Method and system for unknown radio source individual identification based on semi-supervised contrastive learning

Abstract

The application discloses a method and system for unknown radiation source individual identification based on semi-supervised contrast learning, and belongs to the technical field of radiation source individual identification. The method comprises the following steps: obtaining a first sample set and a second sample set; the first sample set contains first radiation source signal samples with labels; the second sample set contains second radiation source signal samples without labels; in a unified latent space, assigning known prototype vectors to each known category corresponding to the labels in the first sample set, and assigning a preset number of unknown prototype vectors; training a feature extraction network through the first sample set and the second sample set; extracting a feature vector of a to-be-identified radiation source signal by using the trained feature extraction network, and determining the category to which the to-be-identified radiation source signal belongs according to the similarity between the feature vector and each prototype vector. The above scheme can improve the known category identification accuracy while realizing the identification and differentiation of unknown new category radiation source individuals.

Description

Technical Field

[0001] This application relates to the field of radiation source individual identification technology, and more specifically, to a method and system for identifying unknown radiation source individuals based on semi-supervised contrastive learning. Background Technology

[0002] Specific Emitter Identification (SEI) utilizes the hardware fingerprint differences between different radiation sources to extract features from physical layer signals, enabling the identification and monitoring of frequency-using devices. With the widespread adoption of wireless communication and IoT devices, the electromagnetic spectrum is becoming increasingly complex, placing higher demands on the identification of legitimate users and the detection of illegitimate users.

[0003] In related technologies, most SEI methods based on deep learning rely on the closed-set assumption, which assumes that all test samples belong to categories already present in the training set. This assumption does not align with real-world open electromagnetic environments, where new and unknown radiation sources constantly emerge. Especially when only limited labeled data is available, coupled with a large amount of unlabeled data and unknown signal categories, existing methods cannot simultaneously handle the accurate identification of known categories and the effective discovery of new, unknown categories. Summary of the Invention

[0004] Therefore, it is necessary to provide a method and system for identifying unknown radiation source individuals based on semi-supervised contrastive learning to address the above problems, improve the accuracy of identifying known categories of radiation source individuals, and realize the identification and classification of unknown new categories of radiation source individuals.

[0005] This application provides a method for identifying individuals of unknown radiation sources based on semi-supervised contrastive learning, including: Acquire a first sample set and a second sample set; the first sample set contains labeled first radiation source signal samples; the second sample set contains unlabeled second radiation source signal samples. In a unified latent space, known prototype vectors are assigned to each known category corresponding to the label in the first sample set, and a preset number of unknown prototype vectors are assigned to represent the potential unknown categories in the second sample set. The feature extraction network is trained using the first and second sample sets. The feature vector of the radiation source signal to be identified is extracted using the trained feature extraction network. The category of the radiation source signal to be identified is determined based on the similarity between the feature vector and each prototype vector. During the training of the feature extraction network, the model parameters and known prototype vectors of the feature extraction network are updated based on the supervised learning results of the first radiation source signal samples; the model parameters and unknown prototype vectors of the feature extraction network are updated based on the unsupervised learning results corresponding to the second sample set.

[0006] In one alternative implementation, the feature extraction network is trained using a first sample set and a second sample set, including: Based on the processing results of the feature extraction network on the first radiation source signal sample and the label of the first radiation source signal sample, the supervised loss function value is obtained; Based on the processing results of the feature extraction network on multiple second radiation source signal samples, the unsupervised loss function value is obtained; The target loss function value is obtained based on the supervised loss function value and the unsupervised loss function value. The parameters of the feature extraction network are updated based on the target loss function value.

[0007] In one optional implementation, a supervised loss function value is obtained based on the processing result of the feature extraction network on the first radiation source signal sample and the label of the first radiation source signal sample, including: The first feature vector is obtained by processing the first radiation source signal sample through a feature extraction network. The first prototype vector is determined based on the similarity between the first feature vector and each known prototype vector; The supervised loss function value is obtained based on the category corresponding to the first prototype vector and the label of the first radiation source signal sample.

[0008] In one optional implementation, the supervised loss function value is obtained based on the category corresponding to the first prototype vector and the label of the first radiation source signal sample, including: The first loss function value is obtained based on the category corresponding to the first prototype vector and the label of the first radiation source signal sample; The second loss function value is constructed based on the similarity between the feature vectors corresponding to samples with the same label in multiple first radiation source signal samples, and the difference between the feature vectors corresponding to samples with different labels. The supervised loss function value is obtained based on the first loss function value and the second loss function value.

[0009] In one optional implementation, the unsupervised loss function value is obtained based on the processing results of the feature extraction network on multiple second radiation source signal samples, including: The signal sample from the second radiation source is enhanced to obtain an enhanced signal; The second radiation source signal sample is processed by a feature extraction network to obtain the second feature vector; The enhanced signal is processed by a feature extraction network to obtain a third feature vector; The second prototype vector is determined based on the similarity between the second feature vector and the known prototype vector / unknown prototype vector; The third loss function value is obtained based on the difference between the second and third feature vectors; The unsupervised loss function value is obtained based on the third and fourth loss function values; the fourth loss function value is used to indicate the difference between the second feature vector and the prototype vectors corresponding to other samples in the same training batch.

[0010] In one optional implementation, before obtaining the unsupervised loss function value based on the third loss function value and the fourth loss function value, the method further includes: The second feature vector is processed by the Softmax function to obtain the first category prediction vector; The third feature vector is processed by the Softmax function to obtain the second category prediction vector; The fourth loss function value is obtained based on the difference between the first category prediction vector and the second category prediction vector.

[0011] In one optional implementation, obtaining the target loss function value based on the supervised loss function value and the unsupervised loss function value includes: The target loss function value is obtained based on the supervised loss function value, the unsupervised loss function value, and the reconstruction loss function value. The reconstruction loss function is obtained based on the difference between the reconstructed signal sample and the original signal sample input to the feature extraction network; the reconstructed signal sample is obtained by reconstructing the feature vector output by the feature extraction network through the decoder.

[0012] In one alternative implementation, the method for identifying unknown radiation sources based on semi-supervised contrastive learning further includes: The prototype separation loss value is obtained based on the similarity between any two prototype vectors; Based on the prototype separation loss value, update all prototype vectors to reduce the similarity between different prototype vectors.

[0013] Furthermore, this application also provides a semi-supervised contrastive learning-based system for identifying individuals of unknown radiation sources, used to implement the aforementioned semi-supervised contrastive learning-based method for identifying individuals of unknown radiation sources. The system includes: The sample acquisition module is used to acquire a first sample set and a second sample set; the first sample set contains labeled first radiation source signal samples; the second sample set contains unlabeled second radiation source signal samples. The vector allocation module is used to allocate known prototype vectors to each known category corresponding to the label in the first sample set in a unified latent space, and to allocate a preset number of unknown prototype vectors to represent the potential unknown categories in the second sample set. The network training module is used to train the feature extraction network using the first sample set and the second sample set; The identification module is used to extract feature vectors of the radiation source signal to be identified using the trained feature extraction network, and to determine the category of the radiation source signal to be identified based on the similarity between the feature vectors and each prototype vector.

[0014] In another aspect, this application also provides an electronic device, comprising: The memory and processor are interconnected and communicate with each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the above-mentioned method for identifying unknown radiation sources based on semi-supervised contrastive learning.

[0015] Compared to existing technologies, the advantages of this application are: This application obtains a labeled first sample set and an unlabeled second sample set, assigns learnable prototype vectors to known categories and a predetermined number of unknown categories in a unified latent space, and jointly trains a feature extraction network using labeled and unlabeled data. The category is determined by the similarity between the feature vector and the prototype vector. This enables accurate identification of known categories and classification of unknown new categories simultaneously, even with only a small amount of labeled known class data. The feature extraction network in this application, through joint optimization of supervised and unsupervised losses, can simultaneously learn category-discriminating features from labeled data and the inherent structure of the data from unlabeled data. This application obtains a first feature vector through the feature extraction network, determines a first prototype vector based on its similarity to the known prototype vector, and then obtains a loss based on the first prototype vector and the label. This allows the known prototype vector to continuously converge towards the feature center of the corresponding category during training, thereby more accurately representing the feature distribution of each category and improving the accuracy of known category identification. Attached Figure Description

[0016] Figure 1 A flowchart of an embodiment of the present application for identifying unknown radiation sources based on semi-supervised contrastive learning; Figure 2 A schematic diagram illustrating the principle of an individual identification method for unknown radiation sources based on semi-supervised contrastive learning, provided in an embodiment of this application; Figure 3 This application provides an embodiment of an analysis chart showing the identification effect under different numbers of unknown new types of radiation sources. Figure 4 This application provides an embodiment of an analysis chart showing the identification effect under different numbers of known types of radiation sources. Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0018] In the field of Specific Emitter Identification (SEI), open electromagnetic space often contains both known and unknown categories of radiating sources. Typically, only a small amount of labeled data exists, while a large amount of unlabeled data contains signals from both known and unknown categories. Traditional deep learning-based SEI methods mostly rely on the closed-set assumption, which assumes all test samples belong to categories already present in the training set. This assumption fails to effectively address the emergence of unknown new categories. Furthermore, when labeled data is limited, the model is prone to overfitting and struggles to improve performance using massive amounts of unlabeled data.

[0019] Therefore, in order to solve the above problems, this application provides a method for identifying unknown radiation sources based on semi-supervised contrastive learning. Figure 1 A flowchart of an embodiment of the present application for identifying individuals of unknown radiation sources based on semi-supervised contrastive learning is shown. Figure 1 As shown, the method includes the following steps: S1. Obtain a first sample set and a second sample set; the first sample set contains labeled first radiation source signal samples; the second sample set contains unlabeled second radiation source signal samples.

[0020] In this embodiment, a receiver acquires radiation source signal data. Specifically, the receiver receives data from known radiation sources and data from unknown, unlabeled radiation sources.

[0021] Known class radiation source data refers to radiation source signals whose individual categories are known, such as transmitter signals of known models A, B, and C; unknown class unlabeled radiation source data refers to radiation source signals whose individual categories are unknown, such as new transmitter signals that have not yet appeared in the database.

[0022] Data augmentation operations are performed on the received data of known radiation sources and unlabeled radiation sources of unknown types to increase sample diversity and improve the model's generalization ability. In one optional implementation, data augmentation operations include, but are not limited to, time shifting and amplitude scaling.

[0023] Subsequently, the enhanced data is normalized. Radiation source signal samples typically contain mutually orthogonal in-phase components (I-path) and quadrature components (Q-path). In this embodiment, the I-path vector and Q-path vector of each sample are normalized independently, using the following normalization formula: ; In the formula, For a single sample, the normalized first-order vector of the I-way or Q-way vector is... One element, For a single sample, the first normalized value in the I-way or Q-way vector is... One element, and These represent the mean and standard deviation of all elements in the I-path or Q-path vector of the sample, respectively. By normalizing the I-path and Q-path vectors separately, the amplitude inconsistency caused by differences in the acquisition channels between the two signals can be eliminated, allowing the model to focus more on the intrinsic fingerprint characteristics of the signal.

[0024] After the above processing, a first sample set and a second sample set are obtained. The first sample set contains labeled first radiation source signal samples, which indicate the known category to which the samples belong. The second sample set contains unlabeled second radiation source signal samples, which may contain samples of known categories and samples of unknown categories that have not yet appeared.

[0025] S2. In a unified latent space, assign known prototype vectors to each known category corresponding to the label in the first sample set, and assign a preset number of unknown prototype vectors to represent the potential unknown categories in the second sample set.

[0026] In this embodiment, a unified latent space is first constructed to represent the features of all samples. This latent space refers to the space containing the feature vectors obtained after mapping the original radiation source signal through a feature extraction network (Encoder). Within the latent space, a learnable unit prototype vector is assigned to each of the known categories and a predetermined number of unknown categories, forming a prototype vector set.

[0027] Specifically, a learnable known prototype vector is assigned to each known category. The known categories are determined by the labels in the first sample set, and the total number of known categories is given by... Then allocate A known prototype vector.

[0028] At the same time, a number of unknown categories is pre-defined in the latent space. , An integer greater than or equal to 1. Allocation. A learnable unknown prototype vector is used to correspond to the unknown categories that may be included in the second sample set, in order to automatically discover and characterize the different potential unknown categories in the second sample set during training.

[0029] Let a single prototype vector be denoted as All prototype vectors (including known and unknown prototype vectors) constitute the prototype vector set. ,in: All prototype vectors are learnable parameters that are updated synchronously with the feature extraction network during training.

[0030] During training, this embodiment uses the von Mises-Fisher (vMF) distribution to model the relationship between feature vectors and prototype vectors, ensuring that the feature vectors output by the feature extraction network are aligned with the prototype vectors of their respective categories. Specifically, for any input sample... ( Indicates the first (one radiation source signal sample), after being processed by the feature extraction network Encoder, yields a feature vector. After L2 normalization, the eigenvectors Both the vectors and their prototype vectors follow a vMF distribution, in which the nth vector... The first category The probability density function for a sample is defined as: ; In the formula: For the first The vMF probability density function values ​​for each category; The lumped parameters of the vMF distribution ( ), used to control the degree to which feature vectors are clustered around the prototype vector; This is the normalization constant.

[0031] According to Bayes' theorem, the sample Corresponding features Category The formula for calculating the posterior probability is: ; In the formula: For the sample Category The posterior probability; For feature vectors With prototype vector The exponential value of the dot product scaled by the temperature coefficient; Temperature control coefficient ( (), used to adjust the smoothness of the probability distribution, The smaller the value, the sharper the probability distribution.

[0032] Therefore, the prediction vector of the final prediction probability can be obtained, and the calculation formula is: ; In the formula: For feature vectors The corresponding category probability distribution vector; This represents the probability of belonging to class 1. For belonging to the first The probability of a class; The total number of categories, ; It is the temperature control coefficient.

[0033] By placing the prototype vectors corresponding to known categories and a predetermined number of unknown categories in the same latent space, and using posterior probability formulas to establish the association between samples and each category, unified modeling of known and unknown categories can be achieved. Furthermore, to avoid the model mapping all samples to the same category, this embodiment also introduces an average entropy maximization regularization term: ; In the formula: This is the mean entropy regularization term; Let be the marginal probability distribution of the two eigenvectors. ; Corresponding features of the data-enhanced samples The predicted vector it represents; The th in the average distribution The probability of each category; This refers to the batch size.

[0034] In summary, this step achieves unified modeling of known categories and potential unknown categories by assigning learnable prototype vectors to known categories and a preset number of unknown categories in a unified latent space, and by using the vMF distribution and posterior probability formula to establish the association between sample features and prototype vectors.

[0035] S3. Train the feature extraction network using the first and second sample sets.

[0036] During the training of the feature extraction network, the model parameters and known prototype vectors of the feature extraction network are updated based on the supervised learning results of the first radiation source signal samples; the model parameters and unknown prototype vectors of the feature extraction network are updated based on the unsupervised learning results corresponding to the second sample set.

[0037] Specifically, the feature extraction network is trained using the first sample set and the second sample set, including: (1) Based on the processing results of the feature extraction network on the first radiation source signal sample and the label of the first radiation source signal sample, the supervised loss function value is obtained.

[0038] Specifically, the first radiation source signal sample is processed through a feature extraction network. Processing is performed to obtain the first feature vector. Then, based on the first eigenvector... With each known prototype vector ( The similarity (e.g., dot product) between samples is used to determine the prototype vector that is most similar to the sample, denoted as the first prototype vector. The category corresponding to the first prototype vector is the predicted category of the model for that sample.

[0039] Next, supervised loss function values ​​are obtained based on the category corresponding to the first prototype vector and the label of the first radiation source signal sample. The supervised loss function values ​​include the first loss function value and the second loss function value.

[0040] The first loss function value is the supervised learning loss, and in this embodiment, the cross-entropy loss function is used. Specifically, the first loss function value is obtained based on the category corresponding to the first prototype vector and the label of the first radiation source signal sample, denoted as . The calculation formula is as follows: ; In the formula: For a subset of labeled samples in a training batch, Represents cross-entropy loss, For the first One-hot encoded label vector for each sample, The category prediction vector is obtained by applying Softmax to the feature vector output by the feature extraction network.

[0041] The second loss function value is the supervised contrastive loss. Specifically, based on the similarity between the feature vectors corresponding to samples with the same label in multiple first radiation source signal samples, and the difference between the feature vectors corresponding to samples with different labels, the second loss function value is constructed, denoted as . The calculation formula is as follows: ; In the formula: For a subset of labeled samples in a training batch, express Middle and the first A set of samples of the same category. This represents the feature vector obtained after passing through the feature extraction network and the projector. This is the temperature control coefficient. For indicator functions (when) (1 if true, 0 otherwise). Supervised contrastive loss brings feature vectors of similar samples closer together in the latent space, while keeping feature vectors of different samples further apart.

[0042] The supervised loss function value is obtained based on the first loss function value and the second loss function value.

[0043] During training, based on the similarity between the first feature vector and the corresponding known prototype vector, the parameters of the feature extraction network and the known prototype vector are updated synchronously through backpropagation, so that the known prototype vector gradually approaches the feature center of its corresponding category.

[0044] (2) Based on the processing results of the feature extraction network on multiple second radiation source signal samples, obtain the unsupervised loss function value.

[0045] Specifically, the signal sample from the second radiation source is first enhanced to obtain an enhanced signal. In this embodiment, data enhancement includes, but is not limited to, time shifting and amplitude scaling operations. Let the original unlabeled sample be denoted as... Two different random enhancements were applied to it, resulting in two enhanced views. and .

[0046] The second radiation source signal samples were processed using a feature extraction network. Processing is performed to obtain the second feature vector. Enhanced signals through feature extraction networks and Processing is performed to obtain the third feature vector. and Based on the second eigenvector The similarity between the prototype vector and all prototype vectors (known prototype vectors / unknown prototype vectors) is used to determine the prototype vector that is most similar to it, which is denoted as the second prototype vector.

[0047] Then, based on the difference between the second and third feature vectors, the third loss function value is obtained. Based on the third and fourth loss function values, the unsupervised loss function value is obtained. The fourth loss function value indicates the difference between the second feature vector and the prototype vectors corresponding to other samples in the same training batch.

[0048] The third loss function is a self-supervised contrastive loss. Specifically, the value of the third loss function is obtained based on the difference between the second and third feature vectors, denoted as . The calculation formula is as follows: ; In the formula: This represents a batch during training that contains both unlabeled samples of known classes and unlabeled samples of unknown classes. and These are two feature vectors obtained from the same sample after different data augmentations, respectively, through a feature extraction network and a projection head. It is the temperature control coefficient.

[0049] The fourth loss function is the self-supervised learning loss, which ensures that different augmented views of the same original sample have a consistent class prediction tendency, denoted as... .

[0050] Specifically, the second feature vector is processed using the Softmax function. Processing is performed to obtain the first category prediction vector. The third eigenvector is processed using the Softmax function. and Processing is performed to obtain the second-class prediction vector. Then, based on the difference between the first-class predicted vector and the second-class predicted vector, the value of the fourth loss function is obtained. Fourth Loss Function The calculation formula is: ; In the formula: This represents the cross-entropy loss.

[0051] During training, based on the similarity between the second feature vector and the second prototype vector, the parameters of the feature extraction network and the second prototype vector are updated synchronously through backpropagation, so that the unknown prototype vector gradually approaches the feature center of its corresponding unknown category sample. For known category samples that may exist in the unlabeled samples, since the supervised learning branch has already updated the known prototype vector, unsupervised learning usually only updates the unknown prototype vector to avoid interference.

[0052] (3) Obtain the target loss function value based on the supervised loss function value and the unsupervised loss function value; update the parameters of the feature extraction network based on the target loss function value. In addition, other auxiliary losses can be introduced in this embodiment.

[0053] In one optional implementation, the target loss function value is obtained based on the supervised loss function value, the unsupervised loss function value, and the reconstruction loss function value. The reconstruction loss function value is obtained based on the difference between the reconstructed signal sample and the original signal sample input to the feature extraction network; the reconstructed signal sample is obtained by reconstructing the feature vector output by the feature extraction network through a decoder. The calculation formula is: ; In the formula: For a batch The Middle One signal sample, For the first reconstructed by the decoder One signal sample.

[0054] In another alternative implementation, a prototype separation loss value is obtained based on the similarity between any two prototype vectors. Based on the prototype separation loss value, all prototype vectors are updated to reduce the similarity between different prototype vectors. Prototype separation loss The calculation formula is: ; In the formula: For temperature parameters, and This represents the different prototype vectors updated in each training round.

[0055] Finally, the total loss function This is obtained by weighted summation of the above losses. For example: ; In the formula: All are adjustable coefficients. This is the regularization term that maximizes the average entropy. Based on the target loss function value... The parameters of the feature extraction network and all prototype vectors are updated through gradient backpropagation.

[0056] S4. Use the trained feature extraction network to extract the feature vector of the radiation source signal to be identified, and determine the category of the radiation source signal to be identified based on the similarity between the feature vector and each prototype vector.

[0057] In this embodiment, after the model training is completed, for any radiation source signal to be identified... First, it is input into the trained feature extraction network. The corresponding feature vectors are obtained. And then perform L2 normalization on the feature vector.

[0058] Then, calculate the eigenvector and all prototype vectors. ( The similarity is calculated between feature vectors (including known and unknown prototype vectors). In one alternative implementation, the similarity is calculated using a dot product (inner product), with a larger value indicating a closer similarity between the feature vector and the prototype vector. The category to which the feature vector belongs can then be calculated using the posterior probability formula. The probability is calculated using the following formula: ; In the formula: For feature vectors Belongs to each category The probability of; This is the temperature control coefficient.

[0059] Finally, the category with the highest posterior probability is output as the recognition result. If the prototype vector corresponding to this category belongs to a known category ( If the corresponding prototype vector belongs to a predefined unknown category, then output the label of that known category; if the corresponding prototype vector belongs to a predefined unknown category ( If the unknown category X is output (X is the index of the prototype), then the unknown radiation source individual can be automatically discovered and finely classified.

[0060] Figure 2 This is a schematic diagram illustrating the principle of an individual identification method for unknown radiation sources based on semi-supervised contrastive learning, provided in an embodiment of the present invention. Figure 2 As shown, for labeled data containing only known classes, data augmentation is first performed. The augmented data is then input into the feature extractor to obtain feature vectors. and .in, Used to calculate supervised comparison loss; The predicted probability is obtained after Softmax processing. The class loss (i.e., supervised learning loss) is calculated by comparing the data with the true labels. For unlabeled data that contains both known and unknown classes, data augmentation is performed in the same way, and the data is input into the same feature extractor to obtain feature vectors. And the corresponding predicted probabilities. Unlabeled data is used in the calculation of self-supervised contrastive loss and self-supervised learning loss. Furthermore, the feature vector output by the feature extractor... It also interacts with learnable prototype vectors to compute prototype losses (including prototype separation loss and posterior probability constraints). Simultaneously, the decoder reconstructs the original signal from the feature space, computing the reconstruction loss. The weighted sum of these various losses (supervised contrastive loss, class loss, self-supervised contrastive loss, self-supervised learning loss, prototype loss, and reconstruction loss) is then used to jointly optimize the feature extractor and all prototype vectors via gradient backpropagation.

[0061] To delve into the performance of the designed model, this embodiment includes performance comparison experiments with different models and performance analysis experiments under different configuration conditions. The evaluation metrics are quantified using three recognition accuracy metrics: known recognition accuracy... Accuracy of unknown category recognition and overall category recognition accuracy The formula for calculating recognition accuracy is as follows: ; In the formula: , , and These represent the true labels, the model-predicted labels, the number of known classes, and the number of new classes, respectively. Represents an unlabeled set The number of samples in the sample; This represents the set of all possible permutations for matching predicted labels with true labels, with the optimal permutation determined using the Hungarian algorithm. Additionally, the set of unlabeled labels is also given. The clustering accuracy of the known classes and the new classes are denoted as follows: and To achieve a more comprehensive evaluation, this embodiment introduces Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI) as performance evaluation metrics.

[0062] Figure 3 This is an analysis chart of the identification effect under different numbers of unknown new types of radiation sources according to an embodiment of the present invention, with reference to... Figure 3 As shown, the unknown radiation source individual identification method based on semi-supervised contrastive learning designed in this invention fully utilizes all available data information. Supervised contrastive learning and self-supervised contrastive learning are applied to labeled and unlabeled data respectively to extract basic representations, and a classifier is jointly trained based on these data. To balance the identification performance of known and unknown new radiation source individuals, a unified prototype representation method is proposed, which models known and new categories together in the same feature space. Simultaneously, a prototype separation loss function is designed to improve the inter-class separation performance of the model. Considering the need to preserve key information of signal data during downsampling, a decoder module is introduced to reconstruct the signal from the feature space, supplemented by a reconstruction loss function constraint.

[0063] Figure 4 This is an analysis chart of the identification effect under different numbers of known radiation source categories provided in an embodiment of the present invention, with reference to... Figure 4 As shown, the number of known radiation source individual categories is fixed at 10, and the accuracy rate of known category identification is... The accuracy of unknown category identification does not change with the increase in the number of unknown radiation source categories. and overall recognition accuracy There was no significant change when the number of unknown radiation source categories was less than or equal to 10, but a downward trend began when the number of unknown radiation source categories exceeded 10. With the number of unknown radiation source categories fixed at 5, when the number of known radiation source categories was greater than 5, the accuracy of all three categories did not change significantly with the number of known radiation source categories. When the number of known radiation source categories was less than or equal to 5, the accuracy of all three categories increased with the increase in the number of known radiation source categories. Experiments on the impact of the number of known and unknown radiation source categories showed that the accuracy of the three categories reached its best effect when the number of unknown and known radiation source categories was equal. However, when the number of unknown radiation source categories exceeded the number of known radiation source categories, the recognition rate of unknown categories and the overall recognition rate decreased significantly. This is because as the number of unknown categories increased, the scale of unlabeled data expanded, making it difficult for the model to fully extract the effective structural information of unknown categories in the unlabeled data under limited known class supervision information, thus blurring the decision boundary.

[0064] Referring to Table 1, this embodiment compares three methods: SimGCD with a trainable parameterized classification head and feature extractor, GCD with no trainable parameterized classification head but a trainable feature extractor, and HDBSCAN without training parameters. The four methods show little difference in accuracy for known classes, indicating that all methods achieve satisfactory performance in distinguishing known classes (i.e., traditional closed-set recognition tasks). However, SemiPC achieves the best performance in distinguishing unknown classes. Since GCD relies on a non-parametric classifier based on semi-supervised K-means, and HDBSCAN directly clusters the feature vectors extracted by the encoder, neither method fully utilizes the underlying features of the data, resulting in significantly weaker unknown class recognition performance compared to the proposed methods. Unlike simGCD, which exhibits performance bias between known and unknown classes, SemiPC alleviates this problem by jointly modeling known and unknown classes in a unified latent space, thus achieving higher unknown class accuracy than simGCD. Meanwhile, SemiPC performed best in terms of clustering quality, with a normalized mutual information (NMI) of 94.71% and an adjusted RAND index (ARI) of 92.58%. This verifies the ability of the proposed method to maintain intra-class compactness and inter-class separability across the entire dataset.

[0065] Table 1 Performance Comparison of Different Methods

[0066] Referring to Table 2, for a practical unknown radiation source identification system, real-time detection is crucial. The method proposed in this invention achieves a good balance between detection performance and model complexity. Although the SemiPC method does not have the lowest number of floating-point operations and parameters among all methods, its lightweight nature is still significantly better than SimGCD. Regarding inference efficiency, since GCD relies on an offline nonparametric K-means algorithm, its inference time is not included in the comparison. SemiPC's inference time is slightly higher than SimGCD due to the introduction of additional prototype normalization operations in the projection head. Nevertheless, in the open-scene SEI task, SemiPC achieves the optimal trade-off between computational overhead and performance.

[0067] Table 2 Comparison of Model Complexity for Different Methods

[0068] Furthermore, this application also provides a semi-supervised contrastive learning-based system for identifying individuals of unknown radiation sources, used to implement the aforementioned semi-supervised contrastive learning-based method for identifying individuals of unknown radiation sources. The system includes: The sample acquisition module is used to acquire a first sample set and a second sample set; the first sample set contains labeled first radiation source signal samples; the second sample set contains unlabeled second radiation source signal samples. The vector allocation module is used to allocate known prototype vectors to each known category corresponding to the label in the first sample set in a unified latent space, and to allocate a preset number of unknown prototype vectors to represent the potential unknown categories in the second sample set. The network training module is used to train the feature extraction network using the first sample set and the second sample set; The identification module is used to extract feature vectors of the radiation source signal to be identified using the trained feature extraction network, and to determine the category of the radiation source signal to be identified based on the similarity between the feature vectors and each prototype vector.

[0069] Reference Figure 5 As shown, Figure 5 This is a schematic diagram of the structure of an electronic device provided in an optional embodiment of the present invention. This electronic device can be a computer device used to execute the above-described method. Figure 5 As shown, the electronic device includes one or more processors 10, a memory 20, and interfaces for connecting the various components, including high-speed interfaces and low-speed interfaces. The various components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processor can process instructions executed within the electronic device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces).

[0070] The processor 10 may further include a hardware chip. This hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0071] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0072] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the use of the electronic device based on the display of a mini-program landing page. Furthermore, the memory 20 may include high-speed random access memory (RAM), and may also include non-transient memory, such as at least one disk storage device, flash memory device, or other non-transient solid-state storage device. The memory 20 may include volatile memory, such as RAM; the memory may also include non-volatile memory, such as flash memory, hard disk, or solid-state drive; the memory 20 may also include combinations of the above types of memory.

[0073] The electronic device also includes a communication interface 30 for communicating with other devices or communication networks.

[0074] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0075] 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.

[0076] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for identifying unknown radiation sources based on semi-supervised contrastive learning, characterized in that, The method includes: Acquire a first sample set and a second sample set; the first sample set contains labeled first radiation source signal samples; the second sample set contains unlabeled second radiation source signal samples. In a unified latent space, known prototype vectors are assigned to each known category corresponding to the label in the first sample set, and a preset number of unknown prototype vectors are assigned to represent the potential unknown categories in the second sample set. A feature extraction network is trained using the first sample set and the second sample set; The feature extraction network after training is used to extract the feature vector of the radiation source signal to be identified. Based on the similarity between the feature vector and each prototype vector, the category to which the radiation source signal to be identified belongs is determined. During the training process of the feature extraction network, the model parameters of the feature extraction network and the known prototype vector are updated based on the supervised learning results of the first radiation source signal samples; the model parameters of the feature extraction network and the unknown prototype vector are updated based on the unsupervised learning results corresponding to the second sample set.

2. The method for identifying unknown radiation sources based on semi-supervised contrastive learning according to claim 1, characterized in that, Using the first sample set and the second sample set, a feature extraction network is trained, including: Based on the processing results of the feature extraction network on the first radiation source signal sample and the label of the first radiation source signal sample, a supervised loss function value is obtained; Based on the processing results of the feature extraction network on multiple second radiation source signal samples, the unsupervised loss function value is obtained; Based on the supervised loss function value and the unsupervised loss function value, the target loss function value is obtained; The parameters of the feature extraction network are updated based on the target loss function value.

3. The method for identifying unknown radiation sources based on semi-supervised contrastive learning according to claim 2, characterized in that, The step of obtaining a supervised loss function value based on the processing result of the feature extraction network on the first radiation source signal sample and the label of the first radiation source signal sample includes: The first radiation source signal sample is processed by the feature extraction network to obtain a first feature vector; The first prototype vector is determined based on the similarity between the first feature vector and each known prototype vector; The supervised loss function value is obtained based on the category corresponding to the first prototype vector and the label of the first radiation source signal sample.

4. The method for identifying unknown radiation sources based on semi-supervised contrastive learning according to claim 3, characterized in that, The step of obtaining the supervised loss function value based on the category corresponding to the first prototype vector and the label of the first radiation source signal sample includes: The first loss function value is obtained based on the category corresponding to the first prototype vector and the label of the first radiation source signal sample; The second loss function value is constructed based on the similarity between the feature vectors corresponding to samples with the same label in the plurality of first radiation source signal samples, and the difference between the feature vectors corresponding to samples with different labels. The supervised loss function value is obtained based on the first loss function value and the second loss function value.

5. The method for identifying unknown radiation sources based on semi-supervised contrastive learning according to claim 2, characterized in that, The step of obtaining unsupervised loss function values ​​based on the processing results of the feature extraction network on multiple second radiation source signal samples includes: The signal sample from the second radiation source is enhanced to obtain an enhanced signal; The second radiation source signal sample is processed by the feature extraction network to obtain a second feature vector; The enhanced signal is processed by the feature extraction network to obtain a third feature vector; The second prototype vector is determined based on the similarity between the second feature vector and the known prototype vector / unknown prototype vector; The third loss function value is obtained based on the difference between the second feature vector and the third feature vector; The unsupervised loss function value is obtained based on the third loss function value and the fourth loss function value; the fourth loss function value is used to indicate the difference between the second feature vector and the prototype vector corresponding to other samples in the same training batch.

6. The method for identifying unknown radiation sources based on semi-supervised contrastive learning according to claim 5, characterized in that, Before obtaining the unsupervised loss function value based on the third loss function value and the fourth loss function value, the method further includes: The second feature vector is processed by the Softmax function to obtain the first category prediction vector; The third feature vector is processed by the Softmax function to obtain the second category prediction vector; The fourth loss function value is obtained based on the difference between the first category prediction vector and the second category prediction vector.

7. The method for identifying unknown radiation sources based on semi-supervised contrastive learning according to any one of claims 2 to 6, characterized in that, The step of obtaining the target loss function value based on the supervised loss function value and the unsupervised loss function value includes: The target loss function value is obtained based on the supervised loss function value, the unsupervised loss function value, and the reconstruction loss function value. The reconstruction loss function value is obtained based on the difference between the reconstructed signal sample and the original signal sample input to the feature extraction network; the reconstructed signal sample is obtained by reconstructing the feature vector output by the feature extraction network through the decoder.

8. The method for identifying unknown radiation sources based on semi-supervised contrastive learning according to any one of claims 2 to 6, characterized in that, The method further includes: The prototype separation loss value is obtained based on the similarity between any two prototype vectors; Based on the prototype separation loss value, all the prototype vectors are updated to reduce the similarity between different prototype vectors.

9. A system for identifying individuals of unknown radiation sources based on semi-supervised contrastive learning, used to implement the method for identifying individuals of unknown radiation sources based on semi-supervised contrastive learning as described in any one of claims 1 to 8, characterized in that, The system includes: A sample acquisition module is used to acquire a first sample set and a second sample set; the first sample set contains labeled first radiation source signal samples; the second sample set contains unlabeled second radiation source signal samples. The vector allocation module is used to allocate known prototype vectors to each known category corresponding to the label in the first sample set in a unified latent space, and to allocate a preset number of unknown prototype vectors to represent the potential unknown categories in the second sample set. The network training module is used to train the feature extraction network using the first sample set and the second sample set; The identification module is used to extract feature vectors of the radiation source signal to be identified using the trained feature extraction network, and to determine the category to which the radiation source signal to be identified belongs based on the similarity between the feature vectors and each prototype vector.

10. An electronic device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the individual identification method for unknown radiation sources based on semi-supervised contrastive learning as described in any one of claims 1 to 8.

Images