Method and apparatus for modulation recognition of signals based on cyclic residual network

Patent metadata

Field	Value
Publication number	US-11909563-B2
Application number	US-202017440120-A
Country	US
Kind code	B2
Filing date	Jul 30, 2020
Priority date	Oct 24, 2019
Publication date	Feb 20, 2024
Grant date	Feb 20, 2024

Abstract

The embodiments of the present application provide a method and apparatus for modulation recognition of signals based on cyclic residual network, the method comprises: obtaining a signal matrix of a to-be-recognized signal, and extracting real part information and imaginary part information of the signal matrix; generating, according to extracted real part information and imaginary part information, a real-and-imaginary-part feature matrix of the to-be-recognized signal; converting, according to a preset matrix conversion method, the real-and-imaginary-part feature matrix into an amplitude-and-phase feature matrix; and inputting the amplitude-and-phase feature matrix into a pre-trained cyclic residual network to obtain a modulation mode corresponding to the to-be-recognized signal. In the embodiments of the present application, the processing of the to-be-recognized signal is simple and easy to operate, in which neither complex algorithms nor manual processing is required, the flexibility of recognition is high, and the result of modulation recognition of the to-be-recognized signal can be accurately obtained.

First claim

What is claimed is:

1. A method for modulation recognition of signals based on a cyclic residual network, comprising:
obtaining a signal matrix of a to-be-recognized signal, and extracting real part information and imaginary part information of the signal matrix; wherein, the to-be-recognized signal is a signal whose modulation is to be recognized;
generating, according to the extracted real part information and the extracted imaginary part information, a real-and-imaginary-part feature matrix of the to-be-recognized signal;
converting, according to a preset matrix conversion method, the real-and-imaginary-part feature matrix into an amplitude-and-phase feature matrix; the amplitude-and-phase feature matrix carries amplitude features and phase features of the to-be-recognized signal, and an amount of information of features carried by the amplitude-and-phase feature matrix varies with an amount of information carried by the to-be-recognized signal; and
inputting the amplitude-and-phase feature matrix into a pre-trained cyclic residual network to obtain a modulation mode corresponding to the to-be-recognized signal; wherein the cyclic residual network is obtained by training according to a preset number of sample feature data items of the to-be-recognized signal and a classification label for the sample feature data items; the sample feature data items comprise a sample amplitude-and-phase feature matrix; and the cyclic residual network comprises: a plurality of gated recurrent units (GRU) configured for processing the amplitude-and-phase feature matrix.

2. The method according to claim 1 , wherein before obtaining the signal matrix of the to-be-recognized signal, the method further comprises:
receiving a plurality of to-be-recognized wireless signals; wherein the plurality of wireless signals are wireless signals received at a plurality of time points in a continuous time period; and
forming the signal matrix with the plurality of wireless signals.

3. The method according to claim 1 , wherein converting, according to the preset matrix conversion method, the real-and-imaginary-part feature matrix into the amplitude-and-phase feature matrix comprises:
converting the real-and-imaginary-part feature matrix into the amplitude-and-phase feature matrix by means of the following formula:

A
=

I
2

+

Q
2

P
=

arctan
⁢

Q
I

wherein, A represents the amplitude of the to-be-recognized signal, P represents the phase of the to-be-recognized signal, I represents the real part of the to-be-recognized signal, and Q represents the imaginary part of the to-be-recognized signal.

4. The method according to claim 1 , wherein the training process of the cyclic residual network comprises:
constructing an initial cyclic residual network; wherein, the initial cyclic residual network comprises: a feature extracting module, a feature fusion module, and a feature classification module; the feature extracting module comprises: a first convolutional layer, a first residual stack, and a second residual stack; each of the first and the second residual stacks comprises: a plurality of residual submodules, each of which comprises: a second convolutional layer, a first batch normalization (BN) layer, and a third convolutional layer; the feature fusion module is configured for performing dimension conversion on feature data output by the feature extracting module; the feature classification module comprises: a plurality of GRUs, a first fully connected (FC) layer and a classifier, each GRU of the feature classification module comprising a plurality of hidden layers and a second BN layer;
inputting the sample feature data items and a classification label for the sample feature data items into the initial cyclic residual network;
obtaining, using the initial cyclic residual network, a classification result for the sample feature data items;
calculating a loss function based on a difference between the classification result and the classification label for the sample feature data items;
minimizing the loss function to obtain a minimized loss function;
determining weight parameters for modules in the initial cyclic residual network by using the minimized loss function; and
updating, based on the weight parameters, parameters in the initial cyclic residual network to obtain the cyclic residual network by training.

5. The method according to claim 1 , wherein the modulation mode comprises one or more of: binary phase shift keying (BPSK), quaternary phase shift keying (QPSK), octal phase shift keying (8PSK), hexadecimal quadrature amplitude modulation (16QAM) and sixty-fourth quadrature amplitude modulation.

6. The method according to claim 4 , wherein the formula of the loss function is as follows:

J
(

θ
)

=

-

1
M

⁢

l
⁡
(
θ
)

=

-

1
M

⁢

∑

m
=
1

M

∑

k
=
1

K

y
k

(
m
)

⁢

log
⁡
(

h

θ
1

⁢

H
k

(

X

AP

Assignees

• Univ Beijing Posts & Telecomm

Inventors

• Feng Zhiyong
• Huang Sai
• Dai Rui
• Zhang Ping
• Ning Fan
• Zhang Qixun
• Zhang Yifan
• Wei Zhiqing

Classifications

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• H04L25/0254 Primary

  using neural network algorithms · CPC title

• G06N3/045

  Combinations of networks · CPC title