Massive MIMO channel state transmission method based on mixed partition and complex valued network

Abstract

The invention provides a Massive MIMO (Multiple Input Multiple Output) channel state transmission method based on mixed partitioning and a complex valued network. The method mainly comprises two parts, wherein the first part provides a channel environment mixed partitioning method; through a dual hybrid partitioning technology based on position information and channel characteristics, the method can effectively reduce the model complexity, significantly improve the system efficiency, expandability and overall performance, and reduce the calculation overhead of the user equipment at the same time. The second part provides a lightweight complex value neural network model, complex value characteristics of a channel state can be fully explored and utilized, and meanwhile high training efficiency and reasoning efficiency of the model are guaranteed through lightweight design. By combining the above two methods, the invention realizes higher accuracy and efficiency of channel state feedback transmission compared with the traditional technology.

Description

Technical Field

[0001] The present invention relates to the field of deep learning and wireless communication physical layer, and in particular to research fields such as channel partition modeling and channel state information transmission based on deep learning. Specifically, it relates to a Massive MIMO channel state transmission method based on hybrid partitioning and complex-valued network, which is used to improve the accuracy of Massive MIMO system in acquiring channel state information (Channel State Information, CSI) and improve the efficiency of CSI compression reconstruction. Background Art

[0002] Massive Multiple-Input Multiple-Output (Massive MIMO) is one of the key technologies in today's 5G communication system and future 6G communication technology. The Massive MIMO system greatly improves the capacity, spectrum efficiency and coverage of the overall system by deploying a large number of antennas (the number of antennas ≥ 64) at the base station (BS), and further improves the data transmission rate and system stability. For the Massive MIMO system, it needs to rely on accurate channel state information (CSI) information to play its role. In the Massive MIMO system using frequency division duplexing (FDD), because the uplink and downlink channel links work at different frequencies, the uplink and downlink channels lose reciprocity, and the base station cannot obtain the CSI characteristics of the downlink channel from the base station to the user equipment (UE). At this time, the CSI of the downlink channel needs to be estimated by the user equipment UE and transmitted back to the base station BS through the uplink channel. However, due to the large number of antennas and the huge amount of channel CSI data in the Massive MIMO system, the user equipment UE generates huge transmission overhead and bandwidth consumption when feeding back the downlink channel CSI to the base station BS. Therefore, it is of great practical significance to study the accurate and efficient transmission method of channel state information CSI.

[0003] At present, the optimization methods for CSI feedback transmission are mainly based on codebook methods and compressed sensing (CS) technology. In the codebook-based method, first, the user equipment UE estimates the CSI of the downlink channel through the reference signal sent by the base station BS, and then the UE selects the optimal precoding matrix from the predefined codebook Codebook according to the CSI, and records the index of the matrix, which is called the precoding matrix indicator (PMI). Next, the UE transmits the PMI and some other channel indicators to the base station BS. The BS further reconstructs the optimal precoding matrix and estimates the CSI of the downlink channel through the feedback information of the UE and the codebook Codebook. This codebook-based CSI transmission method does not directly transmit CSI information, but estimates CSI through the predefined codebook Codebook and PMI. However, when the number of antennas increases, the size of the codebook and the complexity of searching for the optimal precoding matrix will increase significantly, so that when it is applied in the Massive MIMO system, there is still a large time and space overhead. In addition to the codebook-based approach, another CSI transmission optimization method is compressed sensing technology. Compressed sensing technology reduces transmission overhead by compressing information using the sparsity of the CSI matrix. However, the reliability of the CSI matrix sparsity assumption cannot be guaranteed. For example, in a complex communication environment, the sparsity of the CSI matrix is ​​not obvious, resulting in large reconstruction errors. In addition, compressed sensing technology usually involves iterative algorithms with high computational complexity, especially in the scenario of high-dimensional CSI in Massive MIMO systems, where the computational overhead will be very high.

[0004] With the rapid development of deep learning (DL) technology and the upgrading and iteration of hardware, DL technology has gradually been applied to various tasks in the field of wireless communications, such as beamforming, interference management, channel estimation and prediction, etc. In the CSI feedback transmission task, DL technology has also been applied, and it has shown higher CSI recovery accuracy than traditional methods and reduced transmission overhead to a certain extent. However, current DL research has not fully utilized the characteristics of the wireless channel environment. Current research focuses on developing DL models at the granularity level of the BS site to perform CSI compression on the UE side and CSI reconstruction on the BS side. However, within the scope of a site, the channel environment may have different feature distributions and change characteristics in different sub-areas, resulting in poor performance of DL models in sub-areas with significant differences. In addition, traditional DL models cannot fully explore the characteristics of CSI complex-valued elements. From the above perspectives, the accuracy and efficiency of CSI feedback transmission in Massive MIMO systems can be further optimized.

[0005] In summary, it is necessary to design a Massive MIMO channel state transmission method based on hybrid partitioning and complex-valued networks. Summary of the invention

[0006] The purpose of the present invention is to address the problem that in the current Massive MIMO system, the deep learning model does not fully explore the characteristics of the wireless channel environment and the traditional neural network cannot fully utilize the characteristics of the CSI complex-valued elements, resulting in insufficient efficiency and accuracy of CSI feedback transmission. A Massive MIMO channel state transmission method based on hybrid partitioning and complex-valued networks is proposed. The present invention is based on the proposed hybrid partitioning method HZPM (Hybrid Zone Partitioning Method) at a finer granularity level to more fully explore the characteristics of the wireless channel environment, while further reducing the number of parameters and complexity of the model; based on the proposed lightweight complex-valued network LCVCsiNet (Lightweight Complex-Valued CSI Network), the complex-valued characteristics of the channel CSI are more fully utilized to further improve the accuracy of CSI feedback transmission.

[0007] The object of the present invention is achieved in this way:

[0008] On the one hand, the present invention provides a channel hybrid partitioning method HZPM with finer granularity than the site granularity level, which divides a large area within a site range into multiple relatively small sub-areas. In essence, this is to divide a single large and complex model into multiple small and targeted models. This method can effectively reduce the complexity of the model, and the system efficiency, system scalability, and overall performance are significantly improved, while reducing the computational burden on the user equipment UE side. On the other hand, the present invention provides a lightweight complex-valued neural network model LCVCsiNet to fully explore the complex value characteristics of the channel CSI elements. In addition, the lightweight design ensures the high training efficiency and reasoning efficiency of the model. Combining the above two methods, the present invention achieves higher CSI feedback transmission accuracy and efficiency.

[0009] Specifically, the technical solution adopted by the present invention includes the following steps:

[0010] Step 1: Sample and obtain a data set for model training and evaluation, including channel state information CSI and corresponding user equipment UE location information.

[0011] Step 2: First, perform preliminary area division based on location information within the coverage area of ​​the site. Specifically, let the set of user equipment UEs within a given site be X, where the location of each UE is represented by Where m is the dimension of the location information, N is the number of elements in the set X, and the location of the base station BS in the site is expressed as Therefore, the distance from the i-th user equipment UE to the base station BS can be expressed as:

[0012] d i =||x i -x BS ||,i∈{1,2,...,N}

[0013] Where ||·|| represents the Euclidean distance. Further, M threshold values ​​are set, and the entire site coverage is divided into M coarse partitions {C1, C2, ..., C M Each user equipment is allocated to a corresponding coarse zone according to its distance to the base station.

[0014] Step 3: In each coarse partition based on the location information in step 2, further fine-grained partitioning is performed according to the channel state characteristics in the area. This step includes the following sub-steps.

[0015] Step 3.1: Construct a similarity matrix of channel state characteristics. The similarity matrix reflects the degree of similarity of channel states, and can further divide sub-areas by similarity. For a set of user equipment UEs X within the site range, the channel state information of the i-th UE is represents the complex field, N t Indicates the number of BS transmitting antennas, N c represents the number of subcarriers. Then for each coarse partition C j , the channel state information set in its area can be expressed as:

[0016]

[0017] Furthermore, for each coarse partition C j , construct the similarity matrix Where |C j | represents the number of user devices in the corresponding coarse partition. j , the element S in row k and column l j,kl represents the similarity of the channel state information between user equipment k and user equipment l, and is calculated using a Gaussian kernel function.

[0018]

[0019] Among them, σ is a parameter used to control the similarity attenuation, which can be obtained from all channel states H i and H j The median of the Euclidean distances is determined.

