CN114650199A - Deep neural network channel estimation method and system based on data driving - Google Patents

Deep neural network channel estimation method and system based on data driving Download PDF

Info

Publication number
CN114650199A
CN114650199A CN202111637126.6A CN202111637126A CN114650199A CN 114650199 A CN114650199 A CN 114650199A CN 202111637126 A CN202111637126 A CN 202111637126A CN 114650199 A CN114650199 A CN 114650199A
Authority
CN
China
Prior art keywords
data
neural network
deep neural
channel
matrix
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.)
Pending
Application number
CN202111637126.6A
Other languages
Chinese (zh)
Inventor
施毅
孙浩
熊云彩
周唯
沈连丰
燕锋
夏玮玮
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.)
Nanjing Rongzhi Information Innovation Research Institute Co ltd
Original Assignee
Nanjing Rongzhi Information Innovation Research Institute Co ltd
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 Nanjing Rongzhi Information Innovation Research Institute Co ltd filed Critical Nanjing Rongzhi Information Innovation Research Institute Co ltd
Priority to CN202111637126.6A priority Critical patent/CN114650199A/en
Publication of CN114650199A publication Critical patent/CN114650199A/en
Pending legal-status Critical Current

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
    • H04L25/024—Channel estimation channel estimation algorithms
    • H04L25/0254—Channel estimation channel estimation algorithms using neural network algorithms
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L1/00—Arrangements for detecting or preventing errors in the information received
    • H04L1/02—Arrangements for detecting or preventing errors in the information received by diversity reception
    • H04L1/06—Arrangements for detecting or preventing errors in the information received by diversity reception using space diversity
    • H04L1/0618—Space-time coding
    • 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
    • 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
    • H04L25/022—Channel estimation of frequency response
    • 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
    • H04L25/0224—Channel estimation using sounding signals
    • 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
    • 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
    • H04L25/03165—Arrangements for removing intersymbol interference using neural networks

Landscapes

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

Abstract

The invention adopts a data-driven mode to iteratively train a network model, designs a data-driven channel estimation communication system suitable for a deep neural network, adopts various data for data driving, improves the training times of a perceptron, simulates a channel environment and continuously fits real channel distribution, continuously optimizes an extreme value through a forward and reverse algorithm, can acquire channel state information by utilizing data training, and improves the communication quality.

