Radiation source new individual identification method based on signal enhancement double contrast learning

Abstract

The invention discloses a radiation source new individual identification method based on signal enhancement double contrast learning, and belongs to the field of software radio. According to the method, a multi-physical domain enhancement operator sequence is adopted, so that the risk that feature distortion is possibly introduced in traditional data enhancement is overcome, and a high-quality feature discrimination basis is provided for subsequent comparative learning by constructing positive and negative sample pairs with strong correlation. A feature decoupling architecture of double comparative learning is adopted, and the category discrimination advantage of supervised comparative learning and the feature decoupling capability of unsupervised comparative learning are organically fused. Dynamic semi-supervised clustering is realized by adopting a dual decision clustering mechanism, firstly, the number of potential categories is determined by utilizing density peak clustering DPC, secondly, a semi-supervised decision module is introduced to realize known category accurate identification and unknown category adaptive discovery of radiation source individuals, and then new radiation source individual identification is realized. The method is suitable for the field of software radio, and high-precision radiation source new individual identification is realized.

Description

Technical Field

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

[0002] Individual radiator identification is a core technology in the fields of electronic reconnaissance and radio frequency signal analysis. Its goal is to uniquely identify different radiator devices by extracting subtle features of the radiator signal (such as transient ripple, frequency offset, modulation fingerprint, etc.). Traditional methods for individual radiator identification typically rely on manually designed feature extraction algorithms, such as time domain features and frequency domain features. While these methods can achieve certain recognition results in some specific scenarios, their recognition accuracy is often limited in complex and dynamic environments, and their adaptability to unknown radiators is poor. Therefore, improving the accuracy and robustness of individual radiator identification, especially in complex signal environments, is an important direction for current technological development.

[0003] With the rapid development of deep learning technology, its application in the field of individual radiation source identification has gradually attracted widespread attention. Unlike traditional methods, deep learning can automatically extract effective features from raw signals, thus overcoming the heavy reliance on manual feature design. Through end-to-end training, deep learning models can learn complex signal features from large amounts of data, thereby improving the accuracy and robustness of individual radiation source identification. Existing technologies mainly focus on the classification problem of known categories of devices, such as using supervised learning frameworks to perform high-precision identification of radiation sources in a predefined device library. Some studies have attempted to combine semi-supervised or contrastive learning to enhance model robustness by mining the underlying structure of unlabeled data. However, these methods are still limited to closed set assumptions and cannot cope with the dynamic needs of known individual identification and multi-category new individual discovery in real-world scenarios.

[0004] The core bottleneck of current radiation source individual identification technology lies in the inability to detect new individuals in open environments. Existing methods often focus on high-precision classification of known categories in closed-set scenarios, or use isolated anomaly detection mechanisms to perform binary discrimination of unknown individuals (i.e., determine whether a signal belongs to a new individual). However, when multiple new radiation sources emerge simultaneously in complex electromagnetic game scenarios, these methods lack a joint optimization mechanism for learning distinguishable representations between unknown categories and incrementally constructing clustering topologies. This leads to new devices of different models / batches being misclassified into homogeneous clusters due to subtle feature differences. When new individuals are introduced into the model, the lack of a mechanism to protect the feature space of old categories often leads to a significant decline in the recognition performance of known devices. Maintaining the stability of the feature space of known categories while introducing new individuals and avoiding confusion between categories has become a key difficulty in technological development. Therefore, achieving the coordinated optimization of signal enhancement, decoupling of old and new categories, and dynamic identification has become a major challenge in the field of radiation source individual identification. In particular, how to effectively implement the new individual detection task through deep learning methods while protecting the feature space of known categories without relying on prior knowledge is the key to addressing this challenge. 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 and limited accuracy in identifying unknown categories, the present invention provides a new radiation source individual identification method based on signal enhancement and double contrast learning. This is a radiation source individual identification method that integrates signal enhancement and double contrast learning. By designing a multi-physical domain enhancement operator sequence, while strengthening the stability of the known category feature representation, the model's ability to discover unknown radiation sources 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 method for identifying new individual radiation sources based on signal-enhanced dual contrast learning, designs a multi-physical domain enhancement operator sequence (multipath fading reconstruction → phase-frequency joint perturbation → frequency domain selective attenuation), and effectively expands sample diversity while ensuring the integrity of signal fingerprint features. This 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 dual contrast learning by constructing positive and negative sample pairs with strong correlation. Then, a feature decoupling architecture for dual contrast learning is designed, and supervised contrast learning and unsupervised contrast learning based on signal enhancement are collaboratively optimized. Through the dual goals of intra-class compactification and inter-class separability of radiation source fingerprint features, model parameter fine-tuning is achieved, so that the discriminability of known classes and the discoverability of unknown classes are dynamically balanced. A dual decision clustering mechanism is introduced to realize dynamic semi-supervised clustering, and the density peak clustering DPC is used to determine the number of candidate clusters. A semi-supervised decision module is introduced to realize accurate identification of known classes of radiation source individuals and adaptive discovery of unknown classes, thereby realizing the identification of new individual radiation sources. The core innovation of this invention lies in establishing a full-process optimization system from signal preprocessing, feature learning to dynamic clustering. Through the deep integration of signal physical feature enhancement and data-driven learning, it significantly improves the environmental adaptability and model evolution ability of identifying new individual radiation sources in complex electromagnetic environments.

[0008] The present invention discloses a new radiation source individual identification method based on signal enhancement double contrast learning, 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 multiplied by a carrier wave with a 90° phase difference to obtain the radiating source signal data. In-phase and quadrature IQ signals are obtained through a low-pass filter. M consecutive pairs of IQ signal points serve as the network's basic input units, with a dimension of 2*M. The known and unknown classes are determined based on the transmitter serial number, and the basic input units are divided into known and unknown basic input units. Some known basic input units are selected as labeled basic input units, while the remaining known basic input units and all unknown basic input units are designated 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-layer feature extraction network.

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

[0013] Secondly, the neural network based on residual connection is trained: based on the labeled basic input unit obtained in step 1, the 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-layer feature extraction network.

[0014] Step 3: Design a multi-physical domain enhancement operator sequence. Enhance the labeled and unlabeled basic input units to obtain labeled and unlabeled enhanced samples.

[0015] Based on the labeled and unlabeled network basic input units obtained in step 1, a multi-physical domain enhancement operator sequence is designed to enhance them:

[0016] ① Multipath fading reconstruction MFR to eliminate the masking effect of spatial propagation differences on device fingerprints

[0017] Based on the labeled and unlabeled basic input units obtained in step 1, the IQ signal x(t) is directly operated to generate the adversarial perturbation h(t) by sparse multipath channel impulse response, and the IQ signal data x after multipath fading reconstruction enhancement is obtained. MFR (t):

[0018]

[0019] x MFR (t)=x(t)*h(t) (2)

[0020] The path gain a k , time delay τ k , phase θ k It obeys the Rayleigh-Rice mixed distribution, and its parameters are driven by the measured channel data. K represents the number of multipath fading propagation paths, and δ(t-τ k ) represents the Dirac pulse function, which means the signal at time τ k (i.e., the delay time of the kth path) has an impact, and j represents the imaginary number sign.

[0021] ② Phase-frequency joint perturbation to remove the interference of device-level time-varying parameters on steady-state characteristics

[0022] Based on data x MFR (t), simulate the Doppler effect and local oscillator drift, define the joint perturbation model of phase offset Δφ and frequency offset Δf, and obtain the IQ signal data x after phase-frequency joint perturbation enhancement PFP (t):

[0023]

[0024] in, Indicates that the frequency offset Δf obeys a uniform distribution and ranges from [-f max ,f max ]. Indicates that the phase shift Δφ has a mean of 0 and a variance of Gaussian distribution.

[0025] ③ Frequency domain selective attenuation SFA to suppress the pollution of environmental noise on the essential fingerprint

[0026] Based on data x PFP (t), design a channel bandwidth constraint model, implement band-limited filtering through a differentiable frequency domain mask, and obtain the IQ signal data x after frequency domain selective attenuation enhancement SFA (t):

[0027]

[0028] Among them, M LPF is the cutoff frequency f c , a low-pass filter with a transition bandwidth Δf, ⊙ represents a dot product operation, ∈ noise is the band-limited noise injection intensity, which obeys the measured signal-to-noise ratio distribution of the channel, Indicates that the mean is 0 and the variance is σ 2 White Gaussian noise (AWGN).

[0029] Data x SFA (t) is the enhanced sample obtained.

[0030] Step 4: Construct an unsupervised contrast loss function based on labeled and unlabeled enhanced samples.

[0031] For the same labeled or unlabeled basic input unit, any two enhanced samples are considered as positive sample pairs. (abbreviated as ), any two enhanced samples of different basic input units are regarded as negative sample pairs Constructing unsupervised contrastive learning loss based on signal enhancement

[0032]

[0033] Among them, τ unsup is the learnable temperature coefficient, D represents the number of all negative samples, e s(·) Indicates that the similarity scores of positive and negative samples are normalized using softmax for probabilistic modeling of contrastive learning.

[0034] Minimize the loss value of unsupervised contrastive learning based on signal enhancement To achieve unsupervised flow feature space:

[0035]

[0036] δ margin Represents a safety margin, ensuring that the distance between negative samples is at least δ greater than that between positive samples margin .

[0037] Step 5: Construct a supervised contrast loss function based on labeled enhanced samples.

[0038] For basic input units with the same label category, any two enhanced samples are considered as positive sample pairs z p , any two enhanced samples of different categories of basic input units are regarded as negative sample pairs z g , constructing a supervised contrastive learning loss value based on signal enhancement

[0039]

[0040] Among them, z i Represents the feature representation of the current anchor sample, P i is the positive sample set, G(i) is the total sample set including negative samples, τ sup is the learnable temperature coefficient.

[0041] Under the constraints of known category labels, minimize the value of supervised contrastive learning loss based on signal enhancement To achieve the supervised flow feature space:

[0042]

[0043] It means that it holds for all pairs of samples (i, j) with the same label, that is, x i and x j Belong to the same category, z i and z j They are samples x i and x j The feature representation of intra Indicates the preset maximum distance tolerance within a class.

[0044] Step 6. Design a differentiable weight allocator to achieve adaptive fusion of dual contrast loss.

[0045] Design a differentiable weight allocator to achieve adaptive fusion of dual contrast losses:

[0046] α u =σ(ω u ) (9)

[0047] Among them, α uis the fusion weight at step u, which is used to control the relative contribution of the two contrast losses (such as supervised + unsupervised). u Represents a learnable scalar parameter, or a weight dynamically generated by the network. σ(·) represents the Sigmoid function, ensuring that the output is in the (0,1) interval.

[0048] Step 7. Integrate the unsupervised contrast loss function, the supervised contrast loss function and the differentiable weight allocator to construct a dual contrast loss function.

[0049] According to steps 4, 5 and 6, we can get the double contrast loss function.

[0050]

[0051] λ represents a balance coefficient, which is used to control the weight of the temperature consistency constraint in the total loss.

[0052] Step 8. Design a hierarchical gradient update strategy and fine-tune the parameters of the pre-trained multi-layer feature extraction network obtained in step 2 based on the dual contrast loss function to obtain the fine-tuned feature extraction network.

[0053] Through the loss function Fine-tune the multi-layer feature extraction network obtained in step 2 and design a hierarchical gradient update fine-tuning method.

