A method for extracting individual fingerprint features of radiation sources

Abstract

The present invention discloses a method for extracting individual fingerprint features of radiation sources. Based on a recursive invariant risk minimization framework, the method takes pulse sample signals of various types of radiation sources that have undergone time-frequency feature extraction as input, eliminates random modulation samples of each type of radiation source, and constructs a data set with distribution offset between training samples and test samples. The training samples will input each radiation source sample into a deep neural network for training in a certain order. The framework will save the model obtained by training each radiation source sample for the training of the next radiation source sample. The pre-trained model used in each training is the generation model of the previous radiation source sample until all radiation source samples are involved in the generation of the model. This framework can guide the neural network to learn more subtle individual fingerprint features that remove the differences in pulse signal modulation methods. The present invention is of great significance to the generalization and application prospects of radiation source individual identification technology.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing, and in particular relates to a method for extracting individual fingerprint features of a radiation source. Background Art

[0002] The modern information-based battlefield is populated by a wide variety of radio and radar equipment, responsible for numerous tasks, including command and control, communications, intelligence reconnaissance, and electronic surveillance. To achieve superiority, the primary requirement is to acquire and control battlefield information, strategically monitor and conduct comprehensive, timely, and accurate jamming, counter-jamming, and attack operations against the enemy's critical electronic equipment and carriers. However, with the rapid advancement of signal processing technology, signal modulation methods are becoming increasingly diverse, along with the increasing number of high-power devices, and the electromagnetic environment is becoming increasingly complex. Traditional communication modulation identification methods, which extract simple characteristics of signal modulation parameters such as carrier frequency, bandwidth, and symbol rate, are no longer sufficient to identify individual emitters and cannot meet the demands of the modern battlefield.

[0003] Radio frequency fingerprints, a characteristic caused by radio circuits, have been widely used in the field of individual identification of radio sources in recent years. Fingerprint recognition is the process of identifying wireless devices by extracting unintentional modulation characteristics caused by hardware defects in analog circuits. These characteristics are universally present in all wireless devices, and no two devices have the same fingerprint. Therefore, they have great research and application value in the field of individual identification.

[0004] Existing methods for identifying individual emitters have limited research addressing the generalization of identification samples. Due to the specific nature of the research field, available emitter data for identification is scarce, and most research is based on pure simulation or semi-physical simulation data. This type of data satisfies the In-Distribution (ID) condition when acquired. This can easily lead to overfitting of deep models and difficulty generalizing the models to other datasets. In practical applications, the data obtained after model deployment may be Out-of-Distribution (OOD), or distribution-shifted samples. To improve counter-reconnaissance capabilities, battlefield emitters frequently change their modulation mode and parameters for the same emitter. These samples are distribution-shifted compared to the samples used to train the deployed model. Existing methods struggle to achieve good generalization on these samples, resulting in a significant decrease in the rate of individual emitter identification.

[0005] "A Method for Individual Identification of Target Emitters" (CN201810728417.8) and "A Method and System for Individual Identification of Radar Emitters" (CN202110559123.9), respectively, extract time-frequency features and fuzzy function features from the emitter signal. These methods achieve good recognition accuracy on both training and test sets. However, these methods are not tested on samples with distribution shift, and their recognition rates plummet in such samples. "System and Method for Identifying Radio Emitters Based on Weak Fingerprint Features" (CN202211020314.9) also extracts weak fingerprint features, but attempts to interpret subtle fingerprint features based on traditional manual empirical feature extraction methods. This method can only extract fingerprint features with known causes and has limited ability to extract fingerprint features with unknown causes and links. When fingerprint features with unknown causes become dominant, these fingerprint feature extraction methods have poor generalization capabilities, and the accuracy of individual emitter identification decreases. Summary of the Invention

[0006] The present invention aims to accurately extract the subtle fingerprint features of individual radiation sources to improve the generalization of radiation source individual identification classifiers between samples with different distributions.

