A noise-aware axial window attention channel estimation method

Abstract

This invention belongs to the field of wireless communication channel estimation technology, specifically relating to a noise-aware axial window attention channel estimation method. The method includes: acquiring pilot observation signals and constructing an observation tensor; inputting the observation tensor into a convolutional front end for feature extraction to obtain shallow features; processing the shallow features using a main network composed of cascaded self-attention residual blocks from multiple axial windows to obtain deep features; inputting the deep features into a projection head for processing to obtain residual prediction; and reconstructing the residuals based on the residual predictions and the observation tensor to obtain the channel estimation result. This invention can achieve effective weak feature extraction and structural modeling under low SNR and mixed field conditions, improving the stability, robustness, and engineering feasibility of channel estimation under mixed field conditions, and has good application prospects.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication channel estimation technology, specifically relating to a noise-aware axial window attention channel estimation method. Background Technology

[0002] Extremely Large Antenna Array (ELAA) / Extremely Large Multiple Input Multiple Output (XL-MIMO) is one of the key technologies for 6G. To achieve beamforming, link adaptation, and resource scheduling, the receiver (or base station) typically needs to obtain Channel State Information (CSI). Therefore, channel estimation in ELAA / XL-MIMO scenarios has significant engineering implications.

[0003] As the array aperture increases and the operating frequency band rises, the propagation mechanism changes: the near-field spherical wave effect intensifies, and in engineering scenarios, both near-field and far-field scattering paths often coexist, forming a hybrid-field propagation. This makes the channel structure more non-stationary and statistically mismatched. Under low signal-to-noise ratio (SNR) conditions, weak features in pilot observations are more easily submerged by noise, and the aforementioned hybrid structure further exacerbates feature mismatch, leading to a decrease in channel estimation accuracy.

[0004] For the channel estimation problem, existing methods can be broadly divided into two categories: traditional estimation methods based on models / priors and data-driven learning methods.

[0005] (1) Model-based traditional methods: For example, least squares (LS) and minimum mean square error (MMSE) methods usually rely on certain statistical priors or channel model assumptions; in addition, compressed sensing / sparse reconstruction methods (such as OMP and its variants) often utilize the sparsity of millimeter wave / terahertz channels in the angle domain or dictionary domain, and achieve reconstruction by constructing a dictionary and solving for sparse coefficients. These methods have clear mathematical models and interpretability, and can achieve good performance when the model matching or prior is accurate. However, in the ELAA / XL-MIMO hybrid scenario, near-field spherical waves and far-field plane waves coexist, and the number of paths (such as the number of far-field paths) is limited. Near-field path count The sparse support set and its statistical properties change dynamically with the environment, easily leading to dictionary mismatch, prior mismatch, or a sharp increase in parameter dimensionality. Simultaneously, under low SNR conditions, the sparse support set is easily disturbed by noise, resulting in unstable reconstruction. These factors cause traditional methods to often face problems such as decreased accuracy, insufficient robustness, and high computational complexity under complex mixed fields and low SNR.

[0006] (2) Deep learning-based methods: With the development of deep learning, models such as convolutional neural networks (CNNs) have been used to learn channel mapping relationships end-to-end from pilot observations, realizing the integration of feature extraction and estimation output. Compared with traditional methods, they have certain advantages in nonlinear modeling and weak feature extraction. However, in ELAA / XL-MIMO scenarios, the array and channel dimensions are extremely high. Directly using deeper or wider networks will lead to a significant increase in the number of parameters and computation, making it difficult to meet the requirements of resource constraints and real-time deployment. At the same time, the mixed field channel has long-range spatial correlation and non-stationary characteristics, and the limited receptive field of local convolution may be difficult to effectively model. Under low SNR conditions, the learning model may also be sensitive to noise, resulting in feature / attention diffusion leading to mismodeling, and insufficient generalization under different scenarios or different path composition changes. In order to enhance long-range dependency modeling, some works have attempted to introduce self-attention mechanisms, but global self-attention has a large computational cost under ultra-large-scale dimensions, and at low SNR, it is easy to spread noise components through global interaction, thus affecting the estimation stability.

