Pixel attention network-based channel estimation

Abstract

The invention belongs to the field of channel estimation in wireless communication, provides a channel estimation algorithm based on an improved pixel attention network (PAN), and aims to solve the problems of inter-subcarrier crosstalk, inter-subsymbol interference and the like in 5G system signals received by a traditional channel estimation method. According to the method, an SC-PA block and a U-PA block are combined, batch normalization (BN) and a linear rectification function (ReLU) are added after a convolution layer of a feature extraction module of the PAN, and the precision of channel estimation is improved by enhancing the feature expression capability. The method specifically comprises the following steps: firstly, aiming at a block symbol structure of an OFDM symbol, inserting two-dimensionally distributed discrete pilot frequency symbols into a time-frequency grid; and then forming a low-resolution channel time-frequency gain image by using the initial estimation value of the channel at the pilot symbol position, and further processing through the improved PAN network. According to the method, the accuracy of channel estimation is effectively improved, and especially in a 5G system, the problems of crosstalk between subcarriers and interference between sub-symbols can be better solved.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology based on deep learning in channel estimation, and proposes a deep learning channel estimation algorithm based on an improved pixel attention network. Background Art

[0002] Orthogonal frequency division multiplexing (OFDM) is a wireless communication technology that efficiently utilizes spectrum resources. Traditional OFDM channel estimation techniques are mainly divided into blind estimation and non-blind estimation. Non-blind channel estimation is suitable for rapidly changing wireless environments, and commonly used algorithms include least squares (LS) and minimum mean square error (MMSE) algorithms. The LS algorithm has a simple structure and small amount of calculation, but it does not consider noise and interference between subcarriers, and its accuracy is limited. The MMSE algorithm takes noise and interference into account, but it has a large amount of calculation and is limited in practical application. In addition, there are traditional linear estimators that rely on statistical models, have high implementation complexity, and lack robustness in highly dynamic environments.

[0003] The time-varying characteristics of the channel can cause mutual interference between subcarriers, a phenomenon known as inter-carrier interference (ICI). Due to the existence of the Doppler effect, the signal undergoes frequency changes during transmission, resulting in differences between the signal received at the receiving end and the original signal. In order to reduce this interference, Doppler estimation is usually used to compensate the signal. However, since there are certain errors in the Doppler estimation itself, even the compensated signal cannot completely eliminate the interference. This residual Doppler factor will still have a significant impact on the mutual interference between subcarriers, thereby affecting the performance of the entire communication system.

[0004] With the explosive growth of data and the continuous development of artificial intelligence technology, machine learning is gradually unable to meet the current usage needs. Deep Learning (DL) is a machine learning technology based on neural networks. It can handle more complex data structures and has achieved great success in many fields. The core of deep learning lies in the structure of the neural network, which consists of an input layer, a hidden layer, and an output layer. Each neuron node is connected to the neurons in the next layer according to the weight, so it is also called a deep neural network (DNN). Common deep learning algorithms include convolutional neural networks (CNN), recurrent neural networks (RNN), long short-term memory models (LSTM), etc. The literature [H. He, C.-K. Wen, S. Jin, and GY Li, "Deep learning-based channel estimation for beamspace mmWave massive MIMO systems," IEEE Wireless Commun. Lett., vol. 7, no. 5, pp. 852-855, Oct. 2018] uses the learned denoising-based approximate message passing network (LDAMP) to iteratively estimate the millimeter wave channel, and integrates the denoising convolutional neural network (DNCNN) with the iterative sparse signal recovery algorithm, and uses DNCNN to update the estimated channel in each iteration. The paper [Soltani M, Pourahmadi V, Mirzaei A, et al. Deep learning-based channel estimation [J]. IEEE Communications Letters, 2019, 23 (4): 652-655] models the time-frequency grid of the channel response as a low-resolution two-dimensional image with known pilot positions, and proposes a channel estimation method (ChannelNet) consisting of two independent sub-networks cascaded. This method first uses an image super-resolution (SR) algorithm to reconstruct a high-resolution image, and then uses an image restoration (IR) method to remove the influence of noise.The literature [LeCun Y, Bottou L, Bengio Y, et al. Gradient-based learning applied to document recognition [J]. Proceedings of the IEEE, 1998, 86 (11): 2278-2324] first proposed the convolutional neural network structure LeNet-5, which consists of multiple convolutional layers, pooling layers and fully connected layers of different sizes. The LeNet-5 network first performs feature image processing layer by layer through forward propagation to obtain the output result and calculate its error with the label. Then, using the back propagation method, the error is transmitted forward layer by layer from the output layer, the gradient is calculated and the weight parameters are updated. Finally, the loss function converges to the minimum value through repeated iterative calculations. Researchers in the field of wireless communications combine deep learning with traditional technologies and apply them to the physical layer, including channel decoding, signal detection, channel equalization and channel estimation. Deep learning networks can learn layer by layer and extract features from data, and have a stronger ability to deal with nonlinear problems than traditional methods. At the same time, the deep learning model can improve system performance and the efficiency of the communication system through continuous iterative optimization. Summary of the invention

[0005] The present invention aims at the problems of inter-subcarrier crosstalk and inter-subsymbol interference in signal reception of 5G system. In order to further give play to the advantages of image denoising in channel estimation, reduce network complexity and improve the accuracy of channel estimation, the pixel attention mechanism is introduced into channel estimation, and further processing is carried out through an improved PAN network after preliminary estimation.

[0006] The method adopted in the present invention is a deep learning channel estimation method based on an improved pixel attention network. Batch normalization (BN) and linear rectification function (ReLU) are added after the convolution layer of the feature extraction module of the PAN network, and combined with the SC-PA block and the U-PA block, that is, the improved PAN network, a channel estimation method based on an improved pixel attention network, namely CE-PAN, is proposed. The main process of the method is as follows:

[0007] (1) In an orthogonal frequency division multiplexing (OFDM) system, pilot symbols are usually inserted into the time-frequency grid in order to effectively estimate and compensate for the impact of the channel. These pilot symbols are designed to have a block structure to facilitate channel estimation at the receiving end. Specifically, the pilot symbols are distributed in a two-dimensional time-frequency grid so that the channel response information can be obtained at different time and frequency positions.

[0008] (2) At the receiving end, a preliminary channel estimation is first performed on the pilot symbols to obtain the initial estimated value of the channel. Then, the proposed improved pixel attention network is used to further process the initial estimated value. The pixel attention mechanism can focus on the key features in the channel estimation, thereby improving the accuracy of the estimation.

[0009] (3) Based on the pixel attention mechanism, batch normalization (BN) and linear rectification function (ReLU) are added after the convolution layer of the feature extraction module, and the SC-PA block and U-PA block are combined to form the CE-PAN network. The SC-PA block is responsible for modeling the spatial correlation in channel estimation, while the U-PA block handles the temporal correlation in channel estimation. Through this combination, the CE-PAN network can simultaneously consider the spatial and temporal characteristics of channel estimation, further improving the accuracy of estimation.

[0010] (4) During the processing of the CE-PAN network, the network will continuously iterate and optimize, and adjust the network parameters through the back propagation algorithm to minimize the estimation error. Ultimately, the CE-PAN network outputs a more accurate channel estimation result, providing a reliable foundation for subsequent signal demodulation and data recovery.

[0011] The channel estimation algorithm based on the improved pixel attention network proposed in the present invention effectively improves the accuracy of channel estimation through deep learning technology. Especially in 5G systems, it can better deal with the problems of inter-subcarrier crosstalk and inter-subsymbol interference, thereby improving the performance of the entire communication system.

[0012] In practical applications, this CE-PAN-based channel estimation method can significantly improve the performance of OFDM systems, especially in environments with complex and changing channel conditions. Accurate channel estimation can more effectively equalize and demodulate signals, thereby improving the rate and reliability of data transmission. In addition, this method also has a certain degree of adaptability and can adapt to different channel conditions and changes, making it possible to achieve a more robust communication system. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic diagram of the improved PAN network structure.

[0014] Figure 2 is a schematic diagram of an example of the CE-PAN method.

[0015] Figure 3 This is a comparison chart of the test results. DETAILED DESCRIPTION

[0016] The following is a further explanation of a channel estimation algorithm based on an improved pixel attention mechanism network proposed in the present invention in conjunction with the accompanying drawings.

[0017] This study built the PyTorch deep learning framework on the hardware device based on Intel(R) Core(TM) i7-12700HCPU@2.70GHz (16GB running memory) and NVIDIA Geforce GTX3060 GPU under the Windows 11 operating system, and used the Python programming language to implement the training and testing of the Chinese named entity recognition model.

[0018] The present invention uses a MIMO system of a uniform linear array (ULA), such as Figure 1 As shown. The transmitter has N T transmit antennas and The receiving end has N transmit RF chains (RE chains). R receiving antennas and In order to reduce power consumption and cost, in actual applications of millimeter-wave massive MIMO, the number of RF chains is much smaller than the number of antennas. Therefore, the present invention assumes

[0019] The mmWave massive MIMO channel has high path loss for non-line-of-sight (NLOS) signals and has significant spatial angle sparsity, where only a small number of dominant multipaths (usually 3 to 5 multipaths in real environments) are composed of mmWave MIMO multipath channels. The point-to-point mmWave massive MIMO channel can be modeled as formula (1) and formula (2).

[0020]

[0021] Where H is the size of N R ×N T represents the millimeter wave channel between the receiver and the transmitter; ρ represents the average path loss; L represents the number of main paths; assuming that the path amplitude follows the Rayleigh distribution, we have represents the transmission gain of the lth path, is the average power gain. l ∈[0, 2π], Respectively represent the arrival azimuth and departure azimuth (AoD / AoA) of the lth path. R (θ l )and are the response vectors of the antenna array at the receiving end and the transmitting end, respectively, which can be expressed as formula (3) and formula (4).

[0022]

[0023] Where λ is the carrier wavelength and d is the distance between adjacent antennas.

[0024] A R =[a R (θ1), a R (θ2), ..., a R (θ L )](5)

[0025]

[0026] D = diag(α1, α2, ..., α L ) (7)

[0027] (1) In an orthogonal frequency division multiplexing (OFDM) system, pilot symbols are usually inserted into the time-frequency grid in order to effectively estimate and compensate for the impact of the channel. These pilot symbols are designed to have a block structure to facilitate channel estimation at the receiving end. Specifically, the pilot symbols are distributed in a two-dimensional time-frequency grid so that the channel response information can be obtained at different time and frequency positions.

[0028] Assume that the transmitter activates one RF chain in one beam direction to transmit a pilot signal, and the receiver uses all RF chains in different beam directions to combine and receive the pilot signal. The original signal transmitted is S k , then in the baseband of the receiver, the received k-th subcarrier pilot signal can be expressed as formula (8).

[0029]

[0030] Where Z k is additive white Gaussian noise; F k =F RF F BB represents the transmitter precoder, where F RF , F BB are RF and baseband encoders respectively. Similar to the transmitter precoder, W k represents the receiver hybrid combiner, which consists of the RF combiner W RF and baseband combiner W BB composition.

[0031] Assuming that in the worst case W k =W, F k =F, Where P represents the signal transmission power. Formula (8) can be further expressed as formula (9).

[0032]

[0033] (2) At the receiving end, a preliminary channel estimation is first performed on the pilot symbols to obtain the initial estimated value of the channel. Then, the proposed improved pixel attention mechanism network is used to further process the initial estimated value. The pixel attention mechanism can focus on the key features in the channel estimation, thereby improving the estimation accuracy.

[0034] The received pilot signal is processed by the preliminary estimation module through two matrices to obtain a preliminary estimation matrix G k As shown in formula (10).

[0035] G k =B1Y k B2(10)

[0036] in:

[0037] B1=(W H W) -1 W (11)

[0038] B2=(F H F) -1 F (12)

[0039] (3) Based on the pixel attention mechanism, batch normalization (BN) and linear rectification function (ReLU) are added after the convolution layer of the feature extraction module, and the SC-PA block and U-PA block are combined to form the CE-PAN network. The SC-PA block is responsible for modeling the spatial correlation in channel estimation, while the U-PA block handles the temporal correlation in channel estimation. Through this combination, the CE-PAN network can simultaneously consider the spatial and temporal characteristics of channel estimation, further improving the accuracy of estimation.

[0040] Finally, the initial estimate matrix G k As the input data of the deep learning network, the estimated channel matrix is ​​output through the mapping relationship as shown in formula (13).

[0041]

[0042] where θ is the set of parameters of the deep learning network.

[0043] (4) During the processing of the CE-PAN network, the network will continuously iterate and optimize, and adjust the network parameters through the back propagation algorithm to minimize the estimation error. Ultimately, the CE-PAN network outputs a more accurate channel estimation result, providing a reliable foundation for subsequent signal demodulation and data recovery.

[0044] The present invention considers the non-line-of-sight (NLOS) scenario of urban micro-cell (UMi) streets, uses the same training data set, and trains channel data with different signal-to-noise ratios. Figure 3 The normalized mean square error of several channel estimation algorithms, including CE-PAN, Deep CNN, DNCNN, and non-ideal MMSE, is compared in MIMO-OFDM channels. Figure 3 It can be seen that the above-mentioned deep learning channel estimation algorithms are all better than the MMSE algorithm. Among them, the performance of the DNCNN network is better than that of the Deep CNN, and the performance of the CE-PAN network proposed in the present invention is the best.

Claims

1. A channel estimation algorithm based on an improved pixel attention mechanism network, characterized in that: The steps are as follows: (1) In an orthogonal frequency division multiplexing (OFDM) system, pilot symbols are usually inserted into the time-frequency grid in order to effectively estimate and compensate for the impact of the channel. These pilot symbols are designed to have a block structure to facilitate channel estimation at the receiving end. Specifically, the pilot symbols are distributed in a two-dimensional time-frequency grid so that the channel response information can be obtained at different time and frequency positions. (2) At the receiving end, a preliminary channel estimation is first performed on the pilot symbols to obtain the initial estimated value of the channel. Then, the proposed improved pixel attention mechanism network is used to further process the initial estimated value. The pixel attention mechanism can focus on the key features in channel estimation, thereby improving the accuracy of the estimation. (3) Based on the pixel attention mechanism, batch normalization (BN) and linear rectification function (ReLU) are added after the convolution layer of the feature extraction module, and the SC-PA block and U-PA block are combined to form a CE-PAN network. The SC-PA block is responsible for modeling the spatial correlation in channel estimation, while the U-PA block handles the temporal correlation in channel estimation. Through this combination, the improved PAN network can simultaneously consider the spatial and temporal characteristics in channel estimation, further improving the accuracy of estimation. (4) During the processing of the CE-PAN network, the network will continuously iterate and optimize, and adjust the network parameters through the back propagation algorithm to minimize the estimation error. Ultimately, the CE-PAN network outputs a more accurate channel estimation result, providing a reliable foundation for subsequent signal demodulation and data recovery.

2. The channel estimation algorithm based on the pixel attention mechanism network according to claim 1 is characterized in that: The transmitter activates one RF chain in one beam direction to transmit a pilot signal, and the receiver uses all RF chains in different beam directions to combine and receive the pilot signal. The original signal transmitted is S k , then in the baseband of the receiver, the received k-th subcarrier pilot signal can be expressed as: Where Z k is additive white Gaussian noise; F k =F RF F BB represents the transmitter precoder, where F RF , F BB are RF and baseband encoders respectively. Similar to the transmitter precoder, W k represents the receiver hybrid combiner, which consists of the RF combiner W RF and baseband combiner W BB composition. Assuming that in the worst case W k =W,F k =F, Where P represents the signal transmission power. The above formula can be further expressed as:

3. Channel estimation algorithm based on improved pixel attention mechanism network, characterized in that: The received pilot signal is processed by the preliminary estimation module through two matrices to obtain a preliminary estimation matrix G k As shown in formula (10). G k =B1Y k B2 in: B1=(W H IN) -1 IN B2=(F H F) -1 F 4. The channel estimation algorithm based on the improved pixel attention mechanism network according to claim 1, characterized in that: Finally, the initial estimate matrix G k As the input data of the deep learning network, the estimated channel matrix is ​​output through the mapping relationship as shown in the formula. Where θ is the parameter set of the deep learning network. The hidden state of the BiLSTM concatenated at the i-th step Constitutes c i Context-dependent representation of .

5. The channel estimation algorithm based on the improved pixel attention mechanism network according to claim 1, characterized in that: The present invention considers the non-line-of-sight (NLOS) scenario of urban micro-cell (UMi) streets, uses the same training data set, and trains channel data with different signal-to-noise ratios. Figure 3 compares the normalized mean square errors of several channel estimation algorithms CE-PAN, Deep CNN, DNCNN, and non-ideal MMSE under MIMO-OFDM channels. As can be seen from Figure 3, the above-mentioned deep learning channel estimation algorithms are all better than the MMSE algorithm. Among them, the performance of the DNCNN network is better than that of the Deep CNN, and the performance of the CE-PAN network proposed in the present invention is the best.

Images