Radar signal modulation identification method based on time-frequency fusion features and MSANet

Abstract

The invention belongs to the technical field of radar signal processing. Aiming at the problems that a time-frequency image is seriously interfered in a low signal-to-noise ratio environment, and a common convolutional neural network cannot adapt to a variable time-frequency graph, so that the recognition effect is poor, the invention provides a radar signal intra-pulse modulation type recognition method based on time-frequency fusion features and MSANet, and the method comprises the following steps: carrying out three types of time-frequency analysis on a radar signal; preprocessing to obtain a training set, a verification set and a test set; then, constructing an MSANet model, and training the MSANet model by using the training set and the verification set; and after the training of the MSANet model is finished, testing the recognition accuracy and the anti-confusion capability of the MSANet model by using the test set. The method can be used for classifying a pulse internal modulation mode of as many as 12 types of radar signals, and has good robustness under a low signal-to-noise ratio.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal processing, and in particular relates to a radar signal intra-pulse modulation recognition method based on time-frequency fusion features and a multi-scale perception network MSANet. Background Art

[0002] As a core technology in the field of electronic warfare, radar signal modulation recognition plays a vital role in the outcome of modern warfare. It involves analyzing and identifying the signals emitted by enemy radars to reveal their working mechanisms and intentions. With the development of electronic technology, the complexity and diversity of radar signals are increasing, which requires that radar signal modulation recognition technology must continue to improve to adapt to the ever-changing battlefield environment. Traditional methods based on manual decision-making and pattern recognition have obvious defects, such as difficulty in extracting effective features, limited recognition of signal types, slow processing speed, and poor performance in electromagnetic interference environments. These limitations make it difficult to meet the high standards of modern warfare.

[0003] With the continuous advancement of deep learning technology, the signal recognition method based on convolutional neural network provides a promising solution for intra-pulse modulation recognition. This method uses intelligent algorithms to identify complex multi-class radar signals, usually using the time-frequency image of the signal as input. However, in a low signal-to-noise ratio environment, the time-frequency image will be severely interfered, and ordinary convolutional neural networks cannot adapt to the changing time-frequency graph, resulting in poor recognition results. Summary of the invention

[0004] The present invention provides a radar signal pulse intra-modulation recognition method based on time-frequency fusion features and MSANet, so as to improve the interference of noise on the time-frequency image, and enhance the acquisition of global signal dependency information through multi-scale feature extraction, so as to finally accurately complete the radar signal pulse intra-modulation recognition.

[0005] The technical solution of the present invention to solve the above technical problems is as follows:

[0006] A radar signal pulse intra-modulation recognition method based on time-frequency fusion features and MSANet includes the following steps:

[0007] Step 1. Perform Winger-Ville time-frequency analysis, Fourier synchronous compression transform, and smoothed pseudo-Wigner-Ville time-frequency analysis on the radar sampling signal to obtain three types of time-frequency image sets;

[0008] Step 2. Preprocess the time-frequency image set. First, grayscale, binarize, and open and close the three types of time-frequency images respectively. Second, crop the three types of time-frequency images to obtain a suitable size. Then, fuse the three types of time-frequency images in the channel dimension to obtain a three-channel time-frequency fusion feature set.

[0009] The time-frequency fusion feature set is divided into a training set, a validation set, and a test set according to a preset ratio;

[0010] Step 3. Build the MSANet model, use cross-loss entropy as the loss function, use stochastic gradient descent as the optimizer, and then set the training parameters;

[0011] Step 4. Use the training set and validation set to train the MSANet model;

[0012] Step 5. For the MSANet model that has been trained, use the test set to test the recognition accuracy and anti-obfuscation ability of the MSANet model;

[0013] Furthermore, the radar sampling signals in step 1 include CW, LFM, NLFM, BPSK, QPSK, FSK, 4FSK, FRANK and P1~P4, a total of 12 categories; the Winger-Ville time-frequency analysis expression is:

[0014]

[0015] Where t represents time, τ is the integral variable, f represents frequency, (·) * represents the conjugate of the complex number, x(t) represents the received radar signal; the Fourier synchronous compression transform expression is:

[0016]

[0017] Among them, η(t) represents the window function, ω represents the corrected frequency, δ is the impulse response, is the local instantaneous frequency; The expression is:

[0018]

[0019] Among them, RE(·) represents the real part; the smoothed pseudo-Wigner-Ville time-frequency analysis expression is:

[0020]

[0021] Among them, H(f) and g(t) are smoothing window functions, and u is the integration variable.

