A channel estimation method based on global super-resolution denoising

Abstract

The application relates to a channel estimation method based on global super-resolution denoising, and belongs to the technical field of communication. The method comprises the following steps: a global super-resolution denoising neural network is proposed for channel estimation; the input of the neural network is a channel estimation matrix at a pilot without interpolation processing; the matrix is low-dimensional data, and the operation complexity can be greatly reduced. The input data passes through the network, sequentially passes through a global information extraction module, a super-resolution residual error module and a denoising module, and finally a complete channel response estimation matrix after fine processing is obtained. When the global super-resolution denoising neural network is used for channel estimation, the method is not limited by the performance of a pre-interpolation algorithm, and has higher channel estimation accuracy compared with other algorithms.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, and relates to channel estimation in the field of communication, specifically to a method for channel estimation using a global super-resolution denoising neural network. Background Technology

[0002] In non-terrestrial network (NTN) communication scenarios, satellites can be broadly classified into three categories based on their orbital altitude: Low Earth Orbit (LEO), Medium Earth Orbit (MEO), and Geosynchronous Earth Orbit (GEO). In LEO satellite communication systems, satellites orbit the Earth at speeds of approximately 7000 m / s. Under such high speeds, the Doppler frequency shift generated by the satellite relative to terminals on Earth is significantly greater than that generated by terrestrial wireless communication systems. Therefore, compared to terrestrial systems, LEO communication systems possess more complex and challenging wireless channels. Table 1 shows the Doppler frequency shift under different satellite communication scenarios.

[0003] Table 1. Doppler Frequency Offset Range in Satellite Communication

[0004]

[0005] In scenarios where the propagation path between the transmitter and receiver is highly random, or in extremely harsh signal transmission environments, channel estimation at the receiver is crucial in Orthogonal Frequency Division Multiplexing (OFDM) systems. This determines whether the receiver can accurately demodulate the transmitted signal.

[0006] Traditional channel estimation methods in OFDM systems include Least Square (LS) estimation and Linear Minimum Mean Square Error (LMMSE). Traditional LS channel estimation first estimates the channel response at the pilot, then uses an interpolation algorithm to obtain the channel matrix of the complete time-frequency resource block. This method is computationally simple, but its estimation performance is poor because it does not consider the effects of channel noise and residual frequency offset. LMMSE channel estimation utilizes the linear independence of noise between channels to eliminate the influence of noise on channel estimation, resulting in a significant improvement in estimation performance. However, its calculation involves matrix inversion, leading to high computational complexity. In practical engineering, simplified algorithms are usually used, which affects LMMSE performance.

[0007] In the existing technology, there is a technique for channel estimation using a network model called ChannelNet. It treats the channel response matrix as a 2D low-resolution image. The input image is processed by image super-resolution (SRCNN) and denoising (DnCNN) networks to obtain a refined high-resolution image. However, since the input of the ChannelNet network needs to be interpolated to enlarge the channel response matrix to the required size, on the one hand, the performance of channel estimation will be affected by the limitations of the interpolation algorithm. On the other hand, the input of the neural network is high-dimensional data after interpolation, which significantly increases the computational complexity. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide a channel estimation method based on global super-resolution denoising. The method uses a global super-resolution denoising neural network to process the channel estimation matrix, and performs global information extraction, super-resolution residual and denoising to obtain a complete channel response estimation matrix, thereby reducing computational complexity and improving channel estimation accuracy.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A channel estimation method based on global super-resolution denoising is proposed. The method first configures an OFDM time-frequency resource grid to calculate the channel estimation matrix at the pilot. Then, a global super-resolution denoising neural network is constructed to perform channel estimation. The input of the neural network is the uninterpolated channel estimation matrix at the pilot.

[0011] Furthermore, the channel estimation matrix at the pilot is calculated using the least squares method. Specifically, resource blocks are first configured according to the 3GPP 5G standard protocol. Assuming a resource block size of K×N, the channel response matrix of that resource block is expressed as follows: Where K represents the total number of subcarriers in an OFDM symbol, and N represents the total number of OFDM symbols in a resource block; let P represent the number of pilots carried by a single OFDM symbol in a resource block, and S represent the total number of OFDM symbols carrying pilots. Then a resource block contains P×S pilots, and the channel response matrix at the pilot is expressed as: Then use the least squares algorithm to apply H P The channel estimation matrix at the pilot frequency is obtained by estimation.

