An image denoising-based massive MIMO scattering channel estimation method

Abstract

The application discloses a large-scale MIMO scattering channel estimation method based on image denoising, and belongs to the field of large-scale MIMO tropospheric scattering communication. The method comprises the following steps: converting channel matrixes under different time and correlation into gray pictures, constructing a convolutional neural network with a fusion attention mechanism, training the network to denoise the pictures, converting the denoised gray pictures into channel matrixes to obtain the estimation value of the final channel coefficient. The method can overcome the problems of low channel estimation accuracy and high computational complexity of the traditional linear channel estimation method under small signal-to-noise ratio, effectively remove Gaussian noise under small signal-to-noise ratio to improve the estimation accuracy, and reduce the computational complexity by real-time estimation of the channel matrix after training, so that the large-scale MIMO tropospheric scattering communication system can obtain more accurate channel information. The application is simple to implement, low in calculation amount, and applicable to large-scale MIMO tropospheric scattering communication systems under various correlations.

Description

Technical Field

[0001] This invention relates to the field of large-scale MIMO tropospheric scattering communication, and in particular to a large-scale MIMO scattering channel estimation method based on image denoising, which is applicable to various large-scale MIMO tropospheric scattering communication scenarios that require low computational complexity and high channel estimation accuracy. Background Technology

[0002] In the field of communications, massive MIMO technology results in excessively large channel matrices, degrading the performance of traditional algorithms. Therefore, many scholars and research teams have begun to explore applying artificial intelligence techniques from image processing to channel estimation. First, the received signal is treated as a grayscale image, with Gaussian white noise in the image equivalent to the white noise the signal receives after passing through an AWGN channel. Second, a large dataset of labeled grayscale images is generated, and an image denoising network is trained after setting a loss function. Finally, the trained image denoising network is used to denoise the received signal in real time and extract the channel matrix.

[0003] With the development of deep learning, numerous image denoising networks (such as DnCNN, LDAMP, FFDNET, FCDAMP, FFDAMP, and TNLNET) have emerged. Among them, FFDNET has gained widespread attention due to its fast training speed and high estimation accuracy. FFDNET's unique feature is that it segments the input image into sub-images for denoising and then concatenates the denoised sub-images before outputting the result. This method accelerates training speed with almost no loss of accuracy, and has been validated by other research teams as highly effective in removing Gaussian noise from images.

[0004] In the field of channel estimation, FFDNET's estimation performance has been validated by many scholars, demonstrating superior accuracy and lower computational complexity compared to traditional channel estimation algorithms. However, for scattering channels with significant multipath effects, FFDNET performs poorly in channel estimation. This is because FFDNET is designed for natural images containing Gaussian noise, while the noise characteristics of grayscale images converted from received pilot signals are not significant, which is a considerable difference from the target object of FFDNET. Using grayscale images converted from received pilot signals directly as training and testing datasets yields poor results. Some scholars have proposed FFDAMP and TNLNET based on FFDNET. Although FFDAMP combines FFDNET with the AMP algorithm, its channel estimation accuracy is not significantly improved compared to FFDNET. While TNLNET addresses the shortcomings of FFDNET's signal-to-noise ratio estimation part, it sacrifices some computational complexity and increases training time. Summary of the Invention

[0005] The technical problem to be solved by this invention is to overcome the shortcomings mentioned above and provide a large-scale MIMO scattering channel estimation method based on image denoising. This method can improve the estimation accuracy within a small signal-to-noise ratio range while reducing computational complexity. It has the characteristics of fast training speed, good portability, hot-swappability, and wide applicability, and can be applied to large-scale MIMO tropospheric scattering communication systems under various correlations.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for estimating large-scale MIMO scattering channels based on image denoising includes the following steps:

[0008] (1) Obtain the channel matrix H under different correlations of the antenna array using the large-scale MIMO scattering channel model. n The LS channel estimation method is used to obtain a rough estimate of the channel matrix.

