A method and system for individual frequency hopping radio station identification based on multi-branch parallel learning

Abstract

This invention proposes a multi-branch parallel learning method and system for identifying individual frequency-hopping radio stations, achieving effective identification of individual frequency-hopping radio stations. By constructing a deep residual convolutional network with multi-branch parallel learning, data processing is performed based on statistical features and decision thresholds to obtain the final identification result of the individual frequency-hopping radio station. The identification method of this invention yields accurate results and high efficiency, providing a solid foundation for communication security.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic communication technology, particularly to the field of frequency hopping radio individual identification, and more specifically to a multi-branch parallel learning method and system for frequency hopping radio individual identification. Background Technology

[0002] In the digital communication era, the reliance on information is increasing. In military communications, there is a high dependence on the electromagnetic spectrum. Effectively acquiring electromagnetic intelligence from adversaries and constructing a battlefield electromagnetic posture is crucial for the Chinese military to seize electromagnetic dominance and gain the initiative in combat. This has led to increasing focus on the security and anti-jamming capabilities of wireless communication, which is essential for information exchange between combat platforms. Frequency-hopping communication, due to its advantages such as good confidentiality, strong anti-jamming capabilities, low probability of interception, strong multiple access capabilities, and ease of networking, has gradually become the dominant wireless communication technology for anti-jamming and counter-reconnaissance both domestically and internationally. This also presents a significant challenge to the reconnaissance and identification of frequency-hopping communication targets. How to effectively identify individual radio stations has become a major problem and research focus in the field of communication countermeasures. It is the primary step in acquiring electromagnetic intelligence in the battlefield electromagnetic environment, achieving friend-or-foe identification, and determining threat levels, playing a crucial role in ultimately constructing a battlefield electromagnetic posture.

[0003] Current technologies for identifying individual radiation sources primarily target fixed-frequency communication signals or radar pulse signals, and mainly fall into two categories: one uses mathematical modeling methods to simulate and calculate the traditional characteristics of individual radiation sources, then designs a classifier for classification; the second uses neural networks or deep learning methods to extract features based on training with a large number of samples. Existing methods for identifying individual communication radiation sources, such as... Figure 1 As shown, the individual identification method for communication radiation sources mainly consists of three processes: First, data preprocessing, which mainly involves data segmentation and data sample augmentation; second, feature extraction, which first uses a residual network with two shared parameters as the backbone network for feature comparison learning, and then inputs the rectangular integral bispectral features of the augmented samples into the comparison learning module to further learn more discriminative feature representations, thereby enhancing the feature separability between different radiation source samples; third, the extracted new feature identifiers are used for comparison learning at the cluster level to complete the classification and identification task. This method is mainly used for individual identification of fixed-frequency communication radios, but its applicability is poor when dealing with frequency-hopping radios. Existing feature extraction methods are not accurate or efficient enough, and the robustness of differential features is not high enough. Most existing individual identification methods are closed-set identification, meaning that the radios to be identified have already been trained; however, in reality, most cases are open-set, meaning that there may be radios that have not been trained, which brings difficulties to radio identification.

[0004] In the modern battlefield of communications, with the widespread use of frequency hopping radios, a large number of frequency hopping radios are densely distributed in the time, spatial and frequency domains, and the electromagnetic environment is becoming increasingly complex. Under the premise that various frequency hopping radios have basically the same technical system and technical parameters, the methods provided by existing technologies need to be improved in terms of timeliness and accuracy. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method and system for detecting false information based on multi-scale cross-modal feature fusion.

[0006] The first aspect of this invention discloses a multi-branch parallel learning method for identifying individual frequency-hopping radio stations, comprising:

[0007] Step S1: Collect frequency-hopping intermediate frequency sampling data transmitted by multiple frequency-hopping radio stations to generate the original data sample set;

[0008] Step S2: Perform detection, framing, down-conversion, and normalization preprocessing on each intermediate frequency sample data in the original data sample set obtained in step S1 to form baseband data samples;

[0009] Step S3: Construct a deep residual convolutional network with multi-branch parallel learning, train it on the baseband data samples to obtain a solidified convolutional network; use the solidified deep residual convolutional network after training to extract features from each frequency hopping radio station to form a feature vector, and store the feature vector in the target statistical center feature library;

[0010] Step S4: Input the frequency hopping radio data to be identified into the solidified convolutional network, process the data based on statistical features and decision thresholds, and obtain the final individual identification result of the frequency hopping radio.

