attention mechanisms. Background Technology

[0002] Massive Multiple-Input Multiple-Output (MIMO) technology is one of the core technologies of next-generation communication systems. By equipping the base station (BS) with dozens or even hundreds of antennas, it can significantly improve spectral efficiency and energy efficiency. To fully utilize spatial diversity and multiplexing gain, massive MIMO systems need to acquire sufficiently accurate downlink channel state information (CSI). Although time-division duplex operation is commonly used, frequency-division duplex massive MIMO can handle the low latency requirements brought about by standardization and may be much better than time-division duplex solutions. In frequency-division duplex (FDD) massive MIMO systems, the user equipment (UE) must first estimate the downlink CSI and then feed it back to the BS through a feedback link. In traditional MIMO systems, the downlink CSI of FDD systems is first obtained by the UE through downlink pilot estimation and then fed back to the BS. Then, a codebook-based or vector quantization method is used to reduce the feedback overhead. However, this approach is not feasible in massive MIMO because a large number of antennas are used at the BS, which will greatly increase the dimension of the CSI matrix, and significantly increase the amount of codebook feedback and design complexity.

[0003] Traditional compressed sensing methods (LASSO, BM3DA-MP, and TVAL3) suffer from several critical drawbacks, such as heavy reliance on channel sparsity assumptions, low efficiency in iterative signal reconstruction, and insufficient utilization of channel structure. With the rapid development of deep learning (DL) technology, DL methods are increasingly being applied in CSI feedback. CSINet effectively utilizes the channel structure in the training samples, compressing the estimated downlink CSI into codewords at the UE and feeding them back to the BS, where the downlink CSI is then reconstructed. However, previous feedback network designs have neglected the diversity of CSI spatial patterns. Furthermore, existing feedback networks, due to their typically fixed complexity, struggle to adapt to varying resource constraints.

[0004] In the prior art, patent document CN108390706A discloses a deep learning-based method for large-scale MIMO channel state information feedback. This method introduces deep learning technology into the large-scale MIMO feedback scheme, treating the channel matrix as an image without considering the complex physical characteristics of the signal. Its complex phase characteristics cannot be expressed, resulting in the loss of original physical information. Therefore, this method still suffers from the problem of not being able to retain the complete channel information to the greatest extent, even in low feedback overhead and complex outdoor scenarios. Another patent document, CN108847876A, discloses a method for compressing, feeding back, and reconstructing large-scale MIMO time-varying channel state information. This method reduces the feedback overhead of large-scale MIMO channel information, improves reconstruction accuracy, and exhibits excellent robustness to decreases in compression ratio; however, its complexity is excessively high. Summary of the Invention

[0005] In view of this, the present invention provides a MIMO channel state information feedback method based on deep learning and attention mechanism to improve the accuracy of the feedback network, reduce its computational load and parameter count, and thus reduce complexity.

[0006] In a first aspect, the present invention provides a MIMO channel state information feedback method based on deep learning and attention mechanisms, the method comprising:

[0007] Step 1: Calculate the channel matrix of MIMO channel state information in the spatial frequency domain. Perform a two-dimensional discrete Fourier transform to obtain the channel matrix H, which is sparse in the angular delay domain;

[0008] Step 2: Select the first N elements of the channel matrix H. a The non-zero elements in each row are truncated to obtain the truncation matrix H. a ;

[0009] Step 3: Construct the feedback model CANet, which includes an encoder and a decoder; The truncation matrix H... a The encoder is input to obtain low-dimensional codewords; these codewords are then input into the decoder to obtain the reconstructed channel matrix.

[0010] Step 4: Train the feedback model CANet to generate the trained CANet model, and then use the reconstructed channel matrix... Minimize the channel matrix output and minimized truncation matrix H a ;

[0011] Step 5: For the minimized channel matrix By performing zero-padding and a two-dimensional inverse discrete Fourier transform, the original spatial frequency domain channel matrix is ​​recovered.

[0012] Step 6: Use the trained CANet model for compressed sensing and reconstruction of Channel State Information (CSI).

[0013] Optionally, the truncation matrix H a The encoder is input to obtain low-dimensional codewords; these codewords are then input into the decoder to obtain the reconstructed channel matrix. include:

[0014] Step S1: Cut off the matrix H a The encoder serves as the input and outputs compressed codewords; the encoder sequentially includes a first convolutional layer, an attention mechanism, a second convolutional layer, and a fully connected layer.

