Radar interference signal type identification method and device based on deep learning

Abstract

The invention discloses a radar interference signal type identification method based on deep learning, a constructed deep learning network architecture can directly process a smooth pseudo Wigner-Ville distribution time-frequency image data set of a radar interference signal, a convolution kernel can automatically extract composite features in an interference time-frequency image, the features do not need to be manually designed, and the method is easy to implement. Therefore, the classification process is simplified. In addition, the network architecture can automatically adjust parameters in the training process, and is suitable for different types of interference signals. Moreover, the SqueezeNet network of transfer learning is adopted, and the number of parameters of the model can be greatly reduced while high precision is maintained through a plurality of designed Fire modules, so that the demand of computing resources is reduced, and the iteration time is shortened. Therefore, the network is easier to deploy on various hardware platforms, especially resource-limited equipment, and the applicable scene is enlarged.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing, and in particular, relates to a method and device for identifying radar interference signal types based on deep learning. Background Art

[0002] Radar systems play a vital role in detection, aviation, traffic management and other fields. However, due to the continuous development of jamming technology and the continuous improvement of jamming power, radar systems are facing increasingly serious interference effects. These jammers can be roughly divided into two categories: suppression jammers and deception jammers. Suppression jammers aim to reduce the effective detection range of radars or cause radar failure, while deception jammers attempt to mislead the target recognition function of radars.

[0003] In order to improve the detection effectiveness of radar systems, it is crucial to develop anti-interference technology. Among them, accurate classification of interference signals is the basis for implementing anti-interference strategies. Traditional radar interference signal classification methods mainly include likelihood-based methods and feature-based methods. Likelihood-based methods require the establishment of a mathematical model of the interference signal and the identification of the interference type by calculating the likelihood function. Although this method is theoretically feasible, it is often limited by high computational complexity and insufficient prior knowledge in practical applications. Feature-based methods involve manually extracting the time, frequency, or time-frequency domain features of the interference signal and using classifiers (such as support vector machines) to distinguish different types of interference. Although this method is intuitive and easy to understand, the manual feature extraction process is time-consuming and inefficient.

[0004] In recent years, with the rapid development of artificial intelligence technology, especially the advancement of deep learning technology, new solutions have been provided for radar interference signal classification. Deep learning can automatically learn discriminant features from raw data without the need for manual feature design, thus simplifying the classification process; and it can self-adjust parameters during training to adapt to different types of interference signals. Summary of the invention

[0005] The purpose of the present invention is to provide a radar interference signal type recognition method based on deep learning, so as to improve the interference recognition efficiency of the radar system and thus enhance the anti-interference capability.

[0006] To solve the above problems, the technical solution of the present invention is: A radar interference signal type recognition method based on deep learning, comprising: The server generates a radar interference signal set, wherein the radar interference signal set includes at least pure noise interference, suppression interference, deception interference, and composite interference signals; Perform time-frequency analysis on each signal in the radar jamming signal set to generate corresponding time-frequency color images; Perform image processing on each time-frequency color image to obtain a time-frequency grayscale image and generate a corresponding time-frequency grayscale image set; A deep learning network is built based on the SqueezeNet network framework. Based on the time-frequency grayscale image set, an adaptive distance estimation optimizer is used to train the deep learning network to obtain a trained deep learning network. The trained deep learning network is used to classify radar jamming signals, thereby realizing the identification of radar jamming signal types.

[0007] Among them, the radar interference signal set includes pure noise signals, aimed interference signals, blocked interference signals, swept frequency interference signals, intermittent sampling and forwarding interference signals, distance deception interference signals, dense false target interference signals, smart noise interference signals, dense false target interference + smart noise interference signals, distance deception interference + smart noise interference signals, intermittent sampling and forwarding interference + distance deception interference signals and intermittent sampling and forwarding interference + swept frequency interference signals; each type of signal data set contains time series samples of multiple signals.

[0008] A radar interference signal type identification device based on deep learning is also provided, comprising: An interference signal generation module is used for the server to generate a radar interference signal set, wherein the radar interference signal set includes at least pure noise interference, suppression interference, deception interference, and composite interference signals; The time-frequency analysis module is used to perform time-frequency analysis on each signal in the radar jamming signal set and generate corresponding time-frequency color images; An image processing module is used to perform image processing on each time-frequency color image to obtain a time-frequency grayscale image and generate a corresponding time-frequency grayscale image set; The model training module is used to build a deep learning network based on the SqueezeNet network framework, and to train the deep learning network using an adaptive distance estimation optimizer based on a set of time-frequency grayscale images to obtain a trained deep learning network; The interference signal recognition module is used to classify radar interference signals using a trained deep learning network, thereby realizing the identification of radar interference signal types.

[0009] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art: The radar interference signal type recognition method based on deep learning in one embodiment of the present invention can directly process the smoothed pseudo-Wigner-Ville distribution time-frequency image data set of the radar interference signal through the constructed deep learning network architecture, and the convolution kernel can automatically extract the composite features in the interference time-frequency image without the need for manual feature design, thereby simplifying the classification process. In addition, the network architecture can autonomously adjust parameters during the training process and is applicable to different types of interference signals.

[0010] In addition, the SqueezeNet network, which uses transfer learning, can significantly reduce the number of model parameters while maintaining high accuracy through the design of multiple Fire modules, thereby reducing the demand for computing resources and iteration time. This makes it easier to deploy the network on various hardware platforms, especially resource-constrained devices, and increases the applicable scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a flow chart of a method for identifying radar interference signal types based on deep learning in one embodiment of the present invention; Figure 2 A diagram showing the types of radar interference signals in one embodiment of the present invention; Figure 3 is a time-frequency diagram of a radar jamming signal after being processed by smoothed pseudo-distribution in one embodiment of the present invention; Figure 4 A diagram of the training process of a deep learning network in one embodiment of the present invention; Figure 5 A confusion matrix diagram obtained after classifying radar interference signals in one embodiment of the present invention; Figure 6 This is a block diagram of a radar interference signal type identification device based on deep learning in one embodiment of the present invention. DETAILED DESCRIPTION

[0012] The following is a further detailed description of a radar interference signal type identification method and device based on deep learning proposed by the present invention in conjunction with the accompanying drawings and specific embodiments. The advantages and features of the present invention will become more apparent from the following description and claims.

[0013] Please see Figure 1 , the radar interference signal type identification method based on deep learning provided in this embodiment includes the following steps: The server generates a radar interference signal set, wherein the radar interference signal set includes at least pure noise interference, suppression interference, deception interference, and composite interference signals; Perform time-frequency analysis on each signal in the radar jamming signal set to generate corresponding time-frequency color images; Perform image processing on each time-frequency color image to obtain a time-frequency grayscale image and generate a corresponding time-frequency grayscale image set; A deep learning network is built based on the SqueezeNet network framework. Based on the time-frequency grayscale image set, an adaptive distance estimation optimizer is used to train the deep learning network to obtain a trained deep learning network. The trained deep learning network is used to classify radar jamming signals, thereby realizing the identification of radar jamming signal types.

[0014] Through the deep learning network in the above method, interference signals can be identified quickly and effectively, thereby improving the anti-interference capability of the radar system.

[0015] For details, please refer to the radar jammer signal set generated by the server. Figure 2 In this embodiment, four categories of twelve types of interference signals are generated, namely pure noise, suppression interference, deception interference, and composite interference signals, which are specifically divided into pure noise signals, aiming interference signals, blocking interference signals, swept frequency interference signals, intermittent sampling and forwarding interference signals, distance deception interference signals, dense false target interference signals, smart noise interference signals, dense false target interference + smart noise interference signals, distance deception interference + smart noise interference signals, intermittent sampling and forwarding interference + distance deception interference signals, and intermittent sampling and forwarding interference + swept frequency interference signals; and each type of signal data set contains 3000 time series samples of signals, a total of 36000 samples, and the interference-to-signal ratio is 30~60dB.

[0016] In this embodiment, the radar target signal is a linear frequency modulation pulse: Among them, T=20μs is the pulse width, B=10MHz is the bandwidth, and the sampling rate is 20MHz.

[0017] The principle of suppression jamming is to cover the target signal by transmitting radar jamming signals. The main type of suppression jamming is noise jamming, which is based on the center frequency of the jamming signal. and interference signal bandwidth Relative to the center frequency of the radar transmission signal and transmit signal bandwidth The relationship between them can be divided into aimed interference, blocking interference and swept frequency interference.

[0018] Targeted jamming, usually the jamming signal meets the following requirements: Blocking interference, usually the interference signal meets the following requirements: Sweep frequency interference, usually the interference signal meets the following requirements: The working principle of deceptive jamming is to first receive the real target echo signal, and then forward it after a certain modulation. The modulated deceptive jamming has a strong coherence with the real target echo signal, making it impossible for the radar system to distinguish between real and fake targets, and lose the ability to detect and track the real target.

[0019] Intermittent sampling forwarding interference can produce corresponding false target interference effects by intermittently sampling the radar target echo signal. The expression and spectrum of the intermittent sampling interference signal can be expressed as: Where rect is the rectangular window function, τ is the rectangular signal bandwidth, δ is the impulse response function, n is an integer multiple, T s is the symbol period, and the amplitude weighting coefficient is , Represents a convolution operation.

[0020] Range deception uses DRFM to intercept and store radar signals for replication and forwarding with a certain delay. The range false target interference waveform is consistent with the radar target echo waveform, but the target distance detected by the radar is different, which is greater than the radar's range resolution, achieving the effect of range deception. The range deception interference signal is: A j is the interference signal amplitude, and A j >A r , A r Radar transmit signal amplitude, is the forwarding delay, is the transmission signal delay, Ø is the phase, f c is the center frequency, and K is the frequency modulation slope.

[0021] Dense false targets are when the jammer intercepts and stores radar signals, applies different time delays to the captured signals, each delay is equivalent to a false target, and then superimposes these signals together, so that a series of dense false targets can be formed on the radar display. The signal of the jth false target can be expressed as: All false target signals are superimposed together to form the total interference signal S(t): Where N is the total number of false targets.

[0022] The smart noise jammer has a certain correlation with the radar signal in the time domain and frequency domain, but it cannot be exactly the same, so as not to be eliminated by the radar's adaptive signal processing technology. A common method is to convolve the noise signal with the radar signal. The smart noise jammer signal is: , s(t) is the radar signal, n(t) is the white noise signal, Represents a convolution operation.

[0023] Combined interference is achieved through the additive combination of multiple interference signals. Compared with a single type of radar interference, the interference effect achieved by composite interference is better.

[0024] Gaussian noise is added to each of the above signals, and the signal-to-noise ratio is 0dB.

[0025] This embodiment performs time-frequency analysis on each signal in the radar interference signal set to generate a radar interference signal time-frequency grayscale image set. Specifically, the following steps are included: For each signal in the radar jammer signal set, a smooth pseudo-Wigner-Ville distribution is used to perform time-frequency analysis, and a corresponding time-frequency color image is generated accordingly. Please refer to Figure 3 .

[0026] The Wigner-Ville distribution is: in: represents the complex conjugate of the signal x(t); Using the Gaussian function as the smoothing kernel, the smoothed pseudo-Wigner-Ville distribution can be written as: Here, σ is the standard deviation of the Gaussian function and is used to control the degree of smoothing.

[0027] Downsample each time-frequency color image and convert it to grayscale. Create dedicated subdirectories for each signal type and store the corresponding grayscale images in these subdirectories: Each radar jamming signal time-frequency color image is downsampled into a time-frequency color image of size 227*227, and each time-frequency color image is grayed to obtain the time-frequency grayscale image corresponding to the image. A special subdirectory is created for each signal type, and 3000 grayscale images are saved in each subdirectory to form a radar jamming signal time-frequency grayscale image set.

[0028] Based on the time-frequency grayscale image set of radar interference signals, the training set, validation set and test set are generated: 80%, 10% and 10% of the signals are randomly selected from each signal type sub-directory and combined into the training set, validation set and test set.

[0029] In this embodiment, a deep learning network is built based on the SqueezeNet network framework, and an adaptive distance estimation optimizer is used to train the deep learning network based on a time-frequency grayscale image set to obtain a trained deep learning network. The deep learning network includes: 1. Input layer: The input image size and number of channels are 227×227×3.

[0030] 2. Initial convolution layer: 7×7 convolution kernel, stride 2.

[0031] 3. Max pooling layer: 3×3 pooling window with a stride of 2.

[0032] 4. Fire module 1: Squeeze layer: 1×1 convolution kernel.

[0033] Expand layer: 1×1 convolution kernel + 3×3 convolution kernel.

[0034] 5. Fire module 2: Squeeze layer: 1×1 convolution kernel.

[0035] Expand layer: 1×1 convolution kernel + 3×3 convolution kernel.

[0036] 6. Fire module 3: Squeeze layer: 1×1 convolution kernel.

[0037] Expand layer: 1×1 convolution kernel + 3×3 convolution kernel.

[0038] 7. Max pooling layer: 3×3 pooling window with a stride of 2.

[0039] 8. Fire module 4: Squeeze layer: 1×1 convolution kernel.

[0040] Expand layer: 1×1 convolution kernel + 3×3 convolution kernel.

[0041] 9. Fire Module 5: Squeeze layer: 1×1 convolution kernel.

[0042] Expand layer: 1×1 convolution kernel + 3×3 convolution kernel.

[0043] 10. Max pooling layer: 3×3 pooling window with a stride of 2.

[0044] 11. Fire module 6: Squeeze layer: 1×1 convolution kernel.

[0045] Expand layer: 1×1 convolution kernel + 3×3 convolution kernel.

[0046] 12. Fire module 7: Squeeze layer: 1×1 convolution kernel.

[0047] Expand layer: 1×1 convolution kernel + 3×3 convolution kernel.

[0048] 13. Fire module 8: Squeeze layer: 1×1 convolution kernel.

[0049] Expand layer: 1×1 convolution kernel + 3×3 convolution kernel.

[0050] 14. Max pooling layer: 3×3 pooling window with a stride of 2.

[0051] 15. Fire module 9: Squeeze layer: 1×1 convolution kernel.

[0052] Expand layer: 1×1 convolution kernel + 3×3 convolution kernel.

[0053] 16. Final convolutional layer: used for classification tasks, with an output size of 7×7×category.

[0054] 17. Global average pooling layer: reduces the size of the feature map to a single point, and the output size is 1×1×category.

[0055] 18. Output layer: Output category prediction.

[0056] The Fire module is the core of the SqueezeNet network architecture, which processes feature maps through two steps: "Squeeze" and "Expand". The Squeeze layer uses a 1×1 convolution kernel to reduce the number of input channels, thereby reducing computational complexity; while the Expand layer restores the number of channels through 1×1 and 3×3 convolution kernels for the next step of feature extraction.

[0057] When training the deep learning network, read the images of the training set and the validation set, set the learning rate of the training to 0.001, and there are 2 epochs, each epoch has 64 iterations, a total of 128 iterations, and use the adaptive moment estimation (Adam) optimizer to iteratively update the parameters of the deep learning network. The loss function uses the cross entropy loss function to obtain the trained deep learning network. For the training effect, please refer to Figure 4,As can be seen from the figure, in the second epoch, the optimization effect basically reaches the ,expected result.

[0058] Use the trained deep learning network to classify the test set signals, and use the MATLAB built-in function confusionmat to calculate the confusion matrix and visualize the classification performance. For the results, see Figure 5 As can be seen from the figure, the trained deep learning network has a high classification and recognition accuracy for radar interference signals, which enables the radar system to accurately identify up to 12 types of interference signals and greatly improves the anti-interference capability of the radar system.

[0059] Based on the same concept, this embodiment also provides a radar interference signal type recognition device based on deep learning, which is used to implement the above-mentioned radar interference signal type recognition method based on deep learning. Figure 6 , the device comprises: An interference signal generation module is used for the server to generate a radar interference signal set, wherein the radar interference signal set includes at least pure noise interference, suppression interference, deception interference, and composite interference signals; The time-frequency analysis module is used to perform time-frequency analysis on each signal in the radar jamming signal set and generate corresponding time-frequency color images; An image processing module is used to perform image processing on each time-frequency color image to obtain a time-frequency grayscale image and generate a corresponding time-frequency grayscale image set; The model training module is used to build a deep learning network based on the SqueezeNet network framework, and to train the deep learning network using an adaptive distance estimation optimizer based on a set of time-frequency grayscale images to obtain a trained deep learning network; The interference signal recognition module is used to classify radar interference signals using a trained deep learning network, thereby realizing the identification of radar interference signal types.

[0060] The functions and implementation methods of the above modules are similar to the above-mentioned radar interference signal type identification method based on deep learning, and will not be repeated here.

[0061] The embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Even if various changes are made to the present invention, if these changes fall within the scope of the claims of the present invention and their equivalents, they still fall within the protection scope of the present invention.

Claims

1. A radar interference signal type identification method based on deep learning, characterized in that: include: The server generates a radar interference signal set, wherein the radar interference signal set includes at least pure noise interference, suppression interference, deception interference, and composite interference signals; Perform time-frequency analysis on each signal in the radar jamming signal set to generate corresponding time-frequency color images; Perform image processing on each time-frequency color image to obtain a time-frequency grayscale image and generate a corresponding time-frequency grayscale image set; A deep learning network is built based on the SqueezeNet network framework. Based on the time-frequency grayscale image set, an adaptive distance estimation optimizer is used to train the deep learning network to obtain a trained deep learning network. The trained deep learning network is used to classify radar jamming signals, thereby realizing the identification of radar jamming signal types.

2. The radar interference signal type identification method based on deep learning as claimed in claim 1, characterized in that: The radar jamming signal set includes pure noise signals, aimed jamming signals, blocked jamming signals, swept frequency jamming signals, intermittent sampling and forwarding jamming signals, distance deception jamming signals, dense false target jamming signals, smart noise jamming signals, dense false target jamming + smart noise jamming signals, distance deception jamming + smart noise jamming signals, intermittent sampling and forwarding jamming + distance deception jamming signals and intermittent sampling and forwarding jamming + swept frequency jamming signals; each type of signal data set contains time series samples of multiple signals.

3. The radar interference signal type identification method based on deep learning as claimed in claim 2, characterized in that: The aiming jammer signal satisfies the following relationship with respect to the radar transmission signal: ， The blocking interference signal satisfies the following relationship with respect to the radar transmission signal: ， The swept frequency interference signal satisfies the following relationship with respect to the radar transmission signal: ， Among them, f j is the center frequency of the interference signal, Δf j is the interference signal bandwidth, f s is the center frequency of the radar transmitting signal, Δf r is the bandwidth of the radar transmitting signal, and T is the pulse width of the radar target signal.

4. The radar interference signal type identification method based on deep learning as claimed in claim 2, characterized in that: The intermittent sampling and forwarding interference signal is formed by intermittently sampling the radar target echo signal, and the intermittent sampling and forwarding interference signal is: ， ， Where rect is the rectangular window function, τ is the rectangular signal bandwidth, δ is the impulse response function, n is an integer multiple, T s is the symbol period, and the amplitude weighting coefficient is , Represents a convolution operation.

5. The radar interference signal type identification method based on deep learning as claimed in claim 2, characterized in that: The distance deception jamming signal is formed by intercepting and storing the radar transmission signal through DRFM, copying it, and setting a delay for forwarding; the distance deception jamming signal is: A j is the interference signal amplitude, and A j >A r , A r Radar transmit signal amplitude, is the forwarding delay, is the transmission signal delay, Ø is the phase, f c is the center frequency, and K is the frequency modulation slope.

6. The method for identifying the type of radar interference signal based on deep learning according to claim 1, characterized in that: For each signal in the radar jamming signal set, smooth pseudo-Wigner-Ville distribution is used to perform time-frequency analysis and generate corresponding time-frequency color images.

7. The method for identifying the type of radar interference signal based on deep learning according to claim 6, characterized in that: Each time-frequency color image is downsampled and converted into a grayscale image to generate a set of time-frequency grayscale images of radar jamming signals.

8. A radar interference signal type identification device based on deep learning, characterized in that: include: An interference signal generation module is used for the server to generate a radar interference signal set, wherein the radar interference signal set includes at least pure noise interference, suppression interference, deception interference, and composite interference signals; The time-frequency analysis module is used to perform time-frequency analysis on each signal in the radar jamming signal set and generate corresponding time-frequency color images; An image processing module is used to perform image processing on each time-frequency color image to obtain a time-frequency grayscale image and generate a corresponding time-frequency grayscale image set; The model training module is used to build a deep learning network based on the SqueezeNet network framework, and to train the deep learning network using an adaptive distance estimation optimizer based on a set of time-frequency grayscale images to obtain a trained deep learning network; The interference signal recognition module is used to classify radar interference signals using a trained deep learning network, thereby realizing the identification of radar interference signal types.

Images