Unsupervised communication emitter individual identification method based on bispectrum feature contrast learning

Abstract

The application discloses an unsupervised communication radiation source individual identification method based on bispectrum feature contrast learning, and the process of the method is as follows: firstly, a residual network shared by two parameters is used as a backbone network to perform feature contrast learning, then the rectangular integral bispectrum features of augmented samples are input into a contrast learning module to further learn feature representations with better distinguishability, so that the feature separability between different radiation source samples is enhanced; secondly, the extracted new feature representations are used for contrast learning at a clustering cluster level to complete a classification identification task. Through experiments on a measured ultrashort wave communication radio station data set, the method has better identification effect than other unsupervised learning algorithms, and can achieve an identification accuracy of 77.8%.

Description

Technical Field

[0001] This invention relates to the field of radiation source identification technology, and in particular to an unsupervised communication method for individual radiation source identification based on bispectral feature contrast learning. Background Technology

[0002] Specific Emitter Identification (SEI) typically refers to associating intercepted communication signals with a specific type of radio station through signal feature matching, thereby achieving tactical objectives such as reconnaissance and identification. The accuracy of this matching directly impacts the reliability of intelligence gathering and even the overall operational effectiveness of the electronic warfare system (K.C. So, W. Prokopiw, and Y.T. Chan. "Modulation identification of digital signals by the wavelet transform". In: IEE Proceedings-Radar, Sonar and Navigation 147.4 (2002), pp. 169–176 (cit. on p. 1).). In actual battlefield environments, we often face the problem of classifying radiation source signals with a large amount of unlabeled information, i.e., unsupervised identification, which mainly consists of two key stages: feature extraction and classifier design. Among these, how to extract distinguishable data features from these unlabeled radiation source signals significantly affects the performance of downstream classification and identification tasks.

[0003] Traditional unsupervised communication radiation source identification often combines manual features with unsupervised algorithms. After preprocessing the received radiation source signal, subtle features of the signal are extracted manually, such as time-frequency analysis (Yuan Y, Huang Z, Hao W, et al. Specific emitter identification based on Hilbert-Huangtransform-based time-frequency-energy distribution features(J). Communications Iet, 2014, 8(13):2404-2412.), modulation analysis, higher-order spectrum (Zhang XD, Shi Y, Bao ZA new feature vector using selected bispectra for signal classification with application in radar target recognition(J). IEEE Transactions on Signal Processing, 2001, 49(09):1875-1885.), and modeling based on the nonlinear principle of transmitter devices (AC Polak, S. Dolatshahi and DL Goeckel, "Identifying Wireless Users via Transmitter Imperfections," in IEEE Journal on Selected Areas). (e.g., in Communications, vol.29, no.7, pp.1469-1479, August 2011, doi:10.1109 / JSAC.2011.110812.) These artificial features are then used to classify the radiation sources using unsupervised algorithms such as clustering, thus achieving individual identification of the radiation sources. However, these artificial features have significant limitations; they can only reflect some features of the radiation source signal and lack strong distinguishability, often resulting in unsatisfactory identification results. With the continuous research of deep learning, the powerful nonlinear fitting ability of deep neural networks (DNNs) has demonstrated excellent performance in image recognition and face detection, and is beginning to show its potential in the technology of individual identification of communication radiation sources.

[0004] In the literature (China Shipbuilding Industry Corporation 724 Research Institute. A Multi-Model Comprehensive Classification Method for Radar Source Signals Based on Deep Learning: CN202110751828.0(P). 2021-09-07.), Wang Jiaming et al. constructed a comprehensive network model using deep convolutional neural networks and long short-term memory networks, achieving intelligent identification of radar radiation source signals with strong generalization ability; Li Lixin et al. (Northwestern Polytechnical University, Shanghai Satellite Engineering Research Institute. Radio Signal Modulation Recognition Network and Implementation Method Based on Hybrid Neural Network: CN202011368021.0(P). 2021-03-26.) further utilized the temporal and spatial state characteristics of signals by dividing the feature information extracted by the convolutional layer into dimensions and controlling the cyclic threshold, thus improving the accuracy of the classification. Classification performance of modulated signals; Xie Cunxiang et al. (Xie Cunxiang, Zhang Limin, Zhong Zhaogen. Identification of specific radiation sources based on Hilbert-Huang transform and adversarial training (J). Systems Engineering and Electronics, 2021, 43(12):3478-3487.DOI:10.12305 / j.issn.1001-506X.2021.12.08.) established a model that integrates Hilbert-Huang transform and adversarial training in the research process. In the process of processing, the key time and frequency points of the radiation source signal and their corresponding energy values ​​are input into the convolutional neural network for training. Even when the training samples are small, good recognition results can be achieved; The literature (S.Wang, H.Jiang, X.Fang, Y.Ying, J.Li The paper "Radio Frequency Fingerprint Identification Based on Deep Complex Residual Network" by B. Zhang and L. Ying combines feature extraction and classification processes using a deep complex residual network, establishing an end-to-end model suitable for radiation source identification and improving the identification accuracy. The paper "Differential Complex-Valued Convolutional Neural Network-Based Individual Recognition of Communication Radiation Sources" by L. Ying, J. Li and B. Zhang uses a differential complex-valued convolutional neural network to capture the nonlinear features of baseband I / Q signals of 20 similar communication radiation sources, achieving a recognition rate of 99.7%. The paper "Qu Lingzhi, Yang Junan, Liu Hui, Huang Keju" also demonstrates this.A novel approach for identifying communication radiation sources by embedding a two-layer attention mechanism in a residual network is proposed, which improves the identification accuracy and has good stability. (in Chinese) Identification (J). Systems Engineering and Electronics (English Edition), 2022, 33(2):354-359. DOI:10.23919 / JSEE.2022.000037.) First, the network is trained on labeled samples, and then semi-supervised learning is used to detect unlabeled samples and automatically label new samples, realizing the dynamic identification of individuals from unknown radiation sources.

