Multi-modal channel state information compression feedback method and system for very large scale antenna

Abstract

The application discloses a kind of multi-modal channel state information compression feedback methods and systems of super large scale antenna, it is related to deep learning and wireless communication technical field.The application includes: domain transformation is carried out to downlink channel state information and is embedded in block to obtain token sequence;Token sequence is injected into the near-field position coding containing spherical wave second-order curvature term;In self-attention, attention mask determined by visible area is applied;With group-in and group-out combined grouping hierarchical attention coding, potential representation is obtained;Compression ratio is dynamically determined by rate controller, and after quantization feedback, nested code word is truncated, and bit is allocated to each group;Base station reconstructs channel state information for precoding or beamforming.The application improves the reconstruction accuracy of near-field and spatial non-stationary scene, reduces the complexity of self-attention from square level to sub-square level, and supports multiple compression ratios dynamically adaptive with a single model, realizes high-precision, low-complexity and robust compression feedback.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a method and system for compressing and feeding back multimode channel state information for a very large-scale antenna. Background Technology

[0002] With the evolution of mobile communication technology, the antenna scale of base stations has expanded from tens to hundreds of elements to hundreds or even thousands of elements, forming ultra-large-scale antenna arrays. To achieve high-gain beamforming, spatial multiplexing, and interference suppression, base stations need to acquire accurate downlink channel state information. In frequency division duplex systems, due to the different uplink and downlink frequency bands and the lack of channel reciprocity, downlink channel state information is usually measured by user equipment and then fed back to the base station via the uplink. As the antenna dimension and the number of subcarriers increase, the dimension of the original channel state information expands dramatically, causing the feedback overhead and the computational burden on user equipment to become system bottlenecks. Therefore, efficient compression is necessary.

[0003] Among related technologies, a class of deep learning-based channel state information compression feedback methods is widely used. The basic approach involves deploying an encoder on the user equipment side to compress high-dimensional channel state information into low-dimensional codewords and quantize them into a bitstream for feedback. On the base station side, a decoder is deployed to reconstruct the channel state information from the bitstream. The encoder and decoder are jointly trained as an autoencoder. Convolutional autoencoder methods utilize convolution to extract local spatial correlations of the channel; self-attention-based methods utilize global self-attention to capture long-range correlations of the channel in the angular or time-delay domains.

[0004] However, the inventors discovered in their research that when the above method is directly applied to very large-scale antenna arrays, the following technical problems arise because the physical characteristics of the very large aperture array are not fully modeled.

[0005] First, due to the large aperture array, the Rayleigh distance is greatly extended, and a large number of scatterers and users fall into the near field region. The wavefront degenerates from the plane wave assumption to a spherical wave, and the phase across array elements exhibits a second-order curvature change related to distance. Existing methods use the far-field plane wave assumption, and their position encoding only describes the phase that is linearly related to the angle, which cannot characterize the near-field curvature, resulting in the loss of coupling information between angle and distance and a decrease in reconstruction accuracy.

[0006] Secondly, different subarrays on a super-large array observe different scattering clusters, i.e., there is a visible area phenomenon, and the channel exhibits spatial non-stationarity on the array axis. Existing methods perform indiscriminate global self-attention on the entire array, which will establish false correlations between unrelated visible areas, introducing redundancy and amplifying noise, which is particularly serious at low signal-to-noise ratios.

[0007] Third, standard self-attention has quadratic computational complexity for long sequences, and the channel state information of VMI arrays has a very high dimension, resulting in an excessive number of model parameters and computational requirements, making it difficult to deploy in a lightweight manner on user equipment with limited computing power and power consumption.

[0008] Fourth, existing methods typically train a model at a single compression ratio, and cannot dynamically adjust the feedback overhead based on signal-to-noise ratio, movement speed, sparsity of the visible area, and latency budget. Under extremely high compression ratios, the overall performance collapses and robustness is poor.

[0009] Fifth, existing methods only use instantaneous downlink channel state information as input, without utilizing various side information such as location, sensing, uplink angle, and historical channel state information, which limits feedback efficiency and mobility robustness.

[0010] Therefore, it is necessary to propose a new scheme that is specifically designed for the physical characteristics of ultra-large-scale antenna arrays and takes into account accuracy, complexity, adaptability and robustness. Summary of the Invention

[0011] The present invention aims to provide a method and system for compressing and feeding back multimode channel state information of ultra-large scale antennas, so as to overcome the shortcomings of the prior art. The technical problem to be solved by the present invention is achieved through the following technical solution.

