An open-source radar jamming pattern recognition method, apparatus, and electronic device

Abstract

This invention discloses an open-set radar interference pattern recognition method, apparatus, and electronic device. The method includes: matched filtering of the radar received signal to construct a dual-modal input of a one-dimensional range sequence and a two-dimensional time-frequency matrix; feature extraction via a dual-branch network and adaptive fusion based on cosine similarity; introducing cross-modal consistency regularization during the training phase, constructing pseudo-unknown samples by shuffling intra-batch modes, and combining energy constraint loss to compress the energy of known classes and increase the energy of unknown classes to form a clear decision boundary; and adaptively setting a threshold based on the energy distribution of the validation set during the inference phase to achieve accurate identification of known interference and effective rejection of unknown interference. This invention overcomes the performance degradation defects of traditional closed-set methods in the face of unknown interference, and improves the generalization ability, robustness, and engineering practicality of radar interference recognition in complex electromagnetic environments.

Description

Technical Field

[0001] This invention relates to the fields of radar electronic countermeasures, signal processing and artificial intelligence, and more specifically, to an open set radar jamming pattern recognition method, apparatus and electronic equipment that integrates multimodal signal features. Background Technology

[0002] Jamming pattern recognition is a core enabling technology for cognitive radar systems to perceive complex electromagnetic environments and support anti-jamming decisions. Traditional jamming pattern recognition methods are mostly based on the closed-set assumption, which assumes that all jamming types appearing during the testing phase already exist in the training database. However, with the rapid development of advanced active jamming technologies such as digital radio frequency memory, adversaries can flexibly generate new or unseen jamming patterns, making it impossible for existing databases to exhaustively cover all jamming types. In open combat scenarios, traditional closed-set classifiers may incorrectly force unknown jamming to be classified as known jamming, not only causing distorted recognition results but also potentially triggering incorrect anti-jamming strategies, seriously threatening the survivability and mission safety of the radar system.

[0003] Existing research attempts to introduce open set recognition into the field of interference identification, but most methods rely on single-modal features. In complex electromagnetic countermeasures environments, a single mode is insufficient to fully characterize the temporal energy distribution and time-frequency evolution of interference signals, and known and unknown classes are prone to overlap in the feature space, leading to blurred boundaries in open set recognition and low rejection rates for unknown interference. Furthermore, existing discriminative open set recognition methods often rely on confidence thresholds or tail statistical modeling, lacking explicit constraints on the physical consistency of radar signals, and exhibiting insufficient robustness to fluctuations in interference-to-noise ratio and the sudden emergence of unknown patterns. Therefore, there is an urgent need for an open set radar interference pattern recognition method that can fully utilize the multimodal complementary characteristics of radar signals and explicitly construct decision boundaries for known and unknown energy in the feature space. Summary of the Invention

[0004] Purpose of the invention: The present invention aims to overcome the technical defects of existing closed-set interference identification methods, such as significant performance degradation when facing unknown interference, insufficient single-modal feature representation capability, and ambiguous decision boundaries of open sets. It provides an open-set radar interference pattern recognition method and device that integrates multi-modal signal features to improve the generalization ability, robustness, and unknown interference rejection rate of radar interference identification in complex electromagnetic environments.

[0005] Technical Solution: To achieve the above objectives, the present invention provides the following technical solution:

[0006] An open-set radar jamming pattern recognition method includes the following steps:

[0007] Step S1: Acquire the radar received signal and perform matched filtering. Based on the matched filtering output signal, construct a one-dimensional range amplitude sequence and a two-dimensional time-frequency amplitude matrix, respectively, as inputs to the time mode and the visual mode.

[0008] Step S2: Input the temporal modality and visual modality into a dual-branch coding network, extract the temporal modality features and visual modality features respectively, and perform cross-modal feature fusion in the high-level semantic space to obtain the fused feature representation;

[0009] Step S3: During the model training phase, a cross-modal consistency regularization strategy is introduced. Pseudo-out-of-distribution sample pairs are constructed by randomly shuffling the modalities within a batch. An energy upper limit constraint is applied to the known sample pairs with consistent modalities using an energy constraint loss function, and an energy lower limit constraint is applied to the pseudo-out-of-distribution sample pairs with inconsistent modalities, so as to optimize the energy distribution boundary of the feature space.

[0010] Step S4: In the model inference stage, calculate the energy score of the sample to be identified and compare it with the decision threshold determined based on the energy distribution of the known class validation set. Based on the comparison result, output the identification result of the known interference pattern category or the rejection result of the unknown interference pattern.

[0011] Furthermore, in step S1: the impulse response processed by the matched filter is the time-delay conjugate inversion of the radar transmitted signal. ,in The signal is the radar transmission signal, t is the time variable, and * denotes conjugate operation; the one-dimensional range-dimensional amplitude sequence is the amplitude sequence of the matched filter output signal. This is used to characterize the energy distribution and envelope structure of the signal in the distance dimension; the two-dimensional time-frequency amplitude matrix is ​​obtained by performing a short-time Fourier transform on the matched filter output signal to obtain a complex time-frequency matrix. And extract its magnitude matrix It is used to characterize the non-stationary modulation structure and time-frequency evolution of interference signals, among which For frequency variables.

[0012] Furthermore, the dual-branch coding network includes a temporal feature encoder and a visual feature encoder;

[0013] The time feature encoder uses a one-dimensional convolutional neural network to extract the distance dimension energy distribution and envelope structure, and outputs a time modality feature vector of the first preset dimension.

[0014] The visual feature encoder uses a convolutional neural network-Transformer hybrid architecture to extract local spatial features and global time-frequency dependence, and outputs a visual modality feature vector with a second preset dimension.

[0015] When fusing cross-modal features, the features of the two modalities are projected into a unified embedding space, the cosine similarity is calculated and normalized to the modality consistency weight, and the initial fused features after splicing interaction are adaptively modulated to obtain the final fused feature representation.

[0016] Furthermore, the energy-constrained loss function is constructed as follows:

[0017] Calculate sample energy value based on network output Logits ,in Let be the Logits value of the k-th class, T be the temperature parameter, and K be the number of known interference classes;

[0018] Constructing total loss ,in For cross-entropy loss, The loss coefficient, Given the upper limit constraint term for the energy of the known samples, This is a lower bound constraint term for the energy of out-of-distribution samples in the pseudo-distribution. , To predetermine the energy boundary, , The regularization coefficient is used.

[0019] Furthermore, the decision threshold for the reasoning stage is obtained as follows:

[0020] Input the known class validation set into the trained network, calculate the energy score of each sample and form an energy set, sort the samples in ascending order, and take the energy value corresponding to the preset percentile as the decision threshold. ;

[0021] The judgment rule is: if the energy score of the sample to be identified is... If so, it is determined to be a known class and the class corresponding to the maximum Logits is output; if If it does not, it is determined to be an unknown interference mode and rejection is executed.

[0022] To achieve the above objectives, the present invention also provides an open-set radar jamming pattern recognition device that integrates multimodal signal features, comprising:

[0023] The signal preprocessing module is used to acquire radar received signals and perform matched filtering. Based on the matched filtering output signal, a one-dimensional range-dimensional amplitude sequence and a two-dimensional time-frequency amplitude matrix are constructed as inputs to the time mode and the visual mode, respectively.

[0024] The feature extraction and fusion module is used to input the temporal modality and visual modality into a dual-branch coding network, extract temporal modality features and visual modality features respectively, and perform cross-modal feature fusion in a high-level semantic space to obtain a fused feature representation;

[0025] The model training module is used to introduce a cross-modal consistency regularization strategy during the model training phase. It constructs pseudo-out-of-distribution sample pairs by randomly shuffling the modalities within a batch, applies an energy upper limit constraint to known sample pairs with consistent modalities using an energy constraint loss function, and applies an energy lower limit constraint to pseudo-out-of-distribution sample pairs with inconsistent modalities, so as to optimize the energy distribution boundary of the feature space.

[0026] The open set inference module is used to calculate the energy score of the sample to be identified during the model inference stage, and compare it with the decision threshold determined based on the energy distribution of the known class validation set. Based on the comparison result, it outputs the identification result of the known interference pattern category or the rejection result of the unknown interference pattern.

[0027] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the open set radar jamming pattern recognition method described above.

[0028] To achieve the above objectives, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the open set radar jamming pattern recognition method as described above.

[0029] Compared with the prior art, the present invention has the following beneficial effects: (1) By constructing a dual-modal representation of a one-dimensional distance-dimensional amplitude sequence and a two-dimensional time-frequency amplitude matrix, and combining a dual-branch coding and cosine similarity adaptive fusion mechanism, the complementary information of radar signals in the distance domain and time-frequency domain is fully explored, effectively alleviating the problem of category confusion caused by structural similarity interference; (2) A cross-modal consistency regularization strategy is proposed, which uses random shuffling of intra-batch modes to construct pseudo-out-of-distribution samples, explicitly introduces out-of-distribution boundary information, and combines energy constraint loss to compress the energy of known classes and increase the energy of pseudo-unknown classes in the feature space, forming a clear and stable open set decision boundary; (3) The inference stage adopts a threshold adaptively set based on the energy distribution of the known class verification set, without the need to obtain unknown interference samples in advance or introduce additional posterior calibration modules, realizing the energy consistency between training and inference, significantly improving the system's ability to reject unknown interference modes and the recognition accuracy of known classes, and possessing strong robustness and engineering deployment potential in the actual electromagnetic confrontation environment with strong noise and multiple interferences. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the open-set radar interference pattern recognition method that integrates multimodal signal features according to the present invention.

[0031] Figure 2 A schematic diagram of the network structure for temporal and visual feature extraction and fusion;

[0032] Figure 3 This is a flowchart of the open set inference algorithm based on the energy model of this invention;

[0033] Figure 4 A diagram comparing the recognition accuracy of known and unknown classes under different unknown category settings;

[0034] Figure 5 A schematic diagram comparing the energy score distribution under different ablation configurations with the confusion matrix under open set conditions;

[0035] Figure 6 This is a schematic diagram of an experimental scenario where a Ku-band radar system in the actual field is subjected to active interference.

[0036] Figure 7 This is a schematic diagram showing the time-frequency characterization of the matched filtering results and typical interferences for actual field data. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0038] Example 1: Identification Method Flow

[0039] This embodiment provides an open-set radar jamming pattern recognition method that integrates multimodal signal features, such as... Figure 1 As shown, it mainly includes four stages: signal preprocessing and dual-modal construction, dual-branch feature extraction and adaptive fusion, cross-modal consistency regularization training, and open set inference based on energy model.

[0040] Step S1: Signal preprocessing and dual-mode construction

[0041] The baseband signal received by the radar receiver during each pulse repetition interval It can be modeled as a target echo. Active interference signals With additive white Gaussian noise The superposition of signals. First, the received signal is subjected to matched filtering, and the impulse response of the matched filter is set to the time delay conjugate inversion of the radar transmitted signal. ,in The signal is the radar transmission signal, t is the time variable, * indicates conjugate operation, and the filtered output is... ,in This represents the convolution operation.

[0042] Furthermore, a dual-modal input is constructed based on the filtered output:

[0043] (1) One-dimensional distance-dimensional sequence (time mode): Extract the amplitude sequence of the matched filter output. It characterizes the energy distribution and envelope structure of the signal in the distance dimension, and the input dimension can be configured to 1×4096;

[0044] (2) Two-dimensional time-frequency matrix (visual modality): for The complex time-frequency matrix is ​​obtained by performing a short-time Fourier transform. Extract its magnitude matrix The non-stationary modulation structure and time-frequency evolution of the interference signal are characterized, among which... This is a frequency variable. The input dimension can be configured to 1×128×125.

[0045] The two modes represent the signal in a multidimensional way from the one-dimensional distance domain and the two-dimensional time-frequency domain, and have structural complementarity.

[0046] Step S2: Dual-branch feature extraction and adaptive fusion

[0047] like Figure 2 As shown, design a dual-branch coding network:

[0048] The time branch uses a one-dimensional convolutional neural network encoder. right Temporal feature mining is performed. The network consists of a three-layer Conv1D-BatchNorm-ReLU structure, which, after adaptive average pooling, is projected onto a 128-dimensional embedding space through a fully connected layer to output temporal modality feature vectors. .

[0049] The visual branch uses a CNN-Transformer hybrid encoder. right Visual feature mining is performed. First, local spatial features are extracted and downsampled using two layers of Conv2D, then aligned to a uniform resolution via adaptive pooling. Next, patch embedding is applied to project the blocks into the sequence space, which is then input into a multi-layer Pre-Norm Transformer encoder for global dependency modeling. After sequence average pooling and LayerNorm, the output is a visual modality feature vector. .

[0050] Similarity-guided cross-modal fusion: and The cosine similarity between the two is calculated by mapping them to a unified 256-dimensional space using linear projection matrices. Normalization yields modal consistency weights The two features are concatenated and then input into a fusion module consisting of Linear, LayerNorm, and GELU to obtain the initial fused features. The final output is a weighted fusion feature. This mechanism enhances fused representations when modalities are consistent and automatically suppresses inconsistent information when modalities conflict.

[0051] Step S3: Cross-modal consistency regularization training

[0052] During the training phase, two types of sample pairs are constructed for joint optimization:

[0053] (1) Modality consistent real sample pairs: used to learn discriminative features of known interference.

[0054] (2) Pseudo-OOD sample pairs: These are obtained by randomly permuting the visual modality index within the training batch. Construction. Since time-series modes and time-frequency modes come from different physical instances, there are cross-modal semantic conflicts, which can be equivalent to samples outside the unknown interference distribution.

[0055] Define the sample energy function: For an input sample x, the energy value is defined as follows: ,in Let be the Logits value of the k-th class, T be the temperature parameter, and K be the number of known interference classes.

[0056] Constructing an energy-constrained loss: Imposing an energy upper limit constraint on known samples. ,in Given an upper bound on energy; apply a lower bound on energy to the Pseudo-OOD samples. ,in This is a lower bound for the pseudo-unknown energy. The total loss function is... ,in , The regularization coefficient is . For cross-entropy loss, This represents the cross-entropy loss coefficient. This loss optimizes the network, ensuring that known classes are compactly distributed in the energy space, while unknown classes are pushed to higher energy regions. Preferably, , , , .

[0057] Step S4: Open Set Inference Based on Energy Model

[0058] The reasoning phase process is as follows: Figure 3 As shown:

[0059] (1) Threshold estimation: Input the known class validation set into the trained network, calculate the energy score of each sample to form a set. Sort in ascending order and take the first position. Percentiles, preferred As a decision threshold .

[0060] (2) Open set determination: for the test sample Calculate energy .like If it is determined to be a known class, output... ;like It was determined to be unknown interference and was rejected.

[0061] Example 2: Identification Device

[0062] This embodiment provides an open-set radar jamming pattern recognition device that integrates multimodal signal features, including:

[0063] The signal preprocessing module is configured to execute step S1, completing matched filtering and bimodal sequence / matrix construction.

[0064] The feature extraction and fusion module is configured to perform step S2, which extracts and fuses multimodal features through a dual-branch network of 1D-CNN and CNN-Transformer.

[0065] The model training module is configured to execute step S3, which implements batch modal shuffling to construct Pseudo-OOD samples and optimizes network parameters based on energy-constrained loss.

[0066] The open set inference module is configured to execute step S4, which adaptively calculates the threshold based on the energy distribution of the validation set and performs energy determination and category output for the test sample.

[0067] The device can be integrated into a radar signal processor, an edge computing terminal, or an anti-interference decision system, with each module implemented in the form of software algorithms, hardware logic circuits, or a combination of software and hardware.

[0068] Example 3: Computer-readable storage media and electronic devices

[0069] This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0070] This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method described in Embodiment 1. The electronic device may be a signal processing unit of a vehicle-mounted / shipborne / airborne radar or a ground-based electronic countermeasures command and control terminal.

[0071] Example 4: Experimental Verification and Effect Description

[0072] To verify the effectiveness of this invention, systematic tests were conducted using simulation and experimental data. The simulation settings covered eight typical interference patterns, with an interference-to-noise ratio (IRR) ranging from 0 to 30 dB.

[0073] Under the leave-one-class-unknown experimental protocol, when the regularization coefficient is optimized, the experimental results are as follows: Figure 4 As shown, the accuracy of known class recognition in this invention is stable at around 0.97, and the accuracy of unknown class recognition reaches above 0.95 in most cases, indicating that the model has good generalization and rejection capabilities under open set conditions.

[0074] Ablation experiments were designed with SSFJ as an unknown class, comparing the following four configurations: using only temporal modality features, using only visual modality features, removing cross-modal consistency regularization, and the complete method. The confusion matrix results of the energy score distribution and open set conditions under the ablation configurations are as follows: Figure 5 As shown in the figure. Ablation experiments confirm that the combined effect of bimodal fusion and consistency regularization improves the overall performance of the model in open set scenarios, including the accuracy of known class recognition and the ability to reject unknown classes.

[0075] In actual Ku-band radar field countermeasures experiments, the actual field scenarios are as follows: Figure 6 As shown, the radar system used in the experiment operates in the Ku band, and its specific operating parameters are shown in Table 1. In the detection mode, under a detection mode containing target echoes and strong interference, the visualization result of the radar received signal after pulse compression processing is shown below. Figure 7 As shown, each type of interference contains 2048 pulses. The method of this invention achieves an identification accuracy of 0.89 for known interference and an rejection accuracy of 0.95 for unknown interference. These data demonstrate that, through the synergistic effect of multimodal feature complementarity and energy boundary constraints, this method can effectively address signal degradation and the emergence of unknown patterns in practical electromagnetic countermeasures environments, possessing strong engineering practical value.

[0076] Table 1 Operating parameters of the Ku-band radar system

[0077]

[0078] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An open-set radar jamming pattern recognition method, characterized in that, Includes the following steps: S1. Acquire the radar received signal and perform matched filtering. Based on the matched filtering output signal, construct a one-dimensional range-dimensional amplitude sequence and a two-dimensional time-frequency amplitude matrix, respectively, as inputs to the time mode and the visual mode. S2. Input the temporal modality and visual modality into a dual-branch coding network, extract the temporal modality features and visual modality features respectively, and perform cross-modal feature fusion in the high-level semantic space to obtain the fused feature representation; S3. During the model training phase, a cross-modal consistency regularization strategy is introduced. Pseudo-out-of-distribution sample pairs are constructed by randomly shuffling the modalities within a batch. An energy upper limit constraint is applied to the known sample pairs with consistent modalities using an energy constraint loss function, and an energy lower limit constraint is applied to the pseudo-out-of-distribution sample pairs with inconsistent modalities, so as to optimize the energy distribution boundary of the feature space. S4. During the model inference stage, calculate the energy score of the sample to be identified and compare it with the decision threshold determined based on the energy distribution of the known class validation set. Based on the comparison result, output the recognition result of the known interference pattern or the rejection result of the unknown interference pattern.

2. The method according to claim 1, characterized in that, In step S1: The impulse response processed by the matched filter is the time-delay conjugate inversion of the radar transmitted signal. ,in The signal is the radar transmission signal, t is the time variable, and * denotes conjugate operation; the one-dimensional range-dimensional amplitude sequence is the amplitude sequence of the matched filter output signal. This is used to characterize the energy distribution and envelope structure of the signal in the distance dimension; the two-dimensional time-frequency amplitude matrix is ​​obtained by performing a short-time Fourier transform on the matched filter output signal to obtain a complex time-frequency matrix. And extract its magnitude matrix. It is used to characterize the non-stationary modulation structure and time-frequency evolution of interference signals, among which For frequency variables.

3. The method according to claim 1, characterized in that, The dual-branch coding network in step S2 includes a temporal feature encoder and a visual feature encoder, and the cross-modal feature fusion specifically includes: The temporal feature encoder uses a one-dimensional convolutional neural network to perform hierarchical convolutional downsampling and adaptive average pooling on the one-dimensional distance dimension amplitude sequence to extract temporal modality feature vectors. The visual feature encoder uses a convolutional neural network-Transformer hybrid encoder to extract local spatial features and model global dependencies on the two-dimensional time-frequency amplitude matrix, thereby extracting visual modality feature vectors. The temporal modality feature vector and the visual modality feature vector are projected into a unified embedding space, and the cosine similarity between them is calculated and normalized to a modality consistency weight. The two modal feature vectors are concatenated and passed through a fully connected interactive network to obtain the initial fused features. The modal consistency weights are then used to adaptively modulate the initial fused features to obtain the final fused feature representation.

4. The method according to claim 1, characterized in that, Step S3, which involves applying an upper energy limit constraint to known sample pairs with consistent modes and an lower energy limit constraint to pseudo-out-of-distribution sample pairs with inconsistent modes using an energy-constrained loss function, specifically includes: Calculate sample energy value based on network output Logits ,in For the input sample, T is the temperature parameter, and K is the number of known interference categories. The Logits value for the k-th class; Constructing the energy loss function ,in Given the upper limit constraint term for the energy of the sample, Given an upper bound on the energy level, This is a lower bound constraint term for the energy of out-of-distribution samples in the pseudo-distribution. This is a pseudo-unknown energy lower bound. and The regularization coefficient is used. Combined with cross-entropy loss , with total loss function Drive network parameter optimization, This represents the cross-entropy loss coefficient.

5. The method according to claim 1, characterized in that, Step S4, which involves obtaining the decision threshold determined based on the energy distribution of the known class verification set and outputting the recognition or rejection result based on the comparison result, specifically includes: Input the known class validation set into the trained network, calculate the energy score of each sample and form an energy set, sort the samples in ascending order, and take the energy value corresponding to the preset percentile as the decision threshold. ; For the test sample to be identified Calculate its energy score ,like If the value is 0, it is determined to be a known class, and the class corresponding to the largest Logits value is taken as the recognition result; if the value is 0, it is determined to be a known class. If it does not, it is determined to be an unknown interference mode and rejection is executed.

6. An open-set radar jamming pattern recognition device, characterized in that, include: The signal preprocessing module is used to acquire radar received signals and perform matched filtering. Based on the matched filtering output signal, a one-dimensional range-dimensional amplitude sequence and a two-dimensional time-frequency amplitude matrix are constructed as inputs to the time mode and the visual mode, respectively. The feature extraction and fusion module is used to input the temporal modality and visual modality into a dual-branch coding network, extract temporal modality features and visual modality features respectively, and perform cross-modal feature fusion in a high-level semantic space to obtain a fused feature representation; The model training module is used to introduce a cross-modal consistency regularization strategy during the training phase. It constructs pseudo-out-of-distribution sample pairs by randomly shuffling the modalities within a batch and optimizes the energy distribution boundary of the feature space using an energy-constrained loss function. The open set inference module is used to calculate the energy score of the sample to be identified during the inference phase and compare it with the decision threshold determined based on the energy distribution of the known class validation set. Based on the comparison result, it outputs the identification result of the known interference pattern category or the rejection result of the unknown interference pattern.

7. The apparatus according to claim 6, characterized in that, The feature extraction and fusion module is specifically configured to perform the temporal modality feature extraction, visual modality feature extraction, and adaptive cross-modal feature fusion operation based on cosine similarity as described in claim 3.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 5.

Images