CN109379311A - Ultrashort wave signal specific recognition methods based on convolutional neural networks - Google Patents
Ultrashort wave signal specific recognition methods based on convolutional neural networks Download PDFInfo
- Publication number
- CN109379311A CN109379311A CN201811159958.XA CN201811159958A CN109379311A CN 109379311 A CN109379311 A CN 109379311A CN 201811159958 A CN201811159958 A CN 201811159958A CN 109379311 A CN109379311 A CN 109379311A
- Authority
- CN
- China
- Prior art keywords
- layer
- neural networks
- convolutional neural
- signal specific
- signal
- 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
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L27/00—Modulated-carrier systems
- H04L27/0012—Modulated-carrier systems arrangements for identifying the type of modulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—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
Abstract
The invention belongs to radio signal identification technology fields, in particular to a kind of ultrashort wave signal specific recognition methods based on convolutional neural networks, include: Short Time Fourier Transform is carried out to signal specific in sample database, obtain signal time-frequency map, wherein, signal specific is the signal in signal transmitted data frame structure comprising frame swynchronization code；Convolutional neural networks model is trained using time-frequency map；Signal specific in ultra short wave communication is identified using the convolutional neural networks model after training.The present invention analyzes the visual characteristic that signal specific is presented on time-frequency spectrum first, and is trained by convolutional neural networks model, realizes the identification of ultrashort wave signal specific, improves signal identification rate；Finally by emulation experiment, the influence that aliasing interferes on ultrashort wave channel is effectively reduced, it realizes that ultrashort wave signal specific identifies under low signal-to-noise ratio, and interference free performance can be improved by optimization network structure and the increase network number of plies, there is stronger practical application value.
Description
Technical field
The invention belongs to radio signal identification technology field, in particular to a kind of ultrashort wave based on convolutional neural networks
Signal specific recognition methods.
Background technique
Blipology is widely used in radio intelligence, electronic countermeasure and software radio etc., and surpasses
The identification of shortwave signal specific is also a crucial ring therein, becomes the research hotspot in field of signal processing.Ultra short wave communication is
Refer to the communication of the electromagnetic transmission information using 30MHz to 300MHz wave band.But, due to the circulation way of ultra short wave communication,
Ultrashort wave channel is influenced by multipath effect, noise and Doppler effect etc. to a certain extent, so that the signal of transmission exists
Phenomena such as decline, interference and aliasing, become a more complicated channel.And signal specific refers in the data of signal transmission
Signal containing frame swynchronization code (header frame) is present in the time division multiplex communications system such as shortwave, ultrashort wave, satellite more, passes through
The target source of signal can be estimated by identifying this signal specific generally, play significant role to battlefield environment sensing.
It is mainly the time-frequency spectrum for relying on professional to observe reception signal in the signal identification means of previous early stage, relies on
Experience show that the modulation type of signal and the mode of parameter are realized.This manual type has very big limitation, it with
The empirical cumulative situation of scouting personnel has very big correlation.But by the inspiration of this process, some scholars find signal
Identification problem can be solved in the method in application image process field, and the research based on this thought has been achieved for one
Fixed achievement.But from the point of view of current research achievement, existing method is in low signal-to-noise ratio (5dB or less) and strong aliasing disturbed condition
Lower recognition effect is not good enough, because the feature that these methods are extracted cannot characterize the signal by strong aliasing interference effect well.
Summary of the invention
For this purpose, the present invention provides a kind of ultrashort wave signal specific recognition methods based on convolutional neural networks, pass through convolution
Neural network extracts the feature on signal specific time-frequency spectrum and realizes that low signal-to-noise ratio and strong aliasing interfere lower ultrashort wave signal specific
Identification, effectively improve signal identification rate.
According to design scheme provided by the present invention, a kind of ultrashort wave signal specific identification side based on convolutional neural networks
Method includes following content:
Short Time Fourier Transform is carried out to signal specific in sample database, obtains signal time-frequency map, wherein signal specific is
It include the signal of frame swynchronization code in signal transmitted data frame structure；
Convolutional neural networks model is trained and is tested using time-frequency map；
Signal specific in ultra short wave communication is identified using the convolutional neural networks model after training.
Above-mentioned, signal transmitted data frame structure includes frame swynchronization code and data frame, wherein frame swynchronization code is signal transmission
Regular data frame in data frame structure.
Above-mentioned, convolutional neural networks model includes input layer, convolutional layer one, the convolutional layer two, mixed layer set gradually
One, mixed layer two, mixed layer three, convolutional layer three, average pond layer and output layer, wherein convolutional layer one, convolutional layer two, mixing
The convolution kernel of different size size is used in layer one, mixed layer two, mixed layer three, convolutional layer three and average pond layer.
Preferably, convolutional layer one, convolutional layer two, mixed layer one, mixed layer two, mixed layer three, convolutional layer three and average pond
The convolutional layer size for changing layer is successively successively decreased.
Preferably, convolutional layer three connects average pond layer 3*3 convolution kernel using 3*3 convolution kernel.
Preferably, it is averaged pond double-layer structure in mixed layer three using convolution sum.
Preferably, mixed layer one, mixed layer two and mixed layer three are all made of Inception module and optimize.
Preferably, using global average pond and classification output realization full convolutional neural networks end to end in output layer
Training and identification.
