CN113420605A - Small sample enhancement method for communication signal - Google Patents

Small sample enhancement method for communication signal Download PDF

Info

Publication number
CN113420605A
CN113420605A CN202110596466.2A CN202110596466A CN113420605A CN 113420605 A CN113420605 A CN 113420605A CN 202110596466 A CN202110596466 A CN 202110596466A CN 113420605 A CN113420605 A CN 113420605A
Authority
CN
China
Prior art keywords
sample
signal
sample set
original
samples
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202110596466.2A
Other languages
Chinese (zh)
Inventor
李刚
吴麒
乔冠华
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.)
CETC 10 Research Institute
Southwest Electronic Technology Institute No 10 Institute of Cetc
Original Assignee
Southwest Electronic Technology Institute No 10 Institute of Cetc
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 Southwest Electronic Technology Institute No 10 Institute of Cetc filed Critical Southwest Electronic Technology Institute No 10 Institute of Cetc
Priority to CN202110596466.2A priority Critical patent/CN113420605A/en
Publication of CN113420605A publication Critical patent/CN113420605A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Signal Processing (AREA)
  • Noise Elimination (AREA)

Abstract

The invention discloses a small sample enhancement method of a communication signal based on multi-dimensional transformation, which aims to transform an original signal through multiple dimensions, enhance a signal sample and expand sample dimensions. Firstly, judging the type of an acquired original signal according to a standard demodulation and decoding method, and selecting a plurality of different types and signal samples to form an original sample set w; respectively adding noise and labeling labels consistent with the original samples to form a noise sample set; filtering each item of w respectively to form a filtering sample set; respectively carrying out block interception on each block of w, immediately recombining the intercepted blocks, and labeling a label consistent with the original sample to form a recombined sample set; respectively carrying out random frequency shift on each item of w, and labeling a label consistent with the original sample to form a frequency shift sample set; then mixing all sample sets and then disordering and sequencing to form an enhanced sample set; and finally, inputting the enhanced sample set into a deep learning model for training and performing cross validation.