[0012] According to a first aspect of the present invention, a method for compressing and feeding back multimode channel state information of a very large-scale antenna is provided, applied to a very large-scale antenna system including user equipment and a base station, wherein the system comprises an encoder deployed on the user equipment side and a decoder deployed on the base station side, forming an end-to-end jointly trained autoencoder, the method comprising: The user equipment performs domain transformation on the downlink channel state information and embeds the result of the domain transformation into blocks to obtain a token sequence; Inject near-field position codes into the token sequence, the near-field position codes including second-order curvature terms characterizing spherical wave propagation; An attention mask determined by the visible region is applied during self-attention to restrict feature fusion across the visible region in the token sequence; Grouped hierarchical attention, which combines intra-group attention and inter-group attention, is used to encode the token sequence injected with the near-field position encoding to obtain a latent representation; The compression ratio is dynamically determined by the rate controller based on the side information of the channel and the service. The nested codewords corresponding to the potential representation are truncated according to the compression ratio. The truncated codewords are allocated bits group by group and quantized into a bit stream, which is then fed back to the base station via the uplink. The base station dequantizes the bit stream and reconstructs downlink channel state information in a structure symmetrical to the encoder. The reconstructed downlink channel state information is used for downlink precoding or beamforming.

[0013] Preferably, injecting near-field position encoding into the token sequence includes any of the following methods: Method 1: Based on the two-dimensional coordinates of each element in the array, construct a distance-sensitive second-order phase feature, and project the second-order phase feature into the near-field position code through a learnable linear mapping, wherein the second-order phase feature is determined based on the reference angle-distance pair obtained by pilot coarse estimation or codebook enumeration. Method 2 involves introducing learnable quadratic coefficients into the sinusoidal position encoding to learn the curvature of the spherical wave end-to-end from the data source.

[0014] Preferably, the attention mask determined by the visible area is constructed in any of the following ways: In the data-driven approach, the visible region estimation subnetwork takes the received power of each array element as input, outputs the soft membership degree between the array element and the visible region, and generates the attention mask based on the similarity between the soft membership degrees. The geometric prior method uses the physical distance between array elements as a metric to generate a distance-attenuation attention mask that suppresses array elements that are far apart. The signaling-driven method generates the attention mask based on the network-side configuration or the visible area descriptor reported by the user equipment. The attention mask contains a learnable temperature parameter, and when the temperature parameter approaches zero, the attention mask degenerates into global attention.

[0015] Preferably, the encoding of hierarchical attention by combining intra-group attention and inter-group attention includes: The token sequence is divided into multiple spatial groups according to the physical structure of the array, and the boundaries of the spatial groups are aligned with the sub-arrays or visible areas. Perform intra-group attention, which is restricted by the attention mask, within each spatial group; For each spatial group, aggregate to obtain a group representative token, and perform inter-group attention between group representative tokens to obtain the inter-group global context; The inter-group global context is back-broadcast to the tokens within each spatial group to obtain the potential representation.

[0016] Preferably, the nested codewords are obtained and used in the following manner: During the training phase, the latent dimension at the tail of the codeword corresponding to the latent representation is randomly masked, so that the latent dimension at the front can independently reconstruct the downlink channel state information. During the inference phase, the length of the codeword is truncated according to the compression ratio; The method of allocating bits to groups of truncated codewords includes: allocating more bits to groups with active visible regions, allocating fewer bits or zero bits to groups without visible scattering clusters, and the sum of the bits allocated to each group does not exceed the overhead budget.

[0017] Preferably, it further includes: In the encoder, one or more side information from the user equipment's position or attitude, sensed echo or environmental point cloud, uplink channel state information or angle of arrival, and historical channel state information are encoded into additional tokens via cross-modal attention and fused with the token sequence. The decoder uses the same-origin side information available at the base station as a priori for reconstruction. In this process, temporal attention is introduced into the historical channel state information to achieve predictive feedback or differential feedback in mobile scenarios.

[0018] According to a second aspect of the present invention, a multimode channel state information compression feedback system for a very large MIMO antenna is provided, comprising an encoding device deployed on the user equipment side and a decoding device deployed on the base station side, wherein the encoding device and the decoding device are obtained through end-to-end joint training, including: The domain transformation and embedding module is configured to perform domain transformation on the downlink channel state information and embed the results of the domain transformation into blocks to obtain a token sequence. The near-field position coding module is configured to inject near-field position coding into the token sequence, which includes a second-order curvature term characterizing the propagation of spherical waves. The visible region masking module is configured to apply an attention mask determined by the visible region during self-attention to restrict feature fusion across the visible region; The group attention encoding module is configured to encode the token sequence using a group hierarchical attention that combines intra-group attention and inter-group attention to obtain a latent representation; The rate control and quantization module is configured to dynamically determine the compression ratio based on the side information of the channel and the service, truncate the nested codewords corresponding to the potential representation according to the compression ratio, allocate bits in groups, quantize them into a bit stream, and feed it back through the uplink. The reconstruction module is configured to dequantize the bitstream and reconstruct downlink channel state information in a structure symmetrical to the coding device for downlink precoding or beamforming.

[0019] Preferably, the quantization performed by the rate control and quantization module is scalar quantization with a pass-through estimator, vector quantization with a pass-through estimator, or quantization based on a learnable codebook, and entropy coding is superimposed after quantization; the encoding device and the decoding device are jointly trained end-to-end based on a loss function, which includes a normalized mean square error term, a cosine similarity term, a bit rate cost term, and a visible region regularization term.

[0020] According to a third aspect of the present invention, an electronic device is provided, the device comprising: One or more processors; and A storage device is provided for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the above-described multimode channel state information compression feedback method for ultra-large scale antennas.

[0021] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described method for compressing and feeding back multimode channel state information for a very large MIMO antenna.

[0022] The embodiments of the present invention have the following advantages: By injecting near-field position coding containing second-order curvature terms of spherical waves into the token sequence, the model is able to perceive the joint geometry of angle and distance, thereby improving the channel state information reconstruction accuracy in near-field and spatially non-stationary scenarios with the same feedback overhead. By applying an attention mask determined by the visible region during self-attention, spurious correlations and noise amplification between unrelated visible regions are suppressed, further improving the reconstruction quality under low signal-to-noise ratio. By combining intra-group attention with inter-group attention, the computational complexity of self-attention is reduced from the quadratic level of the sequence length to the sub-quadratic level, enabling the encoder to be deployed in a lightweight manner on user equipment with limited computing power and power consumption. By combining rate controller with nested codewords and bit-by-bit allocation, multiple compression ratios can be supported with a single model and dynamically adapted according to channel and service conditions, thus mitigating the performance collapse under extremely high compression ratio and low signal-to-noise ratio. By fusing multiple side information through cross-modal attention, the amount of feedback is further reduced and the predictive feedback capability in mobile scenarios is enhanced without increasing uplink overhead. Attached Figure Description

[0023] Figure 1 A schematic diagram of the overall system framework for channel state information compression feedback according to an embodiment of the present invention is shown; Figure 2A flowchart of a multi-mode channel state information compression feedback method for a very large-scale antenna according to an embodiment of the present invention is shown; Figure 3 A schematic diagram of the near-field sensing codec backbone structure according to an embodiment of the present invention is shown; Figure 4 A schematic diagram illustrating the near-field position encoding principle according to an embodiment of the present invention is shown; Figure 5 A schematic diagram of a visible area mask according to an embodiment of the present invention is shown; Figure 6 A schematic diagram of spatial grouping hierarchical attention according to an embodiment of the present invention is shown; Figure 7 A schematic diagram of the rate control and dynamic bit allocation process according to an embodiment of the present invention is shown; Figure 8 A schematic diagram of a multimodal side information fusion branch according to an embodiment of the present invention is shown; Figure 9 A timing diagram illustrating the end-to-end workflow and signaling interaction according to an embodiment of the present invention is shown; Figure 10 A performance comparison diagram according to an embodiment of the present invention is shown. Detailed Implementation

[0024] 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.

[0025] It should be noted that the above detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0026] In the following description, downlink channel state information refers to data obtained by the user equipment through measurements of the downlink reference signal, used to characterize the complex gain distribution of the downlink channel in the antenna and subcarrier dimensions; token sequence refers to a sequence of several feature vectors obtained by performing domain transformation and block embedding on the downlink channel state information, wherein the feature vectors serve as the basic units for subsequent self-attention processing; visible region refers to the region in a spatially non-stationary channel where consecutive array elements of the same scattering cluster can be observed. Those skilled in the art should understand that the above terms have the same meaning in different embodiments of the present invention.

[0027] Example 1: Channel state information compression feedback for non-line-of-sight near-field scenarios in urban macrocells under frequency division duplex (FDM) system.

[0028] In this embodiment, the ultra-large scale antenna system includes user equipment and a base station. The base station is configured with a two-dimensional uniform array as the transceiver antenna. For example, the uniform array contains 256 elements arranged in 16 rows by 16 columns, with a carrier frequency of 3.5 GHz, 64 subcarriers, and four combinations of dual polarization, resulting in a single complete downlink channel state information having a real dimension of 131,072 dimensions. The system consists of an encoder deployed on the user equipment side and a decoder deployed on the base station side, forming an end-to-end jointly trained autoencoder. See [link to documentation]. Figure 1 .like Figure 1 As shown, the user equipment side includes sequentially performing downlink channel estimation, domain transformation and block embedding, injection near-field position coding, visible area masking and group attention coding, rate control and dynamic bit allocation, and quantization and entropy coding. The base station side includes sequentially performing inverse quantization and entropy decoding, symmetric group attention decoding, inverse domain transformation and reconstruction, and precoding and beamforming. The user equipment side feeds back the bit stream and control domain to the base station side via the uplink.

[0029] like Figure 2 As shown, the overall flow of the channel state information compression feedback method in this embodiment includes steps S101 to S106. The following is a detailed explanation... Figure 3 The codec backbone structure shown is explained step by step.

[0030] In step S101, the user equipment performs domain transformation on the downlink channel state information and embeds the result of the domain transformation into blocks to obtain a token sequence.

