Small sample radiation source individual identification method based on multi-optimization double-channel prototype network

Abstract

The application discloses a small sample radiation source individual identification method based on a multi-optimization double-channel prototype network, which comprises the following steps: obtaining a radiation source signal to be identified; inputting the radiation source signal to be identified into a pre-trained double-channel prototype network; a double-channel convolution network in the pre-trained double-channel prototype network outputs a feature vector of the radiation source signal to be identified; and a classification network in the pre-trained double-channel prototype network determines a signal category to which the radiation source signal to be identified belongs according to the feature vector and clustering centers of various radiation source signals; wherein the double-channel convolution network is trained according to a support set, a query set in a training set containing multiple types of radiation source sample signals and a loss function used for updating the clustering center of the radiation source signal at each time of training and for focusing on the classification error of the sample. The application can improve the identification accuracy.

Description

Technical Field

[0001] This invention belongs to the field of intelligent recognition technology, specifically relating to a method for identifying small-sample radiation source individuals based on a multi-optimized dual-channel prototype network. Background Technology

[0002] Radiation source identification and classification is a method for physical layer authentication of electronic devices. This method categorizes targets based on differences in radio frequency (RF) characteristics within circuits. These RF characteristics originate from manufacturing variations in hardware circuits, making them difficult to counterfeit and unique. Therefore, it is expected that radiation source identification and classification will play a significant role in the Internet of Things (IoT) field.

[0003] With the rapid development of deep learning, many radiation source identification methods based on deep learning networks have been proposed and have achieved very good results. However, these methods require a large number of labeled samples for training, while in reality, training samples are often very limited, leading to generally unsatisfactory classification results. Therefore, researching radiation source identification methods under small sample conditions has great practical significance.

[0004] Training a network requires a large number of labeled samples, and insufficient target samples can lead to overfitting. Therefore, many methods based on scale learning and meta-learning have been proposed for few-shot learning. These methods can be broadly categorized into the following steps: training the feature extraction network with a large amount of irrelevant data; and classifying a small number of target samples using prior knowledge learned earlier. However, these methods still result in low recognition accuracy. Summary of the Invention

[0005] To address the aforementioned problems in related technologies, this invention provides a method for identifying individual radiation sources in small samples based on a multi-optimized dual-channel prototype network. The technical problem to be solved by this invention is achieved through the following technical solution:

[0006] This invention provides a method for identifying individual radiation sources in small samples based on a multi-optimized dual-channel prototype network, comprising:

[0007] Acquire the signal of the radiation source to be identified;

[0008] The radiation source signal to be identified is input into a pre-trained dual-channel prototype network. The dual-channel convolutional network in the pre-trained dual-channel prototype network outputs the feature vector of the radiation source signal to be identified. The classification network in the pre-trained dual-channel prototype network determines the signal category to which the radiation source signal to be identified belongs based on the feature vector and the cluster centers of various types of radiation source signals. The dual-channel convolutional network is trained based on the support set and query set in the training set containing sample signals of multiple types of radiation sources, as well as the loss function used to update the cluster centers of radiation source signals at each training session and to focus on samples with misclassification.

[0009] The present invention has the following beneficial technical effects:

[0010] By designing a dual-channel prototype network based on a dual-channel convolutional network, the neural network learns the nonlinear mapping of the input to the embedding space, thereby acquiring the essential features of individual signals from different radiation sources and ultimately improving the network's recognition accuracy. Furthermore, by designing an optimized loss function to focus on samples with poor classification performance, the trained network gains better generalization ability, further improving the recognition accuracy of individual radiation sources. Finally, by designing an optimized loss function for cluster centers of sample classes, better cluster centers are obtained through iteration, effectively improving the network's recognition accuracy.

[0011] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0012] Figure 1 This is an optional flowchart of a small sample radiation source individual identification method based on a multi-optimized dual-channel prototype network provided in an embodiment of the present invention;

[0013] Figure 2 A schematic diagram of an exemplary dual-channel convolutional network provided in an embodiment of the present invention;

[0014] Figure 3 This is a schematic diagram illustrating an exemplary training and testing process for a dual-channel prototype network, provided as an embodiment of the present invention. Detailed Implementation

[0015] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0016] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0017] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0018] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, disclosure, and appended claims in carrying out the claimed invention. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0019] Currently, some relevant solutions are as follows:

[0020] 1) A decision theory-based method is proposed. This method conducts an in-depth study of the average likelihood ratio function of multi-level frequency shift keying (M-FSK) signals, and finds that averaging it yields an integral expression for the zero-order modified Bessel function. Furthermore, a series expansion can be performed on this expression, leading to a modulation scheme identification algorithm based on higher-order correlation analysis. This successfully solves the M-FSK signal identification problem.

[0021] 2) A machine learning-based method, which is a digital communication signal modulation and recognition algorithm based on high-order cumulants and support vector machines (SVM), achieves accurate recognition of six digital modulation methods.

[0022] 3) A method based on traditional deep learning. This method uses GNU Radio to create a modulated signal dataset and uses a Convolutional Neural Network (CNN) to identify 11 types of modulated signals on the dataset. The recognition accuracy is high when the signal-to-noise ratio (SNR) is greater than 0dB.

[0023] 4) A small-sample-based classifier metric scheme. This method obtains the mapping of samples in the embedding space by training a feature extraction network, and obtains the prototype of the class based on the mean of samples of the same class in the embedding space. Then, the classification task is performed on the sample points by simply finding the nearest class prototype.

[0024] Of the aforementioned approaches, decision theory-based methods suffer from complex models, slow processing speeds, and limited applicability to practical needs. Machine learning-based methods distinguish signals based on the characteristics of different modulated signals, with recognition performance dependent on the selection of modulated signal features and the design of the classifier. Both methods require prior knowledge to build their models, extracting multiple key statistics from the received data to achieve signal recognition. Furthermore, at low signal-to-noise ratios, their recognition performance is poor due to the influence of the channel environment. Decision theory-based and machine learning-based methods are complex and require human intervention, while traditional deep learning methods require massive amounts of data, all of which consume significant resources.

[0025] The purpose of this invention is to achieve accurate classification of individual radiation sources with only a small number of samples, while saving a lot of resources and reducing human intervention.

[0026] Figure 1 This is an optional flowchart of a small-sample radiation source individual identification method based on a multi-optimized dual-channel prototype network provided in an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0027] S101. Obtain the signal of the radiation source to be identified.

[0028] S102. The radiation source signal to be identified is input into the pre-trained dual-channel prototype network. The dual-channel convolutional network in the pre-trained dual-channel prototype network outputs the feature vector of the radiation source signal to be identified. The classification network in the pre-trained dual-channel prototype network determines the signal category to which the radiation source signal to be identified belongs based on the feature vector and the cluster centers of various types of radiation source signals. The dual-channel convolutional network is trained based on the support set and query set in the training set containing sample signals of multiple types of radiation sources, as well as the loss function used to update the cluster centers of radiation source signals during each training and to focus on samples with classification errors.

[0029] Here, the dual-channel convolutional network comprises two sub-networks, each sub-network including: multiple sequentially connected convolutional blocks and a fully connected layer; each convolutional block sequentially includes: a convolutional layer, a batch normalization layer, an activation layer, and a max-pooling layer. For example, as shown... Figure 2 As shown, the two sub-networks are the I-channel and Q-channel, respectively. Each sub-network contains a first convolutional block, a second convolutional block, a third convolutional block, a fourth convolutional block, a fifth convolutional block, a sixth convolutional block, a seventh convolutional block, an eighth convolutional block, and a ninth convolutional block (i.e., nine convolutional blocks with the same structure) connected sequentially, and a fully connected layer. The specific structure of each convolutional block is: a convolutional layer, a batch normalization layer, an activation layer (ReLU activation function), and a max pooling layer. The number of convolutional kernels in each convolutional layer is 64, the kernel size is 1*3, the stride is 1, and the padding is 0. The pooling layer uses max pooling, the window size is 1*2, and the stride is 2. The output dimension of the tenth fully connected layer is 1024.

[0030] Here, the classification network can be any classifier, such as the Euclid classifier.

[0031] Here, before inputting the radiation source signal to be identified into the pre-trained dual-channel prototype network, the feature vectors of the radiation source sample signals of each class can be averaged, and the average value is used as the cluster center of that class of radiation source signals. Each class of radiation source sample signal is a sample signal used to train the pre-trained dual-channel prototype network.

[0032] In some embodiments, prior to S102, the method includes:

[0033] S201. Each time, obtain the current support set containing support samples and the current query set containing query samples from the training set containing sample signals of multiple types of radiation sources.

[0034] For example, the training set can be an ADS-B signal dataset, which consists of 90 signal classes. The training process uses an N-way K-shot L-query episode mechanism, where N classes of modulated signals are randomly selected from the training set each time. Then, K samples are selected from each class of modulated signals as the support set and L samples are selected as the query set. Thus, the support set for each training session contains N*K support samples, and the query set for each training session contains N*L query samples. Each sample has a label (true class).

[0035] S202. Input the support samples and query samples into the dual-channel prototype network to be trained to obtain the first feature vector of each support sample, the confidence level of each query sample belonging to various types of radiation source signals, and the predicted category of each query sample.

[0036] Here, each sample x is mapped to the feature space through a dual-channel convolutional network in the dual-channel prototype network, resulting in a feature vector f(x) with dimension 1024. Here, f(·) is the mapping function of the dual-channel convolutional network, and the mapped representation (feature vector) of the sample can be expressed as: encode=f(x):x∈R 2×4800 →f(x)∈R 1024 .

[0037] Here, the classification network in the dual-channel prototype network can calculate the Euclidean distance between the query sample and the cluster center of each type of radiation source signal based on the feature network of each query sample and the cluster center of each type of radiation source signal. Based on the Euclidean distance, the confidence level of the query sample belonging to each type of radiation source signal is obtained, and the radiation source signal corresponding to the highest confidence level is taken as the predicted category of the query sample.

[0038] Here, when this is the first training session, the cluster centers of various radiation source signals used to calculate the confidence level can be obtained through random initialization, and the dual-channel prototype network to be trained this time is the initial dual-channel prototype network.

[0039] S203. Based on the preset weight coefficient, preset adjustable coefficient, first feature vector, the current cluster center of each type of radiation source signal to which the current support set belongs, the predicted category and true category of each query sample, and the confidence level of each query sample belonging to each type of radiation source signal, determine the current loss value.

[0040] Here, the first loss value can be determined based on the current learning rate, the first feature vector of the support samples of each real class, and the current cluster center of each real class; the second loss value can be determined based on the confidence and predicted class of the real class corresponding to each query sample, as well as the preset adjustable coefficient; and the final loss value can be determined based on the preset weight coefficient, the first loss value, and the second loss value.

[0041] Specifically, the formula for calculating the first loss value in this case is as follows:

[0042]

[0043] Among them, L Center Let y be the first loss value in this case. label Let x be the real category, and x be the number of elements in the current support set that belong to the y-th element.label For each support sample of a radiation signal source, f(W;x) is the first feature vector corresponding to x, and E[.] is the averaging function.

[0044] Here, L Center It can improve the degree of aggregation of inter-class samples toward cluster centers.

[0045] Specifically, the formula for calculating the second loss value in this case is as follows:

[0046]

[0047] Among them, L Focal Let A be the second loss value, and let A be the number of query samples in the current query set. γ represents the confidence level of the true category corresponding to the j-th query sample in this query set, and γ is a preset adjustable coefficient.

[0048] Here, L Focal By focusing the loss on sample types that perform poorly in classification, it can help improve the classification accuracy of difficult-to-classify samples.

[0049] Specifically, the formula for calculating the loss value in this case is as follows:

[0050] L = L Focal +λL Center ;

[0051] Where L is the loss value in this case, L Center L is the first loss value in this case. Focal This is the second loss value for this operation, and λ is a preset weighting coefficient (for example, it can be 0.1).

[0052] S204. Based on the current loss value, update the dual-channel prototype network to be trained and the current cluster center respectively to obtain the next dual-channel prototype network to be trained and the next cluster center of each type of radiation source signal to which the current support set belongs. Iterate in this way until the training ends and the pre-trained dual-channel prototype network is obtained.

[0053] Here, training can be stopped when the number of training iterations reaches a preset value (e.g., 100) or when the network converges, thus obtaining a pre-trained dual-channel prototype network.

[0054] Specifically, the formula for calculating the next cluster center for each type of radiation source signal in this support set is as follows:

[0055]

[0056]

[0057]

[0058] in, For the yth generation of this support set label The cluster centers of the next (t+1) cluster of radiation signal sources. For the yth label The cluster centers of the radiation signal source in this (t-th) iteration, where α is the learning rate for this iteration, B is the number of support samples in the support set for this iteration, and x... i For belonging to the yth label For each support sample of a radiation signal source, f(W; x) i ) is x i The corresponding feature vector, y i For x i The predicted category, y label For x i The true category.

[0059] Here, during network training, you can train only the two-channel convolutional network, or you can train both the two-channel convolutional network and the classification network simultaneously. For example, Figure 3 This is a schematic diagram illustrating the training and testing procedures for a dual-channel prototype network. Figure 3 As shown, the support set and the query set are input together into a dual-channel convolutional network (i.e., Figure 3 After the feature extraction network in the middle, the feature vectors of each support sample and each query sample are obtained respectively; then, L can be calculated based on the feature vector of each support sample. Center ( Figure 3 In the context of Center Loss, the Eulid classifier, on the other hand, obtains the predicted label (predicted category) and confidence score of each query sample belonging to each type of radiation source signal based on the feature vector of each query sample. Then, based on the predicted label, confidence score of each type of radiation source signal, and corresponding true label (true category) of each query sample, L0 can be calculated. Focal ( Figure 3 Focal Loss in L Center and L Focal After obtaining the hybrid loss function (loss value), the network weights and the cluster centers of each type of radiation source signal belonging to the current support set are updated through backpropagation using the hybrid loss function.

[0060] In some embodiments, after obtaining the pre-trained dual-channel prototype network, the network weights can be fixed, and the pre-trained dual-channel prototype network can be tested. For example, the test set has three options: 10 classes, 20 classes, and 30 classes of signals, and the sample classes in the test set do not overlap with the sample classes in the training set. The testing process also adopts an N-way K-shot L-query episode mechanism, that is, in each test, N classes of modulation signals are randomly selected, and K samples are drawn from each class of modulation signals as the support set and L samples as the query set. The testing process is as follows: Figure 3 As shown, each support sample in the support set and each query sample in the query set are mapped to the feature space through a feature extraction network to obtain feature vectors. Then, the feature vectors generated by the same support set are averaged and the average value is used as the cluster center of the class. The Euclid classifier uses the obtained cluster center to perform classification tests on each query sample based on the feature vector of each query sample to obtain the predicted label of each query sample.

[0061] This invention designs a dual-channel prototype network based on a dual-channel convolutional network, enabling the neural network to learn the nonlinear mapping of input to the embedding space, thereby acquiring the essential features of individual signals from different radiation sources and ultimately improving the network's recognition accuracy. Furthermore, it designs an optimized loss function to focus on samples with poor classification performance, thus giving the trained network better generalization ability and further improving the recognition accuracy of individual radiation sources. Finally, it designs an optimized loss function for cluster centers of sample classes, iteratively obtaining better cluster centers and effectively improving the network's recognition accuracy.

[0062] Compared with decision theory-based methods, the model of this invention is simpler, more versatile, and has higher recognition accuracy. Compared with machine learning-based methods, the model of this invention is simpler and does not require manual feature selection, resulting in higher recognition efficiency. Compared with traditional deep learning-based methods, the model of this invention does not require massive amounts of data, resulting in higher recognition efficiency. In its training process, class prototypes are obtained through simple averaging, while in the model of this invention, class prototypes are obtained through iterative training, achieving good accuracy for various types of modulated signals.

[0063] The following simulation experimental data further illustrates the effectiveness of the method described in this invention.

[0064] (1) Experimental conditions

[0065] The hardware platform for the simulation experiment of this invention is as follows: GPU is NVIDIA GeForce RTX 3070, eight-core, with a main frequency of 3.50GHz, memory size is 32GB, and video memory size is 8GB.

[0066] The software platform for the simulation experiment of this invention is Windows 10 operating system.

[0067] (2) Simulation content

[0068] To verify the recognition effect of the method of the present invention under limited sample conditions, the present invention sets up an N-way K-shot small sample mode on the feature extraction network of the dual-channel convolutional neural network.

[0069] In this experiment, the training sample set was selected from 90 modulation signals, totaling 20,566 signals; the test sample set was selected from 10 / 20 / 30 modulation signals, totaling 1,000 / 2,000 / 3,000 signals; the training and testing process adopted an N-way K-shot episode mechanism, that is, in each training and testing, N types of modulation signals were randomly selected, and K samples were extracted from each type of modulation signal as the support set and L samples as the query set. In the experiment, N = [10, 20], K = [5], and L = 5. To verify L Focal and L Center To assess the effectiveness, four experiments were conducted, with loss functions L1, L2, and L3 respectively. Ce L Focal L Ce +L Center and L Focal +L Center These experiments were labeled Experiment 1, Experiment 2, Experiment 3, and Experiment 4, respectively, and their classification results were compared.