[0012] Furthermore, the global super-resolution denoising neural network includes a global information extraction module, a super-resolution residual module, and a denoising module. The channel estimation matrix at the pilot signal first enters the global information extraction module for channel feature extraction and statistics, and dimensionality reduction of the channel data. The dimensionality-reduced channel data serves as the input to the super-resolution residual module, which processes the data to obtain the channel estimation matrix of the complete resource block. This complete resource block channel estimation matrix serves as the input to the denoising module, further filtering out noise from the channel data and improving the accuracy of channel estimation.

[0013] Furthermore, the global information extraction module includes a feature extraction layer I, a global average pooling layer, a compression layer I, an expansion layer I, a multiplication layer, and a compression layer II connected in sequence. The channel estimation matrix is ​​fed into the feature extraction layer I for channel feature extraction. The global average pooling layer performs feature statistics on the channel data through global average pooling operations. The compression layer I and the expansion layer I perform dimensionality reduction and dimensionality increase operations on the channel data, respectively. The multiplication layer is used to merge and multiply the outputs of the feature extraction layer I and the expansion layer I to restore the original channel data dimension. The compression layer II is used to reduce the dimensionality of the channel data output by the multiplication layer.

[0014] Furthermore, the super-resolution residual module includes a feature extraction layer II, a compression layer III, an aggregation layer, an extension layer II, a nonlinear mapping layer, a residual layer, and a deconvolution layer connected in sequence. The feature extraction layer II is used to extract features from the channel data output by the global information extraction module. The compression layer III is used to reduce the dimensionality of the channel data. The aggregation layer is used to extract channel data features from multiple perspectives and concatenate features of multiple different dimensions before outputting them to the extension layer II. The extension layer II performs dimensionality upscaling on the channel data and then performs nonlinear mapping on the channel data through the nonlinear mapping layer. The residual layer is used to calculate the residual between the output of the global information extraction module and the nonlinear mapping layer, accelerating the convergence speed of the neural network model. The output of the residual layer is fed into the deconvolution layer for upsampling to obtain the channel estimation matrix of the complete resource block.

[0015] Furthermore, the denoising module includes three different types of layers: Conv+PReLU, Conv+BN+PReLU, and Conv, as well as a difference structure. The three layers, Conv+PReLU, Conv+BN+PReLU, and Conv, are used to extract the noise contained in the channel estimation matrix of the complete resource block. The difference structure is used to filter out the extracted noise from the channel estimation matrix of the complete resource block to obtain clean channel data.

[0016] The beneficial effects of this invention are as follows:

[0017] (1) Compared with the traditional LS algorithm and the Practical Channel Estimation (PCE) algorithm provided by MATLAB, this invention uses deep learning to learn the general rules and characteristics of data from a large number of channel-related datasets, which has a strong generalization ability and is better than the LS and PCE algorithms.

[0018] (2) As the number of layers in the neural network model increases, the residual structure is used to accelerate the convergence speed of the model, prevent gradient vanishing and gradient explosion, and enable the model to better learn the mapping relationship between input and output, thereby further improving the model estimation performance.

[0019] (3) Compared with ChannelNet (SRCNN+DnCNN), this invention is not limited by the pre-interpolation processing algorithm, and therefore has better performance.

[0020] (4) Compared with ChannelNet (SRCNN+DnCNN) which requires pre-interpolation processing, the global super-resolution denoising neural network constructed in this invention has higher estimation accuracy and significant advantages in computational complexity because the input is only the channel estimate at the pilot.

[0021] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0023] Figure 1 This is a model architecture diagram of the global super-resolution denoising neural network proposed in this invention;

[0024] Figure 2 This is a schematic diagram of the aggregation layer architecture;

[0025] Figure 3 This describes the channel estimation process based on global super-resolution denoising.

[0026] Figure 4 A schematic diagram illustrating the working principle of a global super-resolution denoising neural network;

[0027] Figure 5 The verification flowchart for the global super-resolution denoising neural network;

[0028] Figure 6Generate a flowchart for training data in Matlab;

[0029] Figure 7 This is a schematic diagram illustrating the training effect of a neural network.

[0030] Figure 8 The channel estimation results for each algorithm are shown at a signal-to-noise ratio of 10dB.

[0031] Figure 9 The graph shows the algorithm performance verification results in four NTN scenarios at a frequency offset of 10kHz.

[0032] Figure 10 The figure shows the algorithm performance verification results in four NTN scenarios under a 20kHz frequency offset. Detailed Implementation

[0033] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0034] This invention addresses the problems of existing channel estimation algorithms by proposing a Global Super-Resolution Denoising Neural Network (GIResSRDnNet). The input to this neural network is the uninterpolated channel estimation matrix from the pilot signals, which is low-dimensional data, significantly reducing computational complexity. The input data is sequentially processed through global information extraction, super-resolution residuals, and denoising modules within the neural network, ultimately yielding a refined and complete channel response estimation matrix. Furthermore, this invention utilizes GIRESSRDnNet for channel estimation, unaffected by the performance limitations of pre-interpolation algorithms, resulting in higher channel estimation accuracy compared to other algorithms.

[0035] The proposed GIRESSRDnNet consists of a global information extraction module, a super-resolution residual module, and a denoising module, as detailed below. Figure 1 As shown.

[0036] Compared to traditional neural networks, the global information extraction module has a stronger feature extraction capability for the channel response matrix at the input pilot, enabling it to better distinguish different "feature maps." This module also considers both local and global information, further improving network performance. The global information extraction module sequentially includes a feature extraction layer, a global average pooling layer, a compression layer, an expansion layer, a multiplication layer, and another compression layer. The functions of each layer are as follows: the feature extraction layer extracts channel features from the input data; the global average pooling operation performs feature statistics on the channel data; the compression and expansion layers perform dimensionality reduction and dimensionality increase operations on the channel data, respectively; the outputs of the feature extraction and expansion layers are combined and multiplied to restore the original channel data dimension; and the compression layer reduces the dimensionality of the channel data and uses its output as the input to the super-resolution residual module.

[0037] The super-resolution residual module comprises seven layers: feature extraction, compression, aggregation, expansion, nonlinear mapping, residual, and deconvolution. This module yields the channel estimation matrix for the complete resource block. The functions of each layer are as follows: the feature extraction layer extracts features from the input channel data; the compression layer reduces the dimensionality of the channel data; the aggregation layer has the following structure... Figure 2 As shown, it extracts channel data features from multiple scales through four different paths, concatenates the different dimensional "feature maps" obtained from the four paths, and inputs them into the extension layer; the extension layer completes the dimensionality upscaling of the channel data; the nonlinear mapping layer realizes the nonlinear mapping of the data; the residual structure accelerates the model convergence speed, prevents gradient vanishing and exploding, and enables the model to better learn the mapping relationship between input and output; the deconvolution layer realizes the upsampling processing of the channel data, and finally obtains the channel response estimate of the complete resource block.

[0038] The residual structure constructed between the output of the nonlinear mapping layer in the super-resolution residual module and the original input data can be represented as:

[0039] y = F(x) + x

[0040] Where F(x) represents the output of the nonlinear mapping layer, x represents the original input data of the super-resolution residual module, and y represents the input of the deconvolution layer.

[0041] The denoising module is used to filter out noise in the channel data. The complete channel response estimate obtained after upsampling contains some noise; therefore, the DnCNN denoising module is used to further filter out channel data noise and improve network estimation performance. The denoising module contains three different types of layers: Conv+PReLU, Conv+BN+PReLU, and Conv, and a difference structure. The specific functions of each part are as follows: the three different types of layers, Conv+PReLU, Conv+BN+PReLU, and Conv, can extract the noise contained in the input channel data; the difference structure filters out noise from the noisy input channel data to obtain the final clean channel data.

