A method for identifying a radiation source based on a phase offset constellation diagram

Abstract

The application discloses a radiation source identification method based on a phase offset constellation diagram and relates to the field of radio monitoring.The application provides a radiation source feature extraction method based on a phase offset constellation diagram, which can remove most of the intentional modulation features, realize the radiation source identity identification task under various modulation modes, and solve the technical pain point that current traditional methods are based on one or one kind of modulation mode to realize the radiation source individual identification, and the average accuracy under 0dB signal-to-noise ratio reaches 95.26%. The application also provides a direct supervision contrast network (D-SupCon) model, which can better compare the radiation source features and perform classification. Under a QPSK signal data set, the D-SupCon shows better radiation source classification performance than SupCon network and traditional network. Finally, the application further verifies the effectiveness of the method provided by the application in an actual environment by using measured signals generated by a communication system of a radio device.

Description

Technical Field

[0001] This invention relates to the technical field of radio monitoring, and in particular to a method for identifying radiation sources based on phase-biased constellation diagrams. Background Technology

[0002] With the increasing prevalence of wireless communication devices and the rapid development of software-defined radio technology, wireless communication has become an indispensable part of modern society, widely used in both military and civilian fields. However, while wireless communication frees people from spatial constraints, it also brings many security challenges. Due to its openness and broadcast nature, wireless communication is more vulnerable to malicious attacks than wired communication, making its security a growing concern. Currently, the security of wireless communication networks typically relies on authentication and cryptographic techniques. Once malicious attacks occur or passwords are leaked, communication security is severely threatened. Therefore, existing wireless communication networks urgently need a new security mechanism to enhance communication security.

[0003] Currently, existing technologies have achieved good results in the field of steady-state-based individual radiation source identification. However, all existing technologies are based on signals with one or a class of modulation methods to achieve individual radiation source identification. Traditional methods such as constellation trajectory diagrams and bispectral methods will retain some intentional modulation information, which will reduce the effectiveness of radiation source identification. When signals with new modulation methods appear, the existing radiation source identification methods are extremely ineffective. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a radiation source identification method based on phase-biased constellation diagrams, with the following specific steps:

[0005] S1. Remove the influence of intentional modulation in the signal on radiation source identification; First, remove the modulated baseband information in the received IQ signal. In the phase offset constellation diagram method used in this paper, the baseband information in IQ is eliminated by obtaining the square envelope of the signal, and the signal is subjected to constant removal and normalization.

[0006] S2. Further denoising processing of the IQ signal is required. Since the frequency offset lf varies greatly depending on the individual radiation source and is generally unknown, low-pass filtering is not suitable. This invention uses autocorrelation filtering to extract radiation source features. After the above processing, the signal r is obtained. IQ ;

[0007] S3, Calculate signal r IQ The phase difference, where n takes three values ​​f0 / 10, f0 / 100, and f0 / 1000, yields D1(t), D2(t), and D3(t). nThe signal (t)(n=1,2,3) carries only the parameter λ offset from the carrier frequency and the phase noise j(t), which can completely remove the intentional modulation features in the signal and retain the unintentional modulation features.

[0008] S4. Construct a phase-biased constellation diagram dataset based on D1(t), D2(t), and D3(t); the specific steps include:

[0009] Calculate D n The real and imaginary parts of (t)(n=1,2,3) are separated and converted into IQ signals, and normalization is achieved by removing points with values ​​greater than 1 and less than -1; then D is calculated. n The density distribution of (t)(n=1,2,3) on the phase-biased constellation diagram was obtained, and finally the matrix G was converted into an image to form the phase-biased constellation diagram dataset.

[0010] S5. In the training of the feature extraction part of the D-SupCon network, the parameter settings were optimized. In terms of network structure, the 128D features output by the feature extraction network were directly used for classification, which can effectively preserve the distribution of the feature space and achieve better results. At the same time, since the input data is a phase-biased constellation map with a large image size, multiple strided convolutions were introduced to reduce the amount of computation. Since the horizontal and vertical axes of the phase-biased constellation map are IQ values, the spatial position of the image has practical significance, so data augmentation was canceled.

[0011] S6. Using the trained network structure, the signal to be identified is converted into the corresponding phase-biased constellation diagram and input into the network to obtain the final classification output result, thereby realizing the identification of the radiation source.

[0012] Compared with the prior art, the beneficial effects of the present invention using the above technical solution are as follows:

[0013] This invention proposes a radiation source feature extraction method based on phase-biased constellation diagrams. This method can remove most intentional modulation features, enabling radiation source identification under various modulation schemes and addressing the technical limitations of current traditional methods. It achieves an average accuracy of 95.26% at a 0dB signal-to-noise ratio. Furthermore, a Directly Supervised Contrast Network (D-SupCon) model is proposed, which can effectively compare and classify radiation source features. On a QPSK signal dataset, D-SupCon demonstrates superior radiation source classification performance compared to both the SupCon network and traditional networks. Finally, measured signals generated by a communication system built with wireless equipment are used to further verify the effectiveness of the proposed method in real-world environments. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the individual radiation source identification method based on phase-biased constellation diagram of the present invention;

[0015] Figure 2 This is the phase-biased constellation diagram of the QPSK of the present invention;

[0016] Figure 3 The phase-biased constellation diagrams for the 12 modulation schemes of this invention are shown below.

[0017] Figure 4 These are photographs of the transmitting and receiving devices of the present invention;

[0018] Figure 5 This is a sub-figure comparing the radiation source signal recognition performance under different modulation methods of the present invention;

[0019] Figure 6 This is a comparison chart of the radiation source signal recognition performance under different modulation methods according to the present invention;

[0020] Figure 7 This is a confusion matrix diagram for radiation source identification based on phase offset trajectory diagrams under different modulation modes of the present invention;

[0021] Figure 8 The confusion matrix diagram for radiation source identification based on constellation trajectory diagram under different modulation modes of the present invention.

[0022] Figure 9 This is a confusion matrix diagram for bispectral radiation source identification under different modulation modes of the present invention;

[0023] Figure 10 The diagram shows the radiation source identification accuracy of the three methods under different modulation modes of the present invention.

[0024] Figure 11 This is a diagram of the SupCon network structure of the present invention;

[0025] Figure 12 This is a phase offset constellation diagram of the hackRF one radio frequency signals numbered 1 and 2 of this invention. Detailed Implementation

[0026] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0027] According to embodiments of the present invention, a radiation source identification method based on phase-biased constellation diagrams and its usage method are provided, such as... Figure 1 The diagram shown is a schematic flowchart of a radiation source individual identification method based on a phase-biased constellation diagram provided by an embodiment of the present invention. The method includes:

[0028] S1. Remove the influence of intentional modulation in the signal on radiation source identification. First, remove the modulated baseband information in the received IQ signal. In this paper, the baseband information in IQ is eliminated by calculating the square envelope of the signal in the phase offset constellation diagram method, and signal constant removal and normalization are performed.

[0029] In this step, it is assumed that the radiation source is affected by phase noise and frequency shift at the transmitting end, and by Gaussian white noise during channel transmission. Taking a QPSK modulated signal as an example, the received signal can be expressed as:

[0030] r￠(t)=I￠·cos(2pft)+Q￠·sin(2pft)

[0031] Where I￠ and Q￠ can be represented as:

[0032]

[0033]

[0034] In the above formula, the phase noise signal is added first, followed by the carrier frequency offset. In the IQ signal, the phase noise is modulated to the offset frequency lf, which is consistent with reality. To remove the influence of intentional modulation on radiation source identification, the modulated baseband information in the IQ signal should be removed first. The phase offset constellation diagram eliminates the baseband information in the IQ signal by calculating the squared envelope of the signal. The signal after the squared envelope can be expressed as:

[0035]

[0036] For the formula, since j can be ignored 2 (t), therefore:

[0037]

[0038] It can be simplified to:

[0039]

[0040] in The value is 1 when the signs of I and Q are the same, and -1 when the signs of I and Q are different.

[0041] c(t) = v 2 (t)+2v(t), it should be noted that since v(t) is Gaussian white noise, its power spectrum conforms to a uniform distribution, so it will not have a significant impact on the distribution of j(t).

[0042] Similarly, Q ￠￠ for:

[0043]

[0044] As can be seen from the formula, the IQ signal no longer carries baseband information; however, this method leads to an increase in the intensity of Gaussian white noise when it is present. To better analyze the characteristics of unintentional modulation, the next step is to denoise the IQ signal.

[0045] S2. Further denoising processing of the IQ signal is required. Since the frequency offset lf varies greatly depending on the individual radiation source and is generally unknown, low-pass filtering is not suitable. This invention uses autocorrelation filtering to extract radiation source features. After the above processing, the obtained signal r IQ ;

[0046] Since the frequency offset (lf) varies greatly depending on the individual radiation source and is generally unknown, low-pass filtering is not applicable. This invention uses autocorrelation filtering to extract radiation source features.

[0047] When calculating the autocorrelation function of a discrete sequence of length N, data with n < 0 and n > N cannot be measured, inevitably leading to truncation errors. Therefore, only the autocorrelation function of the sequence can be estimated, which can be expressed as:

[0048]

[0049] The formula is an unbiased autocorrelation estimate for a discrete sequence. When |m| ≤ N-1, we have:

[0050] r xx -E(r xx (m))=0

[0051] After removing constants and normalizing the squared envelope IQ signal, the autocorrelation is estimated. Since Gaussian white noise has cov(c(t),c(t+1))=0 for any k≥1, the noise c(t) is removed. Therefore, the autocorrelation of the IQ signal can be expressed as:

[0052]

[0053] The default values ​​are:

[0054] cov(c(n),I(n+m))=0

[0055] cov(c(n),Q(n+m))=0

[0056] The phase noise j(t) is generally composed of white phase noise, scintillation phase noise, random phase walk noise, scintillation frequency noise, and random frequency walk noise. Among them, random walk noise and scintillation noise are both low-frequency noises. The characteristics of these phase noises will be preserved in the process of calculating autocorrelation. Therefore, calculating the autocorrelation function will not destroy the characteristics of unintentional modulation of the radiation source, thus preventing a decrease in recognition accuracy.

[0057] The receiver outputs a baseband signal with a sampling rate of f0 after passing through an RF front-end, digital down-conversion, and multi-rate data processing. The signal r obtained after the above processing is... IQ .

[0058] S3, Calculate signal r IQ The phase difference, where n takes three values: f0 / 10, f0 / 100, and f0 / 1000, is obtained.

[0059] D1(t), D2(t), D3(t), in D n The signal (t)(n=1,2,3) carries only the parameter λ offset from the carrier frequency and the phase noise j(t), which can completely remove the intentional modulation features in the signal and retain the unintentional modulation features.

[0060] The formula for calculating the phase difference is:

[0061]

[0062] In the formula, H represents the Hilbert transform, since the signal r IQ Calculating the autocorrelation only yields a sinusoidal signal of a single frequency. However, it's difficult to determine the phase difference of the signal using only a sinusoidal signal. Therefore, it's necessary to convert r... IQ Convert to a complex signal. The formula for the Hilbert transform is:

[0063]

[0064] For typical radiation sources, the frequency offset is between a few Hz and several thousand Hz. Therefore, for the formula, n takes three values: f0 / 10, f0 / 100, and f0 / 1000 to prevent different frequency offsets from having the same phase deviation, which could lead to misidentification of the radiation source.

[0065]

[0066] Where M represents the effect of the modulation method on the phase difference, and for QPSK, M = 2.

[0067] D(t) only carries the parameter λ offset from the carrier frequency and the phase noise j(t). Therefore, this formula can completely remove the intentional modulation features in the signal and retain the unintentional modulation features, which also proves the advancement of the radiation source identification method based on the phase offset constellation diagram.

[0068] S4. Construct a phase-biased constellation diagram dataset based on D1(t), D2(t), and D3(t);

[0069] To fully reveal the characteristics of the carrier frequency and phase noise carried in the signal, this invention converts the two-dimensional matrices D1(t), D2(t), and D3(t) into a phase-biased constellation diagram. The specific steps are as follows:

[0070] First, the real and imaginary parts of D(t) are separated and converted into IQ signals, and normalization is achieved by removing points with values ​​greater than 1 and less than -1.

[0071] Next, the density distribution of D(t) on the phase-biased constellation diagram is calculated for an n′n order phase-biased constellation diagram matrix.

[0072]

[0073] have:

[0074]

[0075]

[0076]

[0077]

[0078] Finally, matrix G is converted into an image, forming a phase-biased constellation diagram, such as... Figure 2 As shown.

[0079] The horizontal and vertical axes of the phase-biased constellation diagram correspond to the phase difference between in-phase and quadrature signals of the radiating source signal, respectively. The three regions in the diagram originate from the phase differences at distances of f0 / 10, f0 / 100, and f0 / 1000, representing the fixed phase deviation caused by carrier frequency offset and the phase fluctuation caused by phase noise at different spans, respectively, and can fully represent the characteristics of the radiating source. The phase-biased constellation diagram can adapt to signals with various modulation methods.

[0080] S5. In the feature extraction part of the training network D-SupCon, the parameter settings were optimized. The intentional modulation and channel interference features in the phase-biased constellation diagram of the radiation source signal were eliminated by supervised contrastive learning, with a focus on unintentional radiation source features. In terms of network structure, the 128D features output by the feature extraction network were directly used for classification, which can effectively preserve the distribution of the feature space and achieve better results. At the same time, since the input data is a phase-biased constellation diagram with a large image size, multiple strided convolutions were introduced to reduce the amount of computation. Since the horizontal and vertical axes of the phase-biased constellation diagram are IQ values, the spatial position of the image has practical significance, so data augmentation was canceled.

[0081] The D-SupCon network directly uses the feature layer from the SupCon layer that calculates supervised contrastive loss to calculate cross-entropy loss, without relying on the 2048D feature vector from the previous layer. The logic behind this is that after the feature extraction network is trained using contrastive loss, the final 128D features have a good feature space distribution due to the characteristics of supervised contrastive loss. Using these 128D features for classification ensures that this spatial distribution is preserved during training with cross-entropy. The loss calculation formula for the D-SupCon network is as follows:

[0082]

[0083] Therefore, the backpropagation formula for training a classification network is:

[0084]

[0085] As can be seen from the formula, the features obtained from the supervised contrastive loss are... It was directly used for classification, and the combined training of supervised contrastive loss and cross-entropy improved the classification performance of the network.

[0086] In addition to altering the relationship between cross-entropy loss and supervised contrastive loss, the D-SupCon network also incorporates strided convolutions and eliminates data augmentation based on the characteristics of the input data for individual radiation source identification.

[0087] S6. Using the trained network structure, the signal to be identified is converted into the corresponding phase-biased constellation diagram and input into the network to obtain the final classification output result, thereby realizing the identification of the radiation source.

[0088] Example:

[0089] To fully demonstrate the effectiveness of the proposed radiation source individual identification method based on phase-biased constellation diagrams, this invention conducts experiments using simulated and measured datasets. The evaluation criteria are accuracy, precision, recall, and F-score, and the method is compared and analyzed with bispectral radiation source classification methods and constellation trajectory diagram-based radiation source classification methods.

[0090] This invention uses MATLAB simulations to obtain phase offset constellation diagrams for 12 modulation schemes, including BPSK, AM, QPSK, SSB, BASK, QASK, BFSK, QFSK, MSK, GMSK, 8QAM, and 16QAM. For example... Figure 3As shown, it can be clearly seen that the phase offset constellation diagrams of different modulation methods are basically the same. This indicates that the data feature extraction method of phase offset constellation diagram can remove most of the intentional modulation features of radiation sources, especially modulation mode features and transmitted signal features. At the same time, this method can retain unintentional modulation features such as carrier frequency offset and phase noise. This is the core reason why the method proposed in this invention is superior to the methods in other papers and literature.

[0091] Radiation source identification simulation dataset

[0092] The simulation dataset of this invention was generated using MATLAB simulation code. The main process is as follows: First, IQ signals under 12 modulation schemes, including BPSK, AM, QPSK, SSB, BASK, QASK, BFSK, QFSK, MSK, GMSK, 8QAM, and 16QAM, are generated using random signals. Then, carrier frequency offset and phase noise are added to the IQ signals according to the basic principle of radiation source fingerprinting described in Chapter 2 of this invention. In the simulation experiment, a total of 10 different frequency offsets and phase noises were generated to simulate 10 different radiation sources. The characteristic coefficients of different radiation sources are shown in Table 1. Then, Gaussian white noise is added to the IQ signals according to signal-to-noise ratios of -20dB, -15dB, -10dB, -5dB, 0dB, 5dB, and 10dB to simulate Gaussian white noise channels with different signal-to-noise ratios. Finally, radiation source features are extracted using the method proposed in this invention and the constellation trajectory diagram and bispectral extraction proposed in the literature to construct a radiation source identification simulation dataset.

[0093] It should be noted that, since the distribution of phase noise and carrier frequency offset in actual signals is difficult to generalize, it is difficult to measure accurate values. Therefore, the simulation data of this invention is randomly generated, which also leads to differences between the simulation data and the actual data.

[0094] Table 1 Characteristic coefficients of different radiation sources

[0095]

[0096] Table 2 Characteristic coefficients of different radiation sources

[0097]

[0098] The radiation source identification simulation dataset contains 7 signal-to-noise ratios, 10 types of radiation sources, 12 modulation schemes, and 3 radiation source feature extraction methods. Each dataset consists of 300 samples for training and 100 samples for testing, generating a total of 1,008,000 samples, each with a size of 1*400*400. To distinguish between sub-datasets of different simulation datasets, this invention names the dataset with a signal-to-noise ratio of 0dB and QPSK modulation signals the 0dB-QPSK dataset.

[0099] Radiation source identification test dataset

[0100] The experimental data consisted of IQ signals from 10 HackRF One software radio transmitters of the same manufacturer and batch, collected under real-world experimental conditions and with identical transmission configurations. The transmission frequency was 300MHz, and the bandwidth was 25kHz. The receiver converted the signal to baseband through multiple RF-to-IF conversions and then sampled the signal at a sampling rate of 87500Hz. This resulted in 204800 IQ sample points. The AM-modulated transmission audio and QPSK transmission symbols were completely random to minimize the impact of the transmitted signal on the RF fingerprint. Data acquisition was conducted on an open rooftop (with a direct channel). During IQ data acquisition, the transmitter's position was continuously changed within the room, while the 3900A receiver's position remained fixed. By changing the channel, the influence of the channel on the experiment was reduced. The transmitting and receiving equipment in the experiment were as follows: Figure 4 As shown, radio data is transmitted from the transmitter, passes through the channel, and is finally received by the receiver before reaching the terminal for backend data processing via a network cable.

[0101] For the actual acquired IQ data, some parameters are shown in Table 3. The measured signals of this invention include two modulation methods: QPSK and AM. During feature extraction, the 10 sample data from each hackRF one are converted into a real dataset using a phase-biased constellation diagram, resulting in 2400 400*400 matrices. The obtained phase-biased constellation diagrams are used to construct a dataset at a 4:1 ratio, where the training set is the AM modulated signal and the test set is both the AM and QPSK modulated signals. It should be noted that in actual experiments, multiple data transmissions can damage the transmitter. The transmitter of this invention experienced varying degrees of damage after acquiring AM and QPSK signals; therefore, only AM and QPSK signals are used to generate the real dataset.

[0102] Table 3 Parameters of the measured signal

[0103]

[0104] Network configuration

[0105] In this invention, the image size is specified as 400*400 pixels during network model training. The feature extraction network has a learning rate of 0.001, a batch size of 30, and 20 epochs, using the SGD optimizer. The classification network has a learning rate of 0.01, a batch size of 32, and 10 epochs, also using the SGD optimizer. Both the feature extraction and classification networks use the model from the last epoch as the optimal model. This invention uses the PyTorch framework and an NVIDIA GeForce GTX 2080ti GPU for training.

[0106] Comparison of experimental results of different classification networks

[0107] When using the phase-biased constellation diagram method proposed in this invention to identify radiation sources, the choice of classifier directly affects the identification performance. A better classifier estimates the phase bias and phase distribution patterns in the phase-biased constellation diagram, resulting in better outcomes; conversely, a poorer classifier yields worse results. To demonstrate the superior classification performance of the proposed D-SupCon network for radiation source identification, this invention compares the classification performance of D-SupCon, SupCon, VGG19, VGG16, and Xception networks on the QPSK-0db dataset. Both D-SupCon and SupCon networks utilize ResNet34, ResNet50, and ResNet101 as their network backbones.

[0108] Table 4. Recognition rates of different networks on the QPSK-0db dataset.

[0109] VGG16 86.6％ VGG19 85.9％ Xception 81.4％ SupCon+ResNet34 86.7％ SupCon+ResNet50 86.5％ SupCon+ResNet101 86.7％ D-SupCon+ResNet34 86.8％ D-SupCon+ResNet50 87.0％ D-SupCon+ResNet101 87.3％

[0110] As shown in Table 4, for the QPSK-0db dataset, the D-SupCon network achieves the highest recognition accuracy compared to the traditional SupCon network. Furthermore, the accuracy of the D-SupCon network is superior to that of the traditional VGG16, VGG16, and Xception networks. Therefore, it can be concluded that the D-SupCon network structure proposed in this invention has better performance for radiation source identification tasks.

[0111] Comparison of experimental results of different feature extraction methods

[0112] To test the impact of different radiation source feature extraction methods on the performance of individual radiation source identification tasks, this section generates simulation datasets for 12 modulation schemes, including BPSK, AM, QPSK, SSB, BASK, QASK, BFSK, QFSK, MSK, GMSK, 8QAM, and 16QAM, using MATLAB software. Individual radiation source identification is achieved using the phase-biased constellation diagram, bispectral data, and constellation trajectory diagram proposed in this invention as radiation source feature extraction schemes, and the results are compared.

[0113] The bispectral-based radiation source feature extraction method exhibits good anti-interference capabilities and is not limited by modulation scheme. It extracts carrier frequency offset and phase cumulative change features by calculating the third-order cumulants of the IQ signals; however, this method also retains intentional modulation features. Radiation source individual identification methods based on constellation trajectory diagrams are significantly affected by noise, and constellation trajectory diagrams are only suitable for constant-amplitude signals, showing poor identification results for AM and QAM signals. This invention improves upon this method using the core theory of constellation trajectory diagrams, enabling it to adapt to multiple modulation schemes. This is because existing time-domain-based radiation source identification methods are generally only applicable to one or a single type of modulation scheme, and constellation trajectory diagrams are the inspiration for this invention; therefore, they were chosen as a comparison scheme. However, this also leads to some differences between this method and some results in the literature. Constellation trajectory diagrams can extract the difference values ​​of the IQ signals, representing the carrier frequency offset and phase noise features of the signal; this method also retains intentional modulation features.

[0114] The two radiation source feature extraction methods described above analyze IQ signals in the time and frequency domains, respectively, and basically represent mainstream steady-state-based individual radiation source identification methods. To ensure the rationality of data analysis, this invention uses data extraction to ensure that the feature maps extracted by all methods are of consistent size, and all methods use D-SupCon networks for radiation source classification, comparing only the feature extraction methods. For the simulation dataset of 12 modulation schemes, the recognition results of phase offset constellation diagrams, constellation trajectory diagrams, and bispectral representations are as follows: Figure 5 As shown. From Figure 5 It can be seen that as the signal-to-noise ratio (SNR) increases, the accuracy of radiation source feature extraction methods based on phase-biased constellation diagrams, bispectral methods, and constellation trajectory diagrams also gradually improve. When the SNR is greater than -5dB, for almost all modulation schemes and SNRs, the method based on phase-biased constellation diagrams significantly outperforms the methods based on constellation trajectory diagrams and bispectral methods. When the SNR is between -20dB and -5dB, for most modulation schemes, the accuracy of phase-biased constellation diagrams is also better than that of bispectral methods and constellation trajectory diagrams. The recognition performance of phase-biased constellation diagrams, constellation trajectory diagrams, and bispectral methods on a simulation dataset of 12 modulation schemes is shown below. Figure 5As shown, when the signal-to-noise ratio is greater than -5dB, the radiation source feature extraction method based on the phase-biased constellation diagram achieves a recognition rate of more than 60% for all 12 modulation schemes. When the signal-to-noise ratio is greater than 0dB, the accuracy of individual radiation source identification for all modulation schemes in the bispectral method exceeds 50%. When the signal-to-noise ratio is greater than 10dB, the constellation trajectory diagram achieves individual radiation source identification for signals of most modulation schemes.

[0115] from Figure 5 and 6 It can be observed that under the same signal-to-noise ratio, the signal recognition effect of different modulation methods is different. The main reason for this result is the difference in anti-interference of different modulation methods. In addition, some modulation methods will also introduce different degrees of noise in the feature extraction stage, as shown in Section 3 of this invention.

[0116] Radiation source features exhibit stability and regularity. When a suitable radiation source feature extraction method is selected, given sufficient time, even if the signal-to-noise ratio is so low that communication is impossible, the radiation source features can still be extracted. This section uses experimental comparisons to verify that phase-biased constellation diagrams are more effective than the other two methods for radiation source identification.

[0117] Comparison of experimental results for individual radiation source identification under different modulation methods

[0118] This section primarily tests the recognition performance of different feature extraction methods on signals with novel modulation schemes. In this experiment, features of the QPSK-0db dataset are first extracted using phase-biased constellation diagrams, constellation trajectory diagrams, and bispectral analysis. Then, the D-SupCon network is trained separately using the extracted features. Finally, the adaptability of the three methods to signals with different modulation schemes is verified using data from 11 modulation schemes at 0dB. The confusion matrices for radiation source classification by the three methods are shown below. Figure 7 , 8 As shown in Figure 9, the radiation source feature extraction method based on phase-biased constellation diagrams clearly demonstrates good classification performance for radiation source data with 11 modulation schemes, achieving an average accuracy of 95.26%, significantly outperforming bispectral and constellation trajectory diagrams. This verifies that phase-biased constellation diagrams also provide good classification results for radiation source identification tasks with untrained modulation schemes. The bispectral method can generally identify 8QAM, 16QAM, AM, BASK, BFSK, QASK, QFSK, and SSB signals, but its accuracy is lower for BPSK, GMSK, QPSK, and MSK signals. The radiation source identification task based on constellation trajectory diagrams performs poorly for all 11 modulation schemes, showing good results only for some. The results of the three methods for radiation source identification with 11 modulation schemes are shown in Figure 9. Figure 10As shown, it is clear that the radiation source feature extraction method based on phase-biased constellation diagrams outperforms the bispectral and constellation trajectory diagram methods.

[0119] Actual experimental results

[0120] To test the effectiveness of phase-biased constellation diagrams in identifying actual radiation sources, this invention generated a measured dataset using real equipment. By comparing the phase-biased constellation diagrams of the actual signal and the simulated signal, the rationality of the simulated signal generated by this invention can be demonstrated. Figure 12 As can be seen, different HackRF One radiation sources have different phase-biased constellation diagrams. This proves that different radiation sources can generate phase-biased constellation diagrams and reflect the individual characteristics of different radiation sources. Therefore, it is feasible to use phase-biased constellation diagrams to identify individual radiation sources in practical engineering applications.

[0121] For the experimental dataset, this invention trains the D-supCon network using the QPSK training set, and then validates it using the QPSK validation set and AM validation set. The performance of phase-biased constellation diagrams and the D-supCon network in identifying radiation sources is tested, and the results are shown in Table 2. It can be seen that the radiation source identification method based on phase-biased constellation diagrams proposed in this invention has good performance for radiation source identification tasks where the training and validation data have the same modulation scheme, achieving an accuracy of 86.7%. It also has better classification performance for radiation source identification tasks where the training and validation data have different modulation schemes, achieving an accuracy of 70.06% when the other two methods completely fail. It should also be noted that due to the complexity of the channel in the experimental environment and the difference between the parameters of the real radiation source and the simulated signal, there is a difference between the simulated signal and the actual signal. However, the experimental results show that the method proposed in this invention significantly outperforms bispectral and constellation trajectory diagrams in identifying radiation sources for both signals with the same and different modulation schemes, which fully demonstrates the advancement of the proposed method.

[0122] Table 5 Evaluation metrics for radio frequency fingerprint recognition across different datasets

[0123]

[0124] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, technical solutions formed by replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A radiation source identification method based on phase-biased constellation diagrams, characterized in that, The specific steps are as follows: S1. Remove the influence of intentional modulation in the signal on radiation source identification; First, remove the modulated baseband information in the received IQ signal. In the phase offset constellation diagram method, the baseband information in IQ is eliminated by calculating the square envelope of the signal, and the signal is then subjected to constant removal and normalization. S2. Further noise reduction processing of the IQ signal is required, due to the frequency offset. The characteristics of radiation sources vary greatly depending on the individual source, and are generally unknown. Therefore, low-pass filtering is not suitable. Instead, autocorrelation filtering is used to extract the radiation source features. After the above processing, the signal is obtained. ; S3, Calculate the signal The phase difference, where n takes the value of , , The three values, the Given the sampling rate, we obtain (t), (t), (t), in (t)(n=1,2,3) only carries the parameter offset from the carrier frequency. and phase noise It can completely remove intentionally modulated features from a signal while retaining unintentionally modulated features; S4, according to , , Construct a phase-biased constellation graph dataset; The specific steps include: Calculated The real and imaginary parts of (t)(n=1,2,3) are separated and converted into IQ signals, and normalization is achieved by removing points with values ​​greater than 1 and less than -1; then the calculation is performed. The density distribution of (t)(n=1,2,3) on the phase-biased constellation diagram was obtained, and finally the matrix G was converted into an image to form the phase-biased constellation diagram dataset. S5. In the training of the feature extraction part of the D-SupCon network, the parameter settings were optimized. In terms of network structure, the 128D features output by the feature extraction network were directly used for classification, which can effectively preserve the distribution of the feature space and achieve better results. At the same time, since the input data is a phase-biased constellation map with a large image size, multiple strided convolutions were introduced to reduce the amount of computation. Since the horizontal and vertical axes of the phase-biased constellation map are IQ values, the spatial position of the image has practical significance, so data augmentation was canceled. S6. Using the trained network structure, the signal to be identified is converted into the corresponding phase-biased constellation diagram and input into the network to obtain the final classification output result, thereby realizing the identification of the radiation source.

Images