[0015] Step S2: The decoder receives the compressed codeword sent by the encoder, takes the codeword s as input, and outputs the codeword s along with the channel matrix H. a Same-dimensional reconstructed channel matrix The decoder sequentially includes a fully connected layer, a first convolutional layer, an attention mechanism, a spatially separable attention mechanism, a layer normalization layer, a second convolutional layer, and a Hardsigmoid activation layer.

[0016] Optionally, step S1 includes:

[0017] For the Channel State Information (CSI) complex truncated matrix Transform it into 2×N t ×N a , H a As input to the encoder;

[0018] The first layer of the encoder is a convolutional layer ConvBN, which consists of a CNN with a 5×5 convolutional kernel, a batch normalization layer (BN), and an activation function PReLU, used to extract shallow features and fuse information from the real and imaginary parts;

[0019] The second layer is a coordinate attention (CA) block. Internally, the module decomposes global pooling into a one-to-one one-dimensional feature encoding operation; it then concatenates the feature maps obtained from the width and height of the global receptive field; the concatenated feature map is fed into a shared 3×3 convolutional module to reduce its dimension to C / r; finally, the batch-normalized feature map F1 is fed into the Simgoid activation function to obtain the feature map. The feature map f is divided into two separate tensors along the spatial dimension. and Perform convolutions with a 3×3 kernel on the original height and width to obtain feature maps F with the same number of channels as the original. h and F w After passing through the Sigmoid activation function, the attention weights g of the feature map in height and width are obtained respectively. h and g w By multiplying and weighting the original feature map, a feature map with attention weights in the width and height directions is obtained.

[0020] The third layer is a 3×3 convolutional layer used for feature extraction;

[0021] The fourth layer is a fully connected layer that compresses the vector to M (M < 2N). c N t Quantized codewords of size ) with a compression ratio of

[0022] Optionally, step S2 includes:

[0023] The decoder receives the compressed codeword, decompresses it, and reshapes it into a 2×N format. t ×N a Size of H;

[0024] The first layer of the decoder is a fully connected layer used to decompress and recover the compressed codewords, reconstructing them to a size of 2×N. t ×N a ;

[0025] The second layer is a 5×5 convolutional layer, used to improve reconstruction performance;

[0026] The third layer consists of a Spatially Separable Self-Attention (SSSA) block and a Coordinate Attention (CA) block, designed to minimize reconstruction errors and reduce computational complexity. The SSSA block comprises a layer normalization layer, a linear layer, Gaussian error linearization units, a Global Subsampling Attention Group (GSA), and a Local Self-Attention Group (LSA). Each channel matrix is ​​subdivided into m×m matrices of dimension H×H. And calculate the attention separately for each window; reducing the complexity to A Global Subsampling Attention Group (GSA) is used to capture antenna correlations between windows. The computational complexity of the GSA layer is O(n log n). The total computational complexity of the attention mechanism is

[0027] The fourth layer uses a 3×3 standard convolution kernel to reduce the dimensionality of the channel; the Hardsigmoid function is used for activation.

[0028] Optionally, the trained CANet model generates feedback network parameters, which include the weights and biases of the fully connected layers and the convolutional kernels and biases of the convolutional layers.

[0029] Optionally, step four employs an end-to-end learning approach and an adaptive moment estimation (Adam) optimization approach. The learning rate uses a cosine annealing algorithm to jointly train the encoder and decoder to minimize the cost function. The cost function of the entire feedback model CANet is designed as the channel matrix output by the decoder.

[0030] Thirdly, the present invention provides a computer-readable storage medium comprising a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to execute the MIMO channel state information feedback method based on deep learning and attention mechanism in the first aspect or any possible implementation thereof.

[0031] Fourthly, the present invention provides an electronic device, comprising: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs include instructions that, when executed by the device, cause the device to perform the MIMO channel state information feedback method based on deep learning and attention mechanisms in the first aspect or any possible implementation of the first aspect.

