CN109194595B - A Neural Network-based Channel Environment Adaptive OFDM Reception Method - Google Patents

A Neural Network-based Channel Environment Adaptive OFDM Reception Method Download PDF

Info

Publication number
CN109194595B
CN109194595B CN201811123542.2A CN201811123542A CN109194595B CN 109194595 B CN109194595 B CN 109194595B CN 201811123542 A CN201811123542 A CN 201811123542A CN 109194595 B CN109194595 B CN 109194595B
Authority
CN
China
Prior art keywords
neural network
deep neural
channel
channel estimation
main
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.)
Active
Application number
CN201811123542.2A
Other languages
Chinese (zh)
Other versions
CN109194595A (en
Inventor
姜培文
徐靖
李沙志远
陈翔宇
陈慕涵
金石
温朝凯
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.)
Southeast University
Original Assignee
Southeast 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 Southeast University filed Critical Southeast University
Priority to CN201811123542.2A priority Critical patent/CN109194595B/en
Publication of CN109194595A publication Critical patent/CN109194595A/en
Application granted granted Critical
Publication of CN109194595B publication Critical patent/CN109194595B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L25/00—Baseband systems
    • H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
    • H04L25/0202—Channel estimation
    • H04L25/0204—Channel estimation of multiple channels
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/391—Modelling the propagation channel
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L25/00—Baseband systems
    • H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
    • H04L25/0202—Channel estimation
    • H04L25/024—Channel estimation channel estimation algorithms
    • H04L25/0254—Channel estimation channel estimation algorithms using neural network algorithms
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L25/00—Baseband systems
    • H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
    • H04L25/0202—Channel estimation
    • H04L25/024—Channel estimation channel estimation algorithms
    • H04L25/0256—Channel estimation using minimum mean square error criteria
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L25/00—Baseband systems
    • H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
    • H04L25/03—Shaping networks in transmitter or receiver, e.g. adaptive shaping networks
    • H04L25/03006—Arrangements for removing intersymbol interference
    • H04L25/03159—Arrangements for removing intersymbol interference operating in the frequency domain

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Power Engineering (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • Electromagnetism (AREA)
  • Cable Transmission Systems, Equalization Of Radio And Reduction Of Echo (AREA)

Abstract

The invention discloses a channel environment self-adaptive OFDM receiving method based on a neural network, which comprises the following steps: selecting a plurality of different channel environments to obtain each trained deep neural network; inputting a receiving frequency domain pilot frequency and a local frequency domain pilot frequency to estimate to obtain a noisy channel estimation result, inputting the noisy channel estimation result into a trained main wiener filtering deep neural network and outputting a main channel estimation result, and obtaining each auxiliary channel estimation result, multiplying the auxiliary channel estimation result by weight and adding the auxiliary channel estimation results to obtain a final channel estimation result; and inputting the final channel estimation result and the received frequency domain data to obtain a zero-forcing equalization result, inputting the zero-forcing equalization result into the trained main and auxiliary equalization deep neural networks to obtain output results of the networks, multiplying the output results by a weight coefficient, adding the multiplied output results, and outputting the sum after hard decision to obtain an estimated bit stream. And acquiring the known frequency domain pilot frequency or channel parameters in actual transmission in real time, and dynamically adjusting the weight coefficient. The invention has the advantages of few parameters, high working efficiency, flexible and rapid on-line switching and the like.

Description

一种基于神经网络的信道环境自适应OFDM接收方法A Neural Network-based Channel Environment Adaptive OFDM Reception Method

技术领域technical field

本发明涉及一种基于神经网络的信道环境自适应OFDM接收方法，属于无线通信技术领域。The invention relates to a channel environment adaptive OFDM receiving method based on a neural network, and belongs to the technical field of wireless communication.

背景技术Background technique

智能通信被认为是下一代移动通信的主要研究方向之一。近年来，作为机器学习中的一个主要分支——深度学习技术，已经被广泛地应用于无线通信物理层研究中，并在性能方面取得了巨大地提升。研究方向包括整个通信系统完全由端到端的深度神经网络替代，或者仅由深度神经网络替换通信系统的部分模块，比如编码器、解码器、检测器等。Intelligent communication is considered as one of the main research directions of next-generation mobile communication. In recent years, as a main branch of machine learning, deep learning technology has been widely used in wireless communication physical layer research, and has achieved great improvements in performance. Research directions include the complete replacement of the entire communication system by an end-to-end deep neural network, or only part of the communication system modules, such as encoders, decoders, and detectors, are replaced by deep neural networks.

OFDM接收方法作为4G和5G的关键技术之一，在接收机的信道估计和信道均衡方向均有大量的研究成果。但是，目前采用的方法只是将整个神经网络当作一个完全的黑盒系统，简单替换传统通信的不同模块来取得性能的提升。网络依赖大量仿真或空口采集数据离线训练完成，训练数据的质量严重影响网络实际部署后的可靠性和鲁棒性。并且网络实际运行之后，参数不在变化，无法在不同的环境下均取得较优性能。As one of the key technologies of 4G and 5G, the OFDM reception method has a lot of research results in the channel estimation and channel equalization of the receiver. However, the current method only treats the entire neural network as a complete black-box system, and simply replaces different modules of traditional communication to achieve performance improvement. The network relies on a large amount of simulation or air interface acquisition data to complete offline training, and the quality of the training data seriously affects the reliability and robustness of the network after actual deployment. And after the actual operation of the network, the parameters do not change, and it is impossible to achieve better performance in different environments.

发明内容SUMMARY OF THE INVENTION

本发明所要解决的技术问题在于克服现有技术的不足，提供一种基于神经网络的信道环境自适应OFDM接收方法，解决现有技术存在的网络在实际运行时参数固定，不能根据运行环境自适应地提升性能的问题。The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art, and to provide a channel environment adaptive OFDM receiving method based on a neural network, so as to solve the problem that the existing network in the prior art has fixed parameters during actual operation and cannot be adapted according to the operating environment. to improve performance.

本发明具体采用以下技术方案解决上述技术问题：The present invention specifically adopts the following technical solutions to solve the above-mentioned technical problems:

一种基于神经网络的信道环境自适应OFDM接收方法，包括：A neural network-based channel environment adaptive OFDM receiving method, comprising:

步骤1、在信道估计模块中构建由主维纳滤波深度神经网络和若干个并联的副维纳滤波深度神经网络构成的信道滤波网络，及在信号检测模块中构建由主均衡深度神经网络和若干个并联的副均衡深度神经网络构成的信道均衡网络；Step 1. In the channel estimation module, construct a channel filtering network composed of a main Wiener filtering deep neural network and several parallel auxiliary Wiener filtering deep neural networks, and in the signal detection module, construct a main equalizing deep neural network and several parallel Wiener filtering deep neural networks. A channel equalization network composed of two parallel sub-equalization deep neural networks;

选择若干种不同的信道环境，将其中一种信道环境作为主训练环境及其余信道环境作为副训练环境；利用主训练环境分别训练和确定主维纳滤波深度神经网络和主均衡深度神经网络中参数，以得到训练后的主维纳滤波深度神经网络和主均衡深度神经网络；及利用不同的副训练环境分别训练和确定各副维纳滤波深度神经网络、副均衡深度神经网络中参数，得到训练后的各副维纳滤波深度神经网络和各副均衡深度神经网络；Select several different channel environments, and use one of the channel environments as the main training environment and the other channel environments as the sub-training environments; use the main training environment to train and determine the parameters in the main Wiener filtering deep neural network and the main balanced deep neural network respectively , in order to obtain the main Wiener filter deep neural network and the main balanced deep neural network after training; and use different auxiliary training environments to train and determine the parameters of each auxiliary Wiener filter deep neural network and auxiliary balanced deep neural network respectively, and get the training After each sub-Wiener filter deep neural network and each sub-balanced deep neural network;

步骤2、对于信道估计模块，输入接收频域导频和本地频域导频，利用最小二乘法估计得到带噪信道估计结果

Figure BDA0001811801890000021
将带噪信道估计结果
Figure BDA0001811801890000022
输入训练后的主维纳滤波深度神经网络中并输出主信道估计结果
Figure BDA0001811801890000023
及将主信道估计结果
Figure BDA0001811801890000024
和副维纳滤波深度神经网络得到各副信道估计结果并各自乘以权重系数α后相加，得到最终信道估计结果
Figure BDA0001811801890000025
Step 2. For the channel estimation module, input the received frequency domain pilot and the local frequency domain pilot, and use the least squares method to estimate the noisy channel estimation result
Figure BDA0001811801890000021
The noisy channel estimation results
Figure BDA0001811801890000022
Input the trained main Wiener filter deep neural network and output the main channel estimation result
Figure BDA0001811801890000023
and the main channel estimation result
Figure BDA0001811801890000024
and the sub-Wiener filter deep neural network to obtain the estimation results of each sub-channel and multiply each by the weight coefficient α and add them to obtain the final channel estimation result
Figure BDA0001811801890000025

步骤3、对于信号检测模块，输入信道估计模块所得最终信道估计结果

Figure BDA0001811801890000026
和接收的频域数据Y，利用迫零均衡方法估计得到迫零均衡结果
Figure BDA0001811801890000027
将迫零均衡结果
Figure BDA0001811801890000028
输入训练后的主均衡深度神经网络的结果和各训练后的副均衡深度神经网络乘以权重系数β的结果相加，经过硬判决后输出得到估计的比特流
Figure BDA0001811801890000029
Step 3. For the signal detection module, input the final channel estimation result obtained by the channel estimation module
Figure BDA0001811801890000026
and the received frequency domain data Y, use the zero-forcing equalization method to estimate the zero-forcing equalization result
Figure BDA0001811801890000027
zero-forcing equilibrium result
Figure BDA0001811801890000028
The result of inputting the main balanced deep neural network after training and the result of multiplying the training sub-balanced deep neural network by the weight coefficient β are added, and the estimated bit stream is output after hard decision.
Figure BDA0001811801890000029

步骤4、实时采集实际传输过程中已知的频域导频和频域数据，对所述权重系数α和β动态调整。Step 4: Collect in real time the known frequency domain pilots and frequency domain data in the actual transmission process, and dynamically adjust the weight coefficients α and β.

进一步地，作为本发明的一种优选技术方案，所述步骤2中得到最终信道估计结果

Figure BDA00018118018900000210
为：
Figure BDA00018118018900000211
Further, as a preferred technical solution of the present invention, the final channel estimation result is obtained in the step 2
Figure BDA00018118018900000210
for:
Figure BDA00018118018900000211

其中，

Figure BDA00018118018900000212
为各副信道估计结果；α1、α2…αn为各副信道估计结果的权重系数。in,
Figure BDA00018118018900000212
are the estimation results of each sub-channel; α 1 , α 2 . . . α n are the weight coefficients of the estimation results of each sub-channel.

进一步地，作为本发明的一种优选技术方案，所述步骤1中信道滤波网络的损失函数为：Further, as a preferred technical solution of the present invention, the loss function of the channel filtering network in the step 1 is:

Figure BDA00018118018900000213
Figure BDA00018118018900000213

其中，N是子载波数，H(k)是实际的第k个子载波频域信道，

Figure BDA00018118018900000214
是第k个子载波信道估计结果。where N is the number of subcarriers, H(k) is the actual kth subcarrier frequency domain channel,
Figure BDA00018118018900000214
is the channel estimation result of the kth subcarrier.

进一步地，作为本发明的一种优选技术方案，所述步骤1中信道均衡网络的损失函数为：Further, as a preferred technical solution of the present invention, the loss function of the channel equalization network in the step 1 is:

Figure BDA00018118018900000215
Figure BDA00018118018900000215

其中，B是所要估计的比特数，b(i)是实际发送的第i个比特，

Figure BDA0001811801890000031
是实际发送的第i个比特估计。where B is the number of bits to be estimated, b(i) is the ith bit actually sent,
Figure BDA0001811801890000031
is the ith bit estimate actually sent.

本发明采用上述技术方案，能产生如下技术效果：The present invention adopts the above-mentioned technical scheme, and can produce the following technical effects:

本发明的方法结合传统通信的信道知识，使网络对于特定信道环境进一步提升性能，发挥深度神经网络对于复杂信道环境的学习优化功能。相比于传统的OFDM接收机，以及大量不同环境训练出来的深度学习接收机，均取得更好的效果。The method of the invention combines the channel knowledge of traditional communication to further improve the performance of the network for a specific channel environment, and exerts the learning and optimization function of the deep neural network for the complex channel environment. Compared with traditional OFDM receivers and deep learning receivers trained in a large number of different environments, they have achieved better results.

因此，本发明通过将离线训练和在线学习进行结合，使神经网络针对特定信道环境提升性能的同时，能够在线学习切换，获得良好的鲁棒性。相比于传统接收方法和只做离线训练的接收方法，在提升性能的同时，又保证了系统不会因对特定环境的过度优化，而在不同环境下出现性能恶化的情况。这离线和在线结合的方法，更加适合于系统的实际部署要求。Therefore, by combining offline training and online learning, the present invention enables the neural network to learn and switch online while improving performance for a specific channel environment, thereby obtaining good robustness. Compared with the traditional receiving method and the receiving method that only performs offline training, while improving the performance, it also ensures that the system will not deteriorate in performance in different environments due to over-optimization of specific environments. This method of combining offline and online is more suitable for the actual deployment requirements of the system.

附图说明Description of drawings

图1为本发明基于神经网络的信道环境自适应OFDM接收方法的原理示意图。FIG. 1 is a schematic diagram of the principle of a neural network-based channel environment adaptive OFDM receiving method according to the present invention.

图2为本发明中主维纳滤波深度神经网络结构框图。FIG. 2 is a structural block diagram of the main Wiener filter deep neural network in the present invention.

图3为本发明中导频与数据格式示意图。FIG. 3 is a schematic diagram of pilot and data formats in the present invention.

具体实施方式Detailed ways

下面结合说明书附图对本发明的实施方式进行描述。Embodiments of the present invention will be described below with reference to the accompanying drawings.

如图1所示，本发明设计了一种基于神经网络的信道环境自适应OFDM接收方法，包括：As shown in FIG. 1, the present invention designs a channel environment adaptive OFDM receiving method based on neural network, including:

步骤1、在信道估计模块中构建由主维纳滤波深度神经网络和若干个并联的副维纳滤波深度神经网络构成的信道滤波网络，及在信号检测模块中构建由主均衡深度神经网络和若干个并联的副均衡深度神经网络构成的信道均衡网络；Step 1. In the channel estimation module, construct a channel filtering network composed of a main Wiener filtering deep neural network and several parallel auxiliary Wiener filtering deep neural networks, and in the signal detection module, construct a main equalizing deep neural network and several parallel Wiener filtering deep neural networks. A channel equalization network composed of two parallel sub-equalization deep neural networks;

选择若干种不同的信道环境，将其中一种信道环境作为主训练环境及其余信道环境作为副训练环境；利用主训练环境分别对主维纳滤波深度神经网络和主均衡深度神经网络中参数训练和确定参数，以得到训练后的主维纳滤波深度神经网络和主均衡深度神经网络；及利用不同的副训练环境分别对各副维纳滤波深度神经网络、副均衡深度神经网络中参数训练和确定参数，得到训练后的各副维纳滤波深度神经网络和各副均衡深度神经网络；Several different channel environments are selected, and one of the channel environments is used as the main training environment and the other channel environments are used as sub-training environments; the main training environment is used to train the parameters in the main Wiener filtering deep neural network and the main balanced deep neural network respectively. Determining parameters to obtain the main Wiener filtering deep neural network and the main equalizing deep neural network after training; and using different sub-training environments to train and determine the parameters in each sub-Wiener filtering deep neural network and sub-equalizing deep neural network respectively parameters to obtain the trained sub-Wiener filter deep neural network and each sub-balanced deep neural network;

具体的，分别在主训练环境、副训练环境1、副训练环境2等不同信道环境下，训练信道滤波网络中对应的维纳滤波深度神经网络，如：主训练环境下，各副维纳滤波深度神经网络的权重系数α＝0，使得网络仅对主维纳滤波深度神经网络中的参数训练和确定具体参数；在环境x下，固定权重系数αx＝1,αi＝0(i≠x)，令该环境x下对应的一个副维纳滤波深度神经网络的权重系数为1，其余副维纳滤波深度神经网络的权重系数均为0，训练环境x下对应一个副维纳滤波深度神经网络的参数和确定具体参数，损失函数为最终信道估计结果

Figure BDA0001811801890000041
的均方误差；而在切换其他环境时，同样将对应环境下对应的一个副维纳滤波深度神经网络的权重系数为1，其余副维纳滤波深度神经网络的权重系数均为0，实现各副维纳滤波深度神经网络的训练切换。Specifically, in different channel environments such as the main training environment, the auxiliary training environment 1, and the auxiliary training environment 2, the corresponding Wiener filtering deep neural network in the channel filtering network is trained. For example, in the main training environment, each auxiliary Wiener filtering The weight coefficient α=0 of the deep neural network makes the network only train and determine the specific parameters for the parameters in the main Wiener filter deep neural network; under the environment x, the fixed weight coefficient α x = 1, α i = 0 (i≠ x), let the weight coefficient of a corresponding sub-Wiener filter deep neural network under the environment x be 1, and the weight coefficients of the other sub-Wiener filter deep neural networks are all 0, and the training environment x corresponds to a sub-Wiener filter depth The parameters of the neural network and the specific parameters are determined, and the loss function is the final channel estimation result
Figure BDA0001811801890000041
The mean square error of Training switching for para-Wiener filtering deep neural networks.

同样的，训练不同环境下的信道均衡网络中的均衡深度神经网络具体为：主训练环境下，主均衡深度神经网络的权重系数为1，各副均衡深度神经网络的权重系数β为0，使得网络仅对主均衡深度神经网络中的参数训练和确定具体参数；在环境x下，固定权重系数βx＝1,βi＝0(i≠x)，令该环境x下对应的一个副均衡深度神经网络的权重系数为1，其余副均衡深度神经网络的权重系数均为0，训练环境x下对应一个副均衡深度神经网络的参数和确定具体参数，损失函数为输出估计比特流

Figure BDA0001811801890000042
的均方误差函数。同理，在各种环境下，只训练对应的一个均衡深度神经网络，令其当前环境下的网络权重系数β为1，其余网络的权重系数β为0，优化比特估计性能。Similarly, training the balanced deep neural network in the channel balanced network in different environments is as follows: in the main training environment, the weight coefficient of the main balanced deep neural network is 1, and the weight coefficient β of each secondary balanced deep neural network is 0, so that The network only trains and determines specific parameters for the parameters in the main equalization deep neural network; in the environment x, the fixed weight coefficient β x =1, β i =0 (i≠x), so that a corresponding sub-equalizer under the environment x The weight coefficient of the deep neural network is 1, and the weight coefficients of the other sub-balanced deep neural networks are 0. The training environment x corresponds to the parameters of a sub-balanced deep neural network and determines the specific parameters, and the loss function is the output estimated bit stream.
Figure BDA0001811801890000042
The mean squared error function of . Similarly, in various environments, only one corresponding balanced deep neural network is trained, so that the network weight coefficient β in the current environment is 1, and the weight coefficient β of the other networks is 0, so as to optimize the bit estimation performance.

本实施例中，主运行信道环境选择802.11b指数信道模型环境，副训练环境1选择SUI5室外信道模型，无其它可能运行的信道环境。因而，只需要在α1＝0时，训练信道滤波网络中的主滤波网络，α1＝1时，训练信道滤波网络中的环境1滤波网络。In this embodiment, the 802.11b index channel model environment is selected for the main operating channel environment, the SUI5 outdoor channel model is selected for the sub-training environment 1, and there are no other possible operating channel environments. Therefore, it is only necessary to train the main filter network in the channel filter network when α 1 =0, and train the environment 1 filter network in the channel filter network when α 1 =1.

步骤2、对于信道估计模块，输入接收频域导频和本地频域导频，利用最小二乘法估计得到带噪信道估计结果

Figure BDA0001811801890000043
将带噪信道估计结果
Figure BDA0001811801890000044
的实部虚部串联输入训练后的主维纳滤波深度神经网络中并输出主信道估计结果
Figure BDA0001811801890000045
及将主信道估计结果
Figure BDA0001811801890000046
分别输入训练后的各副维纳滤波深度神经网络得到各副信道估计结果
Figure BDA0001811801890000047
并各自乘以权重系数α后相加，得到最终信道估计结果
Figure BDA0001811801890000051
为
Figure BDA0001811801890000052
其中a1、a2…an为各副信道估计结果的权重系数，可以直接选取数值作为各权重系数。Step 2. For the channel estimation module, input the received frequency domain pilot and the local frequency domain pilot, and use the least squares method to estimate the noisy channel estimation result
Figure BDA0001811801890000043
The noisy channel estimation results
Figure BDA0001811801890000044
The real and imaginary parts are concatenated into the main Wiener filter deep neural network after training and output the main channel estimation result
Figure BDA0001811801890000045
and the main channel estimation result
Figure BDA0001811801890000046
Input the trained sub-Wiener filter deep neural network separately to obtain the estimation results of each sub-channel
Figure BDA0001811801890000047
They are multiplied by the weight coefficient α and added together to obtain the final channel estimation result.
Figure BDA0001811801890000051
for
Figure BDA0001811801890000052
Among them, a 1 , a 2 . . . an n are the weight coefficients of each sub-channel estimation result, and the numerical value can be directly selected as each weight coefficient.

信道估计中各神经网络采用一层无激活函数的全连接层，多个估计网络并联后，乘以相应的切换参数，网络结果表达为：In the channel estimation, each neural network adopts a fully connected layer without activation function. After multiple estimation networks are connected in parallel, they are multiplied by the corresponding switching parameters. The network result is expressed as:

Figure BDA0001811801890000053
Figure BDA0001811801890000053

公式中，

Figure BDA0001811801890000054
Wi为信道估计中的滤波神经网络的乘性参数。初始化时，W主为线性最小均方误差估计实值矩阵的WLMMSE，n1＝0，所有权重参数α＝0。formula,
Figure BDA0001811801890000054
Wi is the multiplicative parameter of the filtering neural network in the channel estimation. During initialization, W is mainly W LMMSE of the linear minimum mean square error estimation real-valued matrix, n 1 =0, and all weight parameters α=0.

该网络训练，在环境x下，只训练对应的乘性参数Wx，此时αx＝1，αi＝0(i≠x)，优化信道估计性能。以主环境和环境1的情况为例，将

Figure BDA0001811801890000055
输入环境1下的维纳滤波神经网络，分别得到估计结果按权重相加，得到最终信道估计结果
Figure BDA0001811801890000056
分别在主环境、环境1下，训练对应的维纳滤波网络：主环境下，固定副维纳滤波深度神经网络的权重系数α1＝0，训练主维纳滤波深度神经网络参数，损失函数为最终信道估计结果
Figure BDA0001811801890000057
的均方误差；环境1下，固定权重系数α1＝1，训练环境1对应的副维纳滤波深度神经网络参数，损失函数为最终信道估计结果
Figure BDA0001811801890000058
的均方误差。In the network training, under the environment x, only the corresponding multiplicative parameter W x is trained, at this time α x =1, α i =0 (i≠x), and the channel estimation performance is optimized. Taking the main environment and environment 1 as an example, set the
Figure BDA0001811801890000055
Input the Wiener filter neural network in environment 1, and add the estimated results separately according to the weights to obtain the final channel estimation result
Figure BDA0001811801890000056
In the main environment and environment 1, the corresponding Wiener filter network is trained: in the main environment, the weight coefficient α 1 =0 of the auxiliary Wiener filter deep neural network is fixed, and the main Wiener filter deep neural network parameters are trained, and the loss function is Final channel estimation result
Figure BDA0001811801890000057
The mean square error of
Figure BDA0001811801890000058
mean squared error.

具体的，损失函数如下：Specifically, the loss function is as follows:

Figure BDA0001811801890000059
Figure BDA0001811801890000059

其中，N是子载波数，H(k)是实际的第k个子载波频域信道，

Figure BDA00018118018900000510
是第k个子载波信道估计结果。where N is the number of subcarriers, H(k) is the actual kth subcarrier frequency domain channel,
Figure BDA00018118018900000510
is the channel estimation result of the kth subcarrier.

步骤3、对于信号检测模块，输入信道估计模块所得最终信道估计结果

Figure BDA00018118018900000511
和接收的频域数据Y，利用迫零均衡方法估计得到迫零均衡结果
Figure BDA00018118018900000512
将迫零均衡结果
Figure BDA00018118018900000513
输入训练后的主均衡深度神经网络和各训练的副均衡深度神经网络中得到各网络的输出结果并各自乘以权重系数β后相加，即β1、β2…βn为各副均衡深度神经网络的权重系数，可以直接选取数值作为各权重系数，相加后的结果经过硬判决后输出得到估计的比特流
Figure BDA00018118018900000514
Step 3. For the signal detection module, input the final channel estimation result obtained by the channel estimation module
Figure BDA00018118018900000511
and the received frequency domain data Y, use the zero-forcing equalization method to estimate the zero-forcing equalization result
Figure BDA00018118018900000512
zero-forcing equilibrium result
Figure BDA00018118018900000513
Input the main balanced deep neural network after training and each trained sub-balanced deep neural network to obtain the output results of each network and multiply them by the weight coefficient β and add them together, that is, β 1 , β 2 . . . β n is the depth of each sub-balanced The weight coefficients of the neural network can be directly selected as the weight coefficients, and the added result is output after hard decision to obtain the estimated bit stream
Figure BDA00018118018900000514

同样的，以主环境和环境1的情况为例，训练主环境下的主均衡深度神经网络时，固定副均衡深度神经网络的权重参数β1＝0，损失函数为输出估计比特流

Figure BDA0001811801890000061
的均方误差函数；训练环境1下的副均衡深度神经网络时，固定该副均衡深度神经网络的权重参数β1＝1，损失函数为输出估计比特流
Figure BDA0001811801890000062
的均方误差函数。Similarly, taking the main environment and environment 1 as examples, when training the main balanced deep neural network in the main environment, the weight parameter β 1 =0 of the sub-balanced deep neural network is fixed, and the loss function is the output estimated bit stream
Figure BDA0001811801890000061
The mean square error function of
Figure BDA0001811801890000062
The mean squared error function of .

具体的，损失函数如下：Specifically, the loss function is as follows:

Figure BDA0001811801890000063
Figure BDA0001811801890000063

其中，B是所要估计的比特数，b(i)是实际发送的第i个比特，

Figure BDA0001811801890000064
是实际发送的第i个比特估计。where B is the number of bits to be estimated, b(i) is the ith bit actually sent,
Figure BDA0001811801890000064
is the ith bit estimate actually sent.

步骤4、实时采集实际信道环境的已知频域导频或频域数据，对所述权重系数α和β动态调整。即部署到实际场景后，所有网络参数不变，而权重系数α,β的具体数值通过实时采集当前传输的OFDM符号中已知导频或训练序列数据，实现在线的动态调整，损失函数为输出估计比特流

Figure BDA0001811801890000065
的均方误差函数cost2。Step 4: Collect known frequency domain pilots or frequency domain data of the actual channel environment in real time, and dynamically adjust the weight coefficients α and β. That is, after being deployed in the actual scene, all network parameters remain unchanged, and the specific values of the weight coefficients α, β can be dynamically adjusted online by collecting the known pilot or training sequence data in the currently transmitted OFDM symbol in real time, and the loss function is the output. estimated bitstream
Figure BDA0001811801890000065
The mean squared error function cost2.

如图2所示，各环境下的维纳滤波深度神经网络均为一层的全连接网络，不采用激活函数，该层网络的乘性参数W主采用线性最小均方误差信道估计的权重矩阵的实值进行初始化，W主初始化为全零，加性参数n初始化为全零，带噪信道估计使用最小二乘法，即

Figure BDA0001811801890000066
并在室内环境将W主为设为可训练的，权重系数α1＝0，此时网络实现的计算为
Figure BDA0001811801890000067
因为最小均方误差信道估计
Figure BDA0001811801890000068
且在深度学习网络中只能进行实值运算，所以当该网络的输入为
Figure BDA0001811801890000069
的实部和虚部串联，网络的乘性参数W初始化为权重矩阵WLMMSE的实值矩阵
Figure BDA00018118018900000610
其中，Re{·}为取实部，Im{·}为取虚部；加性参数n初始化为全零时，网络输出的初始值即为
Figure BDA00018118018900000611
的实部和虚部串联。优化cost1，优化器是Adam优化器，训练时采用小批量梯度下降，每轮内采用50个批量，每个批量的大小为1000条样本，一共训练2000轮，采用动态调整的学习速率，初始学习速率0.001，每100轮降为1/5。As shown in Figure 2, the Wiener filtering deep neural network in each environment is a fully connected network of one layer, without using an activation function, and the multiplicative parameter W of this layer of network mainly adopts the weight matrix of linear minimum mean square error channel estimation Initialize the real value of , W is initialized to all zeros, the additive parameter n is initialized to all zeros, and the noisy channel estimation uses the least squares method, that is,
Figure BDA0001811801890000066
And in the indoor environment, the main W is set to be trainable, and the weight coefficient α 1 =0. At this time, the calculation implemented by the network is:
Figure BDA0001811801890000067
Because of the minimum mean square error channel estimation
Figure BDA0001811801890000068
And in the deep learning network, only real-valued operations can be performed, so when the input of the network is
Figure BDA0001811801890000069
The real and imaginary parts of the network are concatenated, and the multiplicative parameter W of the network is initialized as the real-valued matrix of the weight matrix W LMMSE
Figure BDA00018118018900000610
Among them, Re{·} is the real part, Im{·} is the imaginary part; when the additive parameter n is initialized to all zeros, the initial value of the network output is
Figure BDA00018118018900000611
The real and imaginary parts of . To optimize cost1, the optimizer is the Adam optimizer, using small batch gradient descent during training, using 50 batches in each round, each batch size is 1000 samples, a total of 2000 rounds of training, using a dynamically adjusted learning rate, initial learning Rate 0.001, reduced to 1/5 every 100 rounds.

均衡深度神经网络结构为128×120×16，优化cost2，优化器是Adam优化器，训练时采用小批量梯度下降，每轮内采用50个批量，每个批量的大小为1000条样本。一共训练20000轮，采用动态调整的学习速率，初始学习速率设置为0.001，每1000轮减小为当前的五分之一。The structure of the balanced deep neural network is 128×120×16, and the cost2 is optimized. The optimizer is the Adam optimizer. Mini-batch gradient descent is used during training. 50 batches are used in each round, and the size of each batch is 1000 samples. A total of 20,000 rounds of training were used, and a dynamically adjusted learning rate was used. The initial learning rate was set to 0.001, and every 1,000 rounds was reduced to one-fifth of the current one.

如图3所示，在含有128个子载波的OFDM系统中，数据帧格式为一个导频OFDM符号和一个数据OFDM符号，导频和数据都占其中64个子载波，星座调制方式为QPSK。As shown in Figure 3, in an OFDM system with 128 subcarriers, the data frame format is a pilot OFDM symbol and a data OFDM symbol, both pilot and data occupy 64 subcarriers, and the constellation modulation is QPSK.

故本发明通过将深度学习和通信知识进行结合，离线训练与在线训练结合，为OFDM系统接收机设计提供了一种新方法，在不同环境的鲁棒性上和稳定工作时的BER性能上，相较传统通信接收机和离线训练、在线部署的深度学习接收机，均有很大的提升。并且，具有参数少、工作效率高、在线切换灵活迅速等优点。Therefore, the present invention provides a new method for OFDM system receiver design by combining deep learning with communication knowledge, offline training and online training, in terms of robustness in different environments and BER performance during stable operation, Compared with traditional communication receivers and deep learning receivers that are trained offline and deployed online, they are all greatly improved. In addition, it has the advantages of less parameters, high work efficiency, and flexible and rapid online switching.

上面结合附图对本发明的实施方式作了详细说明，但是本发明并不限于上述实施方式，在本领域普通技术人员所具备的知识范围内，还可以在不脱离本发明宗旨的前提下做出各种变化。The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned embodiments, and can also be made within the scope of knowledge possessed by those of ordinary skill in the art without departing from the purpose of the present invention. Various changes.

Claims (3)

1. A channel environment self-adaptive OFDM receiving method based on a neural network is characterized by comprising the following steps:
step 1, constructing a channel filter network consisting of a main wiener filter deep neural network and a plurality of auxiliary wiener filter deep neural networks connected in parallel in a channel estimation module, and constructing a channel equalization network consisting of a main equalization deep neural network and a plurality of auxiliary equalization deep neural networks connected in parallel in a signal detection module;
selecting a plurality of different channel environments, wherein one channel environment is used as a main training environment and the other channel environments are used as auxiliary training environments; respectively training and determining parameters in a main wiener filtering deep neural network and a main equilibrium deep neural network by using a main training environment so as to obtain a trained main wiener filtering deep neural network and a trained main equilibrium deep neural network; respectively training and determining parameters in each auxiliary wiener filtering deep neural network and each auxiliary equilibrium deep neural network by using different auxiliary training environments to obtain each trained auxiliary wiener filtering deep neural network and each auxiliary equilibrium deep neural network;
step 2, inputting the receiving frequency domain pilot frequency and the local frequency domain pilot frequency for the channel estimation module, and estimating by using a least square method to obtain a noisy channel estimation result
Figure FDA0002721529180000011
Estimating the result of the channel with noise
Figure FDA0002721529180000012
Inputting the trained main wiener filtering deep neural network and outputting a main channel estimation result
Figure FDA0002721529180000013
And estimating the main channel
Figure FDA0002721529180000014
Respectively input eachThe trained sub wiener filtering deep neural network obtains each sub channel estimation result, and each sub channel estimation result is multiplied by a weight coefficient alpha and then is compared with the main channel estimation result
Figure FDA0002721529180000015
Adding to obtain the final channel estimation result
Figure FDA0002721529180000016
The method specifically comprises the following steps:
Figure FDA0002721529180000017
wherein,
Figure FDA0002721529180000018
estimating the result for each sub-channel; alpha is alpha1、α2…αnWeight coefficients for each subchannel estimation result;
step 3, inputting the final channel estimation result obtained by the channel estimation module to the signal detection module
Figure FDA0002721529180000019
And receiving the frequency domain data Y, and estimating by using a zero-forcing equalization method to obtain a zero-forcing equalization result
Figure FDA00027215291800000110
Zero-forcing equalization result
Figure FDA00027215291800000111
Adding the result of the main equalization deep neural network after input training and the result of multiplying the sub equalization deep neural network after each training by the weight coefficient beta, and outputting after hard decision to obtain an estimated bit stream
Figure FDA00027215291800000112
And 4, acquiring known frequency domain pilot frequency and frequency domain data in the actual transmission process in real time, and dynamically adjusting the weight coefficients alpha and beta.
2. The adaptive OFDM receiving method according to claim 1, wherein the loss function of the channel filtering network in step 1 is:
Figure FDA00027215291800000113
where N is the number of subcarriers, H (k) is the actual kth subcarrier frequency domain channel,
Figure FDA0002721529180000021
is the k-th subcarrier channel estimation result.
3. The neural network-based channel environment adaptive OFDM receiving method of claim 1, wherein: the loss function of the channel equalization network in the step 1 is as follows:
Figure FDA0002721529180000022
where B is the number of bits to be estimated, B (i) is the ith bit actually transmitted,
Figure FDA0002721529180000023
is the ith bit estimate that is actually sent.
CN201811123542.2A 2018-09-26 2018-09-26 A Neural Network-based Channel Environment Adaptive OFDM Reception Method Active CN109194595B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811123542.2A CN109194595B (en) 2018-09-26 2018-09-26 A Neural Network-based Channel Environment Adaptive OFDM Reception Method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811123542.2A CN109194595B (en) 2018-09-26 2018-09-26 A Neural Network-based Channel Environment Adaptive OFDM Reception Method

Publications (2)

Publication Number Publication Date
CN109194595A CN109194595A (en) 2019-01-11
CN109194595B true CN109194595B (en) 2020-12-01

Family

ID=64907219

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811123542.2A Active CN109194595B (en) 2018-09-26 2018-09-26 A Neural Network-based Channel Environment Adaptive OFDM Reception Method

Country Status (1)

Country Link
CN (1) CN109194595B (en)

Families Citing this family (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109889462B (en) * 2019-01-21 2021-05-28 东南大学 A neural network-assisted iterative reception method for high-speed visible light communication
CN109831400A (en) * 2019-03-06 2019-05-31 深兰科技（上海）有限公司 A kind of received method and apparatus of OFDM signal
EP3726529A1 (en) * 2019-04-16 2020-10-21 Fraunhofer Gesellschaft zur Förderung der Angewand Method and apparatus for determining a deep filter
CN111431831B (en) * 2020-03-20 2021-06-25 厦门大学 A method and system for adaptive modulation based on multi-dimensional OFDM environment
CN111641570B (en) * 2020-04-17 2021-04-27 浙江大学 A joint device detection and channel estimation method based on deep learning
WO2021262052A1 (en) * 2020-06-25 2021-12-30 Telefonaktiebolaget Lm Ericsson (Publ) A context aware data receiver for communication signals based on machine learning
CN111865863B (en) * 2020-07-20 2021-07-20 山东大学 An OFDM signal detection method based on RNN neural network
CN111884976B (en) * 2020-07-21 2021-09-03 西安电子科技大学 Channel interpolation method based on neural network
CN114465858A (en) * 2020-11-09 2022-05-10 索尼公司 Electronic device, signal processing method, and computer-readable storage medium
CN116455706B (en) * 2023-04-12 2024-05-31 华南理工大学 A design method and system for channel estimation and equalization

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1731779A (en) * 2005-08-05 2006-02-08 西安电子科技大学 Wavelet packet multi-carrier spread spectrum system and control method based on neural network equalizer
CN101867533A (en) * 2010-05-27 2010-10-20 东南大学 Space-division multiple access multi-antenna transmission downlink pilot and channel estimation method
CN104022979A (en) * 2014-06-23 2014-09-03 东南大学 Joint sparse channel estimation method, device and system
CN104022978A (en) * 2014-06-18 2014-09-03 中国联合网络通信集团有限公司 Half-blindness channel estimating method and system
CN108462517A (en) * 2018-03-06 2018-08-28 东南大学 A kind of MIMO link self-adaption transmission methods based on machine learning

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1731779A (en) * 2005-08-05 2006-02-08 西安电子科技大学 Wavelet packet multi-carrier spread spectrum system and control method based on neural network equalizer
CN101867533A (en) * 2010-05-27 2010-10-20 东南大学 Space-division multiple access multi-antenna transmission downlink pilot and channel estimation method
CN104022978A (en) * 2014-06-18 2014-09-03 中国联合网络通信集团有限公司 Half-blindness channel estimating method and system
CN104022979A (en) * 2014-06-23 2014-09-03 东南大学 Joint sparse channel estimation method, device and system
CN108462517A (en) * 2018-03-06 2018-08-28 东南大学 A kind of MIMO link self-adaption transmission methods based on machine learning

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
Implementation of neural network for channel estimation in OFDM network;Ravinder Kaur等;《2016 2nd International Conference on Contemporary Computing and Informatics (IC3I)》;20170504;第437-441页 *
MIMO-OFDM系统信道估计及容量分析技术研究;李喆;《中国优秀硕士学位论文全文数据库 信息科技辑》;20150415;第I136-329页 *
Power of Deep Learning for Channel Estimation and Signal Detection in OFDM Systems;Hao Ye等;《IEEE Wireless Communications Letters》;20180228;第114-117页 *
Timing requirements and scenarios for 5G NR Hybrid ARQ;Nokia等;《3GPP TSG-RAN WG1#86 R1-167299 Gothenburg,Sweden》;20160826;全文 *
基于信号互相关性的MIMO系统盲均衡算法研究;王珊珊;《中国优秀硕士学位论文全文数据库 信息科技辑》;20150715;第I136-203页 *

Also Published As

Publication number Publication date
CN109194595A (en) 2019-01-11

Similar Documents

Publication Publication Date Title
CN109194595B (en) A Neural Network-based Channel Environment Adaptive OFDM Reception Method
CN109617847B (en) OFDM receiving method without cyclic prefix based on model-driven deep learning
Gao et al. ComNet: Combination of deep learning and expert knowledge in OFDM receivers
KR102294156B1 (en) Signal detection method by sc-fde system based on self-adaptive integrated deep learning model
CN101478510B (en) Adaptive equalizer and receiver system using the equalizer
Tang et al. Space-time interference cancellation in MIMO-OFDM systems
CN101499840B (en) Iteration detection method for MIMO system
CN106533590B (en) Uplink channel quality measurement method based on receiving end EVM
CN109818891B (en) Lattice reduction assisted low-complexity greedy sphere decoding detection method
Drakshayini et al. A review of wireless channel estimation techniques: challenges and solutions
CN108156101A (en) A kind of system combined iterative channel estimation of MIMO-SCFDE and iteration equalizing method
Ivanov et al. Smart sorting in massive MIMO detection
Aziz et al. Deep recurrent neural network based detector for OFDM with index modulation
CN111865844B (en) Channel estimation method and device for massive MIMO full-duplex relay system
CN104009824B (en) Pilot aided data fusion method based on differential evolution in a kind of base station collaboration up-line system
CN101958862B (en) Self-adaptive decision feedback equalizer based on superposing structure
Ye et al. Bilinear convolutional auto-encoder based pilot-free end-to-end communication systems
CN106453166A (en) Large-scale MIMO channel estimation method and system
CN107026804A (en) Channel estimation methods based on exponential smoothing in MIMO ofdm systems
CN114614863A (en) MIMO detector selection
CN107836100B (en) Method and apparatus for low-complexity ISI estimation using sparse discontinuous time-domain pilots
CN103297201A (en) Self-adaptive modulation coding selection method under cell interference
CN106789776B (en) High-precision channel estimation method of MIMO-OFDM system
CN103905347B (en) OFDM system fast-change channel estimation method based on double weighing
CN106375065A (en) Iterative MMSE soft detection method based on spatial modulation system with frequency offset

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant