CN113486724A - Modulation identification model based on CNN-LSTM multi-tributary structure and multiple signal representations - Google Patents

Modulation identification model based on CNN-LSTM multi-tributary structure and multiple signal representations Download PDF

Info

Publication number
CN113486724A
CN113486724A CN202110649792.5A CN202110649792A CN113486724A CN 113486724 A CN113486724 A CN 113486724A CN 202110649792 A CN202110649792 A CN 202110649792A CN 113486724 A CN113486724 A CN 113486724A
Authority
CN
China
Prior art keywords
signal
cnn
representation
model
lstm
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
CN202110649792.5A
Other languages
Chinese (zh)
Inventor
张承畅
徐余
余洒
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.)
Chongqing University of Post and Telecommunications
Original Assignee
Chongqing University of Post and Telecommunications
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 Chongqing University of Post and Telecommunications filed Critical Chongqing University of Post and Telecommunications
Priority to CN202110649792.5A priority Critical patent/CN113486724A/en
Publication of CN113486724A publication Critical patent/CN113486724A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • 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
    • G06N3/044—Recurrent networks, e.g. Hopfield 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
    • G06N3/04—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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Signal Processing (AREA)
  • Digital Transmission Methods That Use Modulated Carrier Waves (AREA)

Abstract

Aiming at the problems that the interaction between different depth learning models is mostly ignored in the existing modulation identification method based on the deep learning and the characteristics of signals can be reflected in different layers by different signal representations, the invention provides a modulation identification model based on a CNN-LSTM multi-tributary structure and various signal representations. The time domain signal in the public data set rml2016.10a is first processed into I/Q, a/phi, and the cyclic spectrogram represents the input sample as a model. Secondly, constructing a multi-branch flow model based on CNN-LSTM to extract the characteristics of different signal representations, wherein the first branch flow is responsible for extracting the characteristics of I/Q signal representation, the second branch flow is responsible for extracting the characteristics of A/phi signal representation, and the third branch flow is responsible for extracting the characteristics of cyclic spectrum signal representation. And finally, performing outer product on the features extracted from the three tributaries to obtain a feature matrix, converting the feature matrix into a feature vector through a flatten layer, and classifying by using a neural network based on softmax as a classifier. The invention fully considers the complementarity among different networks and the multi-layer characteristics reflected by different signal representations, utilizes the combination structure of the CNN and the LSTM to extract the related information of the signal space and the signal time, improves the accuracy of modulation identification, has strong realizability and can be well applied to the related engineering of a non-cooperative communication system.