Description

Small sample enhancement method for communication signal
Technical Field
The invention relates to the field of signal characteristic identification, in particular to a signal sample labeling and small sample enhancing processing method.
Technical Field
With the continuous development and application of wireless mobile communication technology, modulation identification systems have been widely applied in the fields of information interception, interference identification, and the like. Modulation pattern recognition is an extremely important technique in the field of uncooperative communications. The modulation identification technology is a technology for judging a modulation mode and other parameters adopted by a signal by analyzing and researching a received signal sample under the condition of lacking prior knowledge. The modulation mode identification technology is mainly used for identifying the modulation mode of a signal, is taken as a research hotspot in the field of communication all the time, and is widely applied to a plurality of military and civil communication fields such as signal monitoring, frequency spectrum management, electronic countermeasure and the like. Modulation pattern recognition is an extremely important technology in the field of uncooperative communication, and the emphasis is on researching what type of communication scheme a received radio signal belongs to, so that a basis is provided for subsequently acquiring communication information and further interfering a sender, and in the era of rapid development of communication technology in the 21 st century, modulation modes are diversified, and channel environments are changed towards complexity. The traditional research on the digital communication signal identification algorithm is generally under the background environment of white gaussian noise, but in the actual living scene, a large amount of non-gaussian noise with spike pulse characteristics is often distributed, and the traditional modulation identification algorithm will lose effect under such a channel environment. With the rapid development of artificial intelligence technology, deep learning theory is also applied to the field of modulation recognition more and more. Compared with the traditional modulation recognition algorithm, the modulation recognition algorithm based on deep learning can automatically learn deeper feature representation from signal data, and better recognition effect is often achieved. The deep learning method has remarkable advantages in the aspects of feature extraction and pattern recognition, and has been applied to the field of signal processing in recent years, and the basis of deep learning is a large amount of abundant and effective labeled data, so how to quickly and efficiently acquire more effective samples is a key point for improving the recognition accuracy. On the premise of high-speed development of the space communication technology, the original space communication target identification (such as modulation mode identification) method cannot distinguish the same type of space communication target individuals in the complex signal environment. Therefore, the detected spatial communication signals must be analyzed by a specific high fidelity communication receiver, and the radio frequency signal characteristics of the spatial communication targets are extracted from the detected spatial communication signals, so as to realize individual identification of the specific spatial communication targets. There are many methods for estimating carrier frequency of communication signal and carrier frequency extracted from modulation parameter, and no matter what modulation mode is adopted, there always exists carrier frequency in radio frequency signal emitted by space communication target. The deviation and stability of the carrier frequency depend on the local oscillation source of the communication transmitting equipment, and different manufacturing precision and debugging processes of the carrier frequency cause the difference of space target individuals. Part of the nonlinear noise of the spatial communication equipment is added to the radio frequency signal in various modulation modes, so that the received signal is slightly changed in the frequency domain, and a large amount of spurious frequency components are generated. Individual difference of spurious characteristics each space communication device will have different nonlinear characteristics due to differences of components, thereby generating different spurious components including intermodulation frequency, harmonic frequency, spurious modulation, etc. The extraction of these spurious features is one of the main bases for individual identification of the spatial communication target. Individual differences of modulation parameters are different due to different components, and the modulation parameters of different spatial communication devices have slight differences (such as the code rate of a PSK signal). The individual recognition is realized by utilizing an analysis method of feature extraction, most of the individual recognition is researched from concept and local features, and no outstanding research result exists due to higher difficulty. However, the practical application of the transient communication signal features has great difficulty, which is mainly caused by the extremely short duration of the transient signal, the difficulty of signal positioning in non-cooperative communication, and the similarity of the transient signal and noise making feature extraction very difficult. In the traditional communication target modulation pattern recognition, the spatial communication target individual recognition needs to perform feature extraction and classification on the fine features of the monitored radio frequency signals. However, these signal subtle feature differences are attached to the actual communication signals and are not easy to monitor in the actual complex signal environment. In the actual space communication target identification process, the difficulty of signal capture and feature extraction is faced in the identification of individuals by using transient signal features. The signal individual feature extraction is to transform digital signals in various different transform domains (time domain, frequency domain, high-order spectral domain and the like), extract features capable of reflecting the essence of target individuals, and fuse the obtained multidimensional signal features to obtain space target individual features suitable for classification; the classification and identification process is to classify the identified targets into a certain specific individual in a feature space through a modern machine learning algorithm and the like; the signal modulation modes can be identified by classifying the waveforms of the signals. And classifying the attributes of the signals, such as classifying the signals emitted by different platforms according to the unintentional modulation of the signals by a circuit and a power supply, and realizing the individual identification of the radiation source. In engineering practice, the number of possible collected effective signals is limited, and the collected signals are decoded and judged and then labeled, so that time and labor are consumed, a large amount of effective sample data is often difficult to obtain, only a small amount of labeled data is usually used as a training set, and the final classification and identification accuracy is low.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a method for enhancing a small sample data set of a communication signal, which is rapid, efficient and easy to realize in engineering, so as to improve the accuracy of signal type identification after deep learning model training.
The above object of the present invention can be achieved by a small sample enhancement method for a communication signal, characterized by comprising the steps of: firstly, judging the type of an acquired original signal according to a standard demodulation decoding method, carrying out data annotation on each signal according to the obtained type, selecting multiple types, and forming an original signal sample set by a small number of signal samples of each type; adding noise to each piece of the original signal sample set respectively, and labeling labels consistent with the original samples to form a noise sample set; filtering each piece of the original sample set respectively, and labeling labels consistent with the original samples to form a filtering sample set; respectively carrying out block interception on each piece of the original signal sample set, carrying out random sequencing and recombination on the intercepted blocks, and labeling a label consistent with the original sample to form a recombined sample set; respectively carrying out random frequency shift transformation on each piece of the original signal sample set, and labeling labels consistent with the original samples to form a frequency shift sample set; then mixing all sample sets and then disordering and sequencing to form an enhanced sample set; and finally, inputting the enhanced sample set into a deep learning model for training, performing cross validation and classification recognition on the validation set, and calculating the recognition accuracy of the signal type.
Compared with the prior art, the invention has the following beneficial effects.
The method comprises the steps of constructing a training model of a small communication signal sample based on multi-dimensional transformation after deep learning training is carried out based on an enhanced signal sample set, mixing and disordering all sample sets obtained through transformation and an original sample set to obtain an enhanced and expanded enhanced sample set, dividing the enhanced sample set into a training set verification set, inputting the training set into a deep neural network for feedback learning and optimization, carrying out cross verification on the verification set after a loss function tends to be stable, and calculating the identification accuracy of the signal type. The signal can be filtered, but the signal modulation mode is not changed. The method comprises the steps of obtaining a signal sample set of a plurality of transform domains by respectively carrying out noise adding, filtering, segmentation and recombination and frequency shift transformation on an original signal, adding the generated signal samples into the original signal sample set to obtain an expanded and enhanced signal sample set, and after deep learning training is carried out on the basis of the enhanced signal sample set, the accuracy of a training model for classifying and identifying the signal is improved by 3% compared with that before the signal is transformed and enhanced. The signal can be used for random frequency shift, but the signal modulation mode is not changed. Noise can also be added by using the signal, but the modulation mode of the signal is not changed.
The main contributions of the invention are: the deep neural network is trained through a multi-class sample set, the debugging mode characteristics of signals are extracted, the generalization of a training model can be enhanced, and the accuracy of classification and identification is improved.
