A deep learning enhancement-based massive MIMO downlink CSI feedback method

Abstract

The application discloses a large-scale MIMO downlink CSI feedback method based on deep learning enhancement, modifies an encoder network, a decoder network and feedback content on the basis of an existing self-encoding network CsiNet, enhances the feature extraction capability of a CSI matrix, reduces network complexity, and makes the input of a deep learning network and the CSI matrix be in the same order of magnitude. A puzzle solving training strategy is added to balance the relationship between compression tasks and identification tasks in deep learning, retains the integrity of physical information of the CSI matrix, does not lose the distinguishability, and enhances the downlink CSI feedback performance.

Description

Technical Field

[0001] This invention belongs to the field of downlink channel state information acquisition in FDD mode of large-scale MIMO system in wireless communication, and particularly relates to a large-scale MIMO downlink CSI feedback method based on deep learning enhancement. Background Technology

[0002] Massive Multiple Input Multiple Output (MIMO) technology is considered a key technology for 5G. The large number of antennas can improve spatial multiplexing gain and diversity gain, and has shown strong advantages in reducing reflection power and improving spectrum efficiency.

[0003] Channel State Information (CSI) in wireless mobile communication systems includes channel quality, interference level, and channel strength. CSI feedback plays a crucial role in massive MIMO systems, helping base stations adjust radio signal parameters, optimize transmission and reception quality, and improve system performance. This is especially important in Frequency Division Duplex (FDD) systems where channel heterogeneity is poor, necessitating feedback to obtain downlink CSI. During downlink CSI feedback training, user equipment acquires CSI and sends it back to the base station (BS) via the feedback link. Traditional feedback methods employ vector quantization or codebook-based approaches to reduce feedback overhead; however, the overhead generated by these methods is positively correlated with the number of transmit antennas, which is prohibited in massive MIMO systems.

[0004] Compressed sensing-based CSI feedback methods require extensive iterative computation and have stringent channel requirements. With the application of deep learning in mobile communications, significant progress has been made in CSI feedback. The autoencoder network CsiNet pioneered a method combining deep learning and CSI feedback to achieve end-to-end wireless communication, demonstrating that compressed feedback significantly outperforms traditional compressed sensing algorithms and bridging the gap between wireless communication and computer vision. Currently, effectively utilizing deep learning techniques to explore the corresponding physical information and optimization metrics of channel state information is a promising research direction.

[0005] Current deep learning-based CSI feedback methods still have certain shortcomings in analyzing data features. Unlike CSI compression tasks, which aim to preserve as much original data information as possible, CSI recognition tasks aim to identify and distinguish the most representative features of the data, retaining the most discriminative information from the original data. Based on this consideration, employing relevant techniques from the field of computer vision is a valuable research area to balance the compression and recognition tasks in CSI feedback and further enhance the downlink CSI feedback performance in large-scale MIMO systems. Summary of the Invention

[0006] The purpose of this invention is to provide a large-scale MIMO downlink CSI feedback method based on deep learning enhancement, which balances the compression and recognition tasks in CSI feedback to solve the technical problem that existing methods cannot preserve the integrity of the physical information of the CSI matrix without losing its distinguishability, resulting in low downlink CSI feedback performance.

[0007] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0008] A deep learning-enhanced large-scale MIMO downlink CSI feedback method includes the following steps:

[0009] Step S1: Construct a downlink CSI feedback model for a large-scale MIMO system under time-division duplex mode, including the encoder network at the user end and the decoder network at the base station end; generate training, validation, and test sets using the CSOT2100 channel model, and the resulting CSI matrix is ​​represented as H. d ;

[0010] Step S2: Construct a spherical CSI feedback network model based on a convolutional neural network, named SphconvNet, and integrate the channel state information (CSI) matrix H. d The power p is extracted and fed back separately, and the remaining part is represented on a sphere as a spherical CSI matrix. CSI matrix H d The encoder and decoder networks are divided into real and imaginary parts, and the output is the CSI matrix recovered after compression feedback.

[0011] Step S3: Employ a puzzle-solving-assisted training strategy to enhance the CSI feedback network model from Step S2, naming it JP-SphconvNet, and then use the spherical CSI matrix obtained in Step S2. After dividing into four equal parts, the order is shuffled. The index is denoted as vector K, and the shuffled CSI matrix is ​​denoted as... The encoder network in step S2 is input, and the output of the encoder network is... Construct the objective function and find suitable parameter sets for the encoder and decoder networks that satisfy the original CSI matrix H. d The matrix after compressed feedback of the SphconvNet network model The difference between them should be minimized, so that the predicted sequence number can be obtained. The error with the original sequence number K is the smallest.

[0012] Furthermore, step S1 specifically includes:

[0013] Step S101: Construct a base station containing N t One antenna, with N subcarriers. f The downlink received signal model at the m-th subcarrier is calculated using the following formula:

[0014]

[0015] in Let represent the channel vector of the m-th subcarrier, (·) H This indicates the conjugate transpose. x represents the beamforming vector at the transmission end. m Indicates the transmitted signal, n m Indicates additive noise;

[0016] Step S102: Stack the channel vectors in the spatial frequency domain to obtain the downlink CSI matrix. in It is the Nth f The channel vectors of each subcarrier; using the two-dimensional Discrete Fourier Transform (DFT), the downlink CSI matrix in the spatial frequency domain is transformed. Transformed into an approximately sparse angular delay domain CSI matrix H t The formula is shown below:

[0017]

[0018] Among them, F c Indicates a size of N f ×N f The DFT transformation matrix, F t Indicates a size of N t ×N t The DFT transformation matrix;

[0019] Step S103: Based on the sparsity property of the angle delay domain matrix, the angle delay domain CSI matrix H obtained in step S102 is... t Cut off the first N t , and we get a size of N t ×N t Downlink CSI matrix Hd A large number of downlink CSI matrices H were generated using the CSOT2100 channel model simulation. d Used as a training set, validation set, and test set.

[0020] Furthermore, step S2 specifically includes:

[0021] Step S201: Construct a spherical CSI feedback network model based on a convolutional neural network, named SphconvNet, and use the downlink CSI matrix H obtained in step S1. d The power p = ||H d The feedback is extracted separately, and the remaining CSI matrix lies on a sphere, hence the name spherical CSI matrix.

[0022] Step S202: Shape the spherical CSI matrix The encoder network, which splits the input into real and imaginary parts, compresses the codewords s, which are then passed through the decoder network and, together with the separately fed-back power p, are used to reconstruct the CSI matrix. The formula for the entire compression feedback process is shown below:

[0023] s = f en (Η d (8)

[0024]

[0025] Where f en (·) represents the encoder network, f de (·) indicates the decoder network.

[0026] Furthermore, step S3 specifically includes:

[0027] Step S301: A jigsaw puzzle-assisted training method is used to enhance the CSI feedback network model, named JP-SphconvNet. This method treats the compression and recovery of the CSI matrix as a jigsaw puzzle game, and uses the spherical CSI matrix obtained in step S2... Divide the matrix into four equal submatrices, then randomly shuffle these submatrices to obtain a sequence vector K. The rearranged CSI matrix is ​​denoted as K.

[0028] Step S302: Rearrange the disordered CSI matrix With the original spherical CSI matrix The input is fed into the encoder network, which, in addition to outputting the compressed codeword 's', also outputs the prediction sequence.

[0029] Step S303: Find a suitable set of parameters for the encoder and decoder networks that satisfies the original CSI matrix H. d The matrix after compressed feedback of the SphconvNet network model To minimize the difference between the two values ​​and also minimize the error in arranging the sequence numbers, the objective function is constructed as follows:

[0030]

[0031] Where λ represents an adjustable hyperparameter, ||·|| represents the 2-norm, and K i This represents the value of the i-th element in the index vector. This represents the value of the i-th element in the prediction sequence vector;

[0032] Finally, using this objective function as the loss function, the JP-SphconvNet network model is trained, and after convergence, the downlink CSI matrix H is input to the test set. d The output is the downlink CSI matrix after compressed feedback. The normalized mean squared error (NMSE) was used to evaluate the model performance.

[0033] The present invention provides a deep learning-enhanced large-scale MIMO downlink CSI feedback method, which has the following advantages:

[0034] 1. This invention enables autoencoded compression feedback of downlink CSI in large-scale MIMO. By separating the power for feedback, the CSI matrix and the input of the deep learning network can be made to be on roughly the same order of magnitude, thus achieving faster gradient descent.

[0035] 2. This invention employs a more complex encoder network than CsiNet, enabling more in-depth capture of CSI matrix feature information.

[0036] 3. The present invention adopts a jigsaw puzzle-assisted training strategy, which can enable the encoder network to retain as much physical information as possible from the sparse parts, while also retaining the most recognizable information of the CSI matrix. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the implementation of this invention;

[0038] Figure 2 This is a schematic diagram of the SphconvNet structure;

[0039] Figure 3 It is a diagram illustrating the principles of solving a jigsaw puzzle.

[0040] Figure 4 This is a schematic diagram of the JP-SphconvNet structure;

[0041] Figure 5 This is a comparison of the NMSE results of the present invention and the existing method CsiNet on the test set;

[0042] Figure 6 This is a grayscale visualization comparison of the present invention and the existing method CsiNet on the test set. Detailed Implementation

[0043] To better understand the purpose, structure, and function of this invention, the following detailed description of a deep learning-enhanced large-scale MIMO downlink CSI feedback method is provided in conjunction with the accompanying drawings.

[0044] Example 1

[0045] See Figures 1-4 This implementation provides a deep learning-enhanced large-scale MIMO downlink CSI feedback method, specifically as follows: Figure 1 As shown, the method includes the following steps:

[0046] Step S1: Construct a downlink CSI feedback model for a large-scale MIMO system under time-division duplex mode, including the encoder network at the user end and the decoder network at the base station end. The base station end contains N t One antenna, with N subcarriers. f The downlink received signal model at the m-th subcarrier is calculated using the following formula:

[0047]

[0048] in Let represent the channel vector of the m-th subcarrier, (·) H This indicates the conjugate transpose. x represents the beamforming vector at the transmission end. m Indicates the transmitted signal, n m This indicates additive noise.

[0049] The downlink CSI matrix is ​​obtained by stacking the channel vectors in the spatial frequency domain. in Indicates the Nth f The channel vectors of each subcarrier. To reduce feedback overhead, based on the sparsity of the angle delay domain CSI matrix, a two-dimensional Discrete Fourier Transform (DFT) is used to transform the spatial frequency domain CSI matrix into an approximately sparse angle delay domain CSI matrix H. t The formula is shown below:

[0050]

[0051] Among them, F c Indicates a size of N f×N f The DFT transformation matrix, F t Indicates a size of N t ×N t The DFT transformation matrix.

[0052] Since the CSI matrix in the angle delay domain, except for the first N... t Apart from the row values, most of the values ​​in the remaining rows are close to zero, resulting in the angular delay domain CSI matrix H. t Cut off the first N t , and we get a size of N t ×N t Downlink CSI matrix H d A large number of downlink CSI matrices H were generated using the CSOT2100 channel model simulation. d Used as a training set, validation set, and test set.

[0053] Step S2: Construct as follows Figure 2 The spherical CSI feedback network model based on convolutional neural networks shown is named SphconvNet. The downlink CSI matrix H obtained in step S1... d The power p = ||H d The feedback is extracted separately, and the remaining CSI matrix lies on a sphere, hence the name spherical CSI matrix.

[0054] Spherical CSI matrix The encoder network, consisting of real and imaginary parts, is input to SphconvNet. The total number of feedback parameters is N, resulting in an M-dimensional codeword s with a compression ratio γ = M / N. This codeword is then processed by a decoder network assembled at the base station, and together with the separately fed-back power p, is reconstructed into a CSI matrix. The formula for the entire compression feedback process is shown below:

[0055] s = f en (Η d (13)

[0056]

[0057] Among them, f en (·) represents the encoder network, f de (·) indicates the decoder network.

[0058] Step S3: Since the encoder easily discards the propagation path delay and angular physical information carried by the CSI matrix during compression, a CSI feedback network JP-SphconvNet with jigsaw puzzle-assisted training is constructed. The CSI matrix compression and recovery is treated as a jigsaw puzzle game, such as... Figure 3 As shown, the spherical CSI matrix obtained in step S2 Divide the matrix into four equal submatrices, then randomly shuffle these submatrices to obtain a sequence vector K. The rearranged CSI matrix is ​​denoted as K. like Figure 4 As shown, the CSI matrix is ​​rearranged from its original order. With the original spherical CSI matrix The input is fed into the encoder network, which, in addition to outputting the compressed codeword 's', also outputs the prediction sequence.

[0059] Finding suitable sets of parameters for both the encoder and decoder networks that satisfy the original CSI matrix H d The matrix after compressed feedback of the SphconvNet network model To minimize the difference between the two values ​​and also minimize the error in arranging the sequence numbers, the objective function is constructed as follows:

[0060]

[0061] Where λ represents an adjustable hyperparameter, ||·|| represents the 2-norm, and K i and These represent the i-th element value of the sequence number and the predicted sequence number vector, respectively.

[0062] Finally, using this objective function as the loss function, the JP-SphconvNet network model is trained, and after convergence, the downlink CSI matrix H is input to the test set. d The output is the downlink CSI matrix after compressed feedback. The formula for the normalized mean square error between the original CSI matrix and the original CSI matrix is ​​as follows:

[0063]

[0064] Where n represents the number of samples in the training set. This represents the k-th CSI matrix in the training set samples. This represents the CSI matrix after the k-th compressed feedback. It is a commonly used metric for CSI feedback effectiveness.

[0065] like Figure 5 As shown, compared with CsiNet's NMSE on the test set, the performance of the present invention is better than that of CsiNet in indoor scenes with different compression ratios, and has a lower NMSE.

[0066] like Figure 6 As shown, ρ represents similarity. Under the same compression ratio, the present invention has higher similarity and compression feedback recovery performance than CsiNet.

[0067] Any aspects of this invention not described in detail are well-known to those skilled in the art.

