WO2024045689A1 - 一种基于深度学习的csi压缩反馈方法 - Google Patents

一种基于深度学习的csi压缩反馈方法 Download PDF

Info

Publication number
WO2024045689A1
WO2024045689A1 PCT/CN2023/093244 CN2023093244W WO2024045689A1 WO 2024045689 A1 WO2024045689 A1 WO 2024045689A1 CN 2023093244 W CN2023093244 W CN 2023093244W WO 2024045689 A1 WO2024045689 A1 WO 2024045689A1
Authority
WO
WIPO (PCT)
Prior art keywords
csi
neural network
dimensional
sequence
base station
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2023/093244
Other languages
English (en)
French (fr)
Inventor
张朝阳
陈子瑞
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University
Original Assignee
Zhejiang University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University filed Critical Zhejiang University
Publication of WO2024045689A1 publication Critical patent/WO2024045689A1/zh
Priority to US18/969,330 priority Critical patent/US12438584B2/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/16—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks using machine learning or artificial intelligence
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/0413—MIMO systems
    • H04B7/0417—Feedback systems
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/06—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station
    • H04B7/0613—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission
    • H04B7/0615—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission of weighted versions of same signal
    • H04B7/0619—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station using simultaneous transmission of weighted versions of same signal using feedback from receiving side
    • H04B7/0621—Feedback content
    • H04B7/0626—Channel coefficients, e.g. channel state information [CSI]
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L5/00—Arrangements affording multiple use of the transmission path
    • H04L5/0001—Arrangements for dividing the transmission path
    • H04L5/0003—Two-dimensional division
    • H04L5/0005—Time-frequency
    • H04L5/0007—Time-frequency the frequencies being orthogonal, e.g. OFDM(A) or DMT

