Radiation source fingerprint identification data enhancement method based on double-constraint GAN

Abstract

The invention provides a radiation source fingerprint identification data enhancement method based on a double-constraint GAN, and relates to the technical field of communication signal processing, and the method comprises the steps: building a Rayleigh fading channel based on a Jakes model, and generating a frequency domain transmission function; constructing a double-constraint generative adversarial network, adopting a least square GAN loss function to constrain the adversarial process of a generator and a discriminator, introducing fingerprint feature matching loss constraint to generate sample feature fidelity, and outputting a superposed channel noise and fading effect by the generator; and balancing the adversarial loss and the feature matching loss until the network converges and outputs an enhanced sample by alternately optimizing parameters of the discriminator and the generator. The technical problems that in the prior art, a data training network obtained through a data enhancement algorithm under the small sample condition is low in recognition accuracy and poor in generalization ability are solved. The technical effects of improving the fingerprint identification accuracy and generalization ability of the radiation source under the small sample condition and keeping the fingerprint features are achieved.

Description

Technical Field

[0001] This invention relates to the field of communication signal processing technology, and in particular to a method for enhancing radiation source fingerprint recognition data based on dual-constraint GAN. Background Technology

[0002] Traditional data augmentation methods can only simulate idealized transmission environments and cannot reproduce real channel characteristics such as multipath fading and Doppler shift under complex electromagnetic interference scenarios, resulting in significant differences between the generated samples and the actual signal distribution.

[0003] These methods, by artificially introducing noise or changing signal parameters, can easily damage the fingerprint characteristics of the radiation source, causing fingerprint feature contamination. This makes it impossible to effectively improve the recognition performance of the generated samples when training the model.

[0004] Furthermore, existing technologies do not establish dynamic matching constraints between generated samples and real sample fingerprint features. This leads to the generator potentially overemphasizing sample realism during adversarial training while neglecting the physical rationality of fingerprint features. Consequently, the generalization ability of the generated samples under non-cooperative reception conditions is significantly reduced, failing to meet the requirements for hardware defect feature stability in individual radiation source identification.

[0005] In summary, existing technologies suffer from the technical problems of low recognition accuracy and poor generalization ability when training networks with data augmentation algorithms under small sample conditions. Summary of the Invention

[0006] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a radiation source fingerprint recognition data augmentation method based on dual-constraint GANs, which solves the technical problems of low recognition accuracy and poor generalization ability of existing technologies when data augmentation algorithms are trained on data under small sample conditions.

[0007] To achieve the above objectives, this invention provides a radiation source fingerprint recognition data augmentation method based on dual-constraint GAN, the method comprising:

[0008] Rayleigh fading channel modeling is performed based on the Jakes channel model to obtain the frequency domain channel transfer function. After acquiring the original radiation source signal, the original radiation source signal data is analytically labeled to obtain labeled radiation source signal data. The normalized labeled radiation source signal data is subjected to segmented preprocessing to obtain standardized training data. A dual-constraint generative adversarial network is pre-constructed, wherein the dual-constraint generative adversarial network uses the loss function of least squares GAN for the adversarial loss of the generator and discriminator, and uses fingerprint feature matching loss as the feature fidelity constraint of the generator. When the generator generates samples, the noise and fading effects simulated by the frequency domain channel transfer function are superimposed. The standardized training data is input into the dual-constraint generative adversarial network, and during the alternating optimization process of the generator and discriminator, the constraints are balanced based on the adversarial loss and fingerprint feature matching loss until the network converges and the generated samples are output.

[0009] In one implementation, the following processing is also performed:

[0010] The generator is pre-embedded with a fingerprint defect knowledge base. During the alternating optimization process of the generator and the discriminator, the fingerprint feature offset is calculated in real time on the original generated signal output by the generator to obtain a fingerprint feature offset vector. Based on the fingerprint feature offset vector, a compensation template is targetedly retrieved from the fingerprint defect knowledge base to perform local feature-targeted repair of the original generated signal, resulting in a repaired generated signal. The repaired generated signal is input into the discriminator, and after calculating the discrimination loss based on the repaired generated signal, the discriminator parameters are updated based on the discrimination loss.

