Radar emitter non-exemplar set incremental identification method based on distribution reconstruction

Abstract

This application relates to the field of radar signal processing technology, and particularly to a paradigm-free incremental identification method for radar radiation sources based on distributed reconstruction. The method includes: preprocessing the pulse sequence of the radar radiation source based on descriptor parameters of the pulse sequence; coupling and mapping the original parameters and dynamic features through multiple views to generate an image-based signal representation; extracting features from the image-based signal representation using a deep learning model and training it with a metric learning loss function based on angular intervals; performing paradigm-free incremental learning when receiving radar data of a new category; merging the radar data of the new category with pseudo-feature samples of the old category for joint training of the deep learning model; and updating the parameters of the deep learning model and the low-dimensional statistical library of the old category. This method can achieve high-performance continuous learning with a balance between new and old categories while ensuring high identification accuracy under harsh conditions and without requiring zero original data storage.

Description

Technical Field

[0001] The embodiments of this application relate to the field of radar signal processing technology, and in particular to a method for incremental identification of radar radiation sources without a paradigm set based on distributed reconstruction. Background Technology

[0002] Specific Emitter Identification (SEI) is the core of electronic warfare systems. With the development of radar technology, radar signals exhibit characteristics such as rapid parameter changes, diverse modulation patterns, and harsh operating environments. Traditional methods based on pulse description word (PDW) statistical histograms are no longer sufficient for high-precision identification. Furthermore, facing a constant stream of new radar types, identification systems must possess "lifelong learning" capabilities, but are often limited by hardware resources and security protocols, preventing the storage of massive amounts of historical raw data.

[0003] Existing radar radiation source identification schemes suffer from the following main problems: First, insufficient feature robustness: In complex electromagnetic environments, radar signals are susceptible to parameter drift, pulse loss, and low signal-to-noise ratio. Traditional methods (such as time-frequency analysis based on wavelet transform) ignore the dynamic laws of signal parameters (such as the rate of frequency change), leading to unstable feature extraction and a significant drop in identification accuracy under adverse conditions. Second, catastrophic forgetting in incremental learning: When a system learns a new category, deep learning models quickly forget old knowledge. The existing patent "Method for Reconstructing Old Category Distribution Based on K-Center Features" only approximates the old category distribution using discrete cluster centers, failing to capture feature covariance information, resulting in distorted distribution reconstruction. Another patent, "Method Based on Parameter Freezing and Wavelet Residual Features," while reducing forgetting by freezing network parameters, severely limits the ability to learn new categories, resulting in poor plasticity. Third, data privacy and storage bottlenecks: In high-security scenarios, original signal data cannot be stored, and incremental learning methods without a paradigm set struggle to accurately reconstruct the old category distribution, causing an imbalance in the recognition performance of new and old categories.

[0004] Therefore, how to achieve high-performance continuous learning that balances new and old categories while ensuring high recognition accuracy under harsh working conditions and without storing zero original data is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] To address the aforementioned technical issues, embodiments of this application propose a paradigmless incremental identification method for radar radiation sources based on distributed reconstruction. This method can achieve high-performance continuous learning with a balance between old and new categories while ensuring high identification accuracy under harsh operating conditions and storing zero original data.

[0006] To achieve the above objectives, embodiments of this application propose a paradigm-free incremental identification method for radar radiation sources based on distributed reconstruction, the method comprising: Receive the pulse sequence from the radar radiation source; Based on the descriptor parameters of the pulse sequence, the pulse sequence of the radar radiation source is preprocessed to obtain the preprocessed original parameters and dynamic characteristics. The original parameters and dynamic characteristics are then coupled and mapped by multiple views to generate a visual signal representation. Using a deep learning model, features are extracted from the image-based signal representation, and a metric learning loss function based on angle interval is used for training to form a compact intra-class and separable inter-class feature distribution on the hypersphere of the feature space; wherein, the compact intra-class and separable inter-class feature distribution reflects that the angle difference of intra-class feature points is less than or equal to a first value, and the angle difference of inter-class feature points is greater than a preset interval. When new category radar data is received, incremental learning without a paradigm is performed; for the old categories that have been learned, the low-dimensional statistics of their features are stored, and the feature distribution of the old categories is reconstructed based on the low-dimensional statistics, and pseudo-feature samples are generated. The radar data of the new category is merged with the pseudo-feature samples of the old category to jointly train the deep learning model. During the training process, knowledge distillation constraints are used to stabilize the feature extraction capability of the old category, and the classifier weights are aligned to eliminate the classification bias between the new and old categories. Update the parameters of the deep learning model and the low-dimensional statistics library of the old categories to complete the current incremental learning phase.

[0007] To achieve the above objectives, embodiments of this application also propose a paradigmless incremental identification system for radar radiation sources based on distributed reconstruction, the system comprising: The receiving module is used to receive the pulse sequence of the radar radiation source; The preprocessing module is used to preprocess the pulse sequence of the radar radiation source based on the descriptor parameters of the pulse sequence, obtain the preprocessed original parameters and dynamic characteristics, and perform multi-view coupling mapping on the original parameters and dynamic characteristics to generate a graphical signal representation. The feature extraction module is used to extract features from the image-based signal representation using a deep learning model, and is trained using a metric learning loss function based on angle intervals to form a compact intra-class and separable inter-class feature distribution on the hypersphere of the feature space. The compact intra-class and separable inter-class feature distribution reflects that the angle difference of intra-class feature points is less than or equal to a first value, and the angle difference of inter-class feature points is greater than a preset interval. The incremental learning module is used to perform paradigm-free incremental learning when new categories of radar data are received. For the old categories that have been learned, the low-dimensional statistics of their features are stored, and the feature distribution of the old categories is reconstructed based on the low-dimensional statistics, and pseudo-feature samples are generated. The joint training module is used to merge radar data of the new category with pseudo-feature samples of the old category to jointly train the deep learning model. During the training process, knowledge distillation constraints are used to stabilize the feature extraction capability of the old category, and the classifier weights are aligned to eliminate the classification bias between the new and old categories. The model update module is used to update the parameters of the deep learning model and the low-dimensional statistics library of the old categories to complete the current incremental learning phase.

[0008] To achieve the above objectives, embodiments of this application also propose an electronic device, including a processor and a memory, wherein the memory stores instructions executable by the processor, and the processor is configured to execute the instructions such that the electronic device can implement the above-described method for incremental identification of radar radiation sources without paradigm sets based on distributed reconstruction.

[0009] To achieve the above objectives, embodiments of this application also propose a computer-readable storage medium storing a computer program that, when executed by a processor, enables the implementation of a paradigmless incremental identification method for radar radiation sources based on distributed reconstruction, as described above.

[0010] This application proposes a paradigm-free incremental identification method for radar radiation sources based on distribution reconstruction. First, it receives the pulse sequence of the radar radiation source. Then, based on the descriptor parameters of the pulse sequence, it preprocesses the pulse sequence to obtain preprocessed original parameters and dynamic features. The original parameters and dynamic features are then coupled and mapped using multiple views to generate an image-based signal representation. Next, a deep learning model is used to extract features from the image-based signal representation, and a metric learning loss function based on angular intervals is used for training to form a compact intra-class and segregated inter-class feature distribution on a hypersphere in the feature space. Finally, when radar data of a new category is received, paradigm-free incremental learning is performed. Next, the radar data of the new category is merged with the pseudo-feature samples of the old category to jointly train the deep learning model. Finally, the parameters of the deep learning model and the low-dimensional statistical library of the old category are updated to complete the current incremental learning stage. Among them, the feature distribution of compact intra-class and separated inter-class reflects that the angle difference of intra-class feature points is less than or equal to the first value, and the angle difference of inter-class feature points is greater than the preset interval. In addition, during training, the feature extraction capability of the old category can be stabilized based on knowledge distillation constraints, and the classifier weights can be aligned to eliminate the classification bias between the new and old categories. Therefore, this scheme can achieve high-performance continuous learning with a balance between the new and old categories while ensuring high recognition accuracy under harsh working conditions and with zero original data storage.

[0011] Optionally, the pulse sequence of the radar radiation source includes: pulse arrival time, carrier frequency, pulse width, and pulse amplitude; The pulse sequence of a radar radiation source is represented by a pulse description word vector, which is represented as follows: ; in, The pulse arrival time, For carrier frequency, The pulse width. The amplitude of the pulse. An index representing the pulse sequence of a radar radiation source; The descriptive word parameters based on the pulse sequence preprocess the pulse sequence of the radar radiation source to obtain the preprocessed original parameters and dynamic characteristics. The original parameters and dynamic characteristics are then coupled and mapped using a multi-view method to generate a graphical signal representation, including: Perform differential operations on the descriptor parameters of the pulse sequence to extract at least first-order differential features; wherein, the differential features include the carrier frequency change rate. and the rate of change of repetition frequency The calculation formula is as follows: ; ; ; in, Indicates the frequency of repetition; The extracted raw parameters and dynamic difference features are normalized and mapped to a preset numerical range, as shown in the following formula: The formula for nonlinearly normalizing the carrier frequency change rate is as follows: ; in, This represents the rate of change of the carrier frequency after nonlinear compression processing; This represents the scaling factor, used to adjust the slope of the hyperbolic tangent function during nonlinear compression, so as to map the carrier frequency change rate to a preset numerical range (0,1). Physically correlated parameters or feature combinations are mapped to different two-dimensional view planes, and pixel-level density accumulation is performed on each view plane to generate multi-channel feature images.

[0012] Optionally, the multi-channel feature image is a three-channel RGB image; the multi-channel feature image enhances contrast through logarithmic transformation and connects broken trajectories caused by pulse loss through Gaussian smoothing; The three-channel RGB includes: The R channel is generated by the joint distribution mapping of the carrier frequency and the repetition frequency; The G channel is generated by the joint distribution mapping of the carrier frequency and the carrier frequency change rate; Channel B is generated by the joint distribution mapping of repetition rate and pulse width; The formula for pixel-level density accumulation is as follows: ; in, These are pixel coordinates; Indicates the first The x-coordinate of the two-dimensional coordinates of each pulse sample point projected onto the view plane. Indicates the first The ordinate of the two-dimensional coordinates of each pulse sample point after it has been projected onto the view plane; right Perform a logarithmic transformation to obtain the transformed features To enhance contrast; the formula for the logarithmic transformation is as follows: .

[0013] Optionally, the deep learning model uses a residual network as the backbone network and embeds a global gating attention module; The process of extracting features from image-based signal representations using a deep learning model includes: By utilizing a global gated attention module, adaptive weighting of feature channels and spatial regions is achieved to suppress noise and enhance the focusing of effective signal trajectories; The global gating attention module includes channel attention. and spatial attention Features are enhanced using the following formula: ; in, It is the original extracted feature map tensor; It is the weighted enhanced feature map tensor of F after applying channel attention and spatial attention; This indicates channel attention.

[0014] Optionally, a metric learning loss function based on angular intervals is used for training to form a compact intra-class and segregated inter-class feature distribution on the hypersphere of the feature space, including: eigenvectors and classifier weights Normalized to the hypersphere, the formula is as follows: ; ; The feature space is compressed using an angle-margin-based metric learning loss function, the formula of which is as follows: ; in, It is the angle between the feature vector and the true class center; is the additive angular interval; N represents the number of samples participating in the current loss calculation; s is the scale factor used to amplify the cosine similarity; Indicates the first The feature vector of the i-th sample and the i-th sample The angle between class centers; a metric learning loss function based on angular intervals is used to force features to form a compact distribution on the hypersphere.

[0015] Optionally, the low-dimensional statistics of the old categories include the class prototype mean. Covariance Matrix It can be calculated using the following formula: ; ; in, Indicates belonging to a category The number of samples; Indicate category Index of internal samples; It is a category The Middle The original input data for each training sample; The generation of pseudo-feature vectors is achieved through multivariate Gaussian sampling: ; in, To prevent singularity in the covariance matrix, small perturbation terms are used to enhance clarity; continuous pseudo-features are generated by sampling to approximate the old class distribution.

[0016] Optionally, based on knowledge distillation constraints to stabilize the feature extraction capability for old categories, and performing alignment operations on classifier weights to eliminate classification bias between new and old categories, including: Using the old model as the teacher network model, the network's feature output is extracted through knowledge distillation to constrain new features; where, the knowledge distillation loss function... It is expressed as follows: ; in, This represents the new feature extraction network to be trained in the current incremental phase; This represents the old feature extraction network whose parameters have been frozen in the previous stage; To eliminate the classification bias between the old and new classes, the weights of the new classifier are aligned. The alignment formula is as follows: ; in, Indicates the weights of the old classifier. Indicates the weights of the new classifier. This represents the arithmetic mean; the weight alignment operation is used to eliminate classification bias caused by sample imbalance.

[0017] Optionally, the parameters of the deep learning model and the low-dimensional statistics library of the old classes are updated to complete the current incremental learning stage, including: merging the corrected new class weights with the old class weights to form the complete classifier for the current stage; recalculating the feature statistics of the new classes and adding them to the prototype library, while supporting the selection of momentum update or maintaining the initial state of the feature statistics of the old classes; freezing the updated model parameters as the teacher network model for the next incremental stage. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies of this application will be briefly introduced below. Obviously, the following drawings are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The drawings described herein are only used to explain this application and are not intended to limit this application.

[0019] Figure 1 This is a flowchart of a paradigm-free incremental identification method for radar radiation sources based on distributed reconstruction, provided in one embodiment of this application; Figure 2 This is a schematic diagram of a class incremental identification process in a reconnaissance system provided in one embodiment of this application; Figure 3 This is a detailed flowchart of another paradigm-free incremental identification method for radar radiation sources based on distributed reconstruction, provided in one embodiment of this application; Figure 4 This is a schematic diagram of the structure of a paradigmless incremental identification system for radar radiation sources based on distributed reconstruction, provided in another embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in another embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. Those skilled in the art will understand that many technical details have been presented in the embodiments of this application to facilitate better understanding. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of this application. The following embodiments can be combined with and referenced by each other without contradiction.

[0021] Specific Emitter Identification (SEI) is the core of electronic warfare systems. With the development of radar technology, radar signals exhibit characteristics such as rapid parameter changes, diverse modulation patterns, and harsh operating environments. Traditional methods based on pulse description word (PDW) statistical histograms are no longer sufficient for high-precision identification. Furthermore, facing the ever-evolving landscape of new radar types, identification systems must possess "lifelong learning" capabilities, but are often limited by hardware resources and security protocols, preventing the storage of massive amounts of historical raw data. Existing technologies propose the following solutions: Patent 1 proposes a feature distribution reconstruction method based on the K-center algorithm. Its core idea is to use a feature extractor trained on an old task during the incremental learning phase to calculate multiple cluster centers (K centers) for old class samples and save these centers as representatives of the old classes. When learning new classes, the saved K centers are used to reconstruct the feature distribution of the old classes, thus replacing the original data for playback. Although this method attempts to solve the problem of no data playback, it essentially still relies on discrete points (K centers) to approximate a continuous feature distribution. For complex multimodal radar signals, the limited number of K centers is insufficient to cover the true distribution boundary, and it ignores the covariance (shape) information of the feature distribution, leading to deviations between the reconstructed pseudo-features and the true distribution, and still carries the risk of forgetting. Furthermore, this method mainly focuses on feature-level reconstruction and lacks the mining of front-end signal dynamic features.

[0022] Patent 2 proposes a cross-domain incremental radiation source identification method based on wavelet residual feature extraction. This invention focuses on incremental identification in cross-domain scenarios. It utilizes continuous wavelet transform to obtain time-frequency maps, which are then input into a residual network to extract multi-scale features. In the incremental stage, a parameter freezing strategy is adopted, freezing all model parameters except for the classifier, and only updating the classifier. While the parameter freezing strategy can preserve old knowledge very well (high stability), it severely limits the model's ability to learn new types of features (poor plasticity). For new radar systems that differ significantly from the old types, the frozen feature extractor cannot extract effective discriminative features, resulting in a low recognition rate for new types. Furthermore, its input is still based on static wavelet time-frequency maps, lacking a dynamic mechanism to combat parameter drift.

[0023] Existing radar radiation source identification schemes suffer from the following main problems: First, insufficient feature robustness: In complex electromagnetic environments, radar signals are susceptible to parameter drift, pulse loss, and low signal-to-noise ratio. Traditional methods (such as time-frequency analysis based on wavelet transform) ignore the dynamic laws of signal parameters (such as the rate of frequency change), leading to unstable feature extraction and a significant drop in identification accuracy under adverse conditions. Second, catastrophic forgetting in incremental learning: When a system learns a new category, deep learning models quickly forget old knowledge. The existing patent "Method for Reconstructing Old Category Distribution Based on K-Center Features" only approximates the old category distribution using discrete cluster centers, failing to capture feature covariance information, resulting in distorted distribution reconstruction. Another patent, "Method Based on Parameter Freezing and Wavelet Residual Features," while reducing forgetting by freezing network parameters, severely limits the ability to learn new categories, resulting in poor plasticity. Third, data privacy and storage bottlenecks: In high-security scenarios, original signal data cannot be stored, and incremental learning methods without a paradigm set struggle to accurately reconstruct the old category distribution, causing an imbalance in the recognition performance of new and old categories.

[0024] Therefore, how to achieve high-performance continuous learning that balances new and old categories while ensuring high recognition accuracy under harsh working conditions and without storing zero original data is a technical problem that urgently needs to be solved.

[0025] In view of this, embodiments of this application provide a paradigmless incremental identification method for radar radiation sources based on distributed reconstruction, which can achieve high identification accuracy under harsh conditions and high-performance continuous learning with a balance between old and new categories without the need for zero original data storage.

[0026] One embodiment of this application proposes a paradigm-free incremental identification method for radar radiation sources based on distributed reconstruction, applied to an electronic device. The electronic device can be a terminal or a server; this embodiment and subsequent embodiments will use a server as an example. The implementation details of the paradigm-free incremental identification method for radar radiation sources based on distributed reconstruction proposed in this embodiment are described below. The following implementation details are provided for ease of understanding and are not essential for implementing this solution.

[0027] The specific process of the paradigmless incremental identification method for radar radiation sources based on distributed reconstruction proposed in this embodiment can be described as follows: Figure 1 As shown, it includes steps 101 to 106.

[0028] Step 101: Receive the pulse sequence from the radar radiation source.

[0029] Understandably, to verify the effectiveness of the embodiments of this application under complex modulation types and dynamic incremental scenarios, a radar radiation source dataset containing multiple systems and covering a wide range of parameters is first constructed. For example, a radar radiation source pulse sequence intercepted by a reconnaissance receiver is simulated.

[0030] The pulse sequence is represented in the form of pulse descriptors, and each pulse sample contains parameters such as time, frequency, width and amplitude.

[0031] In one possible embodiment, the pulse sequence of the radar radiation source includes: pulse arrival time, carrier frequency, pulse width, and pulse amplitude; the pulse sequence of the radar radiation source is represented by a pulse description word vector, which is represented as follows: ; in, The pulse arrival time, For carrier frequency, The pulse width. The amplitude of the pulse. An index representing the pulse sequence of a radar radiation source; To simulate real-world environments, parameter settings cover various radar systems (such as conventional pulse, linear frequency modulation, frequency agility, etc.). The carrier frequency range covers S, C, X, and Ku bands, and the repetition frequency types include fixed, staggered, and jitter modes.

