CN114567528B - Communication signal modulation mode open set recognition method and system based on deep learning - Google Patents
Communication signal modulation mode open set recognition method and system based on deep learning Download PDFInfo
- Publication number
- CN114567528B CN114567528B CN202210095612.8A CN202210095612A CN114567528B CN 114567528 B CN114567528 B CN 114567528B CN 202210095612 A CN202210095612 A CN 202210095612A CN 114567528 B CN114567528 B CN 114567528B
- Authority
- CN
- China
- Prior art keywords
- class
- signal
- modulation
- similarity
- category
- 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.)
- Active
Links
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L27/00—Modulated-carrier systems
- H04L27/0012—Modulated-carrier systems arrangements for identifying the type of modulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Digital Transmission Methods That Use Modulated Carrier Waves (AREA)
Abstract
The invention belongs to the field of radio communication, and particularly relates to a communication signal modulation mode open set identification method and system based on deep learning, which are characterized in that a similarity matrix between a centroid vector and a signal sample feature vector is utilized to train and optimize a deep neural network, and a threshold value for rejecting unknown classes under corresponding modulation classes is set according to similarity distribution between the centroid vector and the signal sample feature vector, wherein the centroid vector is a feature vector of a baseband ideal modulation signal of a known class, and each known modulation class corresponds to one centroid vector; and obtaining a similarity set between the target communication signals in the signal set to be identified and the centroid vector by using the deep neural network, taking the category corresponding to the maximum value in the similarity set as a prediction category, and comparing the maximum value with a set threshold value to judge whether the prediction category is an unknown category. The invention can reject the unknown signal while maintaining the identification performance of the known signal, thereby being convenient for the application in the automatic modulation identification.
Description
Technical Field
The invention belongs to the field of radio communication, and particularly relates to a communication signal modulation mode open set identification method and system based on deep learning.
Background
The non-cooperative communication signal modulation and identification, namely, the method is widely applied to the fields of military, civil use and the like as a third party to analyze and process the intercepted signal and identify the modulation mode of the intercepted signal on the premise that the parameter information of the communication signal is unknown. The method can be used in the fields of communication reconnaissance, electromagnetic countermeasure and the like in military; civil aspects may be used for electromagnetic spectrum management and radio monitoring. At present, signal modulation modes are endless. The MPSK modulation is widely applied to the field of wireless communication, especially in short-wave communication, with the advantages of high spectrum utilization rate, good anti-noise performance and the like. Compared with MPSK, the MQAM modulation is used as a phase and amplitude joint modulation, can meet the requirement of large data information transmission, and is widely applied to the fields of digital cable television transmission, satellite communication and the like. Therefore, the two signal modulation types are of great significance as research objects.
At present, the research of signal modulation recognition is very wide, from a maximum likelihood hypothesis test method to expert priori feature extraction, and then the research is combined with deep learning, researchers continuously try to improve the performance of modulation recognition by using advanced technology, but the research is difficult to apply to real scenes all the time, mainly because the premise of the research is that the modulation mode based on a signal set to be recognized is known. Along with the increasing complexity of electromagnetic environment, the signal modulation mode is continuously updated, and it is difficult to include all signal types in the signal modulation mode, especially based on the modulation identification of a neural network, add a full connection layer at the end of the network, and then perform SoftMax normalization. The neural network can achieve good effect only under the condition that the training set and the testing set are consistent in category. Once an unknown signal appears in the test set, which is not found in the training set, the neural network can only classify the unknown signal into a known class with the highest probability score, so that the open set theory needs to be applied to modulation recognition to improve the limitation of the neural network. Generalized open-set recognition concepts include transfer learning, field adaptation, small sample learning, zero-order learning, and the like, which more or less utilize the association information existing between the training set and the test set, while narrow open-set recognition faces more serious challenges, requiring that not only the known categories in the training set be accurately recognized, but also the unknown categories that do not appear in the training set be directly recognized. Most of the current algorithms convert the problem of identifying unknown classes into directly or indirectly setting a threshold to reject unknown samples.
Disclosure of Invention
Therefore, the invention provides a communication signal modulation mode open set recognition method and system based on deep learning, which can effectively process the problem of communication signal open set recognition, can reject unknown signals while keeping the recognition performance on known signals, and is convenient for application in automatic modulation recognition technology.
According to the design scheme provided by the invention, the open set recognition method of the communication signal modulation mode based on deep learning is provided, and comprises the following contents:
training and optimizing a deep neural network for signal identification by utilizing a similarity matrix between a centroid vector and a signal sample feature vector, and setting a threshold value for rejecting unknown classes under a corresponding modulation class according to similarity distribution between the centroid vector and the signal sample feature vector, wherein the centroid vector is a feature vector of a baseband ideal modulation signal of a known class, and each known modulation class corresponds to one centroid vector;
and acquiring a similarity set between the target communication signals in the signal set to be identified and the centroid vector by using the trained deep neural network, taking the category corresponding to the maximum value in the similarity set as a prediction category, and comparing the maximum value with a set threshold value to judge whether the prediction category is an unknown category.
As the communication signal modulation mode open set identification method based on deep learning, the deep neural network further comprises: the device comprises a feature mapping unit and a nonlinear dimension reduction unit, wherein the feature mapping unit utilizes a residual error contraction module added in a convolution layer to carry out convolution operation on input signal data so as to extract feature information, compresses the feature information by gradually decreasing the number of convolution kernel channels layer by layer, and keeps the input feature dimension of each layer consistent by zero filling operation; feature dimension reduction is performed through a full connection layer in the nonlinear dimension reduction unit, and batch standardization and regularization are used for preventing overfitting in the model training process.
In the deep neural network training optimization, the baseband I/Q data of the baseband ideal modulation signal is used as the input of the deep neural network, the GE2E loss function is used for training, and the average value of the feature vectors output by the deep neural network is used as the centroid vector.
In the deep learning-based communication signal modulation mode open set recognition method, in the network training by utilizing a GE2E loss function, baseband I/Q data of N multiplied by M baseband ideal modulation signals are input into a deep neural network in batches to obtain normalized network output feature vectors, wherein N represents the number of signal modulation categories, and M represents the number of signal samples in each modulation category; and aggregating the same-class feature vectors by using M signal sample average vectors in each modulation class and the similarity between each sample feature vector and the corresponding signal sample average vector, and training and optimizing the network parameters far from the different-class feature vectors.
As the communication signal modulation mode open set identification method based on deep learning, the invention further aims at the feature vector and the centroid vector of the signal sample, the similarity set between the feature vector and the centroid vector is obtained, the similarity set is divided into an intra-class set and an extra-class set by utilizing a maximum inter-class variance method, and a threshold value for rejecting unknown classes under the corresponding modulation class is set according to the variance in the inter-class set.
As the communication signal modulation mode open set identification method based on deep learning, the invention further uses the maximum inter-class variance method to set the corresponding modulation class lower threshold value, sets the intra-class and the outer-class threshold value variables, and obtains the proportion of the total sample number of the intra-class and the outer-class sets and the average value of the similarity by calculating the probability of the similar samples corresponding to the similarity value; the overall similarity mean value is obtained according to the proportion of the samples in the intra-class and the extra-class sets to the total number of the samples and the similarity mean value, the inter-class variance is combined to obtain data corresponding to a threshold variable which enables the inter-class variance to be maximum, and the data corresponding to the threshold variable is used as a threshold value under the corresponding modulation class.
When the method is used for identifying the open set of the communication signal modulation mode based on the deep learning, further judging whether the predicted class is an unknown class or not, comparing the maximum value in the similarity set obtained from the predicted class with a set threshold value, judging that the predicted class is correct and outputting the predicted class if the maximum value is larger than the set threshold value, otherwise, judging that the predicted class is wrong and outputting the unknown class.
Further, the invention also provides a communication signal modulation mode open set identification method based on deep learning, which comprises the following steps: a model construction module and a target recognition module, wherein,
the model construction module is used for training and optimizing the deep neural network for signal identification by utilizing a similarity matrix between a centroid vector and a signal sample feature vector, and setting a threshold value for rejecting unknown classes under corresponding modulation classes according to similarity distribution between the centroid vector and the signal sample feature vector, wherein the centroid vector is a feature vector of a baseband ideal modulation signal of a known class, and each known modulation class corresponds to one centroid vector;
the target recognition module is used for acquiring a similarity set between the target communication signals in the signal set to be recognized and the centroid vector by utilizing the trained deep neural network, taking the category corresponding to the maximum value in the similarity set as a prediction category, and comparing the maximum value with a set threshold value to judge whether the prediction category is an unknown category.
The invention has the beneficial effects that:
compared with selecting the sample feature vector which is predicted correctly in the training set, the baseband ideal modulation signal of the known class is used as the centroid vector in the open space, the influence caused by the difference of training samples can be avoided, and then the neural network is used for training, so that the feature mapping of the similar signal samples on the feature space is close to the centroid vector, and meanwhile, the centroid vector is far away from other centroid vectors, so that the boundary between different classes is more prominent; the known class signals are identified by calculating the similarity between the signal samples and the centroid vector, the self-adaptive threshold is set to reject the unknown signals, the self-adaptive threshold is set for each known class according to the similarity distribution, and the I/Q data is used as network input, so that the method is applicable to common quadrature modulation communication signals. The validity of the unknown signal is refused by a corresponding algorithm in the embodiment of the scheme through experimental analysis, and the unknown signal can be refused while the identification performance of the known signal is maintained; and when the signal-to-noise ratio is 5dB, both indexes reach more than 90%, so that the method has a good application prospect.
Description of the drawings:
FIG. 1 is a schematic diagram of a communication signal modulation scheme open set identification flow based on deep learning in an embodiment;
FIG. 2 is a schematic diagram of a deep neural network structure in an embodiment;
FIG. 3 is a schematic view of similarity distribution in the embodiment;
FIG. 4 is a diagram of network training in an embodiment;
FIG. 5 is a eigenvector schematic of a baseband ideal modulated signal in an embodiment;
FIG. 6 is a schematic representation of a closed set identification confusion matrix in an embodiment;
FIG. 7 is a schematic diagram of a statistical histogram of similarity in an embodiment;
FIG. 8 is a confusion matrix representation of ICS, openMax, and open set identification of algorithms corresponding to embodiments of the present disclosure;
fig. 9 is an illustration of an open set identification performance curve as a function of signal-to-noise ratio in an embodiment.
The specific embodiment is as follows:
the present invention will be described in further detail with reference to the drawings and the technical scheme, in order to make the objects, technical schemes and advantages of the present invention more apparent.
At present, most of communication signal modulation recognition researches are based on the premise that a signal modulation mode is known, so that in order to be closer to a real scene, part of scholars begin to explore and apply an open set theory to the field of communication signal modulation recognition. If the reconstruction discrimination network model is based on the generation of the countermeasure network, the method has better recognition accuracy for the known signal category under the condition of low signal-to-noise ratio, and can reject the unknown signal category, but each signal modulation mode needs to train a reconstruction discrimination network, and when the modulation category is more, the engineering quantity is larger. The embodiment of the invention provides a communication signal modulation mode open set identification method based on deep learning, which is shown in fig. 1 and comprises the following steps:
s101, training and optimizing a deep neural network for signal identification by utilizing a similarity matrix between a centroid vector and a signal sample feature vector, and setting a threshold value for rejecting unknown classes under corresponding modulation classes according to similarity distribution between the centroid vector and the signal sample feature vector, wherein the centroid vector is a feature vector of a baseband ideal modulation signal of a known class, and each known modulation class corresponds to one centroid vector;
s102, acquiring a similarity set between a target communication signal in a signal set to be identified and a centroid vector by using a trained deep neural network, taking a category corresponding to a maximum value in the similarity set as a prediction category, and comparing the maximum value with a set threshold value to judge whether the prediction category is an unknown category.
Compared with selecting a sample feature vector which is predicted correctly in a training set, the baseband ideal modulation signal with a known class is used as a centroid vector in an open space, so that the influence caused by the difference of training samples can be avoided, and then the neural network is used for training, so that the feature mapping of the similar signal samples on the feature space is close to the centroid vector, and meanwhile, the centroid vector is far away from other centroid vectors, so that the boundary between different classes is more prominent; the known class signals are identified by calculating the similarity between the signal samples and the centroid vector, the self-adaptive threshold is set to reject the unknown signals, the self-adaptive threshold is set for each known class according to the similarity distribution, the I/Q data is used as network input, the method is applicable to common quadrature modulation communication signals, and the automatic identification effect of signal modulation is improved.
As the communication signal modulation mode open set identification method based on deep learning in the embodiment of the invention, further, the deep neural network includes: the device comprises a feature mapping unit and a nonlinear dimension reduction unit, wherein the feature mapping unit utilizes a residual error contraction module added in a convolution layer to carry out convolution operation on input signal data so as to extract feature information, compresses the feature information by gradually decreasing the number of convolution kernel channels layer by layer, and keeps the input feature dimension of each layer consistent by zero filling operation; feature dimension reduction is performed through a full connection layer in the nonlinear dimension reduction unit, and batch standardization and regularization are used for preventing overfitting in the model training process. Further, in the training optimization of the deep neural network, baseband I/Q data of a baseband ideal modulation signal is input into the deep neural network, training is performed by utilizing a GE2E loss function, and an average value of feature vectors output by the deep neural network is taken as a centroid vector.
As shown in fig. 2, the network is mainly composed of two parts: a feature mapping part and a nonlinear dimension reduction part. Because the residual error contraction network can effectively inhibit noise in the signal, a residual error contraction module can be added in a convolution layer part to carry out convolution operation on input I/Q data so as to extract characteristic information. In order to extract more critical characteristic information, the following characteristic dimension reduction is also facilitated, and the number of channels of the convolution kernel is reduced layer by layer to perform characteristic compression. In addition, convolution kernels with the sizes of 1×3 and 2×3 can be adopted, so that model convergence can be accelerated while the signal I/Q data structure is reserved to the greatest extent. Since the I/Q data is less informative than the picture pixel data, and also in order to fully preserve the transformed information of the symbol sequence in the signal, no pooling layer is employed in the network. And in order to ensure that the sizes of the input features of each layer are consistent, a zero filling operation is adopted. The dimension reduction part mainly adopts a full connection layer to reduce the dimension of the features, and batch standardization and regularization are used for preventing overfitting. The output of the entire network can be defined as
To facilitate calculation and comparison of feature similarity, the feature vectors extracted from the network are processedLine L2 normalization:
wherein v is ij A feature vector representing the j-th sample of the received i-th modulated signal.
In the method for identifying the open set of the communication signal modulation mode based on the deep learning in the embodiment of the invention, in the network training by utilizing the GE2E loss function, baseband I/Q data of N multiplied by M baseband ideal modulation signals are used as a batch to be input into a deep neural network to obtain normalized network output feature vectors, wherein N represents the number of signal modulation categories, and M represents the number of signal samples in each modulation category; and aggregating the same-class feature vectors by using M signal sample average vectors in each modulation class and the similarity between each sample feature vector and the corresponding signal sample average vector, and training and optimizing the network parameters far from the different-class feature vectors. Further, for the feature vector and the centroid vector of the signal sample, the similarity set is divided into an intra-class set and an extra-class set by acquiring the similarity set between the feature vector and the centroid vector and utilizing a maximum inter-class variance method, and a threshold value for rejecting unknown classes under the corresponding modulation class is set according to the variance in the inter-class set. Further, setting a threshold variable in the corresponding modulation class in a lower threshold of the corresponding modulation class by using a maximum inter-class variance method, and obtaining the proportion of the total number of samples in the intra-class and the extra-class sets and the average value of the similarity by calculating the probability of the similar samples corresponding to the similarity value; the overall similarity mean value is obtained according to the proportion of the samples in the intra-class and the extra-class sets to the total number of the samples and the similarity mean value, the inter-class variance is combined to obtain data corresponding to a threshold variable which enables the inter-class variance to be maximum, and the data corresponding to the threshold variable is used as a threshold value under the corresponding modulation class. Further, when judging whether the predicted class is an unknown class, comparing the maximum value in the similarity set of the acquired predicted class with a set threshold value, if the maximum value is larger than the set threshold value, judging that the predicted class is correct and outputting the predicted class, otherwise, judging that the predicted class is wrong and outputting the unknown class.
The selection of the centroid vector is the key for determining the performance of the algorithm, and the OpenMax algorithm takes the average value of the correctly predicted feature vectors in the training set as the centroid vector, so that the dependence on the training set is relatively large. The NCMC adopts a group of centers obtained by K-means clustering to represent a category, so that the characteristics of the category can be better represented, but the NCMC has large calculation amount and is not suitable for real-time processing. In this embodiment, it may be assumed that the known signal modulation parameters in the signal set to be identified are substantially known, such as modulation type, filter parameters, etc. Therefore, the eigenvector of the baseband ideal modulation signal s (t) of the known class can be selected as the best centroid vector, and can not be influenced by the training samples and the channel interference, and the expression is as follows:
wherein a is k Representing the generated power normalized symbol sequence, T is the symbol interval, and g (T) represents that the channel filter includes a shaping filter, a matched filter, etc.
Let M be the modulation order, the MPSK signal can be expressed as:
the MQAM signal may be expressed as:
a k ＝A mi +jA mq ,m＝1,2,...,M (4)
I/Q data of the baseband ideal modulation signal is sent into a depth network, the GE2E loss function is utilized for training, the average value of the eigenvectors output by the network is taken as a centroid vector, and each known modulation class corresponds to one centroid vector.
Aiming at the similarity distribution between the feature vector and the centroid vector, a certain error caused by function fitting is mainly considered, a function fitting method is not adopted, if the differences among different training samples are large,the similarity distribution may be greatly different, and the scalability of the fitting function is certain, so in the embodiment of the present application, a method based on a threshold may be used to reject unknown classes, which is simple and effective. But the related research of introducing threshold setting is less, P I The SVM estimates the threshold from the posterior probability, but it is obviously not reasonable to use the same threshold for all known classes. The POS-SVM then empirically determines a unique reject threshold for each known class based on an open set risk minimization method.
The similarity between signal feature vectors of the same modulation class and different modulation classes should satisfy two different set distributions, so there is an optimal threshold to reasonably divide the intra-class and the extra-class, as shown in fig. 3. The similarity value between the signal characteristic vector of the same modulation class and the centroid vector of the class is higher and basically larger than the optimal threshold; and the similarity value of the signal characteristic vector of other classes and the centroid vector of the class is lower than the optimal threshold value.
The threshold is determined by referring to an image binarization oxford algorithm, which is also called a maximum inter-class variance method. For the similarity set between the feature vectors and the centroid vectors of all training samples, the similarity set is divided into an intra-class set and an extra-class set by using an Ojin algorithm. The algorithm principle can be described as follows:
let the total number of samples of the similarity set be N, and the number of samples with the similarity value of i be N i The probability is:
assuming that the threshold value of the intra-class and the extra-class of the region is tau, the ratio of the intra-class set (i.e. the similarity is larger than tau) to the total sample number is w 0 ，
The average value of the similarity is mu 0 ：
The proportion of the out-of-class set is w 1 The average similarity is mu 1 ：
w 1 ＝1-w 0 (8)
The mean value of the overall similarity set is μ:
μ＝w 0 μ 0 +w 1 μ 1 (10)
the inter-class variance is g:
g＝w 0 (μ 0 -μ) 2 +w 1 (μ 1 -μ) 2 (11)
substituting formula (10) into formula (11) to obtain:
g＝w 0 w 1 (μ 0 -μ 1 ) 2 (12)
threshold τ for maximizing inter-class variance g * Namely, the optimal threshold value:
training of deep neural networks can be divided into two phases: the first stage is to train the ideal modulation signal of baseband to extract the centroid vector, the second stage is to update the network parameters by using the similarity between the extracted signal feature vector and the centroid vector, so that the signal feature vectors of the same modulation mode are more similar, and the signal feature vector distances of different modulation modes are increased. In the first stage, GE2E can be used as a loss function, the basic principle of which is: taking I/Q data of N×M baseband ideal modulation signals as a batch input depth network (N represents the number of signal modulation classes, M represents the number of signal samples in each modulation class), obtaining normalized network output characteristic v, and thenCalculating an average vector c ' of M signal samples in each modulation class and a feature vector v ' of each sample ' ij Similarity to average vector s':
s' ij,k ＝w×similarity(v' ij ,c' k )+b (15)
where k=1, 2..n, 0.ltoreq.i < N, 0.ltoreq.j < M, w and b are constants.
In training, it is desirable that feature vectors of the same class are mutually aggregated, feature vectors of different classes are mutually distant, and Softmax loss is as follows:
in the second stage training, the centroid vector of the known signal modulation class is:
c k ＝average(v kj ) (18)
inputting N×M received signal I/Q data into depth network, calculating eigenvector v ij Similarity to the centroid vector is:
s ij,k ＝w×similarity(v ij ,c k )+b (19)
as can be seen from fig. 4, in the second stage training, the I/Q waveform data is input to the depth network to extract feature information, the feature maps of the same modulation class are similar, and the feature maps of different classes are different. And forming a similarity matrix by the similarity between the feature vector and the centroid vector, and training to ensure that the similarity value of the color region is larger and the similarity value of the gray region is smaller.
In the test stage, calculating the similarity set between the feature vector and the centroid vector extracted by GE2E,taking similarity s ij The maximum value is taken as a prediction category and is matched with an adaptive threshold tau determined by the Ojin algorithm * And comparing, if the class is larger than tau, judging the class, otherwise judging the class as an unknown class.
Further, based on the above method, the embodiment of the present invention further provides a method for identifying an open set of a communication signal modulation mode based on deep learning, including: a model construction module and a target recognition module, wherein,
the model construction module is used for training and optimizing the deep neural network for signal identification by utilizing a similarity matrix between a centroid vector and a signal sample feature vector, and setting a threshold value for rejecting unknown classes under corresponding modulation classes according to similarity distribution between the centroid vector and the signal sample feature vector, wherein the centroid vector is a feature vector of a baseband ideal modulation signal of a known class, and each known modulation class corresponds to one centroid vector;
the target recognition module is used for acquiring a similarity set between the target communication signals in the signal set to be recognized and the centroid vector by utilizing the trained deep neural network, taking the category corresponding to the maximum value in the similarity set as a prediction category, and comparing the maximum value with a set threshold value to judge whether the prediction category is an unknown category.
To verify the validity of this protocol, the following is further explained in connection with experimental data:
common digital amplitude-phase modulation signals are selected for research, and the modulation types are as follows: 16QAM,32QAM,64QAM,16APSK, BPSK, QPSK,8PSK,16PSK. In the signal transmission process, noise interference is considered, gaussian white noise is added to the modulated baseband ideal signal, the power ratio of the signal to the noise (namely the signal to noise ratio (SNR) of each sampling point) ranges from-5 dB to 10dB, 1000 samples are arranged every 1dB in each modulation category, and the length of each sample is 512 points. To better compare the performance of the corresponding algorithm with other algorithms shown in the embodiment of the present disclosure, the network structure shown in fig. 2 may be used (or the output of the last full-connection layer is modified to the number of modulation classes).
The deep neural network generally utilizes a cross entropy loss function to realize closed set classification and identification, and a similarity method which can be adopted in the embodiment is used for closed set identification: and training a depth network by utilizing the GE2E loss function, calculating the similarity between the extracted feature vector and the centroid vector, and taking the maximum value of the similarity as a prediction category. In the experiment, the calculation of the similarity adopts three modes of Euclidean distance, cosine similarity and weighted combination of Euclidean distance and cosine similarity. Because the euclidean distance has a value in the range of 0, ++ infinity a) of the above-mentioned components, to keep pace with cosine similarity, the Euclidean distance is converted between (0, 1):
where E (x, y) represents the euclidean distance between the two vectors x, y.
The baseband ideal modulated signal is trained with GE2E and the extracted feature vectors are shown in fig. 5. The eigenvectors of the same modulation mode are gathered with each other, and boundaries between different modulation modes are obvious, so that the eigenvector mean value of the baseband ideal modulation signal is used as the centroid vector to be the optimal choice.
Table 1 shows the cross entropy and the overall closed set recognition accuracy (Combined represents a weighted combination of Euclidean distance and cosine similarity 1:1) of the algorithm on the test set in the present embodiment. The closed set identification accuracy of the cross entropy reaches more than 80%, which shows that the designed network has a good modulation identification effect. By contrast, the similarity method is very similar to the cross entropy in the performance of closed set recognition, and the recognition accuracy of the three kinds of similarity is not very different, so that the characteristic vector extracted by using GE2E can well embody the characteristics of a signal modulation mode. The identification accuracy of the cosine similarity is slightly higher than that of the cosine similarity in other two modes, which indicates that the difference of the extracted feature vectors in the direction angle is slightly more prominent than the absolute distance.
TABLE 1 Cross entropy and closed set identification accuracy of the algorithms herein on test sets
Comparing the cross entropy with the closed set identification confusion matrix of cosine similarity, as shown in fig. 6, wherein (a) is the confusion matrix of cross entropy, the effect of identifying 16QAM,32QAM, QPSK,8PSK and 16PSK modulation signals can be found to be relatively good; (b) And the identification effect of the 64QAM and 16APSK modulation signals is slightly excellent for the identification result of the cosine similarity. Because BPSK modulation is relatively simple, both recognition accuracy rates reach 100%. The three groups of signals, namely 16QAM and 64QAM,32QAM and 16APSK,8PSK and 16PSK, are easily mixed, and the subsequent open set identification is very difficult.
Classification recognition and closed set recognition in an open environment are very different, and in order to better describe the open set recognition problem, an open space may be defined by using an openness:
wherein N is TR Representing the number of signal modulation classes, N, contained in the training set TE Representing the number of signal modulation classes contained in the test set, and N TR ≤N TE . When N is TR ＝N TE When the value of openness is 0, the closed set identification is performed; with N TE The value of openness approaches 1, which indicates a higher degree of openness.
Because open set identification needs to consider unknown categories, the evaluation index of closed set identification is no longer applicable. In order to better evaluate the performance of the algorithm in an open environment, two open set identification evaluation indexes Normalized Accuracy and macro-F1 are introduced.
Normalized Accuracy (NA): mainly consists of an accuracy rate (AKS) for identifying known categories and an accuracy rate (AUS) for identifying unknown categories, and the formula is defined as follows:
NA＝λ r AKS+(1-λ r )AUS (25)
wherein lambda is r To balance the coefficient of accuracy between the known class and the unknown class, 0 < lambda r And is less than or equal to 1. In the subsequent comparative experiments, lambda r Taking 0.5.N represents the number of known modulation classes. TP, TN, FP, FN each represent a true example, a false example, and a false example.
F1-Measure: the method is widely applied to evaluating the performance of the classification model, and is a weighted average of Precision and Recall. In the multi-classification task, two specific indexes of micro-F1 and marco-F1 exist. micro-F1 calculates the total Precision and Recall of all categories, then calculates the F1 value, marco-F1 calculates the F1 value of each category, and then takes the average value of them. Since there may be multiple unknown classes in an experiment, the number of samples for all unknown classes will be greater than the number of samples for a single known class, to avoid the effects of sample imbalance, marco-F1 is used, whose formula is defined as:
the emphasis of the two indexes is different, NA is more focused on the discovery of the unknown class, and marco-F1 is used for treating the unknown class as the known class, so that the comprehensive performance of the model can be better evaluated by combining the two indexes.
In experiments, cosine similarity may be used to calculate the similarity between the feature vector and the centroid vector. Fig. 7 is a statistical histogram of the similarity of signal samples to a 16QAM centroid vector. As can be seen from the figure, the similarity of the signal eigenvectors not belonging to 16QAM modulation is relatively small, typically lower than 0.2; the similarity of the signal feature vectors belonging to the 16QAM modulation is relatively large. We use the oxford algorithm to obtain the best threshold value between the inside and outside of the classification.
According to different openness, eight modulation signal categories are randomly divided into a known category and an unknown category; randomly selecting 80% of samples of known modulation classes as a training set; randomly selecting 20% of unknown modulation class samples and the rest 20% of known modulation class samples to form a test set; evaluating the model by using two indexes of NA and marco-F1; on the premise of using the depth network structure designed in the embodiment of the present case, the algorithm in the embodiment of the present case is compared with ICS and OpenMax in performance. Table 2 shows the performance comparisons of different open set recognition algorithms. Under the condition of different openness, compared with the open set identification effect of ICS and OpenMax, the scheme algorithm in the embodiment of the invention is greatly improved, and the scheme algorithm in the embodiment of the invention is excellent in the aspect of processing the open set identification problem, and the ICS and the OpenMax algorithms have super parameters which need to be set manually, and the threshold value of the scheme algorithm in the embodiment of the invention is calculated by the Ojin algorithm without depending on manual experience.
TABLE 2 Algorithm Performance comparison for different openness
The tail size and alpha in the OpenMax algorithm are all optimal values obtained through iterative search, and ρ in the ICS algorithm is 15%. openness=7.42% means that 6 out of 8 modulation classes are selected as known classes, another 2 are unknown classes; openness=12.30% means that 5 out of 8 modulation classes are chosen as known classes, and the other 3 are unknown classes. The values of NA and marco-F1 were averaged over multiple experiments.
FIG. 8 is a confusion matrix of open-ended ICS, openMax, and open-set identification of scheme algorithms in the present embodiment. The 8 kinds of modulation signals are divided into known types {16QAM,64QAM,32QAM, BPSK, QPSK,8PSK } and unknown types {16APSK,16PSK }, because the unknown modulation signals 16APSK,16PSK are respectively similar to the known modulation signals 32QAM,8PSK, noise interference is added, so that the identification effect of ICS and OpenMax on the unknown signals is bad, the scheme algorithm in the embodiment sets an adaptive threshold for each known modulation type, the identification accuracy of the known signals is 78.3%, and compared with the identification accuracy of Cosine Similarity (fig. 6 (b)), the identification accuracy of the unknown signals is 78%.
Fig. 9 is a performance index change curve of openness=7.42%. The marco-F1 index of ICS and OpenMax is higher than the NA index, which indicates that the two algorithms have poor recognition effect on unknown signals. The NA index of ICS can reach 50% under the condition of low signal-to-noise ratio, but along with the improvement of the signal-to-noise ratio, the NA index basically remains at about 50%, even the NA index is somewhat reduced, mainly because the lower the signal-to-noise ratio is, the larger the interference of a modulation signal is, the more difficult the neural network identification is, the prediction score of the low signal-to-noise ratio signal is lower, the atypical sample selected by ICS is basically signal data with lower signal-to-noise ratio, the data is used as a boundary, the open set identification effect of the low signal-to-noise ratio is better, and the identification effect of the unknown signal with high signal-to-noise ratio is poorer. The OpenMax selects a signal with smaller distance from the centroid to fit the weibull function, the higher the signal-to-noise ratio is, the closer the signal-to-noise ratio is to the centroid vector, otherwise, the signal characteristic vector with low signal-to-noise ratio is located at the tail end of the fitting function, so that the known signal predicted value with low signal-to-noise ratio is smaller and is easy to be recognized as an unknown signal by mistake. Along with the increase of the signal-to-noise ratio, two open set identification indexes NA and marco-F1 of the scheme algorithm in the embodiment of the scheme are continuously improved. At 5dB, both exceed 90%, which means that the scheme algorithm in the embodiment of the scheme can effectively process the problem of open set identification of communication signals. The GE2E is used for training the neural network, so that boundaries among categories are more obvious, then an optimal threshold value between the category and the outside of the category is calculated by using an Ojin algorithm, a good open set identification effect is achieved under a high signal-to-noise ratio, and meanwhile, good performance can be maintained under a low signal-to-noise ratio.
The relative steps, numerical expressions and numerical values of the components and steps set forth in these embodiments do not limit the scope of the present invention unless it is specifically stated otherwise.
Finally, it should be noted that: the above examples are only specific embodiments of the present invention, and are not intended to limit the scope of the present invention, but it should be understood by those skilled in the art that the present invention is not limited thereto, and that the present invention is described in detail with reference to the foregoing examples: any person skilled in the art may modify or easily conceive of the technical solution described in the foregoing embodiments, or perform equivalent substitution of some of the technical features, while remaining within the technical scope of the present disclosure; such modifications, changes or substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention, and are intended to be included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (6)
1. The communication signal modulation mode open set identification method based on deep learning is characterized by comprising the following steps:
training and optimizing a deep neural network for signal identification by utilizing a similarity matrix between a centroid vector and a signal sample feature vector, and setting a threshold value for rejecting unknown classes under a corresponding modulation class according to similarity distribution between the centroid vector and the signal sample feature vector, wherein the centroid vector is a feature vector of a baseband ideal modulation signal of a known class, and each known modulation class corresponds to one centroid vector; aiming at the feature vector and the centroid vector of the signal sample, the similarity set is divided into an intra-class set and an extra-class set by acquiring the similarity set between the feature vector and the centroid vector and utilizing a maximum inter-class variance method, and a threshold value for rejecting unknown classes under the corresponding modulation class is set according to the variance in the inter-class set; setting a lower threshold value of a corresponding modulation category by using a maximum inter-category variance method, setting intra-category and extra-category threshold variables, and obtaining the proportion of the intra-category and extra-category sets to the total sample number and the average value of the similarity by calculating the probability of the similar samples corresponding to the similarity value; acquiring an overall similarity mean value according to the proportion of the samples in the intra-class and the extra-class sets to the total number of the samples and the similarity mean value, acquiring data corresponding to a threshold variable which enables the inter-class variance to be maximum by combining the inter-class variance, and taking the data corresponding to the threshold variable as a threshold value for setting a corresponding modulation class;
and acquiring a similarity set between the target communication signals in the signal set to be identified and the centroid vector by using the trained deep neural network, taking the category corresponding to the maximum value in the similarity set as a prediction category, and comparing the maximum value with a set threshold value to judge whether the prediction category is an unknown category.
2. The deep learning-based communication signal modulation scheme open set identification method according to claim 1, wherein the deep neural network comprises: the device comprises a feature mapping unit and a nonlinear dimension reduction unit, wherein the feature mapping unit utilizes a residual error contraction module added in a convolution layer to carry out convolution operation on input signal data so as to extract feature information, compresses the feature information by gradually decreasing the number of convolution kernel channels layer by layer, and keeps the input feature dimension of each layer consistent by zero filling operation; feature dimension reduction is performed through a full connection layer in the nonlinear dimension reduction unit, and batch standardization and regularization are used for preventing overfitting in the model training process.
3. The deep learning-based communication signal modulation scheme open set recognition method according to claim 1 or 2, wherein in the deep neural network training optimization, baseband I/Q data of a baseband ideal modulation signal is input as a deep neural network, training is performed by using a GE2E loss function, and an average value of feature vectors output by the deep neural network is used as a centroid vector.
4. The deep learning-based communication signal modulation mode open set recognition method according to claim 3, wherein in the network training by using a GE2E loss function, baseband I/Q data of N×M baseband ideal modulation signals are input into a deep neural network as a batch to obtain normalized network output feature vectors, wherein N represents the number of signal modulation categories, and M represents the number of signal samples in each modulation category; and aggregating the same-class feature vectors by using M signal sample average vectors in each modulation class and the similarity between each sample feature vector and the corresponding signal sample average vector, and training and optimizing the network parameters far from the different-class feature vectors.
5. The method for identifying an open set of communication signal modulation schemes based on deep learning according to claim 1, wherein when judging whether the predicted class is an unknown class, comparing a maximum value in a similarity set obtained from the predicted class with a set threshold, if the maximum value is larger than the set threshold, judging that the predicted class is correct and outputting the predicted class, otherwise, judging that the predicted class is wrong and outputting the unknown class.
6. A communication signal modulation mode open set recognition system based on deep learning is characterized by comprising: a model construction module and a target recognition module, wherein,
the model construction module is used for training and optimizing the deep neural network for signal identification by utilizing a similarity matrix between a centroid vector and a signal sample feature vector, and setting a threshold value for rejecting unknown classes under corresponding modulation classes according to similarity distribution between the centroid vector and the signal sample feature vector, wherein the centroid vector is a feature vector of a baseband ideal modulation signal of a known class, and each known modulation class corresponds to one centroid vector; aiming at the feature vector and the centroid vector of the signal sample, the similarity set is divided into an intra-class set and an extra-class set by acquiring the similarity set between the feature vector and the centroid vector and utilizing a maximum inter-class variance method, and a threshold value for rejecting unknown classes under the corresponding modulation class is set according to the variance in the inter-class set; setting a lower threshold value of a corresponding modulation category by using a maximum inter-category variance method, setting intra-category and extra-category threshold variables, and obtaining the proportion of the intra-category and extra-category sets to the total sample number and the average value of the similarity by calculating the probability of the similar samples corresponding to the similarity value; acquiring an overall similarity mean value according to the proportion of the samples in the intra-class and the extra-class sets to the total number of the samples and the similarity mean value, acquiring data corresponding to a threshold variable which enables the inter-class variance to be maximum by combining the inter-class variance, and taking the data corresponding to the threshold variable as a threshold value for setting a corresponding modulation class;
the target recognition module is used for acquiring a similarity set between the target communication signals in the signal set to be recognized and the centroid vector by utilizing the trained deep neural network, taking the category corresponding to the maximum value in the similarity set as a prediction category, and comparing the maximum value with a set threshold value to judge whether the prediction category is an unknown category.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210095612.8A CN114567528B (en) | 2022-01-26 | 2022-01-26 | Communication signal modulation mode open set recognition method and system based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210095612.8A CN114567528B (en) | 2022-01-26 | 2022-01-26 | Communication signal modulation mode open set recognition method and system based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114567528A CN114567528A (en) | 2022-05-31 |
| CN114567528B true CN114567528B (en) | 2023-05-23 |
Family
ID=81714050
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210095612.8A Active CN114567528B (en) | 2022-01-26 | 2022-01-26 | Communication signal modulation mode open set recognition method and system based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114567528B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114818839B (en) * | 2022-07-01 | 2022-09-16 | 之江实验室 | Deep learning-based optical fiber sensing underwater acoustic signal identification method and device |
| CN115720184B (en) * | 2022-10-08 | 2024-04-19 | 西安电子科技大学 | Small sample signal modulation type identification method based on characteristic distribution |
| CN115913850B (en) * | 2022-11-18 | 2024-04-05 | 中国电子科技集团公司第十研究所 | Open set modulation identification method based on residual error network |
| CN115728588B (en) * | 2022-12-23 | 2023-06-13 | 广州力赛计量检测有限公司 | Electromagnetic compatibility detection system and method based on big data |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109802905A (en) * | 2018-12-27 | 2019-05-24 | 西安电子科技大学 | Digital signal Automatic Modulation Recognition method based on CNN convolutional neural networks |
| CN111461025A (en) * | 2020-04-02 | 2020-07-28 | 同济大学 | Signal identification method for self-evolving zero-sample learning |
| CN113705446A (en) * | 2021-08-27 | 2021-11-26 | 电子科技大学 | Open set identification method for individual radiation source |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11881287B2 (en) * | 2016-11-10 | 2024-01-23 | Precisionlife Ltd | Control apparatus and method for processing data inputs in computing devices therefore |
-
2022
- 2022-01-26 CN CN202210095612.8A patent/CN114567528B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109802905A (en) * | 2018-12-27 | 2019-05-24 | 西安电子科技大学 | Digital signal Automatic Modulation Recognition method based on CNN convolutional neural networks |
| CN111461025A (en) * | 2020-04-02 | 2020-07-28 | 同济大学 | Signal identification method for self-evolving zero-sample learning |
| CN113705446A (en) * | 2021-08-27 | 2021-11-26 | 电子科技大学 | Open set identification method for individual radiation source |
Non-Patent Citations (2)
| Title |
|---|
| 基于卷积双向长短期神经网络的调制方式识别;谭继远;张立民;钟兆根;;火力与指挥控制(第06期);全文 * |
| 基于频谱特征的深度学习信号检测方法研究;姚朋;中国;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114567528A (en) | 2022-05-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114567528B (en) | Communication signal modulation mode open set recognition method and system based on deep learning | |
| CN109802905B (en) | CNN convolutional neural network-based digital signal automatic modulation identification method | |
| CN114465855B (en) | Automatic modulation recognition method based on attention mechanism and multi-feature fusion | |
| CN112039820B (en) | Communication signal modulation and identification method for quantum image group mechanism evolution BP neural network | |
| CN112749633B (en) | Separate and reconstructed individual radiation source identification method | |
| CN109120563B (en) | Modulation recognition method based on neural network integration | |
| CN112787964B (en) | BPSK and QPSK signal modulation identification method based on range median domain features | |
| CN107612867A (en) | A kind of order of modulation recognition methods of MQAM signals | |
| CN112737992B (en) | Underwater sound signal modulation mode self-adaptive in-class identification method | |
| CN114422311B (en) | Signal modulation recognition method and system combining deep neural network and expert priori features | |
| Zhang et al. | Open set recognition of communication signal modulation based on deep learning | |
| CN116628566A (en) | Communication signal modulation classification method based on aggregated residual transformation network | |
| CN107707497B (en) | Communication signal identification method based on subtraction clustering and fuzzy clustering algorithm | |
| Tan et al. | Specific emitter identification based on software-defined radio and decision fusion | |
| CN113095162B (en) | Spectrum sensing method based on semi-supervised deep learning | |
| CN114398931A (en) | Modulation recognition method and system based on numerical characteristic and image characteristic fusion | |
| Wang et al. | Specific emitter identification based on deep adversarial domain adaptation | |
| CN109274626A (en) | A kind of Modulation Identification method based on planisphere orthogonal scanning feature | |
| CN115913850B (en) | Open set modulation identification method based on residual error network | |
| CN117131436A (en) | Radiation source individual identification method oriented to open environment | |
| CN108494711B (en) | Communication signal map domain feature extraction method based on KL divergence | |
| CN113420817B (en) | Semi-supervised modulation type identification method, device and medium based on network structure characteristic induction | |
| CN111130694A (en) | PDCCH blind detection method and system for low-delay clustering | |
| CN114900406B (en) | Blind modulation signal identification method based on twin network | |
| Hao et al. | Modulation classification using a goodness of fit test |
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 |