[0042] Based on the aforementioned GIRESSRDnNet, this invention provides a detailed explanation of a channel estimation method based on global super-resolution denoising through an embodiment of the present invention, such as... Figure 3 and Figure 4 As shown, the method includes:

[0043] Step 1: OFDM Time-Frequency Resource Grid Configuration. In this embodiment, a resource block contains 14 OFDM symbols, and each OFDM symbol contains 612 subcarriers. The Demodulation Reference Signal (DMRS) type is configured as Type 1, and the number of post-DMRS is 3. Also, within the 14 OFDM symbol resource block, 4 symbols carry pilot information, and each pilot symbol has 306 subcarriers used to carry pilot information. That is, the size of a resource block is 612×14, and the channel response matrix of a resource block is... A resource block contains 306×4 pilots, and the channel response matrix at the pilot is represented as follows:

[0044] Step 2: Use the LS algorithm to obtain the channel estimate at the pilot. This value is then used as the input to the GIRESSRDnNet neural network.

[0045] Step 3 The data then enters the global information extraction module within the neural network. After passing through the global information extraction module, the data dimension remains 306×4.

[0046] Step 4: Use the output of the global information extraction module as the input of the super-resolution residual module. The super-resolution residual module contains a total of 7 layers. After passing through this module, the channel estimation matrix of the complete resource block will be obtained. That is, after passing through this module, the dimension of the input data changes from 306×4 to 612×14.

[0047] Step 5: Use the output of the super-resolution residual module as the input of the denoising module. The input data will not change the data dimension after passing through the denoising module, and the data size will still be 612×14.

[0048] Step Six: After global information extraction, super-resolution residual, and denoising modules, the channel estimation matrix of the complete resource block is obtained as follows: That is when After going through these three modules in sequence, you can... Recovery By comparing the channel estimates The difference between the channel estimation method and the true channel value H is used to measure the performance of this channel estimation method.

[0049] This embodiment is executed in four stages, such as Figure 5 As shown, this embodiment uses Matlab and PyTorch. In Matlab, training and testing data are generated in units of resource blocks. After obtaining the data, the training and testing datasets are input into the PyTorch neural network to train the network model. Finally, the saved neural network model is tested in Matlab to verify its effectiveness. The process includes: Stage 1: Matlab generates training data; Stage 2: PyTorch builds the GIRESSRDnNet model; Stage 3: GIRESSRDnNet model training; Stage 4: GIRESSRDnNet model testing.

[0050] Phase 1: Generating Training Data with Matlab

[0051] During the model training phase, a training set was first generated using MATLAB. To improve the generalizability of the network model, this training set consisted of mixed data with a discrete and uniformly distributed signal-to-noise ratio (SNR) of 0–25 dB. Simultaneously, different frequency offset data, such as 10 kHz and 20 kHz, were incorporated to account for Doppler frequency offset. Then, the LS channel estimates at the pilot frequencies were used... As input to GIRESSRDnNet, the perfect channel estimation function (nrPerfectChannelEstimate) provided by MATLAB will be used to calculate the... The actual channel response values ​​serve as the label values ​​for model training. Since current neural network models do not support complex number operations, the training data must be divided into real and imaginary parts. The specific process for generating the training data is as follows: Figure 6 As shown.

[0052] Phase Two: Building the GIResSRDnNet Model with PyTorch

[0053] The real and imaginary part datasets are fed together into the GIRESSRDnNet model for training. In this embodiment, the hyperparameter configurations of each module of the GIRESSRDnNet network model are as follows:

[0054] 1) The feature extraction layer of the global information extraction module consists of a convolutional layer with 32 filters and 3×3 kernels, and two convolutional layers with the same configuration, each with 64 filters and 3×3 kernels, to extract channel features from the input data. The compression layer consists of a convolutional layer with 32 filters and 3×3 kernels to reduce the dimensionality of the channel data. The expansion layer consists of a convolutional layer with 64 filters and 1×1 kernels to increase the dimensionality of the channel data. The last compression layer of this module consists of a convolutional layer with one filter and 3×3 kernels to reduce the dimensionality of the channel data and use its output as the input to the super-resolution residual module. The specific configuration of each layer of the global information extraction module is shown in Table 2.