[0007] Therefore, existing technologies still generally have the following shortcomings: First, the coexistence of near and far fields in mixed fields leads to complex structures, statistical mismatch, and insufficient estimation robustness; second, weak features are easily submerged by noise at low SNR, resulting in limited denoising and weak feature enhancement capabilities; and third, the ultra-large-scale dimensions bring computational and storage overhead, making global modeling (especially attention-based global interactions) difficult to apply directly or meet real-time requirements.

[0008] In summary, there is an urgent need for a channel estimation method that can effectively extract weak features and model structures under low SNR and mixed field conditions, while also possessing controllable complexity and engineering feasibility. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention proposes a noise-aware axial window attention channel estimation method, which includes:

[0010] S1: Acquire pilot observation signals and construct the observation tensor;

[0011] S2: Input the observation tensor into the convolutional front end to extract features and obtain shallow features;

[0012] S3: Using the method The main network, composed of multi-head self-attention residual blocks connected in series with axial windows, processes shallow features to obtain deep features.

[0013] S4: Input the deep features into the projection head for processing to obtain residual prediction;

[0014] S5: Residual reconstruction is performed based on residual prediction and observation tensor to obtain channel estimation results.

[0015] Preferably, step S1 specifically includes:

[0016] The receiver acquires pilot observation signals in the pilot time slot and forms a frequency domain observation matrix. ;

[0017] The frequency domain observation matrix is ​​split into real and imaginary parts and concatenated in the channel dimension to obtain the observation tensor.

[0018] Preferably, in step S2, the shallow feature representation is obtained as follows:

[0019]

[0020] in, Indicates shallow features. Represents the ReLU activation function. Indicates batch normalization, This represents a two-dimensional convolution operation. Represents the observation tensor. This represents the trainable convolution kernel parameters of the two-dimensional convolutional layer in the convolutional front end.

[0021] Preferably, in step S3, the data processing procedure for the axial window multi-head self-attention residual block includes:

[0022] S31: Adaptively generate noise-aware gating weights based on input features;

[0023] S32: Decompose the noise-sensing gating weights along the height axis and width axis respectively to obtain the height axis sequence and the width axis sequence; concatenate the height axis sequence and the width axis sequence to obtain the axial sequence;

[0024] S33: Divide the axial sequence into multiple window sequences according to a preset window length;

[0025] S34: Perform a linear mapping of multi-head attention for each window to obtain the key, value, and query vector under each attention head;

[0026] S35: Each attention head calculates its output features based on the key, value, and query vector, and introduces a window mask;

[0027] S36: Concatenate the output features of each attention head and obtain the window attention output through linear mapping;

[0028] S37: Concatenate the window attention outputs under all windows in the original window order and restore the two-dimensional feature layout to obtain the output of the axial window multi-head self-attention module;

[0029] S38: The output of the axial window multi-head self-attention module is mapped by two-dimensional convolution, added to the input features, and then processed by ReLU to obtain the output features of the axial window multi-head self-attention residual block.

[0030] Furthermore, the adaptive generation noise-aware gating weights are represented as follows:

[0031]

[0032] in, Indicates the noise-aware gating weight. This represents the Sigmoid function. This represents a two-dimensional convolution operation. This represents the input features of the axial window multi-head self-attention residual block. This represents the trainable convolution kernel parameters of a two-dimensional convolutional layer in a noise-aware gating system.

[0033] Furthermore, the output features of the attention head are calculated as follows:

[0034]

[0035]

[0036] in, Indicates the first Each attention head outputs features. Indicates the first Attention weights for each attention head Indicates the noise-aware gating weight. Indicates the first A query vector with attention heads. Indicates the first A vector of values ​​for each attention head. The dimension representing the attention head, Indicates the first The key vector of an attention head, This represents the window mask matrix.

[0037] Furthermore, the formula for calculating the window mask matrix is ​​as follows:

[0038]

[0039] in, Represents the window mask matrix. Indicates the base mask. This represents the adjustment coefficient. This represents the mapping from gating to mask bias.

