A method and system for cross-domain individual recognition of radiation sources

Abstract

The present application discloses a method and system for cross-domain individual recognition of radiation sources, which relates to the field of radiation source recognition. The method includes obtaining a source domain dataset and a target domain dataset, establishing a teacher model according to the source domain dataset, establishing a student model according to the target domain dataset, optimizing the teacher model and the student model respectively according to the source domain dataset by using a classification loss function, determining the optimized teacher model and the optimized student model, establishing a knowledge transfer mechanism by using a distribution matching loss function according to the source domain feature learning module in the optimized teacher model and the target domain feature learning module in the optimized student model, training the optimized student model and the optimized teacher model, and obtaining the individual recognition result of the radiation source by using the trained student model according to the target domain dataset; the present application can improve the generalization ability and recognition accuracy of the model between different domains on the basis of achieving lightweight.

Description

Technical Field

[0001] This application relates to the field of radiation source identification, and particularly to a method and system for cross-domain individual identification of radiation sources. Background Art

[0002] With the rapid development of modern communication and detection technologies, the individual identification of electromagnetic radiation source signals plays a crucial role in the civilian and public security fields. Especially in complex electromagnetic environments, accurately identifying and classifying different types of radiation sources is of great significance for ensuring the security of transportation and communication systems. Therefore, how to perform individual identity identification on the collected radiation source signals has gradually become an important research direction.

[0003] Traditional radiation source identification methods mainly rely on manually designed feature extraction algorithms, and these methods perform poorly in the face of complex environments. With the development of deep learning technologies, radiation source identification methods based on deep neural networks have gradually become a research hotspot. However, existing methods rarely consider the distribution differences between the source domain and the target domain, resulting in a decline in the performance of the model obtained in the source domain when facing actual applications. Moreover, edge devices have weak computing capabilities and usually require smaller models, and these factors restrict the practical application of existing radiation source methods.

[0004] Regarding the problem of domain distribution differences, some methods have proposed domain adaptation methods to learn the common features of the source domain and the target domain. For example, the "Radiation Source Feature Extraction Method Based on Serial Multi-Adversarial Domain Adaptation Network" with the patent application number 202310590628.0 uses the domain adversarial method to extract the common features of different domains and improve the model's domain adaptation ability. The "Unsupervised Domain Adaptation Radiation Source Individual Identification Method Based on Independence Criterion" with the patent application number 202411523839.3 proposes a distribution distance metric function for cross-domain samples constructed based on the Hilbert-Schmidt independence criterion. These methods measure the differences between domains using appropriate methods to prompt the model to extract the common features of different domains, thereby training a model adapted to the target domain. However, these methods usually use the same network structures for the source domain and the target domain and do not pay attention to the lightweight problem.

[0005] Regarding the problem of designing small models to adapt to edge devices, existing methods have carried out research on knowledge distillation techniques. By transferring the knowledge of large models to small models, knowledge distillation techniques can not only reduce the computational cost and storage requirements of the models, but also retain the high performance of large models. In the field of radiation source identification, some algorithms have also introduced knowledge distillation to obtain lightweight models and accelerate model inference. For example, the "Small-sample Incremental Radiation Source Individual Identification Method Based on Knowledge Distillation and Graph Model" with the patent number CN117496228A designed a small-sample incremental radiation source individual identification method based on knowledge distillation and graph model. First, a backbone network was trained using known radiation source data; then, pseudo-class incremental training was designed, and knowledge distillation was performed on the student model using the teacher model. This method mainly focuses on implementing knowledge distillation based on graph models and does not pay attention to the differences between the source domain and the target domain. Yang Feng introduced a knowledge distillation algorithm in the "Research on Radar Radiation Source Identification Based on Lightweight Neural Networks" in order to deploy the network model on a lightweight edge computing platform. Tu Ting in the "Research on Incremental Learning of Radiation Source Categories Based on Knowledge Distillation" and the "Method for Incremental Radiation Source Individual Identification Based on Knowledge Distillation Mechanism" with the patent number CN114492745A both focused on the lightweight of the model. However, these methods usually assume that the source domain and the target domain have the same distribution and do not pay attention to the performance degradation problem caused by the distribution differences between the two domains.

[0006] Therefore, there is an urgent need to provide a new method or system for cross-domain individual identification of radiation sources, which can improve the generalization ability and identification accuracy of the model between different domains on the basis of achieving lightweight. Summary of the Invention

[0007] The purpose of this application is to provide a method and system for cross-domain individual identification of radiation sources, which can improve the generalization ability and identification accuracy of the model between different domains on the basis of achieving lightweight.

[0008] To achieve the above purpose, this application provides the following solutions:

[0009] In the first aspect, this application provides a method for cross-domain individual identification of radiation sources, and the method for cross-domain individual identification of radiation sources includes:

[0010] Obtain a source domain dataset and a target domain dataset; the source domain dataset includes: simulated electromagnetic wave signal data pairs; the target domain dataset includes: real electromagnetic wave signal data pairs;

[0011] Based on the source domain dataset, a teacher model is established; the teacher model includes: 3 sequentially connected source domain feature learning modules and 1 fully connected layer; the source domain feature learning module includes a first multi-scale feature extraction network; the first multi-scale feature extraction network includes: a first feature extraction unit; the first feature extraction unit includes: a depthwise separable convolution module, a batch normalization module, an activation function module, a pointwise convolution module, and a max pooling layer module;

[0012] Based on the target domain dataset, a student model is established; the student model includes: 3 sequentially connected target domain feature learning modules and 1 fully connected layer; the target domain feature learning module includes a second multi-scale feature extraction network; the second multi-scale feature extraction network includes: a second feature extraction unit; the second feature extraction unit includes: a 1D deformable convolution module, a batch normalization module, an activation function module, a pointwise convolution module, and a max pooling layer module;

[0013] Based on the source domain dataset, using the classification loss function, the teacher model and the student model are optimized respectively to determine the optimized teacher model and the optimized student model;

[0014] Based on the source domain feature learning module in the optimized teacher model and the target domain feature learning module in the optimized student model, a knowledge transfer mechanism is established using the distribution matching loss function; and the optimized student model and the optimized teacher model are trained; the knowledge transfer mechanism is used to minimize the loss value between the feature maps output by the target domain feature learning module and the feature maps output by the source domain feature learning module, so that the trained student model can transfer the knowledge learned by the trained teacher model;

[0015] Based on the target domain dataset, using the trained student model, the individual recognition result of the radiation source is obtained.

[0016] Optionally, the obtaining of the source domain dataset and the target domain dataset specifically includes:

[0017] Using a radiation source simulation device to obtain simulated electromagnetic wave signals; and annotating the simulated electromagnetic wave signals to obtain simulated electromagnetic wave signal data pairs;

[0018] Based on the simulated electromagnetic wave signal data pairs, the source domain dataset is determined;

[0019] Using a radiation source device to obtain real electromagnetic wave signals; and annotating the real electromagnetic wave signals to obtain real electromagnetic wave signal data pairs;

[0020] Based on the real electromagnetic wave signal data pairs, the target domain dataset is determined.

[0021] Optionally, based on the source domain dataset, using a classification loss function, optimize the teacher model and the student model respectively to determine the optimized teacher model and the optimized student model, specifically including:

[0022] Use the formula to determine the classification loss function ; where is the number of samples, is the number of classes; is the sample number, is the class number, is the sample output by the model belongs to class the predicted probability, is the sample the true label, when the sample the true class is then otherwise ;

[0023] Based on the classification loss function, optimize the teacher model and the student model respectively to determine the optimized teacher model and the optimized student model.

[0024] Optionally, based on the source domain feature learning module in the optimized teacher model and the target domain feature learning module in the optimized student model, establish a knowledge transfer mechanism using a distribution matching loss function, specifically including:

[0025] Establish a corresponding distribution matching loss function between the source domain feature learning module in the optimized teacher model and the target domain feature learning module in the corresponding optimized student model;

[0026] Based on the distribution matching loss function, determine the corresponding distribution losses between the source domain feature learning module and the target domain feature learning module respectively;

[0027] Based on the corresponding distribution losses between the source domain feature learning module and the target domain feature learning module, determine the overall distribution loss;

[0028] Based on the overall distribution loss, establish a knowledge transfer mechanism.

[0029] Optionally, based on the distribution matching loss function, determine the corresponding distribution losses between the source domain feature learning module and the target domain feature learning module respectively, specifically including:

[0030] Use the formula to determine the corresponding distribution losses between the source domain feature learning module and the target domain feature learning module;

[0031] where is the The distribution loss between the th source domain feature learning module and the th target domain feature learning module, where is the module number, is the number of channels, is the length of the feature map, is the serial number of the element in the feature map, is the feature map output by the th source domain feature learning module of the optimized teacher model, is the feature map output by the th target domain feature learning module of the optimized student model.

[0032] Optionally, determining the overall distribution loss according to the corresponding distribution loss between the source domain feature learning module and the target domain feature learning module specifically includes:

[0033] Using the formula to determine the overall distribution loss ;

[0034] where is the distribution loss between the th source domain feature learning module and the th target domain feature learning module, and is the module number.

[0035] In a second aspect, the present application provides a radiation source cross-domain individual recognition system, and the radiation source cross-domain individual recognition system includes:

[0036] A dataset acquisition module for acquiring a source domain dataset and a target domain dataset; the source domain dataset includes: simulated electromagnetic wave signal data pairs; the target domain dataset includes: real electromagnetic wave signal data pairs;

[0037] A teacher model establishment module for establishing a teacher model according to the source domain dataset; the teacher model includes: 3 sequentially connected source domain feature learning modules and 1 fully connected layer; the source domain feature learning module includes a first multi-scale feature extraction network; the first multi-scale feature extraction network includes: a first feature extraction unit; the first feature extraction unit includes: a depthwise separable convolution module, a batch normalization module, an activation function module, a point convolution module, and a max pooling layer module;

[0038] A student model building module, which is used to build a student model according to a target domain dataset; the student model includes: three sequentially connected target domain feature learning modules and one fully connected layer; the target domain feature learning module includes a second multi-scale feature extraction network; the second multi-scale feature extraction network includes: a second feature extraction unit; the second feature extraction unit includes: a one-dimensional deformable convolution module, a batch normalization module, an activation function module, a point convolution module, and a max pooling layer module;

[0039] A model optimization module, which is used to optimize the teacher model and the student model respectively according to the source domain dataset by using a classification loss function, and determine the optimized teacher model and the optimized student model;

[0040] A student model training module, which is used to establish a knowledge transfer mechanism according to the source domain feature learning module in the optimized teacher model and the target domain feature learning module in the optimized student model by using a distribution matching loss function; and train the optimized student model and the optimized teacher model; the knowledge transfer mechanism is used to minimize the loss value between the feature map output by the target domain feature learning module and the feature map output by the source domain feature learning module, so that the trained student model can transfer the knowledge learned by the trained teacher model;

[0041] An individual recognition module, which is used to obtain the individual recognition result of the radiation source according to the target domain dataset by using the trained student model.

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

[0043] The present application provides a method and system for cross-domain individual recognition of radiation sources. Aiming at the problem of insufficient feature learning in the existing methods, a multi-scale feature extraction method is designed to learn one-dimensional features, and a teacher model based on depthwise separable convolution is designed. Subsequently, a student model based on deformable convolution is designed to transfer the knowledge learned by the teacher model, and a distribution matching loss function is set to align the distributions of each module; the present application combines cross-domain adaptive technology and knowledge transfer technology, which can improve the generalization ability and recognition accuracy of the model between different domains on the basis of realizing lightweight, and further better meet the needs of edge devices to process radiation source data in unknown cross-domain scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to these drawings.

[0045] Figure 1Schematic flowchart of a method for cross - domain individual recognition of radiation sources in an embodiment of the present application;

[0046] Figure 2 Algorithm structure diagram of a method for cross - domain individual recognition of radiation sources in an embodiment of the present application;

[0047] Figure 3 Schematic diagram of the structure of the source - domain feature learning module in an embodiment of the present application;

[0048] Figure 4 Schematic diagram of the structure of the target - domain feature learning module in an embodiment of the present application. Detailed implementation manners

[0049] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0050] To make the above - mentioned objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0051] In an exemplary embodiment, as Figure 1 shown, a method for cross - domain individual recognition of radiation sources is provided. The method for cross - domain individual recognition of radiation sources includes the following S1 to S6. Among them:

[0052] S1: Obtain a source - domain data set and a target - domain data set.

