A method for cross-domain emitter individual identification based on multi-transform domain feature fusion
By using a deep neural network model that integrates features from multiple transform domains and trains in stages, the problems of low feature utilization and poor noise resistance in cross-domain radiation source signal identification are solved, achieving efficient and accurate identification across different domains.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- UNIV OF ELECTRONICS SCI & TECH OF CHINA
- Filing Date
- 2026-02-12
- Publication Date
- 2026-05-29
AI Technical Summary
Existing radiation source signal identification methods have low feature information utilization and poor noise resistance in cross-domain identification, resulting in low identification accuracy. Furthermore, traditional deep learning models experience performance degradation when faced with domain shifts and environmental adaptations.
A multi-transform domain feature fusion method is adopted. Through rectangular integral bispectral, fuzzy function orthogonal slicing and Hilbert marginal spectrum feature extraction, combined with deep neural networks of ResNet and MLP, a cross-domain radiation source individual identification model is constructed by pre-training and fine-tuning. The model is trained in stages to adapt to the distribution differences of different domains.
It achieves high discriminative recognition across different domains, suppresses sensitivity to changes between domains, and improves the model's recognition accuracy and generalization ability in the target domain.
Smart Images
Figure CN122115966A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of communication signal processing and pattern recognition technology, specifically involving a cross-domain radiation source individual identification technology based on multi-transform domain feature fusion. Background Technology
[0002] In the fields of non-cooperative communication reconnaissance and electronic warfare, Specific Emitter Identification (SEI) technology aims to identify individuals by extracting and recognizing the inherent hardware fingerprints of communication devices, possessing significant military and security value. However, in real electromagnetic environments, training data (source domain) and data to be identified (target domain) often exhibit significant differences in data distribution due to factors such as operating status, transmission channels, and environmental noise—a cross-domain problem that leads to a sharp decline in the performance of traditional identification models. In recent years, deep learning has improved cross-domain robustness to some extent by automatically learning signal representations, but it still faces key challenges such as domain shift and environmental adaptation, urgently requiring more powerful adaptive learning mechanisms to enhance the model's generalization ability.
[0003] Single-type features often only capture one aspect of a device's fingerprint and are extremely sensitive to domain changes. For example, time-frequency features that perform well at a specific signal-to-noise ratio (SNR) may experience a significant decrease in discriminative power as the SNR deteriorates, because the device's core hard fingerprint, such as power amplifier nonlinearity, phase noise, and I / Q imbalance, is overwhelmed by channel effects and noise. Furthermore, blindly piling up high-dimensional features not only introduces a large amount of redundant information and increases the computational burden, but also easily induces overfitting in the source domain, thereby weakening its generalization performance in the target domain. Summary of the Invention
[0004] This invention addresses the problem of low cross-domain recognition accuracy caused by low utilization of feature information and poor noise resistance in existing radiation source signal identification methods. It proposes an integrated deep neural network that integrates features from multiple transform domains of radar radiation source signals, suitable for scenarios such as non-cooperative communication reconnaissance and electronic warfare. This network constructs a more comprehensive feature representation space without blindly piling up high-dimensional features, thereby enabling accurate identification of unknown radiation source targets with different operating states and channel environments.