[0055] Table 2 Hyperparameter Configuration of Global Information Extraction Module

[0056]

[0057] 2) The parameter configurations of each layer in the super-resolution residual module are shown in Table 3. The feature extraction layer consists of a convolutional layer with 32 filters and 3×3 kernels, extracting features from the input channel data. The compression layer consists of a convolutional layer with 16 filters and 3×3 kernels, reducing the dimensionality of the channel data and decreasing the number of data channels from 32 to 16. The hyperparameter configuration of the aggregation layer is shown in Table 4. This layer extracts channel data features from multiple scales through four different paths, concatenates the different dimensional "feature maps" obtained from the four paths, and inputs them into the expansion layer. The expansion layer consists of a convolutional layer with 56 filters and 1×1 kernels, performing dimensionality upscaling of the channel data. The nonlinear mapping layer consists of two identical convolutional layers with 32 filters and 3×3 kernels, performing nonlinear mapping of the data. The deconvolution layer consists of a convolutional layer with one filter and 2×2 kernels, performing upsampling of the channel data to obtain the channel response estimate of the complete resource block.

[0058] Table 3 Hyperparameter Configuration of the Super-Residual Module

[0059]

[0060] Table 4 Hyperparameter Configuration of Aggregation Layer

[0061]

[0062]

[0063] 3) The parameter configurations of each layer in the denoising module are shown in Table 5. The Conv+PReLU layer consists of a convolutional layer with 32 filters and 3×3 convolutional kernels. Following the Conv+PReLU layer are three Conv+BN+PReLU layers with the same hyperparameter configuration, each consisting of a convolutional layer with 64 filters and 3×3 convolutional kernels. Then, a Conv+BN+PReLU layer consisting of 32 filters and 1×1 convolutional kernels is cascaded. The final layer of the denoising module is a Conv+BN+PReLU layer consisting of a single filter and a convolutional layer with 3×3 convolutional kernels.

[0064] Table 5 Hyperparameter Configuration of Noise Reduction Module

[0065] Noise reduction module Layer configuration Conv+PReLU 32×3×3 Conv+BN+PReLU 64×3×3 Conv+BN+PReLU 32×1×1 Conv 1×3×3

[0066] In this embodiment, the activation function chosen is PReLU, as shown in formula (1). Compared with ReLU, PReLU has greater flexibility and adaptability. During network training, PReLU can adaptively adjust its nonlinearity according to different data distributions.

[0067]

[0068] The parameter 'a' here can be continuously learned and changed; it is not a fixed value. This is the key advantage of the PReLU function.

[0069] Phase 3: Training the GIResSRDnNet model

[0070] The input to the GIResSRDnNet model is the channel estimate at the pilot location obtained using LS. Then, the actual channel value is used as the label to train the GIResSRDnNet neural network. The training and testing datasets contain a total of 38,000 samples, with a training set to test set ratio of 7:3. The optimizer is Adam, and the specific training parameters are shown in Table 6.

[0071] Table 6 Training Parameters

[0072]

[0073]

[0074] In this embodiment, the training dataset can be represented as:

[0075]

[0076] in H represents the i-th training sample. i This represents the label corresponding to the i-th training sample.

[0077] Similarly, the test dataset can be represented as:

[0078]

[0079] in, and H i Let represent the i-th test sample and its corresponding label, respectively. The performance of the trained neural network model is validated using the test set data.

[0080] Representing the neural network GIRESSRDnNet as f(·) yields:

[0081]

[0082] in and These represent the input and output of the GIRESSRDnNet network, respectively. Mean Squared Error (MSE) is used as the loss function, as shown below:

[0083]

[0084] Where n represents the total number of samples in a single training batch, H s This represents the true value of the s-th channel response in this batch, i.e., the tag value. This represents the estimated value of the s-th channel response obtained using a neural network.

[0085] The primary function of the loss function is to guide the training process of the network model. When the MSE loss function is minimized, the corresponding model parameters θ represent the optimal values. Throughout the training process, by continuously adjusting the network model parameters to reduce the loss function, the model gradually learns the inherent characteristics and patterns of the data, thereby making more accurate predictions. The final training result of this model is as follows: Figure 7 As shown.

