Method and system for unknown emitter individual identification based on SEIM

Abstract

The application discloses a kind of unknown radiation source individual identification method and system based on SEIM, it is related to radiation source identification technical field, the method includes inputting target radiation source signal into radiation source individual identification model, obtains identification result;If identification result is known class, then output the radiation source individual name of belonging known class, if identification result is unknown class, then output unknown class;Radiation source individual identification model includes feature extraction module and classification identification module;Feature extraction module is used to extract the semantic feature of target radiation source signal;Classification identification module is used to determine identification result according to the semantic feature of the target radiation source signal;Feature extraction module is the module constructed based on convolutional neural network;Attention mechanism is provided in convolutional neural network, and its loss target function is composed of cross-entropy loss function and center loss function.The application solves the problems of unknown radiation source individual identification difficulty and low feature extraction efficiency.

Description

Technical Field

[0001] This invention relates to the field of radiation source identification technology, and in particular to a method and system for identifying unknown radiation sources based on SEIM. Background Technology

[0002] Specific emitter identification (SEI) is the process of extracting features from received radio frequency signals to identify the individual emitter that generated the signal. It plays an important role in fields such as military communications and spectrum security.

[0003] Traditional radiation source identification primarily relies on manually extracted features, including instantaneous frequency and amplitude, signal bispectral density, fractal dimension, fuzzy function slicing, power spectral density coefficient, and channel state information at the receiver. However, in traditional algorithms, manually extracted features are time-consuming and inefficient, and the performance of the identification system depends on the expert's prior knowledge of the signal parameters.

[0004] In recent years, deep learning technology has gained widespread popularity due to its ability to provide end-to-end solutions from feature extraction to classification. Following the success of deep learning models in image recognition and speech recognition, they have also been applied to signal recognition, demonstrating their powerful advantages in target signal feature extraction and classification. Youssef et al. used Support Vector Machines (SVM) and Convolutional Neural Networks (CNN) for individual radiation source identification. Merchant et al. applied deep learning to detect physical layer properties, using complex baseband error signals in the top domain to identify cognitive radio devices. Qu Lingzhi et al. proposed a method for individual radiation source identification combining a two-layer attention mechanism and a residual network to improve model stability under low signal-to-noise ratio conditions. However, these works largely adhere to the closed-set assumption, meaning that the aforementioned deep learning models can only distinguish known signal categories and do not consider situations where training data for certain classes is unavailable. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for identifying individuals of unknown radiation sources based on SEIM.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] This invention provides a method for identifying individuals from unknown radiation sources based on SEIM, comprising:

[0008] Acquire the target radiation source signal;

[0009] The target radiation source signal is input into the radiation source individual identification model to obtain the identification result;

[0010] If the identification result is a known class, the name of the radiation source individual belonging to the known class will be output; if the identification result is an unknown class, the unknown class will be output.

[0011] The radiation source individual identification model includes a feature extraction module and a classification and identification module;

[0012] The feature extraction module is used to extract semantic features of the target radiation source signal; the feature extraction module is a module built on a convolutional neural network; the convolutional neural network is equipped with an attention mechanism; the loss objective function in the convolutional neural network consists of a cross-entropy loss function and a center loss function;

[0013] The classification and recognition module is used to determine the recognition result based on the distance between the semantic features of the target radiation source signal and the center vector of the known class semantic features.

[0014] Optionally, during the training phase, the feature extraction module updates the convolutional neural network parameters and the known class semantic feature center vectors by minimizing the loss objective function.

[0015] Optionally, the convolutional neural network includes low-level hidden layers and high-level hidden layers;

[0016] The low hidden layer includes multiple convolutional layers and multiple pooling layers, and the convolutional layers and pooling layers are distributed alternately; wherein, the first convolutional layer and the last convolutional layer are each provided with a convolutional block attention unit; the convolutional block attention unit includes a spatial attention subunit and a channel attention subunit;

[0017] The high layer includes a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer connected in sequence; the third fully connected layer is used to output the semantic features of the target radiation source signal or the sample radiation source signal; the output layer is used to output the predicted label corresponding to the semantic features of the sample radiation source signal.