Definitions

  • the present invention relates to the field of wireless communications, and in particular, to a CSI compression feedback method based on deep learning.
  • MIMO-input multiple-output (MIMO) antenna arrays and orthogonal frequency division multiplexing (OFDM) are two key technologies in fifth-generation (5G) and beyond (B5G) communication systems.
  • MIMO-OFDM systems in order to bring sufficient performance gains, base stations need accurate downlink channel state information (CSI) for beamforming, subcarrier allocation and power control.
  • CSI channel state information
  • FDD frequency division duplex
  • the downlink CSI needs to be estimated on the user equipment side and fed back to the base station by the user equipment side.
  • the number of antennas and subcarriers is usually very high, the data dimension of CSI is very high, which leads to high CSI feedback cost.
  • MIMO-OFDM CSI is usually expressed mathematically as a matrix, in which row and column indexes correspond to antenna and subcarrier index information respectively. Due to the similarity of electromagnetic wave transmission between different antennas and subcarriers, there is an implicit correlation between elements in the CSI matrix, which makes the entire CSI matrix compressible. Therefore, how to effectively compress CSI to reduce feedback cost has become an important topic in the field of wireless communications.
  • CNN convolutional neural networks
  • the purpose of the present invention is to propose a more efficient CSI compression and reconstruction method based on the shortcomings of the existing technology to better complete the CSI feedback task.
  • a CSI compression feedback method based on deep learning including the following steps:
  • the base station collects CSI data from historical communication data as training data for the neural network
  • the base station builds a two-dimensional sequence-to-sequence structure neural network and randomly initializes the neural network parameters;
  • the two-dimensional sequence-to-sequence structure neural network includes an encoder and a decoder;
  • the encoder part consists of a two-dimensional recurrent neural network, A fully connected network is cascaded;
  • the decoder part is composed of a fully connected network and a two-dimensional recurrent neural network cascaded;
  • the encoder part is used to input the sub-blocks divided by CSI into the corresponding units, and The unit output is spliced into a matrix in the order of the corresponding units and then compressed into a vector;
  • the decoder part is used to reshape the compressed vector into a matrix, divide it into several sub-blocks and input it into the unit, and splice the unit output in the order of the corresponding units.
  • the base station uses the collected CSI data to train the two-dimensional sequence-to-sequence structure neural network in S1.2, and continuously trains the parameters of the two-dimensional sequence-to-sequence neural network through the gradient descent method until convergence;
  • the base station delivers the two-dimensional sequence trained in S1.3 to the encoder part of the sequence structure neural network to the user;
  • S3 The user end uses the encoder part delivered in S2 to compress the downlink CSI and feeds the compressed information back to the base station;
  • the base station uses the decoder part of the two-dimensional sequence-to-sequence structure neural network trained in S1.3 to reconstruct the compressed information fed back by the user end in step S3 to obtain complete downlink CSI.
  • the two-dimensional recurrent neural network uses a two-dimensional long short-term memory network.
  • Each sub-block obtained after MIMO-OFDM CSI division is input into the corresponding unit of the encoder's two-dimensional long short-term memory network.
  • the output size of the unit in its two-dimensional long short-term memory network is the same as the sub-block dimension; all
  • the output of the unit is spliced into a matrix of size N t ⁇ N c in the order of the corresponding units, and the matrix is stretched into a vector of length N t ⁇ N c ; then this vector of length N t ⁇ N c is input to the fully connected layer. , get the compression vector h com .
  • the loss function Loss( ⁇ ) of the neural network is defined as follows:
  • ⁇ is the parameter set of the neural network
  • N is the number of training data
  • H n is the nth original training data
  • It is the output obtained after inputting H n into the two-dimensional sequence into the sequence structure neural network.
  • the neural network training in S1.3 is based on the currently common optimizer based on gradient descent; specifically, the optimizer generates gradients corresponding to the neural network parameters by deriving the loss function and back-propagating the gradient. Gradient descent is used to continuously optimize the parameters of the neural network until the network converges.
  • the beneficial effects of the present invention are: the CSI compression feedback method based on 2D Seq2Seq proposed by the present invention can achieve high-precision CSI feedback while using a feedback vector with fewer dimensions, and solves the problem of other current deep learning-based CSI feedback The performance bottleneck of the method. And the calculation of this plan is complex The relationship between complexity and the number of antennas N t and the number of subcarriers N c is linear, the calculation complexity is low, and it is suitable for actual deployment needs.
  • Figure 1 is a schematic diagram of the overall framework of 2D Seq2Seq deep learning
  • Figure 2 is the overall structure and calculation diagram of 2D LSTM
  • Figure 3 is a schematic diagram of the internal structure and calculation of 2D LSTM cell
  • Figure 4 is a diagram of the sub-block division method of MIMO-OFDM CSI
  • Figure 5 is a comparison diagram between the CSI compression feedback method based on 2D Seq2Seq and other commonly used methods.
  • the base station of the large-scale access system is equipped with 32 antennas, and each user terminal is configured with one antenna.
  • the base station uses the OFDM mode with 256 subcarriers to serve users.
  • this embodiment provides a deep learning-based CSI feedback solution, which includes the following steps:
  • the base station collects 40,000 CSI data from historical communication data as training data for the neural network
  • the base station builds a two-dimensional sequence-to-sequence (2D Seq2Seq) structure neural network based on the structure shown in Figure 1, randomly initializes the neural network parameters, and sets the size of the 2D LSTM hidden layer in 2D Seq2Seq to 128.
  • 2D Seq2Seq two-dimensional sequence-to-sequence
  • the structure of the 2D Seq2Seq model is as follows.
  • the 2D Seq2Seq model consists of two parts: encoder (Encoder) and decoder (Decoder).
  • the encoder part consists of a 2-layer 2D long short-term memory (LSTM) network and a fully connected network cascade.
  • the decoder part consists of a fully connected network and a 2-layer 2D LSTM network cascade.
  • 2-layer 2D LSTM refers to two cascaded stacked 2D LSTM networks, in which the hidden state information dimensions of the two cascaded 2D LSTM networks are equal.
  • the overall and internal structure and calculation diagram of the 2D LSTM network are shown in Figures 2 and 3.
  • 2D LSTM is a network structure in which parameter-shared 2D LSTM cells are expanded and iteratively calculated in two dimensions.
  • Figure 3 shows the internal structure and calculation process of the 2D LSTM cell.
  • the combination of p, q, p-1 and q-1 represents the cell to which the variable belongs
  • s is the status information
  • the output information of the cell x is the input information information, i is the input gate, o is the output gate, f is the forget gate, ⁇ is the weighted gate, c is the unit memory, is the working memory, ⁇ ( ⁇ ) is the sigmoid function, and g( ⁇ ) is the activation function tanh( ⁇ ).
  • i p,q ⁇ (W i x p,q +U i s p-1,q +V i s p,q-1 +b i )
  • f p,q ⁇ (W f x p,q +U f s p-1,q +V f s p,q-1 +b f )
  • W is the weight matrix of the input part
  • U is the weight matrix of the first-dimensional state information
  • V is the weight matrix of the second-dimensional state information
  • b is the bias; subscript Then respectively represent the gate or memory to which the parameters W, U, V, and b belong.
  • each sub-block is input into the corresponding cell of the two-dimensional LSTM.
  • each unit obtains a reconstructed sub-block, and these sub-blocks are spliced in the order of the corresponding cells to form the reconstructed CSI matrix.
  • the calculation process of the 2D Seq2Seq model is described in 2.c and 2.d below.
  • the sub-block size is set to 1 antenna ⁇ 32 carriers.
  • the Encoder part In the Encoder part, first divide the MIMO-OFDM CSI into several sub-blocks according to the method in step 2.b). Each sub-block is then input into the corresponding cell in the two-layer 2D LSTM in the Encoder part, where the output size of the cell in the two-layer 2D LSTM is the same as the sub-block dimension. After two layers of 2D LSTM, the outputs of all cells are spliced into a matrix of N t ⁇ N c size in the order of the corresponding cells, and the matrix is stretched into a vector of N t ⁇ N c length. Next, input this N t ⁇ N c length vector into the full Connect the layers to obtain the compression vector h com . In this way, the Encoder part completes the task of compressing the original MIMO-OFDM CSI into a lower-dimensional vector.
  • the Decoder part completes the task of reconstructing MIMO-OFDM CSI from the compression vector.
  • the base station uses the collected training data to train the neural network in step 2), and continuously trains the parameters of the 2D Seq2Seq neural network through the backpropagation method until convergence;
  • the loss function (Loss) of the neural network is defined as follows:
  • ⁇ is the parameter of the neural network
  • N is the number of training data
  • H n is the nth original training data
  • the training of the neural network is based on the adaptive momentum estimation (Adam) optimizer.
  • the learning rate is set to 0.001, trained for 800 cycles, and the learning rate decays to 1/5 of the previous one every 100 cycles after the 500th training cycle.
  • the base station delivers the encoder (Encoder) part of the 2D Seq2Seq neural network trained in step 3) to the user;
  • the user uses the Encoder delivered in step 4) to compress the downlink CSI and feeds the compressed information back to the base station;
  • the base station uses the decoder part of the 2D Seq2Seq neural network trained in step 3) to reconstruct the compressed information fed back by the user in step 5) to obtain the complete downlink CSI.
  • the CSI feedback scheme based on 2D Seq2Seq neural network proposed by the present invention is better than the traditional compression method based on convolutional neural network at the same compression length; baseline scheme 1CsiNet and the baseline scheme 2AE_CNN have significant accuracy improvements.