Description

Modulation identification model based on CNN-LSTM multi-tributary structure and multiple signal representations
Technical Field
The invention relates to a deep learning related algorithm and a signal processing related theory, belonging to the field of communication signal processing and artificial intelligence.
Background
The development of wireless communication technology is changing greatly, and becoming an indispensable part of human daily life. Under the background of the era with increasingly prominent information demand, the communication environment becomes more and more complex, and the modulation mode becomes more and more complex and diversified. Various wireless services and wireless products are in a great variety, the signal density in the surrounding environment is greatly increased, and each signal modulation mode must be accurately identified to ensure that information can be transmitted at high speed in such a complex environment. Therefore, modulation identification of signals plays an important role in communication systems.
Conventional modulation identification techniques can be broadly classified into two categories: a maximum likelihood ratio recognition method based on hypothesis testing and a pattern recognition method based on feature extraction. The former makes different assumptions aiming at different modulation modes, then carries out likelihood function processing on signals under different assumptions, compares likelihood ratios under different assumptions, and selects the maximum value as a final judgment result. The latter is to extract proper characteristics from the signal to be identified and identify the signal modulation type according to the specific values of the characteristics, and mainly comprises two steps of characteristic extraction and identification classification. In the traditional method, the maximum likelihood ratio identification method based on hypothesis test can obtain the optimal effect theoretically, but has the defects of high calculation complexity, less identification signal quantity, over dependence on prior knowledge and the like; although the pattern recognition method based on feature extraction has great advantages, the method cannot directly process original data, needs a great amount of professional knowledge and engineering technology to manually design features, and when the channel environment is not ideal, the fuzzy features can lead to the unsatisfactory judgment result. In recent years, with the rapid development of wireless communication technology, the wireless communication environment is increasingly severe, the modulation modes of signals are more complicated and diversified, these factors bring unprecedented difficulty and challenge to automatic modulation identification, and the classification accuracy and robustness of the traditional modulation identification method reach the bottleneck period, so that the research of more and more optimal automatic modulation identification methods has a vital significance.
In recent years, the rise of big data and the rapid popularization of high-performance computing devices have promoted unprecedented development of Deep Learning (DL), and made great breakthrough in the fields of machine vision, natural language processing, economics, and the like. The idea of DL comes from deep brain tissue structures and biological neuron mathematical models, which aim to autonomously learn features at different levels from the data, extract low-level features from the raw input, and extract higher-level features based on previous-level feature representations. By the way, different modes can be automatically recognized without manually constructing features, which is also called data-driven feature extraction. Therefore, some researchers have introduced this into the field of modulation identification. Convolutional Neural Networks (CNNs), one of important models for deep learning methods, have also begun to grow in the field of modulation signal recognition, and have achieved certain results. Wang z et al use the signal constellation and a 5-level CNN to achieve QPSK, 8PSK,16QAM, and 64QAM identification. Peng S et al input the constellation of the modulated signal into AlexNet, a classical model of CNN for training, and implement the classification of QPSK, 8PSK,16QAM and 64 QAM. People such as qian liu obtain two-dimensional feature information by extracting an isocontograph of an MPSK signal cyclic spectrum, and train the two-dimensional feature by using a convolutional neural network, so that the classification of BPSK, QPSK and 8PSK is realized. Hauser S C et al propose a hierarchical deep neural network, which utilizes IQ data to obtain a signal time-frequency diagram through FFT transformation, realizes classification of 11 types of signals, and finally achieves 90% of total accuracy, but the identification effect of the network on QPSK and 8PSK is lower than 70%.
Currently, the DL-based modulation identification is based on a single representation of a signal, such as a constellation diagram, an eye diagram, and amplitude/phase, and does not consider that multiple signal representations can reflect the characteristics of the current modulation signal at different levels, and do not consider that different models have complementarity. Therefore, the invention selects the I/Q representation, the A/P representation and the cycle spectrogram representation of the signals as the input data of the CNN-LSTM multi-branch structure convolution neural network, and realizes the modulation classification of the 11 types of signals in the public data set.
Disclosure of Invention
The public data set RML2016.10a is generated by using an open source software Radio platform GNU Radio, a large number of real voice signals are adopted in the generation process, and a dynamic channel model in the GNU Radio is adopted to simulate a plurality of channel effects including frequency offset, phase offset, white Gaussian noise, frequency selective fading and the like, so that the public data set RML2016.10a is more close to an actual communication environment. Therefore, the invention aims to realize automatic modulation identification on a public data set RML2016.10a { BPSK,8PSK,16QAM,64QAM, AM-DSB, AM-SSB, CPFSK, GFSK, PAM4, QPSK and WBFM }.
Based on the common DL modulation recognition algorithm which is discussed above, the characteristics of the current modulation signals can be reflected in different levels by a plurality of characteristics are not considered, and the complementarity among different networks is not fully utilized, the invention selects the I/Q representation, the A/P representation and the cyclic spectrogram representation of the signals as the input data of the CNN-LSTM multi-branch flow network, and realizes the modulation classification of 11 types of signals in the public data set. The structure of the model of the invention is shown in figure 1.
The modulation recognition model based on the CNN-LSTM multi-tributary structure and the multiple signal representations has the following characteristics: the invention considers that different signal expressions can reflect the characteristics of the current modulation signal in different levels, and different networks can learn the characteristics of different levels, namely the convolutional neural network learning space characteristics and the long and short memory network learning time characteristics, so that the invention fully utilizes the complementarity of the two networks to construct a modulation recognition model based on the CNN-LSTM multi-tributary structure and various signal expressions, and improves the recognition rate of 11 types of signals under low signal-to-noise ratio.
The invention discloses a modulation recognition model based on CNN-LSTM and multiple signal representations, which comprises the following steps:
1) signal preprocessing: processing time domain signals in a public data set RML2016.10a into an I/Q signal representation, an A/phi signal representation and a cycle spectrogram representation as input of a network model;
2) constructing a network: constructing a multi-branch flow network model based on CNN-LSTM to identify a modulation mode;
3) sending the I/Q signal representation into a first branch of the CNN-LSTM network model to enable the model to learn characteristics;
4) feeding the A/phi signal representation into a second branch of the CNN-LSTM network model to enable the model to learn characteristics;
5) sending the representation of the cyclic spectrogram into a third branch of the CNN-LSTM network model to enable the model to learn characteristics;
6) and obtaining a characteristic matrix by the outer product of the characteristics extracted from the three tributaries, converting the characteristic matrix into a characteristic vector through a flatten layer, and finally classifying by using a neural network based on softmax as a classifier.
The signal preprocessing in the step 1) is as follows:
Figure BDA0003110686330000031
I/Q and A/phi signal representation
I/Q signal representation refers to the characterization of a signal, r, using the in-phase and quadrature components of the original signalj|→Xj I/Q(ii) a A/phi signal representation refers to representing a signal, r, using amplitude and phase information based on the original signalj|→Xj A/φThe corresponding conversion is as follows:
Figure BDA0003110686330000032
in the formula, XIRepresenting the in-phase component of the signal, XQRepresenting the quadrature component of the signal, Xj I/Q∈R2×N。
Figure BDA0003110686330000033
In the formula, XARepresenting signal amplitude information, XφRepresenting phase information of the carrying signal, Xj A/φ∈R2×N。XAAnd XφThe element (b) can be calculated by the following equation:
Figure BDA0003110686330000034
Figure BDA0003110686330000035
wherein r isInAnd rQnRespectively representing the in-phase component and the quadrature component of the signal, as shown in equation (1).
These two signal representations can learn the original characteristics of the signal directly from the original signal through the I/Q representation on the one hand and the amplitude and phase information of the signal through a/phi on the other hand. In order to more intuitively observe the different signal representations, fig. 2 gives the I/Q representation of the different modulated signals in the 6db case, and fig. 3 gives the a/P representation of the different modulated signals in the 6db case.
Figure BDA0003110686330000036
Cycle spectrogram
Signal xtThe autocorrelation function of (a) is expressed as:
Rx(t,w)＝E[x(t+w/2)]x*(t-w/2) (5)
wherein w represents a time delay, Rx(t, w) is a periodic function with a period P, Rx(t, w) Fourier transform to obtain cyclic autocorrelation function
Figure BDA0003110686330000037
Figure BDA0003110686330000041
Where α is the cycle frequency and R is the cycle frequency if and only if α ═ l/p (l is an integer)x α(w) ≠ 0, for Rx α(w) Fourier transform to obtain cyclic spectrum
Figure BDA0003110686330000042
Namely:
Figure BDA0003110686330000043
in the formula, f represents a spectrum frequency, and α represents a cycle frequency.
Signals with different modulation modes often have different cyclic spectrum characteristics, so that the signals with different modulation modes can be identified according to the different cyclic spectrum characteristics. Fig. 4 shows the cyclic spectrum of the different signals.
