End-to-end radiation source new individual identification method based on signal enhancement entropy regularization

Abstract

The invention discloses an end-to-end radiation source new individual identification method based on signal enhancement entropy regularization, and belongs to the field of software radio. The implementation method comprises the steps of designing an interpretable enhancement operator sequence constrained by an electromagnetic propagation partial differential equation to enhance an original signal, and adopting a prototype classifier and a network fine-tuning end-to-end architecture based on signal enhancement entropy regularization, fusing the category discrimination advantage of supervised prototype clustering and the feature extraction capability of unsupervised prototype clustering; an entropy regularization item based on signal enhancement is adopted, the entropy of a predicted value is calculated, a mean value is solved, and the mean value serves as an additional regularization item to be added into deep clustering loss. When the model encounters samples which are greatly different from known classes, entropy regularization based on signal enhancement forces the model to output more uniform probability distribution. The method is suitable for the field of software radio, and high-precision radiation source new individual identification can be realized.

Description

Technical Field

[0001] The invention relates to a new end-to-end radiation source individual identification method based on signal enhancement entropy regularization, and belongs to the field of software radio. Background Art

[0002] As a key technology in electronic reconnaissance and radio frequency signal analysis, the core goal of individual identification of radiation sources is to uniquely identify the radiation source device by capturing subtle features such as transient ripples, frequency deviation characteristics, and modulation fingerprints contained in the signal. Traditional technical systems often use feature engineering methods based on manual experience and rely on manual feature extraction strategies such as time-frequency domain analysis. Although such methods are feasible in restricted scenarios, their feature characterization capabilities face significant degradation in complex electromagnetic environments, and they have inherent defects such as insufficient generalization performance for unknown devices and limited dynamic adaptability. Therefore, improving the robustness of identification systems in complex signal environments and their adaptability to open environments has become the focus of technological evolution in this field.

[0003] In recent years, deep learning methods have achieved automated feature extraction and mapping from raw signals to device identifiers by building end-to-end learning mechanisms, providing a new paradigm for addressing the limitations of traditional methods. Existing research primarily focuses on achieving high-precision classification within a predefined device library, with some improved solutions exploring the underlying distributional characteristics of unlabeled data through semi-supervised learning or contrastive learning strategies. However, such methods are still limited by the closed set assumption and struggle to cope with the multi-task collaborative requirements of identifying known devices and dynamically discovering unknown devices in real-world scenarios. This is particularly evident in complex adversarial scenarios where new radiation sources appear concurrently.

[0004] The core challenge of current technological development stems from the lack of incremental cognitive capabilities in open environments. Existing systems often use closed-set classification frameworks or new individual binary discrimination mechanisms based on anomaly detection, lacking the ability to decouple features from multiple layers for unknown devices. When multiple models / batches of new devices emerge simultaneously, traditional methods fail to achieve the coordinated optimization of distinguishable feature learning between unknown categories and dynamic end-to-end deep clustering topology, which can easily lead to heterogeneous devices being mistakenly clustered into homogeneous clusters. Current technology faces three bottlenecks: First, the decoupling of new individual detection and clustering tasks weakens the task coupling in dynamic scenarios, making it impossible to feed recognition results back to the network for further network updates; second, if the recognition and clustering frameworks are coupled, that is, an end-to-end framework, the cluster labels of unknown classes will be biased towards known classes; third, there is no systematic solution for the coordinated optimization of decoupling features of new and old categories and dynamic end-to-end deep clustering. Therefore, establishing a joint optimization framework that integrates self-supervised representation learning, incremental and adaptive clustering topology evolution, and realizing coupled end-to-end expression and dynamic cognition of device features without prior constraints has become a key path to breaking through the bottleneck of existing technologies. Summary of the Invention

[0005] In response to the key problems of existing radiation source individual identification algorithms in open environments, such as insufficient sensitivity for new individual identification, limited accuracy for unknown category identification, and the inability of the existing identification decoupling framework to feed back identification results to the network, the purpose of the present invention is to provide an end-to-end radiation source individual identification method that integrates signal enhancement and entropy regularization. By constructing a multi-dimensional signal enhancement space and an end-to-end entropy regularization framework, while strengthening the stability of the known category feature representation, the individual identification model's ability to discover unknown radiation source individuals is significantly improved, thereby realizing the dynamic evolution and accuracy leap of radiation source individual identification in open electromagnetic environments.

[0006] The purpose of the present invention is achieved through the following technical solutions:

[0007] The present invention discloses a new end-to-end radiation source individual identification method based on signal enhancement entropy regularization, constructs an interpretable enhancement operator sequence constrained by the partial differential equation of electromagnetic propagation, and designs a multipath fading reconstruction → phase-frequency joint perturbation → frequency domain selective attenuation operator sequence, which effectively expands the sample diversity while ensuring the integrity of the signal fingerprint feature. The enhancement strategy not only overcomes the risk of feature distortion that may be introduced by traditional data enhancement, but also provides a high-quality feature discrimination basis for subsequent end-to-end dual prototype clustering. Then, a prototype classifier and a network fine-tuning end-to-end architecture based on signal enhancement entropy regularization are designed to organically integrate the category discrimination advantage of supervised prototype clustering with the feature decoupling ability of unsupervised prototype clustering: in the supervised prototype clustering branch, the feature similarity of similar enhanced samples is maximized to strengthen the known category representation; in the unsupervised prototype clustering branch, the potential category discrimination features are mined by decoupling irrelevant feature dimensions, forming a hybrid feature space with both discriminative and generalizable properties. An entropy regularization term based on signal enhancement is introduced to calculate the entropy of the predicted value and find the mean, which is added to the deep clustering loss as an additional regularization term. When the individual recognition model encounters a sample that is quite different from the known class (possibly an unknown class), the entropy regularization based on signal enhancement forces the model to output a more uniform probability distribution (high entropy) rather than forcibly assigning it to the known class. Traditional prototype classifiers may output overconfident probabilities (low entropy). Maximizing entropy is equivalent to assuming that the probability distribution of samples of unknown classes should be close to a uniform distribution (maximum uncertainty), which solves the problem of uneven distribution. The present invention significantly improves the environmental adaptability and model evolution capability of new individual identification of radiation sources in complex electromagnetic environments by establishing a full-process end-to-end optimization system from signal preprocessing, feature learning to dynamic end-to-end deep clustering topology, and by deeply integrating signal physical feature enhancement with data-driven learning.

[0008] The present invention discloses an end-to-end radiator new individual identification method based on signal enhancement entropy regularization, comprising the following steps:

[0009] Step 1: Collect the transmitter signal and down-convert it into an IQ signal to obtain the labeled and unlabeled basic input units;

[0010] The RF signal data emitted by the transmitter is collected, and the RF signal data is multiplied by a carrier with a phase difference of 90° to obtain the radiation source signal data; the in-phase and orthogonal IQ signals are obtained through a low-pass filter, and M consecutive pairs of IQ signal points are the basic input units of the network, whose dimension is 2*M; the known class and the unknown class are determined according to the transmitter serial number, and the basic input unit is divided into known class and unknown class basic input units; some known class basic input units are selected as labeled basic input units, and the remaining known class and all unknown class basic input units are selected as unlabeled basic input units.

[0011] Step 2: Build a residual neural network and use labeled basic input unit pre-training to obtain a multi-residual module feature extraction network;

[0012] First, a neural network based on residual connection is obtained by stacking multiple layers of residual connection modules;

[0013] Secondly, train the neural network based on residual connection: based on the labeled basic input unit obtained in step 1, perform forward propagation and backpropagation processes in sequence until the parameters of the neural network converge, and obtain a neural network that can be used for individual radiation source identification. The pre-training process is completed, and the network parameters obtained by pre-training are saved to obtain a multi-residual module feature extraction network.

[0014] Step 3: Design an interpretable enhancement operator sequence constrained by the electromagnetic propagation partial differential equation; enhance the labeled and unlabeled basic input units to obtain labeled and unlabeled enhanced samples;

[0015] Based on the labeled and unlabeled basic input units x(t) obtained in step 1, a sequence of interpretable enhancement operators constrained by multiple electromagnetic propagation partial differential equations is designed, and x(t) is enhanced through the sequence to obtain the enhanced sample x 增强 (t):

[0016]

[0017] represents the phase-frequency joint correction operator, represents the frequency domain selective attenuation operator, Represents function composition, multipath fading reconstruction operator Expressed as:

[0018]

[0019] Where τ represents the integral variable, which represents the propagation delay and is used to simulate the time delay component of multipath; τ +,-and τ +. / represent the minimum and maximum delay respectively; f0 represents the Doppler frequency shift; j represents the imaginary number symbol; Γ(τ) is the learnable delay spread kernel function, is the hidden layer parameter matrix; represents an adjustable kernel function based on τ, which is composed of learnable parameters control;

[0020] then Expressed as:

[0021]

[0022] in, Indicates function composition, that is, first Continue processing; STFT(·) represents short-time Fourier transform; STFT 11 (·) represents the inverse transform of STFT, which restores the time domain signal; Φ(f) represents the frequency-based weighting function; ⊙ represents the Hadamard product; φ(f) represents the phase function;

[0023] The phase compensation term satisfy:

[0024]

[0025] in, Represents the trainable weight parameter matrix, which belongs to the parameter part of the second-layer phase compensation network;

[0026] Represents frequency domain input features, that is, a set of features related to frequency;

[0027] further Expressed as:

[0028]

[0029] Where N is the number of activated channels, is a learnable wavelet basis, ReLU(·) represents the activation function, represents the convolution operation, represents the PDE-based structural modeling operator;

[0030] From equations (1) and (5), we can get the labeled and unlabeled enhanced samples x 增强 (t);

[0031] Step 4: Based on the pre-trained multi-residual module feature extraction network in step 2, remove the last fully connected layer of the pre-trained network, and connect the multi-layer perceptron (MLP) and the prototype classifier in sequence after the penultimate layer to obtain an end-to-end radiator individual recognition network;

[0032] The prototype classifier described in step 4 has a supervised prototype μ in the feature space c for:

[0033]

[0034] in, is a batch sample of known class c, x , represents the i-th sample belonging to category c, f : (·) represents the feature vector mapped from the feature extraction network to the embedding space;

[0035] Unsupervised Prototyping in Feature Space for:

[0036]

[0037] in, is a set of unlabeled basic input units, x $ Indicates that the jth belongs to is a set of unlabeled basic input units A sample of represents the unsupervised prototype of the previous iteration, represents the indicator function, k represents the number of prototypes;

[0038] According to formulas (6) and (7), the end-to-end radiation source individual recognition network connected to the prototype classifier is obtained.

[0039] Step 5: Based on the end-to-end radiator individual identification network obtained in step 4, obtain the supervised and unsupervised category belonging probabilities;

[0040] Based on the end-to-end radiation source individual recognition network obtained in step 4, the category belonging probability is calculated based on the supervised prototype μ9:

[0041]

[0042] Among them, p @AB (c|x , ) means that in the supervised prototype classifier, the sample x , The probability of belonging to category c, γ represents the temperature coefficient, and C represents the total number of labeled categories;

[0043] By enhancing the sample and pseudo prototype v ＝ The similarity between them is used to calculate the probability of category belonging based on the unsupervised prototype:

[0044]

[0045] Among them, p A-@AB (k|x ,) means that in the unsupervised prototype classifier, the sample x , Assigned to the pseudo-prototype v ＝ The probability of belonging to Represents sample x , , ι represents the temperature parameter, s(·) represents the identity function, and K represents the number of pseudo prototypes currently defined;

[0046] Step 6: Based on the supervised and unsupervised category belonging probabilities obtained in step 5, obtain the supervised and unsupervised prototype clustering losses;

[0047] Based on the p obtained in step 5 @AB (c|x , ) and p A-@AB (k|x , ), and obtain the supervised prototype clustering loss

[0048]

[0049] in, Represents the expectation of the input sample x and its true category label c;

[0050] Unsupervised Prototype Clustering Loss for:

[0051]

[0052] in, Represents the expectation of the input sample x and its pseudo label k;

[0053] Step 7: Design the entropy regularization term;

[0054] Design entropy regularization term for:

[0055]

[0056] Among them, the average probability of all samples belonging to the kth pseudo prototype is for:

[0057]

[0058] Among them, U represents the total number of unlabeled samples;

[0059] Step 8: Integrate steps 6 and 7 to obtain a dual prototype clustering loss based on entropy regularization, fine-tune the end-to-end radiation source individual identification network, and use the fine-tuned radiation source individual identification network to assign category labels to unlabeled basic input units to achieve new radiation source individual identification;

[0060] Based on steps six and seven, we get the dual prototype clustering loss for:

[0061]

[0062] Among them, λ1 represents The scaling factor, λ " express The scaling factor, β represents The scaling factor of

[0063] Utilizing dual prototype clustering loss Based on the end-to-end radiator individual recognition network obtained in step 4, the network parameters are fine-tuned, where only the last residual module is fine-tuned, and the weights of the remaining modules are frozen;

[0064] By fine-tuning the parameters of the radiation source individual identification network, it is possible to learn feature representations for given signal samples, and the end-to-end prototype classifier assigns category labels to unlabeled basic input units to achieve new radiation source individual identification;

[0065] The neural network for identifying new individual radiation sources obtained in step eight is applied to signal processing scenarios. It can identify the RF fingerprint category of unknown RF signals, achieve energy-efficient RF fingerprint recognition tasks, and provide radiation source individual information for subsequent processing of unknown RF signals.

[0066] Beneficial effects:

[0067] 1. This invention discloses a new end-to-end radiator individual identification method based on signal enhancement entropy regularization. It constructs a high-quality IQ input signal representation that can be used for deep learning, improving the expressiveness of signal features. By down-converting the original RF signal into an IQ signal and using M consecutive pairs of IQ samples as the basic input unit, this method effectively preserves the key time and frequency domain information in the transmitter's RF fingerprint. This low-dimensional but high-information signal representation provides a good data foundation for subsequent deep model training, enabling neural networks to more accurately learn and extract device-specific characteristics of the signal.

