Zero sample specific radiation source identification method and system based on metric decoupling learning

Abstract

In order to solve the core problems that an existing SEI method is good in performance under idealized fixed parameters, but recognition performance is reduced due to working parameter changes (especially zero sample parameters) in an actual environment, and RFF features are weak and prone to being interfered by IM information, the invention provides a method for improving the recognition performance of the SEI method. The invention provides a zero-sample specific radiation source identification method and system based on metric decoupling learning, and belongs to the technical field of specific radiation source identification. According to the method, the measurement decoupling representation model is designed, the interference influence of IM information on RFF feature representation is weakened, and the measurement decoupling loss function is designed for the model, so that the causal logic relationship between the RFF and IM information in the RF signal is mined, and the independent RFF features are represented. Meanwhile, through an accumulation measurement recognition method, the model can deduce the radiation source identity of the zero sample working parameters, and the generalization performance of the model under the condition of the zero sample working parameters is ensured. The method is suitable for specific radiation source identification in a complex electromagnetic environment.

Description

Technical Field

[0001] This invention belongs to the field of specific radiation source identification technology, specifically relating to a zero-sample specific radiation source identification method based on metric decoupling learning. Background Technology

[0002] In recent years, deep learning-based Specific Emitter Identification (SEI) methods have made significant progress in improving the recognition performance of SEI tasks. These advancements have driven the exploration and application of more and more SEI algorithms in Electronic Intelligence (ELINT) systems. Early SEI methods mainly focused on identification under single operating parameters, i.e., training and testing were based on constant modulation type or frequency. These standard algorithms have fully demonstrated the effectiveness of deep learning technology in SEI tasks, but their underlying assumptions are too idealistic. Theoretically, even if the transmitter sends signals under different operating parameters, its radio frequency fingerprint (RFF) can still be characterized and used for identification. However, in real-world environments, RFF features are closely related to parameters such as the modulation type of the radio frequency (RF) signal. When the model is trained on one frequency band and tested on another, the recognition accuracy often drops significantly.

[0003] To address the challenges posed by the variable parameters of cognitive radar, an increasing number of scholars are focusing on SEI (Self-Identification and Interference) tasks under variable operating parameters. Due to the randomness and uncertainty of signal forms, this task faces additional challenges, especially when radar operating parameters change and the database lacks such samples (referred to as zero-sample operating parameters), leading to a significant decrease in recognition performance. This has become a key bottleneck in current electronic reconnaissance. The conventional RFF (Radar Frequency) features relied upon by existing ELINT systems cannot provide sufficient discriminative information to meet practical application requirements, thus increasing the difficulty of target radar radiation source identification. Therefore, under data constraints, improving the generalization performance of deep learning-based SEI methods—that is, enhancing the robustness and recognition ability of RFF features under zero-sample conditions—has become one of the important research directions.

[0004] In contrast, SEI tasks under variable or zero-sample operating parameters are still in their early stages. This task aims to learn effective features using limited known samples in a database to identify radiation sources with unseen operating parameters. In 2021, researchers first applied relational networks to zero-sample operating parameter identification, improving model generalization performance through a similarity scoring mechanism. In 2022, researchers proposed a variable modulation radiation source identification method based on domain adaptation, utilizing domain adversarial neural networks and Gaussian encoders to extract domain-invariant features, achieving good results in three variable modulation experiments. In 2023, researchers further proposed a radiation source identification method combining transfer learning and metric learning, reducing feature differences between different modulation types through adversarial domain adaptation and improving identification performance using pre-trained models within a few-shot learning framework.

[0005] Furthermore, accurately and robustly representing the RFF (Radiation Source Fiber) differences of radiation sources remains a key challenge in the SEI (Search Engine Imaging) task. Compared to intentional modulation (IM) information, RFF signals are extremely weak, accounting for only a small portion of the received signal, while IM information exhibits greater variability. This leads the SEI model to tend to capture IM features rather than RFF fingerprint features during the learning process. If IM interference cannot be effectively suppressed, a large amount of irrelevant information will be mixed in during the RFF representation stage, severely impacting recognition performance. Therefore, effectively suppressing IM interference and representing robust RFF features is one of the core challenges of the SEI method.

[0006] In summary, identifying radiation sources under both variable and zero-sample operating parameters remains a highly challenging research problem, especially given limited data. Therefore, this invention further explores a zero-sample SEI method to accurately identify radar radiation sources with unseen parameters. A new generation of ELINT systems with robust identification capabilities under zero-sample conditions will have significant theoretical and engineering implications for enhancing the intelligence and cognitive capabilities of electronic reconnaissance systems. Summary of the Invention

[0007] To address the core challenges of existing SEI methods, which perform well under idealized fixed parameters but suffer from reduced recognition performance in real-world environments due to variations in operating parameters (especially zero-sample parameters), and the weak radio frequency fingerprint (RFF) features susceptible to interference from intentional modulation (IM) information, this invention proposes a zero-sample specific radiation source identification method based on metric decoupling learning.