The network structure in the step 2) includes:
Figure BDA0003110686330000044
long and short term memory network
The core idea of Long Short-Term Memory network (LSTM) is to selectively control information flow by adding a plurality of gating structures in the network, so as to associate information Memory far away, and the basic unit structure is shown in fig. 5.
As can be seen from FIG. 5, the door control structure comprises a forgetting door ftAnd input gate itAnd an output gate ot. The forgetting gate is responsible for determining information which needs to be discarded in the current unit, the input gate is responsible for selecting information which can be added into the current unit, the output gate is responsible for determining the final output information of the LSTM unit, each gate is composed of a sigmoid function, and the function expression of the gate is as follows:
Figure BDA0003110686330000045
the specific calculation process in LSTM can be expressed as:
it＝σ(Wxixt+Whiht-1+bi) (9)
ft＝σ(Wxfxt+Whiht-1+bf) (10)
ot＝σ(Wxoxt+Whoht-1+bo) (11)
c_int＝tanh(Wxcxt+Whcht-1+bc_in) (12)
ct＝ft·ct-1+it·c_int (13)
ht＝ot·tanh(ct) (14)
in the formula, the tanh function expression is:
Figure BDA0003110686330000046
multiplication of corresponding elements, ctAre LSTM cell structure candidate states. h istFor the final output of the LSTM unit, the remaining terms are parameters that need to be learned through training.
Figure BDA0003110686330000051
Convolutional neural network
A Convolutional Neural Network (CNN) is a neural network that is used specifically to process data having a similar mesh structure. It is gradually applied to the communication field in recent years by virtue of good feature extraction characteristics. A typical CNN structure typically includes a convolutional layer, which is responsible for feature extraction, a pooling layer, which is responsible for feature dimension reduction, and a fully-connected layer, which is responsible for mapping the extracted features to the final output.
The convolutional layer is an important component of a convolutional neural network, is mainly responsible for extracting local features of image data and consists of some kernel functions, and the kernel part is convolution operation to obtain convolution features called a feature map after the convolution operation. In general, the convolutional layer output can be expressed as:
Figure BDA0003110686330000052
in the formula, XmRepresents the m-th inputA characteristic diagram is obtained, the convolution calculation is represented,
Figure BDA0003110686330000053
representing the parameters of the convolution kernel, bmIs the bias for each convolutional layer. f (·) is a nonlinear activation function, and a currently commonly used activation function is a Linear rectification unit (ReLU), whose expression is f (x) max (0, x). After the convolution feature map passes through a pooling layer, downsampling to obtain a feature map after dimensionality reduction, namely a pooling feature map. Defining down (-) as a pooling function, the output of the pooling layer can be expressed as:
Xm+1＝f down Xm (16)
in the formula, XmAnd Xm+1Respectively, a convolutional layer feature map and a pooling layer feature map, and f (-) is an activation function.
Figure BDA0003110686330000054
Multi-branch network model based on CNN-LSTM
The invention fully utilizes the characteristics of various signal representations which can represent signals from different layers and the complementarity of different networks, namely CNN learning space characteristics and LSTM learning time characteristics, establishes a modulation identification model based on CNN-LSTM and various signal representations, and the model structure is shown in figure 1.
In the model shown in fig. 1, the input to the model consists of three parts: the I/Q representation of the signal, the A/phi representation of the signal, and the cyclic spectrogram representation of the signal are called three branches of the model. The first and second tributaries are I/Q representations of the signal, the a/phi representations of the signal are respectively used for learning the characteristics of the signal under two different representations, and the two tributaries have the same structure, and both the two tributaries use the combined structure of CNN and LSTM to learn the relevant information characteristics of the signal in space and time, which are based on that the signal time shift, rotation, scale scaling and the like in the modulation process are similar to the displacement, rotation, scaling and the like of the picture in the CNN in the image classification process, and the signal time sequence can be learned through the LSTM network. The cyclic spectrum representation of the third tributary signal uses only CNN learning spatial features.
The first two tributary inputs to the model were 2 × 128, corresponding to each sample signal length in the experimental dataset, each tributary containing 3 convolutional layers Conv1, Conv2 and Conv3, 2 LSTM layers LSTM1 and LSTM2, 3 convolutional layer convolutional kernel numbers 256, 256 and 80, respectively, and corresponding convolutional kernel sizes 1 × 3, 2 × 3 and 1 × 3, respectively. After 3 convolutional layers, the time phase property of the signal is further explored by two LSTM layers, and the time phase property consists of 50 LSTM units and 25 LSTM units respectively. The third tributary of the model contains 5 convolutional layers Conv1, Conv2, Conv3, Conv4 and Conv5, the number of convolutional cores of the 5 convolutional layers is 256, 256 and 80 respectively, and the sizes of the corresponding convolutional cores are 1 × 3, 2 × 3, 1 × 3, 2 × 3 and 1 × 3 respectively. Wherein the activation function between convolutional layers uses the ReLU function and introduces the zero-padding method, 2-bit zeros are supplemented to the data boundaries before each convolutional layer to adjust the output size and furthermore, to prevent model overfitting, the Dropout method is used in convolutional and LSTM layers, where the discard rate p is 0.5.
And finally, performing outer product operation on the learned characteristics of the three branches, namely, interacting the characteristics of the signals in different expression forms to obtain a characteristic matrix containing more information, wherein a specific mathematical expression is as follows:
Figure BDA0003110686330000061
wherein f is1Representing the characteristic function of substream 1, f2Representing a characteristic function of substream 2, f3A characteristic function of the substream 3 is shown,
Figure BDA0003110686330000062
represents the outer product multiplication.
According to the model structure, the feature dimensions extracted by the three branches are all 25 × 1, a 25 × 25 × 25 dimensional feature matrix is obtained after outer product, then the feature matrix is converted into 15625 dimensional feature vectors through a flatten layer, finally, label prediction is carried out by using a neural network based on softmax as a classifier, and the output of the whole model can be expressed as:
Figure BDA0003110686330000063
in the formula, theta is a model parameter obtained by training, F (-) is based on the integral function of the modulation recognition model represented by CNN-LSTM and various signals, and a vector is output
Figure BDA0003110686330000064
Wherein the elements
Figure BDA0003110686330000065
Can be obtained from the following formula:
Figure BDA0003110686330000066
wherein x isjRepresenting input data, K representing the number of modulation classes, thetakParameters representing a k-th type of modulation.
The invention sends a plurality of representation forms of signals into a CNN-LSTM network with a plurality of tributaries, each tributary can learn different characteristics, and finally the learned characteristics of all the tributaries are obtained with richer characteristics through outer products, which is beneficial to improving the recognition rate of the model. The specific training steps of the whole model are as follows:
step 1: initialization of model parameters theta in CNN-LSTM-based multi-tributary structuretThe learning rate is set to 0.001;
step 2: randomly dividing a training set into a plurality of mini-batch;
step 3: randomly selecting a mini-batch Xb；
Step 4: mixing XbThe middle signal is represented by I/Q, A/P and a cycle spectrogram respectively and is transmitted to the model;
step 5: after a sample signal enters a model, extracting signal space characteristics and time sequence characteristics by using a CNN-LSTM combined structure for the first two branches, extracting the signal space characteristics by using a CNN for the third branch, expanding the diversity of the characteristics by outer product of the learned characteristics of the three branches, and finally obtaining a classification result through a softmax layer;
step 6: calculating the error between the predicted probability distribution and the real label distribution;
step 7: updating the parameter t by adopting a random gradient descent algorithm;
step 8: and (5) repeating the steps 2 to 7 until the model performance is not changed any more.
The step 3) is to make the first tributary learn the characteristics of the I-path Q-path signal of the signal.
The step 4) is to make the second tributary learn the amplitude and phase information of the signal.
The step 5) is to let the third tributary learn the spectral characteristics of the signal.
The step 6) is to integrate the learned characteristics of the three tributaries to enrich the characteristics and improve the recognition rate.
Drawings
FIG. 1 is a view showing a structure of a model of the present invention
FIG. 2 is an I/Q representation of different modulation signals of the present invention
FIG. 3 is an A/φ representation of different modulation signals of the present invention
FIG. 4 is a cycle profile representation of different modulation signals of the present invention
FIG. 5 is a basic unit structure of the long/short term memory network according to the present invention
Detailed Description
The invention combines the complementarity of LSTM and CNN, utilizes a plurality of signal representations, and designs a modulation identification model based on the CNN-LSTM multi-tributary structure and the plurality of signal representations. The identification model of the invention is mainly divided into: a signal preprocessing module and a CNN-LSTM model classifier module. Firstly, analysis is carried out according to the modules, and the specific steps are as follows:
step 1: and preprocessing the signal by using the formula to obtain I/Q representation, A/P representation and cycle spectrogram representation of the signal.
Step 2: and constructing a multi-branch network structure based on the CNN-LSTM, wherein the first two branches are combined with the CNN and LSTM structure, and the third branch uses the CNN structure.
And step 3: feeding the I/Q representation of the signal into the first substream for feature extraction
And 4, step 4: feeding the A/P representation of the signal into the second substream for feature extraction
And 5: feeding the cyclic spectrum representation of the signal into the third substream for feature extraction
Step 6: features learned by three tributaries are integrated by using an outer product, and classification is performed by using a neural network based on softmax as a classifier.

Claims (4)

1. The patent proposes a modulation identification model based on a CNN-LSTM multi-tributary structure and various signal representations, and the model can realize higher identification rate for a public data set RML2016.10a under low signal-to-noise ratio. Has more excellent performance than the prior human model. The patent model structure mainly comprises the following steps:
1) signal preprocessing: processing time domain signals in a public data set RML2016.10a into an I/Q signal representation, an A/phi signal representation and a cycle spectrogram representation as input of a network model;
2) constructing a network: constructing a multi-branch flow network model based on CNN-LSTM to identify a modulation mode;
3) sending the I/Q signal representation into a first branch of the CNN-LSTM network model to enable the model to learn characteristics;
4) feeding the A/phi signal representation into a second branch of the CNN-LSTM network model to enable the model to learn characteristics;
5) sending the representation of the cyclic spectrogram into a third branch of the CNN-LSTM network model to enable the model to learn characteristics;
6) and obtaining a characteristic matrix by the outer product of the characteristics extracted from the three tributaries, converting the characteristic matrix into a characteristic vector through a flatten layer, and finally classifying by using a neural network based on softmax as a classifier.
2. Based on the signal pre-processing as claimed in claim 1, it refers to I/Q signal representation, a/phi signal representation, cycle spectrogram representation. The signal preprocessing performed by the predecessor is only of a single type, the characteristic that different signal representations can reflect the modulation signals in different layers is not considered, the signal preprocessing integrates three signal representations, and the characteristics learned by a model are enriched. The three signals are represented as:
1) I/Q signal representation, A/phi signal representation
I/Q signal representation refers to the characterization of a signal, r, using the in-phase and quadrature components of the original signalj|→Xj I/Q(ii) a A/phi signal representation refers to representing a signal, r, using amplitude and phase information based on the original signalj|→Xj A/φThe corresponding conversion is as follows:
Figure FDA0003110686320000011
in the formula, XIRepresenting the in-phase component of the signal, XQRepresenting the quadrature component of the signal, Xj I/Q∈R2×N。
Figure FDA0003110686320000012
In the formula, XARepresenting signal amplitude information, XφRepresenting phase information of the carrying signal, Xj A/φ∈R2×N。XAAnd XφThe element (b) can be calculated by the following equation:
Figure FDA0003110686320000013
Figure FDA0003110686320000014
wherein r isInAnd rQnRespectively representing the in-phase component and the quadrature component of the signal, as shown in equation (1).
2) Representation of cycle spectrogram
Signal xtIs a table of autocorrelation functionsShown as follows:
Rx(t,w)＝E[x(t+w/2)]x*(t-w/2) (5)
wherein w represents a time delay, Rx(t, w) is a periodic function with a period P, Rx(t, w) Fourier transform to obtain cyclic autocorrelation function
Figure FDA0003110686320000021
Figure FDA0003110686320000022
Where α is the cycle frequency and R is the cycle frequency if and only if α ═ l/p (l is an integer)x α(w) ≠ 0, for Rx α(w) Fourier transform to obtain cyclic spectrum
Figure FDA0003110686320000023
Namely:
Figure FDA0003110686320000024
in the formula, f represents a spectrum frequency, and α represents a cycle frequency.
3. The method for constructing a CNN-LSTM-based multi-branch flow network model as claimed in claim 1, wherein the complementarity between different network models is considered, i.e. CNN extracts spatial features, LSTM extracts temporal features, the combination structure of CNN and LSTM is used to effectively extract spatial and temporal features of signals, and the diversity of features is increased by the interaction between different features, thereby improving the modulation recognition accuracy. The first and second branches adopt a combined structure, and the third branch adopts a CNN structure. And then the I/Q signal representation is sent into a first branch of the CNN-LSTM network model, the model is made to learn the characteristics, the A/phi signal representation is sent into a second branch of the CNN-LSTM network model, the model is made to learn the characteristics, and the cycle spectrogram representation is sent into a third branch of the CNN-LSTM network model, and the model is made to learn the characteristics.
4. Obtaining a feature matrix based on the feature outer product extracted from the three tributaries as claimed in claim 1, then converting the feature matrix into a feature vector through a flat layer, and finally classifying by using a neural network based on softmax as a classifier.
The method is characterized in that the learned characteristics of the three branches are integrated, the characteristic diversity is further improved, the accuracy of modulation identification is favorably improved, and the specific integration mode is as follows:
Figure FDA0003110686320000025
wherein f is1Representing the characteristic function of substream 1, f2Representing a characteristic function of substream 2, f3A characteristic function of the substream 3 is shown,
Figure FDA0003110686320000026
represents the outer product multiplication.
CN202110649792.5A 2021-06-10 2021-06-10 Modulation identification model based on CNN-LSTM multi-tributary structure and multiple signal representations Pending CN113486724A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110649792.5A CN113486724A (en) 2021-06-10 2021-06-10 Modulation identification model based on CNN-LSTM multi-tributary structure and multiple signal representations

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110649792.5A CN113486724A (en) 2021-06-10 2021-06-10 Modulation identification model based on CNN-LSTM multi-tributary structure and multiple signal representations

Publications (1)

Publication Number Publication Date
CN113486724A true CN113486724A (en) 2021-10-08

Family

ID=77934638

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110649792.5A Pending CN113486724A (en) 2021-06-10 2021-06-10 Modulation identification model based on CNN-LSTM multi-tributary structure and multiple signal representations

Country Status (1)

Country Link
CN (1) CN113486724A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114465855A (en) * 2022-01-17 2022-05-10 武汉理工大学 Attention mechanism and multi-feature fusion based automatic modulation recognition method

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109818892A (en) * 2019-01-18 2019-05-28 华中科技大学 Construct Cyclic Spectrum characteristic parameter extraction model and signal modulation mode recognition methods
CN110309797A (en) * 2019-07-05 2019-10-08 齐鲁工业大学 Merge the Mental imagery recognition methods and system of CNN-BiLSTM model and probability cooperation
CN110598677A (en) * 2019-10-08 2019-12-20 电子科技大学 Space-time multi-channel deep learning system for automatic modulation recognition
CN111510408A (en) * 2020-04-14 2020-08-07 北京邮电大学 Signal modulation mode identification method and device, electronic equipment and storage medium
CN111832417A (en) * 2020-06-16 2020-10-27 杭州电子科技大学 Signal modulation pattern recognition method based on CNN-LSTM model and transfer learning
CN112132266A (en) * 2020-09-23 2020-12-25 中国人民解放军陆军工程大学 Signal modulation identification system and modulation identification method based on convolution cycle network
CN112241724A (en) * 2020-10-30 2021-01-19 南京信息工程大学滨江学院 Automatic identification method and system based on double-path convolution long-term and short-term neural network

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109818892A (en) * 2019-01-18 2019-05-28 华中科技大学 Construct Cyclic Spectrum characteristic parameter extraction model and signal modulation mode recognition methods
CN110309797A (en) * 2019-07-05 2019-10-08 齐鲁工业大学 Merge the Mental imagery recognition methods and system of CNN-BiLSTM model and probability cooperation
CN110598677A (en) * 2019-10-08 2019-12-20 电子科技大学 Space-time multi-channel deep learning system for automatic modulation recognition
CN111510408A (en) * 2020-04-14 2020-08-07 北京邮电大学 Signal modulation mode identification method and device, electronic equipment and storage medium
CN111832417A (en) * 2020-06-16 2020-10-27 杭州电子科技大学 Signal modulation pattern recognition method based on CNN-LSTM model and transfer learning
CN112132266A (en) * 2020-09-23 2020-12-25 中国人民解放军陆军工程大学 Signal modulation identification system and modulation identification method based on convolution cycle network
CN112241724A (en) * 2020-10-30 2021-01-19 南京信息工程大学滨江学院 Automatic identification method and system based on double-path convolution long-term and short-term neural network

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
ZUFAN ZHANG: "Automatic Modulation Classification Using CNN-LSTM Based Dual-Stream Structure", 《IEEEXPLORE》 *
张斌等: "基于时频分析的深度学习调制识别算法", 《工业控制计算机》 *
翁建新等: "利用并联CNN-LSTM的调制样式识别算法", 《信号处理》 *
谭继远等: "基于卷积双向长短期神经网络的调制方式识别", 《火力与指挥控制》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114465855A (en) * 2022-01-17 2022-05-10 武汉理工大学 Attention mechanism and multi-feature fusion based automatic modulation recognition method
CN114465855B (en) * 2022-01-17 2023-09-01 武汉理工大学 Automatic modulation recognition method based on attention mechanism and multi-feature fusion

Similar Documents

Publication Publication Date Title
Zhang et al. Automatic modulation classification using CNN-LSTM based dual-stream structure
Zheng et al. Spectrum interference-based two-level data augmentation method in deep learning for automatic modulation classification
Liu et al. Deep neural network architectures for modulation classification
Li et al. Robust automated VHF modulation recognition based on deep convolutional neural networks
Wu et al. A deep learning approach for modulation recognition via exploiting temporal correlations
Chen et al. Automatic modulation classification scheme based on LSTM with random erasing and attention mechanism
Huang et al. Visualizing deep learning-based radio modulation classifier
CN110163282B (en) Modulation mode identification method based on deep learning
Yang et al. Deep learning aided method for automatic modulation recognition
CN112418014A (en) Modulation signal identification method based on wavelet transformation and convolution long-short term memory neural network
CN112069883B (en) Deep learning signal classification method integrating one-dimensional two-dimensional convolutional neural network
CN114881092A (en) Signal modulation identification method based on feature fusion
Hou et al. Electromagnetic signal feature fusion and recognition based on multi-modal deep learning
Sang et al. Application of novel architectures for modulation recognition
Lin et al. Modulation recognition using signal enhancement and multistage attention mechanism
CN112910812A (en) Modulation mode identification method for deep learning based on space-time feature extraction
CN114595729A (en) Communication signal modulation identification method based on residual error neural network and meta-learning fusion
CN113486724A (en) Modulation identification model based on CNN-LSTM multi-tributary structure and multiple signal representations
CN115982613A (en) Signal modulation identification system and method based on improved convolutional neural network
CN114615118A (en) Modulation identification method based on multi-terminal convolution neural network
CN111431825A (en) Signal automatic classification and identification method based on deep multi-flow neural network
CN112836755B (en) Sample image generation method and system based on deep learning
Qiu et al. DeepSIG: A Hybrid Heterogeneous Deep Learning Framework for Radio Signal Classification
Essai et al. Automatic Modulation Classification: Convolutional Deep Learning Neural Networks Approaches
Ma et al. Deep learning based cognitive radio modulation parameter estimation

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: 20211008