Channel state information feedback method based on multi-scale physical perception puzzle mechanism

Abstract

The application discloses a channel state information feedback method based on a multi-scale physical perception puzzle mechanism and applied to a frequency division duplex large-scale MIMO system. A two-dimensional discrete Fourier transform is used to convert a space-frequency domain channel matrix into an angle-delay domain sparse matrix and complete preprocessing; a PAM-Net network containing a main branch reconstruction branch and a self-supervised auxiliary branch is constructed, a multi-scale physical perception puzzle mechanism and a physical guided anisotropic dynamic penalty strategy are introduced; a joint loss function is used for end-to-end offline training, and after training, the auxiliary branch is decoupled and stripped, and only the main branch is reserved to perform online compression and reconstruction. The application has the beneficial effects that: under the premise of not increasing the online inference complexity, the multi-scale physical feature perception ability of the channel and the CSI reconstruction accuracy are significantly improved, the feedback overhead is effectively reduced, and the application is suitable for a low-latency and high-reliable communication scene of a 6G super large-scale antenna array.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to a channel state information feedback method based on a multi-scale physical sensing mosaic mechanism. Background Technology

[0002] As mobile communication networks evolve towards sixth generation (6G), massive MIMO technology is rapidly moving towards very large-scale antenna arrays (VMAs) to meet the ever-increasing demands for spectral efficiency and massive connectivity. In widely deployed frequency division duplex (FDD) systems, improved system performance depends on the ability of user equipment (UEs) to provide accurate channel state information (CSI). However, since the physical dimension of downlink CSI is proportional to the number of base station antennas, the continuous expansion of array size has led to a dramatic increase in the amount of feedback data. The massive feedback overhead has become a key challenge restricting system performance improvement. Therefore, exploring more efficient channel feature compression and high-precision reconstruction mechanisms under the constraint of lower feedback overhead has become a critical challenge that urgently needs to be overcome.

[0003] To address this issue, early research proposed codebook-based quantization feedback schemes and compressed sensing (CS)-based feedback techniques. However, codebook methods suffer from the "curse of dimensionality" when the number of antennas increases dramatically, while compressed sensing methods heavily rely on the prior assumption of channel sparsity and have high iterative algorithm complexity, making them difficult to meet real-time requirements. In recent years, deep learning (DL) schemes, represented by CsiNet, have significantly surpassed traditional methods in reconstruction accuracy and speed due to their powerful nonlinear mapping capabilities, establishing their dominance in the field. As research deepens, the focus is shifting from simple network architecture design to innovation in training objectives and optimization methods, further improving model performance by introducing denoising strategies, joint optimization, and data augmentation. To address this issue, early research proposed codebook-based quantization feedback schemes and compressed sensing (CS)-based feedback techniques. These methods discretize the channel space using a pre-defined codebook, but encounter the "curse of dimensionality" when the number of antennas increases dramatically, leading to huge storage and computational overhead. To circumvent the complexity of codebook methods, compressed sensing (CS)-based feedback techniques emerged, utilizing the sparsity of the channel in the angle-delay domain for low-dimensional measurements. Although compressed sensing has achieved theoretical success, its performance is highly dependent on the prior assumption of channel sparsity, and the iterative nature of the reconstruction algorithm introduces high computational complexity and latency, making it difficult to meet the real-time requirements of future communications.

[0004] With the rapid rise of deep learning (DL), the design of CSI feedback networks based on autoencoder architectures has gradually become a research hotspot. CsiNet has demonstrated that neural networks significantly outperform traditional algorithms in feature compression and reconstruction speed. To further improve performance, subsequent research has begun to explore the introduction of more advanced network architectures, such as attention mechanisms and Transformers, as well as physically-defined networks designed for CSI horizontal and vertical paths, aiming to better capture channel features and reduce computational overhead. These evolutions have led to the continuous maturation of deep learning-based CSI feedback schemes, providing new ideas and directions for solving the challenges of high-dimensional CSI feedback.

[0005] However, although these methods overcome the shortcomings of traditional optimization algorithms to some extent, they still face some challenges and limitations in practical applications, specifically in the following aspects: (1) The computational complexity remains high: In order to further improve the representation ability and reconstruction accuracy of channel features, existing deep learning feedback research often relies on introducing global self-attention mechanisms or designing complex spatial computing modules. However, these complex network architectures inevitably increase the online inference complexity during actual deployment. Especially in large-scale MIMO systems, the CSI matrix dimension is extremely high, and the complex feature interaction and spatial computation bring high computational overhead and processing latency, which seriously affects the real-time performance of system feedback and the deployment efficiency on resource-constrained devices. Even with some lightweight methods, how to effectively improve model performance without increasing online inference complexity remains a key challenge in current practical applications.

[0006] (2) The system reconstruction accuracy still needs further improvement: The optimization objectives and auxiliary strategies of existing methods often lack targeted design for the physical heterogeneity of CSI. On the one hand, the mainstream mean square error (MSE) loss function is numerically fitted, which easily smooths the error globally, causing the network to be biased towards high-energy regions and making it difficult to restore the local critical spatial topology and multipath contours. On the other hand, auxiliary tasks such as self-supervision introduced to compensate for the loss of structural information usually only use general spatial processing and equalization of error allocation. Since the physical asymmetry and cross-scale characteristics of CSI in the time delay domain (sparseness) and angle domain (smoothing correlation) are ignored, this homogeneous design lacks physical guidance. The above shortcomings not only easily lead to the loss of critical paths, but also fail to accurately capture the micro-texture boundaries and macro-path clusters, making it difficult to guide the network to deeply understand the multi-scale spatial distribution of the channel, and ultimately limiting the further improvement of the system reconstruction accuracy.

[0007] A prior art method for CSI feedback in large-scale MIMO systems, based on a multi-resolution fusion convolutional feedback network (publication number CN114567359A), employs multi-scale convolution and autoencoders to achieve channel feature compression and reconstruction, thereby reducing feedback overhead and improving reconstruction accuracy. However, it does not utilize a multi-scale physical sensing jigsaw puzzle mechanism or a physically guided anisotropic dynamic penalty strategy. It also does not incorporate self-supervised training based on the angle-delay domain channel physical characteristics, nor does it further improve reconstruction accuracy without increasing online inference complexity through weight sharing and decoupled deployment. Consequently, it cannot fully perceive the multi-scale spatial topology and physical heterogeneity of the channel, and its reconstruction accuracy and physical adaptability remain limited.

[0008] Therefore, the channel state information feedback method of the present invention aims to effectively improve the reconstruction accuracy in FDD massive MIMO systems without increasing the computational complexity of online inference, and overcome the shortcomings of existing deep learning technologies in terms of computational overhead and training methods. Summary of the Invention

[0009] The main objective of this invention is to overcome the shortcomings of existing technologies and propose a channel state information feedback method based on a multi-scale physical sensing jigsaw puzzle mechanism. The system constructed using this method can effectively guide the network to understand the multi-scale physical spatial distribution of the channel without increasing the complexity of online inference computation, ultimately achieving a significant improvement in reconstruction accuracy.

[0010] The technical solution adopted by this invention to achieve the above objectives is: a channel state information feedback method based on a multi-scale physical sensing mosaic mechanism, comprising the following steps: Step 1: Construct a single-cell FDD massive MIMO communication system model. Transform the spatial-frequency domain channel matrix into an angle-delay domain channel matrix through two-dimensional discrete Fourier transform, and perform truncation, decoupling and normalization preprocessing to obtain the network input. Step 2: Construct a PAM-Net network that includes a backbone reconstruction branch and a self-supervised auxiliary branch. In the auxiliary branch, introduce a multi-scale physical perception jigsaw puzzle mechanism and a physical-guided anisotropic dynamic penalty strategy. Step 3: Use the joint loss function for offline end-to-end training. After training, decouple and remove the self-supervised auxiliary branches, and use only the main branch to complete online compression and reconstruction.

[0011] Preferably, in step 1, the domain transformation is performed using a two-dimensional discrete Fourier transform to convert the space-frequency domain channel matrix into an angle-delay domain sparse matrix; the two-dimensional discrete Fourier transform satisfies the following formula: (1) In the formula, For angle-time delay domain sparse matrix, The channel matrix is ​​in the spatial frequency domain. and They represent dimensions as follows: and The orthogonal discrete Fourier transform (DFT) matrix.

[0012] Preferably, the system performs a truncation operation on the angle-time delay domain sparse matrix, retaining only the first part of the matrix. The process begins by filtering out redundant background noise. Then, the truncated complex channel matrix physical map is decoupled into two independent channels, one real and one imaginary, and extreme value scaling is performed to normalize them to a value range of [0,1]. After this preprocessing step, the final input data for the network is obtained. .

