Large-scale MIMO CSI feedback method using lightweight space-time resolution network

Abstract

The invention discloses a large-scale MIMO (Multiple Input Multiple Output) CSI (Channel State Information) feedback method using a lightweight space-time resolution network. The method comprises the following steps: acquiring a CSI matrix of a wireless channel by multi-antenna UE (User Equipment) through channel estimation; discrete Fourier transform is carried out on the CSI matrix, so that a feature vector matrix is sparse and projected to an angle delay domain matrix; separating a real part and an imaginary part of the angle delay domain matrix to form a dual-channel matrix, and compressing the dual-channel matrix at the UE; extracting feature vectors of the two-channel matrix at an encoder by using a multi-head circulation mechanism; the extracted feature vectors are processed through a full connection layer, and compressed code word vectors are output; quantizing the compressed code word vector into a bit stream, and transmitting the bit stream to a decoder end through a feedback link; and at a decoder end, the bit stream is recovered to a codeword vector through inverse quantization processing, and the recovered codeword vector is recovered to a CSI matrix through neural network processing and inverse discrete Fourier transform. According to the invention, the problem of high calculation complexity of a CSI feedback method based on AI can be solved.

Description

Technical Field

[0001] The present invention relates to the fields of CSI feedback and deep learning technology, and in particular to a large-scale MIMO CSI feedback method using a lightweight space-time resolution network. Background Art

[0002] Massive Multiple-Input Multiple-Output (MIMO) systems are a foundational technology in fifth-generation communications (5G) and are expected to play a key role in the upcoming sixth-generation (6G) communications. Acquisition of downlink channel state information (CSI) by the base station (BS) is a key step for effective beam management in massive MIMO systems. Therefore, obtaining accurate and high-resolution CSI is crucial for MIMO system performance. Collaboration between the user equipment (UE) and the BS is essential for CSI feedback. In time-division duplex (TDD) systems, CSI acquisition relies on the reciprocity principle. Frequency-division duplex (FDD) systems enhance coverage by performing CSI estimation at the UE and then feeding it back to the transmitter. Traditional methods have limitations, such as high resource overhead, high computational complexity, and weak interference immunity. Therefore, AI-based approaches have been proposed to mitigate the resource overhead in the CSI feedback process. AI-based networks compress channel data at the UE and then reconstruct it at the BS, a solution designed for CSI. However, the computational complexity of AI-based approaches remains a significant issue that cannot be ignored.

[0003] In early research, the system model used a uniform linear array (ULA) antenna architecture to simplify the workflow. However, as channel models became increasingly complex, the latest specifications introduced the more advanced uniform planar array (UPA) antenna solution. Unlike traditional ULA antennas, which only have a single receive antenna on the user equipment (UE), UPA antenna systems deploy multiple antennas on both the base station (BS) and the UE, forming a complete MIMO architecture.

[0004] CNN-based CSI Feedback Network (CsiNet): Proposed by Wen et al. in 2018, this CSI perception and recovery mechanism is based on a deep convolutional neural network. It can effectively exploit channel structure by learning from training samples. This is the first time that DL has been applied to CSI feedback research. Previous research on CSI feedback has mostly used methods such as compressed sensing (CS).

[0005] The structural principle of CsiNet is that the core structure is an encoder and a decoder, where the values ​​S1×S2×S3 represent the length, width, and number of feature maps, respectively. The first layer of the encoder is a convolutional layer, and the input data is complex numbers. This layer uses a 3×3 dimensional kernel to generate two feature maps. After the convolutional layer, the feature map is reshaped into a vector, and a fully connected layer is used to generate a real-valued vector of size M. The first two layers simulate the projection of compressed sensing and serve as encoders. However, unlike the random projection in compressed sensing, CsiNet converts the extracted feature maps into codewords. The decoder contains two RefineNet units, whose output size is equal to the size of the channel matrix, with the purpose of refinement rather than dimensionality reduction; the identification shortcut connection is introduced to directly pass the data stream to the subsequent layers, in order to avoid the gradient disappearance problem caused by multiple stacked nonlinear transformations.

[0006] The team that proposed CsiNet in 2019 proposed a real-time CSI feedback architecture. This improved the CsiNet model and introduced a long short-term memory (LSTM) network to improve resolution. The proposed model is named CsiNet-LSTM. The LTSM in this paper refers to a decoder with memory that extracts temporal correlations from a previously recovered channel matrix and reconstructs it based on received data. In operation, the system deploys multiple CsiNet encoders with different compression ratios at the UE, while the CsiNet decoder and LSTM are deployed at the base station. Each side has a counter. Initially, CSI data is compressed at a high compression ratio at the UE, reconstructed by the high-compression CsiNet decoder, and initialized by the LSTM at the base station. In the subsequent time step t (2≤t≤T), the data is compressed at a lower compression ratio at the UE, and the resulting codeword incorporates the learned temporal correlation information. The base station's LSTM then concatenates the codeword with the first codeword (the initial high-compression ratio data) and performs an inverse transform. The counter increments with each time step. Similar operations are repeated. When the counter reaches T, the LSTM is reset and the channel group is restored.

[0007] In 2022, Cui et al. proposed a deep learning method for CSI feedback in FDD massive MIMO systems based on Google's Transformer architecture. The core structure of TransNet (single layer) is based on the original Transformer encoder and decoder layer architecture. Transformer is a complex neural network architecture. TransNet has two encoder layers in the TransNet encoder and two decoder layers in the TransNet decoder. Encoder 1 and Encoder 2 have the same structure, and Decoder 1 and Decoder 2 have the same structure. In a complete compression process, the input of the TransNet encoder is first input into Encoder 1 to obtain an output, and then the output is input into Encoder 2 to obtain another output. The output of Encoder 2 is compressed by a fully connected (FC) layer at a fixed ratio to . It is then transmitted via the feedback link and received by the TransNet decoder at the base station. It is assumed that the uplink data feedback is ideal, that is, the data received by the TransNet decoder at the base station is actually the output of the TransNet encoder at the user end.

[0008] The CNN-based CSI feedback network currently has some shortcomings. The most important one is that its performance level is significantly weaker than the current state-of-the-art level. However, its computational complexity is low and the demand for computing resources is not high.

[0009] LSTM-based networks show better performance than CNN-based networks, but the architecture of LSTM-based networks is more complex, resulting in a significant increase in computational complexity and resource requirements.

[0010] Transformer-based networks represent the most advanced DL-based CSI feedback methods, achieving the highest performance levels. However, they also exhibit the highest computational complexity, pose considerable challenges to the training process, and require the largest amount of computational and storage resources.

[0011] In early research, the system model used a ULA antenna architecture to simplify the workflow. However, as channel model complexity continues to increase, the latest specifications have introduced more advanced UPA antenna solutions. However, solutions based on ULA antennas have drawbacks such as low channel modeling accuracy and poor adaptability to complex scenarios.

[0012] In general, CNN-based CSI feedback networks offer low performance and low computational complexity. LSTM-based networks outperform CNN-based networks, but also exhibit significantly higher computational complexity. Transformer-based networks achieve the best performance available, but also exhibit the highest computational complexity. Furthermore, existing solutions based on UPA antennas offer significant performance advantages over those based on ULA antennas. Summary of the Invention

[0013] The present invention aims to solve at least one of the above-mentioned related technical problems to a certain extent.

[0014] The present invention proposes a large-scale MIMO CSI feedback method using a lightweight space-time resolution network, which utilizes the extraction of small-scale spatial correlation and temporal correlation to achieve high-precision CSI reconstruction with spatial and temporal resolution capabilities while maintaining low computational complexity.

[0015] Another object of the present invention is to propose a massive MIMO CSI feedback system using a lightweight space-time resolution network.

[0016] A third object of the present invention is to provide a computer device.

