Large-scale MIMO CSI compression method based on attention fusion auto-encoder

By employing a hybrid training strategy of attention fusion autoencoder and the HAGF/LASA mechanism, the problem of insufficient generalization ability of CSI compression methods in different scenarios is solved, achieving efficient and reliable CSI reconstruction and cross-scenario adaptation, thereby improving the performance of large-scale MIMO systems.

CN120915339APending Publication Date: 2025-11-07SOUTHEAST UNIV +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202510971619.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-15
Publication Date
2025-11-07

AI Technical Summary

Technical Problem

Existing deep learning-based CSI compression methods lack generalization ability across different wireless channel scenarios, resulting in significant performance degradation when a model trained in one scenario is directly applied to another scenario. This increases the complexity of training, storage, and management, and may introduce additional complexity and errors.

Method used

We employ a large-scale MIMO CSI compression method based on attention fusion autoencoders. By using a hybrid training strategy and introducing parallel HAGF attention fusion and LASA self-attention mechanisms, we construct a general CSI compression model that can span different scenarios, including encoder and decoder networks at the user equipment end and base station end for CSI feature extraction, feedback, and reconstruction.

Benefits of technology

It significantly improves CSI reconstruction accuracy and cross-scenario generalization ability, reduces feedback overhead, simplifies model deployment and maintenance, improves system throughput and spectral efficiency, and reduces dependence on specific training environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120915339A_ABST
    Figure CN120915339A_ABST
Patent Text Reader

Abstract

The invention discloses a large-scale MIMO CSI compression method based on an attention fusion auto-encoder. The problem that the generalization ability of an existing CSI compression model based on deep learning is limited in different wireless channel scenes is solved. According to the method, a novel self-encoder network structure is provided, and an encoder effectively extracts and integrates channel dimension and spatial dimension features of CSI through a parallel and adaptive attention fusion mechanism. In particular, a mixed attention gating fusion module and a local perception self-attention module are introduced to accurately capture global contexts and local fine structures of CSI in different scenes. And feature reuse and gradient propagation are optimized by combining gating dense connection. And the cross-scene generalization performance of the model can be remarkably improved by adopting a mixed training strategy. According to the method, under various compression ratios, whether single-scene application or cross-scene application, the CSI reconstruction precision superior to that of an existing reference method is shown, and meanwhile, the method has higher generalization ability for different scene channels.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to a large-scale MIMO CSI compression method based on attention fusion autoencoder, and belongs to the technical field of channel state information compression. BACKGROUND

[0002] Facing the sixth generation (6G) mobile communication system, massive multiple-input multiple-output (MIMO) technology is considered as one of the core drivers to improve spectral efficiency and system capacity. This technology realizes precise beamforming and spatial multiplexing by deploying large-scale antenna arrays at base stations (BS). However, in the widely used frequency division duplex (FDD) mode, the base station needs to rely on the downlink channel state information (CSI) feedback from the user equipment (UE) to perform effective precoding. With the dramatic increase in the number of antennas, the dimension size of the complete CSI matrix grows in a square level, resulting in a huge amount of data required for feedback. This significant feedback overhead not only occupies valuable uplink wireless resources, limiting the timeliness of feedback information, but also may become the main bottleneck restricting the full play of the performance advantages of large-scale MIMO systems. Therefore, developing efficient CSI compression technology is crucial to alleviate the feedback pressure and release the potential of large-scale MIMO.

[0003] Traditional CSI compression methods are mainly based on compressed sensing (CS) theory, utilizing the sparse characteristics of CSI in certain transform domains (such as angle-delay domain) for data dimension reduction. Representative methods include LASSO, approximate message passing (AMP), etc. However, these methods usually rely on strict sparse prior assumptions, while actual wireless channels often do not meet this assumption, leading to performance degradation in complex or dynamically changing channel environments. In addition, CS-based reconstruction algorithms often involve iterative optimization processes, which may bring high computational complexity.

[0004] In recent years, deep learning technology, especially the autoencoder (AE) architecture, has shown great potential in the field of CSI compression. As a pioneering work, CsiNet proves that AE can learn an effective nonlinear mapping of CSI data to low-dimensional code without explicit sparse assumptions, and outperforms traditional CS methods in reconstruction accuracy. Subsequently, researchers have proposed various improved models, such as CRNet using multi-resolution structure, ACRNet introducing network aggregation and parameterized activation function, DCRNet using dilated convolution to expand the receptive field, etc., continuously improving compression performance.