[0013] Preferably, PAM-Net uses a weight-shared encoder, where the self-supervised auxiliary branch and the trunk reconstruction branch share encoder parameters, and the self-supervised auxiliary branch is removed during online deployment.

[0014] Preferably, in step 2, the multi-scale physical perception jigsaw puzzle processing adaptively divides the input matrix into 2×2, 4×4 or 8×8 non-overlapping patches, and after the patches are disordered, the original positions are restored by the multi-scale prediction head.

[0015] Preferably, in step 2, during the physically guided anisotropic dynamic penalty processing, the channel element amplitude satisfies the formula: (4) in, Coordinates are Channel amplitude at that location, and These are the real and imaginary parts of the matrix, respectively.

[0016] Preferably, the local physical complexity of the channel satisfies the formula: (5) in, and These are the penalty sensitivity coefficients in the time delay domain and the angle domain, respectively. Coordinates are Local physical complexity of the channel.

[0017] Preferably, in order to stably map the extracted complexity to penalty weights, this module uses mean normalization and extreme value clipping (Clip) mechanisms to generate weights, ultimately generating the first... Dynamic physics weighting coefficients for each tile : (7) in, This represents the global average complexity of all tiles at the current partitioning scale. and These are the preset lower and upper limits for punishment tolerance, respectively.

[0018] Preferably, in step 3, based on the above network architecture, this paper constructs a joint optimization objective consisting of the main task reconstruction loss and the self-supervised auxiliary task loss, wherein the joint loss function satisfies the formula: (8) (9) Among them, the reconstruction loss of the main task The standard mean squared error (MSE) is used to ensure the numerical similarity between the reconstructed CSI matrix and the original matrix; It is a balance coefficient that decays dynamically with the number of training rounds; For puzzle loss in self-supervised auxiliary tasks.

[0019] Preferably, the cross-entropy loss is calculated using a weighted average of physical dynamic weights, and the weighted cross-entropy loss satisfies the formula: (10) In this formula, This represents the total number of tiles generated under the current multi-scale segmentation strategy; It is a Physics-Guided Anisotropic Dynamic Penalty Module (PG-ADP) for the first Dynamic physics weighting coefficients generated individually for each shuffled tile; The target label for the true original location of the tile; Then represents the first output of the multi-scale prediction head. The tile belongs to the first The predicted probability of each location. By... As a penalty multiplier, the error penalty incurred when the network mispredicts the location of tiles containing complex physical principal path boundaries forces the network to focus on reconstructing the spatial structure. Finally, the physically weighted auxiliary loss is merged with the backbone MSE loss to jointly drive the network's joint gradient backpropagation.

[0020] (1) A single-cell FDD massive MIMO communication system model is constructed. The spatial-frequency domain channel matrix is ​​transformed into the angle-delay domain using two-dimensional discrete Fourier transform (2D-DFT) for accurate modeling. The characteristics of the user-end encoder and the base station decoder are comprehensively considered, and the two work together to form the final channel state information feedback system. At the same time, based on the constraints of feedback overhead and online inference computation complexity, the overall system objective is determined, aiming to maximize the reconstruction accuracy of channel state information. Through this modeling method, the performance of the feedback system under different compression ratio settings can be deeply analyzed, providing a theoretical basis for optimizing deep learning feedback algorithms, thereby improving the communication performance of massive MIMO systems.

[0021] (2) A deep learning algorithm is constructed, and a multi-scale physical perception jigsaw puzzle mechanism and a dynamic penalty strategy are introduced into it. By guiding the network to deeply perceive the physical heterogeneity of the channel and the local spatial topology, the feature representation capability of the model is improved. Based on the constructed system model, experimental data is generated for training, and the model parameters are optimized end-to-end by combining the joint loss function, thereby improving the feature representation capability and reconstruction performance of the algorithm in complex environments.

[0022] The final optimized precoding scheme is obtained based on the deep learning algorithm trained above.