[0005] The technical solution adopted by this invention to solve the above-mentioned technical problems is a method for cross-domain radiation source individual identification based on multi-transform domain feature fusion, comprising:
[0006] Step 1: Obtain the signal of the radiation source to be identified;
[0007] Step 2: Perform multi-transform domain feature extraction on the radiation source signal to obtain three two-dimensional feature images;
[0008] The multi-transform domain feature extraction includes:
[0009] Rectangular integral bispectral feature extraction: Calculate the bispectral of the radiation source signal, integrate the bispectral through multiple preset integration paths to obtain rectangular integral bispectral features and convert them into a first two-dimensional feature image;
[0010] Fuzzy function orthogonal slice feature extraction: Calculate the fuzzy function of the radiation source signal, and extract multiple orthogonal slices along the time delay axis and Doppler frequency offset axis to obtain the fuzzy function orthogonal slice features and convert them into a second two-dimensional feature image;
[0011] Hilbert marginal spectral feature extraction: Empirical mode decomposition and Hilbert transform are performed on the radiation source signal to obtain Hilbert marginal spectral features and convert them into a third two-dimensional feature image;
[0012] Step 3: Input the three types of two-dimensional feature images into a pre-trained deep neural network recognition model to obtain the individual radiation source recognition results;
[0013] The deep neural network recognition model is a fusion model based on ResNet and MLP, including three parallel ResNet base learners and one MLP meta learner. The ResNet base learners are used to extract features from the corresponding input two-dimensional feature images and output high-level feature vectors. The MLP meta learner is used to fuse the high-level feature vectors output by the three ResNet base learners and output the final recognition category.
[0014] Furthermore, the training of the deep neural network recognition model adopts a pre-training-fine-tuning paradigm, including:
[0015] Pre-training phase: Using source domain radiation source signal data, three ResNet base learners are trained respectively to learn the general feature representation of radiation source hardware fingerprint;
[0016] Fine-tuning phase: The pre-trained model is adapted using radiation source signal data from the target domain; during the fine-tuning process, the backbone network parameters of the ResNet base learners are kept fixed, and the parameters of the adaptation layer at the end of each base learner and the MLP meta learner are updated only.
[0017] In this invention, the source domain radiation source signal data and the target domain radiation source signal data exhibit distributional differences in at least one aspect: operating state, transmission channel, or environmental noise. This invention is not limited to a single transform domain, nor does it blindly accumulate high-dimensional features. Instead, it starts from multiple perspectives of the signal, simultaneously extracting various transform domain features, including higher-order cumulants and cyclic spectra, thereby constructing a more comprehensive feature representation space. First, the radiation source signal is processed based on three transform domain methods: bispectral estimation, fuzzy functions, and Hilbert-Huang Transform (HHT). Rectangular integral bispectral features, fuzzy function orthogonal slice features, and Hilbert marginal spectrum features are extracted respectively. These features are then converted into more expressive and interpretable two-dimensional feature images as primary feature representations. These features describe the device fingerprint from different dimensions, such as the signal's statistical properties and cyclic stationarity.
[0018] Subsequently, by designing an effective fusion mechanism, these heterogeneous primary two-dimensional feature images are unified into a high-level representation, aiming to form a super-feature that is highly discriminative of device identity and insensitive to inter-domain changes, thereby fundamentally enhancing the model's ability to cope with unknown domain data. By constructing a fusion recognition model framework of residual network and multilayer perceptron ResNet+MLP, multiple ResNet networks are first used as base learners of the fusion model to perform primary recognition on three transform domain feature datasets, obtaining feature vectors represented by probability in a single transform domain as high-level feature vectors. Finally, the high-level feature vectors are fused and learned through MLP, forming a high-level feature representation for decision-making within the fusion recognition model. Finally, the output layer of the MLP is used to determine the individual category of the radiation source based on this super-feature, obtaining the final signal category information.
[0019] Although the fused super-features are more discriminative, their complex structure (containing three types of heterogeneous sub-features: bispectral, fuzzy function, and HHT) means that the distribution of these three types of sub-features may shift in different forms and to different degrees when transferring from the source domain to the target domain. When training the fusion recognition model, conventional end-to-end training or simple fine-tuning makes it difficult for the model to coordinate this internal imbalanced domain shift, easily leading to the failure of adaptation in some sub-feature branches and weakening the fusion advantage. Therefore, this invention adopts a pre-training-fine-tuning paradigm based on a ResNet network that is deeply bound to the multi-transform domain feature structure. The model is trained in stages: In the pre-training stage, three ResNet base learners, each processing different transform domain features, are independently and sufficiently pre-trained using source domain data. This allows each branch to deeply mine the general representation patterns related to the radiation source hardware fingerprint in its corresponding feature modality, laying a solid and decoupled feature foundation for subsequent fusion and transfer. In the fine-tuning stage, an innovative strategy of trunk freezing and branch collaborative fine-tuning is adopted. Specifically, the ResNet backbone network parameters are kept fixed to preserve general knowledge, and fine-tuning is only performed on the adaptation layers at the end of each branch and the subsequent MLP meta-learners. This design allows the model to learn flexibly how to recalibrate and weight the output contributions of the three heterogeneous feature branches based on the target domain data with a very low number of parameters (avoiding overfitting), and dynamically adjust the fusion strategy to achieve efficient and accurate adaptation to the target domain distribution.
[0020] The beneficial effects of this invention are:
[0021] 1. Parallel extraction of multiple heterogeneous features, including high-order cumulants and cyclic spectra, embodies a novel feature construction approach; and through an effective fusion mechanism, these complementary feature information are organically unified to achieve high discriminative power for device identity while minimizing sensitivity to inter-domain variations.
[0022] 2. A pre-training-fine-tuning paradigm based on ResNet network is adopted to train the model in stages, efficiently transferring the knowledge gained in the source domain to new target environments with different distributions, and finally achieving accurate recognition in the target domain. Attached Figure Description
[0023] Figure 1 This is a framework diagram of the method of the present invention;
[0024] Figure 2 This is a diagram of a multi-transform domain feature fusion recognition model. Detailed Implementation
[0025] The method of the present invention will be further described below with reference to the accompanying drawings.
[0026] like Figure 1 As shown, the specific steps are as follows:
[0027] Step 1: Square Integral Bispectrum (SIB) feature extraction.
[0028] Bispectral processing is a signal processing method based on the double Fourier transform. Its physical definition can be approximated as the decomposition of signal skewness in the frequency domain. Specifically, bispectral calculation utilizes the double Fourier transform of third-order cumulants, where the signal skewness is represented by the third-order cumulant and the variance by the second-order cumulant. When processing radiation source signals, bispectral processing can effectively reflect the nonlinearity of the signal. The expression for bispectral processing is as follows:
[0029]
[0030] in, For the radiation source signal s at the dual-frequency point bispectral values, () represents the coordinate axis point corresponding to the bispectral plane. Let be the third-order cumulative quantity of signal s. All are time delay variables. The imaginary unit, It is a kernel function for two-dimensional Fourier transform, used to map the third-order cumulant in the time domain to the bispectrum in the frequency domain.
[0031] Bispectral estimation effectively preserves the amplitude and phase information of a signal and significantly reduces the adverse effects of Gaussian and non-Gaussian noise. Among all types of higher-order spectra, the bispectral spectrum is the lowest order. Compared to other higher-order spectra, bispectral processing methods are simpler, and its time-invariance, scale-invariance, and phase-preserving properties allow for the analysis of subtle differences between radiation source signals by examining the interactions of signals at different frequencies.
[0032] Due to the properties of bispectral estimation, it has significant advantages in feature extraction of radiation source signals. However, directly applying bispectral estimation for feature extraction requires overcoming its computational complexity. Integral bispectral estimation, a secondary representation technique of bispectral estimation, can overcome the computational complexity problem to some extent. The rectangular integral bispectral method is chosen as the secondary representation of bispectral estimation for radiation source signal feature extraction. The specific operation is as follows:
[0033] (1) Assume that Class of unidentified radiation source signals, each class of signals has The sample number is t, where the sample number at time t is t. The first type of signal Each sample signal is represented as .
[0034] (2) For each sample signal Calculate its bispectral ).
[0035] (3) Calculate the first SIB extraction based on the SIB integral path. The first type of signal The sample signal at the th th Integration path Bispectral features :
[0036] ;
[0037] In the formula, i is the index of the sample signal. Here, L represents the integration path index, and L is the total number of integration paths in the SIB. The first adopted for SIB There are several integration paths.