[0068] 2. The present invention discloses a new end-to-end radiator individual identification method based on signal enhancement entropy regularization, which improves the model's ability to extract complex signal features through pre-training using a residual connection neural network. The introduction of a deep residual neural network structure solves the gradient vanishing and degradation problems that occur in traditional deep networks during the training process. By using a small number of labeled samples for pre-training, the discriminative features of RF fingerprints can be learned in advance and transferable initial model parameters can be generated. This process greatly improves the convergence speed and generalization ability of the model in subsequent tasks, providing a stable and reliable feature extractor for RF signal recognition.

[0069] 3. This invention discloses a new end-to-end radiator identification method based on signal enhancement entropy regularization. This method utilizes an interpretable enhancement operator sequence constrained by the magnetic propagation partial differential equation to enhance the model's robustness and generalization capabilities. A sequence of multipath fading reconstruction, phase-frequency joint perturbation, and frequency-domain selective attenuation operators is designed to construct a variety of signal enhancement samples. This not only simulates the interference and changes in actual communication environments, but also effectively increases the diversity of training data, thereby strengthening the model's focus on key features, reducing sensitivity to irrelevant perturbations, and improving the model's stability and recognition capabilities in unknown environments.

[0070] 4. The present invention discloses a new end-to-end radiation source individual identification method based on signal enhancement entropy regularization, which integrates supervised and unsupervised end-to-end dual prototype clustering mechanisms, and further optimizes the feature space structure and improves the discrimination ability by uniformly predicting the network value based on the entropy regularization term of signal enhancement. The supervised and unsupervised prototype clustering objectives constructed based on signal enhancement act together in the model fine-tuning stage, so that the network can not only strengthen the feature similarity between similar signals, but also significantly widen the feature distance between different classes. This dual prototype clustering mechanism optimizes the structure of the embedding space, enhances the recognition accuracy of the individual recognition model for known classes, and improves the ability to distinguish unknown class samples. While maintaining the accuracy of recognition of known classes, the model also has the ability to discover and cluster unknown class signals. The present invention can realize automatic classification and fingerprint extraction of unknown radio frequency sources, improve the adaptability and intelligence of the system in an open set environment, and provide key individual identification information for target perception and subsequent processing in non-cooperative communication scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 Schematic diagram of the process of the new end-to-end radiation source individual identification method based on signal enhancement entropy regularization disclosed in the present invention;

[0072] Figure 2 Schematic diagram of a pre-trained neural network based on residual connections disclosed in this embodiment;

[0073] Figure 3 Schematic diagram of the end-to-end fine-tuning prototype classification neural network based on residual connection disclosed in this embodiment;

[0074] Figure 4 Schematic diagram of a fine-tuned residual connection neural network based on signal enhancement dual prototype clustering disclosed in this embodiment;

[0075] Figure 5 This is a schematic diagram of the dual prototype clustering process disclosed in this embodiment. DETAILED DESCRIPTION

[0076] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. The technical problems solved by the technical solution of the present invention and the beneficial effects thereof are also described. It should be noted that the described embodiments are only intended to facilitate understanding of the present invention and do not serve to limit the present invention in any way.

[0077] like Figure 1 As shown, the end-to-end radiator new individual identification method based on signal enhancement entropy regularization disclosed in this embodiment is specifically implemented as follows:

[0078] Step 1: Use two sub-datasets (ManyTx and SingleDay) from the open-source WiSig dataset and the ORACLE dataset to implement individual radiator identification. The WiSig dataset is a large-scale WiFi dataset containing 10 million data packets transmitted from 174 off-the-shelf WiFi transmitters and received by 41 USRP receivers over four acquisitions within a month. The dataset includes both raw data capture and a small, convenient pre-processed subset. The ORACLE dataset is the most commonly used dataset in the field of RF fingerprinting. The dataset sizes and data partitioning used are shown in the following table:

[0079] name Number of known classes Unknown number of classes Number of transmitters Number of receivers Collection days ManyTx 120 20 150 18 4 SingleDay 20 5 28 10 1 ORACLE 10 6 16 1 -

[0080] The ManySig and ManyRx preprocessing subsets of the WiSig dataset have a relatively small number of transmitters, making it difficult to demonstrate the advantages of the present invention. Therefore, this embodiment only uses the ManyTx and SingleDay subsets, which have a larger number of transmitters. The ManyTx and SingleDay subsets, as well as the ORACLE dataset, are multi-dimensional lists. In addition to the IQ sequence, they also include dimensions such as transmitter, receiver, acquisition day, and whether the signal is balanced. Because individual radiator identification requires eliminating the influence of factors such as receiver noise, channel noise, and channel variation, multiple dimensions such as receiver, number of signal segments, acquisition day, and whether the signal is balanced are integrated before training. Each IQ sequence segment is 2×256 in size, with two channels, respectively, for the real and imaginary parts. In this embodiment, the neural network input is a 2×256 tensor matrix. The three datasets ultimately utilize 2×256 tensor matrices with 119,463, 48,000, and 96,000 segments, respectively.

[0081] The dataset is divided into known class and unknown class basic input units according to the above table. At the same time, the known class basic input units are split according to the ratio of 8:2, 80% of which are labeled datasets, and the remaining 20% ​​and the unknown class basic input units are unlabeled datasets.

[0082] Step 2: Build a residual neural network such as Figure 2As shown, a multi-residual module feature extraction network is obtained by pre-training the labeled basic input unit;

[0083] First, a neural network based on residual connection is obtained by stacking multiple layers of residual connection modules. There are 4 residual connection modules in total: the first residual connection module is a connection of 1×1 convolution with an output dimension of 64, 3×3 convolution with an output dimension of 64, and 1×1 convolution with an output dimension of 256, repeated 3 times; the second residual connection module is a connection of 1×1 convolution with an output dimension of 128, 3×3 convolution with an output dimension of 128, and 1×1 convolution with an output dimension of 512, repeated 4 times; the third residual connection module is a connection of 1×1 convolution with an output dimension of 256, 3×3 convolution with an output dimension of 256, and 1×1 convolution with an output dimension of 1024, repeated 6 times; the fourth residual connection module is a connection of 1×1 convolution with an output dimension of 512, 3×3 convolution with an output dimension of 512, and 1×1 convolution with an output dimension of 2048, repeated 3 times.

[0084] Next, a residual connection-based neural network is trained: Based on the labeled basic input unit obtained in step 1, forward propagation and backpropagation are performed sequentially until the neural network parameters converge. This results in a neural network that can be used for individual radiation source identification. This completes the pretraining process, and the network parameters obtained from the pretraining are saved to obtain a multi-residual module feature extraction network. This embodiment uses the Adam optimizer during pretraining, with a learning rate set to 0.01, a batch size set to 256, an iteration count set to 20, and a cross-entropy loss function. The individual radiation source identification network is pretrained using gradient descent.

[0085] Step 3: Design an interpretable augmentation operator sequence constrained by the electromagnetic propagation partial differential equation; enhance the labeled and unlabeled basic input units to obtain labeled and unlabeled augmented samples;

[0086] For the labeled and unlabeled network basic input units x(t) in datasets such as the ManyTx and ORACLE datasets and the SingleDay dataset collected from various signal acquisition environments, which are collected by multiple transmitters and exhibit frequency and phase variations, we design an interpretable enhancement operator sequence constrained by multiple electromagnetic propagation partial differential equations and enhance x(t) through the sequence to obtain the enhanced sample x 增强 (t):

[0087]

[0088] represents the phase-frequency joint correction operator, represents the frequency domain selective attenuation operator, Represents function composition, multipath fading reconstruction operator Expressed as:

[0089]

[0090] Where τ represents the integral variable, which represents the propagation delay and is used to simulate the time delay component of multipath; τ +,- and τ +. / represent the minimum and maximum delay respectively; f0 represents the Doppler frequency shift; j represents the imaginary number symbol; Γ(τ) is the learnable delay spread kernel function, is the hidden layer parameter matrix; represents an adjustable kernel function based on τ, which is composed of learnable parameters control;

[0091] then Expressed as:

[0092]

[0093] in, Indicates function composition, that is, first Continue processing; STFT(·) represents short-time Fourier transform; STFT -1 (·) represents the inverse transform of STFT, which restores the time domain signal; Φ(f) represents the frequency-based weighting function; ⊙ represents the Hadamard product; φ(f) represents the phase function;

[0094] The phase compensation term satisfy:

[0095]

[0096] in, Represents the trainable weight parameter matrix, which belongs to the parameter part of the second-layer phase compensation network;

[0097] Represents frequency domain input features, that is, a set of features related to frequency;

[0098] further Expressed as:

[0099]

[0100] Where N is the number of activated channels, is a learnable wavelet basis, ReLU(·) represents the activation function, represents the convolution operation, represents the PDE-based structural modeling operator;

[0101] From equations (1) and (5), we can get the labeled and unlabeled enhanced samples x 增强 (t);

[0102] Step 4: Based on the pre-trained multi-residual module feature extraction network in step 2, remove the last fully connected layer of the pre-trained network and sequentially connect the multi-layer perceptron (MLP) and the prototype classifier after the penultimate layer to obtain an end-to-end radiator individual recognition network.

[0103] The prototype classifier described in step 4 has a supervised prototype μ in the feature space c for:

[0104]

[0105] in, is a batch sample of known class c, x , represents the i-th sample belonging to category c, f : (·) represents the feature vector mapped from the feature extraction network to the embedding space;

[0106] Unsupervised Prototyping in Feature Space for:

[0107]

[0108] in, is a set of unlabeled basic input units, x $ Indicates that the jth belongs to is a set of unlabeled basic input units A sample of represents the unsupervised prototype of the previous iteration, represents the indicator function, k represents the number of prototypes;

[0109] Step 5: Based on the end-to-end radiator individual identification network obtained in step 4, obtain the supervised and unsupervised category belonging probabilities;

[0110] Based on the end-to-end radiation source individual recognition network obtained in step 4, the category belonging probability is calculated based on the supervised prototype μ9:

[0111]

[0112] Among them, p @AB (c|x , ) means that in the supervised prototype classifier, the sample x , The probability of belonging to category c, γ represents the temperature coefficient, and C represents the total number of labeled categories;

[0113] By enhancing the sample and pseudo prototype v ＝ The similarity between them is used to calculate the probability of category belonging based on the unsupervised prototype:

[0114]

[0115] Among them, p A-@AB (k|x , ) means that in the unsupervised prototype classifier, the sample x , Assigned to the pseudo-prototype v ＝ The probability of belonging to Represents sample x , , ι represents the temperature parameter, s(·) represents the identity function, and K represents the number of pseudo prototypes currently defined;

[0116] Step 6: Based on the supervised and unsupervised category belonging probabilities obtained in step 5, obtain the supervised and unsupervised prototype clustering losses;

[0117] Based on the p obtained in step 5 @AB (c|x , ) and p A-@AB (k|x , ), and obtain the supervised prototype clustering loss

[0118]

[0119] in, Represents the expectation of the input sample x and its true category label c;

[0120] Unsupervised Prototype Clustering Loss for:

[0121]

[0122] in, Represents the expectation of the input sample x and its pseudo label k;

[0123] Step 7: Design entropy regularization term;

[0124] Design entropy regularization term for:

[0125]

[0126] Among them, the average probability of all samples belonging to the kth pseudo prototype is for:

[0127]

[0128] Among them, U represents the total number of unlabeled samples;

[0129] Step 8: Integrate steps 6 and 7 to obtain a dual prototype clustering loss based on entropy regularization, and fine-tune the end-to-end radiator individual identification network. Use the fine-tuned radiator individual identification network to assign category labels to unlabeled basic input units to achieve new radiator individual identification;

[0130] Based on steps six and seven, we get the dual prototype clustering loss for:

[0131]

[0132] Among them, λ1 represents The scaling factor, λ " express The scaling factor, β represents The scaling factor of

[0133] Utilizing dual prototype clustering loss In order to prevent overfitting to the fine-tuning set, the network parameters are fine-tuned based on the end-to-end radiator individual recognition network obtained in step 4, where only the last residual module is fine-tuned and the weights of the remaining modules are frozen. The specific network architecture is designed as follows: Figure 3 As shown in Figure 2, during the fine-tuning phase, the last fully connected layer of the pre-trained model is not updated to avoid capturing label-specific information. In other words, during fine-tuning, the second-layer features produced by the fully connected network of the multi-block pre-trained model are directly input into the MLP.

[0134] This example uses the SGD optimizer, the learning rate is set to Cosine annealing for 70 adjustment stages, the momentum is set to 0.9, the batch size is set to 256, the number of iterations is set to 70, the loss function is the cross entropy loss, the optimizer weight decay is set to 0.00005, the double contrast learning balance hyperparameter is set to λ1 = 0.5, and the unsupervised contrast temperature parameter τ is set to 0. A =0.3, the supervision comparison temperature parameter τ9=0.07, and the number of layers of the multi-layer perceptron MLP is 3. The RF fingerprint recognition neural network model is trained using the gradient descent method.

[0135] By fine-tuning the parameters of the radiation source individual identification network, the learning feature representation of a given signal sample can be achieved, and the end-to-end prototype classifier assigns category labels to unlabeled basic input units to realize the identification of new radiation sources; for unlabeled basic input units, by comparing the distance between them and the centers of labeled basic input units, the total intra-class distance (inertia) is minimized and the cluster centers and sample labels are continuously updated, ultimately obtaining a stable clustering result and realizing guided classification of unlabeled basic input units.

[0136] The present invention is compared with the K-means-based method and the semi-supervised K-means-based method. The K-means-based method and the present invention use the same feature extraction network to first pre-train, and fine-tune it through cross-entropy loss, and then apply the classic K-means algorithm to obtain cluster labels to achieve individual identification of radiation sources. The semi-supervised K-means-based method is implemented in the same way as the K-means-based method, except that the K-means algorithm is replaced by a semi-supervised constrained K-means algorithm. Specifically, the class cluster center in the labeled basic input unit is calculated based on the class label, and the unknown class cluster center is gradually selected and iterated for the unlabeled unknown class basic input unit until convergence to obtain the cluster label to achieve individual identification of radiation sources.

[0137] The following table shows the comparison of the effects of the present invention and other comparison methods, summarizing the effects of the present invention and the comparison methods from three dimensions: unlabeled basic input unit, unlabeled known class basic input unit, and unlabeled unknown class basic input unit.

[0138] Table 1 Effects of the present invention and other comparative methods

[0139]

[0140] As shown in the above table, the effects of the present invention are greatly improved in the effects of the unlabeled basic input unit, the unlabeled known class basic input unit, and the unlabeled unknown class basic input unit.

[0141] In summary, the new end-to-end radiator individual identification method based on signal enhancement entropy regularization disclosed in this embodiment first constructs an interpretable enhancement operator sequence constrained by the partial differential equation of electromagnetic propagation, and designs a multipath fading reconstruction → phase-frequency joint perturbation → frequency domain selective attenuation operator sequence. This not only overcomes the risk of feature distortion introduced by traditional data enhancement, but also provides a high-quality feature discrimination basis for subsequent end-to-end prototype clustering. Then, a prototype classifier and a network fine-tuning end-to-end architecture based on signal enhancement entropy regularization are constructed, organically integrating the category discrimination advantage of supervised prototype clustering with the feature decoupling ability of unsupervised prototype clustering to form a hybrid feature space with both discriminative and generalizable properties. Finally, an entropy regularization term based on signal enhancement is innovatively introduced. The algorithm calculates the entropy of the predicted value and finds the mean, which is added to the deep clustering loss as an additional regularization term. When the model encounters a sample that is significantly different from a known class (possibly an unknown class), the entropy regularization based on signal enhancement forces the model to output a more uniform probability distribution (high entropy) rather than forcibly assigning it to a known class. Traditional prototype classifiers may output overconfident probabilities (low entropy). Maximizing entropy is equivalent to assuming that the probability distribution of unknown class samples should be close to uniform distribution (maximum uncertainty) to solve the uneven distribution problem. The core innovation of this invention lies in the establishment of a full-process end-to-end optimization system from signal preprocessing and feature learning to dynamic end-to-end deep clustering topology, which significantly improves the environmental adaptability and model evolution capability of individual radiation source identification in complex electromagnetic environments.

[0142] The above specific description further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A new end-to-end radiator individual identification method based on signal enhancement entropy regularization, characterized by: The following steps are included: Step 1: Collect the transmitter signal and down-convert it into an IQ signal, and construct a training set and a test set containing labeled and unlabeled basic input units; Step 2: Build a residual neural network and use labeled basic input unit pre-training to obtain a multi-residual module feature extraction network; Step 3: Construct an interpretable enhancement operator sequence constrained by the electromagnetic propagation partial differential equation; enhance the labeled and unlabeled basic input units to obtain labeled and unlabeled enhanced samples; Step 4: Based on the pre-trained multi-residual module feature extraction network in step 2, connect the multi-layer perceptron (MLP) and the prototype classifier to obtain an end-to-end radiation source individual recognition network; Step 5: Based on the end-to-end radiator individual identification network obtained in step 4, design supervised and unsupervised prototype classification strategies to obtain supervised and unsupervised category assignment probabilities; Step 6: Based on the supervised and unsupervised category belonging probabilities obtained in step 5, design supervised and unsupervised prototype clustering losses; Step 7: Design the entropy regularization term; Step 8. Design a dual prototype clustering loss based on entropy regularization, fine-tune the end-to-end radiation source individual identification network obtained in step 4, and use the fine-tuned radiation source individual identification network to assign category labels to the test set to realize the identification of new radiation sources.

2. The end-to-end radiator new individual identification method based on signal enhancement entropy regularization according to claim 1 is characterized by: The implementation method of step one is: The RF signal data emitted by the transmitter is collected, and the RF signal data is multiplied by a carrier with a phase difference of 90° to obtain the radiation source signal data; the in-phase and orthogonal IQ signals are obtained through a low-pass filter, and M consecutive pairs of IQ signal points are the basic input units of the network, whose dimension is 2*M; the known class and the unknown class are determined according to the transmitter serial number, and the basic input units are divided into known class and unknown class basic input units; some known class basic input units are selected as labeled basic input units, and the remaining known class and all unknown class basic input units are selected as unlabeled basic input units; all basic input units are used as training sets, and the unlabeled basic input units are also used as test sets.

3. The end-to-end radiator new individual identification method based on signal enhancement entropy regularization according to claim 2, characterized in that: The implementation method of step 2 is: A neural network based on residual connection is obtained by stacking multiple layers of residual connection modules; Training a neural network based on residual connections: Based on the labeled basic input unit obtained in step 1, forward propagation and backpropagation processes are performed in sequence until the parameters of the neural network converge. A neural network that can be used for individual radiation source identification is obtained, completing the pre-training process and saving the network parameters obtained by pre-training to obtain a multi-residual module feature extraction network.

4. The end-to-end radiator new individual identification method based on signal enhancement entropy regularization according to claim 3 is characterized by: The implementation method of step three is: Based on the basic input signal unit x(t) of the known and unknown network obtained in step 1, a sequence of interpretable enhancement operators constrained by multiple electromagnetic propagation partial differential equations is designed to enhance it, and the enhanced sample x is obtained. 增强 (t): The specific design of the interpretable enhancement operator sequence is as follows: Design of multipath fading reconstruction operator Where τ represents the integral variable, which represents the propagation delay and is used to simulate the time delay component of multipath; τ +,- and τ +. / represent the minimum and maximum delay respectively; f0 represents the Doppler frequency shift; j represents the imaginary number symbol; Γ(τ) is the learnable delay spread kernel function, is the hidden layer parameter matrix; represents an adjustable kernel function based on τ, which is composed of learnable parameters control; Signal data Design of joint phase-frequency correction operator in, Indicates function composition, that is, first Continue processing; STFT(·) represents short-time Fourier transform; STFT 11 (·) represents the inverse transform of STFT, which restores the time domain signal; Φ(f) represents the frequency-based weighting function; ⊙ represents the Hadamard product; φ(f) represents the phase function; The phase compensation term satisfy: in, Represents the trainable weight parameter matrix, which belongs to the parameter part of the second-layer phase compensation network; Represents frequency domain input features, that is, a set of features related to frequency; Signal data Design of frequency-domain selective attenuation operator Where N is the number of activated channels, is a learnable wavelet basis, ReLU(·) represents the activation function, represents the convolution operation, represents the PDE-based structural modeling operator; Get labeled and unlabeled enhanced samples x 增强 (t).

5. The end-to-end radiator new individual identification method based on signal enhancement entropy regularization according to claim 4 is characterized by: The implementation method of step four is: Based on the pre-trained multi-residual module feature extraction network in step 2, remove the last fully connected layer of the pre-trained network and connect the multi-layer perceptron (MLP) and prototype classifier after the penultimate layer; For the differentiable prototype classifier, define the supervised prototype μ in the feature space c for: in, is a batch sample of known class c, x , represents the i-th sample belonging to category c, f : (·) represents the feature vector mapped from the feature extraction network to the embedding space, and θ represents the neural network parameters; Defining unsupervised prototypes in feature space for: in, is a set of unlabeled basic input units, x $ Indicates that the jth belongs to is a set of unlabeled basic input units A sample of represents the unsupervised prototype of the previous iteration, represents the indicator function, k represents the number of prototypes; According to formulas (6) and (7), the end-to-end radiation source individual recognition network connected to the prototype classifier is obtained.

6. The end-to-end radiator new individual identification method based on signal enhancement entropy regularization according to claim 5, characterized in that: The implementation method of step five is: Based on the end-to-end radiation source individual recognition network obtained in step 4, a supervised prototype classification strategy is designed, and the category belonging probability is calculated based on the labeled basic input unit prototype: Among them, p @AB (c|x , ) means that in the supervised prototype classifier, the sample x , The probability of belonging to category c, γ represents the temperature coefficient, and C represents the total number of labeled categories; Design an unsupervised prototype classification strategy, by enhancing the similarity between the sample and the pseudo prototype, calculate the probability of belonging to the category of the labeled and unlabeled prototypes: Among them, p A-@AB (k|x , ) means that in the unsupervised prototype classifier, the sample x , Assigned to the pseudo-prototype v ＝ The probability of belonging to Represents sample x , , ι represents the temperature parameter, s(·) represents the identity function, and K represents the number of pseudo prototypes currently defined.

7. The end-to-end radiator new individual identification method based on signal enhancement entropy regularization according to claim 6, characterized in that: The implementation method of step six is: Based on the supervised category belonging probability and the unsupervised category belonging probability obtained in step 5, the supervised prototype clustering loss is designed for: in, Represents the expectation of the input sample x and its true category label c; Designing an unsupervised prototype clustering loss for: in, Represents the expectation of the input sample x and its pseudo label k, z , represents the embedded characteristic, and ρ represents the temperature coefficient.

8. The end-to-end radiator new individual identification method based on signal enhancement entropy regularization according to claim 7, characterized in that: The implementation method of step seven is: Design entropy regularization term for: Among them, the average probability of all samples belonging to the kth pseudo prototype is for: Among them, U represents the total number of unlabeled samples.

9. The end-to-end radiator new individual identification method based on signal enhancement entropy regularization according to claim 8, characterized in that: The implementation method of step eight is: Based on steps 5, 6, and 7, design a dual prototype clustering loss for: Among them, λ1 represents The scaling factor, λ " express The scaling factor, β represents The scaling factor of Utilizing dual prototype clustering loss Based on the end-to-end radiator individual recognition network obtained in step 4, the network parameters are fine-tuned, where only the last residual module is fine-tuned, and the weights of the remaining modules are frozen; By fine-tuning the parameters of the radiation source individual identification network, it is possible to learn feature representations of given signal samples, and the end-to-end prototype classifier assigns category labels to the test set to achieve new individual radiation source identification.

Images