A high-explainability radiation source identification method and system

Abstract

The application relates to the field of radiation source identification, and specifically provides a high-explainability radiation source identification method and system, a training set radar pulse signal is acquired, a multi-domain joint time-varying fingerprint feature is extracted based on physical layer modeling, the fingerprint feature is input into a time-frequency double-branch feature fusion network to obtain time-frequency fusion features; the time-frequency fusion features are subjected to signal-level physical consistency enhancement and feature-level mixed enhancement; based on the enhanced features, an end-to-end training is conducted on the time-frequency double-branch feature fusion network through a hierarchical contrast learning framework containing intra-class contrast and inter-class prototype contrast, a radiation source identification model is obtained, and the trained radiation source identification model is used to identify the individual class of a radar pulse signal to be identified. Through the introduction of the physical layer explainable time-varying fingerprint feature, the adaptive double enhancement mechanism and the hierarchical contrast learning framework, the application realizes high-credibility, strong-adaptation and stable-identification radiation source individual identification.

Description

Technical Field

[0001] This invention relates to the field of radiation source identification, and specifically to a highly interpretable radiation source identification method and system. Background Technology

[0002] With the rapid development of technologies such as communication and radar, the electromagnetic space has become a core carrier of modern information society. From 5G communication to the Internet of Things, the proliferation of various radiation source devices has made the electromagnetic environment increasingly complex, and the competition for spectrum resources and security threats have become more prominent. Against this backdrop, Specific Emitter Identification (SEI) technology has become a research hotspot in the fields of communication and radar signal processing due to its key role in device authentication, spectrum management, and electronic warfare.

[0003] In recent years, deep learning technology has demonstrated tremendous potential in the radar SEI field due to its powerful feature self-learning capabilities. Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and even Transformer architectures have been widely applied, significantly improving recognition accuracy. However, the success of deep learning models heavily relies on the support of large-scale, high-quality labeled datasets. In practical deployment scenarios, this prerequisite is often difficult to meet. First, acquiring a large number of measured signals for a specific radar individual is costly and time-consuming. Second, publicly available datasets are limited in size and have incomplete category coverage. Third, modern radar systems are rapidly updated, with new models constantly emerging, making it difficult to directly transfer models trained on older models to new individuals, while collecting a large number of training samples for each new individual is impractical. Therefore, SEI systems must have the ability to quickly adapt and accurately identify new individuals with only a very small number of samples—this is precisely the problem that Few-Shot Specific Emitter Identification (FS-SEI) aims to solve.

[0004] In recent years, a large number of innovative methods based on technical paths such as metric learning, data augmentation, self-supervised learning, and knowledge fusion have emerged in FS-SEI research. These include methods based on metric learning and prototype networks, methods based on data augmentation and synthesis strategies, methods based on self-supervised and contrastive learning, methods based on multimodal or multi-domain feature fusion, and methods based on model structure innovation and optimization. Although the above five types of methods have made progress in specific directions, they still have several common shortcomings: (1) Poor model interpretability: For example, although structures such as NAS, GAN, and Transformer are effective, they rely solely on deep learning models, and the internal feature extraction situation is unclear, resulting in poor model interpretability; (2) Cross-scene generalization ability problem: Some works only verify the algorithm performance in a single environment, and the generalization ability of the algorithm in different environments is unknown; (3) Feature extraction robustness problem: Some works do not specifically extract features based on the characteristics of radar or communication radiation source signals, but simply rely on deep learning methods to extract features, and the effectiveness of feature extraction needs to be improved. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a highly interpretable radiation source identification method and system. By introducing physically interpretable time-varying fingerprint features, an adaptive dual enhancement mechanism, and a hierarchical contrastive learning framework, it achieves highly reliable, highly adaptive, and stable identification of individual radiation sources.