[0008] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a zero-shot specific radiation source identification method based on metric decoupling learning, the identification method comprising the following steps: Step 1: Collect radio frequency (RF) signal dataset, wherein the RF signal contains intentional modulation information (IM), radio frequency fingerprint (RFF), and ambient noise; Step 2: Construct a metric decoupling representation model and build a loss function for the metric decoupling representation model; Step 3: Input the dataset from Step 1 as the training set into the metric decoupling representation model, train the metric decoupling representation model using the loss function, and obtain the optimal metric decoupling representation model. Step 4: Construct a cumulative metric identification method; Step 5: Input the radio frequency signal (RF) dataset into the optimal metric decoupling representation model to effectively decouple the IM features and RFF features, obtain an independent and stable RFF feature representation, and use the cumulative metric identification method to identify specific radiation sources.

[0009] Furthermore, an RF vector signal generator is used for data acquisition.

[0010] Furthermore, the aforementioned metric decoupling representation model includes an IM feature characterizer, an RFF feature characterizer, and a signal coupler; The IM feature characterizer is used to automatically learn the IM features of each RF signal sample from the input RF signal carrying RFF information through deep learning, thus completing the mapping from the data space to the feature space. The RFF feature characterizer is used to automatically learn the RFF features of each RF signal sample through deep learning from the input RF signal carrying RFF information, thus completing the mapping from the data space to the feature space. The signal coupler is used to couple the decoupled IM features and RFF features to the RF signal in an unsupervised manner in the decoupled feature space, thereby completing the mapping from the feature space to the data space and realizing the effective decoupling of RFF features and IM features.

[0011] Furthermore, the formula for the above IM feature characterizer is:

[0012] The formula for the RFF feature characterizer is as follows:

[0013] The above signal coupler formula is:

[0014] in, The parameters represent the IM feature characterizer. The dimension is ; The parameters represent the RFF feature characterizer. The dimension is ; This represents the parameters of the signal coupler. Indicates the coupling of RF signals. It is a cascaded splicing feature of IM and RFF features, which serves as the input signal of the signal coupler.

[0015] Furthermore, the aforementioned loss functions include the signal coupling loss function, the IM decoupling loss function, and the discriminant metric loss function; The signal coupling loss function is used to ensure that the features of the signal after being characterized by the feature characterizer can still completely couple with the original radio frequency signal; The IM decoupling loss function is used to impose a standardization constraint on IM features, thereby enhancing the independence of RFF features; The discriminant metric loss function is used to reduce the distance between RF signal samples of the same type of radiation source with different operating parameters in the feature space during the learning process, and to increase the distance between RF signal samples of different types of radiation sources in the feature space.

[0016] Furthermore, the aforementioned signal coupling loss function includes both unsupervised learning loss function and orthogonal constraint loss function.

[0017] Furthermore, step 5 above specifically includes: Step 5.1: Connect multiple RF signal pulse segments to form a batch, and input the whole batch into the RFF feature characterizer to characterize the RFF features of each signal sample; Step 5.2: Obtain the accumulated RFF features by averaging the multiple RFF features obtained in the batch dimension; Step 5.3: Establish and load the ID library, which contains the centroids of ID features for each type of radiation source; Step 5.4: Calculate the metric score between the accumulated RFF features and the centroid of the ID features of each type of radiation source in the ID library, and perform metric decision calculation based on the metric score to determine the ID label of the radiation source.

[0018] Secondly, the zero-sample specific radiation source identification method based on metric decoupling learning described in this invention can be entirely implemented using computer software. Therefore, correspondingly, this invention also provides a zero-sample specific radiation source identification system based on metric decoupling learning.

[0019] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the zero-sample specific radiation source identification method based on metric decoupling learning described in any one of the preceding claims.

[0020] Fourthly, the present invention also provides a computer device, the device including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the zero-sample specific radiation source identification method based on metric decoupling learning described in any one of the above-mentioned methods.

[0021] The beneficial effects of this invention are as follows: This invention, starting from decoupling representation theory and combining the formation mechanism of RFF features, constructs a metric decoupling representation model that can effectively decouple RFF and IM features. This model overcomes the limitations of overall feature representation through joint optimization of loss functions in multiple branches. First, in the signal coupling branch, a signal coupling loss function is designed, and a feature-based orthogonal constraint loss function is introduced into it. By constructing orthogonality constraints between the RFF and IM feature branches in the feature space, the two are made independent of each other in the feature space, effectively avoiding the repeated representation of the same information in different features. This maintains the independence of RFF features while preventing IM features from penetrating the RFF feature space. This idea effectively suppresses feature leakage and information redundancy. Then, in the IM decoupling branch, through supervisory labels, the model is guided to automatically represent standardized IM features, thereby further enhancing the independence of RFF features. Secondly, in the discriminative metric branch, a discriminative metric loss function is proposed. This function utilizes metric learning to constrain the model's distribution in the feature space, ensuring that RFF features from the same type of radiation source cluster in the feature space, while keeping RFF features from different radiation sources decoupled. This constraint strengthens the intra-class clustering and inter-class dispersion of RFF features, thereby improving the model's discriminative ability. Finally, this invention proposes an accumulated metric recognition strategy in the recognition stage. By accumulating and averaging multi-pulse RFF features, the model maintains stable recognition performance even under zero-sample conditions. Therefore, through decoupling and metric learning, each RFF feature dimension independently represents the hardware characteristics of the radiation source, avoiding interdependence and information redundancy between features, retaining only RFF information highly relevant to the SEI task; making RFF features insensitive to IM changes, maintaining consistent feature representation under different IM conditions for the same radiation source, changing only with variations in the radiation source hardware; and inferring the radiation source identity under unseen operating parameters through metric learning, ensuring the model can still correctly classify under zero-sample operating parameter conditions.

[0022] This invention solves the problems of insufficient coupling and generalization ability between RFF and IM information, and provides a new technical path and theoretical support for the identification of specific radiation sources in complex electromagnetic environments. Attached Figure Description

[0023] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a flowchart of the zero-shot specific radiation source identification method based on metric decoupling learning described in this embodiment of the invention. Figure 2 This is a structural diagram of the metric decoupling representation model described in the embodiments of the present invention; Figure 3 This is a flowchart of the cumulative metric recognition method described in the embodiments of the present invention; Figure 4 This invention presents a comparison of the recognition accuracy performance of different methods under different modulation types (M1, M2) and zero-sample operating parameters (ZS) when the SNR is 20dB, as described in the embodiments of the present invention. Figure 5 This invention presents a comparison of the recognition accuracy performance of different methods under different modulation types (M1, M2) and zero-sample operating parameters (ZS) when the SNR is 10dB, as described in this embodiment. Figure 6 This invention presents a comparison of the recognition accuracy performance of different methods under different modulation types (M1, M2) and zero-sample operating parameters (ZS) when the SNR is 0dB, as described in the embodiments of the present invention. In this context, MAE represents mask autoencoder, MeDAE represents metric-enhanced denoising autoencoder, Class-DR represents classification decoupling representation model, AMAE represents asymmetric mask autoencoder, SNN represents Siamese neural network, and MDR represents the metric decoupling representation model proposed in this invention. Detailed Implementation

[0025] The specific implementation details (such as experimental setup, operation procedures, data processing steps, and example parameters) of "a zero-sample specific radiation source identification method based on metric decoupling learning" provided in this specification are primarily intended for illustrative purposes rather than limiting definitions, aiming to help those skilled in the art thoroughly understand the principles and implementation of the invention. However, those skilled in the art should understand that these details represent only one feasible embodiment, and the core concept of the invention can be fully implemented through other technical means or alternative solutions not described in detail, without departing from its spirit and essence. Furthermore, the omission of details of conventional experimental methods and apparatus known in the art in the specification is to avoid redundant information interfering with the understanding of the innovation points. This does not mean that these known technologies are not required during implementation, and those skilled in the art should be able to supplement and apply them based on their professional knowledge.

[0026] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. The following examples will help those skilled in the art to further understand the present invention, but do not limit the present invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

[0027] Example 1, Combination Figure 1 This embodiment addresses the core challenges of existing SEI methods, which perform well under idealized fixed parameters but suffer from reduced recognition performance in real-world environments due to variations in operating parameters (especially zero-sample parameters), and the weak radio frequency fingerprint (RFF) features susceptible to intentional modulation (IM) interference. Therefore, it proposes a zero-sample specific radiation source identification method based on metric decoupling learning. This identification method includes the following steps: Step 1: Collect radio frequency (RF) signal dataset, wherein the RF signal contains intentional modulation information (IM), radio frequency fingerprint (RFF), and ambient noise; Step 2: Construct a metric decoupling representation model and build a loss function for the metric decoupling representation model; Step 3: Input the dataset from Step 1 as the training set into the metric decoupling representation model, train the metric decoupling representation model using the loss function, and obtain the optimal metric decoupling representation model. Step 4: Construct a cumulative metric identification method; Step 5: Input the radio frequency signal (RF) dataset into the optimal metric decoupling representation model to effectively decouple the IM features and RFF features, obtain an independent and stable RFF feature representation, and use the cumulative metric identification method to identify specific radiation sources.

[0028] The specific radiation source identification method proposed in this embodiment is as follows: Figure 1As shown, this method starts from decoupling representation theory, combines the RFF feature formation mechanism, and deeply analyzes the positive and negative coupling relationships between RFF features and IM features. It constructs a metric decoupling representation model that can effectively decouple the two types of features, freeing the model from the limitations of holistic feature representation and significantly reducing the interference of IM information on RFF feature representation, thus obtaining RFF feature representations with independence and stability. Based on this, the model, through joint optimization of loss functions in multiple branches, frees the model from the limitations of holistic feature representation. In particular, it introduces a discriminative metric loss function to optimize and train the model parameters, enhancing the intra-class clustering and inter-class separability of RFF features through metric learning, thereby improving the discriminative ability and generalization performance of the features. Furthermore, this method also designs an accumulation metric recognition method in the recognition stage. By accumulating and averaging the RFF features represented by multiple pulse segments, it improves the noise resistance of the feature representation, achieving high-precision radiation source identification under zero-sample conditions.

[0029] The specific radiation source identification method proposed in this embodiment highlights the importance of RFF feature representation independent of IM information. From the perspectives of decoupling, measurement, and accumulation, it provides a new technical approach for RFF feature representation and zero-sample identification in complex electromagnetic environments.

