Method and apparatus for individual radar radiation source identification in scenarios with small sample open sets

Abstract

This application relates to the field of individual radiation source identification, specifically to a method and apparatus for individual radar radiation source identification in small-sample open-set scenarios. One specific implementation of the method includes: inputting the acquired target input signal into a pre-trained optimized feature extraction model to obtain the original target activation vector; determining the target activation distance vector based on the original target activation vector; inputting the target activation distance vector into a preset extreme value theoretical model to obtain a target reliability score; correcting the original target activation vector based on the target reliability score to generate a corrected target activation vector; normalizing the corrected target activation vector based on a preset normalization formula to generate a normalized probability; and determining target identification information based on the normalized probability, a preset threshold, and a preset identification formula. This implementation can achieve individual radiation source identification under small-sample conditions.

Description

Technical Field

[0001] The embodiments of this application relate to the field of individual radiation source identification, specifically to a method and apparatus for individual radar radiation source identification in a small sample open set scenario. Background Technology

[0002] In both civilian and military applications, IoT devices are playing an increasingly important role, but their security has become a serious challenge, even limiting their widespread deployment. IoT devices are typically deployed in dynamic, open environments, making them vulnerable to attacks. Traditional authentication mechanisms based on cryptography or protocol layers are susceptible to cracking, impersonation, or man-in-the-middle attacks. Specific Emitter Identification (SEI) technology, as a physical layer security enhancement method, relies on the radio frequency fingerprint (RFF) generated by inherent defects in the radio frequency front-end hardware (such as power amplifier nonlinearity, filter transient response, crystal oscillator deviation, etc.). This provides a unique and difficult-to-clone additional guarantee for device authentication, thus SEI technology is widely used in the Internet of Things (IoT) and other fields.

[0003] In recent years, deep learning-based methods for identifying specific radiation sources (SEIs) have become a mainstream research direction in this field due to their powerful feature extraction capabilities. These models learn deep discriminative features directly from the raw signals in an end-to-end manner, demonstrating excellent classification performance on closed, fixed-class datasets. Traditional SEI research methods are mostly based on deep neural networks. For example, one-dimensional convolutional neural networks (1D-CNNs) are used to mine local features from the raw I / Q (in-phase orthogonal) time-series signals; two-dimensional convolutional neural networks (2D-CNNs) are used to process the transformed time-frequency maps or constellation diagrams to capture spatial patterns; and recurrent neural networks such as Long Short-Term Memory (LSTM) or Gated Recurrent Units (GRUs) are combined to capture long-range temporal dependencies in the signals. Furthermore, Residual Networks (ResNet) and Transformers have also been introduced to enhance the identification of subtle RF fingerprints. These methods generally learn device-specific fingerprint features from a large amount of labeled signal data through supervised learning, thereby achieving high-precision closed-set identification.

[0004] However, these methods essentially rely on a key assumption: that the set of radiation source categories in the training and testing environments is completely consistent. But the real electromagnetic environment is open and dynamically changing, inevitably receiving a large number of radiation source signals of unknown categories that did not appear during the training phase. If the identification system incorrectly classifies these unknown signals into one of the known categories, it will lead to serious misclassification, threatening the security and reliability of the entire authentication system. To address the problem of misidentification by traditional methods under open-set conditions, open-set identification has been introduced into the SEI (Search Engine Identification) field.

[0005] The core idea of ​​open-set recognition is to enable the model to accurately classify known categories while simultaneously recognizing and rejecting samples from unknown categories. Discriminative models represent the earliest and most direct approach to open-set recognition. Their core focus is not understanding what the unknown is, but rather precisely defining what is known. These methods typically train an excellent closed-set classifier first, and then, based on its output or intermediate features, use statistical tools to define boundaries for known classes, classifying samples outside these boundaries as unknown. Representative methods include extreme value machines and OpenMax (an open-set recognition method for deep learning models). Generative models take a different approach to solving the open-set recognition problem, actively building complete generative models for known category data. A representative work is G-OpenMax (a variant of the algorithm for open-set recognition), which introduces generative adversarial networks to synthesize spurious unknown category samples on top of the original OpenMax. By exposing the model to these pseudo-unknowns during training, it is forced to learn a more compact and robust decision boundary for known classes. With the development of deep learning, researchers are no longer satisfied with breaking down the open set recognition problem into a two-step approach of classification and processing, but are instead pursuing end-to-end solutions. The core of this approach is to design special network structures, loss functions, or training mechanisms that allow the network to directly learn to extract highly discriminative and compact features from the data. For example, introducing open selective kernel residual networks allows the network to adaptively focus on the most discriminative scale features in the signal by dynamically selecting the receptive field. Combining prototype networks with extremum theory allows features of each category to form compact clusters around their prototypes during the training phase through metric learning. Furthermore, by combining the decisions of multiple complementary base classifiers, the overall robustness and generalization ability of the open set recognition system can be improved.