[0054] Frozen layer: retains the network parameters θ of the first N-1 residual blocks of the pre-trained feature extraction network frozen constant

[0055]

[0056] Among them, θ1, θ2,…, θ N-1 Represents the parameters of the 1st to N-1th residual blocks in the feature extraction network. N represents the total number of residual blocks in the feature extraction network.

[0057] Fine-tuning layer: only open the last residual block, and adjust its network parameters θ tunable Perform gradient update

[0058]

[0059] Among them, θ N represents the parameters of the Nth residual block in the feature extraction network, represents the parameter value at step u, represents the parameter value at step u+1, η represents the learning rate to control the step size of parameter update, Represents the loss function θ N The gradient of , that is, the update direction of the current parameters.

[0060] Design a dynamic learning rate adjustment strategy, and η is calculated as follows:

[0061]

[0062] Among them, η0 represents the initial benchmark learning rate, β is the time decay coefficient, which controls the speed at which the learning rate decays over time. Represents the gradient of the pre-trained model parameters in the initial state, ∈ is a small constant used to prevent the denominator from being zero.

[0063] When the last residual block performs gradient update, the update cutoff condition is

[0064]

[0065] After the update cutoff, the fine-tuned feature extraction network is obtained.

[0066] Step 9: Extract embedded features of the unlabeled basic input unit based on the fine-tuned feature extraction network. Dynamic semi-supervised clustering is implemented based on the embedded features through a dual-decision clustering mechanism to achieve the identification of new individual radiation sources.

[0067] Step 9.1: Extract embedded features.

[0068] Based on the unlabeled basic input unit in step 1, the penultimate layer embedded feature F is extracted using the fine-tuned feature extraction network obtained in step 4.

[0069] F = ψ(X; θ * ) (15)

[0070] Among them, ψ(·) represents the feature extraction network, X represents the input sample, and θ * Extract network parameters for the fine-tuned features.

[0071] Step 9.2, initial clustering stage: Based on the density peak detection of the embedded features extracted in step 9.1, the number of candidate clusters is obtained.

[0072] The sample set in feature F is represented as The local density ρ is obtained by formula (16): i With the minimum distance δ i :

[0073]

[0074] f i and f j Represent the feature vectors of the i-th and j-th samples in the embedding space, v represents the bandwidth parameter of the Gaussian kernel, which controls the smoothness of the density estimation, j:ρ j >ρi Represents a set of all sample indices with a density higher than i.

[0075] Based on f i and f j Introducing contrastive learning weight coefficient ω ij :

[0076]

[0077] Based on the local density ρ i , minimum distance δ i And the contrastive learning weight coefficient ω ij Design refactoring decision value γ i :

[0078] γ i =ρ i ·δ i ·max(ω ij (18) Obtain the number of candidate clusters R by screening the maximum reconstruction decision value cand ={r|γ r >ζ·max(γ)}.

[0079] Among them, γ r Represents the reconstruction decision value of the rth sample, ζ represents the threshold control coefficient, which usually takes a value between (0,1) and is used to control the number of candidate centers selected. The larger the value, the stricter the screening. max(γ) represents the maximum reconstruction decision value among all samples.

[0080] Step 9.3, iterative optimization phase: semi-supervised constrained objective function.

[0081] Defining cluster sets Where R is the number of candidate clusters obtained by density peak detection in the initial clustering stage. cand , center of mass μ r Link constraint sets Construct an improved objective function:

[0082]

[0083] Where f∈C r Indicates that the sample feature vector f belongs to the rth cluster, Represents a sample pair in the link constraint, two sample feature vectors f a and f b , ξ is the adaptive balance factor, Represents the indicator function, if the sample pair f a and f bIf they are classified into different clusters, the value is 1, otherwise it is 0. The Lagrange multiplier method is used to optimize (19) to ensure that the constrained sample pairs are forced to be classified into the same cluster.

[0084] Design centroid initialization and iterative optimization strategy:

[0085] Initialization: Based on R cand Generate initial centroids

[0086] Iterative optimization: Calculate the new centroid of the rth cluster in the v+1th iteration

[0087]

[0088] in, Indicates the sample set contained in the rth cluster in the current vth iteration, Represents the sample feature vector currently belonging to the rth cluster, Indicates that it currently belongs to the link constraint set The sample pairs, and express and , θ is the supervision strength factor, and κ(·) is the Kronecker function.

[0089] Step 9.4: Dynamically adjust the strategy: confidence-driven closed-loop update.

[0090] In order to realize the closed-loop dynamic recognition mechanism, a cluster-level confidence evaluation function Conf(C r ) to measure the stability of each cluster, where:

[0091]

[0092] Where τ is the temperature coefficient, which represents the density of sample distribution within the cluster.

[0093] Known Class Identification: When Conf(C r )≥θ high When θ high A high confidence threshold is set to determine whether a cluster can be considered a known class. If the samples in the cluster are highly concentrated, it can be determined to be a stable known class. The label propagation algorithm is activated to confirm the class and update the known class template library.

[0094] New Class Discovery: When Conf(C r )≤θ low When θ lowA low confidence threshold is used to determine whether a cluster needs to be considered as a potential new class. Clusters with dispersed structures and high uncertainty are considered as potential new classes, triggering the semi-supervised clustering mechanism to perform class classification and temporarily storing their features in the unknown class cache pool.

[0095] Combining the dynamic responses of the above two strategies, the real-time update of the category labels after clustering is completed, and the identification and classification of new individuals of communication radiation sources are realized.

[0096] The feature extraction network and two-layer clustering decision method for RF fingerprint identification obtained in steps 8 and 9 are applied to signal processing scenarios. This can discover the RF fingerprint category of unknown RF signals, achieve energy-efficient RF fingerprint identification tasks, and provide individual radiation source information for subsequent processing of unknown RF signals.