[0006] In a first aspect, the technical solution of the present invention provides a highly interpretable radiation source identification method, comprising the following steps: The training set of radar pulse signals is acquired, and multi-domain joint time-varying fingerprint features are extracted based on physical layer modeling. The time-varying fingerprint features include instantaneous frequency deviation, instantaneous amplitude deviation, and phase difference sequence. Multi-domain joint time-varying fingerprint features are input into a time-frequency dual-branch feature fusion network, and time-frequency fusion features are obtained through an attention mechanism; Adaptive enhancement of time-frequency fusion features is performed, including signal-level physical consistency enhancement and feature-level hybrid enhancement. Based on the enhanced features, the time-frequency dual-branch feature fusion network is trained end-to-end through a hierarchical contrastive learning framework that includes intra-class contrast and inter-class prototype contrast, to obtain the radiation source identification model. For the radar pulse signal to be identified, its multi-domain joint time-varying fingerprint features are extracted, and the features are input into the trained radiation source identification model to output the individual category of the identified radiation source.

[0007] Secondly, the technical solution of the present invention provides a highly interpretable radiation source identification system, comprising: The training signal time-varying feature extraction unit is used to acquire radar pulse signals in the training set and extract multi-domain joint time-varying fingerprint features based on physical layer modeling. The time-varying fingerprint features include instantaneous frequency deviation, instantaneous amplitude deviation, and phase difference sequence. The time-frequency feature fusion unit is used to input multi-domain joint time-varying fingerprint features into the time-frequency dual-branch feature fusion network, and obtain time-frequency fused features through an attention mechanism; The recognition model training unit is used to adaptively enhance the time-frequency fusion features, including signal-level physical consistency enhancement and feature-level hybrid enhancement. Based on the enhanced features, the time-frequency dual-branch feature fusion network is trained end-to-end through a hierarchical contrastive learning framework that includes intra-class contrast and inter-class prototype contrast to obtain the radiation source recognition model. The radiation source identification unit is used to extract the multi-domain joint time-varying fingerprint features of the radar pulse signal to be identified, input the features into the trained radiation source identification model, and output the individual category of the identified radiation source.

[0008] As can be seen from the above technical solutions, this application has the following advantages: By starting from the physical layer hardware defect mechanism, a set of joint time-varying fingerprint features with clear physical meaning—instantaneous frequency deviation, instantaneous amplitude deviation, and phase difference sequence—is constructed and extracted. This transforms the input of deep learning models from incomprehensible raw signals or black-box features into physical quantities that directly map core hardware defects such as power amplifier nonlinearity and phase noise. This overcomes the shortcomings of traditional deep learning methods that rely on data-driven approaches and have vague decision-making criteria. By establishing the identification process on a traceable and analyzable physical foundation, the transparency and credibility of model decisions are greatly enhanced, making it more suitable for fields with high interpretability requirements. By employing a dual strategy of signal-level physical consistency enhancement and feature-level hybrid enhancement, the system simulates interference such as phase noise in real channels, enabling the model to have prior adaptability to signal distortion. On the other hand, it effectively expands the feature space with difficult samples, forcing the model to learn clearer decision boundaries. This significantly improves the robustness of the feature extraction network to noise interference and fluctuations in the number of samples, ensuring that it can still stably extract discriminative key features even in harsh electromagnetic environments or when there are only a very small number of samples. By constructing a hierarchical contrastive learning framework encompassing intra-pulse (intra-class) and inter-pulse (inter-class) aspects, the structure of the feature space is systematically optimized. Intra-class contrastive learning focuses the model on stable, essential fingerprint features within individual radiation sources, eliminating reliance on irrelevant changes introduced by data augmentation, thus learning robust representations invariant to scene variations. Inter-class prototypical contrastive learning explicitly widens the spatial distance between features of different categories, forming a compact and well-separated feature distribution. The synergistic effect of these two approaches enables the trained model to accurately generalize to new scenes and environments not encountered during training, leveraging its learned invariance and strong discriminative features. This significantly enhances the practicality and adaptability of the method in complex real-world electromagnetic environments. Attached Figure Description

[0009] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below 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.

[0010] Figure 1 This is a schematic flowchart of a highly interpretable radiation source identification method provided in an embodiment of the present invention.

[0011] Figure 2 This is a graph showing the results of the ablation experiment.

[0012] Figure 3 This is a comparison chart of algorithm performance under different small sample conditions.

[0013] Figure 4 This is a comparison chart of algorithm performance under different signal-to-noise ratio conditions.