[0007] To this end, the purpose of the present invention is to propose a method for extracting fingerprint features of individual radiation sources. In view of the particularity of radiation source samples, a recursive invariant risk minimization framework is adopted to divide samples with different distributions for training and testing respectively, and a recursive invariant risk minimization loss function is set to guide the deep neural network model to learn fingerprint features. A recursive method is also used to retain the model trained for each type of radiation source sample and continue training for the next type of radiation source sample. The model trained under this framework can not only learn the invariant fingerprint features in a single radiation source sample, but also accurately distinguish the differences in fingerprint features between different radiation source individuals. Even if new unknown modulation modes appear in the test set in the future, their individual radiation sources can also be correctly classified.

[0008] To achieve the above-mentioned purpose, the present invention proposes a method for extracting individual fingerprint features of radiation sources, including: obtaining samples through semi-physical simulation of radiation source signals; extracting time-frequency features of the radiation source signal samples, and transforming the original signals into feature matrices or feature vectors, which are easier to input into deep neural network model training; dividing the radiation source samples into distribution offset samples, with one part of the samples used for model training and the other part of the samples used for model generalization ability testing; inputting the samples used for training the model into the deep neural network, saving the model after each training, and continuing to use the next type of radiation source samples recursively, repeating multiple times to obtain a radiation source individual fingerprint feature extraction model. Use the trained model to classify and identify the samples used for testing. The specific plan is as follows:

[0009] A method for extracting individual radio frequency fingerprint features comprises: obtaining pulse signal samples from M radiation sources of the same model, wherein each radiation source sample should contain N types of intra-pulse modulation modes, and samples from different radiation sources should have the same modulation mode and parameters; first, extracting time-frequency features from the radiation source samples, and then inputting them into a recursive invariant risk minimization framework (RIRM); within the RIRM framework, randomly removing samples with different modulation modes from the radiation source samples to obtain training set data and test data with distribution shift; then, sequentially using the sample data from one radiation source at a time for training a neural network model, guiding the model to learn subtle fingerprint features of the radiation source by setting a recursive invariant risk minimization loss function in the model; saving the current model after training each radiation source sample, and recursively training the next radiation source sample; completing model generation when all radiation source samples in the framework have participated in the recursion, wherein the generated model has the ability to extract individual fingerprint features of the radiation source and the ability to identify different individual radiation sources; and recombining the modulation mode data of each radiation source randomly removed during sample input into validation set data for testing the fingerprint feature extraction capability of the model.

[0010] As a further improvement of the present invention, a recursive invariant risk minimization framework RIRM for extracting individual fingerprint features of radiation sources is constructed; the operating logic of the framework is: the first type of radiation source samples in the input samples are randomly eliminated by eliminating one of their modulation modes, and at this time the input samples only contain N-1 modulation modes; the radiation source samples of this type are input into the deep neural network for training, and the recursive invariant risk minimization loss is added to the deep neural network model used for training; after the training of the sample data of the first type of radiation source is completed, the training model is retained; the same sample elimination method is then used to randomly eliminate the second type of radiation source samples in the input samples, and the training is continued using the previously saved training model; the above process is recursively repeated M times in total until the samples of all M radiation sources have participated in multiple recursive generations of the model, the feature extraction process is completed, and the final saved model has a strong generalization ability for extracting individual fingerprint features of radiation sources; the sample data of the single modulation mode eliminated each time are combined into a test set data with a sample size of M to test the model performance.

[0011] As a further improvement of the present invention, before model training, multiple time-frequency feature analysis methods were used to extract two types of features of two different dimensions, signal images and sequences, and feature encoding was completed separately; the image features were extracted using three common time-frequency analysis methods, short-time Fourier transform STFT, continuous wavelet transform CWT, and Hilbert-Huang transform HHT; during feature encoding, the extracted time-frequency graph was first grayscale processed into a two-dimensional grayscale matrix, and the size of the grayscale matrix was unified into an m*m-dimensional feature matrix through bilinear interpolation; sequence features were extracted using discrete wavelet transform DWT and empirical mode decomposition EMD; the wavelet coefficients of the detail component with high-frequency information obtained by DWT, the first-level intrinsic modal component IMF obtained by EMD, and the amplitude sequence of the original signal were selected, and the difference was an equal-length sequence of length n as a multi-channel one-dimensional feature.

[0012] As a further improvement of the present invention, the two extracted features are input into two deep neural networks respectively; the three-channel two-dimensional matrix features are input into the two-dimensional residual network Resnet for training; the three-channel one-dimensional vector features are input into the one-dimensional convolutional neural network for training; the two are trained simultaneously, and the feature vectors obtained by training are combined before the softmax layer to obtain the final classification result;

[0013] The RIRM loss function is added to the deep neural network to solve the generalization problem of the classification model under unknown radiation source modulation mode. The RIRM loss is used to solve the problem that there is still a potential causal relationship between the training and test data even though they belong to different environments or distributions. The potential causal relationship here is the fingerprint feature that needs to be extracted. The expression of RIRM is as follows:

[0014]

[0015] In the loss function, Φ is the data representation, represents the classifier, e represents the current environment, ε tr is the training data, d represents the distance between the current classifier and the classifier of the previous environment. When the training data in a new environment is used for training each time, the smaller the distance between the two classifiers, the higher the reliability of the extracted features. Can induce a predictor that is invariant across environments When there is a Optimal in all environments simultaneously w,Φ are both optimization targets, That is, the optimal classifier in this environment; simplifying the above conditions to a single-variable optimization problem is the RIRM loss function; it consists of two terms, namely the empirical risk minimum term and the invariant risk minimum term, and λ is a hyperparameter that balances the two terms; when λ→∞, the classifier has the optimal solution, and the loss function only uses invariant features.

[0016] A method for extracting fingerprint features from individual radiation sources, as described in embodiments of the present invention, accurately extracts the fingerprint features of individual radiation sources. This method distinguishes different radiation sources by extracting the unique fingerprint features of each source. In real battlefield applications, this method can accurately identify a target radiation source even if it changes its operating mode.

[0017] The beneficial effects of the present invention are:

[0018] 1) A recursive invariant risk minimization framework is proposed, which divides the training and test data sets according to the characteristics of the radiation source samples. During each training, the model continuously adjusts the model weights according to the recursive invariant risk minimization loss to learn the subtle fingerprint characteristics of the radiation source itself. Through multiple recursions, it considers whether the learned fingerprint characteristics can be used to distinguish different categories of radiation source individuals. Finally, the divided test set is used to test the model performance.

[0019] 2) The fingerprint characteristics of individual emitters are unintentional modulation features, belonging to the non-stationary signal component, and are present in the high-frequency portion of the emitter signal. Time-frequency analysis of the emitter signal is the most effective method for extracting the characteristics of non-stationary signal components. Both discrete wavelet coefficients and intrinsic modal components contain both low-frequency and high-frequency components of the signal. To reduce the input feature dimensionality while retaining the high-frequency components that may contain fingerprint characteristics, the high-frequency components of both coefficients are selected for feature combination.

[0020] Additional aspects and advantages of the present invention will be set forth in part in the following description of the drawings and will become apparent in part from the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of the method for extracting individual fingerprint features of radiation sources described in the present invention.

[0022] Figure 2 This is the recursive invariant risk minimization framework diagram proposed by the present invention.

[0023] Figure 3 This is a flow chart of the time-frequency feature extraction described in the present invention. DETAILED DESCRIPTION

[0024] The present invention will be further explained below with reference to the accompanying drawings. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0025] like Figure 1As shown, the method for extracting individual fingerprint features of radiation sources described in the present invention can be roughly divided into semi-physical simulation of radiation source signal samples, time-frequency feature extraction, distribution offset sample division, deep neural network model training, and model generalization ability test.

