CN111695417A - Signal modulation pattern recognition method - Google Patents

Signal modulation pattern recognition method Download PDF

Info

Publication number
CN111695417A
CN111695417A CN202010360390.9A CN202010360390A CN111695417A CN 111695417 A CN111695417 A CN 111695417A CN 202010360390 A CN202010360390 A CN 202010360390A CN 111695417 A CN111695417 A CN 111695417A
Authority
CN
China
Prior art keywords
signal
training
network
signals
digital
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.)
Granted
Application number
CN202010360390.9A
Other languages
Chinese (zh)
Other versions
CN111695417B (en
Inventor
许华
苟泽中
郑万泽
冯磊
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.)
Air Force Engineering University of PLA
Original Assignee
Air Force Engineering University of PLA
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 Air Force Engineering University of PLA filed Critical Air Force Engineering University of PLA
Priority to CN202010360390.9A priority Critical patent/CN111695417B/en
Publication of CN111695417A publication Critical patent/CN111695417A/en
Application granted granted Critical
Publication of CN111695417B publication Critical patent/CN111695417B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/088—Non-supervised learning, e.g. competitive learning
    • 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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Landscapes

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

Abstract

The invention relates to a signal modulation pattern recognition method, which mainly comprises the following steps: generating training data, acquiring and processing signals, manufacturing a signal sample set, extracting neural network characteristics, extracting contrast prediction coding network characteristics, unsupervised pre-training, training a softmax classifier, and testing and adjusting network parameters; the method avoids the increase of calculated amount caused by manually constructing a signal model and signal characteristics; the identification type and the identification performance of the modulation pattern identification can be continuously improved by learning the received actual signal, so that the identification efficiency is improved; by adopting a semi-supervised learning mode, a large amount of label-free data training networks can be effectively utilized, and the modulation pattern classifier can be trained under the condition of less labeled samples.

Description

Signal modulation pattern recognition method
Technical Field
The invention belongs to the technical field of signal identification, and particularly relates to a communication signal modulation pattern identification method.
Background
The automatic modulation pattern recognition technology of communication signals is a key link between signal detection and signal demodulation and is also one of important technologies in the fields of information reconnaissance, communication countermeasure, electromagnetic spectrum monitoring and cognitive radio. The current automatic modulation pattern recognition is mainly based on a statistical mode recognition theory, characteristic parameters are modeled and calculated on known signals through short-time Fourier transform, wavelet transform, high-order cumulant and other methods, approximate ranges of the characteristic parameters corresponding to different modulation patterns are deduced, when the signals are received again, the extracted characteristics are input into a classifier and compared with preset values through preprocessing and characteristic parameter extraction, and finally the similar characteristics are output as recognition results.
The existing automatic modulation pattern recognition method constructs signal features through a method of manually designing the features according to expert experience, but the assumed conditions are looser when the features are manually designed, the difference between the assumed conditions and the actual complex electromagnetic environment is larger, the features extracted from the actual signals are difficult to be matched to a preset feature space, and the stable recognition accuracy cannot be met. And the judgment threshold of the classifier is not easy to set when the characteristics are designed manually, the identification types are few, and the research period is long, so that the real-time processing requirement cannot be met.
The defects of the traditional identification method can be avoided and solved by adopting a deep learning method based on supervised learning, but the method excessively depends on labeled data, for example, an ImageNet library in the field of image identification comprises millions of labeled data which are labeled manually; increasing the number of layers of the neural network in the deep learning method can fit a greater variety of modulation patterns, but the time consumption of training is increased.
A large amount of signal data can be obtained by medium-and-long-term reconnaissance of communication countermeasure, but the large amount of data is difficult to artificially label a large amount of signals due to factors such as high analysis difficulty, large amount, multiple signal types and the like.
Disclosure of Invention
In view of the above technical problem, the present invention provides a method for identifying a signal modulation pattern, including:
step 1: generating training data, namely generating signals of various modulation patterns by using a signal simulation platform and transmitting by using a transmitter;
step 2: signal acquisition processing, namely receiving signals of various modulation modes generated in the step A, and preprocessing the signals to obtain I, Q two paths of digital zero intermediate frequency signals;
step 3, making a signal sample set, namely performing Hilbert transform on I, Q paths of digital zero intermediate frequency signals to obtain instantaneous parameters of amplitude, phase and frequency of the digital zero intermediate frequency signals, taking the instantaneous parameters of all the digital zero intermediate frequency signals as the sample set, not processing 60 percent of samples in the sample set, and taking the instantaneous parameters as unlabeled samples for unsupervised training; labeling 20% of samples to be used as labeled training samples; labeling the rest 20% of data as a test sample;
step 4, extracting neural network characteristics, namely inputting the amplitude, phase and frequency instantaneous parameters obtained in the step 3 into the long-term and short-term memory neural network, extracting scale characteristics of the amplitude, phase and frequency instantaneous parameters at different time step lengths, and connecting the extracted scale characteristics after dimension change into a residual error neural network to extract high-dimensional characteristics;
step 5, extracting the characteristics of the contrast prediction coding network, namely inputting the high-dimensional characteristics extracted in the step 4 into an autoregressive network of the contrast prediction coding network, fitting and predicting according to the data distribution of the high-dimensional characteristics input by the autoregressive network, and constructing a contrast loss function by combining the predicted characteristics output by the autoregressive network;
step 6, unsupervised pre-training, namely using the prediction characteristics obtained in the step 5 and the high-dimensional characteristics extracted in the step 4 to make the loss function in the step E converge by using a gradient descent algorithm;
step 7, training a softmax classifier, namely replacing an autoregressive network of the contrast prediction coding with the softmax classifier, performing supervised training on the labeled sample obtained in the step B, and training softmax by using a cross entropy loss function;
step 8, testing and adjusting network parameters, namely testing the accuracy of the softmax classifier obtained by training by using the test sample obtained in the step 2, and optimizing the network by adjusting the network hyper-parameter so as to achieve the maximum recognition performance of the softmax classifier;
the method avoids the increase of calculated amount caused by manually constructing a signal model and signal characteristics; the identification type and the identification performance of the modulation pattern identification can be continuously improved by learning the received actual signal, so that the identification efficiency is improved; by adopting a semi-supervised learning mode, a large amount of label-free data training networks can be effectively utilized, and the modulation pattern classifier can be trained under the condition of less labeled samples.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a sample flow diagram of the present invention;
FIG. 3 is a diagram of a contrast predictive coding neural network according to the present invention;
FIG. 4 is a schematic diagram of a signal modulation scheme recognition apparatus according to the present invention;
FIG. 5 is a block diagram of a signal modulation scheme identifying apparatus according to an embodiment of the present invention;
fig. 6 is a graph of the identification accuracy of the method of the present invention for 11 modulation pattern signals.
Detailed Description
The invention is further described with reference to the following figures and examples.
The invention provides a signal modulation pattern recognition method, which aims at solving the problem that a communication signal modulation pattern is difficult to recognize under the condition that a training network is lack of a labeled signal. The technical idea is to combine an unsupervised learning framework with a supervised learning framework, and an unsupervised part obtains data characteristics through signal acquisition and made label-free data fitting so as to initialize the data characteristics and network parameters, thereby reducing the search range of the characteristic space in the supervised learning stage. And then mapping the features extracted by the supervised learning to corresponding label categories by using the labeled data, namely training to obtain a modulation signal classifier, and classifying the modulation signals by using the classifier.
The specific steps of the invention are shown in figure 1 and comprise:
generating training data, namely generating signals of various modulation styles by using a signal simulation platform and transmitting the signals by using a transmitter;
step 2: signal acquisition processing, namely receiving signals of various modulation patterns generated in the step 1, and preprocessing the signals to obtain I, Q two paths of digital zero intermediate frequency signals;
and step 3: making a signal sample set, namely performing Hilbert transform on I, Q paths of digital zero intermediate frequency signals to obtain instantaneous parameters of amplitude, phase and frequency of the digital zero intermediate frequency signals, taking the instantaneous parameters of all the digital zero intermediate frequency signals as a sample set, not processing 60% of samples in the sample set, and using the instantaneous parameters as unlabeled samples for unsupervised training; labeling 20% of samples to be used as supervised training samples; labeling the rest 20% of data as a test sample;
and 4, step 4: extracting neural network characteristics, namely inputting the amplitude, phase and frequency instantaneous parameters obtained in the step (3) into a long-term and short-term memory neural network, extracting scale characteristics of the amplitude, phase and frequency instantaneous parameters at different time step lengths, and connecting the extracted scale characteristics after dimension change into a residual error neural network to extract high-dimensional characteristics;
and 5: extracting the characteristics of the contrast prediction coding network, namely inputting the high-dimensional characteristics extracted in the step (4) into an autoregressive network of the contrast prediction coding network, fitting and predicting according to data distribution of the high-dimensional characteristics input by the autoregressive network, and constructing a contrast loss function by combining the predicted characteristics output by the autoregressive network;
step 6: unsupervised pre-training, namely, using the prediction features obtained in the step 5 and the high-dimensional features extracted in the step 4, and using a gradient descent algorithm to make the loss function in the step 5 converge;
and 7: training a softmax classifier, namely replacing an autoregressive network for comparing and predicting codes with the softmax classifier, performing supervised training by using the labeled samples obtained in the step (2), and training the softmax classifier by using a cross entropy loss function;
and 8: testing and adjusting network parameters, namely testing the accuracy of the softmax classifier obtained by training by using the test sample obtained in the step 2, and optimizing the network by adjusting the network hyper-parameter so as to achieve the maximum recognition performance of the softmax classifier;
the step 1 specifically comprises the following steps:
step 1-1: setting a modulation mode, a carrier frequency and a baud rate of a signal to be transmitted on a signal simulation platform;
step 1-2: processing the baseband signal by using a shaping filter, and performing interpolation up-sampling on the baseband signal by using the FPGA to obtain an intermediate frequency signal;
step 1-3: D/A conversion is carried out on the intermediate frequency signal to obtain an analog signal;
step 1-4: amplifying the band-pass filtered radio frequency signal by using a signal amplifier and transmitting the radio frequency signal by using a transmitting antenna;
the step 2 specifically comprises the following steps:
step 2-1: receiving the transmitted signal by using a receiving antenna;
step 2-2: performing band-pass filtering on the received signal by using a band-pass filter;
step 2-3: amplifying the band-pass filtered radio frequency signal by using a signal amplifier;
step 2-4: mixing the amplified radio frequency signal with a signal generated by a local oscillator by using a mixer to generate an intermediate frequency signal;
step 2-5: A/D sampling is carried out on the mixed intermediate frequency signal to obtain a digital signal;
