Limited sample specific radiation source identification method based on Mixup enhancement

Abstract

The invention discloses a finite sample specific radiation source identification method based on Mixup enhancement, and relates to the technical field of radiation source identification. Comprising the following steps: collecting samples and preprocessing; in a training stage, a triple sample is constructed from a training set; whether Mixup data enhancement is carried out on each triple sample or not is selected according to a random strategy, and a mixed sample is generated; inputting the mixed sample into a CVNN-TMN triple network to generate a feature vector; calculating distance difference and back propagation loss between samples based on a triple loss function, and updating a model parameter optimization model; and mapping a test sample to the optimized embedding space, calculating a joint distance measure between the test sample and the training sample, and adopting a nearest neighbor classification strategy to identify the test sample as a radiation source category corresponding to the training sample with the minimum distance from the test sample. According to the method, the problems of insufficient training data, low feature distinction degree, weak model generalization ability and the like under the condition that the sample height is limited are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of radiation source identification, and in particular to a finite sample specific radiation source identification method based on Mixup enhancement. Background Art

[0002] In recent years, Specific Emitter Identification (SEI) technology, as a key means of physical-layer device authentication, has become increasingly important in fields such as the Internet of Things, smart security, aerospace, and 6G secure communications. SEI uniquely identifies different devices by analyzing subtle hardware fingerprints of radio signals, such as manufacturing process deviations of RF front-end components, oscillator frequency offsets, and amplifier nonlinearities. Because signal fingerprints are inherent in the hardware and difficult to forge, SEI is an effective solution for combating identity spoofing and man-in-the-middle attacks.

[0003] However, the current SEI method based on deep learning still faces a number of key technical challenges in actual deployment. First, the similarity of individual features of wireless devices is high, and RF fingerprint data is easily affected by channel changes, multipath fading, and heterogeneous environments. This makes it difficult for the model to extract stable and discriminative signal representations during the learning process. Secondly, deep learning models are highly dependent on the number of labeled samples. In SEI application scenarios, due to the large number of device types, scattered signal collection points, and the high cost of obtaining sample labels, the number of available samples is often very small. When the proportion of training samples is low, it is difficult for commonly used classification networks to construct an embedding space with compactness within the class and discreteness between class intervals in the high-dimensional feature space, resulting in a significant decrease in recognition performance.

[0004] Patent CN202310059098.7 proposes a small sample radiation source identification method based on a twin network. Although the data volume is enhanced through sample combination and signal domain transformation, its sample space expansion method is relatively fixed and lacks smooth modeling of data distribution; the contrast loss does not adequately optimize the discrimination of the embedding space, and the category boundaries are blurred; and this method relies on sample pair input, has low training efficiency, and is sensitive to the imbalance of positive and negative samples.

[0005] Faced with the above problems, how to design a limited-sample specific radiation source identification method that is more suitable for complex-valued signal processing, has a sample enhancement mechanism and high classification and recognition performance in an actual communication environment with highly restricted samples has become the research focus and difficulty of current RF fingerprint recognition technology. Summary of the Invention

[0006] In response to the shortcomings of the above-mentioned existing technologies, the present invention provides a finite sample specific radiation source identification method based on Mixup enhancement, which solves the three major technical problems of insufficient training data, low feature discrimination and weak model generalization ability in the identification of specific radiation sources under limited sample conditions in an actual communication environment with highly restricted samples. By integrating Mixup data enhancement with a complex neural network to construct a CVNN-TMN model, a high recognition accuracy of 89.1% is achieved with a training sample ratio of 5%.

[0007] The specific technical solutions are as follows:

[0008] A method for identifying specific radiation sources using a limited sample based on Mixup enhancement includes the following steps:

[0009] S1. Collect radio signal samples from K radiation sources and preprocess them; divide the preprocessed samples into a training set and a validation set;

[0010] S2. During the training phase, a triplet of samples (a, p, n) is constructed from the training set, where a is a baseline sample, p is a positive sample of the same category, and n is a negative sample of a different category. For each triplet of samples, a random strategy is used to determine whether to perform Mixup data augmentation. If the augmentation conditions are met, a mixed sample (a', p', n') is generated.

