CN115987339A - A Deep Learning-based Encoder-Decoder Decoupling CSI Feedback Method - Google Patents

Abstract

Translated fromChinese

本发明提出一种基于深度学习的编码器解码器解耦CSI反馈方法，包括以下步骤1.基于特征向量的CSI数据集生成；步骤2.在基站端训练编码器解码器；步骤3.在基站终端训练编码器。本发明所述基于深度学习的编码器解码器解耦CSI反馈方法的优越效果在于：使得CSI反馈能够适应大规模MIMO系统，在保证CSI恢复准确性的同时，减小反馈开销，考虑到厂家对模型算法的保护，使得不同厂家之间的模型编码器与解码器能够解耦来搭配使用，在保证CSI恢复性能的同时，使得终端和基站端的模型配置更加灵活。

The present invention proposes an encoder-decoder decoupling CSI feedback method based on deep learning, comprising the following steps: 1. Generation of CSI data sets based on feature vectors; Step 2. Training the encoder-decoder at the base station; Step 3. At the base station Terminal trains the encoder. The superior effect of the deep learning-based encoder-decoder decoupling CSI feedback method of the present invention is that: the CSI feedback can be adapted to a large-scale MIMO system, and the feedback overhead can be reduced while ensuring the accuracy of CSI recovery. The protection of the model algorithm enables the model encoders and decoders of different manufacturers to be decoupled and used together. While ensuring the CSI recovery performance, it makes the model configuration of the terminal and base station more flexible.

Description

Translated fromChinese

一种基于深度学习的编码器解码器解耦CSI反馈方法A CSI feedback method for encoder-decoder decoupling based on deep learning

技术领域Technical Field

本发明涉及解码器解耦信道状态信息反馈方法，更详细地说涉及基于深度学习的编码器解码器解耦信道状态信息反馈方法。The present invention relates to a decoder decoupling channel state information feedback method, and more particularly to an encoder-decoder decoupling channel state information feedback method based on deep learning.

背景技术Background Art

为了理解基于深度学习的编码器解码器解耦的信道状态信息(Channel stateinformation，CSI)反馈方法，有必要先介绍有关CSI反馈技术、大规模多输入多输出(Multiple-Input-Multiple-Output，MIMO)技术以及深度学习自编码器结构的基本理论知识。In order to understand the channel state information (CSI) feedback method of encoder-decoder decoupling based on deep learning, it is necessary to first introduce the basic theoretical knowledge about CSI feedback technology, large-scale multiple-input multiple-output (MIMO) technology and deep learning autoencoder structure.

1.大规模MIMO技术1. Massive MIMO Technology

MIMO技术即多入多出技术，指在发射端和接收端分别使用多个发射天线和接收天线，使信号通过发射端与接收端的多个天线传送和接收，从而改善通信质量。该技术能够充分利用空间资源，通过多个天线实现多发多收，在不增加频谱资源和天线发射功率的情况下，可以成倍的提高系统信道容量，显示出明显的优势，被视为下一代移动通信的核心技术。大规模MIMO技术具有如下的优点：(1)提升频谱效率和空间分辨率，极高的空间自由度可以满足多个用户在同一时频资源上同时通信；(2)提升能量效率，大规模天线阵列的使用提高了阵列增益，能够使用较小的发射功率达到较好的通信质量，使得系统能量效率提升几个数量级；(3)降低系统复杂度；(4)降低系统部署成本，MIMO极高的空间自由度可以降低发射信号的峰均比，从而能够在射频前端采用低线性复杂度、低成本、低功耗的设备，大幅降低部署成本。MIMO technology, or multiple-input multiple-output technology, refers to the use of multiple transmitting antennas and receiving antennas at the transmitting end and the receiving end respectively, so that the signal is transmitted and received through the multiple antennas at the transmitting end and the receiving end, thereby improving the communication quality. This technology can make full use of spatial resources and achieve multiple transmission and multiple reception through multiple antennas. Without increasing the spectrum resources and antenna transmission power, it can increase the system channel capacity exponentially, showing obvious advantages and is regarded as the core technology of the next generation of mobile communications. Large-scale MIMO technology has the following advantages: (1) Improved spectrum efficiency and spatial resolution. The extremely high spatial degree of freedom can meet the simultaneous communication of multiple users on the same time-frequency resources; (2) Improved energy efficiency. The use of large-scale antenna arrays increases the array gain, and can achieve better communication quality with a smaller transmission power, which improves the system energy efficiency by several orders of magnitude; (3) Reduced system complexity; (4) Reduced system deployment cost. The extremely high spatial degree of freedom of MIMO can reduce the peak-to-average ratio of the transmitted signal, so that low linear complexity, low-cost, and low-power devices can be used in the RF front end, greatly reducing deployment costs.

2.CSI反馈技术2.CSI feedback technology

在MIMO无线通信系统中，通过对多根发送天线进行预编码或者波束赋形，能够达到提升传输效率和可靠性的目的。为了能够实现高性能的预编码或者波束赋形，预编码矩阵或波束成矢量需要比较好的匹配信道，这就需要发送端能够较好的获得信道状态信息。因此，CSI反馈是在MIMO系统中实现高性能预编码或波束成型的关键技术。与此同时，进行CSI反馈时，对信道矩阵的量化反馈会带来比较大的反馈开销，特别是支持大带宽的CSI反馈时，因此，反馈开销是限制性能提升的重要问题。而基本的CSI反馈流程由基站端发送导频到终端，终端根据导频进行信道估计得到CSI数据，终端对CSI数据进行压缩和量化，并通过反馈链路传回到基站端。基站端接收到压缩量化的比特流数据后，进行解量化和重构，得到恢复的CSI数据从而进行后续波束赋形等操作。In a MIMO wireless communication system, the transmission efficiency and reliability can be improved by precoding or beamforming multiple transmitting antennas. In order to achieve high-performance precoding or beamforming, the precoding matrix or beamforming vector needs to match the channel well, which requires the transmitter to obtain channel state information well. Therefore, CSI feedback is a key technology for achieving high-performance precoding or beamforming in MIMO systems. At the same time, when performing CSI feedback, the quantization feedback of the channel matrix will bring a relatively large feedback overhead, especially when supporting CSI feedback with a large bandwidth. Therefore, feedback overhead is an important issue that limits performance improvement. The basic CSI feedback process is that the base station sends a pilot to the terminal, and the terminal performs channel estimation based on the pilot to obtain CSI data. The terminal compresses and quantizes the CSI data and transmits it back to the base station through the feedback link. After receiving the compressed and quantized bit stream data, the base station performs dequantization and reconstruction to obtain the recovered CSI data for subsequent beamforming and other operations.

3.深度学习自编码器3. Deep Learning Autoencoder

深度学习属于机器学习的一部分，与机器学习不同的是深度学习在进行数据特征提取时不需要人工，而是依赖搭建的神经网络自动学习和提取数据特征。其中，自编码器是深度学习中经典的模型结构。自编码器属于无监督学习，包括编码器与解码器两部分，是使得输入数据与输出数据尽可能相同的神经网络。其中，编码器用于对输入数据进行压缩，达到对数据进行降维的效果。解码器对压缩的数据进行重构，使得恢复数据尽可能与原始输入数据相同。从自编码器功能来看，与CSI反馈流程大体相同，均是先对数据进行压缩降维，再传输到另一端对数据进行重构恢复。因此，目前将深度学习与CSI反馈技术相结合也成为了无线通信中的研究热点，通过深度学习来提高MIMO系统中CSI反馈的恢复精度，并同时减小反馈开销。Deep learning is part of machine learning. Unlike machine learning, deep learning does not require manual work when extracting data features, but relies on the built neural network to automatically learn and extract data features. Among them, the autoencoder is a classic model structure in deep learning. The autoencoder belongs to unsupervised learning, including an encoder and a decoder. It is a neural network that makes the input data and the output data as similar as possible. Among them, the encoder is used to compress the input data to achieve the effect of reducing the dimension of the data. The decoder reconstructs the compressed data so that the recovered data is as similar as possible to the original input data. From the perspective of the autoencoder function, it is roughly the same as the CSI feedback process. Both compress and reduce the dimension of the data first, and then transmit it to the other end to reconstruct and restore the data. Therefore, the combination of deep learning and CSI feedback technology has also become a research hotspot in wireless communications. Deep learning is used to improve the recovery accuracy of CSI feedback in MIMO systems and reduce feedback overhead at the same time.

4.目前系统不足与缺点4. Current system deficiencies and shortcomings

对于频分复用(Frequency division duplexity,FDD)下的大规模MIMO系统，上行链路和下行链路工作在不同频率上，因此下行CSI由用户端获得，并通过反馈链路传送回基站。而传统的CSI反馈方法无论是基于码本实现的，还是基于压缩感知实现的，都会随着系统中天线数的增加，进而增加计算复杂度，影响实际通信系统的实时性。因此传统的CSI反馈方案无法很好的应用于大规模MIMO系统中。For large-scale MIMO systems under frequency division duplexity (FDD), the uplink and downlink operate at different frequencies, so the downlink CSI is obtained by the user end and transmitted back to the base station through the feedback link. The traditional CSI feedback method, whether based on codebook or compressed sensing, will increase the computational complexity as the number of antennas in the system increases, affecting the real-time performance of the actual communication system. Therefore, the traditional CSI feedback solution cannot be well applied to large-scale MIMO systems.

因此考虑将深度学习与CSI反馈相结合，在提高CSI反馈恢复精度的同时能够有效降低反馈带来的开销。目前，基于深度学习的CSI反馈主要应用自编码器结构，将编码器配置在终端，对数据进行压缩；而解码器配置在基站端，接收终端发送的压缩数据并进行重构。然而，实际中可能遇到终端与基站厂家不同的问题，出于对技术的保护，可能厂家所设计的基于深度学习的模型并不公开，从而使得其他厂家的终端无法及时配置相应的编码器。Therefore, we consider combining deep learning with CSI feedback, which can effectively reduce the overhead caused by feedback while improving the accuracy of CSI feedback recovery. At present, CSI feedback based on deep learning mainly uses the autoencoder structure, where the encoder is configured at the terminal to compress the data; while the decoder is configured at the base station to receive the compressed data sent by the terminal and reconstruct it. However, in practice, the problem of different terminal and base station manufacturers may be encountered. For the protection of technology, the deep learning-based model designed by the manufacturer may not be made public, making it impossible for other manufacturers' terminals to configure the corresponding encoder in time.

发明内容Summary of the invention

针对现有技术存在的缺陷或，本发明提出一种基于深度学习的编码器解码器解耦CSI反馈方法，In view of the defects of the prior art, the present invention proposes a encoder-decoder decoupling CSI feedback method based on deep learning.

所述CSI反馈方法，包括以下步骤：The CSI feedback method comprises the following steps:

步骤1.基于特征向量的CSI数据集生成：Step 1. CSI dataset generation based on feature vector:

步骤1.1进行全信道信息数据子带划分，与全信道信息数据相同，由基站端发送导频到终端，终端根据导频进行理想信道估计，得到全信道信息数据H，信道估计是根据基站端发送的参考信号与终端接收到的信号进行对比进而求得信道信息，而理想信道估计则是完全已知信道信息，能够直接使用作为信道数据；接下来对得到的全信道信息数据H进行子带划分，分为K个子带且每个子带中包含N_sc个子载波，其中的第k个子带的信道数据记作H_k；Step 1.1 performs sub-band division of the full channel information data. The same as the full channel information data, the base station sends a pilot to the terminal, and the terminal performs ideal channel estimation based on the pilot to obtain the full channel information data H. The channel estimation is to obtain the channel information by comparing the reference signal sent by the base station with the signal received by the terminal, while the ideal channel estimation is the complete known channel information, which can be directly used as the channel data; Next, the obtained full channel information data H is sub-band divided into K sub-bands, and each sub-band contains N_sc sub-carriers, and the channel data of the kth sub-band is recorded as H_k ;

步骤1.2求每个子带中信道矩阵的相关阵；并对相关阵进行累加求和求平均，其表达式如下式(1)：Step 1.2 calculates the correlation matrix of the channel matrix in each subband; and accumulates and averages the correlation matrix, which is expressed as follows:

上式(1)中：H_k,i为第k个子带中第i个子载波的信道矩阵，R_k为第k个子带的相关阵，则K个子带的相关阵如下式(2)：In the above formula (1), H_k,i is the channel matrix of the i-th subcarrier in the k-th subband, R_k is the correlation matrix of the k-th subband, and the correlation matrix of the K subbands is as follows (2):

R＝[R₁,R₂,…,R_K]......(2)，R＝[R₁ ,R₂ ,…,R_K ]……(2),

步骤1.3对相关阵进行特征值分解，得到每个子带的特征向量，每个子带的特征向量计算表达式如下式(3)：Step 1.3 performs eigenvalue decomposition on the correlation matrix to obtain the eigenvector of each subband. The calculation expression of the eigenvector of each subband is as follows:

R_kw_k＝λ_kw_k.......(3)，R_k w_k =λ_k w_k .......(3),

上式(3)中：λ_k为第k个子带的最大特征值，w_k为第k个子带的特征，获得基于特征向量的CSI数据，其表达如下式(4)所示：In the above formula (3), λ_k is the maximum eigenvalue of the k-th subband, w_k is the feature of the k-th subband, and CSI data based on the feature vector is obtained, which is expressed as shown in the following formula (4):

步骤2.在基站端训练编码器解码器：Step 2. Train the encoder-decoder on the base station:

步骤2.1基站端根据生成的基于特征向量的CSI数据训练CSI反馈模型，包括编码器f_E(·)与解码器f_D(·)，训练过程中，编码器首先将基于特征向量的CSI数据w作为输入进行压缩，得到压缩后的数据x，对x先进均匀行量化，均匀量化成比特流数据s，再将比特流数据s传入到解码器端，解码器端接收到比特流数据s后，解码器对比特流数据s进行解量化和重构，得到恢复的数据

记为下式(5)：Step 2.1 The base station trains the CSI feedback model based on the generated CSI data based on the feature vector, including the encoder_fE (·) and the decoder_fD (·). During the training process, the encoder first compresses the CSI data w based on the feature vector as input to obtain the compressed data x, and then uniformly quantizes x into bit stream data s, and then transmits the bit stream data s to the decoder. After the decoder receives the bit stream data s, the decoder dequantizes and reconstructs the bit stream data s to obtain the recovered data

It is expressed as the following formula (5):

上式(5)中：编码器与解码器的参数集θ_E和θ_D随着训练过程不断更新，如下式(6)所示：In the above formula (5), the parameter sets_θE and_θD of the encoder and decoder are continuously updated during the training process, as shown in the following formula (6):

上式(6)中：L(·,·)为参数集θ_E和θ_D更新的损失函数，对于CSI反馈采用余弦相似度来评估其恢复准确性的性能，在设计模型训练的损失函数时采用余弦相似度的平方作为损失函数(Square of generalized cosine similarity,SGCS)，表示为下式(7)，In the above formula (6), L(·,·) is the loss function for updating the parameter sets_θE and_θD . For CSI feedback, cosine similarity is used to evaluate the performance of its recovery accuracy. When designing the loss function for model training, the square of cosine similarity (Square of generalized cosine similarity, SGCS) is used as the loss function, which is expressed as the following formula (7):

步骤2.2.训练完成后，将基于特征向量的CSI数据w通过模型的编码器得到压缩待量化的数据x，将w和x一同传入到终端进行下一步训练，同时，将训练好的解码器f_D(·)配置在基站端；Step 2.2. After the training is completed, the CSI data w based on the feature vector is passed through the encoder of the model to obtain the compressed data x to be quantized, and w and x are transmitted to the terminal together for the next step of training. At the same time, the trained decoder f_D (·) is configured at the base station end;

步骤3.基站终端训练编码器：Step 3. The base station terminal trains the encoder:

步骤3.1基站终端根据接收到的基于特征向量的CSI数据w和压缩待量化的数据x来设计编码器h_E(·)，鉴于终端计算能力所限及功耗，在设计h_E(·)时使得模型简单，则考虑以MixerNet编码器或EVCsiNet编码器作为终端的编码；Step 3.1 The base station terminal designs an encoder h_E (·) according to the received CSI data w based on the feature vector and the compressed data to be quantized x. In view of the limitation of the terminal computing power and power consumption, the model is kept simple when designing h_E (·), and MixerNet encoder or EVCsiNet encoder is considered as the terminal encoding;

步骤3.2编码器h_E(·)将基于特征向量的CSI数据w作为原始数据，压缩待量化的数据x作为标签一同输入到编码器h_E(·)中进行训练，编码器h_E(·)输出得到压缩待量化的数据，记为

本过程记为下式(8)：Step 3.2 Encoder h_E (·) takes the CSI data w based on the feature vector as the original data and the compressed data to be quantized x as the label and inputs them into encoder h_E (·) for training. Encoder h_E (·) outputs the compressed data to be quantized, which is recorded as

This process is recorded as the following formula (8):

上式(8)中：编码器的参数集θ_E′随着训练过程不断更新，表达为下式(9)：In the above formula (8), the encoder parameter set θ_E ′ is continuously updated during the training process and is expressed as the following formula (9):

上式(9)中：L′(·)为参数集θ_E′更新的损失函数，编码器h_E(·)的损失函数采用均方误差损失函数(Mean squared error,MSE)，表达式如下式(10)：In the above formula (9), L′(·) is the loss function for updating the parameter set θ_E ′, and the loss function of the encoder h_E (·) adopts the mean squared error loss function (MSE), which is expressed as follows (10):

上式(10)中：N为数据样本个数，x_i表示第i个输入到编码器h_E(·)数据样本对应的标签，

表示第i个输入到编码器h_E(·)数据样本对应的输出，训练完成后，将编码器h_E(·)配置在终端。In the above formula (10), N is the number of data samples,_xi represents the label corresponding to the i-th data sample input to the encoder_hE (·),

It represents the output corresponding to the i-th data sample input to the encoder h_E (·). After the training is completed, the encoder h_E (·) is configured at the terminal.

本发明所述CSI反馈方法具有以下优越技术效果：The CSI feedback method of the present invention has the following superior technical effects:

1.本发明所述CSI反馈方法使用基于深度学习的方法，使得CSI反馈能够适应大规模MIMO系统，在保证CSI恢复准确性的同时，减小反馈开销。1. The CSI feedback method described in the present invention uses a deep learning-based method to enable CSI feedback to adapt to large-scale MIMO systems, while ensuring CSI recovery accuracy and reducing feedback overhead.

2.本发明所述CSI反馈方法所提出的编码器解码器解耦的CSI反馈方法，考虑到厂家对模型算法的保护，使得不同厂家之间的模型编码器与解码器能够解耦来搭配使用，在保证CSI恢复性能的同时，使得终端和基站端的模型配置更加灵活。2. The CSI feedback method of encoder-decoder decoupling proposed in the CSI feedback method of the present invention takes into account the manufacturer's protection of the model algorithm, so that the model encoders and decoders between different manufacturers can be decoupled and used in combination, while ensuring the CSI recovery performance, making the model configuration of the terminal and base station more flexible.

3.本发明所述CSI反馈方法与先前的基于深度学习的CSI反馈模型相比，终端和基站端只能配固定的模型编码器与解码器，而本发明所述CSI反馈方法提出的编码器解码器解耦的CSI反馈方法，使得两端配置模型更加灵活，同时兼顾了厂家对自己模型算法的保护，且终端与基站端所传输的仅是数据，而非模型，也减小了传输带来的开销。3. Compared with the previous CSI feedback model based on deep learning, the CSI feedback method described in the present invention can only be equipped with fixed model encoders and decoders at the terminal and the base station. The CSI feedback method described in the present invention proposes a CSI feedback method with encoder-decoder decoupling, which makes the model configuration method at both ends more flexible, while taking into account the manufacturer's protection of its own model algorithm. Moreover, what is transmitted between the terminal and the base station is only data, not a model, which also reduces the overhead caused by transmission.

附图说明BRIEF DESCRIPTION OF THE DRAWINGS

图1是本发明所述CSI反馈方法流程示意图。FIG1 is a schematic flow chart of the CSI feedback method according to the present invention.

具体实施方式DETAILED DESCRIPTION

为了能够更清楚地理解本发明的上述目的、特征和优点，下面结合附图和具体实施方式对本发明所述CSI反馈方法的具体实施例作进一步的详细描述。In order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the specific embodiments of the CSI feedback method of the present invention are further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

如图1所示，所述CSI反馈方法，将终端训练好的编码器及基站端训练好的解码器分别配置在两侧，即可实现编码器与解码器解耦的CSI反馈方法。实际通信过程中，基站端发送导频到终端，终端通过理想信道估计得到全信道信息数据H，再生成对应的基于特征向量的CSI数据w，将其作为终端编码器h_E(·)的输入，进行数据压缩得到压缩待量化数据

将压缩待量化

进行均匀量化得到比特流数据

并将其通过反馈链路传送给基站端，基站端将比特流数据

作为解码器f_D(·)的输入，对其进行解量化与重构，得到恢复的基于特征向量的CSI数据

用于后续波束赋形的操作。As shown in FIG1 , the CSI feedback method configures the encoder trained by the terminal and the decoder trained by the base station on both sides, respectively, to realize the CSI feedback method of decoupling the encoder and the decoder. In the actual communication process, the base station sends a pilot to the terminal, and the terminal obtains the full channel information data H through ideal channel estimation, and then generates the corresponding CSI data w based on the feature vector, which is used as the input of the terminal encoder h_E (·) to perform data compression to obtain the compressed data to be quantized.

Compression to be quantized

Uniform quantization is performed to obtain bit stream data

The base station transmits the bit stream data to the base station through the feedback link.

As the input of the decoder f_D (·), it is dequantized and reconstructed to obtain the restored CSI data based on the feature vector

Used for subsequent beamforming operations.

本发明并不限于上述实施方式，在不背离本发明实质内容的情况下，本领域技术人员可以想到的任何变形、改进、替换均落入本发明的保护范围。The present invention is not limited to the above-mentioned embodiments. Without departing from the essential content of the present invention, any deformation, improvement and substitution that can be thought of by those skilled in the art shall fall into the protection scope of the present invention.

Claims (4)

1. A encoder and decoder decoupling CSI feedback method based on deep learning is characterized by comprising the following steps:

step 1, generating a CSI data set based on the characteristic vector;

step 2, training a coder decoder at a base station end;

and 3, training an encoder at the base station terminal.

2. The method for decoupling CSI feedback by a deep learning based coder-decoder as claimed in claim 1, wherein in step 1, said CSI data set based on eigenvectors is generated, comprising the steps of:

step 1.1, sub-band division of full channel information data is carried out, the same as the full channel information data, a base station end sends pilot frequency to a terminal, the terminal carries out ideal channel estimation according to the pilot frequency to obtain full channel information data H, the channel estimation is carried out by comparing a reference signal sent by the base station end with a signal received by the terminal to obtain channel information, and the ideal channel estimation is completely known channel information and can be directly used as channel data; then, the obtained full channel information data H is divided into K sub-bands, and each sub-band comprises N_sc Sub-carriers in which channel data of the k-th sub-band is denoted as H_k ；

Step 1.2, solving a correlation matrix of a channel matrix in each sub-band; and carrying out accumulation summation and averaging on the correlation matrix, wherein the expression is as the following formula (1):

in the above formula (1): h_k,i Channel matrix for ith subcarrier in kth subband, R_k The correlation matrix of the K-th subband is then the following equation (2):

R＝[R₁ ,R₂ ,…,R_K ]......(2)，

step 1.3, performing eigenvalue decomposition on the correlation matrix to obtain an eigenvector of each subband, wherein the calculation expression of the eigenvector of each subband is as follows (3):

R_k w_k ＝λ_k w_k .......(3)，

in the above formula (3): lambda [ alpha ]_k Is the maximum eigenvalue of the kth subband, w_k For the features of the kth subband, CSI data based on the feature vector is obtained, which is expressed as the following equation (4):

3. the method for decoupling CSI feedback by deep learning based encoder-decoder as claimed in claim 1, wherein in step 2, said training the encoder-decoder at the base station end comprises the following steps:

step 2.1, the base station end trains a CSI feedback model according to the generated CSI data based on the characteristic vector, and the CSI feedback model comprises an encoder f_E (. O) and decoder f_D In the training process, an encoder firstly compresses CSI data w based on a characteristic vector as input to obtain compressed data x, uniformly quantizes the x into bit stream data s, then transmits the bit stream data s to a decoder, and after the decoder receives the bit stream data s, the decoder dequantizes and reconstructs the bit stream data s to obtain recovered data

As represented by the following formula (5):

in the above formula (5): parameter set theta for encoder and decoder_E And theta_D With the continuous updating of the training process, as shown in the following formula (6):

in the above formula (6): l (·,. Cndot.) is a parameter set θ_E And theta_D The updated loss function adopts cosine similarity for CSI feedback to evaluate the performance of recovering accuracy, adopts the Square of the cosine similarity as the loss function (SGCS) in designing the loss function of model training, and is expressed as the following formula (7),

and 2.2, after training is finished, obtaining compressed data x to be quantized by the CSI data w based on the characteristic vector through a model encoder, transmitting w and x into the terminal together for further training, and meanwhile, transmitting the trained decoder f_D Disposed at the base station side.

4. The deep learning based encoder-decoder decoupling CSI feedback method as claimed in claim 1, wherein in step 3, training an encoder at the base station terminal comprises the following steps:

step 3.1 the base station terminal designs the encoder h according to the received CSI data w based on the eigenvector and the compressed data x to be quantized_E (. H) design h in view of the limited computing power and power consumption of the terminal_E When the model is simple, a MixerNet coder or an EVCsiNet coder is taken as the coding of the terminal;

step 3.2 encoder h_E (. Take the CSI data w based on the feature vector as the original data, compress the waiting quantityThe converted data x is input as a tag to the encoder h_E In (h) training, encoder h_E Output compressed data to be quantized, and record

The training process is represented by the following formula (8):

in the above formula (8): parameter set theta of encoder_E ' with the continuous update of the training process, it is expressed as the following formula (9):

in the above formula (9): l '(. Cndot.) is parameter set θ'_E Updated loss function, encoder h_E The loss function of (DEG) is a Mean squared error loss function (MSE), and is expressed by the following formula (10):

in the above formula (10): n is the number of data samples, x_i Represents the ith input to encoder h_E (ii) a label corresponding to the data sample,

represents the ith input to encoder h_E (v) outputting corresponding data samples, and after the training is finished, enabling the encoder h_E Disposed in the terminal. />

Images