[0018] Optionally, the classification and recognition module includes:

[0019] The feature clustering unit is used to perform clustering processing on the semantic features of the target radiation source signal to obtain the target clustering vector;

[0020] Category differentiation unit, used for:

[0021] Calculate the Euclidean distance between the target clustering vector and the known class semantic feature center vectors;

[0022] When the Euclidean distance is less than a set threshold, the category of the target radiation source signal is marked as a known class, and the identification result is determined as a known class;

[0023] When the Euclidean distance is greater than or equal to the set threshold, the category of the target radiation source signal is marked as unknown, and the identification result is determined as unknown.

[0024] Optionally, the feature clustering unit is further used to perform clustering processing on the semantic features of the target radiation source signal using the T-distribution-random neighbor embedding algorithm to obtain the target clustering vector.

[0025] Optionally, the category differentiation unit is further configured to:

[0026] Calculate the Euclidean distance between the target clustering vector and the semantic feature center vector of each known class;

[0027] When the Euclidean distance is less than a set threshold, the category of the target radiation source signal is marked as a known class, and the identification result is determined as a known class;

[0028] When all the Euclidean distances are greater than or equal to the set threshold, the category of the target radiation source signal is marked as unknown, and the identification result is determined as unknown.

[0029] Optionally, when the Euclidean distance is less than a set threshold, the target radiation source signal is classified as a known class. The class differentiation unit is used to:

[0030] When multiple Euclidean distances are less than a set threshold, the known class corresponding to the smallest Euclidean distance is determined as the category of the target radiation source signal and marked.

[0031] When only one of the Euclidean distances is less than a set threshold, the Euclidean distance less than the set threshold is determined as the minimum Euclidean distance, and then the known class corresponding to the minimum Euclidean distance is determined as the category of the target radiation source signal and marked.

[0032] Optionally, the set threshold is determined based on the three sigma rule of thumb for identifying outliers.

[0033] Optionally, if the identification result is a known class, the name of the radiation source individual belonging to the known class is output, specifically including:

[0034] If the identification result is a known class, output the name of the radiation source individual of the known class corresponding to the minimum Euclidean distance.

[0035] This invention also provides an individual identification system for unknown radiation sources based on SEIM, comprising:

[0036] The signal acquisition module is used to acquire the signal from the target radiation source;

[0037] The identification result determination module is used to input the target radiation source signal into the radiation source individual identification model to obtain the identification result;

[0038] The output module is used to output the name of the radiation source individual belonging to the known class if the identification result is a known class, and to output the unknown class if the identification result is an unknown class.

[0039] The radiation source individual identification model includes a feature extraction module and a classification and identification module;

[0040] The feature extraction module is used to extract semantic features of the target radiation source signal; the feature extraction module is a module built on a convolutional neural network; the convolutional neural network is equipped with an attention mechanism; the loss objective function in the convolutional neural network consists of a cross-entropy loss function and a center loss function;

[0041] The classification and recognition module is used to determine the recognition result based on the distance between the semantic features of the target radiation source signal and the center vector of the known class semantic features.

[0042] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0043] This invention first utilizes a convolutional neural network to automatically extract semantic features from samples. An attention mechanism is added to the convolutional neural network to enhance important information and reduce interference from irrelevant information. The cross-entropy loss function and the center loss function are combined to reduce intra-class distances and increase out-of-class distances of the semantic features of radiation sources, thereby improving the model's classification performance. Then, classification is performed based on the distance metric of the features. This invention solves the problems of difficulty in identifying unknown radiation sources and low feature extraction efficiency. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 A flowchart illustrating the method for identifying unknown radiation sources based on SEIM, as provided in an embodiment of the present invention;

[0046] Figure 2 This is a structural diagram of the radiation source individual identification model provided in an embodiment of the present invention;

[0047] Figure 3This is a structural diagram of the convolutional block attention unit provided in an embodiment of the present invention;

[0048] Figure 4 A structural diagram of a convolutional neural network provided in an embodiment of the present invention;

[0049] Figure 5 The segmented standardized signal diagram provided in the embodiments of the present invention;

[0050] Figure 6 Training loss curves provided for embodiments of the present invention;

[0051] Figure 7 This is a clustering effect diagram during the training phase provided in an embodiment of the present invention;

[0052] Figure 8 The confusion matrix diagram obtained after inputting a test set containing unknown classes into the model, as provided in the embodiments of the present invention; Figure 8 (a) is the confusion matrix obtained after inputting a test set containing one unknown class into the model; Figure 8 (b) is the confusion matrix obtained after inputting a test set containing two unknown classes into the model;

[0053] Figure 9 This is a closed-set identification confusion matrix diagram provided in an embodiment of the present invention;

[0054] Figure 10 A visualization of the test set classification effect provided in an embodiment of the present invention;

[0055] Figure 11 This is a schematic diagram of the structure of an individual identification system for unknown radiation sources based on SEIM provided in an embodiment of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] Zero-shot learning is an approach to addressing the problems described in the background section. It enables the prediction of unknown classes using semantic knowledge and has been widely applied to image classification tasks. Palatucci et al. proposed a zero-shot learning framework that can utilize a semantic knowledge base to predict unknown classes not present in the training set. Existing techniques propose an embedding model that explores the functional relationship between text and image features to transfer inter- and intra-model labels for image annotation. DeViSE et al. extended deep architectures to learn visual and semantic embeddings, enabling the identification of visual objects using labeled image data and semantic information collected from unannotated text. The efficiency of zero-shot learning in image processing mainly benefits from semantic features that can be manually defined through high-level descriptions; however, semantic features of signals are difficult to describe and manually specify. Hao Yunfei et al. proposed an open-set recognition method for signal modulation patterns based on generative adversarial networks, designing a reconstruction discriminant network model suitable for one-dimensional signal data, which can achieve open-set recognition of signal modulation patterns. However, the application of zero-shot learning to individual radiation source identification has been rarely studied.

[0059] To address the challenges of identifying unknown radiation sources and the low efficiency of feature extraction, this invention provides a zero-shot learning-based Specific Emitter Identification Model (SEIM). First, a Convolutional Neural Network (CNN) is used to automatically extract semantic features from the samples. An attention module is added to the CNN to enhance important information and reduce interference from irrelevant information. Combined cross-entropy loss and center loss reduce the intra-class distance of the radiation source semantic features and increase the intra-class distance, thereby improving the model's classification performance. Then, classification is performed based on the distance metric of the features. In experiments, a USRP2974 device was used to collect RF IQ signals from seven walkie-talkies of four different models, and closed-set and open-set recognition tests were conducted. The results show that SEIM can maintain the recognition rate of known classes and identify multiple unknown classes, with an average recognition rate of over 90% for both known and unknown classes.

[0060] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for identifying individuals of unknown radiation sources based on SEIM, which includes the following steps.

[0061] Step 100: Acquire the target radiation source signal.

[0062] Step 200: Input the target radiation source signal into the radiation source individual identification model to obtain the identification result.

[0063] Step 300: If the identification result is a known class, output the name of the radiation source individual belonging to the known class; if the identification result is an unknown class, output the unknown class.

[0064] The radiation source individual identification model includes a feature extraction module and a classification and identification module.

[0065] The feature extraction module is used to extract the semantic features of the target radiation source signal; the feature extraction module is a module built on the basis of a convolutional neural network; the convolutional neural network is equipped with an attention mechanism; the loss objective function in the convolutional neural network consists of a cross-entropy loss function and a center loss function.

[0066] The classification and recognition module is used to determine the recognition result based on the distance between the semantic features of the target radiation source signal and the center vector of the known class semantic features.

[0067] The radiation source individual identification model described in the embodiments of the present invention will be introduced below with reference to the accompanying drawings.

[0068] like Figure 2 As shown, the radiation source individual identification model mainly consists of two parts: a feature extraction module and a classification and recognition module. The feature extraction module is built on a convolutional neural network. During the training phase, the feature extraction module updates the parameters of the convolutional neural network and the known class semantic feature center vectors by minimizing the loss objective function. After the feature extraction module is trained, a test sample radiation source signal is input. The feature extraction module extracts the semantic features of the test sample radiation source signal according to the trained convolutional neural network parameters, and then sends them to the classification and recognition module. The classification and recognition module calculates the distance between the semantic features of the test sample radiation source signal and the known class semantic feature center vectors, and compares this distance with a set threshold to obtain the recognition result, thus achieving the purpose of identifying the radiation source signal.

[0069] The detailed process of the feature extraction module is as follows:

[0070] During training, the feature extraction module uses a convolutional neural network (CNN) to extract semantic features from the radiation source signals of the training samples. The CNN uses a gradient backpropagation algorithm to reduce loss and achieve the optimization objective. To ensure that the semantic features can effectively assist the model in recognizing known and unknown classes, a reasonable loss objective function needs to be designed to control the feature extraction process of the CNN. Therefore, this embodiment of the invention considers the simultaneous use of cross-entropy loss and center loss functions.

[0071] The cross-entropy loss function is used to enhance the discriminative power between different classes. It is expressed as the difference between the true probability distribution and the predicted probability distribution, and its mathematical expression is as follows:

[0072]

[0073] Where, x i For the training sample radiation source signal, F(x) i C(F(x)) represents the semantic features of the radiation source signals in the training samples extracted by the convolutional neural network. i y represents the predicted label output by the convolutional neural network based on the semantic features of the radiation source signals in the training samples. i The true labels for the radiation source signals of the training samples.

[0074] The center loss function is used to encourage semantic features of the same class to approach a center, indirectly separating them from other classes. The extracted semantic features are essentially high-dimensional vectors; therefore, the objective of the center loss function can be expressed as reducing the vector distance between semantic features of the same class and increasing the vector distance between semantic features of different classes during the iteration process. Its mathematical expression is as follows:

[0075]

[0076] in, Indicates label y i The semantic feature center vector of a class, as y i Changes in the semantic features of a class necessitate updates. Ideally, the entire training dataset should be considered, and the semantic features of each class should be averaged in each iteration. In application, based on... Updated in each batch, where 'a' is the learning rate. The calculation method is as follows:

[0077]

[0078] Wherein, δ(y) i =y j ) is a conditional equation; when the condition inside the parentheses is satisfied, δ(y) i =y j =1 otherwise =0, which ensures that the class semantic center is updated only when the labels are the same, i.e., they belong to the same class.

[0079] The total loss L used to train the convolutional neural network is the cross-entropy loss L. ce and central loss L ct The sum of , where λ is a hyperparameter representing the weight of the center loss:

[0080] L = L ct +λL ce (4).

[0081] After designing the loss objective function, a convolutional neural network needs to be built.

[0082] The low-hidden layers of a convolutional neural network consist of four convolutional layers and four pooling layers, with the convolutional and pooling layers alternating. Convolutional Block Attention Modules (CBAMs) are added to the first and fourth convolutional layers. A CBAM is a general-purpose unit containing both channel attention and spatial attention subunits. Its function is to autonomously learn the importance of each channel and then adjust the channel weights based on this importance to improve network performance. Its structure is as follows: Figure 3 As shown:

[0083] The higher layers consist of a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer connected sequentially. The third fully connected layer outputs the semantic features of the target radiation source signal or the sample radiation source signal. The output layer outputs the predicted labels corresponding to the semantic features of the sample radiation source signal, and during training, these are used to calculate the center loss and cross-entropy loss, respectively. The overall structure of the convolutional neural network is as follows: Figure 4 As shown.

[0084] The classification and recognition module includes:

[0085] The feature clustering unit is used to perform clustering processing on the semantic features of the target radiation source signal to obtain the target clustering vector.

[0086] One example is: using the T-distribution-random neighbor embedding algorithm to cluster the semantic features of the target radiation source signal to obtain the target cluster vector.