[0009] The specific method for step 1 is as follows:

[0010] (101) For the number of antenna elements at the transmitting and receiving ends, N are respectively t N r The large-scale MIMO tropospheric scattering communication system uses a tapped delay line model in the frequency domain to generate a spatially independent multipath channel matrix H. iid :

[0011]

[0012] Among them, H i Let τ represent the channel matrix of the i-th path. i P is the time delay of the i-th path. i Let N be the power of the i-th path. d Where f is the maximum number of paths, and f is the carrier frequency;

[0013] (102) Obtain the channel matrix H under different correlations of the antenna array using the large-scale MIMO scattering channel model. n :

[0014]

[0015] Among them, R TX R RX These are the correlation matrices for the sending and receiving ends, respectively, H. iid The multipath channel matrix is ​​free from spatial correlation;

[0016] (103) Use the LS channel estimation method to obtain a rough estimate of the channel matrix.

[0017]

[0018] Among them, X n Y n These are the transmitted signal and the received signal, respectively, and we have:

[0019] Y n =H n X n +W n

[0020] W n This represents a cyclic complex white Gaussian noise matrix, where the mean of each element is 0, and the noise power is determined by the received signal-to-noise ratio (SNR).

[0021] (2) H n The real and imaginary parts are separated, then vertically stitched together, normalized, and mapped to the range of 0-255 to form a grayscale channel image P. n ;Will The real and imaginary parts are separated, then vertically stitched together, normalized, and mapped to the range of 0-255 to form a grayscale channel image.

[0022] (3) Construct an image denoising network that integrates convolutional attention mechanism; the network includes a reversible downsampling operation module, a convolutional block attention model CBAM, a 15-layer convolutional network and an inverse downsampling operation module; the first convolutional layer is composed of "convolution operation Conv + linear rectified ReLU" unit, the middle convolutional layers are composed of the same "convolution operation Conv + batch normalization BN + linear rectified ReLU" unit, one CBAM is inserted between the first and second convolutional layers, and the last convolutional layer uses "convolution operation Conv" unit. The convolutional kernel of each convolutional layer is 3×3, and the zero-padding method is used to keep the dimension of the feature matrix after convolution unchanged.

[0023] The reversible downsampling module is used to reconstruct the N×L dimension received signal into four sub-matrices of N / 2×L / 2 dimension. Then, the four sub-matrices are concatenated with the N / 2×L / 2 dimension noise level map G in the third dimension to form an N / 2×L / 2×5 dimension tensor, which is used as the input of the first convolutional layer. The noise level map G is formed by expanding the dimension by the noise level factor μ, where μ∈(0,50].

[0024] The Convolutional Block Attention (CBAM) model adjusts the feature map output by the reversible downsampling operation module and then passes it to the subsequent 15 convolutional layers. After the last convolutional layer, an N / 2×L / 2×4 dimensional tensor is obtained and passed to the inverse downsampling operation module. The inverse downsampling operation module generates an estimated noise matrix of size N×L. Finally, the estimated noise matrix is ​​removed from the received signal to obtain the estimated channel matrix.

[0025] (4) Using different signal-to-noise ratios and correlations P n Construct sample data s, using the nth input sample and the nth target output sample respectively. n :

[0026]

[0027] Using sample data s n Construct training set U;

[0028] Using the training set U, the image denoising network incorporating convolutional attention mechanisms is trained using the Adam stochastic gradient descent algorithm until the output value of the loss function reaches a predetermined metric; where the loss function is:

[0029]

[0030] Where Θ represents the network parameters, F(·) represents the network output, and F(·) represents the processing of the channel image denoising network that incorporates the convolutional attention mechanism.

[0031] (5) Real-time received signals are acquired at the receiver of a large-scale MIMO tropospheric scattering communication system, and a rough estimate of the channel matrix is ​​obtained using the LS channel estimation method. Grayscale channel images are obtained after normalization mapping. Will Input a pre-trained image denoising network with a fused convolutional attention mechanism, and perform P~ n Remove noise to obtain the denoised grayscale channel image.

[0032] (6) Performing the inverse operations of mapping and normalization yields the final channel matrix estimate.

[0033] The beneficial effects of this invention are as follows:

[0034] 1. The algorithm of this invention is mature, stable, and has good real-time performance.

