Signal automatic modulation identification method and device under communication and inductance integrated system, and medium

Abstract

The invention discloses a signal automatic modulation identification method and device under a communication-inductance integrated system and a medium, and the method comprises the steps: simulating a communication process through Matlab, simulating a real channel effect, and generating signals in different modulation modes; taking a signal real part as an I-path signal and an imaginary part signal as a Q-path signal as original characteristic data; generating a time-frequency image according to the I / Q original signal, performing image processing, signal noise reduction and image feature enhancement to obtain a processed time-frequency image, and performing packaging according to different modulation signals and different signal-to-noise ratio levels to obtain a time-frequency image data set; and inputting the obtained time-frequency image into a pre-trained residual attention network model for identification, performing classification prediction on a signal sample, and identifying a modulation mode of the signal sample. According to the invention, based on the time-frequency image of the signal, the most advanced classification precision is realized on a user-defined mixed data set containing radar and communication signals; especially, features can be extracted from a small number of signals, and the recognition rate is remarkably improved under the condition of small samples.

Description

Technical Field

[0001] The invention belongs to the field of automatic signal modulation recognition, and in particular relates to a method, a device and a medium for automatic signal modulation recognition in a synaesthesia integrated system. Background Art

[0002] Automatic modulation recognition (AMC) plays an important role in intelligent communication receivers as an intermediate step between signal detection and demodulation. AMC has a wide range of applications in military and civilian fields such as spectrum management, electronic reconnaissance, and electronic countermeasures. In particular, identifying the modulation mode of potential attack signals helps in spectrum risk monitoring to discover threat sources, and plays a very important role in non-cooperative communication scenarios.

[0003] Traditional AMC methods mainly include likelihood-based (LB) methods and feature-based (FB) methods. Likelihood methods rely on hypothesis testing theory and construct judgment criteria by analyzing the statistical characteristics of signals, which requires a lot of prior knowledge. Due to the excellent performance of feature-based methods, they have gradually become the mainstream method in the field of AMC.

[0004] However, current models are usually applicable to modulation recognition within a single system, such as a radar system or a communication system. With the rapid development of the wireless communication industry, the fifth generation (5G) and sixth generation (6G) wireless technologies are expected to use higher frequency bands and wider frequency bandwidths than their predecessors. To address the problems of spectrum scarcity and interference, synaesthesia has been widely studied in academia. The emergence of synaesthesia systems highlights the critical need to distinguish between communication and radar signals with similar spectra and characteristics. Existing AMC techniques are insufficient to handle such mixed signals, coupled with the lack of suitable datasets. Therefore, there is an urgent need to develop an efficient automatic modulation recognition method tailored specifically for synaesthesia systems. Summary of the invention

[0005] Purpose of the invention: The present invention provides an automatic modulation recognition method based on time-frequency images for the synaesthesia integration system, which achieves high classification accuracy on a custom mixed data set containing radar and communication signals; in particular, it can extract features from a small amount of signals, and the recognition rate is significantly improved in the case of small samples.

[0006] Technical solution: The method for automatic signal modulation recognition in a synaesthesia integrated system described in the present invention specifically comprises the following steps:

[0007] (1) Use channel simulation technology to simulate the communication process, simulate the real channel impact, and generate signals that contain classic modulation methods used in communication and radar signals;

[0008] (2) The real part of the signal is taken as the I-path signal, and the imaginary part of the signal is taken as the Q-path signal as the original feature data;

[0009] (3) Generate time-frequency images based on the original I-channel and Q-channel signals using the Choi-Williams distribution method;

[0010] (4) Perform image enhancement processing on the time-frequency image to obtain the processed time-frequency image, and encapsulate it according to different modulation signals and different signal-to-noise ratio levels to obtain a time-frequency image dataset;

[0011] (5) The time-frequency image obtained in step (4) is input into the pre-trained residual self-attention network model for recognition, the signal samples are classified and predicted, and their modulation mode is identified.

[0012] Furthermore, the implementation process of step (1) is as follows:

[0013] The communication process is simulated using MATLAB, with random channel delay and Doppler frequency shift added to simulate the real channel impact. At the transmitting end, each signal has unique parameters. Combining the modulation types of communication and radar systems, a data set of 12 modulation types is created, including 7 radar signals and 5 communication signals.

[0014] The signal-to-noise ratio ranges from -18dB to 18dB, increasing in 3dB increments, combined with the influence of AWGN and multipath fading channels, sending signals to the receiving end respectively;

[0015] At the receiving end, the received signal is demodulated and simulated to obtain an IQ signal.

[0016] Furthermore, the implementation process of step (2) is as follows:

[0017] The real part of the signal is extracted as the I-channel data, and the imaginary part is extracted as the Q-channel signal. The data set is divided according to the modulation signal, and the modulation mode and signal-to-noise ratio are used as labels to encapsulate the samples.

[0018] Furthermore, the implementation process of step (3) is as follows:

[0019] The Choi-Williams distribution is used to extract the time-frequency information of the signal, reduce noise, and filter out interleaving terms. The definition of CWD in the time domain is:

[0020]

[0021] in,

[0022]

[0023] The kernel function is:

[0024] Φ(η,τ)=exp[-α(ητ) 2 ]

[0025] Among them, x is the complex number representation of the original signal, x * It is represented as the conjugate signal of x, t is the time variable, f is the frequency variable, η and τ are the frequency and time lag respectively, and α reflects the degree of suppression of the cross term;

[0026] The image obtained by CWD is resized to a uniform size of 224×224 using the bicubic interpolation method; then the pixel values ​​of the image are normalized to obtain a grayscale image.

[0027] Furthermore, the implementation process of step (4) is as follows:

[0028] Image enhancement is performed using global image threshold segmentation and linear stretching. The converted pixel value y is:

[0029]

[0030] Among them, g is the original pixel value of the image, and k is the adjustable threshold;

[0031] Packaging according to different modulation signals and different signal-to-noise ratio levels to obtain a time-frequency image data set;

[0032] Three data sets of different sizes are generated with sample sizes of 30, 100, and 500 in each signal-to-noise ratio stage and each modulation mode, and each data set is divided into a training set, a validation set, and a test set.

[0033] Furthermore, the implementation process of step (5) is as follows:

[0034] The input time-frequency image is first processed using a 7×7 convolution kernel to downsample the input and retain salient features; the resulting feature map of size 56×56 is then fed into the RA module;

[0035] The RA module consists of a residual module and a SE module. The residual module includes two convolutional layers with kernel sizes of 3×3 and 2×2, and an SE module. The SE module includes a global pooling layer, a fully connected layer, a Relu activation layer, a second fully connected layer, and a Sigmoid activation layer. As the layers go deeper, the attention perception function evolves adaptively between different modules.

[0036] After passing through eight layers of RA modules, a feature map of size 7×7 is output, and finally passes through an adaptive pooling layer, a Flatten layer, a fully connected layer, and a Softmax layer to obtain the classification prediction result; among them, a Relu activation layer is added after each convolutional layer.

[0037] A device according to the present invention includes a memory and a processor, wherein:

[0038] A memory for storing computer programs that can be run on the processor;

[0039] The processor is used to execute the steps of the signal automatic modulation recognition method in the synaesthesia integration system as described above when running the computer program.

[0040] The present invention discloses a storage medium having a computer program stored thereon, and when the computer program is executed by at least one processor, the steps of the automatic signal modulation identification method in the synaesthesia integration system as described above are implemented.

[0041] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: the present invention utilizes channel simulation technology to simulate real signals, creates a signal set of a communication and radar coexistence system, obtains the time-frequency image of the signal through the Choi-Willams method, performs image enhancement, extracts the signal time-frequency information, and puts it into an improved convolutional neural network for training, which can effectively extract signal features, thereby improving the recognition accuracy of automatic modulation recognition technology; the present invention adds a residual self-attention module to the network, which can effectively extract effective features in the signal, has fewer parameters and stronger generalization ability than traditional methods, requires fewer training samples, and has higher recognition accuracy; the present invention utilizes MATLAB to simulate the influence of real channels, and generates a data set containing radar and communication modulation methods by simulating radar and communication signals, so that the model can be applicable to the synaesthesia integration system, thereby improving the credibility of the recognition result. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of the signal automatic modulation recognition method under the synaesthesia integration system;

[0043] Figure 2 Schematic diagram of the residual self-attention model proposed in the present invention;

[0044] Figure 3 This is a comparison chart of recognition accuracy of different comparison methods under different training sample sizes of the present invention;

[0045] Figure 4 This is a comparison diagram of the sparse matrix at 0 dB and 6 dB in the present invention. DETAILED DESCRIPTION

[0046] The present invention is further described in detail below with reference to the accompanying drawings.

