A Channel State Information Feedback Method Applicable to Large-Scale MIMO Systems

Abstract

The present invention discloses a channel state information feedback method applicable to large-scale MIMO systems. Utilizing the end-to-end optimization characteristics of deep learning, a channel state information feedback scheme CQ-CsiNet based on the joint optimization of dimension compression and codeword quantization is proposed. First, based on the autoencoder network, an encoder including receptive fields of different convolutional layers and multi-scale features is designed to fully extract and utilize the features of the channel state information matrix; subsequently, by adding quantization and quantization error modules, the CQ-CsiNet feedback method is made more suitable for the storage and transmission requirements of actual systems; finally, a residual module is added in the design of the decoder, and an attention mechanism is introduced to further utilize the feature information of the channel state information, and the recovery of the channel state information matrix is achieved. The present invention comprehensively considers dimension compression and codeword quantization of the channel state information, has high-precision feedback performance and meets the requirements of actual transmission systems for quantization performance.

Description

Technical Field

[0001] The present invention relates to a CSI feedback method for a large-scale MIMO system based on dimension compression and codeword quantization, belonging to the technical field of wireless communication. Background Art

[0002] As one of the core enabling technologies for 5G, the large-scale multiple-input multiple-output (MIMO) system is expected to meet the growing demand for data traffic. It can significantly reduce multi-user interference while providing a multi-fold increase in network throughput, effectively improving link capacity and resource efficiency. However, these gains are based on the premise that the base station (BS) can obtain effective channel state information (CSI) for both the uplink and the downlink. Obtaining CSI for the uplink is relatively easy. The user equipment (UE) only needs to transmit training pilots, and then the BS estimates the channel response of the user terminal based on the received signals. However, obtaining CSI for the downlink is more difficult.

[0003] CSI feedback is the process in which the user compresses and reconstructs the CSI after obtaining it for the downlink and then feeds it back to the base station.

[0004] In the current CSI feedback design, vector quantization or codebook-based methods are usually adopted to extract and feedback channel features to reduce the feedback overhead. However, these methods will lose channel information to a certain extent, and the generated feedback amount is linearly related to the number of transmit antennas. For a large-scale MIMO system, in order to cope with the rapidly changing environment, the UE needs to frequently feedback CSI. The large number of antennas at the BS greatly increases the dimension of the CSI matrix, and the feedback information volume increases sharply, resulting in a significant increase in the codebook design complexity and the corresponding feedback amount, which will cause a large overhead and occupy more bandwidth resources. Therefore, how to accurately obtain high-dimensional CSI with low feedback overhead has become one of the important problems to be solved in the large-scale MIMO system in the FDD mode.

[0005] In recent years, artificial intelligence represented by deep learning has developed rapidly, and deep learning has also become the focus of attention in the field of intelligent communication. In the existing CSI feedback schemes using deep learning, CsiNet, based on the idea of compressed sensing and deep learning methods, takes advantage of the non-linear restoration of the neural network to compress and feedback CSI. However, most UEs directly transmit floating-point compressed codes to the BS without considering the requirements of the actual communication system for the signal transmission form and ignoring the problem of quantization error in the actual wireless communication system. Therefore, in a large-scale MIMO wireless communication system, deep learning methods can be used to perform dimensional compression and codeword quantization on CSI, and further compress CSI by quantizing the compressed vector of CSI. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a CSI feedback method for a large-scale MIMO system based on dimensional compression and codeword quantization.

[0007] The technical solution adopted by the present invention to solve the above technical problem is to provide a CSI feedback method applicable to a large-scale MIMO system, including the following steps:

[0008] 1. A channel state information feedback method applicable to a large-scale MIMO system, characterized by including the following steps:

[0009] S1. Obtain a channel matrix with a dimension of N t ×N r in the spatial-frequency domain of the downlink channel state information CSI at the user equipment side; N t is the number of transmit antennas and receive antennas configured at the base station side, and N r is the number of transmit antennas and receive antennas configured at the user side;

[0010] S2. Perform a two-dimensional discrete Fourier transform on the channel matrix to obtain a CSI matrix H' that is approximately sparse in the angular-delay domain;

[0011] S3. Retain the first n t rows of H' to obtain a truncated matrix H with a dimension of n t ×N r , where n t is the number of rows with non-zero values in H';

[0012] S4. Use the encoder at the user equipment end in the constructed feedback model CQ-CsiNet to perform feature extraction and compression on the truncated matrix H and then input it into the quantization module. The quantization module then compresses to obtain the feedback CSI through quantization and inverse quantization operations. The decoder of the feedback model CQ-CsiNet deployed at the base station end decompresses and recovers the feedback CSI to obtain the reconstructed channel matrix.

[0013] S5. Adopt an end-to-end training method to minimize the error between the original channel matrix and the reconstructed channel matrix. Train the feedback model CQ-CsiNet with the goal of reducing the error between them.

[0014] S6. After the feedback model CQ-CsiNet is trained, use the original CSI matrix H as the input of the encoder of the feedback model CQ-CsiNet. After the decoder of the feedback model CQ-CsiNet outputs the reconstructed channel matrix, perform a two-dimensional inverse discrete Fourier transform on the reconstructed channel matrix to obtain the reconstructed value of the channel matrix in the spatial-frequency domain as the CSI feedback result obtained at the base station end. After that, for the reconstructed channel matrix perform a two-dimensional inverse discrete Fourier transform to obtain the reconstructed value of the channel matrix in the spatial-frequency domain. as the CSI feedback result obtained at the base station end.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] (1) Aiming at the problem of high CSI feedback overhead in large-scale MIMO systems, the present invention utilizes the feature of end-to-end optimization of deep learning, and based on the autoencoder network structure, proposes a CSI feedback network model CQ-CsiNet that includes dimension compression and codeword quantization. Using the great potential of deep learning to extract complex channel features, it models the channel environment and realizes low-overhead and high-precision CSI feedback.

[0017] (2) The present invention quantizes the compression vector of CSI through the quantization module to further compress CSI, and at the same time uses a quantization error recovery module to reduce the impact brought by quantization errors. From the perspective of enhancing the quantization performance in CSI feedback, the CQ-CsiNet feedback scheme better meets the storage and transmission requirements of the actual system and has better robustness against quantization errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic diagram of the large-scale MIMO wireless communication scenario to which the method of the present invention is applied;

[0019] Figure 2 is a schematic diagram of the channel data processing flow;

[0020] Figure 3 is a schematic diagram of the autoencoder network structure for CSI feedback;

[0021] Figure 4 Schematic diagram of the CQ-CsiNet feedback network structure for CSI feedback;

[0022] Figure 5 Schematic diagram of the encoder network structure for CSI dimension compression;

[0023] Figure 6 Schematic diagram of the quantization module network structure for quantizing the CSI compression vector;

[0024] Figure 7 Schematic diagram of the residual unit network structure introducing the attention mechanism in the decoder network;

[0025] Figure 8 Schematic diagram for comparing the simulation results of traditional methods, CsiNet, and this solution at different compression ratios in indoor and outdoor environments;

[0026] Figure 9 Schematic diagram for comparing the simulation results of this solution with different quantization bit settings in indoor and outdoor scenarios. Detailed implementation manners

[0027] To illustrate the present invention in detail, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0028] As Figure 1 shown, the present invention considers a large-scale MIMO wireless communication system model of a simple single-cell downlink in the frequency-division duplex mode. At the base station, N t >> 1 transmit antennas are configured, and at the user, N r ≥ 1 receive antennas are configured. There are N c >> 1 orthogonal subcarriers in the system. Among them, the received signal of the user on the nth receive antenna and the ith subcarrier can be expressed as:

[0029] y n，i = h n,i H v i x i + z n,i

[0030] Among them, and respectively represent the complex channel vector, the complex precoding vector, the complex data transmission signal for downlink transmission, and the complex additive white Gaussian noise or interference.

[0031] Please refer to Figure 2 , the channel data processing flow includes the following steps:

[0032] S1. At the user equipment side of massive MIMO, obtain the downlink CSI with a dimension of N in the spatial-frequency domain t ×N r channel matrix The CSI matrix of the link can be expressed as the stacking of signal vectors on all subcarriers:

[0033]

[0034] where N t and N r respectively represent the number of transmit antennas and receive antennas configured by the BS and the UE, h represents the signal vector on the subcarrier, and the number of parameters to be fed back to the base station at this time is N t ×N r .

