Small sample radiation source individual identification method

Abstract

The invention relates to a small sample radiation source individual identification method. The method comprises the following steps: performing data enhancement on a time domain signal sample of a radiation source radio frequency signal to obtain a first enhanced sample; performing data enhancement on a frequency domain signal sample of the radiation source radio frequency signal, and then converting the frequency domain signal sample back to the time domain to obtain a second enhanced sample; converting the first enhanced sample and the second enhanced sample to a frequency domain to obtain a third enhanced sample and a fourth enhanced sample; pre-training a frequency domain encoder and a time domain encoder by using the obtained enhanced sample by adopting a comparative learning method; constructing a radiation source individual identification model comprising the pre-trained time domain encoder and classifier; performing fine tuning on the radiation source individual identification model by using labeled small sample data of a radiation source radio frequency signal; and inputting a to-be-classified sample into the fine-tuned radiation source individual identification model to obtain a classification result. According to the method, efficient and accurate communication radiation source individual identification can be realized under the condition of limited labeled samples.

Description

Technical Field

[0001] The present invention relates to the technical field of radiation source individual recognition, and particularly to a small-sample radiation source individual recognition method. Background Art

[0002] Traditional radiation source recognition algorithms can be roughly divided into three categories of fingerprint feature extraction methods based on transient signals, based on steady-state signals, and based on mechanism modeling. A conventional recognition system usually includes signal preprocessing, fingerprint feature extraction, and classification and recognition modules. After uniformly preprocessing the signals, radio frequency fingerprint features are defined and extracted based on expert knowledge and combined with signal processing and mathematical tools, and finally the extracted features are used for classification and recognition.

[0003] The fingerprint feature extraction method based on transient signals has the advantages of obvious features and being unaffected by transmission information. However, the disadvantage is that it is necessary to detect the starting point of the signal, and the sample collection is difficult. The fingerprint feature extraction method based on steady-state signals has the advantages of easy sample collection, various feature forms, and good performance in large-scale data sets. The disadvantages are being affected by transmission information, the features are easily submerged, and the computational complexity is high. The fingerprint feature extraction method based on mechanism modeling has the advantages of strong interpretability and clear theoretical boundaries. The disadvantage is that the development of device programmable technology has increased the difficulty of feature extraction. It can be seen that traditional algorithms have a relatively high requirement for professional knowledge level, have limitations in the face of complex feature extraction and multi-dimensional feature fusion, and have weak generalization ability in new scenarios and new devices.

[0004] The rise of deep learning has brought new opportunities for the development of communication radiation source individual recognition technology. In the field of communication radiation source individual recognition technology, the application of deep learning algorithms has achieved significant breakthroughs in recent years. These algorithms have successfully achieved high-precision recognition of different radiation source individuals by deeply learning and finely extracting the feature information in communication signals. However, although these deep learning algorithms show great potential in recognition performance, they are also accompanied by high computational complexity requirements and large training data set requirements. In actual application scenarios, obtaining such a large-scale and completely labeled training data is often a difficult task.

[0005] In view of the above challenges, how to achieve efficient and accurate communication radiation source individual recognition under the condition of limited labeled samples has become the research focus in this field. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a small-sample radiation source individual recognition method, which can achieve efficient and accurate communication radiation source individual recognition under the condition of limited labeled samples.

[0007] The technical solution adopted by the present invention to solve its technical problems is: to provide a small-sample radiation source individual recognition method, including the following steps:

[0008] Preprocess the tagless auxiliary data of the radiation source radio frequency signal to obtain time-domain signal samples, and perform data augmentation on the time-domain signal to obtain the first augmented sample;

[0009] Convert the time-domain signal samples to the frequency domain to obtain frequency-domain signal samples, perform data augmentation on the frequency-domain signal samples and then convert them back to the time domain to obtain the second augmented sample;

[0010] Convert the first augmented sample and the second augmented sample to the frequency domain to obtain the third augmented sample and the fourth augmented sample;

[0011] Adopt the contrastive learning method and use the obtained augmented samples to pre-train the frequency-domain encoder and the time-domain encoder;

[0012] Construct a radiation source individual recognition model including the pre-trained time-domain encoder and the classifier;

[0013] Use the tagged small sample data of the radiation source radio frequency signal to fine-tune the radiation source individual recognition model;

[0014] Input the sample to be classified into the fine-tuned radiation source individual recognition model to obtain the classification result.

[0015] Furthermore, when performing data augmentation on the time-domain signal and the frequency-domain signal, different strategies and intensities are adopted for the I-channel signal and the Q-channel signal respectively to perform data augmentation.

[0016] Furthermore, the adoption of different strategies and intensities for the I-channel signal and the Q-channel signal respectively includes:

[0017] Select one of the operations of only enhancing the I-channel signal, only enhancing the Q-channel signal, enhancing the same positions of the IQ two channels, and enhancing the asymmetric positions of the IQ two channels as the enhancement strategy;

[0018] Adopt different enhancement intensities for the I-channel signal and the Q-channel signal respectively.

[0019] Furthermore, the enhancement operation includes:

[0020] Adopt at least one of local scaling, local distortion, local drift, and local perturbation, and combine local masking for data augmentation.

[0021] Furthermore, the local scaling, local distortion, local drift, local perturbation, and local masking are realized by performing corresponding scaling, distortion, drift, perturbation, and masking operations on local continuous signals by using random numbers.

[0022] Further, the contrastive learning method is adopted to pre-train the frequency-domain encoder and the time-domain encoder using the obtained augmented samples, including:

[0023] The frequency-domain encoder is used to extract features from the third augmented sample and the fourth augmented sample respectively, obtaining a third augmented feature and a fourth augmented feature;

[0024] The time-domain encoder is used to extract features from the first augmented sample and the second augmented sample respectively, obtaining a first augmented feature and a second augmented feature;

[0025] The distances between the first augmented feature and the second augmented feature, the third augmented feature and the fourth augmented feature, the first augmented feature and the third augmented feature, and the second augmented feature and the fourth augmented feature are set as the first distance, the second distance, the third distance, and the fourth distance respectively;

[0026] The loss function of the time-domain encoder is calculated according to the first distance, the third distance, and the fourth distance, and the loss function of the frequency-domain encoder is calculated according to the second distance, the third distance, and the fourth distance. The time-domain encoder and the frequency-domain encoder are updated using the gradient propagation of the loss function.

[0027] Further, the loss function of the time-domain encoder is the sum of the squares of the first distance, the third distance, and the fourth distance.

[0028] Further, the loss function of the frequency-domain encoder is the sum of the squares of the second distance, the third distance, and the fourth distance.

[0029] Further, the use of the labeled small sample data of the radiation source RF signal to fine-tune the radiation source individual recognition model includes:

[0030] The labeled small sample data is input into the radiation source individual recognition model to obtain a predicted label;

[0031] The radiation source individual recognition model is updated using the gradient propagation of the cross-loss entropy of the predicted label.

[0032] Beneficial Effects

[0033] Due to the adoption of the above technical solution, compared with the prior art, the present invention has the following advantages and positive effects:

[0034] (1) By introducing multiplicative enhancement methods (such as local scaling and local distortion) in the individual recognition of small-sample radiation sources, combined with enhancements such as masking, local drift, and local perturbation, the present invention significantly improves the accuracy of RF fingerprint feature extraction; multiplicative enhancement can effectively simulate the changes of RF signals under different environments, hardware conditions, or noise interferences, changing the amplitude or shape of the signals, thereby enhancing the adaptability of the model to complex signals; by transforming the signals in various ways, the model can more accurately extract features from complex RF signals, and further improve the overall accuracy of radiation source individual recognition.

[0035] (2) The time-domain and frequency-domain enhancement strategies provided by the present invention enable the training samples to be extended in multiple dimensions by freely combining different enhancement methods, enriching the diversity of the samples; the asymmetry of time-domain and frequency-domain enhancements, that is, different enhancement intensities, combination numbers, and position processing methods are adopted in the time domain and the frequency domain, further improves the diversity and complexity of the training samples. This diverse enhancement method enables the model to perform more robustly when facing new RF signals, thus effectively improving the generalization ability of the model and reducing the risk of overfitting.

[0036] (3) Aiming at the IQ imbalance problem in RF fingerprint features, the present invention proposes an enhancement strategy for IQ imbalance, and enhances the model's recognition ability for IQ imbalance phenomena through differential processing, thereby improving the accuracy of RF signal feature extraction; in the face of different hardware devices, signal propagation paths, or interference conditions, it can stably and effectively identify signal features, further enhancing the model's recognition ability in complex environments.

[0037] (4) By introducing a frequency-domain feature-guided contrast learning method, calculating the distances (such as Mahalanobis distance, cosine distance, etc.) of samples in the time-frequency domain space, and using these distances as training constraints to guide the optimization process of the feature extraction network; through the guidance of time-frequency domain representation consistency, the model can make more full use of the multi-level information of the signal, learn richer feature representations. This guidance mechanism improves the model's ability to express signal features, makes feature learning more efficient, and further improves the recognition accuracy and training stability of the model. Especially under complex signal conditions, it enhances the robustness and robustness of the model. Description of the Drawings

[0038] Figure 1 is the radiation source signal data enhancement method of the embodiment of the present invention;

[0039] Figure 2 is the schematic diagram of model training principle of the embodiment of the present invention;

[0040] Figure 3 is the schematic flow chart of the embodiment of the present invention. Detailed Embodiments

[0041] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims of this application.

[0042] An embodiment of the present invention relates to a small-sample radiation source individual recognition method based on contrast learning, which adopts time-domain and frequency-domain asymmetry enhancement and IQ imbalance enhancement strategies for data enhancement, and introduces frequency-domain feature-guided contrast learning to achieve radiation source individual recognition under small-sample conditions.

[0043] As Figure 3 shown, in the time-domain and frequency-domain asymmetry enhancement, it includes a combination of one or more operations such as locally scaling a local continuous signal with a random number, locally distorting a local continuous signal with a random number, locally drifting a local continuous signal with a random number, locally perturbing a local continuous signal with a random number, and locally masking a local continuous signal with a random number. The IQ imbalance enhancement strategy is reflected in different ways of processing such as position asymmetry and enhancement intensity asymmetry between the I-channel and Q-channel signals.

[0044] The frequency-domain feature-guided contrast learning calculates the distance of samples in the time-frequency domain space and adds the time-frequency contrast distance as a training constraint condition to the calculation of the loss function.

[0045] As Figure 2 and Figure 3 shown, this embodiment specifically includes the following steps:

[0046] 1. Use the unlabeled auxiliary data set of the radiation source signal for pre-training.

[0047] 1. Preprocess the data: Normalize the signal

[0048] 2. Generate time-domain enhanced samples through the time-domain enhancement library

[0049] (1) Select at least one enhancement operation from local scaling, local distortion, local drift, and local perturbation, and form a random enhancement operation combination with local masking.

[0050] ① Locally scale the sample:

[0051] where and is a random constant.

[0052] ②Perform local distortion on the sample:

[0053] where and are a set of random numbers of the same size as the signal, and are multiplied point - by - point with the signal using the Hadamard product.

[0054] ③Perform local drift on the sample:

[0055] where and is a random constant.

[0056] ④Perform local perturbation on the sample:

[0057] where and are a set of random numbers of the same size as the signal.

[0058] ⑤Perform local masking on the sample:

[0059] where and are a set of random numbers of the same size as the signal.

[0060] (2) After determining the enhancement operation, randomly select one enhancement strategy: enhance only the I - channel signal; enhance only the Q

[0061] - channel signal; enhance the same positions of both the I and Q channels; enhance the asymmetric positions of the I and Q channels:

[0062]

[0063] And use different enhancement intensities:

[0064]

[0065] 3. Convert the signal to the frequency domain through Fourier transform:

[0066] 4. Enhance the sample in the frequency domain through the frequency - domain enhancement library: Select one of local scaling, local distortion, local drift, and local perturbation for enhancement, and randomly combine it with local masking to form different frequency - domain enhancement strategies, and use asymmetric positions and different enhancement intensities for both the I and Q channels.

[0067] 5. After enhancement, convert it back to the time domain through inverse Fourier transform to obtain the frequency - domain enhanced sample

[0068] 6. Through the time - domain encoder E TExtract the features of the enhanced samples:

[0069] 7. Convert the enhanced samples to the frequency domain through Fourier transform:

[0070] 8. Extract the features of the enhanced samples through the frequency domain encoder E F Extract the features of the enhanced samples:

[0071] 9. Calculate the loss function of the time domain encoder:

[0072] 10. Calculate the loss function of the frequency domain encoder:

[0073] 11. Propagate the gradient of the loss function to update the time domain E T and the frequency domain encoder E F .

[0074] 12. Save the time domain encoder E T as the initial value of the encoder for subsequent fine-tuning.

[0075] II. Use the small labeled sample dataset in the fine-tuning stage.

[0076] 1. Add a fully connected classifier G after the pre-trained time domain encoder E T and obtain the predicted labels after inputting the samples

[0077] 2. Calculate the cross entropy:

[0078] 3. Use the cross entropy as the loss function to propagate the gradient and update the time domain encoder E T and the classifier G.

[0079] 4. Model deployment: Input the samples to be classified into the time domain encoder and the classifier to obtain the classification results:

Claims

1. A method for identifying individual radiation sources of a small sample, characterized in that: The following steps are involved: Preprocessing the unlabeled auxiliary data of the radio frequency signal of the radiation source to obtain a time domain signal sample, and performing data enhancement on the time domain signal to obtain a first enhanced sample; Convert the time domain signal sample to the frequency domain to obtain a frequency domain signal sample, perform data enhancement on the frequency domain signal sample and then convert it back to the time domain to obtain a second enhanced sample; Convert the first enhanced sample and the second enhanced sample into the frequency domain to obtain a third enhanced sample and a fourth enhanced sample; The contrastive learning method is adopted to pre-train the frequency domain encoder and the time domain encoder using the obtained enhanced samples; Constructing a radiation source individual recognition model including the pre-trained time domain encoder and classifier; Fine-tune the radiation source individual recognition model using labeled small sample data of the radiation source radio frequency signal; The sample to be classified is input into the fine-tuned radiation source individual recognition model to obtain a classification result.

2. The method according to claim 1, characterized in that When data enhancement is performed on the time domain signal and the frequency domain signal, different strategies and strengths are used for the I-path signal and the Q-path signal respectively to perform data enhancement.

3. The method according to claim 2, characterized in that The different strategies and strengths are respectively adopted for the I-path signal and the Q-path signal, including: Selecting one of performing an enhancement operation on only the I-path signal, performing an enhancement operation on only the Q-path signal, performing an enhancement operation on the same position of the I and Q-path signals, and performing an enhancement operation on an asymmetric position of the I and Q-path signals as an enhancement strategy; Different enhancement strengths are used for the I-channel signal and the Q-channel signal.

4. The method according to claim 3, characterized in that The enhancement operations include: At least one of local scaling, local distortion, local drift and local perturbation is used in combination with local masking for data enhancement.

5. The method according to claim 4, characterized in that The local scaling, local distortion, local drift, local disturbance, and local masking are achieved by using random numbers to perform corresponding scaling, distortion, drift, disturbance, and masking operations on the local continuous signal.

6. The method according to claim 1, characterized in that The contrastive learning method is used to pre-train the frequency domain encoder and the time domain encoder using the obtained enhanced samples, including: Using a frequency domain encoder to perform feature extraction on the third enhanced sample and the fourth enhanced sample respectively, to obtain a third enhanced feature and a fourth enhanced feature; Using a time domain encoder to extract features from the first enhanced sample and the second enhanced sample respectively, to obtain a first enhanced feature and a second enhanced feature; Setting the distances between the first enhancement feature and the second enhancement feature, the distances between the third enhancement feature and the fourth enhancement feature, the distances between the first enhancement feature and the third enhancement feature, and the distances between the second enhancement feature and the fourth enhancement feature to be a first distance, a second distance, a third distance, and a fourth distance, respectively; The loss function of the time domain encoder is calculated according to the first distance, the third distance and the fourth distance, the loss function of the frequency domain encoder is calculated according to the second distance, the third distance and the fourth distance, and the time domain encoder and the frequency domain encoder are updated by using the loss function gradient propagation.

7. The method according to claim 6, characterized in that The loss function of the time domain encoder is the sum of the squares of the first distance, the third distance and the fourth distance.

8. The method according to claim 6, characterized in that The loss function of the frequency domain encoder is the sum of the squares of the second distance, the third distance and the fourth distance.

9. The method according to claim 1, characterized in that: The method of fine-tuning the radiation source individual recognition model using labeled small sample data of the radiation source radio frequency signal includes: Inputting the labeled small sample data into the radiation source individual identification model to obtain a predicted label; The radiation source individual recognition model is updated using the cross-loss entropy gradient propagation of the predicted label.

Images