[0031] Specifically, the user equipment receives a downlink reference signal and estimates downlink channel state information. To enhance compressibility, a domain transformation is performed on the downlink channel state information. In some embodiments, the domain transformation is a two-dimensional discrete Fourier transform along the antenna and subcarrier dimensions, transforming the channel state information to a joint angle and time delay domain, concentrating energy in a few dominant components. In other embodiments, for near-field scenarios, the domain transformation employs a sparsification transformation based on a near-field polar coordinate domain dictionary to adapt to the joint angle and distance sparse structure under spherical wave propagation. For dual-polarization cases, different polarization components are stacked along the polarization dimension and processed uniformly. Subsequently, the result of the domain transformation is divided into several data blocks according to a preset block size, and each data block is linearly embedded and mapped to a feature vector of a fixed dimension, which constitutes a token sequence.

[0032] To facilitate understanding of the near-field characteristics modeling in this embodiment, the signal and channel models are given below. Let the array aperture be D and the wavelength be λ, then the Rayleigh distance is expressed as:

[0033] Where D is the array aperture and λ is the wavelength. When the distance between the scatterer or user and the array is less than the Rayleigh distance, the wavefront is assumed to degenerate from a plane wave to a spherical wave. For the nth element, the elements of its near-field array response are determined by the propagation distance of the spherical wave, rather than the linear phase of the far-field plane wave, and can be expressed as:

[0034] Where a(θ,r) is the near-field array response vector, θ is the angle, r is the distance, and u(θ) is the unit direction vector determined by the angle. Let be the two-dimensional coordinates of the nth element, ‖·‖ denote the L2 norm, and j be the imaginary unit. As the distance approaches infinity, the above equation degenerates into a linear phase in the far field; in the near field, it retains the distance-related second-order curvature term. This embodiment utilizes this second-order curvature term and the visible region structure for specialized modeling.

[0035] Step S102: Inject near-field position coding into the token sequence, the near-field position coding including a second-order curvature term characterizing spherical wave propagation.

[0036] Since the backbone network requires position encoding to inject spatial structure, and related technologies employ sinusoidal position encoding based on the far-field assumption or one-dimensional learnable encoding, they can only express the phase linearly related to the element index and cannot characterize the near-field curvature. Therefore, this embodiment designs a near-field position encoding that explicitly injects the second-order curvature of the spherical wave into the position encoding, enabling attention to perceive the joint geometry of angle and distance. In one implementation (Method 1, physically driven), based on the two-dimensional coordinates of each element in the array, a distance-sensitive second-order phase feature is constructed and projected onto the near-field position encoding via a learnable linear mapping, as shown in the following equation:

[0037] The near-field position code corresponding to the nth array element is denoted as . , Let ψ be a learnable linear mapping, and ψ be a sinusoidal expansion or a multilayer perceptron. and Let n be the two-dimensional coordinates of the nth array element. The square of and The sum of the squares corresponds to the curvature of the spherical wave on the array. The near-field phase is determined by a reference angle and a reference distance, which are obtained by coarse estimation of the pilot or codebook enumeration. In another implementation (Method 2, learnable curvature), learnable quadratic coefficients are introduced based on standard sinusoidal position coding, and the curvature is learned end-to-end from the data source, as shown in the following equation:

[0038] In the above formula, the positional encoding components in even-numbered dimensions are given, where n is the element index, i is the dimension index, d is the dimension of the positional encoding, a is the coefficient of the first-order term, and b is the coefficient of the learnable quadratic term. Odd-numbered dimensions are determined symmetrically using a cosine function. The near-field positional encoding can be applied to the encoder input or injected into the decoder in a shared or symmetrical form to ensure consistent representation of near-field geometry between the encoding and decoding sides. In some embodiments, the dimension of the near-field positional encoding is between 128 and 512. See also... Figure 4 In the far field case, the wavefront is a plane wave, and the phase across array elements changes linearly with the array element number; in the near field case, the wavefront is a spherical wave, and the phase across array elements changes with the array element number in a second-order curvature. The near field position encoding in this embodiment models this second-order curvature term.

[0039] Step S103: Apply an attention mask determined by the visible region to the self-attention to restrict feature fusion across the visible region in the token sequence.

[0040] To address spatial non-stationarity, this embodiment introduces an attention mask determined by the visible region into self-attention, ensuring that each token is strongly correlated and fused only with tokens located in the same or overlapping visible regions, thereby suppressing spurious correlations and noise amplification across visible regions. Self-attention with an applied attention mask can be represented as:

[0041] Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively. Let M be the dimension of the key vector, M be the attention mask, and softmax be the normalized exponential function. The attention mask is constructed in three ways. One is a data-driven approach, where the visible region estimation subnetwork takes the received power or energy envelope of each element as input, outputs the soft membership degree between the element and the visible region, and generates the attention mask based on the similarity between these soft membership degrees, as shown in the following equation:

[0042] in, Let be the element in the i-th row and j-th column of the attention mask, β be the learnable temperature parameter, and sim denote the similarity function. and These represent the visible region membership degrees of the i-th and j-th array elements, respectively. The second approach is a geometric prior method, which uses the physical distance between array elements as a metric to apply soft suppression to array elements that are farther apart, as shown in the following equation:

[0043] in, and These are the coordinates of the i-th and j-th array elements, respectively, and σ is the scale parameter. Thirdly, there is a signaling-driven mode where the block information of the visible area is configured by the network side or determined by the compact visible area descriptor reported by the user equipment, facilitating standardized alignment. The temperature parameter β is a learnable parameter; when the temperature parameter approaches zero, the attention mask degenerates into global attention, thereby ensuring backward compatibility with stable channels. See also... Figure 5 Different subarrays observe different scattering clusters. Correspondingly, the attention mask exhibits a block structure, in which the elements that are allowed to be noticed take the value of zero, and the elements that are masked take the value of negative infinity.

[0044] Step S104: Encode the token sequence injected with the near-field position encoding using grouped hierarchical attention that combines intra-group attention and inter-group attention to obtain a latent representation.

[0045] To reduce the computational complexity of self-attention, this embodiment divides the token sequence into multiple spatial groups according to the physical structure of the array. Each spatial group corresponds to a subarray or candidate visible region, and performs two-level hierarchical attention. See [link to documentation]. Figure 6 Specifically, this includes: intra-group attention, performing local full attention constrained by the attention mask within each spatial group to capture non-stationary substructures and near-field local curvature; group representative aggregation, obtaining group representative tokens by pooling or learningable aggregation for each spatial group; inter-group attention, performing global attention on each group representative token to capture long-range correlations across subarrays, obtaining inter-group global context; and back-injection fusion, broadcasting the inter-group global context back-injection to tokens within each spatial group to complete global information backfeeding, thereby obtaining the latent representation. Through the above two-level hierarchical attention, the computational complexity of self-attention is reduced from the quadratic order of the sequence length to the following order of magnitude:

[0046] Where N is the length of the token sequence and G is the number of spatial groups. When the number of spatial groups is approximately equal to the square root of the sequence length, the computational complexity is close to the order of the sequence length multiplied by its square root, which is significantly reduced compared to quadratic complexity, thus enabling the encoder to be deployed lightweight on user equipment with limited computing power and power consumption. In this embodiment, the number of spatial groups is 16, and the group boundaries are aligned with the 16-column subarray; when the group boundaries are aligned with the visible area structure, the accuracy can be further improved. The group-level attention works in conjunction with the visible area mask and near-field position encoding: the near-field position encoding provides the geometric position, the visible area mask determines the fusion range, and the group-level attention determines the fusion level and computing power allocation.

[0047] In step S105, the rate controller dynamically determines the compression ratio based on the side information of the channel and the service, truncates the nested codewords corresponding to the potential representation according to the compression ratio, and quantizes the truncated codewords into a bit stream by allocating bits in groups, which is then fed back to the base station via the uplink.

[0048] See Figure 7 The rate controller takes side information as input and dynamically determines the overall compression ratio and bit-by-bit allocation based on this side information. This side information includes signal-to-noise ratio, movement speed or coherence time, visible area sparsity, and latency or overhead budget. The output of the rate controller can be expressed as:

[0049] Where CR is the compression ratio. to The number of bits allocated to each spatial group, where RateCtrl represents the rate controller, s represents the side information, and B represents the overhead budget, ensures that the sum of the bits allocated to each spatial group does not exceed the overhead budget. In the group-by-group bit allocation, groups with active visible regions are allocated more bits, while groups without visible scattering clusters are allocated fewer bits or zero bits, thus achieving on-demand feedback. To support multiple compression ratios with a single model, the codewords corresponding to the latent representations are in a nested or progressive form: during the training phase, the latent dimensions at the tail of the codewords are randomly masked, allowing the latent dimensions at the front to independently reconstruct the downlink channel state information; during the inference phase, the codeword length is truncated according to the compression ratio, eliminating the need to train a separate model for each compression ratio.

[0050] Subsequently, the truncated codewords, after being allocated bits group by group, are quantized. In some embodiments, the quantization is scalar quantization or vector quantization with a pass-through estimator; in other embodiments, quantization based on a learnable codebook is used. Entropy coding is superimposed after quantization for further compression. The quantizer is jointly trained with the encoder and decoder, and quantization noise is incorporated into the loss function for optimization. The user equipment feeds back the bitstream via the uplink and carries a compact control domain in the channel state information report. The control domain includes the selected compression ratio index, visible area or grouping indication, quantization level, and model version. In this embodiment, the nested codewords support multiple compression ratios from 1 / 8 to 1 / 256.

[0051] In step S106, the base station dequantizes the bit stream and reconstructs downlink channel state information in a structure symmetrical to the encoder. The reconstructed downlink channel state information is used for downlink precoding or beamforming.

[0052] Specifically, the base station performs entropy decoding and dequantization on the received bitstream, and pads the truncated codewords with zeros to restore the complete latent dimension. Subsequently, the decoder multiplexes near-field position coding and visible region masks symmetrical to the encoding side, obtains features through a group-level attention decoding layer symmetrical to the encoding side, and then performs inverse domain transformation and deblocking to obtain the reconstructed downlink channel state information. The reconstructed downlink channel state information is used for downlink precoding or beamforming. In some embodiments, the base station also updates the rate strategy based on link-adaptive indirect statistical results and performs online fine-tuning and hot version updates of the model. The complete end-to-end workflow and signaling interaction timing are as follows: Figure 9 As shown.

[0053] In this embodiment, the encoder, decoder, quantizer, optional view area estimation subnetwork, and rate controller are jointly trained end-to-end. The loss function used for training consists of three parts: reconstruction fidelity, rate cost, and structure regularization, and can be expressed as:

[0054] Where L is the total loss, H is the original downlink channel state information, the estimate of H is the reconstructed downlink channel state information, E represents the expectation, the subscript F represents the Frobenius norm, the first term is the normalized mean square error to characterize the reconstruction accuracy, the second term is the cosine similarity and weighted by the coefficient γ to directly correlate with beamforming gain, the third term is the bit rate cost and weighted by the coefficient μ to achieve a rate-distortion tradeoff, and the fourth term is the regularization term for the visible area estimation or mask and weighted by the coefficient η. In this embodiment, γ is taken as 1, and μ and η are adapted to different values. To support a single model supporting multiple compression ratios, potential codewords are randomly truncated during training; to improve robustness under low signal-to-noise ratio and high-speed movement, the training samples cover multiple signal-to-noise ratios, multiple movement speeds, and multiple visible area sparsities, and noise injection and curriculum learning are introduced. For example, the training samples cover a signal-to-noise ratio range from 0dB to 30dB and movement speeds of 3 km / h, 30 km / h, and 120 km / h. The online phase supports edge-side fine-tuning based on a small number of samples and hot updates of the model version. Figure 10 The invention demonstrates its comprehensive advantages in maintaining accuracy, controlling complexity, and dynamically adapting as the compression ratio increases. It exhibits the smallest performance drop slope, indicating that a single model can robustly support multiple compression ratios.

[0055] In summary, this invention maintains significant gain even under extreme conditions such as extremely high compression ratio and low signal-to-noise ratio, and combines accuracy, robustness, and deployment feasibility.

[0056] Based on the same inventive concept as the above method, this embodiment also provides a channel state information compression feedback system for very large-scale antenna systems, including an encoding device deployed on the user equipment side and a decoding device deployed on the base station side, wherein the encoding device and the decoding device are jointly trained end-to-end. The system includes: The domain transformation and embedding module is configured to perform step S101, that is, to perform domain transformation on the downlink channel state information and embed it in blocks to obtain a token sequence. The near-field position encoding module is configured to execute step S102, which is to inject near-field position encoding containing a second-order curvature term characterizing spherical wave propagation into the token sequence. The visible region masking module is configured to perform step S103, that is, to apply an attention mask determined by the visible region in the self-attention; The group attention encoding module is configured to execute step S104, that is, to encode the token sequence with a group hierarchical attention that combines intra-group attention and inter-group attention to obtain a potential representation; The rate control and quantization module is configured to execute step S105, which is to dynamically determine the compression ratio based on the side information of the channel and the service, truncate the nested codewords according to the compression ratio, allocate bits group by group and quantize them into a bit stream and feed it back through the uplink. The reconstruction module is configured to perform step S106, which is to dequantize the bit stream and reconstruct the downlink channel state information in a structure symmetrical to the encoding device.

[0057] Each of the above modules corresponds to one of the steps of the method. For relevant explanations, please refer to the method implementation examples, which will not be repeated here.

[0058] Example 2: Position and perception-assisted multimodal prediction feedback.

[0059] This embodiment, based on Embodiment 1, further introduces a cross-modal attention branch into the encoder to fuse multimodal side information. See [link to Embodiment 1]. Figure 8 Specifically, one or more of the following side information—the user equipment's location or attitude, perceived echo or environmental point cloud, uplink channel state information or angle of arrival, and historical channel state information—are encoded into additional tokens by the corresponding modal encoders, and then fused with the token sequence through cross-modal attention.

[0060] In this system, the user equipment's position or attitude can be obtained from a global navigation satellite system or inertial navigation, providing priors for near-field angles and distances, thereby enhancing the initialization of near-field position coding and visible area masks. Sensing echoes or environmental point clouds provide priors for scattering geometry and visible area structure. Uplink channel state information or angle of arrival utilizes the partial reciprocity of frequency division duplexing to provide angle-domain priors. Historical channel state information, after incorporating temporal attention, enables predictive or differential feedback in mobile scenarios, reducing the feedback frequency within the coherence time. The decoder symmetrically utilizes source-side information available from the base station, such as position and uplink angle, as priors for reconstruction, thereby further improving accuracy without increasing uplink overhead. In medium-to-high-speed mobile scenarios, this embodiment can further reduce feedback frequency and overhead.

[0061] Example 3: Joint channel state information compression feedback in non-cellular or distributed cooperative scenarios.

[0062] This embodiment is applied to a non-cellular or distributed VMI (Very Large Scale Antenna) system with multi-access point cooperation. In this scenario, the subarrays of each access point naturally constitute spatial groups and visible areas. Therefore, the inter-group attention of the group-level attention undertakes information fusion across access points, and the visible area mask suppresses spurious associations between non-overlapping access points, thereby achieving joint channel state information compression feedback for the distributed VMI system. The remaining processing in this embodiment can be found in Embodiment 1, and will not be repeated here.