[0097] Beneficial effects:

[0098] 1. The present invention discloses a novel method for identifying individual radiators based on signal-enhanced dual contrast learning. This method constructs high-quality IQ signals suitable for deep learning, improving the expressiveness of signal features. By down-converting the collected RF signal into an IQ signal and using M pairs of consecutive IQ sample points as the basic input unit, it can preserve the key time and frequency domain information of the transmitter's RF fingerprint. This low-dimensional but high-information signal representation provides a good data foundation for subsequent network training, enabling the neural network to more accurately learn and extract the signal's device-specific features.

[0099] 2. The present invention discloses a new method for identifying individual radiation sources based on signal-enhanced dual contrast learning, which implements pre-training of residual connection neural networks and improves the model's ability to extract complex signal features. The introduction of a multi-layer residual neural network structure solves the gradient vanishing and degradation problems that occur in traditional neural networks during training. By using labeled basic input units for pre-training, the discriminative features of RF fingerprints can be learned in advance, and a transferable pre-trained feature extraction network can be obtained. This process significantly improves the convergence speed and generalization ability of the network in subsequent tasks, providing a stable and reliable feature extractor for RF signal recognition.

[0100] 3. The new individual radiator identification method based on signal-enhanced dual contrast learning disclosed in this invention utilizes a multi-physical domain enhancement operator sequence to construct diverse training samples, enhancing the network's robustness and generalization capabilities. A variety of signal enhancement samples are constructed through a physical mechanism sequence of multipath fading reconstruction → phase-frequency joint perturbation → frequency-domain selective attenuation. This not only simulates the interference and changes in actual communication environments, but also effectively increases the diversity of training data. The positive and negative sample pairs constructed on this basis can strengthen the network's focus on key features, reduce sensitivity to irrelevant perturbations, and improve the network's stability and recognition capabilities in unknown environments.

[0101] 4. The present invention discloses a new method for identifying individual radiation sources based on signal-enhanced dual contrast learning. This method integrates a supervised and unsupervised dual contrast learning feature decoupling architecture to further optimize the feature space structure and improve discrimination capabilities. The supervised and unsupervised contrast learning objectives based on signal enhancement work together during the network fine-tuning phase, enabling the network to not only strengthen the feature similarities between similar signals but also significantly widen the feature distances between different classes. The dual contrast mechanism optimizes the structure of the embedding space, enhancing the network's recognition accuracy for known classes while improving its ability to distinguish samples of unknown classes.

[0102] 5. The method for identifying new individual radiation sources based on signal-enhanced dual contrast learning disclosed in the present invention realizes dynamic semi-supervised clustering through a dual-decision clustering mechanism to realize unknown radio frequency signal discovery and improve the system's open set recognition capability. First, density peak clustering DPC is performed to determine the number of candidate clusters. Secondly, a semi-supervised decision module is introduced to realize accurate identification of known categories of radiation sources and adaptive discovery of unknown categories, thereby realizing new individual radiation source identification. While maintaining the accuracy of recognition of known categories, the network also has the ability to discover unknown category signals. The present invention can realize automatic classification and fingerprint extraction of unknown radiation 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

[0103] Figure 1 This is a flow chart of a new method for identifying radiation sources based on signal-enhanced double contrast learning disclosed in the present invention;

[0104] Figure 2 Schematic diagram of the pre-trained feature extraction network based on residual connection disclosed in this embodiment;

[0105] Figure 3 Schematic diagram of a fine-tuning feature extraction network based on signal enhancement dual contrast learning disclosed in this embodiment;

[0106] Figure 4 Schematic diagram of constructing positive and negative sample pairs for the dual contrastive learning disclosed in this embodiment. DETAILED DESCRIPTION

[0107] 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.

[0108] like Figure 1 As shown, the method for identifying new radiation sources based on signal enhancement and double contrast learning disclosed in this embodiment has the following specific implementation steps:

[0109] 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:

[0110] Dataset 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 -

[0111] 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.

[0112] 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.

[0113] Step 2: Build a residual neural network such as Figure 2 As shown, a multi-layer feature extraction network is obtained by pre-training with labeled basic input units;

[0114] 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.

[0115] Next, a residual connection-based neural network is trained: Based on the labeled basic input units obtained in step 1, forward and backward propagation processes are performed sequentially until the neural network parameters converge. This results in a neural network suitable for individual radiation source identification, completing the pre-training process. The network parameters obtained from the pre-training are saved to obtain a multi-layer feature extraction network. This embodiment uses the Adam optimizer during pre-training, 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 pre-trained using gradient descent.

[0116] Step 3: Design a multi-physical domain enhancement operator sequence; enhance the labeled and unlabeled basic input units to obtain labeled and unlabeled enhanced samples;

[0117] Based on the basic input signal units of the known and unknown network obtained in step 1, a multi-physical domain enhancement operator sequence is designed to enhance them:

[0118] ① Multipath fading reconstruction MFR to eliminate the masking effect of spatial propagation differences on device fingerprints

[0119] Based on the known and unknown basic input units obtained in step 1, the IQ signal x(t) is directly operated to generate the sparse multipath channel impulse response to generate the adversarial perturbation h(t), and the IQ signal data x after multipath fading reconstruction enhancement is obtained. MFR (t):

[0120]

[0121] The path gain a k , time delay τ k , phase θ kIt obeys the Rayleigh-Rice mixed distribution, and its parameters are driven by the measured channel data. K represents the number of multipath fading propagation paths, and δ(t-τ k ) represents the Dirac pulse function, which means the signal at time τ k (i.e., the delay time of the kth path) has an impact, and j represents the imaginary number sign.

[0122] ② Phase-frequency joint perturbation to remove the interference of device-level time-varying parameters on steady-state characteristics

[0123] Based on data x MFR (t), simulate the Doppler effect and local oscillator drift, define the joint perturbation model of phase offset Δφ and frequency offset Δf, and obtain the IQ signal data x after phase-frequency joint perturbation enhancement PFP (t):

[0124]

[0125] in, Indicates that the frequency offset Δf obeys a uniform distribution and ranges from [-f max ,f max ]; Indicates that the phase shift Δφ has a mean of 0 and a variance of Gaussian distribution.

[0126] ③ Frequency domain selective attenuation SFA to suppress the pollution of environmental noise on the essential fingerprint

[0127] Based on data x PFP (t), design a channel bandwidth constraint model, implement band-limited filtering through a differentiable frequency domain mask, and obtain the IQ signal data x after frequency domain selective attenuation enhancement SFA (t):

[0128]

[0129] Among them, M LPF is the cutoff frequency f c , a low-pass filter with a transition bandwidth Δf, ⊙ represents a dot product operation, ∈ noise is the band-limited noise injection intensity, which obeys the measured signal-to-noise ratio distribution of the channel, Indicates that the mean is 0 and the variance is σ 2 White Gaussian noise (AWGN).

[0130] Data x SFA (t) is the enhanced sample obtained;

[0131] Step 4: Construct an unsupervised contrast loss function based on labeled and unlabeled enhanced samples;

[0132] For the same basic input unit, any two enhanced samples are considered as positive sample pairs. (abbreviated as ), any two enhanced samples of different basic input units are regarded as negative sample pairs Constructing unsupervised contrastive learning loss based on signal enhancement

[0133]

[0134] Among them, τ unsup is the learnable temperature coefficient, D represents the number of all negative samples, e s(·) Indicates that the similarity scores of positive and negative samples are normalized using softmax for probabilistic modeling of contrastive learning;

[0135] Minimize the loss value of unsupervised contrastive learning based on signal enhancement To achieve unsupervised flow feature space:

[0136]

[0137] δ margin Represents a safety margin, ensuring that the distance between negative samples is at least δ greater than that between positive samples margin .

[0138] Step 5: Construct a supervised contrast loss function based on labeled enhanced samples;

[0139] For basic input units with the same label category, any two enhanced samples are considered as positive sample pairs z p , any two enhanced samples of different categories of basic input units are regarded as negative sample pairs z g , constructing a supervised contrastive learning loss value based on signal enhancement

[0140]

[0141] Among them, z i Represents the feature representation of the current anchor sample, P i is the positive sample set, G(i) is the total sample set including negative samples, τ sup is the learnable temperature coefficient;

[0142] Under the constraints of known category labels, minimize the value of supervised contrastive learning loss based on signal enhancement To achieve the supervised flow feature space:

[0143]

[0144] It means that it holds for all pairs of samples (i, j) with the same label, that is, x i and x j Belong to the same category, z i and z j They are samples x i and x j The feature representation of intra Indicates the preset maximum distance tolerance within the class;

[0145] Step 6: Design a differentiable weight allocator to achieve adaptive fusion of dual contrast loss;

[0146] Design a differentiable weight allocator to achieve adaptive fusion of dual contrast losses:

[0147] α u =σ(ω u )(30)

[0148] Among them, α u is the fusion weight at step u, which is used to control the relative contribution of the two contrast losses (such as supervised + unsupervised); ω u represents a learnable scalar parameter, or a weight dynamically generated by the network; σ(·) represents the Sigmoid function, ensuring that the output is in the (0,1) interval;

[0149] Step 7: Integrate steps 4 to 6 to construct a dual contrast loss function, such as Figure 3 As shown;

[0150] According to steps 4, 5 and 6, we can get the double contrast loss function.

[0151]

[0152] λ represents a balance coefficient, which is used to control the weight of the temperature consistency constraint in the total loss;

[0153] Step 8: Design a hierarchical gradient update strategy and fine-tune the parameters of the pre-trained multi-layer feature extraction network obtained in step 2 based on the dual contrast loss function to obtain a fine-tuned feature extraction network.

[0154] Through the loss function Fine-tune the multi-layer feature extraction network obtained in step 2 and design a hierarchical gradient update fine-tuning method.

[0155] Frozen layer: retains the network parameters θ of the first N-1 residual blocks of the pre-trained feature extraction network frozen constant

[0156]

[0157] Among them, θ1, θ2,…, θ N-1 Represents the parameters of the 1st to N-1th residual blocks in the feature extraction network; N represents the total number of residual blocks in the feature extraction network.

[0158] Fine-tuning layer: only open the last residual block, and adjust its network parameters θ tunable Perform gradient update

[0159]

[0160] Among them, θ N represents the parameters of the Nth residual block in the feature extraction network, represents the parameter value at step u, represents the parameter value at step u+1, η represents the learning rate to control the step size of parameter update, Represents the loss function θ N The gradient of , that is, the update direction of the current parameters.

[0161] Design a dynamic learning rate adjustment strategy, and η is calculated as follows:

[0162]

[0163] Among them, η0 represents the initial benchmark learning rate, β is the time decay coefficient, which controls the speed at which the learning rate decays over time. Represents the gradient of the pre-trained model parameters in the initial state, ∈ is a small constant used to prevent the denominator from being zero.

[0164] When the last residual block performs gradient update, the update cutoff condition is

[0165]

[0166] After the update cutoff, the fine-tuned feature extraction network is obtained.

[0167] In this embodiment, the SGD optimizer is used in the process of fine-tuning network parameters. The learning rate is set to Cosineannealing for a total of 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 cross entropy loss, the optimizer weight decay is set to 0.00005, and the unsupervised comparison temperature parameter τ is used. unsup =0.3, supervision comparison temperature parameter τ sup =0.07. Gradient descent method was used to fine-tune the pre-trained feature extraction network for radiation source individual recognition.

[0168] Step 9: Extract the embedded features of the unlabeled basic input unit based on the fine-tuned feature extraction network; based on the embedded features, implement dynamic semi-supervised clustering through a dual-decision clustering mechanism to achieve new individual radiation source identification;

[0169] Step 9.1: Extract embedded features.

[0170] Based on the unlabeled basic input unit in step 1, the penultimate layer embedded feature F is extracted using the fine-tuned feature extraction network obtained in step 4.

[0171] F = ψ(X; θ * ) (36)

[0172] Among them, ψ(·) represents the feature extraction network, X represents the input sample, and θ * Extract network parameters for fine-tuned features;

[0173] Step 9.2, initial clustering stage: based on the density peak detection of the embedded features extracted in step 9.1, the number of cluster candidates is obtained.

[0174] The sample set in feature F is represented as The local density ρ is obtained by formula (16): i With the minimum distance δ i :

[0175]

[0176] f i and f j Represent the feature vectors of the i-th and j-th samples in the embedding space, v represents the bandwidth parameter of the Gaussian kernel, which controls the smoothness of the density estimation, j:ρ j >ρ i Represents a set of all sample indices with a density higher than i.

[0177] Based on f i and f j Introducing contrastive learning weight coefficient ω ij :

[0178]

[0179] Based on the local density ρ i , minimum distance δ i And the contrastive learning weight coefficient ω ij Design refactoring decision value γ i :

[0180] γ i =ρ i ·δ i ·max(ω ij) (39)

[0181] The number of candidate clusters R is obtained by screening the maximum reconstruction decision value cand ={r|γ r >ζ·max(γ)};

[0182] Among them, γ r Represents the reconstruction decision value of the rth sample, ζ represents the threshold control coefficient, which usually takes a value between (0,1) and is used to control the number of candidate centers selected. The larger the value, the stricter the screening. max(γ) represents the maximum reconstruction decision value among all samples.

[0183] Step 9.3, iterative optimization phase: semi-supervised constrained objective function.

[0184] Defining cluster sets Where R is the number of candidate clusters obtained by density peak detection in the initial clustering stage. cand , center of mass μ r Link constraint sets Construct an improved objective function:

[0185]

[0186] Where f∈C r Indicates that the sample feature vector f belongs to the rth cluster, Represents a sample pair in the link constraint, two sample feature vectors f a and f b , ξ is the adaptive balance factor, Represents the indicator function, if the sample pair f a and f b If they are classified into different clusters, it is 1, otherwise it is 0. The Lagrange multiplier method is used to optimize (19) to ensure that the constrained sample pairs are forced to be classified into the same cluster;

[0187] Design centroid initialization and iterative optimization strategy:

[0188] Initialization: Based on R cand Generate initial centroids

[0189] Iterative optimization: Calculate the new centroid of the rth cluster in the v+1th iteration

[0190]

[0191] in, Indicates the sample set contained in the rth cluster in the current vth iteration, Represents the sample feature vector currently belonging to the rth cluster, Indicates that it currently belongs to the link constraint set The sample pairs, and express and Tags, is the supervision strength factor, k(·) is the Kronecker function;

[0192] Step 9.4: Dynamically adjust the strategy: confidence-driven closed-loop update.

[0193] In order to realize the closed-loop dynamic recognition mechanism, a cluster-level confidence evaluation function Conf(C r ) to measure the stability of each cluster, where:

[0194]

[0195] Where τ is the temperature coefficient, which represents the density of sample distribution within the cluster;

[0196] Known Class Identification: When Conf(C r )≥θ high When θ high A high confidence threshold is used to determine whether a cluster can be considered a known class. If the samples in the cluster are highly concentrated, it can be determined to be a stable known class. The label propagation algorithm is activated to confirm the class and update the known class template library.

[0197] New Class Discovery: When Conf(C r )≤θ low When θ low A low confidence threshold is used to determine whether a cluster needs to be considered as a potential new class. Clusters with dispersed structures and high uncertainty are considered as potential new classes, triggering the semi-supervised clustering mechanism to perform class classification and temporarily storing their features in the unknown class cache pool;

[0198] Combining the dynamic responses of the above two strategies, the real-time update of the category labels after clustering is completed, and the identification and classification of new individuals of communication radiation sources are realized.

[0199] The present invention is compared with the K-means-based method and the semi-supervised K-means-based method. The K-means-based method uses the same feature extraction network as the present invention to first pre-train, and fine-tune it through cross-entropy loss, and then applies 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 centers in the labeled basic input units are calculated based on the class labels, and the unknown class cluster centers are gradually selected and iterated for the unknown class samples in the unlabeled basic input units until convergence, so as to obtain cluster labels and achieve individual identification of radiation sources.

[0200] 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.

[0201] Table 1 Effects of the present invention and other comparative methods

[0202]

[0203] As shown in the above table, the present invention has a significant improvement in the effects of the unlabeled basic input unit, the unlabeled known class basic input unit, and the unlabeled unknown class basic input unit.

[0204] In summary, the method for identifying new individual radiation sources based on signal-enhanced dual contrast learning disclosed in this embodiment mainly describes the use of signal-enhanced dual contrast learning to achieve the recognition of mixed data of known classes and multiple unknown classes. First, a multi-physical domain enhancement operator sequence (multipath fading reconstruction → phase-frequency joint perturbation → frequency domain selective attenuation) is designed, which 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 contrast learning by constructing positive and negative sample pairs with strong correlation. Then, a feature decoupling architecture of dual contrast learning is designed to organically integrate the category discrimination advantages of supervised contrast learning with the feature decoupling capabilities of unsupervised contrast learning. Finally, a dual decision clustering mechanism is introduced to realize dynamic semi-supervised clustering. First, density peak clustering DPC is performed to determine the number of potential classes. Secondly, a semi-supervised decision module is introduced to realize accurate identification of known classes of radiation source individuals and adaptive discovery of unknown classes. The core innovation of the present invention lies in the establishment of a full-process optimization system from signal preprocessing, feature learning to dynamic clustering, which significantly improves the environmental adaptability and model evolution capability of new individual radiation source identification in complex electromagnetic environments.

[0205] 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 method for identifying individual radiation sources based on signal enhancement and double contrast learning, characterized by: The following steps are included: Step 1: Collect the transmitter signal and down-convert it into an IQ signal to obtain the labeled and unlabeled basic input units; Step 2: Build a residual neural network and use labeled basic input units for pre-training to obtain a multi-layer feature extraction network; Step 3: Design a multi-physical domain enhancement operator sequence; enhance the labeled and unlabeled basic input units to obtain labeled and unlabeled enhanced samples; Step 4: Construct an unsupervised contrast loss function based on labeled and unlabeled enhanced samples; Step 5: Construct a supervised contrast loss function based on labeled enhanced samples; Step 6: Design a differentiable weight allocator to achieve adaptive fusion of dual contrast loss; Step 7: Integrate the unsupervised contrast loss function, the supervised contrast loss function and the differentiable weight allocator to construct a dual contrast loss function; Step 8: Design a hierarchical gradient update strategy and fine-tune the parameters of the pre-trained multi-layer feature extraction network obtained in step 2 based on the dual contrast loss function to obtain a fine-tuned feature extraction network. Step 9: Extracting unlabeled basic input unit embedded features based on the fine-tuned feature extraction network; Based on the embedded features, dynamic semi-supervised clustering is implemented through a dual-decision clustering mechanism to realize the identification of new radiation sources.

2. The method for identifying new radiation sources based on signal enhancement and double contrast learning according to claim 1, wherein: 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 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.

3. The method for identifying new radiation sources based on signal enhancement and double contrast learning according to claim 2, wherein: 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 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-layer feature extraction network.

4. The method for identifying new radiation sources based on signal enhancement and double contrast learning as claimed in claim 3, characterized in that: The implementation method of step three is: Based on the labeled and unlabeled network basic input units obtained in step 1, a multi-physical domain enhancement operator sequence is designed to enhance them: ① Multipath fading reconstruction MFR to eliminate the masking effect of spatial propagation differences on device fingerprints; Based on the labeled and unlabeled basic input units obtained in step 1, the IQ signal x(t) is directly operated to generate the adversarial perturbation h(t) by sparse multipath channel impulse response, and the IQ signal data x after multipath fading reconstruction enhancement is obtained. MFR (t): x MFR (t)=x(t)*h(t) (2) The path gain a k , time delay τ k , phase θ k It obeys the Rayleigh-Rice mixed distribution, and its parameters are driven by the measured channel data. K represents the number of multipath fading propagation paths, and δ(t-τ k ) represents the Dirac pulse function, which means the signal at time τ k (i.e., the delay time of the kth path) there is a shock, and j represents the imaginary number sign; ② Joint phase-frequency perturbation to remove the interference of device-level time-varying parameters on steady-state characteristics; Based on data x MFR (t), simulate the Doppler effect and local oscillator drift, define the joint perturbation model of phase offset Δφ and frequency offset Δf, and obtain the IQ signal data x after phase-frequency joint perturbation enhancement PFP (t): in, Indicates that the frequency offset Δf obeys a uniform distribution and ranges from [-f max ,f max ]; Indicates that the phase shift Δφ has a mean of 0 and a variance of Gaussian distribution; ③ Frequency domain selective attenuation (SFA) to suppress the contamination of the essential fingerprint by environmental noise; Based on data x PFP (t), design a channel bandwidth constraint model, implement band-limited filtering through a differentiable frequency domain mask, and obtain the IQ signal data x after frequency domain selective attenuation enhancement SFA (t): Among them, M LPF is the cutoff frequency f : , a low-pass filter with a transition bandwidth Δf, ⊙ represents a dot product operation, ∈ noise is the band-limited noise injection intensity, which obeys the measured signal-to-noise ratio distribution of the channel, Indicates that the mean is 0 and the variance is σ 2 Gaussian white noise AWGN; Data x SFA (t) is the enhanced sample obtained.

5. The method for identifying new radiation sources based on signal enhancement and double contrast learning according to claim 4, characterized in that: The implementation method of step four is: For the same basic input unit, any two enhanced samples are considered as positive sample pairs. (abbreviated as ), any two enhanced samples of different basic input units are regarded as negative sample pairs Constructing unsupervised contrastive learning loss based on signal enhancement Among them, τ unsup is the learnable temperature coefficient, D represents the number of all negative samples, e s(·) Indicates that the similarity scores of positive and negative samples are normalized using softmax for probabilistic modeling of contrastive learning; Minimize the loss value of unsupervised contrastive learning based on signal enhancement To achieve unsupervised flow feature space: δ margin Represents a safety margin, ensuring that the distance between negative samples is at least δ greater than that between positive samples margin .

6. The method for identifying new radiation sources based on signal enhancement and double contrast learning according to claim 5, characterized in that: The implementation method of step five is: For basic input units with the same label category, any two enhanced samples are considered as positive sample pairs z p , any two enhanced samples of different categories of basic input units are regarded as negative sample pairs z g , constructing a supervised contrastive learning loss value based on signal enhancement Among them, z i Represents the feature representation of the current anchor sample, P i is the positive sample set, G(i) is the total sample set including negative samples, τ sup is the learnable temperature coefficient; Under the constraints of known category labels, minimize the value of supervised contrastive learning loss based on signal enhancement To achieve the supervised flow feature space: It means that it holds for all pairs of samples (i, j) with the same label, that is, x i and x j Belong to the same category, z i and z j They are samples x i and x j The feature representation of intra Indicates the preset maximum distance tolerance within a class.

7. The method for identifying new radiation sources based on signal enhancement and double contrast learning according to claim 6, wherein: The implementation method of step six is: Design a differentiable weight allocator to achieve adaptive fusion of dual contrast losses: a u =σ(ω u ) (9) Among them, α u is the fusion weight at step u, which is used to control the relative contribution of the two contrast losses (such as supervised + unsupervised); ω u represents a learnable scalar parameter or a weight dynamically generated by the network; σ(·) represents the Sigmoid function, which ensures that the output is in the (0,1) interval.

8. The method for identifying new radiation sources based on signal enhancement and double contrast learning according to claim 7, wherein: The implementation method of step seven is: According to steps 4, 5 and 6, we can get the double contrast loss function. λ represents a balance coefficient, which is used to control the weight of the temperature consistency constraint in the total loss.

9. The method for identifying new radiation sources based on signal enhancement and double contrast learning as claimed in claim 8, characterized in that: The implementation method of step eight is: Through the loss function Fine-tune the multi-layer feature extraction network obtained in step 2 and design a hierarchical gradient update fine-tuning method; Frozen layer: retains the network parameters θ of the first N-1 residual blocks of the pre-trained feature extraction network frozen constant Among them, θ1,θ , ,…,θ N-1 Represents the parameters of the 1st to N-1th residual blocks in the feature extraction network; N represents the total number of residual blocks in the feature extraction network; Fine-tuning layer: only open the last residual block, and adjust its network parameters θ tunable Perform gradient update Among them, θ N represents the parameters of the Nth residual block in the feature extraction network, represents the parameter value at step u, represents the parameter value at step u+1, η represents the learning rate to control the step size of parameter update, Represents the loss function θ N The gradient of , that is, the update direction of the current parameters; Design a dynamic learning rate adjustment strategy, and η is calculated as follows: Among them, η0 represents the initial benchmark learning rate, β is the time decay coefficient, which controls the speed at which the learning rate decays over time. Represents the gradient of the pre-trained model parameters in the initial state, ∈ is a small constant used to prevent the denominator from being zero; When the last residual block performs gradient update, the update cutoff condition is After the update cutoff, the fine-tuned feature extraction network is obtained.

10. The method for identifying new radiation sources based on signal enhancement and double contrast learning according to claim 9, characterized in that: The implementation method of step nine is: Step 9.1, extract embedded features; Based on the unlabeled basic input unit in step 1, the penultimate layer embedded feature F is extracted using the fine-tuned feature extraction network obtained in step 4. F=ψ(X;θ * ) (15) Among them, ψ(·) represents the feature extraction network, X represents the input sample, and θ * Extract network parameters for fine-tuned features; Step 9.2, initial clustering stage: based on the density peak detection of the embedded features extracted in step 9.1, the number of candidate clusters is obtained; The sample set in feature F is represented as The local density ρ is obtained by formula (16): i With the minimum distance δ i : f i and f j Represent the feature vectors of the i-th and j-th samples in the embedding space, v represents the bandwidth parameter of the Gaussian kernel, which controls the smoothness of the density estimation, j:ρ j >ρ i Represents a set of all sample indices with a density higher than i; Based on f i and f j Introducing contrastive learning weight coefficient ω ij : Based on the local density ρ i , minimum distance δ i And the contrastive learning weight coefficient ω ij Design refactoring decision value γ i : c i =ρ i ·d i ·max(ω ij ) (18) The number of candidate clusters R is obtained by screening the maximum reconstruction decision value cand ={r|γ r >ζ·max(γ)}; Among them, γ r represents the reconstruction decision value of the rth sample, ζ represents the threshold control coefficient, which usually takes a value between (0,1) and is used to control the number of candidate centers selected. The larger the value, the stricter the screening. max(γ) represents the maximum reconstruction decision value among all samples; Step 9.3, iterative optimization phase: semi-supervised constraint objective function; Defining cluster sets Where R is the number of candidate clusters obtained by density peak detection in the initial clustering stage. cand , center of mass μ r Link constraint sets Construct an improved objective function: Where f∈C r Indicates that the sample feature vector f belongs to the rth cluster, Represents a sample pair in the link constraint, two sample feature vectors f a and f b , ξ is the adaptive balance factor, Represents the indicator function, if the sample pair f a and f b If they are classified into different clusters, it is 1, otherwise it is 0. The Lagrange multiplier method is used to optimize (19) to ensure that the constrained sample pairs are forced to be classified into the same cluster; Design centroid initialization and iterative optimization strategy: Initialization: Based on R cand Generate initial centroids Iterative optimization: Calculate the new centroid of the rth cluster in the v+1th iteration in, Indicates the sample set contained in the rth cluster in the current vth iteration, Represents the sample feature vector currently belonging to the rth cluster, Indicates that it currently belongs to the link constraint set The sample pairs, and express and Tags, is the supervision strength factor, κ(·) is the Kronecker function; Step 9.4: Dynamically adjust the strategy: confidence-driven closed-loop update; In order to realize the closed-loop dynamic recognition mechanism, a cluster-level confidence evaluation function Conf(C r ) to measure the stability of each cluster, where: Where τ is the temperature coefficient, which represents the density of sample distribution within the cluster; Known Class Identification: When Conf(C L )≥θ high When θ high A high confidence threshold is used to determine whether a cluster can be considered a known class. If the samples in the cluster are highly concentrated, it can be determined to be a stable known class. The label propagation algorithm is activated to confirm the class and update the known class template library. New Class Discovery: When Conf(C L )≤θ low When θ low A low confidence threshold is used to determine whether a cluster needs to be considered as a potential new class. Clusters with dispersed structures and high uncertainty are considered as potential new classes, triggering the semi-supervised clustering mechanism to perform class classification and temporarily storing their features in the unknown class cache pool; Combining the dynamic responses of the above two strategies, the real-time update of the category labels after clustering is completed, and the identification and classification of new individuals of communication radiation sources are realized.

Images