[0086] Phase 4: GIRESSRDnNet Model Testing

[0087] The neural network model trained using PyTorch is saved, and finally, the channel estimation performance of the neural network model is tested in MATLAB under NTN-TDL-A, B, C, and E scenarios.

[0088] According to the 3GPP 5G standard protocol, it is necessary to simulate the single-transmit and single-receive scenarios of NTN-TDL A, B, C, and E under a signal-to-noise ratio of 0 to 25 dB, and to perform simulation verification based on this. The specific simulation parameter configuration is shown in Table 7.

[0089] Table 7 Simulation Parameters

[0090]

[0091] To simulate a low-orbit satellite communication system, according to formula (6), the Doppler parameters corresponding to a Doppler frequency offset of 20KHz are configured in Table 8. The reasons why the Doppler frequency offset will seriously affect the performance of the communication system are shown in formulas (7) and (8).

[0092]

[0093] In the formula, x(·) and y(·) represent the transmitted signal and the received signal, respectively, and w(·) represents Gaussian white noise.

[0094] Table 8. 20kHz Doppler parameters

[0095]

[0096]

[0097] To visually demonstrate the advantages of GIRESSRDnNet compared to other algorithms, Figure 8 The differences between the channel response estimates and the actual channel response values ​​obtained by various algorithms in the NTN-TDL-E scenario, with a signal-to-noise ratio of 10dB and a Doppler frequency offset of 10kHz, are visualized. It can be seen that in terms of MSE, the traditional LS algorithm has an MSE of 0.2104, the PCE algorithm has an MSE of 0.0995, the SRCNN+DnCNN algorithm requiring pre-interpolation has an MSE of 0.0538, and the proposed GIRESSRDnNet method has an MSE of 0.0248. The MSE comparison clearly shows the advantage of the GIRESSRDnNet method over other methods. Furthermore, the texture differences between the estimated and actual channel response values ​​also show that the texture map corresponding to the GIRESSRDnNet method is closest to the actual channel response texture map. Therefore, the advantages of the proposed algorithm are clearly evident at both the MSE and texture map levels.

[0098] To verify the link performance and universality of the GIRESSRDnNet algorithm, the channel estimation performance of the GIRESSRDnNet method, the traditional LS algorithm, the PCE algorithm, and the SRCNN+DnCNN method were compared in scenarios with Doppler frequency offsets of 10kHz and 20kHz. To further verify the advantages of the proposed algorithm, since the LS and PCE algorithms lack frequency offset suppression, the classic PSS frequency offset estimation algorithm was first used to pre-compensate the resource blocks for frequency offset when verifying their performance. The channel estimates are represented by mean square error in the simulation. The difference between the channel estimate and the true channel value H is used to measure the accuracy of the channel estimation.

[0099] Scenario 1: Doppler frequency deviation of 10kHz

[0100] exist Figure 9 The paper compares the MSE values ​​of four algorithms under different NTN environments with a Doppler frequency offset of 10 kHz. It can be seen that in this scenario, the traditional LS algorithm has the worst channel estimation performance because it does not consider the influence of channel noise. The PCE algorithm, which performs denoising on top of LS, outperforms LS. The SRCNN+DnCNN algorithm, which requires pre-interpolation, and the proposed GIRESSRDnNet algorithm, both employ deep learning to learn the general patterns and features of data from a large amount of channel-related datasets, possess strong generalization ability and outperform LS and PCE algorithms. Furthermore, the GIRESSRDnNet algorithm is not limited by the pre-interpolation algorithm compared to SRCNN+DnCNN, thus offering superior performance. Since the input to the GIRESSRDnNet network model is only the channel estimate at the pilot, it not only has higher estimation accuracy but also significant advantages in computational complexity.

[0101] Under different NTN channel environments, the MSE performance of the GIRESSRDnNet algorithm is improved by 7.83–8.83 dB compared to the LS algorithm; by 5.33–6.06 dB compared to the PCE algorithm; and by 2.73–3.14 dB compared to the SRCNN+DnCNN algorithm.

