Radar interference identification method of multi-representation domain feature interactive fusion and graph structure semantic classifier

Abstract

The invention discloses a radar interference identification method based on multi-representation domain feature interactive fusion and a graph structure semantic classifier, and the method comprises the steps: constructing a multi-representation domain feature interactive fusion module, taking a time domain waveform, a spectrogram and a time-frequency spectrogram as input data, and outputting fusion features; constructing a graph structure semantic classifier, inputting the fused features into the graph structure semantic classifier, analyzing semantic association of different types of signals by the graph structure semantic classifier, and outputting a classification result; dividing the data set into a training set and a test set, and storing network parameters with the best classification effect in the training set according to a classification result; and loading network parameters, and inputting the test set into the radar interference identification network for processing to obtain an identification result. According to the method, the interference signals can be fully represented, the identification and processing capability of the interference signals is improved, robustness representation is carried out on the interference signals, and the accuracy of radar interference identification is improved.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology and relates to a radar interference recognition method using interactive fusion of multi-representation domain features and a graph structure semantic classifier. Background Art

[0002] Radar is the core technical equipment of modern electromagnetic detection systems. However, with the rapid evolution of electronic countermeasure technology, especially the breakthrough development of intelligent active jamming technology, various new jamming methods have posed severe challenges to the target detection and trajectory tracking capabilities of radar systems. There is an urgent need to develop efficient and reliable jamming identification and suppression technologies.

[0003] In radar anti-interference technology, interference signal recognition, a fundamental component, focuses on accurately identifying the presence and pattern of interference within complex electromagnetic environments. This provides a basis for optimizing subsequent anti-interference strategies and improving interference suppression effectiveness. Current mainstream interference identification techniques rely primarily on single-domain feature analysis or multi-domain feature fusion. Single-domain analysis methods focus on extracting features from time-domain waveforms or a single transform domain, and are widely used in engineering due to their computational efficiency. However, the current electromagnetic environment is becoming increasingly complex, and the types of interference faced by radar systems are becoming increasingly diverse and complex. In this context, traditional single-domain feature analysis methods typically focus on single-dimensional representations, such as time-domain waveform changes, frequency-domain energy distribution, or joint time-frequency characteristics. When the interference-noise ratio (JNR) is low, the interference signal weakens across all domains, making it difficult to construct a complete feature representation system using single-domain analysis alone, significantly reducing identification accuracy.

[0004] Multi-domain joint characterization technology can effectively improve system performance indicators by integrating the complementary characteristics of multi-domain information spaces. Given the time-frequency coupling characteristics of radar jamming signals, their different domain representations contain differentiated feature dimensions and multidimensional descriptive information. Existing fusion algorithms often use a fusion architecture that extracts features from different domains followed by feature splicing or linear superposition. This superficial fusion approach has two limitations: first, it fails to fully exploit the nonlinear correlation characteristics between the time and frequency domains; second, the distribution differences between different feature spaces easily lead to feature conflicts, which restricts the improvement of the discriminative ability of the fused features. Summary of the Invention

[0005] The present invention aims to solve the technical problems of insufficient multi-domain feature fusion representation and low recognition accuracy of radar interference signals. The present invention provides a radar interference recognition method that integrates interactive fusion of multi-domain features and a graph-structured semantic classifier. The technical solution adopted is:

[0006] The radar interference recognition method based on interactive fusion of multi-representation domain features and graph structure semantic classifier includes the following steps:

[0007] S1. Construct a multi-domain feature interactive fusion module, which takes time domain waveform, spectrogram and time-spectrogram as input data and outputs fused features;

[0008] S2. Construct a graph structure semantic classifier, input the fusion features into the graph structure semantic classifier, and the graph structure semantic classifier analyzes the semantic associations of different categories of signals and outputs a classification result;

[0009] S3, dividing the data set into a training set and a test set, and saving the network parameters with the best classification effect in the training set according to the classification results;

[0010] S4. Load the network parameters, input the test set into the radar interference recognition network for processing, and obtain a recognition result.

[0011] In one embodiment of the present invention, the multi-representation domain feature interactive fusion module includes: a shallow feature extraction module, a cross-domain feature extraction module and a time-frequency feature extraction module;

[0012] The shallow feature extraction module is used to extract time domain features and frequency domain features;

[0013] The cross-domain feature extraction module uses the cross-domain attention mechanism to achieve information interaction between different domains, extract deep signal features, and output deep time domain features and deep frequency domain features;

[0014] The time-frequency feature extraction module is used to capture the joint characteristics of the signal between time and spectrum and output time-frequency domain features;

[0015] The time domain deep features, the frequency domain deep features and the time-frequency domain features are spliced to form fusion features.

[0016] In one embodiment of the present invention, the shallow feature extraction module is composed of two cascaded residual blocks, each of which integrates two one-dimensional convolution layers with a kernel size of 3×1 and a batch normalization layer, and the residual block connection path is realized by the series connection of a convolution layer with a kernel size of 1×1 and a double residual block.

[0017] In one embodiment of the present invention, it is assumed that the time domain signal is a radar interference signal time series Frequency domain signal is a single-sided spectrum signal N is the number of samples, T is the length of the time domain signal, and F is the length of the frequency domain signal. The output feature of the shallow feature extraction module is expressed as:

[0018]

[0019] In formula (1) and formula (2), and is the corresponding nonlinear mapping function, output feature and They represent time domain features and frequency domain features respectively, and d represents the output feature length.

[0020] In one embodiment of the present invention, the cross-domain feature extraction module outputs time-domain deep features, including:

[0021] The time domain features As the query vector Q t , Q t =Z t W Q ;

[0022] The frequency domain characteristics As the frequency domain key vector K f and frequency domain value vector V f , K f =Z f W K , V f =Z f W V ;

[0023] Calculated by scaling the dot product attention:

[0024]

[0025] In formula (3), Q t is the query vector, K f is the frequency domain key vector, V f is the frequency domain value vector, is the learnable parameter matrix, d k represents the intermediate characteristic length, Used to control the dimension of the dot product;

[0026] The enhanced time domain features obtained through residual connection are expressed as:

[0027] Z t′ =Attention(Q t ,K f ,V f )+Z t (4)

[0028] In formula (4), Z t′ is the enhanced time domain feature;

[0029] After the cross-domain attention, the feedforward network and the residual block are cascaded. The feedforward network consists of two fully connected layers. The feature space is expanded through nonlinear transformation. The residual block is used to extract the deep features of the signal and obtain the deep time domain features of the signal. d mIndicates the output feature length.

[0030] In one embodiment of the present invention, the cross-domain feature extraction module outputs frequency domain deep features, including:

[0031] The frequency domain features As the frequency domain query vector Q f , Q f =Z f W Q ';

[0032] The time domain characteristics As the time domain key vector K t and the time domain value vector V t , K t =Z t W K ', V t =Z t W V ';

[0033] Calculated by scaling the dot product attention:

[0034]

[0035] In formula (5), Q f is the frequency domain query vector, K t is the time domain key vector, V t is the time domain value vector, is the learnable parameter matrix, d k represents the intermediate characteristic length, Used to control the dimension of the dot product.

[0036] The enhanced frequency domain features are obtained through residual connection:

[0037] Z f' =Attention(Q f ,K t ,V t )+Z f (6)

[0038] In formula (6), Z f' is the enhanced frequency domain feature;

[0039] After the cross-domain attention, the feedforward network and the residual block are cascaded. The feedforward network consists of two fully connected layers. The feature space is expanded through nonlinear transformation. The residual block is used to extract the deep features of the signal and obtain the deep frequency domain features of the signal. d m Indicates the output feature length.

[0040] In one embodiment of the present invention, the time-frequency feature extraction module extracts time-frequency domain features, including:

[0041] The time-frequency domain feature module is designed to capture the joint characteristics of the signal between time and spectrum, and output the time-frequency domain features Expressed as:

[0042]

[0043] In formula (7), represents the time-frequency feature extraction module, X tf and w tf are the time-frequency domain signals and network parameters respectively.

[0044] In one embodiment of the present invention, the fusion feature is represented as:

[0045]

[0046] In formula (8), F represents the fusion feature, d f =3·d m .

[0047] In one embodiment of the present invention, step S2 includes:

[0048] said constructing an inter-class relationship graph by utilizing said fusion features;

[0049] Based on the inter-class relationship graph, two symmetric adjacency matrices are defined, namely a comprehensive relationship matrix and a core relationship matrix;

[0050] The comprehensive relationship matrix is expressed as:

[0051]

[0052] In formula (9), is a comprehensive relationship matrix based on the Euclidean distance between the category label embedding vectors, l i and l j are the semantic embedding vectors of the i-th and j-th categories respectively, and η is a hyperparameter used to control the attenuation of distance to similarity;

[0053] The core relationship matrix is expressed as:

[0054]

[0055] In formula (10), is the core relationship matrix, which retains important core relationships by removing edges below the ξ quantile value. Representation matrix ξ percentile value of .

[0056] In one embodiment of the present invention, step S2 further includes:

[0057] The graph structure semantic classifier is constructed by two independent graph convolution functions. and The comprehensive relationship matrix and the core relationship matrix are processed separately and the features of the next layer are calculated, which can be expressed as:

[0058]

[0059] In formula (11) and formula (12), and They are and The degree matrix of and is a trainable weight matrix, σ(·) is the activation function;

[0060] The classification result is obtained by weighted summing the outputs of the comprehensive relationship matrix and the core relationship matrix, and is expressed as:

[0061]

[0062] In formula (13), o is the classification result, ρ is the hyperparameter used to balance the contribution of comprehensive relations and core relations, and L is the number of layers of the graph convolution classifier.

[0063] Beneficial effects of the present invention:

[0064] 1. The radar interference identification method of the present invention, which integrates interactive fusion of multi-domain features with a graph-structured semantic classifier, combines time domain information, frequency domain information, and time-frequency domain information to achieve dynamic complementarity and synergistic enhancement of time, frequency, and time-frequency domain features. This method fully characterizes the interference signal. Time domain information accurately reflects the transient variation characteristics of the interference signal, while frequency domain information clearly displays the frequency composition and energy distribution of the signal. Time-frequency domain information represents the relationship between energy, time, and frequency. By integrating these three types of information, the present invention can more comprehensively and deeply explore the intrinsic characteristics of the interference signal, providing a richer and more accurate data foundation for subsequent analysis and processing, thereby significantly improving the ability to identify and process interference signals.

[0065] 2. The radar interference identification method of the present invention, which integrates the interactive fusion of multi-representation domain features and the graph structure semantic classifier, introduces a graph structure semantic classifier. It uses the characteristics of the graph structure to effectively capture the complex relationship between signals and robustly characterize the interference signals. The graph structure semantic classifier can deeply explore the semantic associations between signals and maintain high classification accuracy and stability when facing complex and changeable interference signal environments. It effectively overcomes the problem of reduced classification performance of existing technologies in complex scenarios and improves the accuracy of radar interference identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a flow chart of a radar interference recognition method using interactive fusion of multi-representation domain features and graph structure semantic classifiers provided by an embodiment of the present invention;

[0067] Figure 2 This is a diagram of the overall network structure provided by an embodiment of the present invention;

[0068] Figure 3 Schematic diagram of the structure of the shallow feature extraction module provided by an embodiment of the present invention;

[0069] Figure 4 Schematic diagram of the structure of the cross-domain feature extraction module provided by an embodiment of the present invention;

[0070] Figure 5 Schematic diagram of an adjacency matrix of a radar jamming signal provided by an embodiment of the present invention;

[0071] Figure 6 is a schematic diagram of a graphical representation of a radar jamming signal provided by an embodiment of the present invention;

[0072] Figure 7 This is the T-SNE visualization result provided by the embodiment of the present invention;

[0073] Figure 8 is the confusion matrix of different methods provided by the embodiments of the present invention under JNR=0dB;

[0074] Figure 9 is the confusion matrix of different methods provided in the embodiment of the present invention under JNR=5dB;

[0075] Figure 10 is a schematic diagram of a difficult-to-distinguish interference signal provided by an embodiment of the present invention;

[0076] Figure 11 is the recognition accuracy rate of difficult-to-distinguish interference signals under the condition of JNR=0dB provided by the embodiment of the present invention;

[0077] Figure 12 This is the recognition accuracy rate of difficult-to-distinguish interference signals under the condition of JNR=5dB provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0078] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0079] This paper proposes a radar interference recognition method that combines interactive fusion of multi-domain features with a graph-structured semantic classifier. The multi-domain interactive fusion framework uses a cross-modal attention mechanism to achieve dynamic complementarity and synergistic enhancement of time, frequency, and time-frequency domain features. The graph-structured semantic classifier utilizes a graph convolutional network to model the comprehensive and core semantic relationships between signal categories, improving the robustness of the classification decision boundary.

[0080] Refer to the attached Figure 1 The radar interference recognition method based on interactive fusion of multi-representation domain features and graph structure semantic classifier includes the following steps:

[0081] S1. Construct a multi-domain feature interactive fusion module, which takes time domain waveform, spectrogram and time-spectrogram as input data and outputs fused features;

[0082] S2. Build a graph structure semantic classifier, fuse the features and input them into the graph structure semantic classifier. The graph structure semantic classifier analyzes the semantic associations of different categories of signals and outputs the classification results.

[0083] S3. Divide the data set into a training set and a test set, and save the network parameters with the best classification effect in the training set according to the classification results;

[0084] S4. Load the network parameters and input the test set into the radar interference recognition network for processing to obtain the recognition results.

[0085] The multi-domain feature interaction fusion module of the present invention includes: shallow feature extraction module, cross-domain feature extraction module and time-frequency feature extraction module; the shallow feature extraction module is used to extract time domain features and frequency domain features; the cross-domain feature extraction module uses the cross-domain attention mechanism to achieve information interaction between different domains, extract signal deep features, and output time domain deep features and frequency domain deep features; the time-frequency feature extraction module is used to capture the joint characteristics of the signal between time and spectrum, and output time-frequency domain features; the time domain deep features, frequency domain deep features and time-frequency domain features are spliced to form fusion features. The overall network structure diagram is shown in the attached figure. Figure 2 shown.

[0086] Aiming at the shallow feature extraction requirements of time domain signals and spectral signals, a general shallow feature extraction module is designed. The shallow feature extraction module consists of two cascaded residual blocks. Each residual block integrates two layers of one-dimensional convolutional layers with a kernel size of 3×1 and a batch normalization layer. The residual block connection path is realized by the concatenation of a convolutional layer with a kernel size of 1×1 and a double residual block, which effectively enhances the feature expression capability. The structure of the shallow feature extraction module is shown in the attached figure. Figure 3 shown.

[0087] Assume that the time domain signal is the radar interference signal time series Frequency domain signal is a single-sided spectrum signal N is the number of samples, T is the length of the time domain signal, F is the length of the frequency domain signal, and the output feature of the shallow feature extraction module is expressed as:

[0088]

[0089] In formula (1) and formula (2), and is the corresponding nonlinear mapping function, output feature and They represent time domain features and frequency domain features respectively, and d represents the output feature length.

[0090] The shallow feature extraction module adopts a dual residual block architecture, enhancing feature expression by cascading two residual units. The residual connection mechanism enables cross-layer information transfer, effectively alleviating the vanishing gradient problem in deep network training. This reduces optimization complexity while providing highly discriminative input features for subsequent feature fusion.

[0091] Based on the conjugate symmetry of Fourier transform, the time domain signal x t With the frequency domain signal x f There is a strict mathematical correspondence: time domain features can characterize the transient response characteristics of the signal, and frequency domain features can reflect the periodicity and energy distribution characteristics of the signal. In order to fully explore the complementarity of time and frequency features, the present invention proposes a cross-domain feature extraction module, as shown in the attached Figure 4 As shown in Figure 2, the core idea of this cross-domain feature extraction module is to achieve information interaction between different domains using the cross-domain attention mechanism. The following example uses frequency domain guidance for time domain feature enhancement to illustrate:

[0092] The time domain features As the query vector Q t , Q t =Z t W Q ; Frequency domain characteristics As the frequency domain key vector K f and frequency domain value vector V f , K f =Z f W K , V f =Z f W V ; Calculated by scaling dot product attention:

[0093]

[0094] In formula (3), Q t is the query vector, K f is the frequency domain key vector, V f is the frequency domain value vector, is the learnable parameter matrix, d k represents the intermediate characteristic length, Used to control the dimension of the dot product.

[0095] The enhanced time domain features obtained through residual connection are expressed as:

[0096] Z t′ =Attention(Q t ,K f ,V f )+Z t (4)

[0097] In formula (4), Z t′ is the enhanced time domain feature.

[0098] In order to further improve the feature expression ability, a feedforward network and a residual block are cascaded after the cross-domain attention. The feedforward network consists of two fully connected layers. The feature space is expanded through nonlinear transformation. The residual block is used to extract the deep features of the signal and obtain the deep time domain features of the signal. d m Indicates the output feature length.

[0099] Similarly, the frequency domain deep features of the signal are obtained

[0100] The frequency domain features As the frequency domain query vector Q f , Q f =Z f W Q ';Time domain characteristics As the time domain key vector K t and the time domain value vector V t , K t =Z t W K ', V t =Z t W V ';Calculate by scaling dot product attention:

[0101]

[0102] In formula (5), Q f is the frequency domain query vector, K t is the time domain key vector, V t is the time domain value vector, is the learnable parameter matrix, d k represents the intermediate characteristic length, Used to control the dimension of the dot product.

[0103] The enhanced frequency domain features are obtained through residual connection:

[0104] Z f' =Attention(Q f ,K t ,V t )+Z f (6)

[0105] In formula (6), Z f' is the enhanced frequency domain feature.

[0106] To further improve the feature expression capability, a feedforward network and a residual block are cascaded after the cross-domain attention. The feedforward network consists of two fully connected layers. The feature space is expanded through nonlinear transformation. The residual block is used to extract the deep features of the signal, and finally the deep frequency domain features of the signal are obtained. d m Indicates the output feature length.

[0107] In addition to extracting signal time domain and frequency domain features, time-frequency domain features are also introduced as a supplement to further improve feature representation capabilities. Time-frequency domain features are extracted through the time-frequency feature extraction module, which aims to capture the joint characteristics of the signal between time and spectrum and output time-frequency domain features. Expressed as:

[0108]

[0109] In formula (7), represents the time-frequency feature extraction module, X tf and w tf are the time-frequency domain signals and network parameters respectively.

[0110] Then the time domain deep feature D t , frequency domain deep features D f and time-frequency domain features D tf The three are spliced together to form a complete fusion feature, which is expressed as:

[0111]

[0112] In formula (8), F represents the fusion feature, d f =3·d m .

[0113] The multi-representation domain feature interactive fusion framework of the present invention realizes the dynamic complementarity and synergistic enhancement of time domain, frequency domain, and time-frequency domain features through a cross-modal attention mechanism. Through specific algorithms and architectures, it promotes the mutual complementation and reinforcement of time domain features, frequency domain features, and time-frequency domain features to form a more representative and discriminative comprehensive feature.

[0114] This paper proposes a graph-structured semantic classifier, whose core goal is to improve the effectiveness of radar jamming signal recognition by analyzing the semantic relationships between different categories. Unlike traditional classification methods, this solution addresses the limitations of traditional methods in capturing potential signal correlations. By introducing a graph convolutional network architecture, this approach leverages the relationship structure between nodes to enhance the model's feature representation capabilities, thereby improving classification accuracy.

[0115] In radar jamming signal scenarios, the correlation between signals is particularly significant, especially in complex interference situations, where signal combinations often contain strong correlations. Each signal sample not only has its own characteristics, but also carries rich pattern information in its relationship with other signals. In view of this, the present invention models the signal as a graph structure entity, using nodes to represent signal samples and edges to describe the degree of correlation between signals. This graph structure modeling method can more accurately capture the dependencies between signals and provides effective structural support for improving signal classification accuracy. By attaching Figure 5 Adjacency matrix and attached Figure 6 Graphical representation of radar jammer signals. The signal diagram is constructed in a way that allows for a more intuitive understanding of the connections between the signals.

[0116] The present invention uses fusion features to construct an inter-class relationship graph. In order to better model the category relationship, the present invention defines two symmetric adjacency matrices based on the inter-class relationship graph, namely the comprehensive relationship matrix and the core relationship matrix.

[0117] The comprehensive relationship matrix is expressed as:

[0118]

[0119] In formula (9), is a comprehensive relationship matrix based on the Euclidean distance between the category label embedding vectors, l i and l j are the semantic embedding vectors of the i-th and j-th categories respectively, and η is a hyperparameter used to control the attenuation of distance to similarity.

[0120] The core relationship matrix is expressed as:

[0121]

[0122] In formula (10), is the core relationship matrix, which retains important core relationships by removing edges below the ξ quantile value. Representation matrix ξ percentile value of .

[0123] During training, Storing comprehensive semantic information between categories can prevent the model from overfitting to local semantic relationships. The two adjacency matrices emphasize important semantic relationships between categories and are used to strengthen key patterns in training. After balanced fusion, these two adjacency matrices guide the training of graph convolutional networks, thereby better transferring inter-class relationships from semantic modalities to classification weights.

[0124] In the specific implementation, the graph structure semantic classifier is implemented through two independent graph convolution functions and Process the comprehensive relationship matrix and the core relationship matrix separately and calculate the features of the next layer, which can be expressed as:

[0125]

[0126] In formula (11) and formula (12), and They are and The degree matrix of and is a trainable weight matrix and σ(·) is the activation function.

[0127] The final classification result is obtained by weighted summing the outputs of the comprehensive relationship matrix and the core relationship matrix, which is expressed as:

[0128]

[0129] In formula (13), o is the classification result, ρ is the hyperparameter used to balance the contribution of comprehensive relations and core relations, and L is the number of layers of the graph convolution classifier.

[0130] The graph structure semantic classifier of the present invention uses a graph convolutional network to model the comprehensive semantic relationships and core semantic relationships between signal categories, thereby improving the robustness of the classification decision boundary and the accuracy of radar interference recognition.

[0131] The dataset was divided into a training set for model parameter optimization and a test set for evaluating the generalization ability of the trained model, with a ratio of 7:3. The network parameters with the best performance on the training set were obtained. The optimal network parameters obtained in step S3 were loaded and the model performance was verified on an independent test set.

[0132] The test results are as follows:

[0133] Experimental parameter settings: All experiments were implemented based on the PyTorch framework. The number of attention heads in the cross-domain attention mechanism was set to 4, ρ in the graph convolution classifier was set to 0.5, and L was set to 2. During model training, ReLU was used as the activation function and Stochastic Gradient Descent (SGD) was used as the optimization algorithm. The optimization parameters were set as follows: the weight decay coefficient was 1×10 -4The initial learning rate was 0.001, which was gradually adjusted with a decay rate of 0.5 during training. The total number of training epochs was set to 100, and the cross-entropy loss function was used as the loss function. The experimental hardware environment included an Intel Core i5-13400F processor, 16GB of memory, and an NVIDIA GeForce RTX 3060 graphics card.

[0134] In order to verify the influence of each component of the present invention on the performance, the present invention conducts an ablation experiment by using a controlled variable method.

[0135] Multi-domain input validity analysis:

[0136] The experimental results of multi-domain input feature ablation are shown in Table 1.

[0137] Table 1. Experimental results of multi-domain input feature ablation (%)

[0138]

[0139] Table 1 compares the recognition accuracy of different input combinations under JNRs of 0dB, 5dB, and 10dB. Table 1 shows that the three-domain combined input has significant robustness advantages in complex noise environments. In particular, under extreme noise conditions of 0dB, its 98.68% accuracy rate is 9.78%, 25.25%, and 9.25% higher than that of single-time domain, single-frequency domain, and single-time-frequency domain input, respectively. This verifies the effectiveness of coupling time, frequency, and time-frequency domain features for noise suppression.

[0140] Furthermore, from the perspective of feature sensitivity, the accuracy of a single frequency domain input under 0dB conditions is 73.43%, significantly lower than the 88.90% of the time domain and 89.43% of the time-frequency domain. This is due to the high sensitivity of frequency domain features to carrier frequency offset and phase noise. It is worth noting that the time-frequency domain input exhibits optimal noise robustness under single-domain conditions, indicating that the short-time Fourier transform effectively disperses noise energy through its time-frequency localization characteristics. The accuracy of the "time domain + frequency domain" combination reaches 91.15%, an increase of 1.72% over the single time domain input, indicating that the combination of the time domain and frequency domain contains more information than the time-frequency domain.

[0141] Finally, by comparing performance degradation trends under different JNRs, we found that when the JNR drops from 10dB to 0dB, the accuracy of the three-domain joint input only drops by 0.7%, while the accuracy of the single time-domain and single frequency-domain inputs drops by 10.91% and 25.88%, respectively. This verifies the mechanism by which multi-domain feature fusion improves model generalization capabilities. This is primarily because the time-domain waveform preserves the signal's temporal correlation, the frequency-domain spectrum reflects the carrier distribution characteristics, and the time-frequency domain provides modulation transient characteristics. The synergistic effect of these three constructs a complete signal representation space, thus proving the effectiveness of the three-domain joint input.

[0142] In order to verify the effectiveness of the graph structure semantic classifier, the classification performance of the graph structure semantic classifier and the fully connected network (FCN) classifier are compared under the condition of a fixed feature extractor model. The performance comparison (%) of the classifiers under different JNR conditions is shown in Table 2.

[0143] Table 2 Comparison of classifier performance under different JNR conditions (%)

[0144]

[0145] Table 2 shows that the graph convolution classifier achieves excellent classification performance under low JNR conditions, particularly at 0 dB, where it achieves an accuracy of 98.68%. This represents a 5.81% improvement over the fully connected network, and a 0.51% improvement at 5 dB. This demonstrates that the graph convolution layer, through a learnable adjacency matrix, adaptively aggregates information from feature nodes, effectively suppressing noise interference. In contrast, the fully connected network's equally connected neurons struggle to capture nonlinear dependencies between features, resulting in insufficient discriminative feature learning in noisy environments. Although the FCN outperforms the FCN by 0.57% under JNR = 10 dB, this may be due to the fact that signal features tend to be linearly separable under high signal-to-noise ratio conditions, where the complex nonlinear mapping of graph convolution introduces a small amount of overfitting noise. However, both networks achieve accuracy above 99%. In terms of parameter count, GCC achieves significant performance improvements over FCN with only 0.05M additional parameters. This demonstrates that the introduction of prior knowledge of graph structure effectively reduces the model's dependence on parameter size and enables more compact decision boundary learning through geometric constraints in the feature space.

[0146] Comparative experiment:

[0147] To demonstrate the effectiveness of the proposed algorithm, comparative experiments were conducted with all methods using the same experimental parameters at JNRs of 0dB, 5dB, and 10dB. The results include a comparative analysis of overall classification performance, classification results for different interference types, and classification performance for difficult-to-distinguish interference signals.

[0148] 1) Overall classification performance comparison

[0149] The classification accuracy of different methods under various JNR conditions is shown in Table 3.

[0150] Table 3 Comparison of classification accuracy with other interference identification methods (%)

[0151]

[0152]

[0153] As can be seen in Table 3, the proposed method outperforms the comparison methods at all JNR levels of 0dB and 5dB, with particularly strong performance at 0dB. Specifically, the proposed method achieves a classification accuracy of 98.68% at 0dB, a 6.85% improvement over the 91.83% achieved by TFA-HFO (Section 3), a 10.46% improvement over JRNet, a 12.34% improvement over FF-CNN, and a 21.19% improvement over DF-CNN, demonstrating significant advantages. At 5dB, while all methods show improved classification performance, the proposed method maintains its lead, exceeding the next-best method, TFA-HFO, by 1.12%, effectively overcoming the limitations of single-scale features in representing complex interference. Its accuracy at 10dB reaches 99.38%, slightly lower than TFA-HFO, but still above 99%.

[0154] The present invention also shows the T-SNE visualization results under the condition of JNR being 0dB, as shown in the attached figure. Figure 7 As shown in the figure, compared to other methods, the MDF-GCC proposed in this paper generates features for different interference types with less overlap in feature space, effectively resolving the problem of misclassification of difficult-to-distinguish interference types. Furthermore, a single interference type contains multiple clusters in the feature space, demonstrating that MDF-GCC, through learning core and general semantics, can achieve adaptive information aggregation of feature nodes, effectively reducing the probability of misclassification of the classifier at the decision boundary. These results qualitatively demonstrate that MDF-GCC can construct more robust and discriminative interference feature representations, thereby forming more precise classification boundaries, highlighting its effectiveness in identifying radar interference in low JNR environments.

[0155] 2) Classification performance of different radar jamming signals

[0156] In order to further analyze the classification performance of different interference types, this study plotted the confusion matrix of the interference signal under the conditions of JNR of 0dB and 5dB. The results are shown in the attached figure. Figure 8 and attached Figure 9 As shown. Figure 8 and attached Figure 9 It can be seen from the results that the method of the present invention is superior to other methods, especially under the condition of 0 dB, the accuracy of most composite interferences reaches 100%, far exceeding other methods, proving the effectiveness of the method of the present invention.

[0157] 3) Classification performance of difficult to distinguish interference signals

[0158] As attached Figure 10As shown in the figure, for the classification problem of difficult-to-distinguish interference types such as NCJ+DFTJ and NCJ+ISFJ, NFMJ and NFMJ+ISFJ, the traditional comparison method has significant limitations, while the method of the present invention achieves accurate discrimination. Figure 11 Experimental data shows that in a 0dB strong noise environment, the classification accuracy of the baseline method is generally less than 90%. Specifically: JRNet's recognition accuracy for NCJ+DFTJ is 71.00%, and FF-CNN's recognition accuracy for NCJ+ISFJ is only 65.00%. TFA-HFO's best performance on the NFMJ+ISFJ category is 89.00%, while DF-CNN's recognition accuracy for NFMJ is 80.00%. In contrast, the classification accuracy of the method of the present invention reached 100% for all the above interference types, showing a significant advantage.

[0159] Attachment Figure 12 Experimental data shows that when the JNR is increased to 5dB, the proposed method still maintains a 100.00% classification accuracy for complex interference types such as NCJ+DFTJ, NCJ+ISFJ, NFMJ+ISFJ, NFMJ, and MFPJ, and its robustness is significantly superior to the comparison methods. Experimental results verify that the synergy between the multi-domain fusion strategy and the graph convolution classifier can effectively enhance the inter-class separability in the feature space, especially achieving breakthrough results in optimizing the decision boundary of difficult-to-distinguish samples.

[0160] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A radar interference recognition method based on interactive fusion of multi-representation domain features and graph structure semantic classifier, characterized by: Including steps: S1. Construct a multi-domain feature interactive fusion module, which takes time domain waveform, spectrogram and time-spectrogram as input data and outputs fused features; S2. Construct a graph structure semantic classifier, input the fusion features into the graph structure semantic classifier, and the graph structure semantic classifier analyzes the semantic associations of different categories of signals and outputs a classification result; S3, dividing the data set into a training set and a test set, and saving the network parameters with the best classification effect in the training set according to the classification results; S4. Load the network parameters, input the test set into the radar interference recognition network for processing, and obtain a recognition result.

2. The radar interference recognition method based on interactive fusion of multi-representation domain features and graph structure semantic classifier according to claim 1 is characterized in that: The multi-representation domain feature interactive fusion module includes: a shallow feature extraction module, a cross-domain feature extraction module and a time-frequency feature extraction module; The shallow feature extraction module is used to extract time domain features and frequency domain features; The cross-domain feature extraction module uses the cross-domain attention mechanism to achieve information interaction between different domains, extract deep signal features, and output deep time domain features and deep frequency domain features; The time-frequency feature extraction module is used to capture the joint characteristics of the signal between time and spectrum and output time-frequency domain features; The time domain deep features, the frequency domain deep features and the time-frequency domain features are spliced to form fusion features.

3. The radar interference recognition method based on interactive fusion of multi-representation domain features and graph structure semantic classifier according to claim 2 is characterized in that: The shallow feature extraction module consists of two cascaded residual blocks, each of which integrates two one-dimensional convolutional layers with a kernel size of 3×1 and a batch normalization layer. The residual block connection path is realized by the series connection of a convolutional layer with a kernel size of 1×1 and a double residual block.

4. The radar interference recognition method based on interactive fusion of multi-representation domain features and graph structure semantic classifier according to claim 3 is characterized in that: Assume that the time domain signal is the radar interference signal time series Frequency domain signal is a single-sided spectrum signal N is the number of samples, T is the length of the time domain signal, and F is the length of the frequency domain signal. The output feature of the shallow feature extraction module is expressed as: In formula (1) and formula (2), and is the corresponding nonlinear mapping function, output feature and They represent time domain features and frequency domain features respectively, and d represents the output feature length.

5. The radar interference recognition method based on interactive fusion of multi-representation domain features and graph structure semantic classifier according to claim 4 is characterized in that: The cross-domain feature extraction module outputs time domain deep features, including: The time domain features As the query vector Q t , Q t =Z t W Q ; The frequency domain characteristics As the frequency domain key vector K f and frequency domain value vector V f , K f =Z f W K , V f =Z f W V ; Calculated by scaling the dot product attention: In formula (3), Q t is the query vector, K f is the frequency domain key vector, V f is the frequency domain value vector, W Q ,W K , is the learnable parameter matrix, d k represents the intermediate characteristic length, Used to control the dimension of the dot product; The enhanced time domain features obtained through residual connection are expressed as: Z t′ =Attention(Q t ,K f ,V f )+Z t (4) In formula (4), Z t′ is the enhanced time domain feature; After the cross-domain attention, the feedforward network and the residual block are cascaded. The feedforward network consists of two fully connected layers. The feature space is expanded through nonlinear transformation. The residual block is used to extract the deep features of the signal and obtain the deep time domain features of the signal. d m Indicates the output feature length.

6. The radar interference recognition method based on interactive fusion of multi-representation domain features and graph structure semantic classifier according to claim 5 is characterized in that: The cross-domain feature extraction module outputs frequency domain deep features, including: The frequency domain features As the frequency domain query vector Q f , Q f =Z f W Q '; The time domain characteristics As the time domain key vector K t and the time domain value vector V t , K t =Z t W K ', V t =Z t W V '; Calculated by scaling the dot product attention: In formula (5), Q f is the frequency domain query vector, K t is the time domain key vector, V t is the time domain value vector, W Q ',W K ', is the learnable parameter matrix, d k represents the intermediate characteristic length, Used to control the dimension of the dot product. The enhanced frequency domain features are obtained through residual connection: Z f' =Attention(Q f ,K t ,V t )+Z f (6) In formula (6), Z f' is the enhanced frequency domain feature; After the cross-domain attention, the feedforward network and the residual block are cascaded. The feedforward network consists of two fully connected layers. The feature space is expanded through nonlinear transformation. The residual block is used to extract the deep features of the signal and obtain the deep frequency domain features of the signal. d m Indicates the output feature length.

7. The radar interference recognition method based on interactive fusion of multi-representation domain features and graph structure semantic classifier according to claim 6 is characterized in that: The time-frequency feature extraction module extracts time-frequency domain features, including: The time-frequency domain feature module is designed to capture the joint characteristics of the signal between time and spectrum, and output the time-frequency domain features Expressed as: In formula (7), represents the time-frequency feature extraction module, X tf and w tf are the time-frequency domain signals and network parameters respectively.

8. The radar interference recognition method based on interactive fusion of multi-representation domain features and graph structure semantic classifier according to claim 7 is characterized in that: The fusion feature is expressed as: In formula (8), F represents the fusion feature, d f =3·d m .

9. The radar interference recognition method based on interactive fusion of multi-representation domain features and graph structure semantic classifier according to claim 1 is characterized in that: The step S2 comprises: said constructing an inter-class relationship graph by utilizing said fusion features; Based on the inter-class relationship graph, two symmetric adjacency matrices are defined, namely a comprehensive relationship matrix and a core relationship matrix; The comprehensive relationship matrix is expressed as: In formula (9), is a comprehensive relationship matrix based on the Euclidean distance between the category label embedding vectors, l i and l j are the semantic embedding vectors of the i-th and j-th categories respectively, and η is a hyperparameter used to control the attenuation of distance to similarity; The core relationship matrix is expressed as: In formula (10), is the core relationship matrix, which retains important core relationships by removing edges below the ξ quantile value. Representation matrix ξ percentile value of .

10. The radar interference recognition method based on interactive fusion of multi-representation domain features and graph structure semantic classifier according to claim 9 is characterized in that: The step S2 further includes: The graph structure semantic classifier is constructed by two independent graph convolution functions. and The comprehensive relationship matrix and the core relationship matrix are processed separately and the features of the next layer are calculated, which can be expressed as: In formula (11) and formula (12), and They are and The degree matrix of and is a trainable weight matrix, σ(·) is the activation function; The classification result is obtained by weighted summing the outputs of the comprehensive relationship matrix and the core relationship matrix, and is expressed as: In formula (13), o is the classification result, ρ is the hyperparameter used to balance the contribution of comprehensive relations and core relations, and L is the number of layers of the graph convolution classifier.

Images