[0032] For example, in frequency-agile radar, the carrier frequency follows a uniform distribution within bandwidth B. Or discrete frequency hopping pattern; For example, for a fixed repetition frequency, ,in, It is a very small timing jitter.

[0033] For example, for uneven repetition rates, PRI is in a set of fixed values. The loop switches between different loops.

[0034] For example, regarding jitter repetition frequency, ,in, As the reference repetition rate value, Simulate the intensity of jitter.

[0035] Complex electromagnetic interference is injected into the pulse sequence, including parameter drift (such as a 10% random offset introduced by carrier frequency and repetition frequency) and pulse loss (probabilistically and randomly discarding pulses) to simulate the harsh conditions of non-cooperative reconnaissance.

[0036] Step 102: Based on the descriptor parameters of the pulse sequence, the pulse sequence of the radar radiation source is preprocessed to obtain the preprocessed original parameters and dynamic characteristics. The original parameters and dynamic characteristics are then coupled and mapped using a multi-view method to generate a graphical signal representation.

[0037] Understandably, by using spatiotemporal-physical coupled dynamic density field mapping, PDW sequences can be converted into image representations, thereby enhancing the robustness of features.

[0038] In one possible embodiment, step 102 above includes: performing a difference operation on the descriptor parameters of the pulse sequence to extract at least first-order difference features. Among them, the differential characteristics include the carrier frequency change rate. and the rate of change of repetition frequency The calculation formula is as follows: ; ; ; in, Indicates the frequency of repetition.

[0039] In physics, It can capture frequency agility patterns (such as frequency hopping step size), has translation invariance, and can resist overall parameter drift.

[0040] The extracted raw parameters and dynamic difference features are normalized and mapped to a preset numerical range, as shown in the following formula: Because the dynamic range of the difference feature is large (e.g. (Up to GHz level), using hyperbolic tangent function compression: Specifically, the formula for nonlinearly normalizing the carrier frequency change rate is as follows: ; in, This represents the rate of change of the carrier frequency after nonlinear compression processing; This represents the scaling factor, used to adjust the slope of the hyperbolic tangent function during nonlinear compression, so as to map the carrier frequency change rate to a preset numerical range (0,1). Physically correlated parameters or feature combinations are mapped to different two-dimensional view planes, and pixel-level density accumulation is performed on each view plane to generate multi-channel feature images.

[0041] In one possible embodiment, the multi-channel feature image is a three-channel RGB image; the multi-channel feature image enhances contrast through logarithmic transformation and connects broken trajectories caused by pulse loss through Gaussian smoothing; The three-channel RGB includes: The R channel consists of the carrier frequency and the repetition frequency ( - The joint distribution mapping is generated; The G channel consists of the carrier frequency and the rate of change of the carrier frequency ( - The joint distribution mapping is generated; Channel B is determined by the repetition rate and pulse width ( - The joint distribution mapping is generated; The formula for pixel-level density accumulation is as follows: ; in, These are pixel coordinates; Indicates the first The x-coordinate of the two-dimensional coordinates of each pulse sample point projected onto the view plane. Indicates the first The ordinate of the two-dimensional coordinates of each pulse sample point after it has been projected onto the view plane; right Perform a logarithmic transformation to obtain the transformed features To enhance contrast; the formula for the logarithmic transformation is as follows: .

[0042] Then, Gaussian smoothing is applied to connect the broken trajectories caused by pulse loss, ultimately generating a 224×224×3 feature image.

[0043] Understandably, the spacetime-physics coupled dynamic density field mapping (ST-PCDM) utilizes the physical invariance of difference features, for example... Insensitive to overall drift and with multi-view complementarity, it converts time-series pulse sequences into static images, preserves dynamic laws, and improves robustness against parameter drift and pulse loss.

[0044] Optionally, although the embodiments of this application use a pulse descriptor word (PDW) sequence as direct input, the embodiments of this application are not limited to this input data. The input data can also be intermediate frequency sampling data of radar signals or orthogonal (I / Q) waveform data, etc. Under this alternative, the first step of the system processing needs to be changed accordingly. For example, conventional pulse detection and parameter estimation can be performed on the intermediate frequency or I / Q waveform data first to obtain the PDW sequence, and then the ST-PCDM process described in this invention can be executed. Another example is to directly perform time-frequency transformation on the waveform data to generate a time-frequency image as network input. The time-frequency transformation method includes, but is not limited to: short-time Fourier transform, continuous wavelet transform, or Cohen-Williams distribution with higher time-frequency clustering, etc.; even if a time-frequency image is used, the above embodiments can still be used as a reference to introduce dynamic information by calculating the inter-frame difference between adjacent time-frequency images.

[0045] Optionally, embodiments of this application use first-order differences (ΔRF, ΔPRI) as dynamic features. To characterize more complex nonlinear modulation laws, the dynamic features can be extended to second-order or even higher-order differences. For example, the acceleration of the carrier frequency change rate (second-order difference, Δ²RF) or the jerk of the repetition rate change (third-order difference) can be further calculated. These higher-order difference features can reveal deeper patterns of signal parameter changes, which is of positive significance for identifying complex modulation radars (such as nonlinear frequency modulation). Therefore, any processing method that introduces the parameter change rate and its higher-order form as a feature dimension should fall within the protection scope of this application.

[0046] Optionally, embodiments of this application will - , - , - Three specific views are mapped to an RGB three-channel image. Multi-view coupled projection can also be replaced by a process of multi-channel tensor stacking and feature concatenation fusion. The original parameters and their differential features (ΔRF, ΔPRI, etc.) can be used as independent feature maps, then stacked along the channel dimension to form a multi-channel tensor input network. For example, a 6-channel input tensor can be represented as: The image generated by the original parameter mapping can be combined with the image generated by the differential feature mapping through channel stitching or early feature fusion. For example, " - "Corresponding view and " - "Corresponding view stitching enables the network to simultaneously perceive the distribution of static parameters and dynamic change patterns."

[0047] Step 103: Use a deep learning model to extract features from the image-based signal representation and train it using a metric learning loss function based on angular intervals to form a compact intra-class and separated inter-class feature distribution on the hypersphere of the feature space.

[0048] Among them, the feature distribution of compact intra-class and separated inter-class reflects that the angle difference of feature points within a class is less than or equal to the first value, and the angle difference of feature points between classes is greater than the preset interval.

[0049] Understandably, the DynMet-GMA-Net feature extractor can be trained to compress the feature distribution through metric learning, laying the mathematical foundation for incremental learning.

[0050] For example, ResNet-18 can be used as the backbone network, the fully connected layers can be removed, and the convolutional layers can be retained as feature extractors to output the feature dimensions. .

[0051] Optionally, ResNet-18 is preferred only for embodiments that balance performance and efficiency. The backbone network can be replaced with other deep network architectures. For example, for higher parameter efficiency, it can be replaced with EfficientNet; for adaptation to computationally limited mobile or embedded deployments, it can be replaced with MobileNet and its variants; for enhanced feature reuse capabilities, it can be replaced with DenseNet; and for better utilization of self-attention mechanisms to capture long-range dependencies in images, it can be replaced with VisionTransformer (ViT) or its lightweight variants.

[0052] Embed a global gated attention module (CBAM) after each residual block: Channel attention Automatically weights the RGB channels based on the signal modulation type (e.g., amplifying the G channel of a frequency-agile signal).

[0053] Spatial attention Generate a two-dimensional mask to suppress background noise pixels and focus on the effective signal trajectory.

[0054] Optionally, the Convolutional Block Attention (CBAM) module used in the embodiments of this application combines channel and spatial attention; any attention mechanism that can achieve adaptive feature enhancement can be used as an alternative. For example, it can be replaced by a squeeze excitation module (SE-Block) that focuses only on inter-channel relationships; or its efficient variant ECA-Net; or it can be replaced by coordinate attention that can simultaneously capture channel relationships and long-distance positional information. Any structure that inserts attention modules between convolutional layers to adaptively weight the feature map channels or spatial dimensions, thereby suppressing invalid background pixels generated by impulse loss or noise, is an equivalent alternative to the embodiments of this application.

[0055] In one possible embodiment, the deep learning model employs a residual network as the backbone network and embeds a global gated attention module; the feature extraction of the image-based signal representation using the deep learning model includes: using the global gated attention module to adaptively weight feature channels and spatial regions to suppress noise and enhance the focusing of effective signal trajectories.

[0056] The global gating attention module includes channel attention. and spatial attention Features are enhanced using the following formula: ; in, It is the original extracted feature map tensor (i.e., the input feature map); It is the weighted enhanced feature map tensor of F after applying channel attention and spatial attention; Indicates channel attention; This indicates element-wise multiplication.

[0057] For example, the additive angular interval of the metric learning loss function with angular interval (i.e., the ArcFace loss function) directly constrains the topological structure of the feature space, making the distribution conform to the single-center Gaussian assumption.

[0058] In one possible embodiment, training is performed using a metric learning loss function based on angular intervals to form a compact intra-class, inter-class separated feature distribution on a hypersphere in the feature space, including: eigenvectors and classifier weights Normalized to the hypersphere, the formula is as follows: ; ; The feature space is compressed using an angle-margin-based metric learning loss function, the formula of which is as follows: ; in, It is the angle between the feature vector and the true class center; It is an additive angular interval (usually 0.5); Indicates the number of samples participating in the current loss calculation; This represents the scale factor (usually set to 64), used to amplify the cosine similarity. Indicates the first The feature vector of the i-th sample and the i-th sample The angle between class centers; a metric learning loss function based on angular intervals is used to force features to form a compact distribution on the hypersphere.

[0059] Understandably, the ArcFace loss function can force the angle difference between feature points of the same class to be less than or equal to a first value (i.e., ... (Very small), the angle difference between feature points in different classes is greater than the preset interval. Specifically, the intra-class angle difference is compressed to near zero, and the inter-class angle difference is at least... This ensures that features are distributed on the hypersphere, resulting in intra-class clustering and inter-class separation.

[0060] Optionally, the ArcFace loss function is one of the effective means to achieve distribution shaping. Of course, any loss function based on cosine similarity and introducing boundaries to compress intra-class distances and separate inter-class distances can be used as the loss function in this application's embodiments. For example, the CosFace loss function (applying an additive margin in cosine space), the SphereFace real function (applying a multiplicative margin in angular space), or the more unified Circle Loss, etc. The common essence of these loss functions is to force similar features to cluster on a hypersphere and separate dissimilar features, thereby providing a mathematical premise for subsequent distribution reconstruction based on the Gaussian assumption.

[0061] Step 104: When new category radar data is received, perform incremental learning without a paradigm set.

[0062] Specifically, for the learned old categories, the low-dimensional statistics of their features are stored, and the feature distribution of the old categories is reconstructed based on the low-dimensional statistics, and pseudo-feature samples are generated.

[0063] Understandably, step 104 is the core of incremental learning, avoiding the storage of raw data through statistical storage and generative replay.

[0064] After each round of training, freeze the feature extractor and compute the old class. Low-dimensional statistics: In one possible implementation, the low-dimensional statistics for the old categories include the class prototype (mean). and characteristic covariance matrix It can be calculated using the following formula: Class prototype (mean): ; Feature covariance matrix ; in, Indicates belonging to a category The number of samples; Indicate category Index of internal samples; It is a category The Middle The original input data for each training sample; Since the feature distribution in step 103 above has been shaped into a Gaussian-like distribution, pseudo-features are generated through multivariate Gaussian sampling: ; in, To prevent singularity in the covariance matrix, small perturbation terms are used to enhance clarity; continuous pseudo-features are generated by sampling to approximate the old class distribution.

[0065] Alternatively, although the embodiments of this application employ efficient Gaussian distribution sampling, to obtain higher quality pseudo-features, the pseudo-feature generation module can be replaced by training a Conditional Generative Adversarial Network (cGAN), a Variational Autoencoder, or a Diffusion Model. For example, in a cGAN, a generator is trained conditioned on the old class labels to synthesize pseudo-features, while a discriminator is trained to distinguish between real and synthesized features. Frozen old model feature extractors can be used as part of the discriminator or to provide discriminative features. As another example, in a Variational Autoencoder, a low-dimensional latent space of the old class feature distribution is learned, and pseudo-features are generated by sampling and decoding from this latent space. As yet another example, in a Diffusion Model, pseudo-features conforming to the old class distribution are generated from random noise through a progressive denoising process.

[0066] Step 105: Merge the radar data of the new category with the pseudo-feature samples of the old category to jointly train the deep learning model.

[0067] Understandably, a dual correction strategy (feature extractor correction and classifier correction) is used to balance the learning of new and old classes and prevent catastrophic forgetting.

[0068] During training, knowledge distillation (i.e., feature extractor correction) constraints are used to stabilize the feature extraction capability for old categories, and classifier weights are aligned (i.e., classifier correction) to eliminate classification bias between new and old categories.

[0069] In one possible embodiment, based on knowledge distillation constraints to stabilize the feature extraction capability for old categories, and performing an alignment operation on classifier weights to eliminate classification bias between new and old categories, the method includes: Using the old model as the teacher network model, the network's feature output is extracted through knowledge distillation to constrain new features; where, the knowledge distillation loss function... It is expressed as follows: ; in, This represents the new feature extraction network to be trained in the current incremental phase; This represents the old feature extraction network whose parameters have been frozen in the previous stage; To eliminate the classification bias between the old and new classes, the weights of the new classifier are aligned. The alignment formula is as follows: ; in, Indicates the weights of the old classifier. Indicates the weights of the new classifier. This represents the arithmetic mean; the weight alignment operation is used to eliminate classification bias caused by sample imbalance.

[0070] Step 106: Update the parameters of the deep learning model and the low-dimensional statistics library of the old categories to complete the current incremental learning phase.

[0071] It is understandable that, upon completing the current phase ( After self-training and double calibration, the algorithm needs to update its state in order to prepare for the next stage. Be prepared. Specifically: In one possible implementation, updating the parameters of the deep learning model and the low-dimensional statistics library of the old categories to complete the current incremental learning phase includes: Classifier integration: The corrected new class weights are combined with the old class weights to form the complete classifier for the current stage.

[0072] For example, the corrected new class classification head weights Weight of old category header Merge to form the complete classifier weights for the current stage. At this point, the model has the ability to identify all the categories that have appeared.

[0073] Prototype library update: Recalculate the feature statistics of the new category and add them to the prototype library, while supporting the option to choose momentum update or keep the initial state for the feature statistics of the old category.

[0074] For example, using the updated feature extractor Recalculate the prototype mean of the newly added category. Covariance Matrix .

[0075] Then, the statistics for the new class are added to the prototype library. .

[0076] It should be noted that the statistics of the old class can be left unupdated (to maintain the initial state and prevent drift accumulation) or updated with momentum (to integrate the features extracted by the current model and adapt to the fine-tuning of the feature space).

[0077] Teacher model replacement: The updated model parameters are frozen and used as the teacher network model for the next incremental stage, i.e. .

[0078] like Figure 2 As shown, Figure 2 This is a schematic diagram of a class incremental identification process in a reconnaissance system provided as an embodiment of this application. Figure 2 This paper describes the entire process from raw signal reception to the generation of advanced threat information (i.e., radar reconnaissance system processing), and integrates a kind of incremental identification process for radiation source signals. In the radar reconnaissance system processing flow, the signal receiver is responsible for intercepting radio frequency signals in space. Then, the pulse detection module initially detects and extracts the radar pulses. Subsequently, the detected pulses are converted from analog to digital by the intermediate frequency sampling signal module to obtain digital signals. Through the parameter measurement module, the descriptor parameters of each pulse, such as arrival time, carrier frequency, pulse width, and amplitude, are determined from the digital signal. The received and sorted individual radiation source pulse sequences are subjected to feature extraction and pattern matching to ultimately determine the information of the radiation source.

[0079] The incremental identification process for radiation source signals includes a training phase and a testing phase. During the training phase, multiple datasets arriving sequentially in task order are received (Task 1 data D1, Task 2 data D2, ..., Task n data Dn). This simulates a technology learning scenario where new radars emerge successively in a real-world context. The system employs an incremental learning strategy. Data from each task (e.g., Task 1 data D1) is used to train a corresponding knowledge module (e.g., Task 1 knowledge in Model 1). When new task data (e.g., Task 2 data D2) arrives, the system, while retaining the learned knowledge (Model 1), uses the new data to learn new knowledge and integrates it to form an updated "comprehensive model." This process iterates continuously, ultimately forming a comprehensive model capable of identifying all categories.

[0080] During the testing phase, a separate "test set" is used to validate the performance of the trained integrated model. The test set also contains samples from multiple tasks (Task 1 to Task n). The test data is input into the "integrated model," and the model outputs the recognition results. The model's generalization ability and incremental learning effectiveness are evaluated by analyzing metrics such as recognition accuracy and confusion matrix.

[0081] Steps 101 to 106 described in the above embodiments can be specifically combined with Figure 3 As shown, the method provided in this application is divided into an initial training phase and an incremental learning phase. In the initial training phase, basic category data is processed to generate image representations and train a feature extractor. Specifically, an "initial dataset" is used as input, containing radar emission source pulse sequences of the first known categories. The data enters the "ST-PCDM preprocessing" module to perform spatiotemporal-physical coupled dynamic density field mapping, converting the pulse descriptor sequences into image representations with explicit physical meaning to enhance the discriminative power and robustness of the features. The preprocessed data enters the basic training (ArcFace+CBAM) module. Here, the image representations are input into a network integrating a global gated attention module and trained using an angular interval-based metric learning loss function to obtain an initial feature extractor and classifier with compact intra-class and inter-class feature distribution capabilities, outputting an initial model. A prototype library is "extracted and saved" from the trained "initial model." This prototype library stores feature statistics (such as class prototypes and covariance matrices) for each known category for subsequent incremental learning without saving the original data.

[0082] During the incremental learning phase, for new category data, continuous learning without a paradigm set can be achieved through distribution reconstruction and dual calibration. Specifically, new radar data can be received and preprocessed using the ST-PCDM module to convert it into an image representation. In parallel, the system reads the old category statistics stored in the prototype library and, based on this, generates a pseudo-feature module through Gaussian sampling to reconstruct the feature distribution samples of the old categories. The true features of the new data and the pseudo-features of the old categories are merged into mixed data (new data + pseudo-features) and fed into the dual calibration training (knowledge distillation + weight alignment) module. After training, the system synchronously updates the model and the prototype library. The model is enhanced to recognize all new and old categories; the prototype library incorporates the statistics of the new categories, expanding its knowledge. If the preset incremental learning cycle has not been completed, or if new categories still appear, the system returns to load new radar data and begins the next cycle.

[0083] This application proposes a paradigm-free incremental identification method for radar radiation sources based on distribution reconstruction. First, it receives the pulse sequence of the radar radiation source. Then, based on the descriptor parameters of the pulse sequence, it preprocesses the pulse sequence to obtain preprocessed original parameters and dynamic features. The original parameters and dynamic features are then coupled and mapped using multiple views to generate an image-based signal representation. Next, a deep learning model is used to extract features from the image-based signal representation, and a metric learning loss function based on angular intervals is used for training to form a compact intra-class and segregated inter-class feature distribution on a hypersphere in the feature space. Finally, when radar data of a new category is received, paradigm-free incremental learning is performed. Next, the radar data of the new category is merged with the pseudo-feature samples of the old category to jointly train the deep learning model. Finally, the parameters of the deep learning model and the low-dimensional statistical library of the old category are updated to complete the current incremental learning stage. Among them, the feature distribution of compact intra-class and separated inter-class reflects that the angle difference of intra-class feature points is less than or equal to the first value, and the angle difference of inter-class feature points is greater than the preset interval. In addition, during training, the feature extraction capability of the old category can be stabilized based on knowledge distillation constraints, and the classifier weights can be aligned to eliminate the classification bias between the new and old categories. Therefore, this scheme can achieve high-performance continuous learning with a balance between the new and old categories while ensuring high recognition accuracy under harsh working conditions and with zero original data storage.

[0084] In summary, the paradigm-free incremental identification method for radar radiation sources based on distributed reconstruction provided by the embodiments of this application has significant application value in various scenarios requiring continuous learning and accurate identification of complex electromagnetic signals, mainly reflected in the following three aspects: First, it is applicable to drone swarms or spaceborne platforms, enabling continuous multi-task learning without storing sensitive raw waveform data, thus resolving the conflict between storage constraints and data security.

[0085] Secondly, it can be deployed in mobile monitoring stations or dedicated signal analysis equipment. Its core advantage lies in its ability to learn and identify newly emerging radio signal regimes within the working area in real time. The system does not require transmitting massive amounts of data back to the central processing node; it can update the identification database online at the edge, significantly shortening the response time from signal detection to identification and classification, and improving the entire monitoring system's adaptability to dynamically changing electromagnetic environments. Third, even in complex environments with strong interference and severe multipath propagation effects, it can still maintain high-precision identification and stable tracking of various non-cooperative radio signals (such as illegally set-up fake base stations and unauthorized black broadcasts).

[0086] The steps described above are for clarity only. In implementation, they can be combined into one step, or some steps can be broken down into multiple steps, as long as they involve the same logical relationship, they are all within the scope of protection of this application. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, without changing the core design of the algorithm and process, are also within the scope of protection of this application.

[0087] Another embodiment of this application proposes a paradigm-free incremental identification system for radar radiation sources based on distributed reconstruction. The details of this paradigm-free incremental identification system based on distributed reconstruction are described below. The following implementation details are provided for ease of understanding and are not essential for implementing this example. Figure 4 This is a schematic diagram of the structure of a paradigmless incremental identification system for radar radiation sources based on distributed reconstruction proposed in this embodiment, including: The receiving module 210 is used to receive the pulse sequence of the radar radiation source; The preprocessing module 220 is used to preprocess the pulse sequence of the radar radiation source based on the descriptor parameters of the pulse sequence, obtain the preprocessed original parameters and dynamic characteristics, and perform multi-view coupling mapping on the original parameters and dynamic characteristics to generate a graphical signal representation. The feature extraction module 230 is used to extract features from the image-based signal representation using a deep learning model, and to train it using a metric learning loss function based on angle intervals, so as to form a compact intra-class and separate inter-class feature distribution on the hypersphere of the feature space; the compact intra-class and separate inter-class feature distribution reflects that the angle difference of intra-class feature points is less than or equal to a first value, and the angle difference of inter-class feature points is greater than a preset interval. The incremental learning module 240 is used to perform paradigm-free incremental learning when receiving radar data of a new category; wherein, for the learned old category, the low-dimensional statistics of its features are stored, and the feature distribution of the old category is reconstructed based on the low-dimensional statistics, and pseudo-feature samples are generated. The joint training module 250 is used to merge radar data of the new category with pseudo-feature samples of the old category to jointly train the deep learning model. During the training process, knowledge distillation constraints are used to stabilize the feature extraction capability of the old category, and the classifier weights are aligned to eliminate the classification bias between the new and old categories. The model update module 260 is used to update the parameters of the deep learning model and the low-dimensional statistics library of the old categories to complete the current incremental learning phase.

[0088] It is not difficult to see that this embodiment is a system embodiment corresponding to the above method embodiments, and this embodiment can be implemented in conjunction with the above method embodiments. The relevant technical details and technical effects mentioned in the above method embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above method embodiments.

[0089] It is worth mentioning that all modules and units involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units do not exist in this embodiment.

[0090] Another embodiment of this application provides an electronic device, such as Figure 5 As shown, it includes a processor 31 and a memory 32. The memory 32 stores instructions that the processor 31 can execute. When the processor 31 is configured to execute the instructions, the electronic device can implement a paradigmless incremental identification method for radar radiation sources based on distributed reconstruction as described in the above method embodiment.

[0091] The memory and processor are connected via a bus, which includes any number of interconnecting buses and bridges, connecting various circuits of one or more processors and the memory. The bus can also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single component or multiple components, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0092] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0093] Another embodiment of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, can implement a paradigmless incremental identification method for radar radiation sources based on distributed reconstruction as described in the above method embodiments.

[0094] That is, those skilled in the art will understand that all or part of the steps in the above method embodiments can be implemented by a program instructing related hardware. The program is stored in a storage medium and includes several instructions to cause a device (such as a microcontroller, chip, etc.) or processor to execute all or part of the steps of the method described in the method embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0095] Those skilled in the art will understand that the above embodiments are specific implementations of this application, and in practical applications, various changes can be made in form and detail without departing from the spirit and scope of this application. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A method for incremental identification of radar radiation sources without a paradigm set based on distributed reconstruction, characterized in that, The method includes: Receive the pulse sequence from the radar radiation source; Based on the descriptor parameters of the pulse sequence, the pulse sequence of the radar radiation source is preprocessed to obtain the preprocessed original parameters and dynamic characteristics. The original parameters and dynamic characteristics are then coupled and mapped by multiple views to generate a visual signal representation. Using a deep learning model, features are extracted from the image-based signal representation, and a metric learning loss function based on angle interval is used for training to form a compact intra-class and separable inter-class feature distribution on the hypersphere of the feature space; wherein, the compact intra-class and separable inter-class feature distribution reflects that the angle difference of intra-class feature points is less than or equal to a first value, and the angle difference of inter-class feature points is greater than a preset interval. When new category radar data is received, incremental learning without a paradigm is performed; for the old categories that have been learned, the low-dimensional statistics of their features are stored, and the feature distribution of the old categories is reconstructed based on the low-dimensional statistics, and pseudo-feature samples are generated. The radar data of the new category is merged with the pseudo-feature samples of the old category to jointly train the deep learning model. During the training process, knowledge distillation constraints are used to stabilize the feature extraction capability of the old category, and the classifier weights are aligned to eliminate the classification bias between the new and old categories. Update the parameters of the deep learning model and the low-dimensional statistics library of the old categories to complete the current incremental learning phase.

2. The method according to claim 1, characterized in that, The pulse sequence of a radar radiation source includes: pulse arrival time, carrier frequency, pulse width, and pulse amplitude; The pulse sequence of a radar radiation source is represented by a pulse description word vector, which is represented as follows: ； in, The pulse arrival time, For carrier frequency, The pulse width. The amplitude of the pulse. An index representing the pulse sequence of a radar radiation source; The descriptive word parameters based on the pulse sequence preprocess the pulse sequence of the radar radiation source to obtain the preprocessed original parameters and dynamic characteristics. The original parameters and dynamic characteristics are then coupled and mapped using a multi-view method to generate a graphical signal representation, including: Perform differential operations on the descriptor parameters of the pulse sequence to extract at least first-order differential features; wherein, the differential features include the carrier frequency change rate. and the rate of change of repetition frequency The calculation formula is as follows: ； ； ； in, Indicates the frequency of repetition; The extracted raw parameters and dynamic difference features are normalized and mapped to a preset numerical range, as shown in the following formula: The formula for nonlinearly normalizing the carrier frequency change rate is as follows: ； in, This represents the rate of change of the carrier frequency after nonlinear compression processing; This represents the scaling factor, used to adjust the slope of the hyperbolic tangent function during nonlinear compression, so as to map the carrier frequency change rate to a preset numerical range (0,1). Physically correlated parameters or feature combinations are mapped to different two-dimensional view planes, and pixel-level density accumulation is performed on each view plane to generate multi-channel feature images.

3. The method according to claim 2, characterized in that, The multi-channel feature image is a three-channel RGB image; the multi-channel feature image enhances contrast through logarithmic transformation and connects broken trajectories caused by pulse loss through Gaussian smoothing; The three-channel RGB includes: The R channel is generated by the joint distribution mapping of the carrier frequency and the repetition frequency; The G channel is generated by the joint distribution mapping of the carrier frequency and the carrier frequency change rate; Channel B is generated by the joint distribution mapping of repetition rate and pulse width; The formula for pixel-level density accumulation is as follows: ； in, These are pixel coordinates; Indicates the first The x-coordinate of the two-dimensional coordinates of each pulse sample point projected onto the view plane. Indicates the first The ordinate of the two-dimensional coordinates of each pulse sample point after it has been projected onto the view plane; right Perform a logarithmic transformation to obtain the transformed features To enhance contrast; the formula for the logarithmic transformation is as follows: 。 4. The method according to claim 3, characterized in that, The deep learning model uses a residual network as the backbone network and embeds a global gating attention module; The process of extracting features from image-based signal representations using a deep learning model includes: By utilizing a global gated attention module, adaptive weighting of feature channels and spatial regions is achieved to suppress noise and enhance the focusing of effective signal trajectories; The global gating attention module includes channel attention. and spatial attention Features are enhanced using the following formula: ； in, It is the original extracted feature map tensor; It is the weighted enhanced feature map tensor of F after applying channel attention and spatial attention; This indicates channel attention.

5. The method according to claim 4, characterized in that, The training process employs a metric learning loss function based on angular intervals to form a compact intra-class and segregated inter-class feature distribution on a hypersphere in the feature space, including: eigenvectors and classifier weights Normalized to the hypersphere, the formula is as follows: ； ； The feature space is compressed using an angle-margin-based metric learning loss function, the formula of which is as follows: ； in, It is the angle between the feature vector and the true class center; It is an additive angular interval; Indicates the number of samples participating in the current loss calculation; It represents the scaling factor, used to amplify the cosine similarity; Indicates the first The feature vector of the i-th sample and the i-th sample The angle between class centers; a metric learning loss function based on angular intervals is used to force features to form a compact distribution on the hypersphere.

6. The method according to claim 5, characterized in that, Low-dimensional statistics for old categories include class prototype mean. Covariance Matrix It can be calculated using the following formula: ； ； in, Indicates belonging to a category The number of samples; Indicate category Index of internal samples; It is a category The Middle The original input data for each training sample; The generation of pseudo-feature vectors is achieved through multivariate Gaussian sampling: ； in, To prevent singularity in the covariance matrix, small perturbation terms are used to enhance clarity; continuous pseudo-features are generated by sampling to approximate the old class distribution.

7. The method according to claim 6, characterized in that, The method of stabilizing feature extraction capability for old categories based on knowledge distillation constraints and aligning classifier weights to eliminate classification bias between old and new categories includes: Using the old model as the teacher network model, the network's feature output is extracted through knowledge distillation to constrain new features; where, the knowledge distillation loss function... It is expressed as follows: ； in, This represents the new feature extraction network to be trained in the current incremental phase; This represents the old feature extraction network whose parameters have been frozen in the previous stage; To eliminate the classification bias between the old and new classes, the weights of the new classifier are aligned. The alignment formula is as follows: ； in, Indicates the weights of the old classifier. Indicates the weights of the new classifier. This represents the arithmetic mean; the weight alignment operation is used to eliminate classification bias caused by sample imbalance.

8. The method according to claim 7, characterized in that, The process of updating the parameters of the deep learning model and the low-dimensional statistics library of the old categories to complete the current incremental learning phase includes: The corrected new class weights are combined with the old class weights to form the complete classifier for the current stage. Recalculate the feature statistics for the new category and add them to the prototype library, while also supporting the option to either update the feature statistics of the old category with momentum or keep them in their initial state. The updated model parameters are frozen and used as the teacher network model for the next incremental stage.

9. An electronic device, characterized in that, include: A processor and a memory, wherein the memory stores instructions executable by the processor, and the processor is configured to, when executing the instructions, enable the electronic device to implement the paradigmless incremental identification method for radar radiation sources based on distributed reconstruction as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it can implement the paradigmless incremental identification method for radar radiation sources based on distributed reconstruction as described in any one of claims 1 to 8.

Images