[0053] The source - domain data set includes: simulated electromagnetic wave signal data pairs; the target - domain data set includes: real electromagnetic wave signal data pairs; both the simulated electromagnetic wave signal data pairs and the real electromagnetic wave signal data pairs include: signals and corresponding identity tags.

[0054] Using a radiation source device, collect the AIS real electromagnetic wave signals of marine ship targets, separate and sample each target signal into a vector with a length of 128, and label the corresponding individual identities to obtain real electromagnetic wave signal data pairs containing (signals , identity tags ), and construct a radiation source signal target - domain data set for cross - domain scenarios according to the real electromagnetic wave signal data pairs . Assume that there are radiation source individuals in the target domain, then a classification problem is formed. The target - domain data set is used as the test set.

[0055] Using a radiation source simulation device, generate A large number of AIS simulation electromagnetic wave signals of radiation source individuals are marked with corresponding individual identities to obtain simulation electromagnetic wave signal data pairs containing (signals , identity tags ), and a radiation source signal source domain dataset for cross-domain scenarios is constructed . The source domain dataset is the training set.

[0056] Based on the source domain dataset and the target domain dataset, the individuals of the radiation source are identified. The algorithm structure diagram of the radiation source cross-domain individual recognition method is as Figure 2 shown.

[0057] S2: According to the source domain dataset, a teacher model is established.

[0058] The teacher model includes: 3 sequentially connected source domain feature learning modules and 1 fully connected layer; the source domain feature learning module includes a first multi-scale feature extraction network; the first multi-scale feature extraction network includes: first feature extraction units of different scales; the first feature extraction unit includes: a depthwise separable convolution module, a batch normalization module, an activation function module, a point convolution module, and a max pooling layer module connected in sequence; the scales of the convolution kernels of the depthwise separable convolution modules in the first feature extraction units of different scales are different.

[0059] The teacher model is composed of 3 sequentially connected source domain feature learning modules BS (Block of Source) and 1 fully connected layer with a length of 2 to fully learn the radiation source signal features and finally connect to a fully connected layer with a length of . The output of this fully connected layer represents the classification of the input signal.

[0060] The input of the BS module is the signal in the source domain dataset or the feature map learned by the previous BS module, and the output is a vector with a scale half of the input. For a signal feature vector with a length of , in order to better extract features, a first multi-scale feature extraction network is designed, and its network structure is as Figure 3As shown. This network consists of 3 scales. Among them, DwConv(k, s, p) represents the depthwise separable convolution module (Depthwise Convolution), k represents the convolution kernel size, and the convolution kernels used in the three scales are 3, 5, and 7 respectively. s represents the stride, p represents the padding pixels. Then, the batch normalization module (BatchNorm) is used for normalization processing, and ReLU is used as the activation function. Then, it passes through a pointwise convolution module PwConv(k, s, p), that is, pointwise convolution Pointwise Convolution. Then, it passes through a max-pooling layer module. Finally, the features of the 3 scales are concatenated to obtain a vector with a length of 1.5* Then, it passes through a 1D convolution Conv1d(k = 1, / 2) and a layer of ReLU activation function, and finally obtains a vector with a length of / 2. The source domain signal passes through 3 BS modules, and finally obtains features with a length of 16. Then, it passes through a fully connected layer with a length of 2 and a fully connected layer with a length of Finally, the classification result of the input signal is obtained.

[0061] In practical applications, the radiation source signal may contain multiple frequency components and changes in time scales, and these changes are crucial for correctly identifying individual radiation sources. The multi-scale feature extraction network can capture feature information at different scales simultaneously. By using convolution kernels of different sizes (3x3, 5x5, 7x7), this network can effectively extract local detail features, medium-scale features, and global context information, thereby fusing information at different scales and enabling subsequent layers to obtain richer feature representations. Subsequently, the 1D convolution Conv1d is used to further compress the feature dimension to / 2 while retaining the key information. Such a design helps to improve the model's understanding depth of the input data and thus enhances the classification accuracy. In the multi-scale network, by combining the depthwise separable convolution module (DwConv) and the pointwise convolution module (PwConv), the model's expressive ability is improved while reducing the number of parameters and computational costs to maintain computational efficiency.

[0062] The design of the teacher model not only takes into account the requirements of the current task but also provides a good foundation for future knowledge distillation. When the teacher model is fully trained, it can fully learn the features related to classification, laying a foundation for passing the relevant knowledge to the student model subsequently. Since the teacher model adopts a complex multi-scale feature extraction mechanism, it can provide more refined guidance for the student model, helping the latter achieve similar or even better performance than the teacher model under a simpler architecture.

[0063] S3: Establish a student model according to the target domain dataset.

[0064] The student model includes: 3 sequentially connected target domain feature learning modules and 1 fully connected layer; the target domain feature learning module includes a second multi-scale feature extraction network; the second multi-scale feature extraction network includes: second feature extraction units of different scales; the second feature extraction unit includes: a one-dimensional deformable convolution module, a batch normalization module, an activation function module, a point convolution module, and a max pooling layer module connected in sequence; the scales of the convolution kernels of the one-dimensional deformable convolution modules in the second feature extraction units of different scales are different.

[0065] The student model consists of 3 target domain feature learning modules BT (Block of Target) and 1 fully connected layer with a length of 2 to fully learn the radiation source signal features, and finally connected to a fully connected layer with a length of The output of this fully connected layer represents the classification of the input signal.

[0066] The input of the BT module is the signal in the target domain dataset or the feature map learned by the previous BT module, and the output is a vector with a scale half of the length of the input image or the feature map learned by the previous BT module. For the signal feature vector with a length of , in order to better extract features, a second multi-scale feature extraction network is designed, and its network structure is as Figure 4 shown. This network contains a total of 2 scales. Among them, DefConv(k, s, p, f) represents one-dimensional deformable convolution (DeformableConvolution), k represents the convolution sum size, the convolution kernels used in the two scales are 3 and 5 respectively, s represents the stride, p represents the padding pixels, f represents the dilation rate, then use the batch normalization module (BatchNorm) for normalization processing, then use ReLU as the activation function, then go through a layer of point convolution module PwConv(k, s, p), that is, Pointwise Convolution, then go through a max pooling layer module, and finally splice the features of the 2 scales to obtain a vector with a length of 0.75* , and then go through a 1D convolution Conv1d(k = 1, / 2) and a layer of ReLU activation function, and finally obtain a vector with a length of / 2.

[0067] By introducing a multi-scale feature extraction network and a one-dimensional deformable convolution module, the student model can improve its learning ability for target domain radiation source signals while maintaining a low computational complexity. Compared with the teacher model, the student model adopts a simpler architecture. Especially in the multi-scale feature extraction part, only two scales of convolutional kernels (3x3, 5x5) are used, and a one-dimensional deformable convolution module (DefConv) is introduced. This simplification not only reduces the number of model parameters but also decreases the demand for computational resources, making the student model more suitable for deployment in resource-constrained environments. In addition, by reducing the number of scales, the student model can further accelerate the inference speed on the premise of ensuring the quality of feature extraction. Embedding the one-dimensional deformable convolution module in the designed feature extraction network allows the convolutional kernel to dynamically adjust its position according to the characteristics of the input signal, which helps the student model capture the non-linear changes and geometric transformations in the signal and fully transfer the knowledge learned by the teacher model.

[0068] The design of the student model fully considers the application scenario of knowledge distillation, ensuring its learning ability while reducing the number of parameters. By mimicking the intermediate layer features of the teacher model, the student model can gradually learn the knowledge of the teacher model during the training process, thereby improving its generalization ability and classification accuracy. In particular, since the teacher model has learned rich feature representations through a complex multi-scale feature extraction mechanism, the student model can learn these features to make up for the deficiencies brought by its simple structure and ultimately achieve performance similar to or even better than that of the teacher model.

[0069] S4: According to the source domain dataset, use the classification loss function to optimize the teacher model and the student model respectively to determine the optimized teacher model and the optimized student model; the classification loss function is used to determine the classification loss.

[0070] Design the classification loss function. For the teacher model and the student model, combine the labeled identity tags , and calculate the corresponding classification loss. Suppose there are samples, then its loss function is as follows:

[0071] .

[0072] Among them, is the classification loss function, is the number of samples, is the number of classes; is the sample number, is the class number, is the predicted probability that the sample belongs to class , is the sample 's true label. When the sample The true category is When , otherwise .

[0073] Using the samples in the source domain dataset and their individual identity labels , train the teacher model, and optimize the teacher model using the classification loss function to determine the optimized teacher model.

[0074] Using the samples in the source domain dataset and their individual identity labels , train the student model, and optimize the student model using the classification loss function to determine the optimized student model.

[0075] S5: According to the source domain feature learning module in the optimized teacher model and the target domain feature learning module in the optimized student model, establish a knowledge transfer mechanism using the distribution matching loss function; and train the optimized student model and the optimized teacher model.

[0076] Based on the optimized teacher model and the optimized student model, design a distribution matching loss function DM (Distribution Match), add a DM module to the output feature maps of each optimized teacher model BS module and optimized student model BT module, whose input is the feature maps output by the optimized teacher model BS module and the optimized student model BT module, and the output is the distribution loss . Assume that the feature maps output by BS and BT of each module are , , and their sizes are all , , , is the number of channels, is the channel number, is the length of the feature map, is the serial number of the element in the feature map, then the loss function ( ) of the th BS module of the teacher model and the is as follows:

[0077] .

[0078] Among them, is the feature map output by the th source domain feature learning module of the optimized teacher model, is the The feature maps output by a target domain feature learning module.

[0079] Since both the teacher model and the student model contain 3 feature learning modules, the overall distribution loss can be denoted as:

[0080] .

[0081] Input source domain samples and target domain samples into the optimized teacher model and the optimized student model simultaneously, and obtain the feature maps output by each module of the optimized teacher model and the optimized student model, which are , , , , and respectively. Subsequently, use the distribution matching loss function to train the optimized teacher model and the optimized student model.

[0082] Repeat S4 - S5. After repeating 500 times, the trained teacher model and student model are obtained. In this application, the Stochastic Gradient Descent (SGD) is used as the optimization function, and the learning rate is set to lr = 0.0002.

[0083] The distribution matching loss function DM (Distribution Match) can establish a deep knowledge transfer mechanism between the teacher model and the student model, ensuring that the student model can not only imitate the behavior of the teacher model at the output level, but also learn a representation similar to the teacher model in the intermediate layer feature space. By adding a DM module to the output feature maps of each teacher model BS module and student model BT module, the feature distributions of the two models at the same level can be directly compared, thereby guiding the student model to better capture the key information learned by the teacher model. The DM loss function forces the student model to generate feature maps similar to the teacher model by minimizing the feature distribution difference between the teacher model and the student model at the same level, which can not only ensure that the student model can imitate the behavior of the teacher model at a high level, but also promote the consistency of low-level features. In addition, since DM acts directly on the intermediate layer features instead of relying only on the error backpropagation of the output layer, it can accelerate the convergence speed during training. By imposing constraints at multiple levels, the DM loss function can also help alleviate the problem of gradient vanishing or explosion, ensuring a more stable training process.

[0084] S6: According to the target domain dataset, use the trained student model to obtain the individual recognition result of the radiation source.

[0085] After training is completed, only keep the trained student model, input the samples in the target domain dataset into the trained student model, and obtain the final individual recognition result of the radiation source.

[0086] In view of the problem that the existing methods do not learn features sufficiently, this application designs a multi-scale feature extraction method to learn one-dimensional features, designs a teacher model based on depthwise separable convolution, and then designs a student model based on deformable convolution to fully transfer the knowledge learned by the teacher model. For scenarios that jointly require transfer and distillation, a distribution matching loss function is designed to align the distributions of each module. The method of this application combines cross-domain adaptation technology and knowledge distillation technology, which can improve the generalization ability and recognition accuracy of the model between different domains.

[0087] Based on the same inventive concept, an embodiment of this application also provides a radiation source cross-domain individual recognition system for implementing the above-mentioned one. The implementation solutions provided by this system to solve problems are similar to those recorded in the above method. Therefore, the specific limitations in one or more embodiments of the radiation source cross-domain individual recognition system provided below can refer to the limitations on radiation source cross-domain individual recognition in the above text, and will not be repeated here.

[0088] In an exemplary embodiment, a radiation source cross-domain individual recognition system is provided. The radiation source cross-domain individual recognition system includes:

[0089] A dataset acquisition module, configured to acquire a source domain dataset and a target domain dataset; the source domain dataset includes: simulated electromagnetic wave signal data pairs; the target domain dataset includes: real electromagnetic wave signal data pairs.

[0090] A teacher model establishment module, configured to establish a teacher model according to the source domain dataset; the teacher model includes: 3 source domain feature learning modules connected in sequence and 1 fully connected layer; the source domain feature learning module includes a first multi-scale feature extraction network; the first multi-scale feature extraction network includes: a first feature extraction unit; the first feature extraction unit includes: a depthwise separable convolution module, a batch normalization module, an activation function module, a point convolution module, and a max pooling layer module.

[0091] A student model establishment module, configured to establish a student model according to the target domain dataset; the student model includes: 3 target domain feature learning modules connected in sequence and 1 fully connected layer; the target domain feature learning module includes a second multi-scale feature extraction network; the second multi-scale feature extraction network includes: a second feature extraction unit; the second feature extraction unit includes: a one-dimensional deformable convolution module, a batch normalization module, an activation function module, a point convolution module, and a max pooling layer module.

[0092] A model optimization module, configured to optimize the teacher model and the student model respectively according to the source domain dataset by using a classification loss function, and determine the optimized teacher model and the optimized student model.

[0093] A student model training module, which is used to establish a knowledge transfer mechanism according to the source domain feature learning module in the optimized teacher model and the target domain feature learning module in the optimized student model by using a distribution matching loss function; and train the optimized student model and the optimized teacher model; the knowledge transfer mechanism is used to minimize the loss value between the feature map output by the target domain feature learning module and the feature map output by the source domain feature learning module, so that the trained student model can transfer the knowledge learned by the trained teacher model.

[0094] An individual recognition module, which is used to obtain the individual recognition result of the radiation source according to the target domain data set by using the trained student model.

[0095] In this article, specific examples are used to elaborate on the principle and implementation manner of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, according to the idea of this application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to this application.

Claims

1. A method for cross-domain individual identification of radiation sources, characterized in that, The radiation source cross-domain individual recognition method includes: Obtain a source domain dataset and a target domain dataset; the source domain dataset includes: simulated electromagnetic wave signal data pairs; the target domain dataset includes: real electromagnetic wave signal data pairs; Based on the source domain dataset, establish a teacher model; the teacher model includes: 3 source domain feature learning modules connected in sequence and 1 fully connected layer; the source domain feature learning module includes a first multi-scale feature extraction network; the first multi-scale feature extraction network includes: a first feature extraction unit; the first feature extraction unit includes: a depthwise separable convolution module, a batch normalization module, an activation function module, a point convolution module, and a max pooling layer module; Based on the target domain dataset, establish a student model; the student model includes: 3 target domain feature learning modules connected in sequence and 1 fully connected layer; the target domain feature learning module includes a second multi-scale feature extraction network; the second multi-scale feature extraction network includes: a second feature extraction unit; the second feature extraction unit includes: a one-dimensional deformable convolution module, a batch normalization module, an activation function module, a point convolution module, and a max pooling layer module; Based on the source domain dataset, use a classification loss function to optimize the teacher model and the student model respectively, and determine the optimized teacher model and the optimized student model; Based on the source domain feature learning module in the optimized teacher model and the target domain feature learning module in the optimized student model, establish a knowledge transfer mechanism using a distribution matching loss function; and train the optimized student model and the optimized teacher model; the knowledge transfer mechanism is used to minimize the loss value between the feature map output by the target domain feature learning module and the feature map output by the source domain feature learning module, so that the trained student model transfers the knowledge learned by the trained teacher model; Based on the target domain dataset, use the trained student model to obtain the individual recognition result of the radiation source.

2. The method for cross-domain individual identification of radiation sources according to claim 1, wherein The obtaining of the source domain dataset and the target domain dataset specifically includes: Use a radiation source simulation device to obtain simulated electromagnetic wave signals; and label the simulated electromagnetic wave signals to obtain simulated electromagnetic wave signal data pairs; Based on the simulated electromagnetic wave signal data pairs, determine the source domain dataset; Use a radiation source device to obtain real electromagnetic wave signals; and label the real electromagnetic wave signals to obtain real electromagnetic wave signal data pairs; Based on the real electromagnetic wave signal data pairs, determine the target domain dataset.

3. The method for cross-domain individual identification of radiation sources according to claim 1, wherein The using of the classification loss function to optimize the teacher model and the student model respectively based on the source domain dataset to determine the optimized teacher model and the optimized student model specifically includes: Determine the classification loss function using the formula ; where is the number of samples, is the number of classes; is the sample number, is the class number, is the predicted probability that the sample output by the model belongs to class , is the true label of the sample , when the true class of the sample is , ; Based on the classification loss function, optimize the teacher model and the student model respectively, and determine the optimized teacher model and the optimized student model.

4. The method for cross-domain individual recognition of radiation sources according to claim 1, characterized in that The establishing of the knowledge transfer mechanism using the distribution matching loss function based on the source domain feature learning module in the optimized teacher model and the target domain feature learning module in the optimized student model specifically includes: Establish a corresponding distribution matching loss function between the source domain feature learning module in the optimized teacher model and the corresponding target domain feature learning module in the optimized student model; According to the distribution matching loss function, respectively determine the corresponding distribution losses between the source domain feature learning module and the target domain feature learning module; According to the corresponding distribution losses between the source domain feature learning module and the target domain feature learning module, determine the overall distribution loss; Establish a knowledge transfer mechanism according to the overall distribution loss.

5. The method for cross-domain individual recognition of radiation sources according to claim 4, wherein The step of respectively determining the corresponding distribution losses between the source domain feature learning module and the target domain feature learning module according to the distribution matching loss function specifically includes: Using the formula to determine the corresponding distribution loss between the source domain feature learning module and the target domain feature learning module; Among them, is the distribution loss between the -th source domain feature learning module and the -th target domain feature learning module, is the module number, is the number of channels, is the channel number, is the length of the feature map, is the serial number of the element in the feature map, is the feature map output by the -th source domain feature learning module of the optimized teacher model, is the feature map output by the -th target domain feature learning module of the optimized student model.

6. The radiation source cross-domain individual recognition method according to claim 4, wherein The step of determining the overall distribution loss according to the corresponding distribution losses between the source domain feature learning module and the target domain feature learning module specifically includes: Using the formula to determine the overall distribution loss ; Among them, is the distribution loss of the th source domain feature learning module and the th target domain feature learning module, where is the module number.

7. A radiation source cross-domain individual recognition system, characterized in that, The radiation source cross-domain individual recognition system includes: A dataset acquisition module, configured to acquire a source domain dataset and a target domain dataset; the source domain dataset includes: simulated electromagnetic wave signal data pairs; the target domain dataset includes: real electromagnetic wave signal data pairs; A teacher model establishment module, configured to establish a teacher model according to the source domain dataset; the teacher model includes: 3 sequentially connected source domain feature learning modules and 1 fully connected layer; the source domain feature learning module includes a first multi-scale feature extraction network; the first multi-scale feature extraction network includes: a first feature extraction unit; the first feature extraction unit includes: a depthwise separable convolution module, a batch normalization module, an activation function module, a point convolution module, and a max pooling layer module; A student model establishment module, configured to establish a student model according to the target domain dataset; the student model includes: 3 sequentially connected target domain feature learning modules and 1 fully connected layer; the target domain feature learning module includes a second multi-scale feature extraction network; the second multi-scale feature extraction network includes: a second feature extraction unit; the second feature extraction unit includes: a one-dimensional deformable convolution module, a batch normalization module, an activation function module, a point convolution module, and a max pooling layer module; A model optimization module, configured to optimize the teacher model and the student model respectively according to the source domain dataset by using a classification loss function, and determine the optimized teacher model and the optimized student model; A student model training module, configured to establish a knowledge transfer mechanism by using a distribution matching loss function according to the source domain feature learning module in the optimized teacher model and the target domain feature learning module in the optimized student model; and train the optimized student model and the optimized teacher model; the knowledge transfer mechanism is used to minimize the loss value between the feature map output by the target domain feature learning module and the feature map output by the source domain feature learning module, so that the trained student model transfers the knowledge learned by the trained teacher model; An individual recognition module, configured to obtain the individual recognition result of the radiation source according to the target domain dataset by using the trained student model.

Images