[0035] S2. Since the channel vector shows sparsity in the time-delay domain, perform a discrete Fourier transform on each column vector of the CSI matrix to obtain a channel matrix that is sparse in the time-delay domain. By performing a discrete Fourier transform on each row vector of the CSI matrix, a channel matrix that is sparse in the angle domain can be obtained. Therefore, by performing a two-dimensional discrete Fourier transform on the CSI matrix in the spatial-frequency domain, an approximately sparse CSI matrix H' in the angle-time-delay domain is obtained, expressed as:

[0036]

[0037] where and respectively represent two discrete Fourier transform matrices of different sizes.

[0038] S3. Due to the finiteness of the delay spread, only the first n t (n t < N t ) rows of H' have non-zero values. Therefore, only the first n t rows of H need to be retained, and thus the truncated matrix is obtained, thereby reducing the feedback amount to N = 2n t N r .

[0039] Please refer to Figure 3 , Figure 3 which is the autoencoder network design for downlink CSI feedback of the present invention, including two parts: an encoder and a decoder:

[0040] S4. Based on the idea of compressive sensing and deep learning methods, an autoencoder network model CQ-CsiNet for channel state information feedback is constructed. The encoder is located at the user equipment side and is used for feature extraction and compression of the channel state information. The quantization module further compresses the feedback channel state information through quantization and inverse quantization operations. The decoder is deployed at the base station side and is used for decompression and recovery of the channel features, and restores the received channel codeword into a channel matrix.

[0041] This invention assumes that the user has obtained perfect CSI, that is, the influence of channel estimation is ignored. The encoder belongs to the user and is used for compression of the CSI matrix. By directly learning the best compression from the CSI matrix to design the codeword, it effectively helps to significantly reduce the amount of data transmitted in the downlink channel. Utilizing the sparse characteristic of the channel matrix, once the user obtains the original CSI matrix H, first compress the original N-dimensional CSI matrix H into a floating-point M-dimensional vector, and then discretize this vector through a quantizer to obtain the codeword s. Its principle can be expressed as:

[0042]

[0043] where ε(·) represents the non-linear compression function represented by the encoder based on the neural network, and Θ ε is the parameter of the encoder network, represents the quantization operation. At this time, the compression ratio used is CR = M / N. The number of feedback bits N bit is calculated through the compression ratio CR and the quantization bit B:

[0044] N bit = L × CR × B

[0045] where L is the dimension of the complete CSI. The decoder belongs to the base station and is used for reconstruction of the CSI matrix, that is, restoring the received codeword s into the original channel matrix Once the base station receives the feedback codeword s from the uplink, it first performs inverse quantization on s, and then uses the neural network to reconstruct the CSI matrix Its principle can be expressed as:

[0046]

[0047] where, represents the reconstruction operation, represents the inverse quantization operation, is the parameter of the decoder network.

[0048] The specific CQ-CsiNet feedback network structure is as Figure 4As shown in the figure, the present invention designs neural network layers and trains them to simulate the performance of the encoder and decoder, namely ε(·) and Among them, the encoder ε(·) is used for feature extraction and compression of CSI, and the quantization module further compresses the feedback CSI through quantization and inverse quantization operations, and the decoder is used for feature decompression and channel recovery.

[0049] As Figure 5 shown, when performing feature extraction on CSI, the present invention is based on the change of the receptive field of the convolutional layer and multi-scale feature fusion. In the design of the encoder network, a CSI dimension compression scheme that can extract multi-scale features while ensuring a certain receptive field and relatively reducing network parameters is considered. Specifically, the original CSI matrix H is used as the input of the encoder. In order to perform feature extraction on the real and imaginary parts of the channel matrix, the input channel image will pass through two parallel paths. One path consists of two convolutional layers with a convolutional kernel size of 7×7, and the other path uses cross convolutional kernels with convolutional kernel sizes of 1×5 and 5×1 respectively. By using convolutional kernels of multiple different sizes, the network can adapt to features of different scales and make selections, increasing the network width and the adaptability of the network to different convolutional scales. Moreover, by using multi-resolution convolutional kernels, the encoder network has a multi-scale receptive field, thereby improving its feature extraction ability. Finally, the two-channel feature maps of CSI output by the two paths are connected and merged through a convolutional kernel with a size of 1×1, realizing linear combination of multiple feature maps and information fusion at the same position of different channels, reducing the number of output data feature maps and the parameters in the network while making the effect of data feature extraction and training results more effective and intuitive, and improving the representation ability of the model to a certain extent. Finally, a fully connected layer compresses the vector with length L into a feedback codeword s with length M and inputs it into the quantization module as an input.

[0050] In an actual digital communication system, the CSI feedback signal is transmitted in the form of a bit stream, that is, the data to be transmitted needs to be discretized. This means that in addition to dimension compression, the CSI of a large-scale MIMO wireless communication system also needs to be encoded and quantized. Therefore, the present invention further quantizes the compressed vector with length M output by the encoder. The network structure of the quantization module used to quantize the CSI compression vector is as Figure 6 shown. In the quantization module, quantization operations are used to represent the compressed measurement values with a finite number of bits and convert them into a bit stream for convenient storage and transmission in the actual system. Inverse quantization The operation is used to restore the bitstream into compressed measurement values. Here, the present invention uses the quantization bit B for uniform quantization, and the output is a bitstream with a length of N bits (N bit = M × B) in the form of a bitstream, and each element in the bitstream is 0 or 1. Once the base station receives the feedback through the feedback link, it will first perform an inverse quantization operation on the bitstream, that is, convert the bitstream form back to floating-point numbers, and then use the decoder to reconstruct the CSI. Some information will be lost during the quantization coding process of the original coding, resulting in quantization errors. The existence of quantization errors will not only have a greater impact on the reconstruction accuracy of the network, but also reduce the training speed of the decoder. Here, the present invention uses a quantization error recovery module to extract features from the feedback coding after inverse quantization, making the network closer to the state without quantization loss. Specifically, in this paper, two fully connected layers are used for the CSI coding after the inverse quantization operation, and after normalization processing, an output with the same value range as the quantization error value range is obtained, and then added to the original coding after inverse quantization in the form of a residual to achieve the effect of reducing the quantization error. If the quantization error is regarded as a kind of data augmentation, then the quantization error repair module can be regarded as a self-supervised learning network. Because there is also a branch that can calculate accurate gradients during the backpropagation process, a better encoder network can be obtained.

[0051] Such as Figure 7As shown, the present invention introduces the idea of a residual network in the decoder network design to help recover more detailed features of the CSI. After non-uniform mapping using a fully connected network, two residual units are used. The input in each residual unit will pass through two paths. One path providing a large-resolution view consists of three convolutional layers with convolutional kernel sizes of 3×3, 1×9, and 9×1 respectively. The other path with a smaller resolution contains two convolutional layers with convolutional kernel sizes of 1×5 and 5×1 respectively. Finally, the outputs of the two paths are connected and merged through a convolutional layer with a convolutional kernel size of 1×1, and the ReLU activation function is used in each convolutional layer. At the same time, inspired by the attention mechanism, an attention module is added to the residual module, predicting a corresponding constant weight for each channel in the feature map, which helps to obtain a decoder network focusing on different convolutional feature maps. When obtaining a feature map of size L×H×W, first use global average pooling to obtain an L×1×1 vector, and then use two convolutional neural networks. First, convert this vector into a C×1×1-dimensional vector to reduce the number of channels of the feature map, and then reconstruct it into an L×1×1-dimensional vector to obtain an L-dimensional weight with the same number of channels as the original feature map. Multiply this weight by the original convolutional feature map to obtain the final feature map. The attention module obtains different attention feature weights through continuous learning of the network, making the weights assigned to the effective feature maps larger, which helps to extract more useful feature information, so that the CSI matrix can be better recovered and the fitting ability of the model can be greatly improved. In addition, a residual connection from the input to the output is constructed in each residual module, which can avoid the problem of network gradient disappearance while retaining more information.