[0032] The technical solution provided by this invention includes a method for processing the channel matrix of MIMO channel state information in the spatial-frequency domain. Perform a two-dimensional discrete Fourier transform to obtain a channel matrix H that is sparse in the angular delay domain; then extract the first N elements of the channel matrix H. a The non-zero elements in each row are truncated to obtain the truncation matrix H. a Construct the feedback model CANet, which includes an encoder and a decoder; truncate the matrix H. a The codewords are input into the encoder to obtain low-dimensional codewords; these codewords are then input into the decoder to obtain the reconstructed channel matrix. The feedback model CANet is trained to generate the trained CANet model, and then the reconstructed channel matrix is ​​used as the basis for this process. Minimize the channel matrix output and minimized truncation matrix H a For the minimized channel matrix By performing zero-padding and a two-dimensional inverse discrete Fourier transform, the original spatial frequency domain channel matrix is ​​recovered. The trained CANet model is used for compressed sensing and reconstruction of Channel State Information (CSI). This method introduces a coordinate attention mechanism into the CSI feedback task for the first time, and incorporates a spatially separable self-attention mechanism into the decoder, improving the quality of channel reconstruction. While the CANet model improves performance by reducing network complexity, CANet performs particularly well in indoor scenarios and has the lowest computational complexity. It exhibits better performance in terms of complexity, reconstruction quality, and runtime. This method improves the accuracy of the feedback network while reducing its computational cost and parameter count, thus lowering overall complexity. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 A flowchart of the MIMO channel state information feedback method provided in an embodiment of the present invention;

[0035] Figure 2 This is a schematic diagram of the encoder provided in an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of the structure of the coordinate attention block provided in an embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram of the decoder provided in an embodiment of the present invention;

[0038] Figure 5 This is a schematic diagram of the structure of a spatially separable self-attention block provided in an embodiment of the present invention;

[0039] Figure 6 A schematic diagram of the CANet feedback network model provided in an embodiment of the present invention.

[0040] Figure 7 A schematic diagram of different network training curves provided in embodiments of the present invention;

[0041] Figure 8 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] It should be understood that the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0044] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention are also intended to include the plural forms unless the context clearly indicates otherwise.

[0045] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0046] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0047] Figure 1 The flowchart of the MIMO channel state information feedback method provided in the embodiments of the present invention is as follows: Figure 1 As shown, the method includes:

[0048] Step 1: Calculate the channel matrix of the multiple-input multiple-output (MIMO) channel state information in the spatial frequency domain. Perform a two-dimensional discrete Fourier transform to obtain the channel matrix H, which is sparse in the angular delay domain.

[0049] Step 2: Extract the first N elements of the channel matrix H. a The non-zero elements in each row are truncated to obtain the truncation matrix H. a .

[0050] Step 3: Construct the feedback model CANet, which includes an encoder and a decoder; truncation matrix H... a The codewords are input into the encoder to obtain low-dimensional codewords; these codewords are then input into the decoder to obtain the reconstructed channel matrix.

[0051] In this embodiment of the invention, the encoder is deployed on the electronic device and the decoder is deployed on the base station.

[0052] In this embodiment of the invention, the truncation matrix H is... a The codewords are input into the encoder to obtain low-dimensional codewords; these codewords are then input into the decoder to obtain the reconstructed channel matrix. include:

[0053] Step S1: Cut off the matrix H a The encoder serves as the input and outputs compressed codewords. The encoder sequentially includes a first convolutional layer, an attention mechanism, a second convolutional layer, and a fully connected layer.

[0054] Figure 2 This is a schematic diagram of the encoder provided in an embodiment of the present invention, as shown below. Figure 2 As shown, step S1 includes:

[0055] For the Channel State Information (CSI) complex truncated matrix To facilitate data processing and training, it is transformed into a 2×N model. t ×N a (That is, the real part and the imaginary part are divided into two different channels), H a As input to the encoder.

[0056] In embodiments of the present invention, such as Figure 2 As shown, this will be the input to the UE encoder. First, it passes through a ConvBN convolutional layer (which consists of a CNN with a 5×5 convolutional kernel, a batch normalization layer (BN), and an activation function PReLU), then through a coordinate attention CA block, then through a 3×3 ConvBN convolutional layer, and finally through a fully connected layer.

[0057] Figure 3 This is a schematic diagram of the structure of the coordinate attention block provided in an embodiment of the present invention, such as... Figure 3As shown, the second layer is a coordinate attention (CA) block, which helps extract information from a deep abstraction and improves overall performance in certain scenarios while reducing computational complexity. The module's internal operations involve decomposing global pooling into a one-to-one one-dimensional feature encoding operation; concatenating the feature maps obtained from the width and height of the global receptive field; feeding the concatenated feature map into a shared 3×3 convolutional module to reduce its dimension to C / r; and then feeding the batch-normalized feature map F1 into the Simgoid activation function to obtain the feature map. The feature map f is divided into two separate tensors along the spatial dimension. and Perform convolutions with a 3×3 kernel on the original height and width to obtain feature maps F with the same number of channels as the original. h and F w After passing through the Sigmoid activation function, the attention weights g of the feature map in height and width are obtained respectively. h and g w By multiplying and weighting the original feature map, a feature map with attention weights in the width and height directions is obtained.

[0058] The third layer is a 3×3 convolutional layer used for feature extraction.

[0059] The fourth layer is a fully connected layer that compresses the vector to M (M < 2N). c N t Quantized codewords of size ) with a compression ratio of

[0060] Step S2: The decoder receives the compressed codeword sent by the encoder, takes the codeword s as input, and outputs the codeword s along with the channel matrix H. a Same-dimensional reconstructed channel matrix The decoder sequentially includes a fully connected layer, a first convolutional layer, an attention mechanism, a spatially separable attention mechanism, a layer normalization layer, a second convolutional layer, and a Hardsigmoid activation layer.

[0061] Figure 4 This is a schematic diagram of the decoder structure provided in an embodiment of the present invention, as shown below. Figure 4 As shown, step S2 includes:

[0062] The decoder receives the compressed codewords via a wireless channel, decompresses the codewords, and reconstructs them into a 2×N format. t ×N a Size H.

[0063] The first layer of the decoder is a fully connected layer used to decompress and recover the compressed codewords, reconstructing them to a size of 2×N. t ×N a .

[0064] The second layer is a 5×5 convolutional layer, used to improve reconstruction performance.

[0065] The third layer consists of a Spatially Separable Self-Attention (SSSA) block and a Coordinate Attention (CA) block, designed to minimize reconstruction errors and reduce computational complexity. Figure 5 As shown, the SSSA block consists of a layer normalization layer, a linear layer, a Gaussian error linear unit, a global subsampling attention group (GSA), and a local self-attention group (LSA). Using windowed attention to reduce the recipient region of attention is a method to reduce computational complexity. Each channel matrix is ​​subdivided into m×m matrices of dimension H×H. And calculate the attention separately for each window; reducing the complexity to Since the windows are fixed and do not communicate with each other, the antenna correlation between windows is lost and cannot be used to compress the channel matrix, resulting in a performance loss in channel reconstruction. Therefore, another layer of local grouped attention is added after LSA to capture the antenna correlation between windows. This can be achieved through a Global Subsampling Attention Group (GSA). The computational complexity of the GSA layer for capturing the antenna correlation between windows is O(log n). The total computational complexity of the attention mechanism is

[0066] The fourth layer uses a 3×3 standard convolution kernel to reduce the dimensionality of the channel; the Hardsigmoid function is used for activation.

[0067] In this embodiment of the invention, to minimize reconstruction error and reduce computational complexity, a parallel method is introduced in the decoder, which has two stems: one composed of Spatially Separable Self-Attention (SSSA) and the other of Computational Attention (CA). The two stems are then refined using 5×5 convolutional kernels concatenated together, and finally, a 3×3 standard convolutional kernel is used to reduce the channel dimension.

[0068] Step 4: Train the feedback model CANet to generate the trained CANet model, and then use the reconstructed channel matrix... Minimize the channel matrix output and minimized truncation matrix H a .

[0069] In this embodiment of the invention, step four adopts an end-to-end learning method and an adaptive moment estimation Adam optimization method. The learning rate adopts the cosine annealing algorithm. The encoder and decoder are jointly trained to minimize the cost function. The cost function of the entire feedback model CANet is designed as the channel matrix output by the decoder.

[0070] Step 5: Minimize the channel matrix By performing zero-padding and a two-dimensional inverse discrete Fourier transform, the original spatial frequency domain channel matrix is ​​recovered.

[0071] Step 6: Use the trained CANet model for compressed sensing and reconstruction of Channel State Information (CSI).

[0072] In this embodiment of the invention, the trained CANet model generates feedback network parameters, which include the weights and biases of the fully connected layers and the convolutional kernels and biases of the convolutional layers.

[0073] In this embodiment of the invention, the feedback network consisting of the encoder and decoder is named the Coordinate Attention Network (CANet), or simply the feedback network CANet. At the electronic device end, the encoder compresses and encodes the channel state information into low-dimensional codewords, which are then transmitted to the base station decoder via the feedback link to reconstruct the channel state information, reducing the feedback overhead of the channel state information. The following describes this method in detail with reference to specific application scenarios.

[0074] In this embodiment of the invention, signal processing is performed using an offline model training and online feedback approach. Data is collected using MATLAB and the proposed method is trained offline using Python. The COST2100 channel model is used in the MATLAB simulation platform to generate datasets. This channel model is a geometry-based stochastic channel model, generating two types of datasets: (1) an indoor microcell scenario operating at 5.3 GHz; and (2) an outdoor scenario operating at 300 MHz. The channel model is configured with N antennas. t =32, the number of subcarriers is N c=1024, then converted to the angular delay domain, generating 100,000 samples as the training set, 30,000 samples as the validation set, and 20,000 samples as the test set. For online prediction, repeated training is unnecessary; the pre-trained model can be used directly. This invention uses an end-to-end approach to obtain all weights and biases in the network. The learning rate of the proposed network uses both cosine annealing and 0.001, with cosine annealing being the default. The number of training iterations and batch size are set to 1000 and 200, respectively, and the network is trained using the Adaptive Moment Estimation (ADAM) optimization method. Compression ratios are 4, 8, 16, 32, and 64, corresponding to codeword lengths M of 512, 256, 128, 64, and 32, respectively. Normalized mean square error (NMSE) is used as the standard for evaluating CSI feedback performance to measure the original H. a and reconstruction distance Complexity is measured in flops (float-point operations per second).

[0075] Figure 6 This is a schematic diagram of the CANet feedback network model provided in an embodiment of the present invention, as shown below. Figure 6 As shown, traditional convolutional blocks are omitted for simplicity. CANet is an end-to-end encoder-decoder framework tailored for the CSI feedback problem. The performance of a CSI feedback scheme largely depends on the compression part; the less information loss during encoder compression, the higher the decompression accuracy. Due to limitations in computing power and storage at the UE end, deep encoder network design is difficult to implement. Therefore, CANet's encoder uses only one coordinate attention module and two convolutional blocks. It implements a low-complexity but information-rich encoder. As the encoder becomes more powerful, the decoder can be correspondingly lighter. Since there is correlation between CSI matrices, extracting this correlation and applying the extracted data to CSI reconstruction will help improve the accuracy of CSI matrix recovery. Attention mechanisms can focus more attention on information that needs attention, acquiring more necessary details and suppressing other useless information. Therefore, this invention utilizes an attention mechanism to extract correlations between CSI matrices for CSI matrix reconstruction and adds a spatially separable self-attention mechanism at the decoder end, which can significantly reduce computational costs.

[0076] The invention was verified following the steps described above, and the same dataset was used to train previously proposed deep learning-based networks such as CLNet, CRNet, CsiNet, DS-NLCsiNet, and Attention-CSI feedback networks. The effectiveness of the invention was demonstrated by comparing these networks. Table 1 shows the normalized mean square error (NMSE), floating-point operations per second (FLOPS), and runtime (T) of the DL-based method under different environments.

[0077] Table 1

[0078]

[0079] Table 1 lists the Normalized Mean Squared Error (NMSE), Floats per Second (FLOPS), and Runtime (T) for the aforementioned CSI feedback networks, with all best results presented in bold. As can be seen from Table 1, the proposed CANet exhibits the lowest computational complexity and the fastest runtime. Specifically, the proposed CANet achieves state-of-the-art performance across various compression ratios in indoor scenarios. Furthermore, in indoor scenarios with CR = 1 / 4, CANet achieves 107.44% of the performance of CLNet, with a FLOPs of 87.7%. Similarly, for CR = 1 / 8, CANet achieves 120.83% of the performance of CLNet, with a FLOPs of 83.1%. CANet also demonstrates good performance in outdoor scenarios. These results show that CANet effectively balances performance and complexity.

[0080] In this embodiment of the invention, Table 2 shows the performance comparison of cosine similarity in indoor scenes.

[0081] Table 2

[0082]

[0083] Table 2 compares the cosine similarity (ρ) of the proposed method and several typical DL-based methods in indoor scenes, with the best results shown in bold. As can be seen from Table 2, CANet achieved the highest ρ value, demonstrating that CANet possesses the best performance.

[0084] Figure 7 This is a schematic diagram of different network training curves provided in embodiments of the present invention, such as... Figure 7 The figure shows the NMSE performance curve as a function of epochs during training in an indoor scene with a compression ratio of 1 / 4. From... Figure 7 As can be seen, the NMSE of all three schemes decreases smoothly and steadily. In particular, CANet's NMSE performance is significantly better than CLNet and CRNet.

[0085] In this embodiment of the invention, Table 3 shows the detailed complexity of the encoder and decoder.

[0086] Table 3

[0087]

[0088] Table 3 shows the computational complexity and number of parameters at the encoder and decoder for CANet, CLNet, and CRNet. At CR = 1 / 4, CANet's encoder and decoder are the lightest, but CANet requires slightly more storage space. At CR = 1 / 64, CANet's decoder is the lightest; a relatively light decoder is also beneficial since BS may need to execute several different models simultaneously.

[0089] In this embodiment of the invention, each step can be performed by an electronic device. For example, electronic devices include, but are not limited to, tablet computers, portable PCs, and desktop PCs.

[0090] The technical solution provided by this invention includes a method for processing the channel matrix of MIMO channel state information in the spatial-frequency domain. Perform a two-dimensional discrete Fourier transform to obtain a sparse channel matrix H in the angular delay domain; truncate the first Na rows of non-zero elements of the channel matrix H to obtain a truncated matrix Htruncation. a Construct the feedback model CANet, which includes an encoder and a decoder; truncate the matrix H. a The codewords are input into the encoder to obtain low-dimensional codewords; these codewords are then input into the decoder to obtain the reconstructed channel matrix. The feedback model CANet is trained to generate the trained CANet model and output the minimized channel matrix. and minimized truncation matrix H a For the minimized channel matrix By performing zero-padding and a two-dimensional inverse discrete Fourier transform, the original spatial frequency domain channel matrix is ​​recovered. The trained CANet model is used for compressed sensing and reconstruction of Channel State Information (CSI). This method introduces a coordinate attention mechanism into the CSI feedback task for the first time, and incorporates a spatially separable self-attention mechanism into the decoder, improving the quality of channel reconstruction. While the CANet model improves performance by reducing network complexity, CANet performs particularly well in indoor scenarios and has the lowest computational complexity. It exhibits better performance in terms of complexity, reconstruction quality, and runtime. This method improves the accuracy of the feedback network while reducing its computational cost and parameter count, thus lowering overall complexity.

[0091] This invention provides a computer-readable storage medium including a stored program, wherein, when the program is running, it controls the electronic device containing the computer-readable storage medium to execute the above-described embodiment of the MIMO channel state information feedback method based on deep learning and attention mechanism.

[0092] Figure 7 A schematic diagram of an electronic device provided in an embodiment of the present invention, such as... Figure 7 As shown, the electronic device 21 includes a processor 211, a memory 212, and a computer program 213 stored in the memory 212 and executable on the processor 211. When the computer program 213 is executed by the processor 211, it implements the MIMO channel state information feedback method based on deep learning and attention mechanism in the embodiment. To avoid repetition, it will not be described in detail here.

[0093] Electronic device 21 includes, but is not limited to, processor 211 and memory 212. Those skilled in the art will understand that... Figure 7 This is merely an example of electronic device 21 and does not constitute a limitation on electronic device 21. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0094] The processor 211 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0095] The memory 212 can be an internal storage unit of the electronic device 21, such as a hard disk or RAM of the electronic device 21. The memory 212 can also be an external storage device of the electronic device 21, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the electronic device 21. Furthermore, the memory 212 can include both internal and external storage units of the electronic device 21. The memory 212 is used to store computer programs and other programs and data required by network devices. The memory 212 can also be used to temporarily store data that has been output or will be output.

[0096] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0097] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for feedback of channel state information based on deep learning and attention mechanisms in multiple-input multiple-output (MIMO) communication, characterized in that, The method includes: Step 1: Calculate the channel matrix of MIMO channel state information in the spatial frequency domain. Perform a two-dimensional discrete Fourier transform to obtain the channel matrix H, which is sparse in the angular delay domain; Step 2: Select the first N elements of the channel matrix H. a The non-zero elements in each row are truncated to obtain the truncation matrix H. a ; Step 3: Construct the feedback model CANet, which includes an encoder and a decoder; The truncation matrix H... a The encoder is input to obtain low-dimensional codewords; these codewords are then input into the decoder to obtain the reconstructed channel matrix. ; Step 4: Train the feedback model CANet to generate the trained CANet model, and then use the reconstructed channel matrix... Minimize the channel matrix output and minimized truncation matrix H a ; Step 5: For the minimized channel matrix By performing zero-padding and a two-dimensional inverse discrete Fourier transform, the original spatial frequency domain channel matrix is ​​recovered. ; Step 6: Use the trained CANet model for compressed sensing and reconstruction of Channel State Information (CSI); The truncated matrix H a The encoder is input to obtain low-dimensional codewords; these codewords are then input into the decoder to obtain the reconstructed channel matrix. include: Step S1: Cut off the matrix H a The encoder serves as the input and outputs compressed codewords; the encoder sequentially includes a first convolutional layer, an attention mechanism, a second convolutional layer, and a fully connected layer. Step S2: The decoder receives the compressed codeword sent by the encoder, takes the codeword s as input, and outputs the codeword s along with the channel matrix H. a Same-dimensional reconstructed channel matrix The decoder sequentially includes a fully connected layer, a first convolutional layer, an attention mechanism, a spatially separable attention mechanism, a layer normalization layer, a second convolutional layer, and a Hardsigmoid activation layer. Step S2 includes: The decoder receives the compressed codeword, decompresses it, and reconstructs it into... Size of H; The first layer of the decoder is a fully connected layer used to decompress and recover the compressed codewords, reshaping them to a size of [size missing]. ; The second layer is a 5×5 convolutional layer, used to improve reconstruction performance; The third layer consists of a Spatially Separable Self-Attention (SSSA) block and a Coordinate Attention (CA) block, designed to minimize reconstruction errors and reduce computational complexity. The SSSA block comprises LayerNorm, a linear layer, Gaussian error linear units, Global Subsampling Attention Groups (GSA), and Local Self-Attention Groups (LSA). Each channel matrix is ​​subdivided into m×m H×H matrices. And compute attention separately for each window; reducing complexity to A Global Subsampling Attention Group (GSA) is used to capture antenna correlations between windows. The computational complexity of the GSA layer is O(n log n). The total computational complexity of the attention mechanism is... ; The fourth layer uses a 3×3 standard convolution kernel to reduce the dimensionality of the channel; the Hardsigmoid function is used for activation.

2. The method according to claim 1, characterized in that, Step S1 includes: For the Channel State Information (CSI) complex truncated matrix Transform it into , will H a As input to the encoder; The first layer of the encoder is a convolutional layer ConvBN, which consists of a CNN with a 5×5 convolutional kernel, a batch normalization layer (BN), and an activation function PReLU, used to extract shallow features and fuse information from the real and imaginary parts; The second layer is a coordinate attention (CA) block. Internally, the module decomposes global pooling into a one-to-one one-dimensional feature encoding operation; it then concatenates the feature maps obtained from the width and height of the global receptive field; the concatenated feature map is fed into a shared 3×3 convolutional module to reduce its dimension to C / r; finally, the batch-normalized feature map F1 is fed into the Simgoid activation function to obtain the feature map. ; the feature map along the spatial dimension Divided into two separate tensors and Perform a 3×3 convolution on the original height and width to obtain feature maps F with the same number of channels as the original. h and F w After passing through the Sigmoid activation function, the attention weights g of the feature map in height and width are obtained respectively. h and g w By multiplying and weighting the original feature map, a feature map with attention weights in the width and height directions is obtained. The third layer is a 3×3 convolutional layer used for feature extraction; The fourth layer is a fully connected layer that compresses the vector into M( Quantized codewords of size ) with a compression ratio of .

3. The method according to claim 1, characterized in that, The trained CANet model generates feedback network parameters, which include the weights and biases of the fully connected layers and the convolutional kernels and biases of the convolutional layers.

4. The method according to claim 1, characterized in that, Step four employs an end-to-end learning approach and an adaptive moment estimation Adam optimization approach. The learning rate uses a cosine annealing algorithm to jointly train the encoder and decoder to minimize the cost function. The cost function of the entire feedback model CANet is designed to be the channel matrix output by the decoder.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the MIMO channel state information feedback method based on deep learning and attention mechanism as described in any one of claims 1 to 4.

6. An electronic device, characterized in that, include: One or more processors; Memory; And one or more computer programs, wherein the one or more computer programs are stored in the memory, the one or more computer programs including instructions that, when executed by the device, cause the device to perform the MIMO channel state information feedback method based on deep learning and attention mechanism as described in any one of claims 1 to 4.

Images