[0014] Figure 5 This is a schematic block diagram of a highly interpretable radiation source identification system provided in an embodiment of the present invention. Detailed Implementation

[0015] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0017] This invention provides a highly interpretable radiation source identification method, namely, a radar radiation source individual identification method based on multi-domain joint time-varying fingerprint features and hierarchical contrastive learning (JTVF-HCL). The architecture of this method includes four main modules: (1) a multi-domain joint time-varying fingerprint feature extraction module: modeling and extracting the essential hardware fingerprint of the radar transmitter from the physical level; (2) a dual-branch time-frequency feature fusion network, capturing complementary representations of the fingerprint from the time domain and the time-frequency domain respectively; (3) an adaptive feature enhancement module, improving the robustness of the model through feature-level and signal-level enhancement strategies; and (4) a hierarchical contrastive learning module, learning highly discriminative feature representations at both intra-pulse and inter-pulse levels. Finally, the network is optimized end-to-end through a joint loss function.

[0018] Based on the above architecture Figure 1 This is a schematic flowchart illustrating a highly interpretable radiation source identification method provided in an embodiment of the present invention. Figure 1 The executing entity can be a highly interpretable radiation source identification system. The highly interpretable radiation source identification method provided in this embodiment is executed by a computer device; correspondingly, the highly interpretable radiation source identification system runs on the computer device. Depending on different requirements, the order of the steps in this flowchart can be changed, and some steps can be omitted.

[0019] like Figure 1 As shown, the method includes the following steps.

[0020] S1, acquire the radar pulse signal of the training set, and extract multi-domain joint time-varying fingerprint features based on physical layer modeling. The time-varying fingerprint features include instantaneous frequency deviation, instantaneous amplitude deviation, and phase difference sequence.

[0021] S2 inputs the multi-domain joint time-varying fingerprint features into the time-frequency dual-branch feature fusion network, and obtains the time-frequency fusion features through the attention mechanism.

[0022] S3 adaptively enhances the time-frequency fusion features, including signal-level physical consistency enhancement and feature-level hybrid enhancement. Based on the enhanced features, the time-frequency dual-branch feature fusion network is trained end-to-end through a hierarchical contrastive learning framework that includes intra-class contrast and inter-class prototype contrast, to obtain the radiation source identification model.

[0023] S4. For the radar pulse signal to be identified, extract its multi-domain joint time-varying fingerprint features, input the features into the trained radiation source identification model, and output the identified individual radiation source category.

[0024] As a refinement and extension of the specific implementation of the above embodiments, in order to fully explain the specific implementation process of this embodiment, the following will provide possible embodiments to describe the specific implementation of the above steps in a non-limiting manner.

[0025] Hardware defects in radar transmitters, such as the nonlinearity of the power amplifier, the phase noise of the local oscillator, and the amplitude-phase imbalance of the modulator, introduce unique, time-varying fingerprint characteristics into the transmitted signal. To accurately characterize this physical phenomenon, the received radar pulse signal is modeled as follows: (1) in, For instantaneous envelope, For ideal linear frequency modulation phase ( For carrier frequency, (frequency modulation slope) and These represent phase noise and additive white Gaussian noise, respectively. The core of the SEI task lies in individual fingerprints. That is, the phase model, which is modeled as the superposition of a polynomial reflecting the nonlinearity of the power amplifier and a sine function reflecting periodic disturbances: (2) in, For the first The coefficients of the first-order polynomial reflect the nonlinearity of the power amplifier. The periodic perturbation parameter reflects the effects of power supply ripple, etc. K is the highest order of the polynomial components, and M is the number of periodic perturbation components. This model can simultaneously characterize the power amplifier nonlinearity (composed of polynomial...). (Description) and periodic disturbances such as power supply ripple (described by the sine term).

