Radiation source identification method and device

Abstract

The present disclosure discloses a method and device for identifying a radiation source, which specifically relates to the field of data processing. The main technical scheme includes: when performing the identification of the radiation source, inputting a noise vector into a generation network in a preset radiation source identification model; outputting a signal sample after the convolution operation of the preset radiation source identification model to blind the discriminant network in the preset radiation source identification model; performing binary classification on the output samples of the generation network and the unlabeled samples based on the discriminant network; and evaluating the binary classification results based on the skewed distribution loss function and the labeled samples. The radiation signal is trained based on the preset radiation element identification model to identify its radiation source, and the gradient imbalance caused by the uneven sample category is offset according to the skewed distribution loss function, thereby increasing the contribution of the minority class samples to the total loss, and realizing the identification of the radiation source of the skewed category distribution data.

Description

Technical Field

[0001] The present disclosure relates to the field of data processing technology, and in particular to a method and device for identifying a radiation source. Background Art

[0002] Specific radiation source identification refers to the technology of extracting subtle differences from received signals to associate a single radiation source. With the large-scale popularization of IoT applications and the rapid growth of the number of IoT access devices, efficient, secure and stable IoT device authentication solutions are the key to preventing malicious attacks and ensuring user privacy security.

[0003] In recent years, deep learning technology has shown great potential in the application of specific radiation source identification due to its powerful abstract feature retrieval capabilities. In the existing technology, specific radiation source identification schemes based on deep learning often perform supervised learning when the number of various samples is basically balanced.

[0004] In practical applications, there is a large amount of imbalance in the number of data between the various categories of radiation source data. In class-imbalanced data sets, a large amount of data belongs to the minority category, while the majority category contains only a small number of samples, and the samples show a skewed distribution. This results in the majority class samples dominating the direction of loss update and gradient descent during the training process of the deep learning model; while for minority class samples, the deep learning model is difficult to obtain sufficient features to achieve nonlinear mapping between input and output, causing the model to tend to identify them as the majority class and escape the penalty caused by misclassification, resulting in a low recognition rate for radiation source data with skewed category distribution. Summary of the invention

[0005] The present disclosure provides a method and device for identifying a radiation source, the main purpose of which is to solve the problem of inaccurate identification of radiation sources caused by skewed category distribution of radiation data.

[0006] According to a first aspect of the present disclosure, a method for identifying a radiation source is provided, comprising:

[0007] Inputting the noise vector into a generation network in a preset radiation source identification model;

[0008] Outputting signal samples after the convolution operation of the preset radiation source recognition model to deceive the discriminant network in the preset radiation source recognition model;

[0009] Based on the discriminant network, the generated network output samples and unlabeled samples are classified into two categories;

[0010] The binary classification results are evaluated based on the skewed distribution loss function and labeled samples.

[0011] Optionally, before inputting the noise vector into a generation network in a preset radiation source identification model, the method further includes:

[0012] A generative adversarial network model is constructed, wherein the generative adversarial network model at least includes: a generative network and a discriminative network, and is used to optimize network parameters of the generative network and the discriminative network through a semi-supervised training method.

[0013] Optionally, the binary classification result is evaluated based on the skew distribution loss function and the labeled samples, and the method further includes:

[0014] Integrating the residual network into the discriminant network;

[0015] Configure the scale topology to fuse the resolution features of signals at different scales.

[0016] Optionally, the method further includes:

[0017] Dividing the discriminant network into an unsupervised part and a supervised part;

[0018] The loss function corresponding to the supervision part is the standard cross entropy between the labeled samples and the binary classification results;

[0019] The standard cross entropy (CE) function for processing balanced data sets is usually expressed as:

[0020] In the formula, y i ∈(0,1) is the confidence probability, is the predicted probability, N is the sample category, and CE loss minimizes the training error by assuming that the loss weight is independent of the frequency of the category to which it belongs.

[0021] Optionally, the evaluating the binary classification result based on the skewed distribution loss function and the labeled samples includes:

[0022] Evaluation is based on the following skewed distribution loss function:

[0023]

[0024]

[0025]

[0026] Where η, χ and μ are hyperparameters. FL is the focal loss, L CI It is the class imbalance loss, which is used to deal with class imbalance data.

[0027] According to a second aspect of the present disclosure, there is provided a radiation source identification device, comprising:

[0028] An input unit, used for inputting a noise vector into a generation network in a preset radiation source identification model;

[0029] An output unit, used to output a signal sample after a convolution operation of the preset radiation source recognition model, so as to deceive a discriminant network in the preset radiation source recognition model;

[0030] The classification unit is used to classify the output samples of the generation network and the unlabeled samples based on the discriminant network;

[0031] The evaluation unit is used to evaluate the binary classification results based on the skew distribution loss function and labeled samples.

[0032] Optionally, the device further comprises:

[0033] A construction unit is used to construct a generative adversarial network model before the input unit inputs the noise vector into the generative network in the preset radiation source identification model. The generative adversarial network model at least includes: a generative network and a discriminative network, which are used to optimize the network parameters of the generative network and the discriminative network through semi-supervised training.

[0034] Optionally, the evaluation unit further includes:

[0035] A fusion module, used to integrate the residual network into the discriminant network;

[0036] The configuration module is used to configure the resolution characteristics of the scale topology fusion signal at different scales.

[0037] Optionally, the device further comprises:

[0038] A division unit, used for dividing the discriminant network into an unsupervised part and a supervised part;

[0039] The loss function corresponding to the supervision part is the standard cross entropy between the labeled samples and the binary classification results;

[0040] The standard cross entropy (CE) function for processing balanced data sets is usually expressed as:

[0041] In the formula, y i ∈(0,1) is the confidence probability, is the predicted probability, N is the sample category, and CE loss minimizes the training error by assuming that the loss weight is independent of the frequency of the category to which it belongs.

[0042] Optionally, the evaluation unit is further configured to perform evaluation based on the following skew distribution loss function:

[0043]

[0044]

[0045]

[0046] Where η, χ and μ are hyperparameters. FL is the focal loss, L CI It is the class imbalance loss, which is used to deal with class imbalance data.

[0047] The radiation source identification method and device provided by the present disclosure, when performing the identification of the radiation source, first, input the noise vector into the generation network in the preset radiation source identification model; secondly, output the signal sample after the convolution operation of the preset radiation source identification model, so as to deceive the discriminant network in the preset radiation source identification model; perform binary classification on the generation network output samples and unlabeled samples based on the discriminant network; finally, evaluate the binary classification results based on the skewed distribution loss function and the labeled samples. Compared with the related art, the embodiment of the present application trains the radiation signal based on the preset radiation element identification model to identify its radiation source, and offsets the gradient imbalance caused by the uneven sample category according to the skewed distribution loss function, thereby improving the contribution of the minority class samples to the total loss, and realizing the identification of the radiation source of the skewed category distribution data.

[0048] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0050] Figure 1 A schematic diagram of a flow chart of a method for identifying a radiation source provided by an embodiment of the present disclosure;

[0051] Figure 2 A schematic diagram of a single-hop communication scenario provided by an embodiment of the present disclosure;

[0052] Figure 3 A flowchart of a method for evaluating classification results provided by an embodiment of the present disclosure;

[0053] Figure 4 A schematic diagram of a discriminant network based on a multi-scale topological structure and a residual unit provided in an embodiment of the present disclosure;

[0054] Figure 5 A schematic diagram of a basic neuron structure provided by an embodiment of the present disclosure;

[0055] Figure 6 A schematic diagram of the structure of a radiation source identification device provided in an embodiment of the present disclosure;

[0056] Figure 7 A schematic diagram of the structure of a radiation source identification device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0057] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0058] The following describes the radiation source identification method and device according to the embodiments of the present disclosure with reference to the accompanying drawings.

[0059] Figure 1 A schematic flow chart of a method for identifying a radiation source provided in an embodiment of the present disclosure.

[0060] Step 101: input the noise vector into a generation network in a preset radiation source identification model.

[0061] The embodiments of the present application are applied to a radiation source identification framework for skewed sample distribution (Semi-supervisedGAN for skewed class distribution, S3CDGAN) to improve the individual classification performance of radiation sources when the sample class distribution is unbalanced and the labels are limited.

[0062] In S3CDGAN, the input of the generator is the noise vector z, and the optimization objective of the generated network is expressed as:

[0063]

[0064] Where V(G,D) represents the function to be optimized, z~P z (z) is the input noise, D(x) represents the correct discrimination probability of the discriminant network for the real sample, G(z) represents the output of the generated network, and Ε(·) represents the expected operator.