[0011] In one implementation, the following processing is also performed:

[0012] The discriminator introduces a defect state transition prediction matrix; after receiving the repair generation signal sent by the generator, the discriminator predicts the defect state transition trajectory based on the defect state transition prediction matrix to obtain an advance compensation instruction; the advance compensation instruction is used to dynamically relax the constraint weights of the discriminator on the fading defects, and to optimize the compensation template for the next round of generator training.

[0013] In one implementation, Rayleigh fading channel modeling is performed based on the Jakes channel model to obtain the frequency domain channel transfer function, and the following processing is also performed:

[0014] Predefine the number of propagation paths, maximum Doppler shift, and maximum frequency spread; model the in-phase component of the Kth propagation path:

[0015]

[0016] Orthogonal component modeling is performed on the Kth propagation path:

[0017]

[0018] in, ω d φ is the maximum Doppler frequency shift on the Kth propagation path. nk Let M be the number of propagation paths, uniformly distributed in the range [0, 2π). Similarly, construct M in-phase components and M quadrature components for each of the M propagation paths. By mapping and superimposing these M in-phase and M quadrature components, synthesize the time-domain channel impulse response function: h(t) = x c (t)+j·x s (t); where x c (t) represents the superposition result of the M in-phase components, x s (t) is the superposition result of the M orthogonal components; perform Fourier transform on the time-domain channel impulse response function to output the frequency-domain channel transfer function: H(f)=F{h(t)}.

[0019] In one implementation, after the original radiation source signal is acquired, the original radiation source signal data is analyzed and labeled to obtain labeled radiation source signal data, and the following processing is also performed:

[0020] The original radiation source signal data is demodulated to output a complex signal; the radiation source identification label is marked according to the nonlinear distortion characteristics of the power amplifier in the original radiation source signal data; the channel parameters of the original radiation source signal data are collected; the original radiation source signal data is marked with the channel parameters, the radiation source identification label and the complex signal to obtain the marked radiation source signal data.

[0021] In one implementation, the discriminant loss function of the discriminant is constructed as follows:

[0022]

[0023] Where, p data For the true sample distribution, p z Let G(z,c) be the noise distribution, and G(z,c) be the sample output by the generator.

[0024] In one implementation, the following processing is also performed:

[0025] The adversarial loss function for the generator is constructed as follows:

[0026]

[0027] The fingerprint feature matching loss function is constructed as follows:

[0028]

[0029] The generator's total loss function is constructed as follows:

[0030]

[0031] Where λ is a preset hyperparameter used to balance adversarial loss and feature matching loss.

[0032] In one implementation, the standardized training data is input into the dual-constraint generative adversarial network. During the alternating optimization of the generator and discriminator, constraint balancing is performed based on the adversarial loss and fingerprint feature matching loss until the network converges and generates samples. The following processing is also performed:

[0033] Step A: After fixing the model parameters of the generator, update the model parameters of the discriminator based on the discriminator loss function; Step B: After unfixing the model parameters of the generator and fixing the model parameters of the discriminator, update the model parameters of the generator based on the generator total loss function; By repeatedly executing steps A to B until the discriminator loss function and the generator total loss function converge, output the generated sample; Superimpose the noise and fading effects simulated by the frequency domain channel transfer function onto the generated sample, and output the updated sample.

[0034] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0035] The method provided in this invention constructs a Rayleigh fading channel based on the Jakes model to generate a frequency domain transfer function; it collects and labels the original radiation source signal, and forms standardized training data after normalization and segmentation; it constructs a dual-constraint generative adversarial network, using a least-squares GAN loss function to constrain the adversarial process between the generator and discriminator, and introduces fingerprint feature matching loss to constrain the fidelity of generated sample features. The generator output is superimposed with channel noise and fading effects; by alternately optimizing the discriminator and generator parameters, the adversarial loss and feature matching loss are balanced until the network converges, and enhanced samples are output. This achieves the technical effect of improving the accuracy and generalization ability of radiation source fingerprint recognition under small sample conditions, preserving fingerprint features, and adapting to complex transmission environments. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 The diagram shows a flowchart of the radiation source fingerprint recognition data enhancement method based on dual-constraint GAN provided by the present invention.