[0005] However, there is a common problem with most current deep learning-based CSI compression models: they are usually designed and trained to be optimized for a specific channel scenario (e.g., indoor-only or outdoor-only scenarios). Due to the significant differences in channel propagation characteristics (such as multipath distribution, delay spread, angle spread, spatial correlation) between different scenarios (e.g., high-frequency indoor vs. low-frequency outdoor, urban complex multipath vs. suburban line-of-sight dominant), the performance of a model trained in one scenario will often be significantly degraded when directly applied to another scenario. To address different scenarios, existing solutions may need to train and deploy a model for each scenario, which not only significantly increases the complexity of training, storage, and management, but also requires accurate and real-time identification of the current communication scenario in practical applications in order to invoke the corresponding model, which itself introduces additional complexity and potential errors. Therefore, developing a general-purpose CSI compression model with strong cross-scenario generalization ability, which can robustly and efficiently process CSI data from diverse wireless environments, has become a key technical challenge in this field. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a large-scale MIMO CSI compression method based on attention fusion autoencoder, which overcomes the limitation of insufficient generalization ability of existing deep learning-based CSI compression methods when crossing different wireless channel scenarios (such as indoor and outdoor).

[0007] The technical problem to be solved by the present application is to provide a large-scale MIMO CSI compression method based on attention fusion autoencoder, which overcomes the limitation of insufficient generalization ability of existing deep learning-based CSI compression methods when crossing different wireless channel scenarios (such as indoor and outdoor).

[0008] The large-scale MIMO CSI compression method based on attention fusion autoencoder comprises the following steps:

[0009] Step 1, obtain a mixed dataset composed of CSI samples of at least two different types of wireless communication scenarios, and pre-process the mixed dataset to obtain a pre-processed mixed dataset, each sample in the pre-processed mixed dataset being a pre-processed CSI feature map; divide the pre-processed mixed dataset into a training set, a validation set, and a test set;

[0010] Step 2, construct an autoencoder network model, the autoencoder network model comprising an encoder network and a decoder network, wherein the encoder network is arranged at a user equipment end, and the decoder network is arranged at a base station end; the encoder network extracts and fuses channel dimension and spatial dimension information of the pre-processed CSI feature map to generate a code word of a predetermined dimension;

[0011] Step 3, the user equipment end feeds back the code word of the predetermined dimension to the base station through an uplink;

[0012] Step 4, at the base station end, the decoder network is used to decompress and reconstruct the code word of the preset dimension to obtain an estimated CSI feature map;

[0013] Step 5, the constructed autoencoder network model is trained and verified by using the training set and the verification set to obtain a trained autoencoder network model;

[0014] Step 6, the preprocessed CSI feature map in the test set is input into the trained autoencoder network model to obtain an estimated CSI feature map; the estimated CSI feature map is converted back to the space-frequency domain by using a two-dimensional inverse discrete Fourier transform to obtain a reconstructed CSI matrix.

[0015] As a preferred scheme of the present application, in step 1, the wireless communication scene at least includes an indoor scene and an outdoor scene, and the CSI sample is an original space-frequency domain CSI matrix of a downlink obtained at a user equipment end wherein, indicates the number of subcarriers, N t indicates the number of antennas of the base station; the CSI samples of different types of wireless communication scenes are mixed at the same proportion to form a mixed data set;

[0016] A two-dimensional discrete Fourier transform is performed on each CSI sample in the mixed data set to obtain an angle-delay domain CSI matrix H, and a truncation threshold N c is set. The first N c rows of H are intercepted to obtain a truncated angle-delay domain CSI matrix The real part and the imaginary part of the matrix H a are separated into two independent channels to form a preprocessed CSI feature map.

[0017] As a preferred scheme of the present application, in step 2, the encoder network includes first to fourth encoder units, and each encoder unit is interconnected through a gated dense connection structure;

[0018] The preprocessed CSI feature map is input into the first encoder unit, and the preprocessed CSI feature map is sequentially subjected to a 1x1 convolution layer and a result after a gating coefficient γ 01 is added to the output of the first encoder unit to obtain a first addition result as the input of the second encoder unit;

[0019] The preprocessed CSI feature map is sequentially subjected to a 1x1 convolution layer and a result after a gating coefficient γ 02 is added to the output of the second encoder unit and a result after a 1x1 convolution layer and a gating coefficient γ 12 of the first addition result to obtain a second addition result as the input of the third encoder unit;

[0020] The preprocessed CSI feature map is sequentially subjected to a 1x1 convolution layer and a gating coefficient γ 03 The result after the gating coefficient γ 13 The result after the gating coefficient γ 23 The result after the gating coefficient γ

[0021] The preprocessed CSI feature map is sequentially subjected to a 1x1 convolution layer and a gating coefficient γ 04 The result after the gating coefficient γ 14 The result after the gating coefficient γ 24 The result after the gating coefficient γ 34 The result after the gating coefficient γ

[0022] The fourth addition result is sequentially subjected to a 1x1 convolution layer and a full connection layer to generate a code word of a preset dimension.

