Multifunctional radar signal sorting method and system based on kernel-driven self-supervised contrast learning

Abstract

The invention discloses a multifunctional radar signal sorting method and system based on kernel-driven self-supervised comparative learning, and the method comprises the following steps: S1, carrying out the data enhancement based on kernel function driving on an intercepted staggered pulse sequence, and obtaining a comparison image set and a sequence set; s2, respectively inputting the comparison image set and the sequence set into parallel networks of a dual-network feature fusion model so as to extract feature vectors of fusion data; s3, performing instance-level and clustering-level head comparison on the feature vectors, and analyzing individual discrimination and group characteristics to realize feature abstraction and mapping of the feature vectors; s4, storing the feature vector in a momentum form, and optimizing negative sample selection by adopting a kernel enhanced memory contrast loss function; and S5, inputting the trained feature vector into a kernel optimization TOA association aggregation algorithm, and outputting a radar signal sorting result. According to the method, the multifunctional radar signals can be accurately sorted under the condition that label data is not used, and the problem of abrupt performance drop caused by serious multi-parameter overlapping can be effectively solved.

Description

Technical Field

[0001] The present invention belongs to the field of radar reconnaissance and electronic warfare technology, and specifically relates to a multi-function radar (MFRs) signal sorting method and system based on kernel-driven self-supervised comparative learning. The method is suitable for unsupervised sorting of radar signals with multi-parameter coupling and dynamic mode switching in complex electromagnetic environments. Background Art

[0002] Radar signal sorting (RSS) is a key technology in electronic warfare and reconnaissance. It aims to accurately separate interleaved or overlapping pulse signals from multiple radar emitters intercepted in complex electromagnetic environments into multiple independent radar signal sequences. The accuracy of RSS directly impacts subsequent tasks such as radar signal identification, jamming decision-making, and threat analysis. However, with the increasing complexity and diversity of signals in modern electronic warfare environments, RSS faces numerous challenges, including diverse patterns, multi-parameter coupling, and difficulty obtaining labels. Therefore, the design of more efficient and robust sorting algorithms is urgently needed.

[0003] Traditional RSS methods primarily rely on timing characteristics such as the pulse repetition interval (PRI) and the combination of multiple parameters to achieve signal sorting by analyzing the regularities within these timing characteristics. However, these methods significantly degrade in sorting performance when faced with complex electromagnetic environments, overlapping parameters, and variable patterns. Furthermore, issues such as noise interference, uneven data distribution, and a lack of flexibility further limit the applicability of these traditional methods.

[0004] The rise of machine learning has enabled the introduction of unsupervised clustering algorithms into the RSS field, providing a new technical path for improving signal sorting performance in complex electromagnetic environments and, to a certain extent, enhancing the ability to jointly process multiple parameters. Furthermore, deep learning, with its powerful feature extraction capabilities, has demonstrated unique advantages in modeling high-dimensional manifolds of radar signals, successfully achieving dynamic capture and adaptive characterization of the complex characteristics of radar signals.

[0005] However, existing machine learning methods still face multiple theoretical challenges: First, the multi-parameter coupling problem causes the algorithm to experience a sharp drop in sorting performance when there is a high degree of similarity in the distribution of signal features. Second, the problem of batch addition and omission caused by the variability of multi-function radar modes is more prominent in scenarios with low-separability feature spaces and complex noise coupling. Third, the strong dependence of the supervised learning paradigm on labeled data limits its generalization performance in unlabeled scenarios. In addition, current research has not yet formed a systematic method for representation learning of the multi-dimensional feature space of pulse descriptor words (PDWs). In particular, there are obvious deficiencies in the in-depth exploration of the spatiotemporal correlation characteristics and local topological structure characteristics contained in radar pulse trains, which directly restricts the robustness and environmental adaptability of the sorting algorithm.

[0006] Therefore, in response to the problems faced by the existing technology in radar signal sorting, such as insufficient multi-parameter decoupling capability, insufficient pattern parsing, and strong label dependence, the present invention provides a sorting method and system based on kernel-driven deep contrast clustering of the self-supervised learning paradigm to overcome the shortcomings of the existing technology and improve the accuracy and anti-interference performance of multi-function radar signal sorting. Summary of the Invention

[0007] In response to the above-mentioned shortcomings of the existing technology, the present invention proposes a multi-function radar signal sorting method and system based on kernel-driven self-supervised contrastive learning, aiming to address the sorting challenges of multi-parameter coupling, diverse modes, and difficult-to-obtain labels, so as to improve the accuracy and performance of multi-function radar signals.

[0008] In order to achieve the above objectives, the present invention adopts the following technical solutions:

[0009] A multifunctional radar signal sorting method based on kernel-driven self-supervised contrastive learning includes the following steps:

[0010] S1. Perform kernel function-driven data enhancement on the intercepted interleaved pulse sequence to obtain a set of contrast images and sequences that can adapt to the radar signal and contrast learning network;

[0011] S2, inputting the contrast image set and sequence set obtained in step S1 as data sets into the parallel network of the dual-network feature fusion model respectively to extract the feature vector of the fused data;

[0012] S3, the feature vector of step S2 is subjected to instance-level and cluster-level comparison for individual identification and group characteristic analysis to achieve feature abstraction and mapping of the feature vector;

[0013] S4. The feature vector obtained in step S3 is stored in momentum form, and the kernel-enhanced memory contrast loss function is used to optimize the selection of negative samples to enhance the optimization of individual and group features; this solves the optimization problem of nonlinear similarity calculation.

[0014] S5. Input the feature vector trained in step S4 into the kernel optimized TOA association aggregation algorithm to solve the problem of fuzzy boundary of the linear classifier, and output the final radar signal sorting result.

[0015] The kernel-driven method mentioned in the present invention refers to driving data enhancement, model optimization, and aggregation optimization through kernel functions.

[0016] Furthermore, in step S1, a kernel-driven method is used to perform data enhancement on the interleaved pulse sequence. The specific steps are as follows:

[0017] Step S11: For the three characteristic parameters of RF, PW and DOA in the PDW data, the original low-dimensional data is mapped to a high-dimensional space through a radial basis kernel function to generate a high-dimensional vector representation with rich nonlinear characteristics.

[0018] Step S12: Reconstruct the mapped high-dimensional pulse features into a two-dimensional structural image by combining them with their temporal and spatial correlations, and construct a reference image with multi-scale characteristics. This representation method enables the joint distribution of high-dimensional parameters to present an interpretable texture pattern.

[0019] Step S13: Based on the spatiotemporal feature neighborhood information contained in the kernel matrix, extract the high inner product representation in the kernel matrix column corresponding to the target pulse, that is, the most adjacent K pulse index sequences, and construct the spatiotemporal correlation matrix of the enhanced image to construct the enhanced image. This step avoids the inner product calculation in high-dimensional space by using the kernel matrix based on the target pulse. By extracting the spatiotemporal feature neighborhood information through the kernel matrix, the spatiotemporal correlation matrix AM of the enhanced image can be constructed. st (i) An image enhancement method that combines neighborhood spatiotemporal feature aggregation to construct an efficient enhanced image. Compared to the baseline image, the enhanced image after adding the perturbation of spatiotemporal feature neighborhood information exhibits blurring. This blurring phenomenon has two implications: First, it reflects the clustering characteristics of PDW data. Blurring can better reflect the similarities and differences within the data and reveal hidden distribution patterns. Second, the information provided by blurring enhances the quality of contrastive learning and strengthens the model's ability to understand subtle changes, thereby improving the model's robustness and generalization capabilities.

[0020] Step S14: The enhanced image and the reference image are combined to form an image pair as a comparison image set. The spatiotemporal correlation matrix AM of each enhanced image is st The corresponding TOA sequences in (i) are used as sequence sets. Through kernel-driven data augmentation, radar sequences of different categories are mapped into image forms suitable for the comparison learning task. Differences in subtle texture features are also presented between the baseline and enhanced images, ensuring semantic consistency in the generated images and enhancing the data representation capability.

[0021] Furthermore, in step S2, the image set is used as the input of the image network branch in the fusion network, and the sequence set is used as the input of the sequence network branch in the fusion network, and features of different characteristics are extracted through different networks. The details are as follows:

[0022] Step S21, the image set formed by RF, PW and DOA parameters is used to extract high-dimensional features using the image feature extraction network ResNet in the dual network feature fusion model. ResNet helps to enhance the extraction of local texture detail features through its deep residual network structure. a ,I' a) respectively input the shared weight ResNet network, through f ResNet (·) Extract its high-level feature vector v a and v' a Since the network shares weights, this approach ensures that the features of the input image pairs are consistent in the same space.

[0023] Step S22, the TOA parameters with time series characteristics are used to capture the time series dependencies between pulses using the long short-term memory sequence network LSTM in the dual network feature fusion model. LSTM can effectively capture long-term dependencies through the gating mechanism, which is helpful for the feature extraction of TOA pulse relationships. i Input LSTM network, through f LSTM (·) to extract its time series feature vector v b , whose gating mechanism captures the dependencies and dynamic characteristics of time series.

[0024] Step S23, the feature vector v generated by LSTM is concatenated by feature concatenation. b Fuse to v a and v' a In the above example, the fused feature vector v is generated. c and v' c , to complete the unified representation and joint modeling of different parameter features, and realize the joint representation of image features and time series features in the feature dimension.

[0025] Furthermore, in step S3, a double contrast mechanism is proposed to extract the feature vectors from the dual fusion network features, and implement the feature abstraction and complex mapping of the feature vectors through the multi-layer nonlinear transformation of the stacked nonlinear perceptron. ins (·) and cluster-level projection f clu (·) By using hierarchical feature decoupling, we improve the joint modeling capability of subtle time-frequency differences and distribution topology of radar signals, thereby achieving the coordinated optimization of generalization performance and representation robustness.

[0026] The instance-level comparison head optimizes the uniqueness of individual samples, strengthening individual discrimination by narrowing the distance between homologous augmented views and widening the distance between heterogeneous samples. The cluster-level comparison head enhances the aggregation of similar samples and the distinction between classes, modeling the distribution characteristics of groups by increasing intra-class aggregation and inter-class separation. The combination of these two approaches not only improves the model's sensitivity to fine-grained feature differences but also strengthens its understanding of global data distribution, thereby enhancing the model's representational capabilities and robustness.

[0027] Furthermore, in step S4, a kernel-enhanced memory contrast loss function is used to solve the optimization problem of nonlinear similarity calculation and negative sample selection. This method can not only alleviate the false negative sample problem, but also enhance the adaptability to uneven data distribution.

[0028] Among them, the global feature library stores global features through the memory module, and the memory module dynamically maintains the global feature library through the momentum update mechanism to provide data support for negative sample optimization.

[0029] For negative sample optimization, a kernel-driven similarity screening criterion is employed. This criterion, by incorporating kernel function-based sample similarity discrimination, can construct a high-confidence negative sample pool. This criterion, while filtering similarity artifacts through the kernel matrix, suppresses false negative samples. Furthermore, it leverages the momentum feature library to overcome batch size limitations and enhance negative sample diversity. The collaborative optimization of these two approaches aims to improve feature space discriminability and model robustness.

[0030] The loss optimization function is designed based on the InfoNCE loss commonly used in contrastive learning and can be expressed as:

[0031]

[0032] Among them, sim(·,·) represents the cosine similarity function, which is used to measure the similarity between vectors. (f i ,f i + ) represents a positive sample pair, (f i ,f i - ) represents the negative sample pair, and τ represents the temperature parameter.

[0033] Combined with the memory pool module based on momentum storage, by dynamically maintaining the feature representation library of historical samples, the momentum update mechanism is used to gradually fuse new and old features, and the fused features are stored in the memory pool to achieve modeling and efficient retrieval of long-term data dependencies.

[0034] Combined with the instance-level and cluster-level dual comparison mechanism in step S3, a kernel optimization comparison loss function is proposed. It can be expressed as:

[0035]

[0036] Among them, f i ∈{z i ,y i}, (f i ,f i ') represents a positive sample pair, N i represents negative samples, K(f i ,f i') represents the inner product of the eigenvector in the high-dimensional space, τ k ∈{τ ins ,τ clu} represent the temperature parameters at the instance level and cluster level respectively. In the above formula, the similarity between feature vectors is measured by the inner product of the kernel function. The overall objective function is:

[0037]

[0038] in, represent The instance-level loss function is represent The cluster-level loss function.

[0039] Furthermore, in step S5, a constrained quadratic programming model is established to characterize the TOA sequence. This model integrates the TOA temporal coherence constraints with the cluster distribution characteristics of the PDW data to construct a kernel-optimized temporal correlation aggregation framework. This method achieves a coordinated optimization of spatiotemporal features by jointly optimizing temporal constraints and kernel spatial similarity metrics, thereby enhancing the discriminative power of pattern boundaries and preventing misclassification.

[0040] Among them, by decoupling the local nonlinearity and global topological characteristics of feature association, while ensuring the continuous temporal evolution law of homologous patterns, the distinguishability between heterogeneous clusters is enhanced, and false alarms and missed detections are significantly suppressed, thereby effectively reducing the "over-batch" and "missed batch" phenomena, and ultimately achieving highly robust signal sorting in non-cooperative complex electromagnetic scenarios.

[0041] The present invention also discloses a multifunctional radar signal sorting system based on kernel-driven self-supervised contrastive learning, which is used to execute the above method and includes the following modules:

[0042] Kernel-driven data enhancement module: performs kernel-driven data enhancement on the intercepted interleaved pulse sequence to obtain a set of contrast images and sequences that can adapt to radar signals and contrast learning networks;

[0043] Dual-network feature extraction and fusion module: The obtained contrast image set and sequence set are input as data sets into the dual-network feature fusion model to extract the feature vector of the fused data;

[0044] Dual-head mapping module: This module uses instance-level and cluster-level comparison heads to analyze individual discrimination and group characteristics of the extracted feature vectors, thereby achieving feature abstraction and complex mapping of the feature vectors.

[0045] Kernel-enhanced memory contrast loss module: This module stores the obtained feature vector in momentum form and uses the kernel-enhanced memory contrast loss function to optimize negative sample selection to enhance the optimization of individual and group features. This module is used to solve the optimization problems of nonlinear similarity calculation and negative sample selection.

[0046] Kernel-Optimized Time of Arrival (TOA) Correlation Aggregation Module: This module inputs the trained feature vectors into the Kernel-Optimized Time of Arrival (TOA) Correlation Aggregation algorithm and outputs radar signal sorting results. This module addresses boundary discrimination errors caused by suboptimal decision surfaces, suppresses misclassification rates, and achieves efficient radar signal sorting.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. This paper proposes an innovative self-supervised deep contrastive clustering paradigm based on kernel-driven enhancement. This unsupervised framework can simultaneously consider the complexity and multimodal distribution of data, providing a new idea for radar signal sorting.

[0049] 2. This paper innovatively proposes an image construction and data enhancement method suitable for PDW data. Combining the nonlinear mapping capability of the kernel function with the spatiotemporal characteristics of the neighborhood structure, it can effectively capture the global and local structural characteristics of PDW data and improve the contrast representation performance.

[0050] 3. The present invention designs a dual-network fusion contrastive learning method that combines the dual contrast head mechanism with kernel optimized memory loss, which can deeply explore the intrinsic characteristics of PDW data through kernel function optimization loss.

[0051] 4. This paper proposes a cluster boundary aggregation algorithm based on TOA temporal characteristics correlation and kernel space enhancement. It effectively reduces the phenomenon of over- and under-information, and significantly improves the robustness and accuracy of radar signal sorting in complex scenarios.

[0052] In summary, in order to address the technical problems faced by existing methods in radar signal sorting, such as insufficient multi-parameter decoupling ability, insufficient pattern parsing, and strong label dependence, which lead to decreased sorting performance, the present invention proposes a multi-function radar signal sorting method and system based on kernel-driven self-supervised contrastive learning. The self-supervised contrastive learning method is introduced into the field of radar signal sorting. Combined with the superiority of kernel functions in capturing nonlinear relationships and constructing high-dimensional feature spaces, the method of kernel-driven data enhancement, kernel-optimized memory contrast, and kernel-enhanced spatiotemporal correlation aggregation solves the problem of insufficient data representation of interleaved pulse sequences composed of multiple radars, and can realize the sorting of multi-function radar signals on a real battlefield without prior information. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1Schematic diagram of the module framework involved in the multifunctional radar signal sorting method based on kernel-driven self-supervised contrastive learning in a preferred embodiment of the present invention;

[0054] Figure 2 This is a flow chart of the kernel-driven data enhancement module in a preferred embodiment of the present invention;

[0055] Figure 3 A visual comparison diagram of the reference image and the enhanced image in a preferred embodiment of the present invention;

[0056] Figure 4 This is a flowchart of the optimization and comparison of core-driven feature extraction of the backbone network in the preferred embodiment of the present invention;

[0057] Figure 5 A comparison diagram of two-dimensional visual distribution diagrams of radar signal sorting achieved using different methods and two-dimensional visual distribution diagrams of radar signal sorting achieved using the present invention;

[0058] Figure 6 This is a block diagram of a multifunctional radar signal sorting system based on kernel-driven self-supervised contrastive learning in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0059] The present invention will elaborate on the implementation of the technical solution through the following preferred embodiments. Based on the technical content disclosed in this specification, those skilled in the art can fully understand the technical advantages and expected effects of the present invention.

[0060] like Figure 1 As shown, this embodiment proposes a multi-function radar signal sorting method based on kernel-driven self-supervised contrastive learning. Taking four radars as an example, two of them are multi-function radars: Radar1 and Radar3 are conventional radars, and Radar2 and Radar4 are multi-function radars, each with three operating modes. The interleaved pulse sequence is the intercepted data, and the sorting results represent the sorting of deinterleaved pulses of different categories. This embodiment method specifically includes the following steps:

[0061] Step S1, using the kernel function driven method to reconstruct and enhance the data of the intercepted interleaved pulse sequence, to obtain a set of contrast images and sequences that can adapt to the radar signal and contrast learning network, such as Figure 2 The specific steps are as follows:

[0062] Step S11: For the three characteristic parameters of RF, PW and DOA in the PDW data, the original low-dimensional data is mapped to a high-dimensional space through the radial basis kernel function. The expression is:

[0063] K(x i ,x j )=exp(-γ||xi -x j || 2 )

[0064] Among them, x i and x j is the original data sample, γ is the kernel bandwidth parameter, which is used to adjust the sensitivity of the feature map. Data can be mapped into high-dimensional space n represents the original data dimension, d corresponds to x i In high-dimensional space, feature dimensions are generated to generate high-dimensional vector representations with rich nonlinear features.

[0065] In step S12, the mapped high-dimensional pulse features are reconstructed into a two-dimensional structural image by combining the temporal and spatial correlation to construct a reference image. Specifically, in this step, d=224 is set to standardize the specifications of the input data.

[0066] The mapped high-dimensional pulse features are combined with spatiotemporal correlation to reconstruct a two-dimensional structural image, and a data representation with multi-scale characteristics is constructed. First, a reference image is generated:

[0067]

[0068] in, d>>n is the mapping operator, N is the number of samples, X represents the total pulse set, Represents the spatiotemporal correlation matrix of the i-th pulse, records the correlation pulse index of the constructed image, the reference image self-correlation, It is used to realize tensor splicing of spatiotemporal feature structures. Figure 3 The left side shows the reference image I a This representation makes the joint distribution of high-dimensional parameters present interpretable texture patterns.

[0069] Step S13, based on the spatiotemporal feature neighborhood information contained in the kernel matrix, extract the high inner product representation in the kernel matrix column corresponding to the target pulse, that is, the K most adjacent pulse index sequences, and construct the spatiotemporal correlation matrix of the enhanced image to construct the enhanced image; specifically, through the ability of the kernel matrix to capture high-order similarities between data, to extract the spatiotemporal feature neighborhood information, and through the image enhancement method of neighborhood spatiotemporal feature aggregation, to construct an efficient enhanced image.

[0070] Among them, the kernel matrix is ​​defined as:

[0071] K ij =<φ(x i ),φ(x j )>

[0072] in, φ(x i ),φ(xj ) is the high-dimensional representation after feature mapping, K ij Is the inner product in high-dimensional space. Based on the target pulse x i The kernel matrix avoids the inner product calculation in high-dimensional space.

[0073] By extracting the neighborhood pulse index represented by the high inner product in the target pulse train of the kernel matrix, the spatiotemporal correlation matrix AM of the enhanced image can be constructed. st (i) To strengthen the consistency between local similarity and global structure between features:

[0074] AM st (i) = argTop d (K ij ),j∈{1,2,...,N}

[0075] Among them, argTop d (K ij ) represents the pulse x i The indices of the d most correlated neighboring pulses. Figure 3 The enhanced image I' is given on the right a , by AM st (i) Substituting back to T i Compared to the baseline image, the enhanced image after perturbation exhibits blurring. This blurring has two implications: First, it reflects the clustering characteristics of PDW data. Blurring can better reflect similarities and differences within the data, revealing hidden distribution patterns. Second, the information provided by blurring enhances the quality of contrastive learning and strengthens the model's ability to understand subtle changes, thereby improving model robustness and generalization.

[0076] In step S14, the enhanced image and the reference image are combined to form an image pair as an image set. The spatiotemporal correlation matrix AM of each enhanced image is st The corresponding TOA sequences in (i) are used as sequence sets. Through kernel-driven data augmentation, radar sequences of different categories are mapped into image forms suitable for the comparison learning task. Differences in subtle texture features are also presented between the baseline and enhanced images, ensuring semantic consistency in the generated images and enhancing the data representation capability.

[0077] In step S2, the contrast image set and the sequence set obtained in step S1 are respectively input into the parallel network of the dual-network feature fusion model as data sets to extract the feature vector of the fused data.

[0078] In this step, the image set is used as a dual network feature fusion module (such as Figure 4The image feature extraction network branch in the left part is used as the input of the sequence set, and the long short-term memory network branch is used as the input to achieve efficient fusion and comparison of the multi-modal features of PDW data through different networks. Figure 4 The dual network fusion part is as follows:

[0079] Step S21: The image set formed by RF, PW and DOA parameters is used to extract high-dimensional features using the image feature extraction network ResNet. ResNet helps to enhance the extraction of local texture detail features through its deep residual network structure. a ,I' a ), respectively input the ResNet network with shared weights, through f ResNet (·) Extract its high-level feature vector v a and v' a Since the network shares weights, this approach ensures that the features of the input image pairs are consistent in the same space.

[0080] v a =f ResNet (I a )

[0081] v' a =f ResNet (I' a )

[0082] Step S22, the TOA parameters with time series characteristics are used to capture the time series dependencies between pulses using a long short-term memory network (LSTM). LSTM can effectively capture long-term dependencies through a gating mechanism, which helps to extract the features of TOA pulse relationships. b Input LSTM network, through f LSTM (·) to extract its time series feature vector v b , whose gating mechanism captures the dependencies and dynamic characteristics of time series.

[0083] v b =f LSTM (I b )

[0084] Step S23, the feature vector v generated by LSTM is concatenated by feature concatenation. b Fuse to v a and v' a In the above example, the fused feature vector v is generated. c and v' c , to complete the unified representation and joint modeling of different parameter features, and realize the joint representation of image features and time series features in the feature dimension.

[0085] v c=cat(v a ,v b )

[0086] v' c =cat(v' a ,v b )

[0087] In step S3, the feature vector extracted by the dual fusion network features is transformed through multiple layers of nonlinear transformation of stacked nonlinear perceptrons (MLPs) to achieve feature abstraction and complex mapping of the feature vector. Based on MLPs, instance-level projections f ins (·) and cluster-level projection f clu (·) By analyzing individual features of instance-level comparison heads and group features of cluster-level comparison heads, hierarchical feature decoupling is achieved to improve the joint modeling capability of subtle time-frequency differences and distribution topology of radar signals, thereby achieving the coordinated optimization of generalization performance and representation robustness.

[0088] The instance-level comparison head optimizes the uniqueness of individual samples, strengthening individual discrimination by narrowing the distance between homologous augmented views and widening the distance between heterogeneous samples. The cluster-level comparison head enhances the aggregation of similar samples and the distinction between classes, modeling the distribution characteristics of groups by increasing intra-class aggregation and inter-class separation. The combination of these two approaches not only improves the model's sensitivity to fine-grained feature differences but also strengthens its understanding of global data distribution, thereby enhancing the model's representational capabilities and robustness.

[0089] z=f ins (v)

[0090] y=f clu (v)

[0091] Where v represents the feature vector obtained in step S2, z and y represent the feature vectors after the instance-level comparison head and cluster-level comparison head, respectively.

[0092] In step S4, the feature vector obtained in step S3 is stored in momentum form, and the kernel-enhanced memory contrast loss function is used to optimize the selection of negative samples to enhance the optimization of individual and group features.

[0093] This step uses the kernel-enhanced memory contrast loss function to solve the optimization problem of nonlinear similarity calculation and the optimization problem of negative sample selection. This method can not only alleviate the problem of false negative samples, but also enhance the adaptability to uneven data distribution. Figure 4 The kernel enhances the memory contrast loss part.

[0094] Among them, the global feature library stores global features through the memory module, and the memory module dynamically maintains the global feature library through the momentum update mechanism to provide data support for negative sample optimization.

[0095] G t =αG t-1 +(1-αz)

[0096] Among them, α∈[0,1] is the momentum coefficient, G t Represents the state of the current feature vector, which is determined by the historical state G t-1 The fusion weight of the current feature z is calculated. This mechanism maintains a dynamic representation of the global data distribution by balancing historical features with real-time updates, thereby improving the model convergence efficiency.

[0097] Among them, for negative sample optimization, a kernel function-driven similarity screening criterion is adopted.

[0098]

[0099] Among them, (z i ,z j ) represents the feature vector obtained in step S3, N i is the negative sample of the i-th sample, K(·,·) is the kernel similarity function, τ is the similarity threshold, and the sample label C i 、C j The pseudo-label class for the target sample is generated by the cluster-level comparison head. This criterion can construct a high-confidence negative sample pool: on the one hand, it filters similarity artifacts through the kernel matrix to suppress false negative samples; on the other hand, it leverages the momentum feature library to overcome batch size limitations and enhance negative sample diversity. The coordinated optimization of these two aims to improve feature space discriminability and model robustness.

[0100] Among them, the loss optimization function is designed based on InfoNCE loss, combining the momentum memory module and the dual contrast mechanism, and proposing a kernel optimization contrast loss function. It can be expressed as:

[0101]

[0102] Where N represents the number of samples, f i ∈{z i ,y i}, (f i ,f i ') represents a positive sample pair, N i represents negative samples, K(f i ,f i ') represents the inner product of the eigenvector in the high-dimensional space, τ k ∈{τ ins ,τ clu} represent the temperature parameters at the instance level and cluster level respectively. In the above formula, the similarity between feature vectors is measured by the inner product of the kernel function. The overall objective function is:

[0103]

[0104] in, represent The instance-level loss function is represent The cluster-level loss function.

[0105] In step S5, a constrained quadratic programming model is established to characterize the TOA sequence. This model integrates the TOA temporal coherence constraints with the cluster distribution characteristics of the PDW data to construct a kernel-optimized temporal correlation aggregation framework. This method achieves collaborative optimization of spatiotemporal features by jointly optimizing temporal constraints and kernel spatial similarity metrics, thereby enhancing the discriminative power of pattern boundaries and preventing misclassification.

[0106] In order to optimize the efficiency and reliability of clustering merging, the feature vectors obtained by network self-training are processed using the t-SNE dimensionality reduction algorithm. The feature vectors obtained in step S4 are preprocessed using the density-based spatial clustering of applications with noise (DBSCAN) clustering algorithm, and the initial cluster structure is generated through its dynamic search capability. Then, based on the time series characteristics, the TOA association criterion between clusters is constructed. The clusters C generated by the pre-clustering are i with C j , define the temporal correlation metric as:

[0107]

[0108] Among them, start and end represent the cluster TOA values ​​of the cluster TOA matrix. in Representative C i The maximum TOA interval of the pulse sequence in the cluster is used to determine cluster C. i With cluster C j This is a potentially homologous radar sequence. This judgment is based on the interpretable boundary of temporal correlation: when the inter-cluster TOA interval is within the dynamic adjustment range of the radar PRI parameters, it can be inferred that they have the characteristics of homologous temporal correlation, that is, the segmented radiation behavior generated by the same radar due to mode switching or parameter agility.

[0109] In order to enhance the reliability of TOA association aggregation, a merging criterion with enhanced kernel similarity is proposed:

[0110]

[0111] in, Represents cluster C i and C j The merging conditions are met, K(Ci ,C j ) is the inter-cluster kernel similarity, and τ is the kernel similarity threshold. This mechanism decouples the local nonlinearity of feature associations from the global topological properties. While ensuring the temporal continuous evolution of homologous patterns, it enhances the distinguishability between heterogeneous clusters and significantly suppresses false alarms and missed detections, effectively reducing the "over-batch" and "missed batch" phenomena, ultimately achieving highly robust signal sorting in non-cooperative complex electromagnetic scenarios.

[0112] According to the above embodiment, the following test was carried out:

[0113] Considering the fact that multi-function radars switch between different working modes in real scenarios, a scenario including four radars was constructed. Among them, Radar 1 and Radar 3 are conventional radars with only one mode; Radar 2 and Radar 4 are multi-function radars with three different working modes respectively, and the parameter value range of each radar is different. The RF, PW, DOA and PRI in PDW are set as the typical characteristics of each radar pulse. They have a variety of modulation types, such as agile, group change, jitter and other modulation styles. Because the algorithm is a self-supervised model, it does not distinguish between training sets and test sets. The data set contains a total of 2500 pulses. The original pulse distribution diagram is shown as follows: Figure 5 As shown in (a).

[0114] The sorting framework of this embodiment was evaluated with the existing sorting framework, and the results are shown in Table 1 and Figure 5 As shown in Table 1, the performance comparison between the kernel-driven self-supervised contrastive learning method and other existing methods in the test scenario is shown in Table 1. Figure 5 is the scene data distribution diagram under different methods).

[0115] Table 1 Performance comparison of various algorithms

[0116]

[0117]

[0118] The above results show that the proposed method KD-DCC shows the highest accuracy under unsupervised conditions. Figure 5Figures (b)-(e) show that the performance bottleneck of label-free algorithms is primarily due to the parameter aliasing effect of multi-function radar signals. In particular, density-sensitive clustering algorithms (DBSCAN and ADPC), whose Euclidean distance-based similarity metrics struggle to effectively distinguish signals in overlapping areas. CNet's network topology also struggles to adapt to sudden changes in signal distribution caused by radar operating mode switching, resulting in mismatches with signal features. This ultimately leads to misclassification of Radar1 or Radar3, resulting in 0% accuracy for some categories and a "missed batch" phenomenon. However, PSO-K partially alleviates the pattern confusion problem by optimizing cluster centers and the preset number of clusters through global search capabilities. Its overall accuracy (65.00%) is higher than DBSCAN (63.60%), DPC (60.00%), and CNet (59.56%). However, dynamic electromagnetic environment mismatch still results in low sorting accuracy.

[0119] On the other hand, SVM uses labeled data to learn data characteristics, which effectively solves the problems of batch addition and batch omission, and improves the accuracy to 81.64%. ResGCN-RSS suppresses inter-batch interference by modeling with a small amount of labeled data and a gated residual network, effectively improving the accuracy to 94.96%. It is worth noting that the method KD-DCC proposed in the present invention further improves the accuracy to 96.73% by using a kernel-driven deep contrast clustering framework without using any labeled data, verifying the effectiveness of the feature decoupling mechanism. Among them, Radar4 has an accuracy slightly lower than that of the supervised model and ADPC due to the high aliasing of multiple signal parameters, but still maintains a greater advantage than other methods. This proves the effectiveness of the method proposed in the present invention in sorting multi-function radar signals in a complex electromagnetic environment where label information is lacking.

[0120] like Figure 6 As shown, this embodiment discloses a multifunctional radar signal sorting system based on kernel-driven self-supervised contrastive learning, which is used to execute the above method, and specifically includes the following modules:

[0121] Kernel-driven data enhancement module: performs kernel function-driven data enhancement on the intercepted interleaved pulse sequence to obtain a set of contrast images and sequences that can adapt to radar signals and contrast learning networks.

[0122] Dual-network feature extraction and fusion module: The obtained contrast image set and sequence set are input as data sets into the dual-network feature fusion model to extract the feature vector of the fused data.

[0123] Dual comparison head mapping module: Based on the characteristics of radar signal data, the extracted feature vectors are analyzed through instance-level and cluster-level comparison heads to analyze individual identification and group characteristics, thereby realizing feature abstraction and mapping of feature vectors.

[0124] Kernel-enhanced memory contrast loss module: The obtained feature vector is stored in momentum form, and the kernel-enhanced memory contrast loss function is used to optimize the selection of negative samples to enhance the optimization of individual and group features. It is used to solve the optimization problems of nonlinear similarity calculation and negative sample selection.

[0125] Kernel-Optimized Time of Arrival (TOA) Correlation Aggregation Module: This module inputs the trained feature vectors into the kernel-optimized TOA correlation aggregation algorithm and outputs radar signal sorting results. This module addresses boundary discrimination errors caused by suboptimal decision surfaces, suppresses misclassification rates, and achieves efficient radar signal sorting.

[0126] For other contents of this embodiment, please refer to the above method embodiment.

[0127] The above is only a preferred embodiment of the present invention and does not limit the scope of protection of the present invention. It should be understood that without departing from the core technical concept of the present invention, the embodiments of the present invention may be subjected to various equivalent replacements or adaptive adjustments, and the specific details described in the specification may also be modified accordingly based on different application scenarios. These modified embodiments all fall within the scope of protection of the claims of the present invention.

Claims

1. A multifunctional radar signal sorting method based on kernel-driven self-supervised contrastive learning, characterized by: The specific steps include: S1. Perform kernel function-driven data enhancement on the intercepted interleaved pulse sequence to obtain a set of contrast images and sequences that can adapt to the radar signal and contrast learning network; S2, inputting the contrast image set and sequence set obtained in step S1 as data sets into the parallel network of the dual-network feature fusion model respectively to extract the feature vector of the fused data; S3, the feature vector of step S2 is subjected to instance-level and cluster-level comparison to analyze individual discrimination and group characteristics, so as to realize feature abstraction and mapping of the feature vector; S4. The feature vector obtained in step S3 is stored in momentum form, and the kernel enhanced memory contrast loss function is used to optimize the negative sample selection; S5. Input the feature vector trained in step S4 into the kernel optimized TOA association aggregation algorithm and output the radar signal sorting result.

2. The multifunctional radar signal sorting method based on kernel-driven self-supervised contrastive learning according to claim 1, characterized in that: In step S1, the kernel function driven method is used to perform data enhancement on the interleaved pulse sequence. The specific steps are as follows: S11, for the three characteristic parameters of RF, PW and DOA in the PDW data, the original low-dimensional data is mapped to a high-dimensional space through the radial basis kernel function; S12, reconstructing the mapped high-dimensional pulse features into a two-dimensional structural image by combining the temporal and spatial correlation to construct a reference image; S13, based on the spatiotemporal feature neighborhood information contained in the kernel matrix, extracting high inner product representations in the kernel matrix columns corresponding to the target pulse, and constructing a spatiotemporal correlation matrix of the enhanced image to construct an enhanced image; S14, the enhanced image and the reference image constitute an image pair as a comparison image set, and the spatiotemporal correlation matrix AM of each enhanced image is st The corresponding TOA sequence in (i) is used as the sequence set.

3. The multifunctional radar signal sorting method based on kernel-driven self-supervised contrastive learning as claimed in claim 2, characterized in that: In step S2, the image set formed by RF, PW and DOA parameters is used to extract high-dimensional features v using the image feature extraction network ResNet in the dual-network feature fusion model. a and v' a ; The TOA sequence is used in the long short-term memory network LSTM in the dual network feature fusion model to capture the temporal dependency between pulses; and the feature vector v generated by LSTM is concatenated by feature splicing b Fuse to v a and v' a In the above example, the fusion feature vector v is generated. c and v' c .

4. The multifunctional radar signal sorting method based on kernel-driven self-supervised contrastive learning as claimed in claim 3, characterized in that: In step S3, the feature vector extracted by the dual fusion network features is transformed through multiple layers of stacked nonlinear perceptrons, and the joint modeling capability of subtle time-frequency differences and distribution topological structures of radar signals is improved through hierarchical feature decoupling, thereby realizing feature abstraction and mapping of the feature vector.

5. The multifunctional radar signal sorting method based on kernel-driven self-supervised contrastive learning as claimed in claim 4, characterized in that: In step S4, a similarity screening criterion driven by the kernel-enhanced memory contrast loss function is adopted: Among them, (z i ,z j ) represents the feature vector obtained in step S3, N i is the negative sample of the i-th sample, K(·,·) is the kernel similarity function, τ is the similarity threshold, and the sample label C i 、C j is the pseudo label class of the target sample, generated by the cluster-level contrast head.

6. The multifunctional radar signal sorting method based on kernel-driven self-supervised contrastive learning as claimed in claim 4, characterized in that: In step S4, the kernel-enhanced memory contrast loss function is adopted, which is expressed as: Where N represents the number of samples, f i ∈{z i ,y i },(f i ,f i ') represents a positive sample pair, N i represents negative samples, K(f i ,f′ i ) represents the inner product of the eigenvector in the high-dimensional space, τ k ∈{τ ins ,τ clu } represent the temperature parameters at the instance level and cluster level respectively. In the formula, the similarity between feature vectors is measured by the inner product of the kernel function; the overall objective function is: in, represent The instance-level loss function is represent The cluster-level loss function.

7. The multifunctional radar signal sorting method based on kernel-driven self-supervised contrastive learning according to claim 6, characterized in that: In step S5, the trained feature vector is processed by t-SNE dimensionality reduction, and the feature vector obtained in step S4 is preprocessed by density-based DBSCAN algorithm, and the initial cluster structure is generated by dynamic search; then, based on the time series characteristics, the TOA association criterion between clusters is constructed; the cluster C generated by pre-clustering is i with C j , define the temporal correlation metric as: Among them, start and end represent the cluster TOA values ​​of the cluster TOA matrix respectively; if in, Representative C i The maximum TOA interval of the pulse sequence in the cluster is used to determine cluster C. i With cluster C j are potentially homologous radar sequences; A merging criterion for kernel similarity enhancement is proposed: in, Represents cluster C i and C j The merging conditions are met, K(C i ,C j ) is the inter-cluster kernel similarity, and τ is the threshold of kernel similarity.

8. A multifunctional radar signal sorting system based on kernel-driven self-supervised contrastive learning, used to execute the method according to any one of claims 1 to 7, characterized in that: Includes the following modules: Kernel-driven data enhancement module: performs kernel-driven data enhancement on the intercepted interleaved pulse sequence to obtain a set of contrast images and sequences that can adapt to radar signals and contrast learning networks; Dual-network feature extraction and fusion module: The obtained contrast image set and sequence set are input as data sets into the dual-network feature fusion model to extract the feature vector of the fused data; Dual-head mapping module: This module uses instance-level and cluster-level comparison heads to analyze individual discrimination and group characteristics of the extracted feature vectors to achieve feature abstraction and mapping. Kernel-enhanced memory contrast loss module: The obtained feature vector is stored in momentum form and the kernel-enhanced memory contrast loss function is used to optimize the selection of negative samples to enhance the optimization of individual and group features; Kernel-optimized TOA correlation aggregation module: inputs the trained feature vector into the kernel-optimized TOA correlation aggregation algorithm and outputs the radar signal sorting results.

Images