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

Abstract

The application provides a Massive MIMO channel state transmission method based on mixed partition and complex value network, and the method mainly comprises two parts: the first part provides a channel environment mixed partition method. Through the double mixed partition technology based on position information and channel characteristics, the method can effectively reduce the model complexity, significantly improve the system efficiency, scalability and overall performance, and reduce the calculation overhead of the user equipment. The second part provides a lightweight complex neural network model, which can fully exploit and utilize the complex value characteristics of the channel state, and the lightweight design ensures the high training efficiency and inference efficiency of the model. Combined with the above two methods, the application realizes higher precision and efficiency of channel state feedback transmission compared with the traditional technology.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of deep learning and wireless communication physical layer, in particular to the research field of channel partition modeling and deep learning-based channel state information transmission, and specifically relates to a Massive MIMO channel state transmission method based on mixed partition and complex-valued network, which is used to improve the accuracy of Massive MIMO system in acquiring channel state information (CSI) and improve the efficiency of CSI compression and reconstruction. BACKGROUND

[0002] Massive Multiple-Input Multiple-Output (Massive MIMO) is one of the key technologies in today's 5G communication systems and future 6G communication technologies. Massive MIMO system deploys a large number of antennas (number of antennas ≥64) at the base station (Base Station, BS) to greatly improve the overall system capacity, spectral efficiency, and coverage, and further improve the data transmission rate and system stability. For Massive MIMO system, it needs to rely on accurate channel state information (Channel State Information, CSI) to play its role. In the Massive MIMO system using Frequency Division Duplexing (Frequency Division Duplexing, FDD), since 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 (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 in the Massive MIMO system and the huge amount of channel CSI data, it results in huge transmission overhead and bandwidth consumption when the user equipment UE feeds back the downlink channel CSI to the base station BS. Therefore, it is of great practical significance to study an accurate and efficient channel state information (CSI) transmission method.

[0003] Currently, the optimization methods of CSI feedback transmission mainly include codebook-based method and compressive 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), then the UE selects the optimal precoding matrix from the predefined codebook and records the index of the matrix, which is called precoding matrix indicator (PMI). Next, the UE transmits the PMI and other channel indicators to the BS. The BS further reconstructs the optimal precoding matrix through the feedback information of the UE and the codebook and estimates the CSI of the downlink channel. This codebook-based CSI transmission method does not directly transmit CSI information, but estimates CSI through the predefined 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, making it still have large time and space overhead when applied in the Massive MIMO system. In addition to the codebook-based method, another optimization method of CSI transmission is the compressive sensing technology. The compressive sensing technology compresses information by utilizing the sparsity of the CSI matrix, thereby reducing the transmission overhead. However, the reliability of the sparsity assumption of the CSI matrix cannot be guaranteed, for example, in a complex communication environment, the sparsity of the CSI matrix is not obvious, thereby causing a large reconstruction error. In addition, the iterative algorithm with high computational complexity is usually involved in the compressive sensing technology, especially in the high-dimensional CSI scenario of the Massive MIMO system, the computational overhead cost will be very high.

[0004] With the rapid development of deep learning (DL) technology and the upgrade iteration of hardware, DL technology has been gradually applied to various tasks in the field of wireless communication, such as beamforming, interference management, channel estimation and prediction, etc. In the task of CSI feedback transmission, DL technology is also applied, and it shows higher CSI recovery accuracy compared to traditional methods and reduces the transmission overhead to a certain extent. However, the 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 for CSI compression at the UE end and CSI reconstruction at the BS end, but within a site range, the channel environment may have different characteristic distributions and variation characteristics in different sub-regions, which causes the DL model to perform poorly in sub-regions with significant differences. In addition, the traditional DL model cannot fully exploit the characteristics of the CSI complex value elements. From the above perspective, the accuracy and efficiency of CSI feedback transmission in the Massive MIMO system can be further optimized.

[0005] In view of the above, it is necessary to design a Massive MIMO channel state transmission method based on hybrid partitioning and complex-valued network. SUMMARY

[0006] The present application aims to solve the problem that the deep learning model does not sufficiently explore the characteristics of the wireless channel environment in the current Massive MIMO system, and the traditional neural network cannot fully utilize the characteristics of the CSI complex 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 network is proposed. Based on the proposed hybrid partitioning method HZPM (Hybrid Zone Partitioning Method) of finer granularity level, the wireless channel environment characteristics are more fully explored, and the parameter quantity and complexity of the model are further reduced. Based on the proposed lightweight complex-valued network LCVCsiNet (Lightweight Complex-Valued CSI Network), the channel CSI complex value characteristics are more fully utilized, and the accuracy of CSI feedback transmission is further improved.

[0007] The purpose of the present application is achieved as follows:

[0008] On the one hand, the present application provides a channel hybrid partitioning method HZPM of finer granularity level than the site granularity level, which divides a large area in the range of a site into a plurality of relatively smaller sub-areas. In essence, this is to divide a single large and complex model into a plurality of small and targeted models, which can effectively reduce the model complexity, significantly improve the system efficiency, system scalability and overall performance, and reduce the computational burden of the user equipment UE side. On the other hand, the present application 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 high training efficiency and inference efficiency of the model. By combining the above two methods, the present application realizes higher accuracy and efficiency of CSI feedback transmission.

[0009] Specifically, the technical scheme adopted by the present application includes the following steps:

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

[0011] Step 2, within the coverage range of a site, first, preliminary area division is performed according to the position information. Specifically, let the user equipment UE set in the given site range be X, where the position of each UE is represented as where m is the dimension of the location information, N is the number of elements of the set X, and the location of the base station BS in the station is denoted as Therefore, the distance of the ith user equipment UE to the base station BS can be denoted as:

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

[0013] where ||·|| denotes the Euclidean distance. Further, M thresholds are set, and according to the distance of each user equipment UE to the BS, the whole station coverage is divided into M coarse partitions {C1, C2,..., CM}. M Each user equipment is assigned to the corresponding coarse partition according to its distance to the base station.

[0014] Step 3, in each location information-based coarse partition in the step 2, further according to the channel state characteristics in the area, the sub-partitions are divided in a finer granularity. 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 the channel state, and further the sub-areas can be divided by similarity. For the set X of user equipments UE in the station range, the channel state information of the ith UE is denotes the complex field, N t denotes the number of BS transmit antennas, N c denotes the number of subcarriers. Then for each coarse partition C j , the set of channel state information in the area can be denoted as:

[0016]

[0017] Further, for each coarse partition C j , a similarity matrix S is constructed, where |C j | denotes the number of user equipments in the corresponding coarse partition. For S j , the element S j,kl in the kth row and the lth column denotes the similarity of the channel state information between the user equipment k and the user equipment l, which is calculated using a Gaussian kernel function.

[0018]

[0019] where σ is a parameter used to control the similarity decay, which can be determined by the median of the Euclidean distance of all channel states H i and H j .

[0020] Step 3.2: Calculate the similarity matrix S j The Laplacian matrix is ​​calculated and its dimensionality is reduced through eigenvalue decomposition. The Laplacian matrix encodes the global structure in the similarity graph. Further eigenvalue decomposition of the Laplacian matrix maps the original high-dimensional channel state features to a low-dimensional space. First, the degree matrix D of the similarity matrix is ​​calculated. j It is a diagonal matrix, where each diagonal element represents a coarse partition C. j The sum of the similarities between each UE and other UEs. Then, based on the degree matrix D... j Construct the Laplace matrix L j , can be represented as:

[0021] L j =D j -S j

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

[0023]

[0024] For the normalized Laplace 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 It is an eigenvalue, u k It is its corresponding eigenvector. Select the top K. j The eigenvectors corresponding to the smallest eigenvalues ​​form a matrix. The i-th row represents the feature embedding representation of the channel state of user equipment i in the low-dimensional space. Furthermore, for matrix U... j Each line u i Normalization is performed to obtain the normalized matrix.

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

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

[0029]

[0030] wherein, c≠c′ and c,c′∈{1,2,...,C}.

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

[0032]

[0033] wherein H is the channel state matrix of the model input, denotes the parameters of the encoder model, denotes the codeword of length L after compression by the UE-side encoder. The compressed codeword is fed back to the BS side through the uplink, and the CSI information is recovered by the decoder model of the BS side. The overall decoder reconstruction process can be represented as follows:

[0034]

[0035] wherein, denotes the parameters of the decoder model. For a specific region c, the codec process of the model can be represented as follows:

[0036]

[0037] wherein, is the parameter set of region c. Finally, the channel transmission model of each region is trained by further optimizing the mean square error loss function. The above steps, through the mixed partition method HZPM and the complex-valued neural network modeling method, fully exploit the characteristics of the wireless channel environment and utilize the characteristics of the CSI complex elements, further improving the accuracy of CSI feedback transmission and reducing the parameter quantity and complexity of the model.

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

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

[0040] (2) The present application proposes a light 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, and the light design ensures high training efficiency and inference efficiency of the model. Further, by combining with the mixed partition method HZPM, the present application realizes higher CSI feedback transmission accuracy and efficiency compared with the traditional technology. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is the overall flowchart of the implementation method of the present application.

[0042] Figure 2 is the partition display diagram of the mixed partition method HZPM.

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

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

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

[0046] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

[0047] As shown in the Figure 1 , the present application is generally divided into three processes. First, sampling is used to obtain a data set for model training and evaluation, including channel state information CSI and corresponding user equipment UE location information. The second step is to divide the sub-regions by the hybrid zoning method HZPM proposed by the present application, including location zoning and channel feature zoning sub-processes. The third step is to establish the CSI transmission model in each region of HZPM by the lightweight complex-valued network model LCVCsiNet proposed by the present application, including data set preprocessing, codec model construction, and model training sub-processes. The steps of the technical solution adopted by the present application are described in detail below.

[0048] Step 1, sampling is used to 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 coverage range of the station, first, according to the location information, preliminary regional division is carried out. Figure 2 is the partitioning display diagram of the hybrid zoning method HZPM, wherein each solid circular ring represents a location region boundary line. Specifically, let the user equipment UE set in a given station range be X, wherein the location of each UE is represented as where m is the dimension of the location information, here m = 2, N is the number of elements of set X, and the location of the base station BS in the station is represented as Therefore, the distance of the i-th user equipment UE to the base station BS can be represented as:

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

[0051] where ||·|| represents the Euclidean distance. Further, let M threshold values {D1, D2,..., DM} be set, where D1 < D2 <... < DM, and the location zoning is carried out according to the following rules: M}, satisfying 0 = D0 <D1<D2<...<D M Based on the distance from each User Equipment (UE) to the Base Station (BS), the entire site coverage area 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] Ultimately, we obtain M coarse partitions {C1, C2, ..., C...}. M Each user equipment i is assigned to a corresponding coarse partition based on its distance to the base station:

[0054] f1(x i If D = j, then D j-1 ≤d i ≤D j

[0055] Step 3: Within each coarse partition based on location information in Step 2, further fine-grained partitioning is performed based on the channel state characteristics within the region. Figure 2 In the diagram, each dashed line represents the boundary of a channel feature subdivision within a coarse location 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 degree of similarity between channel states. Furthermore, sub-regions can be divided using this similarity. The similarity matrix S... j It is a symmetric matrix, where the matrix elements represent the similarity between each data point. For a set of user equipment (UE) X within a site range, the channel state information of the i-th UE is: Represents the complex field, N t N represents the number of BS transmitting antennas. c This represents the number of subcarriers. Then, for each coarse partition C... j The channel state information set within its region can be represented as:

[0057]

[0058] Furthermore, for each coarse partition C j Construct a similarity matrix Where |C j | indicates the number of user devices in the corresponding coarse partition. For S j The element S in the k-th row and l-th column j,klThe similarity of the channel state information between user equipment k and user equipment 1 is denoted as S

[0059]

[0060] where σ is a parameter used to control the decay of similarity, which can be determined by the median of the Euclidean distance of all channel states H i and H j . Based on the above similarity measure, the similarity matrix S j of the coarse partition C j can be obtained as follows:

[0061]

[0062] Step 3.2, compute the Laplacian matrix of the similarity matrix S j and perform eigen-decomposition for dimension reduction. 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 the low-dimensional embedding representation of the channel features. In the mapped low-dimensional space, the structure of the channel feature data is easier to be recognized by traditional clustering algorithms, and then the channel feature sub-region can be divided by clustering. First, the degree matrix D j of the similarity matrix is calculated, which is a diagonal matrix, and each diagonal element can be represented as D j,kk , k = {1, 2,..., |C j |}, which represents the sum of similarities of the k-th UE in C j and other UEs:

[0063]

[0064] where |C j | represents the number of UEs in the coarse partition C j . Then, according to the degree matrix D j , the Laplacian matrix L j is constructed, which can be represented as:

[0065] L j = D j -S j

[0066] Further, the normalized Laplacian matrix L j is constructed by 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 It is an eigenvalue, u k These are its corresponding eigenvectors. Select the top K... j The eigenvectors corresponding to the smallest eigenvalues ​​form a matrix. The i-th row represents the feature embedding representation of the channel state of user equipment i in the low-dimensional space. Furthermore, for matrix U... j Each line u i Normalization is performed to obtain the matrix. Each of the following actions:

[0071]

[0072] Step 3.3: Perform K-Means clustering on low-dimensional features and obtain mixed partitioning results. This step identifies clusters with similar channel features in the low-dimensional space. By clustering channel features in the feature vector space, higher-quality cluster partitioning can be obtained. Specifically, the normalized feature matrix... each line As a new data point, the K-Means method is applied in a low-dimensional space for K... j Clustering of individual clusters. After clustering is complete, the results in coarse partition C can be obtained. j finer-grained channel feature partitioning C j,k k = {1, 2, ..., K} j For user equipment i, its final segment can be determined in the following way.

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

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

[0075]

[0076] Where, f1(x) i ) = j indicates that user equipment i is assigned to coarse partition C based on its distance from the BS. jThe second layer is a subdivision mapping based on channel state characteristics:

[0077]

[0078] Among them, f2(H i ;j)=k indicates that user equipment i is assigned to subdivision C within coarse partition j based on channel state characteristics. j,k .

[0079] Step 4: Preprocess the channel state feature datasets within each sub-region. Preprocessing operations include performing a two-dimensional discrete Fourier transform and partitioning the dataset. In a Massive MIMO system operating in Frequency Division Duplex (FDD) mode, the base station (BS) is configured with N antennas deployed in a Uniform Linear Antenna Array (ULA) configuration. t = 64 transmit antennas, and the number of subcarriers modulated using Orthogonal Frequency Division Multiplexing (OFDM) technology is N. c =256, User Equipment (UE) is configured with N r Root receiving antenna (for the sake of simplifying the model, let N be) r =1, meaning the user equipment is configured with a single antenna. Therefore, 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 This indicates the signal received by the UE. This represents the channel state information matrix for the downlink from the BS to the UE. The channel is represented by Gaussian white noise, where each element is an independent and identically distributed random variable following a complex Gaussian distribution. Since the channel matrix H... i The representation in the angle-delay domain is usually sparse. To reduce subsequent computation and storage overhead, the channel matrix H is... i Transforming the data to the angle-time delay domain using a two-dimensional discrete Fourier transform (2D DFT) can be represented as follows:

[0082]

[0083] in, and These are the DFT matrices for the angle and the time delay direction, respectively. F represents d The conjugate transpose of is physically equivalent to transforming the column directions of the left-hand matrix and converting the frequency domain to the time-delay domain. After transforming to the angle-delay domain, due to the finite path propagation delay, H can be... i Lossless truncation is performed in the time delay direction, retaining only the first N bits containing valid information. c = 64 columns to reduce subsequent storage and processing overhead. Finally, the channel matrix for the i-th user UE is... The user channel matrix dataset within the range of the base station (BS) is Furthermore, through the hybrid partitioning HZPM method, each user UE has one and only one home sub-region, therefore the channel matrix dataset... It can be divided into Non-overlapping subsets of data:

[0084]

[0085] in, c≠c′ and c,c′∈{1,2,...,C}.

[0086] Step 5: Construct the 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. Iterative training of the model parameters for each region is performed using the preprocessed dataset. In LCVCsiNet, complex-valued convolution operations can be performed as follows: Let the convolution kernel of the complex-valued convolutional layer be Q = A + iB, where A and B are real matrices, i is the imaginary unit, and the complex-valued vector be v = x + iy, where 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 performing complex-valued convolution, batch normalization (BN) is used to maintain numerical stability. The batch normalization operation can be represented as:

[0089]

[0090] Here, E(·) represents the calculation of the expected value, and V represents the covariance matrix of the real and imaginary parts of the input x. After complex batch normalization, the LeakyReLU activation function is used to finally obtain the output of the complex-valued convolutional layer. In this way, the complex-valued characteristics of channel information can be fully utilized. In LCVCsiNet, the real and imaginary parts of the complex number are stored separately using real matrices. For example... Figure 3 The structure diagram of the complex-valued feature map is shown below. If the number of channels in the complex-valued feature map is C...chan where the first C chan / 2 channels are used to store the real part information of the complex number, and the last C chan / 2 channels are used to store the imaginary part information of the complex number.

[0091] The proposed lightweight complex-valued neural network LCVC SiNet is divided into two parts: a CSI compression encoder on the UE side and a CSI reconstruction decoder on the BS side. Figure 4 The network structure diagram of the UE-side encoder model is shown. In the encoder part, a lightweight model design is adopted to reduce the number of intermediate feature map channels, and large convolution kernels are decomposed into two orthogonal direction small convolution kernels to reduce the computational overhead. The input of the encoder model is a channel state matrix of 64x64x2, which represents the number of BS transmit antennas, the number of truncated subcarriers, and the real and imaginary parts of the complex number, respectively. The two branches at the entrance of the model are 5x5 complex-valued convolution (replaced by consecutive 1x5 and 5x1 convolution to reduce the amount of calculation) and 1x1x2 three-dimensional convolution, respectively. The 5x5 complex-valued convolution is used to extract the regional correlation features in the input matrix, and the 1x1x2 three-dimensional convolution is used to explore the correlation features of the real and imaginary parts of the complex matrix elements. The MaxPool layer in the model reduces the length and width of the input feature map to 1 / 2. The 1x5 complex-valued convolution and 5x1 complex-valued convolution in the middle of the encoder are used to extract the time delay direction features and angle direction features in the feature map, respectively. Overall, the encoder compression process can be represented as follows:

[0092]

[0093] where, denotes the parameters of the encoder model, denotes the codeword of length L after compression by the UE-side encoder, and thus the compression rate can be defined as:

[0094]

[0095] The compressed codeword is fed back to the BS side through the uplink, and the CSI information is recovered by the decoder model on the BS side. Figure 5The network structure diagram of the BS-side decoder model is shown. The fully connected layer at the entrance of the decoder converts the compressed codeword to a vector of 2048 length (32x32x2), and a 5x5 complex-valued convolution is used to further explore the regional features within the matrix. The decoder model is located on the BS side, which has more computing power and energy resources than the UE side. Therefore, the channel number control parameter N=8 is set in the figure to explore the hidden features to the greatest extent. The 1x5 complex-valued convolution and 5x1 complex-valued convolution in the decoder and the up-sampling layer (UpSample) are used to gradually reconstruct the time delay direction and angle direction information. Finally, a 5x5 complex-valued convolution is used to reconstruct the channel state information (CSI) of (64x64x2). The overall decoder reconstruction process can be represented as follows:

[0096]

[0097] wherein, denotes the parameters of the decoder model. For a specific region c, the coding and decoding process of the model can be represented as follows:

[0098]

[0099] wherein, is the parameter set of region c. Finally, the channel transmission model of each region is trained by further optimizing the mean square error loss function as follows:

[0100]

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

Claims

1. A method for Massive MIMO channel state transmission 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 position information of user equipment (UE); Step 2, within the coverage of a station, preliminary regional division is performed according to the position information, the distance of each user equipment (UE) to a base station (BS) is first calculated, then a series of threshold values are set, the entire station coverage is divided into a series of coarse partitions according to the distance of the user equipment to the base station, and each user equipment is allocated to the corresponding coarse partition according to its distance to the base station; Step 3, in each coarse partition based on position information, further division is performed according to the channel state characteristics in the region, a similarity matrix of the channel state characteristics is first constructed, then a Laplacian matrix of the similarity matrix is calculated and dimensionality reduction is performed through feature decomposition, so as 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 and a mixed partition result of the channel environment is obtained; Step 4, pre-processing is performed on the channel state characteristic data set in each sub-region, the pre-processing operation includes performing two-dimensional discrete Fourier transform of CSI and partitioning the data set; Step 5, a light-weight 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 pre-processed data set is used to iteratively train the model parameters of each region.

2. The method of claim 1, wherein, In the step 2, a set of user equipments, UEs, in a given station range is X, where the position of each UE is denoted as where m is the dimension of the position information, N is the number of elements of the set X, and the position of the base station, BS, in the station is denoted as Thus, the distance of the i-th user equipment, UE, to the base station, BS, can be denoted as: d i =||x i -x BS ||,i∈{1,2,...,N} where ||·|| denotes the Euclidean distance; further, set M thresholds {D1, D2,..., D M}, satisfying 0 = D0< D1< D2<... < DM M According to the distance from each user equipment UE to the BS, the entire site coverage is divided into M rough partitions: C j = {x i |D j-1 ≤ d i ≤ D j , i e {1,2,...,N}, j = 1,2,...,M Finally, M coarse partitions {C1, C2,..., CM} are obtained, each user equipment i being assigned to a respective coarse partition according to its distance to the base station. M}, each user equipment i being assigned to a respective coarse partition according to its distance to the base station.

3. The method of claim 1, wherein, In the step 3, first construct the similarity matrix of channel state features: for a set of user equipments (UEs) X in the range of the station, where the channel state information of the i-th UE is denotes the complex field, N t denotes the number of BS transmit antennas, N c denotes the number of subcarriers; then for each coarse partition C j The set of channel state information in its region can be represented as: H j = {H i | x i ∈ C j , i ∈ {1, 2,..., N}, j = 1, 2,..., M Furthermore, for each coarse partition C j Construct a similarity matrix Where |C j | indicates the number of user devices in the corresponding coarse partition; for S j The element S in the k-th row and l-th column j,kl The similarity of channel state information between user equipment k and user equipment l is calculated using a Gaussian kernel function: where σ is a parameter for controlling the similarity decay, which can be determined by the median of the Euclidean distance of all channel states H i and H j Based on the above similarity, the similarity matrix S j of the coarse partition C j is obtained as follows:

4. The method of claim 1, wherein, In the step 3, the Laplacian matrix of the similarity matrix is calculated and the dimensionality reduction is performed by eigen-decomposition, further obtaining the feature embedding representation of the channel state in the low-dimensional space: first, the degree matrix D of the similarity matrix is calculated j , which is a diagonal matrix, and each diagonal element can be represented as D j,kk , k = {1, 2,..., |C j |} represents the similarity sum of the kth UE in C j and other UEs: where |C j denotes the number of UEs in the coarse partition C j ; then, according to the degree matrix D j , the Laplacian matrix L j is constructed, denoted as: L j = D j - S j Further, by L j Constructing the normalized Laplacian matrix L j,sym : On the normalized Laplacian matrix L j,sym Eigenvalue decomposition is performed to obtain its eigenvalues and eigenvectors: L j,sym u k =λ k u k where λ k are eigenvalues, u k are their corresponding eigenvectors; select the eigenvectors corresponding to the first K j smallest eigenvalues to form a matrix U where the i-th row represents the feature embedding representation of the user equipment i in the low-dimensional space; further, normalize each row u j of the matrix U i to obtain a matrix where each row is as follows:

5. The method of claim 1, wherein, In step 3, K-Means clustering of channel features is performed in the low-dimensional space to obtain the hybrid partitioning results of the channel environment: specifically, the normalized feature matrix is... each line As a new data point, the K-Means method is applied in a low-dimensional space for K... j Clustering of individual clusters; after clustering, the results in coarse partition C can be obtained. j finer-grained channel state feature partitioning C j,k k = {1, 2, ..., K} j Ultimately, the hybrid partitioning method HZPM can be represented as a two-layer partitioning mapping: the first layer is a coarse partitioning mapping based on location information, and the second layer is a finer partitioning mapping based on channel state characteristics.

6. The method of claim 1, wherein, In the step 4, the channel state feature dataset in each sub-region is pre-processed, and the pre-processing operation includes two-dimensional discrete Fourier transform and dataset partitioning: since the channel matrix H i The representation in angle-delay domain is usually sparse, in order to reduce the subsequent calculation and storage overhead, the channel matrix H i The conversion to angle-delay domain through two-dimensional discrete Fourier transform is calculated by the following formula: where, and are the Discrete Fourier Transform matrices in angle and delay direction respectively, denotes the conjugate transpose of F d ; physically, it means transforming the left matrix in column direction and transforming the frequency domain to delay domain; after transforming to angle-delay domain, due to the finiteness of path propagation delay, H i ′ can be losslessly truncated in delay direction, only keeping the first N c ′ columns which contain valid information; set N c ′ = 64 to reduce storage and processing overhead; finally, the channel matrix of the ith user UE is The user channel matrix dataset within the base station BS range is D = {H′1, H′2,..., H′ N}; and through the hybrid zoning HZPM method, each user UE has and only has one home sub-region, thereby dividing the channel matrix dataset D into non-overlapping sub-datasets: D = D (1) ∪ D (2) ∪... ∪ D (C) wherein c≠ c' and c, c' e {1, 2,..., C}.

7. The method of claim 1, wherein, In the step 5, a light-weight 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 pre-processed data set is used to iteratively train the model parameters of each region: for complex-valued convolution operation in LCVCsiNet, the following method is used: let the convolution kernel of the complex-valued convolution layer be Q=A+iB, where A and B are real number matrices, i is the imaginary unit, the complex-valued vector be v=x+iy, and x and y are real number vectors, then the complex-valued convolution operation can be represented as: Q*v=(A*x-B*y)+i(B*x+A*y) After complex-valued convolution, complex-valued batch normalization is used to maintain the stability of the values, and then the LeakyReLU activation function is used to finally obtain the output of the complex-valued convolution layer; The proposed lightweight complex-valued neural network LCVCsiNet is divided into two parts: CSI compression encoder on the UE side and CSI reconstruction decoder on the BS side; In the encoder part, a lightweight model design is adopted to reduce the number of intermediate feature map channels, and large convolution kernels are decomposed into two small convolution kernels in orthogonal directions to reduce the computational overhead; The input of the encoder model is a channel state matrix of 64x64x2, and the two branches at the entrance of the model are complex-valued convolution of 5x5 and three-dimensional convolution of 1x1x2 respectively; Among them, the complex-valued convolution of 5x5 is used to extract the regional correlation features in the input matrix, and continuous 1x5 and 5x1 convolution is used to replace it to reduce the amount of calculation, and the three-dimensional convolution of 1x1x2 is used to explore the correlation features of the real part and the imaginary part of the complex-valued matrix elements; The two branches of 1x5 complex-valued convolution and 5x1 complex-valued convolution in the middle of the encoder are used to extract the features in the time delay direction and the angle direction of the feature map respectively; Overall, the encoder compression process can be represented as follows: s = f en (H; W en ) where W en denotes the parameters of the encoder model, denotes the codeword of length L compressed by the UE-side encoder; the compressed codeword is fed back to the BS side through the uplink, and the CSI information is recovered by the decoder model of the BS side; the full connection layer at the entrance of the decoder converts the compressed codeword to a 2048-length vector, and the 5x5 complex value convolution is used to further explore the regional characteristics in the matrix; the branch 1x5 complex value convolution and 5x1 complex value convolution in the decoder and the up-sampling layer are used to gradually reconstruct the information in the delay direction and the angle direction; overall, the decoder reconstruction process can be represented as follows: where W de denotes the parameters of the decoder model; for a particular region c, the coding process of the model can be represented as follows: wherein, is the parameter set for region c; finally, the channel transmission model for each region is trained further by optimizing the mean square error loss function: The same structure of complex-valued network model is used in all sub-regions, but each sub-region uses its corresponding channel feature data set to train the model; This design can make each model learn the unique channel distribution characteristics in its corresponding region, and when the user equipment UE switches regions, the consistent model structure of each region can enable the UE to quickly obtain and load the new region model parameters for updating, thereby ensuring the accuracy of the model and the efficiency of CSI transmission.

Images