CN114997248A - Model and method for identifying open set interference based on prototype learning - Google Patents

Model and method for identifying open set interference based on prototype learning Download PDF

Info

Publication number
CN114997248A
CN114997248A CN202210909603.8A CN202210909603A CN114997248A CN 114997248 A CN114997248 A CN 114997248A CN 202210909603 A CN202210909603 A CN 202210909603A CN 114997248 A CN114997248 A CN 114997248A
Authority
CN
China
Prior art keywords
prototype
interference
center
class
model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202210909603.8A
Other languages
Chinese (zh)
Other versions
CN114997248B (en
Inventor
陈杰
赵知劲
叶学义
岳克强
姜明
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hangzhou Dianzi University
Original Assignee
Hangzhou Dianzi University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hangzhou Dianzi University filed Critical Hangzhou Dianzi University
Priority to CN202210909603.8A priority Critical patent/CN114997248B/en
Publication of CN114997248A publication Critical patent/CN114997248A/en
Application granted granted Critical
Publication of CN114997248B publication Critical patent/CN114997248B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Landscapes

  • Image Analysis (AREA)

Abstract

The scheme discloses an open set interference recognition model and method based on prototype learning, and the model comprises a PL-Net with a feature extraction network and a classification model, wherein the feature extraction network extracts features from input interference, the classification model performs class recognition according to the distance between a feature vector and a prototype center of an interference class, a loss function for training the PL-Net comprises a closed set training item and an open set training item, a prototype center set of the open set training item comprises an unknown interference class prototype center, a prototype center set of the closed set training item has no unknown interference class prototype center, and a plurality of known interference class prototype centers and an unknown interference class prototype center are obtained through training. The loss function designed by the scheme is based on a prototype learning strategy, and the closed set training and the open set training are combined by assuming an open type prototype center, so that the unknown interference can be identified based on the prototype center, the model can be used for identifying the unknown interference, and the interference identification precision of the model under the open set condition is effectively improved.