[0020] Step 3.2: Calculate the similarity matrix S j The Laplacian matrix of the similarity graph is decomposed and dimensionality is reduced. The Laplacian matrix encodes the global structure in the similarity graph. By further decomposing the Laplacian matrix, the original high-dimensional channel state features can be mapped to a low-dimensional space. First, the degree matrix D of the similarity matrix is ​​calculated. j , which is a diagonal matrix, each diagonal element represents a coarse partition C j The sum of similarities between each UE and other UEs in the . Then, according to the degree matrix D j , construct the Laplace matrix L j , which can be expressed as:

[0021] L j =D j -S j

[0022] Further, through L j Construct the normalized Laplace matrix L j,sym :

[0023]

[0024] The normalized Laplacian matrix L j,sym Perform eigenvalue decomposition to obtain its eigenvalues ​​and eigenvectors:

[0025] L j,sym u k =λ k u k

[0026] where λ k is the eigenvalue, u k is its corresponding eigenvector. Select the first K j The eigenvectors corresponding to the smallest eigenvalues ​​form the matrix The i-th row represents the feature embedding representation of the channel state of user equipment i in the low-dimensional space. j Each row u i Perform normalization to obtain the normalized matrix

[0027] Step 3.3: Perform K-Means clustering of low-dimensional features and obtain mixed partitioning results. This step identifies clusters with similar channel features in the low-dimensional space. By clustering the channel features in the feature vector space, a higher-quality cluster partition can be obtained. Specifically, the normalized feature matrix Each line of As a new data point, K-Means method is applied in low-dimensional space.j After clustering is completed, we can get the coarse partition C j A finer-grained channel state feature partition C under j,k ,k={1,2,..,K j Finally, the hybrid partitioned HZPM can be represented as a two-layer partition mapping as a whole. The first layer is a coarse partition mapping based on location information, and the second layer is a detailed partition mapping based on channel state characteristics.

[0028] Step 4: Preprocess the channel state feature data set in each sub-area. The preprocessing operation includes performing CSI two-dimensional discrete Fourier transform and data set partitioning. Channel state information H i The representation in the angle-delay domain is usually sparse. To reduce the subsequent calculation and storage overhead, the channel state H i Transform to the angle-delay domain H through two-dimensional discrete Fourier transform (2D DFT) i ′. And due to the finiteness of the path propagation delay, H i ′ performs lossless truncation in the delay direction, retaining only some columns containing valid information, thereby reducing storage and processing overhead. Finally, the user channel matrix data set within the range of the base station BS is And through the hybrid partitioning HZPM method, each user UE has and only has one belonging sub-area, so the channel matrix data set Can be divided into non-overlapping sub-datasets:

[0029]

[0030] in, c≠c′ and c,c′∈{1,2,...,C}.

[0031] Step 5. Construct a lightweight complex-valued neural network model LCVCsiNet, which includes the encoder compression model on the UE side and the decoder reconstruction model on the BS side, and use the preprocessed data set to iteratively train the model parameters of each region. In the encoder part, a lightweight model design is adopted to reduce the number of intermediate feature map channels, and the large convolution kernel is decomposed into two small convolution kernels in orthogonal directions to reduce computational overhead. In addition, complex-valued convolution, complex-valued batch normalization, and channel feature multi-dimensional extraction modules (real and imaginary features, angle domain features, and delay domain features) are designed in the encoder to maximize the exploration and utilization of CSI characteristics. Overall, the encoder compression process can be expressed as follows:

[0032]

[0033] Among them, H is the channel state matrix of the model input, represents the parameters of the encoder model, It represents the codeword of length L after being compressed by the UE encoder. The compressed codeword is fed back to the BS via the uplink, and then the CSI information is restored by the decoder model on the BS. The overall decoder reconstruction process can be expressed as follows:

[0034]

[0035] in, Represents the parameters of the decoder model. For a specific region c, the encoding and decoding process of the model can be expressed as follows:

[0036]

[0037] in, is the parameter set of region c. Finally, the channel transmission model of each region is further trained by optimizing the mean square error loss function. The above steps fully explore the characteristics of the wireless channel environment and utilize the characteristics of the CSI complex value element through the hybrid partitioning method HZPM and the complex value neural network modeling, further improving the accuracy of CSI feedback transmission and reducing the number of parameters and complexity of the model.

[0038] The positive effects of the present invention are:

[0039] (1) The present invention proposes a channel hybrid partitioning method HZPM with finer granularity than the site granularity level, which can more fully explore and utilize the characteristics of the wireless channel environment. The HZPM method can effectively reduce the model complexity through the dual hybrid partitioning technology based on location information and channel characteristics, and significantly improve the system efficiency, system scalability, and overall performance, while reducing the computational burden on the user equipment UE side.

[0040] (2) The present invention proposes a lightweight complex-valued neural network model LCVCsiNet, which can fully exploit the complex-valued characteristics of the channel CSI elements, thereby improving the accuracy of CSI compression reconstruction. At the same time, the lightweight design ensures the high training efficiency and reasoning efficiency of the model. Furthermore, by combining with the hybrid partitioning method HZPM, the present invention achieves higher CSI feedback transmission accuracy and efficiency than traditional technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the overall process of the implementation method of the present invention.

[0042] Figure 2 It is a partition display diagram of the hybrid partitioning method HZPM.

[0043] Figure 3 It is the structural diagram of the complex-valued feature map in LCVCsiNet.

[0044] Figure 4 This is the network structure diagram of the UE-side encoder compression model in LCVCsiNet.

[0045] Figure 5 This is the network structure diagram of the BS-side decoder reconstruction model in LCVCsiNet. DETAILED DESCRIPTION

[0046] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings.

[0047] like Figure 1 As shown, the present invention is generally divided into three processes. The first is to sample and obtain a data set for model training and evaluation, including channel state information CSI and the location information of the corresponding user equipment UE. The second step is to divide the sub-areas through the hybrid partitioning method HZPM proposed in the present invention, including sub-processes such as position partitioning and channel feature partitioning. The third step is to establish a CSI transmission model in each area of ​​HZPM through the lightweight complex-valued network model LCVCsiNet proposed in the present invention, including sub-processes such as data set preprocessing, codec model construction and model training. Below, the steps of the technical solution adopted by the present invention are described in detail.

[0048] Step 1: Sample and obtain a data set for model training and evaluation, including channel state information CSI and corresponding user equipment UE location information.

[0049] Step 2: Within the site coverage area, first make a preliminary area division based on the location information. Figure 2 It is a partition display diagram of the hybrid partitioning method HZPM, where each solid circle represents a location area boundary. Specifically, let the set of user equipment UEs within a given site be X, where the location of each UE is represented by Where m is the dimension of the location information, here m=2, N is the number of elements in the set X, and the location of the base station BS in the site is expressed as Therefore, the distance from the i-th user equipment UE to the base station BS can be expressed as:

[0050] d i =||x i -x BS ||,i∈{1,2,...,N}

[0051] Where ||·|| represents the Euclidean distance. Further, set M threshold values ​​{D1, D2, ..., D M}, satisfying 0=D0 <D1<D2<...<D M According to the distance between each user equipment UE and the BS, the entire site coverage is divided into M coarse partitions:

[0052] C j ={x i |D j-1 ≤d i ≤D j ,i∈{1,2,...,N}},j=1,2,...,M

[0053] Finally, we get M coarse partitions {C1,C2,...,C M}, each user equipment i is assigned to a corresponding coarse partition according to its distance to the base station:

[0054] f1(x i )=j, if D j-1 ≤d i ≤D j

[0055] Step 3: In each coarse partition based on the location information in step 2, further fine-grained partitioning is performed according to the channel state characteristics in the area. Figure 2 In , each dotted line represents the boundary line of the channel characteristic subdivision area in the position coarse partition. Specifically, this step includes the following sub-steps.

[0056] Step 3.1: Construct a similarity matrix of channel state features. The similarity matrix reflects the similarity of channel states, and can further divide sub-regions by similarity. The similarity matrix S j is a symmetric matrix, and the matrix elements represent the similarity between each data point. For the user equipment UE set X within the site range, the channel state information of the i-th UE is represents the complex field, N t Indicates the number of BS transmitting antennas, N c represents the number of subcarriers. Then for each coarse partition C j , the channel state information set in its area can be expressed as:

[0057]

[0058] Furthermore, for each coarse partition C j , construct the similarity matrix Where |C j | represents the number of user devices in the corresponding coarse partition. j , the element S in row k and column l j,klrepresents the similarity of the channel state information between user equipment k and user equipment l, and is calculated using a Gaussian kernel function.

[0059]

[0060] Among them, σ is a parameter used to control the similarity attenuation, which can be obtained from all channel states H i and H j Based on the above similarity measure, the coarse partition C can be obtained. j The similarity matrix S j :

[0061]

[0062] Step 3.2: Calculate the similarity matrix S j The Laplacian matrix of the similarity graph is decomposed and dimensionality is reduced by eigenvalue decomposition. The Laplacian matrix encodes the global structure in the similarity graph. By further performing eigenvalue decomposition on the Laplacian matrix, the original high-dimensional channel state features can be mapped to a low-dimensional space to obtain a low-dimensional embedded representation of the channel features. In the mapped low-dimensional space, the structure of the channel feature data is more easily recognized by traditional clustering algorithms, and then the channel feature sub-regions can be divided by clustering. First, the degree matrix D of the similarity matrix is ​​calculated j , which is a diagonal matrix, each diagonal element can be expressed as D j,kk ,k={1,2,...,|C j |}, indicating C j The sum of similarities between the kth UE and other UEs in:

[0063]

[0064] Among them, |C j | indicates coarse partition C j Then, according to the degree matrix D j , construct the Laplace matrix L j , which can be expressed as:

[0065] L j =D j -S j

[0066] Further, through L j Construct the normalized Laplace matrix L j,sym :

[0067]

[0068] The normalized Laplacian matrix L j,symPerform eigenvalue decomposition to obtain its eigenvalues ​​and eigenvectors:

[0069] L j,sym u k =λ k u k

[0070] where λ k is the eigenvalue, u k is its corresponding eigenvector. Select the first K j The eigenvectors corresponding to the smallest eigenvalues ​​form the matrix The i-th row represents the feature embedding representation of the channel state of user equipment i in the low-dimensional space. j Each row u i After normalization, we get the matrix Each of these actions:

[0071]

[0072] Step 3.3: Perform K-Means clustering of low-dimensional features and obtain mixed partitioning results. This step identifies clusters with similar channel features in the low-dimensional space. By clustering the channel features in the feature vector space, a higher-quality cluster partition can be obtained. Specifically, the normalized feature matrix Each line of As a new data point, K-Means method is applied in low-dimensional space. j After clustering is completed, we can get the coarse partition C j A finer-grained channel feature partition C under j,k ,k={1,2,..,K j For user equipment i, the subdivision area to which it finally belongs can be determined in the following manner.

[0073] f2(H i ; j) = k, if H i ∈C j,k ,x i ∈C j

[0074] Finally, if Figure 2 As shown, the hybrid partition HZPM can be represented as a two-layer partition map. The first layer is a coarse partition map based on location information:

[0075]

[0076] Among them, f1(x i ) = j means that user equipment i is allocated to coarse partition C according to the distance from the BS jThe second layer is the segmentation mapping based on channel state characteristics:

[0077]

[0078] Among them, f2(H i j) = k means that user equipment i is allocated to subdivided area C in coarse area j according to channel state characteristics j,k .

[0079] Step 4: Preprocess the channel state feature data set in each sub-area. The preprocessing operation includes two-dimensional discrete Fourier transform and data set partitioning. In the Massive MIMO system working in frequency division duplex FDD mode, the base station BS is set to deploy N uniform linear antenna arrays (ULA). t = 64 transmitting antennas, the number of subcarriers modulated by Orthogonal Frequency Division Multiplexing (OFDM) technology is N c =256, the user equipment UE is configured with N r Root receiving antenna (to simplify the model, set N r =1, that is, the user equipment is configured with a single antenna). Then in the downlink transmission from the BS to the UE, the signal received by the i-th user UE can be represented as:

[0080] y i =H i x+z,i={1,2,...,N}

[0081] in Indicates the signal received by the UE, represents the channel state information matrix of the downlink from BS to UE, represents the Gaussian white noise of the channel, each element is an independent and identically distributed random variable and obeys the complex Gaussian distribution. Since the channel matrix H i The representation in the angle-delay domain is usually sparse. To reduce the subsequent calculation and storage overhead, the channel matrix H i Transformed to the angle-delay domain through a two-dimensional discrete Fourier transform (2D DFT), it can be expressed as follows:

[0082]

[0083] in, and are the DFT matrices in angle and delay directions respectively, Indicates F d The physical meaning of the conjugate transpose is to transform the left matrix in the column direction and transform the frequency domain into the delay domain. After transforming to the angle-delay domain, due to the finiteness of the path propagation delay, H i ′ Perform lossless truncation in the delay direction, retaining only the first N containing valid information c ′=64 columns to reduce subsequent storage and processing overhead. Finally, the channel matrix of the i-th user UE is The user channel matrix data set within the base station BS is And through the hybrid partitioning HZPM method, each user UE has and only has one belonging sub-area, so the channel matrix data set Can be divided into non-overlapping sub-datasets:

[0084]

[0085] in, c≠c′ and c,c′∈{1,2,...,C}.

[0086] Step 5. Construct a lightweight complex-valued neural network model LCVCsiNet, including the encoder compression model on the UE side and the decoder reconstruction model on the BS side, and use the preprocessed data set to iteratively train the model parameters of each region. The convolution operation of complex values ​​in LCVCsiNet can be performed in the following way. Suppose the convolution kernel of the complex-valued convolution layer is Q=A+iB, where A and B are real matrices, i is an imaginary unit, the complex-valued vector is v=x+iy, x and y are real vectors, then the complex-valued convolution operation can be expressed as:

[0087] Q*v=(A*xB*y)+i(B*x+A*y)

[0088] After the complex-valued convolution, complex-valued batch normalization (Batch Normalization, BN) is used to maintain numerical stability. The complex-valued batch normalization operation can be expressed as:

[0089]

[0090] Among them, E(·) represents the expected value of calculation, V represents the covariance matrix of the real and imaginary parts of the input x, and after complex batch normalization, the LeakyReLU activation function is used to finally obtain the output of the complex-valued convolution layer. In the above way, the complex-valued characteristics of the channel information can be fully utilized. In LCVCsiNet, the real and imaginary parts of the complex numbers are stored separately using real matrices. Figure 3 As shown in the structure diagram of the complex-valued feature map, if the number of channels of the complex-valued feature map is Cchan , then the first C chan / 2 channels are used to store the real part of complex numbers, and the latter C chan / 2 channels are used to store the imaginary part information of complex numbers.

[0091] The proposed lightweight complex-valued neural network LCVCsiNet is divided into two parts: the CSI compression encoder on the UE side and the CSI reconstruction decoder on the BS side. Figure 4 The network structure diagram of the encoder model on the UE side is shown. In the encoder part, a lightweight model design is adopted to reduce the number of channels in the intermediate feature map. At the same time, the large convolution kernel is decomposed into two small convolution kernels in orthogonal directions to reduce the computational overhead. The input of the encoder model is a 64×64×2 channel state matrix, which represents the number of BS transmit antennas, the number of truncated subcarriers, and the real and imaginary parts of the complex number. The two branches at the entrance of the model are 5×5 complex-valued convolution (replaced by continuous 1×5 and 5×1 convolutions to reduce the amount of calculation) and 1×1×2 three-dimensional convolution. The 5×5 complex-valued convolution is used to extract regional related features in the input matrix, and the 1×1×2 three-dimensional convolution is used to explore the related features of the real and imaginary parts of the complex-valued matrix elements. The maximum pooling layer (MaxPool) in the model reduces the length and width of the input feature map to 1 / 2. The two branches in the middle of the encoder, 1×5 complex-valued convolution and 5×1 complex-valued convolution, are used to extract the features in the delay direction and the features in the angle direction in the feature map, respectively. Overall, the encoder compression process can be expressed as follows:

[0092]

[0093] in, represents the parameters of the encoder model, represents a codeword of length L after being compressed by the UE-side encoder, so the compression rate can be defined as:

[0094]

[0095] The compressed codeword is fed back to the BS side via the uplink, and the decoder model on the BS side recovers the CSI information. Figure 5The network structure diagram of the decoder model on the BS side is shown. The fully connected layer at the entrance of the decoder converts the compressed codeword into a vector of length 2048 (32×32×2), and the 5×5 complex-valued convolution is used to further explore the regional features in the matrix. The decoder model is located on the BS side, and the computing power and energy resources are more sufficient than the UE side. Therefore, the channel number control parameter N=8 in the figure is set to maximize the exploration and utilization of hidden features. The branched 1×5 complex-valued convolution and 5×1 complex-valued convolution and the upsampling layer (UpSample) in the decoder are used to gradually reconstruct the information in the delay direction and the angle direction. Finally, the channel state information CSI (64×64×2) is reconstructed through the 5×5 complex-valued convolution. The overall decoder reconstruction process can be expressed as follows:

[0096]

[0097] in, Represents the parameters of the decoder model. For a specific region c, the encoding and decoding process of the model can be expressed as follows:

[0098]

[0099] in, is the parameter set of region c. Finally, the channel transmission model of each region is obtained by further training through the following optimized mean square error loss function:

[0100]

[0101] The complex valued network model with the same structure is used in all sub-areas, but each sub-area uses its corresponding channel feature data set for model training. This design allows each model to learn the unique channel distribution characteristics in its corresponding area, and when the user equipment UE switches areas, the consistent model structure of each area allows the UE to quickly obtain the new area model parameters and load and update, thereby ensuring model accuracy and CSI transmission efficiency.

Claims

1. A Massive MIMO channel state transmission method based on hybrid partitioning and complex-valued network, characterized in that: The method comprises: Step 1: Sampling to obtain a data set for model training and evaluation, including channel state information CSI and corresponding location information of user equipment UE; Step 2: Perform preliminary area division within the site coverage area according to the location information. First, calculate the distance between each user equipment UE and the base station BS. Then set a series of threshold values. According to the distance between the user equipment and the base station, divide the entire site coverage area into a series of coarse zones. Each user equipment is assigned to a corresponding coarse zone according to its distance to the base station. Step 3: In each coarse partition based on location information, a finer-grained partition is further performed according to the channel state characteristics in the region. First, a similarity matrix of the channel state characteristics is constructed, and then the Laplacian matrix of the similarity matrix is ​​calculated and eigendecomposition and dimensionality reduction are performed to obtain a feature embedding representation of the channel state in a low-dimensional space. Then, K-Means clustering of the channel characteristics is performed in the low-dimensional space to obtain a mixed partitioning result of the channel environment. Step 4: Preprocess the channel state feature data set in each sub-area, the preprocessing operation includes performing CSI two-dimensional discrete Fourier transform and data set partitioning; Step 5: Construct a lightweight complex-valued neural network model LCVCsiNet, which includes an encoder compression model on the UE side and a decoder reconstruction model on the BS side, and use the preprocessed data set to iteratively train the model parameters of each region.

2. The method according to claim 1, characterized in that In step 2, a set of user equipment UEs within a given site range is X, where the position of each UE is represented by Where m is the dimension of the location information, N is the number of elements in the set X, and the location of the base station BS in the site is expressed as Therefore, the distance from the i-th user equipment UE to the base station BS can be expressed as: d i =||x i -x BS ||,i∈{1,2,...,N} where ||·|| represents the Euclidean distance; further, M threshold values ​​{D1, D2, ..., D M }, satisfying 0=D0 <D1<D2<...<D M , according to the distance between each user equipment UE and the BS, the entire site coverage is divided into M coarse partitions: C j ={x i |D j-1 ≤d i ≤D j ,i∈{1,2,...,N}},j=1,2,...,M Finally, we get M coarse partitions {C1,C2,...,C M }, each user equipment i is allocated to a corresponding coarse partition according to its distance to the base station.

3. The method according to claim 1, characterized in that: In step 3, a similarity matrix of channel state features is first constructed: for a set of user equipment UEs X within the site range, the channel state information of the i-th UE is represents the complex field, N t Indicates the number of BS transmitting antennas, N c represents the number of subcarriers; then for each coarse partition C j , the channel state information set in its area can be expressed as: Furthermore, for each coarse partition C j , construct the similarity matrix Where |C j | represents the number of user devices in the corresponding coarse partition; for S j , the element S in row k and column l j,kl represents the similarity of the channel state information between user equipment k and user equipment l, and is calculated using the Gaussian kernel function: Among them, σ is a parameter used to control the similarity attenuation, which can be obtained from all channel states H i and H j Based on the above similarity measure, the coarse partition C can be obtained. j The similarity matrix S j As follows:

4. The method according to claim 1, characterized in that: In step 3, the Laplacian matrix of the similarity matrix is ​​calculated and the eigendecomposition is performed to reduce the dimension, and the feature embedding representation of the channel state in the low-dimensional space is further obtained: First, the degree matrix D of the similarity matrix is ​​calculated j , which is a diagonal matrix, each diagonal element can be expressed as D j,kk ,k={1,2,...,|C j |}, indicating C j The sum of similarities between the kth UE and other UEs in: Among them, |C j | indicates coarse partition C j The number of UEs in; then, according to the degree matrix D j , construct the Laplace matrix L j , expressed as: L j =D j -S j Further, through L j Construct the normalized Laplace matrix L j,sym : The normalized Laplacian matrix L j,sym Perform eigenvalue decomposition to obtain its eigenvalues ​​and eigenvectors: L j,sym and k =λ k and k where λ k is the eigenvalue, u k is its corresponding eigenvector; select the first K j The eigenvectors corresponding to the smallest eigenvalues ​​form the matrix The i-th row represents the feature embedding representation of user device i in the low-dimensional space. Furthermore, the matrix U j Each row u i After normalization, we get the matrix Each of these actions is as follows:

5. The method according to claim 1, characterized in that In step 3, K-Means clustering of channel features is performed in a low-dimensional space to obtain a mixed partition result of the channel environment: Specifically, the normalized feature matrix Each line of As a new data point, K-Means method is applied in low-dimensional space. j Clustering of clusters; after clustering is completed, we can get the coarse partition C j A finer-grained channel state feature partition C under j,k ,k={1,2,..,K j }; Finally, the hybrid partitioning method HZPM can be represented as a two-layer partition mapping as a whole. The first layer is a coarse partition mapping based on location information, and the second layer is a detailed partition mapping based on channel state characteristics.

6. The method according to claim 1, characterized in that In step 4, the channel state feature data set in each sub-region is preprocessed, and the preprocessing operation includes performing a two-dimensional discrete Fourier transform and data set partitioning: Since the channel matrix H i The representation in the angle-delay domain is usually sparse. To reduce the subsequent calculation and storage overhead, the channel matrix H i Convert to the angle-delay domain through two-dimensional discrete Fourier transform and calculate using the following formula: in, and are the discrete Fourier transform matrices in the angle and delay directions, respectively, Indicates F d The physical meaning of the conjugate transpose is to transform the left matrix in the column direction and transform the frequency domain into the delay domain. After transforming to the angle-delay domain, due to the finiteness of the path propagation delay, H i ′ Perform lossless truncation in the delay direction, retaining only the first N containing valid information c ' column, in the embodiment, N is set c ′=64 to reduce storage and processing overhead; Finally, the channel matrix of the i-th user UE is The user channel matrix data set within the base station BS is And through the hybrid partitioning HZPM method, each user UE has and only has one belonging sub-area, thereby converting the channel matrix data set Divide into non-overlapping sub-datasets: in, c≠c′ and c, 7. The method according to claim 1, characterized in that In the step 5, a lightweight complex-valued neural network model LCVCsiNet is constructed, including an encoder compression model on the UE side and a decoder reconstruction model on the BS side, and the preprocessed data set is used to iteratively train the model parameters of each region: the convolution operation of the complex value in LCVCsiNet is performed in the following manner: suppose the convolution kernel of the complex-valued convolution layer is Q=A+iB, where A and B are both real matrices, i is an imaginary unit, the complex-valued vector is v=x+iy, x and y are real vectors, then the complex-valued convolution operation can be expressed as: Q*v=(A*xB*y)+i(B*x+A*y) After the complex-valued convolution, complex-valued batch normalization is used to maintain numerical stability, and then the output of the complex-valued convolution layer is finally obtained through the LeakyReLU activation function; The proposed lightweight complex-valued neural network LCVCsiNet is divided into two parts: the CSI compression encoder on the UE side and the CSI reconstruction decoder on the BS side. In the encoder part, a lightweight model design is adopted to reduce the number of channels in the intermediate feature map, and the large convolution kernel is decomposed into two small convolution kernels in orthogonal directions to reduce the computational overhead. The input of the encoder model is a 64×64×2 channel state matrix, and the two branches at the entrance of the model are 5×5 complex-valued convolution and 1×1×2 three-dimensional convolution respectively. Among them, the 5×5 complex-valued convolution is used to extract the regional related features in the input matrix, and it is replaced by continuous 1×5 and 5×1 convolutions to reduce the amount of calculation. The 1×1×2 three-dimensional convolution is used to explore the related features of the real and imaginary parts of the complex-valued matrix elements. The two branches in the middle of the encoder, 1×5 complex-valued convolution and 5×1 complex-valued convolution, are used to extract the features in the delay direction and the features in the angle direction in the feature map respectively. Overall, the encoder compression process can be expressed as follows: in, represents the parameters of the encoder model, represents a codeword of length L after compression by the UE encoder; the compressed codeword is fed back to the BS through the uplink, and then the CSI information is restored by the decoder model on the BS side; the fully connected layer at the decoder entrance converts the compressed codeword into a vector of length 2048, and the 5×5 complex-valued convolution is used to further explore the regional features in the matrix; the branched 1×5 complex-valued convolution and 5×1 complex-valued convolution and upsampling layer in the decoder are used to gradually reconstruct the information in the delay direction and the angle direction; overall, the decoder reconstruction process can be expressed as follows: in, Represents the parameters of the decoder model; for a specific region c, the encoding and decoding process of the model can be expressed as follows: in, is the parameter set of region c; finally, the channel transmission model of each region is obtained by further training through the following optimized mean square error loss function: The complex-valued network model with the same structure is used in all sub-areas, but each sub-area uses its corresponding channel feature data set to train the model; such a design enables each model to learn the unique channel distribution characteristics in its corresponding area, and when the user equipment UE switches areas, the consistent model structure of each area allows the UE to quickly obtain the new area model parameters and load and update them, thereby ensuring model accuracy and CSI transmission efficiency.

Images