[0052] S5. The feedback model CQ-CsiNet is trained in an end-to-end training manner to continuously reduce the error between the original channel matrix H and the reconstructed channel matrix and finally optimize the parameters of the channel state information feedback network CQ-CsiNet;

[0053] Specifically, the AdamW optimization algorithm is used to minimize the normalized mean square error loss function to update the parameters of the feedback network model CQ-CsiNet. Among them, the mean square error loss function is:

[0054]

[0055] where N represents the total number of samples of the training data, ‖·‖ 2 is the Euclidean norm, H i is the channel state information matrix approximately sparse in the angular-delay domain corresponding to the i-th sample in the training set, is the reconstructed H iThe estimated value. At the same time, a mini-batch training scheme is introduced, with the batch size set to 200, and the number of training epochs is set to 100 according to the network convergence situation. The learning rate is initially set to 0.001 and then set to 0.0001 after the network is basically converged, so as to obtain a further decrease in the normalized mean square error.

[0056] S6. Perform a two-dimensional inverse discrete Fourier transform on the reconstructed channel matrix to obtain the reconstructed value of the channel matrix in the space-frequency domain ;

[0057] S7. Apply the CQ-CsiNet feedback network model trained in step S5 to the dimension compression, quantization, and reconstruction of channel state information in different scenarios.

[0058] The simulation results of the traditional CSI feedback method, CsiNet, and this scheme under different compression ratios in indoor and outdoor environments are compared as Figure 8 shown. The CQ-CsiNet feedback network proposed in the present invention has obtained the lowest NMSE value, and is significantly better than the method based on compressive sensing at all compression ratios. When the compression ratio is reduced to 1 / 16, the method based on traditional compressive sensing can no longer work, while CsiNet and CQ-CsiNet continue to perform well. Compared with CsiNet, CQ-CsiNet also provides obvious advantages, which is due to the complex deep learning structure in the encoder and decoder. In addition, CSI compression feedback through CQ-CsiNet can be performed with relatively low overhead because CQ-CsiNet only requires a few simple matrix-vector multiplications.

[0059] The simulation results of this scheme under different quantization bit settings in indoor and outdoor scenarios are compared as Figure 9 shown. The simulation results show that when the quantization bits are fixed, the feedback performance of the model improves with the increase of feedback bits. When the feedback bits are the same, the feedback performance of the model decreases with the increase of quantization bits, indicating that the CQ-CsiNet feedback scheme can reduce the influence of quantization errors.

[0060] The above are embodiments for explaining the technical idea of the present invention, and the protection scope of the present invention cannot be limited thereby. Based on the professional knowledge of ordinary technical personnel in the art, on the basis of the above description, other different forms of changes or modifications can be made without departing from the purpose of the present invention. Any modifications, substitutions, and improvements made on the basis of the technical solution proposed in the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A channel state information feedback method applicable to a large-scale MIMO system, characterized in that, it includes the following steps: S1. Obtain the downlink channel state information CSI at the user equipment side, where the dimension in the spatial-frequency domain is N t ×N r channel matrix N t is the number of transmit antennas and receive antennas configured for the base station side, and N r is the number of transmit antennas and receive antennas configured for the user side; S2. Perform a two-dimensional discrete Fourier transform on the channel matrix to obtain a CSI matrix H' that is approximately sparse in the angular-delay domain; S3. Retain the first n rows of H'. t Perform operations to obtain a truncated matrix H of dimension n t ×N r , where n is the number of rows with non-zero values in H'; t ​ S4. The feedback model CQ-CsiNet includes an encoder for feature extraction and compression of CSI, a quantization module, and a decoder for decompression and channel recovery. The encoder located at the user equipment end in the constructed feedback model CQ-CsiNet is used to perform feature extraction and compression on the truncated matrix H and then input it into the quantization module. The quantization module then compresses to obtain the feedback CSI through quantization and inverse quantization operations. The decoder of the feedback model CQ-CsiNet deployed at the base station end decompresses and recovers the feedback CSI to obtain the reconstructed channel matrix Among them, the quantization module performs quantization operations, inverse quantization operations, and quantization error recovery operations. The quantization operation is to convert the input codeword of the feedback CSI into a bit stream. The inverse quantization operation is used to restore the bit stream into the encoding of the feedback CSI. The quantization error recovery module is used to extract the features of the encoding of the feedback CSI after inverse quantization, so that the output of the quantization module is close to the state without quantization loss; The quantization operation is specifically: after uniform quantization using quantization bits, the output is a bit stream; The quantization error recovery is specifically: the encoding of the feedback CSI after the inverse quantization operation is processed using two fully connected layers and then normalized to obtain an output with the same value range as the quantization error value range, and then added to the encoding of the feedback CSI in a residual manner and used as the feedback CSI output of the quantization module; S5. Adopt an end-to-end training method to train the feedback model CQ-CsiNet with the goal of reducing the error between the original channel matrix and the reconstructed channel matrix ; and train the feedback model CQ-CsiNet with the goal of reducing the error between the original channel matrix and the reconstructed channel matrix After the feedback model CQ-CsiNet is trained, the original CSI matrix H is used as the encoder input of the feedback model CQ-CsiNet, and the decoder output of the feedback model CQ-CsiNet reconstructs the channel matrix After that, for the reconstructed channel matrix perform a two-dimensional inverse discrete Fourier transform to obtain the reconstructed value of the channel matrix in the spatial-frequency domain as the CSI feedback result obtained at the base station side.

