CN114943245A

CN114943245A - Automatic modulation recognition method and device based on data enhancement and feature embedding

Info

Publication number: CN114943245A
Application number: CN202210364955.XA
Authority: CN
Inventors: 谭凯文; 闫文君; 于柯远; 凌青; 张立民; 王程昱; 段可欣
Original assignee: School Of Aeronautical Combat Service Naval Aeronautical University Of Pla
Current assignee: School Of Aeronautical Combat Service Naval Aeronautical University Of Pla
Priority date: 2022-04-07
Filing date: 2022-04-07
Publication date: 2022-08-26

Abstract

The invention discloses an automatic modulation identification method and device based on data enhancement and feature embedding, wherein the method comprises the following steps: acquiring original data sets of various modulation signals; wherein the raw data set of the plurality of modulated signals comprises at least raw data set samples of the plurality of modulated signals; inputting an original data set of various modulation signals into a trained deep learning model for data fusion processing; the degree learning model is obtained by training a deep dense generation countermeasure network DD-GAN and a selective kernel convolution neural network SK-CNN; and outputting a discrimination probability matrix of a deep learning model of the original data set of various modulation signals based on data fusion processing to obtain a signal classification result. According to the invention, by setting appropriate extension parameters, the researched scheme based on data enhancement and SPWVD embedding can enrich signal characteristics, and effectively solve the problem of insufficient original signal samples.

Description

Automatic modulation recognition method and device based on data enhancement and feature embedding

Technical Field

The invention relates to the technical field of automatic modulation identification, in particular to an automatic modulation identification method and device based on data enhancement and feature embedding.

Background

Automatic Modulation Recognition (AMR) is the detection of the modulation class of the received signal, which is the basis for the demodulation of the received signal and the further extraction of the information. AMR is critical for spectrum monitoring, cognitive radio, and Electronic Warfare (EW). With the rapid development of information technology, modern electromagnetic environments become more and more complex. The strong noise interference, the complex waveform modulation and the abrupt change of the characteristic parameters of the signal cause the reliability of the conventional AMR method based on five conventional parameters to be sharply reduced. The existing research results are analyzed from the aspects of time domain, frequency domain and time-frequency domain, and various feature extraction methods are provided. In general, early AMR schemes can be divided into two types, conventional likelihood ratio estimation and expert feature extraction. The maximum likelihood method is adopted according to the phase recognition signal modulation mode, and the performance of various modulation constellation classifications is analyzed. A theoretical framework is established under the hierarchical assumption by using the basic accumulated amount and the cyclic accumulated amount of the received signals. This framework is still feasible in several situations (i.e., phase shift, frequency shift, and fading channels). However, the traditional method for manually extracting features mainly relies on expert experience and a great deal of prior knowledge, so that the identification method is poor in robustness and low in fault tolerance. With the rapid development of Deep Learning (DL) in various fields over the past several years, AMR based on deep learning is drawing more and more attention.

DL makes breakthrough progress in image classification, natural language processing and target detection, and has the advantages of automatic feature extraction and strong model generalization capability. Compared to the two methods described above, the DL-based AMR scheme has a lower complexity and does not require manual design features, thus showing a huge potential in complex scenarios where the signal probability density is unknown. A multi-platform fusion recognition system structure is established by applying an ensemble learning framework, deep features of signal time-frequency images are extracted by setting weights of different networks, and the redundant process of manual design and feature recognition is effectively avoided. And denoising the signal under the low signal-to-noise ratio by using an adaptive singular value reconstruction algorithm based on intra-pulse characteristics, and selecting ResNet to identify the time-frequency image after morphological filtering. When the noise ratio is-8 dB, the overall recognition rate of eight signals reaches 94.1%. A sparse connection convolution layer composed of a plurality of basic modules is designed, and loss of signal feature extraction details is effectively avoided when the network depth is increased. With enough training samples, the DL-based AMR scheme is robust to distribute the same data, especially in extracting high-dimensional features of a large amount of data. In practice, it is difficult to obtain large amounts of sample due to environmental factors.

The complexity of the electromagnetic environment and the non-cooperativity of the reconnaissance target. A counter-transfer learning architecture (ATLA) based on DL automodulation classification was developed to reduce the difference between data distributions and implement AMR under limited samples.

Disclosure of Invention

The present invention is directed to solving, at least to some extent, one of the technical problems in the related art.

To this end, the object of the present invention is to propose an automatic modulation recognition method based on data enhancement and feature embedding, first of all a deep dense generation countermeasure network (DD-GAN) was developed to extend the original data set of five modulation signals produced by unsupervised GNU radios. A smooth pseudo-Wigner-Ville distribution (SPWVD) is selected as the time-frequency representation of the signal and embedded in the AMR framework. And inputting the real signal into a lightweight CNN for time correlation extraction, wherein the SK-CNN is used for focusing a useful area in time-frequency image classification. The convergence performance of the DD-GAN was also analyzed and the best hyper-parametric combinations to improve the classification quality were given. The extended and trained DL model has better robustness and recognition performance on an original test set.

Another object of the present invention is to provide an automatic modulation recognition apparatus based on data enhancement and feature embedding.

In order to achieve the above object, an aspect of the present invention provides an automatic modulation identification method based on data enhancement and feature embedding, including:

acquiring original data sets of various modulation signals and original data set samples of various modulation signals; inputting the original data sets of the various modulation signals into a trained deep learning model for data fusion processing; the trained deep learning model is obtained by training a DD-GAN (direct digital-to-GAN) antagonistic network and an SK-CNN (selective kernel convolutional neural network) based on original data set samples of the various modulation signals through deep intensive generation; and outputting a discrimination probability matrix of the deep learning model of the original data set of the various modulation signals based on the data fusion processing so as to obtain a signal classification result.

According to the automatic modulation identification method based on data enhancement and feature embedding, provided by the embodiment of the invention, by setting appropriate extension parameters, the researched scheme based on data enhancement and SPWVD embedding can enrich signal features, and the problem of insufficient original signal samples is effectively solved.

In order to achieve the above object, another aspect of the present invention provides an automatic modulation identification apparatus based on data enhancement and feature embedding, including:

the acquisition module is used for acquiring original data sets of various modulation signals and original data set samples of various modulation signals; the fusion module is used for inputting the original data sets of the various modulation signals into a trained deep learning model for data fusion processing; the trained deep learning model is obtained by training a countermeasure network DD-GAN and a selective kernel convolution neural network SK-CNN through deep intensive generation based on original data set samples of the various modulation signals; and the output module is used for outputting a discrimination probability matrix of the deep learning model of the original data set of the various modulation signals based on the data fusion processing so as to obtain a signal classification result.

According to the automatic modulation recognition device based on data enhancement and feature embedding, disclosed by the embodiment of the invention, by setting appropriate extension parameters, the researched data enhancement and SPWVD embedding-based automatic modulation recognition device can enrich signal features and effectively solve the problem of insufficient original signal samples.

Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.

Drawings

The above and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:

FIG. 1 is a flow chart of an automatic modulation identification method based on data enhancement and feature embedding according to an embodiment of the present invention;

FIG. 2 is an overall block diagram of an AMR system according to an embodiment of the present invention;

fig. 3 is a SPWVD time-frequency diagram corresponding to five signals under the condition of SNR of 5dB according to an embodiment of the present invention;

FIG. 4 is a graph of loss functions for G and D according to an embodiment of the present invention;

FIG. 5 is a schematic diagram of the logarithmic entropy of a DD-GAN according to an embodiment of the present invention;

FIGS. 6(a) and 6(b) are graphs showing the effect of DD-GAN generating a Constant Wave (CW) signal after 1000 and 5000 iterations, respectively, in accordance with an embodiment of the present invention;

FIG. 7 is a diagram of an optional kernel module according to an embodiment of the present invention;

FIG. 8 is a schematic structural diagram of SK-CNN according to an embodiment of the invention;

FIGS. 9(a) and 9(b) are schematic diagrams of SK-CNN and CNN for training on an unexpanded training set and then testing on a test set, respectively, according to an embodiment of the invention;

FIG. 10 is a schematic diagram illustrating an effect of a first ER on AMR frame recognition performance according to an embodiment of the present invention;

FIG. 11 is a schematic diagram illustrating the effect of a second ER on AMR framework recognition performance according to an embodiment of the present invention;

FIG. 12 is a schematic diagram illustrating the effect of a third ER on AMR framework recognition performance according to an embodiment of the present invention;

FIG. 13 is a schematic diagram illustrating the effect of a fourth ER on AMR frame identification performance according to an embodiment of the present invention;

FIG. 14 is a schematic diagram illustrating the effect of a fifth ER on AMR framework recognition performance according to an embodiment of the present invention;

FIG. 15 is a schematic diagram illustrating the recognition performance of different ERs under an AMR framework according to an embodiment of the present invention;

16(a), 16(b), 16(c) and 16(d) are schematic diagrams of data enhancement at feature-embedded confusion matrices and different SNRs and ERs, respectively, according to an embodiment of the invention;

FIG. 17 is a graphical illustration of a comparison of the recognition performance of a feature embedding algorithm with different baseline schemes in accordance with an embodiment of the present invention;

fig. 18 is a schematic structural diagram of an automatic modulation recognition apparatus based on data enhancement and feature embedding according to an embodiment of the present invention.

Detailed Description

It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict. The present invention will be described in detail below with reference to the accompanying drawings in conjunction with embodiments.

In order to make the technical solutions of the present invention better understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, shall fall within the protection scope of the present invention.

An automatic modulation recognition method and apparatus based on data enhancement and feature embedding according to an embodiment of the present invention will be described with reference to the accompanying drawings.

Fig. 1 is a flow chart of an automatic modulation recognition method based on data enhancement and feature embedding according to an embodiment of the present invention.

As shown in fig. 1, the method includes, but is not limited to, the following steps:

s1, acquiring original data sets of various modulation signals and original data set samples of various modulation signals;

s2, inputting the original data set of various modulation signals into a trained deep learning model for data fusion processing; the trained deep learning model is obtained by training a confrontation network DD-GAN and a selective kernel convolution neural network SK-CNN through deep intensive generation based on original data set samples of various modulation signals;

and S3, outputting a discrimination probability matrix of the deep learning model of the original data set of the various modulation signals based on data fusion processing to obtain a signal classification result.

The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

Specifically, the invention provides a scheme for identifying five types of modulation signals based on combination of data expansion and a dynamic convolution network. First, a DD-GAN comprising multiple fully connected layers is designed to extend the data of the original finite sample. They are input into the generative confrontation model and "artificial samples" are generated to enrich the sample characteristics. In order to effectively extract the time domain and frequency domain characteristics of the signal, the complementarity of the data form is fully utilized, and two parallel CNN branches are designed. The input of one CNN is a real I/Q signal, the input of the other branch is SPWVD time-frequency characteristics, and a Selective Kernel (SK) module is introduced to construct a light CNN for image recognition. Finally, the output of the Softmax layer is fused to obtain reliable classification results. The overall model is shown in fig. 2. Samples of five modulated signals (i.e., BPSK, QPSK, BASK, BFSK, and QFSK) are randomly divided into a training set and a test set. And extracting the SPWVD of the modulation signal and the original I/Q sample, and respectively sending the SPWVD to the same classifier for online training. And introducing white Gaussian noise into the enhanced training set, and checking the robustness. And after the training is finished, storing the optimal network models under different Enhancement Rates (ER), and performing offline test on the test set. And finally, outputting the discrimination probability matrix of the DL model of the test signal sample. Experiments show that by setting appropriate extension parameters, the scheme based on data enhancement and SPWVD embedding researched by the invention can enrich signal characteristics and effectively solve the problem of insufficient original signal samples.

And further, constructing a signal model. In practice, supervised modulation recognition can be reduced to an N-class decision problem, in which a modulated discrete sequence signal is input and decision probabilities for different signal samples are output from the system. A GAN-based modulated signal classification framework is proposed (as shown in fig. 2).

For MFSK signals, the original transmitted signal before sampling is

Where the carrier frequency is indicated. For MPSK signals, it can be expressed as:

for the BASK signal, it can be expressed as

In general, the received signal is modeled as:

r(t)＝s(t)*h(t)+n(t) (1)

where s (t) represents the transmitted signal, h (t) is the equivalent channel impulse response, n (t) is additive noise, and r (t) is the complex baseband signal, i.e.:

r(t)＝r _i (t)+jr _q (t) (2)

wherein r is _i (t) and r _q (t) denotes an in-phase component and a quadrature component of the received signal, respectively. Thus, the signal sampled by the receiver can be represented as:

and further, performing time-frequency characteristic analysis. SPWVD is an optimized Cohen time-frequency analysis method based on Wigner-Ville distribution. Compared with Wigner-Ville time frequency distribution, SPWVD can eliminate cross term interference more effectively. The frequency range has high resolution and recognition accuracy for signals in different backgrounds. The analytic signal of the signal is written as:

x(t)＝s(t)+jH[s(t)] (4)

where H [. cndot ] represents the Hilbert transform of the signal, SPWVD is defined as:

wherein

An analytic signal representing s (t),

representing the conjugate analysis signal, h (τ) represents the window function of the time domain filtering, and represents the window function of the frequency domain filtering. Therefore, SPWVD can effectively inhibit cross terms, and the noise of the transformed time-frequency characteristic image is obviously reduced. Image processing methods may then be used to extract the time-frequency characteristics of the signal.

Fig. 3 is a SPWVD time-frequency image showing five signals corresponding to the SNR-5 dB condition, where the signal characteristics are significantly different and easy to classify.

Further, a DD-GAN framework is proposed. Generative challenge networks (GAN) refer to network architectures that continually optimize generative and discriminative models based on challenge training. The method has the advantages that when the prior hypothesis is unknown, the potential distribution rule of the sample can be checked, and new data can be generated through unsupervised learning so as to enrich the characteristics of the sample. The GAN consists of a discriminator and a generator, whose objective function is of the form:

wherein p is _z (z) represents the prior probability density of the potential noise vector, and the noise is mapped to by the generator. D (x) represents the decision probability of the discriminator output, x being derived from the real data.D is trained to effectively distinguish between real data and generated data to maximize V (D, G). The training purpose of G is to fool D. It does not allow D to determine the generated data, but rather minimizes D. When the models of D and GAR are both multi-layer perceptrons (MLPs), the nodes of adjacent fully-connected layers are connected to each other to map the distribution characteristics of the front-end to the label space of the sample. Through the superposition of a plurality of fully connected layers, G and D can effectively enrich the sample characteristics based on the deep mining data characteristics. Thus, in our study, full connectivity layer pairs were used to construct G and D, and DD-GAN. The envelopes of the 5 kinds of modulation signals are directly used as input, and the time sequence relation between adjacent sampling points of the original signal is fully captured.

Further, the structure of DD-GAN. For G, the input is a 1 × 256 gaussian vector and the output is a 1 × 1024 sequence through multiple fully connected layers. Except the last layer, a LeakyRelu function is adopted to activate hidden layer neurons, the derivative of the negative part is increased, silent neurons are reduced, and the problem of gradient disappearance in a fully-communicated structure is solved. The negative activation value of the LeakyRelu function is set to 0.2. And adding a Batch Normalization (BN) layer after the Full Connection (FC) layer to unify the output sequence of the neurons and accelerate the convergence speed of the model. The last output layer is activated by the tanh function to limit the amplitude of the generated signal to [ -1,1 ]. For D, the input of which is true samples with dimension 1 × 1024 and "false samples" generated by G, the discriminant probability of each sample is output after activating the multiple fully-connected layers and the LeakyRelu function, and the activation function of the output layer is set to Sigmoid. The structures of G and D are shown in tables 1 and 2.

TABLE 1 Generator parameters

TABLE 2 discriminator parameters

By choosing a suitable activation function and taking advantage of the good fitting capabilities of MLP, the DD-GAN proposed by the present invention can perform an approximation of any non-linear function, so it can make full use of the temporal correlation between actual input signal samples. The loss function is set to cross entropy, D and G are optimized by Adam, and the learning rate is set to 0.002. In training G, D is disabled from training and the batch size is set to 128.

And further, carrying out convergence analysis. For GAN, since G and D are always in the dynamic countermeasure process, the loss function cannot reach a stable state, and therefore, it is not suitable to determine the convergence timing by the loss function. Therefore, the logarithmic entropy of the generated sample is selected as an index for evaluating the DD-GAN convergence performance. The logarithmic entropy can effectively describe the closeness degree of the generated data and the real data, and is an effective method for measuring the GAN convergence. The expression of the logarithmic entropy is as follows:

wherein x _i Representing received signal samples, G (z) _i ) Is the output of G. As can be seen from fig. 4, since the training schedules of G and D are synchronized, it generally causes the loss value of the generator to increase sharply as the loss value of the discrimination network increases.

Fig. 5 depicts the convergence performance trend with the number of iterations. After 1000 epochs are reached, the curve changes gradually and stably, and the logarithmic entropy value fluctuates between 40 and 80, which indicates that the model has higher convergence speed.

Fig. 6 generation of time domain waveforms of a continuous wave signal fig. 6(a) and fig. 6(b) show the effect of DD-GAN to generate a Constant Wave (CW) signal after 1000 and 5000 iterations, respectively. The initial phase of the original CW signal is 1MHz at the carrier frequency and 1024 samples. The results show that as the number of iterations increases, the predicted value of the generator gradually approaches the envelope of the original signal, and the spurious noise component in the signal is significantly reduced.

Further, the recognition result is analyzed, wherein the kernel structure is selected, and the perception fields (convolution kernels) with different sizes have different influences on the targets with different scales. The SK module is a module that dynamically generates convolution kernels for images of different scales, and the size of the feature map is unchanged. In the selective kernel module (shown in FIG. 7), the feature map is represented by an input size. The operations involved in this architecture include mainly splitting, fusing and selecting. The segmentation is a process of convolving the original feature map with different checks, and the invention sets two types. Fuse refers to an operation that imparts weight to the inner core. Selection is the process of obtaining a new feature map by cores with different weights.

First, the sum of convolution operations is performed using convolution kernels of sizes 3 × 3 and 5 × 5, respectively. Second, a gating mechanism is used to fuse the information of the two channels and add an output feature map. Generating channel statistics by the global pool:

after fc layer dimensionality reduction, the output features can be expressed as:

z＝F _fc (s)＝δ(B(W _S )) (9)

where δ (-) is the Relu activation function and B (-) is the batch normalization. The dimension of z is the number of kernels, W _S ∈R ^d×C And a compression factor r is introduced to determine the effect of the full connection dimension d on the network efficiency as follows:

where L is the minimum of dimension d. Finally, the weight characteristic z after the self-adaptive calibration is loaded to the characteristic by adopting a Softmax function

And

inputting the adaptive calibration into the feature map and outputting the feature V ═[V ₁ ,V ₂ ,...,V _C ]Is provided with

Further, analyzing the network structure, fig. 3 shows SPWVD distributions of five types of signals. Useful information in time-frequency distribution of modulation types is often concentrated in key areas, in order to improve the 'attention' of the network in the key areas, an SK module is introduced into a lightweight CNN to construct an SK-CNN, and a receiving domain can be adaptively changed according to the multi-scale of input information. To accommodate the limited sample case proposed by the present invention, avoiding the occurrence of overfitting, the CNN used for classification consists of only 4 convolutional layers and 3 fully-connected layers, with the number of convolutional kernels in each convolutional layer set to 32, 64, 128, and 256, respectively. The kernel size is set to 3 x 3, the rectifying linear unit (ReLu) as an activation function. After each convolutional layer, the largest pool layer is added, the pool size is 2 × 2, the step size is 2, and the network input is a three-channel time-frequency image with a size of 224 × 224. The proposed SK-CNN structure is shown in fig. 8.

Inside the SK-CNN, the input 224 × 3 time-frequency image passes through the first convolution layer, and a feature map is output with a size of 224 × 32. After the SK module recalibrates the function, the output size remains unchanged. With max pool layers, the output size is 112 × 112 × 32, etc. Finally, through the Softmax layer, a network for 5 modulation signal identification vectors is output.

Further, CNN for I/Q signal classification. The CNN for identifying IQ signals proposed by the present invention is composed of 2 convolutional layers and 4 full-link layers, and the structure is shown in table 3. The activation function of each layer is set to Relu. The Softmax of the last layer is used for outputting a probability distribution matrix for identifying signal samples, and a loss function is set as cross entropy and used for marking the error between an actual label and a predicted label. An exit operation is used instead of a pool to preserve as much of the subtle features of the signal as possible and to some extent effectively avoid overfitting. In order to effectively utilize information in the SPWVD time-frequency image, a score fusion method is adopted to fuse the discrimination probability matrixes output by the two branch networks, and finally a reliable identification result is output.

TABLE 3 IQCNN parameter settings

Further, the effectiveness of the scheme is verified by adopting GNU radio based on an SDR platform and time-sharing transmission of five modulation signals (BPSK, QPSK, BFSK, QFSK and BASK). In addition, white gaussian noise is added to the baseband signal to simulate different channel environments. The signal-to-noise ratio range is set to be-20 dB to +10dB, the step length is 2, and the number of sampling points under each signal-to-noise ratio is 360000. The division ratio of the training set to the test set is 7: 3. In order to improve the generalization performance of the model, a k-fold cross validation method is adopted. In the testing process, a signal sample is divided into three parts, one part is used as a testing set, the second part is used as an original training set, and the third part is directly input into DD-GAN to generate a new sample and expand the original training set. To verify the degree of improvement of recognition performance by generating the confrontation model, the Enhancement Rate (ER) is defined as the ratio of the generated data to the original training data:

wherein G is _z (i) Single output of DD-GAN, X _train Representing the original training data.

The experimental data acquisition is completed on USRP-B210 equipment and a GNU radio platform, and Matlab R2020a is adopted for signal preprocessing and feature extraction. A network is constructed, trained and tested by adopting a GPU driven by NVIDIA GeForce RTX 3080, and the network is realized based on Keras 2.3.1 and Tensorflow 2.2.

Further, the results are discussed. Performance of the raw data: the performance of identifying SPWVD features using different baseline networks is analyzed. To verify the network performance of the present invention, we performed tests on the original data set, the test results are shown in fig. 9. The results of SK-CNN and CNN on the training set (ER ═ 0) are shown in fig. 9(a) and fig. 9(b), respectively. The result shows that when the signal-to-noise ratio is larger than 0db, the classification precision of the initial CNN without the dynamic convolution module on the five types of signals can reach more than 90%. After the SK module is embedded into the lightweight CNN to classify SPWVD images, the identification accuracy of BPSK, QPSK and BFSK signals is remarkably improved, and the SK-CNN has high compatibility with the three signals, however, the compatibility may be at the cost of the identification performance of the other two signals. 9(a) and 9(b) for SK-CNN and CNN for training on the unexpanded training set and then testing on the test set.

Performance of the enhancement data: fig. 10, 11, 12, 13 and 14 show the effect of different ERs on the AMR framework recognition performance. The results of BASK, BFSK, BPSK, QFSK and QPSK in sequence show that the identification rate of the BFSK signal after data enhancement is not obviously improved, the identification performance of other four modulation signals is improved to different degrees under the condition of low signal-to-noise ratio after DD-GAN expansion, but a fixed ER is difficult to find to obtain the highest accuracy. In fig. 14, when ER is 0.5, the performance is worst, probably due to overfitting. When ER is equal to 0.1, compared with data before enhancement, the overall recognition accuracy of the model on the five schemes is improved by 3% -7%, and when ER is equal to 0.2, the accuracy is improved to a certain extent, which shows that the proposed DD-GAN can effectively enrich the characteristics of the data set and expand the samples. The recognition performance of different ERs under the AMR framework is shown in fig. 15.

Further, the confusion matrix and modulation type: the invention has proved that when ER is 0.1 and 0.2, the overall recognition accuracy of the five types of signals is improved after data enhancement. The confusion matrix of the proposed model is as shown in fig. 16(a), fig. 16(b), fig. 16(c) and fig. 16(d) when SNR is-10 dB and 10 dB. In the Confusion matrix, 16(a) has a fusion matrix at SNR of 10dB and ER of 0.2, fig. 16(b) has a fusion matrix at SNR of-10 dB and ER of 0.2, fig. 16(c) has a fusion matrix at SNR of 10dB and ER of 0.1, and fig. 16(d) has an SNR of-10 dB and ER of 0.1.

When the signal-to-noise ratio is 10dB, the framework of the invention can accurately identify five modulation schemes, and the accuracy rate is basically 100%. When the signal-to-noise ratio is extremely low (-10dB), the overall recognition effect of ER 0.2 is worse than that of ER 0.1. Presumably this is because the data generated by the DD-GAN introduces noise. Therefore, when the signal-to-noise ratio is poor, the generated data is limited to improving the recognition performance. But the frame of the invention can make up for the defect of sample shortage to a certain extent and effectively improve the overall identification precision.

And (3) identification precision comparison: FIG. 17 is a comparison of the recognition performance of the proposed feature embedding algorithm with different baseline schemes. The result shows that compared with the scheme of directly sending I/Q signals to CNN and LSTM for identification, the characteristic embedding scheme can make up the deficiency of signal time correlation, and extracts the time-frequency information implicit in the signals as the characteristic suitable for fitting, thereby improving the efficiency of network identification. In the feature analysis scheme, the accuracy of the I/QCNN + SK-CNN framework is 5% higher than that of the I/QCNN + SE-CNN and 15% higher than that of the I/QCNN + Inception Net within the signal-to-noise ratio range of-10 dB to-2 dB.

In summary, the invention provides an AMR system based on data enhancement and time-frequency feature embedding. First, the authentic data collected from the GNU radio is learned using DD-GAN and merged with the original data set. Second, the processed signal samples are sent to two branches: one branch directly uses the 2D convolutional layer to identify the sequence of the received I/Q signal, and the other branch uses the attention mechanism to identify SPWVD time-frequency images in order to reveal the time-frequency relationship implied in the signal. And finally, after the Softmax fusion, summarizing the recognition results of the two branches so as to improve the overall recognition precision. Experimental results on actual data show that when ER is 0.1, the scheme has the strongest generalization capability on five modulation schemes, the precision of the scheme is about 5% higher than that of an original data set, and the AMR problem under the condition of a missing sample or a limited sample can be effectively solved.

According to the automatic modulation identification method based on data enhancement and feature embedding, provided by the embodiment of the invention, by setting appropriate extension parameters, the researched scheme based on data enhancement and SPWVD embedding can enrich signal features and effectively solve the problem of insufficient original signal samples.

In order to implement the above embodiment, as shown in fig. 18, an automatic modulation recognition apparatus 10 based on data enhancement and feature embedding is further provided in the present embodiment, where the apparatus 10 includes: an acquisition module 100, a fusion module 200 and an output module 300.

An obtaining module 100, configured to obtain raw data sets of multiple modulation signals and raw data set samples of the multiple modulation signals;

the fusion module 200 is used for inputting the original data sets of various modulation signals into the trained deep learning model for data fusion processing; the trained deep learning model is obtained by training a confrontation network DD-GAN and a selective kernel convolution neural network SK-CNN through deep intensive generation based on the original data set samples of the various modulation signals;

the output module 300 is configured to output a discrimination probability matrix of a deep learning model of an original data set of multiple modulation signals based on data fusion processing, so as to obtain a signal classification result.

According to the automatic modulation recognition device based on data enhancement and feature embedding, disclosed by the embodiment of the invention, by setting appropriate extension parameters, the researched scheme based on data enhancement and SPWVD embedding can enrich the signal features and effectively solve the problem of insufficient original signal samples.

It should be noted that the foregoing explanation of the embodiment of the automatic modulation identification method based on data enhancement and feature embedding is also applicable to the automatic modulation identification apparatus based on data enhancement and feature embedding of this embodiment, and is not repeated here.

Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality" means at least two, e.g., two, three, etc., unless specifically limited otherwise.

In the description herein, references to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.

Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.

Claims

1. An automatic modulation identification method based on data enhancement and feature embedding is characterized by comprising the following steps:

acquiring original data sets of various modulation signals and original data set samples of various modulation signals;

inputting the original data sets of the various modulation signals into a trained deep learning model for data fusion processing; the trained deep learning model is obtained by training a DD-GAN (direct digital-to-GAN) antagonistic network and an SK-CNN (selective kernel convolutional neural network) based on original data set samples of the various modulation signals through deep intensive generation;

and outputting a discrimination probability matrix of the deep learning model of the original data set of the various modulation signals based on the data fusion processing to obtain a signal classification result.

2. The method according to claim 1, wherein the trained deep learning model is trained by deep dense generation of an antagonistic network DD-GAN and a selective kernel convolutional neural network SK-CNN based on raw data set samples of the plurality of modulation signals, and comprises:

inputting the original data set samples of the various modulation signals into a depth-dense generation countermeasure network DD-GAN for expansion, and dividing the expanded samples into a training set and a test set; wherein, the first and the second end of the pipe are connected with each other,

identifying a sequence of a received I/Q signal by using a Convolutional Neural Network (CNN), identifying a smooth pseudo Wigner-Ville distribution (SPWVD) time-frequency image by using a selective kernel convolutional neural network (SK-CNN), and respectively sending the sequence of the I/Q signal and the SPWVD time-frequency image as the training set to the same Softmax classifier for training; and the number of the first and second groups,

after training is finished, the optimal network models under different enhancement rates ER are stored, and off-line testing is carried out on the test set, so that the trained deep learning model is obtained.

3. The method of claim 2, wherein modeling the plurality of modulated signals comprises:

r(t)＝s(t)*h(t)+n(t) (1)

where s (t) represents the transmitted signal, h (t) is the equivalent channel impulse response, n (t) is additive noise, and r (t) is the complex baseband signal, represented as:

r(t)＝r _i (t)+jr _q (t) (2)

wherein r is _i (t) and r _q (t) representing an in-phase component and a quadrature component of the received signal, respectively; the plurality of modulated signals sampled by the receiver are represented as:

4. the method of claim 3, wherein the analytic signal of the plurality of modulation signals is represented as:

x(t)＝s(t)+jH[s(t)] (4)

wherein H [. cndot. ] represents the Hilbert transform of the signal, SPWVD is defined as:

wherein the content of the first and second substances,

an analytic signal representing s (t),

representing the conjugate analysis signal and h (τ) represents the window function of the temporal filtering.

5. The method of claim 4, further comprising: the form of the goal function for the generative countermeasure network GAN is expressed as:

wherein p is _z (z) represents the prior probability density of the underlying noise vector, and D (x) represents the decision probability of the discriminator output.

6. The method of claim 5, further comprising: adopting a full connection layer to construct G and D and DD-GAN, selecting the logarithmic entropy of a generated sample as an index for evaluating the DD-GAN convergence performance, wherein the expression of the logarithmic entropy is as follows:

wherein x is _i Representing received signal samples, G (z) _i ) Is the output of G.

7. The method of claim 6, further comprising: fusing information of two channels of the selective kernel module by using a gating mechanism, adding output feature mapping, and generating channel statistical information through a global pool:

after fc layer dimensionality reduction, the output characteristic is expressed as:

z＝F _fc (s)＝δ(B(W _S )) (9)

where δ (-) is the Relu activation function, B (-) is the batch normalization, z dimension is the number of kernels, W _S ∈R ^d×C And a compression factor r is introduced to determine the influence of the full-connection dimension d on the network efficiency, as follows:

where L is the minimum value of dimension d.

8. The method of claim 7, further comprising: defining the enhancement rate ER as a ratio of generated data to samples of the original data set:

wherein G is _z (i) Single output, X, of DD-GAN _train Representing a sample of the original data set.

9. The method of claim 8, wherein the 224 x 3 time-frequency image input inside the SK-CNN passes through the first convolution layer, and outputs a signature map with a size of 224 x 32, passes through the max pool layer, outputs a size of 112 x 32, and passes through the Softmax classifier, outputting a trellis for 5 modulated signal identification vectors.

10. An automatic modulation identification apparatus based on data enhancement and feature embedding, comprising:

the acquisition module is used for acquiring original data sets of various modulation signals and original data set samples of various modulation signals;

the fusion module is used for inputting the original data sets of the various modulation signals into a trained deep learning model for data fusion processing; the trained deep learning model is obtained by training a countermeasure network DD-GAN and a selective kernel convolution neural network SK-CNN through deep intensive generation based on original data set samples of the various modulation signals;

and the output module is used for outputting a discrimination probability matrix of the deep learning model of the original data set of the various modulation signals based on the data fusion processing so as to obtain a signal classification result.