Description

Model and method for identifying open set interference based on prototype learning
Technical Field
The invention belongs to the technical field of communication interference signal open recognition, and particularly relates to an open set interference recognition model and method based on prototype learning.
Background
Under the increasingly complex electromagnetic environment of a wireless communication system, and the influence of various interference signals and noises on communication equipment, various communication anti-interference technologies are developed in order to guarantee the reliability of communication, and the efficient and feasible anti-interference technology is particularly important for the anti-interference capability of military communication. However, in practice, an anti-interference strategy is usually effective only for a specific interference type, so that accurate identification of communication interference signals and adoption of corresponding anti-interference measures according to the identified interference signal type have important significance on communication anti-interference capability.
The traditional interference identification method comprises two parts of feature extraction and pattern identification, wherein the features are manually extracted according to time-frequency domain analysis, signal space analysis and the like of signals, and then the classification is carried out by using pattern identification methods such as decision trees, support vector machines and the like in the pattern identification. The methods depend on the effectiveness of manually extracting features, and generally face the problems of high algorithm complexity, incomplete feature information and the like. In order to solve the problem, a feature extraction method based on deep learning is proposed, the method based on deep learning can automatically extract high-dimensional features, can extract deep representation in signals, and avoids dependence on manually identified features. By utilizing the excellent feature extraction capability and classification performance of deep learning, the recognition effect superior to that of the traditional interference recognition algorithm is realized.
However, when new types of interferences are emerging, the existing interference identification method based on deep learning can only identify the known types of interferences, i.e. Closed Set Recognition (CSR). When a new interference mode occurs, the existing method can only identify the error as one of the known interference modes, but cannot judge the error as the unknown interference mode, so that the wrong anti-interference measures are used, and the anti-interference capability of the communication system is reduced. Therefore, how to accurately identify the known interference mode and effectively reject the unknown interference mode is an Open Set Registration (OSR), which is very important for improving the interference rejection capability of the communication system.
Disclosure of Invention
The invention aims to solve the problems and provides an open set interference recognition model and method based on prototype learning.
In order to achieve the purpose, the invention adopts the following technical scheme:
an open set interference recognition model based on prototype learning comprises a PL-Net with a feature extraction network and a classification model, wherein the feature extraction network is used for extracting features of input interference, the classification model is used for performing interference class recognition according to the distance between a feature vector output by the feature extraction network and a prototype center of a corresponding interference class, a loss function used for training the PL-Net comprises a closed set training item and an open set training item, a prototype center set of the open set training item comprises unknown interference class prototype centers, the prototype center set of the closed set training item has no unknown interference class prototype centers, and a plurality of known interference class prototype centers and an unknown interference class prototype center are obtained through training and used for known interference class classification and unknown interference class recognition.
In the above model for identifying interference in an open set based on prototype learning, the prototype center set of the closed set training term is a first prototype set in which all samples correspond to prototypes;
the prototype center set of the open-set training item is a second prototype set of prototypes corresponding to the removed one or more known samples.
In the above model for identifying open set interference based on prototype learning, the classification model calculates the distance between the feature vector of the input interference and the center of each prototype by the following method:
Figure 986173DEST_PATH_IMAGE001
（1）
Figure 187214DEST_PATH_IMAGE002
representing feature vectors and prototype centersThe distance between the two or more of the two or more,
Figure 365386DEST_PATH_IMAGE003
the feature extraction function is represented by a function,
Figure 112762DEST_PATH_IMAGE004
the center of the prototype is represented by the center of the prototype,
Figure 916639DEST_PATH_IMAGE005
，
Figure 941227DEST_PATH_IMAGE006
is a training sample
Figure 149354DEST_PATH_IMAGE007
The label of (a) is used,
Figure 10125DEST_PATH_IMAGE008
is the number of samples.
In the above model for identifying the interference in the open set based on prototype learning, the euclidean distance is used to measure the distance between the feature vector of the input interference and the center of the prototype:
Figure 368425DEST_PATH_IMAGE009
（2）
wherein
Figure 106574DEST_PATH_IMAGE010
For the dimension of the features in the feature space, a prototype center is randomly initialized using gaussian distribution.
In the above model for identifying open set interference based on prototype learning, the loss function includes:
Figure 141395DEST_PATH_IMAGE011
（3）
Figure 738730DEST_PATH_IMAGE012
（4）
wherein
Figure 759775DEST_PATH_IMAGE013
Which is indicative of a parameter of the network,
Figure 241179DEST_PATH_IMAGE014
is a constant number of times, and is,
Figure 197633DEST_PATH_IMAGE015
and K represents a known interference class of class K,
Figure 141319DEST_PATH_IMAGE016
k +1 represents an unknown interference class;
the first term of equation (4) corresponds to closed set training,
Figure 90689DEST_PATH_IMAGE017
representing an unknown interference category prototype in the removal prototype center set, and training an optimization model to enable output to be matched with the known interference category; the second term corresponds to an open-set training,
Figure 678796DEST_PATH_IMAGE018
representing removal of samples in a prototype-centric set
Figure 88043DEST_PATH_IMAGE019
Corresponding real label
Figure 659970DEST_PATH_IMAGE020
Prototype center of
Figure 757239DEST_PATH_IMAGE021
Training the optimization model so that the samples are identified as unknown interference classes by minimizing a loss function
Figure 714699DEST_PATH_IMAGE022
The model has the capability of identifying unknown interference while ensuring the identification accuracy of the known interference category.
In the above model for identifying open set interference based on prototype learning, the feature extraction network includes two branch networks for extracting time domain features and frequency domain features, and the feature extraction network obtains feature vectors of corresponding interference signals by splicing and fusing outputs of the two branch networks.
In the open set interference identification model based on prototype learning, the branch networks of the feature extraction network all adopt a one-dimensional depth residual error network with a ResNet structure;
the pooling layers of the two branch networks are connected to the splicing layer, and the splicing layer performs characteristic splicing output on the output of the two branch networks.
In the open set interference recognition model based on prototype learning, the splicing layer is sequentially connected to the linear layer, the activation function layer and the linear layer to perform nonlinear transformation to obtain a fused feature vector.
An open set interference identification method based on prototype learning comprises the following steps:
s1, extracting a feature vector of input interference by a feature extraction network;
s2, the classification model identifies interference categories according to the distance between the characteristic vectors output by the characteristic extraction network and prototype centers of corresponding interference categories;
and S3, identifying the interference class closest to the input interference as the class to which the corresponding input interference belongs.
In the above method for identifying open set interference based on prototype learning, the feature extraction network and the classification model train and optimize feature extraction capability and class identification capability through the following loss functions:
Figure 872011DEST_PATH_IMAGE023
（1）
Figure 665655DEST_PATH_IMAGE024
（4）
Figure 195643DEST_PATH_IMAGE025
（3）
Figure 148556DEST_PATH_IMAGE026
representing the distance between the feature vector and the center of the prototype,
Figure 617714DEST_PATH_IMAGE027
the feature extraction function is represented by a function,
Figure 147922DEST_PATH_IMAGE028
the center of the prototype is represented by the center of the prototype,
Figure 727939DEST_PATH_IMAGE029
，
Figure 535358DEST_PATH_IMAGE030
is a sample
Figure 926150DEST_PATH_IMAGE031
The label of (a) is used,
Figure 428807DEST_PATH_IMAGE032
the number of samples;
wherein
Figure 327361DEST_PATH_IMAGE033
Is indicative of a parameter of the network,
Figure 395811DEST_PATH_IMAGE034
is a constant number of times, and is,
Figure 331406DEST_PATH_IMAGE035
and K represents a known interference class of class K,
Figure 803583DEST_PATH_IMAGE036
k +1 represents an unknown interference class;
the first term of equation (4) corresponds to closed set training,
Figure 990982DEST_PATH_IMAGE037
representing removed prototype centric setsMerging unknown interference category prototypes, and training an optimization model to enable output to be matched with known interference categories; the second term corresponds to an open-set training,
Figure 507414DEST_PATH_IMAGE038
representing removal of samples in a prototype-centric set
Figure 738544DEST_PATH_IMAGE039
Corresponding real label
Figure 74847DEST_PATH_IMAGE040
Prototype center of
Figure 800358DEST_PATH_IMAGE041
Training the optimization model so that the samples are identified as unknown interference classes by minimizing a loss function
Figure 62974DEST_PATH_IMAGE042
The model has the capability of identifying unknown samples while ensuring the identification accuracy of the known interference types;
after the network training is finished, K +1 prototype centers are obtained, the first K prototype centers are known type prototype centers, the K +1 prototype centers are unknown interference prototype centers, and when interference and the first interference are input
Figure 74792DEST_PATH_IMAGE043
When the distance between the centers of the prototype is the smallest, the sample belongs to the second
Figure 304916DEST_PATH_IMAGE043
Class, when the distance from the center of the K +1 th prototype is minimal, then the input disturbance is an unknown class.
The invention has the advantages that:
1. the loss function designed by the scheme is based on a prototype learning strategy, and the closed set training and the open set training are combined by assuming an open type prototype center, so that the unknown type can be identified based on the prototype center, the model can be used for identifying unknown samples, and the interference identification precision of the model under the condition of the open set is effectively improved;
2. according to the scheme, whether the sample is unknown interference can be directly judged when the open set test is carried out, the open set threshold value is obtained without depending on the information of a training set, and the complexity of recognition is effectively reduced;
3. the scheme integrates the time domain and frequency domain characteristics of the ResNet structure and the interference signal, and the designed characteristic extraction network integrates the time domain and the frequency domain after respectively extracting the characteristics, thereby improving the characteristic extraction capability of the network on the input signal to a great extent;
4. according to the scheme, the known categories can be accurately judged and the unknown categories can be identified through a simpler model and algorithm.
Drawings
FIG. 1 is a diagram of a feature extraction network architecture for an open set interference recognition method based on prototype learning according to the present invention;
FIG. 2 is a flow chart of model training and recognition of the open set interference recognition method based on prototype learning according to the present invention.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.
As shown in fig. 2, an open set interference recognition model PL-Net is constructed to classify known interference and recognize unknown interference, wherein a feature extraction network is shown in fig. 1, the network combines a ResNet structure and characteristics of interference signals, and the specific structure of the designed feature extraction network is as follows:
Figure 817806DEST_PATH_IMAGE044
is an interfering signal of a sample length of 1024,
Figure 308830DEST_PATH_IMAGE045
is and is
Figure 632496DEST_PATH_IMAGE046
White gaussian noise of the same dimension. Signal of time domain
Figure 560701DEST_PATH_IMAGE047
Obtaining frequency domain signal after Fourier transformation
Figure 752648DEST_PATH_IMAGE048
Respectively input into two branch networks
Figure 973545DEST_PATH_IMAGE049
Respectively obtaining signal time domain characteristics and signal frequency domain characteristics, splicing the signal time domain characteristics and the signal frequency domain characteristics, and inputting the spliced signal time domain characteristics and the spliced signal frequency domain characteristics into a linear layer for characteristic fusion to obtain a final characteristic vector.
In which a branch network
Figure 717378DEST_PATH_IMAGE049
Is a one-dimensional depth residual error network adopting a ResNet structure. Each one of which is
Figure 515570DEST_PATH_IMAGE050
The system comprises 33 one-dimensional convolutional layers, 16 residual basic blocks and two one-dimensional pooling layers. The significance of the convolutional layer parameters is respectively as follows: convolution kernel size, convolution layer type, number of convolution kernels and output length variation. For example, the convolutional layer parameters {3x1 conv1d 128,/2} represent a one-dimensional convolution using a convolution kernel size of 3x1, the number of convolution kernels is 128, and the output sequence length is reduced to half of the input sequence. Each branch network has two pooling layers, a maximum pooling layer (max pool) and an adaptive pooling layer (adapt pool).
The function of the cat layer is to characterize the time domain
Figure 917733DEST_PATH_IMAGE051
Sum frequency domain characteristics
Figure 743869DEST_PATH_IMAGE052
Performing feature stitching to obtain output
Figure 674915DEST_PATH_IMAGE053
And performing nonlinear transformation on the spliced features by the subsequent linear layer and relu layer to obtain a fused feature vector.
The design of the feature extraction network integrates the ResNet structure and the time domain and frequency domain characteristics of interference signals, and the feature extraction capability of the network on input signals can be improved to a great extent, so that the subsequent classification capability is guaranteed.
Further, in order to enable the model to distinguish known interference signals and identify unknown interference signals, a unique loss function is designed for PL-Net, the loss function is based on a prototype learning design thought, a set of known interference prototype centers is defined, and then the class to which the interference signals belong is judged by adopting the distance between the interference features and the prototype centers, so that the known interference classification capability of the model can be realized, meanwhile, the prototype center of the unknown interference is trained by assuming that the unknown interference also has the prototype center and removing the prototype corresponding to the known interference, so that the model has the identification capability of the unknown interference at the same time, and the specific implementation mode is as follows:
using the known interference signal as the training data set
Figure 350616DEST_PATH_IMAGE054
Wherein
Figure 884366DEST_PATH_IMAGE055
Is interference
Figure 79855DEST_PATH_IMAGE056
The label of (a) to (b),
Figure 195185DEST_PATH_IMAGE057
in order to train the number of interfering signals in the set,
Figure 843335DEST_PATH_IMAGE058
defining prototype centers for interference signal classes according to a prototype learning strategy
Figure 180776DEST_PATH_IMAGE059
. Obtaining interference signals through a feature extraction network
Figure 214460DEST_PATH_IMAGE060
Is characterized byMeasurement of
Figure 487309DEST_PATH_IMAGE061
For the interference
Figure 481810DEST_PATH_IMAGE062
Belong to the first
Figure 983461DEST_PATH_IMAGE063
The probability of a class is calculated by equation (1):
Figure 747017DEST_PATH_IMAGE064
（1）
wherein
Figure 456347DEST_PATH_IMAGE065
Represents the distance between the feature vector and the prototype,
Figure 62778DEST_PATH_IMAGE066
a feature extraction function is represented, i.e. the input interference is mapped to a feature space. To better express the distance in the feature space, the present embodiment uses euclidean distance to measure the distance between the interference signal feature and the prototype center:
Figure 476442DEST_PATH_IMAGE067
（2）
wherein
Figure 235450DEST_PATH_IMAGE068
Is the dimension of the feature in the feature space. Initializing a prototype center randomly by adopting Gaussian distribution, defining a loss function as (3), and enabling an interference signal feature vector to be as close as possible to the corresponding prototype center so as to train the feature extraction capability and the classification capability of PL-Net at the same time:
Figure 869344DEST_PATH_IMAGE069
（3）
wherein
Figure 838437DEST_PATH_IMAGE070
Representing a network parameter.
Further, the present embodiment further assumes that the unknown interference also has a prototype center
Figure 931158DEST_PATH_IMAGE071
，
Figure 669307DEST_PATH_IMAGE072
Representing an unknown interference class label, the prototype center set is defined as
Figure 969707DEST_PATH_IMAGE073
. Given a well-trained feature extraction network and prototype center, an interference signal is given
Figure 832621DEST_PATH_IMAGE074
And its corresponding real label
Figure 214186DEST_PATH_IMAGE075
For the prototype center set which does not contain the unknown interference class prototype, the feature vector obtained by the sample through the feature extraction network
Figure 72420DEST_PATH_IMAGE076
Should be compatible with the prototype
Figure 28875DEST_PATH_IMAGE077
With a minimum distance. But for the removal of prototypes
Figure 97194DEST_PATH_IMAGE078
The set of remaining prototypes of, the sample
Figure 62876DEST_PATH_IMAGE079
To be considered as unknown samples, feature vectors
Figure 775617DEST_PATH_IMAGE080
Will be centered with unknown prototypes
Figure 385197DEST_PATH_IMAGE081
With the minimum distance, the present embodiment can define the loss function as:
Figure 550599DEST_PATH_IMAGE082
（4）
wherein
Figure 54392DEST_PATH_IMAGE083
Constant, the first term corresponds to closed set training,
Figure 11853DEST_PATH_IMAGE084
representing an unknown class prototype in the removed prototype center set, and training an optimization model to output to match the known class; the second term corresponds to an open-set training,
Figure 169165DEST_PATH_IMAGE085
representing removal of samples in a prototype-centric set
Figure 697229DEST_PATH_IMAGE086
Corresponding real label
Figure 755446DEST_PATH_IMAGE087
Prototype center of
Figure 442779DEST_PATH_IMAGE088
The optimization model is trained such that the samples are identified as unknown classes. By minimizing a loss function
Figure 646359DEST_PATH_IMAGE089
So that the model has a K +1 th prototype center with an unknown class in addition to K prototype centers of known class.
In this embodiment, the prototype center set of the closed set training items is
Figure 786353DEST_PATH_IMAGE090
Removing unknown prototypesThe interference-aware prototype center set comprises a prototype center set of open-set training items
Figure 881217DEST_PATH_IMAGE091
The remaining set of prototype centers for a sample corresponding to a prototype is removed. I.e., the same sample, is considered known and unknown at different archetype set perspectives.
Finally, as shown in fig. 2, the PL-Net model trained by the training set is used to identify the test set interference signals. After the model training is completed, the model training result is obtained
Figure 298423DEST_PATH_IMAGE092
When the model is used for identifying the interference signal, the sample to be identified is substituted into the formula (2) to obtain the sum
Figure 797537DEST_PATH_IMAGE092
Distance from the center of the prototype. The classification rule is expressed as:
Figure 65575DEST_PATH_IMAGE093
when and at the first
Figure 714862DEST_PATH_IMAGE094
When the distance between the centers of the prototype is minimum, the interference is identified as belonging to the second
Figure 111208DEST_PATH_IMAGE095
Class i, due to front
Figure 905858DEST_PATH_IMAGE058
Is a known prototype center, the first
Figure 20444DEST_PATH_IMAGE096
Is an unknown interference prototype center, so that when the sample is compared with the first sample
Figure 473423DEST_PATH_IMAGE097
When the distance from the center of the prototype is minimumThen the sample is identified as unknown interference.
The implementation sets known interference as frequency hopping interference, linear frequency sweep interference and secondary frequency sweep interference, and sets unknown interference as multi-tone interference. The training set is set to known interferers and each interferer contains 2000 samples at each dry to noise ratio from-10 dB to 20dB separated by 2dB for a total of 90000 samples. The test set contains known interference and unknown interference, 500 samples per interference to noise ratio, for a total of 30000 samples. Training the model by using the training set through the construction to obtain 4 prototype centers: the device comprises a frequency hopping interference prototype center, a linear frequency sweep interference prototype center, a secondary frequency sweep interference prototype center and an unknown signal prototype center. And then, testing and verifying the obtained open set interference identification model by using a test set:
after the frequency hopping interference signal is input into the feature extraction network, the distance between the frequency hopping interference signal and the centers of the 4 prototypes is obtained through a formula (2), but the distance between the frequency hopping interference signal and the centers of the frequency hopping interference prototypes is minimum, so that the frequency hopping interference signal is identified to belong to frequency hopping interference. Similarly, after the linear sweep interference signal is input into the feature extraction network, the distance between the linear sweep interference signal and the center of the 4 prototype is obtained through the formula (2), but the distance between the linear sweep interference signal and the center of the linear sweep interference prototype is the minimum, so that the linear sweep interference signal is identified to belong to linear sweep interference. After the secondary frequency sweeping interference signal is input into the feature extraction network, the distance between the secondary frequency sweeping interference signal and the centers of 4 prototypes is obtained through a formula (2), but the distance between the secondary frequency sweeping interference signal and the center of the secondary frequency sweeping interference prototype is the minimum, so that the secondary frequency sweeping interference signal is identified as the secondary frequency sweeping interference. After inputting the multi-tone interference signal into the feature extraction network, obtaining the distance between the multi-tone interference signal and the centers of 4 prototypes through a formula (2), and finding that the distance between the multi-tone interference signal and the centers of unknown prototypes is minimum, so that the multi-tone interference signal is identified as an unknown class. Therefore, the known interference frequency hopping interference, the linear frequency sweeping interference and the secondary frequency sweeping interference can be classified through the construction and the training by using the known interference, and the unknown multi-tone interference can be identified.
The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments, or alternatives may be employed, by those skilled in the art, without departing from the spirit or ambit of the invention as defined in the appended claims.

Claims (10)

1. An open set interference recognition model based on prototype learning is characterized by comprising a PL-Net with a feature extraction network and a classification model, wherein the feature extraction network is used for extracting features of input interference, the classification model is used for performing interference class recognition according to the distance between a feature vector output by the feature extraction network and a prototype center of a corresponding interference class, a loss function used for training the PL-Net comprises a closed set training item and an open set training item, a prototype center set of the open set training item comprises unknown interference class prototype centers, the prototype center set of the closed set training item has no unknown interference class prototype centers, and a plurality of known interference class prototype centers and an unknown interference class prototype center are obtained through training and are used for known interference class classification and unknown interference class recognition.
2. The prototype learning-based open-set interference recognition model of claim 1,
the prototype center set of the closed set training item is a first prototype set of which all samples correspond to prototypes;
the prototype center set of the open-set training item is a second prototype set of prototypes corresponding to the removed one or more known samples.
3. The model of claim 2, wherein the classification model calculates the distance between the feature vector of the input interference and the center of each prototype by:
Figure 784483DEST_PATH_IMAGE001
（1）
Figure 358553DEST_PATH_IMAGE002
representing the distance between the feature vector and the center of the prototype,
Figure 869300DEST_PATH_IMAGE003
the feature extraction function is represented by a function,
Figure 713890DEST_PATH_IMAGE004
the center of the prototype is represented by the center of the prototype,
Figure 42103DEST_PATH_IMAGE005
，
Figure 57464DEST_PATH_IMAGE006
is a sample
Figure 417907DEST_PATH_IMAGE007
The label of (a) is used,
Figure 959747DEST_PATH_IMAGE008
is the number of samples.
4. The model of claim 3, wherein the Euclidean distance is used to measure the distance between the feature vector of the input interference and the prototype center:
Figure 334227DEST_PATH_IMAGE009
（2）
wherein
Figure 115845DEST_PATH_IMAGE010
For the dimension of the features in the feature space, a prototype center is randomly initialized using gaussian distribution.
5. The model of claim 4, wherein the penalty function comprises:
Figure 93028DEST_PATH_IMAGE011
（3）
Figure 630320DEST_PATH_IMAGE012
（4）
wherein
Figure 300336DEST_PATH_IMAGE013
Which is indicative of a parameter of the network,
Figure 70715DEST_PATH_IMAGE014
is a constant number of times, and is,
Figure 726955DEST_PATH_IMAGE015
and K represents a known interference class of class K,
Figure 243387DEST_PATH_IMAGE016
k +1 represents an unknown interference class;
the first term of equation (4) corresponds to closed set training,
Figure 710403DEST_PATH_IMAGE017
representing an unknown interference category prototype in the removal prototype center set, and training an optimization model to enable output to be matched with the known interference category; the second term corresponds to an open-set training,
Figure 718810DEST_PATH_IMAGE018
representing removal of samples in a prototype-centric set
Figure 303375DEST_PATH_IMAGE019
Corresponding real label
Figure 798947DEST_PATH_IMAGE020
Prototype center of
Figure 217290DEST_PATH_IMAGE021
Training the optimization model so that the samples are identified as unknown interference classes by minimizing a loss function
Figure 40890DEST_PATH_IMAGE022
The model has the capability of identifying unknown interference while ensuring the identification accuracy of the known interference category.
6. The model according to claim 1, wherein the feature extraction network comprises two branch networks for extracting time domain features and frequency domain features, and the feature extraction network obtains the feature vector of the corresponding interference signal by splicing and fusing the outputs of the two branch networks.
7. The model of claim 6, wherein the branch networks of the feature extraction network each adopt a one-dimensional depth residual error network of a ResNet structure;
the pooling layers of the two branch networks are connected to the splicing layer, and the splicing layer performs characteristic splicing output on the output of the two branch networks.
8. The model of claim 7, wherein the splice layer is connected to the linear layer, the activation function layer, and the linear layer in sequence and is subjected to nonlinear transformation to obtain a fused feature vector.
9. An open set interference identification method based on prototype learning is characterized by comprising the following steps:
s1, extracting a feature vector of input interference by a feature extraction network;
s2, the classification model identifies the interference categories according to the distance between the characteristic vectors output by the characteristic extraction network and the prototype centers of the corresponding interference categories;
and S3, identifying the interference class closest to the input interference as the class to which the corresponding input interference belongs.
10. The method according to claim 9, wherein the feature extraction network and the classification model train and optimize feature extraction capability and class recognition capability through the following loss functions:
Figure 69893DEST_PATH_IMAGE023
（1）
Figure 295338DEST_PATH_IMAGE024
（4）
Figure 884582DEST_PATH_IMAGE025
（3）
Figure 585691DEST_PATH_IMAGE026
representing the distance between the feature vector and the center of the prototype,
Figure 512058DEST_PATH_IMAGE027
the feature extraction function is represented by a function,
Figure 998535DEST_PATH_IMAGE028
the center of the prototype is represented by the center of the prototype,
Figure 243833DEST_PATH_IMAGE029
，
Figure 42025DEST_PATH_IMAGE030
is a sample
Figure 381871DEST_PATH_IMAGE031
The label of (a) to (b),
Figure 237700DEST_PATH_IMAGE032
the number of samples;
wherein
Figure 27802DEST_PATH_IMAGE033
Which is indicative of a parameter of the network,
Figure 923076DEST_PATH_IMAGE034
is a constant number of times, and is,
Figure 611153DEST_PATH_IMAGE035
and K represents a known interference class of class K,
Figure 665697DEST_PATH_IMAGE036
k +1 represents an unknown interference class;
the first term of equation (4) corresponds to closed set training,
Figure 236487DEST_PATH_IMAGE037
representing an unknown interference category prototype in the removal prototype center set, and training an optimization model to enable output to be matched with the known interference category; the second term corresponds to an open-set training,
Figure DEST_PATH_IMAGE038
representing removal of samples in a prototype-centric set
Figure 196221DEST_PATH_IMAGE039
Corresponding real label
Figure 143449DEST_PATH_IMAGE040
Prototype center of
Figure 209756DEST_PATH_IMAGE041
Training the optimization model so that the samples are identified as unknown interference classes by minimizing a loss function
Figure DEST_PATH_IMAGE042
The model has the capability of identifying unknown samples while ensuring the identification accuracy of the known interference types;
after the network training is finished, K +1 prototype centers are obtained, the first K prototype centers are known type prototype centers, the K +1 prototype centers are unknown interference prototype centers, and when interference and the first interference are input
Figure 748185DEST_PATH_IMAGE043
When the distance from the center of the prototype is minimum, the input interference belongs to the second
Figure 398478DEST_PATH_IMAGE043
Class, when the distance from the center of the K +1 th prototype is minimal, then the input disturbance is an unknown class.
CN202210909603.8A 2022-07-29 2022-07-29 Model and method for identifying open set interference based on prototype learning Active CN114997248B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210909603.8A CN114997248B (en) 2022-07-29 2022-07-29 Model and method for identifying open set interference based on prototype learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210909603.8A CN114997248B (en) 2022-07-29 2022-07-29 Model and method for identifying open set interference based on prototype learning

Publications (2)

Publication Number Publication Date
CN114997248A true CN114997248A (en) 2022-09-02
CN114997248B CN114997248B (en) 2022-11-08

Family

ID=83022076

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210909603.8A Active CN114997248B (en) 2022-07-29 2022-07-29 Model and method for identifying open set interference based on prototype learning

Country Status (1)

Country Link
CN (1) CN114997248B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117689951A (en) * 2023-12-15 2024-03-12 西北农林科技大学 Open set identification method and system based on training-free open set simulator

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109274621A (en) * 2018-09-30 2019-01-25 中国人民解放军战略支援部队信息工程大学 Communication protocol signals recognition methods based on depth residual error network
US20190147336A1 (en) * 2017-11-10 2019-05-16 Fujitsu Limited Method and apparatus of open set recognition and a computer readable storage medium
CN112818893A (en) * 2021-02-10 2021-05-18 北京工业大学 Lightweight open-set landmark identification method facing mobile terminal
CN113055107A (en) * 2021-02-23 2021-06-29 电子科技大学 Interference strategy generation method for radio station with unknown communication mode
CN114004250A (en) * 2021-09-13 2022-02-01 西安电子科技大学 Method and system for identifying open set of modulation signals of deep neural network
CN114239672A (en) * 2021-09-30 2022-03-25 中国人民解放军陆军工程大学 Interference pattern open set identification model and method based on zero sample learning
CN114330522A (en) * 2021-12-22 2022-04-12 上海高德威智能交通系统有限公司 Training method, device and equipment of image recognition model and storage medium

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190147336A1 (en) * 2017-11-10 2019-05-16 Fujitsu Limited Method and apparatus of open set recognition and a computer readable storage medium
CN109274621A (en) * 2018-09-30 2019-01-25 中国人民解放军战略支援部队信息工程大学 Communication protocol signals recognition methods based on depth residual error network
CN112818893A (en) * 2021-02-10 2021-05-18 北京工业大学 Lightweight open-set landmark identification method facing mobile terminal
CN113055107A (en) * 2021-02-23 2021-06-29 电子科技大学 Interference strategy generation method for radio station with unknown communication mode
CN114004250A (en) * 2021-09-13 2022-02-01 西安电子科技大学 Method and system for identifying open set of modulation signals of deep neural network
CN114239672A (en) * 2021-09-30 2022-03-25 中国人民解放军陆军工程大学 Interference pattern open set identification model and method based on zero sample learning
CN114330522A (en) * 2021-12-22 2022-04-12 上海高德威智能交通系统有限公司 Training method, device and equipment of image recognition model and storage medium

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117689951A (en) * 2023-12-15 2024-03-12 西北农林科技大学 Open set identification method and system based on training-free open set simulator

Also Published As

Publication number Publication date
CN114997248B (en) 2022-11-08

Similar Documents

Publication Publication Date Title
CN110826630B (en) Radar interference signal feature level fusion identification method based on deep convolutional neural network
CN111783100A (en) Source code vulnerability detection method for code graph representation learning based on graph convolution network
CN111027378B (en) Pedestrian re-identification method, device, terminal and storage medium
CN111832417B (en) Signal modulation pattern recognition method based on CNN-LSTM model and transfer learning
CN109934269B (en) Open set identification method and device for electromagnetic signals
CN110909760B (en) Image open set identification method based on convolutional neural network
CN110610709A (en) Identity distinguishing method based on voiceprint recognition
CN111050315B (en) Wireless transmitter identification method based on multi-core two-way network
CN113435247A (en) Intelligent identification method, system and terminal for communication interference
CN114997248B (en) Model and method for identifying open set interference based on prototype learning
CN114387627A (en) Small sample wireless device radio frequency fingerprint identification method and device based on depth measurement learning
CN111222442A (en) Electromagnetic signal classification method and device
CN114980122A (en) Small sample radio frequency fingerprint intelligent identification system and method
CN113488060A (en) Voiceprint recognition method and system based on variation information bottleneck
CN115546608A (en) Unmanned aerial vehicle data link electromagnetic interference classification and threat assessment method
CN113052126A (en) Dual-threshold open-set signal modulation identification method based on deep learning
CN117580090B (en) Mobile terminal communication stability testing method and system
CN114520758A (en) Signal modulation identification method based on instantaneous characteristics
Feng et al. FCGCN: Feature Correlation Graph Convolution Network for Few-Shot Individual Identification
CN114793170B (en) DNS tunnel detection method, system, equipment and terminal based on open set identification
CN114626412A (en) Multi-class target identification method and system for unattended sensor system
CN112529035B (en) Intelligent identification method for identifying individual types of different radio stations
CN114998747B (en) Long flight path real-time identification method based on deep learning template matching
Wu et al. Communication Interference Recognition Based on Improved Deep Residual Shrinkage Network
CN117150265B (en) Robust radio frequency signal open set identification method under low signal-to-noise ratio condition

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