[0040] Preferably, in step S4, the residual prediction is expressed as:

[0041]

[0042] in, Indicates residual prediction, Represents a linear mapping layer. This represents a two-dimensional convolution operation. Indicates deep features, This represents the trainable convolution kernel parameters of the two-dimensional convolutional layer in the projector.

[0043] Preferably, in step S5, the residual reconstruction based on the residual prediction and observation tensor is expressed as follows:

[0044]

[0045] in, This represents the channel estimation result. Represents the observation tensor. This indicates residual prediction.

[0046] The beneficial effects of this invention are as follows:

[0047] This invention, through windowed self-attention with axial decomposition, can effectively characterize long-range spatial correlations in the two-dimensional array dimension, reducing the coexistence of near / far fields and changes in path composition (such as...). , The structural / statistical mismatch caused by changes in the field can be mitigated, thereby improving the robustness and stability of channel estimation under mixed field conditions.

[0048] This invention obtains noise priors through noise perception modeling and uses them to constrain window attention interaction and residual fusion. This can suppress the spread of noise and mismodeling during the interaction process under low SNR, improve the ability to extract weak features, and thus improve the estimation accuracy and stability under low SNR conditions.

[0049] This invention replaces global attention interaction with axial decomposition and in-window interaction, which significantly reduces computation and storage overhead while maintaining structural modeling capabilities, making this method more suitable for engineering implementation and real-time processing requirements of ELAA / XL-MIMO high-dimensional channel estimation. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the noise-aware axial window attention channel estimation network structure in this invention;

[0051] Figure 2 This is a schematic diagram of the internal structure of the axial window multi-head self-attention module in this invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] To address the challenges of weak features being easily submerged by noise under low signal-to-noise ratio (SNR) conditions in ELAA / XL-MIMO near-field and far-field hybrid propagation scenarios, the complex channel structure and statistical mismatch caused by the coexistence of near and far fields, and the high computational and storage overhead of global modeling under ultra-large-scale array dimensions, this invention proposes a noise-aware axial window attention channel estimation method. This method constructs a two-dimensional array feature tensor (real / imaginary dual channels) suitable for network processing from the pilot observation signal. First, a convolutional front-end is used to perform preliminary feature extraction on the input to enhance local stability representation. Then, a noise-aware module is introduced to explicitly obtain noise prior information (e.g., noise residuals or gating / confidence weights) from the observed features. This noise prior is embedded in the axial window multi-head self-attention module to constrain and modulate the window mask or attention interaction strength, thereby suppressing noise propagation and mismodeling during the attention interaction process under low SNR conditions. The method includes the following:

[0054] S1: Acquire pilot observation signals and construct observation tensors.

[0055] In a time-division duplex (TDD) system, the receiver (or base station) acquires pilot observation signals in the pilot time slot, forming a frequency domain (or time-frequency domain) observation matrix. In some preferred embodiments of the present invention, the pilot reception process can be described using the following linear observation model:

[0056]

[0057] in, For pilot matrix, For the ELAA / XL-MIMO channel matrix to be estimated, This is the noise term (which can be considered as additive Gaussian noise or equivalent noise). The complex observation matrix... The input is split into real and imaginary parts and then concatenated along the channel dimension to form a two-dimensional multi-channel input.

[0058]

[0059] in, This corresponds to the two dimensions of a two-dimensional array (or its equivalent raster representation). This construction method allows complex observations to be transformed into an input representation suitable for convolutional / attention networks without losing phase information.

[0060] This invention designs a noise-aware axial window attention channel estimation network structure for channel estimation, such as... Figure 1 As shown, its implementation process includes:

[0061] S2: Input the observed tensor into the front end of the convolution to extract features and obtain shallow features.

[0062] observation tensor The input is fed into the convolutional front end for shallow feature extraction to obtain the feature map. :

[0063]

[0064] in, Indicates shallow features. Represents the ReLU activation function. Indicates batch normalization, This represents a two-dimensional convolution operation. This represents the trainable convolutional kernel parameters of the two-dimensional convolutional layer in the convolutional front end. Its specific values ​​are updated through backpropagation during model training. The convolutional front end is used to enhance locally stable features and suppress input distribution drift, providing a more robust feature representation for subsequent long-range correlation modeling.