[0035] 2. This invention employs a Convolutional Attention Module (CBAM) to enhance the saliency of noise features in the image, thereby improving the accuracy of FFDNET in estimating large-scale MIMO scattering channels. Furthermore, CBAM is hot-swappable and versatile, making the magnitude and location of image noise more significant from both channel and spatial dimensions, thus enabling image denoising networks to denoise more accurately.

[0036] 3. This invention uses the fast and flexible deconvolutional network FFDNET as the image denoising network. It employs a downsampling process, which can significantly improve training speed without reducing denoising capability, effectively increase the receptive field without using dilated convolution, and result in a moderate network depth.

[0037] 4. This invention is independent of the data modulation and coding methods in the communication system, and has a wide range of applications and good portability. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of a scattering channel estimation method based on image denoising;

[0039] Figure 2 This is a schematic diagram of an image denoising method that incorporates convolutional attention mechanisms;

[0040] Figure 3 This is a schematic diagram of a 7-path channel model for a certain scattering link;

[0041] Figure 4 This is a schematic diagram of the TDL time-domain model;

[0042] Figure 5 It is P n , And M diagram;

[0043] Figure 6 This is a diagram of an image denoising network that incorporates a convolutional attention mechanism;

[0044] Figure 7 This is a schematic diagram of the CBAM structure;

[0045] Figure 8 This is a structural diagram of the channel attention module;

[0046] Figure 9 This is a structural diagram of the spatial attention module;

[0047] Figure 10 This is a flowchart of the receiving end processing of the present invention;

[0048] Figures 11(a)-11(d)Figure 11(a) shows the comparison between the performance of the present invention and the ideal MMSE algorithm NMSE; where Figure 11(a) shows the case of no correlation, Figure 11(b) shows the case of weak correlation, Figure 11(c) shows the case of moderate correlation, and Figure 11(d) shows the case of strong correlation.

[0049] Figure 12 This is a schematic diagram illustrating one possible way in which the present invention is applied to a practical communication system. Detailed Implementation

[0050] A method for estimating large-scale MIMO scattering channels based on image denoising, the basic principle of which is as follows: Figure 1 As shown, the image denoising method that integrates convolutional attention mechanism is as follows: Figure 2 As shown. Specifically, it includes the following steps:

[0051] (1) Generate the channel matrix H n and H n rough estimate

[0052] The channel matrix H under different correlations of the antenna array is obtained using a large-scale MIMO scattering channel model. n In a large-scale MIMO communication system, the channel matrix H is:

[0053]

[0054] Among them, R TX R RX These are the correlation matrices for the sending and receiving ends, respectively, H. iid It is a multipath channel matrix that does not contain spatial correlation.

[0055] Scattering channels exhibit significant multipath effects, and signals experience small-scale fading. The commonly used 7-path channel model in engineering is sufficient for studying and characterizing scattering channels; therefore, the scattering channel model used in this invention is a 7-path channel, such as... Figure 3 As shown. In selecting the correlation matrix model, the exponential correlation model proposed by academia is adopted. This model has physical rationality: its correlation decreases exponentially with the increase of the distance between receiving antennas, which corresponds to the real physical environment.

[0056] Using the tapped delay line (TDL) model to generate spatially uncorrelated tropospheric scattering fading channels, the TDL time-domain model is as follows: Figure 4 As shown.

[0057] In TDL, the Rayleigh fading generator generates a complex Gaussian variable h0(t), which is then multiplied by the square root of the power of each path to obtain the channel coefficient for that path. For the transmit and receive antenna arrays with N elements... t Nr The large-scale MIMO tropospheric scattering communication system generates a spatially independent 7-path channel matrix H in the frequency domain using TDL. iid As shown in equation (2):

[0058]

[0059] Among them, H i Let t represent the channel matrix of the i-th path. i P is the time delay of the i-th path. i Let N be the power of the i-th path. d This is the maximum diameter.

[0060] The large-scale MIMO scattering system model involved in this method is shown in equation (3):

[0061] Y = HX + W (3)

[0062] Where X and Y are the transmitted and received signals, respectively, and W represents the cyclic complex white Gaussian noise matrix, with the mean of each element being 0, and the noise power... It is determined by the set receive signal-to-noise ratio (SNR).

[0063] A rough estimate of the channel matrix H is obtained using the LS channel estimation method. n As shown in the following formula:

[0064]

[0065] Among them, X n Y n These are the transmitted signal and the received signal, respectively, and we have:

[0066] Y n =H n X n +W n

[0067] W n This represents a cyclic complex white Gaussian noise matrix, where the mean of each element is 0, and the noise power is determined by the received signal-to-noise ratio (SNR).

[0068] The channel matrix H under different correlations is obtained using equation (1). n Using equation (4), a rough estimate of the channel matrix under different correlations can be obtained. Where n represents the sequence number under different correlations.

[0069] (2) Image conversion and dataset construction

[0070] Due to H n and Both are complex matrices, first H nand The real and imaginary parts are separated, then vertically concatenated, and finally normalized and mapped to the range of 0-255 to form a grayscale channel image P. n and In a massive MIMO tropospheric scattering channel, with 16 transmit and receive antennas, a correlation strength of 0.7, and an SNR of 10 dB, P n , and channel noise, such as Figure 5 As shown.

[0071] P under different signal-to-noise ratios and correlations n , Construct sample data s, using them as input samples and target output samples respectively. i Then by s i The dataset S is composed as shown in Equations 5 and 6.

[0072] S is divided into a training set U and a test set V according to a certain ratio. Note that U∩V=Φ.

[0073]

[0074] S = {s1, ..., s} n} (6)

[0075] (3) Construction and training of image denoising network integrating convolutional attention mechanism

[0076] The image denoising network fused with CBAM (CBAM-FFDNET) consists of CBAM and a 15-layer convolutional network, and its structure is as follows: Figure 6 As shown, the network includes a reversible downsampling operation module, a convolutional block attention model (CBAM), a 15-layer convolutional network, and an inverse downsampling operation module. The first convolutional layer consists of a "convolution operation Conv + linear rectified ReLU" unit, the middle convolutional layers consist of the same "convolution operation Conv + batch normalization (BN) + linear rectified ReLU" unit, a CBAM is inserted between the first and second convolutional layers, and the last convolutional layer uses a "convolution operation Conv" unit. The convolutional kernel of each convolutional layer is 3×3, and zero padding is used to keep the dimension of the feature matrix after convolution unchanged.

[0077] The reversible downsampling module is used to reconstruct the N×L dimension received signal into four sub-matrices of N / 2×L / 2 dimension. Then, the four sub-matrices are concatenated with the N / 2×L / 2 dimension noise level map G in the third dimension to form an N / 2×L / 2×5 dimension tensor, which is used as the input of the first convolutional layer. The noise level map G is formed by expanding the dimension by the noise level factor μ, where μ∈(0,50].

[0078] The Convolutional Block Attention (CBAM) model adjusts the feature map output by the reversible downsampling operation module and then passes it to the subsequent 15 convolutional layers. After the last convolutional layer, an N / 2×L / 2×4 dimensional tensor is obtained and passed to the inverse downsampling operation module. The inverse downsampling operation module generates an estimated noise matrix of size N×L. Finally, the estimated noise matrix is ​​removed from the received signal to obtain the estimated channel matrix.

[0079] Convolutional Block Attention (CBAM) model, proposed in 2018 by Woo S, Park J, Lee J, and others, can be added to any CNN architecture to improve model performance. Given an intermediate feature map, the CBAM module infers attention maps sequentially along two independent dimensions (channels and space), and then multiplies the attention maps with the input feature map for adaptive feature optimization. For example... Figure 7 As shown, CBAM operates on feature maps, using channel attention and spatial attention modules to refine the extracted features and improve the model's performance.

[0080] The channel attention module structure is as follows Figure 8 As shown, its function is to compress the spatial dimension while keeping the channel dimension unchanged, so that the extracted high-level features are more comprehensive and richer. In the denoising task, this module focuses on the magnitude of noise in the input image, and the process is shown in Equation (7).