[0030] Example 2, Combination Figures 1 to 3 This embodiment is a detailed description of the zero-shot specific radiation source identification method based on metric decoupling learning described in Embodiment 1 above. Step 1: Collect radio frequency (RF) signal dataset, wherein the RF signal contains intentional modulation information (IM), radio frequency fingerprint (RFF), and ambient noise; Specifically: Three types of RF vector signal generators were selected for data acquisition. The basic principle of a signal generator is the same as that of a transmitter; its internal components (such as power amplifiers) introduce RFF (Radio Frequency Filter) information, which is reflected in the transmitted RF signal. Furthermore, the selected signal generators can be programmed through a simulation platform to import ideal baseband signals and flexibly generate and acquire signals with various operating parameters as needed. In addition, the signal generators employ a pulsed signal transmission method, meaning that the same signal consists of multiple pulses.

[0031] During data acquisition, the baseband ideal signal generated by the simulation platform (with settings for modulation type M, signal bandwidth B, pulse width T, etc.) is first used to generate a corresponding ideal signal dataset, which is then imported into the signal generator. Subsequently, the signal generator modulates the baseband ideal signal into an RF signal (by setting the carrier frequency F and transmit power P). At this point, the RF signal incorporates RFF information and is transmitted into the spatial domain via the antenna. The receiving end uses a signal / spectrum analyzer to receive the RF signal, performs down-conversion and filtering, and finally obtains zero-IF I / Q data, completing the data acquisition process.

[0032] Therefore, the acquired data not only includes the RF signal emitted by the signal generator itself, but also incorporates some environmental noise introduced during its propagation in the airspace. Furthermore, in subsequent processing, Gaussian noise will be further superimposed on the acquired data to simulate different signal-to-noise ratios (SNR).

[0033] Therefore, variations in modulation type M, transmit power P, and carrier frequency F were considered in the collected dataset. The modulation type used was one of two commonly used signal forms: Linear Frequency Modulation (LFM) and Binary Phase Shift Keying (BPSK). The carrier frequency ranged from 1 to 3 GHz with a step size of 0.5 GHz, resulting in 5 frequency points. The transmit power ranged from 10 to 30 dBm with a step size of 5 dBm, resulting in 5 power levels. This resulted in 2 × 5 × 5 = 50 different combinations of operating parameters. Forty combinations were randomly selected and proportionally divided into a variable operating parameter training set and a variable operating parameter test set. The remaining 10 combinations were used as zero-sample operating parameters to construct the zero-sample operating parameter test set. Each set of operating parameter combinations was named in the form "M–F–P". For example, when the modulation type was BPSK, the transmit power was 10 dBm, and the carrier frequency was 1 GHz, it was named "M1–F1–P1".

[0034] Step 2: Construct a metric decoupling representation model and build a loss function for the metric decoupling representation model; Specifically, the design metric decoupling representation model is as follows: By designing a metric decoupling representation model, the effective decoupling of IM features and RFF features can be achieved, thereby obtaining RFF features with independence and stability.

[0035] (1) Input signal sample In the metric decoupling representation model, using Indicates a batch One RF signal sample, Indicates the number of radiation source categories. This indicates the number of RF signal samples from the same type of radiation source.

[0036] (2) Metric decoupling representation model structure The designed metric decoupling representation model includes an IM feature representative. An RFF feature characterizer and a signal coupler .like Figure 2 As shown, the structure diagram of the metric decoupling representation model is illustrated.

[0037] IM Feature Representer: The input is an RF signal carrying RFF information. With the help of deep learning technology, the IM feature characterizer can automatically learn the IM features of each RF signal sample. This enables an efficient mapping from the data space to the feature space.

[0038] In the formula, This represents the parameters of the IM feature characterizer. The dimension is .

[0039] RFF Feature Representer: Input is an RF signal carrying RFF information. With the help of deep learning technology, the RFF feature characterizer can automatically learn the RFF features of each RF signal sample. This enables the mapping from the data space to the feature space:

[0040] In the formula, The parameters represent the RFF feature characterizer. The dimension is .

[0041] To ensure successful decoupling of the RF signal into IM and RFF features, the features characterized by the two feature characterizers are coupled to the RF signal via a signal coupler. The signal coupler can then decouple the features... and An unsupervised approach is used to couple RF signals onto the decoupled feature space, achieving a mapping from the feature space to the data space. The signal coupling process can be described as follows:

[0042] In the formula, This represents the parameters of the signal coupler. This indicates a coupled RF signal. It is a cascaded concatenation of IM and RFF features, used as the input signal for the signal coupler. Through the action of the signal coupler, the RFF and IM features are effectively decoupled without losing the information carried by the data.

[0043] The loss function is designed as follows: Applying appropriate inductive bias constraints to the learning direction of the metric decoupling representation model can guide the model to uncover the causal logical relationship between RFF and IM information in the RF signal. Furthermore, the loss function can be redesigned to optimize the model's learning performance based on the representation requirements of RFF features.

[0044] In the metric-decoupled representation model, two independent feature representations represent two distinct feature representations for each RF signal sample, and are recoupled by concatenating and re-coupling the IM and RFF features. However, this re-coupling alone is insufficient to achieve the desired factorization between the two feature representations.

[0045] Therefore, to achieve effective decoupling, the loss function of the decoupling representation model is decomposed into three branches: 1) a signal coupling branch that promotes the coupling of IM and RFF features (signal coupling loss function); 2) an IM feature decoupling branch aimed at learning decoupling (IM decoupling loss function, which is used to decouple the coupling relationship between the original input RF signals); and 3) a discriminative metric RFF feature representation branch for the SEI task (discriminative metric loss function, which is directly related to the SEI task objective). The three losses are optimized together. When both the discriminative metric loss function and the decoupling loss are optimized simultaneously, the discriminative metric loss function can usually provide guidance for decoupling. Conversely, if a two-stage approach is adopted, i.e., first training the decoupling loss and then applying the decoupling features to the recognition task, the discriminative metric loss function cannot guide the decoupling process.

[0046] The signal coupling loss function for the first branch: To ensure that the features represented by the feature characterizer can still completely couple with the original radio frequency signal, an unsupervised learning loss function based on mean square error is designed under unsupervised conditions. This loss function measures the difference between the original signal and the coupled RF signal generated by the IM feature characterizer, RFF feature characterizer, and signal coupler in the data space. The unsupervised learning loss function is expressed as:

[0047] By minimizing This refers to the mean square error between the original signal and the coupled signal. Under an unsupervised learning framework, the model can automatically adjust its parameters to make the coupled RF signal approximate the original input RF signal as closely as possible, thus ensuring that no information is lost during feature decoupling.

[0048] Then, to further suppress redundancy and information overlap between IM features and RFF features, an orthogonal constraint loss function is proposed. That is, by minimizing the square of the inner product between IM features and RFF features, they are kept approximately orthogonal in the feature space, thereby ensuring that the RFF feature representation is not interfered with by IM information, and achieving effective decoupling of features. The orthogonality constraint loss function is expressed as:

[0049] Therefore, a feature-based orthogonality constraint is introduced into the signal coupling loss function. By constructing orthogonality constraints on the two feature branches, RFF and IM, in the feature space, the two branches are made independent of each other in the feature space. This constraint can effectively reduce the overlapping area of ​​the two feature subspaces and avoid the repeated representation of the same information in different branches. Thus, while maintaining the independence of RFF features, it prevents the infiltration of IM features into the RFF feature space. This mechanism effectively suppresses feature leakage and information redundancy.

[0050] Finally, the unsupervised learning loss function and the orthogonal constraint loss function are combined to form the complete signal coupling loss function:

[0051] in, This represents the weight coefficient of the orthogonal constraint loss function, used to balance the relationship between signal coupling accuracy and feature independence. By optimizing this signal coupling loss function, the model can maintain the accuracy of signal coupling while achieving non-interference between IM and RFF in the feature space, thereby obtaining more independent and stable RFF features.

[0052] The second branch's IM decoupling loss function: The key to decoupling lies in ensuring the solvability of IM and RFF features. To this end, the independence of RFF features is enhanced by imposing normalization constraints on IM features. The normalized representation of IM features is the foundation for ensuring the independence of RFF features. Specifically, ideal IM features are used. As labels, the mean squared error is used to constrain the IM feature characterizer to represent IM characteristics. That is, ideal IM features are used as supervisory labels to constrain the output of the IM feature representation. Therefore, the IM decoupling loss function is expressed as:

[0053] In the formula, It is obtained from an ideal RF signal that does not contain RFF information. This supervised learning makes the IM features identical to the ideal IM features. Notably, the goal of the second branch is to decouple the normalized IM features from the RF signal. At this point, how to obtain Labels are the primary factor for successful decoupling; therefore, we use a pre-trained standard autoencoder model to obtain ideal IM features as labels. At this point, the input to the autoencoder model is designed as ideal RF signals that do not carry RFF information. These ideal RF signals are ideal datasets generated and saved by the simulation platform, i.e., ideal baseband signals designed before being imported into the signal generator. Therefore, these ideal RF signals only contain IM information and not RFF information. Then, the autoencoder model will automatically learn the ideal IM features in the RF signals. And save it. It is worth noting that ideal RF signals are only used to constrain the model during the training phase. During the testing phase, ideal RF signals are not required; simply input the actual acquired RF signals (which contain RFF information and are composed of non-ideal RF signals) into the RFF feature characterizer to characterize the RFF features.

[0054] The third branch's discriminative metric loss function: Ultimately, the SEI task still relies on representing RFF features to identify radiation sources. Therefore, metric learning is used to constrain the target RFF features in the RF signal. A discriminative metric loss function is proposed, which reduces the distance between RF signal samples from the same type of radiation source with different operating parameters in the feature space, and increases the distance between RF signal samples from different types of radiation sources in the feature space during the model's learning process. The calculation steps of the discriminative metric loss function are as follows: Step (1), calculate the centroid of the identity (ID) feature. For each type of radiation source, its ID feature centroid... Defined as the average value of the RFF features of this class, the calculation formula is as follows:

[0055] In the formula, This indicates an average calculation.

[0056] For the Radiation source, which contains Sample RFF features The centroid of the ID feature is calculated as follows:

[0057] In the formula, in the class discrimination, if If the RFF feature and its ID feature centroid belong to different classes (negative samples), then... Conversely, if... If the centroids of the RFF feature and the ID feature belong to the same class (positive samples), then in subsequent calculations... In practice, appropriately deleting a positive sample can help improve the stability of the model.

[0058] Step (2): Calculate the metric score matrix. For each sample, the metric score is formed by the metric score of its RFF feature and the centroid of each category ID feature.

[0059]

[0060] In the formula, This indicates the calculation of the score. The core objective of this metric is to maximize the correlation between RFF features and the centroids of their respective class ID features, while minimizing their correlation with the centroids of features from other classes ID features, thereby achieving intra-class clustering and inter-class dispersion.

[0061] Specifically, in each batch of samples, the RFF feature is calculated and... The Pearson correlation coefficient, a measure of the centroids of the ID features, is calculated as follows:

[0062] In the formula, Let covariance and standard deviation of the corresponding feature be represented respectively. Represents a batch of samples The scores of each metric constitute a metric score matrix.

[0063] Step (3) Calculate the discriminant loss function. The core idea of ​​this loss function is to measure the similarity between the RFF feature of each sample and the centroid of the ID feature of various radiation sources in the feature space, and to regress the similarity to the basic facts through mean squared error. It can be expressed as:

[0064] In the formula, that is, when When, the similarity is equal to 1, when When the similarity is equal to 0, This represents the Pearson correlation coefficient score for each sample.

[0065] Step (4), calculate the discriminant metric loss function for a batch of samples, expressed as:

[0066] In the formula, if Conversely, it equals 0. The discriminant loss function achieves intra-class clustering and inter-class dispersion by calculating the Pearson correlation coefficient between RFF features and class centroids.

[0067] During the training phase, the three sub-networks are jointly trained end-to-end. Through the collaborative constraints of the three branches, the model achieves signal coupling, feature decoupling, and inter-class discrimination in the data space, feature space, and metric space, respectively. Therefore, the metric decoupling loss function of the metric decoupling model is expressed as:

[0068] In the formula, These represent the weights of the IM decoupling loss function, the discriminant metric loss function, and the signal coupling loss function, respectively.

[0069] Step 3: Input the dataset from Step 1 as the training set into the metric decoupling representation model, train the metric decoupling representation model using the loss function, and obtain the optimal metric decoupling representation model. Specifically: The collected training dataset is used as input to the metric decoupled representation model, and a loss function is applied. The training metric decoupled representation model was performed using the Adam optimizer with a learning rate of 0.0001 and a batch size of [missing value]. That is, a batch contains three types of radiation sources, with 20 samples for each type. The trained network is then saved to obtain the parameters of the RF feature characterizer. .

[0070] Step 4: Construct a cumulative metric identification method; Step 5: Input the radio frequency signal (RF) dataset into the optimal metric decoupling representation model to effectively decouple the IM features and RFF features, obtain an independent and stable RFF feature representation, and use the cumulative metric identification method to identify specific radiation sources.

[0071] Specifically: Unlike traditional single-pulse classification and recognition methods, this paper proposes an accumulative metric recognition method in the test recognition phase to further improve the robustness and recognition performance of RFF features. Specifically, this method proposes using multiple RF signal pulse segments, concatenating them to form a batch, and inputting the entire batch into an RFF feature characterizer to represent the RFF features of each signal sample. Subsequently, robust RFF features are obtained by averaging features over the batch dimension, and finally, the radiation source ID label is predicted through metric recognition, rather than relying on a single RFF feature. It is worth noting that this method achieves feature accumulation over the time dimension, rather than feature fusion. This simple and efficient operation requires no additional training and is particularly suitable for scenarios where the training data is noiseless and the test data is noisy.

[0072] Feature averaging calculation: In the testing phase, the RFF feature characterizer is first loaded, and the feature to be identified is... Multiple consecutive RF signal samples emitted by a radiation-like source The input is fed into the RFF feature characterizer to characterize its RFF features. Then, these RFF features are averaged to obtain the accumulated RFF features. This process improves the robustness of the RFF features by averaging features from multiple samples.

[0073]

[0074] like Figure 3 The diagram illustrates the flowchart of the cumulative metric identification method. First, an ID library needs to be built and loaded, containing the centroids of ID features for each type of radiation source. Then, the accumulated RFF features are compared with the centroids of ID features for each type of radiation source in the ID library. The scores are then used to calculate a metric decision based on these scores, thereby determining the ID label of the radiation source.

[0075]

[0076] like Figure 1 The flowchart shown is a zero-sample specific radiation source identification method based on metric decoupling learning according to the present invention. In the metric decoupling representation stage, decoupling learning and metric learning are used to represent RFF features independent of IM features from RF signals, so as to achieve effective decoupling of the feature space. In the accumulation metric identification stage, multi-pulse RF signals are used to represent RFF features through an RFF feature characterizer and metric calculation is performed to achieve high-precision radiation source identification under zero-sample conditions.

[0077] In summary, the zero-shot specific radiation source identification method based on metric decoupling learning proposed in this invention starts from decoupling representation theory and combines it with the formation mechanism of RFF features to construct a metric decoupling representation model that can effectively decouple RFF and IM features. This model overcomes the limitations of overall feature representation through joint optimization of loss functions in multiple branches. First, in the signal coupling branch, a signal coupling loss function is designed, and a feature-based orthogonal constraint loss function is introduced into it. By constructing orthogonality constraints between the RFF and IM feature branches in the feature space, the two are made independent of each other in the feature space, effectively avoiding the repeated representation of the same information in different features. This maintains the independence of RFF features while preventing IM features from penetrating into the RFF feature space. This idea effectively suppresses feature leakage and information redundancy. Then, in the IM decoupling branch, through supervision labels, the model is guided to automatically represent standardized IM features, thereby further enhancing the independence of RFF features. Secondly, in the discriminative metric branch, a discriminative metric loss function is proposed. This function utilizes metric learning to constrain the model's distribution in the feature space, ensuring that RFF features from the same type of radiation source cluster in the feature space, while keeping RFF features from different radiation sources decoupled. This constraint strengthens the intra-class clustering and inter-class dispersion of RFF features, thereby improving the model's discriminative ability. Finally, this invention proposes an accumulated metric recognition strategy in the recognition stage. By accumulating and averaging multi-pulse RFF features, the model maintains stable recognition performance even under zero-sample conditions. Therefore, through decoupling and metric learning, each RFF feature dimension independently represents the hardware characteristics of the radiation source, avoiding interdependence and information redundancy between features, retaining only RFF information highly relevant to the SEI task; making RFF features insensitive to IM changes, maintaining consistent feature representation under different IM conditions for the same radiation source, changing only with variations in the radiation source hardware; and inferring the radiation source identity under unseen operating parameters through metric learning, ensuring the model can still correctly classify under zero-sample operating parameter conditions. This invention solves the problems of insufficient coupling and generalization ability between RFF and IM information, and provides a new technical path and theoretical support for the identification of specific radiation sources in complex electromagnetic environments.

[0078] Example 3, Combination Figures 4 to 6 This embodiment is described in detail. It compares the metric decoupling representation model proposed in this invention with the prior art to fully verify the effectiveness of the metric decoupling representation model.

[0079] Specifically: By simulating different signal-to-noise ratios (SNR), the recognition accuracy performance of different methods under different modulation types (M1, M2) and zero-shot operating parameters (ZS) was compared. Figure 4The comparison of recognition accuracy performance of different methods under different modulation types (M1, M2) and zero-shot operating parameters (ZS) is shown when the SNR is 20dB. Figure 5 The comparison of recognition accuracy performance of different methods under different modulation types (M1, M2) and zero-shot operating parameters (ZS) is shown when the SNR is 10dB. Figure 6 The comparison of recognition accuracy performance of different methods under different modulation types (M1, M2) and zero-sample operating parameters (ZS) is shown when the SNR is 0dB.

[0080] In this context, MAE represents Masked Autoencoding (MAE), MeDAE represents Metric-enhanced Denoising Auto-encoder (MeDAE), Class-DR represents Class-Decoupled Representation (Class-DR), AMAE represents Asymmetric Masked Auto-encoder (AMAE), SNN represents Siamese Neural Network (SNN), and MDR represents the Metric-Decoupled Representation Model proposed in this invention.

[0081] from Figures 4 to 6 It is clear from the data that, regardless of the signal-to-noise ratio (SNR) under any circumstances, the metric decoupling representation model proposed in this invention achieves the highest recognition accuracy under different modulation types (M1, M2) and zero-shot operating parameters (ZS). This fully demonstrates that the zero-shot specific radiation source identification method based on metric decoupling learning proposed in this invention can effectively solve the problems of insufficient coupling and generalization ability between RFF and IM information, and has the highest accuracy compared to existing technologies. It provides a new technical path and theoretical support for the identification of specific radiation sources in complex electromagnetic environments.

[0082] Example 4: The zero-shot specific radiation source identification method based on metric decoupling learning described in any of the above embodiments can be entirely implemented using computer software. Therefore, correspondingly, this embodiment provides a zero-shot specific radiation source identification system based on metric decoupling learning, the system comprising: This is used to collect radio frequency (RF) signal datasets, wherein the RF signals contain intentional modulation information (IM), radio frequency fingerprints (RFF), and a storage device for environmental noise. Storage device for constructing a metric decoupling representation model and for constructing a loss function for the metric decoupling representation model; A storage device for inputting the dataset from step 1 as a training set into the metric decoupling representation model, training the metric decoupling representation model using a loss function, and obtaining the optimal metric decoupling representation model. Storage devices used to construct cumulative metric identification methods; This storage device is used to input radio frequency (RF) signal datasets into the optimal metric decoupling representation model, effectively decouple IM features and RFF features, obtain independent and stable RFF feature representations, and use an accumulative metric recognition method to identify specific radiation sources.

[0083] Example 5: This example provides a computer-readable storage medium storing a computer program. When the computer program is run by a processor, it executes the zero-sample specific radiation source identification method based on metric decoupling learning described in any of the above examples.

[0084] Those skilled in the art will understand that implementing all or part of the processes in the above embodiments can be accomplished by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0085] Example 6: This example provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a zero-sample specific radiation source identification method based on metric decoupling learning as described in any of the above examples.

[0086] This embodiment provides a computer device. This part of the hardware device is a general model and is not shown in the figure. The system includes a processor and a memory. The processor and the memory can be connected by a bus or other means. The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs and modules, as well as corresponding program instructions / modules. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions and modules stored in the memory, so as to realize the zero-sample specific radiation source identification method and steps based on metric decoupling learning in the above method embodiment.

[0087] The memory may include a program storage area and a data storage area, wherein the program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, mobile communication networks, and combinations thereof.

[0088] One or more modules are stored in the memory. When the processor executes, it performs the method steps in the embodiments. In this way, the invention objective can be achieved through the method, apparatus and process of the present invention. The specific details of the computer device described above can be understood by referring to the relevant descriptions and effects in the embodiments, and will not be repeated here.

[0089] The above description of the technical solution provided by the present invention through several specific embodiments is intended to highlight the advantages and benefits of the technical solution provided by the present invention. However, the above-described specific embodiments are not intended to limit the present invention. Any reasonable modifications and improvements to the present invention, reasonable combinations of implementation methods and equivalent substitutions based on the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A zero-shot specific radiation source identification method based on metric decoupling learning, characterized in that, The method is as follows: Step 1: Collect radio frequency (RF) signal dataset, wherein the RF signal contains intentional modulation information (IM), radio frequency fingerprint (RFF), and ambient noise; Step 2: Construct a metric decoupling representation model and build a loss function for the metric decoupling representation model; Step 3: Input the dataset from Step 1 as the training set into the metric decoupling representation model, train the metric decoupling representation model using the loss function, and obtain the optimal metric decoupling representation model. Step 4: Construct a cumulative metric identification method; Step 5: Input the radio frequency signal (RF) dataset into the optimal metric decoupling representation model to effectively decouple the IM features and RFF features, obtain an independent and stable RFF feature representation, and use the cumulative metric identification method to identify specific radiation sources.

2. The zero-shot specific radiation source identification method based on metric decoupling learning according to claim 1, characterized in that, An RF vector signal generator was used for data acquisition.

3. The zero-shot specific radiation source identification method based on metric decoupling learning according to claim 1, characterized in that, The metric decoupling representation model includes an IM feature characterizer, an RFF feature characterizer, and a signal coupler; The IM feature characterizer is used to automatically learn the IM features of each RF signal sample from the input RF signal carrying RFF information through deep learning, thus completing the mapping from the data space to the feature space. The RFF feature characterizer is used to automatically learn the RFF features of each RF signal sample through deep learning from the input RF signal carrying RFF information, thus completing the mapping from the data space to the feature space. The signal coupler is used to couple the decoupled IM features and RFF features to the RF signal in an unsupervised manner in the decoupled feature space, thereby completing the mapping from the feature space to the data space and realizing the effective decoupling of RFF features and IM features.

4. The zero-shot specific radiation source identification method based on metric decoupling learning according to claim 3, characterized in that, The formula for the IM feature characterizer is: The formula for the RFF feature characterizer is: The formula for a signal coupler is: in, The parameters represent the IM feature characterizer. The dimension is ; The parameters represent the RFF feature characterizer. The dimension is ; This represents the parameters of the signal coupler. Indicates the coupling of RF signals. It is a cascaded splicing feature of IM and RFF features, which serves as the input signal of the signal coupler.

5. The zero-shot specific radiation source identification method based on metric decoupling learning according to claim 3, characterized in that, The loss function includes the signal coupling loss function, the IM decoupling loss function, and the discriminant metric loss function; The signal coupling loss function is used to ensure that the features of the signal after being characterized by the feature characterizer can still completely couple with the original radio frequency signal; The IM decoupling loss function is used to impose a standardization constraint on IM features, thereby enhancing the independence of RFF features; The discriminant metric loss function is used to reduce the distance between RF signal samples of the same type of radiation source with different operating parameters in the feature space during the learning process, and to increase the distance between RF signal samples of different types of radiation sources in the feature space.

6. The zero-shot specific radiation source identification method based on metric decoupling learning according to claim 5, characterized in that, Signal coupling loss functions include unsupervised learning loss functions and orthogonal constraint loss functions.

7. The zero-shot specific radiation source identification method based on metric decoupling learning according to claim 3, characterized in that, Step 5 specifically involves: Step 5.1: Connect multiple RF signal pulse segments to form a batch, and input the whole batch into the RFF feature characterizer to characterize the RFF features of each signal sample; Step 5.2: Obtain the accumulated RFF features by averaging the multiple RFF features obtained in the batch dimension; Step 5.3: Establish and load the ID library, which contains the centroids of ID features for each type of radiation source; Step 5.4: Calculate the metric score between the accumulated RFF features and the centroid of the ID features of each type of radiation source in the ID library, and perform metric decision calculation based on the metric score to determine the ID label of the radiation source.

8. A zero-shot specific radiation source identification system based on metric decoupling learning, characterized in that, The system is implemented based on a zero-sample specific radiation source identification method based on metric decoupling learning as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the zero-sample specific radiation source identification method based on metric decoupling learning as described in any one of claims 1-7.

10. A computer device, characterized in that, The device includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a zero-sample specific radiation source identification method based on metric decoupling learning as described in any one of claims 1-7.

Images