Open set emitter individual identification method based on adaptive threshold

Abstract

The application discloses an open set emitter individual identification method based on an adaptive threshold value, and first, based on a known emitter data set, a pre-training model is acquired through deep metric learning and is used as a feature extractor, the emitter data set is processed using the feature extractor, an emitter feature representation library is established, an initial dynamic threshold value and an adaptive threshold value adjustment function for adjusting the dynamic threshold value are calculated, the dynamic threshold value is adjusted in real time according to the adaptive threshold value adjustment function to obtain an adaptive threshold value, and whether a test set is a known emitter or an unknown emitter is judged according to the adaptive threshold value, and finally, according to the identification result, the emitter feature representation library and the dynamic threshold value are updated or newly added in real time, and the next round of identification is performed. The method of the application can adaptively adjust the real-time dynamic discrimination threshold value of each type of emitter, so that it is suitable for open set emitter individual identification under different scenes and unbalanced data, improves the accuracy of identification, enhances the robustness and adaptability, and meets the real-time demand of actual application scenes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wireless communication, and particularly relates to an open set radiation source individual identification method based on an adaptive threshold. BACKGROUND

[0002] Radiation source individual identification technology, also known as radiation source "fingerprint" identification or specific emitter identification (SEI), refers to measuring the characteristics of received electromagnetic signals and determining the individual radiation source that generates the signals according to existing prior information. Therefore, radiation source individual identification has important value in the fields of electronic jamming, device monitoring and tracking. However, in actual application scenarios, radiation sources are diverse and widely distributed. Closed set identification by existing prior information has poor openness, poor real-time performance and slow response speed to emergencies. Therefore, how to effectively identify the unrecorded radiation source individual in the open set is a problem to be solved.

[0003] In the open set radiation source individual identification technology, the selection of the threshold is crucial for the accuracy and robustness of the identification result. A proper threshold can effectively distinguish different radiation source individuals and reduce the probability of misidentification and missed identification. However, due to the complexity and diversity of radiation source signals, it is challenging to determine a suitable threshold. The relatively classic threshold values are experience threshold, fixed threshold, etc. Among them, the experience threshold is manually set based on expert experience and experimental data. The fixed threshold is based on known data sets and partial position data sets, and a statistically optimal threshold is obtained through probability statistics. In addition, machine learning can also be combined to automatically learn and determine the threshold.

[0004] However, the above threshold methods still have certain problems in actual application: (1) Some methods lack adaptability to the signal environment and may not achieve satisfactory identification results in complex and changing environments, requiring additional adjustments or training, thus lacking adaptability and robustness and being unable to perform real-time iterative updates; (2) Some methods have high computational complexity, which may require a large amount of training data and computing resources, affecting real-time performance and ease of use.

[0005] Patent CN 113705446 A discloses an open set radiation source individual identification method, which uses a deep neural network as a feature extractor, constructs an extreme value distribution model with the feature vectors obtained by training data, determines a discrimination threshold to realize a discrimination algorithm, and completes open set identification of the radiation source signal. Since the discrimination threshold used in this method is a fixed threshold, it cannot be updated and self-adaptively adjusted. In the face of complex and variable radiation source signal environment, it needs to be retrained and adjusted multiple times, and cannot adapt to different scenes and environments well. When the data volume is small, the obtained threshold has poor robustness, which affects the accuracy.

[0006] Patent CN 114692665 A discloses an open set radiation source individual identification method based on metric learning, which introduces deep metric learning for model training and feature extraction of the radiation source signal, and performs radiation source discrimination based on this. However, this method does not elaborate on the composition of the discrimination threshold, and does not use the effective feature vectors obtained by metric learning for secondary analysis to obtain a more reliable threshold and threshold adjustment method, lacking robustness and adaptability. SUMMARY

[0007] To solve the above technical problems, the present application provides an open set radiation source individual identification method based on adaptive threshold, which is used to solve the problems of poor threshold robustness, real-time performance and adaptability caused by complex, diverse and variable radiation source signal environment in open set radiation source individual identification.

[0008] The technical scheme of the present application is as follows: an open set radiation source individual identification method based on adaptive threshold, the specific steps are as follows:

[0009] S1, based on the known data set of the radiation source, obtaining a pre-trained model by deep metric learning and using it as a feature extractor;

[0010] S2, using the feature extractor to process the radiation source data set, establishing a radiation source feature representation library, including radiation source labels, feature vectors and feature vector centers;

[0011] S3, based on the radiation source feature representation library established in step S2, calculating an initial dynamic threshold and an adaptive threshold adjustment function for adjusting the dynamic threshold;

[0012] S4, adjusting the dynamic threshold in real time according to the adaptive threshold adjustment function to obtain an adaptive threshold, and judging whether the test set is a known radiation source or an unknown radiation source according to the adaptive threshold;

[0013] S5, updating or adding the radiation source feature representation library and the dynamic threshold in real time according to the identification result, and then performing the next round of identification.

[0014] Further, the step S1 is specifically as follows:

[0015] First, the radiation source signal is collected, and after uniform preprocessing of the collected original signal data, the radiation source IQ signal data is obtained, the preprocessing including: resampling, spectrum shift, low-pass filtering, sliding cutting, normalization;

[0016] Then a deep metric learning model is constructed, the structure of which includes: an input layer, a backbone network layer, a projection layer, and a classification layer. At the same time, a loss function L All is used to measure the difference between the model prediction and the real target to optimize the model parameters, as follows:

[0017] Input layer: the input size that can be received is set to (L, 1, 2);

[0018] Where L represents the signal length.

[0019] Backbone network layer: Resnet18 structure is used, and an m-dimensional feature vector is output through a global average pooling layer GAP;

[0020] Projection layer: multi-layer perceptron (MLP) structure is used, three layers of MLP, each layer has batch normalization (BN), the first two layers have rectified linear unit (ReLU), the output layer has no ReLU, and finally a feature vector of the number of categories is output;

[0021] Classification layer: the multi-classification function takes the Softmax function;

[0022] Loss function L All : L All is the weighted sum of Circle Loss (L Circle ) and CrossEntropy Loss (L CrossEntropy ), which is represented as follows:

[0023]

[0024]

[0025] Where R represents the total number of positive sample signals; K represents the total number of negative sample signals; i and j represent the i-th positive sample and the j-th negative sample, respectively; pt and nt represent the sample as a positive sample and a negative sample, respectively; γ, O pt , O nt , Δpt, and Δnt are hyperparameters, γ represents the proportion coefficient, O pt and O nt represent the similarity target values of positive samples and negative samples, respectively, and Δpt and Δnt represent the threshold values for adjusting positive samples and negative samples, respectively; and O pt , O nt , Δpt, and Δnt are adjusted by a hyperparameter margin, O pt= 1 + margin; O nt = -margin; Ap = 1 - margin; An = margin; respectively represent the similarity score of positive samples and negative samples, respectively represent the similarity comparison value of positive samples and negative samples, representing that when the similarity score of the positive sample is less than the target value of the positive sample, the calculated value is positive; otherwise, it is zero. representing that when the similarity score of the negative sample is greater than the target value of the negative sample, the calculated value is positive; otherwise, it is zero.

[0026]

[0027] where C represents the total number of radiation source signal categories, and c represents a category; y c represents an indicator variable, and if the category c is the same as the category of the sample, y c is 1, otherwise 0; p c represents the predicted probability that the observed sample belongs to the category c.

[0028] L All = μ1*L Circle + μ2*L CrossEntropy

[0029] where μ1 and μ2 are hyperparameters, respectively representing the weights of Circle Loss and CrossEntropy Loss.

[0030] Finally, according to the preprocessed radiation source IQ signal data set, the deep metric learning model is trained until convergence and the accuracy is optimal, and the model at this time is saved as a feature extractor.

[0031] Further, the step S2 is specifically as follows:

[0032] First, based on the preprocessed radiation source IQ signal data set, the corresponding feature vector

[0033] where, represents the feature vector of the i-th sample of the n-th category, x m represents the value of the m-th dimension of the feature vector, and T represents the transpose operation of the vector.

[0034] Then, according to the individual category label n of the radiation source, the feature vectors are respectively stored, and a radiation source feature representation library is established;

[0035] Finally, based on the feature vectors of each category, the center X_center of each category feature vector is calculated nLet X_center represent the nth type of radiation source, where there are q samples. n The calculation formula is as follows:

[0036]

[0037] in, This represents the feature vector corresponding to the i-th sample in the n-th class. The value in the m-th dimension.

[0038] Furthermore, step S3 is specifically as follows:

[0039] S31. Based on the radiation source feature representation library, calculate the distance between the center of each type of feature vector and its similar and dissimilar feature vectors, and obtain the similarity matrix S. same S diff ;

[0040] The distance is calculated using cosine distance, and the formula is as follows:

[0041]

[0042] Where A and B represent two feature vectors, A = (a1, a2, ..., a...). m ) T B = (b1, b2, ..., b m ) T .

[0043] Then the similarity matrix of the nth class The calculation formula is as follows:

[0044]

[0045] Then the heterosimilarity of the nth class The matrix calculation formula is as follows:

[0046]

[0047] Among them, X_center n Represents the eigenvector center of the nth class. Represents the nth class and the Numth... same 1 type of feature vector, Represents the i-th class and the Num-th class. diff N out-of-class feature vectors, Num same Num diff These represent the number of test samples of the same type and the number of test samples of different types, respectively.

[0048] S32, set initial threshold α and precision f according to the similarity matrix of the nth class, step number r, calculate the resolution ACC of the nth class under different threshold k by gradually increasing the threshold k = α + r * f, r = 1, 2,... n , and locate the current best dynamic threshold of the nth class according to the best resolution

[0049] Resolution ACC n The calculation formula is as follows:

[0050] ACC n

[0051] Wherein, The i-th cosine distance of , Num represents the number of correct answers, The i-th cosine distance of , Num represents the number of correct answers, Num is increased by one if the different class recognition is correct, Num is increased by one if the same class recognition is correct, and Num same , Num diff respectively represent the number of test samples of the same class and different class.

[0052] S33, set initial feature vector usage ζ0, precision σ, and extraction times β, gradually increase the feature vector usage ζ = ζ0+ r * σ by r, r = 1, 2,..., that is, extract ζ number of feature vectors from each class in the feature representation library β times, and based on step S2, calculate the feature vector center corresponding to different feature vector usage ζ , and calculate the same class similarity matrix and different class similarity matrix at this time based on step S31;

[0053] S34, based on step S32, calculate the best dynamic threshold of the nth class corresponding to different feature vector usage ζ Until the dynamic threshold tends to be stable, the stable dynamic threshold of the nth class is Calculate the average change of the threshold of all N classes in the feature representation library Finally, fit ζ and to obtain the adaptive threshold adjustment function ΔT of the feature vector usage of the radiation source A = f(ζ);

[0054] Wherein, the fitting method of ζ and is adjusted according to the actual scene to obtain the adaptive threshold adjustment function corresponding to the feature vector usage of the radiation source.

[0055] Further, the step S4 is specifically as follows:

[0056] First, the feature vector X corresponding to the test radiation source signal is extracted using a feature extraction model. test ;

[0057] Calculate the eigenvector X test The maximum similarity S is obtained by measuring the distance between the feature vector centers of all N categories in the feature representation library and the feature vector centers of all categories. max And the corresponding category η, the formula is:

[0058] S max =Max((cos(θ)|X_center) n X test (n = 1, 2, ..., N)

[0059] Among them, X_center n Let cos(θ)|X_center represent the feature vector center of the nth class. n X test X represents test With X_center n The cosine distance between them, where N represents the number of categories in the current feature representation library, which has a total of N categories.

[0060] Then, based on the dynamic threshold of η for that category... And the adaptive threshold adjustment function f(ζ) is used to obtain the adaptive threshold for category η. The calculation formula is as follows:

[0061]

[0062] in, This represents the current adaptive threshold for the ηth class. Let f(ζ) represent the current dynamic threshold of the ηth class, and let f(ζ) represent the value of the adaptive threshold adjustment function when the feature vector usage is ζ.

[0063] Finally, based on the maximum similarity S max Adaptive threshold for the corresponding category The system performs a discrimination test. If the similarity is greater than the value of the test radiation source data, it indicates that the test radiation source data is a known radiation source and the category is the category corresponding to the maximum similarity. If the similarity is less than the value of the test radiation source data, it indicates that the test radiation source data is an unknown radiation source.

[0064] Furthermore, step S5 is specifically as follows:

[0065] If the radiation source in the test set is a known radiation source, its feature vector is added to the corresponding category in the feature representation library, and the corresponding dynamic threshold is updated.

[0066] If the radiation source is unknown, a new class is added to the feature representation library, and a corresponding dynamic threshold for that class is added.

[0067] Repeat steps S3-S5 until all radiation source data in the test set has been processed, completing the identification of individual open-set radiation sources.

[0068] The beneficial effects of this invention are as follows: The method of this invention first uses a known dataset of radiation sources to obtain a pre-trained model through deep metric learning and uses it as a feature extractor. The feature extractor processes the radiation source dataset to establish a radiation source feature representation library. Then, an initial dynamic threshold and an adaptive threshold adjustment function are calculated. Based on the adaptive threshold adjustment function, the dynamic threshold is adjusted in real time to obtain an adaptive threshold, which is used to determine whether the test set contains known or unknown radiation sources. Finally, based on the recognition results, the radiation source feature representation library and dynamic threshold are updated or added in real time before the next round of recognition. This method can adaptively adjust the real-time dynamic discrimination threshold for each type of radiation source, adapting it to the identification of individual open-set radiation sources under different scenarios and imbalanced data, improving recognition accuracy, enhancing robustness and adaptability, while meeting the real-time requirements of practical application scenarios. Attached Figure Description

[0069] Figure 1 This is a flowchart of an open-set radiation source individual identification method based on adaptive threshold according to the present invention.

[0070] Figure 2 This is a diagram of the architecture of the deep metric learning model used in the embodiments of the present invention.

[0071] Figure 3 This is a flowchart illustrating the discrimination process based on an adaptive threshold in an embodiment of the present invention. Detailed Implementation

[0072] The invention will now be further described with reference to the accompanying drawings and embodiments.

[0073] like Figure 1 The flowchart of an open-set radiation source individual identification method based on adaptive thresholding according to the present invention is shown below. The specific steps are as follows:

[0074] S1. Based on the known dataset of radiation sources, a pre-trained model is obtained through deep metric learning and used as a feature extractor;

[0075] S2. Use a feature extractor to process the radiation source dataset and establish a radiation source feature representation library, including radiation source labels, feature vectors, and feature vector centers;

[0076] S3. Based on the radiation source feature representation library established in step S2, calculate the initial dynamic threshold and the adaptive threshold adjustment function used to adjust the dynamic threshold.

[0077] S4. Adjust the dynamic threshold in real time according to the adaptive threshold adjustment function to obtain the adaptive threshold, and use this to determine whether the test set is a known radiation source or an unknown radiation source.

[0078] S5. Based on the identification results, update or add radiation source feature representation library and dynamic threshold in real time, and then conduct the next round of identification.

[0079] In this embodiment, step S1 is specifically as follows:

[0080] First, radiation source signals are acquired. After uniform preprocessing of the acquired raw signal data, radiation source IQ signal data is obtained. The preprocessing includes: resampling, spectrum shifting, low-pass filtering, sliding clipping, and normalization.

[0081] Among these methods, spectrum shifting and low-pass filtering can remove interference signals and improve signal quality, while sliding cropping and normalization can divide the signal into smaller segments and adjust the signal size to a uniform range, which facilitates subsequent feature extraction and analysis and improves the performance of the recognition algorithm.

[0082] Then, a deep metric learning model is constructed, such as... Figure 2 As shown, its structure includes: an input layer, a backbone network layer, a projection layer, and a classification layer. Simultaneously, a loss function L is employed. All Used to measure the difference between model predictions and the actual target, in order to optimize model parameters, as follows:

[0083] Input layer: The acceptable input size is set to (L, 1, 2);

[0084] Where L represents the signal length.

[0085] Backbone network layer: adopts ResNet18 structure, and outputs m-dimensional feature vectors through global average pooling layer GAP;

[0086] Projection layer: A multilayer perceptron (MLP) structure is adopted, with three MLP layers. Each layer has batch normalization (BN). The first two layers have rectified linear units (ReLU), while the output layer does not have ReLU. The final output is a feature vector of the number of categories.

[0087] Classification layer: The softmax function is used for multi-class classification;

[0088] Loss function L All :L All For Circle Loss(L Circle ) and CrossEntropy Loss (L CrossEntropy The weighted sum of ) is expressed as follows:

[0089]

[0090]

[0091] where R represents the total number of positive sample signals; K represents the total number of negative sample signals; i, j represent the ith positive sample and the jth negative sample respectively; pt, nt represent that the sample is a positive sample and a negative sample respectively; γ, O pt nt , Δpt, Δnt are hyperparameters, γ represents a proportional coefficient, O pt nt represent the similarity target values of the positive sample and the negative sample respectively, and Δpt, Δnt represent the threshold values for adjusting the positive sample and the negative sample respectively; and O pt nt , Δpt, Δnt are adjusted by a hyperparameter margin, O pt＝ 1 + margin; O nt = - margin; Δpt = 1 - margin; Δnt = margin; represent the similarity scores of the positive sample and the negative sample respectively, represent the similarity comparison values of the positive sample and the negative sample respectively, represent that when the similarity score of the positive sample is less than the target value of the positive sample, the calculated value is positive; otherwise, it is zero; represent that when the similarity score of the negative sample is greater than the target value of the negative sample, the calculated value is positive; otherwise, it is zero.

[0092]

[0093] where C represents the total number of radiation source signal categories, and c represents a category; y c represents an indicator variable (0 or 1), and if the category c is the same as the category of the sample, y c is 1, otherwise it is 0; p c represents the predicted probability that the observed sample belongs to the category c.

[0094] L All = μ1 * L Circle + μ2 * L CrossEntropy

[0095] where μ1, μ2 are hyperparameters, and represent the weights of the Circle Loss and the CrossEntropy Loss respectively.

[0096] Finally, according to the preprocessed radiation source IQ signal data set, the deep metric learning model is trained until the convergence condition and the accuracy are optimal, and the model at this time is saved as a feature extractor.

[0097] ​​​In the embodiment, the step S2 is specifically as follows:

[0098] Firstly, based on the pre-processed radiation source IQ signal dataset, the corresponding feature vector is obtained by the feature extractor

[0099] wherein, represents the feature vector of the i-th sample of the n-th class, x m represents the value of the m-th dimension of the feature vector, and T represents the transposition operation of the vector.

[0100] Then, the feature vectors are respectively stored according to the individual class label n of the radiation source, and a radiation source feature representation library is established.

[0101] Finally, based on the feature vector of each class, the feature vector center X_center of each class is calculated n to represent the radiation source of this class, and the feature vector center X_center of the n-th class (q samples in total) at this time n The calculation formula is as follows:

[0102]

[0103] wherein, represents the feature vector of the i-th sample of the n-th class The value of the m-th dimension.

[0104] In the embodiment, the step S3 is specifically as follows:

[0105] S31, based on the radiation source feature representation library, the distance between the feature vector center of each class and the feature vectors of the same class and different classes is calculated, and a similarity matrix S is obtained same , S diff ;

[0106] The distance calculation adopts the cosine distance, and the calculation formula is as follows:

[0107]

[0108] wherein, A and B represent two feature vectors respectively, A=(a1, a2,..., a m ) T , B=(b1, b 22 ,..., b m ) T .

[0109] The same class similarity matrix of the n-th class The calculation formula is as follows:

[0110]

[0111] The different class similarity matrix of the n-th class The matrix calculation is as follows:

[0112]

[0113] wherein X_center n represents the eigenvector center of the nth class, represents the Num same th eigenvector of the same class of the nth class, represents the Num diff th eigenvector of the different class of the ith class, Num same , Num diff respectively represent the number of test samples of the same class and the different class.

[0114] S32, according to the similarity matrix of the nth class, set the initial threshold value α and the fineness f respectively, the step number is r, calculate the resolution ACC n of the nth class under different threshold values k by gradually increasing the threshold value k = α + r * f, r = 1, 2,..., and locate the current best dynamic threshold value of the nth class according to the best resolution

[0115] The resolution ACC n is calculated as follows:

[0116] ACC n

[0117] wherein, represents the ith cosine distance of , Num represents the ith cosine distance of , Num represents the correct number, Num is increased by one if the different class recognition is correct, Num is increased by one if the same class recognition is correct; Num same , Num diff respectively represent the number of test samples of the same class and the different class.

[0118] S33, set the initial eigenvector usage ζ0 (generally ζ0 = 1), the fineness σ, and the extraction times β, gradually increase the eigenvector usage ζ = ζ0 + r * σ by r, r = 1, 2,..., that is, extract ζ number of eigenvectors from each class in the feature representation library β times randomly, based on step S2, calculate the eigenvector center corresponding to different eigenvector usage ζ multiple times , and calculate the similarity matrix of the same class and the similarity matrix of the different class at this time based on step S31;

[0119] S34, based on step S32, calculate the best dynamic threshold value of the nth class corresponding to different eigenvector usage ζ T DDynamic Threshold, until the dynamic threshold is stable, the stable dynamic threshold of the nth class is The threshold average change of all N classes in the feature representation library is calculated In this embodiment, finally, the adaptive threshold adjustment function of the radiation source feature vector is obtained by polynomial curve fitting

[0120] Where M represents the highest degree of the polynomial, ζ j represents the jth power of ζ, ω j represents the coefficient of ζ j .

[0121] In this embodiment, the step S4 is specifically as follows:

[0122] As shown in Figure 3 , the specific process of the judgment test set is as follows:

[0123] First, the feature extraction model is used to extract the feature vector X test corresponding to the test radiation source signal.

[0124] The distance between the feature vector X test and the feature vector center of all N classes in the feature representation library is calculated to obtain the maximum similarity S max and the corresponding class η, the formula is:

[0125] S max = Max((cos(θ)|X_center n , X test ), n = 1, 2,... N)

[0126] Where X_center n represents the feature vector center of the nth class, cos(θ)|X_center n , X test represents the cosine distance between X test and X_center n , and N represents the number of classes in the current feature representation library, a total of N classes.

[0127] Then, according to the dynamic threshold of the class η and the adaptive threshold adjustment function , the adaptive threshold of the class η is obtained T A represents the adaptive threshold Adaptive Threshold, and the calculation formula is as follows:

[0128] ​

[0129] wherein, denotes the current adaptive threshold of the η-th class, denotes the current dynamic threshold of the η-th class, and f(z) denotes the value of the adaptive threshold adjustment function corresponding to the feature vector with the amount z.

[0130] Finally, based on the maximum similarity S max and the adaptive threshold of the corresponding class is greater than, it is indicated that the test radiation source data is a known radiation source, and the class is the class corresponding to the maximum similarity, and if it is less than, it is indicated that the test radiation source data is an unknown radiation source.

[0131] In the embodiment, the step S5 is specifically as follows:

[0132] If the radiation source in the test set is a known radiation source, the feature vector thereof is added to the corresponding class in the feature representation library, and the corresponding dynamic threshold is updated;

[0133] If it is an unknown radiation source, a new class is added in the feature representation library, and the corresponding dynamic threshold of the class is added;

[0134] The steps S3-S5 are repeated until all the radiation source data in the test set are processed, and the open set radiation source individual identification is completed.

[0135] In the embodiment, the method of the application is further illustrated by simulation. In the embodiment, the data used in the simulation experiment is collected from 16 different frequency modulation stations. The signals are transmitted by each frequency modulation station, and the USRP 2930 device is used to capture the signals. After the signal collection, pretreatment operations (spectrum shift, low-pass filtering and resampling) are performed on the signals to remove interference signals and improve the signal quality. Next, the processed signals are subjected to sliding cutting and normalization. The sliding cutting operation can divide the signals into smaller segments to facilitate subsequent feature extraction and analysis. The normalization operation adjusts the size of the signals to a unified range, which helps to improve the performance of the identification algorithm. After the above steps, an available data set with data and source labels is obtained. At this time, each class of radiation source has 3000 IQ signal samples, and the sample length is 8192. 13 classes are selected from the 16 classes of radiation sources as the training set, the validation set and the test set (6:2:2), and the remaining 3 classes are used as the open set test set.

[0136] In the simulation experiment in the embodiment, the Circle Loss-based deep metric learning model is trained using the training set, and the training effect reaches the optimum, and the accuracy of the validation set is 99.97%. The model is used as a feature extractor, and the feature vectors of the radiation sources in the training set are extracted The feature vector representing the i-th signal of the n-th type of radiation source has a dimension of m=32. The feature vectors and corresponding radiation source labels are stored to establish a feature representation library, and a feature vector center X_center is calculated n A universal expression for representing the radiation source category.

[0137] Based on the feature representation library, the simulation experiment of the embodiment calculates the intra-class similarity and inter-class similarity matrices of each class of the feature representation library. According to the similarity matrix of the n-th class, an initial threshold α=0 and a precision f=1000 are set, and the current stable dynamic threshold of each class is obtained In addition, an initial feature vector usage ζ0=1, a precision σ=1000, and an extraction frequency β=10 are set, and an adaptive threshold adjustment function f(ζ) that is strongly related to the feature vector usage ζ is calculated. The final adaptive threshold is Num(n) represents the number of feature vectors in the current n-th class library.

[0138] Finally, the simulation experiment results of the embodiment are divided into two parts:

[0139] First, the resolution, that is, the input test set, but the feature representation library and the threshold are not updated, only the judgment of whether it is a known radiation source signal is determined, and finally the closed set resolution is 99.26% and the open set resolution is 95.23%.

[0140] Second, the recognition accuracy, that is, the open set test set is input one by one, the threshold is determined, and it is determined whether it is an unknown radiation source. After the determination, the feature representation library and the corresponding dynamic threshold are updated, and the adaptive threshold adjustment function is used to adjust the dynamic threshold in real time to obtain the adaptive threshold. The threshold and the feature representation library are updated until the test set recognition is completed. Finally, the closed set recognition accuracy is 98.61%, and the open set recognition accuracy is 95.74%.

[0141] In summary, the simulation experiment in the above embodiment shows that based on the adaptive threshold, the method of the present application can adaptively adjust the real-time dynamic discrimination threshold of each type of radiation source to adapt to the open set radiation source individual recognition under different scenarios and unbalanced data, complete the open set radiation source individual recognition, and has certain recognition accuracy, robustness, and adaptability. The method is effective and feasible. At the same time, under the condition of small amount of data, the method of the present application can still maintain good open set radiation source individual recognition performance, and has strong practical value. The method of the present application uses deep metric learning, which aims to make the feature vectors of the same class of radiation sources as close as possible, and the feature vectors of different categories of radiation sources as far apart as possible. Therefore, a highly abstract and high-discrimination feature representation can be obtained, which helps to improve the discrimination ability of the adaptive threshold method in open set radiation source individual recognition, thereby increasing the recognition accuracy. In addition, it can map the original high-dimensional data to a low-dimensional space, thereby reducing the computational complexity and storage requirements.

[0142] Those skilled in the art will appreciate that the foregoing examples are presented for the purpose of aiding the reader in understanding the principles of the present application and should not be construed as limiting the scope of the present application to such specifically enumerated examples and embodiments. Modifications or applications thereof, which are within the scope of the present application, will occur to those skilled in the art upon reading the description of the application.

Claims

1. A method for identifying individuals from open-set radiation sources based on adaptive thresholds, comprising the following steps: S1. Based on the known dataset of radiation sources, a pre-trained model is obtained through deep metric learning and used as a feature extractor; The specific steps of S1 are as follows: First, radiation source signals are acquired. The raw signal data is then preprocessed to obtain the radiation source IQ signal dataset. The preprocessing includes: Resampling, spectrum shifting, low-pass filtering, sliding clipping, normalization; Then, a deep metric learning model is constructed, whose structure includes: an input layer, a backbone network layer, a projection layer, and a classification layer; simultaneously, a loss function is employed. Used to measure the difference between model predictions and the actual target, in order to optimize model parameters; specifically as follows: Input layer: The acceptable input size is set to (L, 1, 2); Where L represents the signal length; Backbone network layer: adopts ResNet18 structure, and outputs m-dimensional feature vectors through global average pooling layer GAP; Projection layer: adopts a multilayer perceptron (MLP) structure, with three MLP layers. Each layer has batch normalization (BN), the first two layers have rectified linear units (ReLU), and the output layer does not have ReLU. The final output is a feature vector of the number of categories. Classification layer: The softmax function is used for multi-class classification; loss function : For Circle Loss With CrossEntropy Loss The weighted sum of is expressed as follows: ； ； in, This represents the total number of positive sample signals; This represents the total number of negative sample signals; , They represent the first The positive sample, the first There are one negative sample; pt and nt represent that the sample is a positive sample and a negative sample, respectively; , , , , For superparameters, Represents the proportionality coefficient. , These represent the target similarity values ​​for positive and negative samples, respectively. , These represent adjusting the thresholds for positive and negative samples, respectively; and , , , It is adjusted by a hyperparameter margin. ; ; ; ; , These represent the similarity scores for positive and negative samples, respectively. , These represent the similarity comparison values ​​for positive and negative samples, respectively. This means that the calculated value is positive when the similarity score of a positive sample is less than the target similarity value of the positive sample; otherwise, it is zero. This means that the calculated value is positive when the similarity score of a negative sample is greater than the target similarity value of the negative sample; otherwise, it is zero. ； Where C represents the total number of radiation source signal categories, and c represents the category; y c This indicates that y is an indicator variable; if the category c and the sample category are the same, then y... c p is 1 if it is positive, otherwise it is 0; c This represents the predicted probability that the observed sample belongs to category c; ； in, , These are hyperparameters, representing the weights of Circle Loss and CrossEntropy Loss, respectively. Finally, based on the preprocessed radiation source IQ signal dataset, the deep metric learning model is trained until convergence and the accuracy is optimal, and the model at this point is saved as a feature extractor. S2. Use a feature extractor to process the radiation source dataset and establish a radiation source feature representation library, including radiation source labels, feature vectors, and feature vector centers; Step S2 is as follows: First, based on the preprocessed IQ signal dataset from the radiation source, the corresponding feature vectors are obtained through a feature extractor. ; in, Let x represent the feature vector of the i-th sample in the n-th class. m Let represent the value of the m-th dimension of the eigenvector, and T represent the transpose operation of the vector; Then, according to the individual category label n of the radiation source, the feature vectors are stored respectively to establish a radiation source feature representation library; Finally, based on the feature vectors of each class, calculate the feature vector center X_center for each class. n Let X_center represent the nth type of radiation source, where there are q samples. n The calculation formula is as follows: ； in, This represents the feature vector corresponding to the i-th sample in the n-th class. The value in the m-th dimension; S3. Based on the radiation source feature representation library established in step S2, calculate the initial dynamic threshold and the adaptive threshold adjustment function used to adjust the dynamic threshold. Step S3 is as follows: S31. Based on the radiation source feature representation library, calculate the distance between the center of each type of feature vector and its similar and dissimilar feature vectors, and obtain the similarity matrix S. same S diff ; The distance is calculated using cosine distance, and the formula is as follows: ； Where A and B represent two feature vectors. , ; Then the similarity matrix of the nth class The calculation formula is as follows: ； Then the heterosimilarity matrix of the nth class The calculation formula is as follows: ； in, Represents the eigenvector center of the nth class. Represents the nth class and the first 1 type of feature vector, Indicates the i-th class and the A heterogeneous feature vector, , These represent the number of test samples of the same type and the number of test samples of different types, respectively. S32. Based on the similarity matrix of the nth class, set the initial threshold α and the precision f respectively, with the number of steps being... By gradually increasing the threshold Calculate the ACC resolution of the nth class under different thresholds k. n And locate the current optimal dynamic threshold for the nth class based on the optimal resolution. ; Resolution ACC n The calculation formula is as follows: ； in, express The i-th cosine distance, express The i-th cosine distance, where Num represents the correct number. If the outlier is correctly identified, Num is incremented by one. If the same type is correctly identified, Num is incremented by one. same Num diff These represent the number of test samples of the same type and the number of test samples of different types, respectively. S33. Set the initial feature vector usage. The precision σ and the number of extractions β are used to gradually increase the amount of feature vectors by r. That is, randomly selecting from each class in the feature representation library β times. Based on step S2, the quantity of feature vectors is calculated multiple times for different feature vector quantities. Corresponding eigenvector center And based on step S31, calculate the similarity matrix of the same class and the similarity matrix of different classes at this time; S34. Based on step S32, calculate the usage of different feature vectors. The corresponding optimal dynamic threshold for the nth class Until the dynamic threshold stabilizes, the optimal dynamic threshold for the nth class is... The average threshold change for all N categories in the feature representation library is calculated as follows: Ultimately passed and By fitting the data, an adaptive threshold adjustment function with strong correlation to the radiation source feature vector is obtained. ; in, and The fitting method is adjusted according to the actual scenario to obtain the corresponding radiation source feature vector using a strongly correlated adaptive threshold adjustment function; S4. Adjust the dynamic threshold in real time according to the adaptive threshold adjustment function to obtain the adaptive threshold, and use this to determine whether the test set is a known radiation source or an unknown radiation source. S5. Based on the identification results, update or add radiation source feature representation library and dynamic threshold in real time, and then conduct the next round of identification.

2. The method for individual identification of open-set radiation sources based on adaptive thresholds according to claim 1, characterized in that, Step S4 is as follows: First, the feature vector X corresponding to the test radiation source signal is extracted using a feature extraction model. test ; Calculate the eigenvector X test The maximum similarity S is obtained by measuring the distance between the feature vector centers of all N categories in the feature representation library and the feature vector centers of all categories. max And the corresponding category η, the formula is: ； Among them, X_center n Represents the eigenvector center of the nth class. X represents test With X_center n The cosine distance between them, where N represents the number of categories in the current feature representation library, which has a total of N categories; Then, based on the dynamic threshold of η for that category... and adaptive threshold adjustment function Obtain the adaptive threshold for category η The calculation formula is as follows: ； in, This represents the current adaptive threshold for the ηth class. This shows the current dynamic threshold for the ηth class. The amount of eigenvector usage is The value of the adaptive threshold adjustment function corresponding to the time; Finally, based on the maximum similarity S max Adaptive threshold for the corresponding category The system performs a discrimination test. If the similarity is greater than the value of the test radiation source data, it indicates that the test radiation source data is a known radiation source and the category is the category corresponding to the maximum similarity. If the similarity is less than the value of the test radiation source data, it indicates that the test radiation source data is an unknown radiation source.

3. The method for individual identification of open-set radiation sources based on adaptive thresholds according to claim 1, characterized in that, Step S5 is as follows: If the radiation source in the test set is a known radiation source, its feature vector is added to the corresponding category in the feature representation library, and the corresponding dynamic threshold is updated. If the radiation source is unknown, a new class is added to the feature representation library, and a corresponding dynamic threshold for that class is added. Repeat steps S3-S5 until all radiation source data in the test set has been processed, completing the identification of individual open-set radiation sources.

Images