A high-speed moving MIMO-OFDM channel estimation method based on an adaptive Transformer network
Patent Information
- Application Number
- CN202611008188.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-08
- Publication Date
- 2026-10-09
AI Technical Summary
研究者们利用卷积神经网络（Convolutional Neural Network, CNN）的局部特征提取优势恢复信道矩阵，通过在空-频域、角-时域等多域进行卷积操作，或拓展至空间-频率-时间三维结构以捕获多维度相关性，部分方案还引入元学习框架提升模型对新环境的适应能力，但标准CNN结构对所有特征通道平等处理，缺乏关键信息辨别能力，在强噪声环境下易受干扰
[0044]（1）与传统最小二乘估计、线性最小均方误差算法相比，本发明通过显式融合信噪比、时延扩展以及多普勒频移物理先验，动态调控特征分布，结合多尺度特征增强与全局建模，显著提高了时变信道的估计准确性；
Smart Images
Figure CN122891398A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wireless communication technology, and in particular to channel estimation technology for multiple-input multiple-output orthogonal frequency division multiplexing systems with time-frequency dual-selective fading and Doppler frequency offset interference in high-speed mobile scenarios. Background Technology
[0002] Multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) technology, with its high spectral efficiency and resistance to multipath fading, has become a key physical layer technology for fifth-generation and future wireless communication systems. In this system, accurate channel state information (CSI) is fundamental to achieving reliable signal detection, efficient beamforming, and adaptive resource allocation; its estimation accuracy directly affects the overall system performance.
[0003] Traditional channel estimation algorithms (such as the least squares method and the linear minimum mean square error method) typically assume that the channel remains quasi-static during the coherence time, making them suitable for low-speed mobile scenarios. However, in high-speed mobile scenarios such as high-speed rail, drones, and vehicle-to-everything (V2X) communication, the high-speed relative motion between the transmitter and receiver causes significant Doppler frequency shift, resulting in strong time-frequency dual-selective fading of the channel. This disrupts the orthogonality between OFDM subcarriers and introduces severe inter-carrier interference, greatly increasing the complexity of channel estimation.
[0004] In recent years, deep learning has provided a new paradigm for breaking through the bottlenecks of traditional methods due to its powerful data-driven nonlinear modeling capabilities and weak dependence on explicit channel statistical priors. Existing research mainly focuses on the effective extraction of channel time-frequency features, the enhanced identification of key features, and the model's adaptability to dynamic environments. Researchers have utilized the local feature extraction advantages of Convolutional Neural Networks (CNNs) to recover the channel matrix. They have performed convolution operations in multiple domains such as the space-frequency domain and the angle-time domain, or extended them to a three-dimensional space-frequency-time structure to capture multi-dimensional correlations. Some schemes have also introduced meta-learning frameworks to improve the model's adaptability to new environments. However, the standard CNN structure treats all feature channels equally, lacks the ability to distinguish key information, and is easily interfered with in noisy environments. To enhance feature selection, researchers have incorporated attention mechanisms into CNNs. This is achieved by embedding attention modules in residual blocks or employing compressed excitation channel attention mechanisms, adaptively adjusting feature channel weights to suppress noise and enhance effective feature representation. However, these methods often rely on static feature correlations to calculate attention weights, resulting in limited adaptive adjustment capabilities to feature distribution in dynamic environments with rapidly changing channels, and their robustness remains to be improved. Addressing the issue of CNNs' local receptive field limiting long-range dependency modeling, related research has introduced the Transformer architecture into channel estimation. By sequentially fusing CNN and Transformer modules, both local detail capture and global context modeling are considered. However, the performance improvement of this approach depends to some extent on weighted training strategies for high signal-to-noise ratio (SNR) data, and robustness to high-frequency noise under low SNR conditions remains challenging. To fundamentally improve the model's adaptability to dynamic and complex channels, recent research has begun to explore explicitly injecting channel statistical features such as signal-to-noise ratio, Doppler shift, and delay spread as prior information into the network. By combining Vision Transformer to process time-frequency features or encoding channel statistics as embedding vectors to guide feature extraction, adaptive channel estimation can be achieved, resulting in certain performance gains in dynamic scenarios. However, most related studies are limited to single-input single-output orthogonal frequency division multiplexing scenarios, making it difficult to directly adapt to the antenna coupling, spatial dimension expansion, and complex time-frequency interference problems of high-speed mobile MIMO-OFDM systems. Moreover, existing methods often adopt implicit strategies of label splicing or embedding addition when fusing prior information, which highly rely on the network itself to learn complex correlation mappings from data. This not only causes computational redundancy but also makes it difficult to achieve accurate and efficient control of intermediate feature distribution. Summary of the Invention
[0005] In view of this, the purpose of this invention is to propose a high-speed mobile MIMO-OFDM channel estimation method based on an adaptive Transformer network. This method designs an adaptive Transformer channel estimation network that integrates a characteristic linear modulation mechanism, using the least-squares initial channel estimation complex matrix corresponding to the pilot position and the channel statistical physical parameter set as network inputs. First, two-dimensional linear upsampling is performed on the initial channel estimation to recover the complete time-frequency grid structure, and then local time-frequency features are extracted through a convolutional feature enhancement module. Second, the channel statistical physical parameter set is used as a priori conditions to dynamically generate scaling and offset coefficients corresponding to the feature layer, and channel-by-channel affine modulation is performed on the local time-frequency features to achieve accurate adaptation to the dynamic changes in channel statistical characteristics. Finally, the modulated feature input is embedded into a two-dimensional position-coded Transformer encoder for global dependency modeling, combined with a multi-scale reconstruction fusion structure, ultimately outputting a high-precision channel estimation matrix across all dimensions.
[0006] To achieve the above objectives, the technical solution adopted by this invention is: a high-speed mobile MIMO-OFDM channel estimation method based on an adaptive Transformer network, comprising the following steps:
[0007] S101: Obtain the least-squares initial channel estimation complex matrix at the pilot. and channel parameters (signal-to-noise ratio) Latency spread and maximum Doppler frequency shift );
[0008] S102: Completion via separable two-dimensional linear upsampling The full-time frequency domain channel estimate is obtained. ;
[0009] S103: The real and imaginary parts of the multi-antenna complex channel are jointly modeled by a complex hybrid layer, and then multi-scale time-frequency features are extracted by a feature enhancement module to obtain a local feature map. ;
[0010] S104: Yes After time-frequency segmentation, the sequences are concatenated along the channel dimension to obtain the Transformer adaptation feature sequence. ;
[0011] S105: Input the channel parameters into the channel adaptive feature modulation module to generate affine parameter modulation. ,get ;
[0012] S106: Transformer encoder combined with two-dimensional learnable position code pair Global modeling is performed, and encoded features are obtained using a self-attention mechanism. ;
[0013] S107: After inverse block division and Residual fusion and convolutional thinning yield the final channel estimation matrix. .
[0014] Based on the above technical solution, in step S101, the least-squares initial channel estimation matrix at the pilot position is obtained. The methods include:
[0015] At the pilot position, based on the known pilot symbol With the received pilot symbol For each pair of transmit and receive antennas, least squares estimation is performed to obtain the initial channel estimation matrix at the pilot position.
[0016] ;
[0017] Simultaneously, save the set of physical parameters related to the current channel statistics, including the signal-to-noise ratio. Latency spread and maximum Doppler frequency shift .
[0018] Based on the above technical solution, in step S102, the initial channel estimation at the pilot position is performed. The channel is decomposed into real and imaginary parts and processed separately. Linear mapping is then performed using the frequency domain mapping matrix and the time domain mapping matrix, respectively, to achieve separable two-dimensional linear upsampling and obtain the full-time-frequency domain channel matrix. .
[0019] Based on the above technical solution, the step S103 involves obtaining a local feature map. The methods specifically include:
[0020] First, a lightweight 1×1 convolutional network is used to... The real and imaginary parts are combined through a full-channel weighted combination to obtain the hybrid feature. ; then the hybrid features The input feature enhancement module performs deep feature extraction;
[0021] The feature enhancement module includes an input projection layer and a multi-scale convolutional block, wherein the input projection layer is used to... Projected onto a high-dimensional feature space; the multi-scale convolutional block includes a standard 3×3 convolutional branch, a dilated 3×3 convolutional branch, and a 1×1 convolutional branch, and fuses multi-scale features in the following way:
[0022] ;
[0023] in, The GELU function is used; deep features are obtained by stacking two layers of the multi-scale convolutional structure. Building upon this, a compression-excitation mechanism is introduced to adaptively calibrate the weight vectors of each feature channel. The weighted features are then mapped back to the original channel dimensions via a projection layer, and the original mixed features are then combined through a global residual connection. The enhanced features are then fused together to ultimately output high-quality local features. .
[0024] Based on the above technical solution, the feature sequence adapted to the Transformer architecture is obtained in step S104. The methods include:
[0025] Use a sliding window with a size of 3×2 Perform non-overlapping block division; for each local block, flatten the 3×2 local matrix in each channel into a vector of length 6, and perform serial splicing in the channel dimension so that the entire OFDM time-frequency grid is reshaped into the feature sequence.
[0026] Based on the above technical solution, in step S105, the following is obtained: The methods specifically include:
[0027] First, three independent multilayer perceptrons are used to measure the signal-to-noise ratio. Latency spread and maximum Doppler frequency shift Embedding encoding is performed and the data are concatenated into a channel state vector. Then the characteristic linear modulation parameter generator will... Mapped to scaling factor and offset factor And by using these two factors to modify the feature sequence output by the block embedding module Perform linear modulation at the feature level:
[0028] ;
[0029] in, This represents the Hada code product.
[0030] Based on the above technical solution, the serialization features in step S106 And perform global modeling and obtain encoded features. Specific methods include:
[0031] First, the input feature sequence Linear projection and two-dimensional learnable positional encoding are performed sequentially to obtain a feature representation with video structure awareness; the linear projection is weighted by a weight matrix. and bias vector The features are mapped to a high-dimensional space; the two-dimensional learnable positional encoding is achieved through separately initialized row embedding parameter matrices. With column embedding parameter matrix Two-dimensional position mask generated by broadcast addition And superimposed on the projected features;
[0032] Secondly, the features after fusing location information The input consists of a Transformer encoder composed of L stacked layers: each layer contains a multi-head self-attention module and a feedforward neural network, and uses residual connections and layer normalization; the multi-head self-attention module performs parallel computation of M attention heads, each head extracts feature relationships through a scaled dot product attention mechanism, and the outputs of all heads are concatenated and linearly projected to obtain the attention output. :
[0033] ;
[0034] Among them, each attention head The calculation method is as follows:
[0035] ;
[0036] in, , , Let be the projection matrices of the query, key, and value of the h-th header, respectively. , , For the corresponding bias matrix, Scaling factor For model dimensions, To output the projection matrix, To output the bias matrix, This indicates a concatenation operation along the feature dimension.
[0037] Finally, the output of each encoder layer is processed sequentially through residual connections, layer normalization, and a feedforward network, ultimately outputting the globally modeled encoded features via the Lth encoder layer:
[0038] ;
[0039] ;
[0040] ;
[0041] in, Representative level normalization, and As a linear projection and As a bias matrix This represents the Gaussian error linear unit (GELU).
[0042] Based on the above technical solution, in step S107, the encoded features output by the encoder are... First, through the weight matrix and bias matrix A linear transformation is performed to recover the target feature dimension, and an inverse block operation is performed on the transformed features to recover the two-dimensional feature map. ; will the The input channel reconstruction module and the output of the feature enhancement module Residual fusion is achieved by element-wise addition to obtain the global channel estimate. Finally, regarding the above... Multi-layer convolution thinning is performed to output the final channel estimation matrix. .
[0043] The present invention has the following beneficial effects:
[0044] (1) Compared with traditional least squares estimation and linear minimum mean square error algorithm, this invention dynamically adjusts the feature distribution by explicitly fusing signal-to-noise ratio, time delay spread and Doppler frequency shift physical prior, and combines multi-scale feature enhancement and global modeling, which significantly improves the estimation accuracy of time-varying channels;
[0045] (2) Compared with existing adaptive methods, this invention achieves efficient fusion of prior information and deep features through a feature linear modulation mechanism, reduces the redundancy of feature modulation, and improves the model's adaptability to dynamic channels;
[0046] (3) Compared with the model without an adaptive module, the present invention reduces the impact of time-frequency dual-selective fading and inter-carrier interference by introducing a channel adaptive module based on the characteristic linear modulation mechanism, thereby improving the estimation stability in high-speed mobile scenarios. Attached Figure Description
[0047] Figure 1 Here is a flowchart of a high-speed mobile MIMO-OFDM channel estimation method based on an adaptive Transformer network proposed in this invention;
[0048] Figure 2 This is a diagram illustrating the overall model architecture of a high-speed mobile MIMO-OFDM channel estimation method based on an adaptive Transformer network proposed in this invention. Detailed Implementation
[0049] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0050] like Figure 1 As shown, the design concept of this invention is to provide a high-speed mobile MIMO-OFDM channel estimation method based on an adaptive Transformer network, including:
[0051] S101: Obtain the least-squares initial channel estimation complex matrix at the pilot. and channel parameters (signal-to-noise ratio) Latency spread and maximum Doppler frequency shift );
[0052] S102: Completion via separable two-dimensional linear upsampling The full-time frequency domain channel estimate is obtained. ;
[0053] S103: The real and imaginary parts of the multi-antenna complex channel are jointly modeled by a complex hybrid layer, and then multi-scale time-frequency features are extracted by a feature enhancement module to obtain a local feature map. ;
[0054] S104: Yes After time-frequency segmentation, the sequences are concatenated along the channel dimension to obtain the Transformer adaptation feature sequence. ;
[0055] S105: Input the channel parameters into the channel adaptive feature modulation module to generate affine parameter modulation. ,get ;
[0056] S106: Transformer encoder combined with two-dimensional learnable position code pair Global modeling is performed, and encoded features are obtained using a self-attention mechanism. ;
[0057] S107: After inverse block division and Residual fusion and convolutional thinning yield the final channel estimation matrix. .
[0058] like Figure 2 The figure shows the overall model architecture of a high-speed mobile MIMO-OFDM channel estimation method based on an adaptive Transformer network.
[0059] In the overall model architecture diagram of the adaptive Transformer network, the least squares initial channel estimation complex matrix corresponding to the pilot position is used. The channel statistical physical parameter set is used as the model input.
[0060] At the pilot position, based on the known pilot symbol With the received pilot symbol For each pair of transmit and receive antennas, least squares estimation is performed to obtain the initial channel estimation matrix at the pilot position:
[0061] ;
[0062] Simultaneously, save the set of physical parameters related to the current channel statistics, including the signal-to-noise ratio. Latency spread and maximum Doppler frequency shift .
[0063] Furthermore, the initial channel estimation at the pilot location is... The channel is decomposed into real and imaginary parts and processed separately. Linear mapping is then performed using the frequency domain mapping matrix and the time domain mapping matrix, respectively, to achieve separable two-dimensional linear upsampling and obtain the full-time-frequency domain channel matrix. .
[0064] Furthermore, a lightweight 1×1 convolutional network is used to... The real and imaginary parts are combined through a full-channel weighted combination to obtain the hybrid feature. ; then the hybrid features The input feature enhancement module performs deep feature extraction. The feature enhancement module includes an input projection layer and multi-scale convolutional blocks: wherein the input projection layer is used to... Projected to a high-dimensional feature space; the multi-scale convolutional block includes a standard 3×3 convolutional branch, a dilated 3×3 convolutional branch, and a 1×1 convolutional branch, and the multi-scale features are fused in the following way:
[0065] ;
[0066] in, The GELU function is used; deep features are obtained by stacking two layers of the multi-scale convolutional structure. Building upon this, a compression-excitation mechanism is introduced to adaptively calibrate the weight vectors of each feature channel. The weighted features are then mapped back to the original channel dimensions via a projection layer, and the original mixed features are then combined through a global residual connection. The enhanced features are then fused together to ultimately output high-quality local features. .
[0067] Furthermore, a sliding window with a size of 3×2 is used for... Perform non-overlapping block division; for each local block, flatten the 3×2 local matrix in each channel into a vector of length 6, and perform serial splicing in the channel dimension so that the entire OFDM time-frequency grid is reshaped into the feature sequence.
[0068] Furthermore, three independent multilayer perceptrons were used to analyze the signal-to-noise ratio. Latency spread and maximum Doppler frequency shift Embedding encoding is performed and the data are concatenated into a channel state vector. Then the characteristic linear modulation parameter generator will... Mapped to scaling factor and offset factor And by using these two factors to modify the feature sequence output by the block embedding module Perform linear modulation at the feature level:
[0069] ;
[0070] in, This represents the Hada code product.
[0071] Furthermore, for the input feature sequence Linear projection and two-dimensional learnable position encoding are performed sequentially to obtain a feature representation with time-frequency structure awareness; the linear projection is processed through a weight matrix. and bias vector The features are mapped to a high-dimensional space; the two-dimensional learnable positional encoding is achieved through separately initialized row embedding parameter matrices. With column embedding parameter matrix Two-dimensional position mask generated by broadcast addition And superimposed on the projected features;
[0072] Furthermore, the features after fusing location information The input consists of a Transformer encoder composed of L stacked layers: each layer contains a multi-head self-attention module and a feedforward neural network, and uses residual connections and layer normalization; the multi-head self-attention module performs parallel computation of M attention heads, each head extracts feature relationships through a scaled dot product attention mechanism, and the outputs of all heads are concatenated and linearly projected to obtain the attention output. :
[0073] ;
[0074] Among them, each attention head The calculation method is as follows:
[0075] ;
[0076] in, , , Let be the projection matrices of the query, key, and value of the h-th header, respectively. , , For the corresponding bias matrix, Scaling factor For model dimensions, To output the projection matrix, To output the bias matrix, This indicates a concatenation operation along the feature dimension.
[0077] Finally, the output of each encoder layer is processed sequentially through residual connections, layer normalization, and a feedforward network, ultimately outputting the globally modeled encoded features via the Lth encoder layer:
[0078] ;
[0079] ;
[0080] ;
[0081] in, Representative level normalization, and As a linear projection and As a bias matrix This represents the Gaussian error linear unit (GELU).
[0082] Furthermore, the encoded features of the encoder output. First, through the weight matrix and bias matrix A linear transformation is performed to recover the target feature dimension, and an inverse block operation is performed on the transformed features to recover the two-dimensional feature map. ; will the The input channel reconstruction module and the output of the feature enhancement module Residual fusion is achieved by element-wise addition to obtain the global channel estimate. Finally, regarding the above... Multi-layer convolution thinning is performed to output the final channel estimation matrix. .
[0083] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.
Claims
1. A high-speed mobile MIMO-OFDM channel estimation method based on adaptive Transformer networks, characterized in that, Includes the following steps: S101: Obtain the least-squares initial channel estimation complex matrix at the pilot. and channel parameters (signal-to-noise ratio) Latency spread and maximum Doppler frequency shift ); S102: Completion via separable two-dimensional linear upsampling The full-time frequency domain channel estimate is obtained. ; S103: The real and imaginary parts of the multi-antenna complex channel are jointly modeled by a complex hybrid layer, and then multi-scale time-frequency features are extracted by a feature enhancement module to obtain a local feature map. ; S104: Yes After time-frequency segmentation, the sequences are concatenated along the channel dimension to obtain the Transformer adaptation feature sequence. ; S105: Input the channel parameters into the channel adaptive feature modulation module to generate affine parameter modulation. ,get ; S106: Transformer encoder combined with two-dimensional learnable position code pair Global modeling is performed, and encoded features are obtained using a self-attention mechanism. ; S107: After inverse block division and Residual fusion and convolutional thinning yield the final channel estimation matrix. .
2. The high-speed mobile MIMO-OFDM channel estimation method based on adaptive Transformer networks according to claim 1, characterized in that: Step S101 is used to obtain the least-squares initial channel estimation matrix at the pilot position. The methods include: At the pilot position, based on the known pilot symbol With the received pilot symbol For each pair of transmit and receive antennas, least squares estimation is performed to obtain the initial channel estimation matrix at the pilot position: ； Simultaneously, save the set of physical parameters related to the current channel statistics, including the signal-to-noise ratio. Latency spread and maximum Doppler frequency shift .
3. The high-speed mobile MIMO-OFDM channel estimation method based on adaptive Transformer networks according to claim 1, characterized in that: In step S102, the initial channel estimation at the pilot position is performed. The channel is decomposed into real and imaginary parts and processed separately. Linear mapping is then performed using the frequency domain mapping matrix and the time domain mapping matrix, respectively, to achieve separable two-dimensional linear upsampling and obtain the full-time-frequency domain channel matrix. .
4. The high-speed mobile MIMO-OFDM channel estimation method based on adaptive Transformer networks according to claim 3, characterized in that: In step S103, local feature maps are obtained. The methods specifically include: First, a lightweight 1×1 convolutional network is used to... The real and imaginary parts are combined through a full-channel weighted combination to obtain the hybrid feature. ; then the hybrid features The input feature enhancement module performs deep feature extraction; The feature enhancement module includes an input projection layer and a multi-scale convolutional block, wherein the input projection layer is used to... Projected onto a high-dimensional feature space; the multi-scale convolutional block includes a standard 3×3 convolutional branch, a dilated 3×3 convolutional branch, and a 1×1 convolutional branch, and fuses multi-scale features in the following way: ； in, The GELU function is used; deep features are obtained by stacking two layers of the multi-scale convolutional structure. Building upon this, a compression-excitation mechanism is introduced to adaptively calibrate the weight vectors of each feature channel. The weighted features are then mapped back to the original channel dimensions via a projection layer, and the original mixed features are then combined through a global residual connection. The enhanced features are then fused together to ultimately output high-quality local features. .
5. The high-speed mobile MIMO-OFDM channel estimation method based on adaptive Transformer networks according to claim 1, characterized in that: In step S104, the feature sequence adapted to the Transformer architecture is obtained. The methods include: Use a sliding window with a size of 3×2 Perform non-overlapping block division; for each local block, flatten the 3×2 local matrix in each channel into a vector of length 6, and perform serial splicing in the channel dimension so that the entire OFDM time-frequency grid is reshaped into the feature sequence.
6. The high-speed mobile MIMO-OFDM channel estimation method based on adaptive Transformer networks according to claim 1, characterized in that: In step S105, the following is obtained The methods specifically include: First, three independent multilayer perceptrons are used to measure the signal-to-noise ratio. Latency spread and maximum Doppler frequency shift Embedding encoding is performed and the data are concatenated into a channel state vector. Then the characteristic linear modulation parameter generator will... Mapped to scaling factor and offset factor And by using these two factors to modify the feature sequence output by the block embedding module Perform linear modulation at the feature level: ； in, This represents the Hada code product.
7. The high-speed mobile MIMO-OFDM channel estimation method based on adaptive Transformer networks according to claim 1, characterized in that: In step S106, the serialization features are... And perform global modeling and obtain encoded features. Specific methods include: First, the input feature sequence Linear projection and two-dimensional learnable position encoding are performed sequentially to obtain a feature representation with time-frequency structure awareness; the linear projection is processed through a weight matrix. and bias vector The features are mapped to a high-dimensional space; the two-dimensional learnable positional encoding is achieved through separately initialized row embedding parameter matrices. With column embedding parameter matrix Two-dimensional position mask generated by broadcast addition And superimposed on the projected features; Secondly, the features after fusing location information The input consists of a Transformer encoder composed of L stacked layers: each layer contains a multi-head self-attention module and a feedforward neural network, and uses residual connections and layer normalization; the multi-head self-attention module performs parallel computation of M attention heads, each head extracts feature relationships through a scaled dot product attention mechanism, and the outputs of all heads are concatenated and linearly projected to obtain the attention output. : ； Among them, each attention head The calculation method is as follows: ； in, , , Let be the projection matrices of the query, key, and value of the h-th header, respectively. , , For the corresponding bias matrix, Scaling factor For model dimensions, To output the projection matrix, To output the bias matrix, This indicates a concatenation operation along the feature dimension; Finally, the output of each encoder layer is processed sequentially through residual connections, layer normalization, and a feedforward network, ultimately outputting the globally modeled encoded features via the Lth encoder layer: ； ； ； in, Representative level normalization, and As a linear projection and As a bias matrix This represents the Gaussian error linear unit (GELU).
8. The high-speed mobile MIMO-OFDM channel estimation method based on adaptive Transformer networks according to claim 1, characterized in that: In step S107, the encoded features output by the encoder are... First, through the weight matrix and bias matrix A linear transformation is performed to recover the target feature dimension, and an inverse block operation is performed on the transformed features to recover the two-dimensional feature map. ; will the The input channel reconstruction module and the output of the feature enhancement module Residual fusion is achieved by element-wise addition to obtain the global channel estimate. Finally, regarding the above... Multi-layer convolution thinning is performed to output the final channel estimation matrix. .
9. The method according to any one of claims 1 to 8, characterized in that, The adaptive Transformer network is trained by minimizing the mean square error between the full-dimensional channel matrix and the true channel matrix.