[0026] The dataset was acquired using hardware-in-the-loop simulations using seven identical software-defined radio transmitters. Each transmitter was configured with multiple identical modulation schemes and parameters. A software-defined radio receiver was then used to collect data.

[0027] like Figure 3 As shown in the figure, time-frequency feature extraction uses a variety of time-frequency analysis methods to analyze the radiation source pulse signal samples. Among them, short-time Fourier transform (STFT), CWT (continuous wavelet transform), and Hilbert-Huang (HHT) methods can be used to extract time-frequency map features. To facilitate training using a three-channel two-dimensional ResNet model, the time-frequency map features are grayscaled and their dimensions are unified using bilinear interpolation. Discrete wavelet transform (DWT) and empirical mode decomposition (EMD) are then used to extract sequence features at different frequency components of the signal. Discrete wavelet coefficients with high high-frequency content, intrinsic modal components, and the signal amplitude sequence are used as input to the three-channel one-dimensional convolutional neural network.

[0028] The sample partitioning process randomly removes data from one pulse modulation method from each radiation source sample, uses the remaining data as training samples, and divides the training set into test sets in an 8:2 ratio. The extracted data is used as the distribution shift sample, i.e., the validation set for the training model.

[0029] After dividing the training, test, and validation sets, the model begins training. During each training phase, the two-dimensional residual network (ResNet) and the one-dimensional convolutional neural network (CNN) are trained in parallel. The resulting feature matrices are combined before the softmax layer, after which the classification results are output. After training for a single radiation source sample, the sample is re-divided and the model continues training. After training for all radiation source samples, the final trained model is obtained.

[0030] The validation set is composed of all the data from a specific type of radiation source that was excluded during each sample partitioning process and included in a specific modulation scheme. This data is then fed into the trained model. The test results are evaluated using methods such as the accuracy of radiation source recognition, confusion matrices, and test loss curves. The accuracy of the recognition results can be used to determine whether the model has learned the subtle fingerprint differences between individual radiation sources.

[0031] This invention discloses a method for extracting individual fingerprint features from radiation sources. Based on a recursive invariant risk minimization framework, it takes pulse sample signals from various radiation sources, after time-frequency feature extraction, as input. Random modulation samples from each radiation source type are eliminated to construct a dataset with distributional shifts between training and test samples. The training samples are fed into a deep neural network in a specific order. The framework saves the model trained for each radiation source sample and uses it to train the next radiation source sample. The pre-trained model used in each training session is the generative model for the previous radiation source sample, and this process continues until all radiation source samples have participated in model generation. This framework guides the neural network to learn more subtle individual fingerprint features beyond those that differ in pulse signal modulation. The proposed radiation source fingerprint feature extraction method can capture subtle, unchanging fingerprint features of the radiation source in real-world scenarios, even when the radiation source target implements counter-reconnaissance strategies, resulting in unknown modulation modes in the signal to be identified. This method can accurately identify the specific ID number of the radiation source. This has significant implications for the generalization and application prospects of individual radiation source identification technology.

[0032] The technical means disclosed in the solutions of the present invention are not limited to those disclosed in the above-mentioned embodiments, but also include technical solutions composed of any combination of the above-mentioned technical features. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for extracting radio frequency individual fingerprint features, characterized in that: The specific method is as follows: Obtain pulse signal samples from M radiation sources of the same model. Each radiation source sample should contain N types of intra-pulse modulation modes, and samples from different radiation sources should satisfy the same modulation mode and parameters. First, extract the time-frequency features of the radiation source samples and then input them into the recursive invariant risk minimization framework (RIRM). Under the RIRM framework, samples of different radiation source samples will first be randomly modulated and eliminated to obtain training set data and test data with distribution offset; then, in sequence, the sample data of one radiation source each time will be used for neural network model training, and the model will be guided to learn the subtle fingerprint features of the radiation source by setting a recursive invariant risk minimization loss function in the model; after the training of each radiation source sample is completed, the current model will be saved and the training of the next radiation source sample will be recursively participated in; when all radiation source samples in the framework have participated in the recursion, the model generation is completed, and the generated model has the ability to extract the fingerprint features of individual radiation sources and the ability to identify different radiation sources; the modulation mode data of each radiation source that is randomly eliminated when the sample is input will be recombined into the verification set data to test the fingerprint feature extraction ability of the model; a method for radiation source is constructed. A recursive invariant risk minimization framework (RIRM) for extracting individual fingerprint features. The operating logic of this framework is as follows: samples of the first type of radiation source in the input sample are randomly eliminated by removing one of its modulation modes. At this time, the input sample only contains N-1 modulation modes. Samples of this type of radiation source are input into a deep neural network for training, and a recursive invariant risk minimization loss is added to the deep neural network model used for training. After the training of the sample data of the first type of radiation source is completed, the training model is retained. The same sample elimination method is then used to randomly eliminate samples of the second type of radiation source in the input sample, and the training is continued using the previously saved training model. This process is recursively repeated M times in total until samples of all M radiation sources have participated in multiple recursive generation of the model. The feature extraction process ends, and the final saved model has a strong generalization capability for extracting individual fingerprint features of radiation sources. The sample data of a single modulation mode eliminated each time are combined into a test set data with a sample size of M to test the model performance.

2. The method for extracting radio frequency individual fingerprint features according to claim 1, wherein: Before model training, multiple time-frequency feature analysis methods were used to extract two types of features, signal images and sequences, of different dimensions, and feature encoding was completed separately. Image features were extracted using three common time-frequency analysis methods: short-time Fourier transform (STFT), continuous wavelet transform (CWT), and Hilbert-Huang transform (HHT). During feature encoding, the extracted time-frequency images were first grayscale processed to form a two-dimensional grayscale matrix. The grayscale matrix was then uniformly resized into an m*m dimensional feature matrix using bilinear interpolation. The sequence features are extracted using discrete wavelet transform DWT and empirical mode decomposition EMD; the wavelet coefficients of the detail component with high-frequency information obtained by DWT, the first-level intrinsic modal component IMF obtained by EMD, and the amplitude sequence of the original signal are selected, and the difference between them is an equal-length sequence of length n as the one-dimensional features of the multi-channel.

3. The method for extracting radio frequency individual fingerprint features according to claim 1, wherein: The two extracted features are fed into two deep neural networks respectively. The three-channel two-dimensional matrix features are fed into the two-dimensional residual network Resnet for training; the three-channel one-dimensional vector features are fed into the one-dimensional convolutional neural network for training. Both are trained simultaneously, and the feature vectors obtained from the training are combined before the softmax layer to obtain the final classification result. The RIRM loss function is added to the deep neural network to solve the generalization problem of the classification model under unknown radiation source modulation mode. The RIRM loss is used to solve the problem that there is still a potential causal relationship between the training and test data even though they belong to different environments or distributions. The potential causal relationship here is the fingerprint feature that needs to be extracted. The expression of RIRM is as follows: In the loss function, Φ is the data representation, represents the classifier, e represents the current environment, ε tr is the training data, d represents the distance between the current classifier and the classifier of the previous environment. When training data in a new environment is used each time, the smaller the distance between the two classifiers, the higher the reliability of the extracted features; data represents Φ: This may lead to a predictor that is invariant across environments. When there is a w: Optimal in all environments simultaneously w,Φ are both optimization targets, That is, the optimal classifier in this environment; simplifying the above conditions to a single-variable optimization problem is the RIRM loss function; it consists of two terms, namely the empirical risk minimum term and the invariant risk minimum term, and λ is a hyperparameter that balances the two terms; when λ→∞, the classifier has the optimal solution, and the loss function only uses invariant features.

Images