A small sample communication modulation recognition preprocessing method and system based on complex denoising network

Abstract

The invention provides a small sample communication modulation recognition preprocessing method and system based on a complex denoising network, which belongs to the field of communication and solves the problem of low communication modulation recognition accuracy caused by noise influence and small sample size in non-cooperative complex electromagnetic scenes. The method comprises: converting a training set into a time-frequency diagram through Choi-Williams distribution; extracting time-frequency diagram features by using CNN and obtaining signal-to-noise ratio categories by using K-means clustering algorithm, and distinguishing high signal-to-noise ratio signals by image information entropy; expanding the number of high signal-to-noise ratio signals by using rotation data enhancement and cyclic time shift enhancement methods to obtain enhanced signals; building CNRN by using complex cross terms to assist CLSTM on the basis of an autoencoder framework; obtaining F-CNRN by training CNRN with enhanced signals; inputting a test set into F-CNRN and outputting a denoised communication modulation signal. The invention reduces the influence of noise pollution on small sample communication modulation recognition and improves the accuracy and robustness of communication modulation recognition.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to a small sample communication modulation recognition preprocessing method and system based on a complex noise reduction network. Background Art

[0002] Automatic modulation recognition technology plays a fundamental and key role in studying the modulation signal system and transmission environment information of wireless communication. However, there are interference sources in complex space electromagnetic environments that bring various random noise pollution to the signal, affecting the accuracy of communication signal modulation recognition. In particular, noise of different intensities has a great impact on the recognition effect. In order to enhance the robustness of communication modulation recognition under different signal-to-noise ratios, it is necessary to carry out noise reduction preprocessing research on communication modulation signals. Under the development trend of intelligent engineering application technology, the application field of deep learning is becoming more and more extensive. Many scholars have begun to study how to make full use of the powerful self-learning ability of neural networks to achieve the purpose of noise reduction of communication signals. As a common deep learning noise reduction method, the denoising autoencoder mainly includes two parts: encoding process and decoding process. In the encoding process, the high-dimensional features of the noisy signal are reduced in dimension to extract hidden variables to obtain the essential features of the signal. The decoding process restores the essential features of the signal in the hidden variables to the data features of the same dimension as the noisy signal features, thereby achieving the purpose of noise reduction and obtaining the noise-reduced signal data.

[0003] As a signal data preprocessing method, good noise reduction effect lays a good foundation for subsequent signal detection estimation, recognition and analysis. As a space electromagnetic signal, the communication modulation signal has also achieved remarkable results in the research on noise reduction methods for this signal. However, the current communication modulation signal noise reduction method based on deep learning, such as the noise reduction autoencoder, still has some shortcomings:

[0004] 1. The non-cooperative communication signals obtained in actual scenarios contain signals with different signal-to-noise ratios. The noise contained in high signal-to-noise ratio signals will not overwhelm the signal features and has little effect on the extraction of the essential features of the signal. The noise contained in low signal-to-noise ratio signals has a significant impact on the extraction of signal features. The modulated signals with poor noise resistance in the signals may even be overwhelmed by the noise. It is impossible to construct samples of low signal-to-noise ratio signals and clean signals, which affects the training of the deep learning denoising network.

[0005] 2. The denoising method based on deep learning needs to rely on a large amount of labeled sample data when training neural networks. However, with the development of communication technology, the modulation styles of communication signals are becoming more and more numerous and complex. Especially in non-cooperative scenarios, when performing denoising on communication modulated signals, there are difficulties such as low interception rate of non-cooperative signals, insufficient information acquisition of some detected signals, and limited manpower and material resources. As a result, the number of high-quality and high-reliability modulation signal label samples is small and cannot support network training.

[0006] 3. After sampling, the data of each sampling point of the real communication modulated signal is complex. When processing the signal, it is split into real and imaginary parts and stored in a two-dimensional real matrix. Most of the existing deep learning-based denoising networks use real networks based on the real number properties of signal storage data. Therefore, the amplitude and phase characteristics of the communication modulated signal are ignored, and it is impossible to extract the more essential characteristics of the signal from the communication modulated signal submerged by noise.

[0007] 4. The existing noise reduction method directly performs noise reduction on high and low signal-to-noise ratio signals together, which is likely to cause quality loss of high signal-to-noise ratio signals, resulting in a decrease in the recognition accuracy of high signal-to-noise ratio signals when the modulation mode is subsequently identified. Summary of the invention

[0008] In view of the complex and diverse interference sources in the space electromagnetic environment in the prior art, the received communication modulation signal is affected by different degrees of noise, and the signal noise reduction preprocessing operation plays an important role in the next step of modulation identification and demodulation analysis. The method of using deep learning to reduce the noise of the signal requires a large number of samples to support network training, but is limited by the number of samples in non-cooperative scenarios. At the same time, most of the existing noise reduction networks are real networks, lacking the characteristic expression of complex signals. In order to solve the problem of low communication modulation recognition accuracy caused by noise influence and small sample size in non-cooperative complex electromagnetic scenarios, the present invention provides a small sample communication modulation recognition preprocessing method and system based on a complex noise reduction network. Through the disclosed technical solution, the pollution of noise in complex electromagnetic environments to communication modulation signals is reduced, so that the robustness of communication modulation recognition under small sample conditions is enhanced, laying a good foundation for subsequent signal detection estimation and recognition analysis.

[0009] To achieve the above object, the present invention proposes a small sample communication modulation recognition preprocessing method based on a complex denoising network, comprising:

[0010] Step 1: Use stratified sampling method to extract noisy signals in the data set according to a preset ratio and divide them into training set and test set;

[0011] Step 2: Convert the noisy signal in the training set into a time-frequency graph through Choi-Williams distribution; use the convolutional neural network CNN to extract the time-frequency graph features and use the K-means clustering algorithm to perform signal-to-noise ratio clustering on the time-frequency graph features to obtain two categories based on the signal-to-noise ratio, and calculate the image information entropy of each signal-to-noise ratio category; identify the high signal-to-noise ratio signal based on the image information entropy;

[0012] Step 3: The number of high signal-to-noise ratio signals is expanded by a rotation data enhancement method and a cyclic time-shift enhancement method to obtain an enhanced signal;

[0013] Step 4: Based on the autoencoder framework, a complex autoencoder denoising network model CNRN is constructed by using complex cross terms to assist the complex long short-term memory network CLSTM, and the enhanced signal is input into the CNRN for automatic noise addition to obtain the noisy signal;

[0014] Step 5. Combine the enhanced signal and the corresponding noisy signal into sample pairs, train CNRN, and obtain the complex denoising network model F-CNRN of the trained small sample communication modulation signal; input the test set into F-CNRN, and output the denoised communication modulation signal.