[0011] S3. Input the mixed sample into the CVNN-TMN triplet network, extract its embedded features through the network structure with shared weights, and generate feature vectors f(a'), f(p'), and f(n');

[0012] S4. Calculate the Euclidean distance difference between samples based on the triple loss function, and use the Mixup interpolation ratio λ a ,λ p ,λ n The triplet loss is weighted to optimize the clustering between similar samples and the discrimination between heterogeneous samples in the embedding space.

[0013] S5. In the recognition phase, the test sample is mapped to the optimized embedding space, a joint distance metric between the test sample and the training sample is calculated, and a nearest neighbor classification strategy is adopted to identify the test sample as the radiation source category corresponding to the training sample with the smallest distance to the test sample.

[0014] S4. Calculate the distance difference between samples based on the triplet loss function and use the Mixup interpolation ratio λ a ,λ p ,λ n The triplet loss is weighted to optimize the clustering between samples of the same class and the discrimination between samples of different classes in the embedding space.

[0015] S5. In the recognition phase, the test sample is mapped to the optimized embedding space, the joint metric distance between the test sample and the training sample is calculated, and the nearest neighbor classification strategy is adopted to identify the test sample as the radiation source category corresponding to the training sample with the smallest distance to the test sample.

[0016] Furthermore, in step S2 of the above scheme, the random strategy is to generate a floating point number between [0, 1] through a random function. If the floating point number is less than 0.5, Mixup data enhancement is performed on the triplet sample; otherwise, Mixup data enhancement is not performed.

[0017] The above solution further performs Mixup data enhancement on the triplet samples, including:

[0018]

[0019] Among them, a', p', n' are randomly selected other benchmark samples, positive samples and negative samples, λ a ,λ p ,λ n It is the Mixup coefficient sampled from Beta(α,α),α>0 distribution, which controls the interpolation weight of different samples. It is a newly generated interpolation sample, which is used to supplement the original triplet sample to participate in subsequent training.

[0020] Furthermore, the CVNN-TMN triplet network consists of three branches with shared weights, each with the same structure, which sequentially includes nine layers of CVConv1D, CVReLU activation function, CVBN batch normalization layer and MaxPooling1D pooling layer, and then outputs the embedded feature vector through the Flatten layer and the fully connected layer.

[0021] Furthermore, in step S4, after the triplet samples are enhanced by Mixup data, the corresponding triplet loss function is:

[0022]

[0023] in, is the triplet loss, f(a), f(p), f(n) are the feature vectors of the benchmark sample, positive sample and negative sample obtained by the model respectively, and denote the Euclidean distance metric between the reference sample and the positive sample, and between the reference sample and the negative sample, respectively. α is a positive hyperparameter used to control the minimum interval between positive and negative samples. [·] +Represents the ReLU operation, that is, when the loss is negative, it takes 0 to ensure that the triplet loss is always non-negative; the loss weighted calculation is:

[0024]

[0025] Among them, λ a ,λ p ,λ n It is the Mixup interpolation ratio, which is used to control the weight distribution of loss in back propagation.

[0026] The above scheme is further improved.

[0027] In the triple loss function, the gradient is back-propagated by the chain rule. The loss function uses the chain rule to calculate the feature vectors f(a), f(p), and f(n). The gradient is The times are:

[0028]

[0029] Furthermore, in step S4, if the triplet samples are not subjected to Mixup data enhancement, the original triplet is used for loss calculation and optimization.