[0023] Compared with existing technologies, the technical solutions adopted in this invention have the following advantages: This invention combines deep learning, multi-scale physical perception jigsaw puzzle mechanisms, physically guided dynamic penalty strategies, and joint training mechanisms to construct a large-scale MIMO channel state information feedback scheme. This scheme can avoid increasing online inference computation overhead, enhance the network's ability to express multi-scale spatial topological features of the channel, and thus improve system feedback efficiency and reconstruction accuracy. Furthermore, this invention integrates multi-scale spatial constraints and anisotropic dynamic penalty mechanisms, effectively improving the algorithm's adaptability to physically heterogeneous features, enabling the system to maintain high reconstruction performance in complex channel environments, and significantly overcoming the defects of structural information loss in traditional training methods. Attached Figure Description

[0024] Figure 1 The overall network architecture of PAM-Net; Figure 2 In Example 2 Example of scale-based tile segmentation and disordering. Detailed Implementation

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] 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, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. 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.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0028] This invention provides a large-scale MIMO channel state information feedback scheme based on a multi-scale physical sensing jigsaw puzzle mechanism. The following detailed description of the invention is provided in conjunction with specific implementation steps.

[0029] Example 1: A channel state information feedback method based on a multi-scale physical sensing mosaic mechanism, comprising the following steps: Step 1: Construct a single-cell FDD massive MIMO communication system model. Transform the spatial-frequency domain channel matrix into an angle-delay domain channel matrix through two-dimensional discrete Fourier transform, and perform truncation, decoupling and normalization preprocessing to obtain the network input. Step 2: Construct a PAM-Net network that includes a backbone reconstruction branch and a self-supervised auxiliary branch. In the auxiliary branch, introduce a multi-scale physical perception jigsaw puzzle mechanism and a physical-guided anisotropic dynamic penalty strategy. Step 3: Use the joint loss function for offline end-to-end training. After training, decouple and remove the self-supervised auxiliary branches, and use only the main branch to complete online compression and reconstruction.

[0030] Example 2: This example is based on Example 1. This example further adaptively optimizes the multi-scale segmentation scale selection, dynamic penalty coefficient, and joint loss balance coefficient. The specific steps are as follows: Step 1: Establish the basic framework of the communication system and define the specific scenarios and preconditions for channel state information feedback. Taking into account the downlink channel physical characteristics in Frequency Division Duplex (FDD) mode and the sparsification transformation process of the channel matrix from the spatial frequency domain to the angle-delay domain, this ensures end-to-end collaborative operation between the user-end feature compression encoder and the base station feature reconstruction decoder. This system model provides the theoretical and data foundation for the subsequent design of the Physical Aware Multi-Scale Network (PAM-Net) and the implementation of the multi-scale mosaic joint training scheme.

[0031] Channel Modeling: The system model in this paper is a standard single-cell frequency division duplex (FDD) massive multiple-input multiple-output (MIMO) communication scenario. In this scenario, the base station is configured with... A large-scale array of transmitting antennas is used to serve end users equipped with a single receiving antenna. The system incorporates Orthogonal Frequency Division Multiplexing (OFDM) technology, activating a total of [number missing] antennas. The number of effective subcarriers. For any number of effective subcarriers... The signal received by the user equipment (UE) on each subcarrier is the result of the combined effects of the downlink channel vector, precoding matrix, transmitted data symbols, and additive white Gaussian noise. In FDD duplex mode, the uplink and downlink occupy different frequency bands, causing complete failure of the channel's physical reciprocity. Therefore, for the base station to perform accurate downlink beamforming, it must rely on the UE to accurately estimate the downlink channel state information (CSI) locally and feed it back to the base station via a dedicated uplink. Let the downlink CSI matrix of the entire spatial frequency domain be denoted as... Considering the application of large-scale antenna arrays and high bandwidth, the physical dimension of this matrix is ​​extremely large. If we choose to directly transmit the uncompressed original matrix back... The amount of feedback data will increase linearly with the size of the antenna and the number of subcarriers, which will inevitably consume or even overwhelm the already limited uplink bandwidth resources.

[0032] Angle-Delay Domain Channel Transformation: To reduce uplink feedback overhead, this scheme utilizes the sparsity of the channel in the angle-delay domain under multipath effects. A two-dimensional discrete Fourier transform (2D-DFT) is used to transform the spatial-frequency domain channel matrix to the angle-delay domain, generating a transformed sparse matrix. The system employs a two-dimensional discrete Fourier transform (2D-DFT) at the user equipment end to transform the original dense space-frequency domain channel matrix into the angle-delay domain, thereby generating a sparse matrix. (1) In the formula, and They represent dimensions as follows: and The orthogonal discrete Fourier transform (DFT) matrix. After the above domain transformation processing, most of the channel's energy is highly compressed and concentrated in the delay domain. In the first row, the amplitude of the data elements in the remaining rows decays sharply to near zero, containing almost no more effective channel structure information. Based on this energy distribution characteristic, the system performs a truncation operation, retaining only the first few rows of the matrix. The process begins by filtering out redundant background noise. Then, the truncated complex channel matrix physical map is decoupled into two independent channels, one real and one imaginary, and extreme value scaling is performed to normalize them to a value range of [0,1]. After this preprocessing step, the final input data for the network is obtained. .

[0033] Problem Modeling: The core objective of this channel feedback system is to effectively compress high-dimensional channel features into low-dimensional codewords under extremely high compression ratio constraints, relying on an end-to-end deep learning architecture, and achieve high-fidelity matrix reconstruction at the base station. The specific processing flow involves processing the preprocessed input tensor... The encoder sent to the user equipment performs dimensionality reduction compression, generating corresponding feedback codewords based on the compression ratio set by the system. (2) In the formula, Represents the learnable parameters within the encoder network. This represents the length of the generated codeword. Through the encoder's feature mapping, the high-dimensional channel matrix is ​​transformed into... One-dimensional vector When the code word After being transmitted via the uplink and arriving at the base station, the base station inputs it to the decoder for reconstruction and decoding: (3) In the formula, This represents the CSI matrix finally reconstructed at the base station. This represents the internal parameters of the decoder network. Addressing the shortcomings of existing deep learning feedback mechanisms that lead to excessive online inference computational overhead in pursuit of reconstruction accuracy, and the structural awareness blind spots in traditional training objectives when extracting physically heterogeneous features, this invention proposes a novel Physically Aware Multi-Scale Network (PAM-Net) architecture. This architecture further optimizes channel feature compression and reconstruction without increasing online inference complexity by introducing a multi-scale jigsaw puzzle joint training mechanism.

[0034] Step 2: This scheme constructs a Physically Aware Multi-Scale Network (PAM-Net), using the angle-delay domain channel matrix as network input. It adaptively generates compressed codewords and performs high-precision reconstruction, thereby improving the overall system performance without increasing online inference overhead. This neural network adopts a joint training architecture, mainly composed of a backbone reconstruction branch and a self-supervised auxiliary branch. The backbone reconstruction branch includes an encoder for feature extraction and compression, and a decoder for feature restoration; the self-supervised auxiliary branch includes a multi-scale physically aware jigsaw puzzle mechanism (MS-JPTS) and a physically guided anisotropic dynamic penalty module (PG-ADP).

[0035] (1) Data splitting and PAM-Net collaborative reconstruction feature extraction like Figure 1 As shown in the overall PAM-Net network architecture, the network first uses a truncated dual-channel angle-delay domain CSI matrix as the raw input. The data flow is divided into tasks according to a pre-defined random triggering mechanism: the backbone reconstruction task participates in the training of each batch, while the self-supervised auxiliary task is dynamically activated only with a specific probability. In the main path, the input CSI matrix first enters the shared encoder for feature extraction and spatial dimensionality reduction, generating low-dimensional compressed codewords. Subsequently, the decoder receives the codewords and maps them back to the high-dimensional feature space, outputting the reconstructed CSI matrix, aiming to guide the network to complete basic numerical fitting through mean squared error (MSE Loss). When the auxiliary branch is activated, the weight sharing mechanism, marked by dashed lines in the figure, ensures that the physical topology awareness learned by the encoder in the auxiliary task can be synchronized to the backbone network in real time. When offline training is completed and deployed to user devices, the auxiliary branch is decoupled and removed, and the network retains only the lightweight backbone reconstruction branch to perform online inference, thereby achieving performance improvement without increasing computational overhead.

