CN-122027404-A - OFDM channel estimation method combining super-resolution and denoising joint model

CN122027404ACN 122027404 ACN122027404 ACN 122027404ACN-122027404-A

Abstract

The invention relates to the technical field of wireless communication, in particular to an OFDM channel estimation method combining a super-resolution and denoising joint model. The method comprises the steps of generating a data set through a constructed multipath Rayleigh fading channel model, converting a channel estimation problem in an OFDM system into an image super-resolution problem, constructing an improved super-resolution module SRCNN, constructing a denoising module DnCNN in a residual error learning mode to predict noise between output of a SRCNN module and a real channel, connecting the super-resolution module and the denoising module through a feature fusion layer to perform information sharing, constructing a joint model, optimizing parameters of the super-resolution module and the denoising module through an end-to-end training strategy, simulating and verifying, setting multiple groups of signal-to-noise ratio parameters through a mean square error evaluation algorithm performance, and verifying performances of different models under different signal-to-noise ratios.

Inventors

LI JUN
JIANG XINYAN
LIN FEI
SHI JUN
HOU MENG
MA GEN

Assignees

齐鲁工业大学(山东省科学院)

Dates

Publication Date: 20260512
Application Date: 20260416

Claims (9)

1. An OFDM channel estimation method combining a super-and denoising joint model, the method comprising: Step 1, generating a data set through a constructed multipath Rayleigh fading channel model; Step 2, according to the step1, converting the channel estimation problem in the orthogonal frequency division multiplexing OFDM system into an image super-resolution problem; step 3, based on step 2, constructing an improved super-resolution module SRCNN for reconstructing a high-resolution image from a low-resolution image; Step 4, constructing a denoising module DnCNN by adopting a residual error learning mode to predict noise between the output of the SRCNN module and a real channel; step 5, connecting the super-resolution module and the denoising module by adopting a feature fusion layer, carrying out information sharing, and constructing a joint model; step 6, optimizing parameters of the super-resolution module and the denoising module by adopting an end-to-end training strategy; and 7, simulation verification, namely evaluating algorithm performance through a mean square error MSE, setting a plurality of groups of signal-to-noise ratio (SNR) parameters, and verifying the performance of different models under different signal-to-noise ratios.
2. The method according to claim 1, wherein step 1 comprises the steps of multipath Rayleigh fading channel model with path gain following complex Gaussian distribution, and introducing exponentially decaying time delay and random Doppler shift to simulate frequency selective and time selective fading; setting system parameters, namely setting the number of subcarriers of an OFDM system as 72 and the number of symbols as 14, wherein the pilot pattern is uniformly distributed, the number of pilot subcarriers is 24, and the number of pilot symbols is 14; Step 12, generating multipath channels, wherein the number P of paths is randomly generated, and 3 to 10 propagation paths are randomly generated for each channel sample, wherein the delay value of each path obeys the exponential distribution, and the probability density function is as follows: ; Wherein, the Τ is the time delay; The delay tau of each path obeys the exponential distribution with the average value of 0.2, the generated delay value is subjected to sorting and normalization processing, the maximum delay value is ensured to be 1, and the expression is as follows: ; the complex gain g of the path follows a standard complex gaussian distribution, wherein both the real and imaginary parts follow an independent gaussian distribution with a mean of 0 and a variance of 1, and is multiplied by an exponential decay factor The Doppler frequency shift v of each path is uniformly and randomly generated within the range of [0, 0.1] to simulate the time variation of a channel; Step 13, calculating the channel response, namely, calculating the frequency response H [ f, t ] of the channel for each subcarrier f and each symbol t, wherein the expression is as follows: H ; wherein P is the number of paths, For the number of sub-carriers, the value is 72, The number of symbols is 14; Step 14, adding noise, namely multiplying the generated ideal channel response h with a random QPSK pilot frequency symbol x to obtain a noiseless receiving signal, calculating noise power according to a set signal-to-noise ratio, and adding complex Gaussian white noise N to obtain a final receiving signal Y; And 15, dividing the data set, namely repeating the process to generate 2800 samples, and randomly dividing the data set into a training set, a verification set and a test set according to the proportion of 2000:400:400.
3. The method according to claim 1, wherein the step 2 comprises: defining a time-frequency channel response matrix of an Orthogonal Frequency Division Multiplexing (OFDM) system as a two-dimensional image, defining an initial channel estimation obtained through pilot symbols as a low-resolution image, and defining a real channel response corresponding to the low-resolution image; Step 21, obtaining initial channel response at pilot position by least squares estimation LS by ignoring noise term Dividing the received signal by the transmitted known pilot symbols by: ; Wherein, the A pilot position index set; for the pilot signal at the transmitting end, For the pilot signal received by the receiving end, Initial channel response at pilot position obtained by LS estimation, namely initial channel estimation value; Step 22, two-dimensional interpolation, adopting bicubic interpolation method to make initial channel response of pilot frequency position Interpolation is carried out on the whole time-frequency grid to obtain an initial channel estimation matrix H; step 23, constructing an input image, separating the real part and the imaginary part of the complex-valued initial channel estimation matrix H as two channels, and jointly forming a low-resolution image X with the size of 72 multiplied by 14 multiplied by 2; the LS estimation value of the pilot frequency position is expanded to the whole time-frequency grid through interpolation, a bicubic interpolation method is adopted to respectively interpolate the real part and the imaginary part of the channel response, and after interpolation is completed, the complex channel matrix is converted into a two-channel image format so as to convert the channel estimation problem into an image super-resolution problem.
