A deep learning modulation recognition method based on cyclic two-dimensional spectrogram deblurring

Abstract

The application relates to a deep learning modulation recognition method based on a cyclic two-dimensional spectrum deblurring method, which comprises the following steps: preparing a template of a cyclic two-dimensional spectrum of an easily-confused signal, and saving the obtained template to a template library; receiving a to-be-recognized signal, and converting the to-be-recognized signal into a cyclic two-dimensional spectrum; matching the cyclic two-dimensional spectrum of the to-be-recognized signal with the templates in the template library; if the matching is successful, using a first recognition model to recognize the cyclic two-dimensional spectrum of the to-be-recognized signal; and if the matching fails, using a second recognition model to recognize the cyclic two-dimensional spectrum of the to-be-recognized signal. The application solves the blurring problem of a cyclic spectrum in modulation recognition.

Description

Technical Field

[0001] This invention relates to the field of communication and information technology, and in particular to a deep learning modulation recognition method based on cyclic two-dimensional spectrogram deblurring. Background Technology

[0002] With the development of information technology, various communication methods and technologies are constantly being updated. In order to better utilize channel resources and meet different user needs, communication signals adopt different modulation methods. Modulation identification is a technology between signal detection and signal demodulation. Its main task is to identify the modulation method used in the intercepted signal.

[0003] Signal modulation recognition is essentially a pattern recognition problem, mainly comprising four parts: information acquisition, preprocessing, feature extraction, and classifier design. The input signal can be an intermediate frequency (IF) signal, which, after carrier parameter estimation, undergoes orthogonal downconversion to obtain a complex baseband signal, or a symbol-synchronized complex signal after synchronous sampling. Preprocessing mainly includes signal processing techniques such as filtering and amplification. Feature extraction can be categorized into methods such as spectral features, statistical measures of instantaneous features (instantaneous amplitude, instantaneous phase, instantaneous frequency), instantaneous feature histograms, and higher-order cumulants. Classifiers include maximum likelihood classifiers, threshold-based tree classifiers, and neural network classifiers. However, these feature extraction methods have some disadvantages, such as the complexity of manually designed features and decreased recognition rate performance at low signal-to-noise ratios, making them difficult to adapt to complex scenarios and requirements.

[0004] With the development of deep learning, leveraging the powerful feature extraction capabilities of neural networks, an increasing number of modulation recognition tasks are adopting end-to-end deep learning network architectures. Commonly used networks in modulation recognition tasks are mainly CNN and LSTM, as well as cascaded combinations of CNN and LSTM. Based on various domestic and international literature, deep learning-based signal modulation recognition methods outperform traditional methods based on manually designed features and classifiers. They can achieve recognition rates exceeding 90% at signal-to-noise ratios (SNR) above 10dB and 80% at SNRs above 0dB, but generally less than 50% at SNRs below 0dB. To address the low recognition rate at low SNRs, some noise-resistant signal processing techniques, such as cyclic spectrum analysis, have been introduced. However, the cyclic spectra of some high-order modulation signals are similar, leading to a decrease in recognition rate at high SNRs due to confusion issues. Summary of the Invention

[0005] This invention provides a deep learning modulation recognition method based on deblurring of cyclic two-dimensional spectra, which solves the ambiguity problem of cyclic spectra in modulation recognition.

[0006] The technical solution adopted by this invention to solve its technical problem is: to provide a deep learning modulation recognition method based on cyclic two-dimensional spectrogram deblurring, comprising the following steps:

[0007] Create templates for cyclic two-dimensional spectra of easily confused signals and save the obtained templates to a template library;

[0008] Receive the signal to be identified and convert it into a cyclic two-dimensional spectrum;

[0009] The cyclic two-dimensional spectrum of the signal to be identified is matched with templates in the template library;

[0010] If a match is successful, the first recognition model is used to identify the cyclic two-dimensional spectrum of the signal to be identified, wherein the first recognition model is trained using sample data of easily confused signals;

[0011] If the matching fails, the second recognition model is used to identify the cyclic two-dimensional spectrum of the signal to be identified, wherein the second recognition model is trained using sample data of signals that are not easily confused.

[0012] The template for creating a cyclic two-dimensional spectrum of easily confused signals specifically includes:

[0013] Obtain multiple cyclic two-dimensional spectra of each modulation style signal at a preset signal-to-noise ratio;

[0014] A template is obtained by performing a pixel-weighted average of multiple cyclic two-dimensional spectra of the same modulation pattern signal at a preset signal-to-noise ratio.

[0015] The process of converting the signal to be identified into a cyclic two-dimensional spectrum specifically includes:

[0016] The signal to be identified is processed to obtain a three-dimensional cyclic spectral density map;

[0017] The three-dimensional cyclic spectral density map is projected into a two-dimensional image by top-down projection to obtain a cyclic two-dimensional spectrum.

[0018] The process of processing the signal to be identified to obtain a three-dimensional cyclic spectral density map specifically includes:

[0019] The autocorrelation function of the signal to be identified is expanded into a Fourier series to obtain the cyclic autocorrelation.

[0020] A three-dimensional cyclic spectral density map is obtained by performing a Fourier transform on the cyclic autocorrelation.

[0021] The step of matching the cyclic two-dimensional spectrum of the signal to be identified with the templates in the template library specifically involves: calculating the root mean square error of the pixels of the cyclic two-dimensional spectrum of the signal to be identified with each template in the template library, and determining whether the cyclic two-dimensional spectrum of the signal to be identified matches the target in the template library by using a set threshold.

[0022] The first recognition model is built based on the MCLDNN network.

[0023] The second recognition model is built on a CNN neural network.

[0024] Beneficial effects

[0025] By adopting the above-mentioned technical solution, this invention has the following advantages and positive effects compared with the prior art: This invention pre-creates the cyclic spectra of these easily confused signals as templates, and then when the signal is input, its cyclic spectra are generated and matched with the template library. If a match is found, it indicates that it belongs to an easily confused category, and it is separated and processed through a dedicated network; otherwise, the cyclic spectra are classified through a traditional network. The above method solves the problem of reduced recognition performance caused by similar cyclic spectra of some high-order modulation signals, and improves the overall recognition performance of the network. Attached Figure Description

[0026] Figure 1 This is a flowchart of a deep learning modulation recognition method based on cyclic two-dimensional spectrogram deblurring according to an embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of the deep learning modulation recognition method based on cyclic two-dimensional spectrogram deblurring according to an embodiment of the present invention.

[0028] Figure 3 This is a template diagram of the cyclic two-dimensional spectra of four easily confused signals in an embodiment of the present invention;

[0029] Figure 4 This is a three-dimensional cyclic spectral density map of the DSSS signal in an embodiment of the present invention;

[0030] Figure 5 This is a two-dimensional projection of the three-dimensional cyclic spectral density of the DSSS signal in an embodiment of the present invention;

[0031] Figure 6 This is a similarity diagram between 13 modulation pattern signals and four templates in the template library;

[0032] Figure 7 This is a schematic diagram of the MCLDNN network structure;

[0033] Figure 8 This is a schematic diagram of the CNN neural network structure;

[0034] Figure 9 This is a comparison chart of the recognition rates of different algorithms;

[0035] Figure 10 This is a confusion matrix diagram with a range of -20 to -18 dB. Detailed Implementation

[0036] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0037] This invention relates to a deep learning modulation recognition method based on deblurring of cyclic two-dimensional spectra. This method pre-creates templates of cyclic spectra for easily confused signals. When a signal is input, its cyclic spectra are generated and matched against templates in a template library. If a match is found, it indicates that the signal belongs to an easily confused category and is separated for processing through a dedicated network; otherwise, the cyclic spectra are classified using a traditional network. For example... Figure 1 and Figure 2 As shown, it includes the following steps:

[0038] Step 1: Create a template for the cyclic two-dimensional spectrum of easily confused signals and save the obtained template to the template library.

[0039] In this embodiment, the cyclic two-dimensional spectra of easily confused signals include four types: QPSK cyclic two-dimensional grayscale spectra, 8PSK cyclic two-dimensional grayscale spectra, 16QAM cyclic two-dimensional grayscale spectra, and 64QAM cyclic two-dimensional grayscale spectra. The specific steps in creating the template for the cyclic two-dimensional spectra of easily confused signals include:

[0040] Obtain multiple cyclic two-dimensional spectra of each modulation style signal at a preset signal-to-noise ratio (SNR), specifically, obtain 100 cyclic two-dimensional spectra of QPSK modulation style signal at an SNR of (0:2:18dB), 100 cyclic two-dimensional spectra of 8PSK modulation style signal at an SNR of (0:2:18dB), 100 cyclic two-dimensional spectra of 16QAM modulation style signal at an SNR of (0:2:18dB), and 100 cyclic two-dimensional spectra of 64QAM modulation style signal at an SNR of (0:2:18dB).

[0041] A template is obtained by pixel-weighted averaging of multiple cyclic two-dimensional spectra of the same modulation pattern signal at a preset signal-to-noise ratio. Specifically, a template for the cyclic two-dimensional spectra of the QPSK modulation pattern signal is obtained by pixel-weighted averaging of 100 cyclic two-dimensional spectra of the QPSK modulation pattern signal at a signal-to-noise ratio of (0:2:18dB). Similarly, a template for the cyclic two-dimensional spectra of the 8PSK modulation pattern signal is obtained by pixel-weighted averaging of 100 cyclic two-dimensional spectra of the 8PSK modulation pattern signal at a signal-to-noise ratio of (0:2:18dB). Templates for the spectral patterns: A pixel-weighted average of 100 cyclic two-dimensional spectra of the 16QAM modulation pattern signal at a signal-to-noise ratio (0:2:18dB) is performed to obtain the template for the cyclic two-dimensional spectra of the 16QAM modulation pattern signal; a pixel-weighted average of 100 cyclic two-dimensional spectra of the 64QAM modulation pattern signal at a signal-to-noise ratio (0:2:18dB) is performed to obtain the template for the cyclic two-dimensional spectra of the 64QAM modulation pattern signal. Thus, templates for the cyclic two-dimensional spectra of the four modulation pattern signals can be obtained (see...). Figure 3 ).

[0042] Step 2: Receive the signal to be identified and convert it into a cyclic two-dimensional spectrum.

[0043] The signal to be identified in this step can be an intermediate frequency signal received by an antenna and down-converted to a fixed frequency by a radio frequency front-end. Converting the signal to be identified into a cyclic two-dimensional spectrum specifically includes: processing the signal to be identified to obtain a three-dimensional cyclic spectral density map; and projecting the three-dimensional cyclic spectral density map into a two-dimensional image using top-view projection to obtain the cyclic two-dimensional spectrum.

[0044] Due to the cyclic stationary nature of communication signals, the autocorrelation function of the signal to be identified is periodic. Since a periodic function can be expanded into a Fourier series, we can obtain... in This indicates the instantaneous autocorrelation R x The magnitude of (t,τ) at frequency α, where α is the frequency corresponding to the instantaneous autocorrelation, is the result of accumulation over instantaneous moments, and is also called the cycle frequency. Defined as cyclic autocorrelation, its cyclic spectral density can be obtained by performing a Fourier transform on it. in, It is the cyclic density spectrum, which is a two-frequency plane function of frequency f and cyclic frequency α. The non-periodicity of the noise signal makes the spectral correlation highly adaptable to noise, and the spectral correlation diagram is very stable at low signal-to-noise ratios.

[0045] Taking a direct sequence spread spectrum (DSSS) signal as an example, its cyclic spectral density like Figure 4 As shown, its cyclic density 2D projection is as follows: Figure 5As shown in the figure. This embodiment also converts the obtained 2D projection image into a single-channel grayscale image, which reduces the amount of data storage.

[0046] Step 3: Match the cyclic two-dimensional spectrum of the signal to be identified with templates in the template library. The specific matching process is as follows: Calculate the mean square error of the pixels of the cyclic two-dimensional spectrum of the signal to be identified with those of each template in the template library, and determine whether the cyclic two-dimensional spectrum of the signal to be identified matches the target in the template library by using a set threshold. In other words, this embodiment uses the mean square error of the pixels of the cyclic two-dimensional spectrum of the signal to be identified with those of each template in the template library as the similarity score. Figure 6 The similarity between 13 modulation pattern signals and four templates in the template library is given. Based on this figure, the threshold is set to 2.5. When the similarity is less than 2.5, it indicates that the signal to be identified belongs to one of the templates in the template library. Otherwise, the signal to be identified does not belong to any of the templates in the template library.

[0047] Step 4: If a match is successful, the first recognition model is used to identify the cyclic two-dimensional spectrum of the signal to be identified. That is, when the similarity is less than 2.5, the cyclic two-dimensional spectrum of the signal to be identified is input into the first recognition model to achieve the identification of the signal to be identified.

[0048] The first recognition model in this step is built based on the MCLDNN network. The MCLDNN network is trained to be a recognition model specifically designed to distinguish four easily confused signals (QPSK, 8PSK, 16QAM, and 64QAM). Its network structure is as follows: Figure 7 As shown, the MCLDNN network integrates CNN, LSTM, and fully connected (FC) deep neural networks into a unified structure, leveraging their complementarity and synergy for spatiotemporal feature extraction and classification.

[0049] Step 5: If the matching fails, the second recognition model is used to identify the cyclic two-dimensional spectrum of the signal to be identified. That is, when the similarity is greater than or equal to 2.5, the cyclic two-dimensional spectrum of the signal to be identified is input into the second recognition model to achieve the identification of the signal to be identified.

[0050] The second recognition model in this step is built on a CNN neural network, which is trained to specifically distinguish between signals that are not easily confused. Its network structure is as follows: Figure 8 As shown, the main task is to extract and classify features from the two-dimensional spectrum of the cyclic spectrum.

[0051] This implementation also included a comparative experiment. The specific parameter settings for the dataset in the comparative experiment are shown in Table 1 below. The dataset was divided into a training set, a validation set, and a test set in a 2:1:1 ratio. The initial learning rate was set to 0.00001, and the Adam gradient descent algorithm was selected.

[0052] Table 1 Dataset Parameter Settings

[0053]

[0054]

[0055] To demonstrate the advantages of the method in this implementation (Template_CNN_MCLDNN) compared to other algorithms, several comparison algorithms were selected: CNN_IQ, LSTM_AmpPhase, and MCLDNN. All experiments were run on the same hardware and software platform (RTX3060 platform based on the Keras deep learning framework). Experimental results are as follows: Figure 9 As shown in the figure, it can be observed that the recognition rate of the method in this embodiment is basically the highest when the signal-to-noise ratio is above 4dB, and it also outperforms other algorithms when the signal-to-noise ratio is below -5dB. The total confusion matrix in the signal-to-noise ratio range of -20 to 18dB is shown in the figure. Figure 10 As shown, it is easy to see that the method of this embodiment solves the problem of decreased recognition performance caused by the similarity of the cyclic spectra of some high-order modulation signals, and improves the overall recognition performance of the network.

Claims

1. A deep learning modulation recognition method based on cyclic two-dimensional spectrogram deblurring, characterized in that, Includes the following steps: Create templates for cyclic two-dimensional spectra of easily confused signals and save the obtained templates to a template library; The cyclic two-dimensional spectrum of the easily confused signal includes the cyclic two-dimensional grayscale image of QPSK, the cyclic two-dimensional grayscale image of 8PSK, the cyclic two-dimensional grayscale image of 16QAM, and the cyclic two-dimensional grayscale image of 64QAM. The template for creating a cyclic two-dimensional spectrum of easily confused signals specifically includes: Obtain multiple cyclic two-dimensional spectra of each modulation style signal at a preset signal-to-noise ratio; A template is obtained by performing a pixel-weighted average of multiple cyclic two-dimensional spectra of the same modulation pattern signal at a preset signal-to-noise ratio. Receive the signal to be identified and convert it into a cyclic two-dimensional spectrum; The cyclic two-dimensional spectrum of the signal to be identified is matched with the templates in the template library. Specifically, the mean square error of the pixels of the cyclic two-dimensional spectrum of the signal to be identified and each template in the template library is calculated, and a threshold is set to determine whether the cyclic two-dimensional spectrum of the signal to be identified matches the target in the template library. If a match is successful, the first recognition model is used to identify the cyclic two-dimensional spectrum of the signal to be identified, wherein the first recognition model is trained using sample data of easily confused signals; If the matching fails, the second recognition model is used to identify the cyclic two-dimensional spectrum of the signal to be identified, wherein the second recognition model is trained using sample data of signals that are not easily confused.

2. The deep learning modulation recognition method based on cyclic two-dimensional spectrogram deblurring according to claim 1, characterized in that, The process of converting the signal to be identified into a cyclic two-dimensional spectrum specifically includes: The signal to be identified is processed to obtain a three-dimensional cyclic spectral density map; The three-dimensional cyclic spectral density map is projected into a two-dimensional image by top-down projection to obtain a cyclic two-dimensional spectrum.

3. The deep learning modulation recognition method based on cyclic two-dimensional spectrogram deblurring according to claim 2, characterized in that, The process of processing the signal to be identified to obtain a three-dimensional cyclic spectral density map specifically includes: The autocorrelation function of the signal to be identified is expanded into a Fourier series to obtain the cyclic autocorrelation. A three-dimensional cyclic spectral density map is obtained by performing a Fourier transform on the cyclic autocorrelation.

4. The deep learning modulation recognition method based on cyclic two-dimensional spectrogram deblurring according to claim 1, characterized in that, The first recognition model is built based on the MCLDNN network.

5. The deep learning modulation recognition method based on cyclic two-dimensional spectrogram deblurring according to claim 1, characterized in that, The second recognition model is built on a CNN neural network.

Images