[0036] (2) MS-JPTS multi-scale spatial segmentation and disordering strategy In the Multi-Scale Physically Sensing Jigsaw Technique (MS-JPTS) mechanism, multi-scale spatial segmentation and physical dynamic weighting are introduced to force the network to actively understand the complex spatial topology relationships in the channel matrix. The CSI matrix not only exhibits asymmetry in the time delay and angle domains, but its physical structure also possesses strong cross-scale characteristics: macroscopically, it presents as a large-scale cluster of signal paths, while microscopically, it manifests as fine multipath cliffs and beam edges. To comprehensively capture these hierarchical spatial features, when the input CSI matrix enters this auxiliary branch, the system adaptively segments it proportionally. , or Non-overlapping meshes. This multi-scale strategy allows the network to operate at both coarse and fine scales. This segmentation learns the macroscopic relative positions between clusters, and can also be used for fine-grained ( The process focuses on the microscopic local texture boundaries. The original matrix is ​​divided into multiple patches and randomly rearranged, forcing the network to actively explore and reconstruct the relative spatial relationships within the channel matrix. The shuffled patches are then input into a shared encoder with identical parameters, and finally into multi-scale heads to infer the original spatial arrangement of each patch.

[0037] (3) PG-ADP Physically Guided Anisotropic Dynamic Penalty Strategy To provide physical prior guidance for the assisted jigsaw puzzle task, PG-ADP first extracts channel amplitude features. This is done considering that the input CSI matrix is ​​separated into its real parts. With the imaginary part Two independent channels, for spatial coordinates in the matrix. The element whose amplitude value Defined as: (4) Secondly, since CSI exhibits a multipath energy cliff in the time delay domain and a smooth transition of beam leakage in the angular domain, traditional isotropic gradients cannot accurately characterize this physical asymmetry. Therefore, this module describes the physical asymmetry along the time delay domain (row direction). ) and angular domain (column direction) Calculate the magnitude difference and assign different weights to the two directions to quantify the local complexity: (5) in, and These are the penalty sensitivity coefficients in the time delay domain and the angle domain, respectively. Finally, multi-scale region pooling and dynamic weight generation are performed. The PG-ADP module spatially aligns the calculated element-level complexity distribution map with the partitioned grid, and calculates the... Each tile average complexity within : (6) To stably map the extracted complexity to penalty weights, this module employs mean normalization and clipping to generate weights, ultimately producing the first... Dynamic physics weighting coefficients for each tile : (7) in, This represents the global average complexity of all tiles at the current partitioning scale. and These represent the preset lower and upper limits of the penalty tolerance, respectively. The resulting dynamic weight set... The tiles will be fed into the tail end synchronously and directly participate in the calculation of cross-entropy loss of multi-scale mosaic branches.

[0038] (4) Joint loss and decoupling optimization strategy Based on the above network architecture, this paper constructs a joint optimization objective consisting of the main task reconstruction loss and the self-supervised auxiliary task loss: (8) (9) Among them, the reconstruction loss of the main task The standard mean squared error (MSE) is used to ensure the numerical similarity between the reconstructed CSI matrix and the original matrix; It is a balance coefficient that dynamically decays with the number of training epochs; the jigsaw puzzle loss for self-supervised auxiliary tasks. The calculation is performed using cross-entropy loss with introduced physical dynamic weights: (10) In this formula, This represents the total number of tiles generated under the current multi-scale segmentation strategy; It is a Physics-Guided Anisotropic Dynamic Penalty Module (PG-ADP) for the first Dynamic physics weighting coefficients generated individually for each shuffled tile; The target label for the true original location of the tile; Then represents the first output of the multi-scale prediction head. The tile belongs to the first... The predicted probability of each location. By... As a penalty multiplier, the error penalty incurred when the network mispredicts the location of tiles containing complex physical principal path boundaries forces the network to focus on reconstructing the spatial structure. Finally, the physically weighted auxiliary loss is merged with the backbone MSE loss to jointly drive the network's joint gradient backpropagation.

[0039] Step 3: After the offline joint training of the neural network is completed, all network parameters (including weights and biases) are determined and saved as the final model. Subsequently, during the online deployment and testing phase, the system performs network decoupling, completely separating the self-supervised auxiliary branch and its related jigsaw puzzle inference and physical penalty modules. The backbone encoder at the user equipment (UE) compresses the input downlink test channel matrix to generate feedback codewords, which are then received by the decoder at the base station (BS) to reconstruct a high-precision angle-delay domain channel state information matrix.

[0040] Through the above steps, this invention constructs a large-scale MIMO channel state information feedback system with low computational complexity and high spectral efficiency, which can achieve high real-time channel feature compression and feedback in complex wireless communication environments, ensuring downlink beamforming accuracy and system communication performance.

[0041] Through the above steps, a large-scale MIMO channel state information feedback system was constructed, which significantly improves reconstruction accuracy without increasing the complexity of online inference computation. This system can achieve high real-time channel feature compression and feedback in complex multipath and spatially heterogeneous wireless communication scenarios, while effectively ensuring the accuracy of downlink beamforming and the overall communication performance of the system.

Claims

1. A channel state information feedback method based on a multi-scale physical sensing mosaic mechanism, characterized in that: Includes the following steps: Step 1: Construct a single-cell FDD massive MIMO communication system model. Transform the spatial-frequency domain channel matrix into an angle-delay domain channel matrix through two-dimensional discrete Fourier transform, and perform truncation, decoupling and normalization preprocessing to obtain the network input. Step 2: Construct a PAM-Net network that includes a backbone reconstruction branch and a self-supervised auxiliary branch. In the auxiliary branch, introduce a multi-scale physical perception jigsaw puzzle mechanism and a physical-guided anisotropic dynamic penalty strategy. Step 3: Use the joint loss function for offline end-to-end training. After training, decouple and remove the self-supervised auxiliary branches, and use only the main branch to complete online compression and reconstruction.

2. The method according to claim 1, characterized in that: In step 1, the domain transformation is performed using a two-dimensional discrete Fourier transform, converting the space-frequency domain channel matrix into a sparse matrix in the angle-delay domain; the two-dimensional discrete Fourier transform satisfies the following formula: (1) In the formula, For angle-time delay domain sparse matrix, The channel matrix is ​​in the spatial frequency domain. and They represent dimensions as follows: and The orthogonal discrete Fourier transform matrix.

3. The method according to claim 2, characterized in that: In step 1, the system performs a truncation operation on the angle-time delay domain sparse matrix, retaining only the first part of the matrix. The truncated complex channel matrix physical map is decoupled into two independent channels, the real part and the imaginary part, and then normalized to the numerical range of [0,1].

4. The method according to claim 1, characterized in that: In step 2, PAM-Net uses a weight-shared encoder, and the self-supervised auxiliary branch and the trunk reconstruction branch share the encoder parameters. The self-supervised auxiliary branch is removed during online deployment.

5. The method according to claim 1, characterized in that: In step 2, the multi-scale physical perception jigsaw puzzle processing adaptively divides the input matrix into 2×2, 4×4 or 8×8 non-overlapping patches. After the patches are disordered, the original positions are restored by the multi-scale prediction head.

6. The method according to claim 1, characterized in that: In step 2, during the physically guided anisotropic dynamic penalty processing, the channel element amplitude satisfies the formula: (4) in, Coordinates are Channel amplitude at that location, and These are the real and imaginary parts of the matrix, respectively.

7. The method according to claim 6, characterized in that: The local physical complexity of the channel satisfies the formula: (5) in, and These are the penalty sensitivity coefficients in the time delay domain and the angle domain, respectively. Coordinates are Local physical complexity of the channel.

8. The method according to claim 7, characterized in that: This module uses mean normalization and extreme value clipping (Clip) mechanisms to generate weights, ultimately generating the first... Dynamic physics weighting coefficients for each tile : (7) in, This represents the global average complexity of all tiles at the current partitioning scale. and These are the preset lower and upper limits for punishment tolerance, respectively.

9. The method according to claim 1, characterized in that: In step 3, the joint optimization objective consists of the main task reconstruction loss and the self-supervised auxiliary task loss, and the joint loss function satisfies the formula: (8) (9) Among them, the reconstruction loss of the main task Standard mean square error is used; It is a balance coefficient that decays dynamically with the number of training rounds; For puzzle loss in self-supervised auxiliary tasks.

10. The method according to claim 1, characterized in that: The cross-entropy loss is calculated using a weighted average of physical dynamic weights, and the weighted cross-entropy loss satisfies the following formula: (10) In this formula, This represents the total number of tiles generated under the current multi-scale segmentation strategy; It is the Physically Guided Anisotropic Dynamic Penalty Module (PG-ADP) for the first... Dynamic physics weighting coefficients generated individually for each shuffled tile; The target label for the true original location of the tile; Then represents the first output of the multi-scale prediction head. The tile belongs to the first... The predicted probability of each location. By... As a penalty multiplier, the error penalty generated when the network mispredicts the location of tiles containing complex physical principal path boundaries forces the network to focus on the reconstruction of spatial structure.

Images