[0006] Despite the positive progress made in open-set SEI research, its success generally rests on an implicit premise: the availability of sufficient and diverse known category samples for model training. However, non-cooperative IoT radiation source signal samples are scarce and labeling is costly, and the continuous addition of new devices to the network makes it difficult to obtain massive amounts of balanced labeled data. Even if experts can provide data labels, subtle differences at the fingerprint level can make manual labeling a challenging or even impossible task. This constraint of small sample sizes makes it difficult for most current open-set recognition models, which require large amounts of data, to learn highly generalizable feature representations, leading to a sharp decline in their performance. Summary of the Invention

[0007] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0008] Some embodiments of this application propose methods, apparatus, computer devices, and computer-readable storage media for individual radar radiation source identification in small sample open set scenarios to solve one or more of the technical problems mentioned in the background section above.

[0009] In a first aspect, some embodiments of this application provide a method for individual radar radiation source identification in a small-sample open-set scenario. The method includes: inputting the acquired target input signal into a pre-trained optimized feature extraction model to obtain the original target activation vector; determining the target activation distance vector based on the original target activation vector; inputting the target activation distance vector into a preset extreme value theoretical model to obtain a target reliability score; modifying the original target activation vector based on the target reliability score to generate a modified target activation vector; normalizing the modified target activation vector based on a preset normalization formula to generate a normalized probability; and determining target identification information based on the normalized probability, a preset threshold, and a preset identification formula.

[0010] Secondly, some embodiments of this application provide a radar radiation source individual identification device in a small-sample open-set scenario. The device includes: a first input unit configured to input the acquired target input signal into a pre-trained optimized feature extraction model to obtain the original target activation vector; a first determination unit configured to determine the target activation distance vector based on the original target activation vector; a second input unit configured to input the target activation distance vector into a preset extreme value theoretical model to obtain a target reliability score; a correction unit configured to correct the original target activation vector based on the target reliability score to generate a corrected target activation vector; a normalization unit configured to normalize the corrected target activation vector based on a preset normalization formula to generate a normalized probability; and a second determination unit configured to determine target identification information based on the normalized probability, a preset threshold, and a preset identification formula.

[0011] Thirdly, this application also provides a computer device, which includes a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein when the computer program is executed by the processor, it implements the method described in any implementation of the first aspect above.

[0012] Fourthly, this application also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the method described in any implementation of the first aspect above.

[0013] The above embodiments of this application have the following beneficial effects: The radar radiation source individual identification method in small-sample open-set scenarios, as described in some embodiments of this application, firstly, constructs an end-to-end FSOSR framework, integrating multi-scale feature extraction, discriminative feature optimization, and unknown class detection into a unified model. This overcomes the error accumulation problem caused by staged processing in traditional methods, significantly improving open-set recognition performance under small-sample conditions. Secondly, a feature extraction architecture based on a selective kernel residual network is designed, and a soft attention mechanism is introduced to dynamically weight and fuse multi-scale radio frequency fingerprint features. Based on this, positive and negative sample pairs are constructed using contrastive learning. By narrowing the distance between similar samples and widening the distance between dissimilar samples, the model's ability to perceive subtle hardware differences and its feature discriminativeness are significantly improved. Furthermore, a multi-task joint optimization method is proposed, combining contrastive learning with center loss to simultaneously optimize intra-class compactness and inter-class separability in the feature space. The OpenMax decision mechanism is embedded into the training process, achieving an organic unity between feature learning and open-set recognition. Therefore, individual radiation source identification under small-sample conditions can be achieved. Attached Figure Description

[0014] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0015] Figure 1 These are flowcharts of some embodiments of the radar radiation source individual identification method in a small sample open set scenario according to this application;

[0016] Figure 2 This is a flowchart illustrating some embodiments of the radar radiation source individual identification method in a small sample open set scenario according to this application;

[0017] Figure 3 This is a schematic diagram of the model structure of a selective kernel residual network for some embodiments of the radar radiation source individual identification method in the small sample open set scenario according to this application;

[0018] Figure 4 These are schematic diagrams of some embodiments of a radar radiation source individual identification device in a small sample open set scenario according to this application;

[0019] Figure 5 This is a schematic diagram of the structure of a computer device suitable for implementing some embodiments of this application;

[0020] Figure 6 This is a schematic diagram of the openness rate parameter;

[0021] Figure 7 This is a diagram illustrating the recognition performance of different methods;

[0022] Figure 8 This is a schematic diagram of the performance of the ablation experiment of the radar radiation source individual identification method in the small sample open set scenario of this application;

[0023] Figure 9 This is a schematic diagram of the experimental results of the radar radiation source individual identification method in the small sample open set scenario of this application under the condition of limited sample size;

[0024] Figure 10 This is a feature visualization of the radar radiation source individual identification method in the small sample open set scenario of this application. Detailed Implementation

[0025] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While some embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this application. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0026] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0027] It should be noted that the concepts of "first" and "second" mentioned in this application are only used to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0028] It should be noted that the terms "a" and "a plurality of" used in this application are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0029] The names of the messages or information exchanged between multiple devices in the embodiments of this application are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0030] The present application will now be described in detail with reference to the accompanying drawings and embodiments.

[0031] Figure 1 The flowchart 100 illustrates some embodiments of a radar source individual identification method in a small-sample open-set scenario according to this application. This radar source individual identification method in a small-sample open-set scenario includes the following steps:

[0032] Step 101: Input the acquired target input signal into the pre-trained optimized feature extraction model to obtain the original target activation vector.

[0033] In some embodiments, the execution entity of the radar radiation source individual identification method in a small sample open set scenario can input the acquired target input signal into a pre-trained optimized feature extraction model to obtain the target's original activation vector. The target input signal can be a signal representing the signal to be identified, acquired from a storage terminal via a wired or wireless connection. The pre-trained optimized feature extraction model can be a pre-trained neural network model that takes the input signal as input and the activation vector as output.

[0034] It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other currently known or future wireless connection methods.

[0035] As an example, a flowchart illustrating the method for identifying individual radar radiation sources in a small-sample open-set scenario is shown below. Figure 2 As shown. Figure 2 In this context, "Self-supervised Pretraining" refers to self-supervised pretraining. This represents the input signal; Dataaugment represents data augmentation. This indicates the first enhanced view. This indicates the second enhanced view. This represents feature extraction, Similarity represents similarity, and Model Fine-tuning with CenterLoss Function represents model optimization combined with the center loss function. Represents cross-entropy loss, The terms represent center loss, OpenMax Calibration, Open set recognition, Known emitter, Unknown emitter, Data accumulation, and Database.

[0036] like Figure 2 As shown, the Few-Shot Open Set Recognition Approach for Specific Emitter Identification (FSOSR-SEI) proposed in this application is an end-to-end framework for addressing the challenge of identifying unknown radiation source categories under small sample conditions. The core objective of this method is to achieve accurate identification of known categories and effective differentiation of unknown categories by fusing multi-scale feature extraction, discriminative feature optimization, and open set probability calibration. The overall process is as follows: Figure 2As shown, the system comprises three core steps: self-supervised pretraining, model optimization incorporating a center loss function, and OpenMax calibration (an open set identification method for deep learning models). In the self-supervised pretraining phase, a contrastive learning approach is first employed. Positive and negative sample pairs are constructed by injecting random noise and scaling the amplitude of the signal data. A Selective Kernel Residual Network (SCR) is used as the backbone feature extractor, forcing the model to learn robust feature representations invariant to data transformations. The core objective of this phase is to provide a highly generalizable feature foundation for subsequent tasks. Furthermore, the system introduces a center loss function, jointly optimizing the model with traditional cross-entropy loss to enhance the intra-class compactness and inter-class separability of features. In the OpenMax calibration phase, based on the trained feature distribution, the system fits the tail distance feature of each known class using a Weibull distribution. For test samples, OpenMax dynamically assesses the probability of unknown classes by calculating their distance to known class centers and the Weibull distribution probability, achieving reliable rejection of open set signals.

[0037] Optionally, prior to step 101, the aforementioned execution entity may also perform the following steps:

[0038] The first step is to select target training samples from the acquired training sample set. In practice, the aforementioned execution entity can randomly select training samples from the acquired training sample set as target training samples. The training samples in the aforementioned training sample set include sample input signals.

[0039] The second step involves performing data augmentation processing on the input signals of the selected target training samples based on a preset augmentation formula, to generate a first augmented view and a second augmented view. The preset augmentation formula can be:

[0040] ,

[0041] in, This indicates the first enhanced view. This indicates the second enhanced view. Indicates the enhancement function, This indicates that Gaussian noise is injected into the signal. This indicates that the signal is being randomly scaled. This represents the sample input signal.

[0042] Therefore, positive sample pairs can be generated by data augmentation of unlabeled radiation source signals, enabling the learning of feature representations robust to noise and perturbations. Two augmented views are generated through random noise injection and time shifting.

[0043] The third step is to determine the initial optimized feature extraction model. In practice, the aforementioned execution entity can use a selective kernel residual network as the initial optimized feature extraction model. As an example, the model structure of a selective kernel residual network is as follows: Figure 3 As shown. Figure 3 In this context, Input represents the input, Kernel 1×3 represents a 1×3 convolution kernel, Kernel 1×5 represents a 1×5 convolution kernel, Kernel 1×7 represents a 1×7 convolution kernel, and SoftMax represents normalization.

[0044] The fourth step involves inputting the first enhanced view and the second enhanced view into the initial optimized feature extraction model to obtain the first feature vector and the second feature vector.

[0045] Fifth, based on a preset loss function, determine the difference between the first feature vector and the second feature vector. The preset loss function can be:

[0046] ,

[0047] in, Indicates the difference value. Indicates comparative loss, Represents cross-entropy loss, Indicates the loss at the center. This indicates the weights used to adjust the center loss, for example, It can be 0.1;

[0048] The contrast loss is:

[0049] ,

[0050] in, Indicates the temperature coefficient, controlling the smoothness of the distribution. Represents the similarity function. Indicates the serial number. Indicates the indicator function, i.e., when The value is 1 if it is true, and 0 otherwise. This indicates the number of training samples in the training sample set;

[0051] The cross-entropy loss is:

[0052] ,

[0053] in, Indicates the number of known categories. Indicates the serial number. Indicates the first Each sample belongs to category The true label, The first prediction made by the optimized feature extraction model represents the... Each sample belongs to category The probability of;

[0054] The central loss is:

[0055] ,

[0056] in, Indicates the serial number. Indicates the first The feature vector of each sample (the feature vector extracted by the network). Indicates the first One sample, Indicates belonging to a category The center vector (the initial value is randomly sampled from the standard normal distribution, and the center vector will be gradually adjusted to the center position of the feature distribution of each category during the training process). Indicates the first Labels for each sample.

[0057] Therefore, the contrastive loss, by maximizing the consistency of positive sample pairs while suppressing similarity with negative samples, forces the model to learn discriminative features sensitive to hardware fingerprints, providing a highly generalizable feature space foundation for subsequent small-sample open-set recognition. Furthermore, this application introduces a center loss function (CL) for model optimization, with the core objective of addressing the open-space risk problem caused by the loose distribution of feature space in small-sample open-set recognition. The open-space risk is represented as:

[0058] ,

[0059] in, Indicates open space, This represents a known space. When the known class features are loosely distributed, the decision boundary expands to... The open space risk, as described above, is the risk of misclassifying samples from the open space as belonging to a known class. To reduce this risk, the distribution range of known class samples in the feature space needs to be reduced. This can be achieved by incorporating a center loss function during training. This function aims to cluster samples of the same class together as much as possible in the feature space, and the degree of clustering is adjustable. The center loss function forces the features of similar samples towards learnable class centers. Clustering, explicitly compressing the known class distribution range as shown in the formula corresponding to the center loss.

[0060] In addition, in the preset loss function These are the weights used to adjust the center loss. After training, the distribution of known classes in the feature space becomes more compact, causing the space that can be labeled as known to shrink. In other words, it becomes more difficult to identify any sample as a known class in an open space.

[0061] Step 6: In response to the determination that the difference value is greater than or equal to a preset difference value, adjust the network parameters in the initial optimized feature extraction model. For example, the difference between the above difference value and the preset difference value can be calculated. Based on this, the network parameters of the initial phase unwrapping model are adjusted using methods such as backpropagation and gradient descent. For example, the preset difference value can be 0.1.

[0062] Optionally, the aforementioned execution entity may also determine the initial optimized feature extraction model as the trained optimized feature extraction model in response to determining that the difference value is less than the aforementioned preset difference value.

[0063] Step 102: Determine the target activation distance vector based on the original target activation vector.

[0064] In some embodiments, the execution entity may determine the target activation distance vector based on the target original activation vector.

[0065] In practice, the input signal data is processed using a network, the activation vectors of correctly classified samples are retained, and activation vector sets for each category are constructed according to their respective categories. The aforementioned execution entity can calculate the target's original activation vector using a preset distance formula. To each category center The Euclidean distance is used to obtain the target activation distance vector. For example, the preset distance formula could be the Euclidean distance formula.

[0066] Step 103: Input the target activation distance vector into the preset extreme value theoretical model to obtain the target reliability score.

[0067] In some embodiments, the execution entity may input the target activation distance vector into a preset extreme value theoretical model to obtain the target reliability score.

[0068] In practice, the aforementioned executing entity can input the target activation distance vector into a preset extreme value theoretical model. In the middle, it was identified as the first kind Target reliability score .

[0069] In practice, the pre-defined mechanism theoretical model is trained through the following steps: First, the sample input signal is processed using an optimized feature extraction model, retaining the activation vectors of correctly classified samples. Activation vector sets for each category are constructed based on their respective categories. (indicating the first) The first step is to calculate the set of activation vectors for all samples in each class. Next, using a predefined distance formula, the distances from all activation vectors within each class to its class center (initially randomly sampled from a standard normal distribution, with the center vectors gradually adjusted to the center of each class's feature distribution during training) are calculated, forming distance sets for each class. Based on these distance sets, K Extreme Value Theory (EVT) models are fitted respectively. These models calculate a reliability score for a sample's classification based on its distance from its class MAV; the greater the distance, the lower the reliability. It should be noted that Extreme Value Theory (EVT) is a statistical framework describing the extreme values ​​of random variables. This application primarily uses the Weibull distribution, whose cumulative distribution function is:

[0070] ,

[0071] in, Indicates the central location of the distribution. Indicates the degree of dispersion of the distribution. Indicates the skewness of the distribution.

[0072] Step 104: Based on the target reliability score, the original target activation vector is modified to generate the target modified activation vector.

[0073] In some embodiments, the executing entity may modify the original target activation vector based on the target reliability score to generate a modified target activation vector. In practice, the modified target activation vector can be generated using the following formula:

[0074] ,

[0075] in, Represents the target modified activation vector. This represents the original activation vector of the target (in the recognition stage, the input signal data is first fed into the feature extraction network to obtain its activation vector). , The calculation process and same), This represents the target reliability score.

[0076] Step 105: Based on the preset normalization formula, normalize the target modified activation vector to generate normalized probabilities.

[0077] In some embodiments, the execution entity may normalize the target modified activation vector based on a preset normalization formula to generate a normalized probability. The preset normalization formula may be:

[0078] ,

[0079] in, Represents the normalized probability. Indicates a known category label. express Target identification information, Indicates the input sample. Indicates the first The target of the class is to modify the activation vector. Indicates the first The target modified activation vector of the class.

[0080] Step 106: Determine the target identification information based on the normalized probability, preset threshold and preset recognition formula.

[0081] In some embodiments, the execution entity may determine the target identification information based on the normalized probability, preset threshold, and preset identification formula. The preset identification formula may be:

[0082] ,

[0083] in, This represents a preset threshold, for example, It can be 0.95.

[0084] Therefore, OpenMax's implementation mainly consists of two steps: fitting an EVT model and calculating the probability of the unknown class. Specifically, for each class, the distance from each feature vector to the center point is calculated, and a certain number of samples with the largest distances are selected to fit an EVT model. Then, for each test sample, its distance to the center point of each class is calculated, and the probability of the unknown class is estimated using the EVT model.

[0085] The above embodiments of this application have the following beneficial effects: The radar radiation source individual identification method in small-sample open-set scenarios, as described in some embodiments of this application, firstly, constructs an end-to-end FSOSR framework, integrating multi-scale feature extraction, discriminative feature optimization, and unknown class detection into a unified model. This overcomes the error accumulation problem caused by staged processing in traditional methods, significantly improving open-set recognition performance under small-sample conditions. Secondly, a feature extraction architecture based on a selective kernel residual network is designed, and a soft attention mechanism is introduced to dynamically weight and fuse multi-scale radio frequency fingerprint features. Based on this, positive and negative sample pairs are constructed using contrastive learning. By narrowing the distance between similar samples and widening the distance between dissimilar samples, the model's ability to perceive subtle hardware differences and its feature discriminativeness are significantly improved. Furthermore, a multi-task joint optimization method is proposed, combining contrastive learning with center loss to simultaneously optimize intra-class compactness and inter-class separability in the feature space. The OpenMax decision mechanism is embedded into the training process, achieving an organic unity between feature learning and open-set recognition. Therefore, individual radiation source identification under small-sample conditions can be achieved.

[0086] Further reference Figure 4 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of a radar radiation source individual identification device in a small sample open set scenario. These embodiments of the radar radiation source individual identification device in a small sample open set scenario are similar to... Figure 1 Corresponding to the method embodiments shown, this radar radiation source individual identification device in the small sample open set scenario can be specifically applied to various electronic devices.

[0087] like Figure 4 As shown, the radar radiation source individual identification device 400 in some embodiments under small sample open set scenarios includes: a first input unit 401, a first determination unit 402, a second input unit 403, a correction unit 404, a normalization unit 405, and a second determination unit 406. The system comprises the following components: a first input unit 401, configured to input the acquired target input signal into a pre-trained optimized feature extraction model to obtain the original target activation vector; a first determination unit 402, configured to determine the target activation distance vector based on the original target activation vector; a second input unit 403, configured to input the target activation distance vector into a preset extreme value theory model to obtain a target reliability score; a correction unit 404, configured to correct the original target activation vector based on the target reliability score to generate a corrected target activation vector; a normalization unit 405, configured to normalize the corrected target activation vector based on a preset normalization formula to generate a normalized probability; and a second determination unit 406, configured to determine target recognition information based on the normalized probability, a preset threshold, and a preset recognition formula.

[0088] It is understandable that the units recorded in the radar radiation source individual identification device 400 in this small sample open set scenario are related to the reference. Figure 1 The steps in the described method correspond to each other. Therefore, the operations, features, and beneficial effects described above are also applicable to the radar radiation source individual identification device 400 and its constituent units in small sample open set scenarios, and will not be repeated here.

[0089] This application also provides a computer device 500. For example... Figure 5 As shown, the computer device 500 includes a bus 501, a processor 502, a memory 503, and a communication interface 504. The processor 502, the memory 503, and the communication interface 504 communicate with each other via the bus 501. The computer device 500 can be a server or a terminal device. It should be understood that this application does not limit the number of processors and memories in the computer device 500.

[0090] Bus 501 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 5 The bus 501 may be represented by a single line, but this does not mean that there is only one bus or one type of bus. The bus 501 may include a path for transmitting information between various components of the computer device 500 (e.g., memory 503, processor 502, communication interface 504).

[0091] Processor 502 may include any one or more processors such as a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP).

[0092] Memory 503 may include volatile memory, such as random access memory (RAM). Memory 503 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0093] The memory 503 stores executable program code, and the processor 502 executes the executable program code to implement the functions of the aforementioned first input unit, first determination unit, second input unit, correction unit, normalization unit, and second determination unit, thereby realizing the radar radiation source individual identification method in the small sample open set scenario. That is, the memory 503 stores instructions for executing the radar radiation source individual identification method in the small sample open set scenario.

[0094] The communication interface 504 uses transceiver modules such as, but not limited to, network interface cards and transceivers to enable communication between the computer device 500 and other devices or communication networks.

[0095] This application embodiment also provides a chip, which includes a processor and a data interface. The processor reads instructions stored in the memory through the data interface to execute the above-described method for individual radar radiation source identification in a small sample open set scenario.

[0096] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium capable of being stored by a computing device, or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct the computing device to execute the radar radiation source individual identification method in the small sample open set scenario described above.

[0097] Combination Figures 6-10Experiments were conducted on the WiSig dataset (a large-scale WiFi (Wireless Network Communication Technology) dataset) to fully verify the performance of the algorithm proposed in this chapter. The WiSig dataset is a publicly available large-scale WiFi dataset containing 10 million data packets collected over one month by 174 WiFi transmitters and 41 USRP (Universal Software Radio Peripheral) receivers. This application selected data from 20 WiFi transmitters collected by the same receiver. 128 samples were randomly selected from each type of radiation source as the training set.

[0098] In the open-set recognition scenario simulation of this application, K=10 types of radiation sources are fixed as known categories, and the recognition model is trained based on the datasets corresponding to these 10 types of radiation sources to simulate the model training process under the condition of prior known categories. In the model recognition testing phase, U newly added unknown radiation sources are mixed with the original K=10 known categories to form the corresponding test datasets. By gradually increasing the number of unknown categories U, the open-set recognition performance of the model when facing samples with different degrees of unknownness is systematically evaluated. The openness ratio is generally used to describe the degree of the number of unknown categories relative to the known categories in the experiment. For this experiment, its calculation formula is:

[0099] ,

[0100] in, This represents the opening rate. The opening rate parameter used in this experiment is as follows: Figure 6 As shown.

[0101] To comprehensively evaluate the overall performance of the model and avoid overly biased evaluations due to relying on a single metric, evaluation metrics such as ACC (accuracy) and F-measure (statistic) are introduced. The calculation formulas are as follows:

[0102] ,

[0103] in, Indicates accuracy rate. Indicates a statistical measure. Indicates a real example, Indicates a false positive example. Indicates a true counterexample. Indicates a false counterexample. And indicates precision, This indicates the recall rate.

[0104] The ACC value is a commonly used evaluation metric in SEI closed-set identification, while the F-measure is a commonly used evaluation metric in open-set identification. ACC represents the model's accuracy, while the F-measure combines recall and precision, taking into account the weights of both.

[0105] The comparative methods used in this experiment are as follows: G-OpenMax: Combines Generative Adversarial Networks (GANs) to synthesize unknown class samples to enhance training data, and integrates OpenMax's probability calibration mechanism to improve the model's ability to distinguish unknown classes. SKResNet: By combining deep feature distance measurement methods and extreme value theory, it models the statistical tail boundary of known class feature distances to achieve accurate identification of known devices and accurate detection of unknown intrusion devices. MASK-CNNLSTM: Constructs an open multi-scale attention kernel CNNLSTM model (a hybrid deep learning model combining Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs)). It enhances the representation ability by increasing the depth and complexity of the feature extraction network and improves it into an open set recognition architecture to achieve accurate identification of unknown transmitting devices under open sets.

[0106] This experiment aims to systematically evaluate the generalization ability and robustness of the proposed FSOSR-SEI (Few-Shot Open-Set Specific Emitter Identification) method in the task of identifying individuals from open-set radiation sources. Based on the WiSig dataset, the experiment simulates real-world identification challenges under different degrees of openness by setting open scenarios with an increasing gradient of the number of unknown classes. Comparison methods include the generative method G-OpenMax and cutting-edge deep learning models SKResNet and MASK-CNNLSTM. Accuracy and F1 (F-measure) scores are used as core evaluation metrics to quantitatively analyze the performance of FSOSR-SEI and verify its effectiveness in small-sample open-set scenarios.

[0107] Experimental results are as follows Figure 7 As shown, Figure 7A schematic diagram illustrating the recognition performance of different methods is shown. As the openness increases, both performance metrics for each method show a downward trend. However, the proposed FSOSR-SEI method exhibits significant superiority: Firstly, it leads in absolute performance values. At all openness levels, FSOSR-SEI's ACC and F1 scores are higher than the comparative methods. When the openness is 0.184, FSOSR-SEI's ACC reaches 89.9%, and its F1 score reaches 90.4%. When the openness is 0, FSOSR-SEI's ACC and F1 scores are 97.8% and 99.1%, respectively. Secondly, its performance degradation is more gradual. During the change in openness from 0 to 10, the decrease in ACC and F1 scores for FSOSR-SEI is smaller than that of the comparative methods, demonstrating its robustness improvement in small-sample open-set scenarios through the combination of self-supervised learning and center loss optimization strategies. This verifies the effectiveness of this technical approach in handling unknown class generalization and small-sample constraints in the task of identifying individuals in an open set of radiation sources.

[0108] To analyze the contribution mechanism of the self-supervised learning module and the central loss function module in the FSOSR-SEI method, this experiment designs an ablation study to verify their synergistic optimization effect. By progressively removing the self-supervised pre-training or central loss function components, method variants are constructed, and the performance differences between each variant and the complete model are compared on the WiSig dataset. The experiment uses accuracy and F1 score as evaluation metrics to verify their necessity in improving the discriminability of individual identification boundaries in small open sets.

[0109] Experimental results are as follows Figure 8 As shown. Figure 8 The diagram illustrates the performance of the ablation experiment of the radar radiation source individual identification method in a small sample open set scenario of this application. Figure 8 In the diagram, FSOSR-SEI represents the complete template, w / o SSL indicates the removal of the self-supervised training module, w / o CL indicates the removal of the center loss module, and w / o ALL indicates the removal of all modules. The results show that both the self-supervised learning module and the center loss function module make significant and complementary contributions to the performance of the FSOSR-SEI method. Removing either module leads to a systematic decline in the accuracy (ACC) and F1 score of the model in scenarios with increasing openness. Specifically, removing the self-supervised learning module has a more significant impact on performance, indicating its crucial role in robust feature representation learning; removing the center loss function mainly weakens the compactness of intra-class features. More importantly, removing both modules simultaneously leads to severe performance degradation. At an openness of 0.184, the ACC and F1 scores of FSOSR-SEI drop to 77.1% and 78.1%, respectively, with a significantly greater decline than removing either module independently. This fully validates the indispensability of self-supervised pre-training and the center loss function in co-optimizing the discriminability of the decision boundary, and together they ensure the generalization ability and robustness of the method in scenarios with small samples and high openness.

[0110] To verify the robustness of the FSOSR-SEI method under limited sample conditions, this experiment limited the number of training samples. With a high openness setting of 10 known classes and 10 unknown classes, the number of samples for the known classes was set to 128, 64 and 32 respectively. The performance changes of the proposed method were compared on the WiSig dataset.

[0111] Experimental results are as follows Figure 9 As shown. Figure 9 A schematic diagram shows the experimental results of the radar radiation source individual identification method in the small sample open set scenario of this application under the condition of limited sample size. Figure 9 In this context, "Number of samples" represents the number of samples, and "Performance" represents the performance. Experimental results show that the FSOSR-SEI method maintains strong robustness even when the number of training samples is drastically reduced. On the WiSig dataset, when the number of known class samples decreases from 128 to 32, the model accuracy (ACC) and F1 score exhibit a stepwise decline, but remain above 85%. Furthermore, the synchronous decline in F1 score and ACC is stable, with a maximum performance drop of approximately 5%, indicating that the FSOSR-SEI method has a certain advantage in sample efficiency. These results validate the adaptability of FSOSR-SEI to scenarios with limited samples.

[0112] To further validate the performance of the FSOSR-SEI method on small-sample open-set tasks, t-SNE (a nonlinear dimensionality reduction algorithm for high-dimensional data visualization) was used to perform visual dimensionality reduction on the features extracted from the test set data. In the visualization, known class samples were marked with circles, unknown class samples were marked with pentagrams, and different categories were distinguished by color. The results are as follows: Figure 10 As shown. Figure 10 The diagram shows a feature visualization of the radar radiation source individual identification method in the small sample open set scenario of this application. Figure 10 In this context, t-SNE Component 1 represents t-SNE component 1, t-SNE Component 2 represents t-SNE component 2, Unknown Class represents unknown class samples, and Class 0, Class 1, Class 2, Class 3, Class 4, Class 5, Class 6, Class 7, Class 8, and Class 9 represent different categories of known class samples.

[0113] Visualization results show that features of individuals in the known class form distinct, compact, and well-defined clusters, while clear inter-class boundaries also exist between the unknown and known classes. This directly benefits from the constraint of center loss on intra-class aggregation and the enhancement of data augmentation invariance through contrastive learning. More importantly, this explicit and compact clustering structure for the known classes naturally leaves ample empty regions in the feature space, making it easier for unknown class samples to be identified as out-of-distribution data during testing, thus significantly improving the model's recognition performance and robustness in open environments.

[0114] Therefore, the few-sample open-set target recognition (FSOSR) method requires the model to simultaneously achieve accurate classification of known classes and effective rejection of unknown classes with a limited number of labeled samples. Existing research has significant limitations: traditional few-sample methods only focus on closed-set classification, ignoring interference from unknown classes; while open-set target recognition methods rely on sufficient data to construct the distribution of unknown classes, failing in small-sample scenarios. Therefore, developing a radiation source-oriented FSOSR framework is of urgent significance for improving the real-time performance and robustness of battlefield electromagnetic target identification.

[0115] This application proposes an innovative Few-Shot Open Set Individual Recognition of Radiation Sources (FSOSR-SEI) framework to address the core challenges of radiation source individual recognition. This framework integrates self-supervised pre-training, feature space optimization based on center loss, and an OpenMax probabilistic calibration mechanism, significantly improving the robustness and environmental adaptability of signal classification. Specifically: first, a self-supervised learning mechanism is used to extract discriminative feature representations; a center loss function is introduced to enhance the intra-class compactness and inter-class separability of the feature space; and OpenMax probabilistic calibration effectively handles unknown signal types in open set scenarios, enhancing the system's generalization ability. Experiments show that this method achieves superior performance on two typical datasets. The FSOSR-SEI framework not only significantly reduces the requirement for labeled data but also improves the system's adaptability in non-steady-state wireless environments, providing effective technical support for real-time intrusion detection, security monitoring, and IoT security.

[0116] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0117] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of this application.

Claims

1. A method for individual radar radiation source identification in a small sample open set scenario, characterized in that, include: The acquired target input signal is input into a pre-trained optimized feature extraction model to obtain the original target activation vector; Based on the original target activation vector, determine the target activation distance vector; The target activation distance vector is input into a preset extreme value theoretical model to obtain the target reliability score; Based on the target reliability score, the original target activation vector is modified to generate a modified target activation vector; Based on a preset normalization formula, the target modified activation vector is normalized to generate a normalized probability. Based on the normalized probability, preset threshold, and preset recognition formula, target recognition information is determined.

2. The method for individual radar radiation source identification in a small sample open set scenario according to claim 1, characterized in that, Before inputting the acquired target input signal into a pre-trained optimized feature extraction model to obtain the original target activation vector, the method further includes: Select target training samples from the acquired training sample set, wherein the training samples in the training sample set include sample input signals; Based on a preset enhancement formula, data augmentation processing is performed on the sample input signals included in the selected target training samples to generate a first enhanced view and a second enhanced view. Determine the initial optimized feature extraction model; The first enhanced view and the second enhanced view are input into the initial optimized feature extraction model to obtain the first feature vector and the second feature vector; Based on a preset loss function, the difference value between the first feature vector and the second feature vector is determined. In response to the determination that the difference value is greater than or equal to the preset difference value, the network parameters in the initial optimized feature extraction model are adjusted.

3. The method for individual radar radiation source identification in a small sample open set scenario according to claim 1, characterized in that, The method further includes: In response to determining that the difference value is less than the preset difference value, the initial optimized feature extraction model is determined as the trained optimized feature extraction model.

4. The method for individual radar radiation source identification in a small sample open set scenario according to claim 2, characterized in that, The preset enhancement formula is: ， in, This indicates the first enhanced view. This indicates the second enhanced view. Indicates the enhancement function, This indicates that Gaussian noise is injected into the signal. This indicates that the signal is being randomly scaled. Indicates the sample input signal; The preset loss function is: ， in, Indicates the difference value. Indicates comparative loss, Represents cross-entropy loss, Indicates the loss at the center. This indicates the weight used to adjust the central loss. The contrast loss is: ， in, Indicates the temperature coefficient, controlling the smoothness of the distribution. Represents the similarity function. Indicates the serial number. Indicates the indicator function, i.e., when The value is 1 if it is true, and 0 otherwise. This indicates the number of training samples in the training sample set; The cross-entropy loss is: ， in, Indicates the number of known categories. Indicates the serial number. Indicates the first Each sample belongs to category The true label, The first prediction made by the optimized feature extraction model represents the... Each sample belongs to category The probability of; The central loss is: ， in, Indicates the serial number. Indicates the first The feature vector of each sample Indicates the first One sample, Indicates belonging to a category The center vector, Indicates the first Labels for each sample.

5. The method for individual radar radiation source identification in a small sample open set scenario according to claim 1, characterized in that, The target modified activation vector is generated using the following formula: ， in, Represents the target modified activation vector. Represents the original activation vector of the target. This represents the target reliability score.

6. The method for individual radar radiation source identification in a small sample open set scenario according to claim 1, characterized in that, The preset normalization formula is: ， in, Represents the normalized probability. Indicates a known category label. express Target identification information, Indicates the input sample. Indicates the first The target of the class is to modify the activation vector. Indicates the first The target modified activation vector of the class.

7. The method for individual radar radiation source identification in a small sample open set scenario according to claim 1, characterized in that, The preset recognition formula is: ， in, This indicates a preset threshold.

8. A radar radiation source individual identification device for small sample open set scenarios, characterized in that, include: The first input unit is configured to input the acquired target input signal into a pre-trained optimized feature extraction model to obtain the original target activation vector; The first determining unit is configured to determine the target activation distance vector based on the target's original activation vector; The second input unit is configured to input the target activation distance vector into a preset extreme value theoretical model to obtain the target reliability score; The correction unit is configured to correct the original target activation vector based on the target reliability score to generate a corrected target activation vector. The normalization unit is configured to normalize the target modified activation vector based on a preset normalization formula to generate a normalized probability. The second determining unit is configured to determine target identification information based on the normalized probability, the preset threshold, and the preset identification formula.

9. A computer device, characterized in that, The computer device includes a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein when the computer program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images