Radar interference signal identification method and system based on deep fusion neural network

Abstract

The invention provides a radar interference signal identification method and system based on a deep fusion neural network, and the method comprises the steps: obtaining an echo signal of a radar, carrying out the generalized S transformation of the echo signal to generate a time-frequency image, carrying out the preprocessing of the time-frequency image, and obtaining a preprocessed time-frequency image; and identifying the preprocessed time-frequency image by using a pre-trained deep fusion neural network model, and outputting an interference signal type. According to the method, the time-frequency image of the radar interference signal with high time-frequency resolution is obtained by adopting generalized S transformation, the time-frequency domain characteristics of the radar interference signal can be flexibly described, and the quality of a learning sample is ensured.

Description

Technical Field

[0001] This invention relates to the field of radar jamming signal identification technology, and in particular to a radar jamming signal identification method based on a deep fusion neural network. Background Technology

[0002] Radar plays a crucial role in the information battlefield; however, the continuous emergence of various new radar jamming technologies seriously threatens the operational effectiveness of radar. Accurate identification of radar jamming signals can provide decision support for friendly electronic defense and enemy attacks, and the level of such identification technology has become an important indicator of the sophistication of electronic warfare equipment.

[0003] Feature extraction-based radar jamming signal identification is currently the most widely used method. This method starts from the generation mechanism of different types of jamming signals, transforms the jamming signals from the time domain to other domains, and then uses professional analysts or classification models to compare the characteristic differences of different types of jamming signals, thereby accurately identifying various types of jamming signals.

[0004] The time-frequency domain provides joint distribution information in the time and frequency domains, clearly describing the relationship between signal frequency and time. Time-frequency analysis of a signal yields a time-frequency image of the signal's energy distribution on both time and frequency scales, maximizing the visualization of the signal's transient characteristics and making it a powerful tool for processing radar jamming signals. However, manually extracting time-frequency image features of radar jamming signals requires analysts to possess solid professional skills and expend considerable effort summarizing and comparing the time-frequency images of jamming signals. The entire process is characterized by high workload, tediousness, time consumption, low recognition accuracy, and the recognition results are easily influenced by subjective experience factors, resulting in poor reliability.

[0005] With the rapid development of classification and recognition technologies such as pattern recognition and machine learning, especially the resurgence of deep learning technology in recent years, more inspiration has been provided for current radar jamming signal identification technology. Improving deep learning models from the field of artificial intelligence and introducing them into the field of radar signal jamming identification will significantly improve both the effectiveness and speed of jamming signal identification.

[0006] Compared to manual feature extraction methods, deep learning technology can automatically learn from time-frequency image data without human intervention. By incorporating prior knowledge and adjusting the model's hierarchical structure and hyperparameters, it can excel in various classification tasks. Therefore, by using time-frequency transformation methods from the field of signal analysis to convert one-dimensional radar jamming signals into two-dimensional time-frequency images, and leveraging deep learning models from the field of artificial intelligence to automatically extract subtle information from these images, along with mature image feature extraction techniques for identifying jamming patterns, radar jamming signal classification tasks can be completed efficiently and accurately.

[0007] However, existing technologies typically employ short-time Fourier transform to convert time-domain signals into time-frequency domain signals, which does not allow for adjustment of the window function length. This results in poor quality of learning samples, complex model structures with numerous hyperparameters, high computational cost, and low robustness, leading to poor recognition performance under low noise-to-interference ratio conditions. Summary of the Invention

[0008] To address the shortcomings of existing technologies that utilize deep learning models for radar interference signal identification, the first objective of this invention is to provide a radar interference signal identification method based on a deep fusion neural network, comprising: acquiring radar echo signals; performing a generalized S-transform on the echo signals to generate a time-frequency image; preprocessing the time-frequency image to obtain a preprocessed time-frequency image; using a pre-trained deep fusion neural network model to identify the preprocessed time-frequency image; and outputting the interference signal type.

[0009] Specifically, the deep fusion neural network model is obtained by training on the preprocessed time-frequency image of the radar jamming signal; the radar jamming signal is obtained by applying active jamming modulation to the radar signal and generating Gaussian white noise superimposed on the jamming signal; the preprocessed time-frequency image is obtained by sequentially performing generalized S-transform and image preprocessing on the radar jamming signal.

[0010] Specifically, the active interference signals include noise amplitude modulation, noise frequency modulation, slice reconstruction, intermittent sampling and forwarding, comb spectrum, and noise convolution.

[0011] Specifically, the preprocessing includes grayscale conversion, filtering and noise reduction, and image scaling.

[0012] The grayscale conversion process involves converting a color time-frequency image into a grayscale image using a weighted average method.

[0013] The filtering and noise reduction process involves removing Gaussian white noise from the time-frequency image using a two-dimensional Wiener filter.

[0014] The image scaling process involves adjusting the size of the time-frequency image using bilinear interpolation.

[0015] Specifically, the deep fusion neural network model includes an input layer, a first convolutional layer, a first Inception structure, a first attention mechanism, a first global average pooling layer, a second convolutional layer, a second Inception structure, a second attention mechanism, a second global average pooling layer, a third convolutional layer, a third Inception structure, a third attention mechanism, a fully connected layer, and a softmax module, all connected in sequence. The activation function of each layer is the ReLU function.

[0016] Specifically, the hyperparameters for initializing the deep fusion neural network model are as follows: the number of convolutional kernels in the convolutional layer is set to 6, the size is 7×7, and the step size for each movement is 1.

[0017] The Inception architecture has 3 branches, and each branch has 4 convolutional kernels with sizes of 1×1, 3×3, and 5×5, respectively.

[0018] The regularization parameter of the attention mechanism is set to 1.5, and the attention weight is 0.6.

[0019] The global average pooling layer has 8 convolutional kernels with a size of 2×2;

[0020] The fully connected layer has 120 neurons;

[0021] The initial learning rate for model training is 0.01, the maximum number of training rounds is 5, and the number of iterations is 100.

[0022] Specifically, the training method for the deep fusion neural network model is as follows:

[0023] Active interference modulation is applied to the transmitted radar signal to obtain an interference signal, and Gaussian white noise is superimposed on the interference signal to generate a radar interference signal.

[0024] A time-frequency image is obtained by performing a generalized S-transform on the radar jamming signal;

[0025] The time-frequency images are preprocessed, and the preprocessed images are divided into training set and validation set according to a preset ratio;

[0026] A deep fusion neural network model is trained using the cross-entropy loss function. The connection weights between each module of the deep fusion neural network model are iteratively updated using the adaptive moment estimation optimization algorithm until the cross-entropy loss function converges.

[0027] The training results are output using a confusion matrix.

[0028] Specifically, the radar interference signal uses the Monte Carlo method to randomly select multiple sets of parameters within a preset parameter value range, and obtains interference samples at equal intervals within a preset interference-to-noise ratio range.

[0029] The second objective of this invention is to provide a radar jamming signal identification system based on a deep fusion neural network, which uses the aforementioned radar jamming signal identification method based on a deep fusion neural network, including:

[0030] The radar signal acquisition module acquires echo signals that include radar interference signals.

[0031] The signal processing module performs a generalized S-transform on the acquired echo signal to generate a time-frequency image;

[0032] The image processing module preprocesses the time-frequency image to obtain a preprocessed time-frequency image;

[0033] The image recognition module uses a pre-trained deep fusion neural network model to recognize the preprocessed time-frequency image and determine the type of interference signal;

[0034] The output module outputs the judgment result.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] 1. This invention uses generalized S-transform to obtain high time-frequency resolution radar interference signal time-frequency images, which not only overcomes the short-time Fourier transform's inability to adjust the window function length, but also has the advantage of wavelet transform's ability to perform multi-resolution analysis on non-stationary signals. It can flexibly characterize the time-frequency domain features of radar interference signals and ensure the quality of learning samples.

[0037] 2. By interspersing multiple attention mechanism modules between the convolutional layer and the Inception structure, this invention enables the model to focus on key regions in the time-frequency image, automatically ignore redundant information in the time-frequency image, avoid the loss of effective features, and improve the feature extraction capability of the model.

[0038] 3. This invention selects a learning strategy that combines the cross-entropy loss function and the adaptive moment estimation optimization operator, which not only reduces the computational cost of the model, but also avoids overfitting and gradient vanishing phenomena, thereby enhancing the model's generalization ability. Attached Figure Description

[0039] Figure 1 This is a flowchart of the present invention;

[0040] Figure 2 This is a flowchart of the model construction process for the present invention;

[0041] Figure 3 The time-frequency image of amplitude-modulated noise interference under a noise-to-interference ratio of 6dB;

[0042] Figure 4 The time-frequency image of the noise frequency modulation interference under a noise-to-interference ratio of 6dB;

[0043] Figure 5 The time-frequency image of the interference reconstructed from slices under a noise-to-interference ratio of 6dB;

[0044] Figure 6 This is a time-frequency image of intermittent sampling and forwarding interference under a noise-to-interference ratio of 6dB;

[0045] Figure 7 The time-frequency image of comb-spectral interference under a noise-to-interference ratio of 6dB;

[0046] Figure 8 This is a time-frequency image of noise convolution interference under a noise-to-interference ratio of 6dB.

[0047] Figure 9 This is a diagram of the deep fusion neural network model architecture of the present invention;

[0048] Figure 10 This is the confusion matrix of the recognition results of the deep fusion neural network model of this invention on the training set;

[0049] Figure 11 This is the confusion matrix of the recognition results of the deep fusion neural network model of this invention on the test set. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. The application principles of this invention will be further described below with reference to the accompanying drawings.

[0051] This invention provides a radar interference signal identification method and system based on a deep fusion neural network. The radar interference signal identification method based on the neural network is as follows: acquiring radar echo signals, performing a generalized S-transform on the echo signals to generate a time-frequency image, preprocessing the time-frequency image to obtain a preprocessed time-frequency image, using a pre-trained deep fusion neural network model to identify the preprocessed time-frequency image, and outputting the interference signal type.

[0052] like Figure 2 As shown, the method for constructing this pre-trained deep fusion neural network model is as follows:

[0053] The first step is to generate simulated radar jamming signals.

[0054] Pulse compression radar not only increases the radar's effective range but also improves range resolution, effectively resolving the contradiction between range and range resolution, thus significantly enhancing radar performance. Linear frequency modulated (LFM) signals have advantages such as high range resolution, wide bandwidth, and strong noise immunity, making them the most commonly used form of pulse compression radar transmission signal. Assuming the radar transmits a linear frequency modulated signal s(t), its expression is:

[0055] ,

[0056] In the formula, Let B be the carrier frequency, T be the pulse repetition period, t be the time, B be the bandwidth, τ be the pulse width, and rect(∙) be the rectangular function.

[0057] Since obtaining real radar jamming signals is difficult and model training requires a large number of training samples, Matlab simulation was used to apply six types of active jamming to the transmitted radar signals: amplitude noise modulation (AMJ), frequency noise modulation (FMJ), slice reconstruction (CIJ), intermittent sampling and forwarding (ISRJ), comb spectrum (CSMJ), and noise convolution (NCJ). A certain degree of Gaussian white noise was superimposed on the simulated jamming signal to simulate the real complex electromagnetic environment.

[0058] In this embodiment, the pulse width of the radar transmitted signal is set to 20 μs, the bandwidth to 30 MHz, the carrier frequency to 20 MHz, and the sampling frequency to 100 MHz. Complete learning samples are constructed by changing the relevant parameters for different interference types. The selection range of relevant parameters for various interference signals is shown in the interference signal parameter value range table. Using the Monte Carlo method, 10 sets of parameters are randomly selected from the range in the table. Simulations are performed at 2 dB intervals under interference-to-noise ratios (IRRs) ranging from -2 to 8 dB to obtain 10 interference samples, thus yielding 600 learning samples for each type of interference signal.

[0059] Interference signal parameter value range table

[0060]

[0061] The second step is to acquire time-frequency images of the simulated radar interference signal.

[0062] Compared to individual time-domain or frequency-domain features, time-frequency domain features encompass the distribution of signal energy in time and frequency space, providing a more comprehensive reflection of the information contained within the signal. The Generalized S-Transform (GST) uses the product of a simple harmonic wave and a Gaussian function as the basic wavelet. By introducing two parameters to adjust the window shape, it performs scaling and translation transformations on the basic wavelet, allowing the window width to adaptively change with the frequency of the radar interference signal. This overcomes the limitation of the Short-Time Fourier Transform in adjusting the window length while retaining the advantage of wavelet transform for multi-resolution analysis of non-stationary signals. It can flexibly characterize the time-frequency domain features of radar interference signals, ensuring the quality of the learning samples.

[0063] For radar jamming signal s(t), the mathematical expression for its generalized S-transform is:

[0064] ,

[0065] In the formula, Let f be the center point of the window function, f be the sampling frequency, λ and p be adjustment factors for the window shape, and t be time.

[0066] The time-frequency images of different types of interference signals after generalized S-transform are shown in the figure, with the horizontal axis representing signal duration, the vertical axis representing signal frequency, and the color representing signal amplitude, all under a noise ratio of 6dB. Figure 3-Figure 8 As shown, the time-frequency image obtained after the generalized S-transform clearly displays the energy distribution characteristics of the interference signal in the time-frequency domain, reflects the transient change law of the interference signal, has good time-frequency focusing, and can provide high-quality time-frequency images for subsequent model training.

[0067] The third step is to preprocess the time-frequency images and divide the dataset.

[0068] To reduce the training cost of the model, the time-frequency images are preprocessed to obtain a preprocessed time-frequency image set. The preprocessing process includes grayscale conversion, filtering and noise reduction, and image scaling. The grayscale conversion process uses a weighted average method to convert the color time-frequency images into grayscale images to reduce the memory footprint of the dataset; the filtering and noise reduction process uses a two-dimensional Wiener filter to remove Gaussian white noise from the time-frequency images, improving the resolution of the time-frequency images; the image scaling process uses bilinear interpolation to adjust the size of the time-frequency images to 100×100 to speed up the model's loading speed.

[0069] Subsequently, the time-frequency image set of radar jamming signals was reordered using a random sampling method, and the disordered time-frequency image set was divided into a training set and a validation set according to a preset ratio. In one embodiment of the invention, the disordered time-frequency image set was divided into a training set and a validation set at a ratio of 5:1, with 500 training samples and 100 validation samples for each type of jamming signal. Although the samples in the training set and the validation set have the same jamming pattern, due to the different simulation parameters, we can still significantly improve the completeness of the training sample set and ensure the generalization ability of the model.

[0070] The fourth step is to build a deep fusion neural network model that incorporates the attention mechanism.

[0071] The Integrating Attention Mechanism Deep Network (AMDN) model constructed in this invention interweaves multiple attention mechanism modules between convolutional layers and Inception structures. It includes, in sequence, an input layer, a first convolutional layer, a first Inception structure, a first attention mechanism, a first global average pooling layer, a second convolutional layer, a second Inception structure, a second attention mechanism, a second global average pooling layer, a third convolutional layer, a third Inception structure, a third attention mechanism, a fully connected layer, and a softmax module. It consists of 3 convolutional layers, 3 Inception structures, 3 attention mechanism modules, 2 global average pooling layers, and 1 fully connected layer. The entire model architecture is as follows: Figure 9 As shown.

[0072] The convolutional layer perceives global features of the time-frequency image through sliding convolutional kernels, and the receptive field of each channel has spatial invariance during perception, which helps the model extract weak features such as edges, textures, and corners in the time-frequency image. The Inception structure can effectively extract features at different scales by paralleling convolutional kernels of different sizes, and uses multi-branch channels to fuse features at different scales to form a richer feature space, which helps the model better handle time-frequency images of different sizes and resolutions. The attention mechanism adjusts the model's attention to different input features, allowing the model to focus on key regions in the time-frequency image, automatically ignoring redundant information in the time-frequency image, avoiding the loss of effective features, and helping the model extract a more discriminative feature set, thereby improving the model's classification accuracy. The global average pooling layer can better fuse image features by averaging the entire time-frequency image, reducing the model's computational load, and avoiding overfitting and gradient vanishing phenomena. The fully connected layer can convert the output of the two-dimensional feature map into a one-dimensional vector, enhancing the consistency between time-frequency image features and categories, and completing the classification of radar interference signals through the Softmax classifier.

[0073] The fifth step is to set the parameters of the network model and train the model using the dataset.

[0074] Initialize the hyperparameters of the AMDN model: set the number of convolutional kernels in the convolutional layers to 6, with a size of 7×7 and a stride of 1; set the Inception structure to have 3 branches, each with 4 convolutional kernels of sizes 1×1, 3×3, and 5×5; set the regularization parameter of the attention mechanism to 1.5 and the attention weight to 0.6; set the number of convolutional kernels in the global average pooling layer to 8, with a size of 2×2; and set the number of neurons in the fully connected layer to 120. Select ReLU as the activation function for each layer, and set the initial learning rate to 0.01, the maximum number of training epochs to 5, and the number of iterations to 100.

[0075] The AMDN model is trained using the cross-entropy loss function, which is formulated as follows:

[0076] ,

[0077] In the formula, y is the actual label value. is the model prediction value, and k is the number of modulation signal categories.

[0078] The adaptive moment estimation optimization algorithm is used to iteratively update the connection weights between each module of the model until the cross-entropy loss function converges, thus obtaining the trained AMDN model. Choosing a learning strategy that combines the cross-entropy loss function and the adaptive moment estimation optimization operator not only reduces the computational cost of the model but also avoids overfitting and gradient vanishing phenomena, enhancing the model's generalization ability.

[0079] The model's recognition results are presented through a confusion matrix. This matrix allows observation of misclassifications due to various types of interference, enabling the network's training parameters to be adjusted appropriately to meet prediction accuracy requirements. For example... Figure 10 and Figure 11 As shown, the confusion matrix displays the recognition results of the model on the training and validation sets. Black represents the recognition accuracy, and gray represents the recognition error rate; the darker the color, the higher the accuracy or error rate. Observing the recognition results in the matrix, it can be seen that the model after parameter updates achieved high recognition accuracy on both the training and test sets, with an overall accuracy exceeding 95%, proving that the trained model has good classification performance.

[0080] This invention also provides a radar interference signal identification system based on a deep fusion neural network, comprising: a radar signal acquisition module for acquiring echo signals containing radar interference signals; a signal processing module for performing a generalized S-transform on the acquired echo signals to generate a time-frequency image; an image processing module for preprocessing the time-frequency image to obtain a preprocessed time-frequency image; an image recognition module for using a pre-trained deep fusion neural network model to recognize the preprocessed time-frequency image and determine the type of interference signal; and an output module for outputting the judgment result.

[0081] This invention deeply analyzes the properties and characteristics of radar jamming signals, simulates and generates waveform data of various jamming methods, and then uses the generalized S-transform to obtain high-resolution time-frequency images of radar jamming signals. The generated time-frequency images are further improved through preprocessing operations such as grayscale conversion, filtering and noise reduction, and image scaling. Finally, multiple attention mechanism modules are interspersed between convolutional layers and the Inception structure to form a deep network model that integrates attention mechanisms. A learning strategy combining cross-entropy loss function and adaptive moment estimation optimization operator is selected to capture local information and perceive global features of the radar jamming signal time-frequency image, thereby achieving intelligent recognition of radar jamming signals.

[0082] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A radar jamming signal identification method based on deep fusion neural networks, characterized in that, include: The radar echo signal is acquired, a generalized S-transform is performed on the echo signal to generate a time-frequency image, the time-frequency image is preprocessed to obtain a preprocessed time-frequency image, a pre-trained deep fusion neural network model is used to identify the preprocessed time-frequency image, and the type of interference signal is output.

2. The radar jamming signal identification method based on deep fusion neural network according to claim 1, characterized in that, The deep fusion neural network model is obtained by training on the preprocessed time-frequency image of the radar jamming signal; the radar jamming signal is obtained by applying active jamming modulation to the radar signal and generating Gaussian white noise superimposed on the jamming signal; the preprocessed time-frequency image is obtained by sequentially performing generalized S-transform and image preprocessing on the radar jamming signal.

3. The radar jamming signal identification method based on deep fusion neural network according to claim 2, characterized in that, The active interference signals include noise amplitude modulation, noise frequency modulation, slice reconstruction, intermittent sampling and forwarding, comb spectrum, and noise convolution.

4. The radar jamming signal identification method based on deep fusion neural network according to claim 1 or 2, characterized in that, The preprocessing includes grayscale conversion, filtering and noise reduction, and image scaling. The grayscale conversion process involves converting a color time-frequency image into a grayscale image using a weighted average method. The filtering and noise reduction process involves removing Gaussian white noise from the time-frequency image using a two-dimensional Wiener filter. The image scaling process involves adjusting the size of the time-frequency image using bilinear interpolation.

5. The radar jamming signal identification method based on deep fusion neural network according to claim 1, characterized in that, The deep fusion neural network model includes, in sequence, an input layer, a first convolutional layer, a first Inception structure, a first attention mechanism, a first global average pooling layer, a second convolutional layer, a second Inception structure, a second attention mechanism, a second global average pooling layer, a third convolutional layer, a third Inception structure, a third attention mechanism, a fully connected layer, and a softmax module, wherein the activation function of each layer is the ReLU function.

6. The radar jamming signal identification method based on deep fusion neural network according to claim 5, characterized in that, The hyperparameters for initializing the deep fusion neural network model are as follows: the number of convolutional kernels in the convolutional layer is set to 6, the size is 7×7, and the step size for each movement is 1. The Inception architecture has 3 branches, and each branch has 4 convolutional kernels with sizes of 1×1, 3×3, and 5×5, respectively. The regularization parameter of the attention mechanism is set to 1.5, and the attention weight is 0.

6. The global average pooling layer has 8 convolutional kernels with a size of 2×2; The fully connected layer has 120 neurons; The initial learning rate for model training is 0.01, the maximum number of training rounds is 5, and the number of iterations is 100.

7. The radar jamming signal identification method based on deep fusion neural network according to claim 1, characterized in that, The training method for the deep fusion neural network model is as follows: Active interference modulation is applied to the transmitted radar signal to obtain an interference signal, and Gaussian white noise is superimposed on the interference signal to generate a radar interference signal. A time-frequency image is obtained by performing a generalized S-transform on the radar jamming signal; The time-frequency images are preprocessed, and the preprocessed images are divided into training set and validation set according to a preset ratio; A deep fusion neural network model is trained using the cross-entropy loss function. The connection weights between each module of the deep fusion neural network model are iteratively updated using the adaptive moment estimation optimization algorithm until the cross-entropy loss function converges. The training results are output using a confusion matrix.

8. The radar jamming signal identification method based on deep fusion neural network according to claim 7, characterized in that, The radar interference signal is obtained by randomly selecting multiple sets of parameters within a preset parameter value range using the Monte Carlo method, and acquiring interference samples at equal intervals within a preset interference-to-noise ratio range.

9. A radar jamming signal identification system based on a deep fusion neural network, using the radar jamming signal identification method based on a deep fusion neural network as described in any one of claims 1-8, characterized in that, include: The radar signal acquisition module acquires echo signals that include radar interference signals. The signal processing module performs a generalized S-transform on the acquired echo signal to generate a time-frequency image; The image processing module preprocesses the time-frequency image to obtain a preprocessed time-frequency image; The image recognition module uses a pre-trained deep fusion neural network model to recognize the preprocessed time-frequency image and determine the type of interference signal; The output module outputs the judgment result.

Images