Radar suppression deception composite interference identification technology under low signal-to-noise ratio condition

Abstract

The invention relates to the technical field of radar signal processing, and discloses a radar suppression deception composite interference identification technology under a low signal-to-noise ratio condition, and the technology comprises the steps: obtaining an interference pulse sequence, carrying out the multi-domain representation of the interference pulse sequence, and generating a time domain sequence image and a time frequency image; extracting deep features of the time domain sequence image and the time frequency image by using a feature extraction network; inputting the time domain sequence image, the time frequency image deep feature and the noise power feature into a cross fusion network, performing deep feature fusion, and outputting a comprehensive feature vector; and determining an interference type according to the comprehensive feature vector. The device comprises a data representation module, a feature extraction module, a cross fusion module and a classification decision module. According to the invention, multi-domain information representation and fusion are carried out on the interference signals, so that the problem that strong noise covers interference characteristics is solved, and the composite interference identification accuracy under the condition of low signal-to-noise ratio is improved.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing technology, specifically to a radar suppression, deception, and composite interference identification technology under low signal-to-noise ratio conditions. Background Technology

[0002] In modern electronic warfare environments, the stable operation of radar systems is crucial for ensuring information superiority and operational effectiveness. However, with the continuous evolution of electronic jamming technology, especially the combination of various jamming patterns to form suppression, deception, and composite jamming, the survivability and detection performance of radars face severe challenges. Therefore, rapid and accurate identification of jamming signals in complex electromagnetic environments is a prerequisite and key step in implementing effective anti-jamming measures.

[0003] Currently, deep learning-based recognition methods have become the mainstream technology in this field. However, existing deep learning methods still face technical bottlenecks when dealing with suppressed deception interference, especially under low signal-to-noise ratio conditions. The fundamental reason lies in the significant power differences between noise and deception interference, and between different types of interference components. To achieve the purpose of deception, the power of deception interference is usually designed to be comparable to the echo of the real target. However, under the dual influence of suppressed interference and environmental noise, its signal characteristics are easily submerged. Whether in the time-domain waveform or in the time-frequency domain image representation, noise and suppressed interference components will have a severe masking effect on the low-power deception interference features, making it difficult for the network model to capture complete and effective discriminative information, thus causing a significant decline in recognition performance.

[0004] Therefore, this invention proposes a radar suppression, deception, and composite interference identification technology under low signal-to-noise ratio conditions to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a radar suppression and deception combined interference identification technology under low signal-to-noise ratio (SNR) conditions. This technology effectively overcomes the masking effect of noise and suppression interference on the characteristics of low-power deception interference under low SNR conditions, thereby significantly improving the identification accuracy of suppression and deception combined interference.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a radar suppression, deception, and composite interference identification technology under low signal-to-noise ratio conditions, comprising the following steps:

[0007] S1. Obtain the interference pulse sequence of the radar echo;

[0008] S2. Perform multi-domain characterization on the interference pulse sequence, and generate in parallel a time-domain sequence image, a time-frequency image, and noise power features corresponding to the interference pulse sequence;

[0009] S3. Based on the time-domain sequence image and the time-frequency image, extract time-domain features and time-frequency domain features respectively using a feature extraction network;

[0010] S4. Input the time-domain features, time-frequency domain features, and noise power features into the cross-fusion network, and the cross-fusion network performs deep feature fusion on the time-domain features, time-frequency domain features, and noise power features to output a comprehensive feature vector;

[0011] S5. Input the comprehensive feature vector into the classifier, and the classifier outputs the finally determined interference category.

[0012] Preferably, in step S2, the step of performing multi-domain characterization on the interference pulse sequence and generating in parallel a time-domain sequence image, a time-frequency image, and noise power features corresponding to the interference pulse sequence includes:

[0013] A short-time Fourier transform is performed on one pulse data in the interference pulse sequence to obtain an initial time-frequency distribution image. The initial time-frequency distribution image is then subjected to histogram equalization. Finally, the time-frequency distribution image after histogram equalization is combined with the initial time-frequency distribution image and subjected to image channel stitching and normalization to generate the time-frequency image.

[0014] A pulse data in the interference pulse sequence is compressed to obtain a compressed time-domain signal. The compressed time-domain signal and the original time-domain signal before pulse compression are modulo-valued, downsampled by max pooling, and plotted into two-dimensional images. The two generated two-dimensional images are then stitched together and normalized to generate the time-domain sequence image.

[0015] The noise power of the time-domain signal after pulse compression is statistically analyzed and used as a noise power characteristic.

[0016] Preferably, in step S4, the step of performing deep feature fusion of time-domain features, time-frequency domain features, and noise power features by the cross-fusion network to output a comprehensive feature vector includes:

[0017] The time-domain deep features and time-frequency domain deep features obtained by the feature extraction network are input into the cross-fusion network for deep feature fusion. The resulting deep fused features are then combined with the noise power features statistically derived from the interference pulse sequence to form the comprehensive feature vector.

[0018] Preferably, the steps of inputting the temporal deep features and time-frequency deep features obtained by the feature extraction network into the cross-fusion network for deep feature fusion include:

[0019] The time-domain features and time-frequency domain features are input into a multilayer perceptron consisting of multiple fully connected layers. The multilayer perceptron performs feature dimensionality reduction and nonlinear mapping to output the fused features.

[0020] Preferably, the feature extraction network, the cross-fusion network, and the classification module structurally constitute a unified network, which includes a feature extraction module and a cross-fusion module.

[0021] The feature extraction module includes a time-domain feature extraction branch for processing the time-domain sequence image and a time-frequency domain feature extraction branch for processing the time-frequency image, so as to generate the time-domain features and time-frequency domain features respectively;

[0022] The cross-fusion module is used to receive the time-domain features and the time-frequency domain features, and perform deep feature fusion.

[0023] Preferably, the temporal feature extraction branch and the time-frequency domain feature extraction branch each include multiple stacked residual convolutional modules for progressively extracting deep features from the temporal sequence image and the time-frequency image.

[0024] Preferably, the cross-fusion module includes one or more fully connected layers, forming a multilayer perceptron.

[0025] Preferably, in step S5, the step of inputting the comprehensive feature vector into the classifier and having the classifier output the finally determined interference category includes:

[0026] The comprehensive feature vector is input into one or more fully connected layers for feature dimension adjustment and nonlinear mapping.

[0027] The features processed by the one or more fully connected layers are input into the final fully connected layer and the Softmax activation function to calculate and output the final interference category of the interference pulse sequence.

[0028] The present invention also provides a radar suppression, deception, and combined interference identification device under low signal-to-noise ratio conditions, the device comprising:

[0029] The data characterization module is used to acquire the interference pulse sequence of the radar echo and perform multi-domain characterization on the interference pulse sequence to generate corresponding time-domain sequence images, time-frequency images and noise power features in parallel.

[0030] The feature extraction module is configured as a feature extraction network, used to extract and generate time-domain features and time-frequency features based on the time-domain sequence image and the time-frequency image, respectively.

[0031] The cross-fusion module, configured as a cross-fusion network, is used to receive the time-domain features, the time-frequency domain features, and the noise power features, and to perform adaptive deep feature fusion on the received features to output a comprehensive feature vector.

[0032] The classification decision module is configured as a classifier to receive the comprehensive feature vector and output the finally determined interference category.

[0033] This invention provides a radar suppression, deception, and composite interference identification technology under low signal-to-noise ratio conditions.

[0034] It has the following beneficial effects:

[0035] 1. This invention effectively enhances the discernibility of weak interference features under low signal-to-noise ratio conditions by performing parallel multi-domain characterization processing on interference pulse sequences. Specifically, pulse compression processing of the time-domain signal can enhance deceptive interference features masked by strong noise, while enhancement processing of the time-frequency image can improve the contrast of different interference components, thereby extracting more comprehensive and reliable low-level features for subsequent accurate identification of interference types.

[0036] 2. This invention employs a dual-branch CNN to extract deep features from the time-domain and time-frequency maps, and utilizes a cross-fusion network for multi-domain feature fusion. Specifically, the time-domain sequence is image-processed and features are extracted using a CNN, effectively leveraging the CNN's ability to extract local image features. The time-frequency distribution reveals the mechanistic differences between different interference signals and preserves the original features of various interferences in the composite interference. The CNN network can effectively extract the features of different interference components in the composite interference. The designed cross-fusion network can adaptively perform deep fusion of multi-domain features. This adaptive mechanism allows the network to dynamically and selectively fuse deep features from the time-domain, time-frequency, and noise power domains based on the specific characteristics of the current input interference signal, thereby ensuring the robustness of the designed network in various complex electromagnetic environments.

[0037] 3. This invention improves the accuracy and robustness of identifying radar suppression and deception complex interference under harsh conditions with low signal-to-noise ratio. Through multi-domain feature representation and enhancement, bi-branch feature extraction, and deep feature fusion, this technical solution effectively overcomes the masking effect of suppression interference or background noise on weak deception interference features, and achieves accurate differentiation of various complex complex interference types. Attached Figure Description

[0038] Figure 1 This is a technical flowchart of a radar suppression, deception and composite interference identification technology under low signal-to-noise ratio conditions according to the present invention;

[0039] Figure 2This is a schematic diagram of the time-domain sequence preprocessing of the present invention;

[0040] Figure 3 This is a schematic diagram of the pulse compression post-time domain sequence preprocessing of the present invention;

[0041] Figure 4 This is a schematic diagram of the overall preprocessing process of the present invention;

[0042] Figure 5 This is an example image of the preprocessed image of the dense false target composite interference of the present invention;

[0043] Figure 6 This is an example image of the preprocessed image of the blocking slice composite interference of the present invention;

[0044] Figure 7 This is an example image of the preprocessed image of the blocking comb spectrum composite interference of the present invention;

[0045] Figure 8 This is an example image of the preprocessed swept-frequency dense false target composite interference of the present invention;

[0046] Figure 9 This is an example image of the preprocessed target comb spectrum composite interference of the present invention;

[0047] Figure 10 This is a schematic diagram of the overall network structure of the present invention;

[0048] Figure 11 This is a comparison curve of the recognition accuracy and loss function of the present invention's technical method and the comparative technical methods;

[0049] Figure 12 This is a confusion matrix diagram of the baseline network identification results using only time-frequency diagrams in this invention;

[0050] Figure 13 This is a confusion matrix diagram of the identification results of the technical method of the present invention;

[0051] Figure 14 This is a graph showing the recognition accuracy under different noise-to-interference ratios of the present invention.

[0052] The module includes: 10. Data representation module; 20. Feature extraction module; 30. Cross-fusion module; and 40. Classification decision module. Detailed Implementation

[0053] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Reference Figure 1 The present invention provides a radar suppression, deception, and composite interference identification technology under low signal-to-noise ratio conditions, which, in one embodiment, may include the following steps:

[0055] S1. Obtain the interference pulse sequence of the radar echo;

[0056] S2. Perform multi-domain characterization on the interference pulse sequence, and generate in parallel a time-domain sequence image, a time-frequency image, and noise power features corresponding to the interference pulse sequence;

[0057] S3. Based on the time-domain sequence image and the time-frequency image, use a feature extraction network to extract and generate time-domain features and time-frequency domain features respectively;

[0058] S4. Input the time-domain features, time-frequency domain features, and noise power features into the cross-fusion network, and the cross-fusion network performs adaptive deep feature fusion on the time-domain features, time-frequency domain features, and noise power features to output a comprehensive feature vector;

[0059] S5. Input the comprehensive feature vector into the classifier, and the classifier outputs the finally determined interference category.

[0060] This invention also provides a radar suppression, deception, and combined interference identification device under low signal-to-noise ratio conditions, the structure of which corresponds to the above-described technical method flow. The device includes:

[0061] Data representation module 10 is used to execute steps S1 and S2;

[0062] Feature extraction module 20, configured as a feature extraction network, is used to execute step S3;

[0063] Cross-fusion module 30, configured as a cross-fusion network, is used to execute step S4;

[0064] The classification decision module 40 is configured as a classifier to execute step S5.

[0065] In one specific embodiment, step S1 is performed by the data characterization module 10, which acquires the raw echo signal received by the radar system and separates the jamming pulse sequence from it. As an example, a dataset can be generated through simulation modeling. This dataset contains 21 types of composite jamming, including three types of suppression jamming (frequency targeting jamming, blocking jamming, and frequency sweeping jamming) and three types of deception jamming (comb spectrum jamming, dense false target jamming, and slice jamming), as well as any combination of two types of jamming.

[0066] Step S2 is also executed by the data representation module 10, and its detailed process is described in [reference needed]. Figure 4 This module performs parallel dual-domain processing on the received interference pulse sequence to generate a time-domain sequence image and a time-frequency image. In the time-frequency domain, the time-frequency image is obtained through short-time Fourier transform, image cropping, histogram equalization, multi-channel stitching, and normalization. In the time domain, the time-domain sequence image is obtained by performing modulo operations, max-pooling downsampling, rendering as a two-dimensional image, and multi-channel stitching and normalization on the original signal and the pulse-compressed signal, respectively. Simultaneously, the data characterization module 10 also calculates and generates the noise power characteristics by statistically analyzing the signal characteristics of the interference pulse sequence. Figure 2 and Figure 3 The preprocessing processes for the time-domain sequence and the pulse-compression time-domain sequence are illustrated respectively. Figures 5 to 9 Examples of images generated after this preprocessing are shown for some types of composite interference.

[0067] Step S3 is performed by a feature extraction network, which physically serves as the front end of a cross-fusion network. This module receives the time-domain sequence image and time-frequency image generated in step S2, and uses its internal time-domain and time-frequency domain feature extraction branches to extract and generate time-domain features and time-frequency domain features, respectively. Both the time-domain and time-frequency domain feature extraction branches employ multiple stacked residual convolutional modules to progressively extract deep features. Preferably, a CBAM attention module can be set at specific locations in the time-frequency domain feature extraction branch to enhance attention to occluded key features.

[0068] Step S4 is performed by the cross-fusion network, the overall structure of which is as follows: Figure 10 This module receives temporal deep features and time-frequency deep features generated by the feature extraction network, as well as noise power features generated by the data representation module. The cross-fusion network uses a multilayer perceptron formed by multiple fully connected layers to perform feature dimensionality reduction and fusion on the temporal and time-frequency deep features. The fused features are then concatenated with the noise power features to form the final comprehensive feature vector.

[0069] Step S5 is performed by the classification decision module 40. This module receives the comprehensive feature vector output by the cross-fusion network, and outputs a classification vector representing the predicted probability of each type of interference through one or more fully connected layers and a final Softmax activation function, thereby determining the final interference category. During the network training phase, the Focal-Loss function can be used to optimize the parameters of the entire network to improve the model's attention to difficult-to-classify samples.

[0070] Through the above technical solution, the present invention can improve the recognition accuracy of suppression deception composite interference under low signal-to-noise ratio conditions. Figure 11 , Figure 12 , Figure 13 and Figure 14 The experimental results demonstrate the performance of the embodiments of the present invention compared to other methods.

[0071] Reference Figure 2 , Figure 3 and Figure 4 These three figures illustrate the time-domain sequence preprocessing, pulse compression time-domain sequence preprocessing, and the overall preprocessing workflow, respectively. The specific steps in this workflow will be explained in detail below.

[0072] Step S2 is performed by the data characterization module 10. This module receives the one-dimensional time-domain interference pulse sequence obtained in step S1 and performs parallel dual-domain processing on it to generate a two-dimensional time-domain sequence image and a two-dimensional time-frequency image.

[0073] To obtain the joint distribution information of the signal in the time and frequency domains, a short-time Fourier transform (STFT) is first performed on the input interference pulse sequence to obtain an initial two-dimensional time-frequency distribution image.

[0074] To enhance the visibility of weak interference features under low signal-to-noise ratio conditions, histogram equalization was performed on the initial time-frequency distribution image to adjust the gray-level dynamic range of the image, thereby obtaining a contrast-enhanced two-dimensional time-frequency distribution image.

[0075] Subsequently, the time-frequency distribution image after histogram equalization is stitched together with the initial time-frequency distribution image along the channel dimension to form a dual-channel two-dimensional image containing both original and enhanced information.

[0076] Finally, to limit the range of pixel values ​​in the image to a uniform interval, each pixel value of the dual-channel 2D image is normalized. The calculation process for normalizing a pixel value is as follows:

[0077]

[0078] In the formula, I′ is the normalized pixel value; I is the original pixel value; Imin I is the minimum value of all pixels in the image; max This represents the maximum value of all pixels in the image. After the above processing, a time-frequency image is finally generated.

[0079] Considering the correlation between repeater-type deception jamming signals and radar transmission signals, gain can be achieved through pulse compression processing, a characteristic typically not found in noise and suppression jamming. Therefore, this embodiment utilizes pulse compression processing to enhance and distinguish different types of jamming characteristics.

[0080] Specifically, the data representation module 10 first uses the known radar transmitted signal as a reference signal to perform pulse compression on the original time-domain interference pulse sequence. The calculation process is as follows:

[0081] x o (t) = IFFT(FFT[x] r (t)]·FFT[s(t) * );

[0082] In the formula, x o (t) represents the time-domain signal after pulse compression; x r (t) represents the received original time-domain signal; s(t) represents the radar transmitted signal; FFT[·] represents the Fast Fourier Transform operation; IFFT(·) represents the Inverse Fast Fourier Transform operation; [·] * This indicates the conjugate operation.

[0083] To extract the amplitude information of the signal, the original time-domain signal x before pulse compression was processed. r (t) and the pulse-compressed time-domain signal x o (t) Modulus value. After pulse compression, the original time-domain signal is compressed into a narrow pulse. To reduce information loss during the size compression process after image processing, max pooling downsampling is performed on the two one-dimensional signals after modulus extraction. The downsampling calculation process is as follows:

[0084] y(k)=max(x((k-1)·S+1),…,x((k-1)·S+P));

[0085] In the formula, y(k) is the one-dimensional signal obtained after downsampling; x is the input one-dimensional signal before downsampling; P is the size of the pooling window; S is the stride of the pooling window; and k is the index of the output signal.

[0086] Subsequently, considering the low dimensionality of the time-domain signal and that its features mainly manifest in local amplitude differences, in order to improve feature extraction efficiency, the two downsampled one-dimensional signals were plotted as two-dimensional images. These two two-dimensional images were then stitched together along the channel dimension to form a dual-channel two-dimensional image containing information about the pulse before and after compression, preparing for subsequent image feature extraction.

[0087] Finally, the same normalization method is used to normalize the dual-channel two-dimensional image, ultimately generating a time-domain sequence image.

[0088] The following is about Figure 1 Step S3 in the illustrated technical method will be described in detail. This step is performed by a feature extraction network, which physically constitutes... Figure 10 The overall network structure shown has two feature extraction branches.

[0089] The network receives the time-domain sequence image and time-frequency image generated in step S2. Its core function is to extract high-dimensional depth features from the images in both domains using a deep convolutional network. Specifically, both the time-domain feature extraction branch and the time-frequency domain feature extraction branch adopt a structure composed of multiple stacked residual convolutional modules (ResBlocks). (Refer to...) Figure 10 The processing flow of a residual convolution module is defined as follows:

[0090] CBR(X)=relu(bn(conv2D(X)));

[0091] CB(X) = bn(conv2D(X));

[0092] ResBlock(X)=relu(CB(CBR(CBR(X)))+CB(X));

[0093] In the formula, X is the input feature map of the module; conv2D is the two-dimensional convolution operation; bn is the batch-normalization operation; relu is the rectified-linear-unit activation function; CBR(X) represents a convolutional batch-normalized activation unit; CB(X) represents a convolutional batch-normalized unit; and ResBlock(X) is the feature map of the final output of the residual convolution module.

[0094] In the time-frequency domain, high-power blocking interference may completely mask the features of low-power spoofing interference. In this case, the time-frequency plot of the composite interference consisting of blocking and spoofing interference is similar to that of single blocking interference. To enable the time-frequency plot branch to adaptively handle this situation and focus only on useful feature information, a CBAM (Convolutional-Block-Attention-Module) attention module is inserted at a specific position in the time-frequency domain feature extraction branch. Its structure is also similar to... Figure 10 This module applies channel attention and spatial attention sequentially, enabling the network to focus on feature channels and spatial regions with richer information. Its processing flow is as follows:

[0095] M c (X)=X·σ(fc(GlobalAvgPool(X))+fc(GlobalMaxPool(X)));

[0096] M s (X)=X·σ(Conv([GlobalAvgPool(X);GlobalMaxPool(X)]));

[0097] CBAM(X)=M s (M c (X));

[0098] In the formula, M c (X) represents the output of the channel attention module; M s (X) represents the output of the spatial attention module; GlobalAvgPool(X) represents the global average pooling operation; GlobalMaxPool(X) represents the global max pooling operation; fc represents the fully connected layer; σ represents the sigmoid activation function; Conv represents the convolution operation; [;] represents the concatenation operation of features along the channel dimension; · represents the element-wise multiplication operation; CBAM(X) represents the final output feature map. This module enables the network to focus on feature channels and spatial regions with richer information by applying channel attention and spatial attention sequentially.

[0099] After processing by these two parallel branches, the final outputs are time-domain depth features and time-frequency domain depth features, which can be used by the subsequent cross-fusion module.

[0100] The following section elaborates on step S4, the adaptive deep feature fusion process. This step is... Figure 10 The cross-fusion module 30 shown is executed. This module receives the time-domain features and time-frequency features generated in step S3, as well as the noise power features generated in step S2.

[0101] To achieve adaptive feature fusion while minimizing model parameters, multiple fully connected layers are used to perform dimensionality reduction and splicing fusion of deep time-domain and time-frequency domain features, dynamically adjusting the feature response intensity of the two branches.

[0102] Finally, the fused features are combined with the noise power features obtained in step S2, for example by concatenation, to form the final integrated feature vector, which is then output to the classification decision module 40.

[0103] The following is about Figure 1 Step S5 in the technical method shown, namely the process of processing the comprehensive feature vector and determining the interference category, will be explained in detail.

[0104] Step S5 is executed by the classification decision module 40, whose function is mainly implemented through the classifier at the end of the network. This module receives the time-domain-time-frequency fusion feature vector generated after adaptive fusion in step S4, and the noise power feature generated in step S2.

[0105] The core task of this module is to integrate the received multiple features and map them to a preset interference category space. Its detailed processing flow is as follows:

[0106] The first step is to combine the time-domain-time-frequency fusion feature vector output from step S4 with the noise power feature, for example, by concatenation, to form a final comprehensive feature vector that contains all the information.

[0107] The second step involves inputting the synthesized feature vector into one or more fully connected layers for final nonlinear mapping and dimensionality transformation. To enhance the model's generalization ability, a dropout operation can be added after the fully connected layers.

[0108] The third step involves inputting the processed vector into the softmax activation function at the end, transforming it into a classification probability vector. Each dimension of this vector corresponds to a predefined predicted probability of an interference category, and the interference category corresponding to the dimension with the highest probability value is determined as the final result of this identification.

[0109] To more clearly illustrate the flow and processing of features at the network's end, the following formula demonstrates the complete calculation process from temporal and time-frequency domain deep features to the final classification probability output:

[0110] F′ seq =dropout(relu(fc1(F seq )));

[0111] F′ img=dropout(relu(fc2(F img )));

[0112] F fusion =dropout(relu(fc3([F′)) img ;F′ seq ])));

[0113] Output = softmax(fc4([F fusion ;F noise ]));

[0114] In the formula, F seq The feature vector output by the temporal feature extraction branch; F img F is the feature vector output by the time-frequency domain feature extraction branch. noise F' represents the noise power feature extracted from the guiding vector; fc1, fc2, fc3, and fc4 are four independent fully connected layers; ReLU is the modified linear unit activation function; dropout is the random deactivation operation; [;] denotes the vector concatenation operation; F′ seq and F′ img These are intermediate features obtained after preliminary compression of the time-domain and time-frequency-domain features, respectively; F fusion is the intermediate feature after fusing time-domain and time-frequency domain features; softmax is the normalization exponential function; Output is the final output classification probability vector.

[0115] Finally, the classifier uses its softmax activation function at the end to transform the processed comprehensive feature vector into a classification probability vector Output. Each dimension of this vector corresponds to a predefined predicted probability of an interference class, and the interference class corresponding to the dimension with the highest probability value is determined as the final result of this recognition.

[0116] The simulation experimental environment, dataset construction process, and model training parameters of a specific embodiment of the present invention will be described in detail below.

[0117] In one embodiment, the dataset used for training and testing is generated through simulation. This dataset is constructed to handle three types of suppression jamming (targeting jamming, blocking jamming, and sweeping jamming) and three types of deception jamming (comb spectrum jamming, dense false target jamming, and slice jamming), as well as composite jamming formed by any two of these six individual jamming types.

[0118] The specific parameter settings used in simulation modeling for various single disturbances are shown in Table 1 below:

[0119] Table 1: Interference Patterns and Parameters

[0120]

[0121]

[0122] Based on the above parameters, 21 types of interference were generated for six single types of interference and their combinations. In the suppression and deception composite interference samples, the interference-to-noise ratio (IRR) of the suppression interference was set to 0–10 dB, and the INR of the deception interference was set to -20–10 dB. 5000 samples were generated for each type of interference, with a signal sampling rate of 200 MHz and a sampling duration of 200 μs. The entire dataset was divided into a training set and a test set in a 4:1 ratio, with the training set containing 84,000 samples and the test set containing 21,000 samples.

[0123] In one embodiment, the model of the cross-fusion module 30 is trained in the following environment:

[0124] Software framework: PyTorch;

[0125] Hardware platform: CPU is Intel i5-12400F; GPU is NVIDIA GeForce RTX 4060; RAM is 16G.

[0126] Optimizer: AdamW optimizer is used.

[0127] Training parameters: initial learning rate set to 0.001, learning rate decay rate set to 0.92, batch size set to 100, and training epochs set to 30.

[0128] In this embodiment, the training set is used to train the cross-fusion module 30. The training process involves inputting data samples from the training set into the network module 30 in batches, and the network module 30 outputs a classification probability vector for each sample.

[0129] The classification probability vector output by the network is compared with the true interference class label of the sample to calculate the loss value. Because the difficulty of recognizing composite interference of different classes varies at different signal-to-noise ratios, the difficulty of samples during training is uneven. To address this issue, this embodiment uses Focal-Loss as the loss function. This loss function reduces the contribution weight of easily classified samples (i.e., samples with high prediction probabilities) to the overall loss, allowing the model to focus more on difficult-to-classify samples during training.

[0130] After obtaining the network's predicted probability of the true class for each sample in a batch, the loss value is calculated using the following focus loss function:

[0131]

[0132] In the formula, L Focal The calculated focus loss value; N is the number of samples in a training batch; n is the index of the sample in the batch; p n y is the probability value predicted by the model for the true class of the nth sample; n α is the true value of the nth sample; α is the weighting factor hyperparameter used to balance the importance of positive and negative samples; γ is the focusing parameter hyperparameter used to adjust the reduction of the weight of easily classified samples.

[0133] Calculated loss value L Focal This is used to update all trainable parameters in the cross-fusion module 30 via the backpropagation algorithm and the optimizer (such as AdamW). This training process is repeated until the model converges.

[0134] Reference Figure 11 , Figure 12 , Figure 13 and Figure 14 These four figures illustrate the experimental results of one embodiment of the present invention. The results are then described and analyzed objectively.

[0135] To verify the effectiveness of the technical solution of the present invention, a simulation experiment was conducted in one embodiment. Figure 11 The accuracy curves of the proposed method and several comparative methods are presented. The validation set accuracy curves show that, under a noise-to-interference ratio (NIR) of -20 to 10 dB, the method described in this invention (combining time-frequency plot enhancement, time-domain pulse compression enhancement, and noise power features) achieves a validation set accuracy of over 97.89% for 21 types of interference. In comparison, the baseline network using only the time-frequency plot has an accuracy of 88.47%; adding noise power features to the baseline network achieves 95.91% accuracy; adding original time-domain sequence features and noise power features to the baseline network achieves 95.79% accuracy; and adding pulse-compressed time-domain sequence features and noise power features to the baseline network achieves 97.71% accuracy. This set of comparative data demonstrates that the multi-domain information representation and fusion strategy employed in this invention effectively improves the accuracy. The loss curves show that the initial loss of the proposed method decreases rapidly, eventually reaching a minimum, indicating that the selected features and their extraction process are efficient and stable, and the model and data are highly matched.

[0136] Figure 12 and Figure 13 The confusion matrices of the recognition results from the baseline network using only time-frequency maps and the method of this invention are shown respectively. In the confusion matrix, the diagonal elements represent the number of correctly classified samples, and the off-diagonal elements represent the number of misclassified samples. Comparing the two figures, it can be seen that... Figure 12 The baseline network in the text exhibits significant confusion when dealing with combined interference of blocking and deception, failing to effectively distinguish between this combined interference and a single blocking interference. Figure 13 In this invention, the confusion matrix exhibits higher numerical concentration on the diagonal and a significant reduction in values ​​off-diagonally. This indicates that the invention effectively solves the recognition ambiguity problem caused by the obscuring of deceptive interference features by blocking interference in the time-frequency domain by fusing time-domain features.

[0137] Figure 14 The results demonstrate the recognition accuracy of the method of this invention on seven test sets with different deception interference interference ratios (-20dB to 10dB) under the condition of fixed interference suppression parameters. The curves in the figure show that the recognition accuracy of the method of this invention remains at a high level across the entire tested interference ratio range. This result indicates that the technical solution of this invention has good adaptability to variations in the power of interference signals.

[0138] In summary, the experimental results objectively demonstrate through multi-dimensional data comparison that the multi-domain representation method, guided vector extraction mechanism, and adaptive deep feature fusion network structure proposed in this invention can effectively improve the recognition accuracy and robustness of various complex radar suppression, deception, and compound interference under low signal-to-noise ratio conditions.

[0139] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A radar suppression, deception, and composite interference identification technology under low signal-to-noise ratio conditions, characterized in that, Includes the following steps: S1. Obtain the interference pulse sequence of the radar echo; S2. Perform multi-domain characterization on the interference pulse sequence, and generate in parallel a time-domain sequence image, a time-frequency image, and noise power features corresponding to the interference pulse sequence; S3. Based on the time-domain sequence image and the time-frequency image, extract time-domain features and time-frequency domain features respectively using a feature extraction network; S4. Input the time-domain features, time-frequency domain features, and noise power features into the cross-fusion network, and the cross-fusion network performs deep feature fusion on the time-domain features, time-frequency domain features, and noise power features to output a comprehensive feature vector; S5. Input the comprehensive feature vector into the classifier, and the classifier outputs the finally determined interference category.

2. The radar suppression, deception, and composite interference identification technology under low signal-to-noise ratio conditions according to claim 1, characterized in that, Step S2, which involves performing multi-domain characterization on the interference pulse sequence and generating in parallel a time-domain sequence image, a time-frequency image, and noise power features corresponding to the interference pulse sequence, includes: A short-time Fourier transform is performed on one pulse data in the interference pulse sequence to obtain an initial time-frequency distribution image. The initial time-frequency distribution image is then subjected to histogram equalization. Finally, the time-frequency distribution image after histogram equalization is combined with the initial time-frequency distribution image and subjected to image channel stitching and normalization to generate the time-frequency image. A pulse data in the interference pulse sequence is compressed to obtain a compressed time-domain signal. The compressed time-domain signal and the original time-domain signal before pulse compression are modulo-valued, downsampled by max pooling, and plotted into two-dimensional images. The two generated two-dimensional images are then stitched together and normalized to generate the time-domain sequence image. The noise power of the time-domain signal after pulse compression is statistically analyzed and used as a noise power characteristic.

3. The radar suppression, deception, and composite interference identification technology under low signal-to-noise ratio conditions according to claim 1, characterized in that, Step S4, which involves the cross-fusion network performing deep feature fusion on time-domain features, time-frequency domain features, and noise power features to output a comprehensive feature vector, includes: The time-domain deep features and time-frequency domain deep features obtained by the feature extraction network are input into the cross-fusion network for deep feature fusion. The resulting deep fused features are then combined with the noise power features statistically derived from the interference pulse sequence to form the comprehensive feature vector.

4. The radar suppression, deception, and composite interference identification technology under low signal-to-noise ratio conditions according to claim 3, characterized in that, The steps for deep feature fusion by inputting the temporal deep features and time-frequency deep features obtained from the feature extraction network into the cross-fusion network include: The time-domain features and time-frequency domain features are input into a multilayer perceptron consisting of multiple fully connected layers. The multilayer perceptron performs feature dimensionality reduction and nonlinear mapping to output the fused features.

5. The radar suppression, deception, and composite interference identification technology under low signal-to-noise ratio conditions according to claim 1, characterized in that, The feature extraction network, the cross-fusion network, and the classification module together structurally constitute a unified network, which includes a feature extraction module and a cross-fusion module. The feature extraction module includes a time-domain feature extraction branch for processing the time-domain sequence image and a time-frequency domain feature extraction branch for processing the time-frequency image, so as to generate the time-domain features and time-frequency domain features respectively; The cross-fusion module is used to receive the time-domain features and the time-frequency domain features, and perform deep feature fusion.

6. The radar suppression, deception, and composite interference identification technology under low signal-to-noise ratio conditions according to claim 5, characterized in that, The temporal feature extraction branch and the time-frequency domain feature extraction branch each include multiple stacked residual convolutional modules, used to progressively extract deep features from the temporal sequence image and the time-frequency image.

7. The radar suppression, deception, and composite interference identification technology under low signal-to-noise ratio conditions according to claim 5, characterized in that, The cross-fusion module includes one or more fully connected layers, forming a multilayer perceptron.

8. The radar suppression, deception, and composite interference identification technology under low signal-to-noise ratio conditions according to claim 1, characterized in that, Step S5, which involves inputting the comprehensive feature vector into the classifier and having the classifier output the final determined interference category, includes: The comprehensive feature vector is input into one or more fully connected layers for feature dimension adjustment and nonlinear mapping. The features processed by the one or more fully connected layers are input into the final fully connected layer and the Softmax activation function to calculate and output the final interference category of the interference pulse sequence.

9. A radar suppression, deception, and combined interference identification device under low signal-to-noise ratio conditions, applied to the radar suppression, deception, and combined interference identification technology under low signal-to-noise ratio conditions as described in any one of claims 1-8, characterized in that, The device includes: The data characterization module is used to acquire the interference pulse sequence of the radar echo and perform multi-domain characterization on the interference pulse sequence to generate corresponding time-domain sequence images, time-frequency images and noise power features in parallel. The feature extraction module is configured as a feature extraction network, used to extract and generate time-domain features and time-frequency features based on the time-domain sequence image and the time-frequency image, respectively. The cross-fusion module, configured as a cross-fusion network, is used to receive the time-domain features, the time-frequency domain features, and the noise power features, and to perform adaptive deep feature fusion on the received features to output a comprehensive feature vector. The classification decision module is configured as a classifier to receive the comprehensive feature vector and output the finally determined interference category.

Images