Channel estimation method based on super-resolution network

Abstract

The invention discloses a channel estimation method based on a super-resolution network. According to the invention, the time-frequency grid of the channel response is modeled as a known 2D image only at the pilot position. A channel grid with a plurality of pilot frequencies is regarded as a low-resolution image, and an estimated channel is regarded as a high-resolution image. And extracting target LR image features by using the constructed convolutional neural network, and performing interpolation processing between known pilot frequency positions, thereby performing super-resolution recovery on a channel image to become an HR image, and realizing channel estimation operation. The method comprises the following steps: constructing a super-resolution neural network; generating a training data set; training the super-resolution neural network; testing the network performance by using the tested data set; according to the method, the super-resolution neural network is used for learning the channel matrix characteristics, the accuracy and robustness of channel estimation are greatly improved, and the performance is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communications, and in particular to a channel estimation method based on a super-resolution network. Background Art

[0002] Emerging deep learning (DL) techniques have attracted great attention in the field of channel estimation. Convolutional neural networks (CNNs) have found great potential in exploiting nonlinear mapping capabilities. One approach involves modeling channel interpolation as a CNN-based image super-resolution method, since channel interpolation can be viewed as reconstructing high-resolution CSI from low-resolution channel estimates at pilot locations. Specifically, the channel matrix is ​​first determined by LS estimation and then passed through a super-resolution convolutional neural network (SRCNN) and a denoising convolutional neural network (DnCNN) to improve the estimation accuracy. In addition, residual learning-based neural networks and super-resolution generative adversarial networks are also used for the interpolation process in channel estimation. In these applications, paired training data (such as low-resolution corrupted observations and high-resolution clean targets) are usually used to train super-resolution models in a supervised training manner.

[0003] However, this type of method usually uses a two-stage training method, which will cause transmission errors in the transmission process between the two networks, increase the difficulty of training, and have poor adaptability to the environment. In addition, unlike images, the information interaction between features in the channel matrix is ​​stronger, and better long-term dependency capture is needed to improve model performance. Summary of the invention

[0004] The purpose of the present invention is to overcome the above-mentioned shortcomings and deficiencies of the prior art and provide a channel estimation method based on a super-resolution network to improve the performance of channel estimation.

[0005] The present invention is achieved through the following technical solutions:

[0006] A channel estimation method based on a super-resolution network comprises the following steps:

[0007] Step 1, construct a super-resolution neural network;

[0008] The super-resolution neural network includes five modules: Conv layer (convolution layer), ReLu activation layer, HorBlock (recursive gated convolution block), UpSample (upsampling module) and Shuffle layer.

[0009] Step 2, generate training data set;

[0010] Step 3, training the super-resolution neural network;

[0011] Step 4: Test the network performance using the test data set.

[0012] Preferably, in the constructed super-resolution neural network, the role of the convolution layer is to extract image features and realize parameter sharing, reduce network parameters, and improve training efficiency, wherein the number of convolution kernels is set to 64, the convolution kernel size is set to 3×3, the step size is set to 1, and the padding size is set to 1.

[0013] The Relu activation layer activates the feature matrix convolved in the previous step to form a new matrix, which is used to add nonlinear factors to the model.

[0014] HorBlock (recursive gated convolution block) uses gated convolution and recursive operations to efficiently implement information interaction of any order. The basic operation trend of visual modeling shows that the expressiveness of the model can be improved by increasing the order of spatial interaction. The features extracted by the convolution layer will pass through the LayerNorm layer and then pass through the g n Conv module, which is based on recursive and gated convolutions. n Conv) enables convolutional neural networks to complete high-order spatial interactions. Suppose the input feature is Gated convolution g n The output of Conv y = g n Conv(x) can be expressed as:

[0015]

[0016]

[0017]

[0018] in It is a linear projection operation to complete the information exchange in the channel dimension. f is the depth-wise convolution, which is to linearly project the input x and divide it according to the channel to obtain p0 and q0. Then, q0 after depth convolution calculation is point-convolved with p0 to obtain p1. p1 will undergo another linear projection to obtain the output y. The above process demonstrates the first-order spatial interaction, and the interaction between adjacent features p0 and q0 is explicitly introduced through element-wise multiplication.

[0019] The above process demonstrates the first-order spatial interaction. The same is true for higher-order spatial interactions. First, use the linear projection function Get a set of projection features p0 and

[0020]

[0021] This formula mainly divides the input features according to the channel, and then inputs them into the gated convolution for recursive operation:

[0022] p k+1 =f k (q k )⊙g k (p k ) / α k=0,1,..,n-1

[0023] The output features in each operation are scaled by 1 / α to make the training more stable. Since the dimensions of the two must be the same during the spatial interaction, g k It is the dimension mapping function during the operation:

[0024]

[0025] Finally, the network outputs the final recursive output q n Input to the projection layer to get g n The final result of Conv. In order to ensure that high-order interactive operations do not introduce too much computational overhead, the number of channels at each order is also constrained:

[0026]

[0027] There are 24 Horblocks in the network.