[0038] Figure 2 The diagram illustrates the process of obtaining labeled radiation source signal data in the radiation source fingerprinting data enhancement method based on dual-constraint GAN provided by the present invention. Detailed Implementation

[0039] This invention provides a radiation source fingerprint recognition data augmentation method based on dual-constraint GAN, which solves the technical problems of low recognition accuracy and poor generalization ability of existing data augmentation algorithms under small sample conditions.

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0041] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0042] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0043] The flowchart of the radiation source fingerprint recognition data enhancement method based on dual-constraint GAN provided in this embodiment of the invention is shown below. Figure 1 The method includes:

[0044] A100: Rayleigh fading channel modeling based on the Jakes channel model yields the frequency domain channel transmission function.

[0045] In one implementation, Rayleigh fading channel modeling is performed based on the Jakes channel model to obtain the frequency domain channel transfer function. The method step A100 provided by this invention includes:

[0046] A110: Predefined number of propagation paths, maximum Doppler shift, and maximum frequency spread.

[0047] A120: Model the in-phase component of the Kth propagation path:

[0048]

[0049] A130: Model the orthogonal components of the Kth propagation path:

[0050]

[0051] in, ω d φ is the maximum Doppler frequency shift on the Kth propagation path. nk M represents a random phase shift uniformly distributed in [0, 2π), and M is the number of propagation paths.

[0052] A140: By analogy, construct M in-phase components and M quadrature components for M propagation paths.

[0053] A150: By mapping and superimposing the M in-phase components and M quadrature components, the time-domain channel impulse response function is synthesized:

[0054] h(t) = x c (t)+j·x s (t);

[0055] Where, x c (t) represents the superposition result of the M in-phase components, x s (t) represents the superposition result of the M orthogonal components.

[0056] A160: Perform a Fourier transform on the time-domain channel impulse response function to output the frequency-domain channel transfer function: H(f)=F{h(t)}.

[0057] Specifically, in this embodiment, the number of propagation paths is predefined as M=3, the maximum Doppler frequency shift is 2Hz, and the maximum frequency extension is 10Hz.

[0058] Since the method for modeling the in-phase and quadrature components of each propagation path is consistent, this embodiment takes the modeling of the in-phase and quadrature components of the Kth propagation path as an example to elaborate on the technical solution in detail.

[0059] Model the in-phase component of the Kth propagation path:

[0060]

[0061] in,

[0062] Orthogonal component modeling is performed on the Kth propagation path:

[0063]

[0064] Where c represents the in-phase component, s represents the quadrature component, t is the time variable, n is the sub-path index, n = 1, 2, ..., M, α nk Let θ be the arrival angle of the nth sub-path of the k-th path, and θ nk For random angle offset, ω d φ is the maximum Doppler frequency shift on the Kth propagation path. nk M represents a random phase shift uniformly distributed in [0, 2π), and M is the number of propagation paths.

[0065] By analogy, construct M in-phase components and M quadrature components corresponding to the number of propagation paths for the M propagation paths.

[0066] By mapping and superimposing the M in-phase components and M quadrature components, the time-domain channel impulse response function is synthesized: h(t) = x c (t)+j·x s (t).

[0067] Where, x c (t) represents the superposition result of the M in-phase components, x s (t) represents the superposition result of the M orthogonal components. The imaginary unit j is the core mathematical tool for constructing complex signals, enabling the model to simultaneously describe the amplitude attenuation caused by multipath effects and the phase shift caused by the Doppler effect. The amplitude attenuation is determined by the real part x. c (t) carries, and the phase shift is determined by the imaginary part x. s (t) Bearing.