[0068] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for CSI feedback of massive MIMO downlink based on deep learning enhancement, characterized in that, The method comprises the following steps: Step S1: Constructing the downlink CSI feedback model of massive MIMO system under the time division duplex system, including the encoder network of the user end and the decoder network of the base station end; generating the training set, the verification set and the test set by using the CSOT2100 channel model, and obtaining the CSI matrix represented as H d ; Step S2: Construct a spherical CSI feedback network model based on a convolutional neural network, named SphconvNet, extract the power p of the channel state information (CSI) matrix H d Separately feed back the remaining part on a sphere, denoted as the spherical CSI matrix Divide the CSI matrix H d Into real and imaginary parts, input into the encoder network and the decoder network, and output the recovered CSI matrix after compression feedback Step S3: using the puzzle solving auxiliary training strategy to enhance the CSI feedback network model in step S2, named JP-SphconvNet, and the spherical CSI matrix obtained in step S2 is input into the encoder network of step S2, and the output of the encoder network is After quartering, the order is disordered, and the order number is recorded as vector K, and the disordered CSI matrix is recorded as Input the encoder network of step S2, and the output of the encoder network is Construct the objective function to find the appropriate parameter set of the encoder network and the decoder network, which can not only satisfy the minimum difference between the original CSI matrix H d And the matrix after the feedback of the SphconvNet network model compression The error between the predicted arrangement number And the original order number K is minimum.

2. The deep learning enhancement based massive MIMO downlink CSI feedback method according to claim 1, characterized in that, The step S1 specifically comprises: Step S101: Constructing the base station end contains N t Root antenna, the number of subcarriers is N f The downlink receiving signal model at the mth subcarrier is calculated as follows: wherein denotes the channel vector of the mthsubcarrier, H denotes the conjugate transpose, denotes the transmit beamforming vector, x m denotes the transmit signal, n m denotes the additive noise; Step S102: Stack the channel vectors in the spatial frequency domain to obtain the downlink CSI matrix. in It is the Nth f Channel vectors of each subcarrier; using the two-dimensional Discrete Fourier Transform (DFT), the downlink CSI matrix in the spatial frequency domain is obtained. Transformed into an approximately sparse angular delay domain CSI matrix H t The formula is shown below: where F c denotes a DFT transform matrix of size N f × N f ; F t denotes a DFT transform matrix of size N t × N t ; Step S103: Based on the sparsity property of the angle delay domain matrix, the angle delay domain CSI matrix H obtained in step S102 is... t Cut off the first N t , and we get a size of N t ×N t Downlink CSI matrix H d A large number of downlink CSI matrices H were generated using the CSOT2100 channel model simulation. d Used as a training set, validation set, and test set.

3. The deep learning enhancement based massive MIMO downlink CSI feedback method according to claim 2, characterized in that, The step S2 specifically comprises: Step S201: Construct a spherical CSI feedback network model based on a convolutional neural network, named SphconvNet, and feed back the power p = ||H d || of the downlink CSI matrix H d obtained in step S1 alone, and the remaining CSI matrix is on a sphere, referred to as a spherical CSI matrix Step S202: Splitting the spherical CSI matrix The encoder network of SphconvNet compresses the real and imaginary parts of the input, and the decoder network recovers the CSI matrix from the code word s and the separately fed back power p The whole compression feedback process is shown in the following formula: s = f en (H d ) (3) where f en (·) denotes an encoder network, f de (·) denotes a decoder network.

4. The deep learning enhancement based massive MIMO downlink CSI feedback method of claim 3, characterized in that, The step S3 specifically comprises: Step S301: using puzzle solving auxiliary training to enhance the CSI feedback network model, named JP-SphconvNet, which regards the CSI matrix compression and recovery as a puzzle solving game, and divides the spherical CSI matrix obtained in step S2 into 4 sub-matrices of equal size, then randomly rearranges the sub-matrices to obtain a sequence vector K, and the CSI matrix after rearrangement is denoted as ​ Step S302: inputting the reordered CSI matrix together with the original spherical CSI matrix to the encoder network, which outputs, in addition to the compressed codeword s, a prediction sequence Step S303: find a suitable and encoded network and decoder network parameter set, both to meet the original CSI matrix H d and the matrix after SphconvNet network model compression feedback The difference between the minimum and the minimum error of the ranking number, and the construction of the objective function is as follows: where λ denotes a tunable hyperparameter, || · || denotes the 2-norm, K i denotes the value of the i-th element of the sequence number vector, denotes the value of the i-th element of the predicted sequence number vector; Finally, the objective function is used as the loss function to train the JP-SphconvNet network model, and after convergence, the downlink CSI matrix H of the test set is input d , and the output is the downlink CSI matrix after compression feedback The normalized mean square error NMSE is used to evaluate the effect of the model.

Images