Method for individual identification of a radiation source in an open environment

Abstract

The application discloses an open environment-oriented radiation source individual identification method and belongs to the technical field of radiation source individual identification. The method comprises the following steps: collecting source domain data samples, setting real classification labels of the samples, collecting target domain data samples, wherein the categories of part of the data samples of the target domain data belong to source domain shared categories, training a radiation source individual identification model based on the source domain and the target domain data sets, collecting signal data of a target radiation source individual, and obtaining the category of the target radiation source individual based on the radiation source individual identification model. The application is based on aligning the source domain signal and the target domain signal in a feature space, maximizing the category distance, and relieving the feature confusion between categories caused by cross-channel. Meanwhile, a multi-pair classifier is used to construct the classification boundary of unknown categories and known categories. The application simultaneously solves the feature confusion caused by cross-channel and the model untrustworthiness problem caused by unknown categories in an open environment.

Description

Technical Field

[0001] This invention belongs to the field of radiation source individual identification technology, specifically relating to a method for identifying radiation source individuals in open environments. Background Technology

[0002] Existing research on radiation source identification based on radio frequency fingerprinting mostly focuses on steps such as signal transformation, feature extraction, and classifier selection, aiming to improve the model accuracy of radiation source identification systems on closed-set identification tasks. However, the task of radiation source identification in real-world open environments is a more complex cross-channel open-set scenario.

[0003] First, due to environmental factors, differences in signal receivers, and interference from other transmitters, the channel information carried in signals varies and becomes confuses with each other in different deployment environments, resulting in differences in the distribution of RF fingerprints collected in different channel environments. This means that the signal data used for training and the signal data used for inference cannot meet the assumption of independent and identically distributed training and test sets required by traditional machine learning, further leading to a significant drop in model performance in cross-channel radiation source identification scenarios. Although deep learning-based SEI (Specific Emitter Identification) research has solved many challenges, most of the work is based on the assumption of closed-set classification, that is, both the training and test set signals come from a known set of transmitters. No matter how large this known set of transmitters is, if any unseen transmitter appears during the inference testing phase, its signal will be misclassified, ultimately leading to security vulnerabilities. A model trained with the RF fingerprint information of four known devices cannot identify or distinguish three other unknown individual devices during the prediction phase, making the model's inference results unreliable. Such a security vulnerability is fatal to any system deployed in the open space because it undermines the system's "autonomy," which is the system's ability to accomplish its objectives with minimal human intervention in complex and unpredictable environments.

[0004] Currently, ensuring the high degree of autonomy of open systems is a pressing issue. SEI models in open environments simultaneously encounter problems such as category information confusion due to channel variations and reduced model reliability caused by unknown categories. Data-driven SEI systems will operate as expected on in-set inputs, but their outputs on near-set inputs are unpredictable, inevitably leading to errors when processing all out-of-set inputs. Furthermore, the impact of out-of-set and near-set data on the task is not a simple additive summation; both act simultaneously, increasing the difficulty of knowledge transfer and open-set identification. Therefore, the final solution cannot be a simple combination of unsupervised domain adaptation and open-set identification.

[0005] Existing research employs independent methods to address the challenges in open environments. Firstly, to address the accuracy degradation of SEI models caused by cross-channel interference in open spaces, existing research has utilized two main strategies. The first strategy is based on signal data, analyzing target signal characteristics and using additional hardware or other channel information to improve the model's generalization ability across channels. In 2019, Sankhe et al. discovered a feedback-driven method for modifying the transmitter end by studying hardware-centric features in the transmission link. This method improves the accuracy of CNN classifiers by using channel estimation from the receiver. However, these signal data-based strategies are affected by the type of signal data itself, transmission protocols, etc., and also utilize redundant information outside the signal, making them difficult to apply on a large scale in complex and variable signal scenarios. Another strategy is Unsupervised Domain Adaptation (UDA). UDA is a subfield of transfer learning, which is an important means of improving model performance in the target domain by transferring knowledge learned from a labeled source domain to an unlabeled target domain, thereby reducing the difference between the two domains. In 2022, Ye et al. proposed an unsupervised Adversarial Domain Adaptation with Wasserstein Distance (ADAW) method to address the decline in accuracy of radiation source identification across time periods. Meanwhile, to address the problem of unknown categories in open spaces, existing research uses generative methods to actively reject unknown categories; other scholars utilize cosine distance to represent the boundary between known and unknown categories. In 2020, Hanna et al.'s research used multiple different open-set identification classification heads to address this issue and found that if the test data and training data were not captured on the same day, the model's recognition performance significantly decreased. Although Hanna et al. recognized that the fingerprint information extracted by the neural network was more likely from the channel, they did not address this problem directly. Summary of the Invention

[0006] The purpose of this invention is to address the aforementioned problems by proposing an open set radiation source individual identification method based on the Maximum Interval Open Boundary (MIOB), which enables the identification of unknown categories while mitigating cross-channel confusion.

[0007] The technical solution adopted in this invention is as follows:

[0008] A method for identifying individual radiation sources in open environments, comprising the following steps:

[0009] Step 1: Collect radiation signal data of various known radiation sources over a period of time using signal receiving equipment. Define n as the number of categories of radiation sources collected.

[0010] The collected radiation signal data were preprocessed to obtain multiple data samples for each radiation source individual and a radiation source individual category label for each data sample. All labeled data samples constituted the source domain dataset D. s ;

[0011] The data preprocessing involves: extracting M seconds of radiation source signal data as a data sample, determining which category of known radiation source the data sample belongs to based on the parameter settings of the signal receiving device, and setting a real classification label for the data sample, where M is a preset value;

[0012] Step 2: Based on the signal receiving device, collect radiation signal data of m+k individual radiation sources over a period of time, and truncate each collected radiation signal data point by a time length of M seconds to obtain multiple data samples for each individual radiation source. All data samples constitute the target domain dataset D. t ;

[0013] Among them, the categories of m types of radiation source individuals are the shared categories of the source domain, i.e., m≤n; the categories of k types of radiation source individuals are unknown categories that have not appeared in the source domain.

[0014] Step 3: Train the radiation source individual identification model based on the source domain and target domain datasets;

[0015] The radiation source individual identification model includes a feature extractor and a dual classifier, wherein the dual classifier includes a closed-set softmax classifier and a one-to-many classifier, and the one-to-many classifier includes n sub-classifiers, and the sub-classifier numbers correspond to the category numbers of the radiation source individuals.

[0016] The feature extractor is used to extract the feature information of the radiation source signal, and then the extracted radiation source signal features are input into the dual classifier;

[0017] The dual classifier is used to output the predicted classification label of the radiation source individual. It takes the class c with the highest classification probability in the Softmax classifier and the output of the c-th sub-classifier in the one-to-many classifier to obtain the probability that the input sample belongs to the c-th class. When the probability is less than a specified threshold, the input sample is considered to belong to the unknown class; otherwise, the c-th class is used as the predicted classification label output by the dual classifier.

[0018] Step 4: Real-time acquisition of signal data of the target radiation source individual at different times as data to be identified, wherein the length of the data to be identified is the same as that of the training set samples;

[0019] The currently collected data to be identified is input into the trained radiation source individual identification model. Based on the output predicted classification label, the identification result of the target radiation source individual is determined, that is, whether the target radiation source individual belongs to a known category or an unknown category.

[0020] Furthermore, in step 3, the feature extractor is a hybrid receptive field deep network. This network uses three sets of convolutional kernels of different sizes to extract signal features under different receptive fields and introduces a self-attention mechanism to apply attention to the convolutional kernels, so that the network can adaptively adjust the proportion of its convolutional kernels of different sizes.

[0021] Furthermore, the feature extractor specifically comprises:

[0022] The input feature map is first processed by three convolutional blocks with different kernel sizes to obtain three feature maps U1, U2 and U3. Each convolutional block includes a depthwise grouped convolution, a batch normalization layer and a GELU activation function in sequence; and the feature maps U1, U2 and U3 have the same dimension.

