Search

CN-122001717-A - Multi-domain channel extrapolation method for ultra-large-scale MIMO system

CN122001717ACN 122001717 ACN122001717 ACN 122001717ACN-122001717-A

Abstract

The invention discloses a multi-domain channel extrapolation method for a super-large-scale MIMO system, which is based on a conditional denoising diffusion probability model construction generation framework and realizes the accurate completion of multi-domain channels through explicit modeling condition priori information. In order to further improve the generation quality and the sampling efficiency, a diffusion reasoning mechanism for generating an countermeasure network by fusing Wasserstein conditions is introduced, and an countermeasure supervision signal is introduced in a combined manner in a training and reasoning stage, so that the fidelity, diversity and convergence speed of a generated result are obviously enhanced. Meanwhile, the invention adopts a channel modeling mode based on blocks, embeds position and time evolution information into block characterization vectors, and designs a mask multi-head attention backbone network to effectively capture local and global structural features in a high-dimensional near-field channel. The invention combines the countermeasure generation mechanism and the attention modeling, and realizes high-precision channel completion and cross-domain generalization of an antenna domain, a frequency domain and a space domain with lower calculation complexity and reasoning time delay.

Inventors

  • HAN YU
  • LI MENGYUAN
  • JIN SHI

Assignees

  • 东南大学

Dates

Publication Date
20260508
Application Date
20260130

Claims (8)

  1. 1. A multi-domain channel extrapolation method for a super-large-scale MIMO system is characterized by comprising the following steps: step 1, establishing a near-field channel response model aiming at a super-large-scale MIMO system, and deducing correlation characterization of channels in three dimensions of an antenna domain, a frequency domain and a space domain; step 2, designing a channel modeling extrapolation strategy in the corresponding dimension based on the correlation characterization of the channel in each dimension and partial channel state information obtained by actual observation; Step 3, constructing a multi-domain channel extrapolation network based on a generated model, taking part of channel state information and noisy channel state information which are actually observed in each dimension as input and equivalent noise predicted in the corresponding dimension as output, wherein the multi-domain channel extrapolation network based on the generated model comprises a generator based on a conditional denoising diffusion implicit model and a WCGAN discriminator, and the generator based on the conditional denoising diffusion implicit model comprises an input block embedding module, a conditional and mask embedding module, a position coding module, a time modulation module and three mask multi-head attention modules which are sequentially connected, wherein the WCGAN discriminator is used for scoring the final predicted complete channel state information; step 4, training a multi-domain channel extrapolation network based on a generative model by adopting a joint loss function, and performing auxiliary supervised learning by using a WCGAN discriminator in the training process; and 5, designing a denoising diffusion hidden mechanism guided by a WCGAN discriminator, combining a step-less non-Markov denoising process, and reversely deducing complete channel state information based on predicted equivalent noise.
  2. 2. The method for multi-domain channel extrapolation for a super-large-scale MIMO system according to claim 1, wherein the specific procedure of step 1 is as follows: configuration for base station end Ultra-large-scale MIMO system with single antenna configured at user end by root uniform area array antenna, for being positioned at position At a frequency of Down to the antenna by the user Channel response therebetween Expressed as: , Wherein, the , Indicating the number of propagation paths and, Is the first The complex gain of the strip propagation path, In units of imaginary numbers, Indicating antenna To the user location or to the diffuser equivalent location Is used for the propagation distance of the beam, Is the speed of light; 1) For the same frequency Same user location Lower two antennas And The correlation of the channel response in the antenna domain is expressed as: , Wherein, the And Respectively representing antennas And Is provided in the position of (a), Respectively represent the first On the first propagation path The distance of the root antenna to the user, , Represent the first The positions of the users or the scatterers corresponding to the propagation paths; 2) For fixed antennas With user location The lower two frequencies And The correlation of the channel response in the frequency domain is expressed as: , 3) For the same frequency Same antenna The next two user positions And The correlation of the channel response in the spatial domain is expressed as: , Wherein, the Respectively represent And Is the first of (2) The propagation path to the first Distance of the root antenna.
  3. 3. The method for multi-domain channel extrapolation for a super-large scale MIMO system as claimed in claim 2, wherein in step 2, the actually observed partial channel state information in the antenna domain is a subset of the total antenna array response, denoted as , The unshielded part of the channel is used for representing the part of the channel state information obtained by actual observation, the shaded part is set to be zero, and the complete two-dimensional channel state information obtained by channel extrapolation is , wherein, The real number domain is represented by the number, , The number of the vertical antenna units and the number of the horizontal antenna units of the uniform area array antenna are respectively; In the frequency domain, the part of the channel state information obtained by actual observation is the channel state information on the part of subcarriers under all antenna arrays, which is recorded as The complete channel state information on all sub-carriers obtained by channel extrapolation is , The number of all subcarriers; in the space domain, the part of channel state information obtained by actual observation is the channel state information of partial space position observation, which is recorded as The complete channel state information at all spatial locations obtained by channel extrapolation is , To the number of spatial locations where channel state information can be observed.
  4. 4. The multi-domain channel extrapolation method for a very large scale MIMO system according to claim 3, wherein in step 3, the input block embedding module is configured to perform a block embedding operation on the partial channel state information actually observed in each dimension, i.e. Dividing into a plurality of space blocks, convoluting, rearranging and normalizing each space block, and mapping into a high-dimensional embedded vector to obtain Corresponding embedded feature tensor , wherein, Representing a set of complex fields, In order to be able to determine the number of batches, For the number of spatial blocks, , The dimensions of the spatial block in the vertical and horizontal directions respectively, Is an embedding dimension; the condition and mask embedding module is used for embedding the characteristic tensor Constructed as a mask matrix The mask value allocated to the observed position in the mask matrix is 0, and the mask value of the unobserved position is minus infinity; The position coding module is used for performing position coding on the mask matrix to obtain a sine position coding matrix For the first The first space block Dimensional component, its position code value Generated by sine and cosine function calculations as follows: , the time modulation module is used for counting the diffusion steps Conversion to a time-embedded vector and a sinusoidal position coding matrix The first masking multi-head attention module is sent together, and block embedding, masking, position coding and time modulation are carried out on noisy channel state information, and the first masking multi-head attention module is sent, and the third masking multi-head attention module outputs predicted equivalent noise; The conditional denoising diffusion implicit model constructs a conditional probability distribution, Time pair Is added in the noise adding process: , Wherein, the Respectively represent The time of day has noisy channel state information, In order to add the noise weight to the signal, , As a result of the gaussian noise, ; The posterior distribution of (2) is expressed as: , , , Wherein, the And The mean and variance of the gaussian distribution respectively, Is that The equivalent noise-plus-weight of the time instant, , Is that The equivalent noise of the time of day, , , For the equivalent noise-plus-weight, Is a unitary matrix, is fitted by a non-deterministic back diffusion process as follows Equivalent noise at time: , Wherein, the Representation of Equivalent noise at time.
  5. 5. The method for multi-domain channel extrapolation for a super-large scale MIMO system as claimed in claim 4, wherein in step 4, the joint loss function is used The method comprises the following steps: , Wherein, the Are the weighting coefficients of the loss function, Representing WCGAN the scoring value of the arbiter for the final predicted complete channel, Equivalent noise for network prediction in back-diffusion process Equivalent to real noise Is used to determine the loss term of (c), Predicted channel obtained by one-step denoising diffusion implicit model for equivalent noise using network prediction With true complete channels Loss term between, predictive channel The calculation is as follows: 。
  6. 6. The method for multi-domain channel extrapolation for a super-MIMO system as set forth in claim 5, wherein in step 5, the conditional transfer distribution between adjacent noise-adding time steps is defined as follows: , Wherein, the For controlling the level of randomness during diffusion, For the initial channel to be the same, Is that The equivalent noise-plus-weight of the time instant, Is that Posterior distribution mean value of (2); in the skip prediction process, the channel at any time before the current time From the channel at the current time Prediction noise And (3) predicting to obtain: , Wherein, the Is that The equivalent noise-plus-weight of the time instant, Is that The equivalent noise-plus-weight of the time instant, As a result of the gaussian noise, For the current time, variance , Is a super parameter.
  7. 7. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor, when executing the computer program, implements the steps of the multi-domain channel extrapolation method for a super-large scale MIMO system according to any one of claims 1 to 6.
  8. 8. A computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the steps of the multi-domain channel extrapolation method for a very large scale MIMO system according to any one of claims 1 to 6.

Description

Multi-domain channel extrapolation method for ultra-large-scale MIMO system Technical Field The invention relates to a multi-domain channel extrapolation method for a super-large-scale MIMO system, and belongs to the technical field of wireless communication. Background With the accelerated evolution of sixth generation mobile communication technology (6G), communication systems are evolving towards higher data rates, more accurate positioning capabilities, and larger scale space division multiplexing capabilities. As one of the key enabling technologies of 6G, a very large-scale multiple input multiple output (XL-MIMO) system has the potential to achieve Tb-level communication rate, centimeter-level positioning accuracy, and support hundreds of users to access in parallel by virtue of its drastic expansion on the order of antennas. However, compared with the traditional far-field MIMO assumption, the XL-MIMO system is always under the near-field propagation condition, and the channel of the XL-MIMO system presents new characteristics such as spherical wavefront, space non-stationarity and the like, so that the Channel State Information (CSI) is obviously improved in the aspects of dimension, structural complexity and sensitivity to the position/gesture of a user, and the difficulty of modeling and estimating the channel of the system is greatly increased. In practical deployment, accurate acquisition of full resolution CSI is critical to achieving high capacity communication, precise positioning and intelligent beamforming for XL-MIMO systems. However, with the rapid increase of the number of antennas, the number of subcarriers and the number of users, the conventional pilot-based channel estimation method faces multiple bottlenecks of prolonged training, large feedback bandwidth, high computational complexity and the like, and severely restricts the expandability and real-time performance of the XL-MIMO system. Therefore, in recent years, a channel extrapolation method is raised, and by utilizing the structural correlation in the existing observation channel information, the estimation and the restoration of the missing CSI are realized on the premise of not directly estimating all channel components, and the system pilot frequency and the calculation overhead are obviously reduced. However, multi-domain channel extrapolation still faces a series of key challenges. The spatial non-stationarity of the antenna domain and the extreme expansion of the array dimension result in huge model training and reasoning cost, the frequency domain is limited by a low-resolution analog-to-digital converter and a limited radio frequency bandwidth, dense sampling of subcarriers is difficult to realize, and the acquisition of CSI (channel state information) in a large-range position of the spatial domain requires extremely high perceived resource and feedback cost, especially when a channel map or a scene map is constructed, the measurement of each spatial point causes an intolerable hardware burden and labor cost. These problems together determine the inefficiency of conventional estimation methods in large-scale systems, and there is an urgent need to develop multi-domain channel extrapolation techniques with low overhead and high accuracy. In view of the above challenges, various extrapolation methods have been proposed by studies from the viewpoints of path geometry modeling, graph structure modeling, differential equation modeling, and the like. For example, the path disturbance model realizes multi-domain extrapolation through sharing geometric parameters, but has poor robustness under dynamic environment or non-line-of-sight conditions, the graph signal modeling method is suitable for dense space domain modeling, but has insufficient expandability in user uneven distribution or multi-domain migration scenes, and the ordinary differential equation method has time continuous modeling capability, but the numerical integration process of the ordinary differential equation method introduces high reasoning time delay and computational complexity and is highly sensitive to initial observation. In recent years, generating a countering network (GAN) and Denoising Diffusion Probability Model (DDPM) exhibits a strong generating capability in channel modeling. Wherein, used to learn a high dimensional channel prior DDPM can generate high fidelity CSI by iterative denoising sampling. However, the GAN method generally has the problems of mode collapse and unstable training and lacks the condition control capability, while the traditional diffusion model has superior performance in terms of fidelity and diversity, the high-dimensional sampling process takes longer time, the actual system deployment requirement is difficult to meet, most of the existing diffusion models only support single domain modeling, and the collaborative extrapolation tasks of antennas, frequencies and spatial domains are difficult to process simultaneously i