Landscapes

  • Engineering & Computer Science (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Radio Transmission System (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本发明公开了一种基于深度学习的信道状态信息(CSI)压缩反馈方案。在大规模多输入多输出(MIMO)系统中，CSI对于基站实现高性能增益至关重要。在频分双工(FDD)模式下，用户设备(UE)需要估计CSI，然后反馈给基站，有效的CSI压缩将大大降低反馈的代价。本发明使用二维的序列到序列(2D Seq2Seq)结构的神经网络来实现有效的CSI压缩和重建。基站利用CSI数据训练该神经网络的参数，并在训练完成后部署在用户端和基站端已进行有效地CSI压缩和重建。本发明能够实现对CSI的高精度压缩和重建，为MIMO系统提供了一种高效稳定的信道状态信息反馈方法。

Description

一种基于深度学习的CSI压缩反馈方法 技术领域
本发明涉及无线通信领域，尤其涉及一种基于深度学习的CSI压缩反馈方法。
背景技术
多输入多输出(MIMO)天线阵列和正交频分复用(OFDM)是第五代(5G)及以上(B5G)通信系统中的两项关键技术。在MIMO-OFDM系统中，为带来充分的性能增益，基站需要准确的下行信道状态信息(CSI)来进行波束赋形、子载波分配和功率控制。而在频分双工(FDD)模式下，需要在用户设备侧估算下行CSI，由用户设备侧反馈给基站。然而，在MIMO-OFDM系统中，由于天线和子载波的数量通常很高多，使得CSI的数据维度很高，这导致了高昂的CSI反馈成本。
MIMO-OFDM CSI在数学上通常表示为一个矩阵，其中行、列索引分别对应天线、子载波索引信息。由于不同天线和子载波之间电磁波传输的相似性，CSI矩阵中各元素之间存在隐式相关性，这使得整个CSI矩阵是可压缩的。因此，如何有效压缩CSI来降低反馈成本已成为无线通信领域的一个重要课题。
近年来，深度学习技术被广泛应用于无线通信领域，展现出了巨大的应用潜力。但是，当前大多数的深度学习方法都是数据驱动的，通常不能够很充分地挖掘深度学习技术在当前任务上的潜力。
目前的基于深度学习的CSI压缩反馈方法大部分都是使用卷积神经网络(CNN)进行CSI的压缩和重建。然而，由于卷积神经网络自身的平移不变和平滑特性，基于CNN方法的CSI压缩重建方法通常会遇到显著的性能瓶颈。如何从MIMO-OFDM CSI的物理特性出发，设计更契合MIMO-OFDM CSI独特性质的神经网络结构以更充分地发挥深度学习在CSI压缩和重建上的潜力，成为了MIMO OFDM系统CSI压缩反馈任务上的重要问题。
发明内容
本发明的目的是基于现有技术的不足，提出一种更高效的CSI压缩和重建方法，以更好地完成CSI反馈任务。
本发明所采用的具体技术方案如下：一种基于深度学习的CSI压缩反馈方法，包括如下步骤：
S1：建立二维序列到序列结构神经网络并训练；
S1.1：基站从历史通信数据中收集CSI数据，作为神经网络的训练数据；
S1.2：基站搭建二维序列到序列结构神经网络，并随机初始化神经网络参数；所述二维序列到序列结构神经网络包括编码器和解码器；编码器部分由一个二维循环神经网络，一个全连接网络级联而成；解码器部分由一个全连接网络，一个二维循环神经网络级联而成；编码器部分用于将CSI划分出的子块输入到对应的单元中，并将单元输出按照对应单元的顺序拼接成矩阵后压缩为向量；解码器部分用于将压缩向量重塑为矩阵，划分成若干个子块并输入到单元中，将单元输出按照对应单元的顺序进行拼接，得到重构的CSI；
S1.3：基站利用收集的CSI数据来训练S1.2中的二维序列到序列结构神经网络，通过梯度下降方法不断训练二维序列到序列神经网络的参数直至收敛；
S2：基站将S1.3中训练完成的二维序列到序列结构神经网络的编码器部分下发给用户端；
S3：用户端使用S2中下发的编码器部分将下行CSI进行压缩，并将压缩后的信息反馈给基站；
S4：基站使用S1.3中训练完成的二维序列到序列结构神经网络的解码器部分对步骤S3中用户端反馈回来的压缩信息进行重建，得到完整的下行CSI。
进一步地，所述二维循环神经网络使用二维长短时记忆网络。
进一步地，S1.2中CSI划分过程为：
对于Nt×Nc维的多输入多输出正交频分复用MIMO-OFDM CSI，其中Nt是基站天线数量，Nc是子载波数，将CSI划分为由若干个子块组成的二维序列。
进一步地，所述编码器的工作过程为：
将MIMO-OFDM CSI划分后得到的每个子块输入到编码器的二维长短时记忆网络与之对应的单元中，其二维长短时记忆网络中单元的输出大小与子块维度相同；将所有单元的输出按照对应单元的顺序拼接成Nt×Nc大小的矩阵，并将矩阵拉伸为一个Nt×Nc长度的向量；之后将这个Nt×Nc长度的向量输入全连接层，得到压缩向量hcom。
进一步地，所述解码器的工作过程为：
首先将编码器端压缩得到的压缩向量hcom输入解码器部分中的全连接层得到一个Nt×Nc长度的向量，再将这个Nt×Nc长度的向量重塑成一个Nt×Nc维度大小的矩阵；接着，将矩阵按照与CSI划分过程相同的方式划分成若干个子块，依次输入解码器部分二维长短时记忆网络中与之对应的单元中，将所有单元的输出按照对应单元的顺序拼接成Nt×Nc大小的矩阵，即得到重建后的MIMO-OFDM CSI。
进一步地，S1.3中的神经网络训练方法中：
神经网络的损失函数Loss(Θ)定义如下：
其中,Θ是神经网络的参数集，N是训练数据的数量，Hn是第n个原始训练数据，是将Hn输入二维序列到序列结构神经网络后得到的输出。
进一步地，S1.3中的神经网络训练基于当前常见的基于梯度下降的优化器进行；具体来说，优化器通过对损失函数求导并将梯度反向传播产生神经网络参数对应的梯度，进行梯度下降，来不断优化神经网络的参数，直至网络收敛。
本发明具有的有益效果是：本发明提出的基于2D Seq2Seq的CSI压缩反馈方法，可以在使用维度较少的反馈向量的情况下实现高精度的CSI反馈，解决了当前其他基于深度学习的CSI反馈方法的性能瓶颈问题。且本方案的计算复 杂度与天线数Nt、子载波数Nc的关系为线性关系，计算复杂度低，适合实际的部署需求。
附图说明
图1是2D Seq2Seq深度学习整体框架示意图；
图2是2D LSTM的总体结构和计算示意图；
图3是2D LSTM cell的内部结构和计算示意图；
图4是MIMO-OFDM CSI的子块划分方式图；
图5是基于2D Seq2Seq的CSI压缩反馈方法与其他常用方法比较图
具体实施方式
本实施例中，大规模接入系统的基站安装32根天线，每个用户终端配置1根天线，基站使用有256个子载波的OFDM模式服务用户。
基于该基站，本实施例提供了一种基于深度学习的CSI反馈方案，其包括如下步骤：
1)基站端从历史通信数据中收集40000个CSI数据，作为神经网络的训练数据；
2)基站根据图1所示的结构搭建二维序列到序列(2D Seq2Seq)结构神经网络，并随机初始化神经网络参数，并设置2D Seq2Seq中2D LSTM隐含层大小均为128。
2.a)2D Seq2Seq模型的结构组成如下。2D Seq2Seq模型由编码器(Encoder)和解码器(Decoder)两部分组成。编码器部分由一个2层的2D长短时记忆(LSTM)网络和一个全连接网络级联组成。解码器部分由一个全连接网络和一个2层的2D LSTM网络级联组成。2层2D LSTM是指两个级联堆叠的2D LSTM网络，其中级联的两个2D LSTM网络的隐状态信息维度相等。2D LSTM网络的总体和内部结构和计算示意图如图2、3所示。如图2，2D LSTM是由参数共享的2D LSTM cell在两个维度上进行展开和迭代计算的网络结构。在图3中展示了2D LSTM cell的内部结构和计算流程，其中每个变量的下标中，p、q、p-1和q-1的组合表示变量所属的cell，s是状态信息，也是cell的输出信息，x是输入信 息，i是输入门，o是输出门，f是遗忘门，λ是加权门，c是单元记忆，是工作记忆，σ(·)是sigmoid函数，g(·)是激活函数tanh(·)。图3中LSTM cell内部的计算流程如下：
ip,q＝σ(Wixp,q+Uisp-1,q+Visp,q-1+bi)
fp,q＝σ(Wfxp,q+Ufsp-1,q+Vfsp,q-1+bf)
op,q＝σ(Woxp,q+Uosp-1,q+Vosp,q-1+bo)
λp,q＝σ(Wλxp,q+Uλsp-1,q+Vλsp,q-1+bλ)