It is above-mentioned, when convolutional neural networks model being trained and identified using time-frequency map before, first to time-frequency figure
Spectrum carries out random cropping and increases the received data scale of convolutional neural networks mode input layer.
Above-mentioned, both mixed layer one and mixed layer three structure introduce great-jump-forward transmitting and carry out residual noise reduction, alleviate training
With gradient disperse when identification.
Beneficial effects of the present invention:
The present invention analyzes the visual characteristic that signal specific is presented on time-frequency spectrum first, and passes through convolutional neural networks mould
Type is trained, and realizes the identification of ultrashort wave signal specific, improves signal identification rate；And finally by emulation experiment, further
The validity that convolutional neural networks are applied to signal specific identification is verified, Robust Performance, operation efficiently, there is stronger reality to answer
With value.
Detailed description of the invention:
Fig. 1 is the ultrashort wave signal specific identification process schematic diagram in embodiment；
Fig. 2 is the frame structure schematic diagram that data are transmitted in embodiment；
Fig. 3 is Traffic Channel header frame schematic diagram in embodiment；
Fig. 4 is that Traffic Channel sends signal time-frequency spectrum in embodiment；
Fig. 5 is LINK11-CLEW signal time-frequency spectrum in embodiment；
Fig. 6 is FM-LINK11 signal time-frequency spectrum in embodiment；
Fig. 7 is convolutional neural networks typical structure schematic diagram in embodiment；
Fig. 8 is signal time-frequency spectrum cutting schematic diagram in embodiment；
Fig. 9 is the time-frequency spectrum in embodiment under strong aliasing interference；
Figure 10 is convolutional neural networks model framework schematic diagram in embodiment；
Figure 11 is the time-frequency spectrum of variety classes signal specific in embodiment；
Figure 12 is point of addition shifted signal schematic diagram in embodiment；
Figure 13 is that channel interference signal schematic diagram is added in embodiment；
Figure 14 is that convolutional neural networks model and penalty values penalty values when InceptionV3 network training are bent in embodiment
Line chart；
Figure 15 is four class signal specific recognition result schematic diagrames under the conditions of signal-to-noise ratio different in embodiment；
Figure 16 is four class signal specific recognition performance comparison diagrams under signal-to-noise ratio different in embodiment；
Figure 17 is the signal time-frequency spectrum of different aliasing ratio situations in embodiment；
Figure 18 be in embodiment under aliasing situation to the performance comparison figure of four class signal specifics.
Specific embodiment:
The present invention is described in further detail with technical solution with reference to the accompanying drawing, and detailed by preferred embodiment
Describe bright embodiments of the present invention in detail, but embodiments of the present invention are not limited to this.
Depth learning technology all made breakthrough progress in fields such as voice, image, natural languages in recent years, cause
The revolutionary change of numerous areas.One of branch as deep learning, convolutional neural networks (Convolution Neural
Network, CNN) do well in field of image recognition, achieved in the major computer vision challenge match in the world it is excellent at
Achievement.For this purpose, the embodiment of the present invention, shown in Figure 1, a kind of ultrashort wave signal specific identification based on convolutional neural networks is provided
Method includes following content:
101, Short Time Fourier Transform is carried out to signal specific in sample database, obtains signal time-frequency map, wherein specific letter
Number for the signal that in signal transmitted data frame structure includes frame swynchronization code；
102, convolutional neural networks model is trained using time-frequency map；
103, signal specific in ultra short wave communication is identified using the convolutional neural networks model after training.
Convolutional neural networks have the characteristics that self study, can realize forcing for complicated function by layer-by-layer nonlinear transformation
Closely, more abstract target signature is extracted implicitly to characterize initial data；Short Time Fourier Transform is done to signal specific to obtain
Then time-frequency spectrum is trained improved convolutional neural networks model using time-frequency spectrum, last test network model, real
Existing ultrashort wave signal specific identification.The feature on signal specific time-frequency spectrum, which is extracted, by convolutional neural networks realizes low signal-to-noise ratio
Lower ultrashort wave signal specific identification is interfered with strong aliasing, the validity and reliability of signal identification is improved, there is stronger reality
Application value.
Shown in Figure 2, in further embodiment of the present invention, signal transmitted data frame structure includes frame swynchronization code and data
Frame, wherein frame swynchronization code is the regular data frame in signal transmission data frame structure.Frame swynchronization code (also referred to as header frame) refers to
In one section of specific regular data frame that signal is sent in frame structure data, generally in the previous section of data frame, play frame synchronization,
The effects of sign synchronization, carrier synchronization.This signal specific is prevalent in shortwave, ultrashort wave, in microwave frequency band, usually by
Produced by time division multiplex communication system.Time-frequency spectrum is that signal passes through Fourier Fourier transformation (abbreviation STFT) method in short-term
It obtains, reflects the energy density of signal at any time with the situation of change of frequency, letter can be obtained by observing time-frequency spectrum
Number time-frequency distributions characteristic.Signal specific can show unique visual characteristic, different frame structures, difference on time-frequency spectrum
The signal specific of modulation system can also produce different visual characteristics.Come by taking iridium communication system and LINK11 as an example below into
Row explanation.
Iridium communication system is a kind of global system for mobile communications, provides mobile phone, paging and low speed data for user
The main businesses such as transmission, the frame structure of Traffic Channel contain as shown in figure 3, the header frame length of Traffic Channel is 54 bits
Multiple duplicate " 1 " bits.These duplicate code words produce unique frequency characteristic after carrying out DQPSK modulation.Industry in Fig. 4
Business channel sends the time-frequency spectrum of signal, it can be seen that the header frame portion of signal is showed different from later dataframe part
Visual characteristic, this visual characteristic can become the feature for identifying this signal.
