A Method and System for Individual Recognition of Open Set Radiation Sources Based on Multiple Gaussian Prototype Learning

By combining multiple Gaussian prototype learning with a one-dimensional deep attention residual network and a joint loss function of generation constraints and discriminative constraints, the problems of low recognition performance and insufficient robustness in individual radiation source identification are solved, especially improving the recognition accuracy in open set scenarios.

CN118171142BActive Publication Date: 2026-11-17Chinese People's Liberation Army Cyberspace Force Information Engineering University
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410265037.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-08
Publication Date
2026-11-17
Estimated Expiration
2044-03-08

AI Technical Summary

Technical Problem

Existing deep learning and prototype learning methods suffer from low recognition performance and insufficient robustness in individual radiation source identification, especially in open set scenarios where they perform poorly in identifying unknown signals and fail to effectively address the class imbalance problem.

Method used

We employ multiple Gaussian prototype learning combined with a one-dimensional deep attention residual network (DAResNet). By adding a channel attention module and a GRU layer to the initial deep attention residual network, we construct the MGPL-DAResNet recognition model. We then train the model using a joint loss function of generation constraints and discriminative constraints to extract latent features of radiation source signals and perform open set recognition.

Benefits of technology

It improves the accuracy and robustness of individual radiation source identification, especially in open set scenarios where it improves identification performance by 13.1% compared to other methods, effectively solving the class imbalance problem.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118171142B_ABST
    Figure CN118171142B_ABST
Patent Text Reader

Abstract

This invention provides a method and system for identifying open-set radiation sources based on multiple Gaussian prototype learning. The method includes: preprocessing the collected radiation source signals to obtain a one-dimensional I / Q sequence; dividing the one-dimensional I / Q sequence into a test set and a training set according to a preset ratio; establishing an MGPL-DAResNet recognition model based on multiple Gaussian prototype learning and a constructed one-dimensional deep attention residual network; training the MGPL-DAResNet recognition model using the training set and iteratively updating the network parameters to obtain the Gaussian prototype and network weights corresponding to each radiation category; and, using a set distance threshold, performing open-set recognition and classification on the test set using the Gaussian prototype and network weights corresponding to each radiation category, and outputting the classification results. This invention solves the technical problems of low recognition performance and insufficient robustness when using existing methods combining deep learning and prototype learning for recognition.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of radiation source individual identification technology, specifically relating to an open set radiation source individual identification method and system based on multiple Gaussian prototype learning. Background Technology

[0002] Specific Emitter Identification (SEI), also known as radiation source fingerprinting or specific radiation source identification, refers to the technique of identifying individual radiation sources by utilizing subtle differences in the modulation process of their signals. These subtle differences are unique to each radiation source and unavoidable, thus making the radiation source fingerprint unique. In recent years, with the rapid development of information technology and increasingly complex electromagnetic interference, traditional signal identification methods can no longer meet the needs of specific radiation source identification in complex environments. SEI methods, by extracting subtle features of individual radiation sources and combining them with traditional analysis methods, can enhance the identification of external devices and improve the security performance of network communications.

[0003] Currently, deep learning technology is developing rapidly, and many scholars are using various deep learning networks such as residual networks and convolutional neural networks (CNNs) to extract features and identify radiation source signals. However, existing SEI methods mainly focus on identification in closed-set scenarios, with limited research on individual radiation source identification in open-set scenarios. In practical applications, receivers are more often exposed to open electromagnetic environments, which means that during the identification of individual radiation sources, some unknown signals outside the categories in the training set may appear. This makes open-set identification more valuable in practical applications. Therefore, researching how to accurately and efficiently identify unknown radiation source signals has become particularly important.

[0004] In recent years, scholars have continuously researched the identification of radiation sources in open-set scenarios. Currently, the main methods in this field include those based on the SoftMax algorithm, those based on the OpenMax algorithm, and improvements based on prototype learning. For example, in 2020, Andrew Draganov et al. proposed an unsupervised class-based distance learning method suitable for open-set identification, which can effectively utilize information from unlabeled data to improve recognition performance. However, this method lacks labeled data and cannot supervise and adjust the training process, potentially leading to overfitting or underfitting of the model.

[0005] Compared to traditional convolutional neural networks, Gaussian prototype networks have better recognition performance, but they do not take into account the class imbalance problem when performing recognition. If the number of samples of a certain class in the test sample set is less than that of other classes, it may lead to poor classification performance for that class.

[0006] Therefore, although Gaussian prototype learning and deep learning networks have made some progress in the field of radiation source individual identification, some technical problems still exist, such as the need for further improvement in recognition performance and insufficient robustness of the models, requiring further research and improvement. Summary of the Invention

[0007] To address the technical problems of low recognition performance and insufficient robustness in existing methods that combine deep learning and prototype learning to identify specific emitters, this invention discloses a method and system for identifying individual open-set radiation sources based on multiple Gaussian prototype learning.

[0008] The technical solution to achieve the purpose of the invention is as follows:

[0009] The first aspect of this invention provides a method for identifying open-set radiation sources based on multiple Gaussian prototype learning, comprising the following steps:

[0010] S100. The collected radiation source signal is preprocessed to obtain a one-dimensional I / Q sequence, and the one-dimensional I / Q sequence is divided into a test set and a training set according to a preset ratio.

[0011] S200. An MGPL-DAResNet recognition model is established based on a one-dimensional deep attention residual network constructed using multiple Gaussian prototype learning. The constructed one-dimensional deep attention residual network includes a channel attention module set at the output end of each residual block in the initial deep attention residual network.

[0012] S300. Input the training set into the MGPL-DAResNet recognition model, extract the latent features of the training set, and calculate the distance between the latent features and each Gaussian prototype corresponding to each radiation source category in the multiple Gaussian prototype learning.

[0013] S400. Based on the joint loss function of generation constraints and discrimination constraints, the MGPL-DAResNet recognition model is iteratively updated according to the distance between the latent features of the training set and the Gaussian prototype, so as to obtain the Gaussian prototype and network weights corresponding to each radiation category.

[0014] S500. By setting a distance threshold, the Gaussian prototype and its network weights corresponding to each radiation category in the trained MGPL-DAResNet recognition model are used to perform open-set recognition and classification on the test set and output the classification results.

[0015] Furthermore, in step S200, the constructed one-dimensional deep attention residual network further includes: setting a GRU layer at the output of the channel attention module of the last residual block, wherein the GRU (Gated Recurrent Unit) layer is a recurrent neural network unit or a gated recurrent unit. The GRU layer plays a role in modeling temporal relationships, extracting features, and understanding context when processing sequence data, extracting more useful and richer temporal features from the sequence, and further improving the performance and expressive power of the network.

[0016] Furthermore, in step S200, the constructed one-dimensional deep attention residual network further includes: setting an average pooling layer and a fully connected layer sequentially at the output of the GRU layer.

[0017] Furthermore, in step S200, the architecture of the multiple Gaussian prototype learning includes an encoder, a decoder, and multiple Gaussian prototypes corresponding to each radiation source category.

[0018] The encoder is used to extract the latent features of the one-dimensional I / Q sequence and the original distribution of the one-dimensional I / Q sequence for encoding, and store the encoding result in the latent space;

[0019] The decoder samples from the latent space and encodes the one-dimensional I / Q sequence according to the network weights in the MGPL-DAResNet recognition model.

[0020] Furthermore, in step S400, the expression for the discriminant constraint is: In Equation (1), l() represents the KL divergence, γ represents the temperature parameter of the smoothed output probability, x represents the latent feature, y represents the radiation source category, m represents the Gaussian prototype, and A is the number of known radiation source categories, with each radiation source category represented by B Gaussian prototypes.

[0021] A second aspect of the present invention provides an open-set radiation source individual identification system based on multiple Gaussian prototype learning, the open-set radiation source individual identification system comprising:

[0022] The signal preprocessing module is used to preprocess the collected radiation source signal to obtain a one-dimensional I / Q sequence, and divide the one-dimensional I / Q sequence into a test set and a training set according to a preset ratio.

[0023] A network construction module is used to set a channel attention module at the output end of each residual block in the initial deep attention residual network to obtain a one-dimensional deep attention residual network.

[0024] An open set recognition network model building module is used to build an MGPL-DAResNet recognition model using multiple Gaussian prototype learning and the one-dimensional deep attention residual network.

[0025] The model training module is used to input the training set into the MGPL-DAResNet recognition model, extract the latent features of the training set, calculate the distance between the latent features and each Gaussian prototype corresponding to each radiation source category in the multiple Gaussian prototype learning, and iteratively update the MGPL-DAResNet recognition model based on the joint loss function of generation constraints and discrimination constraints, according to the distance between the latent features of the training set and the Gaussian prototypes, to obtain the Gaussian prototype and network weights corresponding to each radiation category.

[0026] The radiation source classification module is used to perform open-set recognition and classification on the test set by using the Gaussian prototype and its network weights corresponding to each radiation category in the trained MGPL-DAResNet recognition model, based on a set distance threshold, and output the classification result.

[0027] Furthermore, the network building module is also used to set a GRU layer at the output of the channel attention module of the last residual block.

[0028] Furthermore, the network building module is also used to sequentially set an average pooling layer and a fully connected layer at the output of the GRU layer.

[0029] Compared with the prior art, the beneficial effects that the above-mentioned at least one technical solution adopted in the embodiments of this specification can achieve include at least the following: The open set radiation source individual identification method provided by the present invention firstly designs a one-dimensional deep attention residual network (DAResNet network, English name: Deep Attention Residual). The network (DAResNet) improves feature extraction capabilities by adding channel attention and GRU layers to an initial deep attention residual network to extract latent features from one-dimensional I / Q sequences. Then, it combines the DARESNet network with Multiple Gaussian Prototype Learning (MGPL) to construct the MGPL-DAResNet recognition model. The MGPL-DAResNet model is trained and tested using a preprocessed dataset. During training, multiple Gaussian prototypes are set for different categories of radiation source signals, and the MGPL-DAResNet recognition model is iteratively updated using a joint loss function of generation and discrimination constraints to obtain the Gaussian prototype and network weights corresponding to each radiation category. Finally, during recognition testing using a test set, the model classifies known targets and detects unknown targets by using the trained Gaussian prototypes and network weights corresponding to each radiation category, along with a reasonably selected distance threshold, outputting open-set recognition and classification results.

[0030] By identifying individual radiation sources in both closed-set and open-set scenarios, the results show that the method proposed in this invention has better identification performance than other methods, especially in open-set scenarios, where the method of this invention improves performance by 13.1% compared to other methods. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart of the open set radiation source individual identification method based on multiple Gaussian prototype learning disclosed in an embodiment of the present invention;

[0033] Figure 2 This is a schematic diagram of the architecture principle of multiple Gaussian prototype learning in an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram of the structure of the one-dimensional deep attention residual network constructed in an embodiment of the present invention;

[0035] Figure 4This is a schematic diagram of the recognition process of the MGPL-DAResNet recognition model in an embodiment of the present invention;

[0036] Figure 5 This is a curve showing the average recognition rate of closed-set recognition using the open-set radiation source individual recognition method disclosed in this embodiment of the invention, as a function of the number of Gaussian prototypes.

[0037] Figure 6 This is a curve showing the average recognition rate of open-set recognition using the open-set radiation source individual recognition method as a function of the number of Gaussian prototypes, as disclosed in this embodiment of the invention.

[0038] Figure 7 This is a curve showing the average recognition rate of closed-set recognition using the open-set radiation source individual recognition method as a function of the equilibrium parameter β, as disclosed in an embodiment of the present invention.

[0039] Figure 8 The curve showing the average recognition rate of open-set identification using the open-set radiation source individual identification method disclosed in this embodiment of the invention varies with the equilibrium parameter β. Detailed Implementation

[0040] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0041] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features of the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0042] This invention provides a method for identifying open-set radiation sources based on multiple Gaussian prototype learning. (See also...) Figure 1 As shown, the method for identifying individual open-set radiation sources includes the following steps:

[0043] S100. The collected radiation source signal is preprocessed to obtain a one-dimensional I / Q sequence, and the one-dimensional I / Q sequence is divided into a test set and a training set according to a preset ratio.

[0044] S200. An MGPL-DAResNet recognition model is established based on a one-dimensional deep attention residual network constructed using multiple Gaussian prototype learning. The constructed one-dimensional deep attention residual network includes a channel attention module set at the output end of each residual block in the initial deep attention residual network.

[0045] Specifically, Multiple Gaussian Prototype Learning (MGPL) is based on a framework derived from Bayesian inference. It assigns multiple Gaussian prototypes to samples of the same class, forming a single Gaussian region. These Gaussian regions from different classes constitute a latent space. MGPL enhances the description of class distribution by establishing multiple Gaussian prototypes. For more details, see [link to documentation]. Figure 2 As shown, the MGPL architecture includes an encoder, a decoder, and multiple Gaussian prototypes (Class 1 to Class C) corresponding to each radiation source category. The encoder extracts the latent features of the one-dimensional I / Q sequence and its original distribution, encodes them, and stores the encoding results in the latent space. The decoder samples from the latent space and encodes the one-dimensional I / Q sequence according to the network weights in the MGPL-DAResNet recognition model. In the MGPL method, the Gaussian prototype m is a tuple (μ...). m ,σ m Let ,y), where y is the class label, and the mean and variance of the Gaussian prototype distribution are represented by μ. m and σ m It means that μ m These are trainable parameters, σ m It is an identity matrix.

[0046] Specifically, the one-dimensional deep attention residual network constructed in this invention, abbreviated as DAResNet, has the following structure: Figure 3As shown, the one-dimensional deep attention residual network mainly consists of initial convolutional layers, residual blocks, and channel attention modules. This one-dimensional deep attention residual network adds four residual blocks at the output of each convolutional layer after the initial convolutional layers for further feature extraction and transformation. Each residual block consists of two small convolutional layers. Furthermore, a channel attention module is added at the output of each residual block. The core idea is that DARESNet can automatically learn which features are more important to the current task and give these important features enhanced attention. The channel attention module learns a set of weights and calculates the attention value by weighted summation of the input feature sequence. In this way, DARESNet can adaptively adjust the weights according to the importance of the input features, giving more weight to more critical features. Next, the attention value is multiplied element-wise with the output of the residual block to obtain the output after the attention mechanism. This element-wise multiplication operation allows DARESNet to better focus on latent features and enhance the representation ability of these features. By introducing the attention mechanism, the proposed method exhibits better performance on feature extraction tasks.

[0047] Building upon the above design, to further improve network performance, a GRU (Gated Recurrent Unit) layer is placed at the output of the last channel's attention module. The GRU layer, a recurrent neural network unit or gated recurrent unit, has a memory function; placing it at the output of the last channel's attention module allows for the memorization and cross-layer transfer of previous information. The GRU layer plays a role in modeling temporal relationships, extracting features, and understanding context when processing sequential data, extracting more useful and richer temporal features from the sequence, thereby further improving the network's performance and expressive power.

[0048] Even better, the one-dimensional deep attention residual network further includes: an average pooling layer and a fully connected layer are sequentially set at the output of the GRU layer.

[0049] The structural parameters of the constructed one-dimensional deep attention residual network are shown in Table 1 below: Each convolutional layer has a kernel size of 3, with the initial convolutional layer having a stride and padding count of 1. The number of kernels increases layer by layer, at 16, 32, 64, and 128 respectively, and each attention feature sequence has 4 elements. This design allows for the extraction of deeper features. It is important to note that since the discriminative constraint part of MGPL can directly achieve the network's classification task, no additional fully connected network classifier is added to this network structure.

[0050] Table 1: Parameter Settings for DARESNet

[0051]

[0052] See the MGPL-DAResNet recognition model obtained by combining MGPL and DAResNet in this invention. Figure 4 As shown, Figure 4 The preprocessed one-dimensional I / Q sequence is fed into the encoder, undergoing a series of processes including latent feature extraction, decoding, and reconstruction of the signal I / Q sequence. Furthermore, two constraints exist during the recognition process: discriminative constraints and generative constraints. Discriminative constraints increase inter-class distance, while generative constraints increase intra-class density. During model training using the training set, the network parameters and multiple Gaussian prototypes are updated by minimizing the loss function and using the backpropagation algorithm of the neural network, thereby achieving individual identification of radiation sources in open-set scenarios. This invention describes the specific process of using the established MGPL-DAResNet recognition model for open-set radiation source identification through steps S300 to S500.

[0053] S300. Input the training set into the MGPL-DAResNet recognition model, extract the latent features of the training set, and calculate the distance between the latent features and each Gaussian prototype corresponding to each radiation source category in the multiple Gaussian prototype learning.

[0054] Specifically, after the training set is input into the MGPL-DAResNet recognition model, the encoder extracts latent features from the one-dimensional I / Q sequence. Through these latent features, the distance between each latent feature and each Gaussian prototype corresponding to each radiation source category pre-designed in the multiple Gaussian prototype learning can be calculated.

[0055] S400. Based on the joint loss function of generation constraints and discriminative constraints, the MGPL-DAResNet recognition model is iteratively updated according to the distance between the latent features of the training set and the Gaussian prototype, to obtain the Gaussian prototype and network weights corresponding to each radiation category.

[0056] S500. By setting a distance threshold, the Gaussian prototype and its network weights corresponding to each radiation category in the trained MGPL-DAResNet recognition model are used to perform open-set recognition and classification on the test set and output the classification results.

[0057] Specifically, the generation constraint in the joint loss function is a constraint used to guide the clustering of latent features (also known as latent variables) around the Gaussian prototype of the corresponding separating class. It mainly focuses on the generation task logp given the label y. θ(x|y) represents a constraint on the effectiveness of feature representations during the generation process. Generation constraints help group latent features of the same category together while reserving additional space for latent features of different categories, thereby improving the quality of feature generation. The purpose of generation constraints is to effectively constrain feature representations during the generation process to obtain better generation results.

[0058] The discriminant constraint in the joint loss function is defined as maximizing the posterior probability of the observed variable x falling into the correct class y. Specifically, this is achieved by comparing the features of the input sample x with the Gaussian prototype m of the corresponding class.

[0059] The expression (1) for the discriminant constraint is:

[0060]

[0061] In equation (1), l() represents the KL divergence, γ represents the temperature parameter of the smoothed output probability, x represents the latent feature, y represents the radiation source category, and m represents the Gaussian prototype; A is the number of known radiation source categories, and each radiation source category is represented by B Gaussian prototypes. Analysis of the discriminant constraint expression shows that by minimizing the distance metric, the distance between the latent feature and the correct category's Gaussian prototype will decrease, while the distance between the latent feature and the incorrect category's Gaussian prototype will increase. This helps to further separate prototypes of different categories in the latent space, thereby improving the performance of the open set recognition task.

[0062] Based on the above introduction to generation constraints and discriminant constraints, the two constraints can be combined to form a joint loss function for training the network model. During training, the parameter β is used to weigh the loss of the two parts. The expression (2) of the joint loss function is as follows:

[0063]

[0064] In expression (2), β is the balance parameter. As can be seen from the above expression, the value of the balance parameter β has a certain impact on the recognition performance of the network. If β is too small, the density of the latent variables of the same class around their corresponding Gaussian prototypes is not high enough, making it difficult to achieve the best recognition effect. If β is too large, the intra-class density of the samples will be excessively increased, and the distance between different classes of Gaussian prototypes determined by the discriminant constraint will be too small, which will increase the overfitting degree of the model and thus affect the recognition performance of the network.

[0065] After training with generative and discriminative constraints, the parameters m of the Gaussian prototype distribution corresponding to each category of samples can be obtained, which are also the network weights. This method does not require an additional classifier; the Gaussian prototype can be regarded as a natural discriminator. In layman's terms, when a sample x is input, the encoder first obtains the variable r of its latent space based on x. By comparing the distances of r to all Gaussian prototypes, the closest Gaussian prototype m′ can be obtained. Then, based on the given threshold τ, it is determined whether the sample x belongs to an unknown sample or to the known category corresponding to the closest Gaussian prototype. The expression (3) of the output classification result is as follows:

[0066]

[0067] In the above expression (3), d(r,m′) is the distance from the latent variable (latest feature) r to the Gaussian prototype m′. During the experiment, the value of τ is determined under the condition that most of the data in the validation set can be identified as known classes. The MGPL method used in this paper has both generative and discriminative constraints. This model can effectively constrain the distribution of each feature within a specific region represented by multiple Gaussian prototypes under the same class. Therefore, selecting more than 95% of the samples in the validation set as known samples is effective for threshold estimation.

[0068] The principle of the open-set radiation source individual identification method provided by this invention is as follows: First, a one-dimensional deep attention residual network (DAResNet) is designed. This network incorporates a channel attention mechanism and a GRU layer into the initial deep attention residual network to extract latent features from the one-dimensional I / Q sequence, thereby improving the network's feature extraction capability. Then, the DARESNet network is combined with the Multiple Gaussian Prototype Learning (MGPL) approach to construct the MGPL-DAResNet identification model. The MGPL-DAResNet model is trained and tested using a preprocessed dataset. During training, multiple Gaussian prototypes are set for different categories of radiation source signals. Simultaneously, the MGPL-DAResNet model is iteratively updated using a joint loss function of generation and discrimination constraints to obtain the Gaussian prototype and network weights corresponding to each radiation category. Finally, during identification testing using a test set, the classification of known targets and the detection of unknown targets are completed using the trained Gaussian prototypes and network weights corresponding to each radiation category, along with a reasonably selected distance threshold, outputting the open-set identification classification result.

[0069] This invention uses the following method to construct five categories of radiation source signals, three of which are known and two are unknown, and to construct a simulation dataset of radiation source signals as an example. It then uses the above-mentioned open-set radiation source individual identification method to perform identification in open-set and closed-set scenarios, and examines the impact of different prototype numbers on the identification performance and the impact of different β conditions on the identification performance of the identification method of this invention.

[0070] See Figure 5 and Figure 6 The figure shows the curves illustrating the change in average recognition rate with the number of Gaussian prototypes. The results indicate that as the number of multiple Gaussian prototypes increases from 1 to 7, both the curves showing the change in average recognition rate under closed-set conditions and the curves showing the change in AUC value under open-set conditions exhibit a trend of first increasing and then decreasing. When the number of prototypes is 3, both curves reach their optimal values, indicating that using multiple Gaussian prototypes can improve the model's closed-set classification ability and open-set recognition performance. However, the number of prototypes should not be excessive. When the number of prototypes exceeds 3, both the model's closed-set classification ability and open-set recognition ability decrease, and an excessive number of prototypes also increases the model's complexity. Considering both the model's recognition performance and complexity, a final selection of 3 Gaussian prototypes was made.