其中是哈达玛积，W,U,V,b是神经网络的参数
其中W是输入部分的权重矩阵，U是第一维度状态信息的权重矩阵，V是第二维度状态信息的权重矩阵，b是偏置；下标则分别代表参数W,U,V,b所属于的门或记忆。
2.b)对于Nt×Nc维的多输入多输出正交频分复用MIMO-OFDM CSI(Nt是基站天线数量，Nc是子载波数量)，我们将它划分为有若干个子块组成的二维序列，如图4所示。当整个CSI矩阵被输入到二维LSTM时，每个子块被输入到二维LSTM对应的单元(cell)中。同样，利用二维LSTM输出重构CSI矩阵时，每个单元得到一个重构子块，这些子块按照对应cell的顺序进行拼接，形成重构的CSI矩阵。2D Seq2Seq模型的计算过程如下2.c和2.d所述。这里，子块大小设置为1天线×32载波。
2.c)在Encoder部分，首先按照2.b)步骤中的方式划分MIMO-OFDM CSI为若干个子块。然后将每个子块输入到Encoder部分两层2D LSTM中与之对应的cell中，其中两层2D LSTM中cell的输出大小与子块维度相同。在两层2D LSTM后，将所有cell的输出按照对应cell的顺序拼接成Nt×Nc大小的矩阵，并将矩阵拉伸为一个Nt×Nc长度的向量。紧接着，将这个Nt×Nc长度的向量输入全 连接层，得到压缩向量hcom。这样，Encoder部分就完成了将原始的MIMO-OFDM CSI压缩为一个维度较低的向量的任务。
2.d)在Decoder部分，首先将Encoder端压缩得到的压缩向量hcom输入Decoder部分中的全连接层得到一个Nt×Nc长度的向量，再将这个Nt×Nc长度的向量重塑成一个Nt×Nc维度大小的矩阵。接着，将矩阵按照与步骤2.c)中的子块划分方式相同的划分方法划分成若干个子块，依次输入Decoder部分两层2D LSTM中与之对应的cell中，将所有cell的输出按照对应cell的顺序拼接成Nt×Nc大小的矩阵,即得到重建后的MIMO-OFDM CSI。这样，Decoder部分就完成了从压缩向量重构MIMO-OFDM CSI的任务。
3)基站利用收集的训练数据来训练步骤2)中的神经网络，通过反向传播方法不断训练2D Seq2Seq神经网络的参数直至收敛；
3.a)神经网络的损失函数(Loss)定义如下：
其中,Θ是神经网络的参数，N是训练数据的数量，Hn是第n个原始训练数据，是将Hn输入2D Seq2Seq网络后得到的输出。
3.b)神经网络的训练基于自适应动量估计(Adam)优化器进行。学习率设置为0.001，训练800次循环，并且在第500次训练循环后每100个循环学习率衰减为之前的1/5。
4)基站将步骤3)中训练完成的2D Seq2Seq神经网络的编码器(Encoder)部分下发给用户端；
5)用户使用步骤4)中下发的Encoder将下行CSI进行压缩，并将压缩后的信息反馈给基站；
6)基站端使用步骤3)中训练完成的2D Seq2Seq神经网络的解码器(Decoder)部分对步骤5)中用户反馈回来压缩信息进行重建，得到完整的下行CSI。
通过计算机仿真可以看出：如图5所示，本发明所提出的基于2D Seq2Seq神经网络的CSI反馈方案，在相同的压缩长度下相比于传统的基于卷积神经网络的压缩方法；基线方案1CsiNet和基线方案2AE_CNN有着明显的精度提升。这些优势是因为本方案所使用的2D Seq2Seq结构更加契合MIMO-OFDM CSI的物理结构。因此，本发明所提出的基于2D Seq2Seq神经网络的CSI反馈方案为大规模MIMO通信系统提供了一种高效的CSI反馈方案。
以上所述，仅为本发明较佳的具体实施方式，但本发明的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本发明披露的技术范围内，可轻易想到的变化或替换，都应涵盖在本发明的保护范围之内。因此，本发明的保护范围应该以权利要求书的保护范围为准。

Claims (7)

  1. 一种基于深度学习的CSI压缩反馈方法，其特征在于，包括如下步骤：
    S1：建立二维序列到序列结构神经网络并训练；
    S1.1：基站从历史通信数据中收集CSI数据，作为神经网络的训练数据；
    S1.2：基站搭建二维序列到序列结构神经网络，并随机初始化神经网络参数；所述二维序列到序列结构神经网络包括编码器和解码器；编码器部分由一个二维循环神经网络，一个全连接网络级联而成；解码器部分由一个全连接网络，一个二维循环神经网络级联而成；编码器部分用于将CSI划分出的子块输入到对应的单元中，并将单元输出按照对应单元的顺序拼接成矩阵后压缩为向量；解码器部分用于将压缩向量重塑为矩阵，划分成若干个子块并输入到单元中，将单元输出按照对应单元的顺序进行拼接，得到重构的CSI；
    S1.3：基站利用收集的CSI数据来训练S1.2中的二维序列到序列结构神经网络，通过梯度下降方法不断训练二维序列到序列神经网络的参数直至收敛；
    S2：基站将S1.3中训练完成的二维序列到序列结构神经网络的编码器部分下发给用户端；
    S3：用户端使用S2中下发的编码器部分将下行CSI进行压缩，并将压缩后的信息反馈给基站；
    S4：基站使用S1.3中训练完成的二维序列到序列结构神经网络的解码器部分对步骤S3中用户端反馈回来的压缩信息进行重建，得到完整的下行CSI。
  2. 根据权利要求1所述的一种基于深度学习的CSI压缩反馈方法，其特征在于，所述二维循环神经网络使用二维长短时记忆网络。
  3. 根据权利要求2所述的一种基于深度学习的CSI压缩反馈方法，其特征在于，S1.2中CSI划分过程为：
    对于Nt×Nc维的多输入多输出正交频分复用MIMO-OFDM CSI，其中Nt是基站天线数量，Nc是子载波数，将CSI划分为由若干个子块组成的二维序列。
  4. 根据权利要求3所述的一种基于深度学习的CSI压缩反馈方法，其特征在于，所述编码器的工作过程为：
    将MIMO-OFDM CSI划分后得到的每个子块输入到编码器的二维长短时记忆网络与之对应的单元中，其二维长短时记忆网络中单元的输出大小与子块维度相同；将所有单元的输出按照对应单元的顺序拼接成Nt×Nc大小的矩阵，并将矩阵拉伸为一个Nt×Nc长度的向量；之后将这个Nt×Nc长度的向量输入全连接层，得到压缩向量hcom。
  5. 根据权利要求3所述的一种基于深度学习的CSI压缩反馈方法，其特征在于，所述解码器的工作过程为：
    首先将编码器端压缩得到的压缩向量hcom输入解码器部分中的全连接层得到一个Nt×Nc长度的向量，再将这个Nt×Nc长度的向量重塑成一个Nt×Nc维度大小的矩阵；接着，将矩阵按照与CSI划分过程相同的方式划分成若干个子块，依次输入解码器部分二维长短时记忆网络中与之对应的单元中，将所有单元的输出按照对应单元的顺序拼接成Nt×Nc大小的矩阵，即得到重建后的MIMO-OFDM CSI。
  6. 根据权利要求1所述的一种基于深度学习的CSI压缩反馈方法，其特征在于，S1.3中的神经网络训练方法中：
    神经网络的损失函数Loss(Θ)定义如下：
    其中,Θ是神经网络的参数集，N是训练数据的数量，Hn是第n个原始训练数据，是将Hn输入二维序列到序列结构神经网络后得到的输出。
  7. 根据权利要求1所述的一种基于深度学习的CSI压缩反馈方法，其特征在于，S1.3中的神经网络训练基于当前常见的基于梯度下降的优化器进行；具体来说，优化器通过对损失函数求导并将梯度反向传播产生神经网络参数对应的梯度，进行梯度下降，来不断优化神经网络的参数，直至网络收敛。
PCT/CN2023/093244 2022-09-01 2023-05-10 一种基于深度学习的csi压缩反馈方法 Ceased WO2024045689A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US18/969,330 US12438584B2 (en) 2022-09-01 2024-12-05 CSI compression and feedback method based on deep learning

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211065211.4 2022-09-01
CN202211065211.4A CN115549742B (zh) 2022-09-01 2022-09-01 一种基于深度学习的csi压缩反馈方法

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/969,330 Continuation US12438584B2 (en) 2022-09-01 2024-12-05 CSI compression and feedback method based on deep learning

Publications (1)

Publication Number Publication Date
WO2024045689A1 true WO2024045689A1 (zh) 2024-03-07

Family

ID=84726328

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/093244 Ceased WO2024045689A1 (zh) 2022-09-01 2023-05-10 一种基于深度学习的csi压缩反馈方法

Country Status (3)

Country Link
US (1) US12438584B2 (zh)
CN (1) CN115549742B (zh)
WO (1) WO2024045689A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118972207A (zh) * 2024-07-29 2024-11-15 大连大学 基于压缩感知和深度学习的毫米波频分双工大规模mimo系统信道估计方法
CN120128233A (zh) * 2025-04-01 2025-06-10 广州安会科技有限公司 基于大规模mimo系统的信道状态信息反馈压缩方法
WO2026073391A1 (en) * 2024-10-03 2026-04-09 Qualcomm Incorporated Channel state information compression using neural networks

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115549742B (zh) * 2022-09-01 2024-06-07 浙江大学 一种基于深度学习的csi压缩反馈方法
CN117062001B (zh) * 2023-08-01 2024-07-16 哈尔滨工程大学 一种基于可解释的深度学习的5g nr室内定位方法及系统
CN119030579B (zh) * 2024-07-11 2025-12-09 湖南大学 一种基于深度学习的大规模mimo下行csi反馈方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108390706A (zh) * 2018-01-30 2018-08-10 东南大学 一种基于深度学习的大规模mimo信道状态信息反馈方法
CN109672464A (zh) * 2018-12-13 2019-04-23 西安电子科技大学 基于fcfnn的大规模mimo信道状态信息反馈方法
US20200220593A1 (en) * 2019-01-04 2020-07-09 Industrial Technology Research Institute Communication system and codec method based on deep learning and known channel state information
US20210273707A1 (en) * 2020-02-28 2021-09-02 Qualcomm Incorporated Neural network based channel state information feedback
CN115549742A (zh) * 2022-09-01 2022-12-30 浙江大学 一种基于深度学习的csi压缩反馈方法

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3487246B1 (en) * 2012-07-03 2020-06-10 LG Electronics Inc. Method and device for allocating resource for uplink control channel in wireless communicaiton system
EP3314836B1 (en) * 2015-06-27 2022-10-26 Cohere Technologies, Inc. Orthogonal time frequency space communication system compatible with ofdm
CN111339789B (zh) * 2020-02-20 2023-08-01 北京字节跳动网络技术有限公司 一种翻译模型训练方法、装置、电子设备及存储介质
US20230412230A1 (en) * 2021-10-19 2023-12-21 Samsung Electronics Co., Ltd. Systems, methods, and apparatus for artificial intelligence and machine learning based reporting of communication channel information
CN114051234B (zh) * 2021-12-03 2023-05-02 广东工业大学 基于深度学习的irs辅助通信系统的csi压缩恢复方法和装置
CN114614920B (zh) * 2022-03-10 2023-06-30 南京邮电大学 基于学习因子图的数据与模型联合驱动的信号检测方法
CN118541922A (zh) * 2022-03-15 2024-08-23 联发科技（新加坡）私人有限公司 用于多输入多输出（mimo）信道状态信息（csi）反馈的方法及装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108390706A (zh) * 2018-01-30 2018-08-10 东南大学 一种基于深度学习的大规模mimo信道状态信息反馈方法
CN109672464A (zh) * 2018-12-13 2019-04-23 西安电子科技大学 基于fcfnn的大规模mimo信道状态信息反馈方法
US20200220593A1 (en) * 2019-01-04 2020-07-09 Industrial Technology Research Institute Communication system and codec method based on deep learning and known channel state information
US20210273707A1 (en) * 2020-02-28 2021-09-02 Qualcomm Incorporated Neural network based channel state information feedback
CN115549742A (zh) * 2022-09-01 2022-12-30 浙江大学 一种基于深度学习的csi压缩反馈方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
CHEN ZIRUI; ZHANG ZHAOYANG; XIAO ZHUORAN: "Viewing the MIMO Channel as Sequence Rather than Image: A Seq2Seq Approach for Efficient CSI Feedback", 2022 IEEE WIRELESS COMMUNICATIONS AND NETWORKING CONFERENCE (WCNC), IEEE, 10 April 2022 (2022-04-10), pages 2292 - 2297, XP034123559, DOI: 10.1109/WCNC51071.2022.9771730 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118972207A (zh) * 2024-07-29 2024-11-15 大连大学 基于压缩感知和深度学习的毫米波频分双工大规模mimo系统信道估计方法
WO2026073391A1 (en) * 2024-10-03 2026-04-09 Qualcomm Incorporated Channel state information compression using neural networks
CN120128233A (zh) * 2025-04-01 2025-06-10 广州安会科技有限公司 基于大规模mimo系统的信道状态信息反馈压缩方法

Also Published As

Publication number Publication date
US12438584B2 (en) 2025-10-07
US20250096872A1 (en) 2025-03-20
CN115549742B (zh) 2024-06-07
CN115549742A (zh) 2022-12-30

Similar Documents

Publication Publication Date Title
US12438584B2 (en) CSI compression and feedback method based on deep learning
CN111464220B (zh) 一种基于深度学习的信道状态信息重建方法
CN109672464B (zh) 基于fcfnn的大规模mimo信道状态信息反馈方法
Cai et al. Attention model for massive MIMO CSI compression feedback and recovery
CN108847876B (zh) 一种大规模mimo时变信道状态信息压缩反馈及重建方法
CN108390706B (zh) 一种基于深度学习的大规模mimo信道状态信息反馈方法
CN111555781B (zh) 一种基于深度学习注意力机制的大规模mimo信道状态信息压缩及重建方法
CN116248156A (zh) 基于深度学习的大规模mimo信道状态信息反馈及重建方法
Li et al. A novel CSI feedback approach for massive MIMO using LSTM-attention CNN
CN114553280B (zh) 一种基于深度学习大规模mimo系统的csi反馈方法
CN115865145A (zh) 一种基于Transformer的大规模MIMO信道状态信息反馈方法
CN114067819A (zh) 基于跨层相似性知识蒸馏的语音增强方法
Cheng et al. Swin transformer-based CSI feedback for massive MIMO
CN114884549A (zh) 基于深度学习的大规模mimo信道状态信息反馈方法
CN108259398A (zh) 基于变分贝叶斯推断的复杂混合模型的信道估计方法
Liu et al. Spherical normalization for learned compressive feedback in massive MIMO CSI acquisition
TW201944745A (zh) 基於深度學習作為通道狀態資訊之回饋方法
CN114567359A (zh) 一种大规模mimo系统中基于多分辨率融合卷积反馈网络的csi反馈方法
CN108809389B (zh) 基于Givens旋转矩阵复用的QR分解块对角化预编码实现方法
CN116488696A (zh) 一种非对称网络结构的信道信息压缩反馈方法
Li et al. Jointly sparse signal recovery via deep auto-encoder and parallel coordinate descent unrolling
CN118972207B (zh) 基于压缩感知和深度学习的毫米波频分双工大规模mimo系统信道估计方法
CN119583263A (zh) 一种无线基座大模型的设计与训练方法
Zhao et al. AI for Wireless Physical Layer
Zhang et al. PSA-CsiNet: A CSI feedback algorithm for massive MIMO system

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23858728

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 23858728

Country of ref document: EP

Kind code of ref document: A1