SAR (Synthetic Aperture Radar) active interference identification method and system storage device under low interference-to-signal ratio condition and electronic equipment

Abstract

The invention discloses an SAR (Synthetic Aperture Radar) active interference identification method under a low interference-to-signal ratio condition, a system storage device and electronic equipment. The method comprises the following steps: carrying out normalization processing on a time domain image of interference, and weakening a sidelobe influence through power law transformation after carrying out normalization processing on a time-frequency domain image; taking the preprocessed time-domain image and the time-frequency-domain image as dual-channel input of the network, and extracting multi-scale features by using a diversified branch block and a channel attention mechanism; fusing the extracted time domain features and the extracted time-frequency domain features through a weighted fusion strategy, carrying out dynamic weight distribution on the features of the time domain and the time-frequency domain, and introducing variance parameters of signals and interference to carry out feature optimization under a low interference-to-signal ratio; and inputting the fused feature sequence into a Transform, calculating the attention weight of any position through a Query-Key-Value matrix, capturing a long-distance dependency relationship, and enhancing global feature representation. According to the method, the SAR active interference pattern under the low interference-to-signal ratio can be identified with high precision, and powerful support is provided for a subsequent anti-interference strategy.

Description

Technical Field

[0001] The present application relates to the field of signal processing technology, and in particular to a method, system, computer-readable storage medium, and electronic device for identifying SAR active interference under low signal-to-interference ratio conditions. Background Art

[0002] Synthetic Aperture Radar (SAR), a high-resolution microwave imaging device, holds irreplaceable strategic value in military target identification, battlefield situational awareness, and remote sensing monitoring, thanks to its all-day, all-weather capability and superior penetrating detection of concealed targets. However, with the rapid development of electronic countermeasures, SAR in complex electromagnetic environments faces an increasingly severe interference threat. From suppressive noise interference to high-fidelity deceptive jamming, and from single-system, single-modal jamming to multi-source, composite jamming, the complexity, power dynamic range, and intelligence of jamming signals are increasing exponentially, leading to severe degradation of SAR image quality and even complete failure.

[0003] In the signal reception and processing process of SAR systems, the superposition and mixing of interference signals with target echoes and ground clutter easily overwhelms interference features under low jamming-to-signal ratio (JSR) conditions. Traditional recognition methods based on artificial feature design (such as time-frequency statistical features and polarization features) are insensitive to subtle feature differences and their generalization capabilities are significantly reduced in complex modulation interference scenarios. Although deep learning technology has improved recognition accuracy through end-to-end feature learning, existing methods still face multiple challenges: the weak time-frequency domain energy of interference signals under low JSR leads to insufficient feature extraction in deep models; the overlapping and coupled time-frequency features of multi-source composite interference make feature decoupling and discriminative representation difficult; and the time-varying characteristics of interference parameters (such as frequency agility and polarization modulation) further exacerbate the difficulty of model generalization.

[0004] Interference identification, as a core component of SAR active interference mitigation, directly determines the effectiveness of its anti-interference strategies. Accurately distinguishing between suppressive and deceptive interference allows for targeted filtering suppression or waveform agility. Real-time identification of the components of complex interference allows for dynamic adjustment of signal processing to maximize target information retention. However, current research lacks systematic breakthroughs in feature enhancement mechanisms, multimodal feature fusion strategies, and robust representation methods for complex interference in low JSR scenarios. This limits the engineering applicability of existing algorithms in complex electromagnetic environments.

[0005] In summary, in the face of the new interference threats of "low detectability, high camouflage, and strong confrontation" in modern electronic countermeasures, it is necessary to study the SAR active interference identification method suitable for low JSR conditions, break through the limitations of traditional feature extraction, and construct an intelligent recognition model with both high robustness and strong generalization ability. This is not only a key technical path to improve the reliable operation of SAR systems in complex environments, but also an important support for realizing active perception of electromagnetic situation and enhancing the battlefield survivability of equipment. It has important theoretical significance and engineering value for promoting the practical application of SAR technology in strong confrontation scenarios. Summary of the Invention

[0006] The purpose of this application is to provide a SAR active interference identification method, a system storage device and an electronic device under low interference-to-signal ratio conditions to solve or alleviate the problems existing in the above-mentioned prior art.

[0007] In order to achieve the above objectives, this application provides the following technical solutions:

[0008] The present application provides a SAR active interference identification method under low interference-to-signal ratio conditions, comprising: step S101, simulating a SAR active interference signal; step S102, extracting a time domain image and a time-frequency domain image of the interference; step S103, normalizing the time domain image, and then performing a power law transform on the time-frequency domain image after normalization; step S104, dividing a SAR active interference data set into a training set and a test set; step S105, inputting the training set into a dual-channel feature extraction network for training; step S106, inputting the test set into the trained model to obtain SAR active interference identification results under different interference-to-signal ratios;

[0009] Preferably, in step S101, nine types of SAR active interference signals are simulated, specifically:

[0010] Simulations were conducted for six single interference patterns: noise FM interference, noise AM interference, intermittent sampling interference, frequency shift interference, noise convolution interference, and noise product interference; and three composite interference patterns: intermittent sampling + frequency shift interference, noise product + frequency shift interference, and noise product + intermittent sampling interference. The interference-to-signal ratio ranges for noise FM and noise AM were 0dB, 5dB, and 10dB, while the interference-to-signal ratio ranges for intermittent sampling interference, frequency shift interference, intermittent sampling + frequency shift interference, noise product + frequency shift interference, and noise product + intermittent sampling interference were -10dB, -5dB, and 0dB, respectively. The interference-to-signal ratio ranges for noise convolution interference and noise product interference were -5dB, 0dB, and 5dB, respectively.

[0011] Preferably, in step S102, the time domain image and the time-frequency domain image of the interference are extracted, specifically:

[0012] The nine types of SAR active interference signals are processed to generate the time domain image of the interference, and the interference is subjected to short-time Fourier transform to obtain the time-frequency image of the interference signal. The formula of short-time Fourier transform is:

[0013]

[0014] τ is the time delay, G(τ,f) is the time-frequency signal obtained after short-time Fourier transform, and f is the frequency.

[0015] Preferably, in step S103, the time domain image is normalized and the time-frequency domain image is subjected to power law transformation to obtain processed image data, specifically:

[0016] Normalize the time domain signal and the time-frequency domain signal to standardize the pixel value range between different pixels. The formula is:

[0017]

[0018] Among them, x is the pixel value of the original time domain image, x p is the normalized pixel value, x min is the minimum value of the original image pixel value, x max is the maximum value of the original image pixel value. Then the normalized time-frequency image is subjected to power law transformation to enhance the image contrast and weaken the sidelobe effect. The formula is:

[0019] S=c×x p r

[0020] Among them, S is the time-frequency image after power law transformation, c is the normalization constant, and r is the power law parameter.

[0021] Preferably, in step S105, the training set is input into the dual-channel feature extraction network for training, specifically:

[0022] The processed time domain and time-frequency domain images are used as the two inputs of the dual-channel feature extraction network. Diverse branch block (DBB) is used in both channels to replace the traditional convolution layer. Parallel convolution operations (with different kernel sizes, this method uses four different parallel convolution methods: 1x1, 1x1~kxk, 1x1~avgpool, kxk)) are used to capture multi-scale local details and improve the model's adaptability to complex features. The squeeze-and-excitation (SE) channel attention mechanism is used to dynamically adjust the contribution of each channel, enhance key features and suppress redundant information. In the squeeze stage, global average pooling is used to convert the input feature map X∈R C×H×W Mapped into a channel vector Z∈R C , the formula is:

[0023]

[0024] Among them, H represents the feature map height, W represents the feature map width, and Z C Indicates the description value of the Cth channel, x c (i, j) represents the eigenvalue of the Cth channel at (i, j) in space. Then, in the excitation stage, the importance weight of each channel is learned through the fully connected layer and the Sigmoid activation function. The formula is:

[0025] S=σ(W2δ(W1Z))

[0026] in, and is the weight matrix, t is the reduction rate for reducing complexity, δ represents the ReLU activation function, and σ represents the Sigmoid activation function, which normalizes the weights to the interval [0,1]. Subsequently, the learned channel weights are multiplied by the original feature map X by channel to generate the weighted feature map This process dynamically adjusts the contribution of each channel, thereby enhancing key features and suppressing minor ones.

[0027] After using DBB and SE attention mechanisms to extract local features in the time domain and time-frequency domain, a weighted fusion strategy is used to integrate dual-channel features. The time domain input feature matrix is ​​set as The input feature matrix in the time-frequency domain is N is the sample size, D t and D tf

[0028] are the feature dimensions of time domain and time-frequency domain respectively, then the dynamic weight allocation mechanism is as follows:

[0029]

[0030] Among them, H(F t ) is information entropy, α is temperature parameter, ω tf =1-ω t , ensuring that the weight sum is 1 to avoid energy imbalance. In order to improve the model’s sensitivity to interference features under low JSR, a feature weight optimization term δ is designed for low JSR scenarios, which is expressed as:

[0031]

[0032] where σ s 2 is the variance of the target echo signal, σ j 2 is the variance of the SAR interference signal, σ n2 is the modal independent variance of the ambient noise.

[0033] The fused features not only contain rich local detail information, but also extract global features through the Transformer module, thereby enhancing the global context information. Suppose the input sequence is X = (x1, x2, x3, ... x n ), where x i Represents the i-th input vector. The multi-head attention mechanism can effectively capture long-distance dependencies, and its expression is:

[0034] MultiHead(Q,K,V)=Concat(head1,head2,...head h )W o

[0035] Where Q, K, and V represent query, key, and value matrices, respectively, obtained by applying different linear transformations to the input X; W o Represents the linear transformation matrix of the multi-head attention mechanism output, head i

[0036] The self-attention output corresponding to the i-th head can be expressed as:

[0037]

[0038] Attention refers to the self-attention mechanism, which is formally expressed as:

[0039]

[0040] where d k Indicates the dimension of the key vector.

[0041] After training is completed, save the trained model.

[0042] Preferably, in step S106, according to the trained model obtained in step S105, the test set is input to obtain the SAR active interference recognition result.

[0043] The embodiment of the present application also provides a SAR active interference identification system under low interference-to-signal ratio conditions, including: a data preprocessing unit, including time domain image normalization operations for six types of single interference patterns: noise frequency modulation interference, noise amplitude modulation interference, intermittent sampling interference, frequency shift interference, noise convolution interference, and noise product interference, and three types of composite interference patterns: intermittent sampling + frequency shift composite interference, noise product + frequency shift composite interference, and noise product + intermittent sampling composite interference. After the time-frequency domain image normalization, a power law transformation is performed to improve the image contrast; a model construction unit, configured as a dual-channel feature extraction network, in which each channel replaces the traditional convolution layer with DBB to improve the feature extraction capability, and adds an SE attention mechanism to dynamically adjust the contribution of each channel, thereby enhancing key features and suppressing secondary features. Dynamic feature fusion is performed on the extracted time domain and time-frequency domain features through a dynamic weight strategy, and global context modeling is performed using Transformer after fusion; a model testing unit, configured to input the test set into the dual-channel feature extraction network for pattern recognition based on the optimal network model and parameters to obtain SAR active interference pattern recognition results.

[0044] Embodiments of the present application provide a non-transitory computer-readable storage medium storing an executable program configured to implement any of the above-described methods for identifying active interference with SAR signals under low signal-to-interference ratio conditions. The program can be read and executed by a processor to implement a complete process including data preprocessing, feature extraction, model training, and interference identification.

[0045] Correspondingly, the present application also provides an electronic device comprising a processor and a memory. The executable program stored in the memory, when the processor is running, can execute all the steps of the above-mentioned SAR active interference identification method under low interference signal ratio conditions, specifically including: extracting time domain and frequency domain features of SAR signals, completing feature extraction and pattern classification through a dual-channel neural network, and outputting high-precision interference identification results.

[0046] Beneficial effects:

[0047] The present application first simulates the SAR active interference signal; extracts the original time domain image and time-frequency domain image of each type of interference to obtain the original image data that has not been preprocessed; normalizes the time domain image; normalizes the time-frequency domain image first, and then further enhances the feature details through power law transformation, and finally obtains the preprocessed interference data; divides the preprocessed SAR active interference data set into a training set and a test set in proportion, which are used for network training and testing respectively; inputs the training set into the dual-channel feature extraction network for training, obtains the trained network model through iterative optimization and completes parameter preservation; inputs the test set into the trained model, outputs the SAR active interference recognition results under different interference-to-signal ratio (JSR) conditions, and evaluates the generalization ability of the model in complex electromagnetic environments. Through the present invention, SAR active interference patterns can be identified with high precision, which is conducive to the smooth implementation of subsequent anti-interference strategies, and has important theoretical significance and engineering value for promoting the practical application of SAR technology in strong confrontation scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings and descriptions that constitute part of this application are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. Among them:

[0049] Figure 1 It is a schematic diagram of the process of the present invention;

[0050] Figure 2 The time domain and time-frequency diagrams of the nine types of SAR active interference signals described in the present invention;

[0051] Figure 3 This is a diagram of the SAR active interference identification network structure of the present invention;

[0052] Figure 4 This is a schematic diagram of the overall recognition accuracy of nine types of SAR active jammers under different interference-to-signal ratios according to the present invention. DETAILED DESCRIPTION

[0053] This application will be described in detail with reference to the accompanying drawings and specific embodiments. The examples provided are intended to illustrate rather than limit the scope of protection of this application. Those skilled in the art should understand that various modifications and variations may be made to this application without departing from the essence and scope of protection of the technical solution of this application. For example, the features or structures disclosed in a certain embodiment may be combined with other embodiments to form a new technical solution. Therefore, this application covers all modifications, improvements and equivalent substitutions that are consistent with the inventive concept, which should all fall within the scope of protection of the attached claims and their equivalent technical solutions.

[0054] Exemplary Methods

[0055] like Figure 1 As shown, a SAR active jammer identification method under low interference-to-signal ratio conditions includes:

[0056] Step S101: Simulate nine types of SAR active interference signals, specifically:

[0057] Simulations were conducted for six single interference patterns: noise FM interference, noise AM interference, intermittent sampling interference, frequency shift interference, noise convolution interference, and noise product interference; and three composite interference patterns: intermittent sampling + frequency shift interference, noise product + frequency shift interference, and noise product + intermittent sampling interference. The interference-to-signal ratio ranges for noise FM and noise AM were 0dB, 5dB, and 10dB, while the interference-to-signal ratio ranges for intermittent sampling interference, frequency shift interference, intermittent sampling + frequency shift interference, noise product + frequency shift interference, and noise product + intermittent sampling interference were -10dB, -5dB, and 0dB, respectively. The interference-to-signal ratio ranges for noise convolution interference and noise product interference were -5dB, 0dB, and 5dB, respectively.

[0058] Step S102: In step S102, the time domain image and the time-frequency domain image of the interference are extracted, specifically:

[0059] The nine types of SAR active interference signals are processed to generate the time domain image of the interference, and the interference is subjected to short-time Fourier transform to obtain the time-frequency image of the interference signal. The formula of short-time Fourier transform is:

[0060]

[0061] τ is the time delay, G(τ,f) is the time-frequency signal obtained after short-time Fourier transform, and f is the frequency.

[0062] In step S102, the time domain image and the time-frequency domain image of the interference are extracted;

[0063] Specifically, the nine types of SAR active interference signals are processed to generate a time domain image of the interference, and the interference is subjected to a short-time Fourier transform to obtain a time-frequency image of the interference signal. The formula for the short-time Fourier transform is:

[0064]

[0065] τ is the time delay, G(τ,f) is the time-frequency signal obtained after short-time Fourier transform, and f is the frequency.

[0066] Step S103: normalize the time domain image and perform power law transformation on the time-frequency domain image to obtain the processed image number:

[0067] Specifically, the time domain signal and the time-frequency domain signal are normalized to standardize the pixel value range between different pixels. The formula is:

[0068]

[0069] Among them, x is the pixel value of the original time domain image, x p is the normalized pixel value, x min is the minimum value of the original image pixel value, x max is the maximum value of the original image pixel value. Then the normalized time-frequency image is subjected to power law transformation to enhance the image contrast and weaken the sidelobe effect. The formula is:

[0070] S=c×x p r (4)

[0071] Among them, S is the time-frequency image after power law transformation, c is the normalization constant, and r is the power law parameter

[0072] Step S104: Divide the interference data set into a training set and a test set:

[0073] Specifically, for the nine types of SAR active jammers constructed, the SAR active jammer dataset is divided into a training set and a test set according to the ratio of 80% and 20%.

[0074] Step S105: Input the training set into the dual-channel feature extraction network for training:

[0075] Specifically, the processed time domain and time-frequency domain images are used as the two inputs of the dual-channel feature extraction network. Both channels use the Diverse Branch Block (DBB) to replace the traditional convolution layer. Through parallel convolution operations (different kernel sizes, this method uses four different parallel convolution methods: 1x1, 1x1~kxk, 1x1~avgpool, kxk)), multi-scale local details are captured to improve the model's adaptability to complex features. The squeeze-and-excitation (SE) channel attention mechanism is used to dynamically adjust the contribution of each channel, enhance key features and suppress redundant information. In the squeeze stage, the global average pooling converts the input feature map X∈R C×H×W Mapped into a channel vector Z∈R c , the formula is:

[0076]

[0077] Among them, H represents the feature map height, W represents the feature map width, and Z C Indicates the description value of the Cth channel, x c(i, j) represents the eigenvalue of the Cth channel at (i, j) in space. Then, in the excitation stage, the importance weight of each channel is learned through the fully connected layer and the Sigmoid activation function. The formula is:

[0078] S=σ(W2δ(W1Z))(6)

[0079] in, and is the weight matrix, r is the reduction rate for reducing complexity, δ represents the ReLU activation function, and σ represents the Sigmoid activation function, which normalizes the weights to the interval [0,1]. Subsequently, the learned channel weights are multiplied by the original feature map X by channel to generate the weighted feature map This process dynamically adjusts the contribution of each channel, thereby enhancing key features and suppressing minor ones.

[0080] After using DBB and SE attention mechanisms to extract local features in the time domain and time-frequency domain, a weighted fusion strategy is used to integrate dual-channel features. The time domain input feature matrix is ​​set as The input feature matrix in the time-frequency domain is N is the sample size, D t and D tf

[0081] are the feature dimensions of time domain and time-frequency domain respectively, then the dynamic weight allocation mechanism is as follows:

[0082]

[0083] Among them, H(F t ) is information entropy, α is temperature parameter, ω tf =1-ω t , ensuring that the weight sum is 1 to avoid energy imbalance. In order to improve the model’s sensitivity to interference features under low JSR, a feature weight optimization term δ is designed for low JSR scenarios, which is expressed as:

[0084]

[0085] where σ s 2 is the variance of the target echo signal, σ j 2 is the variance of the SAR interference signal, σ n 2 is the modal independent variance of the ambient noise.

[0086] The fused features not only contain rich local detail information, but also extract global features through the Transformer module, thereby enhancing the global context information. Suppose the input sequence is X = (x1, x2, x3, ... x n ), where x i Represents the i-th input vector. The multi-head attention mechanism can effectively capture long-distance dependencies, and its expression is:

[0087] MultiHead(Q,K,V)=Concat(head1,head2,…head h )W o (9)

[0088] Where Q, K, and V represent query, key, and value matrices, respectively, obtained by applying different linear transformations to the input X; W o Represents the linear transformation matrix of the multi-head attention mechanism output, head i

[0089] The self-attention output corresponding to the i-th head can be expressed as:

[0090]

[0091] Attention refers to the self-attention mechanism, which is formally expressed as:

[0092]

[0093] where d k Indicates the dimension of the key vector.

[0094] After training is completed, save the trained model.

[0095] Step S106: Based on the model saved in step S105, the test set is fed into the optimally trained network for testing:

[0096] Specifically, the time domain image and time-frequency image of the SAR interference signal in the test set are input into the trained dual-channel feature extraction network for testing, and the recognition accuracy of each interference is tested.

[0097] Invention point 1: Using diverse branch blocks (DBB) to replace traditional convolutional layers, extracting multi-granularity local features through parallel multi-scale convolution kernels (1×1, 3×3, etc.), while introducing the squeeze-excitation (SE) attention mechanism, dynamically suppressing redundant channels and enhancing key features (such as the frequency modulation pattern of interference signals) through global average pooling and channel weighting; its technical advantage is that the multi-branch structure of DBB improves the model's adaptability to complex interference patterns (such as the mixed time-frequency features of compound interference). Compared with traditional convolutional layers, it has stronger feature extraction capabilities.

[0098] Invention Point 2: For SAR active interference identification under low signal-to-interference ratio (SIR) conditions, we propose a data-driven dynamic weight allocation mechanism. This mechanism automatically adjusts fusion weights based on the information entropy of time and time-frequency domain features. To improve the model's sensitivity to interference features under low JSR conditions, we design feature weight optimization terms suitable for low JSR scenarios. Compared to fixed-weight fusion, this dynamic strategy can adaptively suppress redundant information and avoid the redundant information interference caused by fixed weights. Weights are automatically learned through end-to-end training, eliminating the need for manual parameter presetting. This adapts to the differences in feature distributions of different interference types, enabling more sensitive capture of subtle key time-frequency domain features, especially under low JSR conditions, improving model generalization.

[0099] Through the above-mentioned inventive concepts, this method achieves an overall recognition accuracy (OA) of 96.69%, an improvement of over 7% over existing state-of-the-art methods. It also maintains a high recognition rate of 89.17% in low JSR (-10dB) scenarios, significantly outperforming traditional methods (e.g., ResNet, which achieves only 75.1% recognition under these conditions) and over 14% higher than traditional convolutional networks (e.g., ResNet). This performance fully demonstrates the method's robust ability to extract weak interference features in complex electromagnetic environments, effectively meeting the real-time anti-interference application requirements of modern SAR systems in strong noise and low signal-to-interference ratio scenarios.

[0100] Exemplary Systems

[0101] The embodiment of the present application also provides a SAR active interference identification system under low interference-to-signal ratio conditions, including: a data preprocessing unit, including time domain image normalization operations for six types of single interference patterns: noise frequency modulation interference, noise amplitude modulation interference, intermittent sampling interference, frequency shift interference, noise convolution interference, and noise product interference, and three types of composite interference patterns: intermittent sampling + frequency shift composite interference, noise product + frequency shift composite interference, and noise product + intermittent sampling composite interference. After the time-frequency domain image normalization, a power law transformation is performed to improve the image contrast; a model construction unit, configured as a dual-channel feature extraction network, in which each channel replaces the traditional convolution layer with DBB to improve the feature extraction capability, and adds an SE attention mechanism to dynamically adjust the contribution of each channel, thereby enhancing key features and suppressing secondary features. Dynamic feature fusion is performed on the extracted time domain and time-frequency domain features through a dynamic weight strategy, and global context modeling is performed using Transformer after fusion; a model testing unit, configured to input the test set into the dual-channel feature extraction network for pattern recognition based on the optimal network model and parameters to obtain SAR active interference pattern recognition results.

[0102] The SAR active interference identification system under low interference signal ratio conditions provided in the embodiments of the present application can implement any of the above-mentioned SAR active interference identification steps and processes under low interference signal ratio conditions and achieve the same technical effects, which will not be described in detail here.

[0103] Exemplary devices

[0104] The present application discloses an electronic device, comprising a processor and a memory coupled thereto. The memory stores an executable program, and when the program is executed by the processor, a method for identifying SAR active interference under low interference signal ratio conditions is implemented.

[0105] Since the steps of a SAR active interference identification method under low interference-to-signal ratio conditions have been introduced in detail in the specific implementation method example, they will not be described in detail here.

[0106] The processor may be a central processing unit (CPU), digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic device to support high-speed signal processing and complex algorithm execution. The memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), or solid-state storage devices for storing trained recognition models, pre-processed interference signal data, and intermediate calculation results.

[0107] The processor can be specifically configured to: simulate the SAR active interference signal; extract the time domain image and time-frequency domain image of the interference; normalize the time domain image and then perform power law transformation on the normalized time-frequency domain image; divide the SAR active interference data set into a training set and a test set; input the training set into a dual-channel feature extraction network for training; input the test set into the trained model to obtain SAR active interference recognition results under different interference-to-signal ratios;

[0108] At the hardware implementation level, the method can be compiled into a hardware description language (HDL) and deployed on a dedicated integrated circuit (ASIC), or as a software program running on a general-purpose computer platform. When the program stored on a non-transitory computer-readable medium (such as a hard disk or flash memory) is loaded by a processor, the general-purpose computing device is transformed into a dedicated recognition platform, supporting real-time signal processing and model inference.

[0109] Those skilled in the art may adjust the hardware architecture based on actual needs, for example, by improving computing efficiency through parallel processing using multi-core processors or expanding data set capacity using distributed storage. The functional units of the device may be physically integrated or distributed across a heterogeneous computing environment, and the implementation method does not affect the core features of the technical solutions of this application.

[0110] It should be noted that the embodiments described in this specification are described in a hierarchical and progressive manner. Common technical features across the various embodiments can be referenced across them, and each embodiment focuses on the differences from other embodiments. In particular, the descriptions of the device and system embodiments are kept brief because their core implementation logic is inherently consistent with the method embodiments. Specific details can be understood in conjunction with the corresponding method embodiments.

[0111] The above device and system embodiments are intended for illustrative purposes only. The independently described functional units can be physically separate hardware modules or logically integrated software components. These units can be deployed on the same device or distributed across heterogeneous network environments. Users can flexibly select module combinations based on actual application requirements, achieving the core functionality of the solution without additional creative effort.

[0112] The protection scope of this application is not limited to the above preferred embodiments.

[0113] Equivalent modification, feature replacement or optimization improvement of the application technical solution, as long as it meets the requirements

[0114] The core concepts and principles of an invention should be included in the scope of patent protection.

Claims

1. A SAR active jammer identification method under low signal-to-interference ratio conditions, characterized in that: include: Step S101, simulating a SAR active interference signal; Step S102: extracting the time domain image and the time-frequency domain image of the interference; Step S103: normalize the time domain image, and then perform power law transformation on the time-frequency domain image after normalization; Step S104: dividing the SAR active interference data set into a training set and a test set; Step S105: input the training set into the dual-channel feature extraction network for training; Step S106: Input the test set into the trained model to obtain SAR active jammer recognition results under different interference-to-signal ratios.

2. The SAR active interference identification method under low interference-to-signal ratio conditions according to claim 1, characterized in that: Specifically, step S101 is as follows: nine types of SAR active interference signals are simulated, including six types of single interference patterns: noise frequency modulation interference, noise amplitude modulation interference, intermittent sampling interference, frequency shift interference, noise convolution interference, and noise product interference, and three types of composite interference patterns: intermittent sampling + frequency shift composite interference, noise product + frequency shift composite interference, and noise product + intermittent sampling composite interference. The interference-to-signal ratio ranges of noise frequency modulation and noise amplitude modulation are 0dB, 5dB, and 10dB; the interference-to-signal ratio ranges of intermittent sampling interference, frequency shift interference, intermittent sampling + frequency shift composite interference, noise product + frequency shift composite interference, and noise product + intermittent sampling composite interference are -10dB, -5dB, and 0dB; the interference-to-signal ratio ranges of noise convolution interference and noise product interference are -5dB, 0dB, and 5dB. 240 interference samples are generated under each type of interference-to-signal ratio.

3. The SAR active interference identification method under low interference-to-signal ratio conditions according to claim 1, characterized in that: In step S102, the time domain image and the time-frequency domain image of the interference are extracted, specifically: The nine types of SAR active interference signals are processed to generate the time domain image of the interference, and the interference is subjected to short-time Fourier transform to obtain the time-frequency image of the interference signal. The formula of short-time Fourier transform is: G(τ,f)=∫ R X(t)h(t-τ)e -j2πft dt Where X(t) represents the time domain signal, t is time, h(t-τ) is the window function, τ is the time delay, G(τ,f) is the time-frequency signal obtained after short-time Fourier transform, and f is the frequency.

4. The SAR active interference identification method based on low interference-to-signal ratio conditions according to claim 1, characterized in that: In step S103, the time domain image is normalized and the time-frequency domain image is subjected to power law transformation to obtain processed image data, specifically: Normalize the time domain signal and the time-frequency domain signal to standardize the pixel value range between different pixels. The formula is: Among them, x is the pixel value of the original time domain image, x p is the normalized pixel value, x min is the minimum value of the original image pixel value, x max is the maximum value of the original image pixel value; then the normalized time-frequency image is subjected to power law transformation to enhance the image contrast and weaken the sidelobe effect. The formula is: S=c×x p r Among them, S is the time-frequency image after power law transformation, c is the normalization constant, and r is the power law parameter.

5. The SAR active interference identification method under low interference-to-signal ratio conditions according to claim 1, characterized in that: In step S104, the interference data set is divided into a training set and a test set, specifically: For the nine categories of SAR active jammers constructed, the SAR active jammer dataset is divided into a training set and a test set according to the ratio of 80% and 20%.

6. The SAR active interference identification method under low interference-to-signal ratio conditions according to claim 1, characterized in that: In step S105, the training set is input into the dual-channel feature extraction network for training, specifically: The processed time domain and time-frequency domain images are used as the two inputs of the dual-channel feature extraction network. Both channels use Diverse branch block (DBB) to replace the traditional convolution layer. The parallel convolution operation captures multi-scale local details and improves the adaptability of the model to complex features. The squeeze-and-excitation (SE) channel attention mechanism is used to dynamically adjust the contribution of each channel, enhance key features and suppress redundant information. In the squeezing stage, the global average pooling will input the feature map X∈R C×H×W Mapped into a channel vector Z∈R c , the formula is: Among them, H represents the feature map height, W represents the feature map width, and Z C Indicates the description value of the Cth channel, x c (i, j) represents the eigenvalue of the Cth channel at (i, j) in space; then in the excitation stage, the importance weight of each channel is learned through the fully connected layer and the Sigmoid activation function, and the formula is: S=σ(W2δ(W1Z)) in, and is the weight matrix, r is the reduction rate for reducing complexity, δ represents the ReLU activation function, and σ represents the Sigmoid activation function, which normalizes the weights to the interval [0,1]. Subsequently, the learned channel weights are multiplied by the original feature map X by channel to generate a weighted feature map. This process dynamically adjusts the contribution of each channel, thereby enhancing key features and suppressing minor features; After using DBB and SE attention mechanisms to extract local features in the time domain and time-frequency domain, a weighted fusion strategy is used to integrate dual-channel features. The time domain input feature matrix is ​​set as The input feature matrix in the time-frequency domain is N is the sample size, D t and D tf are the feature dimensions of time domain and time-frequency domain respectively, then the dynamic weight allocation mechanism is as follows: Among them, H(F t ) is information entropy, α is temperature parameter, ω tf =1-ω t , ensuring that the weight sum is 1 to avoid energy imbalance; in order to improve the model's sensitivity to interference features under low JSR, a feature weight optimization term δ suitable for low JSR scenarios is designed, which is expressed as: where σ s 2 is the variance of the target echo signal, σ j 2 is the variance of the SAR interference signal, σ n 2 is the modal independent variance of the ambient noise; The fused features not only contain rich local detail information, but also perform global feature extraction through the Transformer module, thereby enhancing the global context information; let the input sequence be X = (x1, x2, x3, ... x n ), where x i Represents the i-th input vector. The multi-head attention mechanism can effectively capture long-distance dependencies, and its expression is: MultiHead(Q,K,V)=Concat(head1,head2,.....head h )W o Where Q, K, and V represent query, key, and value matrices, respectively, obtained by applying different linear transformations to the input X; W o Represents the linear transformation matrix of the multi-head attention mechanism output, head i The self-attention output corresponding to the i-th head can be expressed as: head i =Attention(QW i Q ,KW i K ,VW i V ) Attention refers to the self-attention mechanism, which is formally expressed as: where d k Indicates the dimension of the key vector; after training is completed, save the trained model.

7. The SAR active interference identification method under low interference-to-signal ratio conditions according to claim 1, characterized in that: In step S106, based on the model saved in step S105, the test set is fed into the best trained network for testing, specifically: The time domain image and time-frequency image of the SAR interference signal in the test set are input into the trained dual-channel feature extraction network for testing, and the recognition accuracy of each interference is tested.

8. SAR active jammer identification system under low signal-to-interference ratio conditions, including: The data preprocessing unit performs time domain image normalization operations on six types of single interference patterns: sound frequency modulation interference, noise amplitude modulation interference, intermittent sampling interference, frequency shift interference, noise convolution interference, and noise product interference; and three types of composite interference patterns: intermittent sampling + frequency shift composite interference, noise product + frequency shift composite interference, and noise product + intermittent sampling composite interference. After time-frequency domain image normalization, power law transformation is performed to improve image contrast; The model building unit is configured as a dual-channel feature extraction network, in which the traditional convolutional layer is replaced with DBB in each channel to improve feature extraction capabilities, and the SE attention mechanism is added to dynamically adjust the contribution of each channel, thereby enhancing key features and suppressing secondary features; Dynamic feature fusion is performed on the extracted time domain and time-frequency domain features through a dynamic weight strategy, and after fusion, global context modeling is performed using Transformer. The model testing unit is configured to input the test set into the dual-channel feature extraction network for pattern recognition based on the optimal network model and parameters, and obtain the SAR active interference pattern recognition result.

9. A storage device storing a plurality of programs, characterized in that: The program application is loaded and executed by a processor to implement the SAR active interference identification method under low interference-to-signal ratio conditions described in any one of claims 1-7.

10. An electronic device comprising a storage device and a processor; the processor is adapted to execute various programs; and the memory is adapted to store a plurality of programs; characterized in that: When the memory executes the program on the processor, the method for identifying SAR active interference under low interference-to-signal ratio conditions according to any one of claims 1 to 7 is implemented.

Images