step 2-6: performing digital low-pass filtering on the digital signal;
step 2-7: sending the signals subjected to digital low-pass filtering into a down converter to generate I, Q two paths of digital zero intermediate frequency signals;
the step 3 specifically comprises the following steps:
step 3-1: calculating instantaneous parameters of amplitude, phase and frequency of the sample by using the I, Q two paths of digital zero intermediate frequency signals obtained in the step 2-7;
step 3-2, manually labeling 40% of the instantaneous parameter samples of amplitude, phase and frequency obtained in the step 3-1 with the corresponding modulation pattern and signal-to-noise ratio of the instantaneous parameter as the label of the sample, namely labeled samples, and using the rest 60% of the samples as unlabeled training samples for unsupervised training;
and 3-3, taking 50% of the labeled samples obtained in the step 3-2 as labeled training samples, and taking the other half as test samples.
The step 4 specifically comprises the following steps:
step 4-1: taking the amplitude, phase and frequency instantaneous parameters obtained in the step (3) as input tensors, and sending the input tensors into the input end of the long-term and short-term memory neural network for training;
step 4-2: extracting scale features of the input tensor under different time step lengths by using a long-short term memory neural network;
step 4-3: carrying out dimension expansion on the scale features extracted by the medium-long short-term memory neural network in the step 4-2 to enable the scale features to adapt to the input dimension of the cascaded deep residual error network;
step 4-4, sending the dimension expanded scale features into a deep residual error network, and extracting high-dimensional features to be recorded as Z;
the step 5 specifically comprises the following steps:
step 5-1: selecting half of the high-dimensional feature Z' to input into an autoregressive network by using the high-dimensional feature Z extracted in the step 4-4, and training the autoregressive network;
step 5-2: the autoregressive network constructs data representation of a training sample by learning data distribution of Z', and carries out multi-step prediction to obtain a prediction characteristic P of the autoregressive network;
the step 6 specifically comprises the following steps:
step 6-1: constructing a contrast prediction loss function L by using the prediction feature P of the step 5-2 and the high-dimensional feature Z extracted in the step 4-4NThe following formula:
Figure BDA0002474843090000051
wherein
Figure BDA0002474843090000052
Is the predicted value of the t + k position of the autoregressive network,
Figure BDA0002474843090000053
the predicted characteristics of the autoregressive network at the position t + k are obtained, and t and k are independent variables of the high-dimensional characteristic Z. { zl} tableA set of features extracted by a neural network in a batch of samples, p being a sign of a conditional probability;
step 6-2, performing unsupervised pre-training by using the unlabeled training sample in the step 3, and converging the loss function obtained in the step 6-1 by using a gradient descent algorithm to make the weight of the long-short term memory-residual error neural network in the step 4 converge in an optimal solution space range;
the step 7 specifically comprises the following steps:
step 7-1: replacing the autoregressive network in the step 5 with a softmax network;
step 7-2: training a softmax classifier by using the labeled training sample obtained in the step 3-3;
the step 8 specifically comprises:
step 8-1: carrying out accuracy rate test on the softmax by using the test sample obtained in the step 3-3;
step 8-2: the convergence rate of the model is adjusted by adjusting the learning rate of training, the number of samples selected by one-time training and other hyperparameters, and the parameters of the neural network are adjusted according to the convergence condition of the trained loss curve, so that the model achieves the optimal feature extraction effect.
Through the processes of training a model, testing accuracy, adjusting model parameters, training the model and testing, the optimal model is finally obtained.
In a specific embodiment of the present invention, a hardware Radio transceiver USRP (universal software Radio Peripheral) is used, a GNURadio software framework is installed on a ubutu 18.04 operating system platform of a host as a signal generation environment, 1000 signal samples are generated for each type of signal under different signal-to-noise ratios with an interval of 2dB from-4 dB to 18dB, a carrier frequency is 200MHz, and a baud rate is 10M; the data samples are in a 128 x 3 format, where 128 is expressed as the sequence length of the sample and 3 is expressed as instantaneous amplitude, instantaneous phase, instantaneous frequency; then passing through three layers of long-term and short-term memory neural networks; then, a residual error neural network consisting of 5 residual error blocks is used; in the training stage, the long-term and short-term memory neural networks and the residual error network are pre-trained unsupervised by comparing the predictive coding neural network, and finally, the long-term and short-term memory neural networks and the residual error network are classified with softmax through a full link layer.
Step 1 is to generate training data, and specifically comprises:
step 1-1, setting carrier frequency 200MHz, modulation style and baud rate 10M of a signal to be generated on a signal simulation platform;
step 1-2, performing interpolation processing on a baseband signal by using an FPGA of a wireless transmitter;
step 1-3, performing digital-to-analog conversion on a signal by using a 16-bit D/A device of a wireless transmitter to generate an analog signal;
step 1-4, amplifying the signal at the radio frequency end and radiating the signal by using a transmitting antenna;
step 2 is signal acquisition and processing, and specifically comprises the following steps:
step 2-1, after receiving the signal, down-converting the signal to a baseband, then filtering out the out-of-band signal with a low-pass filter, and then performing a/D sampling with a number of samples N, for example, N equals 128, to obtain a digital complex baseband signal r (k), k equals 1,2, …, 128.
Step 2-2, performing band-pass filtering on the received signal by using a band-pass filter with the bandwidth of 30MHz and the central frequency of 200 MHz;
step 2-3: compensating and amplifying the band-pass filtered radio frequency signal by using a signal amplifier so as to achieve the working dynamic range of the receiver;
step 2-4: mixing the amplified radio frequency signal with a signal generated by a local oscillator by using a mixer to generate an intermediate frequency signal;
step 2-5: A/D sampling is carried out on the mixed intermediate frequency signal to obtain a digital signal;
step 2-6: performing digital low-pass filtering on the digital signal;
step 2-7: sending the signals subjected to digital low-pass filtering into a down converter to generate I, Q two paths of digital zero intermediate frequency signals;
step 3 is to make a signal sample set, which specifically comprises:
step 3-1: subjecting the digital zero intermediate frequency signal to Hilbert transform to obtain the corresponding
Figure BDA0002474843090000061
Signal, reconstructing the analysis signal
Figure BDA0002474843090000062
k is 1,2, …, 128. then, instantaneous parameters are calculated and normalized for the real part and imaginary part of the q (k) sequence with the length of N to form a 128 × 3 sample matrix, and the instantaneous parameters comprise:
instantaneous amplitude:
Figure BDA0002474843090000063
instantaneous phase:
Figure BDA0002474843090000064
instantaneous frequency:
Figure BDA0002474843090000065
wherein SbIn order to be able to sample the rate of the signal,
according to norm theory processing normalization, the magnitude vector a is normalized by L2 as:
Figure BDA0002474843090000066
wherein a isiAmplitude values representing the signal sequence;
the phase camber value θ is normalized to:
θ∈(-π,π)→[-1,1]
the instantaneous frequency f is normalized to:
f∈(fmin,fmax)→[-1,1]
normalizing the instantaneous frequency f to between-1 and 1;
step 3-2: manually labeling 40% of the amplitude, phase and frequency instantaneous data samples obtained in the step 3-1 with a modulation pattern and a signal-to-noise ratio corresponding to the instantaneous parameters as labels of the samples, wherein the labels are used as labeled samples, and the rest 60% of the samples are used as unsupervised training samples;
step 3-3: half of the labeled samples were used as labeled training samples and the other half were used as test samples.
As shown in fig. 2, the received radio frequency signal r (k) is down-converted to obtain digital zero intermediate frequency data, the instantaneous parameters of amplitude, phase and frequency are obtained through hilbert conversion, and then the three parameters are spliced into an N × 3 matrix, where N is the time length of a sample. The nx 3 matrix is input to a long short term memory-residual neural network, LSTM-ReNet.
Step 4 is extracting neural network characteristics, and specifically comprises the following steps:
step 4-1: inputting the unlabeled sample with the length of 128 obtained in the step 3 into the long-short term memory neural network;
step 4-2: extracting features of the sample on different time step lengths by using a long-short term memory neural network;
step 4-3: carrying out dimension expansion on the features extracted by the medium and short term memory neural network in the step 4-2 to enable the features to adapt to the input dimension of the subsequent convolutional network;
step 4-4: sending the features obtained in the step 4-3 into a deep residual error network, and extracting high-dimensional features to be recorded as Z;
step 5, extracting the comparative predictive coding network characteristics, which specifically comprises the following steps:
the structure of the contrast predictive coding neural network for unsupervised feature extraction in step 5-1 is shown in fig. 3: the characteristic extraction layer consisting of the long and short term memory-residual error network of the step 4 and the autoregressive network garAn unsupervised representation layer of composition;
step 5-1-1: autoregressive network garInput layer of (3) and feature extraction network output Z of step 4T[z1...,zt-3,zt-2,zt-1,zt,zt+1,zt+2,zt+3,...,z128]And connecting, and adopting a gate control cycle unit as an autoregressive network for prediction.
Step 5-1-2: will z≤tPart of the input into an autoregressive network garBy learning z≤tPartial data distribution, autoregressive network garOutput ct＝gar(z≤t) And to z≥tPartial prediction is noted
Figure BDA0002474843090000071
WkIs the prediction matrix for the t + k position,
Figure BDA0002474843090000072
step 5-2: predicting values from autoregressive networks
Figure BDA0002474843090000073
Sum long and short term memory-residual error neural network characteristic extraction value
Figure BDA0002474843090000074
Establishing mutual information expression of original input and autoregressive network output based on a noise contrast estimation loss function:
Figure BDA0002474843090000075
the above formula indicates that the difference between the predictor variable c and the true value x is measured by the mutual information I (x; c).
Step 6 is unsupervised pre-training, which specifically comprises the following steps:
step 6-1, according to the original input xt+kAnd autoregressive network output ctProportional to autoregressive network prediction
Figure BDA0002474843090000076
And feature extraction value
Figure BDA0002474843090000077
Namely, it is
Figure BDA0002474843090000078
Establishing predictions about autoregressive networks
Figure BDA0002474843090000079
