Fine-grained small sample PDW signal pattern recognition method, device, equipment and medium

Abstract

This invention proposes a fine-grained, small-sample PDW signal pattern recognition method, apparatus, device, and medium, aiming to improve the recognition capability of fine-grained patterns of PDW signals in complex signal environments. Addressing the problem of low recognition rates in real electromagnetic environments due to scarce PDW annotation data, lost PDWs, and interference from spurious PDWs, this invention employs a dual-path prototype contrast enhancement network based on dual-path feature decoupling and attention enhancement mechanisms. A cascade module extracts stable common patterns from PDW sequences, and a projector module adaptively focuses on key difference features between different signal patterns, significantly improving feature discriminativeness. Simultaneously, by collaboratively optimizing classification loss, prototype contrast loss, and feature reconstruction loss, the model maintains excellent generalization ability even with a small number of labeled samples. This invention also exhibits strong anti-interference capabilities in complex signal environments.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology, and in particular to a method, device and medium for fine-grained small sample PDW signal pattern recognition. Background Technology

[0002] This technology enables efficient identification of PDW (Pulse Descriptor) signals in complex signal environments. Specifically, it allows for accurate pattern recognition of PDW signals in non-cooperative receiving scenarios, providing crucial technical support for signal analysis, system status assessment, and even behavioral intent analysis. The field of PDW signal pattern recognition primarily relies on differences in PDW parameters to distinguish between different signal patterns, optimizing recognition performance by extracting effective features and selecting appropriate classifiers.

[0003] In modern complex signal environments, signal environments are becoming increasingly dense and diverse. Many signal sources can flexibly modulate their PDW parameters, such as carrier frequency, pulse width, and pulse repetition time interval, resulting in a large number of fine-grained signal patterns with specific functions and highly similar parameter characteristics. Compared with traditional coarse-grained pattern recognition, fine-grained pattern recognition faces the inherent challenge of "large intra-class differences and small inter-class differences".

[0004] Traditional PDW signal pattern recognition methods heavily rely on the experience of domain experts, using manually defined discrimination rules and logical thresholds for parameters such as PDW repetition interval, pulse width, and frequency for pattern recognition. While these methods offer strong interpretability, their performance largely depends on the completeness and accuracy of prior knowledge, making them ill-suited to the agile parameter changes and complex signal scheduling strategies of modern signal sources.

[0005] In recent years, deep learning-based recognition methods have been widely applied. Researchers have employed marginalized stacked denoising autoencoders to construct multi-layered analysis models, achieving automatic feature extraction. Other studies have proposed hierarchical sequence-to-sequence LSTM networks that can not only identify complex signal patterns but also locate pattern transformation boundaries with PDW-level accuracy. Still others have utilized the global attention mechanism of Transformer networks to effectively capture long-range temporal correlations in signals. However, all of these studies rely on large-scale, complete labeled data, a crucial requirement that is often difficult to meet in real-world non-cooperative reception scenarios.

[0006] Faced with the dual challenges of scarce labeled data and complex signal environments, the few-shot learning paradigm offers an effective technical path. Some researchers have systematically evaluated various deep learning models, establishing the advantages of combining conditional generative adversarial networks (GANs) and residual networks under few-shot conditions. Other studies have explored the technical path of fusing meta-learning and prior knowledge, significantly improving model performance by encoding PDW parameter features and their variation patterns and integrating them into prototype networks and graph convolutional networks. Still other researchers have proposed a refined prototype random walk network that innovatively combines the "general information" of labeled samples with the "manifold information" of unlabeled samples, fully utilizing data distribution characteristics to optimize classification through a three-module architecture.

[0007] However, the generalization and robustness of existing small sample recognition models in complex signal environments containing missing PDWs and spurious PDWs remain a core issue that urgently needs to be explored. Summary of the Invention

[0008] To address the dual challenges of existing technologies, namely the difficulty in learning discriminative features under conditions of scarce labeled data and the difficulty in maintaining stable recognition performance in complex signal environments, this invention proposes a fine-grained small-sample PDW signal pattern recognition method, device, and medium. The aim is to improve the recognition capability of fine-grained patterns of multifunctional PDW signals in complex signal environments and solve the problem of low recognition rate under conditions of scarce labeled data, lost PDWs, and interference from fake PDWs.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] On one hand, the present invention provides a fine-grained small-sample PDW signal pattern recognition method, comprising the following steps:

[0011] Construct a set of PDW sequence samples with PDW signal style category labels as the base class space;

[0012] Multiple training tasks are randomly sampled from the base class category space. Each training task contains a set of non-overlapping support set samples and query set samples.

[0013] For the current training task, feature extraction networks are used to extract feature vectors for each sample;

[0014] Calculate the prototype feature vectors for each category based on the feature vectors of the support set samples;

[0015] The similarity between the feature vectors of the query set samples and the feature vectors of each prototype is calculated. Based on the similarity results, a loss function is constructed to optimize the parameters of the feature extraction network. The current training task is updated and iterated until the model converges, resulting in a trained recognition network model (DPCE-Net).

[0016] For a new recognition task containing multiple new PDW signal pattern categories, multiple support samples are provided for each new category. The support set samples of the new recognition task are input into the trained recognition network model to extract feature vectors and calculate the prototype feature vectors of each new category. The query samples of the new recognition task are input into the trained recognition network model to extract feature vectors and calculate their similarity with the prototype feature vectors of each new category. Classification decisions are made based on the similarity results to complete the PDW signal pattern recognition of the query samples.

[0017] Furthermore, the feature extraction network includes a cascade module and a projector module. The cascade module is used to extract stable common patterns of the PDW sequence. It contains cascaded convolutional processing units and outputs preliminary feature representations. The projector module adaptively focuses on key difference features between different signal patterns and enhances the preliminary feature representations through an attention mechanism to achieve attention residual connections and cross-task feature alignment, thereby obtaining the final discriminative feature vector.

[0018] On the other hand, the present invention provides a fine-grained small-sample PDW signal pattern recognition device, comprising:

[0019] The base class category space construction module is used to construct a set of PDW sequence samples with PDW signal style category labels as the base class category space;

[0020] The task sampling module is used to randomly sample multiple training tasks from the base class category space. Each training task contains a set of non-overlapping support set samples and query set samples.

[0021] The feature extraction module is used to extract feature vectors of each sample for the current training task using a feature extraction network. The feature extraction network includes a cascade module and a projector module. The cascade module is used to extract stable common patterns of the PDW sequence. It contains cascaded convolutional processing units and outputs preliminary feature representations. The projector module adaptively focuses on key difference features between different signal patterns. It uses an attention mechanism to weight and enhance the preliminary feature representations, realizes attention residual connections and cross-task feature alignment, and obtains the final discriminative feature vectors.

[0022] The prototype calculation module is used to calculate the prototype feature vectors of each category based on the feature vectors of the support set samples.

[0023] The training module is used to calculate the similarity between the feature vectors of the query set samples and the feature vectors of each prototype. Based on the similarity results, a loss function is constructed to optimize the parameters of the feature extraction network. The current training task is updated and iterated continuously until the model converges, and a trained recognition network model is obtained.

[0024] The recognition module is used for new recognition tasks containing multiple new PDW signal pattern categories. For each new category, multiple support samples are provided. The support set samples of the new recognition task are input into the trained recognition network model to extract feature vectors and calculate the prototype feature vectors of each new category. The query samples of the new recognition task are input into the trained recognition network model to extract feature vectors and calculate their similarity with the prototype feature vectors of each new category. Based on the similarity results, a classification decision is made to complete the PDW signal pattern recognition of the query samples.

[0025] On the other hand, the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described fine-grained small sample PDW signal pattern recognition method.

[0026] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described fine-grained small sample PDW signal pattern recognition method.

[0027] On the other hand, the present invention provides a computer program product stored on a computer-readable storage medium and including computer instructions that, when executed by a processor, cause a computer device to implement the steps of the fine-grained small sample PDW signal pattern recognition method described above.

[0028] Compared with the prior art, the technical effects of the present invention are as follows:

[0029] The feature extraction network proposed in this invention, based on dual-path feature decoupling and attention enhancement mechanism, can effectively separate the stable common patterns and task-specific differences in PDW sequences, has stronger robustness against PDW loss and spurious PDW interference, and significantly improves the discriminative power of features.

[0030] The proposed dual-supervised meta-learning framework of prototype comparison and self-supervised reconstruction constructs a loss function that is a weighted sum of classification loss, feature reconstruction loss and prototype comparison loss. By co-optimizing classification loss, prototype comparison loss and feature reconstruction loss, it ensures that the model can still maintain excellent generalization ability under the condition of a small number of labeled samples and adapt to the recognition needs of new PDW signal patterns in complex signal environments.

[0031] This invention achieves a balance between feature stability and adaptability through joint optimization of intra-class feature aggregation and cross-task attention projection. In fine-grained recognition scenarios with highly similar signal patterns and scarce labeled data, it has higher recognition accuracy and anti-interference performance compared to traditional few-shot learning methods.

[0032] This invention, based on a meta-learning framework, ensures that the model can maintain excellent generalization ability even with a small number of labeled samples, thus adapting to the recognition requirements of novel PDW signal patterns.

[0033] The fine-grained small-sample PDW signal pattern recognition method proposed in this invention exhibits strong anti-interference capabilities in complex signal environments, providing reliable technical support for practical applications. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0035] Figure 1 This is a flowchart of a fine-grained small-sample PDW signal pattern recognition method in one embodiment;

[0036] Figure 2 This is a schematic diagram of the structure of a recognition network model (Dual Path Prototype Contrast Enhancement Network DPCE-Net) in one embodiment;

[0037] Figure 3 This is a test result diagram of a model under different leak pulse ratios in a 5-way-1-shot task;

[0038] Figure 4 This is a diagram showing the model test results under different spurious impulse ratios in a 5-way-1-shot task;

[0039] Figure 5 This is a diagram showing the model test results under different leak pulse ratios in a 5-way 5-shot task;

[0040] Figure 6 This is a diagram showing the model test results under different spurious pulse ratios in a 5-way 5-shot task. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0042] Fine-grained refers to the fine-grained division of tags. Typically, a signal pattern tag can determine a certain radar operating mode. For example, the signal pattern tag in search mode is called a coarse-grained tag, while a further division relative to the search mode tag, such as velocity search or other search modes, is called a fine-grained tag. Signal patterns under such fine-grained tags often have a high degree of parameter overlap and are difficult to distinguish features, making them more difficult to classify than coarse-grained signal pattern classification problems.

[0043] To address the challenges of existing technologies, such as the difficulty in learning discriminative features under conditions of scarce labeled data and the inability to maintain stable recognition performance in complex signal environments, this paper proposes a method to meet the requirements of fine-grained, small-sample PDW signal pattern recognition. Figure 1 One embodiment proposes a fine-grained small-sample PDW signal pattern recognition method, comprising the following steps:

[0044] Construct a set of PDW sequence samples with PDW signal style category labels as the base class space;

[0045] Multiple training tasks are randomly sampled from the base class category space. Each training task contains a set of non-overlapping support set samples and query set samples.

[0046] For the current training task, feature extraction networks are used to extract feature vectors for each sample;

[0047] Calculate the prototype feature vectors for each category based on the feature vectors of the support set samples;

[0048] Calculate the similarity between the feature vectors of the query set samples and the feature vectors of each prototype. Based on the similarity results, construct a loss function to optimize the parameters of the feature extraction network, update the current training task, and iterate continuously until the model converges to obtain a trained recognition network model.

[0049] For a new recognition task containing multiple new PDW signal pattern categories, multiple support samples are provided for each new category. The support set samples of the new recognition task are input into the trained recognition network model to extract feature vectors and calculate the prototype feature vectors of each new category. The query samples of the new recognition task are input into the trained recognition network model to extract feature vectors and calculate their similarity with the prototype feature vectors of each new category. Classification decisions are made based on the similarity results to complete the PDW signal pattern recognition of the query samples.

[0050] In the above embodiments, the feature extraction network includes a cascade module and a projector module. The cascade module is used to extract stable common patterns of PDW sequences. It contains cascaded convolutional processing units and outputs preliminary feature representations. The projector module adaptively focuses on key difference features between different signal patterns and enhances the preliminary feature representations through an attention mechanism to achieve attention residual connections and cross-task feature alignment, thereby obtaining the final discriminative feature vector.

[0051] To address the issue of insufficient robustness in fine-grained pattern recognition of PDW signals under conditions of scarce labeled samples and complex signal environments, particularly when facing multiple interferences such as PDW loss, spurious PDWs, and parameter measurement errors. The base class space contains a large number of PDW sequence samples with PDW signal pattern labels; for example, this space contains dozens of known radar signal patterns. However, the PDW signal pattern labels in the base class space do not appear in the new recognition task during subsequent meta-testing. In the above embodiments, the PDW sequence samples are time-series data composed of radar pulse descriptor parameters. Each PDW sequence sample includes pulse width, carrier frequency, pulse amplitude, and pulse interval parameters derived from arrival time. Furthermore, each PDW sequence sample is standardized to a uniform format in terms of feature dimensions and time length.

[0052] This invention randomly samples multiple training tasks from the base class space, i.e., an N-way K-shot task. The random sampling from the base class space... i The training task includes: selecting from the base class space. N i There are several different categories. For each selected category, samples are randomly drawn from the corresponding category. K i A support set is constructed from a sample, and samples are randomly selected. Q i Construct a query set from a sample. Q i > K i It also supports sets and query sets that do not overlap.

[0053] The recognition network model in the above embodiments is a dual-path prototype contrast enhancement network (DPCE-Net). It extracts feature vectors from support set samples and query set samples using a feature extraction network employing dual-path feature decoupling and attention enhancement mechanisms. The cascade module extracts stable common patterns from PDW sequence samples, while the projector module adaptively focuses on key difference features between signal patterns. For each category in the training task, based on the feature vectors of the support samples, prototype feature vectors for each category are calculated through intra-class feature aggregation. Specifically, the prototype feature vectors for each category are averaged. Similarity is calculated based on the Euclidean distance between the query sample feature vector and the prototype feature vectors of each category. A triple-supervised loss function is then constructed, comprising classification loss, feature reconstruction loss, and prototype contrast loss. The classification loss is calculated based on the cross-entropy of negative distance; the feature reconstruction loss evaluates feature reconstruction capability and includes attention regularization; and the prototype contrast loss ensures effective separation of prototypes from different categories in the feature space. Backpropagation optimizes the recognition network model parameters, updates the current training task, and iterates until the model converges, resulting in the trained dual-path prototype contrast enhancement network (DPCE-Net). For a new recognition task containing multiple new PDW signal pattern categories, which include multiple new PDW signal pattern label categories that have not appeared in the base class category space, each new PDW signal pattern label provides multiple support samples and multiple query samples; the PDW signal pattern recognition of the query samples in the new recognition task can be achieved by using a trained dual-path prototype contrast enhancement network (DPCE-Net).

[0054] In the meta-training phase, the model is trained using various training tasks. The goal is to enable the feature extraction network to learn how to generalize high-quality feature representations (i.e., prototype feature vectors for each category) from a few samples of a category through a large number of small-sample training tasks. Simultaneously, the similarity between the query sample's feature vector and the prototype feature vectors of each category is calculated using an attention enhancement mechanism and Euclidean distance calculation. All model parameters are adjusted by optimizing a triple-supervised loss function that includes classification loss, feature reconstruction loss, and prototype comparison loss. In the meta-testing phase, when faced with a completely new recognition task with only a few samples, the trained feature extraction network calculates the corresponding prototype feature vectors for the new PDW signal style category based on the provided few samples. Then, it calculates the Euclidean distance between the query sample's feature vector and the prototype feature vectors of each category, converts the distance into probabilities, and outputs a fine-grained recognition result for the unknown signal style of the query sample. Specifically, the negative value of the distance is input into the softmax function to obtain the probability distribution of the query sample belonging to each new PDW signal style category. The category with the highest probability is taken as the recognition result for the unknown signal style of the query sample.

[0055] The above embodiments can be applied to improve the performance of PDW signal pattern recognition in complex signal environments. The technical problems they can solve include: (1) learning discriminative feature representations under the condition of scarce labeled samples; (2) maintaining stable recognition performance under interference conditions such as PDW loss and false PDW; (3) achieving effective generalization from ideal training environment to complex testing environment; and (4) accurate recognition of fine-grained patterns of PDW signals.

[0056] Reference Figure 2 In one embodiment, a recognition network model (Dual Path Prototype Contrast Enhancement Network DPCE-Net) is proposed, including a cascade module, a projector module, and a feature aggregation module. In the figure, FC represents a fully connected layer, BN represents a batch normalization layer, ReLU represents a non-linear activation function, GMP represents global max pooling, GAP represents global average pooling, Conv represents a convolutional layer, and Prototype represents the prototype feature vector. Softmax represents the Softmax function, which is used to convert the output into a probability distribution.

[0057] In one specific embodiment, the cascade module is used to extract stable common patterns from the PDW sequence. It includes three cascaded convolutional processing units, each of which sequentially includes a one-dimensional convolutional layer, a batch normalization layer, a nonlinear activation function, and a pooling operation. The one-dimensional convolutional layer of each convolutional processing unit sequentially extracts and fuses low-level, mid-level, and high-level features from the input PDW sequence samples. Each convolutional layer is followed by batch normalization, nonlinear activation, and pooling operations. Finally, adaptive average pooling is used for global aggregation to fuse multi-layer features and obtain a preliminary feature representation of the PDW sequence samples.

[0058] The projector module, which includes a cross-task attention projection network, receives the initial feature representation output from the cascade module. It generates an attention weight matrix through a learnable fully connected layer and performs attention weighting and residual enhancement on the input initial feature representation to align the attention residual connection with the cross-task features and obtain a fused feature vector. At the same time, the reconstruction network is used to reconstruct the enhanced feature vector through two fully connected layers for self-supervised training.

[0059] The feature aggregation module is used to globally aggregate the features processed by the projector module. It fuses the features of global average pooling with the output features of the projector module through residual connections, and uses them as the final discriminative feature vector of the PDW sequence samples.

[0060] Specifically, in one embodiment, the Dual-Path Prototype Contrast Enhancement Network (DPCE-Net) includes a cascade module, a projector module, and a feature convergence module. The cascade module comprises three cascaded convolutional processing units: the first-level convolutional processing unit uses a 1D convolution with a kernel size of 3 to extract 32-dimensional low-level features from the input channel 3, followed by batch normalization, ReLU activation, and max pooling; the second-level convolutional processing unit uses a 1D convolution with a kernel size of 3 to extract 64-dimensional mid-level features from the 32-dimensional input, followed by batch normalization and ReLU activation; the third-level convolutional processing unit uses a 1D convolution with a kernel size of 3 to extract 128-dimensional high-level features from the 64-dimensional input, followed by batch normalization and ReLU activation; finally, adaptive average pooling compresses the features into a 128-dimensional vector, outputting a preliminary feature vector.

[0061] The projector module adaptively focuses on key differences between various signal patterns. It comprises a convergent network, a projector network, and a reconstructor network: the convergent network performs a non-linear transformation on the initial feature vector output by the cascade through two fully connected layers; the projector network generates a 128-dimensional attention weight mask through fully connected layers and a softmax function; and feature enhancement is achieved by multiplying and summing the attention weight mask with the initial feature vector, resulting in an enhanced feature vector. The reconstructor network reconstructs the enhanced feature vector through two fully connected layers and is used for self-supervised training.

[0062] The feature aggregation module is used to globally aggregate the enhanced feature vectors after feature enhancement by the projector module. It converts the temporal features into fixed-dimensional feature vectors through adaptive average pooling, and obtains the final discriminative feature vectors of the PDW sequence samples, which serve as the input for subsequent processing.

[0063] In one embodiment, calculating the prototype feature vectors of each category based on the feature vectors of the support set samples includes: for any category, calculating the average of the feature vectors of all samples of the same category in the support set to obtain the prototype feature vector of the category.

[0064] In one embodiment, the Euclidean distance between the feature vectors of the query set samples and the prototype feature vectors of each category is calculated to characterize the similarity between the feature vectors of the query set samples and the prototype feature vectors of each category.

[0065] In one embodiment, the loss function comprises three components: classification loss, feature reconstruction loss, and prototype contrast loss. The classification loss is calculated using the cross-entropy function, the feature reconstruction loss includes feature reconstruction error and attention regularization term, and the prototype contrast loss achieves inter-class separation by penalizing prototypes of different classes that are too close in the feature space.

[0066] Specifically, classifying losses The similarity between the feature vectors of the query set samples and the prototype feature vectors of each category is calculated by combining the actual PDW signal style category labels of the query set samples; the specific calculation method is as follows:

[0067] ;

[0068] in, To query the number of samples in the set, For the number of categories, For the first The query set sample belongs to the first... Real labels for each category Indicates the first The feature vectors of each query set sample after feature extraction For the first Prototype feature vectors of each category, For the first The prototype feature vector of the i-th category is obtained by analyzing all elements belonging to the i-th category in the support set. The average of the feature vectors of the samples from each category is obtained. The distance is Euclidean.

[0069] Feature Reconstruction Loss Based on feature reconstruction error and attention regularization term, a formula is constructed to ensure the accuracy of feature reconstruction and the rationality of attention mechanism. The calculation formula is as follows:

[0070] ;

[0071] in, The feature vector output by the projector. The feature vectors reconstructed by the reconstructor network. The mean squared error loss function measures the difference between the original feature vector and the reconstructed feature vector. This represents the attention weight matrix generated by the projector module, used to highlight key features. The entropy of the attention weight matrix is ​​used to avoid excessive attention concentration. This represents the variance of the attention weight matrix, used to avoid excessive attentional distraction. and This is the regularization coefficient, and its specific value is not limited, such as 0.01 and 0.001 respectively.

[0072] Prototype contrast loss By penalizing prototype feature vector pairs that are too close together, we ensure that prototype feature vectors of different categories maintain an appropriate distance in the feature space. The calculation method is as follows: Let all prototype feature vectors be... ,in Set a distance margin for the number of categories. The formula for calculating the prototype contrast loss is:

[0073] ;

[0074] in, This indicates that all distances less than the set distance margin are... The number of prototype feature vector pairs. For the first Types and Number The prototype feature vectors of each category, where .

[0075] Prototype contrast loss The design aims to prevent prototype feature vectors of different categories from becoming too close. If the distance between the prototype feature vectors of two categories is less than... This indicates that they are too similar in the feature space, which increases the difficulty of classification and reduces the model's discriminative ability. By penalizing these prototype feature vector pairs that are too close, they are forced to maintain a certain distance interval (at least 2.0 units).

[0076] Ultimately, the overall loss function The weighted sum of classification loss, feature reconstruction loss, and prototype contrast loss:

[0077] ;

[0078] Based on the above loss function The model parameters are optimized by combining gradient accumulation and backpropagation mechanisms.

[0079] This invention ensures prototype stability through intra-class feature aggregation, enhances feature discriminativeness through attention enhancement, and synergistically optimizes classification performance, feature completeness, and inter-class separability through triple-supervised loss, ultimately achieving high-precision and robust PDW signal pattern recognition under conditions of a small number of labeled samples.

[0080] In one embodiment, step S5 includes:

[0081] For the new PDW signal style category which includes multiple new PDW signal style categories -way- - In the PDW signal pattern recognition task, multiple support samples are provided for each new category. The support set samples of the new recognition task are input into the trained recognition network model to extract feature vectors and calculate the prototype feature vectors of each new category. The query samples of the new recognition task are input into the trained recognition network model to extract feature vectors and calculate the similarity (Euclidean distance) between the feature vectors of the query samples and the prototype feature vectors of each new category. Based on the similarity results, a classification decision is made to complete the PDW signal pattern recognition of the query samples.

[0082] In another embodiment, a fine-grained small-sample PDW signal pattern recognition method is proposed, comprising the following steps:

[0083] Construct a set of PDW sequence samples with PDW signal style labels as the base class category space;

[0084] I N-way-K-shot small sample recognition training tasks are constructed by randomly sampling from the base class category space. For the i-th training task, the following steps are taken: select N different PDW signal pattern label categories from the base class category space; for each selected category, randomly select K samples from all its PDW sequence samples as support set; and randomly select Q samples from the remaining samples of the same category as query set, where Q > K.

[0085] Perform the i-th N-way K-shot training task, and extract the feature vectors of all support samples and query samples in the training task through a feature extraction network based on dual-path feature decoupling and attention enhancement mechanism; for each of these N categories, average the feature vectors of its corresponding K support samples to obtain the prototype feature vector of that category; calculate the Euclidean distance (an N-dimensional distance vector) between the feature vector of each query sample and the prototype feature vectors of these N categories.

[0086] Based on the Euclidean distance, a triple-supervised loss function is constructed: First, the negative value of the Euclidean distance is used as the classification logits, and the cross-entropy classification loss is calculated; simultaneously, the feature reconstruction loss is calculated, including the reconstruction error of enhanced features and the regularization of the attention mask; finally, the prototype contrast loss is calculated to constrain the distance between prototypes of different categories; the three losses are weighted and summed to obtain the total loss. Based on the total loss, the parameters of the feature extraction network model are optimized through backpropagation and gradient accumulation strategies until the current task converges, and then the (i+1)th training task is executed until the I training tasks are completed, resulting in a trained feature extraction network model.

[0087] For a given new N-way-K-shot recognition task, the feature vectors of its support set samples are extracted using a trained feature extraction network, and the prototype feature vectors of N new categories are obtained by averaging. The feature vectors of the query set samples are extracted, and the Euclidean distance between them and the prototype features of each new category is calculated. The distance is negative and then input into the softmax function to obtain the probability distribution of the query sample belonging to each category. The category corresponding to the highest probability is taken as the PDW signal pattern recognition result of the query sample.

[0088] In another embodiment, a fine-grained small-sample PDW signal pattern recognition device is provided, comprising:

[0089] The base class category space construction module is used to construct a set of PDW sequence samples with PDW signal style category labels as the base class category space;

[0090] The task sampling module is used to randomly sample multiple training tasks from the base class category space. Each training task contains a set of non-overlapping support set samples and query set samples.

[0091] The feature extraction module is used to extract feature vectors of each sample for the current training task using a feature extraction network. The feature extraction network includes a cascade module and a projector module. The cascade module is used to extract stable common patterns of the PDW sequence. It contains cascaded convolutional processing units and outputs preliminary feature representations. The projector module adaptively focuses on key difference features between different signal patterns. It uses an attention mechanism to weight and enhance the preliminary feature representations, realizes attention residual connections and cross-task feature alignment, and obtains the final discriminative feature vectors.

[0092] The prototype calculation module is used to calculate the prototype feature vectors of each category based on the feature vectors of the support set samples.

[0093] The training module is used to calculate the similarity between the feature vectors of the query set samples and the feature vectors of each prototype. Based on the similarity results, a loss function is constructed to optimize the parameters of the feature extraction network. The current training task is updated and iterated continuously until the model converges, and a trained recognition network model is obtained.

[0094] The recognition module is used for new recognition tasks containing multiple new PDW signal pattern categories. For each new category, multiple support samples are provided. The support set samples of the new recognition task are input into the trained recognition network model to extract feature vectors and calculate the prototype feature vectors of each new category. The query samples of the new recognition task are input into the trained recognition network model to extract feature vectors and calculate their similarity with the prototype feature vectors of each new category. Based on the similarity results, a classification decision is made to complete the PDW signal pattern recognition of the query samples.

[0095] To verify the effectiveness of the present invention, in one embodiment, a set of PDW sequence samples with PDW signal style labels is constructed as a base class category space, and a 5-way-1-shot training task is randomly sampled from the base class category space. For each 5-way-1-shot training task, samples are selected from the base class category space. N i =5 different PDW signal style label categories; for each selected label category, K samples are randomly drawn from the PDW sequence samples corresponding to that category. i =Construct the support set with 1 sample, and randomly select Q. i =15 samples are used to construct the query set, and the support set and query set do not overlap.

[0096] In another embodiment, a set of PDW sequence samples labeled with PDW signal patterns is constructed as a base class space, and 5-way 5-shot training tasks are randomly selected from the base class space. For each 5-way 5-shot training task, samples are selected from the base class space. N i =5 different PDW signal style label categories; for each selected label category, K samples are randomly drawn from the PDW sequence samples corresponding to that category. i =Construct a support set with 5 samples, and randomly select Q. i =15 samples are used to construct the query set, and the support set and query set do not overlap.

[0097] For the two embodiments described above, the recognition network model was trained based on the method proposed in this invention, and its recognition performance was verified. In the 5-way-1-shot task, facing 70% missed pulses and 80% spurious pulse interference, the recognition accuracy reached 79.58% and 81.92%, respectively. In the 5-way-5-shot task, even facing severe missed pulses and spurious pulse interference of 80%, the recognition accuracy remained at 84.00% and 90.38%, respectively. (Refer to...) Figure 3 The figure shows the model test results under different leak pulse ratios in the 5-way-1-shot mission; refer to Figure 4 The figure shows the model test results under different spurious impulse ratios in the 5-way-1-shot task; refer to Figure 5 The figure shows the model test results under different leak pulse ratios in the 5-way 5-shot mission; refer to Figure 6 The figure shows the model test results under different spurious pulse ratios in the 5-way 5-shot task. The two tasks were set to 5-way 1-shot and 5-way 5-shot respectively. The model can still maintain high recognition accuracy under small sample conditions with 10%-80% missing PDW ratio and 10%-80% spurious PDW signal ratio. This fully demonstrates that the present invention still exhibits strong anti-interference capability in complex signal environments.

[0098] On the other hand, the present invention provides an electronic device including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the fine-grained small-sample PDW signal pattern recognition method provided in any of the above embodiments. The computer device may be a server. The electronic device includes a processor, a memory, a network interface, and a database connected via a system bus. The processor of the electronic device provides computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the electronic device stores sample data. The network interface of the electronic device is used for communication with external terminals via a network connection.

[0099] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the fine-grained small sample PDW signal pattern recognition method provided in any of the above embodiments.

[0100] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0101] Matters not covered in this invention are common knowledge.

[0102] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0103] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

[0104] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fine-grained small-sample PDW signal pattern recognition method, characterized in that, include: Construct a set of PDW sequence samples with PDW signal style category labels as the base class space; Multiple training tasks are randomly sampled from the base class category space. Each training task contains a set of non-overlapping support set samples and query set samples. For the current training task, a feature extraction network is used to extract feature vectors for each sample. The feature extraction network includes a cascade module and a projector module. The cascade module is used to extract stable common patterns of PDW sequences. It contains cascaded convolutional processing units and outputs preliminary feature representations. The projector module adaptively focuses on key difference features between different signal patterns. It uses an attention mechanism to weight and enhance the preliminary feature representations, realizes attention residual connections and cross-task feature alignment, and obtains the final discriminative feature vectors. Calculate the prototype feature vectors for each category based on the feature vectors of the support set samples; The similarity between the feature vectors of the query set samples and the feature vectors of each prototype is calculated. Based on the similarity results, a loss function is constructed to optimize the parameters of the feature extraction network. The current training task is updated and iterated continuously until the model converges, resulting in a trained recognition network model. The loss function includes three components: classification loss, feature reconstruction loss, and prototype contrast loss. For a new recognition task containing multiple new PDW signal pattern categories, multiple support samples are provided for each new category. The support set samples of the new recognition task are input into the trained recognition network model to extract feature vectors and calculate the prototype feature vectors of each new category. The query samples of the new recognition task are input into the trained recognition network model to extract feature vectors and calculate their similarity with the prototype feature vectors of each new category. Based on the similarity results, a classification decision is made to complete the PDW signal pattern recognition of the query samples.

2. The fine-grained small-sample PDW signal pattern recognition method according to claim 1, characterized in that, The PDW sequence samples are time-series data composed of radar pulse descriptor parameters. Each PDW sequence sample includes pulse width, carrier frequency, pulse amplitude, and pulse interval parameters derived from arrival time. Furthermore, each PDW sequence sample is standardized to a uniform format in terms of feature dimensions and time length. Randomly sample the first class from the base class space. i The training task includes: selecting from the base class space. N i There are several different categories. For each selected category, samples are randomly drawn from the corresponding category. K i A support set is constructed from a sample, and samples are randomly selected. Q i Construct a query set from a sample. Q i > K i It also supports sets and query sets that do not overlap.

3. The fine-grained small-sample PDW signal pattern recognition method according to claim 1, characterized in that, The feature extraction network includes a cascade module, a projector module, and a feature convergence module; The cascade module contains three cascaded convolutional processing units. Each convolutional processing unit sequentially includes a one-dimensional convolutional layer, a batch normalization layer, a non-linear activation function, and a pooling operation. The one-dimensional convolutional layer of each convolutional processing unit sequentially extracts and fuses the bottom, middle, and high-level features from the PDW sequence. Each convolutional processing unit is followed by batch normalization, non-linear activation, and pooling operations. Finally, adaptive average pooling is used to perform global convergence to fuse multi-layer features and obtain the preliminary feature representation of the PDW sequence samples. The projector module, which includes a cross-task attention projection network, receives the initial feature representation output from the cascade module. It generates an attention weight matrix through a learnable fully connected layer and performs attention weighting and residual enhancement on the input initial feature representation to align the attention residual connection with cross-task features and obtain a fused feature vector. At the same time, the reconstruction network is used to reconstruct the enhanced feature vector through two fully connected layers for self-supervised training. The feature aggregation module is used to globally aggregate the features processed by the projector module. It fuses the features of global average pooling with the output features of the projector module through residual connections, and uses them as the final discriminative feature vector of the PDW sequence samples.

4. The fine-grained small-sample PDW signal pattern recognition method according to any one of claims 1 to 3, characterized in that, The prototype feature vectors of each category are calculated based on the feature vectors of the support set samples. This includes: for any category, calculating the average of the feature vectors of all samples of the same category in the support set to obtain the prototype feature vector of the category.

5. The fine-grained small-sample PDW signal pattern recognition method according to claim 4, characterized in that, Calculate the Euclidean distance between the feature vectors of the query set samples and the prototype feature vectors of each category to characterize the similarity between the feature vectors of the query set samples and the prototype feature vectors of each category.

6. The fine-grained small-sample PDW signal pattern recognition method according to claim 4, characterized in that, Classification loss The similarity between the feature vectors of the query set samples and the prototype feature vectors of each category is calculated by combining the real PDW signal style category labels of the query set samples. Feature Reconstruction Loss Based on feature reconstruction error and attention regularization term, a formula is constructed to ensure the accuracy of feature reconstruction and the rationality of attention mechanism. The calculation formula is as follows: in, The feature vector output by the projector. The feature vectors reconstructed by the reconstructor network. Let the mean squared error loss function be . This represents the attention weight matrix generated by the projector module; The entropy represents the attention weight matrix; This represents the variance of the attention weight matrix; and The regularization coefficient is used. Prototype contrast loss By penalizing prototype feature vector pairs that are too close together, we ensure that prototype feature vectors of different categories maintain an appropriate distance in the feature space. The calculation method is as follows: Let all prototype feature vectors be... ,in Set a distance margin for the number of categories. The formula for calculating the prototype contrast loss is: in, This indicates that all distances less than the set distance margin are... The number of prototype feature vector pairs. For the first Types and Number The prototype feature vectors of each category, where ; Final overall loss function The weighted sum of classification loss, feature reconstruction loss, and prototype contrast loss: 。 7. A fine-grained small-sample PDW signal pattern recognition device, used to implement the fine-grained small-sample PDW signal pattern recognition method as described in claim 1, 2, 3, 5, or 6, characterized in that, include: The base class category space construction module is used to construct a set of PDW sequence samples with PDW signal style category labels as the base class category space; The task sampling module is used to randomly sample multiple training tasks from the base class category space. Each training task contains a set of non-overlapping support set samples and query set samples. The feature extraction module is used to extract feature vectors of each sample for the current training task using a feature extraction network. The feature extraction network includes a cascade module and a projector module. The cascade module is used to extract stable common patterns of the PDW sequence. It contains cascaded convolutional processing units and outputs preliminary feature representations. The projector module adaptively focuses on key difference features between different signal patterns. It uses an attention mechanism to weight and enhance the preliminary feature representations, realizes attention residual connections and cross-task feature alignment, and obtains the final discriminative feature vectors. The prototype calculation module is used to calculate the prototype feature vectors of each category based on the feature vectors of the support set samples. The training module is used to calculate the similarity between the feature vectors of the query set samples and the feature vectors of each prototype. Based on the similarity results, a loss function is constructed to optimize the parameters of the feature extraction network. The current training task is updated and iterated continuously until the model converges, and a trained recognition network model is obtained. The recognition module is used for new recognition tasks containing multiple new PDW signal pattern categories. For each new category, multiple support samples are provided. The support set samples of the new recognition task are input into the trained recognition network model to extract feature vectors and calculate the prototype feature vectors of each new category. The query samples of the new recognition task are input into the trained recognition network model to extract feature vectors and calculate their similarity with the prototype feature vectors of each new category. Based on the similarity results, a classification decision is made to complete the PDW signal pattern recognition of the query samples.

8. An electronic device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes a computer program to implement the steps of the fine-grained small sample PDW signal pattern recognition method as described in claim 1.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the fine-grained small sample PDW signal pattern recognition method as described in claim 1.

Images