Partially adversarial domain adaptive source individual recognition method based on complex convolution

Abstract

The application provides a partial adversarial domain self-adaptive radiation source individual identification method based on complex convolution. The application uses a weighted domain classifier based on Softmax to weight the features of the source domain and the target domain, and filters out the sample categories in the source domain that do not match the target domain. Then, another domain discriminator and label predictor based on Softmax are constructed, and a Wasserstein distance guided adversarial unsupervised domain adaptation is introduced, so that the distribution alignment between the source domain and the target domain data is realized. Finally, the trained identification model can identify the radiation source equipment in the partial domain self-adaptive scene, and the accuracy of the radiation source identification is greatly improved. Through the comprehensive application of the complex convolution, the adversarial training and the unsupervised domain adaptation and other technologies, the application provides an effective solution for the radiation source individual identification in the domain self-adaptive scene.

Description

Technical Field

[0001] This invention belongs to the field of radiation source individual identification technology, and particularly relates to a partially adversarial domain adaptive radiation source individual identification method based on complex convolution. Background technology:

[0002] Individual emitter identification (SEI) identifies individual emitters by extracting a radio frequency fingerprint (RFF) from the received signal. The RRF fingerprint is extracted from unintentional modulation information (UMI) appended to the signal pulse. This unintentional modulation information stems from unique hardware defects, such as differences in wiring, circuit board materials, and discrepancies between actual and standard values ​​of electronic component parameters. As a key technology in electronic reconnaissance, SEI is widely used in both military and civilian fields. Traditional identification schemes mainly consist of two stages: extracting the RRF fingerprint from the received signal and classifying the features. Currently, deep learning-based methods have achieved groundbreaking progress in SEI. Compared to traditional methods, deep learning significantly improves both identification accuracy and training speed, making deep learning-based SEI methods a mainstream research area.

[0003] However, most deep learning-based methods for identifying radiation sources are data-driven, typically requiring the preservation of the main statistical properties or features of the signal of interest throughout the identification process. That is, the source domain data (SD) and target domain data (TD) need to follow the same distribution. When the source and target domain data have different class spaces and different probability distributions, this is called a Partial Domain Adaptive (PDA) scenario. Models trained directly using existing data may lack generalization performance, resulting in lower classification accuracy in PDA scenarios. Therefore, improving the accuracy of radiation source identification in PDA scenarios has always been a challenge and bottleneck for deep learning-based identification models.

[0004] Therefore, this application focuses on solving the problem of individual radiation source identification in partially adaptive domain scenarios, and proposes a partially adversarial domain adaptive radiation source identification method based on complex convolution. In the proposed method, the two-channel I / Q signals emitted by the radiation source are first converted into one-dimensional sequence complex signals. After adding noise, these signals are directly input into a feature extractor composed of a complex convolutional neural network for feature extraction. A weighted domain classifier filters out samples from the source domain data that do not match the target domain data. Then, adversarial training between the feature extractor and the domain discriminator aligns the probability distributions of the source and target domain data. Finally, the results are input into a label predictor for individual radiation source identification. The partially adversarial domain adaptive radiation source identification method based on complex convolutional neural networks proposed in this application further solves the problem of difficulty in model training due to the mismatch between the probability distributions of training and test data in partially adaptive domain scenarios. Summary of the Invention

[0005] The purpose of this application is to provide a partially adversarial adaptive radiation source individual recognition method based on complex convolution, in order to solve the problems of underfitting, poor generalization performance and poor recognition accuracy of existing deep learning-based radiation source individual recognition methods when directly trained with existing data in partially adaptive scenarios.

[0006] The technical solution of the present invention is as follows:

[0007] A partially adversarial domain adaptive radiation source individual identification method based on complex convolution includes the following steps:

[0008] (1) The two-channel I / Q signals emitted by the radiation source collected in the scene are converted into complex signals in the form of a one-dimensional sequence and directly input into the feature extractor composed of a complex convolutional neural network for feature extraction.

[0009] Specifically, the I / Q signal emitted by the radiation source device is represented in the following form:

[0010] The data emitted by the radiation source is assumed to be an I / Q modulated signal, consisting of two components: an I-carrier and a Q-carrier, representing the in-phase and quadrature components, respectively. The Q-carrier is obtained by rotating the I-carrier by 90 degrees.

[0011]

[0012]

[0013] Therefore, the two-channel I / Q signals emitted by the radiation source are converted into one-dimensional complex signals, and the input I / Q signals are expressed as x = x r +ix i Where x rThe I signal represents the real part of a complex signal sample, x i The Q signal represents the imaginary part of a complex signal sample.

[0014] Feature extraction of the acquired complex I / Q signals is performed using a complex convolutional neural network with a one-dimensional complex convolutional kernel.

[0015] Specifically, the complex convolution process is represented as follows:

[0016] Let the complex sample of the input IQ signal be represented as x = x r +ix i Where i represents the imaginary unit, x r The I signal represents the real part of a complex signal sample, x i The Q signal represents the imaginary part of a complex signal sample. The weights of the complex convolution kernel are represented as w = a + ib, where a represents the real part of the weights and b represents the imaginary part. The convolution operation between the complex IQ signal and the complex convolution kernel is as follows:

[0017] w*x=(a*x r -b*x i )+i(b*x r +a*x i )

[0018] Where * represents the convolution operation, and the real and imaginary parts of the output after complex convolution are a*x, respectively. r -b*x i and b*x r +a*x i .

[0019] (2) Considering that the target domain data has a different distribution than the source domain data in the partial domain adaptation scenario, noise is added to the source domain signal samples to expand the robustness of the model:

[0020] Specifically, the noise-adding processing of the source domain signal samples is represented by the following generator function:

[0021] G(x s ,z,θ G )

[0022] The generator function receives source domain data and noise vectors z ~ Z sampled from a uniform distribution, with the aim of learning a mapping {x}. s ,z}→x f , where x f The data is trained to be similar to the target domain data.

[0023] (3) Based on the traditional domain adversarial training network, a weighted domain classifier is introduced and combined with the similarity measurement method to filter out sample categories that do not match the target domain in the source domain, thereby reducing the impact of domain shift caused by outlier samples in the source domain and improving the quality of training data.

[0024] A softmax-based weighted domain classifier, combined with a similarity measurement method, performs weighted processing on the source domain data:

[0025] Specifically, the weighted domain classifier module mainly consists of two steps: similarity calculation between the source domain data and the target domain data, and weighted processing. The two steps are as follows:

[0026] In actual training, source domain signal samples serve as training data, and target domain signal samples serve as test data. Similarity metrics can characterize the differences between source and target domain signal samples, thereby further ensuring that the categories of samples in the source domain match the categories of samples in the target domain. Therefore, the similarity between source and target domain data can be expressed as:

[0027] D ω (f s,t (x))=p(y=1|f s,t (x))=σ(σ(f s,t (x)))

[0028] Among them, f s,t (x) represents the representative features of the source and target domain samples extracted by the feature extractor. σ is the sigmoid activation function, and y is the label corresponding to the input feature.

[0029] If the probability of a classification feature being labeled as the source domain is "1", then In other words, the probability of features from outlier categories in the source domain is close to "1", while the probability of shared categories between the two domains is close to "0". Outlier categories in the source domain may mismatch with samples in the target domain, leading to domain shift. Therefore, outlier features should be assigned smaller weights to reduce the impact of domain shift on model training. During the weighting stage, the feature weighting function should be... Inversely proportional, the entire weighting process can be expressed as:

[0030]

[0031] Based on traditional adversarial domain adaptation (DA), given f s (x) (corresponding to the source domain sample in DA), for any f t (x) (corresponding to the target domain sample in DA) yields the optimal result. It can be represented as:

[0032]

[0033] Where, p s () represents the data distribution of the source domain samples, p t () represents the data distribution of the target domain samples.

[0034] (4) Construct a domain discriminator and a label predictor, and introduce Wasserstein distance to guide adversarial unsupervised domain adaptation, so that the data distribution in the source domain is aligned with that in the target domain, and obtain the labels corresponding to the samples in the target domain on the label predictor.

[0035] Specifically, in the partial-domain adaptive scenario for individual radiation source identification, the distributions of source and target domain data do not overlap, leading to instability in model training. The optimization objectives of the feature extractor G and the domain discriminator D can be expressed as:

[0036]

[0037] Among them, l wd (D,G) represents the loss function between the domain discriminator and the generator. Let it be the expected function that the samples in the target domain follow. Let λ be the expectation function that the source domain samples follow, and let λ be a hyperparameter that controls the balance between the input signal and the gradient penalty. For the gradient operator, ||..||2 is the 2-norm. Represented as the distribution X from the source domain s and feature extractor distribution X g Data points sampled in the middle:

[0038]

[0039] γ is a hyperparameter that balances the source domain samples and the samples after adding noise, and U[0,1] indicates that the data follows a uniform distribution.

[0040] The cross-entropy loss of the label predictor can be expressed as:

[0041]

[0042] Among them, y s It is the encoding x s The one-hot vector of the label. Finally, the overall optimization objective of the model can be expressed as:

[0043]

[0044] (5) A radiation source identification model is constructed using a complex neural network. Source domain samples and target domain samples are used for model training, and target domain samples are used for model testing.

[0045] (6) Use the radiation source identification model trained in step (5) to identify radiation source devices in the partial domain adaptive scenario.

[0046] The present invention has the following beneficial effects:

[0047] The method proposed in this invention addresses the issues of poor generalization performance and low recognition accuracy that deep learning models often exhibit in partially adaptive domain scenarios. This method converts the two-channel I / Q signals emitted by the acquired radiation source into a one-dimensional complex signal after adding noise, and then applies a complex convolutional neural network for feature extraction. Next, adversarial training is used to align the source domain data with the target domain data, and finally, a label predictor is used to identify the radiation source. Attached Figure Description

[0048] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0049] Figure 1 This is a flowchart illustrating a partially adversarial domain adaptive radiation source individual identification method based on complex convolution according to an exemplary embodiment of the present invention. Detailed Implementation

[0050] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0051] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0052] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0053] Figure 1This is a flowchart illustrating an adaptive radiation source individual identification method based on a partially adversarial domain using complex convolution, according to an exemplary embodiment. Figure 1 As shown, this method, when applied to a terminal, includes the following steps:

[0054] (1) The two-channel I / Q signals emitted by the radiation source collected in the scene are converted into complex signals in the form of a one-dimensional sequence and directly input into the feature extractor composed of a complex convolutional neural network for feature extraction;

[0055] (2) Add noise to the source domain signal samples to expand the robustness of the model.

[0056] (3) Based on the traditional domain adversarial training network, a weighted domain classifier is introduced and combined with the similarity measurement method to filter out the sample categories that do not match the target domain in the source domain, thereby reducing the impact of domain shift caused by outlier samples in the source domain.

[0057] (4) Construct a domain discriminator and a label predictor, introduce Wasserstein distance to guide adversarial unsupervised domain adaptation, align the data distribution in the source domain with that in the target domain, and obtain the label corresponding to the target domain sample on the label predictor;

[0058] (5) A radiation source identification model is constructed using a complex neural network. Source domain samples and target domain samples are used for model training, and target domain samples are used for model testing.

[0059] (6) Use the radiation source identification model trained in step (5) to identify radiation source devices in the partial domain adaptive scenario.

[0060] As can be seen from the above embodiments, this application proposes a partial adversarial domain adaptive radiation source individual recognition scheme based on complex convolution, which solves the problems of poor model generalization performance and poor recognition accuracy of deep learning methods in partial domain adaptive scenarios. It further optimizes the problems of class mismatch and unstable model convergence in the traditional ADA in aligning the data distribution of source domain data and target domain data. The proposed partial adversarial domain adaptive radiation source individual recognition method based on complex convolution can achieve relatively ideal model recognition accuracy in scenarios where the class spaces of training data and target domain data are different.

[0061] While adding noise to the acquired I / Q signals, the signals are also converted into one-dimensional complex signals. This not only further enhances the robustness of the model but also allows for the extraction of mutual coupling information from the I / Q signals through complex network processing, uncovering deeper features. A weighted domain classifier effectively filters out outlier samples from the source domain data, reducing domain shift issues caused by mismatch between source and target domain data. Furthermore, a domain discriminator and a label predictor are constructed, incorporating Wasserstein distance-guided adversarial unsupervised domain adaptation to align the data distributions in the source and target domains, and obtaining the corresponding labels for the target domain samples on the label predictor. The interaction of these four modules enables good recognition accuracy even when the class spaces of the source and target domain data differ.

[0062] In the specific implementation of step (1), a complex convolutional neural network is used to extract features from the acquired complex I / Q signals using a one-dimensional complex convolutional kernel:

[0063] Specifically, the complex convolution process is represented as follows:

[0064] Let the complex sample of the input IQ signal be represented as x = x r +ix i Where i represents the imaginary unit, x r The I signal represents the real part of a complex signal sample, x i The Q signal represents the imaginary part of a complex signal sample. The weights of the complex convolution kernel are represented as w = a + ib, where a represents the real part of the weights and b represents the imaginary part. The convolution operation between the complex IQ signal and the complex convolution kernel is as follows:

[0065] w*x=(a*x r -b*x i )+i(b*x r +a*x i )

[0066] Where * represents the convolution operation, and the real and imaginary parts of the output after complex convolution are a*x, respectively. r -b*x i and b*x r +a*x i In the specific implementation of step (2), considering that the target domain data has a different distribution from the source domain data in the partial domain adaptive scenario, the source domain signal samples are noise-added to expand the robustness of the model;

[0067] Specifically, the noise-adding processing of the source domain signal samples is represented by the following generator function:

[0068] G(x s ,z,θG )

[0069] The generator function receives source domain data and noise vectors z ~ Z sampled from a uniform distribution, with the aim of learning a mapping {x}. s ,z}→x f , where x f The data is trained to be similar to the target domain data.

[0070] The data emitted by the radiation source is assumed to be an I / Q modulated signal, consisting of two components: an I-carrier and a Q-carrier, representing the in-phase and quadrature components, respectively. The Q-carrier is obtained by rotating the I-carrier by 90 degrees.

[0071]

[0072]

[0073] Therefore, the two-channel I / Q signals emitted by the radiation source are converted into one-dimensional complex signals, and the input I / Q signals are expressed as x = x r +ix i Where x r The I signal represents the real part of a complex signal sample, x i The Q signal represents the imaginary part of a complex signal sample.

[0074] In the specific implementation of step (3), a softmax-based weighted domain classifier, combined with a similarity measurement method, is used to weight the source domain data:

[0075] Specifically, the weighted domain classifier module mainly consists of two steps: similarity calculation between the source domain data and the target domain data, and weighted processing. The two steps are as follows:

[0076] (3.1) In the actual training process, source domain signal samples are used as training data, and target domain signal samples are used as test data. Similarity metrics can characterize the differences between source and target domain signal samples, thereby further ensuring that the categories of samples in the source domain match the categories of samples in the target domain. Therefore, the similarity between source and target domain data can be expressed as:

[0077] D ω (f s,t (x))=p(y=1|f s,t (x))=σ(σ(f s,t (x)))

[0078] Among them, f s,t (x) represents the representative features of the source and target domain samples extracted by the feature extractor. σ is the sigmoid activation function, and y is the label corresponding to the input feature.

[0079] (3.2) Assuming the probability that a classification feature is labeled as the source domain is "1", then In other words, the probability of features from outlier categories in the source domain is close to "1", while the probability of shared categories between the two domains is close to "0". Outlier categories in the source domain may mismatch with samples in the target domain, leading to domain shift. Therefore, outlier features should be assigned smaller weights to reduce the impact of domain shift on model training. During the weighting stage, the feature weighting function should be... Inversely proportional, the entire weighting process can be expressed as:

[0080]

[0081] Based on traditional adversarial domain adaptation (DA), given f s (x) (corresponding to the source domain sample in DA), for any f t (x) (corresponding to the target domain sample in DA) yields the optimal result. It can be represented as:

[0082]

[0083] Where, p s () represents the data distribution of the source domain samples, p t () represents the data distribution of the target domain samples.

[0084] In the specific implementation of step (4), a domain discriminator and a label predictor are introduced, and the cross-entropy loss in the domain discriminator is replaced with the Wasserstein distance loss combined with gradient penalty, thus overcoming gradient vanishing.

[0085] Specifically, the original Generative Adversarial Network (GAN) optimizes the model by calculating the Kullback-Leibler (KL) and Jensen-Shannon (JS) divergences between the generator distribution and the actual data distribution. However, when the actual data distribution and the generator distribution do not overlap, the KL and JS divergences are constant, and gradient vanishing occurs. Furthermore, it is difficult to balance the training between the discriminator and the generator, which can lead to training instability. Unlike the KL and JS divergences, the Wasserstein distance provides a smooth distance metric even when there is no overlap between the two distributions. Moreover, using the Wasserstein distance allows us to ignore the training balance between the generator and the discriminator. In the partial domain adaptation scenario for radiation source individual identification, the optimization objectives of the feature extractor G and the domain discriminator D can be expressed as:

[0086]

[0087] Among them, l wd(D,G) represents the loss function between the domain discriminator and the generator. Let it be the expected function that the samples in the target domain follow. Let λ be the expectation function that the source domain samples follow, and let λ be a hyperparameter that controls the balance between the input signal and the gradient penalty. For the gradient operator, ||..||2 is the 2-norm. Represented as the distribution X from the source domain s and feature extractor distribution X g Data points sampled in the middle:

[0088]

[0089] γ is a hyperparameter that balances the source domain samples and the samples after adding noise, and U[0,1] indicates that the data follows a uniform distribution.

[0090] After adversarial training, the distributions of the source and target domain data are aligned. The target domain data is then classified using a label predictor, and the cross-entropy loss of the label predictor can be expressed as:

[0091]

[0092] Among them, y s It is the encoding x s The one-hot vector of the label. Finally, the overall optimization objective of the model can be expressed as:

[0093]

[0094] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0095] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for individual identification of radiation sources based on a partially adversarial domain using complex convolution, characterized in that, Includes the following steps: (1) The two-channel I / Q signals emitted by the radiation source collected in the scene are converted into complex signals in the form of a one-dimensional sequence and directly input into the feature extractor composed of a complex convolutional neural network for feature extraction; (2) Add noise to the source domain signal samples to expand the robustness of the model; (3) Based on the traditional domain adversarial training network, a weighted domain classifier is introduced and combined with the similarity measurement method to filter out the sample categories that do not match the target domain in the source domain, thereby reducing the impact of domain shift caused by outlier samples in the source domain. In step (3), the weighted domain classifier is a softmax-based weighted domain classifier. Combined with a similarity measurement method, the source domain data is weighted. The similarity between the source domain signal sample and the target domain signal sample is defined as: in, This represents the representative features of the source and target domain samples extracted by the feature extractor. It is the sigmoid activation function, and y is the label corresponding to the input feature; assuming that the probability of a categorical feature being labeled as the source domain is "1", then... That is, the probability of features from outlier categories in the source domain is close to "1", while the probability of shared categories between the two domains is close to "0"; outlier features should be assigned smaller weight values ​​to reduce the impact of domain shift; therefore, the feature weighting function and Inversely proportional, the entire weighting process can be expressed as: Based on traditional DA, given For any source domain sample in DA, The optimal sample corresponds to the target domain sample in DA. Represented as: in, This represents the data distribution of the source domain samples. This represents the data distribution of the target domain samples; (4) Construct a domain discriminator and a label predictor, introduce Wasserstein distance to guide adversarial unsupervised domain adaptation, align the data distribution in the source domain with that in the target domain, and obtain the label corresponding to the target domain sample on the label predictor; (5) A radiation source identification model is constructed using a complex neural network. Source domain samples and target domain samples are used for model training, and target domain samples are used for model testing. (6) Use the radiation source identification model trained in step (5) to identify radiation source devices in the partial domain adaptive scenario.

2. The method for individual identification of radiation sources based on partially adversarial domain using complex convolution as described in claim 1, characterized in that, The I / Q signal emitted by the radiation source in step (1) consists of an I carrier and a Q carrier, wherein the Q carrier is obtained by rotating the I carrier by 90 degrees.

3. The method for individual identification of radiation sources based on partially adversarial domain using complex convolution as described in claim 2, characterized in that, The I-carrier corresponds to the real part of the complex signal sample, and the Q-carrier corresponds to the imaginary part of the complex signal sample.

4. The method for individual identification of radiation sources based on partially adversarial domain using complex convolution as described in claim 1, characterized in that, In step (1), the feature extractor constructed from the complex convolutional network uses a one-dimensional complex convolutional kernel for feature extraction; let the complex samples of the input IQ signal be represented as... Where i represents the imaginary unit, The I signal represents the real part of a complex signal sample. The Q signal represents the imaginary part of a complex signal sample, and the weights of the complex convolution kernel are expressed as... Where a represents the real part of the weight and b represents the imaginary part of the weight, the convolution operation between the complex IQ signal and the complex convolution kernel is as follows: Where * represents the convolution operation, and the real and imaginary parts output after complex convolution are respectively and .

5. The method for individual identification of radiation sources based on partially adversarial domain according to claim 1, characterized in that, In step (2), the source domain signal samples are noise-added, and the generating function is expressed as: The generator function receives source domain data and a noise vector sampled from a uniform distribution. The purpose is to learn a mapping ,in The data is trained to be similar to the target domain data.

6. The method for individual identification of radiation sources based on partially adversarial domain according to claim 1, characterized in that, The specific steps of step (4) are as follows: To address the training instability caused by the non-overlapping distributions of source and target domain samples, the cross-entropy loss in the domain discriminator is replaced with a Wasserstein distance loss incorporating gradient penalties, thus overcoming gradient vanishing. The optimization objectives of the feature extractor G and the domain discriminator D are expressed as follows: in, This is expressed as the loss function between the domain discriminator and the generator. Let it be the expected function that the samples in the target domain follow. Let it be the expected function that the samples in the source domain obey. Hyperparameters used to control the balance between the input signal and the gradient penalty For gradient operators, It is a 2-norm. Represented as distribution from the source domain and feature extractor distribution Data points sampled in the middle: To adjust the hyperparameters that balance the source domain samples and the noisy samples, This indicates that the data follows a uniform distribution.

7. The method for individual identification of radiation sources based on a partially adversarial domain using complex convolution as described in claim 6, characterized in that, The cross-entropy loss of the label predictor c in step (4) is expressed as: in, It is encoding The one-hot vector of the label; finally, the overall optimization objective of the model is expressed as: 。

Images