Marking an original signal sample set according to the signal types, and adding random white Gaussian noise to each sample of each type of the original signal sample set to form a noise sample set; respectively performing band-pass filtering on each sample of an original signal sample set aiming at a frequency point Fc and a bandwidth Bw of an acquired signal to form a filtering sample set; dividing each sample of each type of the original signal sample set into a plurality of equal parts, randomly sequencing and recombining to form a recombined sample set; frequency offset transformation is carried out on each sample of each type respectively to form an offset sample set, original signals are transformed in multiple dimensions, useful characteristics of the signal samples are enhanced, sample dimensions are expanded, the problem of insufficient samples caused by difficulty in actual engineering signal acquisition and labeling can be effectively solved, and the classification recognition rate of signal targets can be remarkably improved after deep learning model training. The signal can be used for block interception and random sequencing and recombination, but the signal modulation mode is not changed.
The method comprises the steps of respectively carrying out noise adding, filtering, segmentation and recombination and frequency shift transformation on an original signal sample set w based on multi-dimensional transformation to obtain signal sample sets of a plurality of transformation domains, mixing all the sample sets and then disordering and sequencing to form an enhanced sample set; and finally, inputting the enhanced sample set into a deep learning model for training, performing cross validation and classification recognition on the validation set, and calculating the recognition accuracy of the signal type. The small sample signal set enhancement is realized efficiently, conveniently and easily through engineering, under the condition that the sample amount of the label is small, the unknown label data are classified in a deep neural network training mode, the convergence of a network model can be effectively improved, overfitting is prevented, and the signal type identification accuracy is remarkably improved.
Drawings
FIG. 1 is a flow chart of a small sample enhancement of a communication signal of the present invention;
the process of the present invention will be described in detail with reference to specific examples.
Detailed Description
See fig. 1. According to the invention, firstly, the type of the collected original signal is judged according to a standard demodulation decoding method, each signal is subjected to data marking according to the obtained type, a plurality of different types are selected, and a small amount of signal samples of each type form an original signal sample set; adding noise to each piece of the original signal sample set respectively, and labeling labels consistent with the original samples to form a noise sample set; filtering each piece of the original sample set respectively, and labeling labels consistent with the original samples to form a filtering sample set; respectively carrying out block interception on each piece of the original signal sample set, carrying out random sequencing and recombination on the intercepted blocks, and labeling a label consistent with the original sample to form a recombined sample set; respectively carrying out random frequency shift transformation on each piece of the original signal sample set, and labeling labels consistent with the original samples to form a frequency shift sample set; then mixing all sample sets and then disordering and sequencing to form an enhanced sample set; and finally, inputting the enhanced sample set into a deep learning model for training, performing cross validation and classification recognition on the validation set, and calculating the recognition accuracy of the signal type.
Marking an original signal sample set according to the signal types, and adding random white Gaussian noise to each sample of each type of the original signal sample set to form a noise sample set; respectively performing band-pass filtering on each sample of an original signal sample set aiming at a frequency point Fc and a bandwidth Bw of an acquired signal to form a filtering sample set; dividing each sample of each type of the original signal sample set into a plurality of equal parts to form a block recombination sample set, and carrying out random sequencing recombination; respectively carrying out frequency offset transformation on an offset sample set formed by each sample of each type, respectively carrying out noise adding, filtering, segmentation and recombination and frequency shift transformation on an original signal sample set w based on multi-dimensional transformation to obtain a signal sample set of a plurality of transformation domains, adding the samples generating the ear signals into the original signal sample set to obtain an expanded and enhanced signal sample set, carrying out deep learning training based on the enhanced signal sample set, and constructing a training model thereof, then mixing and disordering all sample sets obtained by transformation and original sample sets to obtain enhanced and expanded enhanced sample sets, dividing the enhanced sample sets into training set verification sets, inputting the training sets into a deep neural network for feedback learning and optimization, and after the loss function tends to be stable, performing cross validation on the validation set, and calculating the identification accuracy of the signal type.
The main process is shown in figure 1. In the specific embodiment, taking a specific signal as an example, other signals can be verified according to the method, which specifically comprises the following steps: the steps are performed in the following order, in labeling the original signal sample set according to the signal type, the original sample set w
Figure RE-GDA0003183114180000041
Wherein m represents the signal types contained in the original sample set w, n represents the number of samples contained in each type of signal, the jth sample of the ith type of signal is represented as q (i, j), and i is a label.
Q when random white Gaussian noise wgn is added to each sample of each class of the original signal sample setnoiseQ (i, j) + wgn, labeled with a label i consistent with the original sample q (i, j) and a jth sample q representing the ith signal after gaussian white noise additionnoise(i, j) to obtain a noisy sample set wnoise