[0102] Scenario 2: Doppler frequency deviation of 20kHz

[0103] Similarly, Figure 10 It can be seen that in various NTN scenarios at 20kHz, the GIResSRDnNet algorithm has significant advantages over LS, PCE, and SRCNN+DnCNN. The MSE performance of the GIResSRDnNet algorithm is improved by 3.37–7.04 dB compared to the LS algorithm; by 2.11–5.75 dB compared to PCE; and by 1.37–4.40 dB compared to SRCNN+DnCNN.

[0104] Using the tic and toc functions provided by MATLAB, the time required to run one round of the SRCNN+DnCNN algorithm and the GIRESSRDnNet algorithm, which require pre-interpolation processing, was recorded on the same computer. The results are shown in Table 9. It can be seen that the algorithm proposed in this invention can reduce the computation time by about 42% compared with the SRCNN+DnCNN algorithm.

[0105] Table 9 Comparison of Algorithm Running Time

[0106]

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A channel estimation method based on global super-resolution denoising, characterized in that: This method constructs a global super-resolution denoising neural network for channel estimation. This neural network includes a global information extraction module, a super-resolution residual module, and a denoising module. The input to the neural network is the channel estimation matrix at the pilot frequencies before interpolation. The channel estimation matrix first enters the global information extraction module for channel feature extraction and statistics, and then performs dimensionality reduction on the channel data. The dimensionality-reduced channel data serves as the input to the super-resolution residual module, which processes the data to obtain the channel estimation matrix for the complete resource block. This complete resource block channel estimation matrix then serves as the input to the denoising module, further filtering out noise from the channel data and improving the accuracy of channel estimation. The method for obtaining the channel estimation matrix at the pilot frequencies includes configuring resource blocks according to the 3GPP 5G standard protocol, assuming a resource block size of... The channel response matrix of this resource block is then expressed as: , K This represents the total number of subcarriers contained in an OFDM symbol. N Represents the total number of OFDM symbols in a resource block; let P This indicates the number of pilot signals carried by a single OFDM symbol in a resource block. S This indicates the total number of OFDM symbols carrying pilot signals; therefore, a resource block contains a total of [number missing]. There are 1 pilot signal, and the channel response matrix at each pilot signal is expressed as follows: Then use the least squares algorithm to... The channel estimation matrix at the pilot frequency is obtained by estimation. ; The global information extraction module includes a feature extraction layer I, a global average pooling layer, a compression layer I, an expansion layer I, a multiplication layer, and a compression layer II connected in sequence. The channel estimation matrix is ​​fed into the feature extraction layer I for channel feature extraction. The global average pooling layer performs feature statistics on the channel data through global average pooling operations. The compression layer I and the expansion layer I perform dimensionality reduction and dimensionality increase operations on the channel data, respectively. The multiplication layer is used to merge and multiply the outputs of the feature extraction layer I and the expansion layer I to restore the original channel data dimension. The compression layer II is used to reduce the dimensionality of the channel data output by the multiplication layer. The super-resolution residual module comprises a feature extraction layer II, a compression layer III, an aggregation layer, an extension layer II, a nonlinear mapping layer, a residual layer, and a deconvolution layer connected in sequence. The feature extraction layer II extracts features from the channel data output by the global information extraction module. The compression layer III performs dimensionality reduction on the channel data. The aggregation layer extracts channel data features from multiple perspectives and concatenates these features before outputting them to the extension layer II. The extension layer II performs dimensionality upscaling on the channel data and then performs nonlinear mapping on the channel data through the nonlinear mapping layer. The residual layer calculates the residual between the outputs of the global information extraction module and the nonlinear mapping layer, accelerating the convergence speed of the neural network model. The output of the residual layer is then fed into the deconvolution layer for upsampling to obtain the channel estimation matrix of the complete resource block. The denoising module includes three different types of layers: Conv+PReLU, Conv+BN+PReLU, and Conv, as well as a difference structure. The three layers, Conv+PReLU, Conv+BN+PReLU, and Conv, are used to extract the noise contained in the channel estimation matrix of the complete resource block. The difference structure is used to filter out the extracted noise from the channel estimation matrix of the complete resource block to obtain clean channel data.

Images