[0023] Feature maps U1, U2, and U3 are summed and fused according to element scale to obtain the feature map. The compressed vector z is obtained by passing it through a global average pooling layer. Attention based on channel scale is then applied to the compressed vector z to obtain the attention weights on each convolutional kernel of the three convolutional blocks.

[0024] Based on the attention weights on each convolutional kernel, the feature maps U1, U2 and U3 are weighted and fused to obtain the feature map V output by the feature extractor.

[0025] Furthermore, in step 3, labeled source domain data is used to supervise the training of the Softmax classifier, where the classification supervision loss of Softmax is:

[0026]

[0027] Among them, D s Represents source domain data, X s ,Y s v represents the source domain sample set and the corresponding true classification label set, respectively. s D represents s The number of samples, x i Let y represent the i-th sample data in the source domain. i x represents i Real category tags, L ce (This represents the cross-entropy loss.) G θ () represents the output of the feature extractor, i.e. the feature map extracted by the feature extractor, and softmax() represents the probability of a sample belonging to a certain category.

[0028] Furthermore, in step 3, the one-to-many classifier uses boundary loss during training, specifically:

[0029] choose The most relevant nearest neighbor class is used as the negative sample class of the sub-classifier in each one-to-many classifier to construct open boundaries;

[0030] The output of each sub-classifier is a two-dimensional vector representing the probability that a sample belongs to a known or unknown class. Let y be... s This represents the true category label of the source domain data. This represents the source domain sample x output by the corresponding subclassifier. s It is the yth s The probability of a class is then determined by the boundary loss constructed using the open boundary construction of the bi-class classifier:

[0031]

[0032] Among them, X s ,Y s Let a represent the source domain sample set and the corresponding true classification label set, respectively. i It belongs to Y s a subset of This represents the source domain sample x output by the corresponding subclassifier. s It is the ath i The probability of a class.

[0033] Furthermore, in step 3, the total loss of the radiation source individual identification model during training is:

[0034] L all =L cls (X s ,Y s )+L obc (X s ,Y s 0+λ×L mi (X t )

[0035] Among them, L cls (X s ,Y s ) represents the classification supervision loss of Softmax, L obc (X s ,Y s L represents the boundary loss of a one-to-many classifier. mi (X t X represents maximizing the class distance loss. s ,Y sLet X represent the source domain sample set and the corresponding true classification label set, respectively. t Let represent the target domain set, and λ represent the balance coefficient for the maximum spacing loss;

[0036] in,

[0037]

[0038] The degree of confusion between category i and any other category j Let W represent the probability of predicting class i in a training batch, where W is a diagonal matrix. express The diagonal elements corresponding to each sample, the diagonal elements of matrix W kk The classifier's predictive certainty regarding the class of the k-th sample is expressed as follows:

[0039]

[0040] Where B represents the size of a batch of training data during target domain training. This represents the probability that the k-th sample in the target domain output by the Softmax classifier is classified into the j-th class. This represents the probability that the k′-th sample in the target domain output by the Softmax classifier is classified into the j-th class.

[0041] The technical solution provided by this invention brings at least the following beneficial effects:

[0042] This invention employs a hybrid receptive field deep network based on a self-attention mechanism, using multiple sets of convolutional kernels of different sizes with adaptive weights to extract signal features under different receptive fields. Compared to the hand-crafted features or ordinary deep features used in existing studies of individual radiation sources, this invention can more fully characterize the radio frequency fingerprint information of the signal. Simultaneously, this invention constructs the classification boundary between unknown and known classes based on a multi-pair classifier, achieving open set recognition through the collaboration of two classifiers. Compared to existing algorithms, this eliminates the need for manually setting the threshold for rejecting unknown samples. Finally, this invention maximizes the class distance by aligning source and target domain signals in the feature space, mitigating feature confusion between classes caused by cross-channel interference. Compared to existing research, this invention can simultaneously solve the problems of feature confusion caused by cross-channel interference and model unreliability caused by unknown classes in an open environment. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of the process for identifying individual radiation sources in open environments, provided in an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of an open set radiation source individual identification model structure based on the maximum spacing open boundary provided by an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of the method for rejecting samples of unknown categories using a dual classifier structure in an example of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0048] This invention provides a method for identifying radiation sources in open environments. Firstly, by aligning source and target domain signals in the feature space, the method maximizes the inter-class spacing, mitigating feature confusion between classes caused by cross-channel interference. Simultaneously, it constructs classification boundaries between unknown and known classes based on a multi-pair classifier. This approach simultaneously addresses feature confusion caused by cross-channel interference and model unreliability issues due to unknown classes in open environments.

[0049] like Figure 1 , Figure 2 As shown in the figure, the specific processing procedure of the radiation source individual identification method for open environments provided by the embodiments of the present invention includes the following steps:

[0050] Step S1: Collect source domain data samples and set the true classification labels for the samples;

[0051] Radiation signal data of n (n>1) known radiation sources were collected over a period of time using a signal receiving device. The collected radiation signal data were preprocessed to obtain multiple data samples for each radiation source and a category label for each data sample. All labeled data samples constituted the source domain dataset D. s ;

[0052] The data preprocessing is as follows: extract M seconds of radiation source signal data as a data sample, and determine which type of known radiation source the data sample belongs to based on the parameter settings of the signal receiving device, and set a real classification label for the data sample, where M is a preset value;

[0053] Step S2: Collect data samples from the target domain, wherein some data samples from the target domain belong to the category shared by the source domain.

[0054] Radiation signal data of m+k individual radiation sources were collected over a period of time using a signal receiving device. Among these, m categories are shared with the source domain (m≤n), and k unknown categories are not present in the source domain. Each collected radiation signal data point is truncated to a duration of M seconds, resulting in multiple data samples for each individual radiation source. All data samples constitute the target domain dataset D. t ;

[0055] Step S3: Train the radiation source individual identification model based on the source domain and target domain datasets;

[0056] The radiation source individual identification model includes a feature extractor and a dual classifier, wherein the dual classifier includes a closed-set softmax classifier and a one-vs-all (OvA) classifier.

[0057] The feature extractor is used to extract the feature information of the radiation source signal, and then the extracted radiation source signal features are input into the dual classifier;

[0058] The dual classifier is used to output the predicted classification label of the radiation source individual. First, the class c with the highest classification probability in the Softmax classifier is selected. Then, the cth class in the OvA classifier group is selected to obtain the probability that the input sample belongs to the cth class. When the probability is less than a specified value (preferably set to 0.5), the input sample is considered to belong to the unknown class. Otherwise, the cth class is used as the predicted classification label output by the dual classifier.

[0059] Step S4: Collect signal data of the target radiation source individual and obtain the category of the target radiation source individual based on the radiation source individual identification model.

[0060] Signal data of individual target radiation sources at different times are collected in real time as data to be identified, and the length of the data to be identified is the same as that of the training set samples.

[0061] The collected data to be identified is input into the radiation source individual identification model trained in step S3 to obtain the radiation source individual predicted classification label, thereby determining whether the target radiation source individual belongs to a known category or an unknown category.

[0062] As one possible implementation, in step S3, the specific training of the radiation source individual identification model is as follows:

[0063] like Figure 2 As shown, for the given feature map X∈R l×C×B Where l is the data length, C is the number of channels in the input feature, and B refers to the batch size of the training data. For all convolution operations, three transformations are performed first:

[0064] F1∶X→U1∈R l′×C′×B ;

[0065] F2:X→U2∈R l′×C′×B ;

[0066] F3:U3∈R l′×C′×B ;

[0067] Among them, F1, F2 and F3 all include depthwise grouped convolution, batch normalization and GELU activation function in turn. GELU activation function is Gaussian error linear unit activation function, which is constructed by combining some properties of dropout, zoneout and ReLU activation function. It has the advantages of higher generalization and faster convergence.

[0068] In this embodiment, the kernel size in the F1 transform is 3, resulting in feature map U1; the kernel size in the F2 transform is 5, resulting in feature map U2; and the kernel size in the F3 transform is 3, resulting in feature map U3. The transformed feature maps have the same dimension, l′×C′×B, where l′ represents the length of the transformed data and C′ represents the number of data channels in the transformed feature map.

[0069] The goal of a feature extractor is to enable neurons to adaptively adjust the proportion of convolutional kernels of different sizes based on feature inputs. The basic idea is based on a self-attention mechanism, mixing information from multiple branches of different receptive fields before outputting it to the next layer of neurons. To achieve this, the network needs to fuse feature data from all different receptive field branches. First, the outputs of multiple branches are simply fused by summing at an element-wise scale:

[0070] Step S3: Then, global average pooling is used to embed global information to generate a compressed vector z, where z∈R. C ′×B More specifically, the k-th element z in vector z. k By compressing and fusing features Calculated using the spatial dimension l′:

[0071]

[0072] Among them, F gp () indicates global average pooling.

[0073] Then, spatial scale information is adaptively selected by applying channel-scale-based attention to the compressed vector z. Specifically, the Softmax operator is applied to the values ​​of each channel, denoted as a. k It is the k-th element of vector a, b k c k Similarly, the final feature map V is obtained through the attention weights on each convolutional kernel:

[0074] V k =a k ·(U1) k +b k ·(U2) k +c k ·(U3) k a k +b k +c k =1,

[0075] Where V = [V1, V2, ..., V C ],V k ∈R l×B

[0076] This design allows the feature extractor to obtain feature information of the signal in different receptive fields through convolutional kernel structures of different sizes. On the other hand, it introduces an attention mechanism on different convolutional kernels, allowing the network to adaptively obtain convolutional kernel weights by increasing a small number of network parameters. Through these two methods, the feature extractor can extract more robust signal features.

[0077] To reject samples of unknown categories, existing algorithms manually set a threshold for rejecting unknown samples based on prior knowledge or a predefined ratio of unknown samples. However, this strategy is impractical in real-world applications. This invention introduces a dual classifier structure, such as... Figure 3 As shown. Considering that the inter-class distance in the source domain is a good threshold for determining the known or unknown values ​​in the target domain, in addition to designing a Softmax closed-set classifier, a one-vs-all classifier was trained for each class using labeled source domain data.

[0078] For the testing phase after model training is completed, the test input samples... The input probabilities are obtained after passing through a feature extractor and two classifiers. The k-th row of the output from the Softmax classifier represents the probability that the input sample belongs to the k-th class.

[0079] In the OvA classifier group, the first value of the k-th classifier represents the probability that the input sample belongs to the k-th class (called the known class score), and the second value represents the probability that it does not belong to the k-th class.

[0080] In the final inference stage: First, the class with the highest classification probability in the Softmax classifier is selected and denoted as class c. Then, the c-th class in the OvA classifier group is selected, and its probability of belonging to class c is obtained (given class scores). If the probability is less than 0.5, the input sample is considered to belong to the unknown class; otherwise, it is considered to belong to class c. This design eliminates the need for manual threshold setting, using only the collaboration of two classifiers to reject samples of unknown class.

[0081] During training, labeled source domain data is used to supervise the training of the Softmax classifier. The specific classification supervision loss of Softmax is:

[0082]

[0083] Among them, D s Represents source domain data, X s ,Y s v represents the source domain sample set and the corresponding label set, respectively. s x represents the number of source domain sample data. i Let y represent the i-th sample data. i L represents the true classification label of the i-th sample data. ce (This represents the cross-entropy loss.) G θ () indicates the output of the feature extractor.

[0084] In supervised learning, cross-entropy loss is often used to measure the difference between the predicted output value and the true value of a classifier. Let y represent the true label, and p(y|x) represent the output of the Softmax classifier, then the cross-entropy loss is expressed as:

[0085]

[0086] Where y represents the label, x represents the sample data, and the subscript i represents the sample number.

[0087] In other works, to construct the open set recognition boundary, the sum of n binary cross entropies (BCEs) is used as the loss for supervised training for the OvA classifier group. However, due to the strong fitting effect of supervised training, the open set recognition performance on unlabeled data in the target domain further deteriorates.

[0088] To improve the efficiency of dual classifiers in collaboratively rejecting unknown categories in the target domain, a method is proposed to construct open classification boundaries for OvA classifier groups using labeled source domain data, thereby achieving more generalizable open set recognition.

[0089] Use labeled source domain data to establish the boundaries between each known class and outlier in the OvA classifier group from scratch.

[0090] Instead of having every source domain sample work for all OvA classifiers, a subset of categories that are different from but more relevant to the sample label are selected as negative categories.

[0091] We weighed the open set recognition performance on the target domain test set against the domain adaptation classification performance, and selected [n / 2] (where n is the number of source domain classes) of the most relevant nearest neighbor classes as the negative sample classes for each OvA subclassifier to construct the open boundary from scratch.

[0092] The Open Set OvA classifier consists of w subclassifiers. The output of each subclassifier is a two-dimensional vector, where each dimension shows the probability that a sample is an in-class value or an outlier.

[0093] The OvA classifier operates on the extracted features, denoted as z. k =w k G θ (x)∈R 2 G θ and w k These represent the network parameters of the feature extractor and the corresponding OvA subclassifier for the k-th class, respectively; each dimension z k ∈R 2 These represent the scores for the known and unknown categories, respectively. This represents the probability that sample x is an in-class value of the k-th class in the OvA classifier, i.e. The loss calculation for open boundary construction based on a dual classifier is as follows:

[0094]

[0095] Among them, X s Y represents the source domain dataset. s Represents the source domain tag set, a i A subset belonging to the source domain tag set, i.e., a i ∈Y s And satisfy

[0096] Inter-class confusion is introduced to describe the uncertainty of classifier inference results:

[0097]

[0098] Among them, C ij Describes the degree of confusion between category i and any other category j. W represents the probability of predicting class i in a training batch; W is a diagonal matrix with diagonal elements W ii This indicates the class predictive certainty of the classifier for the class of the i-th sample.

[0099] W ii The specific calculation process involves first calculating the inverse entropy of the classifier's prediction of the i-th sample, and then calculating the probability output after passing through the Softmax function. This probability represents the class prediction certainty of the i-th sample, specifically expressed as...

[0100]

[0101] Where n is the total number of categories in the source domain, and B is the batch size of the training data during target domain training. This represents the probability that the i-th sample is classified into the j-th class.

[0102] C ij This describes the degree of confusion between category i and any other category j. The objective of this invention is to maximize the distance between each category in the target domain, i.e., to minimize the confusion between categories in the target domain. Therefore, a maximum interval loss is defined:

[0103]

[0104] Among them, X t This represents the target domain dataset.

[0105] Therefore, the overall loss function during the training phase is:

[0106] L all =L cls (X s ,Y s )+L obc (X s ,Y s )+λ×L mi (X t )

[0107] Where λ is the balance coefficient (a hyperparameter) for the maximum spacing loss.

[0108] This invention provides a method for identifying individual radiation sources in open environments. It is highly practical and simultaneously solves the problems of feature confusion caused by cross-channel interference and model unreliability caused by unknown categories in open environments.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0110] The above descriptions are merely some embodiments of the present invention. For those skilled in the art, various modifications and improvements can be made without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for identifying individual radiation sources in open environments, characterized in that, Includes the following steps: Step 1: Collect radiation signal data of various known radiation sources over a period of time using signal receiving equipment. Define n as the number of categories of radiation sources collected. The collected radiation signal data were preprocessed to obtain multiple data samples for each radiation source individual and a radiation source individual category label for each data sample. All labeled data samples constituted the source domain dataset. ; The data preprocessing involves: extracting M seconds of radiation source signal data as a data sample, determining which category of known radiation source the data sample belongs to based on the parameter settings of the signal receiving device, and setting a real classification label for the data sample, where M is a preset value; Step 2, based on signal receiving equipment acquisition The radiation signal data of an individual radiation source over a period of time is collected, and each collected radiation signal data is truncated to a time length of M seconds to obtain multiple data samples for each individual radiation source. All data samples constitute the target domain dataset. ; in, The category of an individual radiation source is a shared category of the source domain, that is... ; The individual radiation source is classified as an unknown category that has not appeared in the source domain; Step 3: Train the radiation source individual identification model based on the source domain and target domain datasets; The radiation source individual identification model includes a feature extractor and a dual classifier, wherein the dual classifier includes a closed-set softmax classifier and a one-to-many classifier, and the one-to-many classifier includes n sub-classifiers, and the sub-classifier numbers correspond to the category numbers of the radiation source individuals. The feature extractor is used to extract the feature information of the radiation source signal, and then the extracted radiation source signal features are input into the dual classifier; The dual classifier is used to output the predicted classification label of the radiation source individual. It takes the class c with the highest classification probability in the Softmax classifier and the output of the c-th sub-classifier in the one-to-many classifier to obtain the probability that the input sample belongs to the c-th class. When the probability is less than a specified threshold, the input sample is considered to belong to the unknown class; otherwise, the c-th class is used as the predicted classification label output by the dual classifier. Step 4: Real-time acquisition of signal data of the target radiation source individual at different times as data to be identified, wherein the length of the data to be identified is the same as that of the training set samples; The currently collected data to be identified is input into the trained radiation source individual identification model, and the identification result of the target radiation source individual is determined based on the output predicted classification label.

2. The method as described in claim 1, characterized in that, The feature extractor of the radiation source individual identification model is a hybrid receptive field deep network. This network uses three sets of convolutional kernels of different sizes to extract signal features under different receptive fields and introduces a self-attention mechanism to apply attention to the convolutional kernels.

3. The method as described in claim 2, characterized in that, In step 3, the feature extractor specifically refers to: The input feature map is first processed by three convolutional blocks with different kernel sizes to obtain three feature maps. , and Each convolutional block consists of depthwise grouped convolutions, batch normalization layers, and a GELU activation function; and the feature map... , and The dimensions are the same; Feature maps are summed and fused at the element scale. , and , to obtain feature map The compressed vector is then obtained through a global average pooling layer. In the compressed vector The attention weights on each convolutional kernel of the three convolutional blocks are obtained using channel-scale-based attention. Based on the attention weights obtained on each convolutional kernel, the feature map , and Weighted fusion is performed to obtain the feature map V output by the feature extractor.

4. The method as described in claim 1, characterized in that, In step 3, labeled source domain data is used to supervise the training of the Softmax classifier, where the classification supervision loss of Softmax is: in, Represents source domain data, These represent the source domain sample set and the corresponding true classification label set, respectively. express The number of samples, This represents the i-th sample data in the source domain. express Real category tags, Represents cross-entropy loss, This represents the output of the feature extractor, i.e., the feature map extracted by the feature extractor. This indicates the probability that a sample belongs to a given category.

5. The method as described in claim 1, characterized in that, In step 3, the one-to-many classifier uses boundary loss during training, specifically: choose The most relevant nearest neighbor class is used as the negative sample class of the sub-classifier in each one-to-many classifier to construct open boundaries; The output of each sub-classifier is a two-dimensional vector representing the probability that a sample belongs to a known or unknown class. This represents the true category label of the source domain data. This represents the source domain samples output by the corresponding subclassifier. It is the first The probability of a class is then used to construct the boundary loss based on the open boundary of the dual classifier: in, These represent the source domain sample set and the corresponding true classification label set, respectively. It belongs to a subset of This represents the source domain samples output by the corresponding subclassifier. It is the first The probability of a class.

6. The method as described in claim 1, characterized in that, In step 3, the total loss of the radiation source individual identification model during training is: in, This represents the classification supervision loss of Softmax. This represents the boundary loss of a one-to-many classifier. This represents maximizing the class distance loss. These represent the source domain sample set and the corresponding true classification label set, respectively. Let represent the target domain set, and let represent the balance coefficient for the maximum spacing loss; in, The degree of confusion between category i and any other category j , Let W represent the probability of predicting class i in a training batch, where W is a diagonal matrix. express The diagonal elements corresponding to each sample, the diagonal elements of matrix W The classifier's predictive certainty regarding the class of the k-th sample is expressed as follows: Where B represents the size of a batch of training data during target domain training. This represents the probability that the k-th sample in the target domain output by the Softmax classifier is classified into the j-th class. The target domain output by the Softmax classifier represents the first... The probability that a sample is classified into the j-th class.

Images