Figure RE-GDA0003183114180000051
Respectively carrying out band-pass filtering on each sample of the original signal sample set aiming at the frequency point Fc and the bandwidth Bw of the acquired signal, and finishing the jth sample q of the ith signal subjected to band-pass filteringfilter(i,j)，
qfilter(i, j) ═ Filter (BandPass, q (i, j)), and labeling the Filter with a label i consistent with the original sample q (i, j), so as to form a Filter sample set wfilter，
Figure RE-GDA0003183114180000052
Wherein BandPass represents the filter coefficient and filter represents the band-pass filter function.
Dividing each sample of each type of the original signal sample set into a plurality of equal parts, randomly sequencing and recombining to finish the jth sample q of the ith type of signal after the signal is randomly recombinedrecon(i, j) is:
qrecon(i,j)＝[q(i,j)(l/3:l*2/3),q(i,j)(l*2/3:l),q(i,j)(1:l/3)]and labeling the original sample with a label i consistent with the original sample q (i, j) to form a recombined sample set wrecon
Figure RE-GDA0003183114180000053
Where l is the data length of each sample, q(i,j)(l/3: l × 2/3) represents a data fragment consisting of l/3 th to l × 2/3 th data bits, cut from the original signal sample q (i, j).
Respectively carrying out frequency shift conversion on each sample of each class, and finishing the j sample q of the ith class signal after frequency shift conversionfre(i, j), namely:
Figure RE-GDA0003183114180000061
and labeling the samples with labels i consistent with the original samples q (i, j) to form a frequency shift sample set wfre
Figure RE-GDA0003183114180000062
Wherein f is0For random frequency shifts, t is the time step and π is the circumference ratio 3.1415926.
Mixing and disordering all sample sets obtained by transformation and original sample sets to obtain enhanced and expanded sample set W [ W, W ]noise,wrecon,wfilter,wfre]Dividing the enhanced sample set W into a training set validation set, inputting the training set into a deep neural network for feedback learning and optimization, performing cross validation on the validation set after a loss function tends to be stable, and calculating the identification accuracy of the signal type; generating a plurality of modulation baseband signals, wherein the modulation baseband signals comprise Binary Phase Shift Keying (BPSK), Quadrature Phase Shift Keying (QPSK), Frequency Modulation (FM), Amplitude Modulation (AM), Binary Frequency Shift Keying (BFSK), Linear Frequency Modulation (LFM) and the like, and taking BPSK as an example, the signal generation function is q (t) ═ m (t) exp [ i (2 pi ft + theta) ofj)]，
Wherein m (t) is the signal amplitudeDegree, default 1, f signal frequency, t signal sampling period, thetajIs the phase. Taking the signal frequency f to be 4-8M, the sampling rate fs to be 40M, the signal bandwidth Bw to be 20M, the signal length T to be 20us, and adding 20dB random noise, wherein each type randomly generates 50 samples, and labels 1-7 according to the modulation type to form a sample set w.
Adding 20dB of white Gaussian noise to 50 samples of each type of modulation signals respectively, and setting the samples as labels No. 1-7 according to the corresponding sequence to form a sample set w of 50 x 7 samplesnoiseEach sample generating function is qnosie(t) wgn (q (t),20), wgn is gaussian white noise function, 50 samples of each type of modulation signal are subjected to band-pass filtering and are set as labels 1-7 in corresponding sequence, and a sample set w of 50 × 7 samples is formedfilterEach sample is represented as follows:
qfilter(t)＝filter(BandPass,q(t))
Figure RE-GDA0003183114180000063
wherein stop1 represents the upper limit cut-off frequency of the band-pass filter, pass1 represents the upper limit band-pass frequency of the band-pass filter, stop2 represents the lower limit cut-off frequency of the band-pass filter, pass2 represents the lower limit band-pass frequency of the band-pass filter, and BandPass is the generated filter coefficient.
Dividing 50 samples of each type of modulation signals into 3 segments at random, randomly sequencing and recombining the 3 segments, setting labels No. 1-7 according to corresponding sequence, and forming a sample set wreconMultiplying 50 samples of each type of modulated signal by
Figure RE-GDA0003183114180000071
Of frequency offset of (a), wherein f0Is random frequency of 100kHz-2MHz and is set as labels No. 1-7 according to the corresponding sequence to form a sample set wfreEach sample is represented as follows:
Figure RE-GDA0003183114180000072
w, wnoise、wrecon、wfilter、wfreThe method comprises the steps of randomly mixing 5 sample sets to form a 50X 7X 5 enhanced sample set W, dividing the enhanced sample set W into a training set verification set according to the proportion of 4:1, inputting the training set into a deep learning network for training, enabling the deep learning network to be composed of a convolutional layer CNN, an activation layer RELU, a full connection layer DNN and a classifier SoftMax, obtaining a prediction sample value X through the deep learning network, comparing the prediction sample value X with a real sample label Y for feedback iteration, training and iterating for 200 times, finally testing the sample identification accuracy by using the verification set, and repeating the steps to generate 100 and 350 different sample quantities for identification accuracy testing.
The above-mentioned embodiments are only examples for illustrating the working procedures and principles of the present invention, and are not intended to limit the embodiments of the present invention, and it will be obvious to those skilled in the art that various modifications may be made on the basis of the above-mentioned embodiments, and it is not intended to list all the embodiments, and all obvious changes or modifications of the technical solution are included in the scope of the present invention.

Claims (10)

1. A method for small sample enhancement of a communication signal, comprising the steps of: firstly, judging the type of an acquired original signal according to a standard demodulation decoding method, carrying out data annotation on each signal according to the obtained type, selecting multiple types, and forming an original signal sample set by a small number of signal samples of each type; adding noise to each piece of the original signal sample set respectively, and labeling labels consistent with the original samples to form a noise sample set; filtering each piece of the original sample set respectively, and labeling labels consistent with the original samples to form a filtering sample set; respectively carrying out block interception on each piece of the original signal sample set, carrying out random sequencing and recombination on the intercepted blocks, and labeling a label consistent with the original sample to form a recombined sample set; respectively carrying out random frequency shift transformation on each piece of the original signal sample set, and labeling labels consistent with the original samples to form a frequency shift sample set; then mixing all sample sets and then disordering and sequencing to form an enhanced sample set; and finally, inputting the enhanced sample set into a deep learning model for training, performing cross validation and classification recognition on the validation set, and calculating the recognition accuracy of the signal type.
2. A method for small sample enhancement of a communication signal as claimed in claim 1, characterized by: marking an original signal sample set according to the signal types, and adding random white Gaussian noise to each sample of each type of the original signal sample set to form a noise sample set; respectively performing band-pass filtering on each sample of an original signal sample set aiming at a frequency point Fc and a bandwidth Bw of an acquired signal to form a filtering sample set; dividing each sample of each type of the original signal sample set into a plurality of equal parts to form a block recombination sample set, and carrying out random sequencing recombination; respectively carrying out frequency offset transformation on an offset sample set formed by each sample of each type, respectively carrying out noise adding, filtering, segmentation and recombination and frequency shift transformation on an original signal sample set w based on multi-dimensional transformation to obtain a signal sample set of a plurality of transformation domains, adding the samples generating the ear signals into the original signal sample set to obtain an expanded and enhanced signal sample set, carrying out deep learning training based on the enhanced signal sample set, and constructing a training model thereof, then mixing and disordering all sample sets obtained by transformation and original sample sets to obtain enhanced and expanded enhanced sample sets, dividing the enhanced sample sets into training set verification sets, inputting the training sets into a deep neural network for feedback learning and optimization, and after the loss function tends to be stable, performing cross validation on the validation set, and calculating the identification accuracy of the signal type.
3. A method for small sample enhancement of a communication signal as claimed in claim 1, characterized by: in labeling an original signal sample set according to signal types, an original sample set w
Figure FDA0003091333950000011
Wherein m represents the signal types contained in the original sample set w, n represents the number of samples contained in each type of signal, the jth sample of the ith type of signal is represented as q (i, j), and i is a label.
4. A method for small sample enhancement of a communication signal as claimed in claim 3, characterized by: q when random white Gaussian noise wgn is added to each sample of each class of the original signal sample setnoiseQ (i, j) + wgn, labeled with a label i consistent with the original sample q (i, j) and a jth sample q representing the ith signal after gaussian white noise additionnoise(i, j) to obtain a noisy sample set wnoise
Figure FDA0003091333950000021
5. A method for small sample enhancement of a communication signal as claimed in claim 1, characterized by: respectively carrying out band-pass filtering on each sample of the original signal sample set aiming at the frequency point Fc and the bandwidth Bw of the acquired signal, and finishing the jth sample q of the ith signal subjected to band-pass filteringfilter(i,j)，
qfilter(i, j) ═ Filter (BandPass, q (i, j)), and labeling the Filter with a label i consistent with the original sample q (i, j), so as to form a Filter sample set wfilter，
Figure FDA0003091333950000022
Wherein BandPass represents the filter coefficient and filter represents the band-pass filter function.
6. A method for small sample enhancement of a communication signal as claimed in claim 1, characterized by: dividing each sample of each type of the original signal sample set into a plurality of equal parts, randomly sequencing and recombining the samples, and after finishing the random recombination of the signalsThe jth sample q of the ith signal of (1)recon(i, j) is:
qrecon(i,j)＝[q(i,j)(l/3:l*2/3),q(i,j)(l*2/3:l),q(i,j)(1:l/3)]and labeling the original sample with a label i consistent with the original sample q (i, j) to form a recombined sample set wrecon
Figure FDA0003091333950000031
Where l is the data length of each sample, q(i,j)(l/3: l × 2/3) represents a data fragment consisting of l/3 th to l × 2/3 th data bits, cut from the original signal sample q (i, j).
7. A method for small sample enhancement of a communication signal as claimed in claim 1, characterized by: mixing and disordering all sample sets obtained by transformation and original sample sets to obtain enhanced and expanded sample set W [ W, W ]noise,wrecon,wfilter,wfre]Dividing the enhanced sample set W into a training set validation set, inputting the training set into a deep neural network for feedback learning and optimization, performing cross validation on the validation set after a loss function tends to be stable, and calculating the identification accuracy of the signal type; generating a plurality of modulation baseband signals, wherein the modulation baseband signals comprise Binary Phase Shift Keying (BPSK), Quadrature Phase Shift Keying (QPSK), Frequency Modulation (FM), Amplitude Modulation (AM), Binary Frequency Shift Keying (BFSK), Linear Frequency Modulation (LFM) and the like, and taking BPSK as an example, the signal generation function is q (t) ═ m (t) exp [ i (2 pi ft + theta) ofj)]，
Wherein m (t) is signal amplitude, default is 1, f is signal frequency, t is signal sampling period, thetajIs the phase. Taking the signal frequency f to be 4-8M, the sampling rate fs to be 40M, the signal bandwidth Bw to be 20M, the signal length T to be 20us, and adding 20dB random noise, wherein each type randomly generates 50 samples, and labels 1-7 according to the modulation type to form a sample set w.
8. A thumbnail of a communication signal as in claim 1The enhancement method is characterized in that: adding 20dB of white Gaussian noise to 50 samples of each type of modulation signals respectively, and setting the samples as labels No. 1-7 according to the corresponding sequence to form a sample set w of 50 x 7 samplesnoiseEach sample generating function is qnosie(t) wgn (q (t),20), wgn is gaussian white noise function, 50 samples of each type of modulation signal are subjected to band-pass filtering and are set as labels 1-7 in corresponding sequence, and a sample set w of 50 × 7 samples is formedfilterEach sample is represented as follows:
qfilter(t)＝filter(BandPass,q(t))
Figure FDA0003091333950000032
wherein stop1 represents the upper limit cut-off frequency of the band-pass filter, pass1 represents the upper limit band-pass frequency of the band-pass filter, stop2 represents the lower limit cut-off frequency of the band-pass filter, pass2 represents the lower limit band-pass frequency of the band-pass filter, and BandPass is the generated filter coefficient.
9. A method for small sample enhancement of a communication signal as claimed in claim 1, characterized by: dividing 50 samples of each type of modulation signals into 3 segments at random, randomly sequencing and recombining the 3 segments, setting labels No. 1-7 according to corresponding sequence, and forming a sample set wreconMultiplying 50 samples of each type of modulated signal by
Figure FDA0003091333950000041
Of frequency offset of (a), wherein f0Is random frequency of 100kHz-2MHz and is set as labels No. 1-7 according to the corresponding sequence to form a sample set wfreEach sample is represented as follows:
Figure FDA0003091333950000042
w, wnoise、wrecon、wfilter、wfreThe 5 sample sets were randomly shuffled to form a 50 x 7 x 5 enhancement sampleThis set W.
10. The method for small sample enhancement of a communication signal according to claim 9, characterized by: dividing an enhanced sample set W into a training set verification set according to the proportion of 4:1, inputting the training set into a deep learning network for training, wherein the deep learning network consists of a convolutional layer CNN, an activation layer RELU, a full connection layer DNN and a classifier SoftMax, comparing a predicted sample value X obtained by the deep learning network with a real sample label Y for feedback iteration, performing training iteration for 200 times, testing the sample identification accuracy by using the verification set, and repeating the steps to generate 100 and 350 different sample quantities for identification accuracy testing.
CN202110596466.2A 2021-05-31 2021-05-31 Small sample enhancement method for communication signal Pending CN113420605A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110596466.2A CN113420605A (en) 2021-05-31 2021-05-31 Small sample enhancement method for communication signal

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110596466.2A CN113420605A (en) 2021-05-31 2021-05-31 Small sample enhancement method for communication signal

Publications (1)

Publication Number Publication Date
CN113420605A true CN113420605A (en) 2021-09-21

Family

ID=77713199

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110596466.2A Pending CN113420605A (en) 2021-05-31 2021-05-31 Small sample enhancement method for communication signal

Country Status (1)

Country Link
CN (1) CN113420605A (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109792346A (en) * 2016-09-28 2019-05-21 高通股份有限公司 To phase noise compensation reference signal design and scrambled enhancing
CN110879989A (en) * 2019-11-22 2020-03-13 四川九洲电器集团有限责任公司 Ads-b signal target identification method based on small sample local machine learning model
US20200143279A1 (en) * 2018-11-06 2020-05-07 DeepSig Inc. Radio frequency band segmentation, signal detection and labelling using machine learning
CN112418347A (en) * 2020-12-09 2021-02-26 浙江工业大学 Countermeasure enhancement method based on radio signal classification

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109792346A (en) * 2016-09-28 2019-05-21 高通股份有限公司 To phase noise compensation reference signal design and scrambled enhancing
US20200143279A1 (en) * 2018-11-06 2020-05-07 DeepSig Inc. Radio frequency band segmentation, signal detection and labelling using machine learning
CN110879989A (en) * 2019-11-22 2020-03-13 四川九洲电器集团有限责任公司 Ads-b signal target identification method based on small sample local machine learning model
CN112418347A (en) * 2020-12-09 2021-02-26 浙江工业大学 Countermeasure enhancement method based on radio signal classification

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
LIANG HUANG 等: ""Data Augmentation for Deep Learning-Based Radio Modulation Classification"", 《IEEE ACCESS (VOLUME: 8)》 *

Similar Documents

Publication Publication Date Title
Ding et al. Specific emitter identification via convolutional neural networks
Zhang et al. An efficient deep learning model for automatic modulation recognition based on parameter estimation and transformation
Wu et al. Robust automatic modulation classification under varying noise conditions
Ureten et al. Wireless security through RF fingerprinting
CN111832417B (en) Signal modulation pattern recognition method based on CNN-LSTM model and transfer learning
CN108234370A (en) Modulation mode of communication signal recognition methods based on convolutional neural networks
Sa et al. Specific emitter identification techniques for the internet of things
Ma et al. Automatic modulation classification under non-Gaussian noise: A deep residual learning approach
CN111049770B (en) Modulation signal identification method based on high-order cumulant
CN112364729A (en) Modulation identification method based on characteristic parameters and BP neural network
Yang et al. One-dimensional deep attention convolution network (ODACN) for signals classification
CN114422311A (en) Signal modulation identification method and system combining deep neural network and expert prior characteristics
CN109274626B (en) Modulation identification method based on constellation diagram orthogonal scanning characteristics
Ying et al. Differential complex-valued convolutional neural network-based individual recognition of communication radiation sources
Ya et al. Modulation recognition of digital signal based on deep auto-ancoder network
Ghauri KNN based classification of digital modulated signals
CN113542180A (en) Frequency domain identification method of radio signal
Hu et al. Feature fusion convolution-aided transformer for automatic modulation recognition
CN113420605A (en) Small sample enhancement method for communication signal
Xu et al. A jamming recognition algorithm based on deep neural network in satellite navigation system
Hazza et al. Robustness of digitally modulated signal features against variation in HF noise model
CN113905383B (en) IFF signal identification method, device and medium based on radio frequency fingerprint
CN115935252A (en) Communication signal modulation pattern recognition method and device based on self-supervision contrast learning
Zhu et al. Specific emitter identification based on visibility graph entropy
Zhang et al. Open Set Domain Adaptation for Automatic Modulation Classification in Dynamic Communication Environments

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20210921