[0023] As a preferred scheme of the present application, the first to fourth encoder units are of the same structure, and each encoder unit comprises a main processing path and a residual path, the main processing path comprising a convolution layer, a batch normalization layer, a PReLU activation function, a hybrid attention gate fusion module and a local perception self-attention module.

[0024] The result F m after the convolution layer, the batch normalization layer and the PReLU activation function in the main processing path of the input of each encoder unit, and the result F r after the 1x1 convolution layer in the residual path of the input of each encoder unit, are both taken as the input of the hybrid attention gate fusion module, the hybrid attention gate fusion module is subjected to the local perception self-attention module to obtain the output of the local perception self-attention module, and the result after the 1x1 convolution layer and the scaling factor α in the residual path of the input of each encoder unit is added to the output of the local perception self-attention module to obtain the output of each encoder unit.

[0025] The convolution kernel size of the convolution layer in the first to fourth encoder units is 1x9, 9x1, 7x7 and 3x3 respectively.

[0026] As a preferred scheme of the present application, the hybrid attention gate fusion module takes the results F m and F rAfter splicing, a 1*1 convolutional layer and a PReLU activation function are sequentially connected to obtain compressed features F comp ;

[0027] The mixed attention gate fusion module further includes a channel attention branch and a spatial attention branch, wherein the channel attention branch includes a global pooling layer, a 1*1 convolutional layer, a PReLU activation function, a 1*1 convolutional layer and a Sigmoid activation function connected in sequence, and the spatial attention branch includes a 3*3 convolutional layer, a PReLU activation function, a 3*3 convolutional layer and a Sigmoid activation function connected in sequence;

[0028] The compressed features F comp The channel attention map M c and 1-M c are output through the channel attention branch. comp The spatial attention map M s and 1-M s are output through the spatial attention branch. c 1-M s and 1-M r are multiplied by F c to obtain a first multiplication result M s is multiplied by M m and F comp to obtain a second multiplication result, the first gate result is obtained by adding the second multiplication result and the first multiplication result through a gate mechanism β1, and the compressed features F local are obtained through a gate mechanism β2, and the second gate result is added to the first gate result as the output of the mixed attention gate fusion module.

[0029] As a preferred scheme of the present application, the local perception self-attention module includes a local convolution branch and a global self-attention branch, the local convolution branch includes a 3*3 convolutional layer, a GELU activation function and a 1*1 convolutional layer connected in sequence, and the input of the local perception self-attention module is processed through the local convolution branch to obtain features F local ;

[0030] The global self-attention branch includes three paths, in the first path, the input of the local perception self-attention module is processed through a 1*1 convolutional layer, and the convolution result is deformed to obtain a query matrix; in the second path, the input of the local perception self-attention module is processed through a 1*1 convolutional layer, and the convolution result is deformed to obtain a key matrix; in the third path, the input of the local perception self-attention module is processed through a 1*1 convolutional layer, and the convolution result is deformed to obtain a value matrix; after transposing the query matrix, the query matrix is multiplied by the key matrix to obtain a third multiplication result, the third multiplication result is processed through a Softmax activation function to obtain an attention feature map, the attention feature map is multiplied by the value matrix to obtain a fourth multiplication result, and the fourth multiplication result is deformed to obtain the output F of the global self-attention branch.att ;

[0031] F att and F local After dynamic weighted fusion, the residual connection is performed with the input of the local perception self-attention module to obtain the output of the local perception self-attention module.

[0032] As a preferred scheme of the present application, in step 4, the decoder network comprises, in sequence, a full connection layer, a first decoder unit, a second decoder unit, a hybrid attention gate fusion module, a local perception self-attention module and a Sigmoid activation function; the code word of the preset dimension is taken as the input of the full connection layer, and the Sigmoid activation function outputs the estimated CSI feature map.

[0033] The first and second decoder units have the same structure and each comprises, in sequence, a first to fourth CBP module, each of which comprises a convolution layer, a batch normalization layer and a PReLU activation function; the input of each decoder unit is sequentially input into the first, second, third and fourth CBP modules, and the output of the fourth CBP module is added to the input of each decoder unit as the output of each decoder unit.

[0034] As a preferred scheme of the present application, in step 6, the estimated CSI feature map is zero-padded to restore the padded CSI feature map to the original spatial-frequency domain dimension, and the padded CSI feature map is converted back to the spatial-frequency domain by using a two-dimensional inverse discrete Fourier transform to obtain the final reconstructed CSI matrix.

[0035] A computer device comprises a memory, a processor, and a computer program stored in the memory and capable of running on the processor, and the processor implements the steps of the large-scale MIMO CSI compression method based on the attention fusion autoencoder when executing the computer program.

[0036] A computer readable storage medium stores a computer program, and the computer program implements the steps of the large-scale MIMO CSI compression method based on the attention fusion autoencoder when executed by a processor.