[0015] In step 1, the method of dividing the training set and the test set is:

[0016] The numpy.random.choice function in the computer python language is used to extract the noisy signals in the data set according to the preset ratio and divide them into training set and test set.

[0017] In step 2, the Choi-Williams distribution is:

[0018]

[0019] Where P(t,f) represents the time-frequency distribution, y(·) represents the signal, * represents the conjugate, u represents the integral variable, τ represents the time shift, Φ(τ,ν) is the kernel function value, ν represents the frequency shift, j represents the imaginary unit, t represents the time, and f represents the frequency;

[0020] The calculation formula of the kernel function value Φ(τ,ν) is as follows:

[0021] Φ(τ,ν)=exp(-(2πντ) 2 / c);

[0022] Where c is the scale factor.

[0023] In step 2, the K-means clustering algorithm is:

[0024]

[0025] Among them, G k-means ((X,d),(C 1 ,…,C K )) is the clustering result, i.e., the signal-to-noise ratio category, and d represents the distance; the clustering loss function in the K-means clustering algorithm is defined as cluster C k From sample x to center point μ k The mean square distance of ; K represents the total number of clusters, k represents the cluster number and k∈{1,2,…,K}, sample x belongs to sample set X;

[0026] Among them, the clustering loss function is:

[0027]

[0028] ψ is the clustering loss function value.

[0029] In step 2, the image information entropy is calculated as follows:

[0030]

[0031] Among them, H represents the image information entropy, a represents the gray value of the pixel, b represents the neighborhood gray mean, D represents the scale of the image, P a,b It indicates the proportion of pixels whose gray value is a and whose neighborhood gray mean is b.

[0032] In step 3, the rotation data enhancement method is:

[0033]

[0034] Among them, (R I ; R Q ) represents a dataset with a limited number of original signal label samples, (R I ′; R Q ′) represents the data set after rotation data enhancement, where the subscripts I and Q indicate the real part and the imaginary part respectively, θ represents the counterclockwise rotation angle, and the rotation angle is selected according to the expansion multiple N and is evenly divided into 360°.

[0035] In step 3, the cyclic time-shift enhancement method is:

[0036]

[0037] Among them, r IQ is an IQ format signal sample matrix of length L, where L represents the length of the window added when processing the intercepted signal, and is also the signal sample length, l represents the length of the cyclic left shift, and the number of bits of the cyclic time shift is evenly divided into L according to the expansion multiple N, s I () represents the real part of the signal sample s, s Q () represents the imaginary part of the signal sample s.

[0038] In step 4, the complex cross term is:

[0039] I n ′=I n w r-n -Q n w i-n

[0040] Q n ′=I n w i-n +Q n wr-n

[0041] Among them, I n ′, Q n ′ represents the output data of the nth layer, I n , Q n represents the input data of the nth layer, n∈{1,2,3,4,5} represents the number of network layers; w r-n represents the real network weight of the nth layer, w i-n Represents the imaginary network weight of the nth layer.

[0042] In step 5, the method for training CNRN includes:

[0043] S1, use the enhanced signal and the corresponding noisy signal to form a sample pair, where the enhanced signal is used as the label of the noisy signal and the noisy signal is used as the input of the CNRN;

[0044] S2, initialize the network parameters of CNRN, input the noisy signal into CNRN to obtain the predicted signal, calculate the mean square error loss function value between the predicted signal and the label, iterate S2 until the mean square error loss function value is minimum, stop training, and obtain the trained complex denoising network model F-CNRN of the small sample communication modulation signal; where the mean square error loss function is:

[0045]

[0046] In the formula, loss MSE is the mean square error loss function value, z(m) is the enhanced signal; z′(m) represents the predicted signal, M is the total number of signal sampling points, that is, the signal length, m is the sequence number of the signal sampling point and m∈{0,1,…,M-1}.

[0047] The present invention also provides a small sample communication modulation recognition preprocessing system based on a complex noise reduction network, comprising:

[0048] A partitioning module is used to extract noisy signals in the data set according to a preset ratio by using a stratified sampling method, and divide the data set into a training set and a test set;

[0049] The unsupervised classification module is used to convert the noisy signals in the training set into time-frequency graphs through Choi-Williams distribution; use the convolutional neural network CNN to extract the time-frequency graph features and use the K-means clustering algorithm to perform signal-to-noise ratio clustering on the time-frequency graph features to obtain two categories based on the signal-to-noise ratio, and calculate the image information entropy of each signal-to-noise ratio category; and identify high signal-to-noise ratio signals based on the image information entropy;

[0050] A data enhancement module, used to expand the number of high signal-to-noise ratio signals by a rotation data enhancement method and a cyclic time-shift enhancement method to obtain an enhanced signal;

[0051] The CNRN module is used to build a complex autoencoder denoising network model CNRN based on the autoencoder framework by using complex cross terms to assist the complex long short-term memory network CLSTM, and input the enhanced signal into the CNRN for automatic denoising to obtain the noisy signal;

[0052] The F-CNRN module is used to form a sample pair of the enhanced signal and the corresponding noisy signal, train the CNRN, and obtain the complex denoising network model F-CNRN of the trained small sample communication modulation signal; input the test set into the F-CNRN, and output the denoised communication modulation signal.

[0053] Beneficial effects of the present invention:

[0054] 1. The present invention solves the problem that it is difficult to obtain communication modulation signal samples in non-cooperative scenarios, the influence of noises of different intensities on signals is too different to achieve high-quality sample pair construction, and it is difficult to support deep learning network training;

[0055] 2. Cluster the time-frequency diagrams by taking advantage of the fact that they contain more information than signal sequences. Use image information entropy to classify the clustering results and identify high signal-to-noise ratio signals, provide high-quality training sample pairs, and improve the accuracy of subsequent signal modulation mode recognition.

[0056] 3. Through the rotation data enhancement method and the cyclic time-shift enhancement method, the data of high signal-to-noise ratio signals is expanded to provide a sufficient number of high-quality sample pairs for network training;

[0057] 4. Taking full account of the complex properties, CLSTM and complex cross terms are used to learn the time domain and complex domain features of the communication modulation signal, which can not only maintain the essential characteristics of the communication modulation signal but also effectively reduce the impact of noise in the spatial electromagnetic environment on the signal data, thereby realizing the extraction of the essential characteristics of the signal.

[0058] 5. The F-CNRN disclosed in the present invention reduces the impact of noise and enhances the robustness of communication modulation recognition under small sample conditions. It has theoretical value for complex electromagnetic signal processing methods and has positive reference significance for enhancing the robustness of communication modulation signal recognition methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is the time domain waveform of the BPSK original signal.

[0060] Figure 2 It is the power spectrum of the original BPSK signal.

[0061] Figure 3 It is the time-domain waveform diagram of the BPSK signal after being rotated by 180°.

[0062] Figure 4 It is the power spectrum diagram of the BPSK signal after being rotated by 180°.

[0063] Figure 5 It is the time-domain waveform diagram of the BPSK signal after being circularly shifted left by 64 sampling points.

[0064] Figure 6 It is the power spectrum diagram of the BPSK signal after being circularly shifted left by 64 sampling points.

[0065] Figure 7 It is the time-frequency diagram of the BPSK signal before noise reduction.

[0066] Figure 8 It is the time-frequency diagram of the BPSK signal after noise reduction without data augmentation.

[0067] Fig. 9 It is the time-frequency diagram of the BPSK signal after noise reduction with data augmentation.

[0068] Fig.10 It is the time-frequency diagram of the QAM64 signal before noise reduction.

[0069] Fig.11 It is the time-frequency diagram of the QAM64 signal after noise reduction without data augmentation.

[0070] Fig.12 It is the time-frequency diagram of the QAM64 signal after noise reduction with data augmentation.

[0071] Fig.13 It is the time-frequency diagram of the CPFSK signal before noise reduction.

[0072] Fig.14 It is the time-frequency diagram of the CPFSK signal after noise reduction without data augmentation.

[0073] Fig.15 It is the time-frequency diagram of the CPFSK signal after noise reduction with data augmentation.

[0074] Fig.16 It is the schematic diagram of the SNR comparison of the noise reduction indexes of the BPSK signal by different methods.

[0075] Fig.17 It is the schematic diagram of the MSE comparison of the noise reduction indexes of the BPSK signal by different methods.

[0076] Fig.18 It is the comparison of the recognition rates of different noise reduction methods under three recognition networks.

[0077] Fig.19 It is the schematic diagram of the recognition rate comparison of different noise reduction methods under CNN.

[0078] Fig. 20 This is a schematic diagram comparing the recognition rates of different denoising methods under LSTM.

[0079] Fig.21 This is a schematic diagram comparing the recognition rates of different denoising methods under CLDNN.

[0080] Fig. 22 This is the network structure diagram of CNRN. DETAILED DESCRIPTION

[0081] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0082] The embodiment of the present invention provides a small sample communication modulation recognition preprocessing method based on a complex denoising network, comprising:

[0083] Step 1: Use stratified sampling method to extract noisy signals in the data set according to a preset ratio and divide them into training set and test set;

[0084] Step 2: Convert the noisy signal in the training set into a time-frequency graph through the Choi-Williams distribution; Use the convolutional neural network CNN to extract the time-frequency graph features and use the K-means clustering algorithm to perform signal-to-noise ratio clustering on the time-frequency graph features to obtain two categories based on the signal-to-noise ratio, and calculate the image information entropy of each signal-to-noise ratio category respectively; According to the image information entropy, the high signal-to-noise ratio signal is identified; Step 2: In order to solve the problem that it is difficult to find the corresponding noisy signal for the intercepted low signal-to-noise ratio signal, which makes it difficult to obtain sample pairs and cannot meet the training conditions of the deep learning network, the signal-to-noise ratio unsupervised classification is performed to obtain the high signal-to-noise ratio signal;

[0085] Step 3: The number of high signal-to-noise ratio signals is expanded by the rotation data enhancement method and the cyclic time-shift enhancement method to obtain enhanced signals; Step 3 solves the problem of limited sample number of the end-to-end denoising network in non-cooperative scenarios;

[0086] Step 4: Based on the autoencoder framework, a complex autoencoder denoising network model CNRN is constructed by using a complex cross-term assisted complex long short-term memory network CLSTM. The enhanced signal is input into CNRN for automatic denoising to obtain a noisy signal. CNRN is as follows: Fig. 22 As shown in Table 6; Fig. 22 In, in 、s out They represent the noisy signal and the denoised signal respectively, and I and Q represent the real part and imaginary part of the signal circulating in the network respectively.

[0087] In order to solve the problem that the real network fails to fully extract the complex characteristics of the transmission signal in the complex electromagnetic space, the complex autoencoder denoising network CNRN built in step 4 can fully mine the amplitude and phase characteristics of the communication modulation signal. The network parameters of CNRN are shown in Table 1.

[0088] Table 1

[0089]

[0090]

[0091] Step 5. Combine the enhanced signal and the corresponding noisy signal into sample pairs, train CNRN, and obtain the trained few-shot communication modulation signal complex noise reduction network model (Few-shot Communication Modulation Signals Complex Noise Reduction Network, F-CNRN); input the test set into F-CNRN, and output the communication modulation signal after noise reduction. The performance of the technical solution disclosed in the present invention is analyzed according to the relevant evaluation indicators before and after noise reduction. Step 5 realizes the enhancement of the robustness of the modulation recognition algorithm with different signal-to-noise ratios under small sample conditions.

[0092] In step 1, the method of dividing the training set and the test set is:

[0093] The numpy.random.choice function in the computer python language is used to extract the noisy signals in the data set according to the preset ratio and divide them into training set and test set.

[0094] In step 2, Choi-Williams distribution is a type of Cohen-type time-frequency distribution, which has the characteristics of high time-frequency resolution and effective suppression of cross-term influence. Based on the characteristics that the time-frequency diagram contains more information than the digital sequence, the Choi-Williams distribution is used to perform time-frequency transformation to convert the noisy signal into a time-frequency diagram. Choi-Williams distribution is:

[0095]

[0096] Where P(t,f) represents the time-frequency distribution, y(·) represents the signal, * represents the conjugate, u represents the integral variable, τ represents the time shift, Φ(τ,ν) is the kernel function value, ν represents the frequency shift, j represents the imaginary unit, t represents the time, and f represents the frequency;

[0097] The calculation formula of the kernel function value Φ(τ,ν) is as follows:

[0098] Φ(τ,ν)=exp(-(2πντ) 2 / c);

[0099] Where c is the scale factor, and its value affects the resolution and cross-term suppression effect. Preferably, c is 3.6.

[0100] In step 2, the K-means clustering algorithm is a basic and commonly used clustering algorithm that can find a solution for dividing k clusters in an unsupervised iterative manner until the clustering loss function is minimized to obtain the division result. In step 2, the K-means clustering algorithm is:

[0101]

[0102] Among them, G k-means ((X,d),(C 1 ,…,C K )) is the clustering result, i.e., the signal-to-noise ratio category, and d represents the distance; the clustering loss function in the K-means clustering algorithm is defined as cluster C k From sample x to center point μ k The mean square distance of ; K represents the total number of clusters, k represents the cluster number and k∈{1,2,…,K}, sample x belongs to sample set X;

[0103] Among them, the clustering loss function is:

[0104]

[0105] ψ is the clustering loss function value.

[0106] In step 2, the image information entropy is calculated as follows:

[0107]

[0108] Among them, H represents the image information entropy, a represents the gray value of the pixel, b represents the neighborhood gray mean, D represents the scale of the image, P a,b It indicates the proportion of pixels whose gray value is a and whose neighborhood gray mean is b.

[0109] In step 3, in order to solve the problem that the number of high-quality and high-reliability label samples is insufficient due to non-cooperative scenarios, making it difficult to train the network of the denoising method based on deep learning, we propose to use the rotation data enhancement method and the cyclic time-shift enhancement method to expand the number of extracted high signal-to-noise ratio signals to obtain enhanced signals, and then use the enhanced signals to train and learn the network. In step 3, the rotation data enhancement method is:

[0110]

[0111] Among them, (R I ; R Q ) represents a dataset with a limited number of original signal label samples, (R I′; R Q ′) represents the data set after rotation data enhancement, where the subscripts I and Q indicate the real part and the imaginary part respectively, θ represents the counterclockwise rotation angle, and the rotation angle is selected according to the expansion multiple N and is evenly divided into 360°.

[0112] In step 3, the cyclic time-shift enhancement method is:

[0113]

[0114] Among them, r IQ is an IQ format signal sample matrix of length L, where L represents the length of the window added when processing the intercepted signal, and is also the signal sample length, l represents the length of the cyclic left shift, and the number of bits of the cyclic time shift is evenly divided into L according to the expansion multiple N, s I () represents the real part of the signal sample s, s Q () represents the imaginary part of the signal sample s.

[0115] In step 4, the complex cross term is:

[0116] I n ′=I n w r-n -Q n w i-n

[0117] Q n ′=I n w i-n +Q n w r-n

[0118] Among them, I n ′, Q n ′ represents the output data of the nth layer, I n , Q n represents the input data of the nth layer, n∈{1,2,3,4,5} represents the number of network layers; w r-n represents the real network weight of the nth layer, w i-n Represents the imaginary network weight of the nth layer.

[0119] In step 5, the method for training CNRN includes:

[0120] S1, use the enhanced signal and the corresponding noisy signal to form a sample pair, where the enhanced signal is used as the label of the noisy signal and the noisy signal is used as the input of the CNRN;

[0121] S2, initialize the network parameters of CNRN, input the noisy signal into CNRN to obtain the predicted signal, calculate the mean square error loss function value between the predicted signal and the label, iterate S2 until the mean square error loss function value is minimum, stop training, and obtain the trained complex denoising network model F-CNRN of the small sample communication modulation signal; where the mean square error loss function is:

[0122]

[0123] In the formula, loss MSE is the mean square error loss function value, z(m) is the enhanced signal; z′(m) represents the predicted signal, M is the total number of signal sampling points, that is, the signal length, m is the sequence number of the signal sampling point and m∈{0,1,…,M-1}.

[0124] The embodiment of the present invention further provides a small sample communication modulation recognition preprocessing system based on a complex noise reduction network, comprising:

[0125] A partitioning module is used to extract noisy signals in the data set according to a preset ratio by using a stratified sampling method, and divide the data set into a training set and a test set;

[0126] The unsupervised classification module is used to convert the noisy signals in the training set into time-frequency graphs through Choi-Williams distribution; use the convolutional neural network CNN to extract the time-frequency graph features and use the K-means clustering algorithm to perform signal-to-noise ratio clustering on the time-frequency graph features to obtain two categories based on the signal-to-noise ratio, and calculate the image information entropy of each signal-to-noise ratio category; and identify high signal-to-noise ratio signals based on the image information entropy;

[0127] A data enhancement module, used to expand the number of high signal-to-noise ratio signals by a rotation data enhancement method and a cyclic time-shift enhancement method to obtain an enhanced signal;

[0128] The CNRN module is used to build a complex autoencoder denoising network model CNRN based on the autoencoder framework by using complex cross terms to assist the complex long short-term memory network CLSTM, and input the enhanced signal into the CNRN for automatic denoising to obtain the noisy signal;

[0129] The F-CNRN module is used to form a sample pair of the enhanced signal and the corresponding noisy signal, train the CNRN, and obtain the complex denoising network model F-CNRN of the trained small sample communication modulation signal; input the test set into the F-CNRN, and output the denoised communication modulation signal.

[0130] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0131] In order to facilitate those skilled in the art to understand the present invention, specific examples are now provided to illustrate the technical solutions of the present invention in detail.

[0132] Take the communication modulation signal data set shown in Table 2 as an example:

[0133] Table 2

[0134]

[0135] Example 1:

[0136] (1) Use internal evaluation methods to evaluate clustering effects based on the internal distribution characteristics of clustering result samples:

[0137] 1) Calinski-Harabasz index (CH) represents the ratio of inter-cluster distance to intra-cluster distance. The smaller the intra-cluster distance and the larger the inter-cluster distance, the larger the CH. The larger the CH, the better the clustering effect.

[0138]

[0139] Where K represents the total number of clusters, k represents the cluster number and k∈{1,2,…,K}, and x represents the kth cluster C k The sample in μ k represents the cluster center corresponding to the kth cluster, sum k represents the number of samples in the kth cluster, sum represents the number of all samples, and μ represents the global cluster center.

[0140] 2) Silhouette coefficient (SC) represents the ratio of the average distance from a sample point in a cluster to other sample points in the cluster to the average distance to the sample point in the nearest cluster structure. The larger the SC, the better the clustering effect.

[0141]

[0142] Where sum represents the number of sample points, nem represents the sequence number of the sample points and nem∈{1,2,…,sum}, A′(nem) represents the mean distance from the nemth sample to other sample points in the cluster, and B′(nem) represents the mean distance from the nemth sample to the sample points in the nearest cluster.

[0143] 3) The Davies-Bouldin index (DB) represents the average similarity between the two most similar clusters among all clusters, that is, the average ratio of the intra-cluster diameter to the inter-cluster distance. The smaller the DB, the better the clustering effect.

[0144]

[0145] Where K represents the number of clusters, Respectively represent the kth 1 clusters and the kth 2 The average distance from the sample points in a cluster to the cluster center is the cluster diameter. Indicates the kth 1 clusters and the kth 2 The distance between cluster centers is the distance between clusters.

[0146] (2) By calculating the three indicators of CH, SC and DB, the different clustering effects of the time-frequency diagram and the original signal sequence under the K-means clustering algorithm are compared, and the clustering results are visualized to verify that the time-frequency diagram provides more feature information and is more accurate for signal-to-noise ratio clustering. Then, the high and low signal-to-noise ratio classification is quickly achieved by calculating the information entropy of the two categories of images after clustering.

[0147] 1) Verification of clustering effect of time-frequency graph: Table 3 is the comparison of clustering effect of signal-to-noise ratio of modulation signal expression form;

[0148] Table 3

[0149]

[0150] From Table 3, we can see that the CH value based on the clustering of the time-frequency graph is significantly higher than the CH value based on the signal sequence clustering. The maximum CH value based on the clustering of the time-frequency graph is 345.05, and the maximum CH value based on the signal sequence is 120.37. The CH ratios are all greater than 2, and the differences are all greater than 100. The CH ratio of the 8PSK signal time-frequency graph and the signal sequence reaches a maximum of 4.2, and the difference reaches a maximum of 228.79; the SC value based on the clustering of the time-frequency graph is significantly higher than the SC value based on the signal sequence clustering, and the difference is greater than 0.1; the DB value based on the clustering of the time-frequency graph is significantly lower than the DB value based on the signal sequence clustering. The DB indexes of the time-frequency graph and sequence clustering of the GFSK signal are both the smallest, which are 2.52 and 1.66, respectively, with a difference of 0.86. The larger the CH and SC, the smaller the DB, and the better the clustering effect. From the comparison of the three indicators, it can be seen that the signal-to-noise ratio clustering method based on the time-frequency graph is better than the signal-to-noise ratio clustering method based on the signal sequence.

[0151] 2) Image information entropy classification effect verification:

[0152] The image information entropy is calculated to classify the signal-to-noise ratio clustering results into high and low signal-to-noise ratios. The image information entropy can be used to indicate the degree of disorder of the image. The average image information entropy of the clustering results is calculated. The larger the information entropy, the more disordered the class is, indicating that the class is more affected by noise. Therefore, the class with large information entropy is the low signal-to-noise ratio class, and the class with small information entropy is the high signal-to-noise ratio class. It can also be seen from the time-frequency diagram example that image information entropy is effective in signal-to-noise ratio classification. The classification results of the signal-to-noise ratio of eight modulation signals based on image information entropy are shown in Table 4.

[0153] Table 4

[0154]

[0155]

[0156] Example 2

[0157] (1) Verification of the effectiveness of enhanced signals:

[0158] According to the relevant characteristics of the communication signal in time domain, frequency domain, and time-frequency domain, the effectiveness of the enhanced signal obtained by data enhancement is verified by comparing the time domain waveform and power spectrum of the signal sample after data enhancement with the original signal sample. The comparison effect is as follows: Figures 1 to 6 shown.

[0159] The experiment plots the spatial orthogonal signals of complex communication modulation signals on the same plane for easy observation and judgment. Figure 1 It is the time domain waveform of the original signal. The blue and yellow lines in the figure represent the sum components of the complex signal respectively. Figure 3 It is the time domain waveform of the sample generated after the original signal is rotated 180° counterclockwise. Figure 1 and Figure 3 It can be seen that after the signal is rotated, the highest point of the sum changes from (38, 0.002) and (63, 0.011) to the lowest point (38, -0.002) and (63, -0.011), and the ordinate value of the feature point is flipped in the plane diagram. Figure 5 It is the time domain waveform of the sample generated after the original signal is cyclically shifted to the left by 64 sampling points. Figure 1 and Figure 5 It can be seen from the figure that the highest points of the signal after time shift change from (38, 0.002) and (63, 0.011) to (102, 0.002) and (127, 0.011), and the horizontal coordinate values ​​of the feature points are all cycled around 64 bits. Figure 2 , Figure 4 and Figure 6 It can be seen that the characteristic distribution of the generated samples obtained by rotation and cyclic time shifting of the signal samples changes, but the essential characteristics of the signal modulation type are still consistent with the original signal, and can be used as expanded samples to train the DL-based denoising network.

[0160] (2) Verification of the effectiveness of data enhancement:

[0161] The data enhancement ablation experiment further verifies in the signal denoising experiment that the data enhancement method based on rotation and time shift can solve the small sample problem in non-cooperative scenarios. The denoising effect of the communication modulation signal is visualized by comparing the time-frequency diagram before and after denoising. The effectiveness of data enhancement on signal denoising is verified by visualizing the time-frequency diagram after denoising with and without data enhancement. Three modulation signals, BPSK, QAM64, and CPFSK, are selected respectively. The time-frequency diagrams before and after denoising with and without data enhancement are compared. Figures 7 to 15 shown.

[0162] Figures 7 to 15 The time-frequency diagram of the α-D signal directly reflects the effect of data enhancement on the noise reduction effect of different modulation signals. Figure 7 , Fig. 9 )and( Fig.10 , Fig.12 )and( Fig.13 , Fig.15 )These three pairs of time-frequency diagrams before and after denoising with data enhancement show that the signal frequency and envelope are basically consistent. While removing the noise, the characteristic information of the signal is not affected, and some details covered by the noise can even be restored. However, as ( Figure 7 , Figure 8 )、( Fig.10 , Fig.11 )and( Fig.13 , Fig.14 ) As shown in the three pairs of time-frequency diagrams, when data enhancement processing is not performed on the samples, although the messy frequency points representing noise in the figure are reduced, the regular frequency point positions and envelope shapes representing the signal are changed, which violates the principle of not affecting the signal while removing noise. At the same time, because the signal has undergone qualitative changes, the modulation recognition results cannot be used to carry out ablation experiments in the future. The denoising results show that directly using small samples for denoising network training results in fewer features learned by the network, that is, small samples cannot meet the requirements of deep learning network training. Therefore, in the denoising of communication modulation signals in complex electromagnetic environments in non-cooperative scenarios, data enhancement methods based on rotation and cyclic time shift are of great significance.

[0163] Example 3

[0164] (1) Noise reduction effect verification:

[0165] The signal-to-noise ratio (SNR) is an important indicator for measuring signal quality. The performance of the denoising network is evaluated by comparing the signal-to-noise ratio of the signal before and after denoising. The higher the SNR, the better the signal quality and the better the denoising effect. At the same time, the mean square error (MSE) can be used to estimate the difference between the denoised signal sig′(m) and the clean signal sig(m) (i.e., the signal without noise interference). The smaller the MSE, the smaller the difference between the denoised signal and the clean signal, and the better the denoising performance. SNR and MSE are defined as direct evaluation indicators of the signal denoising effect as follows:

[0166]

[0167] Signal denoising is an effective means to improve the communication modulation recognition performance. The modulation recognition accuracy of the signal before and after denoising can be used to indirectly evaluate the denoising effect of the communication modulation signal. The higher the signal recognition accuracy after denoising, the better the denoising performance. The indirect evaluation index Acc is defined as follows:

[0168]

[0169] Among them, TP represents a correctly predicted positive example; TN represents a correctly predicted negative example; FP represents an incorrectly predicted positive example; and FN represents an incorrectly predicted negative example.

[0170] The proposed denoising method F-CNRN, wavelet denoising (wavelet), smoothing denoising (smooth), convolutional denoising autoencoder (DAE_CNN) and LSTM denoising autoencoder (DAE_LSTM) are directly compared with five denoising methods, such as Fig.16 and Fig.17 As shown in Figure 2, two of the denoising autoencoders based on machine learning are real network structures. Fig.16 and Fig.17 In the comparison of three denoising methods, F-CNRN, convolutional denoising autoencoder and LSTM denoising autoencoder, it can be seen from the two indicators of SNR and MSE that the denoising method F-CNRN proposed in the present invention is significantly better than the other two real number denoising methods, which further illustrates the effectiveness of the complex number denoising network proposed in the present invention. Fig.16 It can be clearly seen that when the input SNR is less than -6dB, the output SNR of F-CNRN denoising is significantly higher than that of wavelet denoising, smoothing denoising, convolution denoising autoencoder and LSTM denoising autoencoder, and -20dB is greatly improved to 0dB. However, when the input SNR is greater than -6dB, the output SNR of F-CNRN denoising is lower than that of wavelet denoising and smoothing denoising. Even when the input SNR is greater than 13dB, the output SNR is lower than the original SNR. This is because the DL-based denoising method is a process of calculating the optimal value, and the denoising will be concentrated near one value, reflecting the state where the difference between the maximum and minimum values ​​is small. Fig.17 It can also be clearly seen that the MSE of the signal after F-CNRN denoising is smaller than the MSE of the other four methods, indicating that the denoising effect of F-CNRN is significantly better than the other four methods. The comparison of the two indicators of SNR and MSE shows that the denoising method proposed in the present invention is better than the four denoising methods of wavelet denoising, smoothing denoising, convolution denoising autoencoder and LSTM denoising autoencoder.

[0171] (2) Verification of recognition effect:

[0172] Modulation recognition based on deep learning is mostly improved on the basis of CNN and LSTM. Therefore, CNN, LSTM, and convolutional long short-term memory fully connected deep neural network (CLDNN) are used to perform modulation recognition on the denoised signal to indirectly verify the effectiveness of the denoising network F-CNRN. The specific recognition network structure parameters are shown in Table 5.

[0173] Table 5

[0174]

[0175]

[0176] Three typical modulation recognition networks are used to compare the recognition effects of F-CNRN before and after denoising, and the recognition accuracy is shown in Table 6. By comparing the modulation recognition Acc before and after signal denoising, it is proved that the denoising network F-CNRN is effective in removing the noise of non-cooperative communication modulation signals.

[0177] Table 6

[0178]

[0179] ①Comparative analysis of -20~-2dB signal before and after noise reduction:

[0180] pass Figures 18 to 21 It can be seen from the recognition rate curve that when the signal-to-noise ratio is -20 to -2dB, the recognition accuracy of the three modulation recognition networks is significantly improved after F-CNRN denoising. According to Table 6, the average recognition accuracy of CNN, LSTM and CLDNN before and after denoising at -20 to -2dB is compared. It can be seen that the recognition accuracy of CNN before denoising is 28.10%, which is increased by 33.68% after denoising to 61.78%; the recognition accuracy of LSTM before denoising is 34.94%, and the accuracy is increased by 37.41% after denoising to 72.35%; the recognition accuracy of CLDNN before denoising is 33.80%, and it reaches 68.53% after denoising, and the accuracy is increased by 34.73%. Because LSTM has a stronger ability to pay attention to the characteristics of time series signals, the recognition accuracy of LSTM after denoising at low signal-to-noise ratio is more significantly improved than that of CNN and CLDNN. Verify the effectiveness of the method proposed in the present invention for denoising low signal-to-noise ratio signals.

[0181] ②Comparative analysis of 0-18dB signal noise reduction before and after:

[0182] according to Figures 18 to 21 The recognition rate curves of different recognition networks before and after F-CNRN denoising at 0-18dB are analyzed. Combined with the recognition accuracy at 0-18dB in Table 6, it can be seen that the recognition effects of CNN and LSTM are significantly enhanced after denoising, among which the recognition rate of CNN reaches 90.75%, which is 15.69% higher than before denoising, and the recognition accuracy of LSTM increases by 19.96% compared with before denoising, reaching 97.91%; the recognition rate of CLDNN network increases by 4.29% compared with before denoising, and the recognition rate reaches 97.45%. It shows that the method proposed in the present invention has a better denoising effect on high signal-to-noise ratio signals.

[0183] ③-20~18dB signal noise reduction before and after comparison analysis:

[0184] By comparing the overall recognition rates of CNN, LSTM and CLDNN before and after noise reduction in Table 6, it can be seen that the signal recognition accuracy of -20 to 18dB before CNN noise reduction is 51.59%, and after noise reduction, it reaches 76.26%, an increase of 24.67%; the overall recognition accuracy of LSTM after noise reduction reaches 85.13%, an increase of 21.20%; the overall recognition accuracy of CLDNN after noise reduction is increased by 19.51%, reaching 82.99%. The recognition rates of the three recognition networks are significantly improved after the modulation signal F-CNRN is denoised, which indirectly proves that the noise reduction method proposed in the present invention can effectively solve the influence of noise in the recognition process of communication modulation signals.

[0185] Three typical modulation recognition networks are used to compare the recognition effects of five methods before and after denoising: F-CNRN, wavelet denoising (wavelet), smooth denoising (smooth), convolutional denoising autoencoder (DAE_CNN) and LSTM denoising autoencoder (DAE_LSTM). The recognition rate curves of different denoising methods under three recognition networks are shown in Figure 2. Figures 18 to 21 As shown. Fig.18 It can be seen that the recognition accuracy of signals with a signal-to-noise ratio of -20 to -2dB after F-CNRN denoising is significantly higher than that of wavelet denoising, smoothing denoising and convolutional denoising autoencoder denoising, while it is not much different from the recognition accuracy of LSTM denoising autoencoder denoising, or even slightly lower. Figures 19 to 21 It can be seen that the recognition accuracy of signals with a signal-to-noise ratio of 0 to 18dB after F-CNRN denoising is significantly higher than that of the other four denoising methods. In particular, the part where F-CNRN is higher than the LSTM denoising autoencoder is significantly greater than the disadvantage of -20 to -2dB. Overall, F-CNRN is better than the LSTM denoising autoencoder. Figures 19 to 21 It can be clearly seen that the F-CNRN represented by the red curve has an average recognition accuracy of -20 to 18 dB under the three recognition networks of CNN, LSTM, and CLDNN, which is significantly better than the other four noise reduction methods. This shows that the method proposed in the present invention can effectively maintain signal quality while removing noise without affecting the modulation recognition accuracy.

[0186] Beneficial effects of the embodiments of the present invention:

[0187] 1. The present invention solves the problem that it is difficult to obtain communication modulation signal samples in non-cooperative scenarios, the influence of noises of different intensities on signals is too different to achieve high-quality sample pair construction, and it is difficult to support deep learning network training;

[0188] 2. Cluster the time-frequency diagrams by taking advantage of the fact that they contain more information than signal sequences. Use image information entropy to classify the clustering results and identify high signal-to-noise ratio signals, provide high-quality training sample pairs, and improve the accuracy of subsequent signal modulation mode recognition.

[0189] 3. Through the rotation data enhancement method and the cyclic time-shift enhancement method, the data of high signal-to-noise ratio signals is expanded to provide a sufficient number of high-quality sample pairs for network training;

[0190] 4. Taking full account of the complex properties, CLSTM and complex cross terms are used to learn the time domain and complex domain features of the communication modulation signal, which can not only maintain the essential characteristics of the communication modulation signal but also effectively reduce the impact of noise in the spatial electromagnetic environment on the signal data, thereby realizing the extraction of the essential characteristics of the signal.

[0191] 5. The F-CNRN disclosed in the present invention reduces the impact of noise and enhances the robustness of communication modulation recognition under small sample conditions. It has theoretical value for complex electromagnetic signal processing methods and has positive reference significance for enhancing the robustness of communication modulation signal recognition methods.

Claims

1. A small sample communication modulation recognition preprocessing method based on complex denoising network, characterized in that: include: Step 1: Use stratified sampling method to extract noisy signals in the data set according to a preset ratio and divide them into training set and test set; Step 2: Convert the noisy signal in the training set into a time-frequency graph through Choi-Williams distribution; use the convolutional neural network CNN to extract the time-frequency graph features and use the K-means clustering algorithm to perform signal-to-noise ratio clustering on the time-frequency graph features to obtain two categories based on the signal-to-noise ratio, and calculate the image information entropy of each signal-to-noise ratio category; identify the high signal-to-noise ratio signal based on the image information entropy; Step 3: The number of high signal-to-noise ratio signals is expanded by a rotation data enhancement method and a cyclic time-shift enhancement method to obtain an enhanced signal; Step 4: Based on the autoencoder framework, a complex autoencoder denoising network model CNRN is constructed by using complex cross terms to assist the complex long short-term memory network CLSTM, and the enhanced signal is input into the CNRN for automatic noise addition to obtain the noisy signal; Step 5. Combine the enhanced signal and the corresponding noisy signal into sample pairs, train CNRN, and obtain the complex denoising network model F-CNRN of the trained small sample communication modulation signal; input the test set into F-CNRN, and output the denoised communication modulation signal.

2. A small sample communication modulation recognition preprocessing method based on a complex denoising network as claimed in claim 1, characterized in that: In step 1, the method of dividing the training set and the test set is: The numpy.random.choice function in the computer python language is used to extract the noisy signals in the data set according to the preset ratio and divide them into training set and test set.

3. A small sample communication modulation recognition preprocessing method based on a complex denoising network as claimed in claim 1, characterized in that: In step 2, the Choi-Williams distribution is: Among them, P(t,f) represents the time-frequency distribution, y( · ) represents the signal, * represents the conjugate, u represents the integral variable, τ represents the time shift, Φ(τ,ν) is the kernel function value, ν represents the frequency shift, j represents the imaginary unit, t represents the time, and f represents the frequency; where, The calculation formula of the kernel function value Φ(τ,ν) is as follows: Φ(τ,ν)=exp(-(2πντ) 2 / c); Where c is the scale factor.

4. A small sample communication modulation recognition preprocessing method based on a complex denoising network as claimed in claim 1, characterized in that: In step 2, the K-means clustering algorithm is: Among them, G k-means ((X,d),(C1, … ,C K )) is the clustering result, i.e., the signal-to-noise ratio category, and d represents the distance; the clustering loss function in the K-means clustering algorithm is defined as cluster C k From sample x to center point μ k The mean square distance of K represents the total number of clusters, k represents the cluster number and k∈{1,2, … ,K}, sample x belongs to sample set X; Among them, the clustering loss function is: ψ is the clustering loss function value.

5. The method for preprocessing small sample communication modulation recognition based on complex denoising network as claimed in claim 1, characterized in that: In step 2, the image information entropy is calculated as follows: Among them, H represents the image information entropy, a represents the gray value of the pixel, b represents the neighborhood gray mean, D represents the scale of the image, P a,b It indicates the proportion of pixels whose gray value is a and whose neighborhood gray mean is b.

6. A small sample communication modulation recognition preprocessing method based on a complex denoising network as claimed in claim 1, characterized in that: In step 3, the rotation data enhancement method is: Among them, (R I ; R Q ) represents a dataset with a limited number of original signal label samples, (R I ′; R Q ′) represents the data set after rotation data enhancement, where the subscripts I and Q indicate the real part and the imaginary part respectively, θ represents the counterclockwise rotation angle, and the rotation angle is selected according to the expansion multiple N and is evenly divided into 360°.

7. A small sample communication modulation recognition preprocessing method based on a complex denoising network as claimed in claim 1, characterized in that: In step 3, the cyclic time-shift enhancement method is: Among them, r IQ It is an IQ format signal sample matrix with a length of L. L represents the length of the window added when processing the intercepted signal, which is also the signal sample length. l represents the length of the cyclic left shift. The number of bits of the cyclic time shift is evenly divided into L according to the expansion multiple N. s I () represents the real part of the signal sample s, s Q () represents the imaginary part of the signal sample s.

8. The method for preprocessing small sample communication modulation recognition based on complex denoising network as claimed in claim 1, characterized in that: In step 4, the complex cross term is: I n ′=I n w r-n -Q n w i-n Q n ′=I n w i-n +Q n w r-n Among them, I n ′, Q n ′ represents the output data of the nth layer, I n , Q n represents the input data of the nth layer, n∈{1,2,3,4,5} represents the number of network layers; w r-n represents the real network weight of the nth layer, w i-n Represents the imaginary network weight of the nth layer.

9. A small sample communication modulation recognition preprocessing method based on a complex denoising network as claimed in claim 1, characterized in that: In step 5, the method for training CNRN includes: S1, use the enhanced signal and the corresponding noisy signal to form a sample pair, where the enhanced signal is used as the label of the noisy signal and the noisy signal is used as the input of the CNRN; S2, initialize the network parameters of CNRN, input the noisy signal into CNRN to obtain the predicted signal, calculate the mean square error loss function value between the predicted signal and the label, iterate S2 until the mean square error loss function value is minimum, stop training, and obtain the trained complex denoising network model F-CNRN of the small sample communication modulation signal; where the mean square error loss function is: In the formula, loss MSE is the mean square error loss function value, z(m) is the enhanced signal; z′(m) represents the predicted signal, M is the total number of signal sampling points, i.e., the signal length, m is the sequence number of the signal sampling point and m∈{0,1, … ,M-1}.

10. A small sample communication modulation recognition preprocessing system based on complex denoising network, characterized in that: include: A partitioning module is used to extract noisy signals in the data set according to a preset ratio by using a stratified sampling method, and divide the data set into a training set and a test set; The unsupervised classification module is used to convert the noisy signals in the training set into time-frequency graphs through Choi-Williams distribution; use the convolutional neural network CNN to extract the time-frequency graph features and use the K-means clustering algorithm to perform signal-to-noise ratio clustering on the time-frequency graph features to obtain two categories based on the signal-to-noise ratio, and calculate the image information entropy of each signal-to-noise ratio category; and identify high signal-to-noise ratio signals based on the image information entropy; A data enhancement module, used to expand the number of high signal-to-noise ratio signals by a rotation data enhancement method and a cyclic time-shift enhancement method to obtain an enhanced signal; The CNRN module is used to build a complex autoencoder denoising network model CNRN based on the autoencoder framework by using complex cross terms to assist the complex long short-term memory network CLSTM, and input the enhanced signal into the CNRN for automatic denoising to obtain the noisy signal; The F-CNRN module is used to form a sample pair of the enhanced signal and the corresponding noisy signal, train the CNRN, and obtain the complex denoising network model F-CNRN of the trained small sample communication modulation signal; input the test set into the F-CNRN, and output the denoised communication modulation signal.

Images