4. The method according to claim 1, wherein the step 3 comprises: the improved super resolution module SRCNN includes three core stages of operation: Step 31, feature extraction layer Feature extraction, which activates the input low resolution image by 9×9 convolution, 256 filters, batch normalizes BN and ReLU, extracts overlapping image blocks and represents each block as a high-dimensional feature vector, expressed as: ; Wherein, the Convolution weights of 9×9×256, BN for batch normalization operations, reLU for activation functions; Step 32, nonlinear mapping layer Non-LINEAR MAPPING, which nonlinearly maps the high-dimensional feature vector output by the first layer to another high-dimensional feature vector to learn nonlinear combination among different features, and adopts two convolution layers to strengthen mapping capability in order to deepen the mapping process in the middle, wherein the expression is: ; ; Wherein, the Is 5×5×128 convolution weights; Is 5×5×64 convolution weights; one of the two convolution layers sequentially comprises 5×5 convolutions, 128 filters and batch normalized BN and ReLU activation, and the other of the two convolution layers sequentially comprises 5×5 convolutions, 64 filters and batch normalized BN and ReLU activation; and 33, reconstructing layer Reconstruction, namely adopting 5×5 convolution, 2 filters, linearly activating, outputting 2 channels, reconstructing the information of the feature domain back to the image domain corresponding to the channel response of the real part and the imaginary part after Reconstruction, and aggregating the front high-dimensional features by the Reconstruction layer to generate a final high-resolution image, namely a noise image.
5. The method according to claim 1, wherein the structure of the denoising module DnCNN in step 4 comprises: Step 41, the first layer is an input layer, the structure of the first layer is 3×3 convolution, 128 filters are sequentially adopted, and batch normalization BN and ReLU are activated; Step 42, the middle 15 layers are deep feature extraction layers, the structures of which are 3×3 convolutions, 128 filters, and batch normalization BN and ReLU activation; Step 43, the last layer is an output layer, the structure of the last layer is 3×3 convolution, 2 filters are activated linearly, the number of output channels is 2, the real part and the imaginary part of the corresponding noise residual are corresponding, and a final denoising result is obtained through residual subtraction; The noise image output by the super resolution module SRCNN is used as input of the denoising module DnCNN, a 17-layer depth network is adopted to increase network depth, 128 filters are used for all hidden layers in the middle, L2 regularization constraint weights are used in convolution layers to improve model generalization capability, residual images, namely prediction noise, are obtained by an output layer, and the prediction noise output is subtracted by input through residual learning to obtain the denoised image.
6. The method according to claim 1, wherein the step 5 comprises: step 51, an input layer receives a low resolution image; Step 52, SRCNN, extracting features of the model, namely outputting a final reconstruction result, and reserving a feature map of a middle layer for subsequent fusion; Step 53, feature fusion layer a. Multi-level feature extraction, wherein feature graphs are extracted from different depths of SRCNN modules, and include shallow features and deep features; b. cross-module feature transfer, namely directly transferring the features of the SRCNN modules to corresponding layers of the DnCNN modules through jump connection; c. the feature fusion operation, in which features of different sources are fused by adopting a channel splicing mode; d. self-adaptive weight learning, namely automatically learning importance weights of different feature sources to realize self-adaptive fusion; and step 54, a denoising branch of DnCNN models, which takes the fused features as input.
7. The method according to claim 1, wherein the step 6 comprises: And an end-to-end joint training mechanism is adopted, and a single loss function is utilized for end-to-end optimization, wherein the expression is as follows: ; Wherein, the Representing a loss function, N represents the number of training samples in one Batch, Representing the real channel matrix for i samples, Representing the estimated channel matrix for the i-th sample, The Frobenius norm square representing the matrix, i.e. the sum of all the element squares; Sets of parameter values representing SRCNN and DnCNN modules, respectively; and in the process of counter propagation, the gradient simultaneously flows through two modules, namely SRCNN module and DnCNN module by adopting a gradient co-propagation mechanism, wherein the expression is as follows: ; ; wherein L represents a loss function, Representing the gradient of the loss function to the joint output, Representing the final output of the joint model, The output of the SRCNN module is represented, Representing DnCNN the noise predicted by the module.
8. A computer readable storage medium, characterized in that the computer readable storage medium comprises a stored program, wherein the program when run controls a device in which the computer readable storage medium is located to perform the OFDM channel estimation method combining the super and denoising joint model according to any one of claims 1 to 7.
9. An electronic device comprising one or more processors, memory, and one or more computer programs, wherein the one or more computer programs are stored in the memory, the one or more computer programs comprising instructions that, when executed by the device, cause the device to perform the method of OFDM channel estimation incorporating the joint model of superresolution and denoising of any of claims 1 to 7.