[0005] Existing deep learning-based methods for identifying radiation sources in communication typically involve supervised learning on labeled datasets. However, in real-world non-cooperative communication, the captured radiation source data often lacks prior information, thus limiting the performance of supervised learning algorithms. In unsupervised learning algorithms, clustering algorithms can group feature vectors into different clusters without any labels, such as sparse embedded k-means clustering and multi-kernel k-means clustering with matrix-induced regularization. However, because these methods rely on manually generated features, most algorithms produce poor results on complex datasets. To address the problem of insufficient feature representation, deep clustering utilizes neural networks to extract representative information from images to obtain more discriminative feature representations, thus facilitating downstream clustering tasks (Caron M, Bojanowski P, Joulin A, et al. Deep clustering for unsupervised learning of visual features(C) / / Proceedings of the European conference on computer vision(ECCV).2018:132-149.). However, these algorithms often iteratively group features and use subsequent assignments to update the deep network. The alternation between feature representation learning and clustering often leads to error accumulation, which in turn affects clustering performance.

[0006] In summary, deep neural networks perform well in classification and recognition problems on labeled datasets, but often fail to achieve satisfactory results on unlabeled data. Summary of the Invention

[0007] The purpose of this invention is to address the problems of difficulty in extracting features from unlabeled communication radiation source data and low classification accuracy. By applying contrastive learning theory, this invention proposes an unsupervised method for individual identification of communication radiation sources based on bispectral feature contrastive learning.

[0008] The technical solution to achieve the purpose of this invention is: an unsupervised communication radiation source individual identification method based on bispectral feature contrast learning, comprising the following steps:

[0009] Step 1: Build the network model. Input the dataset χ, set the number of training iterations E, batch size N, and hyperparameter τ. I τ C Number of categories M;

[0010] Step 2, Data Preprocessing: Normalize the time-domain signal data and extract data samples;

[0011] Step 3, Data Augmentation: For each signal sample, randomly crop, add noise, and flip the sample. Apply each data augmentation method independently with a set probability to generate positive samples.

[0012] Step 4, Bispectral Feature Extraction: Extract bispectral features for each sample after data augmentation, transforming the one-dimensional time-domain signal into a two-dimensional feature matrix;

[0013] Step 5: Select a batch of data from dataset χ Two data augmentation methods T are randomly selected. a ,T b Calculate the sample-to-sample contrast loss L ins ;

[0014] Step 6: Calculate the cluster contrast loss L clu The overall loss value L is calculated, and the network f,g is updated by minimizing L. I ,g C Parameters;

[0015] Step 7: For each sample x in the dataset, extract features using h = f(x) and c = argmaxg C (h) Calculate the cluster assignment for each sample;

[0016] Step 8: Output the one-hot encoding of each cluster to complete the identification of individual radiation sources in unsupervised communication.

[0017] Furthermore, by utilizing the similarity between the original signal sample x and its positive examples, and the differences between negative examples, comparative learning is performed to extract more discriminative feature representations h. a ,h b Used for downstream classification and recognition tasks;

[0018] According to h a ,h b It is a feature representation obtained from the same data sample through two augmentation methods. When clustering, it should belong to the same class. A clustering loss function is designed, and comparative learning is carried out at the cluster level to complete the classification and recognition task.

[0019] Positive sample pairs are defined as two augmented samples of the same signal sample, and other sample pairs are defined as negative sample pairs.

[0020] Furthermore, the network model in step 1 is as follows:

[0021] The one-dimensional time series is transformed into a two-dimensional feature matrix and input into the contrastive learning module for training. The residual network is selected as the backbone network of the contrastive learning.

[0022] The residual network is constructed from residual blocks, each consisting of multiple cascaded convolutional layers and a short circuit. The outputs of both layers are summed, and then activated using the ReLU function to obtain the final output. These residual blocks are then concatenated to form deeper networks. Experiments are used to optimize the network parameters and determine the network model for feature comparison learning.

[0023] When training the model, the bispectral feature maps of 5 VHF radio stations were used as the input to the network. 1500 data samples were taken from each station. The bispectral features of each sample were calculated after data augmentation. Each sample contained 4096 data points, and the feature matrix dimension was set to 128×128. 60% of the data was randomly selected as the training set, 20% as the validation set, and 20% as the test set.

[0024] ResNet-18 and ResNet-34 were selected as the backbone networks, respectively. The Adam method was used for network optimization. The initial learning rate was set to 0.001, the number of training iterations E was set to 400, and the batch data size N was set to 32. The recognition results corresponding to different network depths were obtained. Finally, ResNet-34 was selected as the backbone network for comparative feature extraction.

[0025] Furthermore, in step 5, the sample pair contrast loss L is calculated. ins The details are as follows:

[0026] For a given signal sample x i Two data augmentation methods T are randomly selected with a set probability. a ,T b Construct sample pairs to obtain two related samples. Represented as Two related samples obtained by augmenting the same sample are denoted as a positive sample pair;

[0027] By inputting a pair of samples into a deep neural network f(·) with shared parameters for comparative training, features are extracted to obtain new feature representations. Represented as

[0028] Stack a two-layer nonlinear fully connected layer g I (·),pass Mapping the feature matrix to a subspace where contrastive loss is applied, in and The loss is calculated by comparison.

[0029] Pairwise similarity is measured by cosine distance, i.e.

[0030]

[0031] Where k1,k2∈{a,b},i,j∈[1,N];

[0032] Extract N segments of radiation source signal samples, and for each sample x i Two augmentation methods were applied to obtain 2N signal samples. For a specific sample There are 2N-1 sample pairs, and the augmented samples associated with them will be... Let 'positive' be the sample, and we get the positive sample pairs. The remaining 2N-2 pairs are all recorded as negative sample pairs;

[0033] In the entire dataset of radiation sources, specific samples are defined to optimize pairwise similarity. The comparative loss takes the following form:

[0034]

[0035] Where, τ I is a hyperparameter used to identify all positive pairs in the entire dataset. For each augmented sample, a sample-pair contrast loss is calculated, i.e.:

[0036]

[0037] Further, in step 6, the cluster contrast loss L is calculated. clu The details are as follows:

[0038] Characterize the new features obtained Input into clustering network g C In (·), They are grouped into the same category;

[0039] Set the dimensions of the network output matrix Y to satisfy Y a ∈R N×M Where N is the number of samples in each training batch, M is the number of clusters, and Y is the number of clusters. a Y b Outputs for each batch of samples under two data augmentations; since each sample belongs to only one cluster, the rows of Y should resemble a one-hot distribution;

[0040] When a data sample is projected into a space with dimension equal to the number of clusters, the i-th element of a feature is considered the probability of belonging to the i-th cluster, the i-th column of Y is considered the representation of the i-th cluster, and all columns should be distinct from each other.

[0041] Use another two-layer fully connected layer g C (·) Features Mapped to an M-dimensional subspace, represented as in It is matrix Ya The i-th row is the sample x i The output of the clustering network after augmentation method A;

[0042] Let matrix Y a The i-th column is That is, the representation of the i-th cluster after the first augmentation of the data sample; similarly, and Combining to form positive clusters The other 2M-2 cluster pairs are considered as negative cluster pairs;

[0043] Cosine distance is used to measure the similarity between cluster pairs, i.e.:

[0044]

[0045] Where k1,k2∈{a,b},i,j∈[1,M];

[0046] The following loss function is used to cluster With and Distinguish from all other clusters:

[0047]

[0048] Where, τ C These are the hyperparameters of the clustering network;

[0049] By traversing all clusters, and to avoid assigning most samples to the same class, the cluster contrastive loss Lcluster is used. clu The definition is as follows:

[0050]

[0051] in, It is the probability of cluster assignment. The entropy.

[0052] Further, in step 6, the overall loss value L is calculated, as follows:

[0053] By simultaneously optimizing the feature comparison and cluster comparison loss functions, the entire unsupervised communication radiation source classification and identification network is optimized, namely:

[0054] L = L ins +L clu

[0055] Compared with the prior art, the significant advantages of this invention are: (1) First, the residual network with two shared parameters is used as the backbone network for feature comparison learning. Then, the rectangular integral bispectral features of the augmented sample are input into the comparison learning module to further learn more discriminative feature representations, thereby enhancing the feature separability between different radiation source samples; (2) The extracted new feature representations are used to perform comparison learning at the cluster level to complete the classification and recognition task. Through experiments on the measured ultra-shortwave communication radio station dataset, this method has better recognition effect than other unsupervised learning algorithms and can achieve a recognition accuracy of 77.8%. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the principle of contrastive learning.

[0057] Figure 2 This is a schematic diagram of the augmentation of radiation source signal sample data.

[0058] Figure 3 These are bispectral diagrams of signal samples from three radio stations of the same model.

[0059] Figure 4 This is a schematic diagram of the network structure of a communication radiation source individual identification algorithm based on contrastive learning.

[0060] Figure 5 This is a flowchart of an algorithm for individual identification of communication radiation sources based on contrastive learning.

[0061] Figure 6 This is a schematic diagram of the frequency interface of the signal acquisition device.

[0062] Figure 7 This is a schematic diagram of the time-domain waveforms of some signals from two of the VHF radio stations.

[0063] Figure 8 This is a schematic diagram of the training results of the VHF radio dataset.

[0064] Figure 9 This is a schematic diagram of the test set confusion matrix.

[0065] Figure 10 This is a diagram comparing the recognition rates of three different features.

[0066] Figure 11 This is a diagram showing the average recognition rate (%) when three features are compared and learned to extract features of different dimensions. Detailed Implementation

[0067] To enable the direct and effective extraction of fingerprint information from a large number of unlabeled radiation source signals using deep neural networks, thereby facilitating downstream classification and recognition tasks, this invention introduces the concept of contrastive learning. A contrastive learning module is constructed using two structurally identical residual networks (ResNet). The rectangular integrated bispectral (SIB) features of each signal sample and its augmented data are input into the network for contrastive learning, in order to learn more discriminative feature representations and thus improve the overall recognition performance of the network.

[0068] 1. Basic Principles of Individual Identification Algorithm for Communication Radiation Sources Based on Comparative Learning

[0069] 1.1 Comparative Learning

[0070] Contrastive learning (CL) is an unsupervised learning algorithm. Its main idea is to construct positive and negative sample pairs through data augmentation, and then map the data to a feature representation space to maximize the similarity of positive pairs and minimize the similarity of negative pairs, thereby training the network to learn more discriminative feature representations.

[0071] In the identification of communication radiation sources, for a given signal sample x, two augmentation methods T are randomly selected from a series of data augmentation methods. a and T b Two related samples x were obtained. a and x b We treat these as positive sample pairs, and all other sample pairs as negative sample pairs. Then, we pass each pair of samples through two identical neural network encoders f(·) to obtain the encoded representations h of the two samples. a ,h b The output of encoder f(·) is then fed into a set of nonlinear fully connected layers g(·) to transform the data into another space, represented as z. a ,z b This additional step avoids information loss caused by contrastive loss, thereby improving network performance. Finally, by setting the contrastive loss function to maximize the feature similarity of positive sample pairs and minimize the feature similarity of negative sample pairs, the network is trained using input training samples to further extract more discriminative feature representations. The principle and process are as follows: Figure 1 As shown:

[0072] In the figure, Sim(z) a ,z bCosine similarity is a metric used to measure the similarity of sample pairs. In our experiments, we found that the contrastive learning module exhibits better radiation source identification performance when the extracted feature dimension is higher. While Euclidean distance is easily affected by feature dimension, cosine similarity, in contrast, still follows the property of being 1 when vectors are the same, 0 when they are orthogonal, and -1 when they are opposite in high dimensions. Therefore, we use cosine distance to measure the similarity of sample pairs, i.e.:

[0073]

[0074] 1.2 Sample Data Augmentation

[0075] Data augmentation is a technique for expanding a dataset by creating more samples from a smaller set. Currently, it's widely used in image recognition. For example, rotating, flipping, or cropping an image can augment it into at least four images. If every image in a dataset is augmented, the total number of images can quadruple. The core idea is to obtain similar data samples by performing operations on the original samples without altering their essential characteristics. In contrastive learning of radiation sources, the purpose of data augmentation is not to expand the sample size, but to treat augmented samples of the same signal as belonging to the same class, creating "pseudo-labels" to construct positive and negative sample pairs for comparative training to further extract feature representations.

[0076] Unlike image data, the communication radiation source signals we collect are one-dimensional sequence data, with temporal relationships between the sampling points in the samples. This requires that the temporal relationship between the newly generated samples and the original samples be taken into account during the augmentation operation. For example... Figure 2 As shown, in our experiments, we mainly adopted methods such as flipping, stretching, and adding noise to ensure semantic invariance between the beginning and end of the sequence.

[0077] 1.3 Bispectral Feature Analysis

[0078] When analyzing signals from actual radiation sources, many non-Gaussian problems are often encountered. Comparisons show that higher-order statistics can extract more statistical signal features and provide richer information compared to second-order statistics. Bispectral analysis is the lowest-order higher-order spectrum, with simple processing methods, and it includes amplitude and phase information of the signal, while also having a certain noise suppression effect. Rectangular integral bispectral analysis (SIB) is currently the best performing of the four integral bispectral methods, avoiding issues of duplicate or missing bispectral values. Therefore, in our experiments, we considered using the rectangular integral bispectral features of signal samples for comparative learning.

[0079] The bispectral can be represented as:

[0080]

[0081] In the formula: |B(ω1,ω2)| and φB(ω1,ω2) represent the amplitude and phase of the bispectral spectrum, respectively.

[0082] Let the continuous signal be b(t), the bispectral density can be represented as:

[0083]

[0084] In the formula: c 3b (τ1,τ2) is the third-order autocorrelation function of the signal b(t).

[0085]

[0086] SIB can be defined as:

[0087]

[0088] In the formula: S l For any integration path of SIB.

[0089] A sample was randomly selected from each of five identical FM VHF radios. The bispectral characteristics of the sample signals were analyzed using the bispectral estimation algorithm described above. The bispectral feature maps of three of these radios are then presented as follows: Figure 3 .

[0090] from Figure 3 It is known that even among different radio stations of the same model, there are certain differences in their bispectral characteristics. Therefore, this feature can be selected to identify individual radiation sources. Based on practical experience, the advantages of extracting signal features through bispectral transformation are: it preserves the phase and amplitude information of individual signals; it is insensitive to the choice of the time starting point; and it effectively suppresses additive Gaussian noise.

[0091] 2. Algorithm for Individual Identification of Communication Radiation Sources Based on Contrastive Learning

[0092] 2.1 Network Model

[0093] The basic idea of ​​the experiment is to use the similarity between the original signal sample x and its positive examples, and the differences between negative examples, to perform comparative learning and extract more discriminative feature representation h. a ,h b Used for downstream classification and recognition tasks. Then, based on h a ,h b It consists of feature representations obtained from the same data sample through two augmentation methods. During clustering, these representations should belong to the same class. A clustering loss function is designed, and comparative learning is performed at the cluster level to complete the classification and recognition task. In the experiment, we define positive sample pairs as two augmented samples of the same signal sample, and other sample pairs as negative sample pairs. The network structure diagram of this method is shown below. Figure 4As shown, the algorithm flow is as follows: Figure 5 As shown.

[0094] In this method, we transform the one-dimensional time series into a two-dimensional feature matrix and input it into the contrastive learning module for training. Convolutional Neural Networks (CNNs) have significant advantages in feature extraction and are often used to process two-dimensional and three-dimensional data. However, directly stacking shallow CNNs into deep networks can lead to a "degradation" problem. To leverage the powerful nonlinear feature fitting capabilities of deep networks, we consider selecting a residual network as the backbone network for contrastive learning.

[0095] Proposed by Kaiming He et al. in 2015, residual networks are constructed from residual building blocks, each consisting of multiple cascaded convolutional layers and a shortcut connection. The outputs of these convolutional layers are summed, and then activated using the ReLU function to obtain the final output. By concatenating residual blocks, deeper networks can be formed, thus addressing the vanishing gradient problem common in deep networks. Two main typical structures are ResNet-18 and ResNet-34. To select a superior network model for feature comparison learning, this invention optimizes the network parameters through experiments.

[0096] When training the model, the bispectral feature maps of five VHF radio stations were used as the network input. Each station had 1500 data samples, and each sample underwent data augmentation before bispectral feature calculation. Each sample contained 4096 data points, and the feature matrix dimension was set to 128×128. 60% of the data was randomly selected as the training set, 20% as the validation set, and 20% as the test set.

[0097] ResNet-18 and ResNet-34 were selected as the backbone networks, respectively. The Adam method was used for network optimization. The initial learning rate was set to 0.001, the number of training iterations was set to 400, and the batch size was 32. The recognition results for different network depths are shown in Table 1.

[0098] Table 1. Impact of different network depths on the accuracy of radiation source identification.

[0099]

[0100] The above experiments show that ResNet-34 extracts more discriminative feature representations in the comparative training of radiation source data, achieving better radiation source recognition performance compared to ResNet-18. This indicates that the deeper residual network did not suffer from network "degradation" during the training of radiation source data. Therefore, we selected ResNet-34 as the backbone network for comparative feature extraction, and its network parameter settings are shown in Table 2.

[0101] Table 2 ResNet34 Network Structure Diagram

[0102]

[0103] 2.2 Feature Comparison Learning Module

[0104] For a given signal sample x i We randomly select two data augmentation methods T with a certain probability. a ,T b Construct sample pairs to obtain two related samples. It can be represented as Two related samples obtained by augmenting the same sample are denoted as a positive sample pair. Then, this pair of samples is input into a deep neural network f(·) with shared parameters for comparative training to extract features and obtain new feature representations. It can be represented as To mitigate the information loss that might be caused by contrastive loss, we do not directly modify the feature matrix. Instead of performing comparative training, it stacks a two-layer nonlinear fully connected layer g. I (·),pass Mapping the feature matrix to a subspace where contrastive loss is applied, in and The contrastive loss is calculated to maximize the similarity of positive samples while minimizing the similarity of negative sample pairs. Pairwise similarity is measured using cosine distance, i.e.

[0105]

[0106] Where k1,k2∈{a,b},i,j∈[1,N].

[0107] In the experiment, since there was no usable label information for the radiation source signals under unsupervised conditions, positive and negative sample pairs could only be constructed based on "pseudo-labels" generated by data augmentation. Specifically, in the experiment, we extracted N segments of radiation source signal samples, and for each sample x... i Two augmentation methods were applied to obtain 2N signal samples. For a specific sample There are 2N-1 sample pairs, and the augmented samples associated with them will be... Let 'positive' be the sample, and we get the positive sample pairs. The remaining 2N-2 pairs are all recorded as negative sample pairs.

[0108] In order to optimize pairwise similarity within the entire dataset of radiation sources, specific samples can be defined. The comparative loss takes the following form:

[0109]

[0110] Where, τ I is a hyperparameter. Since we want to identify all positive pairs in the entire dataset, we compute the sample-pair contrast loss for each augmented sample, i.e.:

[0111]

[0112] 2.3 Clustering Network Module

[0113] After extracting features through the above comparative training, the resulting new feature representations will be... Input into clustering network g C In (·), ideally, should be Clustered into the same category. Based on this understanding, we set the dimension of the network output matrix Y to satisfy Y a ∈R N×M Where N is the number of samples in each training batch, M is the number of clusters, and Y is the number of clusters. a Y b The outputs are shown twice for each batch of samples after data augmentation. Since each sample belongs to only one cluster, ideally, the rows of Y should resemble a one-hot distribution. When the data samples are projected onto a space of dimension equal to the number of clusters, the i-th element of its features can be considered the probability of belonging to the i-th cluster. In this sense, the i-th column of Y can be seen as a representation of the i-th cluster, and all columns should be distinct from each other. g is used in the feature comparison. I Similarly, we use another two-layer fully connected layer g. C (·) Features Mapped to an M-dimensional subspace, represented as in, It is matrix Y a The i-th row can be considered as sample x i The output of the clustering network is obtained after augmentation method A.

[0114] Let matrix Y a The i-th column is That is, the representation of the i-th cluster after the first augmentation of the data sample; similarly, it is compared with... Combining to form positive clusters The remaining 2M-2 cluster pairs are considered negative cluster pairs. Cosine distance is still used to measure the similarity between cluster pairs, i.e.:

[0115]

[0116] Where k1,k2∈{a,b},i,j∈[1,M]. The following loss function is used to cluster... With and Distinguish from all other clusters:

[0117]

[0118] Where, τ C The hyperparameters of the clustering network are determined by traversing all clusters, and to avoid assigning the majority of samples to the same class, the clustering contrastive loss Lc is used. clu The definition is as follows:

[0119]

[0120] in, It is the probability of cluster assignment. The entropy.

[0121] By simultaneously optimizing the feature comparison and cluster comparison loss functions, the entire unsupervised communication radiation source classification and identification network is optimized, namely:

[0122] L = L ins +L clu (11)

[0123] 2.4 Algorithm Steps

[0124] The specific steps of the unsupervised communication radiation source individual identification algorithm based on contrastive learning proposed in the above analysis are shown in Table 3:

[0125] Table 3. Main steps of the unsupervised communication-based individual radiation source identification algorithm based on contrastive learning.

[0126]

[0127] 3. Experimental Results and Analysis

[0128] To evaluate the feasibility and effectiveness of the unsupervised communication radiation source identification algorithm based on contrastive learning, this section presents extensive experiments on a dataset of signals from five VHF radio stations of the same model. To more intuitively analyze the experimental results, we compare our proposed method (SIB / CL) with traditional methods combining artificial features and unsupervised learning algorithms: K-means clustering, density spatial clustering (DBSCAN), and density peak clustering (DPC).

[0129] We set up four sets of experiments. The first set was an unsupervised radiation source identification experiment, which mainly verified the feasibility of the proposed method (SIB / CL) by identifying unlabeled radiation source datasets. The second set was a comparison experiment with different pre-features. After augmenting the radiation source signal sample data, we extracted different artificial features and input them into the network for training to verify the difference in recognition performance of the algorithm under different artificial pre-feature extraction. The third set was a comparison experiment with different feature dimensions to verify the impact of extracting features of different dimensions through a contrastive learning network on the radiation source identification performance. The fourth set was a comparison experiment with different unsupervised algorithms. By comparing with several classic unsupervised communication radiation source identification algorithms, we verified the superiority of our algorithm.

[0130] 3.1 Experimental Data Acquisition

[0131] In the experiment, voice signals from five FM VHF radios of the same model were collected. These voice signals were data from conversations between fixed individuals. The radios operated at center frequencies of 35, 55, and 85 MHz, respectively, and were collected in "low power" mode. The acquisition scenario involved 50m diffraction, meaning there was a tall building obstructing the view between the receiver and the radio station at a distance of 50 meters. During the acquisition of the radio voice signals, the receiver acquired the zero intermediate frequency (IF) I / Q signal. Figure 6 Provides the signal frequency interface corresponding to the data acquisition device:

[0132] The receiver's parameter settings are shown in Table 4:

[0133] Table 4 Receiver parameter settings

[0134]

[0135] The time-domain signal waveforms, composed of 40,000 data points from two radio stations, are shown below. Figure 7 :

[0136] 3.2 Unsupervised Identification Experiment of Radiation Sources

[0137] To verify the feasibility of the SIB / CL algorithm, our experiments were conducted on signal data from five VHF radio stations of the same model. 1500 data samples were extracted from each station's signal, and the algorithm was trained 400 times with a batch size of 32. The accuracy curve as a function of the number of training iterations is shown below. Figure 8 .

[0138] After training, the network model was saved. The saved neural network model was then used to test the data that was not used in the training. The results are shown in Table 5.

[0139] Table 5 Experimental Results of the VHF Radio Station

[0140]

[0141] from Figure 9 It can be seen that the overall recognition rate of the five types of radiation sources is relatively evenly distributed. Although there is still room for improvement in the recognition rate, the method of further extracting feature representations through bispectral feature comparison learning is feasible for the unsupervised recognition of communication radiation sources.

[0142] 3.3 Comparison Experiment with Different Features

[0143] Because the conventional Fourier transform cannot describe the frequency variation of a signal over time, it cannot fully extract the subtle features of the signal. To understand the relationship between frequency and time, time-frequency analysis methods are needed. Wavelet transform, developed from the Fourier transform, is commonly used to analyze the time and frequency domain information of signals and has wide applications in radiation source feature extraction. Wavelet transform can transform a one-dimensional time-domain signal to a two-dimensional time-frequency plane. Additionally, time-domain waveform concatenation can be used to transform one-dimensional signal data into a two-dimensional matrix to adapt to the input dimension of a network.

[0144] In this experiment, we extracted time-domain waveforms, wavelet transforms, and rectangular integral bispectral features from communication data samples from five VHF radio stations. Each feature was then input into a contrastive learning module for 400 training iterations. 1500 samples were extracted from each radio station's communication data. When extracting bispectral features, each segment contained 4096 data points, with a data overlap ratio of 5%, a segment length of 128, and 128 FFT points. The results were processed as follows: Figure 10 :

[0145] As shown in the figure, with the increase of the number of experiments, inputting bispectral features into the contrastive learning network can extract more discriminative feature representations, and the effect is the best in radiation source identification. The time-frequency graph features obtained by wavelet transform can improve the feature representation by contrastive networks to a certain extent compared with the original time-domain waveform features, but the improvement in recognition performance is limited. This may be because bispectral features are more sensitive to signal augmentation transforms.

[0146] 3.4 Comparison Experiment of Different Feature Dimensions

[0147] To verify how the radiation source identification performance changes when the dimensions of the new features extracted through contrastive learning are altered, this invention classifies 32, 64, 96, 128, 160, 192, 224, and 256-dimensional features further extracted through the contrastive learning network using three different feature extraction methods during network testing. Ten experiments were conducted for each method to calculate the average recognition rate. The experimental results are shown in Table 6 and... Figure 11 As shown.

[0148] Table 6 shows the average recognition rate (%) of the three feature extraction methods when extracting different feature dimensions through comparative learning.

[0149]

[0150] From Table 6 and Figure 11 It can be observed that the recognition rate of the contrastive learning algorithm varies with the feature extraction dimension. Generally speaking, when the feature dimension is 32, 64, and 96, the algorithm's recognition performance does not show a significant advantage. However, when the feature dimension output by the contrastive learning network exceeds 160, the algorithm's recognition performance tends to stabilize, and the overall recognition rate significantly improves. This is because lower-dimensional network output features are insufficient to distinguish different radiation source signals, while higher-dimensional features can easily lead to overfitting. Therefore, in the experiment, we should select an appropriate feature extraction dimension for radiation source classification and recognition.

[0151] 3.5 Comparative Experiments of Different Unsupervised Algorithms

[0152] This group of experiments evaluated the adopted network model on an ultra-shortwave dataset and compared it with three previous clustering algorithms used for unsupervised communication radiation source identification, including K-means clustering, density spatial clustering (DBSCAN), and density peak clustering (DPC). In the experiments, we extracted bispectral features of different dimensions from signal samples and conducted radiation source identification experiments on the three clustering algorithms. The comparison with the SIB / CL algorithm using bispectral features of corresponding dimensions as network input is shown in Table 7. In this group of experiments, we set the training times for the contrastive learning network to 400 times, and the average recognition rate was taken for 10 trials in each group of experiments.

[0153] Table 7 Comparison of SIB / CL algorithm with different unsupervised algorithms

[0154]

[0155] The experimental results show that this method does not have a significant advantage when using contrastive learning networks to extract low-dimensional features. This may be because contrastive learning networks do not have significant feature discrimination when extracting low-dimensional features. However, as the feature dimension increases, we can see that the powerful nonlinear fitting ability of neural networks comes into play, and the recognition rate is significantly higher than other unsupervised algorithms.

[0156] 4. Conclusion

[0157] Deep neural networks possess powerful nonlinear fitting capabilities, offering significant advantages in feature extraction. This invention introduces a contrastive learning algorithm, utilizing a residual neural network (ResNet34) to construct a feature contrastive learning module for re-extracting bispectral features from communication radiation source data. Simultaneously, contrastive learning is further employed at the cluster level to enhance clustering performance. Experimental results demonstrate that on an unlabeled VHF / UHF communication radio dataset, the algorithm achieves a 77.8% recognition accuracy. Compared to other algorithms combining manual features with unsupervised learning, the proposed algorithm exhibits superior recognition performance, indicating its feasibility for solving the problem of individual communication radiation source identification without prior information. This algorithm employs a deep residual neural network in constructing the contrastive learning backbone network, demonstrating mature technology, stable recognition results, and high application value.

Claims

1. An unsupervised communication radiation source individual identification method based on bispectral feature contrast learning, characterized in that, Includes the following steps: Step 1: Build the network model and input the dataset. Set the number of training sessions Batch data volume hyperparameters , Number of categories M; Step 2, Data Preprocessing: Normalize the time-domain signal data and extract data samples; Step 3, Data Augmentation: For each signal sample, perform random cropping, random noise addition, and flipping transformation, and independently apply each data augmentation method with a set probability to generate positive samples; Step 4, Bispectral Feature Extraction: Extract bispectral features for each sample after data augmentation, transforming the one-dimensional time-domain signal into a two-dimensional feature matrix; Step 5: From the dataset Select a batch of data Two data augmentation methods were randomly selected. , Calculate the sample-to-comparison loss ; Step 6: Calculate the cluster contrast loss And calculate the overall loss value. By minimizing To update the network , , Parameters; Step 7: For each sample in the dataset ,pass Extract features, through Calculate the cluster assignment for each sample; Step 8: Output the results of each cluster. Encoding is used to identify individual radiation sources in unsupervised communication. In step 5, the contrast loss of the sample pairs is calculated. The details are as follows: Take a pair of samples , Input to a deep neural network with shared parameters Contrast training is performed to extract features and obtain new feature representations. , , represented as , ; Stack a two-layer non-linear fully connected layer ,pass , Map the feature matrix to a subspace to which the contrastive loss is applied; Define specific samples The comparative loss takes the following form: in, , , Indicates pairwise similarity; This is a hyperparameter, and the sample-pair contrast loss is calculated as follows: In step 6, the cluster contrast loss is calculated. The details are as follows: Characterize the new features obtained , Input into clustering network middle; Set the network output matrix Dimensions satisfy ,in It represents the number of samples in each training batch. It is the number of clusters. , These represent the outputs of each batch of samples after two data augmentations; Let matrix The Listed as That is, the first augmentation of the data sample The representation of each cluster, similarly, will and Combining to form positive clusters , will another Each cluster pair is considered a negative cluster pair; The following loss function is used to cluster With and Distinguish from all other clusters: in, These are the hyperparameters of the clustering network; , , Indicates the similarity between cluster pairs; Clustering contrast loss The definition is as follows: in, It is the probability of cluster assignment. Entropy; In step 6, the overall loss value is calculated. The details are as follows: 。 2. The method for individual identification of unsupervised communication radiation sources based on bispectral feature contrast learning according to claim 1, characterized in that, Using raw signal samples By comparing the similarity with positive examples and the differences with negative examples, more discriminative feature representations can be extracted. , Used for downstream classification and recognition tasks; according to , It is a feature representation obtained from the same data sample through two augmentation methods. When clustering, it should belong to the same class. A clustering loss function is designed, and comparative learning is carried out at the cluster level to complete the classification and recognition task. Positive sample pairs are defined as two augmented samples of the same signal sample, and other sample pairs are defined as negative sample pairs.

3. The method for individual identification of unsupervised communication radiation sources based on bispectral feature contrast learning according to claim 1, characterized in that, The network model in step 1 is as follows: The one-dimensional time series is transformed into a two-dimensional feature matrix and input into the contrastive learning module for training. The residual network is selected as the backbone network of the contrastive learning. The residual network is constructed from residual blocks, which in turn consist of multiple cascaded convolutional layers and a short circuit. The output values ​​of the two are accumulated and then activated using the ReLU function to obtain the output result. By concatenating the residual blocks, a deeper network is formed. The network parameters were optimized through experiments, and the network model was determined to perform feature comparison learning. When training the model, the bispectral feature maps of five VHF radio stations were used as the network input. Each station had 1500 data samples, and each sample underwent data augmentation before bispectral feature calculation. Each sample contained 4096 data points, and the feature matrix dimension was set to [missing value]. ; Randomly select 60% of the data as the training set, 20% as the validation set, and 20% as the test set; ResNet-18 and ResNet-34 were selected as the backbone networks, respectively. The Adam method was used for network optimization, and the initial learning rate was set to 0.

001. The number of training iterations was [number missing]. Set to 400 times, batch data volume By setting the value to 32, we obtained the recognition results corresponding to different network depths. Finally, we selected ResNet-34 as the backbone network for comparative feature extraction.

Images