[0026] This model can simultaneously characterize the power amplifier nonlinearity (derived from the polynomial). The model includes periodic disturbances such as power supply ripple (described by a sinusoidal term) and power supply ripple. Based on this model, we define and extract three key instantaneous fingerprint features: instantaneous frequency deviation, instantaneous frequency offset, and periodic disturbances such as power supply ripple (described by a sinusoidal term). Instantaneous Frequency Deviation (IFD) (Instantaneous Amplitude Deviation, IAD) and phase difference sequences (Phase Difference Sequence, PDS), which are defined as follows: (3) (4) (5) in, For an ideal constant envelope amplitude, choose... and As core fingerprint features, their fundamental reason is that they can directly and sensitively map to the inherent physical defects of radar transmitter hardware and have excellent time-varying characterization capabilities. Defined as the instantaneous derivative of the individual fingerprint phase. It directly reflects the short-term stability of the local oscillator phase noise, the time-domain characteristics of the modulator clock jitter, and the modulation effect of the power amplifier nonlinearity on the signal phase. Since frequency is the rate of change of phase, it has extremely high sensitivity to dynamic and weak phase fluctuations caused by hardware defects, and can capture transient details that are ignored by traditional steady-state characteristics (such as carrier frequency deviation). This directly measures the distortion of the signal envelope relative to the ideal state, which is mainly traced back to the nonlinear amplitude response of the power amplifier and the power supply modulation effect. The gain compression characteristics of the power amplifier under different input powers are highly device-dependent, and this nonlinearity directly imprints a unique amplitude fingerprint on the time-domain envelope of the transmitted signal. By characterizing the minute phase changes between adjacent sampling points, it effectively captures transient phase disturbances introduced by hardware, such as local oscillator phase noise and clock jitter, and has high sensitivity and stability in distinguishing individual radiation sources.

[0027] To fully utilize fingerprint information, a dual-branch feature fusion network was designed. The time-frequency dual-branch feature fusion network includes: The temporal branch inputs the multi-domain joint time-varying fingerprint feature sequence into a one-dimensional convolutional neural network to extract temporal features. The frequency domain branch extracts frequency domain features by performing a short-time Fourier transform on the original radar pulse signal and then using a two-dimensional convolutional neural network. The fusion module uses time-domain features as queries and frequency-domain features as keys and values, and performs weighted fusion through a scaling dot product attention mechanism to obtain the time-frequency fusion features.

[0028] Specifically, the temporal branch inputs the extracted instantaneous feature sequence into a one-dimensional convolutional neural network to learn its temporal local dependencies and obtain temporal features. As shown in the following formula: (6) The frequency domain branch performs a short-time Fourier transform on the original signal to obtain a time-frequency diagram. : (7) Then, frequency domain features are extracted using a lightweight 2D convolutional network. This feature has a stronger ability to capture subtle differences in the frequency dimension. Finally, the time-domain and frequency-domain features are fused. To achieve effective information complementarity, an attention fusion module is introduced. Using the time-domain feature as the query and the frequency-domain feature as the key and value, the contribution weight of the frequency-domain feature to the time-domain feature is calculated through a scaled dot product attention mechanism, and then weighted and fused. (8) in, They are respectively composed of and Through linear projection, this fusion strategy of "time domain as the main focus and frequency domain as the auxiliary focus" enables the network to adaptively focus on the most discriminative time-frequency region.

[0029] To address the issues of sample scarcity and signal-to-noise ratio fluctuations in real-world scenarios, an adaptive feature enhancement module was designed. This module comprises two levels: signal-level physical consistency enhancement and feature-level hybrid enhancement.

[0030] Signal-level physical consistency enhancement is achieved by applying controllable phase modulation based on a phase noise model to the original radar pulse signal. Specifically, to simulate real phase noise disturbances, controllable phase modulation is applied to the original signal: (9) in, For modeling phase noise, This is the noise intensity coefficient.

[0031] Feature-level fusion enhancement refers to generating enhanced features by linearly interpolating the time-frequency fusion features of two different samples in the feature space, where the interpolation coefficients follow a Beta distribution. Specifically, linear interpolation is performed on the features of the two samples in the feature space: (10) in, This can efficiently generate difficult samples that fall between the two classes, improving the clarity of the decision boundary.

[0032] To learn more compact intra-class features and more separated inter-class features, this embodiment proposes a two-level contrastive learning framework, which includes hierarchical contrastive learning of intra-class contrast and inter-class prototype contrast.

[0033] Intra-class comparison involves constructing different enhanced versions of the same radar pulse signal as positive sample pairs, and optimizing them using an intra-class feature comparison loss function, which is an InfoNCE-based loss function. Specifically, different enhancements (such as cropping or adding noise) are applied to the same class of pulse signals to construct positive sample pairs. The InfoNCE loss function is then used for optimization. This forces the network to ignore the changes brought about by the enhancement and focus on the stable fingerprint features within the pulse, thereby improving the robustness of the features. Loss function Defined as: (11) in, and These represent the characteristics of different enhanced versions of the same pulse. Different enhanced versions refer to both signal-level physical enhancement and feature-level hybrid enhancement. For cosine similarity, For temperature hyperparameters, The batch size, This is an indicator function. The function value is 1 when the condition is true, and 0 otherwise. In other words, this term is only included in the summation when j is not equal to i.

[0034] Inter-class prototype comparison includes: calculating the mean of the time-frequency fusion features of all samples for each radiation source category to obtain the prototype vector of that category; and optimizing using an inter-class prototype comparison loss function, which is used to drive the time-frequency fusion features of samples to be closer to the prototype vector of their own category and further away from the prototype vectors of other categories.

[0035] First, calculate a prototype vector for each category, specifically the prototype vector for the k-th class. for: (12) in, For the first A sample set of classes For the first Time-frequency fusion features of each sample. Prototype contrastive loss. This encourages sample features to be close to the prototype of their own class and far away from the prototypes of other classes. This effectively utilizes the distribution information of all samples, narrowing intra-class distances and widening inter-class distances. Prototype contrastive loss. Defined as: (13) in, For the true label of sample i, The total number of categories, This is the cosine similarity function.

[0036] In this embodiment, the JTVF-HCL algorithm uses a joint loss function during end-to-end training, which is composed of weighted sums of multiple sub-items, as follows:

[0037] in, Cross-entropy classification loss is used to ensure that the model has basic classification capabilities; For intra-class feature contrast loss, For inter-class prototype comparison loss, and Used to improve feature quality; This is a weight regularization term to prevent overfitting; To balance hyperparameters, one specific implementation is as follows: , , Through joint optimization, the model can simultaneously achieve accurate classification and high-quality feature learning.

[0038] After the model has been trained, it can be deployed in a real-world system for identifying individuals from unknown radiation sources. The specific process for this identification phase is as follows: Receive the radar pulse signal to be identified, and calculate the instantaneous frequency deviation, instantaneous amplitude deviation and phase difference sequence of the signal according to the physical layer modeling and processing process that is completely consistent with the training phase, to form the multi-domain joint time-varying fingerprint feature vector of the signal. The time-varying fingerprint feature vector obtained in the previous step is input into the pre-trained time-frequency dual-branch feature fusion network. This network automatically performs time-domain convolution and frequency-domain transformation, and outputs a highly abstract fusion feature through its internal fixed attention fusion module. The fused features are input into the classification head (i.e., output layer) of the trained radiation source identification model. The classification head calculates the probability that the fused features belong to each known radiation source category based on the category prototype or decision boundary learned by the model during the training phase. The model selects the category with the highest probability as the recognition result and outputs the corresponding individual category identifier of the radiation source to complete the recognition task.

[0039] The experimental results and analysis are presented below.

[0040] To closely resemble real-world applications and ensure experimental reproducibility, a semi-physical simulation dataset was constructed. The signal acquisition system is based on the Ettus Research USRP X310 software-defined radio, coupled with the UBX-160 RF daughterboard and the OctoClock-G clock synchronization module. The specific hardware configuration is shown in Table 1. Considering that individual samples of real radar radiation sources are subject to confidentiality restrictions and are difficult to obtain on a large scale, a hybrid method of "high-fidelity radiation source modeling + USRP channel impairment injection" was adopted to generate the dataset.

[0041] Table 1: Hardware Configuration Parameters

[0042] At the radiation source end, five LFM radar transmitters with unique hardware fingerprints (including individual characteristics such as power amplifier nonlinearity and phase noise) were simulated using a physical model, with parameters set at a carrier frequency of 10 GHz, bandwidth of 50 MHz, and pulse width of 10 μs. At the receiver end, the simulated signals were replayed through the DAC module of a USRP X310 and then reacquired through the ADC module of the same USRP. To realistically simulate different environmental conditions, data was acquired in a shielded anechoic chamber environment, and a controllable AWGN (Additive White Gaussian Noise, AWGN) was superimposed to simulate different signal-to-noise ratio (SNR) conditions, with the SNR range set to -5 to 25 dB (achieved by adjusting the USRP gain and noise power). All acquired complex baseband signals were preprocessed and input into a multi-domain joint feature extraction module to generate time-varying fingerprint features for subsequent small-sample recognition tasks. The ratio of training set, validation set, and test set was 7:2:1.

[0043] To demonstrate the superiority of the proposed algorithm over existing algorithms, comparative experiments were conducted with a signal-to-noise ratio of 11 dB. The selected comparative experimental subjects were typical algorithms from the five categories of existing algorithms in the relevant work. The recognition accuracy of the algorithm under different small sample conditions was tested, as shown in Table 2.

[0044] Table 2: Comparative Experimental Results

[0045] Analysis of Table 2 reveals that the JTVF-HCL algorithm exhibits multi-dimensional advantages over existing methods. Firstly, at the feature representation level, the triple time-varying fingerprint feature based on physical layer modeling can simultaneously capture the hardware-essential characteristics of power amplifier nonlinearity and phase noise. Compared to existing deep metric learning models and self-supervised clustering methods, it offers stronger physical interpretability and greater adaptability to channel distortion. Secondly, in the feature fusion architecture design, the time-frequency dual-branch network based on an attention mechanism significantly improves feature discriminativity through cross-domain complementary information aggregation. Compared to multi-view attention fusion methods, its network structure better reflects the time-varying characteristics of the radiation source signal. Furthermore, the proposed dual enhancement strategy (feature-level mixing + signal-level physical consistency enhancement) for the small sample size challenge, compared to fuzzy oversampling methods, increases data diversity while maintaining physical constraints, effectively alleviating the overfitting problem caused by sample scarcity. Finally, the hierarchical contrastive learning module significantly improves the separability of the feature space distribution through the synergistic effect of prototype contrast and feature contrast, achieving an average improvement of approximately 3% in various small sample scenarios compared to single-layer optimization methods based on phase decoupling. Comparison with existing algorithms fully demonstrates the advantages of JTVF-HCL in processing radar radiation source data.

[0046] The JTVF-HCL algorithm mainly consists of four core modules: a multi-domain joint time-varying fingerprint feature extraction module, a dual-branch time-frequency feature fusion network, an adaptive feature enhancement module, and a hierarchical contrastive learning module. To verify the contribution of each module to the algorithm's performance, ablation experiments were conducted on each module. Figure 2 As shown in the figure, condition 1 represents the removal of the multi-domain joint time-varying fingerprint feature extraction module, condition 2 represents the removal of the dual-branch time-frequency feature fusion network, condition 3 represents the removal of the dual-branch time-frequency feature fusion network, and condition 4 represents the removal of the hierarchical contrastive learning module. It can be seen that after removing each module, the JTVF-HCL algorithm exhibits varying degrees of performance degradation, decreasing by 20.4%, 8.3%, 6.8%, and 36.7% respectively, demonstrating the effectiveness of the design of each sub-module.

[0047] To evaluate the generalization ability of the proposed method under small sample conditions, the number of samples for each type of radar radiation source in the training set was systematically adjusted (1, 3, 5, 7, 9, 11, 13, and 15 samples per type) under a fixed test set, and the recognition accuracy of the model under each setting was recorded. Figure 3 As shown, experimental results demonstrate that even with extremely low sample sizes (e.g., only 7 samples per class), the proposed algorithm maintains high recognition accuracy, significantly outperforming traditional deep learning baseline methods. This verifies that the proposed method possesses good generalization ability and sample utilization efficiency in data-scarce scenarios, making it suitable for SEI tasks where obtaining a large number of labeled samples is difficult in practice.

[0048] To further examine the algorithm's adaptability in complex electromagnetic environments, different signal-to-noise ratio conditions (ranging from...) were introduced during the testing phase. (5dB to 25dB) to simulate the degree of signal interference in a real environment. For example... Figure 4 As shown in the figure, the red dashed line at the bottom represents the random guessing line, and the green dashed line at the top represents the 90% threshold line. Experimental results show that in extremely low signal-to-noise ratio (SNR) environments, the algorithm's recognition rate is approximately 20% (random guessing results for a 5-class classification task). However, the proposed method significantly improves its recognition rate under conditions greater than 3 dB, achieving an accuracy of approximately 70% under 7 dB environmental conditions. This result demonstrates that the designed algorithm exhibits strong generalization ability to noise and can effectively support radar radiation source identification applications in harsh electromagnetic environments.

[0049] The above text describes in detail an embodiment of a highly interpretable radiation source identification method. Based on the highly interpretable radiation source identification method described in the above embodiment, this invention also provides a highly interpretable radiation source identification system corresponding to the method.

[0050] Figure 5 This is a schematic block diagram of a highly interpretable radiation source identification system provided in an embodiment of the present invention. In this embodiment, the highly interpretable radiation source identification system 500 can be divided into multiple functional units according to the functions it performs. The unit referred to in this invention is a series of computer program segments that can be executed by at least one processor and can perform a fixed function, and which are stored in memory.

[0051] The training signal time-varying feature extraction unit 510 is used to acquire the radar pulse signal of the training set and extract multi-domain joint time-varying fingerprint features based on physical layer modeling. The time-varying fingerprint features include instantaneous frequency deviation, instantaneous amplitude deviation, and phase difference sequence.

[0052] The time-frequency feature fusion unit 520 is used to input multi-domain joint time-varying fingerprint features into the time-frequency dual-branch feature fusion network and obtain time-frequency fused features through an attention mechanism.

[0053] The recognition model training unit 530 is used to adaptively enhance the time-frequency fusion features, including signal-level physical consistency enhancement and feature-level hybrid enhancement. Based on the enhanced features, the time-frequency dual-branch feature fusion network is trained end-to-end through a hierarchical contrastive learning framework that includes intra-class contrast and inter-class prototype contrast to obtain the radiation source recognition model.

[0054] The radiation source identification unit 540 is used to extract the multi-domain joint time-varying fingerprint features of the radar pulse signal to be identified, input the features into the trained radiation source identification model, and output the individual category of the identified radiation source.

[0055] The highly interpretable radiation source identification system of this embodiment is used to implement the aforementioned highly interpretable radiation source identification method. Therefore, the specific implementation of this system can be found in the embodiment section of the highly interpretable radiation source identification method above. Thus, its specific implementation can be referred to the description of the corresponding embodiments, and will not be elaborated here.

[0056] Furthermore, since the highly interpretable radiation source identification system in this embodiment is used to implement the aforementioned highly interpretable radiation source identification method, its function corresponds to that of the method described above, and will not be repeated here.

[0057] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A highly interpretable method for identifying radiation sources, characterized in that, Includes the following steps: The training set of radar pulse signals is acquired, and multi-domain joint time-varying fingerprint features are extracted based on physical layer modeling. These time-varying fingerprint features include instantaneous frequency deviation, instantaneous amplitude deviation, and phase difference sequence. Specifically, the instantaneous frequency deviation is calculated based on a phase model characterizing the individual hardware fingerprint of the radiation source. ,in, For instantaneous frequency deviation, The phase model is modeled as a superposition of a polynomial reflecting the nonlinearity of the power amplifier and a sine function reflecting periodic disturbances. The multi-domain joint time-varying fingerprint features and the original radar pulse signal are input into a time-frequency dual-branch feature fusion network, and the time-frequency fusion features are obtained through an attention mechanism. The original radar pulse signal undergoes signal-level physical consistency enhancement, and the time-frequency fusion features undergo feature-level hybrid enhancement to obtain enhanced features. Based on the enhanced features, the time-frequency dual-branch feature fusion network is trained end-to-end using a hierarchical contrastive learning framework that includes intra-class contrast and inter-class prototype contrast to obtain a radiation source identification model. Intra-class contrast includes constructing different enhanced versions of the same radar pulse signal as positive sample pairs and optimizing them using an intra-class feature contrast loss function, which is an InfoNCE-based loss function. Different enhanced versions refer to signal-level physical enhancement and feature-level hybrid enhancement. Inter-class prototype contrast includes calculating the mean of the time-frequency fusion features of all samples for each radiation source category to obtain the prototype vector of that category, and optimizing it using an inter-class prototype contrast loss function. The inter-class prototype contrast loss function is used to drive the time-frequency fusion features of samples to be closer to the prototype vector of their own category and further away from the prototype vectors of other categories. For the radar pulse signal to be identified, its multi-domain joint time-varying fingerprint features are extracted. The features and the corresponding original radar pulse signal are input into the trained time-frequency dual-branch feature fusion network to obtain time-frequency fusion features. The time-frequency fusion features are then input into the classification head of the trained radiation source identification model to output the individual category of the identified radiation source.

2. The highly interpretable radiation source identification method according to claim 1, characterized in that, The time-frequency dual-branch feature fusion network includes: The temporal branch inputs the multi-domain joint time-varying fingerprint feature sequence into a one-dimensional convolutional neural network to extract temporal features. The frequency domain branch extracts frequency domain features by performing a short-time Fourier transform on the original radar pulse signal and then using a two-dimensional convolutional neural network. The fusion module uses time-domain features as queries and frequency-domain features as keys and values, and performs weighted fusion through a scaling dot product attention mechanism to obtain the time-frequency fusion features.

3. The highly interpretable radiation source identification method according to claim 1, characterized in that, Signal-level physical consistency enhancement is achieved by applying controllable phase modulation based on a phase noise model to the original radar pulse signal.

4. The highly interpretable radiation source identification method according to claim 1, characterized in that, Feature-level hybrid enhancement refers to generating enhanced features by linearly interpolating the time-frequency fusion features of two different samples in the feature space, where the interpolation coefficients follow a Beta distribution.

5. The highly interpretable radiation source identification method according to claim 1, characterized in that, The inter-class prototype comparison loss function is expressed as follows: in, For the first Time-frequency fusion features of each sample For the first The true class label of each sample Let be the prototype vector of the k-th class. The total number of categories, The cosine similarity function is used. This refers to temperature hyperparameters.

6. The highly interpretable radiation source identification method according to claim 1, characterized in that, The joint loss function used in end-to-end training is expressed as: in, Cross-entropy classification loss, For intra-class feature contrast loss, For inter-class prototype comparison loss, For weight regularization, To balance the hyperparameters.

7. A highly interpretable radiation source identification system, characterized in that, include: The training signal time-varying feature extraction unit is used to acquire radar pulse signals from the training set and extract multi-domain joint time-varying fingerprint features based on physical layer modeling. These time-varying fingerprint features include instantaneous frequency deviation, instantaneous amplitude deviation, and phase difference sequence. Specifically, the instantaneous frequency deviation is calculated based on a phase model characterizing the individual hardware fingerprint of the radiation source. ,in, For instantaneous frequency deviation, The phase model is modeled as a superposition of a polynomial reflecting the nonlinearity of the power amplifier and a sine function reflecting periodic disturbances. The time-frequency feature fusion unit is used to input multi-domain joint time-varying fingerprint features and original radar pulse signals into the time-frequency dual-branch feature fusion network, and obtain time-frequency fused features through an attention mechanism; The identification model training unit is used to enhance the signal-level physical consistency of the original radar pulse signal and to enhance the time-frequency fusion features at the feature level, resulting in enhanced features. Based on the enhanced features, the time-frequency dual-branch feature fusion network is trained end-to-end through a hierarchical contrastive learning framework that includes intra-class contrast and inter-class prototype contrast, resulting in a radiation source identification model. Intra-class contrast includes constructing different enhanced versions of the same radar pulse signal as positive sample pairs and optimizing them using an intra-class feature contrast loss function, which is an InfoNCE-based loss function. Different enhanced versions refer to signal-level physical enhancement and feature-level hybrid enhancement. Inter-class prototype contrast includes calculating the mean of the time-frequency fusion features of all samples for each radiation source category to obtain the prototype vector of that category, and optimizing it using an inter-class prototype contrast loss function, which drives the time-frequency fusion features of samples to be closer to the prototype vector of their own category and further away from the prototype vectors of other categories. The radiation source identification unit is used to extract multi-domain joint time-varying fingerprint features from the radar pulse signal to be identified, input the features and the corresponding original radar pulse signal into the trained time-frequency dual-branch feature fusion network to obtain time-frequency fusion features, input the time-frequency fusion features into the classification head of the trained radiation source identification model, and output the individual category of the identified radiation source.

Images