[0017] A fourth object of the present invention is to provide a non-transitory computer-readable storage medium.

[0018] To achieve the above objectives, the present invention proposes, on one hand, a method for massive MIMO CSI feedback using a lightweight space-time resolution network, comprising:

[0019] The multi-antenna UE obtains the CSI matrix of the wireless channel through channel estimation;

[0020] Performing a discrete Fourier transform on the CSI matrix to sparse the eigenvector matrix and project it onto the angular delay domain matrix;

[0021] Separate the real and imaginary parts of the angle delay domain matrix to form a dual-channel matrix and perform compression at the UE;

[0022] In the encoder, a multi-head loop mechanism is used to extract the eigenvectors of the dual-channel matrix;

[0023] The extracted feature vector is processed through the fully connected layer to output the compressed codeword vector;

[0024] Quantize the compressed codeword vector into a bit stream and transmit it to the decoder through a feedback link;

[0025] At the decoder end, the bit stream is restored to a codeword vector through inverse quantization processing, and the restored codeword vector is restored to a CSI matrix through neural network processing and inverse discrete Fourier transform.

[0026] The massive MIMO CSI feedback method using a lightweight space-time resolution network in an embodiment of the present invention may also have the following additional technical features:

[0027] In one embodiment of the present invention, a multi-head loop mechanism is used in an encoder to extract a feature vector of a dual-channel matrix, including:

[0028] The STBlock layer is used to perform spatiotemporal transformation on the input dual-channel matrix to extract temporal and spatial features. The feedforward layer is used to perform nonlinear transformation on the extracted features, and the GroupNorm layer is used to normalize the features to obtain normalized features.

[0029] A multi-head recurrence mechanism is used to process the normalized features, integrating temporal and spatial resolutions to extract and separate feature vectors.

[0030] In one embodiment of the present invention, the extracted feature vector is processed through a fully connected layer to output a compressed codeword vector, including:

[0031] The extracted feature vector is input into the fully connected layer for linear transformation, and the compressed codeword vector D is output, with a size of 1×2JN T / β, where β is the compression factor.

[0032] In one embodiment of the present invention, quantizing the compressed codeword vector into a bit stream includes:

[0033] A unified scalar quantization strategy is used to map each element to a new quantized element. The final codeword is a vector of all discrete elements: Convert the codeword vector D of data type float32 into a length of 2BJN T / β and sends the bitstream to the decoder.

[0034] In one embodiment of the present invention, at the decoder side, the bit stream is restored to a codeword vector through inverse quantization processing, and the restored codeword vector is restored to a CSI matrix through neural network processing and inverse discrete Fourier transform, further comprising:

[0035] At the decoding end, the received feedback bits are dequantized to recover the channel data, and the vector D is reconstructed through the dequantization process to obtain a size of 1×2JN T / β vector

[0036] By having {1×2JN T / β,1×2JNT The FC layer, STBlock layer, feedforward layer and GroupNorm layer of the unit process vector features Get the feature vector The processed eigenvector is converted from the angle delay domain back to the time-frequency domain and restored to the CSI matrix.

[0037] In one embodiment of the present invention, the method further includes:

[0038] The decoder performs inverse quantization on the received feedback bits, recovers D, and reconstructs the data through the reverse process of the encoder. The training optimization process is expressed as:

[0039]

[0040] where Θ en and Θ de Represent the training parameters of the encoder and decoder respectively.

[0041] To achieve the above objectives, the present invention further proposes a method for massive MIMO CSI feedback using a lightweight space-time resolution network, comprising:

[0042] The matrix acquisition module is used to obtain the CSI matrix of the wireless channel through channel estimation of the multi-antenna UE;

[0043] A feature processing module, configured to perform discrete Fourier transform on the CSI matrix to sparse the eigenvector matrix and project it onto the angular delay domain matrix;

[0044] A feature separation module is used to separate the real and imaginary parts of the angle delay domain matrix to form a dual-channel matrix and compress it at the UE;

[0045] Feature extraction module, used to extract the feature vector of the dual-channel matrix using a multi-head loop mechanism in the encoder;

[0046] The compression output module is used to process the extracted feature vector through the fully connected layer and output the compressed codeword vector;

[0047] Quantization transmission module, used to quantize the compressed codeword vector into a bit stream and transmit it to the decoder end through the feedback link;

[0048] The decoding recovery module is used to recover the bit stream into a codeword vector through inverse quantization processing at the decoder end, and to recover the recovered codeword vector into a CSI matrix through neural network processing and inverse discrete Fourier transform.

[0049] The large-scale MIMO CSI feedback method and system using a lightweight space-time resolution network in the embodiments of the present invention are optimized for data processing of UPA antenna systems. Compared with traditional solutions based on ULA antennas, they have advantages such as better adaptability to complex scenarios and higher channel modeling accuracy. The proposed LSTCNet network combines a multi-head parallel mechanism with a loop mechanism. Compared with the traditional Transformer multi-head complete separation and the traditional recurrent neural network single-head full recursion (loop), the multi-head recursion compresses data with less information loss.

[0050] To achieve the above-mentioned purpose, the third aspect embodiment of the present application proposes a computer device, comprising: a processor and a memory; wherein the processor runs a program corresponding to the executable program code by reading the executable program code stored in the memory, so as to implement the large-scale MIMO CSI feedback method using a lightweight space-time resolution network as described in the first aspect embodiment.

[0051] To achieve the above-mentioned purpose, the fourth embodiment of the present application proposes a non-temporary computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, it implements the large-scale MIMO CSI feedback method using a lightweight space-time resolution network as described in the first embodiment.

[0052] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0054] Figure 1 is a flowchart of a massive MIMO CSI feedback method using a lightweight space-time resolution network according to an embodiment of the present invention;

[0055] Figure 2 1 is a structural diagram of the encoder and decoder structure of LSTCNet according to an embodiment of the present invention;

[0056] Figure 3 are two architecture diagrams that have been applied in the field of CSI feedback according to embodiments of the present invention;

[0057] Figure 4 is a structural diagram of a massive MIMO CSI feedback system using a lightweight space-time resolution network according to an embodiment of the present invention;

[0058] Figure 5 is a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0060] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0061] The following describes a massive MIMO CSI feedback method, system device, and storage medium using a lightweight space-time resolution network according to an embodiment of the present invention with reference to the accompanying drawings.

[0062] The following are technical terms that may be used in the present invention:

[0063] 3GPP Third Generation Partnership Project

[0064] BS Base Station

[0065] CNN Convolutional Neural Network

[0066] CR Compression Ratio

[0067] CS Cosine Similarity

[0068] CSI Channel State Information Channel State Information

[0069] FDD Frequency Division Duplexing

[0070] FLOPs Floating Point Operations

[0071] KPI Key Performance Indicator

[0072] LSTM Long Short-Term Memory

[0073] MIMO Multiple-Input Multiple-Output

[0074] MSE Mean Squared Error

[0075] OFDM Orthogonal Frequency Division Multiplexing

[0076] SGCS Squared Generalized Cosine Similarity

[0077] TDD Time Division Duplexing

[0078] Figure 1 is a flow chart of a massive MIMO CSI feedback method using a lightweight space-time resolution network according to an embodiment of the present invention. Figure 1 Shown, including:

[0079] S1, obtain the CSI matrix of the wireless channel through channel estimation of the multi-antenna UE;

[0080] S2, performs discrete Fourier transform on the CSI matrix to sparse the eigenvector matrix and project it onto the angle delay domain matrix;

[0081] S3, separate the real and imaginary parts of the angle delay domain matrix to form a dual-channel matrix and compress it at the UE;

[0082] S4, in the encoder, a multi-head loop mechanism is used to extract the eigenvectors of the dual-channel matrix;

[0083] S5, processes the extracted feature vector through a fully connected layer to output a compressed codeword vector;

[0084] S6, quantizes the compressed codeword vector into a bit stream and transmits it to the decoder through the feedback link;

[0085] S7, at the decoder end, the bit stream is restored to a codeword vector through inverse quantization processing, and the restored codeword vector is restored to a CSI matrix through neural network processing and inverse discrete Fourier transform.

[0086] It is understandable that the most commonly used system model is a BS equipped with N t antennas, one UE is equipped with one antenna, forming an FDD massive MIMO model unit. In addition, there are N csubcarriers, the signal on the nth subcarrier can be described as:

[0087]

[0088] in represent the channel vector, precoding vector, transmitted symbol and additive noise on the nth subcarrier respectively.

[0089] The total channel matrix can be defined as:

[0090]

[0091] in represents the CSI in the frequency domain. The data preprocessing method is to transform Sparse and project onto the angular delay domain matrix H.

[0092] The basic process of CSI feedback involves encoding and decoding the channel. The complex matrix H is used as the processing target of the encoder-decoder studied in this invention. During system operation, the real and imaginary parts of the matrix H are separated to establish two channel data. The encoder-decoder f ed (.) First, compress the two channel data at the UE and then reconstruct them at the BS matrix.

[0093] In early research work, the system model used a uniform linear array (ULA) antenna architecture to simplify the workflow. However, as the complexity of the channel model continues to increase, the latest specifications have introduced a more advanced uniform planar array (UPA) antenna solution. Unlike traditional ULA antennas that only have a single receiving antenna on the user equipment (UE), the UPA antenna system is equipped with N antennas on the base station (BS). T transmit antennas, and N are deployed at the UE end. R The method proposed in this paper is specifically optimized for data processing in UPA antenna systems and has significant performance advantages over traditional solutions based on ULA antennas. Therefore, Equation (2) can be restated as:

[0094]

[0095] in represents the downlink channel for the jth subband, where J represents the number of subbands, and each subband consists of 4 resource blocks (RBs). Data preprocessing includes eigenvalue decomposition according to industry standards. Therefore, the channel matrix is ​​projected into the eigenspace:

[0096]

[0097] where λ j and Respectively represent the maximum eigenvalue and corresponding eigenvector of the jth subband. The codec processes the eigenvector matrix composed of J eigenvectors. During system operation, the system extracts the real and imaginary parts of the eigenvector matrix to form a real matrix

[0098] like Figure 2 Figure 2 shows the encoder and decoder structures of the LSTCNet of the present invention. The parallel structure of the Transformer can extract global spatial features from the CSI matrix, while the loop structure of the RetNet allows capturing the temporal correlation within the CSI matrix. Figure 2 The shown STBlock integrates these two architectures, first extracting global features and then leveraging overlapping temporal heads to enhance compression performance.

[0099] like Figure 3 As shown, Figure 3 (a) Attention and Figure 3 (b) The retained structures in the STBlock are the core structures of TransNet and RetNet, respectively. Each head within the STBlock initially computes matrices Q, K, and V based on the input data. Compared to the attention and retention mechanisms, the STBlock relies primarily on basic matrix operations. It also integrates Hadamard and GroupNorm to enhance its functionality. The weights of the values ​​are calculated as follows:

[0100] Q=(XW Q ⊙e iθ ),K=(XW K ⊙e -iθ ),V=XW V #(5)

[0101] in is a learnable matrix. Here, d represents the size of each head. is derived from the matrix F through the embedding operation. In addition, STBlock adopts causal mask and exponential decay to promote the recursive operation, which is defined as:

[0102]

[0103] STBlock(X)=Concat(Q1S1,…,Q h S h )#(6)

[0104] in is an exponential decay factor, and h represents the number of heads.

[0105] The computational complexity of the network increases significantly with the increase in the number of CSI matrix processing steps and the number of single-step operations. The multi-head loop mechanism can maintain high throughput, low latency, and minimal memory usage for long sequences. Based on this, the present invention proposes LSTCNet, whose overall structure is as follows Figure 2 The specific steps are as follows:

[0106] LSTCNet codec f ed (.) is divided into two parts: an encoder and a decoder. After obtaining the eigenvector matrix F, the system concatenates the real and imaginary parts of the data to create a two-channel matrix as f ed (.) input.

[0107] It can be understood that through the system model, the CSI matrix H can be obtained, and then the real part and the imaginary part of H are separated to obtain a two-channel matrix F. This two-channel matrix is ​​the input of the codec fed(.).

[0108] For the encoder, the input data passes through the STBlock layer, the feedforward layer, and the GroupNorm layer. LSTCNet only uses a single layer of STBlock (typically, multiple layers of the same structure are stacked within the network), effectively controlling the overall computational complexity of the network.

[0109] Then, the fully connected (FC) layer output size is 1×2JN T =A vector D of 1 / β, where β represents the compression factor. After encoding at the UE, the amount of data is significantly reduced, resulting in a compression ratio typically ranging from 1 / 4 to 1 / 64. However, compressed data still poses deployment challenges for massive MIMO systems.

[0110] Quantization is essential, and this step converts the codeword into a bit stream. The network uses a unified scalar quantization strategy, where each element is mapped to a new quantization element, and the final codeword is a vector of all discrete elements. For B-bit quantization, the encoder converts D into a 2BJN T / β and sends the bit stream to the decoder.

[0111] At the decoding end, the network dequantizes the received feedback bits to recover the channel data, and reconstructs the vector D through the dequantization process to obtain a size of 1×2JN T / β vector Then, By having {1×2JN T / β,1×2JN T} unit’s FC layer, STBlock layer, feedforward layer, and GroupNorm layer. Finally, the feature vector is obtained after post-processing

[0112] LSTCNet not only leverages the advanced features of RetNet, but also ensures its potential for future maintenance thanks to its basic architecture. LSTCNet has a modular structure, and although it has achieved a compact and lightweight design, each component can be individually configured and divided for training according to specific circumstances. The decoder performs inverse quantization on the received feedback bits, recovers D, and reconstructs the data through the reverse process of the encoder. Based on the description provided, the training optimization process can be expressed as:

[0113]

[0114] where Θ en and Θ de represent the trainable parameters of the encoder and decoder respectively.

[0115] The experimental configuration is as follows:

[0116] Table 1

[0117]

[0118] The present invention adopts two channel models to evaluate the performance of different networks. LSTCNet is trained and tested using the NVIDIA GeForce RTX 4090 GPU. During the statistical process, CsiNet is used as the benchmark for ratio calculation under the COST 2100 channel model, and EVCsiNet is used as the benchmark for ratio calculation under the 3GPP link-level channel model.

[0119] First, the COST 2100 channel model is used to compare and analyze the LSTCNet of the present invention with the established network. The indoor microcell scenario is 5.3GHz, and the outdoor rural scenario is 300MHz. In addition, the present invention sets h = 4, d model = 64, and the cosine similarity loss function and the default adaptive momentum (Adam) optimizer are used. The epoch, batch size and learning rate are set to 1000, 200 and 6×10 respectively. -4 The training, validation, and test datasets contain 100,000, 30,000, and 20,000 samples, respectively.

[0120] The present invention further evaluates the performance of LSTCNet by using the link-level channel model according to 3GPP TR 38.901. The basic parameters are listed in Table 1. The present invention considers CDL-A and CDL-C with delay spreads of 30ns and 300ns. T= 32 and J = 12 UPA antennas produce a matrix F size of 12 × 64. The number of quantization bits, epoch, batch size and learning rate are set to 2, 1000, 100 and 6 × 10 respectively. -4 The training, validation, and test datasets contain 80,000, 10,000, and 10,000 samples, respectively.