[0071] See Figure 7 and Figure 8 The figure shows a schematic diagram of the average recognition rate as a function of the balance parameter β. The results indicate that: (See Figure 1) Figure 7 As shown, in closed-set scenarios, the impact of β on the average recognition rate fluctuates within a relatively small range and remains essentially constant. However, it can be observed that the average recognition rate reaches its optimal value when β = 0.02, and the average recognition rate shows a downward trend after the β value exceeds 0.02. Figure 8 In the open-set scenario, as the balance parameter is gradually increased from a very small value, the AUC value corresponding to the open-set environment also increases within a certain range as the parameter β increases. However, after β = 0.02, the AUC value decreases as the parameter β increases. Figure 7 and Figure 8 Analysis of the curve changes reveals that a larger balancing parameter negatively impacts the model's classification ability, reducing its recognition performance. Specifically, increasing the value of β leads to excessively strong generation constraints, causing the model to focus too much on the generation task, resulting in decreased classification performance. Furthermore, the generated feature representations may fail to effectively distinguish between known and unknown categories, thus lowering the recognition rate. Therefore, according to... Figure 7 and Figure 8 As a result, 0.02 was ultimately selected as the appropriate balance parameter.

[0072] In this embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-described methods for identifying open-set radiation sources based on multiple Gaussian prototype learning.

[0073] Specifically, the computer device can be a computer terminal, a server, or a similar computing device.

[0074] In this embodiment, a computer-readable storage medium is provided, which stores a computer program that executes any of the above-described methods for identifying open-set radiation sources based on multiple Gaussian prototype learning.

[0075] Specifically, computer-readable storage media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer-readable storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable storage media does not include transient media, such as modulated data signals and carrier waves.

[0076] Based on the same inventive concept, this invention also provides an open-set radiation source individual identification system based on multiple Gaussian prototype learning, as described in the following embodiments. Since the principle of the open-set radiation source individual identification system based on multiple Gaussian prototype learning is similar to that of the open-set radiation source individual identification method based on multiple Gaussian prototype learning, the implementation of the above-described open-set radiation source individual identification system can refer to the implementation of the above-described open-set radiation source individual identification method, and repeated details will not be elaborated further. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0077] The open-set radiation source individual identification system based on multiple Gaussian prototype learning provided in this embodiment of the invention includes:

[0078] The signal preprocessing module is used to preprocess the collected radiation source signal to obtain a one-dimensional I / Q sequence, and divide the one-dimensional I / Q sequence into a test set and a training set according to a preset ratio.

[0079] A network construction module is used to set a channel attention module at the output end of each residual block in the initial deep attention residual network to obtain a one-dimensional deep attention residual network.

[0080] An open set recognition network model building module is used to build an MGPL-DAResNet recognition model using multiple Gaussian prototype learning and the one-dimensional deep attention residual network.

[0081] The model training module is used to input the training set into the MGPL-DAResNet recognition model, extract the latent features of the training set, calculate the distance between the latent features and each Gaussian prototype corresponding to each radiation source category in the multiple Gaussian prototype learning, and iteratively update the MGPL-DAResNet recognition model based on the joint loss function of generation constraints and discrimination constraints, according to the distance between the latent features of the training set and the Gaussian prototypes, to obtain the Gaussian prototype and network weights corresponding to each radiation category.

[0082] The radiation source classification module is used to perform open-set recognition and classification on the test set by using the Gaussian prototype and its network weights corresponding to each radiation category in the trained MGPL-DAResNet recognition model, based on a set distance threshold, and output the classification result.

[0083] Furthermore, the network building module is also used to set a GRU layer at the output of the channel attention module of the last residual block.

[0084] Furthermore, the network building module is also used to sequentially set an average pooling layer and a fully connected layer at the output of the GRU layer.

[0085] The embodiments of the present invention achieve the following technical effects: by identifying individual radiation sources of five types of radiation source signals in closed-set and open-set scenarios, the results show that the method proposed in this invention has better identification performance than other methods, especially in open-set scenarios, where the method of the present invention improves performance by 13.1% compared to other methods.

[0086] Obviously, those skilled in the art should understand that the modules or steps of the above-described embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

[0087] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying individual open-set radiation sources based on multiple Gaussian prototype learning, characterized in that, include: The collected radiation source signals are preprocessed to obtain a one-dimensional I / Q sequence, and the one-dimensional I / Q sequence is divided into a test set and a training set according to a preset ratio; An MGPL-DAResNet recognition model is established based on a one-dimensional deep attention residual network constructed using multiple Gaussian prototype learning. The constructed one-dimensional deep attention residual network includes a channel attention module set at the output of each residual block in the initial deep attention residual network. The training set is input into the MGPL-DAResNet recognition model to extract the latent features of the training set, and the distance between the latent features and each Gaussian prototype corresponding to each radiation source category in the multiple Gaussian prototype learning is calculated. Based on the joint loss function of generation constraints and discriminative constraints, the MGPL-DAResNet recognition model is iteratively updated according to the distance between the latent features of the training set and the Gaussian prototype, so as to obtain the Gaussian prototype and network weights corresponding to each radiation category. By setting a distance threshold, the Gaussian prototype and its network weights corresponding to each radiation category in the trained MGPL-DAResNet recognition model are used to perform open set recognition and classification on the test set and output the classification results. The constructed one-dimensional deep attention residual network further includes: setting a GRU layer at the output of the channel attention module of the last residual block, and setting an average pooling layer and a fully connected layer in sequence at the output of the GRU layer; The architecture of the multiple Gaussian prototype learning includes an encoder, a decoder, and multiple Gaussian prototypes corresponding to each radiation source category; The encoder is used to extract the latent features of the one-dimensional I / Q sequence and the original distribution of the one-dimensional I / Q sequence for encoding, and store the encoding result in the latent space; The decoder samples from the latent space and encodes the one-dimensional I / Q sequence according to the network weights in the MGPL-DAResNet recognition model.

2. The method for identifying individual open-set radiation sources based on multiple Gaussian prototype learning according to claim 1, characterized in that, The expression for the discriminant constraint is: In equation (1), l Denotes KL divergence, The temperature parameter represents the smoothed output probability, x represents the latent feature, y represents the radiation source category, m represents the Gaussian prototype, and A is the number of known radiation source categories, with each radiation source category represented by B Gaussian prototypes.

3. A system for identifying open-set radiation sources based on multiple Gaussian prototype learning, characterized in that, The system for implementing the method of claim 1 or 2 includes: The signal preprocessing module is used to preprocess the collected radiation source signal to obtain a one-dimensional I / Q sequence, and divide the one-dimensional I / Q sequence into a test set and a training set according to a preset ratio. A network construction module is used to set a channel attention module at the output end of each residual block in the initial deep attention residual network to obtain a one-dimensional deep attention residual network. An open set recognition network model building module is used to build an MGPL-DAResNet recognition model using multiple Gaussian prototype learning and the one-dimensional deep attention residual network. The model training module is used to input the training set into the MGPL-DAResNet recognition model, extract the latent features of the training set, calculate the distance between the latent features and each Gaussian prototype corresponding to each radiation source category in the multiple Gaussian prototype learning, and iteratively update the MGPL-DAResNet recognition model based on the joint loss function of generation constraints and discrimination constraints, according to the distance between the latent features of the training set and the Gaussian prototypes, to obtain the Gaussian prototype and network weights corresponding to each radiation category. The radiation source classification module is used to perform open-set recognition and classification on the test set by using the Gaussian prototype and its network weights corresponding to each radiation category in the trained MGPL-DAResNet recognition model, based on a set distance threshold, and output the classification result.

4. The open-set radiation source individual identification system based on multiple Gaussian prototype learning according to claim 3, characterized in that, The network building module is also used to set a GRU layer at the output of the channel attention module of the last residual block.

5. The open-set radiation source individual identification system based on multiple Gaussian prototype learning according to claim 4, characterized in that, The network building module is also used to sequentially set an average pooling layer and a fully connected layer at the output of the GRU layer.