[0037] Compared with the prior art, the above technical scheme has the following technical effects:

[0038] 1、The application can capture the internal structure and key features of CSI data more deeply and accurately by introducing parallel HAGF attention fusion and global / local LASA self-attention mechanism. Even if trained and tested in a single scene, its reconstruction accuracy (measured by NMSE) is significantly better than existing advanced benchmark models such as ACRNet. With normalized mean square error (NMSE) as the measurement standard, AiANet can reduce up to 3.42dB compared to existing autoencoder-based methods, effectively reducing feedback overhead and providing key technical support for efficient and reliable large-scale MIMO systems.

[0039] 2、The application combines a hybrid training strategy, and AiANet can learn a general channel feature representation across different scenes while retaining the ability to adapt to scene-specific changes. Compared with models trained only in a single scene, which suffer from a sharp decline in performance when tested across scenes, the application method exhibits strong robustness and generalization performance, effectively processing CSI data from different environments and significantly reducing the model's dependence on a specific training environment.

[0040] 3、The application proposes a gated dense connection (GDC) structure that dynamically adjusts feature flow through a learnable gating mechanism, promoting the transmission and reuse of effective features in the deep network while suppressing redundant information, which helps to improve network expression ability and training stability.

[0041] 4、The method proposed by the application has higher CSI reconstruction accuracy and excellent generalization ability. Higher CSI reconstruction accuracy provides a foundation for base stations to achieve more accurate downlink precoding and beamforming, which helps to improve the throughput and spectral efficiency of actual communication systems. Excellent generalization ability greatly simplifies model deployment and maintenance, reduces the need to develop and manage multiple specialized models for different deployment environments, and improves the economic efficiency and scalability of the technical solution. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a CSI compression workflow schematic based on an autoencoder provided by the application;

[0043] Figure 2 is an AiANet encoder network structure schematic provided by the application;

[0044] Figure 3 is an internal structure schematic of an encoder unit provided by the application;

[0045] Figure 4 is a structure schematic of a hybrid attention gate fusion (HAGF) module provided by the application;

[0046] Figure 5is a structural schematic diagram of a local sensing self-attention (LASA) module proposed by the application;

[0047] Figure 6 is a structural schematic diagram of an AiANet decoder network proposed by the application;

[0048] Figure 7 is a comparison diagram of CSI visual reconstruction effects of the method (AiANet) proposed by the application and the existing method (ACRNet) under hybrid training conditions when the compression ratio η = 1 / 16, wherein (a) is an original CIS matrix, (b) is a CIS matrix reconstructed by the ACRNet method, and (c) is a CIS matrix reconstructed by the AiANet method. DETAILED DESCRIPTION

[0049] Embodiments of the application are described in detail below with reference to the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the application, and cannot be explained as a limitation of the application.

[0050] As shown in Figure 1 , the application proposes a large-scale MIMO CSI compression method based on an attention fusion autoencoder, and the specific implementation process is as follows:

[0051] Step 1: CSI acquisition and preprocessing

[0052] At the user equipment (UE) end, first, the original spatial-frequency domain CSI matrix H of the downlink is acquired Herein, represents the number of subcarriers, N t represents the number of antennas of the base station. In order to utilize the sparsity or structural characteristics of the channel in the angle-delay domain, the H r is applied to two-dimensional discrete Fourier transform (2D-DFT) to convert it to the angle-delay domain to obtain the matrix H Considering that the multipath delay spread of the actual channel is usually limited, most of the energy is concentrated in the first part of the delay domain, therefore, according to the maximum expected delay spread, a truncation threshold N c (for example, N c = 32, and ) is selected, only the first n c rows of H are retained, and the truncated angle-delay domain CSI matrix H is obtained. This is intended to reduce the amount of data for subsequent processing and remove noise or non-significant components. Subsequently, the real part and the imaginary part of the complex matrix H a are separated into two independent channels to form a real-valued tensor with dimensions of 2×N c ×N t(2x32x32 in this embodiment). This tensor is the pre-processed feature map, which is the input of the subsequent encoder network.

[0053] Step two: encoder processing and compression

[0054] The pre-processed CSI feature map with dimension 2x32x32 is input into the encoder network (AiANet) designed by the present application, as shown in Figure 2 . The encoder network adopts a deep hierarchical structure, which is composed of four cascaded encoder units, for gradually extracting and abstracting CSI features.

[0055] To capture channel features of different scales and directions, different sizes of convolution kernels can be used inside each encoder unit. For example, the first encoder unit can use a 1x9 convolution kernel (output channel number 16) to capture narrowband vertical features; the second uses a 9x1 convolution kernel (output channel number 32) to capture narrowband horizontal features; the third uses a 7x7 convolution kernel (output channel number 64) to detect larger scattering cluster distribution; the fourth uses a 3x3 convolution kernel (output channel number 128) to extract fine local details. The number of channels gradually increases with the depth of the network (2→16→32→64→128), while the spatial resolution (32x32) remains unchanged in the present application.

[0056] Each encoder unit (structure as shown in Figure 3 ) contains a main processing path and a residual path inside. The main processing path usually contains convolution layers, batch normalization (BatchNorm) layers and activation functions (such as PReLU). Crucially, the main processing path integrates two attention mechanism modules proposed by the present application for feature refinement:

[0057] Hybrid Attention Gate Fusion (HAGF) module (as shown in Figure 4 ): receives the main path (spatial convolution) features and residual features of the encoder unit respectively, calculates the channel attention map (Mc) and the spatial attention map (Ms) in parallel, applies the attention map to the main features and the complementary residual features through the gating mechanism (parameters β1, β2), and fuses the compressed features F comp , to realize dynamic spatial-channel feature fusion.

[0058] Local-aware Self-Attention (LASA) module (as shown in Figure 5 ): receives input features x in , through a parallel global self-attention branch (the middle three parallel paths, calculating F att ) and a local convolution branch (the uppermost path, calculating F local), respectively capturing long-range dependencies and local patterns. Both outputs are adaptively weighted by a learnable channel gating function σ(ω) and fused with the original input via a residual connection (scaling factor γ) to balance global and local information.

[0059] To enhance the information flow among individual encoder units, gated dense connections (GDC) are used in the design of the encoder architecture. Specifically, for any one encoder unit in the encoder, its input not only receives the output of the previous unit, but also fuses features from all previous stages (including the initial input H a and the outputs of all previous encoder units). As shown in FIG. 2B, these features from previous stages are each connected via an independent residual connection. On this connection, the feature first goes through a 1x1 convolutional block for channel transformation, and then is weighted by a learnable gating coefficient γ. Finally, all weighted features are fused with the feature on the main path in an additive manner to serve as the input of the current unit. This mechanism allows the model to dynamically control the amount of historical information flowing in. The process can be expressed in the following equation: Figure 3

[0060]

[0061] where A k represents the set of operations in the kth encoder module; T i (k) (·) utilizes a 1x1 convolutional operation to change the number of channels, so that different encoder modules can be connected via a residual connection; W i (k) is the gating coefficient used to balance the output weights of each encoder module in the dense connection; and represents the weight of the input tensor. This connection allows the network to dynamically and selectively reuse features from the shallow layers, enhancing information flow and helping to alleviate the vanishing gradient problem.

[0062] After being processed by the four encoder units with different convolutional kernels, the high-dimensional feature map (128x32x32) is input to a fully connected (FC) layer via a final feature integration layer (e.g., a 1x1 convolutional layer, optionally followed by global average pooling or flattening operation). The FC layer maps the features to a low-dimensional code vector s with dimension M = 2xN c x N t x η = 2048η, where η is a pre-set CSI compression rate (e.g., η = 1 / 4, 1 / 16, 1 / 32, 1 / 64).

[0063] Step three: code vector feedback

[0064] ​The generated low-dimensional codeword s can be optionally quantized to further reduce the bit number according to system requirements, then modulated, and finally transmitted to the base station (BS) through the finite-capacity uplink feedback channel. In the simulation of the present embodiment, for the sake of simplifying the analysis, it is generally assumed that the feedback channel is error-free and does not need to be quantized, and the codeword s is directly passed to the decoder.

[0065] Step four: Decoder processing and reconstruction

[0066] At the base station (BS) end, the received codeword s (dimension M) is input to the AiANet decoder network, as shown in Figure 6 The task of the decoder is to restore the compressed codeword to an estimate of the original CSI feature map.

[0067] The decoder first projects and reshapes the low-dimensional codeword s back to the spatial dimension (2x32x32) of the preprocessed CSI feature map through a fully connected (FC) layer. This provides an initial basis for reconstruction.

[0068] Then, the initial reconstructed feature map is sequentially passed through two cascaded decoder units. Each decoder unit can internally include a series of convolutional layers (e.g., using different convolutional kernels such as 3, 5, 7 for multi-scale processing) and residual connections, aiming to gradually refine the feature representation and restore the lost details.

[0069] In order to further improve the reconstruction quality and balance the efficiency of information flow, the decoder integrates two attention mechanism modules: an HAGF module and a LASA module. These modules enhance the discriminability of features and suppress noise and artifacts by adaptively calibrating and modeling the dependence of channels and spaces in the decoding process.

[0070] The last layer of the decoder network is an output activation function layer, which can use the Sigmoid function to normalize the input value range to (0, 1) (if the input CSI data is normalized to this range). The output is the estimated CSI feature map with a dimension of 2x32x32.

[0071] Step five: Decompression post-processing

[0072] The estimated CSI feature map output by the decoder (containing two channels of real and imaginary parts) is combined into a complex matrix Then, it is zero-padded in the delay domain (dimension of rows) to restore its dimension to Finally, the zero-padded matrix is applied with a two-dimensional inverse discrete Fourier transform (2D-IDFT) to convert it from the angle-delay domain back to the frequency-space domain, obtaining the final reconstructed CSI matrix

[0073] Step six: Model training

[0074] To achieve superior cross-scenario generalization performance, the embodiment adopts a hybrid training strategy.

[0075] Dataset construction: Prepare a CSI sample dataset containing at least two representative wireless scenarios. For example, data can be generated using a widely recognized channel model such as COST2100, containing an indoor scenario (5.3 GHz frequency band is used in the embodiment, N t = 32) and an outdoor scenario (300 MHz frequency band is used in the embodiment, N t = 32). Mix the data of the two scenarios in approximately equal proportions (e.g., 1:1) to form a training set (e.g., 100k samples), a validation set (e.g., 30k samples), and a test set (e.g., 20k samples). Each sample is a preprocessed CSI feature map H a .

[0076] Training settings: Implement the AiANet model using a deep learning framework (such as PyTorch). Use the AdamW optimizer to update the model parameters, with an initial learning rate η0= 1 × 10 -4 , β1= 0.9, and β2= 0.999. The weight decay can be set as needed (1 × 10 -4 or disabled in the embodiment). Use the cosine annealing learning rate scheduling strategy to dynamically adjust the learning rate during training, η min = 1 × 10 -6 , T max = 1000 (total number of training rounds). Use PReLU as the main activation function. The loss function for training is the normalized mean square error (NMSE) between the original preprocessed CSI feature map H a and the decoder reconstruction . The goal is to minimize this loss function. The training process uses batch processing, with 200 samples per batch, and the training data is randomly shuffled before each epoch.

[0077] Step seven: Performance evaluation

[0078] After training, use an independent test set to evaluate the model performance.

[0079] Evaluation indicators: Mainly use NMSE:

[0080]

[0081] and cosine similarity (ρ):

[0082]

[0083] The NMSE is used to quantify the consistency between the reconstructed CSI and the original CSI. The lower the NMSE, the better, and the closer the value of p to 1, the better.

[0084] Evaluation scenarios:

[0085] Mixed test: The average performance of the model is evaluated using a mixed test set containing indoor and outdoor samples.

[0086] Scene-specific test: The model performance is evaluated on pure indoor and pure outdoor test sets respectively to test its specific performance and generalization ability in different scenarios.

[0087] Comparative analysis: The performance of AiANet is compared with existing advanced benchmark models (such as ACRNet, DCRNet, etc.) under the same training (including mixed training) and testing settings. The experimental results show that under mixed training, the NMSE of AiANet on indoor and outdoor test sets is significantly lower than that of ACRNet, proving its stronger learning and generalization ability.

[0088] Ablation experiment: To verify the effectiveness of the key components (HAGF, LASA, GDC) in the present application, an ablation experiment can be conducted: remove or replace these modules respectively, and compare the changes in model performance. The ablation experiment results show that removing any key component will cause performance degradation, proving the contribution of these designs.

[0089] Visual comparison: Visualize the reconstructed CSI matrix, as shown in (a)-(c) of FIG. 1, and compare it with the original CSI and the CSI reconstructed by other methods, to intuitively show the advantages of the present application in restoring fine structures and sparse patterns. Figure 7

[0090] The present application fully demonstrates the large-scale MIMO CSI compression method based on attention fusion autoencoder by detailing the network structure of AiANet (including encoder / decoder units, HAGF, LASA, GDC, and other key components), the end-to-end CSI compression and reconstruction process, and the recommended mixed training strategy. This method can effectively improve the accuracy and cross-scene generalization ability of CSI compression, and has important theoretical significance and practical application value.

[0091] Based on the same inventive concept, the embodiments of the present application provide a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the computer program to implement the steps of the large-scale MIMO CSI compression method based on attention fusion autoencoder.

[0092] ​Based on the same inventive concept, the embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the large-scale MIMO CSI compression method based on the attention fusion autoencoder.

[0093] Those skilled in the art will appreciate that embodiments of the application can be supplied as a method, a system, or a computer program product. Therefore, the application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer usable program code.

[0094] The application is described with reference to the flowcharts and / or block diagrams of the methods, apparatus (systems), and computer program products according to the embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus generate a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for carrying out the functions specified in the flowcharts and / or block diagrams.

[0095] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer readable memory produce a product including an instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for carrying out the functions specified in the flowcharts and / or block diagrams.

[0096] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for carrying out the functions specified in the flowcharts and / or block diagrams.

[0097] The above examples only illustrate the technical idea of the present application, and cannot be used to limit the protection scope of the present application. Any modification made according to the technical idea of the present application on the basis of the technical scheme falls within the protection scope of the present application.

Claims

1. A large-scale MIMO CSI compression method based on attention fusion autoencoder, characterized in that, The method comprises the following steps: Step 1, obtaining a mixed dataset composed of a group of CSI samples of at least two different types of wireless communication scenarios, and preprocessing the mixed dataset to obtain a preprocessed mixed dataset, each sample in the preprocessed mixed dataset being a preprocessed CSI feature map; The preprocessed mixed dataset is divided into a training set, a validation set and a test set; Step 2, constructing an autoencoder network model, the autoencoder network model comprising an encoder network and a decoder network, wherein the encoder network is arranged at a user equipment end, and the decoder network is arranged at a base station end; the encoder network extracts and fuses channel dimension and spatial dimension information of the preprocessed CSI feature map to generate a code word of a preset dimension; Step 3, the user equipment end feeds back the code word of the preset dimension to the base station through an uplink; Step 4, at the base station end, the decoder network is used to decompress and reconstruct the code word of the preset dimension to obtain an estimated CSI feature map; Step 5, the training set and the validation set are used to train and verify the constructed autoencoder network model to obtain a trained autoencoder network model; Step 6, the preprocessed CSI feature map in the test set is input into the trained autoencoder network model to obtain an estimated CSI feature map; the estimated CSI feature map is converted back to a spatial-frequency domain by using a two-dimensional inverse discrete Fourier transform to obtain a reconstructed CSI matrix.

2. The attention-based fusion autoencoder-based massive MIMO CSI compression method according to claim 1, wherein, The wireless communication scene in the step 1 comprises at least an indoor scene and an outdoor scene, and the CSI sample is an original spatial-frequency domain CSI matrix of downlink obtained at a user equipment end wherein, N represents the number of subcarriers t N represents the number of antennas of the base station; the CSI samples of different types of wireless communication scenes are mixed at the same proportion to form a mixed data set; A two-dimensional discrete Fourier transform is performed on each CSI sample in the mixed data set to obtain an angle-delay domain CSI matrix H, and a truncation threshold N is set c , The first N c rows of H are intercepted to obtain a truncated angle-delay domain CSI matrix The real part and the imaginary part of the matrix H a are separated into two independent channels to form a preprocessed CSI feature map. 3.The large-scale MIMO CSI compression method based on attention fusion auto-encoder according to claim 1, wherein, In the step 2, the encoder network comprises first to fourth encoder units, and each encoder unit is interconnected through a gated dense connection structure; The preprocessed CSI feature map is taken as an input of a first encoder unit, and the preprocessed CSI feature map is sequentially subjected to a 1x1 convolution layer and a gating coefficient γ 01 The result after the gating coefficient is added to the output of the first encoder unit to obtain a first addition result as an input of a second encoder unit; The preprocessed CSI feature map is sequentially added to the output of the second encoder unit and the first addition result, and the result after the 1x1 convolution layer and the gating coefficient γ 02 The preprocessed CSI feature map is sequentially added to the output of the second encoder unit and the first addition result, and the result after the 1x1 convolution layer and the gating coefficient γ 12 The preprocessed CSI feature map is sequentially added to the output of the second encoder unit and the first addition result, and the result after the 1x1 convolution layer and the gating coefficient γ The preprocessed CSI feature map is sequentially added to the output of the third encoder unit, the first addition result, the result after the 1x1 convolution layer and the gating coefficient γ 03 The second addition result, the result after the 1x1 convolution layer and the gating coefficient γ 13 The third addition result, the result after the 1x1 convolution layer and the gating coefficient γ 23 The third addition result, the result after the 1x1 convolution layer and the gating coefficient γ The preprocessed CSI feature map sequentially passes through a 1x1 convolution layer and a gating coefficient γ 04 The output of the fourth encoder unit, the first addition result, sequentially passes through a 1x1 convolution layer and a gating coefficient γ 14 The second addition result, sequentially passes through a 1x1 convolution layer and a gating coefficient γ 24 The third addition result, sequentially passes through a 1x1 convolution layer and a gating coefficient γ 34 The fourth addition result is obtained by adding the result after the fourth addition result, the first addition result, the second addition result, the third addition result, and the third addition result. The fourth addition result is sequentially subjected to a 1x1 convolution layer and a full connection layer to generate the code word of the preset dimension.

4. The attention-based fusion autoencoder-based massive MIMO CSI compression method according to claim 3, characterized in that, The first to fourth encoder units have the same structure, and each encoder unit comprises a main processing path and a residual path, the main processing path comprising a convolution layer, a batch normalization layer, a PReLU activation function, a hybrid attention gate fusion module and a local perception self-attention module; The input of each encoder unit is sequentially processed through the convolutional layer, batch normalization layer, and PReLU activation function on the main processing path, resulting in the output F. m And the result F of the input of each encoder unit through the 1×1 convolutional layer on the residual path. r All of these are used as inputs to the hybrid attention-gated fusion module. The hybrid attention-gated fusion module passes through the local perception self-attention module to obtain the output of the local perception self-attention module. The inputs of each encoder unit are sequentially passed through a 1×1 convolutional layer on the residual path and the scaling factor α. The result is added to the output of the local perception self-attention module to obtain the output of each encoder unit. The convolution kernel sizes of the convolution layers in the first to fourth encoder units are 1x9, 9x1, 7x7 and 3x3 respectively.

5. The attention-based fusion autoencoder-based massive MIMO CSI compression method according to claim 4, characterized in that, The mixed attention gate fusion module is used for obtaining a result F m and F r After splicing, a 1*1 convolution layer and a PReLU activation function are sequentially used to obtain compressed features F comp ; The hybrid attention gate fusion module further comprises a channel attention branch and a spatial attention branch, wherein the channel attention branch comprises a global pooling layer, a 1x1 convolution layer, a PReLU activation function, a 1x1 convolution layer and a Sigmoid activation function connected in sequence, and the spatial attention branch comprises a 3x3 convolution layer, a PReLU activation function, a 3x3 convolution layer and a Sigmoid activation function connected in sequence; compressed feature F comp The channel attention branch outputs a channel attention map M c and 1-M c , the compressed feature F comp The spatial attention branch outputs a spatial attention map M s and 1-M s , 1-M c and 1-M s and F r The first multiplication result M c and M s and F m The second multiplication result is obtained, and the second multiplication result is added to the first multiplication result. After the gating mechanism β1, the first gating result is obtained, the compressed feature F comp The second gating result is obtained through the gating mechanism β2, and the second gating result is added to the first gating result as the output of the hybrid attention gating fusion module. 6.The large-scale MIMO CSI compression method based on attention fusion auto-encoder of claim 4, wherein, The local perception self-attention module comprises a local convolution branch and a global self-attention branch, the local convolution branch comprises a 3*3 convolution layer, a GELU activation function and a 1*1 convolution layer connected in sequence, and the input of the local perception self-attention module obtains the feature F through the local convolution branch local ; The global self-attention branch includes three paths. In the first path, the input of the local perception self-attention module is subjected to a 1x1 convolution layer, and then the convolution result is deformed to obtain a query matrix. In the second path, the input of the local perception self-attention module is subjected to a 1x1 convolution layer, and then the convolution result is deformed to obtain a key matrix. In the third path, the input of the local perception self-attention module is subjected to a 1x1 convolution layer, and then the convolution result is deformed to obtain a value matrix. After transposing the query matrix, the query matrix is multiplied by the key matrix to obtain a third multiplication result. The third multiplication result is subjected to a Softmax activation function to obtain an attention feature map. The attention feature map is multiplied by the value matrix to obtain a fourth multiplication result. The fourth multiplication result is deformed to obtain the output F of the global self-attention branch att ​ To F att and F local After dynamic weighted fusion, the residual connection is performed with the input of the local perception self-attention module to obtain the output of the local perception self-attention module. 7.The large-scale MIMO CSI compression method based on attention fusion auto-encoder of claim 4, characterized in that, In the step 4, the decoder network comprises a full connection layer, a first decoder unit, a second decoder unit, a hybrid attention gate fusion module, a local perception self-attention module and a Sigmoid activation function connected in sequence; the code word of the preset dimension is input into the full connection layer, and the Sigmoid activation function outputs the estimated CSI feature map; The first and second decoder units are identical in structure, and each comprises first to fourth CBP modules connected in sequence, each CBP module comprising a convolution layer, a batch normalization layer and a PReLU activation function; the input of each decoder unit is sequentially input into the first, second, third and fourth CBP modules, and the output of the fourth CBP module is added to the input of each decoder unit as the output of each decoder unit. 8.The large-scale MIMO CSI compression method based on attention fusion auto-encoder of claim 2, wherein, In step 6, the estimated CSI feature map is zero-padded, the padded CSI feature map is restored to the original spatial-frequency domain dimension, the padded CSI feature map is converted back to the spatial-frequency domain by using a two-dimensional inverse discrete Fourier transform, and a final reconstructed CSI matrix is obtained.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The processor executes the computer program to implement the steps of the large-scale MIMO CSI compression method based on the attention fusion autoencoder according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the steps of the large-scale MIMO CSI compression method based on the attention fusion autoencoder according to any one of claims 1 to 8.

Citation Information

Patent Citations

  • Large-scale MIMO system CSI feedback method based on long-short-term attention mechanism

    CN110912598A

  • Large-scale MIMO CSI feedback method fusing self-attention mechanism

    CN119152340A