[0070] Simulation results:

[0071] The experimental accuracy is shown in Table 1. As can be seen from Table 1, when K = 5, that is, when the number of samples in each class of the support set is 5, the loss function is L. Focal L Ce +L Center and L Focal +L Center The network classification results are compared with the loss function L. Ce The network classification results were improved in most cases, with L... Center The most significant improvement is in the accuracy of network recognition. And L Focal +L Center The network classification results are the best.

[0072] The experimental results are largely in line with expectations. Furthermore, the results show that the loss function L... Focal and L Center Both have a positive effect on network training.

[0073]

[0074] Table 1

[0075] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for identifying individual radiation sources in small samples based on a multi-optimized dual-channel prototype network, characterized in that, include: Acquire the signal of the radiation source to be identified; The radiation source signal to be identified is input into a pre-trained dual-channel prototype network. The dual-channel convolutional network in the pre-trained dual-channel prototype network outputs a feature vector of the radiation source signal to be identified. The classification network in the pre-trained dual-channel prototype network determines the signal category to which the radiation source signal to be identified belongs based on the feature vector and the cluster centers of various types of radiation source signals. The dual-channel convolutional network is trained based on the support set and query set in the training set containing sample signals of multiple types of radiation sources, as well as the loss function used to update the cluster centers of the radiation source signals at each training session and to focus on samples with misclassification. The dual-channel convolutional network includes two sub-networks, namely I-channel and Q-channel. Each sub-network includes: multiple sequentially connected convolutional blocks and a fully connected layer; each convolutional block includes: a convolutional layer, a batch normalization layer, an activation layer, and a max pooling layer. The method further includes, before inputting the radiation source signal to be identified into the pre-trained dual-channel prototype network: Each time, a support set containing support samples and a query set containing query samples are obtained from the training set containing sample signals from multiple types of radiation sources. The support samples and query samples are input into the dual-channel prototype network to be trained to obtain the first feature vector of each support sample, the confidence of each query sample belonging to various types of radiation source signals, and the predicted category of each query sample. The first loss value is determined based on the learning rate, the first feature vector of the support samples of each real class, and the cluster center of each real class in this instance. The second loss value is determined based on the confidence level of the true category and the predicted category corresponding to each query sample, as well as the preset adjustable coefficient. The current loss value is determined based on the preset weighting coefficients, the first loss value, and the second loss value. Based on the current loss value, update the dual-channel prototype network to be trained and the current cluster center respectively to obtain the next dual-channel prototype network to be trained and the next cluster center of each type of radiation source signal to which the current support set belongs. This process is repeated iteratively until the pre-trained dual-channel prototype network is obtained at the end of training.

2. The method for identifying individual radiation sources from small samples based on a multi-optimized dual-channel prototype network according to claim 1, characterized in that, The formula for calculating the first loss value is as follows: ； in, This is the first loss value in this case. For the real category, The support for this event is concentrated in the first... Each support sample of the radiation signal source, for The corresponding first feature vector, It is an average function.

3. The method for identifying individual radiation sources in small samples based on a multi-optimized dual-channel prototype network according to claim 1, characterized in that, The formula for calculating the second loss value is as follows: ； in, This is the second loss value in this case. This represents the number of query samples in this query set. This is the first query in this query set. The confidence level of the true category corresponding to each query sample. This is a preset adjustable coefficient.

4. The method for identifying individual radiation sources in small samples based on a multi-optimized dual-channel prototype network according to claim 1, characterized in that, The formula for calculating the loss value in this case is as follows: ； in, This is the loss value for this instance. This is the first loss value in this case. This is the second loss value in this case. These are preset weighting coefficients.

5. The method for identifying individual radiation sources in small samples based on a multi-optimized dual-channel prototype network according to claim 1, characterized in that, The formula for calculating the next cluster center for each type of radiation source signal in the current support set is as follows: ； ； ； in, This is the first support set to which this group belongs. The next cluster center of the radiation-like signal source. For the first The cluster centers of this radiation-like signal source. For this learning rate, This represents the number of supporting samples in this support set. For belonging to the first Each support sample of the radiation signal source, for The corresponding feature vector, for Prediction categories, for The true category.

6. The method for identifying individual radiation sources in small samples based on a multi-optimized dual-channel prototype network according to claim 1, characterized in that, Before inputting the radiation source signal to be identified into the pre-trained dual-channel prototype network, the method includes: The feature vectors of the sample signals from each type of radiation source are averaged, and the average value is used as the cluster center of that type of radiation source signal.

Images