[0081]

[0082] Where F represents the input feature map, M C (F) represents the channel attention weights, and Ψ(·) represents the sigmoid function.

[0083] Spatial attention modules, such as Figure 9 As shown, its function is to keep the spatial dimension unchanged while compressing the channel dimension. In the denoising task, this module focuses on the location information of the noise, and the process is shown in equation (8).

[0084]

[0085] Where F represents the input feature map, M S (F) represents the spatial attention weight.

[0086] An image denoising network using a fused convolutional attention mechanism reduces or even removes noise from the channel image. The loss function is set as shown in Equation (9). U is used to train the subnetwork, and V is used to measure its performance. The Adam stochastic gradient descent algorithm is used for training, and the maximum number of iterations and training objective should be adjusted according to the actual situation.

[0087]

[0088] Where Θ represents the network parameters, F(·) represents the network output, and F(·) represents the processing of the channel image denoising network that incorporates the convolutional attention mechanism.

[0089] The denoising operation of the denoising network that incorporates the attention mechanism is shown in the following equation:

[0090]

[0091] The noise level map G is formed by expanding the noise level factor μ, where μ∈(0,50).

[0092] In this formula, the L1 regularization term Φ(·) can constrain the computational complexity of the model. The noise level factor μ affects the trade-off between image noise removal and detail preservation; a higher noise level factor indicates stronger denoising capabilities.

[0093] (4) Inverse operations of mapping and normalization

[0094] Image denoising network using fused convolutional attention mechanism Mapped After removing noise, a denoised grayscale channel image is obtained. As shown in the following formula:

[0095]

[0096] Ultimately, for By performing the inverse operations of mapping and normalization, a more accurate estimate of the channel matrix is ​​obtained.

[0097] The processing flowchart of the receiver in this method is as follows: Figure 10 As shown.

[0098] The following section demonstrates the effectiveness of the channel estimation method of this invention by combining the process of channel estimation using an image denoising network with fused convolutional attention mechanism in a large-scale MIMO scattering channel estimation communication system with 16×16 antennas at both ends.

[0099] Under different correlation and signal-to-noise ratios, the estimation performance of the image denoising network estimation method fused with convolutional attention mechanism (CBAM-FFDNET) is compared with that of the commonly used ideal MMSE algorithm. The accuracy of the estimation method is measured by NMSE. Figures 11(a)-11(d)As shown, α = 0, 0.2, 0.5, and 0.8 represent the cases of no correlation, weak correlation, moderate correlation, and strong correlation between antenna elements, respectively. It can be seen that the accuracy of this method is superior to the ideal MMSE algorithm under low signal-to-noise ratio (SNR) conditions (SNR less than 10dB) and all correlation levels. When the SNR is low, the noise proportion is high. The advantage of this method lies in improving the channel estimation accuracy by removing the influence of noise, while the error of the ideal MMSE algorithm is positively correlated with the noise power; the higher the noise proportion, the lower the accuracy of the ideal MMSE algorithm.

[0100] Finally, one possible way to apply this method to a practical communication system is shown in [the document / example]. Figure 12 middle.

[0101] In summary, this invention adds a step of converting the channel matrix into a grayscale image to the traditional channel estimation method LS algorithm, so that the channel matrix estimation problem can be transformed into a denoising problem of the grayscale image of the channel matrix. In order to obtain better denoising effect, a convolutional attention module CBAM is integrated on the basis of FFDNET. CBAM makes the noise features more obvious, so that FFDNET can remove the influence of Gaussian noise on the channel matrix.

[0102] This invention overcomes the problems of low channel estimation accuracy and high computational complexity of traditional linear channel estimation methods at low signal-to-noise ratios (SNRs). It effectively removes Gaussian noise at low SNRs, thereby improving estimation accuracy. The trained network can estimate the channel matrix in real time, reducing computational complexity and enabling large-scale MIMO tropospheric scattering communication systems to obtain more accurate channel information. This invention is simple to implement, computationally inexpensive, and applicable to large-scale MIMO tropospheric scattering communication systems under various correlation conditions.

Claims

1. A method for estimating large-scale MIMO scattering channels based on image denoising, characterized in that, Includes the following steps: (1) Obtain the channel matrix H under different correlations of the antenna array using the large-scale MIMO scattering channel model. n The LS channel estimation method is used to obtain a rough estimate of the channel matrix. The specific method for step 1 is as follows: (101) For the number of antenna elements at the transmitting and receiving ends, N are respectively t N r The large-scale MIMO tropospheric scattering communication system uses a tapped delay line model in the frequency domain to generate a spatially independent multipath channel matrix H. iid : Among them, H i Let τ represent the channel matrix of the i-th path. i P is the time delay of the i-th path. i Let N be the power of the i-th path. d Where f is the maximum number of paths, and f is the carrier frequency; (102) Obtain the channel matrix H under different correlations of the antenna array using the large-scale MIMO scattering channel model. n : Among them, R TX R RX These are the correlation matrices for the sending and receiving ends, respectively, H. iid The multipath channel matrix is ​​free from spatial correlation; (103) Use the LS channel estimation method to obtain a rough estimate of the channel matrix H~ n : Among them, X n Y n These are the transmitted signal and the received signal, respectively, and we have: Y n =H n X n +W n W n This represents a cyclic complex white Gaussian noise matrix, where the mean of each element is 0, and the noise power is determined by the received signal-to-noise ratio (SNR). (2) H n The real and imaginary parts are separated, then vertically stitched together, normalized, and mapped to the range of 0-255 to form a grayscale channel image P. n ;Will The real and imaginary parts are separated, then vertically stitched together, normalized, and mapped to the range of 0-255 to form a grayscale channel image. (3) Construct an image denoising network that integrates convolutional attention mechanism; the network includes a reversible downsampling operation module, a convolutional block attention model CBAM, a 15-layer convolutional network and an inverse downsampling operation module; the first convolutional layer is composed of "convolution operation Conv + linear rectified ReLU" unit, the middle convolutional layers are composed of the same "convolution operation Conv + batch normalization BN + linear rectified ReLU" unit, one CBAM is inserted between the first and second convolutional layers, and the last convolutional layer uses "convolution operation Conv" unit. The convolutional kernel of each convolutional layer is 3×3, and the zero-padding method is used to keep the dimension of the feature matrix after convolution unchanged. The reversible downsampling module is used to reconstruct the N×L dimension received signal into four sub-matrices of N / 2×L / 2 dimension. Then, the four sub-matrices are concatenated with the N / 2×L / 2 dimension noise level map G in the third dimension to form an N / 2×L / 2×5 dimension tensor, which is used as the input of the first convolutional layer. The noise level map G is formed by expanding the dimension by the noise level factor μ, where μ∈(0,50]. The Convolutional Block Attention (CBAM) model adjusts the feature map output by the reversible downsampling operation module and then passes it to the subsequent 15 convolutional layers. After the last convolutional layer, an N / 2×L / 2×4 dimensional tensor is obtained and passed to the inverse downsampling operation module. The inverse downsampling operation module generates an estimated noise matrix of size N×L. Finally, the estimated noise matrix is ​​removed from the received signal to obtain the estimated channel matrix. (4) Using different signal-to-noise ratios and correlations P n Construct sample data s, using the nth input sample and the nth target output sample respectively. n : Using sample data s n Construct training set U; Using the training set U, the image denoising network incorporating convolutional attention mechanisms is trained using the Adam stochastic gradient descent algorithm until the output value of the loss function reaches a predetermined metric; where the loss function is: Where Θ represents the network parameters, F(·) represents the network output, and F(·) represents the processing of the channel image denoising network that incorporates the convolutional attention mechanism. (5) Real-time received signals are acquired at the receiver of a large-scale MIMO tropospheric scattering communication system, and a rough estimate of the channel matrix is ​​obtained using the LS channel estimation method. Grayscale channel images are obtained after normalization mapping. Will Input the trained image denoising network with fused convolutional attention mechanism, and... Remove noise to obtain the denoised grayscale channel image. (6) Performing the inverse operations of mapping and normalization yields the final channel matrix estimate.

Images