A deep learning-based massive MIMO downlink CSI feedback method

Abstract

The present application relates to the technical field of wireless communication, and in particular to a large-scale MIMO downlink CSI feedback method based on deep learning, a COST2100 channel model is used to construct a data set, a training set, a verification set and a test set are generated, a deep neural network model is constructed, the deep neural network model comprises an encoder and a decoder, the deep neural network model is trained by using the training set, and the verification set is used to assist in adjusting the training strategy in the training process, the trained deep neural network model is used to test the model performance on the test set, the trained deep neural network model is saved, and the encoder is deployed on a user side, and the decoder is deployed on a base station. On the basis of lower model complexity, higher feedback precision can be realized, the CSI feedback performance is greatly improved, and an efficient and feasible method is provided for large-scale MIMO system downlink CSI feedback.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, and in particular to a large-scale MIMO downlink CSI feedback method based on deep learning. BACKGROUND

[0002] Massive multiple-input multiple-output (MIMO) technology is a key technology of the new generation of wireless communication systems. By deploying a large-scale antenna array at the base station, the system spatial diversity gain and antenna array gain are utilized to greatly reduce multi-user interference and significantly improve the spectral efficiency and energy efficiency. For a large-scale MIMO system, the premise of achieving these goals is that the base station can accurately obtain the downlink channel state information. In a time division duplex (TDD) large-scale MIMO system, the channel state information (CSI) can be obtained by utilizing the reciprocity of the uplink and downlink channels. However, in a frequency division duplex (FDD) mode, since the operating frequency band of the uplink communication link is often different from that of the downlink communication link, the base station cannot obtain accurate CSI in real time through channel reciprocity. Therefore, the base station needs to first send a downlink pilot to the user end, and the user end obtains the CSI by channel estimation on the pilot, and then feeds back to the base station end through the feedback link. The amount of CSI feedback data increases with the increase in the number of large-scale MIMO transceiving antennas, and the contradiction between feedback overhead and system performance gradually intensifies. The traditional CSI feedback algorithm based on compressed sensing (CS) requires prior assumptions of the channel and is highly dependent on the channel sparsity characteristics. Due to the differences between the prior knowledge and the actual situation, the accuracy of the CSI feedback reconstruction is greatly affected. In addition, the CS recovery algorithm involves a large number of matrix iterative calculations, and the computational complexity increases sharply with the increase in the number of antennas, thereby affecting the real-time performance of the communication system.

[0003] In recent years, deep learning (DL) has achieved great success in the fields of computer vision and natural language processing. The data-driven approach based on deep learning provides a new way of thinking for the design of CSI feedback algorithms. Many schemes successfully introduce deep learning techniques into the CSI feedback task. CsiNet is the first deep learning model designed for CSI feedback, which compresses the CSI at the user end by introducing an autoencoder structure, and then reconstructs the CSI at the base station. Compared with traditional CSI feedback methods, CsiNet has better feedback accuracy and lower complexity. Based on CsiNet, researchers have proposed many DL-based CSI feedback models. For example, CRNet first introduced a multi-resolution convolutional architecture, used larger convolution kernels, and adopted residual learning, which improved the feedback performance to some extent. ACRNet uses network aggregation technology and PReLU activation function, and for the first time combines binarization with vector quantization, making the model have good feedback performance.

[0004] However, the above methods cannot balance the model complexity and feedback performance, and the method of increasing the model complexity to improve the feedback accuracy is unacceptable in large-scale MIMO communication systems. SUMMARY

[0005] The purpose of the present application is to provide a large-scale MIMO downlink CSI feedback method based on deep learning, which solves the problem that the method of increasing the model complexity to improve the feedback accuracy is unacceptable in large-scale MIMO communication systems.

[0006] To achieve the above purpose, the present application provides a large-scale MIMO downlink CSI feedback method based on deep learning, comprising the following steps:

[0007] The COST2100 channel model is used to construct a data set, and the data set is preprocessed and divided into a training set, a validation set and a test set;

[0008] A deep neural network model is constructed, which includes an encoder and a decoder. The encoder is deployed at the user end, compresses the CSI matrix into code words according to the preset compression rate, and feeds back to the base station through the feedback link. The decoder is deployed at the base station, and the received code words are reconstructed into the original channel matrix;

[0009] The training set is used to train the deep neural network model, and the validation set is used to assist in adjusting the training strategy during the training process. The trained deep neural network model is tested on the test set to test the model performance.

[0010] The trained deep neural network model is saved, and the encoder is deployed at the user end and the decoder is deployed at the base station.

[0011] The deep neural network model is constructed, which includes an encoder and a decoder, and the step further comprises:

[0012] The encoder is composed of a PF layer, a transpose convolution layer, a convolution layer and a fully connected layer. The first layer PF layer of the encoder contains a PoolFormer module. After the PF layer, the feature mapping is upsampled using transpose convolution. After the transpose convolution layer, the feature mapping is restored using convolution, and then a second convolution is used for feature extraction. Finally, the code word is generated through the fully connected layer.

[0013] The deep neural network model is constructed, which includes an encoder and a decoder, and the step further comprises:

[0014] The decoder is composed of a full connection layer, a PF layer, two convolutional layers and a Sigmoid layer, the full connection layer reconstructs the received code word into the original channel state matrix size, the matrix is input into the first convolutional layer, the features are extracted through convolution, the PF layer composed of three same PoolFormer modules is used after the first convolutional layer for feature extraction and CSI matrix reconstruction, residual learning is applied after each PoolFormer module to improve the overall performance of the model, the processed matrix is subjected to feature fusion and reconstruction through the second convolutional layer composed of transposed convolution and convolution, and finally the reconstructed matrix is output after being processed by the Sigmoid activation function.

[0015] The PoolFormer module is composed of three different types of modules: a normalization layer, an average pooling layer and a multi-layer perceptron.

[0016] The multi-layer perceptron is composed of two linear layers and a Gaussian error linear unit, and the Gaussian error linear unit is arranged between the two linear layers.

[0017] The large-scale MIMO downlink CSI feedback method based on deep learning provided by the application uses the COST2100 channel model to construct a data set, generates a training set, a validation set and a test set, constructs a deep neural network model, the deep neural network model includes an encoder and a decoder, the encoder is arranged at the user end, compresses the CSI matrix into a code word according to a preset compression rate, and feeds back to the base station through a feedback link, and the decoder is arranged at the base station, and reconstructs the received code word into an original channel matrix; the training set is used to train the deep neural network model, and the validation set is used to assist in adjusting the training strategy during the training process, the trained deep neural network model is used to test the model performance on the test set; the trained deep neural network model is saved, and the encoder is arranged at the user end, and the decoder is arranged at the base station. It can realize higher feedback accuracy on the basis of lower model complexity, greatly improve the CSI feedback performance, and provide an efficient and feasible method for large-scale MIMO system downlink CSI feedback. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiment or prior art description will be briefly introduced below.

[0019] Figure 1 is a flowchart of the large-scale MIMO downlink CSI feedback method based on deep learning of the application.

[0020] Figure 2 is a general structure schematic diagram of the PFNet deep learning model of the application.

[0021] Figure 3is a structural schematic diagram of a PoolFormer module of the present application.

[0022] Figure 4 is a structural schematic diagram of an MLP module of the present application.

[0023] Figure 5 is a step diagram of the deep learning-based massive MIMO downlink CSI feedback method of the present application. DETAILED DESCRIPTION

[0024] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, the embodiments described below by reference to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0025] The first embodiment of the present application is:

[0026] Please refer to Figures 1 to 5 , wherein, Figure 1 is a flowchart of the deep learning-based massive MIMO downlink CSI feedback method of the present application. Figure 2 is a general structural schematic diagram of the PFNet deep learning model of the present application. Figure 3 is a structural schematic diagram of a PoolFormer module of the present application. Figure 4 is a structural schematic diagram of an MLP module of the present application. Figure 5 is a step diagram of the deep learning-based massive MIMO downlink CSI feedback method of the present application. The present application provides a deep learning-based massive MIMO downlink CSI feedback method, comprising the following steps:

[0027] S101: Use the COST2100 channel model to construct a data set, and pre-process the data set into a training set, a validation set and a test set;

[0028] Specifically, the COST2100 channel model is used to generate training sets, validation sets and test sets for indoor and outdoor scenes respectively, each containing 100,000, 30,000 and 20,000 data. The specific method is as follows:

[0029] The channel model used in this embodiment deploys N t transmission antennas at the base station and 1 receiving antenna at the user end, and the channel model has N c subcarriers. The signal received by the user end at the Nth subcarrier can be represented as:

[0030]

[0031] In the formula, respectively, represent channel vector and precoding vector respectively, in FDD system, the base station designs the precoding vector based on the received channel matrix. respectively, represent data symbol of downlink transmission and noise of Nth subcarrier. Downlink CSI matrix can be expressed as:

[0032]

[0033] The user end feeds back to the base station, the total number of feedback parameters is 2N t N c In order to reduce the feedback overhead, two-dimensional discrete Fourier transform (DFT) compression method is used to make the channel matrix sparse, as follows:

[0034]

[0035] In the formula And are two DFT matrices, for the angle delay domain channel matrix H, only the first non-zero rows of H are retained, and the remaining most of the rows with zero parameters are deleted. The truncated matrix is expressed as H a The total number of feedback parameters is reduced to 2N t N c The size of the channel matrix is N t *N c , and the number of channels is 2. Through the above steps, H a is fed back instead of H, which can reduce the feedback overhead. H a Data generated by COST2100 channel model is used as a data set for training and testing of the model, wherein N t = 32, N c = 32.

[0036] S102: Construct a deep neural network model, the deep neural network model includes an encoder and a decoder, the encoder is deployed at the user end, compresses the CSI matrix into code words according to a preset compression rate, and feeds back to the base station through a feedback link, and the decoder is deployed at the base station, reconstructs the received code words into the original channel matrix;

[0037] Specifically, a deep learning neural network model PFNet is built, which includes an encoder and a decoder. The decoder is deployed at the user end, used to compress the CSI matrix H to generate code words S and transmit them to the base station through a feedback link, and the decoder is deployed at the base station, reconstructs the code words from the user end into the original CSI matrix The specific method is as follows:

[0038] The encoder in the neural network model consists of a PF layer, a transpose convolution layer, a convolution layer, and a fully connected layer. The matrix H a The feedback codeword s is generated as a dual-channel matrix input encoder. The first layer of the encoder, the PF layer, contains a PoolFormer module. After the PF layer, the feature map is upsampled using a transpose convolution layer with a kernel size of 4*4. After the transpose convolution layer, a convolution with a kernel size of 4*4 is used to restore the feature map, and then a convolution with a kernel size of 5*5 is used for feature extraction. In these operations, the channels of the feature map remain 2. Finally, a fully connected layer is used to generate the feedback codeword s, which is then transmitted to the decoder of the base station. The encoder processing process is defined as follows:

[0039] s = f en (H a )

[0040] where f en and H a represent the encoder model and input parameters, respectively, and s is the compressed codeword of the encoder. The compression rate of the neural network is η = S / 2N t N c .

[0041] The decoder consists of a fully connected layer, a PF layer, two convolution layers, and a Sigmoid layer. The fully connected layer reconstructs the codeword transmitted from the encoder into the original channel state matrix size. The matrix is input into the first convolution layer, which performs 5*5 convolution for feature extraction. After the first convolution layer, a PF layer consisting of three identical PoolFormer blocks is used for feature extraction and CSI matrix reconstruction. After each PoolFormer module, residual learning is applied to improve the overall performance of the model. Subsequently, the processed matrix is passed through a second convolution layer consisting of a 4*4 transpose convolution and a 4*4 convolution for feature fusion and reconstruction. The CSI is continuously reconstructed through these processes, and finally the reconstructed CSI matrix is output through the Sigmoid layer. The decoder processing process is defined as follows:

[0042]

[0043] where f de and s are the function and input parameters of the decoder, is the reconstructed channel matrix of the decoder.

[0044] The PoolFormer module is shown in the right subgraph of Figure 3 , by specifying the average pooling operation as MetaFormer (as Figure 3The token mixer (shown in the left subgraph) obtains PoolFormer, which is composed of three different types of blocks: a normalization layer (LayerNorm), an average pooling layer (AvgPooling), and a multi-layer perceptron (MLP). The MLP module is composed of two linear layers and a Gaussian error linear unit (GELU), as shown in Figure 4

[0045] S103: Train the deep neural network model with the training set, and use the validation set to assist in adjusting the training strategy during the training process. Test the performance of the trained deep neural network model on the test set.

[0046] Specifically, the neural network is trained, and the model compression rate η is set. In this embodiment, the compression rates are set to 1 / 4, 1 / 8, 1 / 16, 1 / 32, and 1 / 64, respectively. Learning rate, batch size, and training round number, etc. Hyperparameters are set. The AdamW optimization algorithm is used as the optimizer, and the model is continuously trained through back propagation to minimize the model loss function. The mean square error is used as the loss function Loss, which is specifically defined as follows:

[0047]

[0048] where N is the number of training set data, H i is the i-th data in the training set, is the corresponding network model output with H i as input.

[0049] The trained model is saved, and the model performance is tested using the test set. The normalized mean square error (NMSE) and the floating point operation number (FLOPs) are used as the evaluation index of the feedback performance, and the specific definition of NMSE is as follows:

[0050]

[0051] where H i is the i-th sample channel matrix in the data set, is the channel matrix compressed and reconstructed by the encoder and the decoder with H i as the model input. Through a large number of experiments, it can be obtained that the neural network algorithm based on deep learning of the PFNet in this embodiment is obviously superior to existing methods in terms of performance of downlink CSI feedback of large-scale MIMO systems. As shown in Table 1, the PFNet proposed in this embodiment has better feedback performance in indoor and outdoor scenarios under different compression ratios, which is specifically embodied in that the PFNet has the lowest FLOPs and NMSE.

[0052] ​Table 1: Performance comparison of the embodiment and other CSI feedback algorithms

[0053]

[0054]

[0055] S104: Save the trained deep neural network model, and deploy the encoder at the user end and the decoder at the base station.

[0056] Specifically, the trained deep neural network model is used, which has better performance and lower computational complexity compared with other deep learning-based feedback ideas, and provides a high-performance method for downlink CSI feedback of large-scale MIMO communication systems.

[0057] The above only discloses one or more preferred embodiments of the present application, and cannot limit the scope of the rights of the present application. Those skilled in the art can understand that the implementation of all or part of the above embodiments, and the equivalent changes made according to the claims of the present application, still belong to the scope covered by the present application.

Claims

1. A deep learning based massive MIMO downlink CSI feedback method, characterized in that, The method comprises the following steps: A dataset is constructed using the COST2100 channel model, and the dataset is preprocessed and divided into a training set, a validation set, and a test set; A deep neural network model is constructed, which comprises an encoder and a decoder. The encoder is deployed at the user end, compresses the CSI matrix into code words according to a preset compression rate, and feeds back the code words to the base station through a feedback link. The decoder is deployed at the base station, and reconstructs the received code words into the original channel matrix. The encoder comprises a PF layer, a transpose convolution layer, a convolution layer, and a fully connected layer. The first layer PF layer of the encoder comprises a PoolFormer module. After the PF layer, the feature mapping is upsampled using transpose convolution. After the transpose convolution layer, the convolution is used to restore the feature mapping, and then the second convolution is used for feature extraction. Finally, the fully connected layer is used to generate the code word. The decoder comprises a fully connected layer, a PF layer, two convolution layers, and a Sigmoid layer. The fully connected layer reconstructs the received code word into the original channel state matrix size. The matrix is input into the first convolution layer, and the features are extracted through convolution. After the first convolution layer, the PF layer composed of three identical PoolFormer modules is used for feature extraction and CSI matrix reconstruction. After each PoolFormer module, residual learning is applied to improve the overall performance of the model. The processed matrix is subjected to feature fusion and reconstruction through the second convolution layer composed of transpose convolution and convolution. Finally, the reconstructed matrix is output after being processed by the Sigmoid activation function. The deep neural network model is trained using the training set, and the validation set is used to assist in adjusting the training strategy during the training process. The trained deep neural network model is used to test the model performance on the test set. The trained deep neural network model is saved, and the encoder is deployed at the user end, and the decoder is deployed at the base station.

2. The deep learning-based massive MIMO downlink CSI feedback method of claim 1, wherein The PoolFormer module is composed of three different types of modules: a normalization layer, an average pooling layer, and a multilayer perceptron.

3. The deep learning-based massive MIMO downlink CSI feedback method of claim 2, wherein The multilayer perceptron is composed of two linear layers and a Gaussian error linear unit. The Gaussian error linear unit is deployed between the two linear layers.

Images