[0030] The above scheme is further combined with the distance metric function sim(f i ,f j )for:

[0031] sim(f i ,f j )=β·euclidean_distance(f i ,f j )+(1-β)·cos_distance(f i ,f j ),

[0032] Where β is a weight factor between 0 and 1, euclidean_distance(f i ,f j )=||f i -f j ||2, cos_distance(f i ,f j )=1-(f i ·f j ) / (||f i ||2·||f j ||2), where cos_similarity∈[-1,1].

[0033] Furthermore, in step S5, all samples of the training set and the validation set are encoded through the CVNN-TMN triplet network to obtain the corresponding embedding vectors.

[0034]

[0035] Among them, ε train Represents the set of embedded vectors corresponding to the training set samples, including N train samples; ε val The set of embedding vectors corresponding to the validation set samples contains N val samples;

[0036] For each validation sample f(x m ), calculate its difference with all training samples ε train The joint distance sim(f(x m ),f(x k )), select the one with the minimum joint distance The training samples of are classified as the nearest neighbors, which are expressed as follows:

[0037]

[0038] Among them, k * Representation and verification sample f(x m ) the most recent training sample index, is the predicted class label, i.e., the class of the most similar training sample.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] The present invention performs Mixup data enhancement on triple samples through a random strategy, effectively expanding the finite sample data set and improving the model training effect by increasing sample diversity; at the same time, the random strategy can avoid excessive enhancement and balance the effect and risk of data enhancement; by combining the triplet loss to construct the CVNN-TMN ternary twin network, the clustering of similar samples and the discrimination of different samples are enhanced, and the embedding space structure is optimized; and the interpolation proportional correlation loss function generated by Mixup is weighted, so that in the back propagation process, the network parameters can be adjusted more accurately according to the sample enhancement situation, the clustering of similar samples and the discrimination of heterogeneous samples are enhanced, and the classification accuracy of the model is improved; the joint distance metric is combined with the nearest neighbor classification strategy for sample classification, which improves the recognition accuracy of test samples, and ultimately significantly improves the accuracy and reliability of specific radiation source identification of finite samples, and enhances the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the steps of the method of the present invention.

[0042] Figure 2 It is a model construction diagram of the solution of the present invention.

[0043] Figure 3 It is the visual contrast effect of distinguishability enhancement. DETAILED DESCRIPTION

[0044] The following is a further detailed description of the embodiments of the invention in conjunction with the accompanying drawings to make the objectives, technical solutions and technical effects of the invention more clearly presented.

[0045] like Figure 1 As shown, the present invention discloses a method for identifying specific radiation sources with limited samples based on Mixup enhancement, which specifically includes the following steps:

[0046] S1. Collect radio signal samples from K radiation sources and perform preprocessing. Divide the preprocessed samples into a training set and a validation set. Preprocessing involves standard sampling data processing, such as normalization, to improve data quality, eliminate noise, and standardize the data format. The training set and validation set ratio is set at 70%:30%.

[0047] S2. During the training phase, a triplet sample (a, p, n) is constructed from the training set, where a is the baseline sample, p is a positive sample of the same category, and n is a negative sample of a different category. For each triplet sample, a random strategy is used to determine whether to perform Mixup data augmentation. If the augmentation conditions are met, a mixed sample (a', p', n') is generated. Here, the random strategy is to generate a floating-point number between [0, 1] using a random function. If the floating-point number is less than 0.5, Mixup data augmentation is performed on the triplet sample; otherwise, Mixup data augmentation is not performed. That is, a random strategy is selected for each triplet sample, or each sampled sample, to determine whether to perform Mixup data augmentation. This avoids the negative effects of blind augmentation or missed augmentation opportunities, making the decision more objective. This allows the data sample to be expanded with a certain probability during training, increasing data diversity, while preventing over-augmentation that causes the model to learn incorrect feature relationships, thus balancing the effectiveness and risks of data augmentation.

[0048] S3. Input the mixed sample into the CVNN-TMN triplet network, extract its embedded features through the shared weight network structure, and generate feature vectors f(a'), f(p'), and f(n'). The CVNN-TMN triplet network consists of three branches with shared weights. Each branch has the same structure, sequentially containing nine layers of CVConv1D, CVReLU activation function, CVBN batch normalization layer, and MaxPooling 1D pooling layer. The embedded feature vector is then output through the Flatten layer and the fully connected layer.

[0049] S4. Calculate the distance difference between samples based on the triplet loss function and use the Mixup interpolation ratio λ a ,λ p ,λ n The triplet loss is weighted to optimize the clustering between samples of the same class and the discrimination between samples of different classes in the embedding space.

[0050] Here, the triplet samples are mixed up for data enhancement, including:

[0051]

[0052] Among them, a', p', n' are randomly selected other benchmark samples, positive samples and negative samples, λ a ,λ p ,λ n It is the Mixup coefficient sampled from Beta(α,α),α>0 distribution, which controls the interpolation weight of different samples. It is a newly generated interpolation sample, which is used to supplement the original triplet sample to participate in subsequent training.

[0053] In step S4, after the triplet samples are enhanced by Mixup data, the corresponding triplet loss function is:

[0054]

[0055] in, is the triplet loss, f(a), f(p), f(n) are the feature vectors of the benchmark sample, positive sample and negative sample obtained by the model respectively, and denote the Euclidean distance metric between the reference sample and the positive sample, and between the reference sample and the negative sample, respectively. α is a positive hyperparameter used to control the minimum interval between positive and negative samples. [·] + Represents the ReLU operation, that is, when the loss is negative, it takes 0 to ensure that the triplet loss is always non-negative; the loss weighted calculation is:

[0056]

[0057] Among them, λ a ,λ p ,λ n It is the Mixup interpolation ratio, which is used to control the weight distribution of loss in back propagation.

[0058] In the triple loss function, the gradient is back-propagated by the chain rule. The loss function uses the chain rule to calculate the feature vectors f(a), f(p), and f(n). The gradient is The times are:

[0059]

[0060] Step S4 calculates the distance difference between samples based on the triplet loss function to measure the similarity between the baseline sample and the positive and negative samples. After the triplet samples undergo Mixup data enhancement, the triplet loss is weighted in combination with the Mixup mixed interpolation ratio. The weight distribution of the loss in back propagation is adjusted according to the sample enhancement situation to strengthen the clustering of similar samples in the embedding space and the distinction between heterogeneous samples. If the triplet samples do not undergo Mixup data enhancement, the original triplet is used for loss calculation and optimization. In this way, the gradient is used to perform back propagation calculations through the chain rule, the network parameters are updated, and the embedding space is continuously optimized, so that the model can better distinguish different types of radiation sources and improve recognition accuracy.

[0061] In step S4, if the triplet sample is not subjected to Mixup data enhancement, the original triplet is used for loss calculation and optimization. To ensure that the model can still perform loss calculation and parameter optimization normally without Mixup data augmentation, the original triples are explicitly used for loss calculation and optimization. This allows the model to maintain a stable training process under different data augmentation strategies, enhancing the flexibility and robustness of the model training method.

[0062] S5. In the recognition phase, the test sample is mapped to the optimized embedding space, a joint distance metric between the test sample and the training sample is calculated, and a nearest neighbor classification strategy is adopted to identify the test sample as the radiation source category corresponding to the training sample with the smallest distance to the test sample.

[0063] Among them, the joint distance metric function sim(f i ,f j )for:

[0064] sim(f i ,f j )=β·euclidean_distance(f i ,f j )+(1-β)·cos_distance(f i ,f i ),

[0065] Where β is a weight factor between 0 and 1, euclidean_distance(f i ,f j )=||f i -f j ||2, cos_distance(f i ,fj )=1-(f i ·f j ) / (||f i ||2·||f j ||2), where cos_similarity∈[-1,1].

