A method and system for emitter individual identification based on removing receiver dyeing effect

Abstract

The application discloses a radiation source individual identification method and system based on removal of receiver dyeing effect, relates to the field of radiation source individual identification, and comprises the following steps: preprocessing time domain signal data of a to-be-processed radiation source to obtain frequency domain features; obtaining category domain invariant features based on a DSC-DANN model; screening the category domain invariant features based on feature importance scores of a random forest to obtain radiation source classification sensitive features; and inputting the radiation source classification sensitive features into a random forest classifier to output a radiation source classification result. The application adopts the DSC-DANN model to obtain the category domain invariant features, adopts the feature importance scores of the random forest to screen, and obtains the radiation source classification result, so that the application can be suitable for cross-domain radiation source individual identification, gradually reduces the influence of the receiver dyeing effect on classification and identification, effectively removes the receiver dyeing effect, and overcomes the disadvantage of increased calculation resources caused by the adversarial training.

Description

Technical Field

[0001] This application relates to the field of radiation source individual identification, and in particular to a method and system for radiation source individual identification based on removing receiver staining effects. Background Technology

[0002] Specific Emitter Identification (SEI) is a technique for identifying individual radio radiation sources by measuring the external characteristics of intercepted electromagnetic signals. Traditional SEI methods typically fall into two categories: those based on transient signals and those based on steady-state signals. While transient signals are readily identifiable and have obvious features, they require high-quality intercepted signals and have poor resistance to noise. Steady-state signals offer advantages such as high concealment and strong environmental adaptability, but suffer from difficulties in signal extraction. With the increasing complexity of electromagnetic signals and noise and the diversification of radar systems, traditional parameter matching methods, which identify radiation sources by comparing the received signal with the parameter characteristics of signals already in a database, suffer from drawbacks such as strong database dependence, poor adaptability, susceptibility to noise interference, and poor performance in handling nonlinear and multipath effects.

[0003] Due to factors such as nonlinear distortion, internal thermal noise and shot noise, clock jitter, and sampling distortion, the receiver's output signal inevitably carries hardware distortion information, i.e., the signal is "colored" by the receiver. This effect alters the characteristics of the received signal, masking or distorting signal information that could originally reflect the individual characteristics of the radiation source, thus severely interfering with radiation source identification. However, current research focuses on the transmitter, primarily on extracting and learning fingerprint features from different types of radiation sources, neglecting the study of receiver coloring effects. This results in SEI being highly tied to the receiver used for training, making it difficult to adapt to cross-receiver data, severely limiting the technology's versatility across receiving systems, application platforms, and even different locations. In practical applications, it is preferable to adapt a model obtained from a single acquisition and training to different receivers, rather than building a model for each receiver.

[0004] Currently, there are four main approaches to eliminating receiver coloration effects, each with its own drawbacks: (1) Using IQ (In-phase and Quadrature) compensation helps to unify the signal characteristics of different receiving channels and reduce the impact of hardware differences on radiation source identification. However, IQ compensation for multi-receiver systems requires processing multi-channel data, resulting in high computational complexity and a large demand for hardware resources. (2) Perform intrinsic time-scale decomposition on the received signals of each receiver and extract non-Gaussian features from the decomposed signals. However, the extracted non-Gaussian features are affected by prior knowledge and are often not the most sensitive features for radiation source classification. (3) Set up an additional radiation source or receiver as a reference standard and calibrate the radiation source or receiver. However, the reference equipment also has non-ideal distortion, and alignment with the reference does not truly eliminate the hardware influence. (4) Estimate the distortion parameters of the receiver devices and perform correction. However, this method is limited by the understanding of the distortion model and can only be applied to specific devices and cannot characterize the complex coupling relationship of all devices.

[0005] In fact, the universality problem of radiation source fingerprinting technology across receiver platforms can be understood as a domain adaptation problem in transfer learning, that is, mapping the feature vectors of data from different receiver platforms to a maximally domain-invariant space to minimize the influence of receiver staining effects. Currently, SEI transfer learning methods mainly have the following three mainstream approaches: (1) SEI transfer learning method based on domain adaptation: In order to reduce the data distribution difference between the source domain and the target domain, the features of the source domain and the target domain can be mapped to a common RKHS (Reproducing Kernel Hilbert Space) based on a certain difference metric. The feature transformation is learned by minimizing the metric of the distribution difference between the domains, thereby achieving the distribution alignment between the source domain and the target domain.

[0006] (2) Reconstruction-based SEI transfer learning method: The reconstruction-based method refers to the method of extracting transferable features using an autoencoder. The domain adaptation process will destroy the discriminative information of the samples, while the reconstruction method can reduce the loss of information. The reconstruction method can decouple features into domain-specific features and domain-invariant features. The domain-invariant features are used to transfer knowledge, while the domain-specific features are used to complete the target domain task.

[0007] (3) SEI Transfer Learning Method Based on Adversarial Transfer Learning: The domain adaptation method based on adversarial transfer learning introduces the idea of ​​adversarial generative networks into the domain adaptation problem, usually based on DANN (Domain Adversarial Neural Network). The training process of adversarial domain adaptation is a game between the feature extractor and the domain discriminator: the domain discriminator learns to distinguish between source domain features and target domain features, while the feature extractor learns domain-invariant features to confuse the domain discriminator. After training, the network can extract feature representations that are both categorical and domain-invariant.

[0008] In practical research, it was found that, compared with the three methods, the SEI transfer learning method based on adversarial transfer often performs better. The core reasons are: (1) This method uses dynamic nonlinear alignment, which is more suitable for handling the complex distribution shift of radiation source signals and has better generalization performance; (2) This method has efficient adaptability in unsupervised situations and can still maintain high accuracy when the target domain label is scarce; (3) This method has end-to-end optimization compatibility and is easy to integrate with existing deep learning models. However, the SEI transfer learning method based on adversarial transfer requires additional training of the discriminator, which increases the consumption of computational resources and thus limits the practical application of this method to a certain extent. Therefore, how to reduce the consumption of computational resources while maintaining high recognition accuracy is a key issue that needs to be addressed. Summary of the Invention

[0009] The purpose of this application is to provide a method and system for individual radiation source identification based on the removal of receiver coloring effects. Based on the classic DANN model, it employs the DSC-DANN (Depthwise Separable Convolution Domain Adversarial Neural Network) model to process frequency domain features into categorical domain-invariant features. While maintaining good coloring effect removal, it uses deep separable convolutional layers to achieve high recognition accuracy and reduce computational costs. Furthermore, it uses feature importance scores based on random forests for filtering to obtain radiation source classification results. This effectively improves SEI recognition accuracy, suppresses noise feature interference, reduces the impact of receiver coloring effects on classification, and is adaptable to cross-domain individual radiation source identification. It gradually reduces the impact of receiver coloring effects on classification, effectively removes receiver coloring effects, and overcomes the drawback of increased computational resources caused by adversarial training.

[0010] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for individual identification of radiation sources based on removing receiver coloration effects, including: The time-domain signal data of the radiation source to be processed is preprocessed to obtain the frequency-domain characteristics; The frequency domain features are used to obtain categorical domain-invariant features based on the DSC-DANN model. The DSC-DANN model includes a feature extractor, a radiation source category classifier, and a receiver category classifier. The feature extractor is composed of a multi-layer deep separable convolutional network and is used to generate categorical domain-invariant features. The receiver category classifier is connected to the feature extractor through a gradient inversion layer, so that the categorical domain-invariant features generated by the feature extractor cannot distinguish between different receiver domains. The radiation source category classifier enables the categorical domain-invariant features generated by the feature extractor to distinguish between different radiation source categories. Categorical domain-invariant features are filtered based on feature importance scores from a random forest to obtain radiation source classification-sensitive features; The radiation source classification sensitive features are input into a random forest classifier, and the radiation source classification results are output.

[0011] Secondly, this application provides a radiation source individual identification system based on removing receiver coloration effects, comprising: a preprocessing module, a model processing module, a filtering module, and a classifier output module; the preprocessing module is used to preprocess the time-domain signal data of the radiation source to be processed to obtain frequency-domain features; the model processing module is used to obtain categorical domain-invariant features based on the frequency-domain features using a DSC-DANN model; the DSC-DANN model includes a feature extractor, a radiation source category classifier, and a receiver category classifier, wherein the feature extractor is composed of multiple deep separable convolutional networks and is used to generate categorical domain-invariant features, the receiver category classifier is connected to the feature extractor through a gradient inversion layer, so that the categorical domain-invariant features generated by the feature extractor cannot distinguish between different receiver domains, and the radiation source category classifier enables the categorical domain-invariant features generated by the feature extractor to distinguish between different radiation source categories; the filtering module is used to filter the categorical domain-invariant features based on the feature importance score of a random forest to obtain radiation source classification sensitive features; the classifier output module is used to input the radiation source classification sensitive features into the random forest classifier and output the radiation source classification result.

[0012] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the radiation source individual identification method based on removing receiver staining effects as described above.

[0013] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the radiation source individual identification method based on removing receiver staining effects as described above.

[0014] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the radiation source individual identification method based on removing receiver staining effects as described above.

[0015] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method and system for individual radiation source identification based on removing receiver coloration effects. This method converts the time-domain signal data of the radiation source into frequency-domain features through preprocessing, enabling a more intuitive capture of the modulation characteristics of radar signals. It is particularly effective for complex modulation signals (LFM (Linear Frequency Modulation), phase coding, etc.) due to their superior spectral sparsity and separability, which helps deep learning models extract key discriminative features. To maintain good coloration effect removal while addressing the issues of low recognition accuracy and high computational cost caused by the use of linear layers in the feature extractor, classifier, and neighborhood discriminator networks of the original DANN model, a DSC-DANN model is used to process the frequency-domain features, obtaining category-invariant features. This achieves high recognition accuracy while reducing computational cost. To effectively improve SEI recognition accuracy, suppress noise feature interference, and reduce the impact of receiver coloring effect on classification, categorical domain-invariant features are filtered based on the feature importance score of random forest to obtain radiation source classification sensitive features. These sensitive features are then input into a random forest classifier to output radiation source classification results, thus completing the adaptive cross-domain radiation source individual recognition. This gradually reduces the impact of receiver coloring effect on classification, effectively removes receiver coloring effect, and overcomes the drawback of increased computational resources caused by adversarial training. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is an application environment diagram of a radiation source individual identification method based on removing receiver staining effects according to an embodiment of this application; Figure 2 A flowchart illustrating a method for individual radiation source identification based on removing receiver coloration effects, provided in an embodiment of this application; Figure 3A DSC-DANN model structure diagram for a radiation source individual identification method based on removing receiver staining effects provided in an embodiment of this application; Figure 4 A flowchart illustrating the secondary screening of a random forest based on the weighted index Gini impurity index, provided in an embodiment of this application; Figure 5 A detailed flowchart of a radiation source individual identification method based on removing receiver staining effects provided in an embodiment of this application; Figure 6 A schematic diagram of the functional modules of a radiation source individual identification system based on removing receiver coloration effects, provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] The radiation source individual identification method based on removing receiver staining effects provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send the time-domain signal data of the radiation source to be processed to server 104. Server 104 receives the time-domain signal data of the radiation source to be processed, preprocesses the time-domain signal data of the radiation source to obtain frequency domain features, and then uses the frequency domain features to obtain categorical domain-invariant features based on the DSC-DANN model. The DSC-DANN model includes a feature extractor, a radiation source category classifier, and a receiver category classifier. The feature extractor is composed of a multi-layer deep separable convolutional network and is used to generate categorical domain-invariant features. The receiver category classifier is connected to the feature extractor through a gradient inversion layer, so that the categorical domain-invariant features generated by the feature extractor cannot distinguish between different receiver domains. The radiation source category classifier enables the categorical domain-invariant features generated by the feature extractor to distinguish between different radiation source categories. The categorical domain-invariant features are then filtered based on the feature importance score of a random forest to obtain radiation source classification-sensitive features. The radiation source classification-sensitive features are input into the random forest classifier, and the radiation source classification result is output. Server 104 can feed back the obtained radiation source classification result to terminal 102. Furthermore, in some embodiments, the radiation source individual identification method based on removing receiver coloration effects can also be implemented separately by the server 104 or the terminal 102. For example, the terminal 102 can directly preprocess the time-domain signal data of the radiation source to be processed, and then obtain the radiation source classification result through the DSC-DANN model, the feature importance score of the random forest, and the random forest classifier. Alternatively, the server 104 can preprocess the obtained time-domain signal data of the radiation source to be processed from the data storage system, and then obtain the radiation source classification result through the DSC-DANN model, the feature importance score of the random forest, and the random forest classifier.

[0021] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0022] In one exemplary embodiment, such as Figure 2As shown, a method for individual radiation source identification based on removing receiver coloration effects is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 100 to 400. Among them: Step 100: Preprocess the time-domain signal data of the radiation source to be processed to obtain frequency-domain features.

[0023] Step 200: Obtain categorical domain-invariant features from the frequency domain features based on the DSC-DANN model. The DSC-DANN model includes a feature extractor, a radiation source category classifier, and a receiver category classifier. The feature extractor consists of multiple deep separable convolutional networks used to generate categorical domain-invariant features. The receiver category classifier is connected to the feature extractor through a gradient inversion layer, ensuring that the categorical domain-invariant features generated by the feature extractor cannot distinguish between different receiver domains. The radiation source category classifier enables the categorical domain-invariant features generated by the feature extractor to distinguish between different radiation source categories.

[0024] Step 300: The categorical domain-invariant features are filtered based on the feature importance scores of random forest to obtain radiation source classification sensitive features.

[0025] Step 400: Input the radiation source classification sensitive features into the random forest classifier and output the radiation source classification results.

[0026] By implementing steps 100 to 400 above, this application preprocesses the time-domain signal data of the radiation source into frequency-domain features, which can more intuitively capture the modulation characteristics of the radar signal, especially for complex modulation signals (LFM, phase coding, etc.) with better spectral sparsity and separability, which helps deep learning models extract key discriminative features. To maintain good coloring effect removal while addressing the problems of low recognition accuracy and high computational cost caused by the use of linear layers in the feature extractor, classifier, and neighborhood discriminator networks of the original DANN model, a DSC-DANN model is used to process the frequency-domain features into category-specific neighborhood-invariant features. To effectively improve SEI recognition accuracy, suppress noise feature interference, and reduce the impact of receiver coloring effects on classification, the category-specific neighborhood-invariant features are filtered based on the feature importance scores of a random forest to obtain radiation source classification-sensitive features; these features are then input into a random forest classifier to output the radiation source classification results.

[0027] In another exemplary embodiment of this application, considering that frequency domain features are generally more discriminative than time domain features and help improve the accuracy of radiation source identification, the time domain signal data of the radiation source to be processed is preprocessed to obtain frequency domain features. In this case, step 100 is replaced by steps 101 to 102: Step 101: Perform a discrete Fourier transform on the time-domain signal data of the radiation source to be processed to obtain the transformed frequency-domain features.

[0028] The specific time-domain signal data of the radiation source to be processed is as follows: Assuming the received radar radiation source signal has been sorted into one-dimensional time-domain radiation source signal data, the physical layer characteristics of the radiation source mainly originate from the nonlinear system response of the power amplifier, and the high-frequency signal to be radiated input to the power amplifier. for: (1).

[0029] in, The signal is a modulated signal from baseband. and Here, j represents the carrier frequency and sampling frequency, respectively, j is the imaginary unit, and n is the number of cycles.

[0030] Taylor polynomials are introduced to describe the specificity of the power amplifier, and the power amplifier output in a single-hop scenario is discussed. for: (2).

[0031] in, These are the Taylor polynomial coefficients. Let m be the order of the polynomial, and m be the order of the Taylor expansion.

[0032] Because the receiver intercepts the sampled signal from the i-th radiation source for: (3).

[0033] in, The fading coefficient from the transmitter to the receiver is denoted as . This is additive noise in the channel.

[0034] Based on the above, it can be seen that the receiver intercepts the sampled signal from the i-th radiation source. for: (4).

[0035] In the deep learning framework for radiation source identification, data preprocessing is an important step. The Discrete Fourier Transform (DFT) is an efficient method for converting time-domain signals into frequency-domain signals.