2. The method according to claim 1, characterized in that, In step S1, the channel matrix is a stack of signal vectors on all subcarriers.

3. The method according to claim 1, characterized in that, In step S2, each column vector of the CSI matrix is subjected to a discrete Fourier transform to obtain a channel matrix that is sparse in the time delay domain, and each row vector of the CSI matrix is subjected to a discrete Fourier transform to obtain a channel matrix that is sparse in the angle domain. The CSI matrix H' that is approximately sparse in the angle-time delay domain is expressed as: Among them, and respectively represent two different discrete Fourier transform matrices.

4. The method according to claim 1, characterized in that, In step S4, the encoder compresses the input N-dimensional CSI matrix H into a floating-point M-dimensional vector by using the sparse characteristics of the channel matrix, and then discretizes the vector through a quantizer to obtain the codeword s of the feedback CSI: where ε(·) represents a non-linear compression function based on the encoder representation of the neural network, and Θ ε are the parameters of the encoder, represents the quantization operation, and the compression ratio CR = M / N used, and the number of feedback bits N bit is: N bit = L × CR × B where L is the dimension of the complete CSI, B is the quantization bit, and CR is the compression ratio; After receiving the codeword s of the feedback CSI, the decoder restores the codeword s of the feedback CSI to the reconstructed channel matrix Among them, represents a reconstruction operation, represents an inverse quantization operation, are the parameters of the decoder network.

5. The method according to claim 4, characterized in that, The encoder processes the input channel matrix through two feature extraction paths. One path consists of two convolutional layers with a convolutional kernel size of 7×7, and the other path first passes through a convolutional layer with a convolutional kernel size of 3×3 and then uses cross convolutional kernels with convolutional kernel sizes of 1×5 and 5×1 respectively; the two-channel CSI feature maps output by the two feature extraction paths are connected and merged through a convolutional kernel with a size of 1×1, and finally the merged vector is compressed into a codeword s of the feedback CSI with a length of M through a fully connected layer.

6. The method according to claim 5, characterized in that, The decoder uses a fully connected network to perform non-uniform mapping on the input and uses two residual units with an attention mechanism; Each residual unit contains two paths. One path provides a large-resolution view and consists of three convolutional layers with kernel sizes of 3×3, 1×9, and 9×1 respectively. The other path with a smaller resolution contains two convolutional layers with kernel sizes of 1×5 and 5×1 respectively. Finally, the outputs of the two paths are connected and merged through a convolutional layer with a kernel size of 1×1. Each convolutional layer uses the ReLU activation function. The merged vector is first subjected to global average pooling, and then two convolutional layers with kernel sizes of 1×1 are used to reduce the number of channels of the vector, obtaining weights with the same number of channels as the original CSI matrix. This weight is multiplied by the original CSI matrix to obtain the final reconstructed channel matrix Each residual unit also includes a residual connection from the input to the output to avoid the problem of network gradient vanishing.

Images