[0066] In step S5, all samples in the training set and the validation set are encoded through the CVNN-TMN triplet network to obtain the corresponding embedding vectors:

[0067]

[0068] Among them, ε train Represents the set of embedded vectors corresponding to the training set samples, including N train samples; ε val The set of embedding vectors corresponding to the validation set samples contains N val samples.

[0069] For each validation sample f(x m ), calculate its difference with all training samples ε train The joint distance sim(f(x m ),f(x k )), select the one with the minimum joint distance The training samples of are classified as the nearest neighbors, which are expressed as follows:

[0070]

[0071] Among them, k * Representation and verification sample f(x m ) the most recent training sample index, is the predicted class label, i.e., the class of the most similar training sample.

[0072] During the recognition phase, all samples in the training and validation sets can be encoded using a CVNN-TMN triplet network to obtain corresponding embedding vectors and construct a feature space for the samples. For each validation sample, a joint metric distance is calculated between it and all training samples. This joint distance metric function comprehensively considers different distance calculation methods and balances them with a weighting factor β, enabling a more accurate measurement of the similarity between samples. Using a nearest neighbor classification strategy, the test sample is identified as the radiation source category corresponding to the training sample with the smallest distance to it, thereby achieving effective recognition of specific radiation sources within a limited sample. The trained model can then be applied to actual radiation source classification tasks, outputting reliable recognition results.

[0073] like Figure 2As shown in FIG, the model is constructed for the solution of the present invention, which mainly includes data preprocessing, Mixup data enhancement, model training using the CVNN-TMN triplet network, and nearest neighbor classification using a joint distance metric.

[0074] Table 1 CVNN-TMN network architecture

[0075]

[0076]

[0077] The basic neural network architecture of the SEI method based on the CVNN-TMN is shown in Table 1. It consists of nine convolutional layers and two fully connected layers, using a small convolution kernel with a kernel size of 3. To compensate for the limited feature extraction capability of the small convolution kernel, the kernel layer is increased in depth by adding layers. Furthermore, each convolution is followed by a max pooling operation to reduce the feature dimensionality. These two techniques are used to construct the basic neural network to ensure SEI performance. Each sample is sequentially fed into nine convolutional modules, each consisting of a convolution block, batch normalization, a pooling layer, and a flatten layer. Each convolution layer extracts higher-order features. Each layer is followed by an activation function and batch normalization layer to accelerate convergence and prevent vanishing gradients. The input data is processed through nine complex convolutional layers to extract multi-level features. The feature maps are then converted into one-dimensional vectors using the flatten operation to generate embedded features.

[0078] Experimental setup

[0079] The CVNN-TMN model is based on the PyTorch framework. All neural network models were run on a Linux system with an Intel(R) E5-2678 64-bit CPU and an NVIDIA GeForce GTX 1080Ti GPU. The experiments used an open-source ADS-B dataset. The ADS-B system is widely used to monitor aircraft status and features large amounts of data, easy labeling, and open source. The ADS-B dataset used in this section operates in the 1090 MHz frequency band. The signal is a baseband complex signal, divided into two I / Q channels, with 4800 points per channel. In the pre-training phase, the sample types were 10 categories, totaling 3091 samples, with uneven sample size across categories. The samples were divided into training and validation sets, with training sets accounting for 5%, 7%, 10%, and 20%, respectively. The remaining data was used as validation sets. The test set was not partitioned and was directly loaded from external test data.

[0080] Four advanced SEI schemes are selected to compare their performance with the proposed CVNN-TMN method in terms of recognition accuracy. The selected schemes are SlimCVNN, SSRCNN and two MAT schemes MAT_PA and MAT_CL.

[0081] Table 2 compares the five methods in terms of usage conditions, usage methods and performance.

[0082] Table 2 Comparison of usage of different methods

[0083]

[0084]

[0085] Table 3 presents the performance comparison results of the five methods. These results show that the proposed CVNN-TMN scheme significantly outperforms the other methods in accuracy at all training set ratios. In particular, at training set ratios of 5%, 7%, 10%, and 20%, CVNN-TMN achieves accuracies of 89.10%, 92.81%, 93.2%, and 95.05%, respectively. This result demonstrates that CVNN-TMN not only demonstrates strong recognition capabilities even with relatively low training data volumes, but also maintains a relatively stable improvement in accuracy with increasing training set ratios, demonstrating its robustness across varying data volumes. In contrast, the SlimCVNN method consistently lags behind CVNN-TMN in accuracy at the same training set ratio, particularly at smaller training set ratios, where its accuracy falls below 70%. Furthermore, although MAT_PA and MAT_CL also perform relatively well at higher training set ratios, their accuracy still falls short of CVNN-TMN, particularly at training set ratios of 5% and 7%, where they are significantly weaker than our method.

[0086] Table 3 Comparison of the accuracy of different methods

[0087]

[0088] Figure 3The classification visualization results of different methods are shown: (a) CVNN, (b) MAT_CL, (c) CVNN-TMN, and (d) the present invention. As can be seen, CVNN-TMN demonstrates significant advantages in enhancing discriminability. The distribution of points in each category is clear and independent, with large separations between categories, demonstrating that CVNN-TMN can effectively distinguish samples from different categories. CVNN-TMN demonstrates a more pronounced separation between closely spaced categories, making the boundaries between different categories more distinct. In comparison, the other methods perform mediocrely in terms of classification discrimination. SlimCVNN's category distribution is rather chaotic, especially for categories 0, 1, and 9, where there is significant overlap between categories, resulting in low discernibility in its classification results. MAT-CL and SSRCNN show some improvement, but some overlap still exists in the separation of some categories, such as between categories 4 and 5.

[0089] The above results show that the proposed CVNN-TMN method can effectively improve the classification accuracy while ensuring the clear separation of various categories and showing the best classification discrimination, which can verify the effectiveness of the CVNN-TMN method in improving the discrimination.

[0090] Ablation experiments

[0091] To effectively verify the contribution of different modules to model performance, an ablation experiment was designed to verify the role of Mixup enhancement and triplet twinning in enhancing SEI accuracy in the model. In the experiment, the training set ratio was set to 5%. The experimental results are shown in Table 4.

[0092] Table 4 Ablation experiment results

[0093]

[0094] This shows that the combination of Mixup and TMN structures significantly improves the recognition accuracy of SEI tasks. Removing any one component will lead to performance degradation, and removing both simultaneously will lead to severe performance degradation. This further proves the effectiveness and necessity of each module of the proposed method.

[0095] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of the patent application of the present invention. Any equivalent changes, equivalent replacements or modified changes within the technical spirit and principles suggested by the present invention should be included in the scope of patent protection covered by the present invention.

Claims

1. A finite sample specific radiation source identification method based on Mixup enhancement, characterized in that: The following steps are involved: S1. Collect radio signal samples from K radiation sources and preprocess them; divide the preprocessed samples into a training set and a validation set; S2. During the training phase, a triplet of samples (a, p, n) is constructed from the training set, where a is a baseline sample, p is a positive sample of the same category, and n is a negative sample of a different category. For each triplet of samples, a random strategy is used to determine whether to perform Mixup data augmentation. If the augmentation conditions are met, a mixed sample (a', p', n') is generated. S3. Input the mixed sample into the CVNN-TMN triplet network, extract its embedded features through the network structure with shared weights, and generate feature vectors f(a'), f(p'), and f(n'); S4. Calculate the distance difference between samples based on the triplet loss function and use the Mixup interpolation ratio λ a ,λ p ,λ n Weighting the triplet loss to optimize the clustering between samples of the same class and the discrimination between samples of different classes in the embedding space; S5. In the recognition phase, the test sample is mapped to the optimized embedding space, the joint metric distance between the test sample and the training sample is calculated, and the nearest neighbor classification strategy is adopted to identify the test sample as the radiation source category corresponding to the training sample with the smallest distance to the test sample.

2. The method for identifying specific radiation sources using a limited sample based on Mixup enhancement according to claim 1, characterized in that: In step S2, the random strategy is to generate a floating point number between [0, 1] through a random function. If the floating point number is less than 0.5, Mixup data enhancement is performed on the triplet sample; otherwise, Mixup data enhancement is not performed.

3. The method for identifying specific radiation sources using a limited number of samples based on Mixup enhancement according to claim 2, characterized in that: Mixup data enhancement is performed on the triplet samples, including: Among them, a', p', n' are randomly selected other benchmark samples, positive samples and negative samples, λ a ,λ p ,λ n It is the Mixup coefficient sampled from Beta(α,α),α>0 distribution, which controls the interpolation weight of different samples. It is a newly generated interpolation sample, which is used to supplement the original triplet sample to participate in subsequent training.

4. The method for identifying specific radiation sources using a limited number of samples based on Mixup enhancement according to claim 1, characterized in that: The CVNN-TMN triplet network consists of three branches with shared weights. Each branch has the same structure, which contains 9 layers of CVConv1D, CVReLU activation function, CVBN batch normalization layer and MaxPooling1D pooling layer in sequence, and then outputs the embedded feature vector through the Flatten layer and the fully connected layer.

5. The method for identifying specific radiation sources using a limited sample based on Mixup enhancement according to claim 3, characterized in that: In step S4, after the triplet samples are enhanced by Mixup data, the corresponding triplet loss function is: in, is the triplet loss, f(a), f(p), f(n) are the feature vectors of the benchmark sample, positive sample and negative sample obtained by the model respectively, and denote the Euclidean distance metric between the reference sample and the positive sample, and between the reference sample and the negative sample, respectively. α is a positive hyperparameter used to control the minimum interval between positive and negative samples. [·] + Represents the ReLU operation, that is, when the loss is negative, it takes 0 to ensure that the triplet loss is always non-negative; the loss weighted calculation is: Among them, λ a ,λ p ,λ n It is the Mixup interpolation ratio, which is used to control the weight distribution of loss in back propagation.

6. The method for identifying specific radiation sources using a limited number of samples based on Mixup enhancement according to claim 5, characterized in that: In the triple loss function, the gradient is back-propagated by the chain rule. The loss function uses the chain rule to calculate the feature vectors f(a), f(p), and f(n). The gradient is The times are:

7. The method for identifying specific radiation sources using a limited sample based on Mixup enhancement according to claim 1, characterized in that: In step S4, if the triplet sample is not subjected to Mixup data enhancement, the original triplet is used for loss calculation and optimization.

8. The method for identifying specific radiation sources using a limited sample based on Mixup enhancement according to claim 1, characterized in that: Joint distance metric function sim(f i ,f j )for: sim(f i ,f j )=β·euclidean_distance(f i ,f j )+(1-β)·cos_distance(f i ,f j ), Where β is a weight factor between 0 and 1, euclidean_distance(f i ,f j )=||f i -f j ||2, cos_distance(f i ,f j )=1-(f i ·f j ) / (||f i ||2·||f j ||2), where cos_similarity∈[-1,1].

9. The method for identifying specific radiation sources using a limited sample based on Mixup enhancement according to claim 1, characterized in that: In step S5, all samples of the training set and the validation set are encoded through the CVNN-TMN triplet network to obtain the corresponding embedding vectors. Among them, ε train Represents the set of embedded vectors corresponding to the training set samples, including N train samples; ε val The set of embedding vectors corresponding to the validation set samples contains N val samples; For each validation sample f(x m ), calculate its difference with all training samples ε train The joint distance sim(f(x m ),f(x k )), select the one with the minimum joint distance f(x k* ) is used as the nearest neighbor for classification, which is expressed as follows: Among them, k * Representation and verification sample f(x m ) the most recent training sample index, is the predicted class label, i.e., the class of the most similar training sample.

Images