CN115720184A - Small sample signal modulation type identification method based on characteristic distribution - Google Patents
Small sample signal modulation type identification method based on characteristic distribution Download PDFInfo
- Publication number
- CN115720184A CN115720184A CN202211222724.1A CN202211222724A CN115720184A CN 115720184 A CN115720184 A CN 115720184A CN 202211222724 A CN202211222724 A CN 202211222724A CN 115720184 A CN115720184 A CN 115720184A
- Authority
- CN
- China
- Prior art keywords
- signal
- modulation type
- matrix
- training
- distribution
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 71
- 238000012549 training Methods 0.000 claims abstract description 87
- 238000000605 extraction Methods 0.000 claims abstract description 51
- 238000004364 calculation method Methods 0.000 claims abstract description 25
- 239000011159 matrix material Substances 0.000 claims description 128
- 238000012360 testing method Methods 0.000 claims description 43
- 238000011176 pooling Methods 0.000 claims description 40
- 238000010606 normalization Methods 0.000 claims description 24
- 230000004913 activation Effects 0.000 claims description 16
- 238000004590 computer program Methods 0.000 claims description 11
- 239000013598 vector Substances 0.000 claims description 7
- 238000004422 calculation algorithm Methods 0.000 claims description 4
- 239000000523 sample Substances 0.000 description 60
- 238000002474 experimental method Methods 0.000 description 15
- 238000010586 diagram Methods 0.000 description 8
- 230000000694 effects Effects 0.000 description 5
- 230000000052 comparative effect Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000001149 cognitive effect Effects 0.000 description 1
- 239000012468 concentrated sample Substances 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Image Analysis (AREA)
- Digital Transmission Methods That Use Modulated Carrier Waves (AREA)
Abstract
The invention discloses a small sample signal modulation type identification method based on characteristic distribution, which comprises the following steps: constructing a signal modulation type identification network, wherein the signal modulation type identification network comprises a signal feature extraction module, a distribution calculation module and a classifier module which are sequentially connected; acquiring a training support set and a training query set; training the signal modulation type recognition network by utilizing the training support set and the training query set to obtain a trained signal modulation type recognition network model; and identifying the modulation type of the original signal by using the trained signal modulation type identification network to obtain the modulation type of the original signal. The method has better robustness on the premise of achieving higher identification accuracy, and the constructed identification network can obtain good identification capability without a large number of labeled samples belonging to modulation types to be identified, so that the label work is greatly reduced.
Description
Technical Field
The invention belongs to the technical field of signal modulation type identification, and particularly relates to a small sample signal modulation type identification method based on characteristic distribution.
Background
Signal modulation type identification (automatic modulation classification) is a method of studying the identification of the modulation type of a radio signal. The automatic modulation classification refers to the automatic identification of the modulation type of the received signal, and plays an important role in the aspects of cognitive radio, spectrum monitoring, situation perception and the like.
Conventional automatic modulation classification methods include likelihood-based methods and feature-based methods. The likelihood-based method requires a large amount of assumptions, the feature-based method has high requirements on the selected features, and both methods need to consume a large amount of manpower and material resources, are low in efficiency, cannot adapt to variable scenes, and are poor in universality.
Recently, with the wide application of deep learning in various fields, it has also been successful in the automatic modulation classification problem. The traditional deep learning method needs a large number of training samples to enable the network to obtain good generalization performance, however, the process needs to consume a large amount of resources and cost to perform labeling work on the samples. Especially for samples with strong specialization, technical personnel are required to label the samples, and human and material resources are consumed. Therefore, how to correctly classify the signals to be identified through a small number of labeled signal samples has very important significance in the field of signal modulation classification.
Disclosure of Invention
In order to obtain the capability of classifying signals through a small number of labeled samples belonging to the type to be identified, the invention provides a small sample signal modulation type identification method based on characteristic distribution. The technical problem to be solved by the invention is realized by the following technical scheme:
the invention provides a small sample signal modulation type identification method based on feature distribution, which comprises the following steps:
s1: constructing a signal modulation type identification network, wherein the signal modulation type identification network comprises a signal feature extraction module, a distribution calculation module and a classifier module which are sequentially connected, and the signal feature extraction module is used for extracting the features of the original signal of the modulation type to be identified to obtain a feature matrix of the signal; the distribution calculation module is used for modeling the distribution of the characteristic matrix to obtain a distribution matrix of the characteristic matrix; the classifier module is used for identifying the modulation type of the signal according to the distribution matrix to obtain a prediction type label of the original signal;
s2: acquiring a training support set and a training query set;
s3: training the signal modulation type recognition network by utilizing the training support set and the training query set to obtain a trained signal modulation type recognition network model;
s4: and identifying the modulation type of the original signal by using the trained signal modulation type identification network to obtain the modulation type of the original signal.
In an embodiment of the present invention, the signal feature extraction module includes an I-path signal feature extraction unit and a Q-path signal feature extraction unit, which are respectively configured to perform feature extraction on an I-path signal and a Q-path signal of an original signal, wherein,
the I-path signal feature extraction unit comprises a first one-dimensional convolution layer, a first batch of normalization layers, a first ReLU activation layer, a first maximum pooling layer, a second one-dimensional convolution layer, a second batch of normalization layers, a second ReLU activation layer, a second maximum pooling layer, a third one-dimensional convolution layer, a third batch of normalization layers, a third ReLU activation layer and a third maximum pooling layer which are connected in sequence;
the Q-path signal feature extraction unit comprises a fourth one-dimensional convolution layer, a fourth normalization layer, a fourth ReLU activation layer, a fourth maximum pooling layer, a fifth one-dimensional convolution layer, a fifth normalization layer, a fifth ReLU activation layer, a fifth maximum pooling layer, a sixth one-dimensional convolution layer, a sixth normalization layer, a sixth ReLU activation layer and a sixth maximum pooling layer which are sequentially connected.
In an embodiment of the present invention, the convolution kernel sizes of the first one-dimensional convolution layer, the second one-dimensional convolution layer, the third one-dimensional convolution layer, the fourth one-dimensional convolution layer, the fifth one-dimensional convolution layer, and the sixth one-dimensional convolution layer are all 1 × 3, the number of convolution kernels is all 8, and the step length is all 1; the pooling core sizes of the first largest pooling layer, the second largest pooling layer, the third largest pooling layer, the fourth largest pooling layer, the fifth largest pooling layer, and the sixth largest pooling layer are all 1 × 3, and the step size is 2.
In one embodiment of the present invention, the S2 includes:
s2.1: selection of training set from predetermined open source signal data set
Comprises A type modulation signals, and each type modulation signal comprises B tr Sample-label pairs;
s2.2: randomly selecting N types of modulation signals from the training set, and randomly selecting K signal samples from each type of modulation signals to form a training support set, wherein N is less than or equal to A, and K is less than B tr ；
S2.3: the rest N (B) in the N-type modulation signals tr -K) signal samples constitute a training query set.
In one embodiment of the present invention, the S3 includes:
s3.1: setting the maximum iteration time T of training, and initializing the current iteration time T =0;
s3.2: acquiring signal sample data of a training support set and a training query set of a current iteration stage;
s3.3: sequentially inputting the signal sample data of the training support set and the training query set to the signal modulation type identification network, performing network training and updating network parameters;
s3.4: and judging whether the iteration cutoff requirement T = T is met, if not, updating the current iteration time T = T +1, repeatedly executing S3.2 and S3.3, and if so, outputting the trained signal modulation type identification network.
In one embodiment of the present invention, said S3.3 comprises:
s3.3.1: sequentially inputting the signal sample data in the training support set into the signal feature extraction module to map the signal sample data to a feature space, and obtaining a feature matrix corresponding to the signal sample data;
s3.3.2: inputting the obtained feature matrix into the distribution calculation module to obtain a distribution matrix corresponding to the current feature matrix, wherein the distribution matrix calculation formula is as follows:
wherein F represents a feature matrix output by the signal feature extraction module,
a matrix representing all 1 s, superscript T represents the transpose of the matrix,
representing the operation of a root number on each element in the matrix, G (F) representing the Euclidean distance matrix between two row vectors of the feature matrix F, G dtri (F) Representing the matrix after normalization of the matrix G (F), a representing the dimension of the feature matrix F;
s3.3.3: calculating a prototype c of each modulation type through a distribution matrix of a feature matrix corresponding to the N types of modulation types in the training support set i ：
Wherein i =1,2 i Representing signal samples belonging to modulation type i in the supporting set, x representing signal samples, | S i I represents the number of samples in the support set, flatten (·) represents a one-dimensional tiling operation on the matrix,
representing taking only the normalized matrix G dtri The upper triangle part data and the diagonal line data;
s3.3.4: calculating a distribution matrix of a feature matrix of the samples in the training and query set, obtaining classification probability distribution of the current sample according to the distance between the distribution matrix and the modulation type prototype, and taking the modulation type corresponding to the maximum probability as the modulation type of the current sample;
s3.3.5: and calculating cross entropy loss according to the predicted modulation type and the real label of the signal sample in the training query set, and updating the network parameters in the signal modulation type recognition network through a back propagation algorithm.
In one embodiment of the present invention, the S3.3.4 comprises:
obtaining a distribution matrix G of samples x in a training query set dtri To modulation type prototype c i The distance of (a) is:
||G dtri (f θ (x)),c i || 2 ，
wherein, c i A prototype representing each modulation type;
obtaining the probability that the sample x belongs to the modulation type i according to the distance as follows:
in an embodiment of the present invention, after step S3, the method further includes:
and acquiring a test support set and a test query set, and evaluating the performance of the trained signal modulation type recognition network by using the test support set and the test query set.
Another aspect of the present invention provides a storage medium, in which a computer program is stored, the computer program being configured to execute the steps of the method for identifying a modulation type of a small sample signal based on a feature distribution according to any one of the above embodiments.
Yet another aspect of the present invention provides an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor, when calling the computer program in the memory, implements the steps of the small sample signal modulation type identification method based on feature distribution as described in any one of the above embodiments.
Compared with the prior art, the invention has the beneficial effects that:
1. the invention provides a small sample signal modulation type identification method based on characteristic distribution, which designs a mode for calculating the distribution of a characteristic matrix and takes the distribution of the sample characteristic matrix as a classification basis. The distributed calculation mode provided by the invention has small calculation amount, and the small sample signal modulation type identification method can realize signal modulation type identification only by a few labeled samples belonging to the type to be identified, thereby greatly reducing the label cost; moreover, experiments on the RML2016.10.A data set show that compared with the method for directly using feature matrix classification, the method has better robustness on the premise of achieving higher identification accuracy.
2. The signal modulation type identification network designed by the invention maps signals to a characteristic space insensitive to signal-to-noise ratio to obtain a characteristic matrix thereof, thereby realizing correct modulation type identification, wherein the network maps I, Q two paths of signals respectively and splices the signals to obtain a final characteristic matrix; the signal feature extraction network of the invention uses a full convolution structure, and can adapt to the input of samples with different size parameters without manually changing the network structure parameters due to the characteristic that the network does not use a full connection layer.
3. The distribution calculation module of the invention identifies the modulation type of the signal by calculating the distribution of the signal characteristic matrix, and has better robustness while achieving high identification accuracy compared with a mode of directly classifying the signal through the characteristic matrix.
4. The small sample signal modulation type identification method based on characteristic distribution can obtain good identification capability without a large number of labeled samples belonging to modulation types to be identified, and label work is greatly reduced.
5. The network structure of the signal modulation type recognition network is simpler, the parameter quantity of the network is smaller, the signal modulation type recognition method does not need to adjust other hyper-parameters except the learning rate, the network is easier to converge and train, the network parameter quantity is smaller, the structure is simpler, the network is easier to train, and the result of modulation type recognition is more stable.
The present invention will be described in further detail with reference to the accompanying drawings and examples.
Drawings
Fig. 1 is a flowchart of a small sample signal modulation type identification method based on feature distribution according to an embodiment of the present invention;
fig. 2 is a schematic block diagram of a signal modulation type identification network according to an embodiment of the present invention;
fig. 3 is a schematic structural diagram of a signal feature extraction module according to an embodiment of the present invention;
fig. 4 is a schematic diagram of a training process of a signal modulation type recognition network according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of a distribution calculation process of a distribution calculation module according to an embodiment of the present invention;
fig. 6 is a histogram of experimental accuracy of a small sample signal modulation type identification method based on feature distribution according to an embodiment of the present invention.
Detailed Description
To further illustrate the technical means and effects of the present invention adopted to achieve the predetermined object, the following detailed description is provided with reference to the accompanying drawings and specific embodiments for a small sample signal modulation type identification method based on feature distribution according to the present invention.
The foregoing and other technical matters, features and effects of the present invention will be apparent from the following detailed description of the embodiments, which is to be read in connection with the accompanying drawings. The technical means and effects of the present invention adopted to achieve the predetermined purpose can be more deeply and specifically understood through the description of the specific embodiments, however, the attached drawings are provided for reference and description only and are not used for limiting the technical scheme of the present invention.
It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that an article or device that comprises a list of elements does not include only those elements but may include other elements not expressly listed. Without further limitation, an element defined by the phrase "comprising a … …" does not exclude the presence of another identical element in an article or device that comprises the element.
Referring to fig. 1, fig. 1 is a flowchart of a small sample signal modulation type identification method based on feature distribution according to an embodiment of the present invention. The small sample signal modulation type identification method comprises the following steps:
s1: and constructing a signal modulation type identification network.
Referring to fig. 2, fig. 2 is a schematic block diagram of a signal modulation type identification network according to an embodiment of the present invention, where the signal modulation type identification network includes a signal feature extraction module 101, a distribution calculation module 102, and a classifier module 103, which are connected in sequence, where the signal feature extraction module 101 is configured to extract features of an original signal of a modulation type to be identified, so as to obtain a feature matrix of the signal; the distribution calculation module 102 is configured to model the distribution of the feature matrix to obtain a distribution matrix of the feature matrix; the classifier module 103 is configured to identify a modulation type of the signal according to the distribution matrix, and obtain a prediction type tag of the original signal.
Further, the signal feature extraction module 101 includes a plurality of one-dimensional convolution layers, a plurality of batch normalization layers, a plurality of ReLU activation layers, and a plurality of max pooling layers. Specifically, please refer to fig. 3, where fig. 3 is a schematic structural diagram of a signal feature extraction module according to an embodiment of the present invention. The signal feature extraction module combines the characteristics that a signal sample is I, Q two paths of signals, constructs I, Q two paths of signals for feature extraction, and then performs matrix splicing to output a feature matrix. The signal feature extraction module 101 of this embodiment includes an I-path signal feature extraction unit and a Q-path signal feature extraction unit, which are respectively configured to perform feature extraction on an I-path signal and a Q-path signal of an original signal, where the I-path signal feature extraction part includes a first one-dimensional convolution layer, a first batch of normalization layers, a first ReLU active layer, a first maximum pooling layer, a second one-dimensional convolution layer, a second batch of normalization layers, a second ReLU active layer, a second maximum pooling layer, a third one-dimensional convolution layer, a third batch of normalization layers, a third ReLU active layer, and a third maximum pooling layer, which are sequentially connected; the Q-path signal feature extraction unit comprises a fourth one-dimensional convolution layer, a fourth normalization layer, a fourth ReLU activation layer, a fourth maximum pooling layer, a fifth one-dimensional convolution layer, a fifth normalization layer, a fifth ReLU activation layer, a fifth maximum pooling layer, a sixth one-dimensional convolution layer, a sixth normalization layer, a sixth ReLU activation layer and a sixth maximum pooling layer which are sequentially connected.
Preferably, the convolution kernel sizes of the first one-dimensional convolution layer, the second one-dimensional convolution layer, the third one-dimensional convolution layer, the fourth one-dimensional convolution layer, the fifth one-dimensional convolution layer, and the sixth one-dimensional convolution layer of this embodiment are all 1 × 3, the number of convolution kernels is all 8, and the step length is all 1; the pooling cores of the first largest pooling layer, the second largest pooling layer, the third largest pooling layer, the fourth largest pooling layer, the fifth largest pooling layer and the sixth largest pooling layer are all 1 × 3, and the step size is 2.
S2: a training support set and a training query set are obtained, wherein the training support set and the training query set are both obtained from a training set.
Step S2 of the present embodiment includes:
s2.1: training set selection from a predetermined open source signal data set
Comprises A type modulation signals, and each type modulation signal comprises B tr Sample-label pairs;
s2.2: randomly selecting N types of modulation signals from the training set, and randomly selecting K signal samples from each type of modulation signals to form a training support set, wherein N is less than or equal to A, and K is less than B tr ；
S2.3: the rest N (B) in the N-type modulation signals tr -K) signal samples constitute a training query set.
Specifically, the signal samples in the training set of this embodiment are selected from the source data set rml2016.10.A, each signal sample includes I, Q two paths of information, the sample size is 2 × 128, and the signal length is 128. For a small sample task of N-way K-shot, the training support set contains N modulation types, each modulation type contains K signal samples, where K is typically set to 1 or 5.
In this embodiment, the training set
Comprises A type modulation signal, and each type modulation signal comprises B tr A sample-label pair, wherein N is less than or equal to A and K is less than B tr ；
A set of sample-label pairs belonging to modulation type i in the training set is represented. Randomly selecting N types of modulation signals from a training set, randomly selecting K signal samples from each type of modulation signals to form a training support set, and randomly selecting the rest N (B) in the N types of modulation signals tr -K) signal samples constitute a training query set.
S3: and training the signal modulation type recognition network by utilizing the training support set and the training query set to obtain a trained signal modulation type recognition network model.
Specifically, referring to fig. 4, fig. 4 is a schematic diagram of a training process of a signal modulation type recognition network according to an embodiment of the present invention, where the training process of the signal modulation type recognition network according to the embodiment includes the following steps:
s3.1: setting the maximum iteration number T of training to be more than or equal to 10000, and initializing the current iteration number T =0;
s3.2: acquiring signal sample data of a training support set and a training query set of a current iteration stage;
s3.3: and sequentially inputting the signal sample data of the training support set and the training query set to the signal modulation type identification network, performing network training and updating network parameters.
In this embodiment, step 3.3 specifically includes the following sub-steps:
s3.3.1: and sequentially inputting the signal sample data in the training support set into the signal feature extraction module 101 to map the signal sample data to a feature space, so as to obtain a feature matrix corresponding to the signal sample data.
Specifically, the signal feature extraction module 101 is denoted as
Wherein the content of the first and second substances,
and
sub-modules for extracting I, Q paths of signal characteristics, namely an I path signal characteristic extraction unit, a Q path signal characteristic extraction unit and a theta path signal characteristic extraction unit I 、θ Q The parameters of the I-path signal feature extraction unit and the Q-path signal feature extraction unit are respectively. A signal sample x is input to the signal feature extraction module 101, and the output feature matrix can be represented as:
wherein, the first and the second end of the pipe are connected with each other,
denotes the dimension of the set, concat () denotes the splicing operation on the matrix.
S3.3.2: the obtained feature matrix is input into the distribution calculation module 102, so as to obtain a distribution matrix corresponding to the current feature matrix.
Specifically, please refer to fig. 5, wherein fig. 5 is a schematic diagram of a distribution calculation process of a distribution calculation module according to an embodiment of the present invention. In particular, the feature matrix f for one signal sample x θ (x) Briefly denoted as F, the distribution calculation formula is expressed as:
wherein, F = F θ (x) A feature matrix representing the output of the signal feature extraction module 101,
a matrix representing all 1 s, the superscript T representing the transpose of the matrix,
representing the operation of a root number on each element in the matrix, G (F) representing the Euclidean distance matrix between two row vectors of the feature matrix F, G dtri (F) Represents the matrix G (F) after normalization, i.e. the mean value of G (F) is normalized to 0, and the matrix G after normalization dtri (F) (ii) a a represents the dimension of the feature matrix F.
Furthermore, for the sake of simplifying the description, G (F) is denoted as G, G dtri (F) Is G dtri . The resulting distribution matrix
Since the matrix G is a symmetric matrix, G is satisfied T = G, therefore G dtri Also a symmetric matrix, i.e.
Since the data contained in the upper triangular part and the lower triangular part of the symmetric matrix are the same, only G is taken to reduce the data redundancy in the calculation process dtri The upper triangle part data and the diagonal line data are taken as classification bases and recorded as
The physical meaning of the calculation process is explained next. For two row vectors
The Euclidean distance of the two can be written as follows:
similarly, matrix G represents the euclidean distance matrix between every two row vectors of matrix F, which also explains the physical meaning of matrix G as a symmetric matrix; g1 denotes the sum matrix of each row of elements of the matrix G,
an average value matrix representing the sum of the elements of each row of the matrix G; 1*G represents the sum matrix of each column element of matrix G,
an average value matrix representing the sum of each column element of the matrix G; 1G 1 denotes the sum matrix of all elements of the matrix G,
representing the matrix of the average of all the elements of matrix G. Equation (2)) The physical significance of (1) is that the mean value of the normalized matrix G is 0, and the matrix G after normalization is obtained dtri 。
S3.3.3: calculating a prototype c of each modulation type through a distribution matrix of a feature matrix corresponding to the N types of modulation types in the training support set i ：
Wherein i =1,2 i Representing signal samples in the supporting set belonging to class i, x representing signal samples, | S i I represents the number of samples in the support set, flatten (·) represents a one-dimensional tiling operation on the matrix,
representing taking only the normalized matrix G dtri The upper triangular portion data and the diagonal data.
In addition, the prototype c i The method is characterized in that a class center is obtained according to a distribution matrix of a sample feature matrix in each class support set, and for an unknown sample, the unknown sample can be predicted to belong to the most similar class by comparing the similarity of the unknown sample and various class prototypes.
S3.3.4: and calculating a distribution matrix of the feature matrix of the samples in the training query set, obtaining the classification probability distribution of the samples according to the distance between the distribution matrix and the modulation type prototype, and taking the modulation type corresponding to the maximum probability as the modulation type of the current sample.
Specifically, for a signal sample x in a training query set, the distribution matrix G of its feature vectors is used dtri To calculate a classification probability vector p 1 ,p 2 ,...,p N ]. Training a distribution matrix G of samples x in a query set dtri To modulation type prototype c i The distance of (a) is: i G dtri (f θ (x)),c i || 2 Wherein, c i A prototype of each modulation type is represented. The probability p that the sample x thus obtained belongs to modulation type i i Comprises the following steps:
wherein, y j A real label representing the specimen.
S3.3.5: and calculating cross entropy loss according to the predicted modulation type and the real label of the signal sample in the training query set, and updating the network parameters in the signal feature extraction module through a back propagation algorithm.
Wherein, the calculation formula of the cross entropy loss L is as follows:
wherein x is j Represents the jth sample, y, in the training iteration j The true label representing the sample, g represents the number of samples in the training iteration.
S3.4: and judging whether the iteration cutoff requirement T = T is met, if not, updating the current iteration time T = T +1, repeatedly executing S3.2 and S3.3, and if so, outputting the trained signal modulation type identification network.
S4: and identifying the modulation type of the modulation signal to be identified by utilizing the trained signal modulation type identification network to obtain the modulation type of the signal.
Specifically, the modulation type of the modulation signal can be obtained by inputting the modulation signal to be recognized into the signal modulation type recognition network after training.
Further, this embodiment further includes, after step S3:
and evaluating the performance of the trained signal modulation type recognition network by utilizing the test support set and the test query set.
First, a test support set and a test query set are obtained, wherein the test support set and the test query set are obtained from the test set, which is also from the source data set rml2016.10.a and has different data than the training set.
In particular, test sets
Includes N types of modulation signals, each type of modulation signal includes B te A pair of sample labels is provided for each sample,
a set of sample-label pairs belonging to modulation type i in the test set is represented. Randomly selecting K signal samples from each class of N modulation signals of a test set to form a test support set, wherein the rest N (B) in the N modulation signals te -K) signal samples constituting a test query set, wherein K < B te 。
Then, the trained signal modulation type recognition network is tested by using the test support set and the test query set, and the specific test process comprises the following substeps:
step (1): setting the repetition times Q of the test process, wherein the identification accuracy rate is related to the representativeness of the signal sample, so that generally setting Q > 1, and repeatedly testing the network performance for multiple times to obtain convincing evaluation parameters; initializing a current test round q =0;
step (2): obtaining a test support set and a test query set corresponding to the current test round;
and (3): obtaining a modulation type prototype according to the test support set, and referring to the step 3.3.1-the step 3.3.3; obtaining the predicted modulation type of the sample according to the test query set, and referring to step 3.3.4; calculating the prediction accuracy of the current test query set according to the real modulation type;
and (4): judging whether a test stopping condition Q = Q is met; if yes, outputting the test accuracy of each test process; if not, continuously repeating the testing step 4.2 to the step 4.4 to obtain the identification accuracy under different support set conditions;
and (5): and calculating the average value, the maximum value, the minimum value and the standard deviation of the identification accuracy of the Q times of test processes as performance evaluation parameters of the signal modulation identification network.
In addition, the effect of the small sample signal modulation type identification method based on feature distribution provided by the embodiment of the present invention can be further illustrated and verified by the following specific examples.
(1) Conditions of the experiment
CPU Intel (R) Core (TM) i7-10875H, eight cores, with a dominant frequency of 2.30GHz; the memory size is 16GB; GPU is NVIDIA GeForce RTX 2060, and video memory size is 6GB.
(2) Emulation content
In order to verify the identification effect of the small sample signal modulation type identification method, the method provided by the embodiment of the invention is tested.
The modulation signal in RML2016.10a was selected for the experiments. Wherein, signals of 8 modulation types including BPSK, CPFSK, GFSK, PAM4, QAM16, QAM64, QPSK and WBFM are selected as training sets; taking signals of 3 modulation types of 8PSK, AM-DSB and AM-SSB as modulation types to be identified; wherein, each modulation type contains 10 signal-to-noise ratios of 0 tr ＝8，B tr =10000, test set B te =1000; of each signal sample
Wherein 2 represents I, Q two-way signals, and 128 represents the sample length; the maximum iteration number T =10000 in the training stage; the number of repeated tests of the test phase Q =1000; features extracted by a modulated signal feature extraction network
I.e. a =8,b =32; distribution matrix of characteristic matrix
The network updates the network weights using Adam optimization algorithm, learning rate lr =0.005.
For the small sample task of N-way K-shot, each modulation type in the supporting set only contains K signal samples, K is generally set to 1 or 5, and since the modulation type to be identified contains 3, the small sample task in this experiment is defined as 3-way 1-shot or 3-way 5-shot.
To demonstrate the effectiveness of the method of the present invention, the results of comparative experiments using the existing signature matrix identification and identification using the method of the present invention are presented in tables 1 and 2.
Specifically, in experiment 1, a full connection layer is connected to each of two paths of feature extraction networks (i.e., an I-path signal feature extraction unit and a Q-path signal feature extraction unit), a feature matrix is one-dimensionally tiled, then mapped to a specific length, and finally spliced to obtain a category prototype, that is:
wherein, the linear I (. Cndot.) and Linear Q (. Cndot.) represents the fully connected layer after connecting to the I-path signal feature extraction unit and the Q-path signal feature extraction unit respectively.
Experiment 2 is a method based on an embodiment of the present invention.
The statistics of the classification accuracy of the test results of experiment 1 and experiment 2, which were tested 1000 times respectively, are given in table 1, and the maximum value, the minimum value, the average value and the standard deviation of the two values are compared. It can be seen that, no matter for the experiments of 3-way 1-shot or 3-way 5-shot, the method of the embodiment of the invention can obviously improve the minimum value and the average value, greatly reduce the standard deviation, and the maximum values of the two experiments are basically the same. The method provided by the embodiment of the invention can greatly improve the network performance of small sample classification, enhance the classification stability and prove the effectiveness of the method.
Table 2 shows the average confusion matrix obtained by testing 1000 times with the method of the embodiment of the present invention, and it can be seen that good identification accuracy can be achieved for experiments of 3-way 1-shot and 3-way 5-shot. Fig. 6 shows a distribution histogram of the experimental accuracy of 1000 tests according to the method of the embodiment of the present invention, and it can be seen that the identification capability of the network depends on the representativeness of the supporting concentrated samples to some extent, but the classification accuracy of the network is maintained at a high level as a whole, which also proves the effectiveness and stability of the method of the embodiment of the present invention.
TABLE 1 Classification accuracy statistics for comparative experiments
Table 2 experiment 2 (invention) average confusion matrix (units/%)
The embodiment of the invention provides a small sample signal modulation type identification method based on characteristic distribution, and designs a mode for calculating characteristic matrix distribution. Wherein the distribution of the sample feature matrix is used as a classification basis. The distributed calculation mode provided by the invention has small calculation amount, and can realize signal modulation type identification only by a few labeled samples belonging to types to be identified, thereby greatly reducing the label cost; moreover, experiments on the RML2016.10.A data set show that compared with the method for directly using feature matrix classification, the method provided by the embodiment of the invention has better robustness on the premise of achieving higher identification accuracy.
The invention can obtain good identification capability without a large number of labeled samples belonging to modulation types to be identified, thereby greatly reducing label work. The network structure of the invention is simpler, and the parameter quantity of the network is smaller; besides, the learning rate of the method provided by the invention needs to be adjusted, no other hyper-parameters need to be adjusted, and the network is easier to converge and train. The invention has the advantages of smaller network parameter, simpler structure, easier network training and more stable modulation type identification result.
Yet another embodiment of the present invention provides a storage medium, in which a computer program is stored, the computer program being used for executing the steps of the small sample signal modulation type identification method based on feature distribution in the above embodiment. Yet another aspect of the present invention provides an electronic device, which includes a memory and a processor, where the memory stores a computer program, and the processor implements the steps of the method for identifying modulation type of small sample signal based on feature distribution as described in the above embodiment when calling the computer program in the memory. Specifically, the integrated module implemented in the form of a software functional module may be stored in a computer readable storage medium. The software functional module is stored in a storage medium and includes several instructions to enable an electronic device (which may be a personal computer, a server, or a network device) or a processor (processor) to execute some steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: various media capable of storing program codes, such as a usb disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk, or an optical disk.
The foregoing is a more detailed description of the invention in connection with specific preferred embodiments and it is not intended that the invention be limited to these specific details. For those skilled in the art to which the invention pertains, several simple deductions or substitutions can be made without departing from the spirit of the invention, and all shall be considered as belonging to the protection scope of the invention.
Claims (10)
1. A small sample signal modulation type identification method based on feature distribution is characterized by comprising the following steps:
s1: constructing a signal modulation type identification network, wherein the signal modulation type identification network comprises a signal feature extraction module, a distribution calculation module and a classifier module which are sequentially connected, and the signal feature extraction module is used for extracting the features of the original signal of the modulation type to be identified to obtain a feature matrix of the signal; the distribution calculation module is used for modeling the distribution of the characteristic matrix to obtain a distribution matrix of the characteristic matrix; the classifier module is used for identifying the modulation type of the signal according to the distribution matrix to obtain a prediction type label of the original signal;
s2: acquiring a training support set and a training query set;
s3: training the signal modulation type recognition network by utilizing the training support set and the training query set to obtain a trained signal modulation type recognition network model;
s4: and identifying the modulation type of the original signal by using the trained signal modulation type identification network to obtain the modulation type of the original signal.
2. The method for identifying the modulation type of the small sample signal based on the feature distribution as claimed in claim 1, wherein the signal feature extraction module comprises an I-path signal feature extraction unit and a Q-path signal feature extraction unit for performing feature extraction on the I-path signal and the Q-path signal of the original signal respectively, wherein,
the I-path signal feature extraction unit comprises a first one-dimensional convolution layer, a first batch of normalization layers, a first ReLU activation layer, a first maximum pooling layer, a second one-dimensional convolution layer, a second batch of normalization layers, a second ReLU activation layer, a second maximum pooling layer, a third one-dimensional convolution layer, a third batch of normalization layers, a third ReLU activation layer and a third maximum pooling layer which are connected in sequence;
the Q-path signal feature extraction unit comprises a fourth one-dimensional convolution layer, a fourth batch of normalization layers, a fourth ReLU activation layer, a fourth maximum pooling layer, a fifth one-dimensional convolution layer, a fifth batch of normalization layers, a fifth ReLU activation layer, a fifth maximum pooling layer, a sixth one-dimensional convolution layer, a sixth batch of normalization layers, a sixth ReLU activation layer and a sixth maximum pooling layer which are connected in sequence.
3. The method according to claim 2, wherein the convolution kernel sizes of the first one-dimensional convolution layer, the second one-dimensional convolution layer, the third one-dimensional convolution layer, the fourth one-dimensional convolution layer, the fifth one-dimensional convolution layer, and the sixth one-dimensional convolution layer are all 1 × 3, the number of convolution kernels is 8, and the step length is 1; the pooling core sizes of the first largest pooling layer, the second largest pooling layer, the third largest pooling layer, the fourth largest pooling layer, the fifth largest pooling layer, and the sixth largest pooling layer are all 1 × 3, and the step size is 2.
4. The method for identifying the modulation type of the small sample signal based on the characteristic distribution according to claim 2, wherein the S2 comprises:
s2.1: selection of training set from predetermined open source signal data set
Comprises A type modulation signals, and each type modulation signal comprises B tr A sample-tag pair;
s2.2: randomly selecting N types of modulation signals from the training set, and randomly selecting K signal samples from each type of modulation signals to form a training support set, wherein N is less than or equal to A, and K is less than B tr ；
S2.3: the rest N (B) in the N-type modulation signals tr -K) signal samples constitute a training query set.
5. The method for identifying the modulation type of the small sample signal based on the characteristic distribution according to claim 2, wherein the step S3 comprises:
s3.1: setting the maximum iteration time T of training, and initializing the current iteration time T =0;
s3.2: acquiring signal sample data of a training support set and a training query set of a current iteration stage;
s3.3: sequentially inputting the signal sample data of the training support set and the training query set to the signal modulation type identification network, performing network training and updating network parameters;
s3.4: judging whether an iteration cutoff requirement T = T is met, if not, updating the current iteration time T = T +1, and repeatedly executing S3.2 and S3.3; if yes, outputting the trained signal modulation type identification network.
6. The feature distribution based small sample signal modulation type identification method according to claim 5, wherein said S3.3 comprises:
s3.3.1: sequentially inputting the signal sample data in the training support set into the signal feature extraction module to map the signal sample data to a feature space, and obtaining a feature matrix corresponding to the signal sample data;
s3.3.2: inputting the obtained feature matrix into the distribution calculation module to obtain a distribution matrix corresponding to the current feature matrix, wherein the distribution matrix calculation formula is as follows:
wherein F represents a feature matrix output by the signal feature extraction module,
a matrix representing all 1 s, the superscript T representing the transpose of the matrix,
representing the operation of a root number on each element in the matrix, G (F) representing the Euclidean distance matrix between two row vectors of the feature matrix F, G dtri (F) Representation matrix G (F) returnA matrix after normalization, a represents the dimension of the feature matrix F;
s3.3.3: calculating a prototype c of each modulation type through a distribution matrix of a feature matrix corresponding to the N types of modulation types in the training support set i ：
Wherein i =1,2 i Representing signal samples belonging to modulation type i in the supporting set, x representing signal samples, | S i L represents the number of samples in the support set, flatten (·) represents a one-dimensional tiling operation on the matrix,
representing taking only the normalized matrix G dtri The upper triangle part data and the diagonal line data;
s3.3.4: calculating a distribution matrix of a feature matrix of the samples in the training query set, obtaining classification probability distribution of the current sample according to the distance between the distribution matrix and the modulation type prototype, and taking the modulation type corresponding to the maximum probability as the modulation type of the current sample;
s3.3.5: and calculating cross entropy loss according to the predicted modulation type and the real label of the signal sample in the training query set, and updating the network parameters in the signal modulation type recognition network through a back propagation algorithm.
7. The method according to claim 6, wherein the S3.3.4 comprises:
obtaining a distribution matrix G of samples x in a training query set dtri To modulation type prototype c i The distance of (a) is:
||G dtri (f θ (x)),c i || 2
wherein, c i A prototype representing each modulation type;
obtaining the probability that the sample x belongs to the modulation type i according to the distance as follows:
8. the method for identifying the modulation type of the small sample signal based on the characteristic distribution according to claim 1, further comprising after step S3:
and acquiring a test support set and a test query set, and evaluating the performance of the trained signal modulation type recognition network by using the test support set and the test query set.
9. A storage medium, characterized in that the storage medium stores therein a computer program for executing the steps of the feature distribution based small sample signal modulation type identification method according to any one of claims 1 to 8.
10. An electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method for identifying modulation type of small sample signal based on feature distribution according to any one of claims 1 to 8 when calling the computer program in the memory.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211222724.1A CN115720184B (en) | 2022-10-08 | 2022-10-08 | Small sample signal modulation type identification method based on characteristic distribution |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211222724.1A CN115720184B (en) | 2022-10-08 | 2022-10-08 | Small sample signal modulation type identification method based on characteristic distribution |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115720184A true CN115720184A (en) | 2023-02-28 |
| CN115720184B CN115720184B (en) | 2024-04-19 |
Family
ID=85253856
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211222724.1A Active CN115720184B (en) | 2022-10-08 | 2022-10-08 | Small sample signal modulation type identification method based on characteristic distribution |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115720184B (en) |
Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2008178072A (en) * | 2006-12-22 | 2008-07-31 | Toshiba Corp | Modulation type determination apparatus and modulation type determination method |
| CN107979554A (en) * | 2017-11-17 | 2018-05-01 | 西安电子科技大学 | Radio signal Modulation Identification method based on multiple dimensioned convolutional neural networks |
| CN108805039A (en) * | 2018-04-17 | 2018-11-13 | 哈尔滨工程大学 | The Modulation Identification method of combination entropy and pre-training CNN extraction time-frequency image features |
| CN110728230A (en) * | 2019-10-10 | 2020-01-24 | 江南大学 | Signal modulation mode identification method based on convolution limited Boltzmann machine |
| CN113141325A (en) * | 2021-04-14 | 2021-07-20 | 北京邮电大学 | Training method, identification method and device for optical OFDM signal subcarrier modulation format identification model |
| CN113420817A (en) * | 2021-06-25 | 2021-09-21 | 西安电子科技大学 | Semi-supervised modulation type identification method, device and medium based on network structure characteristic induction |
| US20220014401A1 (en) * | 2019-11-06 | 2022-01-13 | Soochow University | Fast modulation recognition method for multilayer perceptron based on multimodally-distributed test data fusion |
| CN114118339A (en) * | 2021-11-12 | 2022-03-01 | 吉林大学 | Radio modulation signal identification and classification method based on cuckoo algorithm improved ResNet |
| CN114567528A (en) * | 2022-01-26 | 2022-05-31 | 中国人民解放军战略支援部队信息工程大学 | Communication signal modulation mode open set identification method and system based on deep learning |
| CN114580484A (en) * | 2022-04-28 | 2022-06-03 | 西安电子科技大学 | Small sample communication signal automatic modulation identification method based on incremental learning |
| CN114881092A (en) * | 2022-06-16 | 2022-08-09 | 杭州电子科技大学 | Signal modulation identification method based on feature fusion |
| CN114925720A (en) * | 2022-04-20 | 2022-08-19 | 西安电子科技大学 | Small sample modulation signal identification method based on space-time mixed feature extraction network |
-
2022
- 2022-10-08 CN CN202211222724.1A patent/CN115720184B/en active Active
Patent Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2008178072A (en) * | 2006-12-22 | 2008-07-31 | Toshiba Corp | Modulation type determination apparatus and modulation type determination method |
| CN107979554A (en) * | 2017-11-17 | 2018-05-01 | 西安电子科技大学 | Radio signal Modulation Identification method based on multiple dimensioned convolutional neural networks |
| CN108805039A (en) * | 2018-04-17 | 2018-11-13 | 哈尔滨工程大学 | The Modulation Identification method of combination entropy and pre-training CNN extraction time-frequency image features |
| CN110728230A (en) * | 2019-10-10 | 2020-01-24 | 江南大学 | Signal modulation mode identification method based on convolution limited Boltzmann machine |
| US20220014401A1 (en) * | 2019-11-06 | 2022-01-13 | Soochow University | Fast modulation recognition method for multilayer perceptron based on multimodally-distributed test data fusion |
| CN113141325A (en) * | 2021-04-14 | 2021-07-20 | 北京邮电大学 | Training method, identification method and device for optical OFDM signal subcarrier modulation format identification model |
| CN113420817A (en) * | 2021-06-25 | 2021-09-21 | 西安电子科技大学 | Semi-supervised modulation type identification method, device and medium based on network structure characteristic induction |
| CN114118339A (en) * | 2021-11-12 | 2022-03-01 | 吉林大学 | Radio modulation signal identification and classification method based on cuckoo algorithm improved ResNet |
| CN114567528A (en) * | 2022-01-26 | 2022-05-31 | 中国人民解放军战略支援部队信息工程大学 | Communication signal modulation mode open set identification method and system based on deep learning |
| CN114925720A (en) * | 2022-04-20 | 2022-08-19 | 西安电子科技大学 | Small sample modulation signal identification method based on space-time mixed feature extraction network |
| CN114580484A (en) * | 2022-04-28 | 2022-06-03 | 西安电子科技大学 | Small sample communication signal automatic modulation identification method based on incremental learning |
| CN114881092A (en) * | 2022-06-16 | 2022-08-09 | 杭州电子科技大学 | Signal modulation identification method based on feature fusion |
Non-Patent Citations (4)
| Title |
|---|
| LIHUI WANG等: ""Wireless signal modulation identification method based on RF I/Q data "", SCIENTIFC REPORTS, 1 November 2021 (2021-11-01) * |
| 吴婉婷: ""基于深度学习的通信信号调制识别研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》, 15 April 2022 (2022-04-15) * |
| 张茜茜: ""基于深度学习的自动调制识别方法综述"", 《无线电通信技术》, vol. 48, no. 4, 9 June 2022 (2022-06-09) * |
| 洪先成;张国毅;李冬梅;王长宇;: ""多特征融合的雷达信号脉内调制识别"", 信息与电子工程, vol. 9, no. 5, 31 October 2011 (2011-10-31) * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115720184B (en) | 2024-04-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109063112B (en) | Rapid image retrieval method, model and model construction method based on multitask learning deep semantic hash | |
| CN105930402A (en) | Convolutional neural network based video retrieval method and system | |
| CN110097003A (en) | Check class attendance method, equipment, storage medium and device neural network based | |
| CN114580484B (en) | Small sample communication signal automatic modulation identification method based on incremental learning | |
| CN110059157A (en) | A kind of picture and text cross-module state search method, system, device and storage medium | |
| CN110837846A (en) | Image recognition model construction method, image recognition method and device | |
| CN111931505A (en) | Cross-language entity alignment method based on subgraph embedding | |
| CN110457677B (en) | Entity relationship identification method and device, storage medium and computer equipment | |
| CN109766942A (en) | A kind of small-sample learning image-recognizing method based on attention neural network | |
| CN110866564B (en) | Season classification method, system, electronic device and medium for multiple semi-supervised images | |
| CN111723874A (en) | Sound scene classification method based on width and depth neural network | |
| CN114584440A (en) | Small sample AMC method based on Transformer coding network | |
| CN112241920A (en) | Investment and financing organization evaluation method, system and equipment based on graph neural network | |
| CN114528876A (en) | Small sample signal automatic modulation identification method based on improved prototype network | |
| CN115272692A (en) | Small sample image classification method and system based on feature pyramid and feature fusion | |
| CN105701227B (en) | A kind of across media method for measuring similarity and search method based on local association figure | |
| CN112836007B (en) | Relational element learning method based on contextualized attention network | |
| CN115720184B (en) | Small sample signal modulation type identification method based on characteristic distribution | |
| CN113743079A (en) | Text similarity calculation method and device based on co-occurrence entity interaction graph | |
| CN112163603A (en) | Zero sample image identification method and device, electronic equipment and storage medium | |
| CN111508528A (en) | No-reference audio quality evaluation method and device based on natural audio statistical characteristics | |
| CN115019175B (en) | Pest identification method based on migration element learning | |
| CN110119465B (en) | Mobile phone application user preference retrieval method integrating LFM potential factors and SVD | |
| CN112445939A (en) | Social network group discovery system, method and storage medium | |
| CN116664240B (en) | Shopping behavior prediction method, system, storage medium and terminal based on graphic neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |