Zero-shot radio fingerprinting method based on triple network and unsupervised clustering

Abstract

The application provides a kind of zero sample radio frequency fingerprint identification method based on triple network and unsupervised clustering, comprising the following steps: collecting the radio frequency fingerprint of several radio equipment, and making radio frequency fingerprint dataset;The triple network is trained by the radio frequency fingerprint dataset, and the optimized triple network is obtained;The radio frequency fingerprint data to be measured is preprocessed, and the preprocessed radio frequency fingerprint data is input into the optimized triple network, and the feature sample is extracted;According to the feature sample, the estimated equipment number is estimated, and the radio frequency fingerprint identification result is calculated by clustering algorithm.The application can improve the accuracy of zero sample radio frequency fingerprint identification.

Description

Technical Field

[0001] This invention relates to the field of radio frequency fingerprint recognition, and in particular to a zero-shot radio frequency fingerprint recognition method based on triplet networks and unsupervised clustering. Background Technology

[0002] Due to individual hardware variations and manufacturing processes, subtle differences inevitably exist between any two individual radios. Even radios of the same model and from the same production line will emit signals with unique individual characteristics, known as radio frequency fingerprints. As modern warfare rapidly evolves into information warfare, various electronic devices, represented by radios, undertake crucial tasks such as communication, surveillance, and electronic jamming. Therefore, the ability to identify enemy radios before the enemy is a vital aspect of electronic warfare and, to a certain extent, determines the initiative on the battlefield. Consequently, radio frequency fingerprinting technology has attracted widespread attention from researchers worldwide in recent years.

[0003] Generally, radio frequency (RF) fingerprint recognition can be viewed as a classification problem, and most of the state-of-the-art RF fingerprint classifiers are based on deep learning technology. Thanks to their powerful feature extraction and pattern recognition capabilities, deep learning-based classifiers can achieve remarkably high accuracy. However, they require a large number of training samples during offline training to ensure satisfactory performance and avoid overfitting. In practical applications, obtaining sufficient labeled training samples is difficult. Furthermore, in non-cooperative communication scenarios such as electronic warfare, unknown devices may exist outside the training set, making it even more challenging to obtain labeled samples from these unknown devices as training data. Therefore, solving the zero-shot RF fingerprint recognition problem—that is, accurately classifying unknown devices without labeled samples—is a pressing issue in the field, yet existing research has rarely addressed it.

[0004] To address this issue, some literature proposes using deep causal convolutional networks for feature extraction, followed by dimensionality reduction using the UMAP algorithm and clustering using the DBSCAN algorithm, thereby achieving zero-shot RFID fingerprint recognition with some success. This approach is known as the DML scheme. However, the large number of parameters in the deep causal convolutional networks in the DML scheme results in long computation times. Furthermore, the DML scheme lacks discussion on estimating the number of devices, so the final cluster count often does not match the actual number of devices.

[0005] Therefore, existing zero-sample radio frequency fingerprint recognition solutions still have significant room for improvement in accuracy. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a zero-shot radio frequency fingerprinting method based on triplet networks and unsupervised clustering, which can improve the accuracy of radio frequency fingerprinting.

[0007] This invention provides a zero-shot radio frequency fingerprinting method based on triplet networks and unsupervised clustering, comprising the following steps:

[0008] Collect radio frequency fingerprints from several radio stations and create an radio frequency fingerprint dataset;

[0009] The triplet network is trained using the aforementioned radio frequency fingerprint dataset to obtain an optimized triplet network;

[0010] The RF fingerprint data to be tested is preprocessed, and the preprocessed RF fingerprint data is input into the optimized triplet network to extract feature samples;

[0011] The estimated number of devices is obtained based on the feature samples, and the radio frequency fingerprint recognition result is calculated using a clustering algorithm.

[0012] Furthermore, the process of collecting radio frequency signals from several radio stations to create a radio frequency fingerprint dataset specifically includes:

[0013] A sliding window is set to sample the radio frequency signal. The sampling start point is determined by calculating the posterior probability that each sampling point in the sliding window is the starting point of the power-on transient signal. Starting from the sampling start point, a radio frequency signal of a preset length is extracted as the power-on transient signal.

[0014] The power-on transient signal is regularized to obtain a regular transient signal, and the envelope of the regular transient signal is extracted by Hilbert transform as the radio frequency fingerprint of the radio equipment.

[0015] The radio frequency fingerprints of several radio stations are collected and combined to form the radio frequency fingerprint dataset.

[0016] Furthermore, the triplet network includes three one-dimensional convolutional layers, each of which is followed by a batch normalization layer and a max pooling layer; wherein, the convolutional layers are used to extract feature vectors of radio frequency fingerprints at different scales.

[0017] Furthermore, the step of training the triplet network using the radio frequency fingerprint dataset to obtain an optimized triplet network specifically includes:

[0018] The samples in the radio frequency fingerprint dataset are preprocessed, and the preprocessed samples are input into the triplet network. The training feature vector is extracted by the triplet network.

[0019] Based on the training feature vector, the training loss value is calculated using a preset loss function;

[0020] Based on the training loss value, the partial derivative of the training loss value with respect to each trainable parameter in the triplet network is calculated using the backpropagation algorithm. Then, the trainable parameters are optimized using a preset optimization algorithm to obtain optimized network parameters, thereby obtaining an optimized triplet network.

[0021] Preferably, the preprocessing of samples in the radio frequency fingerprint dataset specifically includes:

[0022] A preset number of radio frequency fingerprints are selected from the radio frequency fingerprint dataset as the original sample set, and the size M of the original sample set and the number N of original samples used for each fusion are determined.

[0023] Construct a fusion label table; wherein the fusion label table includes all sample combinations of N samples selected from the original sample set;

[0024] Each sample in the aforementioned sample combination is fused to obtain several enhanced samples; wherein, each enhanced sample corresponds to a sample combination, and the enhanced sample is obtained by averaging the samples in the corresponding sample combination;

[0025] The original samples and the augmented samples are combined into an augmented training set, and each sample in the augmented training set is preprocessed based on variational mode decomposition.

[0026] Preferably, the further optimization of the triplet network specifically includes:

[0027] After updating the optimized network parameters to the triplet network, the samples of the verification set in the radio frequency fingerprint dataset are input into the triplet network after the parameters are updated, and the verification feature vector is obtained.

[0028] The verification loss value is calculated based on the verification feature vector using the preset loss function;

[0029] When the verification loss value is greater than the preset loss value threshold, the optimized network parameters are optimized again according to the radio frequency fingerprint dataset using a preset optimization algorithm, and the optimized network parameters are updated in the triplet network, and then verified again using the verification set.

[0030] When the verification loss value is less than the preset loss value threshold, the corresponding network parameters are used as the optimal network parameters and input into the triplet network to obtain the optimized triplet network.

[0031] Furthermore, the radio frequency fingerprint data to be tested is preprocessed, and the preprocessed radio frequency fingerprint data is input into the optimized triplet network to extract feature samples, specifically including:

[0032] The radio frequency fingerprint data to be tested is preprocessed based on variational mode decomposition. The preprocessed radio frequency fingerprint data includes five vectors VP1 to VP5. Among them, VP1 is the low-frequency component of the original signal sample; VP2 and VP3 are the real and imaginary parts of the fast Fourier transform of VP1, respectively; VP4 is the instantaneous frequency of VP1; and VP5 is the first mode of the difference sequence of VP1 after variational mode decomposition.

[0033] The preprocessed radio frequency fingerprint data is input into the optimized triplet network, and the corresponding feature samples are extracted.

[0034] Furthermore, estimating the number of radio devices based on the feature samples specifically includes:

[0035] Determine the search range for the number of radio devices, wherein the search range includes several estimated device counts;

[0036] The feature samples are clustered using the K-means++ clustering algorithm, and the values ​​of the several estimated devices are used as the pre-input K values ​​of the K-means++ clustering algorithm to obtain several clustering results.

[0037] The silhouette coefficient and Calinski-Harabaz index of the clustering results are calculated respectively; wherein, one estimated device value corresponds to one silhouette coefficient and one Calinski-Harabaz index;

[0038] Based on the contour coefficient and the Calinski-Harabaz index, the evaluation scores of the several estimated device counts are calculated respectively, and the estimated device count corresponding to the highest evaluation score is taken as the estimated number of devices.

[0039] Preferably, the step of calculating the evaluation scores of the plurality of estimated device values ​​based on the contour coefficient and the Calinski-Harabaz index specifically includes:

[0040] Based on several profile coefficients and several Calinski-Harabaz indices corresponding to several estimated device values, an evaluation score is calculated for each estimated device value using an evaluation formula; wherein, the evaluation formula is specifically:

[0041]

[0042] Where s is the evaluation score corresponding to the current estimated number of device values, s SC The profile coefficient corresponding to the current estimated number of equipment values, min(s) SC ) represents the minimum value among several contour coefficients, max(s) SC ) represents the maximum value among several contour coefficients, s CH Let min(s) be the Calinski-Harabaz exponent corresponding to the current estimated number of devices. CH ) represents the minimum value among several Calinski-Harabaz exponents, max(s) CH ) is the maximum value among several Calinski-Harabaz indices.

[0043] Furthermore, the calculation of the radio frequency fingerprint recognition result through the clustering algorithm specifically includes:

[0044] The estimated number of devices is set as the pre-input K value of the K-means++ clustering algorithm, and the feature samples are clustered by the K-means++ clustering algorithm to obtain the radio frequency fingerprint recognition result.

[0045] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0046] To address the high cost and low accuracy issues of zero-shot RFID fingerprinting, this invention proposes a novel zero-shot RFID fingerprinting method based on triplet networks and unsupervised clustering. On the one hand, reasonable data preprocessing and loss function construction reduce neural network parameters, thereby reducing method complexity. On the other hand, a K-value estimation algorithm is proposed, which, combined with the K-means++ clustering algorithm, can improve the reliability of clustering while accurately estimating the number of devices, thus improving the accuracy of RFID fingerprinting. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a zero-shot radio frequency fingerprinting method based on triplet networks and unsupervised clustering, provided as an embodiment of the present invention.

[0048] Figure 2 This is a schematic diagram of a triplet network structure provided in an embodiment of the present invention.

[0049] Figure 3 This is a flowchart illustrating a preferred embodiment of a zero-shot radio frequency fingerprinting method based on triplet networks and unsupervised clustering provided by an embodiment of the present invention.

[0050] Figure 4This is an overall framework diagram of a zero-shot radio frequency fingerprinting method based on triplet networks and unsupervised clustering, provided as an embodiment of the present invention.

[0051] Figure 5 This is a schematic diagram of the experimental results of comparative experiment one provided in an embodiment of the present invention.

[0052] Figure 6 This is a schematic diagram of the experimental results of comparative experiment two provided in an embodiment of the present invention.

[0053] Figure 7 This is a schematic diagram of the experimental results of comparative experiment three provided in an embodiment of the present invention. Detailed Implementation

[0054] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0055] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0056] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0057] Reference Figure 1 The following is a flowchart illustrating a zero-shot radio frequency fingerprinting method based on triplet networks and unsupervised clustering, according to an embodiment of the present invention, comprising the following steps:

[0058] S1: Collect the radio frequency fingerprints of several radio station devices and create an radio frequency fingerprint dataset;

[0059] S2: Train the triplet network using the aforementioned radio frequency fingerprint dataset to obtain an optimized triplet network;

[0060] S3: Preprocess the RF fingerprint data to be tested, and input the preprocessed RF fingerprint data into the optimized triplet network to extract feature samples;

[0061] S4: Estimate the number of devices based on the feature samples, and calculate the radio frequency fingerprint recognition result using a clustering algorithm.

[0062] For step S1, specifically, the process of collecting radio frequency signals from several radio stations to create a radio frequency fingerprint dataset includes:

[0063] A sliding window is set to sample the radio frequency signal. The sampling start point is determined by calculating the posterior probability that each sampling point in the sliding window is the starting point of the power-on transient signal. Starting from the sampling start point, a radio frequency signal of a preset length is extracted as the power-on transient signal.

[0064] The power-on transient signal is regularized to obtain a regular transient signal, and the envelope of the regular transient signal is extracted by Hilbert transform as the radio frequency fingerprint of the radio equipment.

[0065] The radio frequency fingerprints of several radio stations are collected and combined to form the radio frequency fingerprint dataset.

[0066] In a preferred embodiment, the step of acquiring radio frequency signals includes:

[0067] S11: Set a sliding window to extract the power-on transient signal of the radio equipment. First, calculate the posterior probability that each sampling point in the sliding window is the starting point of the power-on transient signal. Then, determine the sampling point with the largest posterior probability as the sampling starting point. Finally, extract a fixed length L of radio frequency signal from the sampling starting point as the power-on transient signal.

[0068] S12: Regularize the extracted transient signal, i.e., divide it by its root mean square value, to eliminate the influence of received signal power variations. The specific calculation expression is as follows:

[0069]

[0070] Where a[v] and a′[v] are the v-th instantaneous values ​​of the initial transient signal and the regular transient signal, respectively, and L is the length of the transient signal;

[0071] S13: Extract the envelope b[v] of the canonical transient signal a′[v] using Hilbert transform, and use b[v] as the RF fingerprint of the corresponding radio station equipment. The specific calculation expression is as follows:

[0072]

[0073] in, Here, is the Hilbert transform function, and j represents the imaginary unit.

[0074] S14: Collect the radio frequency fingerprints of several radio stations respectively, and combine these radio frequency fingerprints into a radio frequency fingerprint dataset. The radio frequency fingerprint dataset is further divided into a training set, a validation set, and a test set.

[0075] For step S2, specifically, the triplet network includes three one-dimensional convolutional layers, each of which is followed by a batch normalization layer and a max pooling layer; wherein, the convolutional layers are used to extract feature vectors of radio frequency fingerprints at different scales.

[0076] In a preferred embodiment, refer to Figure 2This is a schematic diagram of a triplet network provided in an embodiment of the present invention. The triplet network used in the present invention includes three one-dimensional convolutional layers; wherein, each convolutional layer is followed by a batch normalization (BN) layer and a max pooling layer.

[0077] In the triplet network, each convolutional layer is mainly responsible for extracting radio frequency fingerprint features at different scales, with output channel dimensions of C. out，1 C out，2 C out，3 The convolutional kernel sizes are 17, 9, and 5, with padding sizes of 8, 4, and 2, respectively. Following the convolutional layers is a Batch Normalization (BN) layer with a ReLU activation function to mitigate the vanishing and exploding gradient problems during training, accelerating neural network convergence. The max-pooling layers after the BN layers are used to reduce the time channel size while preserving key features, thus reducing complexity and accelerating convergence. The filter sizes of the three max-pooling layers are 8, 16, and 16, respectively. The final layer outputs C... out,3 The dimensional vector is the feature vector extracted from the input preprocessed signal.

[0078] For step S2, specifically, training the triplet network using the radio frequency fingerprint dataset to obtain an optimized triplet network includes:

[0079] The samples in the radio frequency fingerprint dataset are preprocessed, and the preprocessed samples are input into the triplet network. The training feature vector is extracted by the triplet network.

[0080] Based on the training feature vector, a training loss value is calculated using a preset loss function to determine the training discrimination of feature extraction; wherein, the magnitude of the training loss value is inversely proportional to the magnitude of the training discrimination; based on the training loss value, the partial derivative of the training loss value with respect to each trainable parameter in the triplet network is calculated using a backpropagation algorithm, and then the trainable parameters are optimized using a preset optimization algorithm to obtain optimized network parameters, thereby obtaining an optimized triplet network.

[0081] Furthermore, as a preferred embodiment, the preprocessing of samples in the radio frequency fingerprint dataset specifically includes:

[0082] A preset number of radio frequency fingerprints are selected from the radio frequency fingerprint dataset as the original sample set, and the size M of the original sample set and the number N of original samples used for each fusion are determined.

[0083] Construct a fusion label table; wherein the fusion label table includes all sample combinations of N samples selected from the original sample set;

[0084] Each sample in the aforementioned sample combination is fused to obtain several enhanced samples; wherein, each enhanced sample corresponds to a sample combination, and the enhanced sample is obtained by averaging the samples in the corresponding sample combination;

[0085] The original samples and the augmented samples are combined into an augmented training set, and each sample in the augmented training set is preprocessed based on variational mode decomposition.

[0086] Furthermore, as a preferred embodiment, the further optimization of the triplet network specifically includes:

[0087] After updating the optimized network parameters to the triplet network, the samples of the verification set in the radio frequency fingerprint dataset are input into the triplet network after the parameters are updated, and the verification feature vector is obtained.

[0088] Based on the verification feature vector, the verification loss value is calculated using the preset loss function, and the verification discrimination of the feature extraction is determined.

[0089] When the verification discrimination is less than the preset discrimination threshold, the optimized network parameters are optimized again according to the radio frequency fingerprint dataset using the preset optimization algorithm, and the optimized network parameters are updated to the triplet network, and then verified again using the verification set.

[0090] When the verification discrimination is greater than the preset discrimination threshold, the corresponding network parameters are used as the optimal network parameters and input into the triplet network to obtain the optimized triplet network.

[0091] In a preferred embodiment, the triplet network needs to be trained using training set samples from the RF fingerprint dataset. Before training the network, the training set samples are further fused and enhanced. The fusion and enhancement process is as follows:

[0092] First, determine the original sample set S. o The size M and the number of original samples N used in each fusion;

[0093] Then, according to S o Select all combinations of N samples to construct a fusion label table.

[0094] Then, when generating a new sample for the i-th time, according to T i From the original sample set S o=[S o,1 ,…,S o,M Select N samples from [the sample list]. in Then, the samples are fused to obtain new samples. The specific formula for calculating the new samples is as follows:

[0095]

[0096] Finally, the augmented training set is obtained, which contains the original M samples and One enhanced sample.

[0097] Furthermore, the input values ​​of the enhanced training set are used to extract training feature vectors from the triplet network. To quantitatively measure the discriminative power of the extracted features, this invention constructs a triplet loss function, the specific calculation expression of which is as follows:

[0098]

[0099] Among them, F a Indicates anchor point characteristics; F p Indicates and F a Positive features belonging to the same category; F n Indicates and F a Negative features belonging to different classes; φ scope and φ margin The loss function hyperparameters set by the user.

[0100] Obviously, when the distance between features extracted from the same device is small, while the distance between features extracted from different devices is large, the loss function can approach the minimum value, and the discriminative power of the extracted features tends to the maximum value.

[0101] Then, the partial derivatives of the loss function with respect to each trainable parameter in the neural network are systematically calculated using the backpropagation algorithm. Based on this, the neural network parameters are updated using a pre-defined optimization algorithm, thereby improving the performance of the triplet network feature extractor and achieving high-discrimination feature extraction. In this method, the Adam algorithm is used, and the learning rate is set to 1×10⁻⁶. -3 .

[0102] After the updated neural network parameters are input into the triplet network, the performance of the triplet network is tested using validation set samples from the RF fingerprint dataset. Specifically, this includes:

[0103] The samples in the validation set are input into the triplet network after the neural network parameters are updated, and the validation feature vector is extracted.

[0104] Based on the obtained validation feature vector, the validation loss value is calculated using the loss function.

[0105] The verification loss value is compared with a preset loss value threshold. When the verification loss value is greater than the preset loss value threshold, the optimized network parameters are optimized again using a preset optimization algorithm, and the optimized network parameters are updated in the triplet network for verification again.

[0106] When the verification loss value is less than the preset loss value threshold, the corresponding network parameters are determined to be the optimal network parameters and input into the triplet network to finally obtain the optimized triplet network.

[0107] For step S3, specifically, the radio frequency fingerprint data to be tested is preprocessed, and the preprocessed radio frequency fingerprint data is input into the optimized triplet network to extract feature samples, specifically including:

[0108] The radio frequency fingerprint data to be tested is preprocessed based on variational mode decomposition. The preprocessed radio frequency fingerprint data includes five vectors VP1 to VP5. Among them, VP1 is the low-frequency component of the original signal sample; VP2 and VP3 are the real and imaginary parts of the fast Fourier transform of VP1, respectively; VP4 is the instantaneous frequency of VP1; and VP5 is the first mode of the difference sequence of VP1 after variational mode decomposition.

[0109] The preprocessed radio frequency fingerprint data is input into the triplet network, and the corresponding feature samples are extracted.

[0110] In a preferred embodiment, after preprocessing the RF fingerprint data to be tested using variational mode decomposition (VMD), the original signal S can be decomposed into Q modes, such that the sum of the estimated bandwidths of each of the Q modes is minimized under the constraint that the sum is S, thereby decomposing the original signal into signal components of different frequency scales. The specific constraint variational expression is as follows:

[0111]

[0112] Among them, {u q} and {ω q} represent the set of modes and their set of center frequencies, respectively; δ represents the Dirac distribution; * represents the convolution operator.

[0113] Considering the significant differences in the rising gradient of the power-on transient signals of different devices, the sample after VMD data preprocessing in this invention will contain five vectors VP1 to VP5. VP1 is the first mode of the transient signal after VMD processing, i.e., the low-frequency component of the original signal sample; VP2 and VP3 are the real and imaginary parts of the Fast Fourier Transform of VP1, respectively; VP4 is the instantaneous frequency of VP1; and VP5 is the first mode of the difference sequence of VP1 after VMD processing. After VMD preprocessing, the triplet network feature extractor can extract more discriminative features and reduce the influence of noise and interference.

[0114] Finally, the preprocessed RF fingerprint data is input into the optimized triplet network, and the corresponding feature samples are obtained after extraction.

[0115] For step S4, specifically, estimating the number of radio devices based on the feature samples includes:

[0116] Determine the search range for the number of radio devices, wherein the search range includes several estimated device counts;

[0117] The feature samples are clustered using the K-means++ clustering algorithm, and the values ​​of the several estimated devices are used as the pre-input K values ​​of the K-means++ clustering algorithm to obtain several clustering results.

[0118] The silhouette coefficient and Calinski-Harabaz index of the clustering results are calculated respectively; wherein, one estimated device value corresponds to one silhouette coefficient and one Calinski-Harabaz index;

[0119] Based on the contour coefficient and the Calinski-Harabaz index, the evaluation scores of the several estimated device counts are calculated respectively, and the estimated device count corresponding to the highest evaluation score is taken as the estimated number of devices.

[0120] As a preferred embodiment, the step of calculating the evaluation scores of the plurality of estimated device values ​​based on the contour coefficient and the Calinski-Harabaz index specifically includes:

[0121] Based on several profile coefficients and several Calinski-Harabaz indices corresponding to several estimated device values, an evaluation score is calculated for each estimated device value using an evaluation formula; wherein, the evaluation formula is specifically:

[0122]

[0123] Where s is the evaluation score corresponding to the current estimated number of device values, s SC The profile coefficient corresponding to the current estimated number of equipment values, min(s) SC ) represents the minimum value among several contour coefficients, max(s) SC ) represents the maximum value among several contour coefficients, s CH Let min(s) be the Calinski-Harabaz exponent corresponding to the current estimated number of devices. CH ) represents the minimum value among several Calinski-Harabaz exponents, max(s) CH ) is the maximum value among several Calinski-Harabaz indices.

[0124] In a preferred embodiment, for the feature samples extracted from the RF fingerprint data to be tested, the K-value estimation method proposed in this invention needs to be used to estimate the number of radio station devices, specifically including:

[0125] First, determine the search range for the number of devices [K0,K0+1,…,K0+κ-1];

[0126] Then, the number of devices was set to the K value of the K-means++ clustering algorithm, and clustering was performed. Subsequently, the silhouette coefficient and Calinski-Harabaz (CH) index of the clustering results under the current K value were calculated to form two 1×κ vectors s. SC With s CH , representing the silhouette coefficient and CH index of the clustering results at each K value, respectively;

[0127] Finally, the evaluation score for the number of each device is calculated using the following formula:

[0128]

[0129] Where s is the evaluation score corresponding to the current estimated number of device values, s SC The profile coefficient corresponding to the current estimated number of equipment values, min(s) SC ) represents the minimum value among several contour coefficients, max(s) SC ) represents the maximum value among several contour coefficients, s CH Let min(s) be the Calinski-Harabaz exponent corresponding to the current estimated number of devices. CH ) represents the minimum value among several Calinski-Harabaz exponents, max(s) CH ) is the maximum value among several Calinski-Harabaz indices.

[0130] The number of devices corresponding to the largest element in s is the estimated number of devices.

[0131] Using this method to evaluate the score of each device's number of devices can limit the score to between 0 and 1, and also increases the adaptability and stability of the evaluation score compared to using only a single indicator.

[0132] For step S4, specifically, the calculation of the radio frequency fingerprint recognition result through the clustering algorithm includes:

[0133] The estimated number of devices is set as the pre-input K value of the K-means++ clustering algorithm, and the feature samples are clustered by the K-means++ clustering algorithm to obtain the radio frequency fingerprint recognition result.

[0134] In summary, the flowchart of one specific embodiment of the zero-shot RFID fingerprinting method based on triplet networks and unsupervised clustering provided by the present invention is shown below. Figure 3 As shown, the overall framework diagram of the method is as follows: Figure 4 As shown.

[0135] Below, embodiments of the present invention will be presented with reference to the accompanying drawings, and four comparative experiments will be conducted to verify the differences between the technical solutions provided by the present invention and the prior art.

[0136] Experiment 1

[0137] In Experiment 1, ten identical KELIX DP485-01 DMR walkie-talkies were used as the radio devices to be identified. To avoid the influence of differences in walkie-talkie parameter settings on RF fingerprint recognition, all walkie-talkies were uniformly set to a center frequency of 141.825MHz, a bandwidth of 12.5kHz, and a modulation method of 4FSK. The walkie-talkie's transmitted signal was transmitted to a data acquisition card via a shielded cable for acquisition. The data acquisition card used had a sampling rate of 400 MS / s and a sampling precision of 14 bits.

[0138] Each walkie-talkie collects 155 signal blocks, and the 1550 transient power-on signals obtained after preprocessing are used as the RF fingerprint dataset. Of the sampled and preprocessed dataset, 50 samples are used as the training set to train the triplet network feature extractor, and the remaining 1500 samples are used as the test set.

[0139] To test the performance of the VMD preprocessing described in this invention, as well as the performance of the triplet network feature extractors with and without preprocessing and with partial preprocessing, multiple triplet networks were trained using the aforementioned 50 training set samples under different preprocessing conditions. After training, 1500 test set samples were input into each triplet network to obtain feature sets under different preprocessing conditions. Subsequently, Principal Component Analysis (PCA) was performed on each feature set to reduce its dimensionality to 2D graphs, so as to visually demonstrate the feature extraction capabilities of the feature extractors trained under each preprocessing condition.

[0140] Experimental results are as follows Figure 5 As shown above, the sample after VMD data preprocessing in this invention will contain 5 vectors VP1 to VP5, where VP1 is the first mode of the transient signal after VMD processing, that is, the low-frequency component of the original signal sample; VP2 and VP3 are the real and imaginary parts of the fast Fourier transform of VP1, respectively; VP4 is the instantaneous frequency of VP1; and VP5 is the first mode of the difference sequence of VP1 after VMD processing.

[0141] As can be seen, when no preprocessing is performed and the feature extractor is not trained, such as Figure 5 As shown in (a), the features extracted by different devices overlap and are difficult to distinguish; and the results are equally poor if the feature extractor is trained without preprocessing. Figure 5 As shown in (b), the overlapping area of ​​features is still relatively large; and according to Figure 5 (c) to Figure 5 (e) We can see that with the help of VMD preprocessing, the extracted features are significantly easier to distinguish. Furthermore, different preprocessing vectors contribute differently to distinguishing individual devices; therefore, the complete VMD preprocessing, which combines different preprocessing vectors, has the best feature discrimination. Finally, by further applying random integration (RI) data augmentation to the training set samples in step S4, combined with the complete VMD preprocessing, the trained triplet feature extractor can achieve relatively optimal feature extraction performance, such as... Figure 5 As shown in (f).

[0142] Experiment 2

[0143] Experiment 2 primarily investigates the performance comparison of the proposed scheme and the control scheme in zero-sample RF fingerprint recognition under different numbers of unknown devices. In Experiment 2, the number of known devices was fixed at 5, and the number of unknown devices was set to 0, 1, 2, 3, 4, and 5, respectively. The number of test samples for each device was fixed at 150. The control schemes were the TRSN scheme for solving small-sample RF fingerprint recognition and the existing zero-sample RF fingerprint recognition scheme DML.

[0144] The experimental results of Experiment 2 are as follows: Figure 6 As shown in the figure. The results indicate that, under different numbers of unknown devices, the proposed scheme has the optimal adjusted Land coefficient compared to the control scheme. Furthermore, thanks to the K-value estimation algorithm proposed in this invention, the average number of clusters in the proposed scheme is closest to the actual number of devices, thus achieving the highest accuracy in estimating the number of devices.

[0145] In contrast, the TRSN scheme, being a small-sample rather than zero-sample identification scheme, always results in the same number of clusters as the number of devices in the training set when there is no support set. Therefore, its device count estimation accuracy is 1 when there are no unknown devices, and 0 when there are 1 or more unknown devices. The DML scheme, on the other hand, suffers from information loss due to its UMAP dimensionality reduction and inaccurate automatic cluster count estimation using the DBSCAN algorithm. Consequently, its average number of clusters is always greater than the actual number of devices, resulting in a relatively low device count estimation accuracy.

[0146] Experiment 3

[0147] Experiment 3 primarily investigates the performance comparison of the proposed scheme and the control scheme in zero-sample RF fingerprint recognition under different test set sample numbers. In Experiment 3, the number of known devices and the number of unknown devices were both fixed at 5, and the number of test samples for each device was set to 30, 60, 90, 120, and 150, respectively. The control schemes are the TRSN scheme for solving small-sample RF fingerprint recognition and the existing zero-sample RF fingerprint recognition scheme DML.

[0148] The experimental results of Experiment 3 are as follows Figure 7 As shown in the figure. The results indicate that, under different test sample numbers, the proposed scheme has the optimal adjusted Land coefficient compared to the control scheme. Furthermore, thanks to the K-value estimation algorithm proposed in this invention, the average number of clusters in the proposed scheme is closest to the actual number of devices, thus achieving the highest accuracy in estimating the number of devices.

[0149] In contrast, the TRSN scheme, being a small-sample rather than zero-sample identification scheme, always results in the same number of clusters as the number of devices in the training set when there is no support set. Therefore, its device count estimation accuracy is zero when unknown devices are present. The DML scheme, due to information loss caused by UMAP dimensionality reduction and the inaccurate automatic clustering estimation of the DBSCAN algorithm, has an average cluster count greater than the actual number of devices under different test sample sizes, resulting in a relatively low device count estimation accuracy.

[0150] For the proposed scheme, under the condition of a small number of test samples, the distribution of device signal sample characteristics is unstable and highly random. Therefore, the accuracy of device number estimation is reduced. However, when the number of test samples for each device is greater than 90, the accuracy of device number estimation can approach 1, which proves the effectiveness of the proposed scheme.

[0151] Experiment 4

[0152] In Experiment 4, the number of known devices was fixed at 5, the number of unknown devices was also fixed at 5, the number of test samples for each device was set to 150, and the total number of test samples was 1500. Under these conditions, Experiment 4 compared the time and space complexity of the proposed scheme with the control scheme, and the results are shown in the table below.

[0153] Radio frequency fingerprint recognition solution Recognition time (seconds) Number of neural network parameters TRSN 173.869 45345892 DML 114.487 1365600 The proposed solution 9.985 25320

[0154] The results show that, thanks to the preprocessing steps and triplet loss function used in the technical solution provided by this invention, the triplet network used in the solution can simplify the neural network structure and reduce the size of the neural network while maintaining excellent feature extraction performance, thereby significantly reducing the number of parameters in the neural network. Therefore, relying on the highly efficient and lightweight triplet network feature extractor, the technical solution provided by this invention has the lowest recognition time and the lowest number of parameters, thus exhibiting better real-time performance and practicality compared to the control scheme, and is more suitable for real-world application scenarios.

[0155] In summary, this invention provides a zero-shot RFID fingerprint intelligent recognition method suitable for practical applications. This method introduces random fusion data augmentation and VMD to achieve effective data preprocessing, ensuring recognition performance under extremely small sample conditions. Furthermore, the proposed triplet network feature extractor can extract highly discriminative RFID fingerprint features, and compared to existing solutions, it has lower complexity and higher real-time performance.

[0156] In addition, this invention proposes a K-value estimation method applicable to the K-means++ unsupervised clustering algorithm, which can ensure that the proposed scheme can accurately estimate the number of equipment categories, and significantly improve the reliability, effectiveness and security of the scheme.

[0157] In summary, the technical solution provided by this invention is more suitable for practical application scenarios than existing solutions, alleviating the problem that current radio frequency fingerprint intelligent recognition is not very practical in practical application scenarios, especially in non-cooperative communication scenarios.

[0158] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A zero-shot radio frequency fingerprinting method based on triplet networks and unsupervised clustering, characterized in that, Includes the following steps: Collect radio frequency fingerprints from several radio stations and create an radio frequency fingerprint dataset; The triplet network is trained using the aforementioned radio frequency fingerprint dataset to obtain an optimized triplet network. The triplet network includes three one-dimensional convolutional layers, each of which is followed by a batch normalization layer and a max pooling layer. The convolutional layers are used to extract feature vectors of radio frequency fingerprints at different scales. The RF fingerprint data to be tested is preprocessed, and the preprocessed RF fingerprint data is input into the optimized triplet network to extract feature samples, specifically including: The RF fingerprint data to be tested is preprocessed based on variational mode decomposition. The preprocessed RF fingerprint data includes 5 vectors. ;in, This refers to the low-frequency component of the original signal sample; and They are respectively The real and imaginary parts of the fast Fourier transform; for The instantaneous frequency; for The first mode after the difference sequence is processed by variational mode decomposition; The preprocessed radio frequency fingerprint data is input into the optimized triplet network, and the corresponding feature samples are extracted. The estimated number of devices is obtained based on the feature samples, and the radio frequency fingerprint recognition result is calculated using a clustering algorithm.

2. The zero-shot radio frequency fingerprinting method based on triplet networks and unsupervised clustering according to claim 1, characterized in that, The process of collecting radio frequency fingerprints from several radio stations to create a radio frequency fingerprint dataset specifically includes: A sliding window is set to sample the radio frequency signal. The sampling start point is determined by calculating the posterior probability that each sampling point in the sliding window is the starting point of the power-on transient signal. Starting from the sampling start point, a radio frequency signal of a preset length is extracted as the power-on transient signal. The power-on transient signal is regularized to obtain a regular transient signal, and the envelope of the regular transient signal is extracted by Hilbert transform as the radio frequency fingerprint of the radio equipment. The radio frequency fingerprints of several radio stations are collected and combined to form the radio frequency fingerprint dataset.

3. The zero-shot radio frequency fingerprinting method based on triplet networks and unsupervised clustering according to claim 1, characterized in that, The step of training the triplet network using the radio frequency fingerprint dataset to obtain an optimized triplet network specifically includes: The samples in the radio frequency fingerprint dataset are preprocessed, and the preprocessed samples are input into the triplet network. The training feature vector is extracted by the triplet network. Based on the training feature vector, the training loss value is calculated using a preset loss function; Based on the training loss value, the partial derivative of the training loss value with respect to each trainable parameter in the triplet network is calculated using the backpropagation algorithm. Then, the trainable parameters are optimized using a preset optimization algorithm to obtain optimized network parameters, thereby obtaining an optimized triplet network.

4. The zero-shot radio frequency fingerprinting method based on triplet networks and unsupervised clustering according to claim 3, characterized in that, The preprocessing of samples in the radio frequency fingerprint dataset specifically includes: A preset number of radio frequency fingerprints are selected from the radio frequency fingerprint dataset as the original sample set, and the size M of the original sample set and the number N of original samples used for each fusion are determined. Construct a fusion label table; wherein the fusion label table includes all sample combinations of N samples selected from the original sample set; Each sample in the aforementioned sample combination is fused to obtain several enhanced samples; wherein, each enhanced sample corresponds to a sample combination, and the enhanced sample is obtained by averaging the samples in the corresponding sample combination; The original samples and the augmented samples are combined into an augmented training set, and each sample in the augmented training set is preprocessed based on variational mode decomposition.

5. The zero-shot radio frequency fingerprinting method based on triplet networks and unsupervised clustering according to claim 3, characterized in that, The further optimization of the triplet network specifically includes: After updating the optimized network parameters to the triplet network, the samples of the verification set in the radio frequency fingerprint dataset are input into the triplet network after the parameters are updated, and the verification feature vector is obtained. The verification loss value is calculated based on the verification feature vector using the preset loss function; When the verification loss value is greater than the preset loss value threshold, the optimized network parameters are optimized again according to the radio frequency fingerprint dataset using a preset optimization algorithm, and the optimized network parameters are updated in the triplet network, and then verified again using the verification set. When the verification loss value is less than the preset loss value threshold, the corresponding network parameters are used as the optimal network parameters and input into the triplet network to obtain the optimized triplet network.

6. The zero-shot radio frequency fingerprinting method based on triplet networks and unsupervised clustering according to claim 1, characterized in that, The step of estimating the number of devices based on the feature samples specifically includes: Determine the search range for the number of radio devices, wherein the search range includes several estimated device counts; The feature samples are clustered using the K-means++ clustering algorithm, and the values ​​of the several estimated devices are used as the pre-input K values ​​of the K-means++ clustering algorithm to obtain several clustering results. The silhouette coefficient and Calinski-Harabaz index of the clustering results are calculated respectively; wherein, one estimated device value corresponds to one silhouette coefficient and one Calinski-Harabaz index; Based on the contour coefficient and the Calinski-Harabaz index, the evaluation scores of the several estimated device counts are calculated respectively, and the estimated device count corresponding to the highest evaluation score is taken as the estimated number of devices.

7. The zero-shot radio frequency fingerprinting method based on triplet networks and unsupervised clustering according to claim 6, characterized in that, The step of calculating the evaluation scores for the several estimated device values ​​based on the contour coefficient and the Calinski-Harabaz index specifically includes: Based on several profile coefficients and several Calinski-Harabaz indices corresponding to several estimated device values, an evaluation score is calculated for each estimated device value using an evaluation formula; wherein, the evaluation formula is specifically: ； in, This represents the evaluation score corresponding to the current estimated number of devices. The profile coefficient corresponding to the current estimated number of equipment values. It is the minimum value among several contour coefficients. The maximum value among the several contour coefficients, The Calinski-Harabaz index corresponds to the current estimated number of devices. It is the minimum value among several Calinski-Harabaz exponents. It is the maximum value among several Calinski-Harabaz indices.

8. The zero-shot radio frequency fingerprinting method based on triplet networks and unsupervised clustering according to claim 1, characterized in that, The radio frequency fingerprint recognition result calculated by the clustering algorithm specifically includes: The estimated number of devices is set as the pre-input K value of the K-means++ clustering algorithm, and the feature samples are clustered by the K-means++ clustering algorithm to obtain the radio frequency fingerprint recognition result.

Images