[0028] UpSample is a module for increasing the resolution of images. Its function is to rescale the input image to a desired size according to certain rules. UpSample consists of conv layer and shuffle. Shuffle here refers to PixelShuffle, which is a specific upsampling method. UpSample is implemented by doubling the number of channels through convolution; then using PixelShuffle, the feature maps of two channels are inserted into each other to double the size, so as to reconstruct the final superpixel of the image. For example, first, a feature map (n, 64, 64, 256) with doubled channel number is obtained through convolution, and then the feature map is cut into several parts, and the pixels of each part (n, 64, 64, 4) are rearranged and reshaped into (n, 64, 64, 2, 2), then reshaped into (n, 64, 2, 64, 2), and finally reshaped into (n, 128, 128, 1), and then these are spliced ​​together to obtain a feature map (n, 128, 128, 64).

[0029] Further, the training dataset is generated considering the use of a single antenna at the transmitter and receiver. For channel modeling and pilot transmission, we used the widely used LTE simulator developed at the University of Vienna, the Vienna LTE-A Simulator. The training, test, and validation sets consist of 32000, 4000, and 4000 channels, respectively, where each frame consists of 14 time slots and 72 subcarriers. The radio channel model for VehicularA (VehA) with a carrier frequency of 2.1GHz, a bandwidth of 1.6MHz, and a UE (User Equipment) speed of 50km / h.

[0030] Furthermore, the super-resolution neural network training method is as follows:

[0031] Let Θ represent the set of all network parameters. The input of the network is the pilot value vector The output is the estimated channel matrix, expressed as

[0032]

[0033] The overall loss function of the network is the mean square error (MSE) between the estimated channel response and the actual channel response, calculated as follows:

[0034]

[0035] where Γ is the set of all training data, H is the perfect channel, and ‖Γ‖ is the size of the training set.

[0036] Similar to image-based techniques, the optimal weights of the network depend on the value of the SNR; therefore, to get a complete solution, the network must be retrained for each SNR value. Training the network for several SNR values ​​(12dB and 22dB in this method) still leads to good performance.

[0037] In order to enable the network's feature extraction module LR space to better extract features, the real part, imaginary part and amplitude value of the data in the training set need to be extracted and spliced ​​together before training. The original training set size is (32000, 72, 14, 1), and the training set size after splitting should be (96000, 72, 14, 1). The optimizer of all networks is Adam, the training learning rate is set to 0.001, the batch size (Batch_size) is 128, and the training epochs is 300.

[0038] Furthermore, the network performance is tested using the test data set using the normalized mean square error (NMSE) to measure the channel estimation performance, which is defined as:

[0039]

[0040] Compared with the prior art, the present invention has the following advantages and effects:

[0041] The present invention is a one-stage model, which has better robustness than the two-stage training model;

[0042] The present invention uses a recursive gated convolution block to replace the traditional residual convolution block, taking into account the long-distance information interaction between the features in the channel matrix, and extracting the real part, imaginary part and amplitude value of the data in the data set during the training phase, thereby increasing the feature quantity of the data and facilitating better learning of the network;

[0043] Compared with traditional channel estimation methods, the present invention utilizes a super-resolution neural network to learn channel matrix characteristics, which greatly improves the accuracy and robustness of channel estimation and achieves performance improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Schematic diagram of the general flow of the channel estimation method based on the super-resolution network.

[0045] Figure 2 This is the structure of the super-resolution neural network.

[0046] Figure 3 It is the structure of HorBlock (recursive gated convolution block). DETAILED DESCRIPTION

[0047] The present invention is further described in detail below in conjunction with specific embodiments.

[0048] like Figure 1 The channel estimation method based on the super-resolution network of the present invention is performed in the following steps:

[0049] S1. Build a super-resolution neural network with the following structure: Figure 2 As shown;

[0050] S11. Set the number of convolution kernels of the convolution layer to 64, the convolution kernel size to 3×3, the step size to 1, and the padding size to 1.

[0051] S12. Use the Relu activation layer to activate the feature matrix convolved in the previous step to form a new matrix to add nonlinear factors to the model.

[0052] S13, using HorBlock (recursive gated convolution block), with the help of gated convolution and recursive operations, we can efficiently realize information interaction of any order. The basic operation trend of visual modeling shows that the expressiveness of the model can be improved by increasing the order of spatial interaction. The specific structure is as follows Figure 3As shown, the features extracted by the convolutional layer will pass through the Layer Norm layer and then through the g n Conv module, which is based on recursive and gated convolutions. n Conv) enables convolutional neural networks to complete high-order spatial interactions. Suppose the input feature is Gated convolution g n The output of Conv y = g n Conv(x) can be expressed as:

[0053]

[0054]

[0055]

[0056] in It is a linear projection operation to complete the information exchange in the channel dimension. f is the depth-wise convolution, which is to linearly project the input x and divide it according to the channel to obtain p0 and q0. Then, q0 after depth convolution calculation is point-convolved with p0 to obtain p1. p1 will undergo another linear projection to obtain the output y. The above process demonstrates the first-order spatial interaction, and the interaction between adjacent features p0 and q0 is explicitly introduced through element-wise multiplication.

[0057] The above process demonstrates the first-order spatial interaction. The same is true for higher-order spatial interactions. First, use the linear projection function Get a set of projection features p0 and

[0058]

[0059] This formula mainly divides the input features according to the channel, and then inputs them into the gated convolution for recursive operation:

[0060] p k+1 =f k (q k )⊙g k (p k ) / α k=0,1,..,n-1

[0061] The output features in each operation are scaled by 1 / α to make the training more stable. Since the dimensions of the two must be the same during the spatial interaction, g k It is the dimension mapping function during the operation:

[0062]

[0063] Finally, the network outputs the final recursive output q n Input to the projection layer to get g n The final result of Conv. In order to ensure that high-order interactive operations do not introduce too much computational overhead, the number of channels at each order is also constrained:

[0064]

[0065] The network includes 24 Horblocks in total;

[0066] S14. Use UpSample (up-sampling module) to expand the image resolution. Its function is to rescale the input image to a desired size according to certain rules. UpSample (up-sampling module) consists of conv layer and shuffle. Here, shuffle refers to PixelShuffle, which is a specific up-sampling method. The implementation of UpSample is to double the number of channels through convolution; then use PixelShuffle to insert the feature maps of the two channels into each other to double the size, so as to reconstruct the final superpixel of the image. For example, first obtain the feature map (n, 64, 64, 256) with doubled channel number through convolution, then cut the feature map into several parts, rearrange the pixels of each part (n, 64, 64, 4), reshape into (n, 64, 64, 2, 2), then reshape into (n, 64, 2, 64, 2), and finally reshape into (n, 128, 128, 1), and then splice them together to obtain the feature map (n, 128, 128, 64);

[0067] S2. Generate the training dataset, considering the use of a single antenna at the transmitter and receiver. For channel modeling and pilot transmission, we use the widely used LTE simulator developed at the University of Vienna, the Vienna LTE-A Simulator. The training, test, and validation sets consist of 32000, 4000, and 4000 channels, respectively, where each frame consists of 14 time slots and 72 subcarriers. The wireless channel model for VehicularA (VehA) with a carrier frequency of 2.1GHz, a bandwidth of 1.6MHz, and a UE (User Equipment) speed of 50km / h.

[0068] S3. The method for training super-resolution neural network is as follows:

[0069] Let Θ represent the set of all network parameters. The input of the network is the pilot value vector The output is the estimated channel matrix, expressed as

[0070]

[0071] The overall loss function of the network is the mean square error (MSE) between the estimated channel response and the actual channel response, calculated as follows:

[0072]

[0073] where Γ is the set of all training data, H is the perfect channel, and ‖Γ‖ is the size of the training set.

[0074] Similar to image-based techniques, the optimal weights of the network depend on the value of the SNR; therefore, to get a complete solution, the network must be retrained for each SNR value. Training the network for several SNR values ​​(12dB and 22dB in this method) still leads to good performance.

[0075] In order to enable the network's feature extraction module LR space to better extract features, the real part, imaginary part and amplitude value of the data in the training set need to be extracted and spliced ​​together before training. The original training set size is (32000, 72, 14, 1), and the training set size after splitting should be (96000, 72, 14, 1). The optimizer of all networks is Adam, the training learning rate is set to 0.001, the batch size (Batch_size) is 128, and the training epochs is 300.

[0076] S4. The network performance is tested using the test data set. The normalized mean square error (NMSE) is used to measure the channel estimation performance, which is defined as:

[0077]

[0078] This example is a one-stage model, which has better robustness than the two-stage training model;

[0079] This example uses a recursive gated convolution block to replace the traditional residual convolution block, taking into account the long-distance information interaction between the features in the channel matrix, and extracting the real part, imaginary part and amplitude value of the data in the data set during the training phase, increasing the feature volume of the data and facilitating better learning of the network;

[0080] Compared with traditional channel estimation methods, this example uses a super-resolution neural network to learn channel matrix features, which greatly improves the accuracy and robustness of channel estimation and achieves performance improvement.

[0081] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A channel estimation method based on a super-resolution network, characterized in that: The steps include: Step 1, construct a super-resolution neural network; The super-resolution neural network includes five modules: Conv convolution layer, ReLu activation layer, HorBlock recursive gated convolution block, UpSample upsampling module and Shuffle layer; Step 2, generate training data set; Step 3, training the super-resolution neural network; Step 4: Test the network performance using the test data set.

2. The channel estimation method based on super-resolution network according to claim 1, characterized in that: The Conv convolution layer is used to extract image features and realize parameter sharing, reduce network parameters, and improve training efficiency, wherein the number of convolution kernels is set to 64, the convolution kernel size is set to 3×3, the step size is set to 1, and the padding size is set to 1.

3. The channel estimation method based on super-resolution network according to claim 1, characterized in that: The Relu activation layer activates the feature matrix convolved in the previous step to form a new matrix, which is used to add nonlinear factors to the model.

4. The channel estimation method based on super-resolution network according to claim 1, characterized in that: The HorBlock recursive gated convolution block uses gated convolution and recursive operations to achieve information interaction of any order; the special features extracted by the convolution layer will pass through the Layer Norm layer and then pass through the g n Conv module, which is based on recursive and gated convolutions Recursive GatedConvolutions (g n Conv) enables convolutional neural networks to complete high-order spatial interactions; let the input feature be Gated convolution g n The output of Conv y = g n Conv(x) can be expressed as: in It is a linear projection operation to complete the information exchange in the channel dimension. f is the depth-wise convolution, which is to linearly project the input x and split it according to the channel to get p0 and q0. Then, q0 after the depth convolution calculation is convolved with p0 to get p1. p1 will be output y after another linear projection. The above process demonstrates the first-order spatial interaction, and the interaction between adjacent features p0 and q0 is explicitly introduced through element-wise multiplication. The above process demonstrates the first-order spatial interaction. The same is true for higher-order spatial interactions. First, use the linear projection function Get a set of projection features p0 and This formula divides the input features into channels and then inputs them into the gated convolution for recursive operation: p k+1 =f k (q k )⊙g k (p k ) / α k=0,1,..,n-1 The output features in each operation are scaled by 1 / α to make the training more stable. Since the dimensions of the two must be the same during the spatial interaction, g k It is the dimension mapping function during the operation: Finally, the network outputs the final recursive output q n Input to the projection layer to get g n The final result of Conv; in order to reduce the computational cost of high-order interactive operations, the number of channels at each order is also constrained:

5. The channel estimation method based on super-resolution network according to claim 1, characterized in that: The UpSample upsampling module is a module for increasing the resolution of an image, and its function is to rescale the input image to a desired size according to certain rules; the UpSample upsampling module consists of a conv layer and a shuffle, where shuffle refers to PixelShuffle, which is a specific upsampling method. The implementation of UpSample is to double the number of channels through convolution; then use PixelShuffle to insert the feature maps of the two channels into each other to double the size, so as to reconstruct the final superpixel of the image.

6. The channel estimation method based on super-resolution network according to claim 1, characterized in that: The generation of the training dataset considers the use of a single antenna at the transmitter and receiver; for channel modeling and pilot transmission, a widely used LTE simulator developed at the University of Vienna, the Vienna LTE-A Simulator, is used; The training, test and validation sets consist of 32000, 4000 and 4000 channels respectively, where each frame consists of 14 time slots and 72 subcarriers; for the wireless channel model of VehicularA with a carrier frequency of 2.1GHz, a bandwidth of 1.6MHz and a UE speed of 50km / h.

7. The channel estimation method based on super-resolution network according to claim 1, characterized in that: The specific steps of training the super-resolution neural network are as follows: Let Θ represent the set of all network parameters. The input of the network is the pilot value vector The output is the estimated channel matrix, expressed as The overall loss function of the network is the mean squared error between the estimated channel response and the actual channel response, calculated as follows: Where Γ is the set of all training data, H is the perfect channel, and ‖Γ‖ is the size of the training set; In order to enable the network's feature extraction module LR space to better extract features, the real part, imaginary part and amplitude value of the data in the training set need to be extracted and spliced ​​together before training. The original training set size is (32000, 72, 14, 1), and the training set size after splitting should be (96000, 72, 14, 1). The optimizer of all networks is Adam, the training learning rate is set to 0.001, the batch size (Batch_size) is 128, and the training epochs is 300.

8. The channel estimation method based on super-resolution network according to claim 7, characterized in that: The network performance is tested using the test data set, specifically using the normalized mean square error to measure the channel estimation performance, which is defined as:

Images