[0022] Furthermore, in step 2, the binarization method adopted is adaptive threshold binarization; the cropping method used is bicubic interpolation; the specific process of establishing the time-frequency fusion feature set is:

[0023] In the range of -14dB to 8dB, there are 12 signal-to-noise ratios with an interval of 2dB. Under each signal-to-noise ratio, a total of 500 signal samples are generated for 12 types of radar intra-pulse modulation signals, of which 300 are divided into training samples, 100 are divided into verification samples, and 100 are divided into test samples.

[0024] Furthermore, the MSANet model established in step 3 includes 5 convolutional layers, 1 pooling layer, 3 dilated convolution compression structures, 1 flattening layer, and 1 fully connected layer; the dilated convolution compression structure first uses convolution with 3 dilation rates, channel information compression, and spatial information compression to obtain multi-scale information, and then fuses the feature information through activation functions, point multiplications, and direct addition operations to obtain multi-scale fusion information.

[0025] Furthermore, in step 4, the MSANet model is trained using training samples with various signal-to-noise ratios, and then the MSANet model is tested using the validation set, and the training is continued until the loss curve of the MSANet model converges.

[0026] The advantage of the present invention is that a radar signal intra-pulse modulation recognition method based on time-frequency fusion features and MSANet is proposed to address the problems of serious time-frequency image pollution and poor performance of ordinary convolutional neural networks under low signal-to-noise ratios. By fusing multiple time-frequency features, more useful information is provided, and the impact of time-frequency image pollution on accuracy is significantly reduced. MSANet obtains time-frequency information at multiple scales by using convolutions with multiple expansion rates, and adaptively obtains important information through the attention mechanism, which more effectively deals with the problems of fast changes in the time-frequency ridges, narrow curves and large spans of radar signals. Experimental results show that for 12 types of intra-pulse modulation radar signals, the overall recognition accuracy of the present invention is 92.99% when the signal-to-noise ratio is as low as -12dB, and it has strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flow chart of the method of the present invention;

[0028] Figure 2 Schematic diagram of the time-frequency distribution of Fourier synchronous compression transform of 12 types of radar signals in the present invention;

[0029] Figure 3 It is a schematic diagram of pretreatment in the present invention;

[0030] Figure 4 It is a schematic diagram of MSANet in the present invention;

[0031] Figure 5 Schematic diagram of the dilated convolution compression structure in the present invention;

[0032] Figure 6A schematic diagram showing a comparison of activation mapping analysis effects in the present invention;

[0033] Figure 7 It is a schematic diagram of the classification correct recognition rate curve in the present invention;

[0034] Figure 8 Schematic diagram of the classification confusion matrix in the present invention;

[0035] Fig. 9 It is a schematic diagram comparing the accuracy recognition rates of the four recognition methods in the present invention;

[0036] Fig.10 It is a schematic diagram of the accurate recognition rate of different time-frequency analysis methods in the present invention. Specific implementation methods

[0038] Below, we will explain the present invention in more detail with reference to the embodiments and drawings. It should be noted that the embodiments of the present invention are not limited to this example.

[0039] Example 1.

[0040] Refer to the appendix of the manual Figure 1 As shown, the specific implementation of Example 1 is divided into the following steps:

[0041] Step 1. Perform three time-frequency analyses on the radar sampling signal, where the Winger-Ville time-frequency analysis expression is:

[0042]

[0043] Where t represents time, τ is the integral variable, f represents frequency, (·) * represents the conjugate of the complex number, x(t) represents the received radar signal; the Fourier synchronous compression transform expression is:

[0044]

[0045] Among them, η(t) represents the window function, ω represents the corrected frequency, δ is the impulse response, is the local instantaneous frequency; The expression is:

[0046]

[0047] Among them, RE(·) represents the real part; the smoothed pseudo-Wigner-Ville time-frequency analysis expression is:

[0048]

[0049] Wherein, H(f) and g(t) are smooth window functions, and u is an integral variable. In the embodiment, three types of time-frequency analysis are performed on 12 types of radar signals to obtain three types of time-frequency image sets. Figure 2 The figure is a schematic diagram of the time-frequency distribution of 12 types of radar signals obtained by Fourier synchronous compression transform. The 12 types of signals are CW, LFM, NLFM, BPSK, QPSK, FSK, 4FSK, FRANK and P1~P4;

[0050] Step 2. As shown in the appendix of the manual Figure 3 As shown in the figure, the obtained time-frequency image is grayed, adaptively thresholded, binarized, opened and closed, and then cropped using the bicubic interpolation method to obtain an image with a length and width of 224. After that, the three types of time-frequency images are fused in the channel dimension to obtain a time-frequency fusion feature map with 3 channels and a length and width of 224. The specific process of establishing the time-frequency fusion feature set is:

[0051] In the range of -14dB to 8dB, there are 12 signal-to-noise ratios with an interval of 2dB. Under each signal-to-noise ratio, 500 signal samples are generated for 12 types of radar intra-pulse modulation signals, of which 300 are divided into training samples, 100 are divided into verification samples, and 100 are divided into test samples. Then, the signal is processed through steps 1 and 2 to obtain the time-frequency fusion feature set; the parameters of the radar signal vary randomly within a certain range, as shown in Table 1.

[0052] Table 1: Parameters of radar signals with different intra-pulse modulation methods

[0053]

[0054]

[0055] Step 3. Build the MSANet model, use cross-loss entropy as the loss function, use stochastic gradient descent as the optimizer, and then set the training parameters. Figure 4 The MSANet model includes 5 convolutional layers, 1 pooling layer, 3 dilated convolutional compression structures, 1 flattening layer, and 1 fully connected layer; the dilated convolutional compression structure is shown in the appendix of the manual. Figure 5 As shown in the figure, convolution with three expansion rates, channel information compression and spatial information compression are first used to obtain multi-scale information, and then the feature information is fused through activation function, point multiplication and direct addition operations to obtain multi-scale fusion information.

[0056] Step 4. Set model training parameters;

[0057] The present invention adopts PyTorch 2.2.1 framework, Windows 11 operating system and NVIDIA GeForce RTX4060Laptop GPU to train the network; the experimental parameters are uniformly set as follows: the initial learning rate is 0.01, the batch processing size is 64, the optimization algorithm is the stochastic gradient descent algorithm, the number of iterations is 50, and the loss function is the cross loss entropy.

[0058] Step 5. Use the training set and validation set to train the MSANet model, and save the MSANet model after training and stabilization;

[0059] During model training, we first initialize the parameters of the model. Then, we calculate the value of the overall loss function through forward propagation. This process continues until the maximum number of iterations is reached. In each iteration, we use the stochastic gradient descent algorithm to update the parameters of the model through backpropagation to gradually reduce the loss during training. In addition, the training set and validation set are used simultaneously during training to evaluate the model performance. When both the training error and the validation error show a convergence trend, it indicates that the model has reached a stable state.

[0060] Step 6. For the trained MSANet model, use the test set to test the recognition accuracy and anti-obfuscation ability of the MSANet model;

[0061] The test set was classified and identified using the MSANet model, and the class activation maps of 12 types of radar intra-pulse modulation signals were obtained; specifically:

[0062] Extract the feature map before the fully connected layer for visualization analysis, and superimpose and merge the original time-frequency images of 12 radar signals and their feature heat maps into one image to obtain the appendix of the manual. Figure 6 , where the higher the brightness, the more important the area is, that is, the area that the network focuses on; use the test set as the input of the MSANet model, classify 12 types of pulse modulation radar signals, including CW, LFM, NLFM, BPSK, QPSK, FSK, 4FSK, FRANK and P1~P4, record the classification results, and obtain the appendix of the manual Figure 7 and Figure 8 ; Figure 7 Indicates the accurate recognition rate of various signals, Figure 8 Indicates the type of each signal classified by the MSANet model. The values ​​on the diagonal are the number of correctly classified samples, and the values ​​on the off-diagonal are the number of confused samples.

[0063] Furthermore, simulation experiment 1 was conducted to build ResNet50, MobileNetV3, and ShuffleNetV2, and the training set and validation set constructed in this paper were used for training, and then the test set was used for classification test to compare with the MSANet model of the present invention. The experimental results are shown in the appendix of the specification. Fig. 9 As shown, the experimental results illustrate the good performance of the MSANet model of the present invention;

[0064] Further, simulation experiment 2 was conducted, and the Fourier synchronous compression transform, Winger-Ville time-frequency analysis and smoothed pseudo-Winger-Ville time-frequency analysis were used to train the MSANet model in the same way as the method of the present invention. At this time, the number of input channels of the first convolutional layer of the MSANet model needs to be adjusted to 1; then, the test sets of the corresponding time-frequency methods were used for classification tests and compared with the time-frequency fusion feature method of the present invention. The experimental results are shown in the appendix of the specification. Fig.10 As shown, the experimental results illustrate that the time-frequency fusion feature of the present invention has better stability and can provide better recognition performance.

[0065] The basic principle of the present invention is to identify the intra-pulse modulation mode of radar signals based on time-frequency fusion features and the MSANet model. Due to the serious pollution of time-frequency images at low signal-to-noise ratios, a single time-frequency image cannot provide enough information to support accurate signal recognition. The time-frequency fusion features of the present invention fuse multiple time-frequency analysis methods as the object of neural network learning, which has stronger robustness; at the same time, the MSANet model uses an expanded convolution compression structure to overcome the rapid changes in the time-frequency images of radar signals with different modulation modes, extract signals of multiple scales and perform information compression.

[0066] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A radar signal pulse intra-modulation recognition method based on time-frequency fusion features and MSANet, characterized in that: The steps include: Step 1. Perform Winger-Ville time-frequency analysis, Fourier synchronous compression transform, and smoothed pseudo-Wigner-Ville time-frequency analysis on the radar sampling signal to obtain three types of time-frequency image sets; Step 2. Preprocess the time-frequency image set. First, grayscale, binarize, and open and close the three types of time-frequency images respectively. Second, crop the three types of time-frequency images to obtain a suitable size. Then, fuse the three types of time-frequency images in the channel dimension to obtain a three-channel time-frequency fusion feature set. The time-frequency fusion feature set is divided into a training set, a validation set, and a test set according to a preset ratio; Step 3. Build the MSANet model, use cross-loss entropy as the loss function, use stochastic gradient descent as the optimizer, and then set the training parameters; Step 4. Use the training set and validation set to train the MSANet model; Step 5. For the MSANet model that has completed training, use the test set to test the recognition accuracy and anti-confusion ability of the MSANet model.

2. According to claim 1, a radar signal pulse intra-modulation recognition method based on time-frequency fusion features and MSANet is characterized in that: The radar sampling signals in step 1 include CW, LFM, NLFM, BPSK, QPSK, FSK, 4FSK, FRANK and P1~P4; the Winger-Ville time-frequency analysis expression is: Where t represents time, τ is the integral variable, f represents frequency, (·) * represents the conjugate of the complex number, x(t) represents the received radar signal; the Fourier synchronous compression transform expression is: Among them, η(t) represents the window function, ω represents the corrected frequency, δ is the impulse response, is the local instantaneous frequency; The expression is: Among them, RE(·) represents the real part; the smoothed pseudo-Wigner-Ville time-frequency analysis expression is: Among them, H(f) and g(t) are smoothing window functions, and u is the integration variable.

3. According to claim 1, a radar signal pulse intra-modulation recognition method based on time-frequency fusion features and MSANet is characterized in that: In step 2, the binarization method used is adaptive threshold binarization; the cropping method used is bicubic interpolation; the specific process of establishing the time-frequency fusion feature set is: In the range of -14dB to 8dB, there are 12 signal-to-noise ratios with an interval of 2dB. Under each signal-to-noise ratio, a total of 500 signal samples are collected for 12 types of radar intra-pulse modulation signals, of which 300 are divided into training samples, 100 are divided into verification samples, and 100 are divided into test samples.

4. According to claim 1, a radar signal pulse intra-modulation recognition method based on time-frequency fusion features and MSANet is characterized in that: The MSANet model established in step 3 includes 5 convolutional layers, 1 pooling layer, 3 dilated convolution structures, 1 flattening layer, and 1 fully connected layer; the dilated convolution compression structure first uses convolution with 3 dilation rates, channel information compression, and spatial information compression to obtain multi-scale information, and then fuses the feature information through activation functions, point multiplication, and direct addition operations to obtain multi-scale fusion information.

5. According to claim 1, a radar signal pulse intra-modulation recognition method based on time-frequency fusion features and MSANet is characterized in that: In step 4, the MSANet model is trained using training samples with various signal-to-noise ratios, and then the validation set is used to test the MSANet model until the loss curve of the MSANet model converges.

Images