And feature extraction value
Figure BDA00024748430900000710
Loss function of (2):
Figure BDA00024748430900000711
wherein the likelihood function L is lostNExpressed as predictive features
Figure BDA0002474843090000081
And the conditional probability of the actual feature z,
Figure BDA0002474843090000082
is the predicted value of the t + k position of the autoregressive network,
Figure BDA0002474843090000083
the predicted characteristics of the autoregressive network at the position t + k are obtained, and t and k are independent variables of the high-dimensional characteristic Z. { zlDenotes the set of features extracted by the neural network in a batch of samples, p is the sign of the conditional probability;
step 6-2, enabling the autoregressive network to learn the best expression characteristics of the sample in the unlabeled sample by minimizing a contrast loss function, and initializing parameters of the long-term and short-term memory-residual error network;
step 7 is training the softmax classifier, and specifically comprises the following steps:
step 7-1, replacing an autoregressive network of a contrast prediction coding network by a softmax classifier;
step 7-2 trains the classifier softmax with a number N of labeled data x, y,
Figure BDA0002474843090000084
wherein
Figure BDA0002474843090000085
Presentation classifier
Figure BDA0002474843090000086
And an initial value of θ*Feature extractor of
Figure BDA0002474843090000087
The formed mixed function finds the optimal parameter by supervised learning
Figure BDA0002474843090000088
N represents the number of labeled training samples;
step 8 is the test adjusts the network parameter, including:
step 8-1, the accuracy rate of the softmax is tested by using the test sample obtained in the step 3-3, and the method comprises the following steps:
and 8-1-1, inputting the data part of the test sample to the trained joint neural network to obtain an expected output vector. Taking the index of the maximum value in the expected output vectors of the samples to obtain the identification result of the signal modulation pattern;
step 8-1-2, comparing the identification result of the step 8-1-1 with the label of the sample to obtain the identification accuracy of different modulation patterns under different signal to noise ratios;
step 8-2, adjusting the parameters of the training hyperparameters and the neural network according to the test results, and then training;
parameters of the neural network are adjusted according to the trained loss curve convergence condition, so that the model achieves the optimal feature extraction effect, and finally the softmax classifier can achieve 90% accuracy under the condition of 0dB signal-to-noise ratio.
The invention also provides a signal modulation mode recognition device, which comprises a training data generation module, a signal receiving module, a digitization module, a data processing module, a model training module and a power supply module, wherein the training data generation module can generate various types of actual modulation signals through a signal simulation platform to generate abundant training data for training a neural network, so that a trained model is closer to an actual environment; the signal receiving module receives the modulation signal, and converts the modulation signal into two paths of analog signals after the modulation signal is subjected to band-pass filtering, amplification, frequency mixing and low-pass filtering, and then sends the two paths of analog signals to the digitizing module, the digitizing module performs analog-to-digital conversion on the two paths of analog signals and then stores the two paths of analog signals into the cache unit and then sends the two paths of analog signals into the data processing module, the FPGA board card of the data processing module stores a program for realizing the method, and outputs instantaneous parameters subjected to data processing to the model training module; the GPU board card used for training reads data used for training and operates a neural network model to perform data fitting; the power supply module is connected with other modules and provides voltage required by the other modules during working.
As shown in fig. 5, the training data generation module includes a signal simulation platform, an FPGA signal processor, a high-speed digital-to-analog converter, an amplifier, a broadband radio frequency front end, and a transmitting antenna, wherein the signal simulation platform generates various baseband modulation signals by using a GNURadio framework program interface, the baseband modulation signals are sent to the FPGA signal processor for interpolation to obtain intermediate frequency signals, the intermediate frequency signals are converted by the high-speed digital-to-analog converter to obtain analog signals, and then the analog signals are input to the broadband radio frequency front end by the amplifier for amplification to obtain radio frequency signals, and the radio frequency signals are transmitted to an external space by the transmitting antenna.
The signal receiving module comprises a receiving antenna, a broadband radio frequency front end, a band-pass filter, an amplifier, a mixer, a local oscillator and a low-pass filter, wherein the signal receiving antenna receives a modulation signal transmitted by the transmitting antenna to the broadband radio frequency front end, the broadband radio frequency front end outputs a radio frequency signal to the band-pass filter, the band-pass filter receives a signal in a carrier frequency range and then outputs the signal to the amplifier, the amplifier amplifies the signal and sends the signal to the mixer, the mixer is connected with the local oscillator to output the input radio frequency signal into I, Q two paths of analog intermediate frequency signals, and I, Q two paths of analog intermediate frequency signals are output to the digitizing module through the low-;
the digitization module comprises a high-speed analog-to-digital converter and a high-speed memory, wherein I, Q paths of externally input analog intermediate frequency signals are discretized by the high-speed analog-to-digital converter to obtain two paths of orthogonal digital signals, and the two paths of orthogonal digital signals are sequentially stored in the high-speed memory in a paired mode;
the data processing module is an FPGA board card and is responsible for reading two paths of orthogonal digital signals received in the high-speed memory, performing Hilbert transform and calculating instantaneous parameters of the signals, and then sending the instantaneous parameters to the GPU to participate in model training;
the model training module is a GPU board card, loads a neural network model constructed by a Pythrch frame program interface and signal data output by an FPGA through a PCIE interface of a mainboard, and analyzes and identifies data samples;
in a specific embodiment of the present invention, the signal generation module creates a GNURadio signal simulation platform on the Ubuntu operating system, and generates multiple types of actual modulation signals by using a signal simulation program interface provided by GNURadio in cooperation with a USRP (universal software Radio Peripheral) wireless transmission device.
The signal receiving module is a USRP receiver and comprises a receiving antenna, a band-pass filter, a signal amplifier, a frequency mixer, a local oscillator and a low-pass filter, wherein the signal receiving antenna receives a modulation signal transmitted by the transmitting antenna, the band-pass filter receives a signal in a carrier frequency range, the signal amplifier is sequentially connected with the signal receiving antenna, the signal amplifier is connected with the frequency mixer to convert a radio-frequency signal into an intermediate-frequency signal, and the frequency mixer is connected with the local oscillator to output two paths of I/Q analog signals which are output to the digitizing module through the low-pass filter;
the digitization module comprises a 16-bit high-speed analog-to-digital converter and a 64G high-speed memory, wherein external input signals are respectively accessed into the high-speed analog-to-digital converter to obtain two paths of orthogonal digital signals, and the two paths of orthogonal digital signals are sequentially stored in the high-speed memory unit in a paired mode;
the data processing module is an FPGA board card and is responsible for carrying out Hilbert transform on the received I/Q data, and then sending the calculated instantaneous parameters to the model training module to participate in model training;
the model training module is a Ubuntu operating system with a deep learning framework, a host with an NVIDIA P4000 GPU board card, a neural network training program is constructed by using the Pythrch deep learning framework, and the GPU board card is responsible for operating a neural network model and signal data output by an FPGA and analyzing and identifying data samples; the host can be an embedded platform or a computer.
The power supply module uses the LTM4644 to form a single-channel output dc/dc power supply, is responsible for supplying power to the signal receiving module, and uses the LTM4620 to supply power to the data processing module where the FPGA board card is located.
In order to verify the technical effect of the method, a GNURADio frame is used for building a signal transceiving platform, a deep learning Pythrch frame is used for building a neural network model in the method, and an NVIDIA P4000 GPU is used for building a training environment. FIG. 6 is a graph comparing the recognition performance of the method of the present invention with that of the method using supervised learning only, in which the solid line is the method of the present invention, the dotted line is the method of supervised learning, the test is performed under-2 dB, 0dB, and 10dB, respectively, and the test is plotted according to the change of the number of training samples, the numerical value of the horizontal axis represents the number of labeled signal samples used in recognition, and the numerical value of the vertical axis represents the recognition accuracy. As can be seen from FIG. 6, the recognition accuracy of the method of the present invention is significantly higher than that of supervised learning. When the labeled samples of 200-300 types of signals are used, the recognition accuracy of the semi-supervised learning method can achieve the best effect of supervised learning. When the number of the labeled data is only 5, the recognition accuracy of 11 modulation signals is close to 60% only through unsupervised pre-training, and the method shows that the effective characteristics of the signals can be learned through comparing a loss function by using unlabeled samples.
The method can be used for a modulation pattern recognition task by setting up a general neural network signal simulation platform; compared with the traditional communication modulation pattern recognition equipment, the invention provides an intelligent recognition framework based on deep learning, and for modulation patterns not used in the experiment, other modulation patterns can be recognized only by retraining the steps of the method; compared with the traditional method for manually designing the signal characteristics, the method avoids the increase of calculated amount caused by manually constructing the signal model and the signal characteristics; the invention can continuously improve the identification type and the identification performance of the modulation pattern identification by learning the received actual signal, thereby improving the identification efficiency; the invention adopts a semi-supervised learning mode, can effectively utilize a large amount of label-free data to train the network, and can train the modulation pattern classifier under the condition of less labeled samples.