[0036] Assume the received signal (i.e., the time-domain signal data of the radiation source to be processed) is (Length is N, Its frequency domain characteristics after discrete Fourier transform are as follows: (5).

[0037] in, It is the first frequency domain signal Complex components at each frequency point.

[0038] To prevent spectral leakage, frequency domain features are typically windowed. This application uses a Hamming window, defined as follows: (6).

[0039] The windowing process involves processing the time-domain signal data of the radiation source to be processed. With window function Multiplying point by point, the windowed time-domain signal data is obtained as follows: (7).

[0040] Therefore, the frequency domain characteristics of the windowed time-domain signal data after discrete Fourier transform are as follows: (8).

[0041] This is the final signal form obtained in the data preprocessing stage. During preprocessing, considering that frequency domain features are generally more discriminative than time domain features and help improve the accuracy of radiation source identification, the frequency domain features of the signal are quickly extracted using DFT, facilitating the training and inference of deep learning models. To address the interference of spectral leakage on feature extraction, windowing is applied to provide more robust frequency domain input.

[0042] Performing a DFT transform on the input one-dimensional radiation source time-domain signal data converts the time-domain signal data into frequency-domain features, which can more intuitively capture the modulation characteristics of radar signals. In particular, it is better for the spectral sparsity and separability of complex modulation signals (LFM, phase coding, etc.), which helps deep learning models extract key discriminative features.

[0043] Step 102: Perform Combat de-batch processing on the transformed frequency domain features to obtain frequency domain features with preliminary elimination of platform difference features.

[0044] The Combat de-batch processing specifically includes: Step 112: Standardize the transformed frequency domain features so that the mean of all samples of the transformed frequency domain features is 0 and the standard deviation is 1, thus obtaining the standardized frequency domain features.

[0045] Step 122: Calculate the mean and standard deviation of the standardized frequency domain features for each batch; wherein the batches are formed due to their origin from different receiver platforms.

[0046] Step 132: Based on the mean and standard deviation of each batch, perform batch adjustment on the samples of the standardized frequency domain features of each batch to obtain the batch-processed frequency domain features; the batch adjustment is: the difference between the sample and the mean of the batch is calculated, and then compared with the standard deviation of the batch.

[0047] Step 142: The frequency domain features after batch adjustment are adjusted globally using the global mean and global standard deviation to obtain frequency domain features with preliminary elimination of platform difference features; the global distribution adjustment is: multiply all samples of the frequency domain features after batch adjustment by the global standard deviation, and then sum them with the global mean.

[0048] In one specific implementation, the Combat de-batch processing consists of the following steps: (1) Data standardization Standardize all frequency domain features after discrete Fourier transform so that their mean is 0 and their standard deviation is 1 across all samples: (9).

[0049] in, It is the global mean of all samples. It is the global standard deviation of all samples. To complete the standardized data, This is the original data.

[0050] (2) Batch effect modeling Assuming batch The mean and standard deviation can be expressed as adjustments to the global mean and standard deviation: (10); (11).

[0051] in, It is a sample In batch The mean of the middle, It is a sample In batch The standard deviation in and It is a batch Adjust the parameters.

[0052] (3) Batch effect adjustment By adjusting the mean and standard deviation of each batch, the data is normalized to a global distribution: (12).

[0053] The function of this formula is to divide the batch mean Adjust to global mean Batch Standard deviation Adjust to global standard deviation , This is an adjustment value for the global distribution.

[0054] (4) Parameter estimation parameter and Estimate using the following formula: (13); (14).

[0055] In summary, the final adjustment formula is: (15).

[0056] This application employs Combat de-batch processing for data preprocessing in the SEI task. Its core objective is to eliminate device-related batch bias by aligning the feature distributions of data collected from different receivers, thus initially removing receiver coloring effects from the data. This method estimates and corrects differences in mean and standard deviation between receiver platforms, achieving consistent data distribution across receivers while preserving radiation source category discrimination information. This improves the generalization performance of the classification model in heterogeneous device scenarios and avoids the decrease in classification accuracy caused by receiver differences.

[0057] In the SEI task, the DANN model, through feature decoupling, can effectively learn domain-independent general features, making the model more robust in cross-domain tasks and applicable to scenarios with different data distributions. Simultaneously, through adversarial training, the DANN model can effectively reduce distributional differences between domains, improving its generalization performance in cross-domain tasks. Based on these advantages, the DANN model is widely used in the field of transfer learning.

[0058] However, DANN models still suffer from computational inefficiencies that make them unsuitable for real-world applications. Their training process requires simultaneous optimization of both the classifier and feature extractor, leading to a significant increase in computational complexity. The classifier structure is relatively simple and fixed, typically a fully connected layer architecture, while the feature extractor has room for improvement. The performance of a DANN model largely depends on the quality of its feature extraction. Therefore, it is crucial to reduce computational and storage overhead while maintaining the expressive power of feature extraction.

[0059] In another exemplary embodiment of this application, addressing the drawbacks of low recognition accuracy and high computational cost caused by the use of linear layers in the feature extractor, classifier, and domain discriminator networks of the original DANN model, this application obtains categorical domain-invariant features based on the frequency domain features using the DSC-DANN model. By utilizing DSC (Depthwise Separable Convolution) to improve the network structure, the computational and storage overhead is significantly reduced while maintaining feature expressive power. Step 200 is replaced by steps 201 to 203: Step 201: In the DSC-DANN model, the feature extractor performs local feature extraction on the frequency domain features through deep convolution, and then fuses the information of different channels through pointwise convolution to generate categorical domain-invariant features.

[0060] This application improves the structure of the DANN model by replacing the original fully connected layers with depthwise separable convolutional layers, and proposes the network architecture of the DSC-DANN model. The model structure diagram is shown below. Figure 3 As shown.

[0061] Feature extractor The main body consists of a three-layer DSC network. DSC is an efficient convolution operation that decomposes standard convolution into DWConv (Depthwise Convolution) and PWConv (Pointwise Convolution), thus significantly reducing the number of model parameters and computational complexity. When extracting features, DSC extracts local features through DWConv and then fuses information from different channels through PWConv. This helps the model learn the non-linear combination relationships between features, improving the efficiency of traditional convolution operations and overcoming the shortcomings of insufficient feature extraction capabilities caused by the use of fully connected layers in ordinary DANN networks. The following section details the DSC feature extraction process using mathematical formulas.

[0062] For any layer of the DSC, DWConv processing is performed first, assuming the input of that layer... (in For sequence length, (Number of input channels); convolution kernel , The kernel size; output ( Similar to the number of input channels, DWConv does not change the number of channels. Therefore, the input / output formula for DWConv processing is: (16).

[0063] in, For the input location index, ; For output channel index, ; This is the kernel position index. .

[0064] Next, PWConv processing is performed, assuming the input of this layer is... (in For sequence length, (Number of input channels); convolution kernel ( convolution, (Number of output channels); Output The input / output formulas for PWConv processing are then: (17).

[0065] in, For the input location index, d This is the output channel index.

[0066] In terms of computational cost, compared with ordinary convolutional layers and fully connected layers, depthwise separable convolution (DSC) also demonstrates a significant advantage. The network parameter configurations are shown in Table 1, and the parameter complexity comparison is shown in Table 2. Analysis of the data in the tables shows that, theoretically, compared with fully connected layers and ordinary convolutional layers, the parameter size of DSC is reduced to 1% and 33.9% respectively, effectively solving the problem of high computational cost.

[0067] Table 1 Network parameter configurations

[0068] Table 2 Comparison of parameter complexity of various networks used in feature extractors

[0069] Step 202: The receiver category classifier is connected to the feature extractor through a gradient inversion layer. The gradient inversion layer reverses the gradient direction, making the categorical domain-invariant features generated by the feature extractor unable to distinguish between different receiver domains.

[0070] Step 203: The radiation source category classifier is used to constrain the pure features of radiation sources so that the categorical domain-invariant features generated by the feature extractor can distinguish different radiation source categories.

[0071] Feature extractor Generated features Input radiation source category classifier and receiver category classifier Radiation source category classifier and receiver category classifier It adopts a fully connected layer architecture (such as...) Figure 3 As shown), in the receiver category classifier The pre-embedded layer is a GRL (Gradient Reversal Layer). The GRL reverses the gradient direction, making the feature extractor... Generated features Unable to distinguish between different receiver domains, the receiver category classifier By distinguishing the source, the receiver-domain-specific coloring features are obscured, forcing the features to have cross-receiver domain invariance.

[0072] To remove receiver coloration effects while preventing the removal of radiation source purity features, a radiation source category classifier is used. Constraints preserve the purity characteristics of the radiation source.

[0073] By calculating the loss function value and backpropagating to optimize the parameters, the trained DSC-DANN model is finally obtained. .

[0074] Therefore, the loss function can be expressed as: (18).

[0075] in, For the first One input sample, The total number of samples, These are the internal parameters of the feature extractor. These are the internal parameters of the radiation source category classifier. These are the internal parameters of the receiver category classifier; Classify and label radiation sources. Classify and label receivers by category; The output vector of the feature extractor network. The output vector of the radiation source category classification network. The output vector of the receiver category classification network; Classifying network losses by radiation source category Receiver category classification network loss.

[0076] This application employs DSC-DANN adversarial training to deceive the receiver category classifier in the receiver feature domain, aiming to confuse features from different receiver domains and thus remove receiver coloring effects. To prevent the removal of radiation source classification-sensitive features while removing receiver coloring effects, pure radiation source features are preserved through classifier loss constraints.

[0077] Random forest classification has wide applications in SEI tasks. By integrating the voting results of multiple decision trees, random forests can effectively reduce the risk of overfitting and improve classification accuracy, especially suitable for identifying radiation source signals in complex electromagnetic environments. Meanwhile, radiation source signals may be affected by noise interference or feature fluctuations; the random feature selection and bagging mechanism of random forests can enhance model robustness and reduce interference from outlier samples.

[0078] In another exemplary embodiment of this application, after removing receiver coloring effects in the DSC-DANN model to obtain pure categorical domain-invariant features, in order to obtain more accurate radiation source classification results, it is necessary to screen radiation source classification-sensitive features before classifier training. Therefore, step 300 is replaced by steps 301-302: Step 301: Calculate the feature importance score of each feature when splitting the decision tree node of the random forest based on the weighted index Gini impurity for the categorical domain-invariant features.

[0079] The Gini impurity S, a computational metric for node fission in traditional random forests, is defined as: (19).

[0080] Analysis of Gini impurity reveals two significant drawbacks: first, when class distributions are severely imbalanced, Gini impurity may fail to effectively distinguish the impact of splitting on a minority of classes; second, the sensitivity of Gini impurity to probability distribution is fixed (depending solely on...). This approach cannot adapt to the needs of different data distributions. Therefore, this application improves the Gini impurity index and proposes a weighted index Gini impurity.

[0081] In one specific implementation, the categorical domain-invariant features are used to calculate the feature importance score of each feature when splitting at the decision tree node of the random forest based on the weighted Gini impurity index. Specifically, this includes: Step 311: Calculate the weighted exponential Gini impurity of each feature when it splits at nodes in all decision trees, based on the category discrimination difficulty weights of the distance metric. Step 312: The split contribution of each feature in a single decision tree is obtained from the weighted exponential Gini impurity that each feature reduces when splitting at a node. Step 313: The total contribution of each feature in the random forest is obtained by summing the weighted exponential Gini impurities that each feature reduces during all node splits. Step 314: Obtain the feature importance score based on the split contribution of each feature in a single decision tree and its total contribution in the random forest.

[0082] This application aims to assign higher weights to a minority of classes and lower weights to the majority of classes. To this end, this application uses the DICDW (Distance Induced Class Discriminative Weight) metric, assuming the dataset contains K classes, and the sample set for each class k is... Its sample mean is The covariance matrix is For categories The sample mean and covariance are calculated as follows: (20-1).

[0083] (20-2).

[0084] in, x For the sample.

[0085] Next, the inter-class distance matrix is ​​calculated. Since Mahalanobis distance takes into account the covariance structure between classes and is more sensitive to data distribution, this application uses Mahalanobis distance. : (twenty one).

[0086] Define category Difficulty of differentiation The average distance between this category and all other categories: (twenty two).

[0087] in It is a category and The Mahalanobis distance between them.

[0088] To translate the difficulty of differentiation into weights, it's natural to want the difficulty of differentiation to be greater (the smaller the average distance) and the higher the weight. The higher the value (because harder-to-distinguish categories need more attention), the better. Additionally, weight normalization is ensured so that the sum of the weights for all categories is 1. Based on this, the final DICDW definition is: (twenty three).

[0089] For the class distribution of a node ,remember It is a category The probability, It represents the total number of categories. WEGI (Weighted Exponential Gini Impurity) is defined as follows: (twenty four).

[0090] Among them, the index Parameters used to control sensitivity to probability distributions. When WEGI is more sensitive to probability differences when When it degenerates into Gini impurity, when It reduces sensitivity, making it suitable for scenarios that are more robust to noise. And... It is a category The weights are used to adjust for the effects of class imbalance.

[0091] Feature importance reflects the degree to which each feature contributes to the prediction results in the random forest model. Specifically, it is calculated based on the total contribution of each feature to reducing impurity when splitting nodes across all decision trees. During the training of each decision tree, the split contribution of a feature refers to the reduction in WEGI value when that feature splits, while the total contribution of each feature is the sum of its reduction in WEGI value across all splits.

[0092] The formula for calculating the contribution of each feature to the splitting of a single decision tree is as follows: (25).

[0093] in, The total classification contribution of all nodes in the decision tree; The formula for calculating the total contribution of each feature in the random forest is as follows: (26).

[0094] in, As a feature, Features exist T The total contribution of each decision tree Features Contribution to splitting in a single decision tree; The formula for calculating feature importance score is: (27).

[0095] in, im_score For feature importance scores, D The total number of features.

[0096] Step 302: Based on the preset number of feature selections, select the features with the highest feature importance scores, and use the selected features as sensitive features for radiation source classification.

[0097] Optionally, the preset number of features to be selected can be set to 100. The top 100 features with the highest feature importance scores are selected as the sensitive features for radiation source classification. The classifier is then retrained using these sensitive features. The specific flowchart is as follows: Figure 4 As shown in the figure. Then, the radiation source classification sensitive features are input into the random forest classifier, and the radiation source classification results are output.

[0098] To effectively improve SEI recognition accuracy and suppress noise interference, this application, based on the aforementioned WEGI-RF (Weighted Exponential Gini Impurity-Random Forest) feature selection method, addresses the shortcomings of the original Gini impurity-based random forest feature selection method, namely its inability to adapt to class imbalance and its poor adaptability to changes in probability distribution. This application proposes a weighted exponential Gini impurity index, improving upon the original Gini impurity formula. To further enhance the method's robustness to class imbalance, a DICDW index is defined for the weighting factors. The WEGI-RF feature selection method is used to filter features extracted by the DSC-DANN feature extractor during model training, obtaining features sensitive to radiation source classification, thus providing better features for the final radiation source classification. Finally, the selected features are used to retrain the classifier using the WEGI-RF feature selection method, and the model is saved for subsequent model validation.

[0099] This application validates the method on a validation set using a pre-trained classifier, observing its classification performance. To further verify the effectiveness of the proposed method, ablation experiments, robustness experiments for differences in the number of radiation sources, validation experiments for imbalanced class numbers, gradient inversion intensity factor optimization experiments, WEGI exponential factor comparison experiments, and feature selection number optimization experiments were conducted. Experimental results on a dataset of self-collected real radar radiation source individual signals demonstrate that the proposed method achieves competitive performance on the aforementioned tasks.

[0100] The flowchart of the radiation source individual identification method based on removing receiver staining effects is as follows: Figure 5 As shown. Assume the original training dataset is... ,in Indicates sample features, This indicates the label to which the sample belongs. There are a total of K radiation sources in the data. This represents the total number of samples in the training set, where i is the sample index. The test set data is... ,in Indicates sample features, This indicates the label to which the sample belongs. There are a total of K radiation sources in the data. This represents the total number of samples in the test set. For and There are two types of datasets, each of which can be further divided into three subsets, corresponding to onshore platform receivers (Platform A), offshore platform receivers (Platform B), and aerial platform receivers (Platform C). For example, the test set... Can be divided into , and The molecular sets are labeled A, B, and C. Furthermore, the training and test sets do not overlap.

[0101] The overall framework of the radiation source individual identification method based on removing receiver staining effects includes a data preprocessing stage, a DSC-DANN model training stage, and a WEGI-RF method feature selection and classifier training stage.

[0102] First, using the original training set Data training network model. In the data preprocessing stage, the input is the actual radar radiation source time-domain signal data. The frequency domain features of the signal are extracted using DFT. Standardization is used to transform the frequency domain features into a distribution with a mean of 0 and a standard deviation of 1. Combat de-batching is then used to initially remove receiver coloration effects, resulting in a pre-cleaned and corrected training set. The data normalizer and Combat de-batch processing model are saved. During the DSC-DANN model training phase, the corrected dataset obtained in the previous phase is first processed... Through feature extractor Feature extraction, in the feature extractor In this study, considering the limitations of traditional DANN models that use fully connected layers for feature extraction (resulting in weak feature capture and slow computation speed), depthwise separable convolution was employed. Depthwise separable convolution is an effective convolution operation that can significantly reduce the number of model parameters and computational complexity. The features extracted by the feature extractor... Input into the radiation source category classifier and receiver category classifier In the receiver category classifier The pre-embedding incorporates a GRL, which reverses the gradient direction, thus improving the feature extractor. Generated features Unable to be classified by receiver category classifier By differentiating sources, receiver-domain-specific coloring features are obscured, forcing features to exhibit cross-receiver domain invariance. To remove receiver coloring effects while simultaneously preventing the removal of source purity features, a source category classifier is used. Constraining the purity features of the radiation source ensures that while removing receiver taint features, the purity features of the radiation source are preserved. By calculating the loss function value and backpropagating to optimize the parameters, the trained DSC-DANN model is finally obtained. In the feature selection and classifier training stages of the WEGI-RF method, the features output by the DSC-DANN network feature extractor are... The WEGI-RF method is used to calculate feature importance scores, and features sensitive to radiation source classification are selected based on the importance scores. Finally, the selected features are used to train a classifier and the results are saved.

[0103] Secondly, use the test set Test data, test set data The method is passed sequentially through a data normalizer, a Combat model, a DSC-DANN model, and a classifier model to output the recognition accuracy and confusion matrix, thereby verifying the effectiveness of the proposed method.

[0104] Based on the same inventive concept, this application also provides a system for implementing the aforementioned radiation source individual identification system based on removing receiver coloration effects. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the radiation source individual identification system based on removing receiver coloration effects provided below can be found in the limitations of the radiation source individual identification method based on removing receiver coloration effects described above, and will not be repeated here.

[0105] In one exemplary embodiment, such as Figure 6 As shown, a radiation source individual identification system based on removing receiver staining effects is provided, including: a preprocessing module 1, a model processing module 2, a screening module 3, and a classifier output module 4.

[0106] Preprocessing module 1 preprocesses the time-domain signal data of the radiation source to be processed to obtain frequency-domain features; model processing module 2 is used to obtain categorical domain-invariant features based on the frequency-domain features using a DSC-DANN model; the DSC-DANN model includes a feature extractor, a radiation source category classifier, and a receiver category classifier. The feature extractor is composed of multiple deep separable convolutional networks and is used to generate categorical domain-invariant features. The receiver category classifier is connected to the feature extractor through a gradient inversion layer, so that the categorical domain-invariant features generated by the feature extractor cannot distinguish between different receiver domains, while the radiation source category classifier enables the categorical domain-invariant features generated by the feature extractor to distinguish between different radiation source categories; filtering module 3 filters the categorical domain-invariant features based on the feature importance score of a random forest to obtain radiation source classification-sensitive features; classifier output module 4 inputs the radiation source classification-sensitive features into the random forest classifier and outputs the radiation source classification result.

[0107] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 7 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data for individual radiation source identification based on removing receiver coloration effects. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for individual radiation source identification based on removing receiver coloration effects.

[0108] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0109] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0110] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0111] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0112] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0113] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0114] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0115] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0116] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for individual identification of radiation sources based on removing receiver coloration effects, characterized in that, include: The time-domain signal data of the radiation source to be processed is preprocessed to obtain the frequency-domain characteristics; The frequency domain features are used to obtain category-specific, domain-invariant features based on the DSC-DANN model; The DSC-DANN model includes a feature extractor, a radiation source category classifier, and a receiver category classifier. The feature extractor is composed of multiple deep separable convolutional networks and is used to generate categorical domain-invariant features. The receiver category classifier is connected to the feature extractor through a gradient inversion layer, so that the categorical domain-invariant features generated by the feature extractor cannot distinguish between different receiver domains. The radiation source category classifier enables the categorical domain-invariant features generated by the feature extractor to distinguish between different radiation source categories. The following steps are taken to filter categorical domain-invariant features based on their feature importance scores in a random forest to obtain radiation source classification-sensitive features: First, the feature importance score of each categorical domain-invariant feature is calculated at the decision tree node splits in the random forest using a weighted exponential Gini impurity index. Second, features with the highest feature importance scores are selected based on a preset number of feature selections, and these selected features are used as radiation source classification-sensitive features. The calculation of the categorical domain-invariant feature importance score at the decision tree node splits in the random forest using a weighted exponential Gini impurity index is further taken. This includes: First, the weighted exponential Gini impurity of each feature at node splits in all decision trees is calculated based on the class discrimination difficulty weight of a distance metric. Third, the split contribution of each feature in a single decision tree is obtained from the reduction in weighted exponential Gini impurity at node splits. Fourth, the total contribution of each feature in the random forest is obtained from the sum of the reductions in weighted exponential Gini impurity at all node splits. Fifth, the feature importance score is obtained based on the split contribution of each feature in a single decision tree and its total contribution in the random forest. The formula for calculating the class discrimination difficulty weight of the distance metric is as follows: ; ; ; ; ;in, Difficulty weights are used to differentiate categories. For category The difficulty of distinguishing them For category The difficulty of distinguishing them The total number of categories, For category and Mahalanobis distance between them For category The sample set, For category The sample mean, For category The covariance matrix, where x is the sample. For category The sample mean, For category The covariance matrix; the formula for calculating the weighted exponential Gini impurity is: Wherein, WEGI is the weighted index Gini impurity. For category The probability, Here are the parameters used to control sensitivity to the probability distribution; the formula for calculating the split contribution of each feature in a single decision tree is: ;in, This represents the total classification contribution of all nodes in the decision tree; the formula for calculating the total contribution of each feature in the random forest is: ;in, As a feature, Features The total contribution of T decision trees Features The formula for calculating the feature importance score in the splitting contribution of a single decision tree is as follows: Where im_score is the feature importance score, and D is the total number of features; The radiation source classification sensitive features are input into a random forest classifier, and the radiation source classification results are output.

2. The method for individual identification of radiation sources based on removing receiver staining effects according to claim 1, characterized in that, The time-domain signal data of the radiation source to be processed is preprocessed to obtain frequency-domain features, specifically including: Perform a Discrete Fourier Transform on the time-domain signal data of the radiation source to be processed to obtain the transformed frequency-domain characteristics; The transformed frequency domain features are subjected to Combat de-batch processing to obtain frequency domain features with preliminary elimination of platform difference features.

3. The method for individual radiation source identification based on removing receiver staining effects according to claim 1, characterized in that, The frequency domain features are used to obtain categorical, domain-invariant features based on the DSC-DANN model, specifically including: In the DSC-DANN model, the feature extractor extracts local features from the frequency domain through deep convolution, and then fuses information from different channels through pointwise convolution to generate categorical domain-invariant features. The receiver category classifier is connected to the feature extractor through a gradient inversion layer. The gradient inversion layer reverses the gradient direction, making the categorical domain-invariant features generated by the feature extractor unable to distinguish between different receiver domains. The radiation source category classifier is used to constrain the pure features of radiation sources, so that the categorical, domain-invariant features generated by the feature extractor can distinguish different radiation source categories.

4. A radiation source individual identification system based on removing receiver staining effects, characterized in that, include: The preprocessing module is used to preprocess the time-domain signal data of the radiation source to be processed to obtain frequency-domain features; The model processing module is used to obtain categorical domain-invariant features from the frequency domain features based on the DSC-DANN model; The DSC-DANN model includes a feature extractor, a radiation source category classifier, and a receiver category classifier. The feature extractor is composed of multiple deep separable convolutional networks and is used to generate categorical domain-invariant features. The receiver category classifier is connected to the feature extractor through a gradient inversion layer, so that the categorical domain-invariant features generated by the feature extractor cannot distinguish between different receiver domains. The radiation source category classifier enables the categorical domain-invariant features generated by the feature extractor to distinguish between different radiation source categories. The filtering module is used to filter categorical domain-invariant features based on the feature importance scores of a random forest to obtain radiation source classification-sensitive features. Specifically, this includes: calculating the feature importance score of each feature when splitting at a decision tree node in the random forest using a weighted Gini impurity index; filtering features with the highest feature importance scores based on a preset number of feature selections, and using these selected features as radiation source classification-sensitive features; and calculating the feature importance score of each feature when splitting at a decision tree node in the random forest using a weighted Gini impurity index. The importance score specifically includes: calculating the weighted exponential Gini impurity of each feature when splitting nodes across all decision trees based on the class discrimination difficulty weights of the distance metric; obtaining the splitting contribution of each feature in a single decision tree from the reduction in weighted exponential Gini impurity when splitting nodes; obtaining the total contribution of each feature in the random forest from the sum of the reduction in weighted exponential Gini impurity of each feature across all node splits; and obtaining the feature importance score based on the splitting contribution of each feature in a single decision tree and its total contribution in the random forest. The formula for calculating the class discrimination difficulty weights of the distance metric is as follows: ; ; ; ; ;in, Difficulty weights are used to differentiate categories. For category The difficulty of distinguishing them For category The difficulty of distinguishing them The total number of categories, For category and Mahalanobis distance between them For category The sample set, For category The sample mean, For category The covariance matrix, where x is the sample. For category The sample mean, For category The covariance matrix; the formula for calculating the weighted exponential Gini impurity is: Wherein, WEGI is the weighted index Gini impurity. For category The probability, Here are the parameters used to control sensitivity to the probability distribution; the formula for calculating the split contribution of each feature in a single decision tree is: ;in, This represents the total classification contribution of all nodes in the decision tree; the formula for calculating the total contribution of each feature in the random forest is: ;in, As a feature, Features The total contribution of T decision trees Features The formula for calculating the feature importance score in the splitting contribution of a single decision tree is as follows: Where im_score is the feature importance score, and D is the total number of features; The classifier output module is used to input the radiation source classification sensitive features into the random forest classifier and output the radiation source classification results.

5. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the radiation source individual identification method based on removing receiver staining effects as described in any one of claims 1-3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the individual identification method for radiation sources based on removing receiver staining effects as described in any one of claims 1-3.

7. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the individual identification method for radiation sources based on removing receiver staining effects as described in any one of claims 1-3.

Images