[0121] The multi-antenna UE of the present invention obtains CSI through channel estimation, and then uses the LSTCNet network proposed in the present invention to complete the feature extraction, compression and quantization process of the CSI matrix. The quantized codeword is then transmitted to the decoding end, and finally the CSI is recovered through processes such as inverse quantization, neural network processing and IDFT.

[0122] In summary, this paper proposes a new lightweight network LSTCNet to address the CSI feedback problem in massive MIMO FDD systems. This network has spatial and temporal resolution capabilities. By optimizing the design for the specific requirements of the CSI feedback task and utilizing simple matrix operations and a complementary multi-head loop mechanism, the computational complexity is significantly reduced. Experimental results show that while maintaining a recovery accuracy of more than 70% of the Transformer network, the computational complexity of LSTCNet is less than 14%, making it suitable for a wide range of communication scenarios.

[0123] According to an embodiment of the present invention, a large-scale MIMO CSI feedback method using a lightweight space-time resolution network utilizes the extraction of small-scale spatial and temporal correlations to achieve high-precision CSI reconstruction, with spatial and temporal resolution capabilities while maintaining low computational complexity. LSTCNet uses more residual connections, combined with a scaling mechanism, and is more stable than existing methods. Furthermore, the present invention uses a symmetrical, reconfigurable, and pluggable design, which is more flexible than existing methods and can adapt to a variety of system scenarios.

[0124] In order to implement the above embodiment, Figure 4 As shown, this embodiment also provides a massive MIMO CSI feedback system 10 using a lightweight space-time resolution network, including:

[0125] The matrix acquisition module 100 is configured to obtain the CSI matrix of the wireless channel through channel estimation for the multi-antenna UE;

[0126] The feature processing module 200 is used to perform discrete Fourier transform on the CSI matrix to sparse the eigenvector matrix and project it onto the angle delay domain matrix;

[0127] A feature separation module 300 is used to separate the real and imaginary parts of the angle delay domain matrix to form a dual-channel matrix and perform compression at the UE;

[0128] A feature extraction module 400 is used to extract feature vectors of a dual-channel matrix using a multi-head loop mechanism in the encoder;

[0129] The compression output module 500 is used to process the extracted feature vector through a fully connected layer and output a compressed codeword vector;

[0130] Quantization transmission module 600, used to quantize the compressed codeword vector into a bit stream and transmit it to the decoder end through the feedback link;

[0131] The decoding recovery module 700 is used to recover the bit stream into a codeword vector through inverse quantization processing at the decoder end, and recover the recovered codeword vector into a CSI matrix through neural network processing and inverse discrete Fourier transform.

[0132] It is understandable that the massive MIMO CSI feedback method using a lightweight space-time resolution network in the embodiment of the present invention is applied to a massive MIMO CSI feedback system using a lightweight space-time resolution network, and therefore no redundant description is given.

[0133] According to an embodiment of the present invention, a massive MIMO CSI feedback system using a lightweight space-time resolution network utilizes the extraction of small-scale spatial and temporal correlations to achieve high-precision CSI reconstruction, with spatial and temporal resolution capabilities while maintaining low computational complexity. LSTCNet uses more residual connections, combined with a scaling mechanism, making it more stable than existing methods. Furthermore, the present invention utilizes a symmetrical, reconfigurable, and pluggable design, which is more flexible than existing methods and can adapt to a variety of system scenarios.

[0134] In order to implement the method of the above embodiment, the present invention also provides a computer device, such as Figure 5 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein the processor 602 runs a program corresponding to the executable program code by reading the executable program code stored in the memory 601, so as to implement each step of the method described above.

[0135] In order to implement the above embodiments, the present application further proposes a non-transitory computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the method described in the above embodiments is implemented.

[0136] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0137] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

Claims

1. A massive MIMO CSI feedback method using a lightweight space-time resolution network, characterized in that: include: The multi-antenna UE obtains the CSI matrix of the wireless channel through channel estimation; Performing a discrete Fourier transform on the CSI matrix to sparse the eigenvector matrix and project it onto the angular delay domain matrix; Separate the real and imaginary parts of the angle delay domain matrix to form a dual-channel matrix and perform compression at the UE; In the encoder, a multi-head loop mechanism is used to extract the eigenvectors of the dual-channel matrix; The extracted feature vector is processed through the fully connected layer to output the compressed codeword vector; Quantize the compressed codeword vector into a bit stream and transmit it to the decoder through a feedback link; At the decoder end, the bit stream is restored to a codeword vector through inverse quantization processing, and the restored codeword vector is restored to a CSI matrix through neural network processing and inverse discrete Fourier transform.

2. The method according to claim 1, characterized in that The encoder uses a multi-head loop mechanism to extract the feature vector of the dual-channel matrix, including: The STBlock layer is used to perform spatiotemporal transformation on the input dual-channel matrix to extract temporal and spatial features. The feedforward layer is used to perform nonlinear transformation on the extracted features, and the GroupNorm layer is used to normalize the features to obtain normalized features. A multi-head recurrent mechanism is used to process the normalized features, integrating temporal and spatial resolutions to extract and separate feature vectors.

3. The method according to claim 1, characterized in that The extracted feature vector is processed through the fully connected layer to output the compressed codeword vector, including: The extracted feature vector is input into the fully connected layer for linear transformation, and the compressed codeword vector D is output, with a size of 1×2JN T / β, where β is the compression factor.

4. The method according to claim 3, characterized in that Quantizing the compressed codeword vector into a bit stream includes: A unified scalar quantization strategy is used to map each element to a new quantized element. The final codeword is a vector of all discrete elements: Convert the codeword vector D of data type float32 into a length of 2BJN T / β and sends the bitstream to the decoder.

5. The method according to claim 4, characterized in that At the decoder side, the bit stream is restored to a codeword vector through inverse quantization, and the recovered codeword vector is restored to a CSI matrix through neural network processing and inverse discrete Fourier transform, which also includes: At the decoding end, the received feedback bits are dequantized to recover the channel data, and the vector D is reconstructed through the dequantization process to obtain a size of 1×2JN T / β vector By having {1×2JN T / β,1×2JN T The FC layer, STBlock layer, feedforward layer and GroupNorm layer of the unit process vector features Get the feature vector The processed eigenvector is converted from the angle delay domain back to the time-frequency domain and restored to the CSI matrix.

6. The method according to claim 5, characterized in that The method further comprises: The decoder performs inverse quantization on the received feedback bits, recovers D, and reconstructs the data through the reverse process of the encoder. The training optimization process is expressed as: where Θ en and Θ de Represent the training parameters of the encoder and decoder respectively.

7. A massive MIMO CSI feedback system using a lightweight space-time resolution network, characterized in that: include: The matrix acquisition module is used to obtain the CSI matrix of the wireless channel through channel estimation of the multi-antenna UE; A feature processing module, configured to perform discrete Fourier transform on the CSI matrix to sparse the eigenvector matrix and project it onto the angular delay domain matrix; A feature separation module is used to separate the real and imaginary parts of the angle delay domain matrix to form a dual-channel matrix and compress it at the UE; Feature extraction module, used to extract the feature vector of the dual-channel matrix using a multi-head loop mechanism in the encoder; The compression output module is used to process the extracted feature vector through the fully connected layer and output the compressed codeword vector; Quantization transmission module, used to quantize the compressed codeword vector into a bit stream and transmit it to the decoder end through the feedback link; The decoding recovery module is used to recover the bit stream into a codeword vector through inverse quantization processing at the decoder end, and to recover the recovered codeword vector into a CSI matrix through neural network processing and inverse discrete Fourier transform.

8. A computer device, characterized in that: including processor and memory; The processor runs a program corresponding to the executable program code by reading the executable program code stored in the memory, so as to implement the large-scale MIMO CSI feedback method using a lightweight space-time resolution network as described in any one of claims 1-6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the massive MIMO CSI feedback method using a lightweight space-time resolution network is implemented.

Images