[0063] As alternative implementation methods, in terms of system type, besides frequency division duplex, this invention can also be used for differential or incremental feedback under time division duplex reciprocity assistance; in terms of dynamic compression, besides nested codewords, multi-branch or hypernetworks can be used to generate parameters according to the compression ratio, or reinforcement learning can be used to train the rate controller; in terms of application objectives, the reconstruction objective can be changed from complete channel state information to a precoding matrix or feature vector, and correspondingly, the loss function can be changed to a proxy indicator of link throughput or spectral efficiency. All of the above alternative implementation methods fall within the protection scope of this invention.

[0064] An embodiment of the present invention provides a computing device, which may be the aforementioned user equipment or base station. The computing device includes a computing unit, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory or a computer program loaded from a storage unit into a random access memory. The random access memory may also store various programs and data required for the operation of the computing device. The computing unit, the read-only memory, and the random access memory are interconnected via a bus. Input / output interfaces are also connected to the bus. Multiple components in the computing device are connected to the input / output interfaces, including: input units, such as a keyboard, mouse, etc.; output units, such as various types of displays, speakers, etc.; storage units, such as disks, optical disks, etc.; and communication units, such as network interface cards, modems, wireless transceivers, etc. The communication unit allows the computing device to exchange information or data with other devices through computer networks such as the Internet and / or various telecommunications networks. The computing unit executes the various methods and processes described above, such as the methods described in steps S101 to S106.

[0065] It should be understood that the methods described in the embodiments of the present invention can be implemented by hardware, software, or a combination of hardware and software. The above steps can be performed by a computing unit executing instructions stored on a machine-readable medium. The computing unit may include, but is not limited to, a central processing unit, a graphics processing unit, various dedicated artificial intelligence computing chips, various computing units running machine learning model algorithms, a digital signal processor, and any suitable processor, controller, or microcontroller. Considering that the encoder is deployed on the user equipment side with limited computing power and power consumption, the grouped hierarchical attention of the present invention reduces the computational complexity of self-attention from O(N²) to a sub-square order of O(N√N), thereby enabling the encoder to run in real time on the computing unit with lower computational overhead.

[0066] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the methods described in any of the above embodiments. The computer-readable storage medium may be a tangible medium that may contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. The computer-readable storage medium may include, but is not limited to, electrical storage media, magnetic storage media, optical storage media, electromagnetic storage media, semiconductor storage media, or any suitable combination thereof. More specific examples of machine-readable storage media may include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, optical fibers, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0067] 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.

[0068] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein.

[0069] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0070] For ease of description, spatial relative terms such as "above," "on top of," "on the upper surface of," "above," etc., are used herein to describe the spatial positional relationship of a device or feature as shown in the figures to other devices or features. It should be understood that spatial relative terms are intended to encompass different orientations in use or operation beyond the orientation of the device as described in the figures. For example, if the device in the figures were inverted, a device described as "above" or "on top of" other devices or structures would subsequently be positioned as "below" or "under" other devices or structures. Thus, the exemplary term "above" can include both "above" and "below." The device may also be positioned in other different ways, such as rotated 90 degrees or in other orientations, and the spatial relative descriptions used herein will be interpreted accordingly.

[0071] In the detailed description above, reference has been made to the accompanying drawings, which form part of this document. In the drawings, similar symbols typically identify similar parts unless the context otherwise indicates otherwise. The illustrated embodiments described in the detailed specification, drawings, and claims are not intended to be limiting. Other embodiments may be used and other changes may be made without departing from the spirit or scope of the subject matter presented herein.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-mode channel state information compression feedback method for ultra-large scale antennas, applied to an ultra-large scale antenna system including user equipment and base stations, wherein the system consists of an encoder deployed on the user equipment side and a decoder deployed on the base station side, forming an end-to-end jointly trained autoencoder, characterized in that... The method includes: The user equipment performs domain transformation on the downlink channel state information and embeds the result of the domain transformation into blocks to obtain a token sequence; Inject near-field position codes into the token sequence, the near-field position codes including second-order curvature terms characterizing spherical wave propagation; The near-field position encoding of the token sequence is encoded by a grouped hierarchical attention that combines intra-group attention and inter-group attention to obtain a latent representation; wherein, an attention mask determined by the visible region is applied in the intra-group attention to restrict feature fusion across the visible region in the token sequence. The compression ratio is dynamically determined by the rate controller based on the side information of the channel and the service. The nested codewords corresponding to the potential representation are truncated according to the compression ratio. The truncated codewords are allocated bits group by group and quantized into a bit stream, which is then fed back to the base station via the uplink. The base station dequantizes the bit stream and reconstructs downlink channel state information in a structure symmetrical to the encoder. The reconstructed downlink channel state information is used for downlink precoding or beamforming.