[0087] Category differentiation unit, used for:

[0088] Calculate the Euclidean distance between the target clustering vector and the known class semantic feature center vectors;

[0089] When the Euclidean distance is less than a set threshold, the category of the target radiation source signal is marked as a known class, and the identification result is determined as a known class;

[0090] When the Euclidean distance is greater than or equal to the set threshold, the category of the target radiation source signal is marked as unknown, and the identification result is determined as unknown.

[0091] One example is: the category differentiation unit is further used for:

[0092] Calculate the Euclidean distance between the target clustering vector and the semantic feature center vector of each known class;

[0093] When the Euclidean distance is less than a set threshold, the category of the target radiation source signal is marked as a known class, and the identification result is determined as a known class;

[0094] For example: when multiple Euclidean distances are less than a set threshold, the known class corresponding to the smallest Euclidean distance is determined as the category of the target radiation source signal and marked; when only one Euclidean distance is less than the set threshold, the Euclidean distance less than the set threshold is determined as the smallest Euclidean distance, and then the known class corresponding to the smallest Euclidean distance is determined as the category of the target radiation source signal and marked.

[0095] When all the Euclidean distances are greater than or equal to the set threshold, the category of the target radiation source signal is marked as unknown, and the identification result is determined as unknown.

[0096] Based on this, the embodiment of the present invention, which outputs the name of the radiation source individual belonging to the known class if the identification result is of a known class, specifically includes:

[0097] If the identification result is a known class, output the name of the radiation source individual of the known class corresponding to the minimum Euclidean distance.

[0098] The detailed process of the classification and recognition module is as follows:

[0099] The classification and recognition module is used to distinguish between known and unknown classes based on the semantic features extracted by the feature extraction module. The semantic features extracted by the feature extraction module are high-dimensional vectors. The model uses the T-distributed stochastic neighbor embedding (TSNE) algorithm to visualize and form clusters. The optimal performance of the model is obtained by observing the distribution of the clusters, and then classification is achieved based on the distance metric.

[0100] The semantic features F(x) of each radiation source signal i ) and known class semantic feature center vector The Euclidean distance is:

[0101]

[0102] After calculating the semantic features of the radiation source signal and the distances to the semantic feature centers of several known classes, the radiation source signal can be distinguished as belonging to a known class or an unknown class by comparing the distances with a set threshold θ. The threshold is set based on the three Sigma rule of thumb, which is often used to identify outliers. As shown in Table 1, μ is the mean and σ is the standard deviation. Generally, it can be assumed that the values ​​of a set of detection data are almost entirely concentrated in the interval (μ-3σ, μ+3σ), and data outside this range can be considered outliers. Therefore, the threshold θ is set as θ = 3σ, where σ is the standard deviation of the semantic features of the sample radiation source signal and its class semantic center.

[0103] Table 1. Three Sigma Rules of Experience

[0104]

[0105] When d < θ, the radiation source signal is classified as a known class, and its label can be obtained further by searching for the minimum distance:

[0106]

[0107] Conversely, if the radiation source signal is classified as unknown, a new label representing the unknown class is created for it. and class semantic feature center vector The data is then added to the label set and the known class semantic feature center vector set. The overall steps of the determination process are shown in Table 2.

[0108] Table 2. Judgment Process of Classification and Recognition Module

[0109]

[0110] To evaluate the feasibility and effectiveness of the radiation source individual identification model proposed in this embodiment of the invention, the experiment used measured walkie-talkie signals as test objects, and set up two open set identification tasks with one unknown class and two unknown classes, as well as a closed set identification task. During model training, the PyTorch framework under the Windows 11 platform was used for training and testing on an NVIDIA RTX A6000.

[0111] The data used in the experiment consisted of radio frequency (IQ) signals from seven walkie-talkies of four different models, collected by a single USRP device. The USRP-2974 receiving device was connected to the host computer via an antenna. The radiation source signal acquisition platform built using LabVIEW software could directly collect and store the radio frequency (IQ) signals of the walkie-talkies during operation. The seven walkie-talkies were one BAOFENG model, two SHIDA models, one TCL model, and three TYT models. All walkie-talkies used FM operation mode with a frequency set to 450MHz.