LINK11 data-link is a kind of tactical data communication mode.The time-frequency spectrum of conventional LINK11 signal is as shown in Figure 5.
It can be seen that LINK11 is a kind of multiple signals, header frame is two audios, and data frame is made of 16 audios.
LINK11 data-link is transmitted using hybrid multiplex modulation system.In ultrashort wave band, LINK11_CLEW signal uses FM mode tune
It makes on carrier wave.LINK11_CLEW signal time-frequency spectrum of (referred to as FM_LINK11) after FM is as shown in Figure 6.FM_ in figure
LINK11 signal header frame also shows the visual characteristic different from data frame.
Fourier Fourier transformation STFT is the basic methods of time frequency analysis in short-term, applies STFT in the embodiment of the present invention
Method analyzes the visual characteristic of signal specific, not only can effectively avoid the transformation bring cross term problem such as WVD, but also unlike small echo becomes
It changes with excessive calculation amount, especially under strong aliasing disturbed condition, also can clearly embody the time-frequency characteristic of signal, be applicable in very much
It is analyzed in signal specific.Therefore, the embodiment of the present invention finds that signal specific has apparent visual characteristic on time-frequency spectrum, no
Same frame structure, different modulating mode signal specific different visual characteristics can be showed on time-frequency spectrum, using the side STFT
Method can clearly reflect that visual characteristic of the signal specific on time-frequency spectrum, this visual characteristic can become identification signal specific
Feature；By extracting this visual characteristic to signal specific application STFT method, and then realize signal specific identification.
Convolutional neural networks realize approaching for complicated function by layer-by-layer nonlinear transformation, can effectively be depicted multiple
The information of miscellaneous image, to extract profound characteristics of image.The typical structure of convolutional neural networks is as shown in fig. 7, convolution is refreshing
It is generally made of input layer, convolutional layer, pond layer, full articulamentum and output layer through network.The basic module of convolutional neural networks
For four kinds of operations such as convolution streams, including convolution, Chi Hua, activation primitive and batch normalization.(1), convolution, which refers to, utilizes convolution kernel
Input picture is handled, the higher characteristic of robustness may be learned.Convolution operation can reduce unnecessary weight connection,
Induce one sparse or part connection, and bring weight sharing policy greatly reduces parameter amount, to mitigate overfitting problem.Simultaneously
Convolution operation has translation invariance, for the position due to caused by frequency deviation and time delay on ultrashort wave signal specific time-frequency spectrum
Offset and deformation can be corrected effectively, so that the feature extracted has generalization and robustness.(2), pond refers to down-sampled
Operation takes a specific value output that is, in regional area.Substantially, pondization operation executes the polymerization of space or characteristic type,
Spatial Dimension is reduced, over-fitting risk is effectively relieved, while portraying translation invariance.Common pond mode have maximum pond,
Average pond, norm pond and log probability pond etc..For ultrashort wave signal specific, time-frequency spectrum can be effectively reduced in pondization operation
The influence of noise on figure.(3), activation primitive is a kind of nonlinear function, so that each level result carries out Nonlinear Mapping.Core
It is the compound non-linear ability of portraying to promote whole network by level Nonlinear Mapping (activation primitive).Otherwise, multiple
The linear combination of level is still linear approximation mode, and the ability for characterizing complex characteristics is limited.Common activation primitive has modified line
Property unit R eLU (accelerating convergence, inside accumulate sparsity), Softmax (calculating probability respondence for output layer), Sigmoid system (biography
The core for neural network of uniting, including Logistic-Sigmoid function and Tanh-Sigmoid function).Depth model generally uses
ReLU activation primitive, because it has the prominent characteristics such as unilateral inhibition, sparse activity, saturability.Letter specific for ultrashort wave
Number, layer-by-layer nonlinear transformation can extract the feature of ultrashort wave signal specific profound level to describe signal specific.(4), return in batches
One change is optimization operation, so that the mean value of each level result is 0, variance 1.With the intensification of network level, network convergence meeting
Very slow, the training time can be too long, will lead to gradient disperse problem.By batch normalization operation, distribution of results can be made to swash
In the linearly interval of function living, accelerate the convergence rate in training process, while avoiding the case where falling into local optimum.For super
The identification of shortwave signal specific, convolutional neural networks are very suitable to extract the visual characteristic on signal specific time-frequency spectrum, part
Connection, the shared and down-sampled characteristic of weight have it to positional shift, change in shape of the signal specific in time-frequency spectrum
Very high tolerance, and can extract more to be abstracted with comprehensive characteristic information in time-frequency spectrum and characterize signal specific, together
When the influence of noise and interference on time-frequency spectrum can be effectively reduced, therefore, the embodiment of the present invention by convolutional neural networks come pair
The time-frequency spectrum of signal specific is identified, the low signal-to-noise ratio and strong aliasing disturbed condition on ultrashort wave channel are overcome, and is realized super
The identification of shortwave signal specific.Convolutional neural networks mould in another embodiment of the present invention, for the identification of ultrashort wave signal specific
Type frame structure is as shown in Figure 10, and convolutional neural networks model includes the input layer set gradually, convolutional layer one, convolutional layer two, mixing
Layer one, mixed layer two, mixed layer three, convolutional layer three, average pond layer and output layer, wherein convolutional layer one, mixes convolutional layer two
Close the convolution kernel that different size size is used in layer one, mixed layer two, mixed layer three, convolutional layer three and average pond layer.For
Ultrashort wave signal specific identifies network model, may be selected to use relatively new Google InceptionV3 network, using not
With the convolution kernel of size, more fully feature can be extracted, signal specific under strong jamming is relatively suitble to identify.But its structure ratio
It is more complex, the identification of ultrashort wave signal specific is directly applied to, exists and restrained slow, the too long problem of training.The present invention is implemented
In example, the characteristics of according to ultrashort wave signal specific, it is optimized, its complexity is reduced, to improve training effectiveness.It is excellent
It is as follows to change content:
(1) increase training data scale.Mode input layer is designed and sized to 299*299.The number of ultrashort wave signal specific
It is little according to amount, it is difficult to meet training requirement.In the embodiment of the present invention, by large-sized input picture carry out random cropping come
More data samples are obtained, while in order to obtain whole signal specific features by cutting, the size of input picture can not
It is excessive.Fig. 8 is that iridium satellite DQPSK signal spectrogram can with random cropping is carried out to it having a size of 299*299 box when its is excessive
Incomplete signal specific feature can be obtained, therefore the signal specific spectrogram having a size of 320*320 may be selected in the present embodiment and make
For input picture, the random cropping having a size of 299*299 is carried out to it to increase training data scale, improves the extensive energy of model
Power.