Claims (7)

1. A signal modulation pattern recognition method, comprising:
generating training data, namely generating signals of various modulation styles by using a signal simulation platform and transmitting the signals by using a transmitter;
step 2: signal acquisition processing, namely receiving signals of various modulation patterns generated in the step 1, and preprocessing the signals to obtain I, Q two paths of digital zero intermediate frequency signals;
and step 3: making a signal sample set, namely performing Hilbert transform on I, Q two paths of digital zero intermediate frequency signals to obtain instantaneous parameters of amplitude, phase and frequency of the digital zero intermediate frequency signals, and taking the instantaneous parameters of all the digital zero intermediate frequency signals as a sample set to divide the instantaneous parameters into a label-free sample, a supervised training sample and a test sample;
and 4, step 4: extracting neural network characteristics, namely inputting the amplitude, phase and frequency instantaneous parameters obtained in the step (3) into a long-term and short-term memory neural network, extracting scale characteristics of the amplitude, phase and frequency instantaneous parameters at different time step lengths, and connecting the extracted scale characteristics after dimension change into a residual error neural network to extract high-dimensional characteristics;
and 5: extracting the characteristics of the contrast prediction coding network, namely inputting the high-dimensional characteristics extracted in the step (4) into an autoregressive network of the contrast prediction coding network, fitting and predicting according to data distribution of the high-dimensional characteristics input by the autoregressive network, and constructing a contrast loss function by combining the predicted characteristics output by the autoregressive network;
step 6: unsupervised pre-training, namely, using the prediction features obtained in the step 5 and the high-dimensional features extracted in the step 4, and using a gradient descent algorithm to make the loss function in the step 5 converge;
and 7: training a softmax classifier, namely replacing an autoregressive network for comparing and predicting codes with the softmax classifier, performing supervised training by using the labeled samples obtained in the step (2), and training the softmax classifier by using a cross entropy loss function;
and 8: and (3) testing and adjusting network parameters, namely testing the accuracy of the trained softmax classifier by using the test sample obtained in the step (2), and optimizing the network by adjusting the network hyper-parameters.
2. A signal modulation pattern recognition method according to claim 1, characterized in that:
the step 1 specifically comprises the following steps:
step 1-1: setting a modulation mode, a carrier frequency and a baud rate of a signal to be transmitted on a signal simulation platform;
step 1-2: processing the baseband signal by using a shaping filter, and performing interpolation up-sampling on the baseband signal by using the FPGA to obtain an intermediate frequency signal;
step 1-3: D/A conversion is carried out on the intermediate frequency signal to obtain an analog signal;
step 1-4: amplifying the band-pass filtered radio frequency signal by using a signal amplifier and transmitting the radio frequency signal by using a transmitting antenna;
the step 2 specifically comprises the following steps:
step 2-1: receiving the transmitted signal by using a receiving antenna;
step 2-2: performing band-pass filtering on the received signal by using a band-pass filter;
step 2-3: amplifying the band-pass filtered radio frequency signal by using a signal amplifier;
step 2-4: mixing the amplified radio frequency signal with a signal generated by a local oscillator by using a mixer to generate an intermediate frequency signal;
step 2-5: A/D sampling is carried out on the mixed intermediate frequency signal to obtain a digital signal;
step 2-6: performing digital low-pass filtering on the digital signal;
step 2-7: sending the signals subjected to digital low-pass filtering into a down converter to generate I, Q two paths of digital zero intermediate frequency signals;
the step 3 specifically comprises the following steps:
step 3-1: calculating instantaneous parameters of amplitude, phase and frequency of the sample by using the I, Q two paths of digital zero intermediate frequency signals obtained in the step 2-7;
step 3-2, manually labeling 40% of the instantaneous parameter samples of amplitude, phase and frequency obtained in the step 3-1 with the corresponding modulation pattern and signal-to-noise ratio of the instantaneous parameter as the label of the sample, namely labeled samples, and using the rest 60% of the samples as unlabeled training samples for unsupervised training;
3-3, taking 50% of the labeled samples obtained in the step 3-2 as supervised training samples, and taking the other half as test samples;
the step 4 specifically comprises the following steps:
step 4-1: taking the amplitude, phase and frequency instantaneous parameters obtained in the step (3) as input tensors, and sending the input tensors into the input end of the long-term and short-term memory neural network for training;
step 4-2: extracting scale features of the input tensor under different time step lengths by using a long-short term memory neural network;
step 4-3: carrying out dimension expansion on the scale features extracted by the medium-long short-term memory neural network in the step 4-2 to enable the scale features to adapt to the input dimension of the cascaded deep residual error network;
step 4-4, sending the dimension expanded scale features into a deep residual error network, and extracting high-dimensional features to be recorded as Z;
the step 5 specifically comprises the following steps:
step 5-1: selecting half of the high-dimensional feature Z' to input into an autoregressive network by using the high-dimensional feature Z extracted in the step 4-4, and training the autoregressive network;
step 5-2: the autoregressive network constructs data representation of a training sample by learning data distribution of Z', and carries out multi-step prediction to obtain a prediction characteristic P of the autoregressive network;
the step 6 specifically comprises the following steps:
step 6-1: constructing a contrast prediction loss function L by using the prediction feature P of the step 5-2 and the high-dimensional feature Z extracted in the step 4-4NThe following formula:
Figure FDA0002474843080000031
wherein
Figure FDA0002474843080000032
Is the predicted value of the t + k position of the autoregressive network,
Figure FDA0002474843080000033
is the prediction characteristic of the autoregressive network at the position t + k, and t and k are high dimensionsIndependent variable of characteristic Z, { ZlDenotes the set of features extracted by the neural network in a batch of samples, p is the sign of the conditional probability;
step 6-2, performing unsupervised pre-training by using the unlabeled training sample in the step 3, and converging the loss function obtained in the step 6-1 by using a gradient descent algorithm to make the weight of the long-short term memory-residual error neural network in the step 4 converge in an optimal solution space range;
the step 7 specifically comprises the following steps:
step 7-1: replacing the autoregressive network in the step 5 with a softmax network;
step 7-2: training a softmax classifier by using the labeled training sample obtained in the step 3-3;
the step 8 specifically comprises:
step 8-1: carrying out accuracy rate test on the softmax by using the test sample obtained in the step 3-3;
step 8-2: and adjusting the convergence rate of the hyperparametric regulation model by adjusting the learning rate of training, the number of samples selected by one-time training and the like, and adjusting the parameters of the neural network according to the convergence condition of the trained loss curve.
3. A signal modulation pattern recognition method according to claim 2, characterized in that:
step 1-1, setting carrier frequency 200MHz, modulation style and baud rate 10M of a signal to be generated on a signal simulation platform;
step 1-2, performing interpolation processing on a baseband signal by using an FPGA of a wireless transmitter;
step 1-3, performing digital-to-analog conversion on a signal by using a 16-bit D/A device of a wireless transmitter to generate an analog signal;
step 1-4, amplifying the signal at the radio frequency end and radiating the signal by using a transmitting antenna;
step 2-1, after receiving a signal, down-converting the signal to a baseband, filtering out an out-of-band signal by using a low-pass filter, and performing a/D sampling, wherein the number of sampling points is N, N is 128, and a digital complex baseband signal r (k) is obtained, k is 1,2, …, 128;
step 2-2, performing band-pass filtering on the received signal by using a band-pass filter with the bandwidth of 30MHz and the central frequency of 200 MHz;
step 2-3: compensating and amplifying the band-pass filtered radio frequency signal by using a signal amplifier so as to achieve the working dynamic range of the receiver;
step 2-4: mixing the amplified radio frequency signal with a signal generated by a local oscillator by using a mixer to generate an intermediate frequency signal;
step 2-5: A/D sampling is carried out on the mixed intermediate frequency signal to obtain a digital signal;
step 2-6: performing digital low-pass filtering on the digital signal;
step 2-7: sending the signals subjected to digital low-pass filtering into a down converter to generate I, Q two paths of digital zero intermediate frequency signals;
step 3-1: subjecting the digital zero intermediate frequency signal to Hilbert transform to obtain the corresponding
Figure FDA0002474843080000041
Signal, reconstructing the analysis signal
Figure FDA0002474843080000042
And then calculating instantaneous parameters for the real part and the imaginary part of a q (k) sequence with the length of N and carrying out normalization processing to form a 128 × sample matrix, wherein the instantaneous parameters comprise:
instantaneous amplitude:
Figure FDA0002474843080000043
instantaneous phase:
Figure FDA0002474843080000044
instantaneous frequency:
Figure FDA0002474843080000045
wherein SbIn order to be able to sample the rate of the signal,
according to norm theory processing normalization, the magnitude vector a is normalized by L2 as:
Figure FDA0002474843080000046
wherein a isiAmplitude values representing the signal sequence;
the phase camber value θ is normalized to:
θ∈(-π,π)→[-1,1]
the instantaneous frequency f is normalized to:
f∈(fmin,fmax)→[-1,1]
normalizing the instantaneous frequency f to between-1 and 1;
step 3-2: manually labeling 40% of the amplitude, phase and frequency instantaneous data samples obtained in the step 3-1 with a modulation pattern and a signal-to-noise ratio corresponding to the instantaneous parameters as labels of the samples, wherein the labels are used as labeled samples, and the rest 60% of the samples are used as unsupervised training samples;
step 3-3: taking one half of the labeled samples as labeled training samples, and taking the other half of the labeled samples as test samples;
the received radio frequency signal r (k) is subjected to down-conversion to obtain digital zero intermediate frequency data, instantaneous parameters of amplitude, phase and frequency are obtained through Hilbert conversion, then the three parameters are spliced into an Nx 3 matrix, wherein N is the time length of a sample, and the Nx 3 matrix is input into a long-short term memory-residual error neural network;
step 4-1: inputting the unlabeled sample with the length of 128 obtained in the step 3-2 into the long-short term memory neural network;
step 4-2: extracting features of the sample on different time step lengths by using a long-short term memory neural network;
step 4-3: carrying out dimension expansion on the features extracted by the medium and short term memory neural network in the step 4-2 to enable the features to adapt to the input dimension of the subsequent convolutional network;
step 4-4: sending the features obtained in the step 4-3 into a deep residual error network, and extracting high-dimensional features to be recorded as Z;
the comparison prediction coding neural network structure extracted from the unsupervised features in the step 5-1 comprises a long-term memory and short-term memory-residual error networkFeature extraction layer and autoregressive network g ofarAn unsupervised representation layer of composition;
step 5-1-1: autoregressive network garInput layer of (3) and feature extraction network output Z of step 4T[z1...,zt-3,zt-2,zt-1,zt,zt+1,zt+2,zt+3,...,z128]Connecting, and predicting by adopting a gate control cycle unit as an autoregressive network;
step 5-1-2: inputting the z ≦ t portion to the autoregressive network garBy learning z≤tPartial data distribution, autoregressive network garOutput ct＝gar(z≤t) And to z≥tPartial prediction is noted
Figure FDA0002474843080000051
WkIs the prediction matrix for the t + k position,
Figure FDA0002474843080000052
step 5-2: predicting values from autoregressive networks
Figure FDA0002474843080000053
Sum long and short term memory-residual error neural network characteristic extraction value
Figure FDA0002474843080000054
Establishing mutual information expression of original input and autoregressive network output based on a noise contrast estimation loss function:
Figure FDA0002474843080000055
the above formula represents that the difference between the predicted variable c and the true value x is measured by mutual information I (x; c);
step 6-1, according to the original input xt+kAnd autoregressive network output ctProportional to autoregressive network prediction
Figure FDA0002474843080000056
And feature extraction value
Figure FDA0002474843080000057
Namely, it is
Figure FDA0002474843080000058
Establishing predictions about autoregressive networks
Figure FDA0002474843080000059
And feature extraction value
Figure FDA00024748430800000510
Is used to determine the loss function of (c),
Figure FDA00024748430800000511
wherein the likelihood function L is lostNExpressed as predictive features
Figure FDA00024748430800000512
And the conditional probability of the actual feature z,
Figure FDA00024748430800000513
is the predicted value of the t + k position of the autoregressive network,
Figure FDA00024748430800000514
for the predicted features of the autoregressive network at the t + k positions, t, k are the independent variables of the high-dimensional feature Z, { ZlDenotes the set of features extracted by the neural network in a batch of samples, p is the sign of the conditional probability;
step 6-2, enabling the autoregressive network to learn the best expression characteristics of the sample in the unlabeled sample by minimizing a contrast loss function, and initializing parameters of the long-term and short-term memory-residual error network;
step 7-1, replacing an autoregressive network of a contrast prediction coding network by a softmax classifier;
step 7-2 trains the classifier softmax with a number N of labeled data x, y,
Figure FDA0002474843080000061
wherein
Figure FDA0002474843080000062
Presentation classifier
Figure FDA0002474843080000063
And an initial value of θ*Feature extractor of
Figure FDA0002474843080000064
The formed mixed function finds the optimal parameter by supervised learning
Figure FDA0002474843080000065
N represents the number of labeled training samples;
step 8-1, the accuracy rate of the softmax is tested by using the test sample obtained in the step 3-3, and the method comprises the following steps:
step 8-1-1, inputting the data part of the test sample to the trained joint neural network to obtain an expected output vector, and obtaining the identification result of the signal modulation pattern by taking the index of the maximum value in the expected output vectors of the samples;
step 8-1-2, comparing the identification result of the step 8-1-1 with the label of the sample to obtain the identification accuracy of different modulation patterns under different signal to noise ratios;
and 8-2, adjusting the parameters of the training hyperparameters and the neural network according to the test results, then training, and adjusting the parameters of the neural network according to the convergence condition of the trained loss curve.
4. The utility model provides a signal modulation mode recognition device, includes training data production module, signal reception module, digital module, data processing module, model training module, power module, its characterized in that: the training data generation module generates various types of modulation signals through a signal simulation platform; the signal receiving module receives the modulation signal, and converts the modulation signal into two paths of analog signals after the modulation signal is subjected to band-pass filtering, amplification, frequency mixing and low-pass filtering and then sends the two paths of analog signals to the digitization module; the digital module performs analog-to-digital conversion on the two paths of analog signals, stores the analog signals into the cache unit and then sends the analog signals into the data processing module; the FPGA board card of the data processing module stores a program for realizing the method of the preceding claim and outputs the instantaneous parameters subjected to data processing to the model training module; the GPU board card reads data for training and operates a neural network model to perform data fitting; the power supply module is connected with other modules and provides voltage required by the other modules during working.
5. A signal modulation scheme identification apparatus according to claim 4, wherein:
the training data generation module comprises a signal simulation platform, an FPGA signal processor, a high-speed digital-to-analog converter, an amplifier, a broadband radio frequency front end and a transmitting antenna, wherein the signal simulation platform generates various baseband modulation signals by utilizing a GNURADio frame program interface, the baseband modulation signals are sent to the FPGA signal processor for interpolation to obtain intermediate frequency signals, the intermediate frequency signals are converted by the high-speed digital-to-analog converter to obtain analog signals, then the analog signals are input to the broadband radio frequency front end by the amplifier for amplification to obtain radio frequency signals, and the radio frequency signals are transmitted to an external space by the transmitting antenna;
the signal receiving module comprises a receiving antenna, a broadband radio frequency front end, a band-pass filter, an amplifier, a mixer, a local oscillator and a low-pass filter, wherein the signal receiving antenna receives a modulation signal transmitted by the transmitting antenna to the broadband radio frequency front end, the broadband radio frequency front end outputs a radio frequency signal to the band-pass filter, the band-pass filter receives a signal in a carrier frequency range and then outputs the signal to the amplifier, the amplifier amplifies the signal and sends the signal to the mixer, the mixer is connected with the local oscillator to output the input radio frequency signal into I, Q two paths of analog intermediate frequency signals, and I, Q two paths of analog intermediate frequency signals are output to the digitizing module through the low-;
the digitization module comprises a high-speed analog-to-digital converter and a high-speed memory, wherein I, Q paths of externally input analog intermediate frequency signals are discretized by the high-speed analog-to-digital converter to obtain two paths of orthogonal digital signals, and the two paths of orthogonal digital signals are sequentially stored in the high-speed memory in a paired mode;
the data processing module is an FPGA board card and is responsible for reading two paths of orthogonal digital signals received in the high-speed memory, performing Hilbert transform and calculating instantaneous parameters of the signals, and then sending the instantaneous parameters to the GPU to participate in model training;
the model training module is a GPU board card, loads a neural network model constructed by a Pythrch frame program interface and signal data output by the FPGA through a PCIE interface of the mainboard, and analyzes and identifies data samples.
6. A signal modulation scheme identification apparatus according to claim 5, wherein:
the signal generation module is used for generating various types of actual modulation signals by building a GNURADio signal simulation platform on a Ubuntu operating system and matching with USRP (Universal Software Radio Peripheral) wireless transmitting equipment through a signal simulation program interface provided by GNURADio;
the signal receiving module is a USRP receiver and comprises a receiving antenna, a band-pass filter, a signal amplifier, a frequency mixer, a local oscillator and a low-pass filter, wherein the signal receiving antenna receives a modulation signal transmitted by the transmitting antenna, the band-pass filter receives a signal in a carrier frequency range, the signal amplifier is sequentially connected with the signal receiving antenna, the signal amplifier is connected with the frequency mixer to convert a radio-frequency signal into an intermediate-frequency signal, the frequency mixer is connected with the local oscillator to output two paths of I/Q analog signals, and the two paths of I/Q analog signals are output to the digitizing module through;
the digitization module comprises a 16-bit high-speed analog-to-digital converter and a 64G high-speed memory, wherein external input signals are respectively accessed into the high-speed analog-to-digital converter to obtain two paths of orthogonal digital signals, and the two paths of orthogonal digital signals are sequentially stored in the high-speed memory unit in a paired mode;
the data processing module is an FPGA board card and is responsible for carrying out Hilbert transform on the received I/Q data, and then sending the calculated instantaneous parameters to the model training module to participate in model training;
the model training module is a host which runs a Ubuntu operating system with a deep learning framework and is provided with an NVIDIA P4000 GPU board card, a neural network training program is constructed by using the Pythrch deep learning framework, and the GPU board card is responsible for running a neural network model and signal data output by the FPGA and analyzing and identifying data samples;
the power supply module uses the LTM4644 to form a single-channel output dc/dc power supply, is responsible for supplying power to the signal receiving module, and uses the LTM4620 to supply power to the data processing module where the FPGA board card is located.
7. A signal modulation scheme identification apparatus according to claim 6, wherein:
the host of the model training module is an embedded platform or a computer.
CN202010360390.9A 2020-04-30 2020-04-30 Signal modulation pattern recognition method Active CN111695417B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010360390.9A CN111695417B (en) 2020-04-30 2020-04-30 Signal modulation pattern recognition method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010360390.9A CN111695417B (en) 2020-04-30 2020-04-30 Signal modulation pattern recognition method

Publications (2)

Publication Number Publication Date
CN111695417A true CN111695417A (en) 2020-09-22
CN111695417B CN111695417B (en) 2023-03-14

Family

ID=72476810

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010360390.9A Active CN111695417B (en) 2020-04-30 2020-04-30 Signal modulation pattern recognition method

Country Status (1)

Country Link
CN (1) CN111695417B (en)

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112215293A (en) * 2020-10-20 2021-01-12 平安国际智慧城市科技股份有限公司 Plant disease and insect pest identification method and device and computer equipment
CN112382291A (en) * 2020-11-23 2021-02-19 北京百度网讯科技有限公司 Voice interaction processing method and device, electronic equipment and storage medium
CN112491677A (en) * 2020-11-06 2021-03-12 厦门大学 CAN bus identification method and device based on physical layer characteristic fingerprint
CN112887245A (en) * 2021-01-08 2021-06-01 鹏城实验室 Symbol detection method, device, equipment and storage medium of high dynamic channel
CN113052013A (en) * 2021-03-08 2021-06-29 中国人民解放军63891部队 Radio frequency fingerprint identification method for radio station modulation signals
CN113095378A (en) * 2021-03-26 2021-07-09 重庆邮电大学 Wireless network device identification method, computer device and readable storage medium
CN113239788A (en) * 2021-05-11 2021-08-10 嘉兴学院 Mask R-CNN-based wireless communication modulation mode identification method
CN113341382A (en) * 2021-05-31 2021-09-03 东南大学 Detection, identification and development integrated platform for intelligent cognitive radio
CN113518050A (en) * 2021-06-24 2021-10-19 华东交通大学 Modulation identification method, system, readable storage medium and device
CN113723556A (en) * 2021-09-08 2021-11-30 中国人民解放军国防科技大学 Modulation mode identification method based on entropy weighting-multi-mode domain antagonistic neural network
CN113869227A (en) * 2021-09-29 2021-12-31 西南交通大学 Signal modulation mode identification method, device, equipment and readable storage medium
CN114189416A (en) * 2021-12-02 2022-03-15 电子科技大学 Digital modulation signal identification method based on consistency regularization
CN115296758A (en) * 2022-07-07 2022-11-04 中山大学 Method, system, computer device and storage medium for identifying interference signal
CN115348140A (en) * 2022-08-17 2022-11-15 西安理工大学 Multi-channel communication signal implementation method based on FPGA
CN115664908A (en) * 2022-10-19 2023-01-31 西安电子科技大学 Cross-layer intelligent countermeasure method, system, medium and equipment for communication signal modulation recognition
CN115728588A (en) * 2022-12-23 2023-03-03 广州力赛计量检测有限公司 Electromagnetic compatibility detection system and method based on big data
CN115913850A (en) * 2022-11-18 2023-04-04 中国电子科技集团公司第十研究所 Open set modulation identification method based on residual error network
CN116016070A (en) * 2022-12-05 2023-04-25 天津光电通信技术有限公司 Digital signal blind demodulation system based on category increment learning
WO2024040451A1 (en) * 2022-08-24 2024-02-29 北京京东方技术开发有限公司 Data processing device, office equipment, office method and apparatus, medium, and product
CN117914661A (en) * 2024-03-20 2024-04-19 长春通视光电技术股份有限公司 Signal identification system based on instantaneous eigenvalue

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018176889A1 (en) * 2017-03-27 2018-10-04 华南理工大学 Method for automatically identifying modulation mode for digital communication signal
CN110163282A (en) * 2019-05-22 2019-08-23 西安电子科技大学 Modulation Mode Recognition method based on deep learning
CN110276256A (en) * 2019-05-17 2019-09-24 中国人民解放军战略支援部队信息工程大学 Based on the low signal-to-noise ratio Modulation Recognition of Communication Signal method and device for adjusting ginseng accidental resonance

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018176889A1 (en) * 2017-03-27 2018-10-04 华南理工大学 Method for automatically identifying modulation mode for digital communication signal
CN110276256A (en) * 2019-05-17 2019-09-24 中国人民解放军战略支援部队信息工程大学 Based on the low signal-to-noise ratio Modulation Recognition of Communication Signal method and device for adjusting ginseng accidental resonance
CN110163282A (en) * 2019-05-22 2019-08-23 西安电子科技大学 Modulation Mode Recognition method based on deep learning

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
苟泽中等: "基于半监督联合神经网络的调制识别算法", 《信号处理》 *

Cited By (29)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112215293A (en) * 2020-10-20 2021-01-12 平安国际智慧城市科技股份有限公司 Plant disease and insect pest identification method and device and computer equipment
CN112491677A (en) * 2020-11-06 2021-03-12 厦门大学 CAN bus identification method and device based on physical layer characteristic fingerprint
CN112382291A (en) * 2020-11-23 2021-02-19 北京百度网讯科技有限公司 Voice interaction processing method and device, electronic equipment and storage medium
CN112382291B (en) * 2020-11-23 2021-10-22 北京百度网讯科技有限公司 Voice interaction processing method and device, electronic equipment and storage medium
CN112887245A (en) * 2021-01-08 2021-06-01 鹏城实验室 Symbol detection method, device, equipment and storage medium of high dynamic channel
CN112887245B (en) * 2021-01-08 2022-10-11 鹏城实验室 Symbol detection method, device, equipment and storage medium of high dynamic channel
CN113052013A (en) * 2021-03-08 2021-06-29 中国人民解放军63891部队 Radio frequency fingerprint identification method for radio station modulation signals
CN113052013B (en) * 2021-03-08 2023-03-10 中国人民解放军63891部队 Radio frequency fingerprint identification method for radio station modulation signals
CN113095378B (en) * 2021-03-26 2022-04-05 重庆邮电大学 Wireless network device identification method, computer device and readable storage medium
CN113095378A (en) * 2021-03-26 2021-07-09 重庆邮电大学 Wireless network device identification method, computer device and readable storage medium
CN113239788A (en) * 2021-05-11 2021-08-10 嘉兴学院 Mask R-CNN-based wireless communication modulation mode identification method
CN113341382A (en) * 2021-05-31 2021-09-03 东南大学 Detection, identification and development integrated platform for intelligent cognitive radio
CN113518050A (en) * 2021-06-24 2021-10-19 华东交通大学 Modulation identification method, system, readable storage medium and device
CN113723556A (en) * 2021-09-08 2021-11-30 中国人民解放军国防科技大学 Modulation mode identification method based on entropy weighting-multi-mode domain antagonistic neural network
CN113869227A (en) * 2021-09-29 2021-12-31 西南交通大学 Signal modulation mode identification method, device, equipment and readable storage medium
CN114189416A (en) * 2021-12-02 2022-03-15 电子科技大学 Digital modulation signal identification method based on consistency regularization
CN114189416B (en) * 2021-12-02 2023-01-10 电子科技大学 Digital modulation signal identification method based on consistency regularization
CN115296758A (en) * 2022-07-07 2022-11-04 中山大学 Method, system, computer device and storage medium for identifying interference signal
CN115296758B (en) * 2022-07-07 2023-08-11 中山大学 Method, system, computer equipment and storage medium for identifying interference signals
CN115348140A (en) * 2022-08-17 2022-11-15 西安理工大学 Multi-channel communication signal implementation method based on FPGA
CN115348140B (en) * 2022-08-17 2024-04-19 西安理工大学 Multi-channel communication signal implementation method based on FPGA
WO2024040451A1 (en) * 2022-08-24 2024-02-29 北京京东方技术开发有限公司 Data processing device, office equipment, office method and apparatus, medium, and product
CN115664908A (en) * 2022-10-19 2023-01-31 西安电子科技大学 Cross-layer intelligent countermeasure method, system, medium and equipment for communication signal modulation recognition
CN115664908B (en) * 2022-10-19 2024-04-05 西安电子科技大学 Cross-layer intelligent countermeasure method, system, medium and equipment for communication signal modulation recognition
CN115913850A (en) * 2022-11-18 2023-04-04 中国电子科技集团公司第十研究所 Open set modulation identification method based on residual error network
CN115913850B (en) * 2022-11-18 2024-04-05 中国电子科技集团公司第十研究所 Open set modulation identification method based on residual error network
CN116016070A (en) * 2022-12-05 2023-04-25 天津光电通信技术有限公司 Digital signal blind demodulation system based on category increment learning
CN115728588A (en) * 2022-12-23 2023-03-03 广州力赛计量检测有限公司 Electromagnetic compatibility detection system and method based on big data
CN117914661A (en) * 2024-03-20 2024-04-19 长春通视光电技术股份有限公司 Signal identification system based on instantaneous eigenvalue

Also Published As

Publication number Publication date
CN111695417B (en) 2023-03-14

Similar Documents

Publication Publication Date Title
CN111695417B (en) Signal modulation pattern recognition method
Wu et al. Robust automatic modulation classification under varying noise conditions
CN111160176B (en) Fusion feature-based ground radar target classification method for one-dimensional convolutional neural network
CN114465855B (en) Automatic modulation recognition method based on attention mechanism and multi-feature fusion
CN106443632B (en) The radar target identification method of multitask Factor Analysis Model is kept based on label
CN114157539B (en) Data-aware dual-drive modulation intelligent identification method
CN111126332B (en) Frequency hopping signal classification method based on contour features
CN111342867A (en) MIMO iterative detection method based on deep neural network
CN114239749B (en) Modulation identification method based on residual shrinkage and two-way long-short-term memory network
CN112910812B (en) Modulation mode identification method for deep learning based on space-time feature extraction
CN111815806A (en) Method for preprocessing flight parameter data based on wild value elimination and feature extraction
CN114881093B (en) Signal classification and identification method
CN112749633A (en) Separate and reconstructed individual radiation source identification method
Zhang et al. Open set recognition of communication signal modulation based on deep learning
CN111551893A (en) Deep learning and neural network integrated indoor positioning method
CN114757224A (en) Specific radiation source identification method based on continuous learning and combined feature extraction
CN113095162B (en) Spectrum sensing method based on semi-supervised deep learning
CN116720062A (en) Intelligent satellite signal monitoring and identifying method and device
CN113240081B (en) High-resolution range profile target robust identification method aiming at radar carrier frequency transformation
CN115883301A (en) Signal modulation classification model based on sample recall increment learning and learning method
Lin et al. Features fusion based automatic modulation classification using convolutional neural network
CN114118151A (en) Intelligent spectrum sensing method with environment adaptive capacity
CN113642457A (en) Cross-scene human body action recognition method based on antagonistic meta-learning
CN111988252A (en) Signal modulation mode identification method based on deep learning
Fan et al. A Hybrid Approach for Signal Modulation Recognition Using Deep Learning Methods

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant