Dimension separation attention and fusion vector transformer twin network multi-radar jamming signal small sample classification method
Patent Information
- Application Number
- CN202611163443.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-03
- Publication Date
- 2026-10-16
AI Technical Summary
[0006]为解决现有雷达干扰识别方法在小样本条件下泛化能力差、多雷达观测数据因简单融合导致信息利用不充分、以及难以有效提取信号长距离时序依赖特征的问题，本发明提供维度分离注意力与融合向量的Transformer孪生网络多雷达干扰信号小样本分类方法，包括：
[0024] This invention trains the model using positive and negative sample pairs, enabling the model to learn the similarities and differences between interference signals, thus reducing the reliance on a large number of labeled training samples. It also improves feature representation by extracting fused features from multiple radar observation signals through a dimensionality separation attention and fusion vector cross-attention feature extraction network. Finally, it achieves radar interference signal classification under small sample conditions by using Siamese network difference measurement and maximum similarity classification.
Smart Images
Figure CN122926083A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radar jamming signal identification and classification technology, specifically to a small-sample classification method for multiple radar jamming signals using Transformer Siamese networks based on dimensionality separation attention and vector fusion. Background Technology
[0002] With the increasing complexity of the electromagnetic environment, radar received signals may simultaneously contain various interference patterns, such as suppression interference, deception interference, and combined interference. The identification results of radar interference signals directly affect the selection of subsequent anti-interference measures. Therefore, how to accurately classify radar interference signals in complex electromagnetic environments is an important problem that needs to be solved in the field of radar signal processing.
[0003] Existing radar interference identification methods mostly employ manual feature extraction followed by a classifier, or use ordinary deep neural networks for direct multi-class classification. Manual features rely on expert experience and are easily affected by noise, sampling conditions, and variations in interference patterns; ordinary multi-class classification networks typically require a large number of training samples for each interference pattern, and when the number of training samples is small, the model is prone to overfitting and insufficient generalization ability.
[0004] In multi-radar observation scenarios, the same interference signal can be acquired by multiple radars. While multi-radar observation information provides a richer signal description, differences exist in amplitude, phase, observation angle, and local noise between data from different radars. Simply stitching, averaging, or performing single-stage fusion of multiple radar observation data can easily overlook the local correlations within the same radar observation data and the complementary correlations between data from different radar observation nodes, leading to insufficient utilization of multi-radar information.
[0005] Siamese networks, by constructing similarity relationships between sample pairs and learning samples, can transform multi-class classification problems into class-specific or class-specific discrimination problems, making them suitable for small-sample classification tasks with limited training samples. Transformer structures can model long-range correlations in sequence features, making them suitable for extracting global correlation information from radar observation signals. Therefore, it is necessary to propose a small-sample classification method for radar interference signals that combines positive and negative sample pair construction, multi-radar dimension-separated self-attention feature extraction, fusion vector cross-attention, and Siamese network similarity discrimination to improve classification capabilities in multi-radar small-sample scenarios. Summary of the Invention
[0006] To address the problems of poor generalization ability under small sample conditions, insufficient information utilization due to simple fusion of multi-radar observation data, and difficulty in effectively extracting long-range temporal dependence features of signals in existing radar interference identification methods, this invention provides a small-sample classification method for multi-radar interference signals using Transformer Siamese networks with dimensionality separation attention and fusion vectors, including:
[0007] Step 1: Obtain a multi-radar jamming signal dataset. The multi-radar jamming signal dataset contains multiple types of radar jamming patterns, and each jamming pattern includes complex signal data collected under radar observation nodes. Each jamming pattern is structured according to radar observation nodes, sampling sequences, complex channels, and pulse periods, and the batch dimension is expanded to form a five-dimensional structured data, which provides standardized input for subsequent multi-radar dimension separation feature extraction.
[0008] Step 2: Construct training sample pairs based on five-dimensional structured data. The training sample pairs include: positive sample pairs consisting of two samples with the same radar jamming pattern and negative sample pairs consisting of two samples with different radar jamming patterns, providing supervised samples for the similarity discrimination learning of the Siamese network.
[0009] Step 3: Input the two samples in the training sample pair into the parameter-shared dimension separation attention and fusion vector cross attention feature extraction network. First, perform self-attention encoding on the sequence features within the same radar observation node, then perform cross-node self-attention encoding on the features between different radar observation nodes. Finally, fuse the token average feature and the learnable fusion vector to generate the first sample level fusion feature vector and the second sample level fusion feature vector, so that the features of the two types of samples are in the same comparable representation space.
[0010] Step 4: Based on the Siamese network structure, measure the difference between the first-sample level fused feature vector and the second-sample level fused feature vector, output the similarity between the two samples belonging to the same interference pattern, and use the loss function and optimizer to iteratively update the Siamese network parameters to obtain the trained classification model.
[0011] Step 5: Combine the sample to be identified with each candidate sample to form a sample pair input to the trained classification model. After obtaining the similarity of each sample pair, take the radar interference pattern of the candidate sample with the highest similarity as the classification result of the sample to be identified. Accurate classification in small sample scenarios is achieved based on similarity comparison.
[0012] Furthermore, in step one, the dimensions of the five-dimensional structural data are, in order, batch dimension, radar observation node dimension, sampling sequence dimension, complex channel dimension, and pulse period dimension.
[0013] Furthermore, in step two, the label of positive sample pairs is set to 1, and the label of negative sample pairs is set to 0. For each type of radar interference pattern, the samples of this type are divided into training samples, validation samples, and test samples in a ratio of 1:1:10. The samples in the training set, validation set, and test set do not overlap. During the training process, positive sample pairs and negative sample pairs are constructed alternately so that each training batch contains both similarity association information of samples of the same type and difference association information of samples of different types.
[0014] Furthermore, in step three, the feature extraction network for dimensionality separation attention and fusion vector cross attention includes, in sequence: a periodic dimension mapping layer, a signal block encoding layer, a position encoding layer, a dimensionality separation self-attention encoding module, a fusion vector cross attention module, and a feature fusion layer;
[0015] The periodic dimension mapping layer maps the pulse period dimension to a single-period representation;
[0016] The signal block coding layer uses one-dimensional convolution to block and tokenize the sampling sequence corresponding to each radar observation node;
[0017] The location coding layer adds a learnable location code to the token within each radar observation node;
[0018] The dimension separation self-attention encoding module first performs self-attention encoding on the token sequence within the same radar observation node, and then performs self-attention encoding on the token features between different radar observation nodes.
[0019] The cross-attention module of the fusion vector uses the learnable fusion vector as the query and the Token feature after dimensional separation and attention encoding as the key and value to update the learnable fusion vector;
[0020] The feature fusion layer concatenates and linearly maps the average features of the tokens with the updated learnable fusion vector to obtain a sample-level fused feature vector.
[0021] Furthermore, the difference measurement method in step four is as follows: the absolute difference of each element of the first sample-level fused feature vector and the second sample-level fused feature vector is obtained to obtain the difference vector. The difference vector is input into the linear discriminant layer to obtain the same-class discrimination score. Then, the same-class discrimination score is mapped to the interval of 0~1 through the Sigmoid function to obtain the similarity between two samples belonging to the same interference pattern. During training, binary cross-entropy loss, Focal loss or mean squared error loss are used as loss functions, and the AdamW optimizer and cosine annealing learning rate scheduling strategy are used to update the model parameters.
[0022] Furthermore, in step five, an N-way one-shot recognition task is used to achieve classification: in each test round, one sample to be identified is combined with one candidate sample of the same type and N-1 candidate samples of different types to form N pairs of samples to be identified. After calculating the similarity between the N pairs of samples, the radar interference pattern of the candidate sample with the highest similarity value is taken as the final classification result of the sample to be identified.
[0023] The beneficial effects of this invention are:
[0024] This invention trains the model using positive and negative sample pairs, enabling the model to learn the similarities and differences between interference signals, thus reducing the reliance on a large number of labeled training samples. It also improves feature representation by extracting fused features from multiple radar observation signals through a dimensionality separation attention and fusion vector cross-attention feature extraction network. Finally, it achieves radar interference signal classification under small sample conditions by using Siamese network difference measurement and maximum similarity classification. Attached Figure Description
[0025] Figure 1 This is the overall flowchart of the Transformer Siamese Network method for small sample classification of multi-radar interference signals based on dimension separation attention and fusion vector as described in this invention;
[0026] Figure 2 This is a schematic diagram of the positive and negative sample pair construction process described in this invention;
[0027] Figure 3 This is a schematic diagram illustrating the difference measurement process between the dimension separation attention and fusion vector cross-attention feature extraction network and the Siamese network described in this invention.
[0028] Figure 4 This is a schematic diagram of the maximum similarity classification process described in this invention.
[0029] illustrate: Figure 4 In this context, Query represents the sample to be identified, Cn represents the nth candidate sample, Sn represents the similarity between Query and Cn, and Smax represents the maximum similarity. Detailed Implementation
[0030] The technical solution of the present invention will be further described below with reference to embodiments, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention. In the following embodiments, process equipment or devices not specifically specified are all conventional equipment or devices in the art. Unless specifically specified, the technical means used in the embodiments of the present invention are all conventional means well known to those skilled in the art.
[0031] Example 1, combined with Figure 1This embodiment describes a Transformer Siamese network method for small-sample classification of multi-radar interference signals using dimensionality-separated attention and fusion vectors, including:
[0032] Step 1: Obtain a multi-radar jamming signal dataset. The multi-radar jamming signal dataset contains multiple types of radar jamming patterns, and each jamming pattern includes complex signal data collected under radar observation nodes. Each jamming pattern is structured according to radar observation nodes, sampling sequences, complex channels, and pulse periods, and the batch dimension is expanded to form a five-dimensional structured data, which provides standardized input for subsequent multi-radar dimension separation feature extraction.
[0033] Step 2: Construct training sample pairs based on five-dimensional structured data. The training sample pairs include: positive sample pairs consisting of two samples with the same radar jamming pattern and negative sample pairs consisting of two samples with different radar jamming patterns, providing supervised samples for the similarity discrimination learning of the Siamese network.
[0034] Step 3: Input the two samples in the training sample pair into the parameter-shared dimension separation attention and fusion vector cross attention feature extraction network. First, perform self-attention encoding on the sequence features within the same radar observation node, then perform cross-node self-attention encoding on the features between different radar observation nodes. Finally, fuse the token average feature and the learnable fusion vector to generate the first sample level fusion feature vector and the second sample level fusion feature vector, so that the features of the two types of samples are in the same comparable representation space.
[0035] Step 4: Based on the Siamese network structure, measure the difference between the first-sample level fused feature vector and the second-sample level fused feature vector, output the similarity between the two samples belonging to the same interference pattern, and use the loss function and optimizer to iteratively update the Siamese network parameters to obtain the trained classification model.
[0036] Step 5: Combine the sample to be identified with each candidate sample to form a sample pair input to the trained classification model. After obtaining the similarity of each sample pair, take the radar interference pattern of the candidate sample with the highest similarity as the classification result of the sample to be identified. Accurate classification in small sample scenarios is achieved based on similarity comparison.
[0037] Specifically, this invention first normalizes the complex interference signals collected by each node of multiple radars into a five-dimensional standardized input based on radar observation node, sampling sequence, complex channels, pulse period, and batch dimension. Then, it transforms the traditional multi-class classification task into a similarity discrimination task. By constructing a dimension separation attention and fusion vector cross-attention network with shared positive and negative sample pairs as driving parameters, it successively mines the temporal correlation features within a single radar node and the complementary correlation features across radar nodes. It fuses the token average feature and the learnable fusion vector to generate sample-level fusion features in the same representation space. After training with difference measurement and loss optimization, a classification model is obtained. Finally, by comparing the similarity between the sample to be identified and each candidate sample, it achieves accurate classification of multi-radar interference patterns that only rely on a small number of labeled samples.
[0038] Furthermore, in step one, the dimensions of the five-dimensional structural data are, in order, batch dimension, radar observation node dimension, sampling sequence dimension, complex channel dimension, and pulse period dimension.
[0039] Specifically, when partitioning the one-to-many radar jamming signal dataset in step one, random partitioning is performed for each type of radar jamming pattern. First, all sample indices for that radar jamming pattern are obtained. Then, the sample indices are randomly shuffled, and training samples, validation samples, and test samples are selected sequentially according to a preset ratio. The training set is used to construct positive and negative sample pairs and update the Siamese network parameters. The validation set is used to observe the model's similarity discrimination ability during training. The test set is used to finally evaluate the radar jamming pattern classification performance.
[0040] In this embodiment, the multi-radar jamming signal dataset can include fourteen types of radar jamming patterns, including five types of suppression jamming: narrowband aiming jamming (noise amplitude modulation), wideband blocking jamming (noise frequency modulation), frequency sweeping jamming (linear frequency sweeping), clutter jamming, and comb spectrum jamming; four types of deception jamming: intermittent sampling and forwarding jamming, range deception jamming, dense false target jamming, and velocity deception jamming; and five types of combined jamming: narrowband aiming + dense false target jamming, wideband blocking + dense false target jamming, sinusoidal frequency sweeping + dense false target jamming, narrowband aiming + intermittent sampling and forwarding jamming, and intermittent sampling and forwarding + dense false target jamming. Each sample includes complex signal data collected from five radar observation nodes. The data from each radar observation node includes eight pulse periods, and each pulse period includes a real part channel and an imaginary part channel. When inputting into the model, each sample can be organized into five-dimensional data of shape [B,V,L,C,P], where B represents the batch size, V represents the number of radar observation nodes, L represents the sampling sequence length, C represents the number of complex channels, and P represents the number of pulse periods. In a specific implementation, V is 5, L is 2000, C is 2, and P is 8.
[0041] The number of radar jamming patterns, radar observation nodes, sampling sequence length, number of channels, and number of pulse periods described above are illustrative and do not constitute a limitation on the scope of protection of this invention. Those skilled in the art can adjust these quantities according to specific radar systems, sampling configurations, and jamming patterns.
[0042] Furthermore, in step two, the label of positive sample pairs is set to 1, and the label of negative sample pairs is set to 0. For each type of radar interference pattern, the samples of this type are divided into training samples, validation samples, and test samples in a ratio of 1:1:10. The samples in the training set, validation set, and test set do not overlap. During the training process, positive sample pairs and negative sample pairs are constructed alternately so that each training batch contains both similarity association information of samples of the same type and difference association information of samples of different types.
[0043] Specifically, Figure 2 The complete process of generating positive and negative sample pairs based on training set samples is demonstrated. It clarifies the logic of constructing positive sample pairs (labeled 1) by randomly selecting two samples from the same interference pattern, constructing negative sample pairs (labeled 0) by randomly selecting samples from different interference patterns, and finally combining the two types of sample pairs into training sample pairs.
[0044] For positive sample pairs, two samples are randomly selected from the training samples corresponding to the same radar jamming pattern to form the first sample and the second sample, and the label of this sample pair is set to the same class label. For negative sample pairs, two different radar jamming patterns are first randomly selected, and then one sample is randomly selected from the training samples corresponding to each of these two jamming patterns to form the first sample and the second sample, and the label of this sample pair is set to the different class label.
[0045] During training, positive and negative sample pairs can be constructed alternately, ensuring that each training batch simultaneously includes similarity relationships between samples of the same class and difference relationships between samples of different classes. Through this sample pair construction method, the Siamese network does not directly learn fixed category numbers, but rather learns whether two radar interference samples belong to the same interference pattern, thereby improving its ability to distinguish the similarity of new samples under small sample conditions.
[0046] Furthermore, in step three, the feature extraction network for dimensionality separation attention and fusion vector cross attention includes, in sequence: a periodic dimension mapping layer, a signal block encoding layer, a position encoding layer, a dimensionality separation self-attention encoding module, a fusion vector cross attention module, and a feature fusion layer;
[0047] The periodic dimension mapping layer maps the pulse period dimension to a single-period representation;
[0048] The signal block coding layer uses one-dimensional convolution to block and tokenize the sampling sequence corresponding to each radar observation node;
[0049] The location coding layer adds a learnable location code to the token within each radar observation node;
[0050] The dimension separation self-attention encoding module first performs self-attention encoding on the token sequence within the same radar observation node, and then performs self-attention encoding on the token features between different radar observation nodes.
[0051] The cross-attention module of the fusion vector uses the learnable fusion vector as the query and the Token feature after dimensional separation and attention encoding as the key and value to update the learnable fusion vector;
[0052] The feature fusion layer concatenates and linearly maps the average features of the tokens with the updated learnable fusion vector to obtain a sample-level fused feature vector.
[0053] Specifically, Figure 3 The paper presents the processing of a single sample in the feature extraction network of dimension separation attention and fusion vector cross attention, as well as the logic of difference measurement and similarity output of feature vectors of two samples. Specifically, it includes the complete path of periodic dimension mapping, signal block encoding, position encoding, dimension separation self-attention encoding, fusion vector cross attention update, feature fusion to generate sample-level feature vectors, and finally outputting similarity through the difference vector input to the similarity discrimination layer.
[0054] For any sample, the dimension separation attention and fusion vector cross-attention feature extraction network processes the data according to the following procedure to achieve hierarchical feature extraction and global fusion of multiple radar interference signals:
[0055] The input samples are five-dimensional data organized according to [B,V,L,C,P], where B represents the batch dimension, V represents the radar observation node dimension, L represents the sampling sequence dimension, C represents the complex channel dimension, and P represents the pulse period dimension. The period dimension mapping layer first performs a linear mapping on the pulse period dimension P, compressing the observation information of multiple pulse periods into a single period representation, resulting in intermediate data organized according to [B,V,L,C]. This processing can significantly reduce the dimensionality complexity of subsequent token encoding while retaining multi-pulse period observation information. In the specific implementation, the number of pulse periods P is set to 8, the period dimension mapping layer adopts a linear layer, the input dimension is 8, and the output dimension is 1.
[0056] The data after periodic dimension mapping is adjusted to a shape suitable for one-dimensional convolution processing, so that the data of each radar observation node uses complex channels as input channels and the sampling sequence as the convolution direction. The signal block coding layer uses one-dimensional convolution to block and encode the sampling sequence of each radar observation node, outputting the token sequence corresponding to each radar observation node. In a specific implementation, the number of input channels of the one-dimensional convolution is 2, corresponding to the real part channel and the imaginary part channel respectively; the number of output channels is 128; the sampling sequence length L is 2000; the block size is 32; the stride is 16, and the number of blocks can be directly calculated from the sampling sequence length, block size, and stride. After signal block coding, the data can be represented as [B,V,T,D], where T represents the number of tokens corresponding to each radar observation node, and D represents the token feature dimension.
[0057] The location encoding layer adds learnable location codes to the tokens in each radar observation node, enabling the model to distinguish the sequential positions of different tokens within the same radar observation node, while maintaining the independent structural information of the radar observation node dimension and the token dimension. The dimensions of the learnable location codes correspond one-to-one with the number of radar observation nodes, the number of tokens, and the token feature dimension, and can be optimized together with the network parameters.
[0058] The dimensionality-separated self-attention encoding module completes feature modeling in two stages: the first stage is intra-view token association modeling, where self-attention encoding is performed on the token sequence corresponding to each radar observation node, enabling the model to capture long-range dependencies and local combined features between different sampling segments within the same radar observation node; the second stage is cross-radar observation node token association modeling, where view-dimensional self-attention encoding is performed on token features from different radar observation nodes, enabling the model to capture complementary and consistent information between different radar observation nodes. By separating intra-view association modeling and cross-view association modeling, the weakening of structural information caused by simply mixing the radar observation node dimension and the token dimension can be avoided. In a specific implementation, the feature dimension D of the dimensionality-separated self-attention encoding module is 128, the number of attention heads is 8, the hidden layer dimension ratio is 2, the number of encoding module stacks is 4, the dropout rate is 0.1, and each encoding block consists of a dimensionality-separated self-attention layer and a multilayer perceptron layer connected in series.
[0059] To enhance the global representation capability at the sample level, this scheme sets up a learnable fusion vector. The learnable fusion vector is copied to a shape corresponding to the batch size B in each batch, and iteratively updated through a cross-attention module after each encoding block of the attention encoding module is dimensionally separated. In the cross-attention module, the learnable fusion vector serves as the query, and the token features after attention encoding are dimensionally separated as the key and value. Through cross-attention calculation, the learnable fusion vector aggregates global information from multi-radar token features to form a fusion representation oriented towards the entire sample.
[0060] After encoding, the token features after dimensionality separation and attention encoding are first averaged along the token dimension to obtain the average token features of each radar observation node; then, they are averaged again along the radar observation node dimension to obtain the global average token features. Simultaneously, the learnable fusion vector updated through cross-attention is flattened to obtain the fusion vector features. Finally, the average token features and the fusion vector features are concatenated and input into a linear fusion layer for dimensionality mapping to obtain the final sample-level fusion feature vector. This sample-level fusion feature vector simultaneously contains statistical distribution information of multiple radar tokens and global semantic information obtained by aggregating the learnable fusion vector, providing a highly discriminative feature foundation for subsequent similarity measurements.
[0061] Furthermore, the difference measurement method in step four is as follows: the absolute difference of each element of the first sample-level fused feature vector and the second sample-level fused feature vector is obtained to obtain the difference vector. The difference vector is input into the linear discriminant layer to obtain the same-class discrimination score. Then, the same-class discrimination score is mapped to the interval of 0~1 through the Sigmoid function to obtain the similarity between two samples belonging to the same interference pattern. During training, binary cross-entropy loss, Focal loss or mean squared error loss are used as loss functions, and the AdamW optimizer and cosine annealing learning rate scheduling strategy are used to update the model parameters.
[0062] The difference between the first-sample-level fused feature vector FA and the second-sample-level fused feature vector FB is measured based on the Siamese network structure. Specifically, the absolute difference between FA and FB is calculated element-by-element to obtain the difference vector |FA-FB|.
[0063] The difference vector is input into a linear discriminant layer and converted into a similarity score indicating that two samples belong to the same interference pattern. The higher the similarity score, the higher the probability that the two samples belong to the same radar interference pattern; the lower the similarity score, the higher the probability that the two samples belong to different radar interference patterns.
[0064] During training, the similarity discrimination results are constrained based on the labels of positive and negative sample pairs, so that the model gradually narrows the distance between sample-level fused feature vectors of similar samples and widens the distance between sample-level fused feature vectors of dissimilar samples.
[0065] Furthermore, in step five, an N-way one-shot recognition task is used to achieve classification: in each test round, one sample to be identified is combined with one candidate sample of the same type and N-1 candidate samples of different types to form N pairs of samples to be identified. After calculating the similarity between the N pairs of samples, the radar interference pattern of the candidate sample with the highest similarity value is taken as the final classification result of the sample to be identified.
[0066] Specifically, Figure 4 The paper demonstrates the reasoning process for classifying radar interference patterns in small-sample scenarios based on a trained Siamese network. It clarifies the "N-way one-shot" identification logic, which involves forming sample pairs with multiple candidate samples, inputting them into the Siamese network to calculate the similarity of each sample pair, and then selecting the interference pattern of the candidate sample with the highest similarity as the final classification result. This demonstrates the accurate classification capability of the solution under small-sample conditions.
[0067] After completing the training of the Siamese network, the maximum similarity principle is used to classify the samples to be identified. For a sample to be identified, it is used as a query sample. Candidate samples are selected from the candidate radar interference patterns, and the query sample is paired with each candidate sample to form a sample pair to be identified.
[0068] The multiple pairs of samples to be identified are input into the trained radar jamming signal small sample classification model to obtain the similarity between the query sample Query and each candidate sample. The nth candidate sample is denoted as Cn, the similarity between Query and Cn is denoted as Sn, and the maximum similarity is denoted as Smax. When Smax corresponds to the kth candidate sample, the radar jamming pattern of the kth candidate sample is determined as the radar jamming pattern of the sample to be identified.
[0069] In an N-way one-shot recognition task, each test round includes one sample to be identified, one candidate sample of the same class, and N-1 candidate samples of different classes. The sample to be identified is paired with each of the N candidate samples to form N pairs of samples to be identified, and N similarities are calculated for each pair. If the similarity score of the candidate sample of the same class is the highest among the N similarities, then the recognition in that test round is considered correct.
[0070] In a specific experiment, fourteen types of radar jamming patterns were used, with N set to 14 in the N-way model and 1000 test rounds. Training parameters included a learning rate of 0.0005, weight decay of 0.0005, and 150 training rounds. Experimental results showed that the optimal test accuracy of the method in this invention reached 95.70% in 14-way one-shot testing, and the average test accuracy reached 94.15% during the stable training phase. This result demonstrates that this invention can effectively identify multiple radar jamming patterns under small sample conditions.
[0071] The method described in this invention can be applied to datasets containing various radar jamming patterns. For each jamming pattern, only a small number of training samples are needed to construct positive and negative sample pairs and train a Siamese network; during the testing phase, the classification result can be output by comparing the similarity between candidate samples and samples to be identified.
Claims
1. A small-sample classification method for multi-radar interference signals using Transformer Siamese networks with dimensionality separation attention and vector fusion, characterized in that... include: Step 1: Obtain a multi-radar jamming signal dataset. The multi-radar jamming signal dataset contains multiple types of radar jamming patterns, and each jamming pattern includes complex signal data collected under radar observation nodes. Each jamming pattern is structured according to radar observation nodes, sampling sequences, complex channels, and pulse periods, and the batch dimension is expanded to form a five-dimensional structured data, which provides standardized input for subsequent multi-radar dimension separation feature extraction. Step 2: Construct training sample pairs based on five-dimensional structured data. The training sample pairs include: positive sample pairs consisting of two samples with the same radar jamming pattern and negative sample pairs consisting of two samples with different radar jamming patterns, providing supervised samples for the similarity discrimination learning of the Siamese network. Step 3: Input the two samples in the training sample pair into the parameter-shared dimension separation attention and fusion vector cross attention feature extraction network. First, perform self-attention encoding on the sequence features within the same radar observation node, then perform cross-node self-attention encoding on the features between different radar observation nodes. Finally, fuse the token average feature and the learnable fusion vector to generate the first sample level fusion feature vector and the second sample level fusion feature vector, so that the features of the two types of samples are in the same comparable representation space. Step 4: Based on the Siamese network structure, measure the difference between the first-sample level fused feature vector and the second-sample level fused feature vector, output the similarity between the two samples belonging to the same interference pattern, and use the loss function and optimizer to iteratively update the Siamese network parameters to obtain the trained classification model. Step 5: Combine the sample to be identified with each candidate sample to form a sample pair input to the trained classification model. After obtaining the similarity of each sample pair, take the radar interference pattern of the candidate sample with the highest similarity as the classification result of the sample to be identified. Accurate classification in small sample scenarios is achieved based on similarity comparison.
2. The method for small-sample classification of multiple radar interference signals using Transformer Siamese networks with dimensionality separation attention and fusion vectors as described in claim 1, characterized in that, In step one, the dimensions of the five-dimensional structured data are, in order, batch dimension, radar observation node dimension, sampling sequence dimension, complex channel dimension, and pulse period dimension.
3. The method for small-sample classification of multiple radar interference signals using Transformer Siamese networks with dimensionality separation attention and fusion vectors according to claim 2, characterized in that, In step two, the label of positive sample pairs is set to 1, and the label of negative sample pairs is set to 0. For each type of radar interference pattern, the samples of this type are divided into training samples, validation samples and test samples in a ratio of 1:1:
10. The samples in the training set, validation set and test set do not overlap. During the training process, positive sample pairs and negative sample pairs are constructed alternately so that each training batch contains both the similarity association information of samples of the same type and the difference association information of samples of different types.
4. The method for small-sample classification of multiple radar interference signals using Transformer Siamese networks with dimension-separated attention and fused vectors according to claim 3, characterized in that, In step three, the feature extraction network for dimensionality separation attention and fusion vector cross attention includes, in sequence: periodic dimension mapping layer, signal block encoding layer, position encoding layer, dimensionality separation self-attention encoding module, fusion vector cross attention module, and feature fusion layer; The periodic dimension mapping layer maps the pulse period dimension to a single-period representation; The signal block coding layer uses one-dimensional convolution to block and tokenize the sampling sequence corresponding to each radar observation node; The location coding layer adds a learnable location code to the token within each radar observation node; The dimension separation self-attention encoding module first performs self-attention encoding on the token sequence within the same radar observation node, and then performs self-attention encoding on the token features between different radar observation nodes. The cross-attention module of the fusion vector uses the learnable fusion vector as the query and the Token feature after dimensional separation and attention encoding as the key and value to update the learnable fusion vector; The feature fusion layer concatenates and linearly maps the average features of the tokens with the updated learnable fusion vector to obtain a sample-level fused feature vector.
5. The method for small-sample classification of multiple radar interference signals using Transformer Siamese networks with dimension-separated attention and fused vectors according to claim 4, characterized in that, The difference measurement method in step four is as follows: the absolute difference of each element of the first-sample-level fused feature vector and the second-sample-level fused feature vector is obtained to obtain the difference vector. The difference vector is input into the linear discriminant layer to obtain the same-class discrimination score. Then, the same-class discrimination score is mapped to the interval of 0~1 by the Sigmoid function to obtain the similarity of two samples belonging to the same interference pattern. During training, binary cross-entropy loss, Focal loss or mean squared error loss is used as the loss function, and the AdamW optimizer and cosine annealing learning rate scheduling strategy are used to update the model parameters.
6. The method for small-sample classification of multiple radar interference signals using Transformer Siamese networks with dimensionality separation attention and fusion vectors according to claim 5, characterized in that, In step five, an N-way one-shot recognition task is used to achieve classification: In each test round, one sample to be identified is combined with one candidate sample of the same type and N-1 candidate samples of different types to form N pairs of samples to be identified. After calculating the similarity between the N pairs of samples, the radar interference pattern of the candidate sample with the highest similarity value is taken as the final classification result of the sample to be identified.