[0065] S3: Using the method The main network, composed of multi-head self-attention residual blocks connected in series with axial windows, processes shallow features to obtain deep features.

[0066] shallow features Input to by indivual( This is a positive integer, which can be set according to the estimation accuracy, computational complexity, and deployment resource requirements; in some embodiments of the present invention, The range of values ​​is In the main network composed of cascaded multi-head self-attention residual blocks with axial windows, features are iteratively updated:

[0067]

[0068] in, For the first Multi-head self-attention residual block mapping with axial windows. Residual connections are used to stabilize the training process and preserve the original feature information.

[0069] Multi-head self-attention residual block of any axial window The data processing procedure is as follows:

[0070] like Figure 2As shown, this invention designs an axial window multi-head self-attention module. This module is used to model the long-range spatial correlation of a two-dimensional array under controllable complexity, and can combine noise-aware gating to constrain the window interaction strength. The specific implementation process is as follows: S31~S36:

[0071] S31: Adaptively generate noise-aware gating weights based on input features.

[0072] To suppress noise propagation through attention interactions under low SNR conditions, noise-aware gating weights are adaptively generated from input features within the residual block:

[0073]

[0074] in, Indicates the noise-aware gating weight. This represents the Sigmoid function. This represents a two-dimensional convolution operation. Input features representing axial window multi-head self-attention residual blocks , This represents the trainable convolutional kernel parameters of the two-dimensional convolutional layer in the noise-aware gating system; their specific values ​​are adaptively learned during training. Gating weights. Modulation of subsequent window masks or attention scores to reduce the intensity of unreliable interactions in noisy regions.

[0075] S32: Decompose the noise-sensing gating weights along the height axis and the width axis respectively to obtain the height axis sequence and the width axis sequence; concatenate the height axis sequence and the width axis sequence to obtain the axial sequence.

[0076] Will The two-dimensional attention computation is decomposed into two sets of one-dimensional sequence attention computations by performing decompositions along both the height and width axes. Taking the height axis as an example:

[0077]

[0078] The same applies to the width axis. By decomposing along the axis, the computational complexity of 2D global self-attention can be reduced while maintaining the ability to model long-range dependencies along the row / column directions.

[0079] Then, the height axis sequence and the width axis sequence are spliced ​​together to obtain the axial sequence.

[0080] S33: Divide the axial sequence into windows according to a preset window length to obtain multiple window sequences.

[0081] For axial sequences by window length Divide the window into sections (which may overlap or not). The sequence within any window is denoted as:

[0082]

[0083] Windowed interaction avoids the high computational overhead of global attention and reduces the risk of noise propagating over long distances at low SNR.

[0084] S34: Perform a linear mapping of multi-head attention for each window to obtain the key, value, and query vector under each attention head.

[0085] For each window Perform a linear mapping for multi-head attention, for the ... One point of focus is:

[0086]

[0087] in, For a trainable parameter matrix, , and They represent the first Each attention head has a key vector, a value vector, and a query vector.

[0088] S35: Each attention head calculates its output features based on the key, value, and query vector, and introduces a window mask.

[0089] Calculate the scaled dot product score for each attention head: Introduce a window mask and obtain the attention weights via Softmax, represented as:

[0090]

[0091] In some preferred embodiments of the present invention, gating is used. Modulate the window mask or attention score to suppress noise propagation:

[0092]

[0093] in, Represents the window mask matrix. Indicates the base mask. The adjustment factor representing the gate mask bias can be preset according to the noise level or the performance of the validation set; This represents the mapping from gating to mask bias.

[0094] In some other preferred embodiments of the invention, the attention weights are further modulated:

[0095]

[0096] in, Indicates the first Attention weights for each attention head Indicates the noise-aware gating weight. The dimension representing the attention head, This represents the window mask matrix.

[0097] The above modulation suppresses attentional interactions at locations with high noise and preserves effective long-range modeling at locations with low noise, thereby improving stability under low SNR conditions.

[0098] The attention head output features are processed to obtain the first... Each attention head output feature:

[0099]

[0100] S36: Concatenate the output features of each attention head and obtain the window attention output through linear mapping.

[0101] The output features of each attention head are concatenated along the channel dimension and then linearly mapped to obtain the attention output of a single window:

[0102]

[0103] in, This represents the multi-head attention splicing feature. Indicates the number of attention heads. This represents the trainable weight parameters of the output linear mapping layer. This represents the output features of a single window after multi-head attention computation.

[0104] S37: Concatenate the window attention outputs under all windows in the original window order and restore the two-dimensional feature layout to obtain the output of the axial window multi-head self-attention module;

[0105] Output characteristics for all windows The features are then stitched together in the original window order, folded back to the axial sequence position, and then restored to the two-dimensional feature layout to obtain the entire input feature set. The corresponding axial window multi-head self-attention module output:

[0106]

[0107] in, This indicates that all windows will be output. The operation of restoring the set to a two-dimensional feature layout according to the spatial order before window partitioning.

[0108] S38: The output of the axial window multi-head self-attention module is mapped by two-dimensional convolution, added to the input features, and then processed by ReLU to obtain the output features of the axial window multi-head self-attention residual block.

[0109] Within the residual block, Residual features are obtained through two-dimensional convolution mapping. Then Block input features The summation is then applied to the ReLU activation function to obtain the output characteristics of the residual block. :

[0110]

[0111]

[0112] in, Indicates the first Input features of multi-head self-attention residual blocks with axial windows This represents the residual features after being mapped by an axial window multi-head self-attention module and a two-dimensional convolution. This represents the trainable convolution kernel parameters of the two-dimensional convolutional layer within the residual block. The specific values ​​are obtained adaptively through backpropagation during model training. or This indicates the output characteristics of the residual block.

[0113] This invention employs an attention computation method combining axial decomposition (along the array height / width direction) and windowed interaction. This characterizes long-range spatial correlations on a two-dimensional array with controllable complexity, avoiding the high computational overhead of global self-attention in ultra-large-scale dimensions, and improving estimation stability under varying near / far-field path compositions. Finally, residual prediction is output through a projection head, and a residual reconstruction method is used to obtain the channel estimation result, thereby achieving robust recovery of low SNR channels in mixed fields, improving estimation accuracy, stability, and engineering feasibility.

[0114] S4: Input the deep features into the projection head for processing to obtain residual predictions.

[0115] deep features Input the projection head, output the residual prediction in the same dimension as the input. :

[0116]

[0117] in, Indicates residual prediction, This represents a linear mapping layer used to project features onto the observation tensor. Dimensionally consistent real / imaginary dual-channel residual representation; This represents a two-dimensional convolution operation. Indicates deep features, This represents the trainable convolution kernel parameters of the two-dimensional convolutional layer in the projector; its specific values ​​are automatically updated during the network training process.

[0118] S5: Residual reconstruction is performed based on residual prediction and observation tensor to obtain channel estimation results.

[0119] The residual output from the projection head With input observation tensor The final output is obtained by performing residual reconstruction. (or its equivalent representation), thus obtaining the final output estimate of the network:

[0120]

[0121] in, This represents the channel estimation result. Represents the observation tensor. This indicates residual prediction.

[0122] In summary, this invention, through an overall process of "convolutional front-end feature extraction—noise-aware prior generation—noise-constrained axial window multi-head self-attention modeling—projection head residual prediction—residual reconstruction output," achieves long-range spatial correlation characterization and weak feature enhancement of ELAA / XL-MIMO near-field and far-field hybrid propagation channels without relying on explicit near / far-field path numbers or other prior parameters. Axial decomposition and windowing interaction effectively reduce the computational and storage overhead of global modeling, while the modulation of attention interaction by the noise prior suppresses noise propagation and mismodeling under low signal-to-noise ratio (SNR) conditions. This improves the stability, robustness, and engineering feasibility of channel estimation under hybrid field conditions, demonstrating promising application prospects.

[0123] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A noise-aware axial window attention channel estimation method, characterized in that, Includes the following steps: S1: Acquire pilot observation signals and construct the observation tensor; S2: Input the observation tensor into the convolutional front end to extract features and obtain shallow features; S3: Using the method The main network, composed of multi-head self-attention residual blocks connected in series with axial windows, processes shallow features to obtain deep features. S4: Input the deep features into the projection head for processing to obtain residual prediction; S5: Residual reconstruction is performed based on residual prediction and observation tensor to obtain channel estimation results.

2. The noise-aware axial window attention channel estimation method according to claim 1, characterized in that, Step S1 specifically includes: The receiver acquires pilot observation signals in the pilot time slot and forms a frequency domain observation matrix. ; The frequency domain observation matrix is ​​split into real and imaginary parts and concatenated in the channel dimension to obtain the observation tensor.

3. The noise-aware axial window attention channel estimation method according to claim 1, characterized in that, In step S2, the shallow feature representation is obtained as follows: ； in, Indicates shallow features. Represents the ReLU activation function. Indicates batch normalization, This represents a two-dimensional convolution operation. Represents the observation tensor. This represents the trainable convolution kernel parameters of the two-dimensional convolutional layer in the convolutional front end.

4. The noise-aware axial window attention channel estimation method according to claim 1, characterized in that, In step S3, the data processing procedure for the axial window multi-head self-attention residual block includes: S31: Adaptively generate noise-aware gating weights based on input features; S32: Decompose the noise-sensing gating weights along the height axis and width axis respectively to obtain the height axis sequence and the width axis sequence; concatenate the height axis sequence and the width axis sequence to obtain the axial sequence; S33: Divide the axial sequence into multiple window sequences according to a preset window length; S34: Perform a linear mapping of multi-head attention for each window to obtain the key, value, and query vector under each attention head; S35: Each attention head calculates its output features based on the key, value, and query vector, and introduces a window mask; S36: Concatenate the output features of each attention head and obtain the window attention output through linear mapping; S37: Concatenate the window attention outputs under all windows in the original window order and restore the two-dimensional feature layout to obtain the output of the axial window multi-head self-attention module; S38: The output of the axial window multi-head self-attention module is mapped by two-dimensional convolution, added to the input features, and then processed by ReLU to obtain the output features of the axial window multi-head self-attention residual block.

5. The noise-aware axial window attention channel estimation method according to claim 4, characterized in that, The adaptive generation noise-aware gating weights are represented as follows: ； in, Indicates the noise-aware gating weight. This represents the Sigmoid function. This represents a two-dimensional convolution operation. This represents the input features of the axial window multi-head self-attention residual block. This represents the trainable convolution kernel parameters of a two-dimensional convolutional layer in a noise-aware gating system.

6. The noise-aware axial window attention channel estimation method according to claim 4, characterized in that, The calculated attention head output features are represented as follows: ； ； in, Indicates the first Each attention head outputs features. Indicates the first Attention weights for each attention head Indicates the noise-aware gating weight. Indicates the first A query vector with attention heads. Indicates the first A vector of values ​​for each attention head. The dimension representing the attention head, Indicates the first The key vector of an attention head, This represents the window mask matrix.

7. The noise-aware axial window attention channel estimation method according to claim 6, characterized in that, The formula for calculating the window mask matrix is: ； in, Represents the window mask matrix. Indicates the base mask. This represents the adjustment coefficient. This represents the mapping from gating to mask bias.

8. The noise-aware axial window attention channel estimation method according to claim 1, characterized in that, In step S4, the residual prediction is expressed as follows: ； in, Indicates residual prediction, Represents a linear mapping layer. This represents a two-dimensional convolution operation. Indicates deep features, This represents the trainable convolution kernel parameters of the two-dimensional convolutional layer in the projector.

9. The noise-aware axial window attention channel estimation method according to claim 1, characterized in that, In step S5, the residual reconstruction based on the residual prediction and observation tensor is expressed as follows: ； in, This represents the channel estimation result. Represents the observation tensor. This indicates residual prediction.

Images