[0068] Perform a Fourier transform on the time-domain channel impulse response function to output the frequency-domain channel transfer function: H(f)=F{h(t)}, which is used for subsequent superposition of channel effects in the frequency domain.

[0069] A200: After the original radiation source signal is acquired, the original radiation source signal data is analyzed and labeled to obtain labeled radiation source signal data;

[0070] In one implementation, see Figure 2 After acquiring the original radiation source signal, the original radiation source signal data is analyzed and labeled to obtain labeled radiation source signal data. The method step A200 provided by this invention includes:

[0071] A210: Perform complex signal demodulation on the original radiation source signal data to output a complex signal.

[0072] A220: Label the radiation source with a nonlinear distortion characteristic of the power amplifier in the original radiation source signal data.

[0073] A230: Channel parameters for acquiring the original radiation source signal data.

[0074] A240: The original radiation source signal data is labeled using the channel parameters, radiation source identification tags, and complex signals to obtain the labeled radiation source signal data.

[0075] Specifically, in this embodiment, the original radiation source signal (I / Q signal) is converted into a complex form through quadrature demodulation, which fully preserves the amplitude, phase and frequency information of the signal. The complex signal is composed of real (In-phase) and imaginary (Quadrature) components, which are the basic data carrier for subsequent fingerprint feature extraction.

[0076] The nonlinear distortion, including amplitude / phase distortion, generated by the Taylor polynomial response of the power amplifier is defined as the core hardware fingerprint of the radiation source. Based on the specificity of the power amplifier response, each signal is labeled with a unique identifier to distinguish different radiation sources, thus obtaining the labeled radiation source identifier.

[0077] Record the channel parameters during signal acquisition, including physical layer transmission characteristic parameters such as the number of propagation paths, maximum Doppler shift, and maximum frequency spread, for subsequent channel modeling and conditional constraints for generative adversarial networks.

[0078] Using complex signals as the main data subject, a structured dataset is generated by combining radiation source identification labels and channel parameters to form labeled samples that simultaneously carry signal waveforms, hardware fingerprints, and transmission scenario characteristics, thus obtaining the labeled radiation source signal data.

[0079] A300: Perform segmented preprocessing on the normalized labeled radiation source signal data to obtain standardized training data.

[0080] The continuous radiation source signal is divided into equal-length segments of fixed length to meet the input dimension requirements of generative adversarial networks. During the segmentation process, it is necessary to ensure that each segment contains complete signal periodic characteristics, such as carrier period and modulation symbols, while eliminating invalid signal segments such as silent periods or strong noise interference segments.

[0081] The segmented signal fragments are standardized to eliminate dimensional differences, ultimately generating standardized data samples that meet the training requirements of dual-constraint generative adversarial networks, ensuring that the network can stably extract radiation source fingerprint features and channel transmission characteristics from the segmented signals.

[0082] A400: A pre-constructed dual-constraint generative adversarial network, wherein the dual-constraint generative adversarial network uses the loss function of least squares GAN for the adversarial loss of the generator and the fingerprint feature matching loss as the feature fidelity constraint of the generator, and the noise and fading effects simulated by the frequency domain channel transfer function are superimposed when the generator generates samples.

[0083] In one implementation, the discriminator loss function for constructing the discriminator is as follows:

[0084]

[0085] Where, p data For the true sample distribution, p z Let G(z,c) be the noise distribution, and G(z,c) be the sample output by the generator.

[0086] In one implementation, the method further includes:

[0087] A431: The adversarial loss function for the generator is constructed as follows:

[0088] A432: The fingerprint feature matching loss function is constructed as follows:

[0089] A433: The generator's total loss function is constructed as follows:

[0090] Where λ is a preset hyperparameter used to balance adversarial loss and feature matching loss.

[0091] Specifically, in this embodiment, the dual-constraint generative adversarial network ensures the authenticity and task adaptability of the generated data through a dual-constraint mechanism. The generator receives a noise vector z and a condition parameter c, where the condition parameter c includes the radiation source identification label and channel parameters, and outputs an initial generated sample G(z,c). To simulate a real channel environment, the generated sample needs to be superimposed with noise and fading effects simulated by the channel transfer function H(f), ensuring that the generated sample retains both the fingerprint characteristics of the radiation source, such as the nonlinear distortion of the power amplifier, and reflects the multipath fading and Doppler effects of the wireless channel, such as amplitude time-varying and frequency offset.