[0038] Step 2: Fuzzy Function (AF) Orthogonal Slice Feature Extraction.
[0039] The ambiguity function depends entirely on the signal waveform emitted by the radiation source. The ambiguity function is unique for each radiation source signal, and the allocation of the signal ambiguity volume is entirely determined by the signal itself. The ambiguity function extracts the variation characteristics of signals with different time delays and frequency shifts by convolving them, thereby distinguishing different individual radiation sources.
[0040] Radiation source signal fuzzy function It can be defined as:
[0041] ;
[0042] In the formula, The delay represents the time lag between the signal copy and the original signal. Doppler frequency offset represents the frequency shift of the signal replica relative to the original signal. For time, for The conjugate function of the signal; The fuzzy function is determined by its complex envelope. The decision was made.
[0043] The slices with zero frequency offset and time delay of the ambiguity function are extracted as the main characterization of the radar radiation source signal. Based on the symmetry of AF, the slices are cut and combined along the half-axis of frequency offset and half-axis of time delay of the ambiguity function to obtain orthogonal AF slices.
[0044] At that time, Yan At that time, the correlation between the signal and itself (without time delay) across all Doppler frequency offsets ξ is calculated, which reflects a slice of the signal's frequency resolution characteristics:
[0045] ;
[0046] At that time, Yan At that time, the cutting is performed perpendicular to the amplitude axis along the frequency deviation axis. For a non-zero delay, the slice can be represented by the following function:
[0047] ;
[0048] This slice is equivalent to the input signal being... The frequency-shifted echo signal passes through a matched filter and is shifted along the frequency shift axis. Expand the output value at each time step.
[0049] When frequency shift At that time, the correlation between the signal and its delayed replica at zero Doppler frequency offset is calculated, which reflects a slice of the signal's range (delay) resolution characteristics:
[0050] ;
[0051] At that time, Yan At that time, the cutting is performed perpendicular to the amplitude axis along the frequency deviation axis. For non-zero Doppler frequency offset, the slice can be represented by the following function.
[0052] ;
[0053] The AF slice represented by the above formula is equivalent to the input signal being The time-delayed echo signal passes through a matched filter and in The output value expanded along the time delay at any given moment.
[0054] Step 3: Hilbert marginal spectrum feature extraction.
[0055] The Hilbert Huang Transform (HHT) consists of two parts: Empirical Mode Decomposition (EMD) and Hilbert Spectral Analysis (HAS). EMD is a method for stabilizing non-stationary signals. Essentially, it involves locally decomposing the signal and iterating repeatedly to break down the original signal into multiple Intrinsic Mode Functions (IMFs), and then using the IMFs to further decompose the signal sequence.
[0056] When the EMD process ends, the radiation source signal s(t) is decomposed into n intrinsic mode functions (IMFs) and a residual term:
[0057]
[0058] The number of intrinsic modes, One IMF, For intrinsic mode function (IMF) index, This is the final residual component;
[0059] For the signal in the above formula After performing the Hilbert transform, we can obtain :
[0060] ;
[0061] s(t) and m(t) form a complex conjugate pair, then the analytic signal of s(t) is... Represented as: ; It is a natural exponential function;
[0062] In the formula, a(t) is the instantaneous amplitude function. Let be the instantaneous phase function. Defined as:
[0063] ;
[0064] ;
[0065] The instantaneous frequency can be obtained from the derivative of the instantaneous phase. Definition:
[0066] ;
[0067] For each eigenmode function obtained from the above empirical mode decomposition By performing the Hilbert transform, the corresponding analytic signal can be obtained. ;
[0068] and These represent the instantaneous amplitude and instantaneous phase of the k-th IMF, and the instantaneous frequency of the k-th IMF, respectively. ;
[0069] The radiation source signal s(t) can be expressed as the sum of a series of analytic signal components whose frequency and amplitude both vary with time:
[0070] ;
[0071] This indicates taking the real part.
[0072] Hilbert spectrum Defined as integrating over time the joint representation of the signal's energy distribution in the time-frequency plane, it integrates the instantaneous amplitude of each component k. and instantaneous frequency Mapped onto the time-frequency plane, it is represented as:
[0073]
[0074] The Hilbert spectrum reflects the evolution of signal energy over time and frequency. (Hilbert marginal spectrum) for:
[0075]
[0076] Where T is the total duration of the signal, and the Hilbert marginal spectrum represents the total amplitude value corresponding to each frequency. Statistically, it reflects the energy accumulation distribution of the signal at various frequencies over the entire time span, and more accurately reflects the frequency components of the actual signal. Therefore, the Hilbert marginal spectrum is more suitable for processing radiation source signals with non-stationary characteristics.
[0077] Step 4: Building a Radiation Source Signal Identification Model
[0078] After multi-transform domain processing, the radiation source signal was processed to extract rectangular integral bispectral features, fuzzy function orthogonal slice features, and Hilbert marginal spectrum features, which were then transformed into a two-dimensional image. Given the unique advantages of deep neural networks in image processing, a ResNet+MLP network model was constructed to classify and recognize the two-dimensional feature images obtained by the above methods. The multi-transform domain feature fusion recognition model is shown in the figure below. Figure 2 As shown.
[0079] The ResNet network primarily serves as the base learner for the recognition model, achieving initial recognition of radiation source signals through a multi-transform domain feature dataset and obtaining feature vectors represented in probabilistic form. Then, an MLP is used as a meta-learner to fuse the feature vectors, improving recognition accuracy and obtaining the final recognition result.
[0080] The pre-training phase takes place in the source domain. We choose ResNet-50 as the base network architecture. During this phase, all layers of the network, from the initial convolutional layers responsible for extracting basic patterns to the final fully connected layers performing classification, have parameters that are learnable. Using all labeled data from the source domain, the network is trained end-to-end with the objective of classifying radiation source individuals. The goal of this process is to train the model's lower-layer networks into general feature extractors, such as learning to identify basic structures related to hardware fingerprints, like edges and textures, from the time-frequency plot of a signal, while simultaneously teaching the higher-layer networks how to combine these basic features to achieve accurate individual discrimination.
[0081] After the pre-trained model converges and obtains a good set of initial parameters, we enter the crucial fine-tuning stage. To adapt it to the target domain, we first strategically modify the network structure. Specifically, we retain the core feature extraction part of the pre-trained ResNet model, i.e., all layers except the last fully connected layer, and freeze the parameters of these layers to keep them unchanged during subsequent fine-tuning training. This is to preserve the general feature knowledge that the model has learned and prevent it from being corrupted or forgotten by a small amount of target domain data. At the same time, we replace the original classifier at the end of the network, i.e., the last fully connected layer, modifying its output dimension to the number of classes of individuals in the target domain radiation source, and randomly initialize this new layer.
[0082] Subsequently, fine-tuning training is performed using a small amount of labeled data from the target domain. During this process, the gradients calculated by the backpropagation algorithm cease updating the parameters of the frozen layers, and the model's optimization focuses solely on the newly introduced, randomly initialized final classification layer. This means that the model relies on its powerful, general feature extraction capabilities learned in the source domain to generate effective feature representations for target domain samples, and then quickly adjusts the final classifier layer to learn how to map these features to specific categories in the target domain. In this way, the model can efficiently transfer and adapt its knowledge acquired in the source domain to new, differently distributed target environments while avoiding overfitting to a small amount of target domain data, ultimately achieving accurate recognition in the target domain.
Claims
1. A method for cross-domain radiation source individual identification based on multi-transform domain feature fusion, characterized in that, Includes the following steps: Step 1: Obtain the signal of the radiation source to be identified; Step 2: Perform multi-transform domain feature extraction on the radiation source signal to obtain three two-dimensional feature images; The multi-transform domain feature extraction includes: Rectangular integral bispectral feature extraction: Calculate the bispectral of the radiation source signal, integrate the bispectral through multiple preset integration paths to obtain rectangular integral bispectral features and convert them into a first two-dimensional feature image; Fuzzy function orthogonal slice feature extraction: Calculate the fuzzy function of the radiation source signal, and extract multiple orthogonal slices along the time delay axis and Doppler frequency offset axis to obtain the fuzzy function orthogonal slice features and convert them into a second two-dimensional feature image; Hilbert marginal spectral feature extraction: Empirical mode decomposition and Hilbert transform are performed on the radiation source signal to obtain Hilbert marginal spectral features and convert them into a third two-dimensional feature image; Step 3: Input the three types of two-dimensional feature images into a pre-trained deep neural network recognition model to obtain the individual radiation source recognition results; The deep neural network recognition model is a fusion model based on ResNet and MLP, including three parallel ResNet base learners and one MLP meta learner. The ResNet base learners are used to extract features from the corresponding input two-dimensional feature images and output high-level feature vectors. The MLP meta learner is used to fuse the high-level feature vectors output by the three ResNet base learners and output the final recognition category.
2. The method as described in claim 1, characterized in that, The training of the deep neural network recognition model adopts a pre-training-fine-tuning paradigm, including: Pre-training phase: Using source domain radiation source signal data, three ResNet base learners are trained respectively to learn the general feature representation of radiation source hardware fingerprint; Fine-tuning phase: The pre-trained model is adapted using radiation source signal data from the target domain; during the fine-tuning process, the backbone network parameters of the ResNet base learners are kept fixed, and the parameters of the adaptation layer at the end of each base learner and the MLP meta learner are updated only. Among them, the source domain radiation source signal data and the target domain radiation source signal data have distribution differences in at least one aspect, such as operating status, transmission channel, or environmental noise.
3. The method as described in claim 2, characterized in that, The fine-tuning stage specifically includes: modifying the output dimension of the last fully connected classification layer of each ResNet base learner to be consistent with the number of target domain categories, and training the fully connected layer and the MLP meta-learner using the target domain data.
4. The method as described in claim 1, characterized in that, The rectangular integral bispectral feature extraction specifically includes: For time t, the... The first type of signal The signal from each radiation source is represented as Calculate its bispectral ), ( ) represents the coordinate axis point corresponding to the bispectral plane; calculate the integral bispectral features on each path through L preset integration paths. : ； Where i is the index of the sample signal. Here, L represents the integration path index, and L is the total number of integration paths in the SIB. For the first One integration path; All The resulting feature vectors are converted into a first two-dimensional feature image.
5. The method as described in claim 1, characterized in that, The orthogonal slice feature extraction of the fuzzy function specifically includes: Calculate the radiation source signal at time t fuzzy function ,in, For time delay, This is due to Doppler frequency shift; Extract when slices and slices at time ; Extract at least one slice corresponding to a non-zero delay and at least one non-zero Doppler frequency offset slices ; All extracted slices are combined into the orthogonal slice features of the fuzzy function and converted into a second two-dimensional feature image.
6. The method as described in claim 1, characterized in that, The Hilbert marginal spectrum feature extraction specifically includes: Empirical mode decomposition (EMD) is performed on the radiation source signal s(t) at time t to obtain n eigenmode functions. and a residual term Intrinsic mode function index : Perform a Hilbert transform on each intrinsic mode function to obtain the corresponding instantaneous amplitude. and instantaneous frequency ; Calculate the Hilbert spectrum: ;in, Represents the imaginary unit. It is a natural exponential function; Integrating the Hilbert spectrum over time yields the Hilbert marginal spectrum. Where T is the total duration of the signal; The Hilbert marginal spectrum is converted into a third two-dimensional feature image.