Description

OFDM channel estimation method combining super-resolution and denoising joint model Technical Field The invention relates to the technical field of wireless communication, in particular to an OFDM channel estimation method combining an ultra-wideband and denoising joint model. Background With the wide application of Orthogonal Frequency Division Multiplexing (OFDM) technology in 5G-Advanced and 6G communication systems, high-precision channel estimation becomes a key challenge for improving the spectral efficiency and reliability of the system. The OFDM system effectively resists frequency selective fading through frequency domain multi-carrier transmission, but under the high mobility scene and the low signal to noise ratio environment, the traditional channel estimation method has difficulty in meeting the requirements of precision and instantaneity. Especially in a large-scale MIMO-OFDM system, the dimension of a channel matrix is obviously increased, and the existing method faces the problems of high computational complexity and accumulated estimation errors. In the prior art, the traditional least square estimation (LS) combined linear interpolation method is simple and easy to realize, but is sensitive to noise and ignores the channel structure characteristics, so that the performance is rapidly deteriorated under low signal to noise ratio, while the cascade deep learning-based method can improve the estimation precision, but has the error propagation problem caused by two-stage training, and the super-resolution reconstruction and denoising module lacks cooperative optimization, so that the overall performance improvement is limited. At present, a channel estimation method based on super-resolution network is disclosed in the patent document CN119996119A, target LR image characteristics are extracted by utilizing a constructed convolutional neural network, channel estimation operation is realized by performing super-division recovery on a channel image to form an HR image, the structure is simple, the processing capacity of channel estimation is weak, the performance of the channel estimation is required to be improved continuously, a model can be improved continuously, a wireless fading channel estimation method based on depth dense Residual network is disclosed in the patent document CN114363129A, the depth dense network DeDNN and the depth Residual network ReDNN are respectively constructed by improving the depth neural network DNN by utilizing the dense network DENSENETS and the Residual network ResNets, gradient explosion and disappearance problems in network training are restrained by forming the structure DeReNet in series, but the problem of super-parameters and easy loss of spatial information exists in a fully-connected network structure, and the error propagation problem caused by two-stage training is solved by using a method based on cascade depth learning (such as SRCNN and DnCNN series connection) for the international conference paper Residual LEARNING MEETS OFDM Channel Estimation, and the overall performance of the super-resolution reconstruction and denoising module is limited. Disclosure of Invention In view of this, the present invention provides an OFDM channel estimation method combining the super-resolution and denoising joint models, which is used to realize high-precision and low-overhead channel estimation, and improve applicability and engineering practical value. In a first aspect, the present invention provides an OFDM channel estimation method combining a joint model of super-resolution and denoising, the method comprising: Step 1, generating a data set through a constructed multipath Rayleigh fading channel model; Step 2, according to the step1, converting the channel estimation problem in the orthogonal frequency division multiplexing OFDM system into an image super-resolution problem; step 3, based on step 2, constructing an improved super-resolution module SRCNN for reconstructing a high-resolution image from a low-resolution image; Step 4, constructing a denoising module DnCNN by adopting a residual error learning mode to predict noise between the output of the SRCNN module and a real channel; step 5, connecting the super-resolution module and the denoising module by adopting a feature fusion layer, carrying out information sharing, and constructing a joint model; step 6, optimizing parameters of the super-resolution module and the denoising module by adopting an end-to-end training strategy; and 7, simulation verification, namely evaluating algorithm performance through a mean square error MSE, setting a plurality of groups of signal-to-noise ratio (SNR) parameters, and verifying the performance of different models under different signal-to-noise ratios. Optionally, the step 1 comprises the steps that a multipath Rayleigh fading channel model has path gain conforming to complex Gaussian distribution, and exponentially decaying time delay and random Doppler frequen