2. The multi-mode channel state information compression and feedback method for ultra-large scale antennas according to claim 1, characterized in that, The injection of near-field position encoding into the token sequence includes any of the following methods: Method 1: Based on the two-dimensional coordinates of each element in the array, construct a distance-sensitive second-order phase feature, and project the second-order phase feature into the near-field position code through a learnable linear mapping, wherein the second-order phase feature is determined based on the reference angle-distance pair obtained by pilot coarse estimation or codebook enumeration. Method 2 involves introducing learnable quadratic coefficients into the sinusoidal position encoding to learn the curvature of the spherical wave end-to-end from the data source.

3. The multi-mode channel state information compression and feedback method for ultra-large scale antennas according to claim 1, characterized in that, The attention mask determined by the visible region is constructed in any of the following ways: In the data-driven approach, the visible region estimation subnetwork takes the received power of each array element as input, outputs the soft membership degree between the array element and the visible region, and generates the attention mask based on the similarity between the soft membership degrees. The geometric prior method uses the physical distance between array elements as a metric to generate a distance-attenuation attention mask that suppresses array elements that are far apart. The signaling-driven method generates the attention mask based on the network-side configuration or the visible area descriptor reported by the user equipment. The attention mask constructed using the data-driven approach or the geometric prior approach includes a learnable temperature parameter. When the temperature parameter approaches zero, the attention mask degenerates into global attention.

4. The multi-mode channel state information compression and feedback method for ultra-large scale antennas according to claim 1, characterized in that, The encoding of hierarchical attention by combining within-group attention and between-group attention includes: The token sequence is divided into multiple spatial groups according to the physical structure of the array, and the boundaries of the spatial groups are aligned with the sub-arrays or visible areas. Perform intra-group attention, which is restricted by the attention mask, within each spatial group; For each spatial group, aggregate to obtain a group representative token, and perform inter-group attention between group representative tokens to obtain the inter-group global context; The inter-group global context is back-broadcast to the tokens within each spatial group to obtain the potential representation.

5. The multi-mode channel state information compression feedback method for ultra-large scale antennas according to claim 1, characterized in that, The nested codewords are obtained and used in the following manner: During the training phase, the latent dimension at the tail of the codeword corresponding to the latent representation is randomly masked, so that the latent dimension at the front can independently reconstruct the downlink channel state information. During the inference phase, the length of the codeword is truncated according to the compression ratio; The method of allocating bits to groups of truncated codewords includes: allocating more bits to groups with active visible regions, and allocating fewer bits or zero bits to groups without active visible regions, and the sum of the bits allocated to each group does not exceed the overhead budget.

6. The multi-mode channel state information compression feedback method for ultra-large scale antennas according to claim 1, characterized in that, Also includes: In the encoder, one or more side information from the user equipment's position or attitude, sensed echo or environmental point cloud, uplink channel state information or angle of arrival, and historical channel state information are encoded into additional tokens via cross-modal attention and fused with the token sequence. The decoder uses the same-origin side information available at the base station as a priori for reconstruction. In this process, temporal attention is introduced into the historical channel state information to achieve predictive feedback or differential feedback in mobile scenarios.

7. A multi-mode channel state information compression feedback system for a very large-scale antenna, comprising an encoding device deployed on the user equipment side and a decoding device deployed on the base station side, wherein the encoding device and the decoding device are jointly trained end-to-end, characterized in that, include: The domain transformation and embedding module is configured to perform domain transformation on the downlink channel state information and embed the results of the domain transformation into blocks to obtain a token sequence. The near-field position coding module is configured to inject near-field position coding, which includes a second-order curvature term characterizing the propagation of spherical waves, into the token sequence; The visible region masking module is configured to apply an attention mask determined by the visible region during self-attention to restrict feature fusion across the visible region; The group attention encoding module is configured to encode the token sequence using a group hierarchical attention that combines intra-group attention and inter-group attention to obtain a latent representation; The rate control and quantization module is configured to dynamically determine the compression ratio based on the side information of the channel and the service, truncate the nested codewords corresponding to the potential representation according to the compression ratio, allocate bits in groups, quantize them into a bit stream, and feed it back through the uplink. The reconstruction module is configured to dequantize the bitstream and reconstruct downlink channel state information in a structure symmetrical to the coding device for downlink precoding or beamforming.

8. The multi-mode channel state information compression feedback system for ultra-large-scale antennas according to claim 7, characterized in that, The quantization performed by the rate control and quantization module is scalar quantization with a pass-through estimator, vector quantization with a pass-through estimator, or quantization based on a learnable codebook, and entropy coding is superimposed after quantization; The encoding device and the decoding device are jointly trained end-to-end based on a loss function, which includes a normalized mean square error term, a cosine similarity term, a bit rate cost term, and a visible region regularization term.

9. An electronic device, characterized in that, The device includes: One or more processors; and A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the multimode channel state information compression feedback method for ultra-large MIMO antennas according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the multi-mode channel state information compression feedback method for ultra-large scale antennas according to any one of claims 1 to 6.

Images