Nothing Special   »   [go: up one dir, main page]

CN111683023B - Model-driven large-scale equipment detection method based on deep learning - Google Patents

Model-driven large-scale equipment detection method based on deep learning Download PDF

Info

Publication number
CN111683023B
CN111683023B CN202010305189.0A CN202010305189A CN111683023B CN 111683023 B CN111683023 B CN 111683023B CN 202010305189 A CN202010305189 A CN 202010305189A CN 111683023 B CN111683023 B CN 111683023B
Authority
CN
China
Prior art keywords
matrix
base station
iteration
representing
equipment
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
CN202010305189.0A
Other languages
Chinese (zh)
Other versions
CN111683023A (en
Inventor
邵晓丹
陈晓明
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University ZJU
Original Assignee
Zhejiang University ZJU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University ZJU filed Critical Zhejiang University ZJU
Priority to CN202010305189.0A priority Critical patent/CN111683023B/en
Publication of CN111683023A publication Critical patent/CN111683023A/en
Application granted granted Critical
Publication of CN111683023B publication Critical patent/CN111683023B/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
    • H04L25/0224—Channel estimation using sounding signals
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • 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
    • 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
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W24/00—Supervisory, monitoring or testing arrangements
    • H04W24/02—Arrangements for optimising operational condition
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W88/00—Devices specially adapted for wireless communication networks, e.g. terminals, base stations or access point devices
    • H04W88/08—Access point devices

Landscapes

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

Abstract

The invention discloses a model-driven large-scale equipment detection method based on deep learning. In the cellular internet of things, a base station equipped with a large-scale antenna array simultaneously serves a large number of single-antenna devices. Only a portion of the devices are active at any given time slot, while the other devices are in a dormant state. The invention adopts an authorization-free random access protocol, namely, the activation equipment simultaneously sends a known pilot frequency sequence to the base station. The base station constructs a deep learning network based on a vector approximation message transfer model, trains parameters of the deep learning network in a back propagation mode, detects the equipment state with less training data and estimates the channel information of the equipment. And finally, the base station performs data interaction with the activation equipment by using the estimated channel information. The invention has strong robustness on the distribution of the equipment state matrix and the ill-conditioned pilot frequency matrix, and provides an efficient equipment detection and channel estimation method for the cellular Internet of things with large-scale equipment access.

Description

Model-driven large-scale equipment detection method based on deep learning
Technical Field
The invention relates to the field of wireless communication, in particular to large-scale active device detection and channel estimation in a cellular internet of things.
Background
Enhanced mobile broadband, large-scale machine communication and ultra-reliable low-delay communication are three main scenes of a 6G wireless network. Particularly, with the rise and development of the internet of things and artificial intelligence, a future wireless network needs to support simultaneous access of massive terminal devices, so that large-scale machine communication becomes a hotspot for research in academia and industry. An important feature of large-scale machine communication is that the activation mode of the terminal devices in the network is usually intermittent. In each time slot, only a portion of the terminals are active and need to communicate with the base station. Other devices are temporarily in a dormant state to conserve energy, and they are only activated when triggered by an external event. The activated terminal equipment simultaneously transmits pilot frequency sequences to the base station at the beginning stage of each time slot, and the base station learns which terminals are in an activated state and obtains corresponding channel state information thereof through an activation detection and channel estimation algorithm. And then, in the remaining duration of each time slot, performing uplink and downlink data interaction between the base station and the activated terminal equipment.
Large-scale machine communication requires a communication system to have the capability of processing a large amount of wireless data, and to accurately recognize and dynamically detect the device status and realize high-speed information interaction. Therefore, terminal activation detection in future wireless communication systems is necessarily highly intelligent. As a commonly used artificial intelligence method, machine learning, especially deep learning, has enjoyed great success in the field of computer vision and natural language processing. Recently, machine learning has been applied to wireless communications such as physical layer communications, resource allocation, and intelligent traffic control. However, most existing deep learning networks are data-driven, using standard neural networks like a black box structure, for extensive data training. Which is often scarce in wireless communication networks. And due to the huge potential equipment in the large-scale machine communication system and the use of large-scale antenna arrays, the dimension of each training data is very high, so that the training process consumes a lot of time.
Most of the current large-scale terminal detection algorithms are only suitable for the situation that the pilot matrixes are gaussian distributions which are independent and distributed, and the performance of the algorithms is seriously reduced due to ill-conditioned pilot matrixes. In addition, these detection algorithms generally assume that the device state matrix conforms to a bernoulli-gaussian distribution, however this assumption is limiting because the channel in a real environment typically exhibits more complex statistical properties. In general, it is more reasonable to assume that the device state matrix obeys a bernoulli-mixed gaussian distribution. How to use the characteristics of the channels, the pilot matrix, the intermittent activation of the equipment and the like to construct a new detection algorithm. How to construct a deep network based on a detection algorithm so as to train parameters in the deep network in a back propagation mode and how to realize large-scale active device detection with less training data and shorter training time. These problems become particularly critical in future large-scale machine communication.
Disclosure of Invention
The invention aims to solve the problems that when a base station is provided with a large-scale antenna array, in the existing large-scale access system based on machine learning, the parameter training time of a terminal activation detection and channel estimation scheme is long, the calculation complexity is high and the distribution of a required pilot matrix is limited, and provides a model-driven large-scale equipment detection method based on deep learning.
The invention adopts the following specific technical scheme:
a model-driven deep learning-based large-scale equipment detection method comprises the following steps:
1) at the beginning stage of each time slot with the length of T, all activated terminal equipment simultaneously sends pilot sequences with the length of L to a base station;
2) after receiving the pilot frequency sequence, the base station maps the received signal from a high-dimensional space to a low-dimensional space based on a data decomposition method so as to reduce the algorithm complexity;
3) in a low-dimensional space, a base station constructs an activation detector based on deep learning based on a vector approximation message transfer model;
4) after the base station obtains the activated detector in the step 3), training unknown parameters in the model layer by using a back propagation mode;
5) the base station substitutes the trained parameters into the activation detector in the step 3), detects the activated terminal equipment and estimates the channel state information of the activated equipment;
6) and in the time length with the remaining length of T-L of each time slot, the base station performs uplink and downlink data interaction with the activation equipment by using the channel estimation value.
On the basis of the technical scheme, each step can be further realized by adopting the following specific mode.
Preferably, the data decomposition method in step 2) is as follows:
firstly, the base station carries out singular value decomposition on a received signal Y:
Figure GDA0003062337740000031
wherein SsdIs a unitary matrix, VsdIs a matrix of singular values and is,
Figure GDA0003062337740000032
is a unitary matrix; then obtain
Figure GDA0003062337740000033
Wherein
Figure GDA0003062337740000034
Is SsdFront r ofeThe columns of the image data are,
Figure GDA0003062337740000035
is composed of VsdUpper left corner r ofe×reA square matrix of elements, where reIs the rank of the unknown signal that needs to be detected; followed by taking
Figure GDA0003062337740000036
Front r ofeObtaining U; data decomposition satisfies
Figure GDA0003062337740000037
And V has a rank re，
Figure GDA0003062337740000038
And UUHI, where I is the identity matrix and M is the number of antennas of the base station.
Preferably, the activation detector based on deep learning in step 3) is:
a) setting the training parameters to
Figure GDA0003062337740000039
And
Figure GDA00030623377400000310
wherein
Figure GDA00030623377400000311
The power adjusting parameter of the nth terminal device in the T-layer network, T represents the transposition operation,
Figure GDA00030623377400000312
representing the training parameters at the t-th iteration, wherein
Figure GDA00030623377400000313
The probability value of the jth component in the bernoulli-gaussian mixture distribution of the nth terminal device in the t-th layer network,
Figure GDA00030623377400000314
the Gaussian distribution variance of the jth component in the Bernoulli-Gaussian mixture distribution of the nth terminal equipment in the t-th network is obtained; the subscript N ∈ {1,2, …, N } represents the nth device, N represents the total number of terminal devices, J ∈ {1,2, …, J } represents the jth component in the Bernoulli-Gaussian mixture, J is the total number of components; upper bound of iteration number is tau1,max，τ2,maxAnd Tmax；
3.b) initializing the number of internal iterations τ to 0 for any
Figure GDA00030623377400000315
Updating non-linear estimation values of elements in state matrix of terminal equipment of tau iteration in sequence
Figure GDA00030623377400000316
Intermediate variables
Figure GDA00030623377400000317
Sum noise accuracy
Figure GDA00030623377400000318
First of all update
Figure GDA0003062337740000041
Wherein enVariable representing activation probability of nth device
Figure GDA0003062337740000042
Is calculated by
Figure GDA0003062337740000043
Figure GDA0003062337740000044
Representing the r-th noise-accuracy value, an intermediate variable, of the t-th iteration
Figure GDA0003062337740000045
Representing variables
Figure GDA0003062337740000046
Meet the mean of 0 and variance of
Figure GDA0003062337740000047
Normal distribution, intermediate variable of
Figure GDA0003062337740000048
Representing variables
Figure GDA0003062337740000049
Meet the mean of 0 and variance of
Figure GDA00030623377400000410
Normal distribution of (2);
then, update
Figure GDA00030623377400000411
Wherein
Figure GDA00030623377400000412
Is an intermediate variable, function g'1,nrIs calculated by
Figure GDA00030623377400000413
Then, update
Figure GDA00030623377400000414
Wherein||.||2Represents a two-norm of the signal,
Figure GDA00030623377400000415
is the intermediate variable(s) of the variable,
Figure GDA00030623377400000416
is formed by
Figure GDA00030623377400000417
A vector of components; updating iteration time τ ← τ +1, and then repeating the next iteration until τ ═ τ ← τ +11,maxStopping the circulation and finally outputting an estimated value
Figure GDA00030623377400000418
Wherein
Figure GDA00030623377400000419
Is formed by
Figure GDA00030623377400000420
A matrix of compositions;
3, c) to any
Figure GDA00030623377400000421
Non-linear estimation of state matrix for sequentially updating t-th iteration
Figure GDA00030623377400000422
Accuracy of noise
Figure GDA00030623377400000423
Accuracy of noise
Figure GDA00030623377400000424
And intermediate variables
Figure GDA00030623377400000425
Figure GDA00030623377400000426
Figure GDA00030623377400000427
Figure GDA00030623377400000428
Figure GDA0003062337740000051
3, d) updating the linear estimates of the state matrix of the t-th iteration in turn according to the following formula
Figure GDA0003062337740000052
Intermediate variables
Figure GDA0003062337740000053
Accuracy of noise
Figure GDA0003062337740000054
Accuracy of noise
Figure GDA0003062337740000055
And intermediate variables
Figure GDA0003062337740000056
Figure GDA0003062337740000057
Figure GDA0003062337740000058
Figure GDA0003062337740000059
Figure GDA00030623377400000510
Figure GDA00030623377400000511
Wherein A (. beta.) ist)＝Adiag(βt) Where A represents the pilot matrix, diag (β)t) The representation forms a vector betatIs a diagonal matrix of diagonal elements, σ2Representing the variance of the noise, INRepresenting a diagonal matrix of dimension NxN with diagonal elements of 1, vrRepresenting the r-th column of the low-dimensional spatial reception matrix V.
Preferably, the method for training the unknown parameters in the model layer by using the back propagation method in the step 4) comprises the following steps:
initializing training parameters
Figure GDA00030623377400000512
And
Figure GDA00030623377400000513
are each beta0And Ω0Setting TmaxSetting a network layer number identifier t as 0 for training the upper bound of the layer number, and starting parameter learning of the t-th iteration by using 3.b), 3.c) and 3.d) in step 3:
first fix it
Figure GDA00030623377400000514
Study omegatTo achieve a cost function that minimizes the denoising step
Figure GDA00030623377400000515
(ii) wherein | |. calorifiesFRepresenting the F norm, wherein S is the true value of a state matrix of equipment in the system;
then learn ΩtAnd
Figure GDA00030623377400000516
to achieve a cost function that minimizes the denoising step
Figure GDA00030623377400000517
The object of (a);
then fix omegatAnd
Figure GDA0003062337740000061
learning betatTo achieve a cost function that minimizes the linear least squares step
Figure GDA0003062337740000062
The object of (a); matrix array
Figure GDA0003062337740000063
Is composed of a vector
Figure GDA0003062337740000064
A matrix of formations;
finally study
Figure GDA0003062337740000065
To achieve a minimum linear least squares step cost function
Figure GDA0003062337740000066
The object of (a);
after the iteration of one layer of network is completed, updating external iteration times T ← T +1, and repeating the parameter learning of the next layer of network again until T ═ TmaxAnd stopping the circulation when the model is in-1, and finishing the training of unknown parameters in the model.
Preferably, the method for detecting device activation and channel estimation in step 5) comprises: equivalently considering the number of training layers and the number of iterations, performing the following iterations:
a) initializing the intermediate variables of the r-th column with the number of external iterations t equal to 0
Figure GDA0003062337740000067
Sum noise accuracy
Figure GDA0003062337740000068
Initial value of (2)Are respectively arranged as
Figure GDA0003062337740000069
And
Figure GDA00030623377400000610
maximum number of iterations is Tmax1The base station substitutes the model parameters trained in the step 4) into the activation detector constructed in the step 3);
5.b) performing said 3.b), 3.c) and 3.d) one time;
5, c) to any
Figure GDA00030623377400000611
Updating the external iteration time T ← T +1, and then re-performing the next iteration, namely performing the step 5.b), until T ═ T ← T +max1Time-out loop, and finally output the estimated value of the state matrix
Figure GDA00030623377400000612
D) using the activation criterion:
Figure GDA00030623377400000613
to determine which terminal devices are in an active state, where k is a terminal device identifier, v is an adjustable parameter,
Figure GDA00030623377400000614
is composed of
Figure GDA00030623377400000615
The (c) th row of (a),
Figure GDA00030623377400000616
a set of identities representing the detected active devices; reuse relational expression
Figure GDA00030623377400000617
Recovering the signal estimation value of the original high-dimensional space, thereby obtaining the specific channel estimation value of the active device
Figure GDA00030623377400000618
Wherein
Figure GDA00030623377400000619
Represents an estimate of the unknown state vector in the high-dimensional space,
Figure GDA00030623377400000620
representing and getting
Figure GDA00030623377400000621
Neutralization of
Figure GDA00030623377400000622
Corresponding partial row xikIs the transmitted energy of the pilot.
The invention has the beneficial effects that: the deep learning-based large-scale terminal activation detection and channel estimation method provided by the invention can realize more accurate terminal activation detection and channel estimation by using less training data, and solves a series of problems of high computational complexity, low detection accuracy and the like in the traditional large-scale terminal activation detection and channel estimation problems. The method is suitable for a wider pilot matrix, and can effectively reduce the problem of detection performance reduction caused by a sick pilot matrix.
Drawings
FIG. 1 is a diagram of a deep learning ensemble framework;
FIG. 2 is a block diagram of a per-layer learning network;
FIG. 3 is a graph of normalized mean square error of channel estimation versus the number of deep learning layers when the pilot matrix obeys Gaussian distribution with mean 1 and variance 1, comparing the deep learning-based large-scale terminal channel estimation method of the present invention with other commonly used terminal channel estimation methods;
fig. 4 is a relationship between a detection error rate and a signal-to-noise ratio when comparing the large-scale terminal detection method based on deep learning of the present invention with other common terminal detection methods.
Detailed Description
In this embodiment, a base station of a large-scale access system is provided with M antennas, each terminal is configured with 1 antenna, only a small number of terminals are randomly activated to communicate with the base station in each time slot, and other terminals are temporarily in a sleep state. And the activated terminal can directly access the network without being authorized by the base station. That is, the active terminal transmits pilot sequence to the base station at the same time in the beginning of each time slot, and the base station obtains which terminals are in the active stage and obtains corresponding channel state information through large-scale terminal detection and channel estimation algorithm. And in the rest part of each time slot, activating the terminal to perform data interaction with the base station.
Based on the base station, the embodiment provides a model-driven deep learning-based large-scale device detection method, which includes the following steps:
1) at the beginning of each time slot of length T, all active terminal devices simultaneously transmit pilot sequences of length L to the base station.
2) After receiving the pilot sequence, the base station maps the received signal from a high-dimensional space to a low-dimensional space based on a data decomposition method to reduce the algorithm complexity.
In this step, the data decomposition method is:
firstly, the base station carries out singular value decomposition on a received signal Y:
Figure GDA0003062337740000071
wherein SsdIs a unitary matrix, VsdIs a matrix of singular values and is,
Figure GDA0003062337740000072
is a unitary matrix; then obtain
Figure GDA0003062337740000073
Wherein
Figure GDA0003062337740000074
Is SsdFront r ofeThe columns of the image data are,
Figure GDA0003062337740000081
is composed of VsdUpper left corner r ofe×reA square matrix of elements, where reIs the rank of the unknown signal that needs to be detected; followed by taking
Figure GDA0003062337740000082
Front r ofeObtaining U; data decomposition satisfies
Figure GDA0003062337740000083
And V has a rank re，
Figure GDA0003062337740000084
And UUHI, where I is the identity matrix and M is the number of antennas of the base station.
3) In a low-dimensional space, the base station constructs an activation detector based on deep learning based on a vector approximation message transfer model.
In this step, the vector approximation message transmission detection method is developed according to the outer iteration identification, i.e. each iteration t is a layer of network, and each layer of network input is
Figure GDA0003062337740000085
And
Figure GDA0003062337740000086
the output is
Figure GDA0003062337740000087
And
Figure GDA0003062337740000088
in total of TmaxThe layer networks are connected in series to form a deep learning network. The deep learning-based activation detector specifically comprises:
a) setting the training parameters to
Figure GDA0003062337740000089
And
Figure GDA00030623377400000810
wherein
Figure GDA00030623377400000811
The power adjusting parameter of the nth terminal device in the T-layer network, T represents the transposition operation,
Figure GDA00030623377400000812
representing the training parameters at the t-th iteration, wherein
Figure GDA00030623377400000813
The probability value of the jth component in the bernoulli-gaussian mixture distribution of the nth terminal device in the t-th layer network,
Figure GDA00030623377400000814
the Gaussian distribution variance of the jth component in the Bernoulli-Gaussian mixture distribution of the nth terminal equipment in the t-th network is obtained; the subscript N ∈ {1,2, …, N } represents the nth device, N represents the total number of terminal devices, J ∈ {1,2, …, J } represents the jth component in the Bernoulli-Gaussian mixture, J is the total number of components; upper bound of iteration number is tau1,max，τ2,maxAnd Tmax. t and τ are the outer loop variable and the inner loop variable, respectively.
3.b) initializing the number of internal iterations τ to 0 for any
Figure GDA00030623377400000815
Updating non-linear estimation values of elements in state matrix of terminal equipment of tau iteration in sequence
Figure GDA00030623377400000816
Intermediate variables
Figure GDA00030623377400000817
Sum noise accuracy
Figure GDA00030623377400000818
First of all update
Figure GDA00030623377400000819
Wherein enVariable representing activation probability of nth device
Figure GDA00030623377400000820
Is calculated by
Figure GDA00030623377400000821
Figure GDA00030623377400000822
For the variance of the jth component of the device n for t iterations,
Figure GDA00030623377400000823
representing the r-th noise-accuracy value, an intermediate variable, of the t-th iteration
Figure GDA00030623377400000824
Representing variables
Figure GDA00030623377400000825
Meet the mean of 0 and variance of
Figure GDA0003062337740000091
Normal distribution, intermediate variable of
Figure GDA0003062337740000092
Representing variables
Figure GDA0003062337740000093
Meet the mean of 0 and variance of
Figure GDA0003062337740000094
Normal distribution of (2);
then, update
Figure GDA0003062337740000095
Wherein
Figure GDA0003062337740000096
Is an intermediate variable, function g'1,nrIs calculated by
Figure GDA0003062337740000097
Then, update
Figure GDA0003062337740000098
Wherein | |. calo | |)2Represents a two-norm of the signal,
Figure GDA0003062337740000099
is the intermediate variable(s) of the variable,
Figure GDA00030623377400000910
is formed by
Figure GDA00030623377400000911
A vector of components; updating iteration time τ ← τ +1, and then repeating the next iteration until τ ═ τ ← τ +11,maxStopping the circulation and finally outputting an estimated value
Figure GDA00030623377400000912
Wherein
Figure GDA00030623377400000913
Is formed by
Figure GDA00030623377400000914
A matrix of compositions;
3, c) to any
Figure GDA00030623377400000915
Non-linear estimation of state matrix for sequentially updating t-th iteration
Figure GDA00030623377400000916
Accuracy of noise
Figure GDA00030623377400000917
Accuracy of noise
Figure GDA00030623377400000918
And intermediateMeasurement of
Figure GDA00030623377400000919
Figure GDA00030623377400000920
Figure GDA00030623377400000921
Figure GDA00030623377400000922
Figure GDA00030623377400000923
3, d) updating the linear estimates of the state matrix of the t-th iteration in turn according to the following formula
Figure GDA00030623377400000924
Intermediate variables
Figure GDA00030623377400000925
Accuracy of noise
Figure GDA00030623377400000926
Accuracy of noise
Figure GDA00030623377400000927
And intermediate variables
Figure GDA00030623377400000928
Figure GDA0003062337740000101
Figure GDA0003062337740000102
Figure GDA0003062337740000103
Figure GDA0003062337740000104
Figure GDA0003062337740000105
Wherein A (. beta.) ist)＝Adiag(βt) Where A represents the pilot matrix, diag (β)t) The representation forms a vector betatIs a diagonal matrix of diagonal elements, σ2Representing the variance of the noise, INRepresenting a diagonal matrix of dimension NxN with diagonal elements of 1, vrRepresenting the r-th column of the low-dimensional spatial reception matrix V.
4) After the base station obtains the activated detector in the step 3), the unknown parameters in the model are trained layer by using a back propagation mode.
In this step, the method for training the unknown parameters in the model layer by using the back propagation mode comprises the following steps:
initializing training parameters
Figure GDA0003062337740000106
And
Figure GDA0003062337740000107
are each beta0And Ω0Setting TmaxSetting a network layer number identifier t as 0 for training the upper bound of the layer number, and starting parameter learning of the t-th iteration by using 3.b), 3.c) and 3.d) in step 3:
first fix it
Figure GDA0003062337740000108
Study omegatTo achieve a cost function of minimizing the denoising stepNumber of
Figure GDA0003062337740000109
(ii) wherein | |. calorifiesFRepresenting the F norm, wherein S is the true value of a state matrix of equipment in the system;
then learn ΩtAnd
Figure GDA00030623377400001010
to achieve a cost function that minimizes the denoising step
Figure GDA00030623377400001011
The object of (a);
then fix omegatAnd
Figure GDA00030623377400001012
learning betatTo achieve a cost function that minimizes the linear least squares step
Figure GDA00030623377400001013
The object of (a); matrix array
Figure GDA00030623377400001014
Is composed of a vector
Figure GDA00030623377400001015
A matrix of formations;
finally study
Figure GDA0003062337740000111
To achieve a minimum linear least squares step cost function
Figure GDA0003062337740000112
The object of (a);
after the iteration of one layer of network is completed, updating external iteration times T ← T +1, and repeating the parameter learning of the next layer of network again until T ═ TmaxAnd stopping the circulation when the model is in-1, and finishing the training of unknown parameters in the model.
5) And the base station substitutes the trained parameters into the activation detector in the step 3), detects the activated terminal equipment and estimates the channel state information of the activated equipment.
In this step, the deep learning model of the invention is divided into an off-line training stage and an on-line detection stage, wherein the off-line training stage is to obtain the value of the trained parameter, and t is the t-th network; in the on-line detection stage, the trained parameters are substituted into the algorithm structure, and t is equivalently regarded as the t-th iteration. Therefore, the device activation detection and channel estimation method is specifically as follows:
equivalently considering the number of training layers and the number of iterations, performing the following iterations:
a) initializing the intermediate variables of the r-th column with the number of external iterations t equal to 0
Figure GDA0003062337740000113
Sum noise accuracy
Figure GDA0003062337740000114
Are respectively set as
Figure GDA0003062337740000115
And
Figure GDA0003062337740000116
maximum number of iterations is Tmax1The base station substitutes the model parameters trained in the step 4) into the activation detector constructed in the step 3);
5.b) performing said 3.b), 3.c) and 3.d) one time;
5, c) to any
Figure GDA0003062337740000117
Updating the external iteration time T ← T +1, and then re-performing the next iteration, namely performing the step 5.b), until T ═ T ← T +max1Time-out loop, and finally output the estimated value of the state matrix
Figure GDA0003062337740000118
D) using the activation criterion:
Figure GDA0003062337740000119
to determine which terminal devices are in an active state, where k is a terminal device identifier, v is an adjustable parameter,
Figure GDA00030623377400001110
is composed of
Figure GDA00030623377400001111
The (c) th row of (a),
Figure GDA00030623377400001112
a set of identities representing the detected active devices; reuse relational expression
Figure GDA00030623377400001113
Recovering the signal estimation value of the original high-dimensional space, thereby obtaining the specific channel estimation value of the active device
Figure GDA00030623377400001114
Wherein
Figure GDA00030623377400001115
Represents an estimate of the unknown state vector in the high-dimensional space,
Figure GDA00030623377400001116
representing and getting
Figure GDA00030623377400001117
Neutralization of
Figure GDA00030623377400001118
Corresponding partial row xikIs the transmitted energy of the pilot.
6) And in the time length with the remaining length of T-L of each time slot, the base station performs uplink and downlink data interaction with the activation equipment by using the channel estimation value.
The deep learning overall framework and the schematic diagram of each layer of learning network are respectively shown in fig. 1 and fig. 2. As can be seen by computer simulation: as shown in fig. 3, the pilot matrix is set to obey gaussian distribution with mean 1 and variance 1, and compared with the conventional channel estimation scheme, i.e., the approximate message passing algorithm, the large-scale terminal channel estimation scheme of the present invention has significantly improved estimation accuracy of the vector approximate message passing algorithm. Fig. 4 shows that the detection accuracy of the large-scale detection method provided by the invention is obviously improved compared with the detection accuracy of the traditional orthogonal matching pursuit algorithm, and the influence of the estimation value of the state matrix rank on the detection accuracy is within an acceptable range. The advantages are that on the basis of inheriting the original advantages of the traditional vector approximation message transfer algorithm, the scheme combines more accurate Bernoulli-Gaussian mixture distribution of the state matrix, combines deep learning to optimize the network structure of the vector approximation message transfer algorithm, and effectively trains system parameters. Therefore, the terminal activation detection and channel estimation scheme provided by the invention can provide an efficient terminal activation detection and channel estimation method for a large-scale communication system.

Claims (4)

1. A model-driven large-scale equipment detection method based on deep learning is characterized by comprising the following steps:
1) at the beginning stage of each time slot with the length of T, all activated terminal equipment simultaneously sends pilot sequences with the length of L to a base station;
2) after receiving the pilot frequency sequence, the base station maps the received signal from a high-dimensional space to a low-dimensional space based on a data decomposition method so as to reduce the algorithm complexity;
3) in a low-dimensional space, a base station constructs an activation detector based on deep learning based on a vector approximation message transfer model;
4) after the base station obtains the activated detector in the step 3), training unknown parameters in the model layer by using a back propagation mode;
5) the base station substitutes the trained parameters into the activation detector in the step 3), detects the activated terminal equipment and estimates the channel state information of the activated equipment;
6) in the time length of the rest length of each time slot being T-L, the base station utilizes the channel estimation value to carry out uplink and downlink data interaction with the activation equipment;
the activation detector based on deep learning in the step 3) is as follows:
a) setting the training parameters to
Figure FDA0002957375400000011
And
Figure FDA0002957375400000012
wherein
Figure FDA0002957375400000013
The power adjusting parameter of the nth terminal device in the T-layer network, T represents the transposition operation,
Figure FDA0002957375400000014
representing the training parameters at the t-th iteration, wherein
Figure FDA0002957375400000015
The probability value of the jth component in the bernoulli-gaussian mixture distribution of the nth terminal device in the t-th layer network,
Figure FDA0002957375400000016
the Gaussian distribution variance of the jth component in the Bernoulli-Gaussian mixture distribution of the nth terminal equipment in the t-th network is obtained; the subscript N ∈ {1,2, …, N } represents the nth device, N represents the total number of terminal devices, J ∈ {1,2, …, J } represents the jth component in the Bernoulli-Gaussian mixture, J is the total number of components; upper bound of iteration number is tau1,max，τ2,maxAnd Tmax；
3.b) initializing the number of internal iterations τ to 0 for any
Figure FDA0002957375400000021
Updating non-linear estimation values of elements in state matrix of terminal equipment of tau iteration in sequence
Figure FDA0002957375400000022
Intermediate variables
Figure FDA0002957375400000023
Sum noise accuracy
Figure FDA0002957375400000024
First of all update
Figure FDA0002957375400000025
Wherein enVariable representing activation probability of nth device
Figure FDA0002957375400000026
Is calculated by
Figure FDA0002957375400000027
Figure FDA0002957375400000028
Representing the r-th noise-accuracy value, an intermediate variable, of the t-th iteration
Figure FDA0002957375400000029
Representing variables
Figure FDA00029573754000000210
Meet the mean of 0 and variance of
Figure FDA00029573754000000211
Normal distribution, intermediate variable of
Figure FDA00029573754000000212
Representing variables
Figure FDA00029573754000000213
Meet the mean of 0 and variance of
Figure FDA00029573754000000214
Normal distribution of (2);
then, update
Figure FDA00029573754000000215
Wherein
Figure FDA00029573754000000216
Is an intermediate variable, function g'1,nrIs calculated by
Figure FDA00029573754000000217
Then, update
Figure FDA00029573754000000218
Wherein | |. calo | |)2Represents a two-norm of the signal,
Figure FDA00029573754000000219
is the intermediate variable(s) of the variable,
Figure FDA00029573754000000220
is formed by
Figure FDA00029573754000000221
A vector of components; updating iteration time τ ← τ +1, and then repeating the next iteration until τ ═ τ ← τ +11,maxStopping the circulation and finally outputting an estimated value
Figure FDA00029573754000000222
Wherein
Figure FDA00029573754000000223
Is formed by
Figure FDA00029573754000000224
A matrix of compositions;
3, c) to any
Figure FDA00029573754000000225
Non-linear estimation of state matrix for sequentially updating t-th iteration
Figure FDA00029573754000000226
Accuracy of noise
Figure FDA00029573754000000227
Accuracy of noise
Figure FDA00029573754000000228
And intermediate variables
Figure FDA00029573754000000229
Figure FDA00029573754000000230
Figure FDA0002957375400000031
Figure FDA0002957375400000032
Figure FDA0002957375400000033
3, d) updating the linear estimates of the state matrix of the t-th iteration in turn according to the following formula
Figure FDA0002957375400000034
Intermediate variables
Figure FDA0002957375400000035
Accuracy of noise
Figure FDA0002957375400000036
Accuracy of noise
Figure FDA0002957375400000037
And intermediate variables
Figure FDA0002957375400000038
Figure FDA0002957375400000039
Figure FDA00029573754000000310
Figure FDA00029573754000000311
Figure FDA00029573754000000312
Figure FDA00029573754000000313
Wherein A (. beta.) ist)＝Adiag(βt) Where A represents the pilot matrix, diag (β)t) The representation forms a vector betatIs a diagonal matrix of diagonal elements, σ2Representing the variance of the noise, INRepresenting a diagonal matrix of dimension NxN with diagonal elements of 1, vrRepresenting the r-th column of the low-dimensional spatial reception matrix V.
2. The model-driven deep learning-based large-scale equipment detection method according to claim 1, wherein the data decomposition method in step 2) is as follows:
firstly, the base station carries out singular value decomposition on a received signal Y:
Figure FDA00029573754000000314
wherein SsdIs a unitary matrix, VsdIs a matrix of singular values and is,
Figure FDA00029573754000000315
is a unitary matrix; then obtain
Figure FDA00029573754000000316
Wherein
Figure FDA00029573754000000317
Is SsdFront r ofeThe columns of the image data are,
Figure FDA00029573754000000318
is composed of VsdUpper left corner r ofe×reA square matrix of elements, where reIs the rank of the unknown signal that needs to be detected; followed by taking
Figure FDA00029573754000000319
Front r ofeObtaining U; data decomposition satisfies
Figure FDA00029573754000000320
And V has a rank re，
Figure FDA00029573754000000321
And UUHI, where I is the identity matrix and M is the number of antennas of the base station.
3. The model-driven deep learning-based large-scale equipment detection method according to claim 1, wherein the method for training the unknown parameters in the model layer by layer in the back propagation manner in step 4) comprises:
initializing training parameters
Figure FDA0002957375400000041
And
Figure FDA0002957375400000042
are each beta0And Ω0Setting TmaxSetting a network layer number identifier t as 0 for training the upper bound of the layer number, and starting parameter learning of the t-th iteration by using 3.b), 3.c) and 3.d) in step 3:
first fix it
Figure FDA0002957375400000043
Study omegatTo achieve a cost function that minimizes the denoising step
Figure FDA0002957375400000044
(ii) wherein | |. calorifiesFRepresenting the F norm, wherein S is the true value of a state matrix of equipment in the system;
then learn ΩtAnd
Figure FDA0002957375400000045
to achieve a cost function that minimizes the denoising step
Figure FDA0002957375400000046
The object of (a);
then fix omegatAnd
Figure FDA0002957375400000047
learning betatTo achieve a cost function that minimizes the linear least squares step
Figure FDA0002957375400000048
The object of (a); matrix array
Figure FDA0002957375400000049
Is composed of a vector
Figure FDA00029573754000000410
A matrix of formations;
finally study
Figure FDA00029573754000000411
To achieve a minimum linear least squares step cost function
Figure FDA00029573754000000412
The object of (a);
after the iteration of one layer of network is completed, updating external iteration times T ← T +1, and repeating the parameter learning of the next layer of network again until T ═ TmaxAnd stopping the circulation when the model is in-1, and finishing the training of unknown parameters in the model.
4. The model-driven deep learning based large-scale device detection method as claimed in claim 1, wherein the device activation detection and channel estimation method in step 5) is: equivalently considering the number of training layers and the number of iterations, performing the following iterations:
a) initializing the intermediate variables of the r-th column with the number of external iterations t equal to 0
Figure FDA00029573754000000413
Sum noise accuracy
Figure FDA00029573754000000414
Are respectively set as
Figure FDA00029573754000000415
And
Figure FDA00029573754000000416
maximum number of iterations is Tmax1The base station substitutes the model parameters trained in the step 4) into the activation detector constructed in the step 3)Performing the following steps;
5.b) performing said 3.b), 3.c) and 3.d) one time;
5, c) to any
Figure FDA0002957375400000051
Updating the external iteration time T ← T +1, and then re-performing the next iteration, namely performing the step 5.b), until T ═ T ← T +max1Time-out loop, and finally output the estimated value of the state matrix
Figure FDA0002957375400000052
D) using the activation criterion:
Figure FDA0002957375400000053
to determine which terminal devices are in an active state, where k is a terminal device identifier, v is an adjustable parameter,
Figure FDA0002957375400000054
is composed of
Figure FDA0002957375400000055
The (c) th row of (a),
Figure FDA0002957375400000056
a set of identities representing the detected active devices; reuse relational expression
Figure FDA0002957375400000057
Recovering the signal estimation value of the original high-dimensional space, thereby obtaining the specific channel estimation value of the active device
Figure FDA0002957375400000058
Wherein
Figure FDA0002957375400000059
Represents an estimate of the unknown state vector in the high-dimensional space,
Figure FDA00029573754000000510
representing and getting
Figure FDA00029573754000000511
Neutralization of
Figure FDA00029573754000000512
Corresponding partial row xikIs the transmitted energy of the pilot.
CN202010305189.0A 2020-04-17 2020-04-17 Model-driven large-scale equipment detection method based on deep learning Active CN111683023B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010305189.0A CN111683023B (en) 2020-04-17 2020-04-17 Model-driven large-scale equipment detection method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010305189.0A CN111683023B (en) 2020-04-17 2020-04-17 Model-driven large-scale equipment detection method based on deep learning

Publications (2)

Publication Number Publication Date
CN111683023A CN111683023A (en) 2020-09-18
CN111683023B true CN111683023B (en) 2021-08-24

Family

ID=72451623

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010305189.0A Active CN111683023B (en) 2020-04-17 2020-04-17 Model-driven large-scale equipment detection method based on deep learning

Country Status (1)

Country Link
CN (1) CN111683023B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113766541B (en) * 2021-09-07 2023-10-17 北京交通大学 Active device in MMTC scene and detection method of using channel thereof
CN113766669B (en) * 2021-11-10 2021-12-31 香港中文大学（深圳） Large-scale random access method based on deep learning network

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107276934A (en) * 2017-06-30 2017-10-20 东南大学 A kind of extensive up Robust Detection Method of mimo system multi-user
CN107483089A (en) * 2017-08-15 2017-12-15 南京林业大学 The navigation system and design method of a kind of multiple-input multiple-output broadcast

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017067574A1 (en) * 2015-10-19 2017-04-27 Huawei Technologies Co., Ltd. Mitigating inter-cell pilot interference via network-based greedy sequence selection and exchange
US11025456B2 (en) * 2018-01-12 2021-06-01 Apple Inc. Time domain resource allocation for mobile communication
CN109714086B (en) * 2019-01-23 2021-09-14 上海大学 Optimized MIMO detection method based on deep learning
CN110336631B (en) * 2019-06-04 2020-10-13 浙江大学 Signal detection method based on deep learning

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107276934A (en) * 2017-06-30 2017-10-20 东南大学 A kind of extensive up Robust Detection Method of mimo system multi-user
CN107483089A (en) * 2017-08-15 2017-12-15 南京林业大学 The navigation system and design method of a kind of multiple-input multiple-output broadcast

Also Published As

Publication number Publication date
CN111683023A (en) 2020-09-18

Similar Documents

Publication Publication Date Title
CN111641570B (en) Joint equipment detection and channel estimation method based on deep learning
KR102034955B1 (en) Method and apparatus for controlling transmit power in wireless communication system based on neural network
CN113194548B (en) Intelligent super-surface-assisted large-scale passive random access method
CN105790813B (en) Code book selection method based on deep learning under a kind of extensive MIMO
CN110912598A (en) Large-scale MIMO system CSI feedback method based on long-time attention mechanism
Bai et al. Deep learning based fast multiuser detection for massive machine-type communication
CN110139392B (en) LTE electric wireless private network random access channel multiple conflict detection method
CN111683023B (en) Model-driven large-scale equipment detection method based on deep learning
CN110177062B (en) Terminal activation detection and channel estimation method
CN112637094A (en) Multi-user MIMO receiving method based on model-driven deep learning
CN112910806B (en) Joint channel estimation and user activation detection method based on deep neural network
CN110971547B (en) Millimeter wave/terahertz-based broadband large-scale terminal detection and channel estimation method
CN116192307A (en) Distributed cooperative multi-antenna cooperative spectrum intelligent sensing method, system, equipment and medium under non-Gaussian noise
Shi et al. Sparse signal processing for massive device connectivity via deep learning
Deng et al. Joint flexible duplexing and power allocation with deep reinforcement learning in cell-free massive MIMO system
CN113965881B (en) Millimeter wave integrated communication and sensing method under shielding effect
Sery et al. A sequential gradient-based multiple access for distributed learning over fading channels
CN114204971B (en) Iterative aggregate beam forming design and user equipment selection method
Yu et al. Complex-Valued Neural Network Based Federated Learning for Multi-User Indoor Positioning Performance Optimization
CN107181705A (en) A kind of half-blind channel estimating method and system
CN107426748B (en) Method for estimating performance of multiple sensors in wireless network control system
Alvi et al. Federated learning cost disparity for IoT devices
CN115361258A (en) Large-scale MIMO sparse channel estimation method and related equipment
Zhou et al. Over-the-air computation assisted hierarchical personalized federated learning
Shi et al. Automatic Neural Network Construction-Based Channel Estimation for IRS-Aided Communication Systems

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