(2) simplify Inception structure.The core of Google InceptionV3 network is to have used Inception knot
Structure.Inception structure uses different size of convolution kernel, can extract the different characteristic of image, realizes and is reducing net
Network depth is promoted while network parameter, improves the computational efficiency of network.But for the identification of ultrashort wave signal specific,
The Inception structure of Google InceptionV3 network is more, and the training time is longer, therefore model has mainly used three
Inception structure.Convolution kernel size is bigger to extract more abstract signal specific feature, therefore mixed layer 4 uses
The 7*7 convolution kernel of large scale extracts more abstract signal specific feature；Mixed layer 3 and 5 structure of mixed layer are by inducing one simultaneously
Great-jump-forward transmits skip connection thought, carries out residual noise reduction, alleviates gradient and diffuses problem.
(3) deep layer network structure is improved.Signal specific on ultrashort wave channel makes it possible to often by strong aliasing interference effect
The characteristic information for measuring relatively weak signal specific is unobvious, as shown in Figure 9.In this case, the maximum pond of deep layer network
Change will cause the characteristic information for only extracting strongly disturbing feature and ignoring signal specific, influence recognition effect.Therefore model than
Maximum pond layer in deeper mixed layer 5 replaces with average pond layer.
(4) optimize logistic regression layer.Google InceptionV3 is replaced in logistic regression layer with the average pond layer of the overall situation
Full articulamentum realizes full convolutional network end to end, greatly reduces network parameter.It is but dry for aliasing strong in Fig. 9
Disturb the identification of lower signal specific, the average pondization of oversized dimensions can protrude strong jamming and off-energy it is weaker signal specific it is special
Sign.Therefore, the convolution kernel connection 3*3 that 3*3 can be selected in the embodiment of the present invention is averaged pond layer to replace 8*8 to be averaged pond.
In Figure 10, network model includes three convolutional layers, and three mixed layers, a mean value pond layer, mixed layer uses
Inception module.The major parameter information of network model may be designed as shown in table 1:
The major parameter information of 1 network model of table
For the validity for verifying the present invention program, explanation is further explained below by emulation experiment data:
Emulation experiment is carried out for ultrashort wave signal specific recognition methods, has chosen iridium satellite DQPSK, MIL-STD
These four types of representative signal specifics of SOQPSK, LINK4A and FM-LINK11 are as data set, wherein sample rate
1MHz, carrier frequency are 200KHz, and specifying information is as shown in table 2:
2 data set information of table
Time-frequency spectrum generates under MATLAB environment, in order to balance the temporal resolution and frequency resolution of time-frequency spectrum,
The basic parameter of comprehensive ultrashort wave signal specific, obtains more significant visual characteristic, carries out 1024 points short to signal specific
When Fourier convert to obtain time-frequency spectrum, the time-frequency spectrum according to above-mentioned model interception 320*320 size as input picture,
Its frequency resolution is 1KHz.Different types of signal specific has different visual characteristics, major embodiment on time-frequency spectrum
In the shape, size and amplitude of header frame, as shown in figure 11.It (a) be iridium satellite DQPSK spectrogram, (b) is MIL-STD SOQPSK
Spectrogram, (c) are LINK4A spectrogram, (d) is FM-LINK11 spectrogram.In view of ultrashort wave signals wide band scan situation, in specific letter
Number time-frequency spectrum in increase random horizontal-shift (delay) and vertical shift (frequency deviation) and various ultrashort wave aliasings dry
It disturbs.By taking iridium satellite DQPSK signal as an example, additive effect is as shown in Figure 12,13.Original signal in comparison diagram 11, signal in Figure 12
It is added to random site offset, signal is added to random ultrashort wave channel disturbance in Figure 13.
Network model training and test are to call the TensorFlow depth of Google publication under Anaconda3 platform
For learning database come what is completed, programming language uses Python.In the range of signal-to-noise ratio is -9dB~21dB, every 3dB
2000 signal specific spectrograms are generated, wherein training set and test set are all 1000, are in total all 11000.Learning rate is
0.1, study attenuation rate is 0.99, regularization coefficient 0.003, and batch standardization decays to 0.99, frequency of training 2000 times, is carried out
20 experiments.Penalty values when network model and InceptionV3 network training are as shown in figure 14.It can be sent out from Figure 14
It is existing, to the convolutional neural networks model ratio InceptionV3 network ladder in the training process of signal specific, in the embodiment of the present invention
Faster, faster, the penalty values at the end of training are smaller for convergence rate for degree decline.It demonstrates in the embodiment of the present invention using convolution mind
Validity through network model.
Four class signal specific recognition results are as shown in figure 15 under the conditions of different signal-to-noise ratio.From Figure 15 it can be found that with
The increase of signal-to-noise ratio, the discrimination of signal specific also promoted.When signal-to-noise ratio is greater than 0dB, four class signal specifics all reach
95% or more discrimination, when signal-to-noise ratio is -5dB, part signal specific can also reach 90% discrimination, on the whole originally
Literary method can achieve the effect that feel quite pleased.The discrimination of iridium satellite DQPSK signal without other three classes signal specifics discrimination that
It is high.Because the more complicated signal specific of header frame shape is less susceptible to be disturbed influence under low signal-to-noise ratio.As shown in figure 11,
The header frame part of iridium satellite DQPSK signal is relatively simple, becomes unobvious vulnerable to interference effect, reduces recognition effect.Pass through
Experimental data can obtain overall average discrimination in the embodiment of the present invention and reach 98.1%, further illustrate convolutional neural networks model pair
Signal specific identification has preferable performance.
To verify this paper recognition methods with preferable noise robustness, by square with classical principal component analysis (PCA)
Method and independent component analysis (ICA) method compare.Both of which is extracted 100 features, has selected classical branch
Vector machine (SVM) is held as classifier.The experiment simulation condition of the above method is identical.Figure 16 is these three methods in different noises
Than the lower performance comparison figure to four class signal specifics.With the increase of signal-to-noise ratio, the knowledge of three kinds of methods it can be seen from Figure 16
Rate is not all being promoted, and wherein the performance of method is obviously better than other two methods in the embodiment of the present invention.When signal-to-noise ratio is greater than
When 6dB, the average recognition rate of method basically reaches 100% in the embodiment of the present invention, average to identify when signal-to-noise ratio reaches 0dB
Rate remains to reach 97%.Scheme disclosed in the embodiment of the present invention indicated above has good noise resisting ability, can be suitably used for
The Classification and Identification of signal specific under low signal-to-noise ratio.
There is preferable anti-aliasing jamming performance to verify the recognition methods of this paper, different aliasing interference journeys of having analyzed and researched
Spend the discrimination of lower signal specific.The unknown strong jamming for adding certain white Gaussian noise and aliasing different proportion is believed in the experiment
Number and generate signal specific time-frequency spectrum be used as sample data, aliasing interference ratio be 10%~100% in the range of, often
10% generates 1000 sample datas as test set, is total up to 10000.Training set is 2000 non-strongly disturbing spies of aliasing
Determine signal time-frequency spectrum, other conditions are same as above.Illustrate different aliasing ratio situations by taking iridium satellite DQPSK signal as an example, in noise
Under conditions of for 0dB, the average power ratio of interference signal and echo signal is 3:1, it is 10% that aliasing, which interferes ratio, 40%,
Spectrogram when 70% is as shown in figure 17.Scheme uses the stronger width of mean power in the embodiment of the present invention as can see from Figure 17
Band unknown signaling carries out interference aliazing effect to signal specific, as aliasing ratio increases, is increasingly difficult to object observing signal.
Scheme in the embodiment of the present invention is respectively trained heterogeneous networks model under aliasing disturbed condition, recognition result such as table 3
It is shown:
The performance of 3 heterogeneous networks model of table compares
| Model | With the difference of model of the embodiment of the present invention | Average recognition rate |
| 1 | Model of this embodiment of the present invention | 0.8079 |
| 2 | Mixed layer 5 uses maximum pond layer | 0.7592 |
| 3 | Mixed layer 4 replaces 7*7 convolution kernel using 3*3 | 0.7514 |
| 4 | The 7th layer of average pond using 8*8 | 0.66 |
From table 3 it is observed that model of this embodiment of the present invention is more preferable compared to other model performances, demonstrate for strong mixed
The folded model optimization for interfering lower signal specific identification is effective.Figure 18 is for distinct methods to the specific letter of four classes under aliasing situation
Number performance comparison figure.From Figure 18 it can be found that with aliasing ratio increase, the discrimination of three kinds of methods all declining.
In general, the performance of recognition methods is better than other methods in this embodiment of the present invention.Have when aliasing is less than 50%
90% or more discrimination, when aliasing is more than 60%, discrimination just gradually glides.Different from other two methods, convolution mind
There is the ability of local shape factor through network, signal specific partial visual feature can be extracted, therefore discrimination is higher.More than
Showing scheme in this embodiment of the present invention has certain anti-aliasing ability, can be suitably used for part aliasing and interferes lower signal specific
Classification and Identification.
In the embodiment of the present invention, by the visual characteristic on research ultrashort wave signal specific time-frequency spectrum, find not of the same race
There are bigger othernesses for the signal specific time-frequency spectrum of class, image recognition are applied to field of signal identification, and introduce
Convolutional neural networks break tional identification algorithmic procedure, while carrying out feature extraction and Classification and Identification.The experimental results showed that this
The influence that aliasing interferes on ultrashort wave channel can be effectively reduced in disclosure of the invention scheme, realize ultrashort wave signal specific under low signal-to-noise ratio
Identification, and interference free performance can be improved by optimization network structure and the increase network number of plies.In conclusion being based on time-frequency spectrum
The ultrashort wave signal specific recognition methods of figure and convolutional neural networks can effectively realize the category identification of signal specific.The party
The ability that there is method preferable anti-noise jamming and anti-part aliasing to interfere improves the accuracy rate of identification and general compared with conventional method
Change ability is a kind of effective signal specific recognition methods.
Each embodiment in this specification is described in a progressive manner, the highlights of each of the examples are with other
The difference of embodiment, the same or similar parts in each embodiment may refer to each other.For device disclosed in embodiment
For, since it is corresponded to the methods disclosed in the examples, so being described relatively simple, related place is said referring to method part
It is bright.
The unit and method and step of each example described in conjunction with the examples disclosed in this document, can with electronic hardware,
The combination of computer software or the two is realized, in order to clearly illustrate the interchangeability of hardware and software, in above description
In generally describe each exemplary composition and step according to function.These functions are held with hardware or software mode
Row, specific application and design constraint depending on technical solution.Those of ordinary skill in the art can be to each specific
Using using different methods to achieve the described function, but this realization be not considered as it is beyond the scope of this invention.
Those of ordinary skill in the art will appreciate that all or part of the steps in the above method can be instructed by program
Related hardware is completed, and described program can store in computer readable storage medium, such as: read-only memory, disk or CD
Deng.Optionally, one or more integrated circuits also can be used to realize, accordingly in all or part of the steps of above-described embodiment
Ground, each module/unit in above-described embodiment can take the form of hardware realization, can also use the shape of software function module
Formula is realized.The present invention is not limited to the combinations of the hardware and software of any particular form.
The foregoing description of the disclosed embodiments makes professional and technical personnel in the field can be realized or use the application.
Various modifications to these embodiments will be readily apparent to those skilled in the art, as defined herein
General Principle can be realized in other embodiments without departing from the spirit or scope of the application.Therefore, the application
It is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein phase one
The widest scope of cause.
Claims (10)
1. a kind of ultrashort wave signal specific recognition methods based on convolutional neural networks, which is characterized in that include following content:
Short Time Fourier Transform is carried out to signal specific in sample database, obtains signal time-frequency map, wherein signal specific is signal
It include the signal of frame swynchronization code in transmitting data frame structure；
Convolutional neural networks model is trained using time-frequency map；
Signal specific in ultra short wave communication is identified using the convolutional neural networks model after training.
2. the ultrashort wave signal specific recognition methods according to claim 1 based on convolutional neural networks, which is characterized in that
Signal transmitted data frame structure includes frame swynchronization code and data frame, wherein frame swynchronization code is that signal is sent in data frame structure
Regular data frame.
3. the ultrashort wave signal specific recognition methods according to claim 1 based on convolutional neural networks, which is characterized in that
Convolutional neural networks model includes the input layer set gradually, convolutional layer one, convolutional layer two, mixed layer one, mixed layer two, mixing
Layer three, convolutional layer three, average pond layer and output layer, wherein convolutional layer one, mixed layer one, mixed layer two, mixes convolutional layer two
Close the convolution kernel that different size size is used in layer three, convolutional layer three and average pond layer.
4. the ultrashort wave signal specific recognition methods according to claim 3 based on convolutional neural networks, which is characterized in that
The convolution kernel size of convolutional layer one, convolutional layer two, mixed layer one, mixed layer two, mixed layer three, convolutional layer three and average pond layer
Successively successively decrease.
5. the ultrashort wave signal specific recognition methods according to claim 3 based on convolutional neural networks, which is characterized in that
Convolutional layer three connects average pond layer 1*1 convolution kernel using 3*3 convolution kernel.
6. the ultrashort wave signal specific recognition methods according to claim 3 based on convolutional neural networks, which is characterized in that
It is averaged pond double-layer structure in mixed layer three using convolution sum.
7. the ultrashort wave signal specific recognition methods according to claim 3 based on convolutional neural networks, which is characterized in that
Mixed layer one, mixed layer two and mixed layer three are all made of Inception module and optimize.
8. the ultrashort wave signal specific recognition methods according to claim 3 based on convolutional neural networks, which is characterized in that
The training and identification of full convolutional neural networks end to end are realized in output layer using global average pond and classification output.
9. the ultrashort wave signal specific recognition methods according to claim 1 based on convolutional neural networks, which is characterized in that
Before when convolutional neural networks model being trained and/or identified using time-frequency map, time-frequency map is cut out at random first
It cuts and increases the received data scale of convolutional neural networks mode input layer.
10. the ultrashort wave signal specific recognition methods according to claim 1 based on convolutional neural networks, feature exist
In both mixed layer one and mixed layer three structure introduce great-jump-forward transmitting and carries out residual noise reduction, gradient when alleviating training and identification
Disperse.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811159958.XA CN109379311B (en) | 2018-09-30 | 2018-09-30 | Ultra-short wave specific signal identification method based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811159958.XA CN109379311B (en) | 2018-09-30 | 2018-09-30 | Ultra-short wave specific signal identification method based on convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109379311A true CN109379311A (en) | 2019-02-22 |
| CN109379311B CN109379311B (en) | 2021-08-17 |
Family
ID=65403407
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811159958.XA Active CN109379311B (en) | 2018-09-30 | 2018-09-30 | Ultra-short wave specific signal identification method based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109379311B (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110048977A (en) * | 2019-03-14 | 2019-07-23 | 中国人民解放军战略支援部队信息工程大学 | The recognition methods of short-wave signal system and device based on the detection of gray level co-occurrence matrixes textural characteristics |
| CN110336638A (en) * | 2019-04-17 | 2019-10-15 | 四川大学 | A kind of short-term burst signal detecting method based on time-frequency figure |
| CN110927706A (en) * | 2019-12-10 | 2020-03-27 | 电子科技大学 | Convolutional neural network-based radar interference detection and identification method |
| CN111049595A (en) * | 2019-11-29 | 2020-04-21 | 西安冉科信息技术有限公司 | Signal identification and acquisition system |
| CN111464468A (en) * | 2020-03-02 | 2020-07-28 | 南方科技大学 | Signal modulation mode identification method, device, equipment and readable storage medium |
| CN111562597A (en) * | 2020-06-02 | 2020-08-21 | 南京敏智达科技有限公司 | Beidou satellite navigation interference source identification method based on BP neural network |
| CN112039818A (en) * | 2020-08-31 | 2020-12-04 | 中原工学院 | Carrier synchronization method based on frequency domain shift average periodogram and neural network decoding |
| CN112634927A (en) * | 2020-12-03 | 2021-04-09 | 电子科技大学 | Short wave channel voice enhancement method |
| CN113312996A (en) * | 2021-05-19 | 2021-08-27 | 哈尔滨工程大学 | Detection and identification method for aliasing short-wave communication signals |
| CN114157544A (en) * | 2021-12-07 | 2022-03-08 | 中南大学 | Frame synchronization method, device and medium based on convolutional neural network |
| CN114205821A (en) * | 2021-11-30 | 2022-03-18 | 广州万城万充新能源科技有限公司 | Wireless radio frequency anomaly detection method based on depth prediction coding neural network |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101662328A (en) * | 2009-09-23 | 2010-03-03 | 成都林海电子有限责任公司 | Performance detector of portable ultrashort wave frequency hopping station |
| CN106546892A (en) * | 2016-11-10 | 2017-03-29 | 华乘电气科技（上海）股份有限公司 | The recognition methodss of shelf depreciation ultrasonic audio and system based on deep learning |
| CN107122738A (en) * | 2017-04-26 | 2017-09-01 | 成都蓝色起源科技有限公司 | Automatic Communication Signals Recognition based on deep learning model and its realize system |
| WO2017200782A1 (en) * | 2016-05-19 | 2017-11-23 | Microsoft Technology Licensing, Llc | Permutation invariant training for talker-independent multi-talker speech separation |
| CN108538011A (en) * | 2018-03-07 | 2018-09-14 | 北京宇航时代科技发展有限公司 | A kind of laser radar fire detection method |
| CN108599863A (en) * | 2018-04-26 | 2018-09-28 | 王晓楠 | Overhead transmission line wind based on shallow-layer CNN disaggregated models waves monitoring and pre-warning system and method |
-
2018
- 2018-09-30 CN CN201811159958.XA patent/CN109379311B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101662328A (en) * | 2009-09-23 | 2010-03-03 | 成都林海电子有限责任公司 | Performance detector of portable ultrashort wave frequency hopping station |
| WO2017200782A1 (en) * | 2016-05-19 | 2017-11-23 | Microsoft Technology Licensing, Llc | Permutation invariant training for talker-independent multi-talker speech separation |
| CN106546892A (en) * | 2016-11-10 | 2017-03-29 | 华乘电气科技（上海）股份有限公司 | The recognition methodss of shelf depreciation ultrasonic audio and system based on deep learning |
| CN107122738A (en) * | 2017-04-26 | 2017-09-01 | 成都蓝色起源科技有限公司 | Automatic Communication Signals Recognition based on deep learning model and its realize system |
| CN108538011A (en) * | 2018-03-07 | 2018-09-14 | 北京宇航时代科技发展有限公司 | A kind of laser radar fire detection method |
| CN108599863A (en) * | 2018-04-26 | 2018-09-28 | 王晓楠 | Overhead transmission line wind based on shallow-layer CNN disaggregated models waves monitoring and pre-warning system and method |
Non-Patent Citations (6)
| Title |
|---|
| CHRISTIAN SZEGEDY等: "Going Deeper with Convolutions", 《2015 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)》 * |
| KAIMING HE等: "Deep Residual Learning for Image Recognition", 《IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| KAREN SIMONYAN等: "Very Deep Convolutional Networks for Large-Scale Image Recognition", 《HTTP://ARXIV.ORG/PDF/1409.1556.PDF》 * |
| LIGANG GAO等: "Demonstration of Convolution Kernel Operation on Resistive Cross-Point Array", 《IEEE ELECTRON DEVICE LETTERS》 * |
| MATTHEW D. ZEILER等: "Stochastic Pooling for Regularization of Deep Convolutional Neural Networks", 《HTTPS：//ARXIV.ORG/ABS/1301.3557.PDF》 * |
| 雪饼: "大话CNN经典模型：GoogLeNet（从Inception v1到v4的演进）", 《HTTPS://MY.OSCHINA.NET/U/876354/BLOG/1637819》 * |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110048977A (en) * | 2019-03-14 | 2019-07-23 | 中国人民解放军战略支援部队信息工程大学 | The recognition methods of short-wave signal system and device based on the detection of gray level co-occurrence matrixes textural characteristics |
| CN110048977B (en) * | 2019-03-14 | 2022-03-01 | 中国人民解放军战略支援部队信息工程大学 | Short wave signal system identification method and device based on gray level co-occurrence matrix texture feature detection |
| CN110336638A (en) * | 2019-04-17 | 2019-10-15 | 四川大学 | A kind of short-term burst signal detecting method based on time-frequency figure |
| CN111049595A (en) * | 2019-11-29 | 2020-04-21 | 西安冉科信息技术有限公司 | Signal identification and acquisition system |
| CN110927706A (en) * | 2019-12-10 | 2020-03-27 | 电子科技大学 | Convolutional neural network-based radar interference detection and identification method |
| CN110927706B (en) * | 2019-12-10 | 2022-05-24 | 电子科技大学 | Convolutional neural network-based radar interference detection and identification method |
| CN111464468A (en) * | 2020-03-02 | 2020-07-28 | 南方科技大学 | Signal modulation mode identification method, device, equipment and readable storage medium |
| CN111562597A (en) * | 2020-06-02 | 2020-08-21 | 南京敏智达科技有限公司 | Beidou satellite navigation interference source identification method based on BP neural network |
| CN112039818A (en) * | 2020-08-31 | 2020-12-04 | 中原工学院 | Carrier synchronization method based on frequency domain shift average periodogram and neural network decoding |
| CN112634927A (en) * | 2020-12-03 | 2021-04-09 | 电子科技大学 | Short wave channel voice enhancement method |
| CN112634927B (en) * | 2020-12-03 | 2022-07-29 | 电子科技大学 | Short wave channel voice enhancement method |
| CN113312996A (en) * | 2021-05-19 | 2021-08-27 | 哈尔滨工程大学 | Detection and identification method for aliasing short-wave communication signals |
| CN114205821A (en) * | 2021-11-30 | 2022-03-18 | 广州万城万充新能源科技有限公司 | Wireless radio frequency anomaly detection method based on depth prediction coding neural network |
| CN114205821B (en) * | 2021-11-30 | 2023-08-08 | 广州万城万充新能源科技有限公司 | Wireless radio frequency anomaly detection method based on depth prediction coding neural network |
| CN114157544A (en) * | 2021-12-07 | 2022-03-08 | 中南大学 | Frame synchronization method, device and medium based on convolutional neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109379311B (en) | 2021-08-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109379311A (en) | Ultrashort wave signal specific recognition methods based on convolutional neural networks | |
| CN102279390B (en) | Intra-pulse modulation and recognition method of low signal-to-noise radar radiation source signal | |
| CN109344772A (en) | Ultrashort wave signal specific reconnaissance method based on spectrogram and depth convolutional network | |
| Qu et al. | Radar signal intra-pulse modulation recognition based on convolutional denoising autoencoder and deep convolutional neural network | |
| Gao et al. | Fusion image based radar signal feature extraction and modulation recognition | |
| CN109887047B (en) | Signal-image translation method based on generation type countermeasure network | |
| Xiao et al. | Radar signal recognition based on transfer learning and feature fusion | |
| CN109543643A (en) | Carrier signal detection method based on one-dimensional full convolutional neural networks | |
| CN113033473B (en) | ST2DCNN + SE-based radar overlapping signal identification method | |
| Jiang et al. | Continuous IFF response signal recognition technology based on capsule network | |
| CN110276256A (en) | Based on the low signal-to-noise ratio Modulation Recognition of Communication Signal method and device for adjusting ginseng accidental resonance | |
| Qin et al. | Radar waveform recognition based on deep residual network | |
| Ristea et al. | Estimating the magnitude and phase of automotive radar signals under multiple interference sources with fully convolutional networks | |
| Sun et al. | Radar emitter individual identification based on convolutional neural network learning | |
| CN113780521B (en) | Radiation source individual identification method based on deep learning | |
| CN116047427A (en) | Small sample radar active interference identification method | |
| Zhu et al. | A probe-feature for specific emitter identification using axiom-based grad-CAM | |
| Xia et al. | Radar waveform recognition based on a two‐stream convolutional network and software defined radio | |
| CN115951315B (en) | Radar spoofing interference identification method and system based on improved wavelet packet energy spectrum | |
| CN116449328A (en) | Radar working mode identification method and system based on time-frequency analysis feature fusion | |
| CN115600101A (en) | Unmanned aerial vehicle signal intelligent detection method and device based on priori knowledge | |
| CN114627889A (en) | Multi-sound-source sound signal processing method and device, storage medium and electronic equipment | |
| Liu et al. | Radar signal waveform recognition based on convolutional denoising autoencoder | |
| Huang et al. | A method for extracting fingerprint feature of communication satellite signal | |
| CN115409058B (en) | Anti-disturbance generation method and system for automatic modulation recognition depth network |
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 |