Description

Deep neural network channel estimation method and system based on data driving
Technical Field
The invention relates to the technical field of communication, in particular to a deep neural network channel estimation method and system based on data driving.
Background
With the development of communication and signal fields and the continuous improvement of communication systems, the wireless communication technology can realize low-delay mutual communication through the propagation of radio signals at any time, but the traditional communication technology still has serious intersymbol interference, low spectrum utilization rate, multipath effect and other influences, which all include the design of a plurality of system bottom layer reasons and the corresponding selection of frequency, bandwidth resource allocation and channel model technology. In the twenty-first century, with the introduction of discrete fourier transform (OFDM) and inverse discrete fourier transform (ifft) into Orthogonal Frequency Division Multiplexing (OFDM) and Multiple Input Multiple Output (MIMO) systems, the problem of improving the spectrum utilization and reducing inter-symbol interference can be truly achieved, and the transmission capacity of a communication system can be improved by transmission and reception of Multiple antennas. The introduction of the MIMO-OFDM system improves the system capacity and the spectrum efficiency to a certain extent, but the performance of the communication system is an important index for measuring the development progress of the communication system, and solving the signal distortion and recovering the check received signal in the wireless communication system is an important field needing to be researched at present.
In the development of communication technology, in order to solve the problems that signals at a receiving end cannot be recovered, signal distortion, received signal precision improvement and the like, the technical field of channel estimation is facing the era of rapid development. Whether the channel estimation accuracy is high or not can determine the quality of the whole MIMO-OFDM communication system to a great extent, the application fields can be millimeter wave large-scale MIMO system and 5G communication MIMO communication system, and the channel state information is realized by corresponding channel estimation algorithm. At present, the conventional channel estimation algorithm mainly includes a data-aided channel estimation algorithm, which inserts a certain number of pilot channels and regular pilot channels into a transmission channel, and then estimates channel characteristics through pilot signal information by using a corresponding algorithm, but this requires other data to help estimate the channel and is easy to generate pilot pollution, resulting in a large system overhead, and other problems; the blind channel estimation algorithm is used for counting channel characteristics through a large number of data sets, pilot signal overhead can not be generated, but the blind channel estimation algorithm has the problems of too much data quantity requirement, low efficiency and the like; a semi-blind channel estimation algorithm, which mainly includes a Least Square (LS) algorithm and a Minimum Mean Square Error (MMSE) by a channel estimation method combining blind channel estimation and data-aided channel estimation, but the performance of the algorithm still cannot achieve a good effect, noise interference exists in an actual channel, and a serious inter-symbol interference problem exists when channel state information is estimated, so a new channel estimation method and system are needed to solve the problem at present.
With the gradual maturity of deep learning technology and the specific practical application of advanced neural network technology in computer vision, speech recognition and other directions in recent years, the achievements of deep learning in channel estimation, signal detection, feedback of channel state information and other aspects verify that the technology has more obvious advantages compared with the field of traditional channel estimation algorithm, has strong learning capability to extract features, analyze a system model and have flexible network structure, and the features can effectively estimate channel state information and improve estimation precision, compared with the traditional channel estimation method, the channel estimation algorithm based on the deep neural network can test training data by using a data model to achieve the purpose of estimating the channel state information, wherein the deep neural network based on data driving can firstly monitor and fit the channel characteristics through the input of a data set, the learning processes signal distortion in a supervision mode, continuously iterates and updates data to prevent accuracy loss to the maximum extent, and then the obtained channel characteristics can be directly used in actual channel estimation, so that the actual characteristics can be better met, more accurate channel state information is output, and the performance of a communication system is improved.
Disclosure of Invention
The invention provides a deep neural network channel estimation method and system based on data driving, aiming at the problem that the performance, system accuracy and frequency spectrum utilization rate of a channel estimation algorithm are seriously insufficient. The antenna array technology is realized by utilizing a space-time coding technology, so as to solve the influence of multipath effect in a communication system and establish a multi-input multi-output orthogonal frequency division multiplexing system based on channel frequency domain response vectors, simulation data and input, the system is used as a transmission channel among signal streams, a deep neural network training iterative network model based on data driving is adopted, channel characteristics are learned, a propagation algorithm and a rejection strategy are utilized to optimize a minimum extreme value and check and recover received data, a product of a [0, 1] vector value and a neuron output value is randomly generated through a Bernoulli function to determine the activation state of a neuron, then the network model carries out coefficient matrix and bias vector calculation on an antenna receiving end channel parameter estimation module of a channel estimation system, an estimated optimal value in an error threshold range is updated after iterative training, and finally, according to an output channel matrix, compensating for channel characteristics. The invention adopts a data-driven mode to iteratively train a network model, designs a data-driven channel estimation communication system suitable for a deep neural network, adopts various data for data driving, improves the training times of a perceptron, simulates a channel environment and continuously fits real channel distribution, continuously optimizes an extreme value through a forward and reverse algorithm, can acquire channel state information by utilizing data training, and improves the communication quality.
In order to effectively illustrate the above object of the present invention, the present invention provides a deep neural network channel estimation method and system based on data driving, which is characterized by comprising the following steps:
(1) training test data for obtaining channel response data and simulation data set
(2) Creating data-driven based deep neural network communication system
(3) Propagation algorithm iterative optimization threshold range optimization extremum
(4) Refusing strategy to optimize neural network activation state and raising iteration efficiency
The deep neural network channel estimation method and system based on data driving are realized by the following technical scheme in the steps of (1) and (2):
in a MIMO OFDM system, multi-antenna transmission and multi-antenna reception are achieved by using antenna array technologyThe method comprises the steps that a driven deep neural network carries out channel estimation to obtain channel state information, the source of training data comprises compiling DeepMIMO simulation data set testing data, sampling collection of data is carried out in a limited testing training space through a deep neural learning framework, and a data set Ddata＝[h1，h2，...，hn-1，hn]Where H ∈ HT*L，HT*LIn order to simulate a training data matrix, H '═ C (H x F + theta x), H' is H matrix, real part and imaginary part are extracted and then convolution operation is carried out, theta, x are data set input parameters, F is convolution vector, and the training collection process of the data set is divided into the following steps: firstly, randomly dividing original data into m data sets which are not equal to each other, inputting the m data sets into parameters, randomly extracting n data subsets from the m data sets, putting the n data subsets into a training frame, and then carrying out training iteration on the divided subsets for n times by different methods and then carrying out mean value calculation to obtain the data; subsequently saving the obtained data to Ddata(ii) a Channel response data is obtained by using minimum mean square error algorithm, wherein MMSE objective function is
Figure RE-GDA0003635672930000031
FMMSE(θ)＝min(F(θ)，L{M})＝min L{(M-θHx)(M-θHx)H}
Figure RE-GDA0003635672930000032
Where θ is the matrix to be solved by the algorithm:
Figure RE-GDA0003635672930000033
Figure RE-GDA0003635672930000034
calculating theta', F by taking its derivativeMMSEThe value of (theta) can be found by setting its first derivative to 0MMSE(θ)：
Figure RE-GDA0003635672930000035
Figure RE-GDA0003635672930000041
Wherein KMNFor values at the inserted pilot, KMN＝KHHIn which K isHHIs calculated for the autocorrelation matrix
Figure RE-GDA0003635672930000042
Figure RE-GDA0003635672930000043
The simulation data set and the channel response data are obtained through the technology, and the two parts of data are input into the deep neural network together to carry out the next iterative training and learn the channel characteristics, so that the channel state information is better estimated;
the deep neural network channel estimation method and system based on data driving, wherein the (3) comprises the following technical scheme:
after the data samples and the data set are used as a deep neural network supervision and prediction channel model, fitting nonlinear data and outputting the fitted data as the input of the neuron of the next layer through continuously learning channel characteristics and parameters of a hidden layer, and continuously and repeatedly learning to obtain a coefficient matrix tauiAnd the offset vector omega, and then outputting a final result. Wherein the data sample is a channel response matrix obtained by a minimum mean square error algorithm
Figure RE-GDA0003635672930000044
And the data set is DdataThe two parts are input into a deep neural network as a whole data, and fit a nonlinear relation through weighted summation of neurons in a hidden layer and an activation function between the output of a lower-layer node and the input of an upper-layer node, wherein the step is a forward propagation algorithmProcess, for the (n, m) th neuron output of the hidden layer:
Figure RE-GDA0003635672930000045
in the formula tauiIs a coefficient matrix, ω is an offset vector, i
Figure RE-GDA0003635672930000046
Outputting a data value for the (n, m) th neuron,
Figure RE-GDA0003635672930000047
as an activation function:
Figure RE-GDA0003635672930000048
mapping the characteristics of the neuron by an activation function, and inputting a data output end to an input end of another layer so as to strengthen the mapping capability of a nonlinear model of each layer of the neuron, a sparse network and solve the problem of gradient disappearance;
coefficient matrix tau obtained by front line propagation algorithmiThe sum offset vector ω is passed through a loss function to calculate the corresponding difference value for τiThe sum omega is close to the actual output of the sample by the minimum difference value, the minimum extreme value is optimized by using the loss function, and then the result of optimizing the final extreme value by using the back propagation algorithm is realized by utilizing the multiple iterations of the gradient descent algorithm; wherein the corresponding loss function:
Figure RE-GDA0003635672930000051
where r is the actual output of the neuron, a is the desired output, τiCoefficient matrix, ω bias vector, first order bias derivative is taken for coefficient matrix and bias vector:
Figure RE-GDA0003635672930000052
Figure RE-GDA0003635672930000053
will iterate τiCoefficient matrix, ω bias vector solution:
τi＝-η(δ(τix+ω)-a)δ′(τix+ω)·x-τ
ω＝-ηδ′(τix+ω)(δ(τix+ω)-a)+ω
will tauiAfter the coefficient matrix and the omega bias vector are obtained, the output value theta of the ith neuron of the output layer is solved through the valuel：
Figure RE-GDA0003635672930000054
In the formula riIs a target value, ziThe hidden layer output, x is the input vector,
Figure RE-GDA0003635672930000055
vector matrix, with w +1 layers
Figure RE-GDA0003635672930000056
As a vector:
Figure RE-GDA0003635672930000061
updating thetal：
Figure RE-GDA0003635672930000062
Figure RE-GDA0003635672930000063
Where E is the sum of squares error and yi，riRespectively, the output value and the target expected value, and finally obtaining the ith output of the w layer of the output layer:
Figure RE-GDA0003635672930000064
the deep neural network channel estimation method and system based on data driving, wherein the implementation technical scheme of the (4) is as follows: after the data set is input into a front-line propagation algorithm of the deep neural network for training and testing, the obtained result is subjected to error calculation and then input into the neural network again for inverse propagation iteration to update a coefficient matrix and a bias vector, and in order to prevent over-fitting, a rejection strategy (Dropout) is used for solving the problem. Generation of random probability vectors p using Bernoulli functionsiValue p ofi∈[0，1]Introduction of loss rate Ki，Ki＝pi(1-pi) The Dropout strategy has a regularization effect, with the expectation that individual neurons are deleted being:
Figure RE-GDA0003635672930000065
in the formula tauiIs a matrix of coefficients, xiFor inputting vectors, the number of activated neurons in deep neural network training is solved through loss rate parameters, and the problems of overlarge resource consumption and overfitting of a network system are solved.
The invention has the following beneficial effects:
the learning characteristic of the deep neural network is applied to a channel estimation algorithm, the strong learning capability of the neural network is used for carrying out characteristic extraction and analysis on a system model, the neural network is trained and iterated to output all layers of neurons through a flexible network structure, training test data of the neural network is derived from channel response data and a data set obtained by a minimum mean square error algorithm, two parts of data are input into the network to achieve supervision and fitting of the channel characteristic, then the neural network is supervised in a channel transmission process to process signal distortion, and continuously iterated and updated data are lost with the maximum precision, so that the real characteristic can be better met, the performance of a communication system is improved, and the frequency spectrum utilization rate and the signal transmission effectiveness are improved.
Drawings
FIG. 1 is an algorithmic flow chart of the present invention
FIG. 2 is a deep neural network communication system architecture of the present invention
FIG. 3 is a model of a neuron perceptron of the present invention
FIG. 4 is a deep neural network architecture of the present invention
Detailed Description
The deep neural network channel estimation method and system based on data driving provided by the invention fully utilize the characteristics of the deep neural network supervision training data, flexible network structure and learning channel characteristics, and take two parts of data as input, thereby achieving the purposes of training a learning neuron model and fitting channel distribution. On the premise of the technology of rapid development in the aspects of channel estimation, signal detection, feedback of channel state information and the like by utilizing deep learning, data streams in the MIMO-OFDM communication system are transmitted, so that channel parameters are estimated at a receiving end, and the communication performance is improved by algorithm optimization. The deep neural network based on data driving can firstly carry out supervision fitting on the characteristics of a channel through the input of a data set, learn to process signal distortion in a supervision mode, continuously and iteratively update data to lose the maximum precision, and then can directly use the obtained channel characteristics in actual channel estimation, so that the deep neural network can better accord with the actual characteristics, and more accurate channel state information is output.
The detailed description is made with reference to the implementation steps of the channel estimation algorithm shown in fig. 1, which mainly includes the following steps:
(1) obtaining training test data
(2) Iterative updating coefficient matrix and bias vector of propagation algorithm
(3) The training depth is corrected by the rejection strategy, and the training efficiency is improved
The array antenna technology utilizing space-time coding can solve the problems of low multipath effect and low frequency spectrum utilization rate in system model communication, a multi-input multi-output communication system is built by the technology to serve as a transmission channel between signal streams, and training data are collected by using a simulation data setSet of Ddata＝[h1，h2，...，hn-1，hn]，h∈HT*L，HT*LFor simulating a training data matrix, H '═ C (H x F + theta x), H' is H matrix, real part and imaginary part are extracted, convolution operation is carried out, theta, x are data set input parameters, and F is convolution vector; the second part of data obtains approximate channel state information by a method of the lowest mean square error, FMMSE(θ)＝min(F(θ)，L{M})＝min L{(M-θHx)(M-θHx)H}，
Figure RE-GDA0003635672930000081
Where θ is the matrix to be solved by the algorithm:
Figure RE-GDA0003635672930000082
is calculated to obtain
Figure RE-GDA0003635672930000083
As shown in fig. 2;
the method obtains the simulation data set and the channel response data, and the two parts of data are input into the deep neural network together to carry out the next iterative training and learn the channel characteristics, so that the channel state information is better estimated. The deep neural network is composed of an input layer, a hidden layer and an output layer, a data set and a channel matrix are used as data input, coefficient matrix and bias vector calculation are carried out through the hidden layer, wherein the hidden layer comprises a plurality of neurons, each neuron can be called a perceptron, the output and the output of each perceptron are formed by continuous iteration optimization of the previous layer of iteration and minimum extreme values obtained in a threshold range, as shown in figure 3, input data X belongs to { X ∈ { X } X1，x2，...，xn}，
Figure RE-GDA0003635672930000084
I-th perceptron of l-th layer, yiIn order to be output, the output is,
Figure RE-GDA0003635672930000085
continuously learning channel characteristics through hidden layersFitting nonlinear data and outputting the fitted data as the input of the next layer of neuron, and continuously and repeatedly learning to obtain coefficient matrix tauiAnd a bias vector ω, which outputs for the (n, m) th neuron of the hidden layer:
Figure RE-GDA0003635672930000086
Figure RE-GDA0003635672930000087
outputting a data value for the (n, m) th neuron,
Figure RE-GDA0003635672930000088
as an activation function:
Figure RE-GDA0003635672930000089
the characteristics of the neurons are mapped out through an activation function, a complex communication system model comprises a plurality of data transmission types, a plurality of nonlinear data are input in the neural network training process, and in order to fit the data, the data which are more consistent with channel distribution are output, and the problem is solved by adopting the activation function. The method inputs a data output end to an input end of another layer so as to strengthen the mapping capability of a nonlinear model of each layer of the neuron, sparsely network and solve the problem of gradient disappearance, and a coefficient matrix tauiThe sum offset vector ω is passed through a loss function to calculate the corresponding difference value for τiThe sum omega is close to the actual output of the sample by the minimum difference value, the minimum extreme value is optimized by using the loss function, and then the result of optimizing the final extreme value by using the back propagation algorithm is realized by utilizing the multiple iterations of the gradient descent algorithm; wherein corresponding loss function
Figure RE-GDA00036356729300000810
δ(τix + ω) ═ r, r actual output of the neuron, a desired output, τiCoefficient matrix, ω bias vector, first order bias derivative is taken for coefficient matrix and bias vector:
Figure RE-GDA00036356729300000811
will iterate τiCoefficient matrix, ω bias vector solution: tau isi＝-η(δ(τix+ω)-a)δ′(τix+ω)·x-τ，ω＝-ηδ′(τix+ω)(δ(τix + ω) -a) + ω, and riAfter the coefficient matrix and the omega bias vector are obtained, the output value theta of the ith neuron of the output layer is solved through the valuel：
Figure RE-GDA0003635672930000091
riIs a target value, ZiThe hidden layer output, X is the input vector,
Figure RE-GDA0003635672930000092
vector matrix, with w +1 layers
Figure RE-GDA0003635672930000093
As a vector:
Figure RE-GDA0003635672930000094
updating thetal：
Figure RE-GDA0003635672930000095
E is the sum of squares error, yi，riRespectively, the output value and the target expected value, and finally obtaining the ith output of the W layer of the output layer:
Figure RE-GDA0003635672930000096
calculating each layer output by using a forward propagation algorithm of a deep neural network, solving a difference value by using a reverse solution algorithm of gradient descent and optimizing an extreme value, and randomly generating [0, 1] by using a rejection strategy (Dropout strategy) and a Bernoulli function in order to prevent overfitting when data is fitted]The product of the vector value and the output value of the neuron determines the leaving of the neuron, and the Bernoulli function generates a random probability vector piValue p ofi∈[0，1]Introduction of lossRate Ki，Ki＝pi(1-pi) The Dropout strategy has a regularization effect, with the expectation that individual neurons are deleted being:
Figure RE-GDA0003635672930000097
τiis a matrix of coefficients, xiThe number of neurons activated in deep neural network training is solved by loss rate parameters for inputting vectors, so that the neural network can have a deeper and more efficient network model, as shown in fig. 4. The channel state information is estimated by using a deep neural network, various data are driven and trained by using data, the training times of a perceptron are increased, the channel environment is simulated, real channel distribution is continuously fitted, the channel state information is obtained through data training, and the communication quality is improved.

Claims (6)

1. The deep neural network channel estimation method and system based on data driving are characterized by comprising the following steps:
(1) in the algorithm of channel estimation, the deep neural network is used for estimating and calculating channel estimation state information in consideration of the key effectiveness and reliability of test data on neural network training, in order to supplement the shortage of data quantity, the deep neural network based on data driving is adopted, and the data source of the deep neural network is based on channel response data obtained by a DeepMIMO simulation data set and Minimum Mean Square Error (MMSE) algorithm;
(2) acquiring a data set through simulation data and using the data set as deep neural network training data, inputting the deep neural network training data into neurons, performing iterative training, and solving channel data distortion and learning channel characteristics in a supervision mode after multiple iterative training;
(3) inputting the channel response data and the simulation data set into a deep neural network for training and learning, monitoring the channel characteristics, and outputting a neuron coefficient matrix and a bias vector after sequentially passing through a hidden layer perceptron;
(4) in order to optimize and reduce estimation errors, a forward propagation algorithm is used for continuously outputting an optimized coefficient matrix and an optimized bias vector, and a loss function and a gradient descent algorithm are used for calculating a difference value and then gradient descent iteration is carried out again to obtain an extremum value of optimization minimization;
(5) in the neural network training process, in order to avoid the situation that data seriously deviates from the real characteristics due to an excessive phenomenon occurring in the fitting process, a discarding strategy (Dropout) is adopted to prevent the excessive fitting;
(6) and outputting a channel matrix and performing channel equalization.
2. The data-driven deep neural network channel estimation method and system according to claim 1, wherein the method comprises the following steps (1):
on an MIMO-OFDM (Multiple-Input Multiple-Output-Orthogonal Frequency Division Multiplexing) system, an antenna array technology is utilized to achieve the purposes that a plurality of antennas transmit and receive signals, a deep neural network based on data driving is used for channel estimation, channel state information is obtained, a training data source of the deep MIMO simulation data set comprises compiling DeepMIMO simulation data set testing data, sampling collection of the data is carried out in a limited testing training space through a deep neural learning framework, and a data set is obtained
Figure DEST_PATH_IMAGE001
=
Figure 600318DEST_PATH_IMAGE002
Wherein
Figure DEST_PATH_IMAGE003
，
Figure 19798DEST_PATH_IMAGE004
In order to simulate the matrix of training data,
Figure DEST_PATH_IMAGE005
，
Figure 636724DEST_PATH_IMAGE006
is composed of
Figure DEST_PATH_IMAGE007
The matrix extracts the real part and the imaginary part and then carries out convolution operation,
Figure 458049DEST_PATH_IMAGE008
x is the input parameter of the data set, F is the convolution vector, and the training collection process of the data set is divided into: firstly, randomly dividing original data into m data sets which are not equal to each other, inputting the m data sets into parameters, randomly extracting n data subsets from the m data sets, putting the n data subsets into a training frame, and then carrying out training iteration on the divided subsets for n times by different methods and then carrying out mean value calculation to obtain the data; subsequently saving the obtained data to
Figure 357522DEST_PATH_IMAGE001
(ii) a Channel response data is obtained by using minimum mean square error algorithm, wherein MMSE objective function is
Figure DEST_PATH_IMAGE009
:
Figure DEST_PATH_IMAGE011
Figure DEST_PATH_IMAGE013
Wherein
Figure 998719DEST_PATH_IMAGE014
For the matrix to be solved by the algorithm:
Figure DEST_PATH_IMAGE015
Figure DEST_PATH_IMAGE017
calculated by taking its derivative
Figure 340708DEST_PATH_IMAGE018
，
Figure DEST_PATH_IMAGE019
The value can be found by setting its first derivative to 0
Figure 282119DEST_PATH_IMAGE019
：
Figure DEST_PATH_IMAGE021
Figure 864410DEST_PATH_IMAGE022
L{
Figure DEST_PATH_IMAGE023
Figure DEST_PATH_IMAGE025
Wherein
Figure 524061DEST_PATH_IMAGE026
In order to insert the value at the pilot,
Figure DEST_PATH_IMAGE027
wherein
Figure 171206DEST_PATH_IMAGE028
Is calculated for the autocorrelation matrix
Figure 498282DEST_PATH_IMAGE009
：
Figure 985895DEST_PATH_IMAGE030
The simulation data set and the channel response data are obtained through the steps, and the two parts of data are input into the deep neural network together to carry out the next iterative training and learn the channel characteristics, so that the channel state information is better estimated.
3. The data-driven deep neural network channel estimation method and system according to claim 1, wherein the method comprises the steps of (2) (4):
after the data samples and the data sets are used as a deep neural network supervision and prediction channel model, fitting nonlinear data and outputting the fitted data as the input of the neuron of the next layer through continuously learning channel characteristics and parameters of a hidden layer, and continuously and repeatedly learning to obtain a coefficient matrix
Figure DEST_PATH_IMAGE031
And an offset vector
Figure 867263DEST_PATH_IMAGE032
Then outputting a final result; wherein the data sample is a channel response matrix obtained by a minimum mean square error algorithm
Figure 160841DEST_PATH_IMAGE009
And a data set of
Figure 794954DEST_PATH_IMAGE001
The two parts are input into a deep neural network as a whole data, and fit a nonlinear relation through weighted summation of neurons in a hidden layer and an activation function between the output of a lower layer node and the input of an upper layer node, wherein the step is a forward propagation algorithm process, and the first part of the hidden layer is subjected to the second step
Figure DEST_PATH_IMAGE033
The individual neurons output:
Figure DEST_PATH_IMAGE035
in the formula
Figure 719047DEST_PATH_IMAGE031
In the form of a matrix of coefficients,
Figure 822133DEST_PATH_IMAGE032
in order to be a vector of the offset,
Figure 653822DEST_PATH_IMAGE036
is a first
Figure 158753DEST_PATH_IMAGE033
The individual neurons output a data value that is,
Figure DEST_PATH_IMAGE037
as an activation function:
Figure DEST_PATH_IMAGE039
the characteristic of the neuron is mapped out through an activation function, and a data output end is input to an input end of another layer, so that the mapping capability of a nonlinear model of each layer of the neuron is enhanced, a network is sparse, and the problem of gradient disappearance is solved.
4. The data-driven deep neural network channel estimation method and system according to claim 1, wherein the method comprises the steps of (3) (4):
coefficient matrix obtained by using front line propagation algorithm
Figure 204813DEST_PATH_IMAGE031
And an offset vector
Figure 857511DEST_PATH_IMAGE032
Calculating the corresponding difference value by a loss function in order to
Figure 430575DEST_PATH_IMAGE031
And
Figure 586750DEST_PATH_IMAGE032
the minimum difference value is close to the actual output of the sample, the minimum extreme value is optimized by using the loss function, and then the result of optimizing the final extreme value by using the back propagation algorithm is realized by utilizing the multiple iterations of the gradient descent algorithm; wherein the corresponding loss function:
Figure 587067DEST_PATH_IMAGE040
，
Figure DEST_PATH_IMAGE041
where the actual output of the r neuron, a desired output,
Figure 179591DEST_PATH_IMAGE031
the matrix of coefficients is a matrix of coefficients,
Figure 556346DEST_PATH_IMAGE042
and (3) calculating a first-order partial derivative of the coefficient matrix and the bias vector:
Figure 832606DEST_PATH_IMAGE044
Figure 269404DEST_PATH_IMAGE046
will be iterated
Figure 568798DEST_PATH_IMAGE031
The matrix of coefficients is a matrix of coefficients,
Figure 545982DEST_PATH_IMAGE042
solving bias vectors:
Figure 365164DEST_PATH_IMAGE048
Figure 769601DEST_PATH_IMAGE050
5. the data-driven deep neural network channel estimation method and system according to claim 4, wherein the method comprises:
will be provided with
Figure 556291DEST_PATH_IMAGE031
The matrix of coefficients is a matrix of coefficients,
Figure 71586DEST_PATH_IMAGE042
after the bias vector is obtained, the value is used to solve the second layer of the output layer
Figure DEST_PATH_IMAGE051
Output value of each neuron
Figure 728964DEST_PATH_IMAGE052
：
Figure DEST_PATH_IMAGE053
，
Figure 756831DEST_PATH_IMAGE054
In the formula
Figure DEST_PATH_IMAGE055
In order to achieve the target value,
Figure 296397DEST_PATH_IMAGE056
the output of the hidden layer is output,
Figure DEST_PATH_IMAGE057
in order to input the vector, the vector is input,
Figure 818645DEST_PATH_IMAGE058
vector matrix of
Figure DEST_PATH_IMAGE059
Layer(s)
Figure 127267DEST_PATH_IMAGE060
As a vector:
Figure DEST_PATH_IMAGE061
update
Figure 107728DEST_PATH_IMAGE052
：
Figure DEST_PATH_IMAGE063
Figure 869011DEST_PATH_IMAGE064
=
Figure DEST_PATH_IMAGE065
In the formula, E is the sum of squares error,
Figure 194950DEST_PATH_IMAGE066
，
Figure 92499DEST_PATH_IMAGE055
respectively as output value and target desired value, and finally obtaining the output layer
Figure DEST_PATH_IMAGE067
First of a layer
Figure 462169DEST_PATH_IMAGE051
And (3) outputting:
Figure DEST_PATH_IMAGE069
6. the data-driven deep neural network channel estimation method and system based on claim 1, comprising the following steps (5):
after a data set is input into a front-line propagation algorithm of a deep neural network for training and testing, the obtained result is subjected to error calculation and then input into the neural network again for carrying out back propagation iteration to update a coefficient matrix and a bias vector, and in order to prevent over-fitting, a rejection strategy (Dropout) is used for solving the problem; generation of random probability vectors using bernoulli functions
Figure 976327DEST_PATH_IMAGE070
Value of
Figure DEST_PATH_IMAGE071
Introduction of loss rate
Figure 105957DEST_PATH_IMAGE072
，
Figure DEST_PATH_IMAGE073
The Dropout strategy has a regularization effect, with the expectation that individual neurons are deleted being:
Figure 123592DEST_PATH_IMAGE074
in the formula
Figure 165628DEST_PATH_IMAGE031
In the form of a matrix of coefficients,
Figure DEST_PATH_IMAGE075
for inputting the vector, the number of the neurons activated in deep neural network training is solved through loss rate parameters, and the problems of overlarge resource consumption and overfitting of a network system are solved.
CN202111637126.6A 2021-12-30 2021-12-30 Deep neural network channel estimation method and system based on data driving Pending CN114650199A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111637126.6A CN114650199A (en) 2021-12-30 2021-12-30 Deep neural network channel estimation method and system based on data driving

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111637126.6A CN114650199A (en) 2021-12-30 2021-12-30 Deep neural network channel estimation method and system based on data driving

Publications (1)

Publication Number Publication Date
CN114650199A true CN114650199A (en) 2022-06-21

Family

ID=81994005

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111637126.6A Pending CN114650199A (en) 2021-12-30 2021-12-30 Deep neural network channel estimation method and system based on data driving

Country Status (1)

Country Link
CN (1) CN114650199A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115438592A (en) * 2022-11-08 2022-12-06 成都中科合迅科技有限公司 Industrial research and development design data modeling method based on system engineering
CN116822089A (en) * 2023-07-07 2023-09-29 深圳希哈科技有限公司 Data-driven motor internal disturbance analysis and modeling method
CN117236900A (en) * 2023-10-23 2023-12-15 广东电网有限责任公司 Individual tax data processing method and system based on flow automation

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040208254A1 (en) * 2003-04-21 2004-10-21 Samsung Electronics Co., Ltd. Method and apparatus for estimating a signal sequence in a MIMO-OFDM mobile communication system
CN109391315A (en) * 2018-09-13 2019-02-26 东南大学 A kind of MIMO receiver of data model double drive
CN112182961A (en) * 2020-09-23 2021-01-05 中国南方电网有限责任公司超高压输电公司 Large-scale fading modeling prediction method for wireless network channel of converter station
CN112600772A (en) * 2020-12-09 2021-04-02 齐鲁工业大学 OFDM channel estimation and signal detection method based on data-driven neural network
CN113472706A (en) * 2021-07-12 2021-10-01 南京大学 MIMO-OFDM system channel estimation method based on deep neural network

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040208254A1 (en) * 2003-04-21 2004-10-21 Samsung Electronics Co., Ltd. Method and apparatus for estimating a signal sequence in a MIMO-OFDM mobile communication system
CN109391315A (en) * 2018-09-13 2019-02-26 东南大学 A kind of MIMO receiver of data model double drive
CN112182961A (en) * 2020-09-23 2021-01-05 中国南方电网有限责任公司超高压输电公司 Large-scale fading modeling prediction method for wireless network channel of converter station
CN112600772A (en) * 2020-12-09 2021-04-02 齐鲁工业大学 OFDM channel estimation and signal detection method based on data-driven neural network
CN113472706A (en) * 2021-07-12 2021-10-01 南京大学 MIMO-OFDM system channel estimation method based on deep neural network

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
赵鹏;赵昆;贺英;: "MIMO-OFDM系统中信道估计方法研究", 黑龙江科技信息, no. 17, pages 88 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115438592A (en) * 2022-11-08 2022-12-06 成都中科合迅科技有限公司 Industrial research and development design data modeling method based on system engineering
CN115438592B (en) * 2022-11-08 2023-01-24 成都中科合迅科技有限公司 Industrial research and development design data modeling method based on system engineering
CN116822089A (en) * 2023-07-07 2023-09-29 深圳希哈科技有限公司 Data-driven motor internal disturbance analysis and modeling method
CN117236900A (en) * 2023-10-23 2023-12-15 广东电网有限责任公司 Individual tax data processing method and system based on flow automation
CN117236900B (en) * 2023-10-23 2024-03-29 广东电网有限责任公司 Individual tax data processing method and system based on flow automation

Similar Documents

Publication Publication Date Title
CN114650199A (en) Deep neural network channel estimation method and system based on data driving
CN113472706B (en) MIMO-OFDM system channel estimation method based on deep neural network
CN111464465B (en) Channel estimation method based on integrated neural network model
CN109993280B (en) Underwater sound source positioning method based on deep learning
WO2020253690A1 (en) Deep learning beam domain channel estimation method based on approximate message passing algorithm
CN108566257B (en) Signal recovery method based on back propagation neural network
Ozpoyraz et al. Deep learning-aided 6G wireless networks: A comprehensive survey of revolutionary PHY architectures
CN112637093B (en) Signal detection method based on model-driven deep learning
CN108768585B (en) Multi-user detection method of uplink signaling-free non-orthogonal multiple access (NOMA) system based on deep learning
CN110336594B (en) Deep learning signal detection method based on conjugate gradient descent method
CN106059972A (en) Modulation identification method under MIMO related channel based on machine learning algorithm
CN108964725B (en) Sparse estimation method of channel parameters in time-varying large-scale MIMO network
CN111614584B (en) Transform domain adaptive filtering channel estimation method based on neural network
CN110808932B (en) Multi-layer sensor rapid modulation identification method based on multi-distribution test data fusion
CN112054832B (en) Deep learning multiple-input multiple-output detection method
CN112422208B (en) Signal detection method based on antagonistic learning under unknown channel model
Zhang et al. Designing tensor-train deep neural networks for time-varying MIMO channel estimation
Zhang et al. On the performance of deep neural network aided channel estimation for underwater acoustic OFDM communications
CN116192209A (en) Gradient uploading method for air computing federal learning under MIMO channel
CN113037668B (en) Millimeter wave point-to-point communication channel equalization method
CN113285902B (en) Design method of OFDM system detector
CN113794660B (en) Model-driven deep neural network method for multi-input multi-output detection
CN111786915B (en) Wireless communication method and system based on deep learning under nonlinear distortion condition
CN114759997B (en) MIMO system signal detection method based on data model double driving
Yıldırım et al. Deep receiver design for multi-carrier waveforms using cnns

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
WD01 Invention patent application deemed withdrawn after publication
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20220621