[0047] like Figure 1 As shown, the present invention proposes a signal automatic modulation recognition method under a synaesthesia integrated system, comprising the following steps:

[0048] Step 1: Use channel simulation technology to simulate the communication process, add random channel delays and Doppler frequency shifts within a reasonable range, simulate the impact of real channels, generate signals containing classic modulation methods in communication and radar signals, and create a data set for the communication and radar coexistence system.

[0049] The communication process is simulated using MATLAB to simulate the real channel impact. At the transmitting end, each signal has unique parameters. Combined with commonly used modulation types, a data set of 12 modulation types is created, including 7 radar signals (Rect, LFM, Costas, T1, T2, P1, P2) and 5 communication signals (BPSK, 16QAM, CPFSK, GFSK, FM). The signal-to-noise ratio ranges from -18dB to 18dB, increasing in 3dB increments. Combined with the fading channel impact such as AWGN and multipath effects, the signals are sent to the receiving end respectively. At the receiving end, the received signal is demodulated and simulated to obtain the original IQ signal.

[0050] Step 2: The real part of the received signal is taken as the I-channel signal, and the imaginary part of the received signal is taken as the Q-channel signal as the original signal data;

[0051] The real part of the signal is extracted as the I-channel data, and the imaginary part is extracted as the Q-channel signal. The data set is divided according to the modulation signal, and the modulation mode and signal-to-noise ratio are used as labels to encapsulate the samples.

[0052] Step 3: Generate a time-frequency image based on the original I / Q signal, and use the CWD method to extract the signal time-frequency features.

[0053] Choi-Williams distribution (CWD) is used to extract the time-frequency information of the signal, which has high clarity, can reduce noise, and filter out interleaving terms. Among them, CWD is defined in the time domain as:

[0054]

[0055] in,

[0056]

[0057] The kernel function is:

[0058] Φ(η,τ)=exp[-α(ητ) 2 ]

[0059] Among them, x is the complex number representation of the original signal, x * It is expressed as the conjugate signal of x, t is the time variable, f is the frequency variable, η and τ are the frequency and time lag respectively, and α reflects the degree of suppression of the cross term.

[0060] The image obtained by CWD is resized to 224 × 224 using the bicubic interpolation method. Subsequently, the pixel values ​​of the image are normalized to fall within the range of [0, 255] to obtain a grayscale image.

[0061] Step 4: Perform image processing on the time-frequency image to extract the time-frequency features in the signal, reduce signal noise, enhance image features, and obtain a processed time-frequency image.

[0062] Image enhancement is performed using global image threshold segmentation and linear stretching to improve image quality, thereby showing finer details and enhancing image contrast. The converted pixel value y is:

[0063]

[0064] Among them, g is the original pixel value of the image, and k is the adjustable threshold.

[0065] The time-frequency image dataset was obtained by encapsulating it according to different modulation signals and different signal-to-noise ratio levels. Three datasets of different sizes were generated with the sample sizes of 30, 100, and 500 in each signal-to-noise ratio stage and each modulation mode, and each dataset was divided into training set, validation set, and test set according to the ratio of 6:2:2.

[0066] Step 5: Input the obtained time-frequency image into the pre-trained residual attention network model, use the residual self-attention model to extract features of the image, classify and predict the signal, and output the modulation method.

[0067] The residual self-attention model (RA-CNN) is as follows Figure 2 As shown in Figure 1, the input time-frequency image is first processed using a 7×7 convolution kernel to downsample the input and retain salient features. The resulting feature map of size 56×56 is then fed into the RA module.

[0068] The RA module consists of a residual module and a SE module, where the residual module includes two convolutional layers with kernel sizes of 3×3 and 2×2, and an SE module, which includes a global pooling layer, a fully connected layer, a Relu activation layer, a second fully connected layer, and a Sigmoid activation layer. As the layers go deeper, the attention perception function evolves adaptively between different modules.

[0069] After eight layers of RA modules, the output feature map is 7×7, and finally passes through the adaptive pooling layer, Flatten layer, fully connected layer and Softmax layer to obtain the classification prediction result. A Relu activation layer is added after each convolution layer.

[0070] In addition, the present invention provides a device, including a memory and a processor, wherein: the memory is used to store a computer program that can be run on the processor; the processor, when running the computer program, executes the steps of the signal automatic modulation identification method under the synaesthesia integration system as described above.

[0071] The present invention also provides a storage medium having a computer program stored thereon. When the computer program is executed by at least one processor, the steps of the signal automatic modulation identification method in the synaesthesia integration system as described above are implemented.

[0072] The present invention adopts the cross entropy loss function, the optimizer adopts the Adam optimizer, the learning rate is set to 0.001, the skin size of the experimental gradient update is set to 100, and the other parameters are the default parameters in Pytorch. The results are shown in Figure 3 , Figure 4 As shown, Figure 3 From left to right are the comparison charts of classification accuracy when the sample size is 30, 100, and 500, respectively, where RA-CNN is an example proposed by the present invention. Figure 4 : In the figure is the confusion matrix of the present invention when the sample size is 500 and the signal-to-noise ratio is 0dB and 6dB respectively. The horizontal axis represents the predicted label and the vertical axis represents the actual correct label.

[0073] All models are tested using datasets with different sample sizes. Figure 3 As shown in the figure, the present invention achieves higher accuracy than other models. This advantage stems from the residual self-attention network and the time-frequency image processing module, whose combination can effectively identify signal features, allowing the present invention to achieve an accuracy of nearly 80% even when the sample size is 30. In addition, when the sample size is increased to 500, the present invention outperforms other models by about 5% in the scene, with an accuracy of up to 95%. These results demonstrate the effectiveness of the present invention in modulation recognition, highlighting its superiority over other methods.

[0074] like Figure 3 As shown in the figure, in the three cases, when the signal-to-noise ratio is greater than -6dB, the present invention can achieve higher accuracy. This result is attributed to the time-frequency image processing, which enhances the signal characteristics and enables the present invention to better identify the modulation mode. However, in the case of low signal-to-noise ratio, the noise will blur the signal and it becomes challenging to extract meaningful features from the image, resulting in a decrease in performance at this time.

[0075] The results show that the proposed method can successfully identify most modulation modes. The main confusion occurs in BPSK and QAM, because the main difference between BPSK and QAM is the phase, which is difficult to represent on the time-frequency image. However, the proposed method performs well in identifying the remaining modulation modes, showing the applicability of the model.

[0076] In general, the residual self-attention network based on time-frequency image processing proposed in the present invention is used for the automatic modulation recognition task in the synaesthesia integrated system. The present invention integrates the time-frequency image processing with the residual self-attention network framework. In order to verify the feasibility of the present invention, a mixed data set containing communication and radar modulation signals is constructed. It is worth noting that the present invention shows robustness even with a limited number of signal samples.

[0077] Experimental results show that the effectiveness of the proposed method in identifying joint communication and radar modulation is significantly improved compared with other networks (such as CNN, CLDNN, AWN). Specifically, when the signal samples are limited, the recognition performance of the proposed method is significantly better than other methods.

[0078] The technical solutions provided by the embodiments of the present invention are described in detail above. The principles and implementation methods of the embodiments of the present invention are described in this patent using specific examples. The above embodiments and features in the embodiments can be combined with each other without conflict. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for automatic signal modulation recognition in a synaesthesia integrated system, It is characterized in that The following steps are involved: (1) Use channel simulation technology to simulate the communication process, simulate the real channel impact, and generate signals that contain classic modulation methods used in communication and radar signals; (2) The real part of the signal is taken as the I-path signal, and the imaginary part of the signal is taken as the Q-path signal as the original feature data; (3) Generate time-frequency images based on the original I-channel and Q-channel signals using the Choi-Williams distribution method; (4) Perform image enhancement processing on the time-frequency image to obtain the processed time-frequency image, and encapsulate it according to different modulation signals and different signal-to-noise ratio levels to obtain a time-frequency image dataset; (5) The time-frequency image obtained in step (4) is input into the pre-trained residual self-attention network model for recognition, the signal samples are classified and predicted, and their modulation mode is identified.

2. The signal automatic modulation recognition method in the synaesthesia integrated system according to claim 1, It is characterized in that The implementation process of step (1) is as follows: Use MATLAB to simulate the communication process, add random channel delay and Doppler frequency shift to simulate the real channel impact. At the transmitting end, each signal has unique parameters. Combine the modulation types of communication and radar systems to create data sets with different modulation types. The signal-to-noise ratio ranges from -18dB to 18dB, increasing in 3dB increments, combined with the influence of AWGN and multipath fading channels, sending signals to the receiving end respectively; At the receiving end, the received signal is demodulated and simulated to obtain an IQ signal.

3. The signal automatic modulation recognition method in the synaesthesia integrated system according to claim 1, It is characterized in that The implementation process of step (2) is as follows: The real part of the signal is extracted as the I-channel data, and the imaginary part is extracted as the Q-channel signal. The data set is divided according to the modulation signal, and the modulation mode and signal-to-noise ratio are used as labels to encapsulate the samples.

4. The signal automatic modulation recognition method in the synaesthesia integrated system according to claim 1, It is characterized in that The implementation process of step (3) is as follows: The Choi-Williams distribution is used to extract the time-frequency information of the signal, reduce noise, and filter out interleaving terms. The definition of CWD in the time domain is: in, The kernel function is: Φ(η,τ)=exp[-α(ητ) 2 ] Among them, x is the complex number representation of the original signal, x * It is represented as the conjugate signal of x, t is the time variable, f is the frequency variable, η and τ are the frequency and time lag respectively, and α reflects the degree of suppression of the cross term; The image obtained by CWD is resized to a uniform size using the bicubic interpolation method; then the pixel values ​​of the image are normalized to obtain a grayscale image.

5. The signal automatic modulation recognition method in the synaesthesia integrated system according to claim 1, It is characterized in that The implementation process of step (4) is as follows: Image enhancement is performed using global image threshold segmentation and linear stretching. The converted pixel value y is: Among them, g is the original pixel value of the image, and k is the adjustable threshold; Packaging according to different modulation signals and different signal-to-noise ratio levels to obtain a time-frequency image data set; Three data sets of different sizes are generated with sample sizes of 30, 100, and 500 in each signal-to-noise ratio stage and each modulation mode, and each data set is divided into a training set, a validation set, and a test set.

6. The signal automatic modulation recognition method in the synaesthesia integration system according to claim 1, It is characterized in that The implementation process of step (5) is as follows: The input time-frequency image is first processed using a 7×7 convolution kernel to downsample the input and retain salient features; the resulting feature map of size 56×56 is then fed into the RA module; The RA module consists of a residual module and a SE module. The residual module includes two convolutional layers with kernel sizes of 3×3 and 2×2, and an SE module. The SE module includes a global pooling layer, a fully connected layer, a Relu activation layer, a second fully connected layer, and a Sigmoid activation layer. As the layers go deeper, the attention perception function evolves adaptively between different modules. After passing through eight layers of RA modules, a feature map of size 7×7 is output, and finally passes through an adaptive pooling layer, a Flatten layer, a fully connected layer, and a Softmax layer to obtain the classification prediction result; among them, a Relu activation layer is added after each convolutional layer.

7. The signal automatic modulation recognition method in the synaesthesia integrated system according to claim 2, It is characterized in that There are 12 data sets of different modulation types, including 7 radar signals and 5 communication signals.

8. The signal automatic modulation recognition method in the synaesthesia integrated system according to claim 4, It is characterized in that The uniform size is 224×224.

9. A device, It is characterized in that comprising a memory and a processor, wherein: A memory for storing computer programs that can be run on the processor; A processor is used to execute the steps of the signal automatic modulation recognition method in the synaesthesia integration system as described in any one of claims 1 to 8 when running the computer program.

10. A storage medium, It is characterized in that The storage medium stores a computer program, which, when executed by at least one processor, implements the steps of the signal automatic modulation identification method in the synaesthesia integration system as claimed in any one of claims 1 to 8.

Images