CN107203782A - Communication interference signals recognition methods under Larger Dynamic signal to noise ratio based on convolutional neural networks - Google Patents
Communication interference signals recognition methods under Larger Dynamic signal to noise ratio based on convolutional neural networks Download PDFInfo
- Publication number
- CN107203782A CN107203782A CN201710370299.3A CN201710370299A CN107203782A CN 107203782 A CN107203782 A CN 107203782A CN 201710370299 A CN201710370299 A CN 201710370299A CN 107203782 A CN107203782 A CN 107203782A
- Authority
- CN
- China
- Prior art keywords
- interference
- layer
- neural networks
- convolutional neural
- sample
- 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
Links
Classifications
-
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Feedback Control In General (AREA)
Abstract
Communication interference signals recognition methods under Larger Dynamic signal to noise ratio based on convolutional neural networks, is related to interference signal type identification field.The present invention is to extract the features such as statistic and Higher Order Cumulants using artificial to solve the existing type identification mode to interference signal, and existing characteristics extract that difficulty is big, and form is complicated, and the accuracy rate classified to interference signal it is low the problem of.The interference signal that the present invention can be recognized is 5 kinds of common interference signals and the combination of two of this 5 kinds of interference signals, it is 15 kinds of interference signals altogether, 5 kinds of interference signals are respectively audio disturbances, with the interference of frequency range arrowband, Sweeping nonlinearity, rectangular pulse interference and spread spectrum interference, have the feature of strong robustness under Larger Dynamic signal to noise ratio by building convolutional neural networks model extraction interference signal；By building the support vector machine classifier of 15 classification, this 15 kinds of interference signals are classified, it is used to classify to interference signal.
Description
Technical field
The present invention relates to communication interference signals recognition methods under the Larger Dynamic signal to noise ratio based on convolutional neural networks.Belong to dry
Disturb radar recognition field.
Background technology
Because modern communication technology is using open communications channel, communication system probably suffer from intentional or nothing
The interference of meaning.Therefore, either civilian or military systems are all had to using effective anti-interference method come in suppression system
Interference.However, the interference signal complexity of practical communication system is various, some interference protection measures are inevitably to useful signal
Produce infringement.Moreover, a kind of disturbance restraining method is generally only effective to a kind of interference type, due to interference in practical communication confrontation
Means complexity is various, receives in signal and there may be unknown one or more interference types, therefore, practicable interference suppression
System must be docked collect mail number in interference type carry out accurate Classification and Identification in real time, increase the purpose and effectively of AF panel
Property.
The type identification of interference signal is a kind of technology classified to interference signal.At present, interference type identification master
It is divided into two parts of feature extraction and application tagsort, characteristic extraction part can be divided into artificial extraction feature and automatic again
Feature is extracted, conventional method is mostly manually to extract the features such as statistic and Higher Order Cumulants, its feature extraction difficulty is big,
Form is complex.
The content of the invention
The present invention be in order to solve the existing type identification mode to interference signal use it is artificial extract statistic with
And the feature such as Higher Order Cumulants, existing characteristics extract that difficulty is big, and form is complicated, and the accuracy rate classified to interference signal is low
Problem.Communication interference signals recognition methods under Larger Dynamic signal to noise ratio based on convolutional neural networks is now provided.
Communication interference signals recognition methods under Larger Dynamic signal to noise ratio based on convolutional neural networks, it comprises the following steps：
Step 1: collecting the interference signal in any SNR ranges, interference sample is used as；
Step 2: the interference sample in step one is filtered, sample is disturbed from convolutional neural networks by filtered
In input layer travel to softmax layers, the error between sample classification result is disturbed in softmax layer of contrast and step one,
Round-off error value, obtaining one group is used in the adjacent layer of classification of disturbance nerve in connection parameter and each layer between neuron elements
The weights offset parameter of first unit, wherein softmax layers is used to classify to data；
Contain the connection parameter between neuron elements in each layer and each Step 3: interference signal to be sorted is input to
In layer in the convolutional neural networks of the weights offset parameter of neuron elements, by the full articulamentum output characteristic of convolutional neural networks
Value, this feature value is sent into the support vector machine classifier classified for interference signal, according to support vector machine classifier
Output valve divides the classification of interference signal to be sorted, so as to realize the type identification to interference signal to be sorted.
Preferably, in step 2, the generation of the error in softmax layers of contrast and step one between interference sample classification result
Valency function J (W, b) is realized,
In formula, the parameter that λ decays for weight, λ=0, J (W, b) is least mean-square error, W is the weight matrix of neuron,
B offsets for the weights of each neuron, and m indicates the quantity of training sample, hW,b(x(i)) it is in the case where input sample is x
The output valve of current convolutional neural networks, y(i)For the preferable output valve of convolutional neural networks, nlFor the number of plies, slFor l layers of nerve
The quantity of the unit of member,To couple parameter between l layers of jth unit and l+1 layers of i-th cell.
Preferably, in step 2, using stochastic gradient descent method round-off error value, obtaining one group is used for the phase of classification of disturbance
The weights offset parameter of neuron elements in connection parameter and each layer in adjacent bed between neuron elements：
In formula, lrFor learning rate, ρ is momentum,For the weight drift of l layers of i-th cell.
Preferably, in step 2, convolutional neural networks have 15 layers, are arranged in order as input layer, 2 convolutional layers, 1 pond
Change layer, 1 convolutional layer, 1 pond layer, 1 full articulamentum, the 1 batch of standardization layer, 1 full articulamentum, 1 criticize the layer that standardizes,
1 full articulamentum, 1 batch of standardization layer, 1 full articulamentum, 1 batch of standardization layer and 1 softmax layers,
Disturb sample by input layer input to be entered in 3 convolutional layers, adjacent convolutional layer partly to connect between neuron elements,
All connected between neuron elements in adjacent full articulamentum, the output end of the output end of each convolutional layer and each full articulamentum
An activation primitive is connected respectively,
Convolutional layer is used to extract the characteristic parameter in interference sample, and pond layer is used to enter the characteristic parameter that convolutional layer is extracted
Row dimensionality reduction, full articulamentum is used to receive the data after dimensionality reduction, and parameter extraction is carried out again,
Criticizing standardization layer is used in each stochastic gradient descent method, and standardization behaviour is to corresponding activation primitive output valve
Make so that the average of each dimension of output signal is 0, and variance is 1.
Preferably, in step 3, the class of interference signal to be sorted is divided according to the output valve of support vector machine classifier
Not, so that the detailed process for realizing the type identification to interference signal to be sorted is：
Disturb sample to include k classification, a support vector machine classifier, k class are set between any two classes sample
Other interference sample needs k (k-1)/2 support vector machine classifier,
Each support vector machine classifier uses Radial basis kernel function k (x1,x2)：
In formula, x1、x2For two input parameters of Radial basis kernel function, σ=1/15,
Each support vector machine classifier introduces slack variable ε in trainingi, slack variable εiRestrictive condition is：
yi(uTxi+b)≥1-εiFormula five；
In formula, xiFor sample, yiFor the tag along sort of sample, value is that 1 or -1, T is vectorial transposition,
Make optimal function J (u) in formula yi(uTxi+b)≥1-εiRestrictive condition under get minimum value：
One group of parameter u to be optimized is found,
In formula, C is penalty factor,
From the output function f (x) of each support vector machine classifier：
F (x)=k (u, x)+b formula seven；
In formula, k (u, x) is Radial basis kernel function, and input parameter x is characterized value, and b is biasing,
It is a class to obtain numerical value identical in the output valve of each support vector machine classifier, output valve, according to multiple values
It is middle the classification that identical value determines the sample of interference signal occur.
Preferably, in step 2, the interference sample in step one is filtered and realized using FIR bandpass filters.
Beneficial effects of the present invention are：
The interference signal in any SNR ranges is collected, as known sample, with known sample training convolutional nerve
Network, obtains the unknown parameter in convolutional neural networks：It is refreshing in connection parameter and each layer in adjacent layer between neuron elements
Weights offset parameter through first unit, interference signal to be sorted is input to train contain each layer in neuron elements it
Between connection parameter and each layer in neuron elements weights offset parameter convolutional neural networks in, the characteristic value of output feeding
In support vector machine classifier, a support vector machine classifier is designed between any two kinds of samples, according to SVMs point
The output valve of class device divides the classification of interference signal to be sorted, so as to realize the type identification to interference signal to be sorted.
The interference signal that the present invention can be recognized is 5 kinds of common interference signals and the group two-by-two of this 5 kinds of interference signals
Close, be 15 kinds of interference signals altogether, 5 kinds of interference signals are respectively audio disturbances, with the interference of frequency range arrowband, Sweeping nonlinearity, rectangle
Impulse disturbances and spread spectrum interference.
The depth model being made up of multilayer convolutional neural networks, with it is automatic obtain by it is rudimentary to it is senior, by simple to answering
It is miscellaneous, general arrive special feature the characteristics of.Level is lower, and the feature extracted is simpler, more general, and progressively extracts target times
The feature for correlation of being engaged in.
Interference signal identification based on convolutional neural networks (CNN) automatically extracts signal characteristic using CNN, most believes at last
Number feeding support vector machine classifier (SVM classifier) form for being classified, reduces traditional complicated artificial extraction spy
The part levied, considerably increases convenient degree, and the more traditional method of classification accuracy improves more than 40%.
Brief description of the drawings
Fig. 1 is communication interference signals under the Larger Dynamic signal to noise ratio based on convolutional neural networks described in embodiment one
The flow chart of recognition methods；
The structure chart of convolutional neural networks when Fig. 2 is training；
The structure chart of convolutional neural networks when Fig. 3 is non-training.
Embodiment
Communication interference signals recognition methods under Larger Dynamic signal to noise ratio based on convolutional neural networks, it comprises the following steps：
Step 1: collecting the interference signal in any SNR ranges, interference sample is used as；
Step 2: the interference sample in step one is filtered, sample is disturbed from convolutional neural networks by filtered
In input layer travel to softmax layers, the error between sample classification result is disturbed in softmax layer of contrast and step one,
Round-off error value, obtaining one group is used in the adjacent layer of classification of disturbance nerve in connection parameter and each layer between neuron elements
The weights offset parameter of first unit, wherein softmax layers is used to classify to data；
Contain the connection parameter between neuron elements in each layer and each Step 3: interference signal to be sorted is input to
In layer in the convolutional neural networks of the weights offset parameter of neuron elements, by the full articulamentum output characteristic of convolutional neural networks
Value, this feature value is sent into the support vector machine classifier classified for interference signal, according to support vector machine classifier
Output valve divides the classification of interference signal to be sorted, so as to realize the type identification to interference signal to be sorted.
The effect of present embodiment is：The interference signal in any SNR ranges is collected, as known sample, with this
Know sample training convolutional neural networks, obtain the unknown parameter in convolutional neural networks：In adjacent layer between neuron elements
Couple the weights offset parameter of neuron elements in parameter and each layer, interference signal to be sorted is input to containing of training
The convolutional neural networks of the weights offset parameter of neuron elements in connection parameter and each layer in each layer between neuron elements
In, classified in the characteristic value feeding support vector machine classifier of output.
The depth model being made up of multilayer convolutional neural networks, with it is automatic obtain by it is rudimentary to it is senior, by simple to answering
It is miscellaneous, general arrive special feature the characteristics of.Level is lower, and the feature extracted is simpler, more general, and progressively extracts target times
The feature for correlation of being engaged in.
Interference signal identification based on convolutional neural networks (CNN) automatically extracts signal characteristic using CNN, most believes at last
Number form classified of feeding SVM classifier, reduces the part of traditional complicated artificial extraction feature, considerably increases
Convenient degree, and the more traditional method of classification accuracy improves more than 40%.
In preferred embodiment, in step 2, the mistake in softmax layers of contrast and step one between interference sample classification result
Difference using least mean-square error cost function J (W, b) realize,
In formula, the parameter that λ decays for weight, λ=0, J (W, b) is least mean-square error, W is the weight matrix of neuron,
B offsets for the weights of each neuron, and m indicates the quantity of training sample, hW,b(x(i)) it is in the case where input sample is x
The output valve of current convolutional neural networks, y(i)For the preferable output valve of convolutional neural networks, nlFor the number of plies, slFor l layers of nerve
The quantity of the unit of member,To couple parameter between l layers of jth unit and l+1 layers of i-th cell.
In present embodiment, the interference sample in step one is known sample, and carrying out classification to the sample carries out label.
The effect of present embodiment is：Section 1 is square item behind equal sign in formula one, and Section 2 is weight attenuation term,
So, using cost function J (W, b) error between contrast batch standardization layer and input the layer data mesh of least mean-square error
Be to prevent over-fitting.
In preferred embodiment, in step 2, using stochastic gradient descent method round-off error value, obtain one group be used for disturb divide
The weights offset parameter of neuron elements in connection parameter and each layer in the adjacent layer of class between neuron elements：
In formula, lrFor learning rate, ρ is momentum,For the weight drift of l layers of i-th cell.
The effect of present embodiment is：For parameter W update methods,For parameter B update method, and change every time
For when, upset training sample, 128 training samples of every batch of taking-up, using randomly select 128 samples calculate gradient with
Machine gradient descent method (SGD).During this model training, learning rate lrValue be 0.001.In gradient updating, use
Momentum (momentum) method is optimized to update mode, i.e. ρ in formula, and value is 0.3.
In preferred embodiment, in step 2, convolutional neural networks have 15 layers, are arranged in order as input layer, 2 convolution
Layer, 1 pond layer, 1 convolutional layer, 1 pond layer, 1 full articulamentum, 1 criticize the layer that standardizes, 1 full articulamentum, 1 criticize
Standardize layer, 1 full articulamentum, 1 batch of standardization layer, 1 full articulamentum, 1 batch of standardize layer and 1 softmax layers,
Disturb sample by input layer input to be entered in 3 convolutional layers, adjacent convolutional layer partly to connect between neuron elements,
All connected between neuron elements in adjacent full articulamentum, the output end of the output end of each convolutional layer and each full articulamentum
An activation primitive is connected respectively,
Convolutional layer is used to extract the characteristic parameter in interference sample, and pond layer is used to enter the characteristic parameter that convolutional layer is extracted
Row dimensionality reduction, full articulamentum is used to receive the data after dimensionality reduction, and parameter extraction is carried out again,
Criticizing standardization layer is used in each stochastic gradient descent method, and standardization behaviour is to corresponding activation primitive output valve
Make so that the average of each dimension of output signal is 0, and variance is 1.
The effect of present embodiment is：Each layer effect of convolutional neural networks when training：
Full articulamentum：Feature is extracted, with the characteristic connected entirely between neuron, computation complexity is big, and efficiency is low.
Convolutional layer：It is also to extract feature, but is partly connected between neuron, each implicit unit can only connects defeated
Enter a part for unit, reduce the complexity of calculating.
Criticize standardization layer (Normalization layers of Batch) (BN layers)：In each stochastic gradient descent method (SGD),
Standardized operation is done to corresponding activation primitive output valve so that the average of each dimension of output signal is 0, and variance is 1, every layer
The activation primitive of neutral net is index linear unit (ELU), and the effect of activation primitive is：To linear model introduce it is non-linear because
Element, strengthens the ability to express of model.
Softmax layers：Logistic regression is promoted how classificatory, is a polytypic grader.
By the convolutional neural networks structure of multilayer, by it is rudimentary to it is senior, from simple to complex, by it is general to it is special by
Step extracts characteristic value, reduces traditional complicated artificial extracting mode, considerably increases convenient degree.
In preferred embodiment, in step 3, interference to be sorted is divided according to the output valve of support vector machine classifier and believed
Number classification so that the detailed process for realizing the type identification to interference signal to be sorted is：
Disturb sample to include k classification, a support vector machine classifier, k class are set between any two classes sample
Other interference sample needs k (k-1)/2 support vector machine classifier,
Each support vector machine classifier uses Radial basis kernel function k (x1,x2)：
In formula, x1、x2For two input parameters of Radial basis kernel function, σ=1/15,
Each support vector machine classifier introduces slack variable ε in trainingi, slack variable εiRestrictive condition is：
yi(uTxi+b)≥1-εiFormula five；
In formula, xiFor sample, yiFor the tag along sort of sample, value is that 1 or -1, T is vectorial transposition,
Make optimal function J (u) in formula yi(uTxi+b)≥1-εiRestrictive condition under get minimum value：
One group of parameter u to be optimized is found,
In formula, C is penalty factor,
From the output function f (x) of each support vector machine classifier：
F (x)=k (u, x)+b formula seven；
In formula, k (u, x) is Radial basis kernel function, and input parameter x is characterized value, and b is biasing；
It is a class to obtain numerical value identical in the output valve of each support vector machine classifier, output valve, according to multiple values
It is middle the classification that identical value determines the sample of interference signal occur.
The effect of present embodiment is：It will be combined, classified with support vector machine classifier based on convolutional neural networks
The more traditional method of accuracy rate improve 40%.
In preferred embodiment, in step 2, the interference sample in step one is filtered real using FIR bandpass filters
It is existing.
The effect of present embodiment is：The interference signal in any frequency range is filtered using FIR bandpass filters.
Embodiment：
Fig. 1 shows communication interference signals recognition methods under the Larger Dynamic signal to noise ratio based on convolutional neural networks in embodiment
Flow chart.Communication interference signals recognition methods is used to extract interference letter under the Larger Dynamic signal to noise ratio based on convolutional neural networks
Number, then interference signal is classified, so as to carry out effective suppression to corresponding interference type.
Fig. 2 shows the structure chart of convolutional neural networks when being trained in embodiment, the concrete structure of convolutional neural networks
Parameter is as shown in table 1.
The structure chart of convolutional neural networks when Fig. 3 shows non-training in embodiment,
The convolutional neural networks network architecture parameters of table 1
Referring to figs. 1 to Fig. 3, communication interference signals are known under the Larger Dynamic signal to noise ratio based on convolutional neural networks of the present embodiment
Other method includes step one to step 3.
Step 1: collecting every kind of interference under 15 kinds of interference signals that SNR ranges are -10dB~10dB, each signal to noise ratio
900 samples are had, 15 kinds of interference signals have 15 × 21 × 900 interference samples in the range of -10dB~10dB signal to noise ratio
This；
Step 2: all interference samples of 15 kinds of interference signals in step one are separately input into FIR bandpass filters
In, the Structural assignments that the interference sample that FIR bandpass filters are exported is entered in convolutional neural networks, convolutional neural networks are successively
For input layer, 2 convolutional layers, 1 pond layer, 1 convolutional layer, 1 pond layer, 1 full articulamentum, 1 batch of standardization layer, 1
Individual full articulamentum, 1 BN layers, 1 full articulamentum, 1 BN layers, 1 full articulamentum, 1 BN layers and softmax layers, using most
(W b) contrasts the error between interference sample classification result in softmax layers and step one to the cost function J of small mean square error：
In formula, the parameter that λ decays for weight, λ=0, J (W, b) is least mean-square error, W is the weight matrix of neuron,
B offsets for the weights of each neuron, and m indicates the quantity of training sample, hW,b(x(i)) it is in the case where input sample is x
The output valve of current convolutional neural networks, y(i)For the preferable output valve of convolutional neural networks, nlFor the number of plies, slFor l layers of nerve
The quantity of the unit of member,To couple parameter between l layers of jth unit and l+1 layers of i-th cell,
Using stochastic gradient descent method round-off error value：
Obtaining one group is used for the connection parameter in the adjacent layer of classification of disturbance between neuron elementsWith nerve in each layer
The weights offset parameter of first unit
In formula, lrFor learning rate, value is 0.001；ρ is momentum, and value is 0.3；J (W, b) is least mean-square error,
Drifted about for the weight of l layers of i-th cell,
Step 3: 15 kinds of interference signals to be sorted are separately input to containing the connection between neuron elements in each layer
In parameter and each layer in the convolutional neural networks of the weights offset parameter of neuron elements, by the full articulamentum of convolutional neural networks
15 characteristic values are exported altogether, and 15 characteristic values are sent into the support vector machine classifier for 15 kinds of interference signal classification,
A support vector machine classifier is designed between any two kinds of samples, disturbing signal for 15 kinds needs 105 support vector cassifications
Device,
Each support vector machine classifier uses Radial basis kernel function k (x1,x2)：
In formula, x1、x2For two input parameters of Radial basis kernel function, σ=1/15,
Each support vector machine classifier introduces slack variable ε in trainingi, its restrictive condition is：
yi(uTxi+b)≥1-εiFormula five；
In formula, xiFor sample, yiFor the tag along sort of sample, value is that 1 or -1, T is vectorial transposition,
According to the cost function J (u) that need to optimize：
One group of parameter u to be optimized is found,
In formula, C is penalty factor, and penalty factor value is 1.0, εiFor slack variable,
Make optimal function J (u) in formula yi(uTxi+b)≥1-εiRestrictive condition under get minimum value,
From the output function f (x) of each support vector machine classifier：
F (x)=k (u, x)+b formula seven；
In formula, k (u, x) is Radial basis kernel function, and input parameter x is characterized value, and b is biasing,
It is a class to obtain numerical value identical in the output valve of each support vector machine classifier, output valve, according to multiple values
It is middle the classification that identical value determines the sample of 15 kinds of interference signals occur.
Claims (6)
1. communication interference signals recognition methods under the Larger Dynamic signal to noise ratio based on convolutional neural networks, it is characterised in that it includes
Following steps：
Step 1: collecting the interference signal in any SNR ranges, interference sample is used as；
Step 2: the interference sample in step one is filtered, by filtered interference sample from convolutional neural networks
Input layer travels to the error between interference sample classification result, amendment in softmax layers, softmax layers of contrast and step one
Error amount, obtaining one group is used in the adjacent layer of classification of disturbance neuron list in connection parameter and each layer between neuron elements
The weights offset parameter of member, wherein softmax layers is used to classify to data；
Step 3: interference signal to be sorted is input to containing in the connection parameter and each layer between neuron elements in each layer
In the convolutional neural networks of the weights offset parameter of neuron elements, by the full articulamentum output characteristic value of convolutional neural networks,
This feature value is sent into the support vector machine classifier classified for interference signal, according to the output of support vector machine classifier
Value divides the classification of interference signal to be sorted, so as to realize the type identification to interference signal to be sorted.
2. communication interference signals identification side under the Larger Dynamic signal to noise ratio according to claim 1 based on convolutional neural networks
Error in method, it is characterised in that in step 2, softmax layers of contrast and step one between interference sample classification result is used
Least mean-square error cost function J (W, b) realize,
In formula, λ is the parameter that weight decays, and (W b) is least mean-square error, W is the weight matrix of neuron, and b is to λ=0, J
The weights skew of each neuron, m indicates the quantity of training sample, hW,b(x(i)) be in the case where input sample is x when
The output valve of preceding convolutional neural networks, y(i)For the preferable output valve of convolutional neural networks, nlFor the number of plies, slFor l layers of neuron
Unit quantity,To couple parameter between l layers of jth unit and l+1 layers of i-th cell.
3. communication interference signals identification side under the Larger Dynamic signal to noise ratio according to claim 1 based on convolutional neural networks
Method, it is characterised in that in step 2, using stochastic gradient descent method round-off error value, obtaining one group is used for the phase of classification of disturbance
The weights offset parameter of neuron elements in connection parameter and each layer in adjacent bed between neuron elements：
In formula, lrFor learning rate, ρ is momentum,For the weight drift of l layers of i-th cell.
4. communication interference signals are recognized under the Larger Dynamic signal to noise ratio based on convolutional neural networks according to claim 1 or 3
Method, it is characterised in that in step 2, convolutional neural networks have 15 layers, be arranged in order for input layer, 2 convolutional layers, 1
Pond layer, 1 convolutional layer, 1 pond layer, 1 full articulamentum, 1 batch of standardization layer, 1 full articulamentum, 1 batch of standardization
Layer, 1 full articulamentum, 1 batch of standardization layer, 1 full articulamentum, 1 batch of standardization layer and 1 softmax layers,
Disturb sample by input layer input to be entered in 3 convolutional layers, adjacent convolutional layer partly to connect between neuron elements, it is adjacent
All connected between neuron elements in full articulamentum, the output end difference of the output end of each convolutional layer and each full articulamentum
An activation primitive is connected,
Convolutional layer is used to extract the characteristic parameter in interference sample, and pond layer is used to drop the characteristic parameter that convolutional layer is extracted
Dimension, full articulamentum is used to receive the data after dimensionality reduction, and parameter extraction is carried out again,
Criticizing standardization layer is used in each stochastic gradient descent method, and standardized operation is done to corresponding activation primitive output valve,
So that the average of each dimension of output signal is 0, variance is 1.
5. communication interference signals identification side under the Larger Dynamic signal to noise ratio according to claim 1 based on convolutional neural networks
Method, it is characterised in that in step 3, the class of interference signal to be sorted is divided according to the output valve of support vector machine classifier
Not, so that the detailed process for realizing the type identification to interference signal to be sorted is：
Sample is disturbed to include k classification, one support vector machine classifier of setting between any two classes sample, k classification
Interference sample needs k (k-1)/2 support vector machine classifier,
Each support vector machine classifier uses Radial basis kernel function k (x1,x2)：
In formula, x1、x2For two input parameters of Radial basis kernel function, σ=1/15,
Each support vector machine classifier introduces slack variable ε in trainingi, slack variable εiRestrictive condition is：
yi(uTxi+b)≥1-εiFormula five；
In formula, xiFor sample, yiFor the tag along sort of sample, value is that 1 or -1, T is vectorial transposition,
Make optimal function J (u) in formula yi(uTxi+b)≥1-εiRestrictive condition under get minimum value：
One group of parameter u to be optimized is found,
In formula, C is penalty factor,
From the output function f (x) of each support vector machine classifier：
F (x)=k (u, x)+b formula seven；
In formula, k (u, x) is Radial basis kernel function, and input parameter x is characterized value, and b is biasing,
It is a class to obtain numerical value identical in the output valve of each support vector machine classifier, output valve, is gone out according in multiple values
Existing identical value determines the classification of the sample of interference signal.
6. communication interference signals identification side under the Larger Dynamic signal to noise ratio according to claim 1 based on convolutional neural networks
Method, it is characterised in that in step 2, the interference sample in step one is filtered and realized using FIR bandpass filters.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710370299.3A CN107203782A (en) | 2017-05-23 | 2017-05-23 | Communication interference signals recognition methods under Larger Dynamic signal to noise ratio based on convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710370299.3A CN107203782A (en) | 2017-05-23 | 2017-05-23 | Communication interference signals recognition methods under Larger Dynamic signal to noise ratio based on convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107203782A true CN107203782A (en) | 2017-09-26 |
Family
ID=59906125
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710370299.3A Pending CN107203782A (en) | 2017-05-23 | 2017-05-23 | Communication interference signals recognition methods under Larger Dynamic signal to noise ratio based on convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107203782A (en) |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107832737A (en) * | 2017-11-27 | 2018-03-23 | 乐普（北京）医疗器械股份有限公司 | Electrocardiogram interference identification method based on artificial intelligence |
| CN108135003A (en) * | 2017-12-25 | 2018-06-08 | 广东海格怡创科技有限公司 | The construction method and system of interference type identification model |
| CN108197545A (en) * | 2017-12-25 | 2018-06-22 | 广东海格怡创科技有限公司 | The recognition methods of interference type and system |
| CN108509911A (en) * | 2018-04-03 | 2018-09-07 | 电子科技大学 | Interference signal recognition methods based on convolutional neural networks |
| CN109407067A (en) * | 2018-10-13 | 2019-03-01 | 中国人民解放军海军航空大学 | Radar moving targets detection and classification integral method based on time-frequency figure convolutional neural networks |
| CN109493847A (en) * | 2018-12-14 | 2019-03-19 | 广州玛网络科技有限公司 | Sound recognition system and voice recognition device |
| CN110047506A (en) * | 2019-04-19 | 2019-07-23 | 杭州电子科技大学 | A kind of crucial audio-frequency detection based on convolutional neural networks and Multiple Kernel Learning SVM |
| CN110197127A (en) * | 2019-05-06 | 2019-09-03 | 安徽继远软件有限公司 | Wireless signal detection and electromagnetic interference categorizing system and method based on deep learning |
| CN110557209A (en) * | 2019-07-19 | 2019-12-10 | 中国科学院微电子研究所 | Broadband signal interference monitoring method |
| CN110611627A (en) * | 2018-06-15 | 2019-12-24 | 维沃移动通信有限公司 | Signal detection method and receiving end |
| CN111049615A (en) * | 2018-10-15 | 2020-04-21 | 华为技术有限公司 | Method and apparatus for processing signals |
| CN111160317A (en) * | 2020-01-06 | 2020-05-15 | 西南电子技术研究所（中国电子科技集团公司第十研究所） | Weak signal blind extraction method |
| CN111245455A (en) * | 2020-02-19 | 2020-06-05 | 北京紫光展锐通信技术有限公司 | Dynamic interference suppression method for receiver, receiver system and storage medium |
| CN111539222A (en) * | 2020-05-20 | 2020-08-14 | 北京百度网讯科技有限公司 | Training method and device for semantic similarity task model, electronic equipment and storage medium |
| CN112435356A (en) * | 2020-11-04 | 2021-03-02 | 南京航天工业科技有限公司 | ETC interference signal identification method and detection system |
| CN112711984A (en) * | 2020-12-09 | 2021-04-27 | 北京航空航天大学 | Fixation point positioning method and device and electronic equipment |
| CN114301498A (en) * | 2021-12-30 | 2022-04-08 | 中国电子科技集团公司第五十四研究所 | Behavior identification method and system for frequency hopping communication radiation source |
| CN115276847A (en) * | 2022-07-12 | 2022-11-01 | 山东大学 | Radio frequency modulation fuse interference signal identification method based on convolutional neural network |
| WO2023045926A1 (en) * | 2021-09-23 | 2023-03-30 | 中兴通讯股份有限公司 | Interference signal avoidance method and apparatus, and base station and storage medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102982373A (en) * | 2012-12-31 | 2013-03-20 | 山东大学 | OIN (Optimal Input Normalization) neural network training method for mixed SVM (Support Vector Machine) regression algorithm |
| CN105320961A (en) * | 2015-10-16 | 2016-02-10 | 重庆邮电大学 | Handwriting numeral recognition method based on convolutional neural network and support vector machine |
| CN106599941A (en) * | 2016-12-12 | 2017-04-26 | 西安电子科技大学 | Method for identifying handwritten numbers based on convolutional neural network and support vector machine |
-
2017
- 2017-05-23 CN CN201710370299.3A patent/CN107203782A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102982373A (en) * | 2012-12-31 | 2013-03-20 | 山东大学 | OIN (Optimal Input Normalization) neural network training method for mixed SVM (Support Vector Machine) regression algorithm |
| CN105320961A (en) * | 2015-10-16 | 2016-02-10 | 重庆邮电大学 | Handwriting numeral recognition method based on convolutional neural network and support vector machine |
| CN106599941A (en) * | 2016-12-12 | 2017-04-26 | 西安电子科技大学 | Method for identifying handwritten numbers based on convolutional neural network and support vector machine |
Non-Patent Citations (3)
| Title |
|---|
| 刘念组主编: "《智能信息管理研究》", 31 December 2010, 立信会计出版社 * |
| 温媛媛等: "一种新型的干扰识别与检测算法", 《空间电子技术》 * |
| 胡志坚主编: "《技术前瞻与评价 第1卷 第2辑 2015版》", 31 October 2015, 科学技术文献出版 * |
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107832737A (en) * | 2017-11-27 | 2018-03-23 | 乐普（北京）医疗器械股份有限公司 | Electrocardiogram interference identification method based on artificial intelligence |
| CN107832737B (en) * | 2017-11-27 | 2021-02-05 | 上海优加利健康管理有限公司 | Artificial intelligence-based electrocardiogram interference identification method |
| CN108135003A (en) * | 2017-12-25 | 2018-06-08 | 广东海格怡创科技有限公司 | The construction method and system of interference type identification model |
| CN108197545A (en) * | 2017-12-25 | 2018-06-22 | 广东海格怡创科技有限公司 | The recognition methods of interference type and system |
| CN108509911A (en) * | 2018-04-03 | 2018-09-07 | 电子科技大学 | Interference signal recognition methods based on convolutional neural networks |
| CN108509911B (en) * | 2018-04-03 | 2020-06-12 | 电子科技大学 | Interference signal identification method based on convolutional neural network |
| CN110611627B (en) * | 2018-06-15 | 2021-03-19 | 维沃移动通信有限公司 | Signal detection method and receiving end |
| CN110611627A (en) * | 2018-06-15 | 2019-12-24 | 维沃移动通信有限公司 | Signal detection method and receiving end |
| CN109407067A (en) * | 2018-10-13 | 2019-03-01 | 中国人民解放军海军航空大学 | Radar moving targets detection and classification integral method based on time-frequency figure convolutional neural networks |
| CN111049615A (en) * | 2018-10-15 | 2020-04-21 | 华为技术有限公司 | Method and apparatus for processing signals |
| CN109493847A (en) * | 2018-12-14 | 2019-03-19 | 广州玛网络科技有限公司 | Sound recognition system and voice recognition device |
| CN109493847B (en) * | 2018-12-14 | 2019-10-18 | 广州一玛网络科技有限公司 | Sound recognition system and voice recognition device |
| CN110047506A (en) * | 2019-04-19 | 2019-07-23 | 杭州电子科技大学 | A kind of crucial audio-frequency detection based on convolutional neural networks and Multiple Kernel Learning SVM |
| CN110047506B (en) * | 2019-04-19 | 2021-08-20 | 杭州电子科技大学 | Key audio detection method based on convolutional neural network and multi-core learning SVM |
| CN110197127A (en) * | 2019-05-06 | 2019-09-03 | 安徽继远软件有限公司 | Wireless signal detection and electromagnetic interference categorizing system and method based on deep learning |
| CN110197127B (en) * | 2019-05-06 | 2022-10-18 | 安徽继远软件有限公司 | Wireless signal detection and electromagnetic interference classification system and method based on deep learning |
| CN110557209A (en) * | 2019-07-19 | 2019-12-10 | 中国科学院微电子研究所 | Broadband signal interference monitoring method |
| CN110557209B (en) * | 2019-07-19 | 2021-08-31 | 中国科学院微电子研究所 | Broadband signal interference monitoring method |
| CN111160317A (en) * | 2020-01-06 | 2020-05-15 | 西南电子技术研究所（中国电子科技集团公司第十研究所） | Weak signal blind extraction method |
| CN111160317B (en) * | 2020-01-06 | 2023-03-28 | 西南电子技术研究所（中国电子科技集团公司第十研究所） | Weak signal blind extraction method |
| CN111245455A (en) * | 2020-02-19 | 2020-06-05 | 北京紫光展锐通信技术有限公司 | Dynamic interference suppression method for receiver, receiver system and storage medium |
| CN111539222A (en) * | 2020-05-20 | 2020-08-14 | 北京百度网讯科技有限公司 | Training method and device for semantic similarity task model, electronic equipment and storage medium |
| CN111539222B (en) * | 2020-05-20 | 2023-05-23 | 北京百度网讯科技有限公司 | Training method, device, equipment and storage medium of semantic similarity task model |
| CN112435356A (en) * | 2020-11-04 | 2021-03-02 | 南京航天工业科技有限公司 | ETC interference signal identification method and detection system |
| CN112711984A (en) * | 2020-12-09 | 2021-04-27 | 北京航空航天大学 | Fixation point positioning method and device and electronic equipment |
| WO2023045926A1 (en) * | 2021-09-23 | 2023-03-30 | 中兴通讯股份有限公司 | Interference signal avoidance method and apparatus, and base station and storage medium |
| CN114301498A (en) * | 2021-12-30 | 2022-04-08 | 中国电子科技集团公司第五十四研究所 | Behavior identification method and system for frequency hopping communication radiation source |
| CN115276847A (en) * | 2022-07-12 | 2022-11-01 | 山东大学 | Radio frequency modulation fuse interference signal identification method based on convolutional neural network |
| CN115276847B (en) * | 2022-07-12 | 2023-11-24 | 山东大学 | Radio frequency modulation fuze interference signal identification method based on convolutional neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107203782A (en) | Communication interference signals recognition methods under Larger Dynamic signal to noise ratio based on convolutional neural networks | |
| CN108509911B (en) | Interference signal identification method based on convolutional neural network | |
| CN108984745A (en) | A kind of neural network file classification method merging more knowledge mappings | |
| CN110532932B (en) | Method for identifying multi-component radar signal intra-pulse modulation mode | |
| CN106373397B (en) | Remote sensing images road situation analysis method based on fuzzy neural network | |
| CN107633255A (en) | A kind of rock lithology automatic recognition classification method under deep learning pattern | |
| CN107885853A (en) | A kind of combined type file classification method based on deep learning | |
| CN108346293B (en) | Real-time traffic flow short-time prediction method | |
| CN109711314B (en) | Radar radiation source signal classification method based on feature fusion and SAE | |
| CN109116312A (en) | Radar signal classification method based on QMFB and convolutional neural networks | |
| CN115438708A (en) | Classification and identification method based on convolutional neural network and multi-mode fusion | |
| CN104318246A (en) | Depth self-adaption ridgelet network based polarimetric SAR (Synthetic Aperture Radar) image classification | |
| CN101241181A (en) | Non-library target range image discrimination method | |
| CN110048978A (en) | A kind of signal modulate method | |
| CN112765607A (en) | Neural network model backdoor attack detection method | |
| CN111709329A (en) | Unmanned aerial vehicle measurement and control signal high-speed identification method based on deep learning | |
| CN110764152B (en) | Device and method for rapid detection and identification of unmanned aerial vehicle | |
| CN105913090A (en) | SAR image object classification method based on SDAE-SVM | |
| CN115186712A (en) | Modulated signal identification method and system | |
| CN114980122A (en) | Small sample radio frequency fingerprint intelligent identification system and method | |
| CN113109782A (en) | Novel classification method directly applied to radar radiation source amplitude sequence | |
| CN113759326A (en) | Radar signal interference prediction method combining sparse autoencoder and improved LSTM | |
| Yin et al. | Co-channel multi-signal modulation classification based on convolution neural network | |
| CN117648948A (en) | CNN-LSTM network signal modulation identification method based on signal periodicity characteristics | |
| CN103020657A (en) | License plate Chinese character recognition method |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170926 |