WO2021077841A1 - 一种基于循环残差网络的信号调制识别方法及装置 - Google Patents

一种基于循环残差网络的信号调制识别方法及装置 Download PDF

Info

Publication number
WO2021077841A1
WO2021077841A1 PCT/CN2020/105749 CN2020105749W WO2021077841A1 WO 2021077841 A1 WO2021077841 A1 WO 2021077841A1 CN 2020105749 W CN2020105749 W CN 2020105749W WO 2021077841 A1 WO2021077841 A1 WO 2021077841A1
Authority
WO
WIPO (PCT)
Prior art keywords
signal
matrix
identified
amplitude
residual network
Prior art date
Application number
PCT/CN2020/105749
Other languages
English (en)
French (fr)
Inventor
冯志勇
黄赛
戴蕊
宁帆
张奇勋
张轶凡
尉志青
Original Assignee
北京邮电大学
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 北京邮电大学 filed Critical 北京邮电大学
Priority to US17/440,120 priority Critical patent/US11909563B2/en
Publication of WO2021077841A1 publication Critical patent/WO2021077841A1/zh

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L25/00—Baseband systems
    • H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
    • H04L25/0202—Channel estimation
    • H04L25/024—Channel estimation channel estimation algorithms
    • H04L25/0254—Channel estimation channel estimation algorithms using neural network algorithms
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L27/00—Modulated-carrier systems
    • H04L27/0012—Modulated-carrier systems arrangements for identifying the type of modulation
    • 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

Definitions

  • This application relates to the field of wireless communication technology, and in particular to a method and device for signal modulation recognition based on a cyclic residual network.
  • the Internet of Things connects networks to things that exist in the real world, which has attracted widespread attention.
  • the development of the Internet of Things has exacerbated the problem of the shortage of spectrum resources and has led to difficulties in spectrum allocation.
  • cognitive radio is used for dynamic spectrum detection.
  • the simple spectrum sensing method aims to ensure the correct use of the spectrum of cognitive radio technology, and the modulation method recognition technology is used to identify the modulation method of the signal under noise and interference.
  • the process of adopting the existing signal modulation method to identify is: firstly, perform time-frequency analysis on the signal in the sample library, convert the time-frequency spectrum of the signal into a gray-scale image, and then use the gray-scale image to perform the deep residual network model After training, the trained deep residual network model is used to detect and identify the specific signal in the transmission process, and the specific signal is the signal for targeted training during the model training process.
  • time-frequency analysis is performed on the signal in the sample library, and the realization process of converting the time-frequency spectrum of the signal into a gray-scale image is: short-time Fourier transform is used as a time-frequency analysis method, The observation signal is intercepted by a centrally symmetric sliding window, and Fourier transform is performed on the signal in the sliding window to convert the time-spectrogram of the signal into a gray-scale image.
  • short-time Fourier transform is used as a time-frequency analysis method
  • the observation signal is intercepted by a centrally symmetric sliding window, and Fourier transform is performed on the signal in the sliding window to convert the time-spectrogram of the signal into a gray-scale image.
  • the length of the input signal in practical applications is not fixed, and in the above method of identifying the signal modulation mode, the deep residual network model trained and used is affected by the characteristics of the deep residual network model and can only be processed
  • the fixed length of the input signal makes the recognition method less flexible.
  • the purpose of the embodiments of the present application is to provide a signal modulation recognition method and device based on a cyclic residual network to solve the problems of complex input signal processing and poor recognition flexibility in the prior art.
  • the specific technical solutions are as follows:
  • an embodiment of the present application provides a signal modulation recognition method based on a cyclic residual network, the method including:
  • the signal to be identified is a signal to be modulated and identified;
  • the real part and imaginary part characteristic matrix is converted into an amplitude-phase characteristic matrix;
  • the amplitude-phase characteristic matrix carries the amplitude characteristic and phase characteristic of the signal to be identified, and the amplitude-phase characteristic matrix
  • the amount of characteristic information carried by the characteristic matrix changes with the amount of information carried by the signal to be identified;
  • the cyclic residual network is based on the preset number of sample features of the signal to be identified Data, the category label corresponding to the sample feature data is obtained through training; the sample feature data includes a sample amplitude and phase feature matrix; the cyclic residual network includes: a plurality of gate cyclic units GRU, the GRU is used to The amplitude and phase characteristic matrix is processed.
  • the method before the acquiring the signal matrix of the signal to be identified, the method further includes:
  • the multiple wireless signals are: wireless signals received at multiple time points in a continuous time period;
  • the multiple wireless signals are combined into a signal matrix.
  • the step of converting the real and imaginary feature matrix into an amplitude and phase feature matrix according to a preset matrix conversion method includes:
  • A represents the amplitude of the signal to be identified
  • P represents the phase of the signal to be identified
  • I represents the real part of the signal to be identified
  • Q represents the imaginary part of the signal to be identified.
  • the training process of the cyclic residual network includes:
  • the initial cyclic residual network includes: a feature extraction module, a feature fusion module, and a feature classification module;
  • the feature extraction module includes: a first convolutional layer, a first residual stack, and a first The second residual stack, the residual stack includes: a plurality of residual sub-modules, the residual sub-modules include: a second convolutional layer, a first batch of normalized BN layer and a third convolutional layer;
  • the feature fusion module is used to perform dimensional conversion on the feature data output by the feature extraction module;
  • the feature classification module includes: multiple GRUs, a first fully connected FC layer and a classifier, and the GRU includes multiple hidden layers And the second BN layer;
  • the parameters in the initial cyclic residual network are updated based on the weight parameters, and the cyclic residual network is obtained by training.
  • the modulation method includes: binary phase shift keying BPSK, quaternary phase shift keying QPSK, octal phase shift keying 8PSK, hexadecimal quadrature amplitude modulation 16QAM, and sixty-fourth quadrature One or more of amplitude modulation.
  • the expression of the loss function is:
  • M represents the number of the sample characteristic data
  • K represents the number of modulation methods
  • ⁇ represents all the parameters in the initial cyclic residual network
  • X AP(m) represents the m-th sample feature data
  • It represents the feature data obtained by the feature data of the m-th sample after passing through the feature extraction module
  • l( ⁇ ) represents the likelihood function of the parameter ⁇
  • H k represents the k-th modulation mode.
  • an embodiment of the present application provides a signal modulation recognition device based on a cyclic residual network, the device including:
  • the extraction module is used to obtain the signal matrix of the signal to be identified, and extract the real part information and the imaginary part information of the signal matrix; the signal to be identified is the signal to be modulated and identified;
  • a generating module configured to generate the real and imaginary feature matrix of the signal to be identified according to the extracted real and imaginary information
  • the conversion module is configured to convert the real part and imaginary part characteristic matrix into an amplitude-phase characteristic matrix according to a preset matrix conversion method; the amplitude-phase characteristic matrix carries the amplitude characteristic and phase characteristic of the signal to be identified, And the characteristic information amount carried by the amplitude and phase characteristic matrix changes with the change of the information amount carried by the signal to be identified;
  • the recognition module is used to input the amplitude and phase characteristic matrix into the pre-trained cyclic residual network to obtain the modulation mode corresponding to the signal to be recognized; the cyclic residual network is based on the pre-trained signal to be recognized It is assumed that the number of sample feature data and the category label corresponding to the sample feature data are obtained through training; the sample feature data includes a sample amplitude and phase feature matrix; the cyclic residual network includes: a plurality of gate cyclic units GRU, the GRU It is used to process the amplitude and phase characteristic matrix.
  • the device further includes:
  • the receiving module is configured to receive multiple wireless signals to be identified; the multiple wireless signals are: wireless signals received at multiple time points in a continuous time period;
  • the combination module is used to combine the multiple wireless signals into a signal matrix.
  • the conversion module is specifically used for:
  • A represents the amplitude of the signal to be identified
  • P represents the phase of the signal to be identified
  • I represents the real part of the signal to be identified
  • Q represents the imaginary part of the signal to be identified.
  • the device further includes:
  • the construction module is used to construct an initial cyclic residual network; wherein, the initial cyclic residual network includes: a feature extraction module, a feature fusion module, and a feature classification module; the feature extraction module includes: a first convolutional layer, a first A residual stack and a second residual stack, the residual stack includes: a plurality of residual sub-modules, the residual sub-modules include: a second convolutional layer, a first batch normalized BN layer, and a third volume Layers; the feature fusion module is used to perform dimensional conversion on the feature data output by the feature extraction module; the feature classification module includes: a plurality of GRUs, a first fully connected FC layer and a classifier, the GRU includes Multiple hidden layers and the second BN layer;
  • the first training module is configured to input the sample feature data and the category label corresponding to the sample feature data into the initial cyclic residual network;
  • the first obtaining module is configured to use the initial cyclic residual network to obtain a classification result corresponding to each of the sample feature data;
  • a calculation module configured to calculate a loss function based on the difference between the classification result and the category label corresponding to the sample feature data
  • the second obtaining module is used to minimize the loss function to obtain the minimized loss function
  • the determining module is used to determine the weight parameters of each module in the initial cyclic residual network according to the minimized loss function
  • the second training module is configured to update the parameters in the initial cyclic residual network based on the weight parameter, and train to obtain the cyclic residual network.
  • the modulation method includes: binary phase shift keying BPSK, quaternary phase shift keying QPSK, octal phase shift keying 8PSK, hexadecimal quadrature amplitude modulation 16QAM, and sixty-fourth quadrature One or more of amplitude modulation.
  • the expression of the loss function is:
  • M represents the number of the sample characteristic data
  • K represents the number of modulation methods
  • ⁇ represents all the parameters in the initial cyclic residual network
  • the category label representing the characteristic data of the m-th sample is the k-th modulation method
  • X AP(m) represents the characteristic data of the m-th sample
  • It represents the feature data obtained by the feature data of the m-th sample after passing through the feature extraction module
  • l( ⁇ ) represents the likelihood function of the parameter ⁇
  • H k represents the k-th modulation mode.
  • an embodiment of the present application also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, where the processor, the communication interface, and the memory complete mutual communication through the communication bus;
  • Memory used to store computer programs
  • the processor is configured to implement the method steps of the signal modulation recognition method based on the cyclic residual network described in the first aspect when executing the program stored in the memory.
  • the embodiments of the present application also provide a computer-readable storage medium that stores instructions in the computer-readable storage medium, which when run on a computer, causes the computer to execute the one described in the first aspect.
  • the embodiments of the present application also provide a computer program product containing instructions.
  • the computer program product runs on a computer, the computer executes the cycle-based residuals described in the first aspect.
  • the method steps of the network signal modulation identification method are described in the first aspect.
  • the embodiment of the application provides a signal modulation recognition method and device based on a cyclic residual network, which obtains the signal matrix of the signal to be recognized, extracts the real part information and the imaginary part information of the signal matrix, and generates the real part and imaginary part of the signal to be recognized Part characteristic matrix, according to the preset matrix conversion method, transform the real part and imaginary part characteristic matrix into the amplitude-phase characteristic matrix, and then use the amplitude-phase characteristic matrix as the input of the pre-trained cyclic residual network, and the modulation method of the signal to be identified Identify it.
  • the signal matrix of the signal to be identified is processed to obtain the amplitude and phase characteristic matrix of the signal to be identified.
  • the realization process is simple and easy to operate, and does not require complex algorithms or manual processing; and, a pre-trained cyclic residual network
  • the GRU contained in it can process the amplitude and phase characteristic matrix that the characteristic information carried by the signal to be identified changes with the change of the information carried by the signal to be identified, so that the cyclic residual network can modulate the corresponding modulation method for any length of the signal to be identified
  • the recognition flexibility is high; in addition, the pre-trained cyclic residual network is used to extract and classify the features of the signal to be recognized, which can accurately obtain the recognition result of the modulation mode of the signal to be recognized.
  • FIG. 1 is a schematic flowchart of a signal modulation recognition method based on a cyclic residual network according to an embodiment of the application;
  • FIG. 2 is a schematic flowchart of a signal matrix composition method based on the embodiment shown in FIG. 1;
  • FIG. 3 is a flowchart of a network training implementation manner provided by an embodiment of this application.
  • FIG. 4 is a schematic diagram of a cyclic residual network structure provided by an embodiment of this application.
  • FIG. 5 is a schematic structural diagram of a residual sub-module provided by an embodiment of the application.
  • FIG. 6 is a schematic diagram of another structure of a residual sub-module provided by an embodiment of the application.
  • FIG. 7 is a schematic diagram of a recognition simulation result under different signal lengths according to an embodiment of the application.
  • FIG. 8 is a schematic diagram of simulation results under different network models provided by an embodiment of this application.
  • FIG. 9 is a schematic structural diagram of a signal modulation identification device based on a cyclic residual network provided by an embodiment of the application.
  • FIG. 10 is a schematic structural diagram of another signal modulation identification device based on a cyclic residual network provided by an embodiment of the application;
  • FIG. 11 is a schematic structural diagram of an electronic device provided by an embodiment of the application.
  • an embodiment of the present application provides a signal based on a cyclic residual network. Modulation identification method, device, electronic equipment, computer readable storage medium and computer program product.
  • the signal modulation identification method based on the cyclic residual network provided by the embodiment of the present application can be applied to any electronic device that needs to identify the signal modulation mode, for example, it can be a signal receiver, a processor, and the like.
  • electronic equipment For clarity of description, hereinafter referred to as electronic equipment.
  • FIG. 1 is a schematic flowchart of a method for signal modulation identification based on a cyclic residual network provided by an embodiment of the application, and the method may include:
  • S101 Obtain a signal matrix of a signal to be identified, and extract real part information and imaginary part information of the signal matrix.
  • the electronic device may receive the signal to be identified transmitted by the base station or any signal sending device capable of transmitting signals, and perform the modulation mode identification on the received signal to be identified.
  • the signal to be identified is a signal to be modulated and identified, that is, information that needs to be modulated to be identified.
  • the signal to be identified may include, but is not limited to, a wireless digital signal, a wireless analog signal, and the like.
  • the electronic device After receiving the signal to be identified, the electronic device can obtain a signal matrix of the signal to be identified.
  • the signal matrix is used to represent the signal to be identified, and the signal to be identified may be in the form of a complex signal. Then, the electronic device can extract the real part information and the imaginary part information of the signal matrix.
  • the signal to be identified may include multiple signals, and the electronic device may extract the real part information and the imaginary part information corresponding to each signal in the signal matrix.
  • each signal in the acquired signal matrix can be represented in the form of the following complex signal:
  • f represents a signal
  • a represents the real part of the signal
  • b represents the imaginary part of the signal
  • i represents the imaginary unit
  • the real part a and the imaginary part b of each signal in the acquired signal matrix can be extracted, the real part is regarded as a row, and the imaginary part is regarded as a row, and then the real part and imaginary part characteristic matrix of the signal to be identified is generated.
  • the number of rows of the real part and imaginary part characteristic matrix corresponding to each signal is 2, then the rows of the real part and imaginary part characteristic matrix of the generated signal to be identified The number is twice the number of rows of the acquired signal matrix.
  • the amplitude and phase characteristic matrix carries the amplitude and phase characteristics of the signal to be identified, and the amount of characteristic information carried by the amplitude and phase characteristic matrix changes as the amount of information carried by the signal to be identified changes.
  • the amplitude feature describes the maximum deviation of the signal to be identified from the equilibrium position
  • the phase feature describes the measurement of the waveform change of the signal to be identified.
  • the real part information and the imaginary part information of each signal in the real and imaginary part characteristic matrix of the signal to be identified are converted into an amplitude and phase characteristic matrix containing the amplitude characteristic information and phase characteristic information of the signal.
  • the amplitude and phase characteristic matrix is the same size as the real and imaginary characteristic matrix, the number of rows is twice the modulation mode to be recognized for the received signal to be recognized, and the number of columns is the number of signals to be recognized in each modulation mode received.
  • the following expression can be used to convert the real part and imaginary part characteristic matrix into the amplitude and phase characteristic matrix:
  • A represents the amplitude of the signal to be identified
  • P represents the phase of the signal to be identified
  • I represents the real part of the signal to be identified
  • Q represents the imaginary part of the signal to be identified.
  • S104 Input the amplitude and phase characteristic matrix into the pre-trained cyclic residual network to obtain a modulation mode corresponding to the signal to be identified.
  • the obtained amplitude and phase characteristic matrix of the signal to be identified is input into a pre-trained cyclic residual network.
  • the cyclic residual network is based on a preset number of sample feature data and sample characteristics of the signal to be identified.
  • the category label corresponding to the data, the standard sample feature data corresponding to each sample feature data, and the category label corresponding to the standard sample feature data are obtained through training, and then the modulation mode corresponding to the signal to be identified is obtained.
  • the sample characteristic data corresponds to the sample amplitude and phase characteristic matrix
  • the standard sample characteristic data corresponds to the standard sample amplitude and phase characteristic matrix
  • the sample characteristic data includes a sample amplitude and phase characteristic matrix.
  • the cyclic residual network may include: multiple GRUs (Gated recurrent units). The GRU is used to process the amplitude and phase characteristic matrix, and the characteristic information of the amplitude and phase characteristic matrix can be further extracted. Specifically, the training process of the cyclic residual network is described in detail below.
  • the modulation mode corresponding to the signal to be identified may include, but is not limited to: BPSK (Binary Phase Shift Keying), QPSK (Quadrature Phase Shift Keying, quaternary phase shift keying), 8PSK (Octal phase shift keying, octal phase shift keying), 16QAM (Hexadecimal Quadrature Amplitude Modulation, hexadecimal quadrature amplitude modulation), sixty-fourth quadrature amplitude modulation, etc.
  • the embodiment of the application provides a signal modulation recognition method based on a cyclic residual network, which obtains the signal matrix of the signal to be recognized, and extracts the real part information and the imaginary part information of the signal matrix, and generates the real and imaginary part characteristics of the signal to be recognized Matrix:
  • the real part and imaginary part characteristic matrix is converted into the amplitude phase characteristic matrix, and then the amplitude and phase characteristic matrix is used as the input of the pre-trained cyclic residual network to identify the modulation mode of the signal to be identified .
  • the signal matrix of the signal to be identified is processed to obtain the amplitude and phase characteristic matrix of the signal to be identified.
  • the realization process is simple and easy to operate, and does not require complex algorithms or manual processing; and, a pre-trained cyclic residual network
  • the GRU contained in it can process the amplitude and phase characteristic matrix that the characteristic information carried by the signal to be identified changes with the change of the information carried by the signal to be identified, so that the cyclic residual network can modulate the corresponding modulation method for any length of the signal to be identified
  • the recognition flexibility is high; in addition, the pre-trained cyclic residual network is used to extract and classify the features of the signal to be recognized, which can accurately obtain the recognition result of the modulation mode of the signal to be recognized.
  • a signal modulation recognition method based on a cyclic residual network in an embodiment of the present application may further include:
  • S201 Receive multiple wireless signals to be identified.
  • multiple wireless signals to be identified may be received, and the multiple wireless signals may be: wireless signals received at multiple time points within a certain time period.
  • wireless signals in the same modulation mode or in different modulation modes may be received at multiple time points in a time period such as 30 seconds or 2 minutes.
  • any number of wireless signals in the same modulation mode can be received to identify the modulation mode of the signal in the modulation mode.
  • receiving multiple wireless signals to be identified may include: receiving multiple wireless signals in the same modulation mode within a certain period of time.
  • the implementation of the above step S103 may be: extract the real part information and the imaginary part information of each signal in the signal matrix, and then generate the to-be-signature information according to the real part information and the imaginary part information of each signal in the extracted signal matrix. Identify the real part and imaginary part feature matrix of the signal.
  • the real part and imaginary part feature matrix one line is the real part information, the other line is the imaginary part information, and each column represents a signal.
  • the size of the real part and imaginary part characteristic matrix of the generated signal to be identified is 2 rows and N columns.
  • any number of wireless signals in different modulation modes can be received to identify the modulation mode of the signal in each modulation mode.
  • receiving multiple wireless signals to be identified may include: receiving multiple wireless signals in each modulation mode within a certain period of time.
  • the implementation of the above step S103 may be: extract the real part information and imaginary part information of each signal in the signal matrix, and then generate different information according to the real part information and imaginary part information of each signal in the extracted signal matrix.
  • the size of the characteristic matrix of the real part and the imaginary part of the generated signal to be identified in the different modulation modes is 2W rows and N columns.
  • W can be any value from 1 to 10
  • N can be 128, or 256, or 512, or 1024, and so on.
  • the received multiple wireless signals are combined into a signal matrix.
  • the N wireless signals are combined into a 1 ⁇ N signal matrix, one row indicates that the modulation mode of the received wireless signal is the same modulation mode, and N columns Indicates that the number of received wireless signals is N; after receiving N wireless signals under W different modulation modes, combine the N wireless signals under W different modulation modes into a W ⁇ N signal matrix, W The row indicates that the modulation method of the received wireless signal is W modulation methods, and the N column indicates that the number of wireless signals in each modulation method received is N.
  • the W modulation modes can be completely different or partly different.
  • the foregoing steps S201 and S202 may be located before step S101 in the embodiment shown in FIG. 1, wherein the implementation process of the foregoing steps S203-206 is the same as the implementation process of the foregoing steps S101-S104, please refer to the above description. The examples are not repeated here.
  • the embodiment of the application provides a signal modulation identification method based on a cyclic residual network, which receives multiple wireless signals to be identified, combines the multiple wireless signals into a signal matrix, and then obtains the signal matrix of the signal to be identified, and extracts the signal Real part information and imaginary part information of the matrix, generate the real part and imaginary part characteristic matrix of the signal to be identified, according to the preset matrix conversion method, convert the real part and imaginary part characteristic matrix into the amplitude and phase characteristic matrix, and then the amplitude and phase characteristic matrix
  • the modulation mode of the signal to be identified is identified.
  • the signal matrix of the signal to be identified is processed to obtain the amplitude and phase characteristic matrix of the signal to be identified.
  • the realization process is simple and easy to operate, and does not require complex algorithms or manual processing; and, a pre-trained cyclic residual network
  • the GRU contained in it can process the amplitude and phase characteristic matrix that the characteristic information carried by the signal to be identified changes with the change of the information carried by the signal to be identified, so that the cyclic residual network can modulate the corresponding modulation method for any length of the signal to be identified
  • the recognition flexibility is high; in addition, the pre-trained cyclic residual network is used to extract and classify the features of the signal to be recognized, which can accurately obtain the recognition result of the modulation mode of the signal to be recognized.
  • the training process of the cyclic residual network may include:
  • the constructed initial cyclic residual network model is shown in FIG. 4, and may include: a feature extraction module, a feature fusion module 400, and a feature classification module.
  • the feature extraction module includes: a first convolutional layer 410 and multiple residual piles.
  • the multiple residual piles may include a first residual pile 411 and a second residual pile 412, where each residual pile may include: A plurality of residual sub-modules are connected, and the residual sub-module includes: a second convolutional layer, a first BN (Batch Normalization, batch normalization) layer, and a third convolutional layer that are sequentially connected in sequence.
  • the feature fusion module 400 is used to perform dimensional conversion on the feature data output by the feature extraction module.
  • the feature classification module includes multiple GRUs 413, a first FC (Fully Connected, fully connected) layer 414, and a classifier 415.
  • the GRU includes multiple hidden layers (not shown in FIG. 4) and a second BN layer (in FIG. 4). Not shown).
  • the samples used for training the cyclic residual network may be a preset number of sample feature data corresponding to the signal to be identified, and a category label corresponding to the sample feature data.
  • the method for determining the preset number of sample feature data corresponding to the signal to be identified may be: modulate the signal to be identified by using different modulation methods to obtain a preset number of signals, and then obtain the signal corresponding to the signal to be identified in different modulation methods
  • Matrix that is, the signal matrix of the aforementioned predetermined number of signals, the real part information and the imaginary part information of each signal in the signal matrix are further extracted, and then the real part and imaginary part characteristic matrix of the predetermined number of signals is generated, and The generated real part and imaginary part characteristic matrix is converted into an amplitude and phase characteristic matrix, and the amplitude and phase characteristic matrix is used as the sample characteristic data.
  • the preset number of signals in each modulation mode can be used as a sample, that is, in the sample amplitude and phase characteristic matrix corresponding to the sample characteristic data, each row represents a sample, which represents the preset in a modulation mode Data signals, the preset number can be 128, or 256, or 512, or 1024, etc. Specifically, those skilled in the art can set it according to actual needs.
  • the category labels corresponding to the sample feature data may be identifiers of different modulation methods. Specifically, numbers or characters may be used to indicate different modulation methods, that is, as identifiers of different modulation methods.
  • the training set formed by the sample feature data and the category labels corresponding to the sample feature data can be expressed as:
  • y represents the category label set corresponding to the sample feature data
  • (x AP(m) , y (m) ) represents the m-th labeled sample data set
  • x AP(m) represents the m-th Sample feature data, which can correspond to the sample amplitude and phase feature matrix
  • M represents the number of sample feature data
  • K represents the category number of the modulation method, when the category label of the m-th sample feature data x AP(m)
  • the value of is 1, and the other items are 0.
  • S303 Use the initial cyclic residual network to obtain a classification result corresponding to the feature data of each sample.
  • the preset number of sample feature data of the signal to be identified and the category label corresponding to the sample feature data are input into the initial cyclic residual network as shown in FIG. 4.
  • the convolutional layer is used to extract the feature information of the sample feature data
  • the BN layer can speed up the training speed of the entire model by reducing the dependence of the gradient on the parameters
  • the FC layer can integrate the features.
  • the size of the sample amplitude and phase feature matrix corresponding to the sample feature data is 2 ⁇ N.
  • a first ReLU Rectified Linear Unit, linear rectification function
  • the number n1 of the residual piles may be two, and the specific number of residual piles can be set by those skilled in the art according to actual needs.
  • the residual stack may include multiple residual sub-modules.
  • the number of residual sub-modules may be 4.
  • the specific number of residual sub-modules can be set by those skilled in the art according to actual needs.
  • the model of the residual sub-module can be seen in Figure 5.
  • the 16 2 ⁇ N feature matrices are sequentially input into the second convolution layer 510 with a convolution kernel size of 1 ⁇ 5 in each residual sub-module in the first residual pile, a first BN layer 511 and A second convolution layer 512 with a convolution kernel size of 1 ⁇ 5. Then input a second ReLU layer (not shown in Figure 5) and the first BN layer (not shown in Figure 5) in turn, and further extract more detailed features in the feature matrix.
  • the number of feature matrices It is 32 2 ⁇ N feature matrices.
  • the model of the residual sub-module can be as shown in Figure 6.
  • 16 2 ⁇ N feature matrices are sequentially input into the first residual pile.
  • Each residual sub-module has a convolution kernel with a size of 1 ⁇
  • a second ReLU layer 613 and a first BN layer 614 are sequentially input, and more detailed features in the feature matrix are further extracted.
  • the number of feature matrices is 32 2 ⁇ N feature matrices.
  • 64 2 ⁇ N 3-dimensional feature matrices are input to the feature fusion module, and the feature data in the feature matrix is dimensionally transformed to obtain 64 2 ⁇ N 2-dimensional feature matrices.
  • the 3-dimensional feature matrix can be expressed in the form of (x, y, z), and the feature data of the y and z dimensions are glued as one-dimensional feature data, so that the feature matrix becomes 2-dimensional, and then, Exchange the feature data of the x dimension and the glued one dimension to obtain 64 2 ⁇ N 2-dimensional feature matrices.
  • each GRU can contain 64 hidden layers and a second BN layer.
  • 64 2 ⁇ N 2-dimensional feature matrices are sequentially input into 64 hidden layers, a second BN layer, and a third ReLU layer.
  • the 64 hidden layers, a second BN layer, and a third ReLU layer perform feature extraction on the 2-dimensional feature matrix to obtain a K ⁇ 1 feature matrix, where K is the number of types of modulation methods in training.
  • the classification result corresponding to the sample characteristic data may be a category label corresponding to the sample characteristic data, and the category label may identify the type of modulation mode of the signal corresponding to the sample characteristic data, and each type corresponds to a type of modulation.
  • the category label corresponding to the category can be 1, 2, 3...K, etc.
  • the classifier may include, but is not limited to, a softmax classifier.
  • the classification result corresponding to the sample feature data output by the initial cyclic residual network can be expressed as:
  • h ⁇ ( ⁇ ) represents the cyclic residual network
  • x AP(m) represents the feature data of the m-th sample
  • It represents the feature extraction value of the k-th modulation mode (ie, the classification result corresponding to the feature data of the m-th sample)
  • ⁇ represents all the parameters in the initial cyclic residual network.
  • a loss function is calculated based on the difference between the classification result and the category label corresponding to the sample feature data.
  • the expression of the loss function used may be:
  • M represents the number of sample feature data
  • K represents the number of types of modulation methods
  • ⁇ represents all the parameters in the initial cyclic residual network.
  • the category label of the m-th sample feature data is the k-th modulation method
  • X AP(m) represents the m-th sample feature data
  • It represents the feature data obtained by the feature data of the m-th sample after the feature extraction module
  • l( ⁇ ) represents the likelihood function of the parameter ⁇
  • H k represents the k-th modulation method.
  • the loss function can be minimized by using a likelihood function method.
  • the likelihood function l( ⁇ ) can be expressed as:
  • the corresponding category label set is the joint distribution rate of y
  • x AP(m) represents the feature data of the m-th sample
  • y(m) (k) represents the category label of the m-th sample feature data as k.
  • the parameters in the initial cyclic residual network are updated based on the weight parameters, and the cyclic residual network is obtained by training.
  • the weight parameters of each module in the initial cyclic residual network are determined according to the minimized loss function, and finally, the parameters in the initial cyclic residual network are updated by using the weight parameters to train to obtain the cyclic residual network.
  • a cyclic residual network can be trained by using a gradient descent algorithm, a stochastic gradient descent algorithm, etc., which will not be repeated here in the embodiment of the application.
  • the trained cyclic residual network is tested. Taking the received signal as F sample signals of a single antenna system as an example, the F sample signals are processed to obtain their corresponding amplitude and phase. Feature matrix, input the obtained amplitude and phase feature matrix into the trained cyclic residual network for classification and identification, and define the test statistics to be used as modulation classification, expressed as:
  • the value of these K elements indicates the probability that the modulation method of the sample signal is the corresponding modulation method, and the larger the probability indicates that the modulation method of the sample signal is
  • H k represents the kth modulation mode
  • FIG. 7 is a schematic diagram of a recognition simulation result corresponding to different signal lengths according to an embodiment of the application, and FIG. 7 shows the classification of signals of different signal lengths by the method of the embodiment of the application.
  • Recognition accuracy rate where N is the signal length, that is, the number of signals included in the signal to be recognized, which are 128, 256, 512, and 1024, respectively. It can be seen that the accuracy rate increases with the increase of the signal length, which indicates that the method of the embodiment of the present application is progressive. In particular, when the signal-to-noise ratio reaches 4dB and the signal length is fixed at 128, the classification and recognition accuracy rate reaches 90%.
  • FIG. 8 is a schematic diagram of simulation results under a different network model provided by an embodiment of the application. Taking whether there is a GRU in the cyclic residual network model used in the embodiment of the application, and the received signal lengths are 256 and 512 respectively, as an example, it can be seen from Figure 8 that the performance of the cyclic residual network with a GRU layer is better and accurate The rate is higher. The performance of the cyclic residual network with GRU layer at 256 sample points is better than that of the cyclic residual network without GRU layer at 512 sample points.
  • the cyclic residual network with a GRU layer has a signal-to-noise ratio of 2dB lower than the cyclic residual network without a GRU layer and achieves a classification and recognition accuracy of 95%.
  • 256 sample points and 512 sample points indicate that the signal length is 256 and 512, respectively.
  • an embodiment of the present application provides a signal modulation recognition device based on a cyclic residual network.
  • the device may include:
  • the extraction module 401 is used to obtain a signal matrix of a signal to be identified, and extract real part information and imaginary part information of the signal matrix; the signal to be identified is a signal to be modulated and identified.
  • the generating module 402 is configured to generate the real and imaginary feature matrix of the signal to be identified according to the extracted real and imaginary information.
  • the conversion module 403 is used to convert the real part and imaginary part characteristic matrix into an amplitude-phase characteristic matrix according to a preset matrix conversion method; the amplitude-phase characteristic matrix carries the amplitude characteristic and phase characteristic of the signal to be identified, and the amplitude-phase characteristic matrix is The amount of characteristic information carried varies with the amount of information carried by the signal to be identified.
  • the recognition module 404 is used to input the amplitude and phase characteristic matrix into the pre-trained cyclic residual network to obtain the modulation mode corresponding to the signal to be recognized; the cyclic residual network is based on the preset number of sample characteristic data of the signal to be recognized, The class label corresponding to the sample feature data is obtained through training; the sample feature data includes the sample amplitude and phase feature matrix; the cyclic residual network includes: multiple gated cyclic units GRU, and the GRU is used to process the amplitude and phase feature matrix.
  • the embodiment of the application provides a signal modulation recognition device based on a cyclic residual network, which obtains the signal matrix of the signal to be recognized, and extracts the real part information and the imaginary part information of the signal matrix, and generates the real and imaginary part characteristics of the signal to be recognized Matrix:
  • the real part and imaginary part characteristic matrix is converted into the amplitude phase characteristic matrix, and then the amplitude and phase characteristic matrix is used as the input of the pre-trained cyclic residual network to identify the modulation mode of the signal to be identified .
  • the signal matrix of the signal to be identified is processed to obtain the amplitude and phase characteristic matrix of the signal to be identified.
  • the realization process is simple and easy to operate, and does not require complex algorithms or manual processing; and, pre-trained cyclic residual network
  • the GRU contained in it can process the amplitude and phase characteristic matrix that the characteristic information carried changes with the change of the information carried by the signal to be identified, so that the cyclic residual network can modulate the corresponding modulation method for any length of the signal to be identified
  • the recognition flexibility is high; in addition, using a pre-trained cyclic residual network to perform feature extraction and classification of the signal to be recognized, the recognition result of the modulation mode of the signal to be recognized can be accurately obtained.
  • the device in the embodiment of the present application is a device corresponding to the signal modulation recognition method based on the cyclic residual network shown in FIG. 1, and the signal modulation recognition method based on the cyclic residual network shown in FIG. 1 All the embodiments of the method are applicable to the device and can achieve the same beneficial effects.
  • the foregoing apparatus may further include:
  • the receiving module 501 is configured to receive multiple wireless signals to be identified; the multiple wireless signals are: wireless signals received at multiple time points in a continuous time period.
  • the combination module 502 is used to combine multiple wireless signals into a signal matrix.
  • the conversion module 403 is specifically used for:
  • A represents the amplitude of the signal to be identified
  • P represents the phase of the signal to be identified
  • I represents the real part of the signal to be identified
  • Q represents the imaginary part of the signal to be identified.
  • the foregoing device may further include:
  • the building module is used to build the initial cyclic residual network;
  • the initial cyclic residual network includes: feature extraction module, feature fusion module, feature classification module;
  • feature extraction module includes: first convolutional layer, first residual stack, The second residual stack, the residual stack includes: multiple residual sub-modules, the residual sub-modules include: the second convolutional layer, the first batch of normalized BN layer and the third convolutional layer;
  • feature fusion module using To perform dimensional conversion on the feature data output by the feature extraction module;
  • the feature classification module includes: multiple GRUs, a first fully connected FC layer and a classifier, and the GRU includes multiple hidden layers and a second BN layer.
  • the first training module is used to input the sample feature data and the category label corresponding to the sample feature data into the initial cyclic residual network.
  • the first obtaining module is used to obtain the classification result corresponding to the characteristic data of each sample by using the initial cyclic residual network.
  • the calculation module is used to calculate the loss function based on the difference between the classification result and the category label corresponding to the sample feature data.
  • the second obtaining module is used to minimize the loss function to obtain the minimized loss function.
  • the determining module is used to determine the weight parameters of each module in the initial cyclic residual network according to the minimized loss function.
  • the second training module is used to update the parameters in the initial cyclic residual network based on the weight parameters, and train to obtain the cyclic residual network.
  • the above-mentioned modulation methods may include: binary phase shift keying BPSK, quaternary phase shift keying QPSK, octal phase shift keying 8PSK, hexadecimal quadrature amplitude modulation 16QAM, and sixty-fourth quadrature One or more of amplitude modulation.
  • the expression of the above loss function can be:
  • M represents the number of sample feature data
  • K represents the number of modulation methods
  • ⁇ represents all the parameters in the initial cyclic residual network.
  • the category label of the m-th sample feature data is the k-th modulation method
  • X AP(m) represents the m-th sample feature data
  • It represents the feature data obtained by the feature data of the m-th sample after the feature extraction module
  • l( ⁇ ) represents the likelihood function of the parameter ⁇
  • H k represents the k-th modulation method.
  • the embodiment of the present application also provides an electronic device, as shown in FIG. 11, including a processor 601, a communication interface 602, a memory 603, and a communication bus 604.
  • the processor 601, the communication interface 602, and the memory 603 pass through the communication bus 604. Complete the communication between each other,
  • the memory 603 is used to store computer programs
  • the processor 601 is configured to implement the steps of a signal modulation recognition method based on a cyclic residual network provided in an embodiment of the present application when executing a program stored in the memory 603.
  • An electronic device obtains a signal matrix of a signal to be identified, extracts the real part information and imaginary part information of the signal matrix, generates a characteristic matrix of the real and imaginary parts of the signal to be identified, and transforms it according to a preset matrix
  • the real part and imaginary part feature matrix is converted into the amplitude and phase feature matrix, and then the amplitude and phase feature matrix is used as the input of the pre-trained cyclic residual network to identify the modulation mode of the signal to be identified.
  • the signal matrix of the signal to be identified is processed to obtain the amplitude and phase characteristic matrix of the signal to be identified.
  • the realization process is simple and easy to operate, and does not require complex algorithms or manual processing; and, a pre-trained cyclic residual network
  • the GRU contained in it can process the amplitude and phase characteristic matrix that the characteristic information carried by the signal to be identified changes with the change of the information carried by the signal to be identified, so that the cyclic residual network can modulate the corresponding modulation method for any length of the signal to be identified
  • the recognition flexibility is high; in addition, the pre-trained cyclic residual network is used to extract and classify the features of the signal to be recognized, which can accurately obtain the recognition result of the modulation mode of the signal to be recognized.
  • the communication bus mentioned in the above electronic device may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc.
  • PCI Peripheral Component Interconnect
  • EISA Extended Industry Standard Architecture
  • the communication bus can be divided into address bus, data bus, control bus and so on. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.
  • the communication interface is used for communication between the above-mentioned electronic device and other devices.
  • the memory may include random access memory (Random Access Memory, RAM), and may also include non-volatile memory (Non-Volatile Memory, NVM), such as at least one disk storage.
  • NVM non-Volatile Memory
  • the memory may also be at least one storage device located far away from the foregoing processor.
  • the above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (Network Processor, NP), etc.; it can also be a digital signal processor (Digital Signal Processing, DSP), a dedicated integrated Circuit (Application Specific Integrated Circuit, ASIC), Field-Programmable Gate Array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.
  • CPU central processing unit
  • NP Network Processor
  • DSP Digital Signal Processing
  • ASIC Application Specific Integrated Circuit
  • FPGA Field-Programmable Gate Array
  • FPGA Field-Programmable Gate Array
  • a computer-readable storage medium stores a computer program.
  • the computer program is executed by a processor, any one of the above-mentioned The steps of the signal modulation recognition method of the cyclic residual network to achieve the same technical effect.
  • a computer program product containing instructions, which when run on a computer, causes the computer to execute any of the above-mentioned embodiments of signal modulation based on the cyclic residual network Identify the steps of the method to achieve the same technical effect.
  • the computer may be implemented in whole or in part by software, hardware, firmware, or any combination thereof.
  • software it can be implemented in the form of a computer program product in whole or in part.
  • the computer program product includes one or more computer instructions.
  • the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.
  • the computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium.
  • the computer instructions may be transmitted from a website, computer, server, or data center.
  • the computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server or a data center integrated with one or more available media.
  • the usable medium may be a magnetic medium (for example, a floppy disk, a hard disk, and a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium (for example, a solid state disk (SSD)).

Landscapes

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

Abstract

本申请实施例提供了一种基于循环残差网络的信号调制识别方法及装置，所述方法包括：获取待识别信号的信号矩阵，提取所述信号矩阵的实部信息和虚部信息；根据所提取的实部信息和虚部信息，生成所述待识别信号的实部虚部特征矩阵；根据预设的矩阵转换方法，将所述实部虚部特征矩阵转换为幅度相位特征矩阵；将所述幅度相位特征矩阵输入预先训练好的循环残差网络中，得到与所述待识别信号对应的调制方式。本申请实施例，对待识别信号的处理过程简单、易操作，并不需要复杂的算法也不需要人工处理，识别灵活性较高，且能够准确的得到待识别信号调制方式的识别结果。

Description

一种基于循环残差网络的信号调制识别方法及装置
本申请要求于2019年10月24日提交中国专利局、申请号为201911017496.2发明名称为“一种基于循环残差网络的信号调制识别方法及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及无线通信技术领域，特别是涉及一种基于循环残差网络的信号调制识别方法及装置。
背景技术
近年来，随着物联网的高速发展，物联网将网络连接到真实世界中存在的事物，引起了广泛关注。然而，物联网发展的同时加剧了频谱资源短缺的问题，并导致了频谱分配的困难。为了缓解频谱资源短缺及频谱分配困难问题，认知无线电被用来进行动态频谱检测。简单的频谱感知方法旨在保证认知无线电技术的频谱正确使用，而调制方式识别技术被用来在噪声和干扰下识别信号的调制方式。
采用现有对信号调制方式进行识别的过程为：首先，对样本库中信号进行时频分析，将信号的时频谱图转换成灰度图像，然后，利用灰度图像对深度残差网络模型进行训练，再通过训练后的深度残差网络模型对传输过程中的特定信号进行检测识别，该特定信号即为模型训练过程中进行针对性训练的信号。
上述对信号调制方式进行识别的方法中，对样本库中信号进行时频分析，将信号的时频谱图转换成灰度图像的实现过程是：通过短时傅里叶变换作为时频分析手段，利用中心对称的滑动窗截取观测信号，对滑动窗内信号进行傅里叶变换，将信号的时频谱图转换成灰度图像，然而，发明人发现该过程的实现较为复杂，且需要人工对信号进行处理。另外，实际应用中输入信号的长度是不固定的，而上述对信号调制方式进行识别的方法中，训练及使用的深度残差网络模型，受该深度残差网络模型特性的影响，只能处理固定长度的输入信号，使得识别方法的灵活性较差。
发明内容
本申请实施例的目的在于提供一种基于循环残差网络的信号调制识别方法及装置，用以解决现有技术中对输入信号处理过程复杂，识别灵活性差的问题。具体技术方案如下：
第一方面，本申请实施例提供了一种基于循环残差网络的信号调制识别方法，所述方法包括：
获取待识别信号的信号矩阵，提取所述信号矩阵的实部信息和虚部信息；所述待识别信号为待进行调制识别的信号；
根据所提取的实部信息和虚部信息，生成所述待识别信号的实部虚部特征矩阵；
根据预设的矩阵转换方法，将所述实部虚部特征矩阵转换为幅度相位特征矩阵；所述幅度相位特征矩阵中携带有所述待识别信号的幅度特征和相位特征，且所述幅度相位特征矩阵所携带的特征信息量随所述待识别信号所携带的信息量变化而变化；
将所述幅度相位特征矩阵输入预先训练好的循环残差网络中，得到与所述待识别信号对应的调制方式；所述循环残差网络是根据所述待识别信号的预设数量个样本特征数据、所述样本特征数据对应的类别标签训练得到的；所述样本特征数据包括样本幅度相位特征矩阵；所述循环残差网络包括：多个门循环单元GRU，所述GRU用于对所述幅度相位特征矩阵进行处理。
可选地，所述获取待识别信号的信号矩阵之前，所述方法还包括：
接收待识别的多个无线信号；所述多个无线信号为：在连续的时间段内的多个时间点接收的无线信号；
将所述多个无线信号组合成信号矩阵。
可选地，所述根据预设的矩阵转换方法，将所述实部虚部特征矩阵转换为幅度相位特征矩阵的步骤，包括：
使用如下表达式，将所述实部虚部特征矩阵转换为幅度相位特征矩阵：
Figure PCTCN2020105749-appb-000001
Figure PCTCN2020105749-appb-000002
其中，A表示所述待识别信号的幅度，P表示所述待识别信号的相位，I表 示所述待识别信号的实部，Q表示所述待识别信号的虚部。
可选地，所述循环残差网络的训练过程，包括：
构建初始循环残差网络；其中，所述初始循环残差网络包括：特征提取模块、特征融合模块、特征分类模块；所述特征提取模块包括：第一卷积层、第一残差堆、第二残差堆，所述残差堆包括：多个残差子模块，所述残差子模块包括：第二卷积层、第一批量归一化BN层和第三卷积层；所述特征融合模块，用于对所述特征提取模块输出的特征数据进行维度转换；所述特征分类模块包括：多个GRU、第一全连接FC层和分类器，所述GRU包括多个隐含层和第二BN层；
将所述样本特征数据，以及所述样本特征数据对应的类别标签，输入所述初始循环残差网络；
利用所述初始循环残差网络，得到各所述样本特征数据对应的分类结果；
基于所述分类结果与所述样本特征数据对应的类别标签的差异，计算损失函数；
对损失函数进行最小化处理，得到最小化损失函数；
根据最小化损失函数，确定初始循环残差网络中各模块的权重参数；
基于所述权重参数对所述初始循环残差网络中的参数进行更新，训练得到所述循环残差网络。
可选地，所述调制方式包括：二进制相移键控BPSK、四进制相移键控QPSK、八进制相移键控8PSK、十六进制正交幅度调制16QAM及六十四进制正交幅度调制中的一种或几种。
可选地，所述损失函数的表达式为：
Figure PCTCN2020105749-appb-000003
其中，M表示所述样本特征数据的个数，K表示调制方式的个数，θ表示所述初始循环残差网络中所有的参数，
Figure PCTCN2020105749-appb-000004
表示第m个样本特征数据的类别标签为第k种调制方式，X AP(m)表示第m个样本特征数据，
Figure PCTCN2020105749-appb-000005
表示 第m个样本特征数据经过所述特征提取模块后得到的特征数据，l(θ)表示参数θ的似然函数，H k表示第k种调制方式。
第二方面，本申请实施例提供了一种基于循环残差网络的信号调制识别装置，所述装置包括：
提取模块，用于获取待识别信号的信号矩阵，提取所述信号矩阵的实部信息和虚部信息；所述待识别信号为待进行调制识别的信号；
生成模块，用于根据所提取的实部信息和虚部信息，生成所述待识别信号的实部虚部特征矩阵；
转换模块，用于根据预设的矩阵转换方法，将所述实部虚部特征矩阵转换为幅度相位特征矩阵；所述幅度相位特征矩阵中携带有所述待识别信号的幅度特征和相位特征，且所述幅度相位特征矩阵所携带的特征信息量随所述待识别信号所携带的信息量变化而变化；
识别模块，用于将所述幅度相位特征矩阵输入预先训练好的循环残差网络中，得到与所述待识别信号对应的调制方式；所述循环残差网络是根据所述待识别信号的预设数量个样本特征数据、所述样本特征数据对应的类别标签训练得到的；所述样本特征数据包括样本幅度相位特征矩阵；所述循环残差网络包括：多个门循环单元GRU，所述GRU用于对所述幅度相位特征矩阵进行处理。
可选地，所述装置还包括：
接收模块，用于接收待识别的多个无线信号；所述多个无线信号为：在连续的时间段内的多个时间点接收的无线信号；
组合模块，用于将所述多个无线信号组合成信号矩阵。
可选地，所述转换模块，具体用于：
使用如下表达式，将所述实部虚部特征矩阵转换为幅度相位特征矩阵：
Figure PCTCN2020105749-appb-000006
Figure PCTCN2020105749-appb-000007
其中，A表示所述待识别信号的幅度，P表示所述待识别信号的相位，I表 示所述待识别信号的实部，Q表示所述待识别信号的虚部。
可选地，所述装置还包括：
构建模块，用于构建初始循环残差网络；其中，所述初始循环残差网络包括：特征提取模块、特征融合模块、特征分类模块；所述特征提取模块包括：第一卷积层、第一残差堆、第二残差堆，所述残差堆包括：多个残差子模块，所述残差子模块包括：第二卷积层、第一批量归一化BN层和第三卷积层；所述特征融合模块，用于对所述特征提取模块输出的特征数据进行维度转换；所述特征分类模块包括：多个GRU、第一全连接FC层和分类器，所述GRU包括多个隐含层和第二BN层；
第一训练模块，用于将所述样本特征数据，以及所述样本特征数据对应的类别标签，输入所述初始循环残差网络；
第一获得模块，用于利用所述初始循环残差网络，得到各所述样本特征数据对应的分类结果；
计算模块，用于基于所述分类结果与所述样本特征数据对应的类别标签的差异，计算损失函数；
第二获得模块，用于对损失函数进行最小化处理，得到最小化损失函数；
确定模块，用于根据最小化损失函数，确定初始循环残差网络中各模块的权重参数；
第二训练模块，用于基于所述权重参数对所述初始循环残差网络中的参数进行更新，训练得到所述循环残差网络。
可选地，所述调制方式包括：二进制相移键控BPSK、四进制相移键控QPSK、八进制相移键控8PSK、十六进制正交幅度调制16QAM及六十四进制正交幅度调制中的一种或几种。
可选地，所述损失函数的表达式为：
Figure PCTCN2020105749-appb-000008
其中，M表示所述样本特征数据的个数，K表示调制方式的个数，θ表示所述初始循环残差网络中所有的参数，
Figure PCTCN2020105749-appb-000009
表示第m个样本特征数据的类别 标签为第k种调制方式，X AP(m)表示第m个样本特征数据，
Figure PCTCN2020105749-appb-000010
表示第m个样本特征数据经过所述特征提取模块后得到的特征数据，l(θ)表示参数θ的似然函数，H k表示第k种调制方式。
第三方面，本申请实施例还提供了一种电子设备，包括处理器、通信接口、存储器和通信总线，其中，处理器，通信接口，存储器通过通信总线完成相互间的通信；
存储器，用于存放计算机程序；
处理器，用于执行存储器上所存放的程序时，实现上述第一方面所述的一种基于循环残差网络的信号调制识别方法的方法步骤。
第四方面，本申请实施例还提供了一种计算机可读存储介质，所述计算机可读存储介质中存储有指令，当其在计算机上运行时，使得计算机执行上述第一方面所述的一种基于循环残差网络的信号调制识别方法的方法步骤。
第五方面，本申请实施例还提供了一种包含指令的计算机程序产品，当所述计算机程序产品在计算机上运行时，使得所述计算机执行上述第一方面所述的一种基于循环残差网络的信号调制识别方法的方法步骤。
本申请实施例有益效果：
本申请实施例提供的一种基于循环残差网络的信号调制识别方法及装置，获取待识别信号的信号矩阵，并提取信号矩阵的实部信息和虚部信息，生成待识别信号的实部虚部特征矩阵，根据预设的矩阵转换方法，将实部虚部特征矩阵转换为幅度相位特征矩阵，进而将幅度相位特征矩阵作为预先训练好的循环残差网络的输入，对待识别信号的调制方式进行识别。其中，对待识别信号的信号矩阵进行处理，得到该待识别信号的幅度相位特征矩阵实现过程简单、易操作，并不需要复杂的算法也不需要人工处理；并且，预先训练好的循环残差网络中包含的GRU，可以对携带的特征信息量随待识别信号所携带的信息量变化而变化的幅度相位特征矩阵进行处理，使得该循环残差网 络可以对任意长度的待识别信号对应的调制方式进行识别，识别灵活性较高；另外，使用预先训练好的循环残差网络，对待识别信号进行特征提取、分类，能够准确的得到待识别信号调制方式的识别结果。
当然，实施本申请的任一产品或方法并不一定需要同时达到以上所述的所有优点。
附图说明
为了更清楚地说明本申请实施例和现有技术的技术方案，下面对实施例和现有技术中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1为本申请实施例提供的一种基于循环残差网络的信号调制识别方法的流程示意图；
图2为基于图1所示实施例的信号矩阵组成方式的一种流程示意图；
图3为本申请实施例提供的一种网络训练实施方式流程图；
图4为本申请实施例提供的一种循环残差网络结构示意图；
图5为本申请实施例提供的一种残差子模块结构示意图；
图6为本申请实施例提供的残差子模块另一种结构示意图；
图7为本申请实施例提供的一种不同信号长度下识别仿真结果示意图；
图8为本申请实施例提供的一种不同网络模型下仿真结果示意图；
图9为本申请实施例提供的一种基于循环残差网络的信号调制识别装置的结构示意图；
图10为本申请实施例提供的另一种基于循环残差网络的信号调制识别装置的结构示意图；
图11为本申请实施例提供的一种电子设备的结构示意图。
具体实施方式
为使本申请的目的、技术方案、及优点更加清楚明白，以下参照附图并举实施例，对本申请进一步详细说明。显然，所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本 申请保护的范围。
为了解决现有对信号调制方式进行识别的方法中，对接收信号进行分析处理的过程复杂，且识别方法的灵活性较差的问题，本申请实施例提供了一种基于循环残差网络的信号调制识别方法、装置、电子设备、计算机可读存储介质及计算机程序产品。
下面首先对本申请实施例所提供的一种基于循环残差网络的信号调制识别方法进行介绍。本申请实施例所提供的一种基于循环残差网络的信号调制识别方法可以应用于任意需要对信号调制方式进行识别的电子设备，例如，可以为信号接收机、处理器等。为了描述清楚，以下称为电子设备。
如图1所示，图1为本申请实施例提供的一种基于循环残差网络的信号调制识别方法的流程示意图，该方法可以包括：
S101，获取待识别信号的信号矩阵，提取信号矩阵的实部信息和虚部信息。
本申请实施例中，电子设备可以接收基站或任何能够发送信号的信号发送设备所发射的待识别信号，并对接收的待识别信号进行调制方式的识别。该待识别信号为待进行调制识别的信号，也就是需要进行调制方式识别的信息。示例性的，该待识别信号可以包括但不限于无线数字信号、无线模拟信号等。
在接收到待识别信号后，电子设备可以获取该待识别信号的信号矩阵，信号矩阵用于表示待识别信号，该待识别信号可以是复信号的形式。然后，电子设备可以提取该信号矩阵的实部信息和虚部信息。在一种实施方式中，待识别信号可以包括多个信号，那么电子设备可以提取该信号矩阵中每一信号对应的实部信息和虚部信息。
S102，根据所提取的实部信息和虚部信息，生成待识别信号的实部虚部特征矩阵。
示例性的，所获取的信号矩阵中每一信号可以使用如下复信号的形式进行表示：
f＝a+bi
其中，f表示一个信号，a表示信号的实部，b表示信号的虚部，i表示虚数 单位。
示例性的，可以提取所获取的信号矩阵中每一信号的实部a和虚部b，将实部作为一行，虚部作为一行，进而生成待识别信号的实部虚部特征矩阵。其中，由于每一信号均具有实部a和虚部b，所以每一信号对应的实部虚部特征矩阵的行数为2，那么所生成的待识别信号的实部虚部特征矩阵的行数为所获取的信号矩阵行数的2倍。
S103，根据预设的矩阵转换方法，将实部虚部特征矩阵转换为幅度相位特征矩阵。
其中，幅度相位特征矩阵中携带有待识别信号的幅度特征和相位特征，且幅度相位特征矩阵所携带的特征信息量随待识别信号所携带的信息量变化而变化。幅度特征描述了待识别信号偏离平衡位置的最大值，相位特征描述了待识别信号波形变化的度量。
本申请实施例中，将待识别信号的实部虚部特征矩阵中每一信号的实部信息与虚部信息，转换为包含该信号的幅度特征信息和相位特征信息的幅度相位特征矩阵。该幅度相位特征矩阵与实部虚部特征矩阵的大小相同，行数为接收的待识别信号所要识别的调制方式的2倍，列数为接收的每一调制方式下待识别信号的个数。
作为本申请实施例一种实施方式，可以使用如下表达式，将实部虚部特征矩阵转换为幅度相位特征矩阵：
Figure PCTCN2020105749-appb-000011
Figure PCTCN2020105749-appb-000012
其中，A表示待识别信号的幅度，P表示待识别信号的相位，I表示待识别信号的实部，Q表示待识别信号的虚部。
S104，将幅度相位特征矩阵输入预先训练好的循环残差网络中，得到与待识别信号对应的调制方式。
本申请实施例中，将得到的待识别信号的幅度相位特征矩阵输入预先训练好的循环残差网络中，该循环残差网络，是根据待识别信号的预设数量个样本特征数据、样本特征数据对应的类别标签、各样本特征数据对应的标准 样本特征数据、以及标准样本特征数据对应的类别标签训练得到的，进而得到与待识别信号对应的调制方式。
其中，样本特征数据对应样本幅度相位特征矩阵，标准样本特征数据对应标准样本幅度相位特征矩阵。另一种实施方式中，样本特征数据包括样本幅度相位特征矩阵。该循环残差网络中可以包括：多个GRU(Gated recurrent unit，门循环单元)，GRU用于对幅度相位特征矩阵进行处理，可以进一步提取幅度相位特征矩阵的特征信息。具体的，循环残差网络的训练过程在下文详细介绍。
本申请实施例中，待识别信号对应的调制方式可以包括但不限于：BPSK(Binary Phase Shift Keying，二进制相移键控)，QPSK(Quadrature Phase Shift Keying，四进制相移键控)，8PSK(Octal phase shift keying，八进制相移键控)，16QAM(Hexadecimal Quadrature Amplitude Modulation，十六进制正交幅度调制)，六十四进制正交幅度调制等。
本申请实施例提供的一种基于循环残差网络的信号调制识别方法，获取待识别信号的信号矩阵，并提取信号矩阵的实部信息和虚部信息，生成待识别信号的实部虚部特征矩阵，根据预设的矩阵转换方法，将实部虚部特征矩阵转换为幅度相位特征矩阵，进而将幅度相位特征矩阵作为预先训练好的循环残差网络的输入，对待识别信号的调制方式进行识别。其中，对待识别信号的信号矩阵进行处理，得到该待识别信号的幅度相位特征矩阵实现过程简单、易操作，并不需要复杂的算法也不需要人工处理；并且，预先训练好的循环残差网络中包含的GRU，可以对携带的特征信息量随待识别信号所携带的信息量变化而变化的幅度相位特征矩阵进行处理，使得该循环残差网络可以对任意长度的待识别信号对应的调制方式进行识别，识别灵活性较高；另外，使用预先训练好的循环残差网络，对待识别信号进行特征提取、分类，能够准确的得到待识别信号调制方式的识别结果。
在图1所示实施例的基础上，如图2所示，本申请实施例中一种基于循环残差网络的信号调制识别方法还可以包括：
S201，接收待识别的多个无线信号。
本申请实施例中，可以接收待识别的多个无线信号，该多个无线信号可以为：在一定时间段内的多个时间点接收的无线信号。示例性的，可以在30秒或2分钟等时间段内的多个时间点接收同一调制方式或不同调制方式下的无线信号。
作为本申请实施例一种可选的实施方式，可以接收同一调制方式下的任意数量个无线信号，以对该调制方式下信号的调制方式进行识别。具体的，接收待识别的多个无线信号可以包括：接收某一时间段内的同一调制方式下的多个无线信号。
对应的，上述步骤S103的实施方式可以为：提取该信号矩阵中每一信号的实部信息和虚部信息，再根据提取的信号矩阵中每一信号的实部信息和虚部信息，生成待识别信号的实部虚部特征矩阵，该实部虚部特征矩阵中一行为实部信息，一行为虚部信息，每一列表示一个信号。示例性的，当接收同一调制方式下的N个信号时，生成的待识别信号的实部虚部特征矩阵大小为2行N列。
作为本申请实施例一种可选的实施方式，可以接收不同调制方式下的任意数量个无线信号，以对每一调制方式下信号的调制方式进行识别。具体的，接收待识别的多个无线信号可以包括：接收某一时间段内的每一调制方式下的多个无线信号。
对应的，上述步骤S103的实施方式可以为：提取该信号矩阵中每一信号的实部信息和虚部信息，再根据提取的信号矩阵中每一信号的实部信息和虚部信息，生成不同调制方式下的待识别信号的实部虚部特征矩阵，该实部虚部特征矩阵中一行为实部信息，一行为虚部信息，每一列表示一个信号。
示例性的，当接收W种不同调制方式下的N个信号时，生成的不同调制方式下的待识别信号的实部虚部特征矩阵大小为2W行N列。示例性的，W可以是1到10中任一数值，N可以是128，或256，或512，或1024，等等。
S202，将多个无线信号组合成信号矩阵。
本申请实施例中，在接收同一调制方式或不同调制方式下的多个无线信号之后，将所接收的多个无线信号组合成信号矩阵。示例性的，在接收同一调制方式下的N个无线信号之后，将该N个无线信号组合成一个1×N的信 号矩阵，1行表示接收的无线信号的调制方式是同一调制方式，N列表示接收的无线信号的数量为N个；在接收W种不同调制方式下的N个无线信号之后，将该W种不同调制方式下的N个无线信号组合成一个W×N的信号矩阵，W行表示接收的无线信号的调制方式是W种调制方式，N列表示接收的每一调制方式下无线信号的数量为N个。其中，W种调制方式可以是完全不同的，也可以是部分不同的。
上述步骤S201和S202可以位于图1所示实施例中的步骤S101之前，其中，上述步骤S203-206的实现过程分别与上述步骤S101-S104的实现过程相同，可参见上文描述，本申请实施例在此不再赘述。
本申请实施例提供的一种基于循环残差网络的信号调制识别方法，接收待识别的多个无线信号，将多个无线信号组合成信号矩阵，再获取待识别信号的信号矩阵，并提取信号矩阵的实部信息和虚部信息，生成待识别信号的实部虚部特征矩阵，根据预设的矩阵转换方法，将实部虚部特征矩阵转换为幅度相位特征矩阵，进而将幅度相位特征矩阵作为预先训练好的循环残差网络的输入，对待识别信号的调制方式进行识别。其中，对待识别信号的信号矩阵进行处理，得到该待识别信号的幅度相位特征矩阵实现过程简单、易操作，并不需要复杂的算法也不需要人工处理；并且，预先训练好的循环残差网络中包含的GRU，可以对携带的特征信息量随待识别信号所携带的信息量变化而变化的幅度相位特征矩阵进行处理，使得该循环残差网络可以对任意长度的待识别信号对应的调制方式进行识别，识别灵活性较高；另外，使用预先训练好的循环残差网络，对待识别信号进行特征提取、分类，能够准确的得到待识别信号调制方式的识别结果。
作为本申请实施例一种可选的实施方式，如图3所示，循环残差网络的训练过程，可以包括：
S301，构建初始循环残差网络。
作为本申请实施例一种可选的实施方式，所构建的初始循环残差网络模型如图4所示，可以包括：特征提取模块、特征融合模块400和特征分类模块。特征提取模块包括：第一卷积层410以及多个残差堆，多个残差堆可以 包括第一残差堆411、第二残差堆412，其中，各残差堆可以包括：依次顺序连接的多个残差子模块，残差子模块包括：依次顺序连接的第二卷积层、第一BN(Batch Normalization，批量归一化)层和第三卷积层。特征融合模块400用于对特征提取模块输出的特征数据进行维度转换。特征分类模块包括：多个GRU413、第一FC(Fully Connected，全连接)层414和分类器415，GRU包括多个隐含层(图4中未示出)和第二BN层(图4中未示出)。
S302，将样本特征数据，以及样本特征数据对应的类别标签，输入初始循环残差网络。
本申请实施例中，训练循环残差网络所采用的样本，可以是待识别信号对应的预设数量个样本特征数据，以及样本特征数据对应的类别标签。其中，待识别信号对应的预设数量个样本特征数据的确定方式可以是：采用不同调制方式对待识别信号进行调制，得到预设数量个信号，然后，获取不同调制方式下待识别信号对应的信号矩阵，也就是上述预设数量个信号的信号矩阵，进一步提取该信号矩阵中每一信号的实部信息和虚部信息，进而生成该预设数量个信号的实部虚部特征矩阵，并将生成的实部虚部特征矩阵转换为幅度相位特征矩阵，将该幅度相位特征矩阵作为样本特征数据。
示例性的，可以将每一调制方式下的预设数量个信号作为一个样本，即在样本特征数据对应的样本幅度相位特征矩阵中，每一行表示一个样本，代表一种调制方式下的预设数据个信号，该预设数量可以是128，或256，或512，或1024等等，具体的，本领域技术人员可根据实际需求进行设置。该样本特征数据对应的类别标签可以是不同的调制方式的标识，具体的，可以使用数字或字符等来表示不同的调制方式，也就是作为不同的调制方式的标识。
示例性的，样本特征数据以及样本特征数据对应的类别标签所构成的训练集可以表示为：
Figure PCTCN2020105749-appb-000013
其中，
Figure PCTCN2020105749-appb-000014
表示样本特征数据集，y表示样本特征数据对应的类别标签集，(x AP(m),y (m))表示第m个已打标签的样本数据集，x AP(m)表示第m个样本特征数据，该样本特征数据可以对应样本幅度相位特征矩阵，M表示样本特征数据的数量，
Figure PCTCN2020105749-appb-000015
表示样本特征数据x AP(m)对应的类别 标签，K表示调制方式的类别数，当第m个样本特征数据x AP(m)的类别标签
Figure PCTCN2020105749-appb-000016
的类别为k时，
Figure PCTCN2020105749-appb-000017
的值为1，其他各项为0。
S303，利用初始循环残差网络，得到各样本特征数据对应的分类结果。
示例性的，将待识别信号的预设数量个样本特征数据、样本特征数据对应的类别标签，输入如图4所示的初始循环残差网络中。其中，卷积层用于提取样本特征数据的特征信息，BN层通过降低梯度对参数的依赖程度，可加快整个模型的训练速度，FC层可对特征进行整合。
对于一个样本特征数据，该样本特征数据对应的样本幅度相位特征矩阵大小为2×N。首先，将该样本特征数据输入到一个卷积核大小为1×7的第一卷积层，再输入一个第一ReLU(Rectified Linear Unit，线性整流函数)层，对样本特征数据进行特征提取，得到16个2×N的特征矩阵。
接下来，将得到的16个2×N的特征矩阵输入残差堆中。示例性的，该残差堆的个数n1可以为2个，具体残差堆的个数本领域技术人员可根据实际需求进行设置。残差堆可以包括多个残差子模块，示例性的，该残差子模块的个数可以为4个，具体残差子模块的个数本领域技术人员可根据实际需求进行设置。
残差子模块的模型可参见图5所示。将16个2×N的特征矩阵依次输入到第一个残差堆中每一残差子模块的一个卷积核大小为1×5的第二卷积层510，一个第一BN层511和一个卷积核大小为1×5的第二卷积层512。之后再依次输入一个第二ReLU层(图5中未示出)和第一BN层(图5中未示出)，进一步提取特征矩阵中更细致的特征，经过该残差堆，特征矩阵数量为32个2×N的特征矩阵。
具体的，残差子模块的模型可以如图6所示，将16个2×N的特征矩阵依次输入到第一个残差堆中每一残差子模块的一个卷积核大小为1×5的第二卷积层610，一个第一BN层611和一个卷积核大小为1×5的第二卷积层612。之后再依次输入一个第二ReLU层613和第一BN层614，进一步提取特征矩阵中更细致的特征，经过该残差堆，特征矩阵数量为32个2×N的特征矩阵。
再将32个2×N的特征矩阵依次输入到第二个残差堆中每一残差子模块的一个卷积核大小为1×5的第二卷积层，一个第一BN层和一个卷积核大小 为1×5的第二卷积层，之后再依次输入一个第二ReLU层和第一BN层，进一步提取特征矩阵中更细致的特征，经过该残差堆，特征矩阵数量为64个2×N的特征矩阵，且特征矩阵的维数为3维。
再次，将64个2×N的3维特征矩阵输入特征融合模块，对特征矩阵中的特征数据进行维度转换，得到64个2×N的2维特征矩阵。示例性的，3维特征矩阵可以表示为(x,y,z)形式，将y和z维度的特征数据进行粘合作为一个维度的特征数据，使得特征矩阵变成2维的，然后，再将x维度和粘合后的一个维度的特征数据进行交换，得到输出64个2×N的2维特征矩阵。
最后，将64个2×N的2维特征矩阵输入GRU，示例性的，该GRU的个数n2可以为2个，具体GRU的个数本领域技术人员可根据实际需求进行设置。每个GRU可以包含64个隐含层和一个第二BN层。具体的，将64个2×N的2维特征矩阵依次输入64个隐含层、一个第二BN层和一个第三ReLU层。该64个隐含层、一个第二BN层和一个第三ReLU层对2维特征矩阵进行特征提取，得到1个K×1的特征矩阵，K为训练中调制方式的种类数。
再将1个K×1的特征矩阵输入到第一FC层，进一步对特征进行整合，进而将该特征矩阵输入到分类器中，得到样本特征数据对应的分类结果。示例性的，该样本特征数据对应的分类结果，可以是该样本特征数据对应的类别标签，该类别标签可以标识该样本特征数据所对应的信号的调制方式的种类，每一种类对应一种调制方式，该种类对应的类别标签可以是1、2、3……K等，具体形式本申请实施例在此不作限定。示例性的，分类器可以包括但不限于softmax分类器。
示例性的，初始循环残差网络输出的样本特征数据对应的分类结果可以表示为：
Figure PCTCN2020105749-appb-000018
其中，h θ(·)表示循环残差网络，x AP(m)表示第m个样本特征数据，
Figure PCTCN2020105749-appb-000019
表示第k类调制方式的特征提取值(即第m个样本特征数据对应的分类结果)，θ表示初始循环残差网络中所有的参数。
参见图3，S304，基于分类结果与样本特征数据对应的类别标签的差异， 计算损失函数。
本申请实施例中，所采用的损失函数的表达式可以为：
Figure PCTCN2020105749-appb-000020
其中，M表示样本特征数据的个数，K表示调制方式的种类数，θ表示初始循环残差网络中所有的参数，
Figure PCTCN2020105749-appb-000021
表示第m个样本特征数据的类别标签为第k种调制方式，X AP(m)表示第m个样本特征数据，
Figure PCTCN2020105749-appb-000022
表示第m个样本特征数据经过特征提取模块后得到的特征数据，l(θ)表示参数θ的似然函数，H k表示第k种调制方式。
S305，对损失函数进行最小化处理，得到最小化损失函数。
作为本申请实施例一种可选的实施方式，可以采用似然函数的方法对损失函数进行最小化处理，示例性的，似然函数l(θ)可以表示为：
Figure PCTCN2020105749-appb-000023
其中，
Figure PCTCN2020105749-appb-000024
表示样本特征数据集
Figure PCTCN2020105749-appb-000025
对应的类别标签集为y的联合分布率，x AP(m)表示第m个样本特征数据，
Figure PCTCN2020105749-appb-000026
表示第m个样本特征数据x AP(m)的类别标签，
Figure PCTCN2020105749-appb-000027
表示第k类调制方式的特征提取值，y(m) (k)表示第m个样本特征数据的类别标签为k。
S306，根据最小化损失函数，确定初始循环残差网络中各模块的权重参数。
S307，基于权重参数对初始循环残差网络中的参数进行更新，训练得到 循环残差网络。
上述步骤S306～S307中，根据最小化损失函数，确定初始循环残差网络中各模块的权重参数，最后，利用权重参数对初始循环残差网络中的参数进行更新，训练得到循环残差网络。此过程的详细实现过程，可参见现有技术的实现，例如，可以采用梯度下降算法、随机梯度下降算法等训练得到循环残差网络，本申请实施例在此不再赘述。
示例性的，本申请实施例中对所训练的循环残差网络进行测试，以接收的信号为单天线系统的F个样本信号为例，对F个样本信号进行处理，得到其对应的幅度相位特征矩阵，将得到的幅度相位特征矩阵输入训练好的循环残差网络中进行分类识别，定义测试统计量用作调制分类，表示为：
Figure PCTCN2020105749-appb-000028
其中，
Figure PCTCN2020105749-appb-000029
表示预测的结果向量，由K个元素组成，分别表示K种调制方式，这K个元素值的大小表示样本信号的调制方式为对应的调制方式的概率，概率大则表示样本信号的调制方式为该种调制方式，H k表示第k种调制方式，向量
Figure PCTCN2020105749-appb-000030
表示为第k类调制方式的测试统计量，
Figure PCTCN2020105749-appb-000031
为循环残差网络输出的结果。
可以使用如下表达式来表示识别的调制方式的结果：
Figure PCTCN2020105749-appb-000032
示例性的，如图7所示，图7为本申请实施例提供的一种不同信号长度对应的识别仿真结果示意图，图7中显示了本申请实施例的方法对不同信号长度的信号的分类识别准确率，其中，N为信号长度，也就是待识别信号包括的信号的数量，分别取128、256、512和1024。可见，该准确率随信号长度的增加而有所提升，表明了本申请实施例的方法具有渐进性。特别地，当信噪比到达4dB且信号长度固定在128时，分类识别准确率达到了90％。
示例性的，如图8所示，图8为本申请实施例提供的一种不同网络模型下仿真结果示意图。以本申请实施例所采用的循环残差网络模型中是否有GRU，接收信号长度分别为256和512为例，由图8可知，循环残差网络有GRU层时比无GRU层性能好，准确率更高。带有GRU层的循环残差网络在256样本点的情况下性能好于循环残差网络无GRU层在512样本点的情况。特别来讲，当样本点数固定到512时，带有GRU层的循环残差网络比没有GRU层的循环残差网络同达到95％的分类识别准确率时信噪比要小2dB。其中，256样本点和512样本点分别表示信号长度为256和512。
相应于上述方法实施例，本申请实施例提供了一种基于循环残差网络的信号调制识别装置，如图9所示，该装置可以包括：
提取模块401，用于获取待识别信号的信号矩阵，提取信号矩阵的实部信息和虚部信息；待识别信号为待进行调制识别的信号。
生成模块402，用于根据所提取的实部信息和虚部信息，生成待识别信号的实部虚部特征矩阵。
转换模块403，用于根据预设的矩阵转换方法，将实部虚部特征矩阵转换为幅度相位特征矩阵；幅度相位特征矩阵中携带有待识别信号的幅度特征和相位特征，且幅度相位特征矩阵所携带的特征信息量随待识别信号所携带的信息量变化而变化。
识别模块404，用于将幅度相位特征矩阵输入预先训练好的循环残差网络中，得到与待识别信号对应的调制方式；循环残差网络是根据待识别信号的预设数量个样本特征数据、样本特征数据对应的类别标签训练得到的；样本特征数据包括样本幅度相位特征矩阵；循环残差网络包括：多个门循环单元GRU，GRU用于对幅度相位特征矩阵进行处理。
本申请实施例提供的一种基于循环残差网络的信号调制识别装置，获取待识别信号的信号矩阵，并提取信号矩阵的实部信息和虚部信息，生成待识别信号的实部虚部特征矩阵，根据预设的矩阵转换方法，将实部虚部特征矩阵转换为幅度相位特征矩阵，进而将幅度相位特征矩阵作为预先训练好的循环残差网络的输入，对待识别信号的调制方式进行识别。其中，对待识别信 号的信号矩阵进行处理，得到该待识别信号的幅度相位特征矩阵实现过程简单、易操作，并不需要复杂的算法也不需要人工处理；并且，预先训练好的循环残差网络中包含的GRU，可以对携带的特征信息量随待识别信号所携带的信息量变化而变化的幅度相位特征矩阵进行处理，使得该循环残差网络可以对任意长度的待识别信号对应的调制方式进行识别，识别灵活性较高；另外，使用预先训练好的循环残差网络，对待识别信号进行特征提取、分类，能够准确的得到待识别信号调制方式的识别结果。
需要说明的是，本申请实施例的装置是与图1所示的一种基于循环残差网络的信号调制识别方法对应的装置，图1所示的一种基于循环残差网络的信号调制识别方法的所有实施例均适用于该装置，且均能达到相同的有益效果。
可选地，在图9的基础上，如图10所示，上述装置还可以包括：
接收模块501，用于接收待识别的多个无线信号；多个无线信号为：在连续的时间段内的多个时间点接收的无线信号。
组合模块502，用于将多个无线信号组合成信号矩阵。
可选地，转换模块403，具体用于：
使用如下表达式，将实部虚部特征矩阵转换为幅度相位特征矩阵：
Figure PCTCN2020105749-appb-000033
Figure PCTCN2020105749-appb-000034
其中，A表示待识别信号的幅度，P表示待识别信号的相位，I表示待识别信号的实部，Q表示待识别信号的虚部。
可选地，上述装置还可以包括：
构建模块，用于构建初始循环残差网络；其中，初始循环残差网络包括：特征提取模块、特征融合模块、特征分类模块；特征提取模块包括：第一卷积层、第一残差堆、第二残差堆，残差堆包括：多个残差子模块，残差子模块包括：第二卷积层、第一批量归一化BN层和第三卷积层；特征融合模块，用于对特征提取模块输出的特征数据进行维度转换；特征分类模块包括：多个GRU、第一全连接FC层和分类器，GRU包括多个隐含层和第二BN层。
第一训练模块，用于将样本特征数据，以及样本特征数据对应的类别标签，输入初始循环残差网络。
第一获得模块，用于利用初始循环残差网络，得到各样本特征数据对应的分类结果。
计算模块，用于基于分类结果与样本特征数据对应的类别标签的差异，计算损失函数。
第二获得模块，用于对损失函数进行最小化处理，得到最小化损失函数。
确定模块，用于根据最小化损失函数，确定初始循环残差网络中各模块的权重参数。
第二训练模块，用于基于权重参数对初始循环残差网络中的参数进行更新，训练得到循环残差网络。
可选地，上述调制方式可以包括：二进制相移键控BPSK、四进制相移键控QPSK、八进制相移键控8PSK、十六进制正交幅度调制16QAM及六十四进制正交幅度调制中的一种或几种。
可选地，上述损失函数的表达式可以为：
Figure PCTCN2020105749-appb-000035
其中，M表示样本特征数据的个数，K表示调制方式的个数，θ表示初始循环残差网络中所有的参数，
Figure PCTCN2020105749-appb-000036
表示第m个样本特征数据的类别标签为第k种调制方式，X AP(m)表示第m个样本特征数据，
Figure PCTCN2020105749-appb-000037
表示第m个样本特征数据经过特征提取模块后得到的特征数据，l(θ)表示参数θ的似然函数，H k表示第k种调制方式。
本申请实施例还提供了一种电子设备，如图11所示，包括处理器601、通信接口602、存储器603和通信总线604，其中，处理器601，通信接口602，存储器603通过通信总线604完成相互间的通信，
存储器603，用于存放计算机程序；
处理器601，用于执行存储器603上所存放的程序时，实现本申请实施例 所提供的一种基于循环残差网络的信号调制识别方法的步骤。
本申请实施例提供的一种电子设备，获取待识别信号的信号矩阵，并提取信号矩阵的实部信息和虚部信息，生成待识别信号的实部虚部特征矩阵，根据预设的矩阵转换方法，将实部虚部特征矩阵转换为幅度相位特征矩阵，进而将幅度相位特征矩阵作为预先训练好的循环残差网络的输入，对待识别信号的调制方式进行识别。其中，对待识别信号的信号矩阵进行处理，得到该待识别信号的幅度相位特征矩阵实现过程简单、易操作，并不需要复杂的算法也不需要人工处理；并且，预先训练好的循环残差网络中包含的GRU，可以对携带的特征信息量随待识别信号所携带的信息量变化而变化的幅度相位特征矩阵进行处理，使得该循环残差网络可以对任意长度的待识别信号对应的调制方式进行识别，识别灵活性较高；另外，使用预先训练好的循环残差网络，对待识别信号进行特征提取、分类，能够准确的得到待识别信号调制方式的识别结果。
上述电子设备提到的通信总线可以是外设部件互连标准(Peripheral Component Interconnect，PCI)总线或扩展工业标准结构(Extended Industry Standard Architecture，EISA)总线等。该通信总线可以分为地址总线、数据总线、控制总线等。为便于表示，图中仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。
通信接口用于上述电子设备与其他设备之间的通信。
存储器可以包括随机存取存储器(Random Access Memory，RAM)，也可以包括非易失性存储器(Non-Volatile Memory，NVM)，例如至少一个磁盘存储器。可选的，存储器还可以是至少一个位于远离前述处理器的存储装置。
上述的处理器可以是通用处理器，包括中央处理器(Central Processing Unit，CPU)、网络处理器(Network Processor，NP)等；还可以是数字信号处理器(Digital Signal Processing，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现场可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。
在本申请提供的又一实施例中，还提供了一种计算机可读存储介质，该 计算机可读存储介质内存储有计算机程序，所述计算机程序被处理器执行时实现上述任一一种基于循环残差网络的信号调制识别方法的步骤，以达到相同的技术效果。
在本申请提供的又一实施例中，还提供了一种包含指令的计算机程序产品，当其在计算机上运行时，使得计算机执行上述实施例中任一一种基于循环残差网络的信号调制识别方法的步骤，以达到相同的技术效果。
在上述实施例中，可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。所述计算机程序产品包括一个或多个计算机指令。在计算机上加载和执行所述计算机程序指令时，全部或部分地产生按照本申请实施例所述的流程或功能。所述计算机可以是通用计算机、专用计算机、计算机网络、或者其他可编程装置。所述计算机指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，所述计算机指令可以从一个网站站点、计算机、服务器或数据中心通过有线(例如同轴电缆、光纤、数字用户线(DSL))或无线(例如红外、无线、微波等)方式向另一个网站站点、计算机、服务器或数据中心进行传输。所述计算机可读存储介质可以是计算机能够存取的任何可用介质或者是包含一个或多个可用介质集成的服务器、数据中心等数据存储设备。所述可用介质可以是磁性介质，(例如，软盘、硬盘、磁带)、光介质(例如，DVD)、或者半导体介质(例如固态硬盘Solid State Disk(SSD))等。
需要说明的是，在本文中，诸如第一和第二等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者设备中还存在另外的相同要素。
本说明书中的各个实施例均采用相关的方式描述，各个实施例之间相同 相似的部分互相参见即可，每个实施例重点说明的都是与其他实施例的不同之处。尤其，对于装置、电子设备、计算机可读存储介质及计算机程序产品实施例而言，由于其基本相似于方法实施例，所以描述的比较简单，相关之处参见方法实施例的部分说明即可。
以上所述仅为本申请的较佳实施例，并不用以限制本申请，凡在本申请的精神和原则之内，所做的任何修改、等同替换、改进等，均应包含在本申请保护的范围之内。

Claims (11)

  1. 一种基于循环残差网络的信号调制识别方法，其特征在于，所述方法包括：
    获取待识别信号的信号矩阵，提取所述信号矩阵的实部信息和虚部信息；所述待识别信号为待进行调制识别的信号；
    根据所提取的实部信息和虚部信息，生成所述待识别信号的实部虚部特征矩阵；
    根据预设的矩阵转换方法，将所述实部虚部特征矩阵转换为幅度相位特征矩阵；所述幅度相位特征矩阵中携带有所述待识别信号的幅度特征和相位特征，且所述幅度相位特征矩阵所携带的特征信息量随所述待识别信号所携带的信息量变化而变化；
    将所述幅度相位特征矩阵输入预先训练好的循环残差网络中，得到与所述待识别信号对应的调制方式；所述循环残差网络是根据所述待识别信号的预设数量个样本特征数据、所述样本特征数据对应的类别标签训练得到的；所述样本特征数据包括样本幅度相位特征矩阵；所述循环残差网络包括：多个门循环单元GRU，所述GRU用于对所述幅度相位特征矩阵进行处理。
  2. 根据权利要求1所述的方法，其特征在于，所述获取待识别信号的信号矩阵之前，所述方法还包括：
    接收待识别的多个无线信号；所述多个无线信号为：在连续的时间段内的多个时间点接收的无线信号；
    将所述多个无线信号组合成信号矩阵。
  3. 根据权利要求1所述的方法，其特征在于，所述根据预设的矩阵转换方法，将所述实部虚部特征矩阵转换为幅度相位特征矩阵的步骤，包括：
    使用如下表达式，将所述实部虚部特征矩阵转换为幅度相位特征矩阵：
    Figure PCTCN2020105749-appb-100001
    Figure PCTCN2020105749-appb-100002
    其中，A表示所述待识别信号的幅度，P表示所述待识别信号的相位，I表示所述待识别信号的实部，Q表示所述待识别信号的虚部。
  4. 根据权利要求1所述的方法，其特征在于，所述循环残差网络的训练过程，包括：
    构建初始循环残差网络；其中，所述初始循环残差网络包括：特征提取模块、特征融合模块、特征分类模块；所述特征提取模块包括：第一卷积层、第一残差堆、第二残差堆，所述残差堆包括：多个残差子模块，所述残差子模块包括：第二卷积层、第一批量归一化BN层和第三卷积层；所述特征融合模块，用于对所述特征提取模块输出的特征数据进行维度转换；所述特征分类模块包括：多个GRU、第一全连接FC层和分类器，所述GRU包括多个隐含层和第二BN层；
    将所述样本特征数据，以及所述样本特征数据对应的类别标签，输入所述初始循环残差网络；
    利用所述初始循环残差网络，得到各所述样本特征数据对应的分类结果；
    基于所述分类结果与所述样本特征数据对应的类别标签的差异，计算损失函数；
    对损失函数进行最小化处理，得到最小化损失函数；
    根据最小化损失函数，确定初始循环残差网络中各模块的权重参数；
    基于所述权重参数对所述初始循环残差网络中的参数进行更新，训练得到所述循环残差网络。
  5. 根据权利要求1-4任一所述的方法，其特征在于，所述调制方式包括：二进制相移键控BPSK、四进制相移键控QPSK、八进制相移键控8PSK、十六进制正交幅度调制16QAM及六十四进制正交幅度调制中的一种或几种。
  6. 根据权利要求4所述的方法，其特征在于，所述损失函数的表达式为：
    Figure PCTCN2020105749-appb-100003
    其中，M表示所述样本特征数据的个数，K表示调制方式的个数，θ表示所述初始循环残差网络中所有的参数，
    Figure PCTCN2020105749-appb-100004
    表示第m个样本特征数据的类别标签为第k种调制方式，X AP(m)表示第m个样本特征数据，
    Figure PCTCN2020105749-appb-100005
    表示 第m个样本特征数据经过所述特征提取模块后得到的特征数据，l(θ)表示参数θ的似然函数，H k表示第k种调制方式。
  7. 一种基于循环残差网络的信号调制识别装置，其特征在于，所述装置包括：
    提取模块，用于获取待识别信号的信号矩阵，提取所述信号矩阵的实部信息和虚部信息；所述待识别信号为待进行调制识别的信号；
    生成模块，用于根据所提取的实部信息和虚部信息，生成所述待识别信号的实部虚部特征矩阵；
    转换模块，用于根据预设的矩阵转换方法，将所述实部虚部特征矩阵转换为幅度相位特征矩阵；所述幅度相位特征矩阵中携带有所述待识别信号的幅度特征和相位特征，且所述幅度相位特征矩阵所携带的特征信息量随所述待识别信号所携带的信息量变化而变化；
    识别模块，用于将所述幅度相位特征矩阵输入预先训练好的循环残差网络中，得到与所述待识别信号对应的调制方式；所述循环残差网络是根据所述待识别信号的预设数量个样本特征数据、所述样本特征数据对应的类别标签训练得到的；所述样本特征数据包括样本幅度相位特征矩阵；所述循环残差网络包括：多个门循环单元GRU，所述GRU用于对所述幅度相位特征矩阵进行处理。
  8. 根据权利要求7所述的装置，其特征在于，所述装置还包括：
    接收模块，用于接收待识别的多个无线信号；所述多个无线信号为：在连续的时间段内的多个时间点接收的无线信号；
    组合模块，用于将所述多个无线信号组合成信号矩阵。
  9. 一种电子设备，其特征在于，包括处理器、通信接口、存储器和通信总线，其中，处理器，通信接口，存储器通过通信总线完成相互间的通信；
    存储器，用于存放计算机程序；
    处理器，用于执行存储器上所存放的程序时，实现权利要求1-6任一所述的方法步骤。
  10. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质内存储有计算机程序，所述计算机程序被处理器执行时实现权利要求1-6任一 所述的方法步骤。
  11. 一种包含指令的计算机程序产品，其特征在于，当所述计算机程序产品在计算机上运行时，使得所述计算机执行权利要求1-6任一所述的方法步骤。
PCT/CN2020/105749 2019-10-24 2020-07-30 一种基于循环残差网络的信号调制识别方法及装置 WO2021077841A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/440,120 US11909563B2 (en) 2019-10-24 2020-07-30 Method and apparatus for modulation recognition of signals based on cyclic residual network

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911017496.2A CN110798417B (zh) 2019-10-24 2019-10-24 一种基于循环残差网络的信号调制识别方法及装置
CN201911017496.2 2019-10-24

Publications (1)

Publication Number Publication Date
WO2021077841A1 true WO2021077841A1 (zh) 2021-04-29

Family

ID=69441288

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/105749 WO2021077841A1 (zh) 2019-10-24 2020-07-30 一种基于循环残差网络的信号调制识别方法及装置

Country Status (3)

Country Link
US (1) US11909563B2 (zh)
CN (1) CN110798417B (zh)
WO (1) WO2021077841A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113947151A (zh) * 2021-10-20 2022-01-18 嘉兴学院 一种海上复杂环境下的无线通信信号自动调制识别方法
CN114239749A (zh) * 2021-12-23 2022-03-25 桂林电子科技大学 基于残差收缩及双向长短期记忆网络的调制识别方法
CN115392325A (zh) * 2022-10-26 2022-11-25 中国人民解放军国防科技大学 一种基于CycleGan的多特征降噪调制识别方法
CN115913850A (zh) * 2022-11-18 2023-04-04 中国电子科技集团公司第十研究所 一种基于残差网络的开集调制识别方法

Families Citing this family (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110798417B (zh) 2019-10-24 2020-07-31 北京邮电大学 一种基于循环残差网络的信号调制识别方法及装置
CN111464469B (zh) * 2020-03-12 2021-11-05 南京航空航天大学 基于神经网络的混合数字调制方式识别方法
CN115001937B (zh) * 2022-04-11 2023-06-16 北京邮电大学 面向智慧城市物联网的故障预测方法及装置
CN115442192B (zh) * 2022-07-22 2024-02-27 西安电子科技大学 一种基于主动学习的通信信号自动调制识别方法及装置
CN115238749B (zh) * 2022-08-04 2024-04-23 中国人民解放军军事科学院系统工程研究院 一种基于Transformer的特征融合的调制识别方法
CN115604061B (zh) * 2022-08-30 2024-04-09 电子科技大学 一种基于外部注意力机制的射频信号调制方式识别方法
CN115994303A (zh) * 2023-03-24 2023-04-21 中国人民解放军军事科学院国防科技创新研究院 一种残差神经网络模型及其信号调制识别方法
CN116593971A (zh) * 2023-07-13 2023-08-15 南京誉葆科技股份有限公司 一种瞬时频率特征的雷达信号调制识别方法
CN117807526B (zh) * 2023-12-29 2024-05-14 中国人民解放军军事科学院系统工程研究院 一种基于循环谱特征选择与融合机制的电磁信号识别方法
CN117857270B (zh) * 2024-03-07 2024-05-14 四川广播电视监测中心 一种无线电通信信号调制识别方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060188032A1 (en) * 2005-02-18 2006-08-24 Samsung Electronics Co., Ltd. Dual carrier modulator for modulating OFDM multi-carrier, OFDM transmitter therewith, and a method thereof
CN108521387A (zh) * 2018-03-30 2018-09-11 中国电子科技集团公司第三十六研究所 一种信号调制样式识别方法和装置
CN110048980A (zh) * 2019-04-19 2019-07-23 中国电子科技集团公司第三十六研究所 一种数字通信盲解调方法和装置
CN110071893A (zh) * 2019-05-15 2019-07-30 山东大学 基于信号空间分集的正交空间调制系统的工作方法
CN110309854A (zh) * 2019-05-21 2019-10-08 北京邮电大学 一种信号调制方式识别方法及装置
CN110798417A (zh) * 2019-10-24 2020-02-14 北京邮电大学 一种基于循环残差网络的信号调制识别方法及装置

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CA2260336A1 (en) * 1999-02-15 2000-08-15 Robert Inkol Modulation recognition system
CN107276938A (zh) * 2017-06-28 2017-10-20 北京邮电大学 一种数字信号调制方式识别方法及装置
CN108282263B (zh) * 2017-12-15 2019-11-26 西安电子科技大学 基于一维深度残差轻量网络的编码调制联合识别方法
CN108616470A (zh) * 2018-03-26 2018-10-02 天津大学 基于卷积神经网络的调制信号识别方法
CN109274625B (zh) * 2018-11-12 2020-06-19 北京邮电大学 一种信息调制方式确定方法、装置、电子设备及存储介质
CN109361635B (zh) * 2018-11-23 2021-06-04 泰山学院 基于深度残差网络的水下通信调制方式识别方法及系统

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060188032A1 (en) * 2005-02-18 2006-08-24 Samsung Electronics Co., Ltd. Dual carrier modulator for modulating OFDM multi-carrier, OFDM transmitter therewith, and a method thereof
CN108521387A (zh) * 2018-03-30 2018-09-11 中国电子科技集团公司第三十六研究所 一种信号调制样式识别方法和装置
CN110048980A (zh) * 2019-04-19 2019-07-23 中国电子科技集团公司第三十六研究所 一种数字通信盲解调方法和装置
CN110071893A (zh) * 2019-05-15 2019-07-30 山东大学 基于信号空间分集的正交空间调制系统的工作方法
CN110309854A (zh) * 2019-05-21 2019-10-08 北京邮电大学 一种信号调制方式识别方法及装置
CN110798417A (zh) * 2019-10-24 2020-02-14 北京邮电大学 一种基于循环残差网络的信号调制识别方法及装置

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113947151A (zh) * 2021-10-20 2022-01-18 嘉兴学院 一种海上复杂环境下的无线通信信号自动调制识别方法
CN113947151B (zh) * 2021-10-20 2024-05-24 嘉兴学院 一种海上复杂环境下的无线通信信号自动调制识别方法
CN114239749A (zh) * 2021-12-23 2022-03-25 桂林电子科技大学 基于残差收缩及双向长短期记忆网络的调制识别方法
CN114239749B (zh) * 2021-12-23 2024-04-05 桂林电子科技大学 基于残差收缩及双向长短期记忆网络的调制识别方法
CN115392325A (zh) * 2022-10-26 2022-11-25 中国人民解放军国防科技大学 一种基于CycleGan的多特征降噪调制识别方法
CN115392325B (zh) * 2022-10-26 2023-08-18 中国人民解放军国防科技大学 一种基于CycleGan的多特征降噪调制识别方法
CN115913850A (zh) * 2022-11-18 2023-04-04 中国电子科技集团公司第十研究所 一种基于残差网络的开集调制识别方法
CN115913850B (zh) * 2022-11-18 2024-04-05 中国电子科技集团公司第十研究所 一种基于残差网络的开集调制识别方法

Also Published As

Publication number Publication date
CN110798417B (zh) 2020-07-31
CN110798417A (zh) 2020-02-14
US11909563B2 (en) 2024-02-20
US20220191066A1 (en) 2022-06-16

Similar Documents

Publication Publication Date Title
WO2021077841A1 (zh) 一种基于循环残差网络的信号调制识别方法及装置
CN108600135B (zh) 一种信号调制方式的识别方法
CN108171203B (zh) 用于识别车辆的方法和装置
CN105608179B (zh) 确定用户标识的关联性的方法和装置
CN112508044A (zh) 人工智能ai模型的评估方法、系统及设备
CN112966712A (zh) 语言模型训练方法、装置、电子设备和计算机可读介质
WO2021051917A1 (zh) 人工智能ai模型的评估方法、系统及设备
CN110738235A (zh) 肺结核判定方法、装置、计算机设备及存储介质
WO2023143016A1 (zh) 特征提取模型的生成方法、图像特征提取方法和装置
WO2023016111A1 (zh) 键值匹配方法、装置、可读介质及电子设备
CN111784401A (zh) 下单率预测方法、装置、设备及可读存储介质
WO2022017082A1 (zh) 用于检测虚假交易订单的方法和装置
CN113781478B (zh) 油罐图像检测方法、装置、电子设备和计算机可读介质
CN113592033B (zh) 油罐图像识别模型训练方法、油罐图像识别方法和装置
CN114925720A (zh) 基于时空混合特征提取网络的小样本调制信号识别方法
WO2021143126A1 (zh) 一种调制方式确定方法及装置
CN114528876A (zh) 基于改进原型网络的小样本信号自动调制识别方法
WO2020258509A1 (zh) 终端设备异常访问的隔离方法和装置
CN112329692B (zh) 一种有限样本条件下跨场景人体行为无线感知方法及装置
WO2020215911A1 (zh) 混合信号的分类方法、装置及电子设备
CN115311664A (zh) 图像中文本类别的识别方法、装置、介质及设备
CN111310823B (zh) 目标分类方法、装置和电子系统
CN114913513A (zh) 一种公章图像的相似度计算方法、装置、电子设备和介质
CN113609018A (zh) 测试方法、训练方法、装置、设备、介质和程序产品
CN113989632A (zh) 一种遥感图像桥梁检测方法、装置、电子设备及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20879596

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20879596

Country of ref document: EP

Kind code of ref document: A1