[0065] Before executing this step, a radiation source signal model is established to train a preset radiation source recognition model; a communication scenario is established with at least one simplex radiation source, a channel and a single receiver. For ease of understanding of the single-hop communication scenario described in the embodiment of the present application, please refer to Figure 2, the radiation source signal in this scenario is received in time-sharing, so there is no need to consider blind source separation. The physical layer characteristics of the radiation source mainly come from the nonlinear system response of the power amplifier network. The high-frequency radiation signal s(nT) input to the power amplifier can be expressed as:

[0066]

[0067] Where m(nT) represents the modulated signal from the baseband, f c and f s Denote the carrier frequency and sampling frequency respectively. Taylor polynomials are introduced to describe the specificity of the power amplifier. The power amplifier output T in a single-hop scenario (i) (·) can be expressed as:

[0068]

[0069] In the formula is the Taylor polynomial coefficient, M is the order of the polynomial, and the sampling signal R(i)(nT) from the i-th radiation source intercepted by the receiver can be expressed as:

[0070] R (i) (nT) = T (i) (s(nT))*h (i) (nT)+w (i) (nT),n=1,2,...,N,i=1,2,...,K

[0071] Where h (i) (nT) is the channel fading coefficient from the transmitter to the receiver, w (i) (nT) is the channel additive noise, and the received signal R(i)(nT) can be expressed as:

[0072]

[0073] Step 102, outputting a signal sample after the convolution operation of the preset radiation source recognition model to deceive the discriminant network in the preset radiation source recognition model.

[0074] In S3CDGAN, the input of the generator is a noise vector z, and after the convolution operation, the output signal sample is used to deceive the discriminator. The discriminator network receives the generator output, the labeled sample (x, y) ~ Pdata (x, y) and the unlabeled sample x ~ Pdata (x) at the same time. The supervised part and the unsupervised part of the discriminator network share weights and parameters except the output layer to make full use of the feature distribution of the unlabeled samples to update their own parameters.

[0075] Step 103: Binary classification is performed on the generated network output samples and the unlabeled samples based on the discriminant network.

[0076] The discriminant network continuously improves its ability to distinguish between real samples and false samples through training, so the optimization goal is

[0077]

[0078] (6) x~P data (x) means that it obeys the real sample distribution, and D(G(z)) means the probability that the discriminant network predicts the output sample of the generative network as the real sample. The essence of network optimization is the game between the two, and its loss function can be expressed as:

[0079]

[0080] The generative network and the discriminative network are continuously adjusted and optimized during the confrontation process. When the output of the generative network matches the real sample, Nash equilibrium is reached. At this time, V(G,D) is the global optimal solution.

[0081] The discriminant network receives the output of the generative network, the labeled samples (x, y)~Pdata(x, y) and the unlabeled samples x~Pdata(x) at the same time. The supervised and unsupervised parts of the discriminant network share weights and parameters except the output layer to fully utilize the feature distribution of the unlabeled samples to update their own parameters. In the semi-supervised classification task with K radiation sources, for the generative network output G(z) and the unlabeled samples x~Pdata(x), the preset radiation source recognition model uses the Sigmoid function of the output layer to complete the binary classification:

[0082]

[0083] In the formula, xi is the output element in the binary classification vector. In the supervision part, the generated network output G(z) is taken as the K+1th class, and the Sofmtax layer is used together with the labeled samples of class K to implement K+1-dimensional supervised classification, converting the output vector into class probability:

[0084]

[0085] Where xi is the output in the K-dimensional classification vector.

[0086] Step 104 , evaluating the binary classification results based on the skew distribution loss function and the labeled samples.

[0087] In particular, the preset reflection source generation network uses the output G(z) of the generation network as the “pseudo sample” of the K+1th class. In the preset radiation source recognition model, the optimization function L of the generation network is G It can be expressed as maximizing the probability that G(z) is judged as the K+1th class

[0088]

[0089] In order to improve the stability of the generated network and prevent over-training of the current discriminant network, we introduce the feature matching loss L into the above formula: FM

[0090]

[0091] Where f(·) represents the feature extraction layer in the discriminant network, and the improved optimization function can be expressed as

[0092] Where τ is defined as the matching coefficient. The optimization function L of the discriminant network D Divided into the unsupervised part V(D) uns and supervision part V(D) sup , the loss function of the unsupervised part is expressed as:

[0093]

[0094] The supervised loss function V(D) under the balanced distribution of sample categories sup is the standard cross entropy between the true class label and the predicted class label:

[0095]

[0096] The overall objective function of the discriminant network is:

[0097]

[0098] Therefore, the loss function of the preset radiation source identification model can be expressed in the form of a zero-sum game:

[0099]

[0100] The radiation source identification method provided by the present disclosure, when performing the identification of the radiation source, first, input the noise vector into the generation network in the preset radiation source identification model; secondly, output the signal sample after the convolution operation of the preset radiation source identification model, so as to blind the discriminant network in the preset radiation source identification model; based on the discriminant network, the generation network output samples and the unlabeled samples are classified into two categories; finally, the binary classification results are evaluated based on the skewed distribution loss function and the labeled samples. Compared with the related art, the embodiment of the present application trains the radiation signal based on the preset radiation element identification model to identify its radiation source, and offsets the gradient imbalance caused by the uneven sample category according to the skewed distribution loss function, thereby improving the contribution of the minority class samples to the total loss, and realizing the identification of the radiation source of the skewed category distribution data.

[0101] As an extension of the above-mentioned application embodiment, when constructing a preset radiation source recognition model, the following method can be adopted: the generative adversarial network model (GAN) includes at least: a generator network and a discriminant network, which are used to optimize the network parameters of the generator network and the discriminant network through semi-supervised training. Since the original GAN ​​can only perform unsupervised expansion on real samples and cannot effectively utilize the correlation information between labeled data and unlabeled data, mode collapse in the adversarial training process is difficult to avoid. For this reason, the application embodiment introduces semi-supervised learning into GAN to construct a preset radiation source recognition model.

[0102] Figure 3 A flow chart of a method for evaluating classification results provided in an embodiment of the present application includes:

[0103] Step 201, integrating the residual network into the discriminant network. :

[0104] Step 202: configure the scale topology structure to fuse the resolution features of the signal at different scales.

[0105] See also Figure 4 , Figure 4 This is a schematic diagram of a discriminant network based on a multi-scale topological structure and a residual unit provided in the implementation of this application. This embodiment of the application introduces a residual network into the discriminant network to improve the efficiency of the front-end feature transfer, and designs a multi-scale topological structure to fuse the resolution features of the signal at different scales. Figure 4As shown in the figure, the input of the discriminant network is a 1×1000-dimensional time domain signal. After the fully connected layer and dimension transformation, it outputs a 2×512×1 three-dimensional tensor for 2D convolution operation. For the one-dimensional signal envelope R(i)(nT), its own additional nonlinear power amplifier distortion makes it have cross-correlation characteristics in the time domain. In the multi-scale topological structure, by setting one-dimensional convolution kernels of different sizes (1×7, 1×5 and 1×3), the multi-scale topological structure can extract the time cross-domain correlation information of R(i)(nT) from multiple scales vertically, and the three added parallel branches enrich the channel information from the horizontal dimension, thereby improving the multi-dimensional representation ability of the classification network; the cross connection directly transmits the initial features of the front end to the receptive field of the back end, transmits the detailed information of the original time domain signal, and avoids the loss of subtle features and gradient explosion caused by the stacking of convolutional layers. Each residual unit is composed of three convolutional layers. Batch Normalization (BN) and Rectified Linear Unit (ReLU) layers are added after the convolutional layers to prevent gradient explosion. The convolution step size is set to 1 to retain high-dimensional feature maps. The number of convolution kernels of the four residual units is 128, 64, 32, and 16 respectively. After the feature vector is flattened, the supervised classification probability and unsupervised discrimination probability of the discriminant network are output by the Softmax layer and the Sigmoid layer respectively. The input of the generated network is 1×100 sampled Gaussian white noise. Batch Normalization and Leaky Rectified Linear Unit (LReLU) layers are added after each convolutional layer. The convolution kernel size is 1×3 and the step size is set to 1. The addition of convolution layer 6 is to reduce the dimensionality of the front-end feature map and reduce the number of connected nodes.

[0106] As an implementable manner of the embodiment of the present application, when evaluating the binary classification result based on the skew distribution loss function and the labeled samples, the skew distribution loss function can be used for evaluation. The standard cross entropy function for processing a balanced data set is usually expressed as:

[0107]

[0108] Where y i ∈(0,1) is the confidence probability, is the prediction probability, and N is the sample category. CE loss minimizes the training error by assuming that the loss weight is independent of the frequency of the category to which it belongs, that is, it has equal weights for the majority class samples that are easy to classify and the minority class samples that are difficult to classify. However, this will cause the majority class to mislead the direction of gradient descent, so it is necessary to have sufficiently balanced category frequency samples to achieve the generalization of the inter-category separation boundary model. For this reason, we innovatively propose the skewed category distribution loss, which is defined as:

[0109]

[0110] Where η, χ and μ are hyperparameters. Focus loss L FL and class imbalance loss L CI Used to handle class imbalanced data, respectively expressed as:

[0111]

[0112] Where υ and λ represent hyper parameters respectively.

[0113] It should be noted that the above offset distribution loss function is a differentiable function. The minimum value or saddle point of the loss function is obtained through iteration, thereby determining the optimal nonlinear mapping corresponding to the current data set. Consider a network with i neurons, see Figure 5 , Figure 5 A schematic diagram of a basic neuron structure provided in an embodiment of the present application, in which b represents a bias term, σ(·) represents an activation function, and the output z of the i-th neuron i It can be expressed as:

[0114]

[0115] where w ij Represents the jth weight value of the ith neuron, and selects the output a after Softmax activation i It is expressed as:

[0116]

[0117] In the preset radiation source recognition model, the supervision part of the discriminant network uses skew distribution loss to achieve K+1-dimensional classification. According to the chain derivation rule, the gradient of the loss function L for the output of the i-th neuron can be expressed as:

[0118]

[0119] And there is

[0120] Through the above two formulas, we can get:

[0121]

[0122] make Then there is

[0123] It can be expressed as

[0124]

[0125] In the formula It has nothing to do with i and can be regarded as a constant c; and Mi =y i ·(P i +Q i ), so it can be simplified to:

[0126] Where a i represents the predicted value, η and c represent constants, is about the predicted value a i The function is used to adjust the rate of change of weights caused by samples of different frequency classes. Therefore, the gradient of the network for the output of the neuron exists, that is, by comparing the true value y i and the network prediction value a i The network can converge to the global optimal solution by back propagation of gradients.

[0127] In summary, the core of the preset radiation source recognition model is to make full use of unlabeled samples to train the discriminant network to achieve the sharing of front-end weights. Therefore, the parameter setting of the discriminant network has an important influence on the recognition performance. The embodiment of the present application selects the recognition accuracy to measure the network performance. For example, the loss function is set to CE, the batch size is 128, the signal-to-noise ratio is 16dB, and the label ratio of the training set is 0.6. Among them, when the number of convolutional layers in the residual unit is 2, the recognition accuracy rate increases significantly with the increase of the number of units; and when the number of convolutional layers is 3 or 4, the recognition accuracy rate reaches 89% and 90% respectively. At this time, the change in the number of units has little effect on the network performance. This shows that the recognition accuracy of S3CDGAN (radiation source recognition framework for skewed sample distribution) is nonlinearly proportional to the number of convolutional layers and the number of units; and when the number of convolutional layers is 3, the discriminator can obtain the best feature mapping capability. Therefore, the parameter combination setting of 3 convolutional layers and 4 units is adopted in the subsequent.

[0128] Deep neural networks often need to learn the distribution of existing sample features from a large amount of labeled data. The embodiment of the present application performs uniform random sampling on the standard training set to verify the impact of the number of training samples NT on network performance. For example, the training set labeling ratio is 0.6, and the loss function is set to CE. According to the simulation results, the recognition rate under each SNR increases steadily with the increase of NT. When the SNR is 0dB~10dB, the recognition accuracy under different NT conditions is not much different. The recognition rate when NT is 2400 is about 10% higher than NT=1200, that is, the recognition accuracy is reduced by about 10% when NT is missing 50%; when the SNR is higher than 12dB, this indicator rises to 20% at most, mainly because the number of convolutional layers of the discriminator is too many, and overfitting occurs in the case of a small number of samples of NT=1200. It can be considered to improve the recognition rate by adjusting the network parameters. In general, the algorithm can utilize the existing network structure and show strong robustness under different NT conditions, which is suitable for real electronic reconnaissance scenarios.

[0129] To verify the recognition performance of the preset radiation source recognition model proposed in the embodiment of the present application under different numbers of labeled samples NL, the label ratio is first defined as the ratio of the number of labeled samples NL to the number of training samples NT, and the labeled samples are randomly selected from the standard training set and the training set three according to the label ratio to train the semi-supervised classifier, and finally the supervised recognition accuracy is obtained on the standard test set. For example, the loss function is set to SDL, and the SNR is selected as 16dB and 20dB respectively. According to the results, the recognition accuracy of the preset radiation source recognition model under the three types of label ratios in the standard training set is 95.4%, 94.8% and 90.9% respectively, and the recognition accuracy under the three types of label ratios in the training set three is 75.8%, 74.2% and 72.4% respectively, indicating that the increase in the number of labeled samples NL can improve the recognition performance of the network to a certain extent. When the number of labeled samples differs by 40%, the recognition accuracy of the two datasets only differs by 4.5% and 3.4%, indicating that S3CDGAN (radiant source recognition framework for skewed sample distribution) can achieve higher recognition performance with less label information, and the network can converge quickly even when NL is missing, showing strong robustness to label information.

[0130] In the embodiment of the present application, on the standard training set, the recognition accuracy of CE loss, FL loss and SDL loss is not much different, and FL loss and SDL loss do not show obvious advantages. This is because the sample categories in the standard training set are uniformly distributed, so FL loss and SDL loss cannot weight the gradient according to the confidence probability, and the fluctuation of recognition rate mainly comes from the random parameter update of the network. For the four-category skewed data set, the recognition performance of FL loss is higher than that of CE loss, and SDL loss shows the highest recognition accuracy in most cases, which verifies that both FL loss and SDL loss can adaptively adjust the loss value through the confidence probability of difficult and easy classification samples, and add a downward weighted rate for easy-to-classify multi-class samples. On the four-category skewed data set, the overall average recognition accuracy of SDL loss is 5.34% and 2.69% higher than that of CE loss and FL loss, respectively, which verifies the effectiveness of the proposed SDL loss in processing skewed distribution samples.

[0131] Corresponding to the above-mentioned radiation source identification method, the present invention also provides a radiation source identification device. Since the device embodiment of the present invention corresponds to the above-mentioned method embodiment, details not disclosed in the device embodiment can be referred to the above-mentioned method embodiment, and will not be repeated in the present invention.

[0132] Figure 6 A schematic diagram of a radiation source identification device provided in an embodiment of the present disclosure is shown in FIG. Figure 6 As shown, including:

[0133] An input unit 31, used to input the noise vector into a generation network in a preset radiation source identification model;

[0134] An output unit 32, configured to output a signal sample after the convolution operation of the preset radiation source recognition model, so as to deceive the discriminant network in the preset radiation source recognition model;

[0135] A classification unit 33, configured to perform binary classification on the generated network output samples and the unlabeled samples based on the discriminant network;

[0136] The evaluation unit 34 is used to evaluate the binary classification result based on the skew distribution loss function and the labeled samples.

[0137] Furthermore, in a possible implementation of this embodiment, as Figure 7 As shown, the device also includes:

[0138] The construction unit 35 is used to construct a generative adversarial network model before the input unit inputs the noise vector into the generative network in the preset radiation source identification model. The generative adversarial network model at least includes: a generative network and a discriminative network, which are used to optimize the network parameters of the generative network and the discriminative network through semi-supervised training.

[0139] The radiation source identification device provided by the present disclosure, when performing the identification of the radiation source, first, inputs the noise vector into the generation network in the preset radiation source identification model; secondly, outputs the signal sample after the convolution operation of the preset radiation source identification model, so as to deceive the discriminant network in the preset radiation source identification model; based on the discriminant network, the generation network output samples and the unlabeled samples are binary classified; finally, the binary classification results are evaluated based on the skewed distribution loss function and the labeled samples. Compared with the related art, the embodiment of the present application trains the radiation signal based on the preset radiation element identification model to identify its radiation source, and offsets the gradient imbalance caused by the uneven sample category according to the skewed distribution loss function, thereby increasing the contribution of the minority class samples to the total loss, and realizing the identification of the radiation source of the skewed category distribution data.

[0140] Furthermore, in a possible implementation of this embodiment, as Figure 7 As shown, the evaluation unit 34 includes:

[0141] An integration module 341, used to integrate the residual network into the discriminant network;

[0142] The configuration module 342 is used to configure the resolution characteristics of the scale topology fusion signal at different scales.

[0143] Furthermore, in a possible implementation of this embodiment, as Figure 7 As shown, the device also includes:

[0144] A division unit 36, used for dividing the discriminant network into an unsupervised part and a supervised part;

[0145] The loss function corresponding to the supervision part is the standard cross entropy between the labeled samples and the binary classification results;

[0146] The standard cross entropy (CE) function for processing balanced data sets is usually expressed as:

[0147] In the formula, y i ∈(0,1) is the confidence probability, is the predicted probability, N is the sample category, and CE loss minimizes the training error by assuming that the loss weight is independent of the frequency of the category to which it belongs.

[0148] Furthermore, in a possible implementation of this embodiment, as Figure 7 As shown, the evaluation unit 34 is further used to perform evaluation based on the following skew distribution loss function:

[0149]

[0150]

[0151]

[0152] Where η, χ and μ are hyperparameters. FL is the focal loss, L CI It is the class imbalance loss, which is used to deal with class imbalance data.

[0153] It should be noted that the above explanation of the method embodiment is also applicable to the device of this embodiment, and the principle is the same, which is not limited in this embodiment.

[0154] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0155] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A method for identifying a radiation source, characterized in that: include: Inputting the noise vector into a generation network in a preset radiation source identification model; Outputting signal samples after the convolution operation of the preset radiation source recognition model to deceive the discriminant network in the preset radiation source recognition model; Based on the discriminant network, the generated network output samples and unlabeled samples are classified into two categories; Evaluate the binary classification results based on the skewed distribution loss function and labeled samples; The evaluation of the binary classification results based on the skewed distribution loss function and the labeled samples includes: Based on the following skew distribution loss function L SDL To conduct an assessment: Where η, X and μ are hyperparameters, y i ∈(0,1) is the confidence probability, is the predicted probability, and N is the sample category.

2. The identification method according to claim 1, characterized in that: Before inputting the noise vector into the generation network in the preset radiation source identification model, the method further includes: A generative adversarial network model is constructed, wherein the generative adversarial network model at least includes: a generative network and a discriminative network, and is used to optimize network parameters of the generative network and the discriminative network through a semi-supervised training method.

3. The identification method according to claim 1, characterized in that: The binary classification result is evaluated based on the skew distribution loss function and the labeled samples, and the method further includes: Integrating the residual network into the discriminant network; Configure the scale topology to fuse the resolution features of signals at different scales.

4. The identification method according to claim 1, characterized in that: The method further comprises: Dividing the discriminant network into an unsupervised part and a supervised part; The loss function corresponding to the supervision part is the standard cross entropy between the labeled samples and the binary classification results; The standard cross entropy (CE) function for processing balanced data sets is usually expressed as: In the formula, y i ∈(0,1) is the confidence probability, is the predicted probability, N is the sample category, and CE loss minimizes the training error by assuming that the loss weight is independent of the frequency of the category to which it belongs.

5. A radiation source identification device, characterized in that: include: An input unit, used for inputting the noise vector into a generation network in a preset radiation source identification model; An output unit, used to output a signal sample after a convolution operation of the preset radiation source recognition model, so as to deceive a discriminant network in the preset radiation source recognition model; A classification unit, used for performing binary classification on the generated network output samples and the unlabeled samples based on the discriminant network; An evaluation unit, used to evaluate the binary classification results based on the skew distribution loss function and labeled samples; The evaluation unit is also used to evaluate the skew distribution loss function L SDL To conduct an assessment: Where η, χ and μ are hyperparameters, y i ∈(0,1) is the confidence probability, is the predicted probability, and N is the sample category.

6. The identification device according to claim 5, characterized in that: The device also includes: A construction unit is used to construct a generative adversarial network model before the input unit inputs the noise vector into the generative network in the preset radiation source identification model. The generative adversarial network model at least includes: a generative network and a discriminative network, which are used to optimize the network parameters of the generative network and the discriminative network through semi-supervised training.

7. The identification device according to claim 5, characterized in that: The evaluation unit further comprises: A fusion module, used to integrate the residual network into the discriminant network; The configuration module is used to configure the resolution characteristics of the scale topology fusion signal at different scales.

8. The identification device according to claim 5, characterized in that: The device also includes: A division unit, used for dividing the discriminant network into an unsupervised part and a supervised part; The loss function corresponding to the supervision part is the standard cross entropy between the labeled samples and the binary classification results; The standard cross entropy (CE) function for processing balanced data sets is usually expressed as: In the formula, y i ∈(0,1) is the confidence probability, is the predicted probability, N is the sample category, and CE loss minimizes the training error by assuming that the loss weight is independent of the frequency of the category to which it belongs.

Images