[0112] Each device acquired three segments of RF IQ signal, each segment consisting of approximately one minute of signal sampling points with a length of 2,931,000. Before using these signals, segmentation and standardization processing is required, i.e., slicing and normalizing these continuous sampling points. Each sliced ​​signal is considered a sample. The segmented and standardized data is as follows: Figure 5 As shown.

[0113] The processed dataset consists of 1000 samples per device, with each sample measuring 8192*2 pixels. Details are shown in Table 3. The training and test sets are randomly divided in an 8:2 ratio. In the experiment, different devices are selected as the unknown class based on the classification task. When a device is in the unknown class, its samples are not allocated to the training set; only 20% of the samples are selected as the test set.

[0114] Table 3. Walkie-talkie signal data set

[0115]

[0116] The partitioned dataset is fed into the model for training. The training iterations are set to 200, the learning rate to 0.001, and the weight λ of the center loss is set to 0.3. Figure 6 The cross-entropy loss, center loss, and the total loss curve formed by the two are given during the training process.

[0117] To obtain the optimal performance of the feature extraction module during training, model parameters were saved for every 25 neurons in the experiment, and semantic features extracted from the training set were obtained. Then, the TSNE visualization was used to observe the classification effect of known classes, and the test set was input into the model with the best classification effect for recognition. Figure 7 The clustering manifolds respectively demonstrate the clustering effect of the six known categories in the training set during the training phase, given the existence of one unknown class.

[0118] The training process showed that intra-class aggregation and inter-class separation reached relative optimality when the number of neurons was 125. Therefore, subsequent experiments used the model saved when the number of neurons was 125 for testing.

[0119] The experiment identified signal data from 7 walkie-talkies, representing 7 types of radiation sources. Tests were conducted with 1 unknown class and 2 known classes. The class containing multiple devices of the same model was selected as the unknown class to verify the model's performance. In addition, closed-set recognition was performed on a regular convolutional neural network for comparison. Table 4 shows the relevant recognition results, in which the unknown class is labeled.

[0120] Table 4 Recognition Results

[0121]

[0122]

[0123] Figure 8 (a) and Figure 8 (b) shows the confusion matrix obtained after inputting the model with a test set containing unknown classes, corresponding to the cases with 1 and 2 unknown classes, respectively. Figure 8 As can be seen, the recognition rate decreases when there are two unknown classes compared to when there is one unknown class. This is because the radiation source individual recognition model uses semantic features as the classification basis. When more unknown classes appear, the semantic features of the corresponding more unknown data have not undergone loss control during training, resulting in poor intra-class clustering and inter-class separation. This affects the recognition rate of other known classes. Figure 10 This is also reflected in the clustering manifold.

[0124] Figure 9 For the closed-set recognition test results, the same direct training data method was used, employing a traditional convolutional neural network for recognition and classification. With all classes known, the average recognition rate was 95.57%. In comparison, although the radiation source individual recognition model differs from the traditional convolutional neural network classification method, in the presence of one unknown class, the average recognition rate for known classes reaches 96.5%. Even with unknown classes present, the average recognition rate for known classes remains above 90%, verifying that the model proposed in this embodiment of the invention can feasiblely distinguish different categories based on the vector distance of semantic features. The radiation source individual recognition model can identify unknown classes while maintaining the recognition rate for known classes.

[0125] Figure 10The results demonstrate the clustering manifold formed by TSNE dimensionality reduction after semantic feature extraction from the test set. It can be seen that in the case of one unknown class, the clustering of the unknown class SHIDA2 separates well from the known classes, achieving a 100% recognition rate when judging based on distance. In the case of two unknown classes (SHIDA1 and TYT2), the clustering of unknown class TYT2 (unknown2) overlaps with the clustering of known class TYT3 (6), resulting in recognition rates of 81% and 86% respectively. These rates are lower than when the TYT model walkie-talkie is considered an unknown class. This can be explained by the similarity of signal data from devices of the same model. The closed-set recognition results show that individuals of the TYT model are more difficult to distinguish than those of the SHIDA model. The radiation source individual identification model, trained on known TYT1 and TYT3 data, is used to extract features from the unknown class TYT2. However, because TYT2 data is similar to TYT1 and TYT3, and TYT2 was not used in the initial training, the feature extraction module cannot use loss penalty to find subtle differences between TYT2 and TYT1 and TYT3. This results in similar extracted semantic features, affecting the discrimination. Nevertheless, the radiation source individual identification model still maintains a high recognition rate for the unknown class. Experiments show that the average recognition rate of the radiation source individual identification model for the unknown class reaches over 90%, demonstrating its effectiveness in identifying the unknown class.

[0126] This invention provides a zero-shot learning-based model for identifying unknown radiation sources. First, signal data from seven walkie-talkies of four different models were collected to form a dataset, which was then divided into unknown classes according to the experimental task. Next, a convolutional neural network (CNN) incorporating cross-entropy loss, center loss, and an attention mechanism was used to train the known class data samples. The trained CNN was then used to extract semantic features from the known and unknown classes in the test set. Finally, the identification of known and unknown classes was achieved based on the clustering manifold and distance metric derived from these semantic features. Experiments show that the zero-shot learning-based radiation source classification and identification model proposed in this invention is feasible and maintains an overall recognition rate of over 90% even when multiple unknown classes exist.

[0127] Example 2

[0128] In order to implement the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, an individual identification system for unknown radiation sources based on SEIM is provided below.

[0129] like Figure 11 As shown, this embodiment of the invention provides an individual identification system for unknown radiation sources based on SEIM, comprising:

[0130] Signal acquisition module 1 is used to acquire the target radiation source signal.

[0131] The identification result determination module 2 is used to input the target radiation source signal into the radiation source individual identification model to obtain the identification result.

[0132] Output module 3 is used to output the name of the radiation source individual belonging to the known class if the identification result is a known class, and to output the unknown class if the identification result is an unknown class.

[0133] The radiation source individual identification model includes a feature extraction module and a classification and identification module;

[0134] The feature extraction module is used to extract the semantic features of the target radiation source signal; the feature extraction module is a module built on the basis of a convolutional neural network; the convolutional neural network is equipped with an attention mechanism; the loss objective function in the convolutional neural network consists of a cross-entropy loss function and a center loss function.

[0135] The classification and recognition module is used to determine the recognition result based on the distance between the semantic features of the target radiation source signal and the center vector of the known class semantic features.

[0136] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0137] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for identifying individuals of unknown radiation sources based on SEIM, characterized in that, include: Acquire the target radiation source signal; The target radiation source signal is input into the radiation source individual identification model to obtain the identification result; If the identification result is a known class, the name of the radiation source individual belonging to the known class will be output; if the identification result is an unknown class, the unknown class will be output. The radiation source individual identification model includes a feature extraction module and a classification and identification module; The feature extraction module is used to extract the semantic features of the target radiation source signal. The feature extraction module is built upon a convolutional neural network (CNN). The CNN incorporates an attention mechanism. The objective loss function of the CNN consists of a cross-entropy loss function and a center loss function. The CNN includes low-level hidden layers and high-level hidden layers. The low-level hidden layers include multiple convolutional layers and multiple pooling layers, which are alternately distributed. Both the first and last convolutional layers contain convolutional block attention units. Each convolutional block attention unit includes a spatial attention subunit and a channel attention subunit. The high-level hidden layers consist of a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer connected sequentially. The third fully connected layer outputs the semantic features of the target radiation source signal or the sample radiation source signal. The output layer outputs the predicted labels corresponding to the semantic features of the sample radiation source signal. The classification and recognition module is used to determine the recognition result based on the distance between the semantic features of the target radiation source signal and the center vector of the known class semantic features.

2. The method for identifying individuals of unknown radiation sources based on SEIM according to claim 1, characterized in that, During the training phase, the feature extraction module updates the convolutional neural network parameters and the known class semantic feature center vectors by minimizing the loss objective function.

3. The method for identifying individuals of unknown radiation sources based on SEIM according to claim 1, characterized in that, The classification and recognition module includes: The feature clustering unit is used to perform clustering processing on the semantic features of the target radiation source signal to obtain the target clustering vector; Category differentiation unit, used for: Calculate the Euclidean distance between the target clustering vector and the known class semantic feature center vectors; When the Euclidean distance is less than a set threshold, the category of the target radiation source signal is marked as a known class, and the identification result is determined as a known class; When the Euclidean distance is greater than or equal to the set threshold, the category of the target radiation source signal is marked as unknown, and the identification result is determined as unknown.

4. The method for identifying individuals of unknown radiation sources based on SEIM according to claim 3, characterized in that, The feature clustering unit is further used to perform clustering processing on the semantic features of the target radiation source signal using the T-distribution-random neighbor embedding algorithm to obtain the target clustering vector.

5. The method for identifying unknown radiation sources based on SEIM according to claim 3, characterized in that, The category differentiation unit is further used for: Calculate the Euclidean distance between the target clustering vector and the semantic feature center vector of each known class; When the Euclidean distance is less than a set threshold, the category of the target radiation source signal is marked as a known class, and the identification result is determined as a known class; When all the Euclidean distances are greater than or equal to the set threshold, the category of the target radiation source signal is marked as unknown, and the identification result is determined as unknown.

6. The method for identifying individuals of unknown radiation sources based on SEIM according to claim 5, characterized in that, When the Euclidean distance is less than a set threshold, the target radiation source signal is classified as a known class. The class differentiation unit is used to: When multiple Euclidean distances are less than a set threshold, the known class corresponding to the smallest Euclidean distance is determined as the category of the target radiation source signal and marked. When only one of the Euclidean distances is less than a set threshold, the Euclidean distance less than the set threshold is determined as the minimum Euclidean distance, and then the known class corresponding to the minimum Euclidean distance is determined as the category of the target radiation source signal and marked.

7. The method for identifying unknown radiation sources based on SEIM according to claim 3, characterized in that, The set threshold is determined based on the three Sigma rule of thumb for identifying outliers.

8. The method for identifying individuals of unknown radiation sources based on SEIM according to claim 6, characterized in that, If the identification result is a known class, then output the name of the radiation source individual belonging to the known class, specifically including: If the identification result is a known class, output the name of the radiation source individual of the known class corresponding to the minimum Euclidean distance.

9. A system for identifying individuals from unknown radiation sources based on SEIM, characterized in that, include: The signal acquisition module is used to acquire the signal from the target radiation source; The identification result determination module is used to input the target radiation source signal into the radiation source individual identification model to obtain the identification result; The output module is used to output the name of the radiation source individual belonging to the known class if the identification result is a known class, and to output the unknown class if the identification result is an unknown class. The radiation source individual identification model includes a feature extraction module and a classification and identification module; The feature extraction module is used to extract the semantic features of the target radiation source signal. The feature extraction module is built upon a convolutional neural network (CNN). The CNN incorporates an attention mechanism. The objective loss function of the CNN consists of a cross-entropy loss function and a center loss function. The CNN includes low-level hidden layers and high-level hidden layers. The low-level hidden layers include multiple convolutional layers and multiple pooling layers, which are alternately distributed. Both the first and last convolutional layers contain convolutional block attention units. Each convolutional block attention unit includes a spatial attention subunit and a channel attention subunit. The high-level hidden layers consist of a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer connected sequentially. The third fully connected layer outputs the semantic features of the target radiation source signal or the sample radiation source signal. The output layer outputs the predicted labels corresponding to the semantic features of the sample radiation source signal. The classification and recognition module is used to determine the recognition result based on the distance between the semantic features of the target radiation source signal and the center vector of the known class semantic features.

Images