[0092] The discriminator employs Least Squares Adversarial Loss (LSGAN), and its loss function is defined as:

[0093]

[0094] p data For the true sample distribution, p zLet G(z,c) represent the noise distribution, G(z,c) be the sample output from the generator, and 0.5 be the optimal balance coefficient. The function is used to calculate the expected value of the distribution of real samples, where x represents the sampled from the real dataset, i.e., the labeled radiation source signal data. This function replaces the cross-entropy loss of traditional GANs and aims to solve the mode collapse problem.

[0095] D(x,c) is the discriminator's output for the input real sample x and the conditional parameter c. The goal is to output a value close to 1 so that the sample is considered a real sample.

[0096] (D(x,c)-1) 2 For the loss term of the real samples, calculate the squared error between the discriminator output and the target value of 1, and drive the discriminator's output for the real samples to approach 1. To determine the noise distribution p z The expected value is calculated, where z is a random noise vector sampled from the noise distribution.

[0097] G(z,c) is the fake sample generated by the generator based on the noise z and the conditional parameter c. D(G(z,c),c) is the output of the discriminator on the generated sample G(z,c) and the conditional parameter c. The goal is to have an output close to 0 so that it is judged as a fake sample.

[0098] (D(G(z,c),c)) 2 To generate the loss term for the sample, the squared error between the discriminator output and the target value of 0 is calculated, driving the discriminator to approach 0 for the output of the generated sample.

[0099] This function requires the discriminator to output close to 1 for real samples and close to 0 for generated samples, thus avoiding the mode collapse problem and improving the generalization of generated samples.

[0100] Construct the adversarial loss function for the generator:

[0101] in, It is the expected value calculation, which means sampling and averaging the random noise vector z in the noise distribution, p z The input noise distribution for the generator is used to generate diverse fake samples.

[0102] The fingerprint feature matching loss function is constructed as follows:

[0103] The generator's total loss is weighted and integrated using the hyperparameter λ to overcome the dual constraints. The specific generator total loss function is constructed as follows:

[0104] Here, λ is used to balance the adversarial loss and the feature matching loss.

[0105] Building upon this foundation, this embodiment pre-embeds a fingerprint defect knowledge base in the generator. This allows typical hardware distortion patterns to be incorporated as prior knowledge into the generation process. After the generator outputs the original generated signal, it calculates the fingerprint feature offset in real time, such as the difference between amplitude / phase distortion and standard fingerprints. Local features are then repaired by targeting compensation templates in the knowledge base, generating a repaired signal. This process ensures that the generated sample strictly conforms to the physical defect characteristics of the radiation source hardware, avoiding fingerprint feature drift during the generation process.

[0106] Building upon this foundation, this embodiment introduces a defect state transition prediction matrix into the discriminator, constructing a defect evolution trajectory based on historical training data, such as a time series of feature offsets. Upon receiving a repair signal, the discriminator uses the matrix to predict the defect state transition direction within the next k steps, generating a forward compensation command.

[0107] This instruction dynamically relaxes the constraint weights on fading defects, optimizes the compensation template selection strategy for the next round of generator training, and forms a closed-loop feedback loop of defect repair, state prediction, and constraint optimization. This improves the discriminator's adaptability to complex defect states and enhances the physical plausibility of the generated samples.

[0108] The optimization process of the dual-constraint generative adversarial network will be explained in the following description of this embodiment.

[0109] A500: The standardized training data is input into the dual-constraint generative adversarial network. During the alternating optimization of the generator and discriminator, constraint balancing is performed based on the adversarial loss and fingerprint feature matching loss until the network converges and generates samples.

[0110] In one implementation, the standardized training data is input into the dual-constraint generative adversarial network. During the alternating optimization of the generator and discriminator, constraint balancing is performed based on the adversarial loss and fingerprint feature matching loss until the network converges and generates samples. The method steps A500 provided by this invention include:

[0111] Step A: After fixing the model parameters of the generator, update the model parameters of the discriminator based on the discriminator loss function.

[0112] Step B: After removing the model parameters of the generator and fixing the model parameters of the discriminator, update the model parameters of the generator based on the total loss function of the generator.

[0113] By repeatedly executing steps A to B until the discriminator loss function and the generator total loss function converge, the generated sample is output.

[0114] The noise and fading effects simulated by the frequency domain channel transfer function are superimposed on the generated samples to output updated samples.

[0115] In one implementation, the method further includes:

[0116] A5011: The generator pre-embeds a fingerprint defect knowledge base.

[0117] A5012: During the alternating optimization process of the generator and the discriminator, the fingerprint feature offset is calculated in real time on the original generated signal output by the generator to obtain the fingerprint feature offset vector.

[0118] A5013: Based on the fingerprint feature offset vector, a compensation template is targetedly retrieved from the fingerprint defect knowledge base, and the local features of the original generated signal are targeted for repair to obtain the repaired generated signal.

[0119] A5014: Input the repaired signal into the discriminator, calculate the discrimination loss based on the repaired signal, and update the discriminator parameters according to the discrimination loss.

[0120] In one implementation, the method further includes:

[0121] A5021: Introduce a defect state transition prediction matrix into the discriminator.

[0122] A5022: After receiving the repair generation signal sent by the generator, the discriminator performs defect state migration trajectory prediction based on the defect state migration prediction matrix to obtain the advance compensation instruction.

[0123] A5023: The discriminator's constraint weights on the fading defects are dynamically relaxed using the aforementioned advance compensation instruction, and the compensation template for the next round of generator training is optimized.

[0124] Specifically, during the training process of the dual-constraint generative adversarial network, the discriminator and the generator gradually converge through an alternating optimization strategy to generate samples that both conform to the real data distribution and retain the fingerprint characteristics of the radiation source.

[0125] First, after fixing the generator's parameters, the discriminator updates its own parameters using Least Squares Adversarial Loss (LSGAN). The discriminator aims to distinguish between real and generated samples, and its loss function is designed to output close to 1 for real samples and close to 0 for generated samples. By fixing the generator, the discriminator adjusts its parameters only based on the features of the currently generated samples, thereby improving its ability to distinguish between the two types of samples. This process drives the discriminator to more accurately identify the authenticity of samples by calculating the squared error between real and generated samples, while avoiding the pattern collapse problem caused by excessively strong adversarial loss for generated samples.

[0126] After the generator outputs the original generated signal, the fingerprint feature offset vector needs to be calculated in real time. This offset is obtained by comparing the difference between the generated sample and the standard fingerprint features, reflecting the degree of deviation of the generated signal in fingerprint features.

[0127] This calculation relies on the output of the discriminator feature extraction layer. The difference between the feature vectors of the generated sample and the real sample is quantified by the L1 norm. The real-time nature of the offset vector ensures that the generator can dynamically perceive feature deviations and provide a basis for subsequent repair.

[0128] Based on the offset vector, the generator targets and retrieves a compensation template from the fingerprint defect knowledge base to repair local features of the original generated signal. The fingerprint defect knowledge base stores typical hardware distortion patterns, and the generator retrieves the closest compensation template by matching the offset vector with the defect patterns in the knowledge base.

[0129] The compensation template contains correction parameters for specific defects, which are applied to local features of the generated sample, such as amplitude or phase distortion regions, to generate a repaired signal. This process ensures that the generated sample strictly conforms to the physical defect characteristics of the radiation source hardware, avoiding fingerprint feature drift.

[0130] The repaired generated signal is used as input to the discriminator. By recalculating the discrimination loss and updating the discriminator parameters, the repaired signal must satisfy the condition that the discriminator output approaches 0 and the loss term approaches 0. The discriminator parameters are adjusted through gradient descent to enhance its ability to distinguish the repaired generated samples. This step allows the discriminator to adapt to the features repaired by the generator, forming dynamic feedback and promoting a balance between the authenticity and fingerprint fidelity of the generated samples.

[0131] After the generator parameters are unfixed, the generator updates its own parameters based on the total loss function. The generator's total loss is composed of a weighted sum of adversarial loss and fingerprint feature matching loss. Specifically, the adversarial loss drives the generated samples to be identified as real, and the fingerprint feature matching loss constrains the generated samples to be consistent with the feature vectors of real samples through L1 norm.

[0132] The hyperparameter λ is used to balance the weights of both, preventing a single constraint from dominating and causing sample distortion or feature loss. The generator adjusts the parameters through backpropagation to generate samples that are both realistic and retain key features.

[0133] After receiving the repair-generated signal, the discriminator predicts the defect state migration trajectory within the next k steps based on the defect state migration prediction matrix. This defect state migration matrix is ​​constructed using historical training data and records the time-series evolution of fingerprint feature offsets.

[0134] Based on the prediction results, advance compensation instructions are generated to guide the generator in dynamically adjusting the selection strategy of the compensation template in the next round of training. For example, if it is predicted that a certain type of defect will gradually decay, the generator can reduce the repair intensity for that defect to avoid feature distortion caused by overcompensation.

[0135] The advance compensation instruction optimizes the selection of compensation templates for the next round of generator training by dynamically relaxing the constraint weights of the discriminator on fading defects. The relaxation mechanism adjusts the constraint strength of the discriminator on specific defects based on the prediction results, such as reducing the weight of defects that are about to fade, so that the generator can allocate repair resources more efficiently. This closed-loop feedback mechanism improves the physical rationality of the generated samples and ensures the adaptability of the generator under complex defect conditions.

[0136] By iteratively optimizing the discriminator and generator, until their loss functions converge, specifically, discriminator loss convergence indicates that its ability to distinguish between real and generated samples has reached a balance, while generator loss convergence indicates that the generated sample has reached the optimal balance between realism and fingerprint fidelity.

[0137] The final output samples need to be superimposed with Rayleigh fading and Doppler effects simulated by the Jakes channel model. Noise and fading effects are superimposed through the frequency domain channel transfer function to make it closer to the actual wireless transmission scenario.

[0138] This process enhances the generalization ability of generated samples under non-cooperative reception conditions by introducing the multipath fading coefficient and Doppler frequency shift in the channel transfer function, thus providing high-quality training data for radiation source fingerprinting networks.

[0139] This embodiment achieves the technical effect of improving the accuracy and generalization ability of radiation source fingerprint recognition under small sample conditions, preserving fingerprint features and adapting to complex transmission environments.

[0140] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A radiation source fingerprint recognition data augmentation method based on dual-constraint GAN, characterized in that, include: Rayleigh fading channel modeling is performed based on the Jakes channel model to obtain the frequency domain channel transfer function; After the original radiation source signal is acquired, the original radiation source signal data is analyzed and labeled to obtain labeled radiation source signal data. The normalized labeled radiation source signal data is subjected to segmented preprocessing to obtain standardized training data. A pre-constructed dual-constraint generative adversarial network is provided, wherein the dual-constraint generative adversarial network uses the loss function of least squares GAN for the adversarial loss of the generator and the fingerprint feature matching loss is used as the feature fidelity constraint of the generator, and the noise and fading effects simulated by the frequency domain channel transfer function are superimposed when the generator generates samples. The standardized training data is input into the dual-constraint generative adversarial network. During the alternating optimization of the generator and discriminator, constraint balancing is performed based on the adversarial loss and fingerprint feature matching loss until the network converges and generates samples.

2. The radiation source fingerprint recognition data augmentation method based on dual-constraint GAN as described in claim 1, characterized in that, Also includes: The generator is pre-embedded with a fingerprint defect knowledge base; During the alternating optimization process of the generator and the discriminator, the fingerprint feature offset is calculated in real time on the original generated signal output by the generator to obtain the fingerprint feature offset vector; Based on the fingerprint feature offset vector, a compensation template is targetedly retrieved from the fingerprint defect knowledge base, and the local features of the original generated signal are targeted for repair to obtain the repaired generated signal. The repaired signal is input into the discriminator. After calculating the discrimination loss based on the repaired signal, the discriminator parameters are updated according to the discrimination loss.

3. The radiation source fingerprint recognition data augmentation method based on dual-constraint GAN as described in claim 2, characterized in that, Also includes: A defect state transition prediction matrix is ​​introduced into the discriminator; After receiving the repair generation signal sent by the generator, the discriminator predicts the defect state migration trajectory based on the defect state migration prediction matrix to obtain the advance compensation instruction. The discriminator's constraint weights on fading defects are dynamically relaxed using the aforementioned advance compensation instruction, and the compensation template for the next round of generator training is optimized.

4. The radiation source fingerprint recognition data augmentation method based on dual-constraint GAN as described in claim 1, characterized in that, Rayleigh fading channel modeling based on the Jakes channel model yields the frequency domain channel transfer function, including: Predefined number of propagation paths, maximum Doppler shift, and maximum frequency spread; Model the in-phase component of the Kth propagation path: Orthogonal component modeling is performed on the Kth propagation path: in, ω d φ is the maximum Doppler frequency shift on the Kth propagation path. nk The random phase shift is uniformly distributed in [0, 2π), and M is the number of propagation paths; By analogy, construct M in-phase components and M quadrature components for M propagation paths; By mapping and superimposing the M in-phase components and M quadrature components, the time-domain channel impulse response function is synthesized: h(t)=x c (t)+j·x s (t); Where, x c (t) represents the superposition result of the M in-phase components, x s (t) represents the superposition result of the M orthogonal components; Perform a Fourier transform on the time-domain channel impulse response function to output the frequency-domain channel transfer function: H(f)=F{h(t)}.

5. The radiation source fingerprint recognition data augmentation method based on dual-constraint GAN as described in claim 1, characterized in that, After the raw radiation source signal is acquired, the raw radiation source signal data is analyzed and annotated to obtain annotated radiation source signal data, including: The original radiation source signal data is demodulated to output a complex signal. The radiation source identification label is marked according to the nonlinear distortion characteristics of the power amplifier in the original radiation source signal data; Channel parameters for acquiring the original radiation source signal data; The original radiation source signal data is labeled using the channel parameters, radiation source identification tags, and complex signals to obtain the labeled radiation source signal data.

6. The radiation source fingerprint recognition data augmentation method based on dual-constraint GAN as described in claim 1, characterized in that, The discriminator loss function for constructing the discriminator is as follows: Where, p data For the true sample distribution, p z Let G(z,c) be the noise distribution, and G(z,c) be the sample output by the generator.

7. The radiation source fingerprint recognition data augmentation method based on dual-constraint GAN as described in claim 6, characterized in that, Also includes: The adversarial loss function for the generator is constructed as follows: The fingerprint feature matching loss function is constructed as follows: The generator's total loss function is constructed as follows: Where λ is a preset hyperparameter used to balance adversarial loss and feature matching loss.

8. The radiation source fingerprint recognition data augmentation method based on dual-constraint GAN as described in claim 7, characterized in that, The standardized training data is input into the dual-constraint generative adversarial network (GAN). During the alternating optimization of the generator and discriminator, constraint balancing is performed based on the adversarial loss and fingerprint feature matching loss until the network converges, outputting generated samples, including: Step A: After fixing the model parameters of the generator, update the model parameters of the discriminator based on the discriminator loss function; Step B: After removing the model parameters of the generator and fixing the model parameters of the discriminator, update the model parameters of the generator based on the total loss function of the generator; By repeatedly executing steps A to B until the discriminator loss function and the generator total loss function converge, the generated sample is output. The noise and fading effects simulated by the frequency domain channel transfer function are superimposed on the generated samples to output updated samples.

Images