[0011] According to the multi-branch parallel learning method for individual identification of frequency-hopping radios as described in the first aspect of the present invention, step S1 of collecting frequency-hopping intermediate frequency sampling data transmitted by multiple frequency-hopping radios includes: during the data collection process, the sampling frequency of the collection device is a fixed value, and the center frequency of the frequency-hopping radio is changed so that the collected data sample set covers the operating frequency of the frequency-hopping radio.

[0012] According to the multi-branch parallel learning method for individual identification of frequency-hopping radio stations as described in the first aspect of the present invention, the construction of a deep residual convolutional network for multi-branch parallel learning in step S3 includes: setting an input layer; setting a multi-branch parallel structure, wherein the parallel structure includes multiple parallel single-layer learning networks, each single-layer learning network consisting of multiple residual network units and an attention unit, and each residual network unit consisting of three convolutional units, one pooling unit and one shortcut connection unit, wherein the kernel size, stride and pooling parameters of the convolutional units on each branch are different, and the feature correlation degree of each branch is also different; setting a fully connected layer and an output layer.

[0013] According to the multi-branch parallel learning method for individual frequency hopping radio identification as described in the first aspect of the present invention, the differential target loss function in the deep residual convolutional network is: Based on cross-entropy, a target feature constraint is introduced, and the optimized target loss function is:

[0014]

[0015] st|c m -c n | 2 ≥δ,m,n=1,2,..,M,m≠n

[0016] Among them, c m c n Let f be the N-dimensional feature statistical center vector of the m-th and n-th targets. m,k Let c be the N-dimensional feature vector extracted by the network from the k-th baseband data sample of the m-th target. m f m,k As variables, optimization is performed iteratively. δ is the feature constraint value, α and β are weight factors that satisfy α+β=1. α, β and δ are global variables that do not participate in training but gradually increase or decrease with the number of iterations. H(p, q) is the cross-entropy function of the convolutional network, p represents the true distribution, q represents the predicted distribution, M represents the number of target N-dimensional feature statistical center vectors, and K represents the total number of baseband data samples.

[0017] According to the multi-branch parallel learning method for individual frequency hopping radio identification as described in the first aspect of the present invention, step S4 specifically includes:

[0018] Step S41: Input the frequency hopping radio data to be identified into the solidified convolutional network for feature extraction to obtain a feature vector;

[0019] Step S42: Input the feature vector extracted in step S41 into the variational autoencoder network, and calculate the feature distribution probability p and the reconstruction error e;

[0020] Step S43: Compare the feature vector extracted in step S41 with the target statistical center feature library for differentiation:

[0021] d c =min||xx c || 2 c∈C,

[0022] Where, d c C is the feature differentiation value; C is the target set; x is the extracted feature vector; x c The feature vector is the corresponding feature vector in the target statistical center feature library;

[0023] Step S44: Combine the feature distribution probability p, reconstruction error e, and feature difference value d. c The target is compared with the decision threshold vector τ. If the target is less than the threshold value corresponding to the decision threshold vector τ, the target identification result is output; otherwise, the target identity is output as unknown.

[0024] According to the multi-branch parallel learning method for individual identification of frequency hopping radio stations as described in the first aspect of the present invention, the construction of the deep residual convolutional network for multi-branch parallel learning in step S3 further includes: in the multiple branches of the deep residual convolutional network, the features extracted by each branch are superimposed and integrated, and the effectiveness of the features extracted by different branches is compared and analyzed, and the common and different features of each branch are combined to obtain the final signal features.

[0025] The second aspect of the present invention discloses a multi-branch parallel learning frequency hopping radio individual identification system, the system being used to implement the multi-branch parallel learning frequency hopping radio individual identification method described in the first aspect.

[0026] The third aspect of the present invention discloses an electronic device, including a memory and a processor. The memory stores a computer program, and the processor is used to execute the program in the memory to implement the multi-branch parallel learning frequency hopping radio individual identification method described in the first aspect.

[0027] The fourth aspect of the present invention discloses a computer-readable storage medium storing a computer program that implements the multi-branch parallel learning frequency hopping radio individual identification method described in the first aspect.

[0028] In summary, the proposed solution, under the premise of comprehensively considering and measuring hardware computing power and model training speed, designs a deep convolutional network with multiple branches to obtain stable, rich and distinguishable subtle features of the radiation source carried by each burst signal of the radio station. By constructing an individual recognition framework based on statistical features and decision thresholds, the accuracy of recognition is improved. The method provided by this invention ensures the effectiveness and accuracy of recognition. Attached Figure Description

[0029] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0030] Figure 1 This is the process for identifying individual communication radiation sources in existing technologies;

[0031] Figure 2 This is a flowchart of the multi-branch parallel learning frequency hopping radio individual identification method of the present invention;

[0032] Figure 3 This is a diagram of the deep residual convolutional network structure for multi-branch parallel learning in this invention;

[0033] Figure 4 This is a diagram of the individual identification framework based on statistical features and decision thresholds of the present invention;

[0034] Figure 5 This is a diagram of the individual identification framework based on statistical features and decision thresholds of the present invention.

[0035] Figure 6 (a)-(d) are schematic diagrams of the individual identification results of three radio stations in Embodiment 1 of the present invention;

[0036] Figure 7 (a)-(d) are schematic diagrams of the identification results of six individual radio stations in Embodiment 1 of the present invention;

[0037] Figure 8 This is a schematic diagram of the identification results of six radio station verification datasets when SNR=30dB in Embodiment 2 of the present invention;

[0038] Figure 9 This is a schematic diagram of the dimensionality reduction result of the features extracted from the verification data of six radio stations when SNR=30dB in Embodiment 2 of the present invention;

[0039] Figure 10 This is a schematic diagram comparing the differences between the verification data characteristics and the central statistical characteristics of radio station 1 when SNR = 30dB in Embodiment 2 of the present invention.

[0040] Figure 11 This is a schematic diagram of the identification results of six radio station verification datasets when SNR=15dB in Embodiment 2 of the present invention;

[0041] Figure 12 This is a schematic diagram of the dimensionality reduction result of the features extracted from the verification data of six radio stations when SNR=15dB in Embodiment 2 of the present invention;

[0042] Figure 13 This is a schematic diagram comparing the differences between the verification data characteristics and the central statistical characteristics of radio station 1 when SNR=15dB in Embodiment 2 of the present invention.

[0043] Figure 14 This is the dimensionality reduction clustering diagram of radio station features under open set conditions in Embodiment 3 of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Figure 2 A flowchart of a multi-branch parallel learning method for individual frequency hopping radio identification provided by this invention includes the following steps:

[0046] S1: Collect intermediate frequency (IF) sampling data transmitted by multiple frequency-hopping radios to generate an original data sample set. During data collection, the sampling frequency of the acquisition equipment is fixed, and the sample set should be designed to cover the operating frequencies of the frequency-hopping radios as much as possible by varying the center frequency of the radios. For example, assuming the operating frequencies of the frequency-hopping radios are 225MHz to 400MHz, the operating frequencies of the collected dataset should be evenly distributed within this range. This method ensures that the acquired sample data fully characterizes the operating status of the FM radios, resulting in higher accuracy in the final individual identification.

[0047] S2:Y ij Let X be the j-th intermediate frequency (IF) data sample of the i-th frequency hopping radio station. For each IF data sample, after preprocessing such as detection, framing, down-conversion, and normalization, the baseband data sample is X. im .

[0048] Baseband data samples can be processed using common data detection, framing, downconversion, and normalization methods in this field for subsequent data analysis.

[0049] Baseband data sample X imThe number of samples depends on the hopping speed of the frequency-hopping radio and the data acquisition time. Generally, one intermediate frequency (IF) data sample will generate hundreds or thousands of baseband data samples, with each baseband data sample having a size of (15000, 2). 80% of the obtained baseband data samples are randomly selected to form the training sample set; half of the remaining 20% ​​are randomly selected to form the validation sample set; and the remaining samples form the test sample set.

[0050] S3: Construct a deep residual convolutional network with multi-branch parallel learning, and train it on the training sample set. Its basic structure diagram is shown below. Figure 3 As shown, it specifically includes the following structure:

[0051] S31 sets up residual convolutional network units. Each residual network unit consists of 3 convolutional units, 1 pooling unit, and 1 shortcut connection unit. The residual network can solve the degradation phenomenon caused by excessive network depth.

[0052] S32 employs a single-layer learning network, consisting of multiple residual network units and one attention unit. To fully leverage the model's powerful autonomous learning and feature extraction capabilities, and to capture as many stable, rich, and discriminative subtle radiation source features as possible from each burst signal of the radio station, the single-layer learning network contains 56 cascaded residual network units. The addition of the attention unit enables the network structure to select information more critical to the current task objective from a wealth of data, improving the robustness of feature selection.

[0053] S33 constructs a multi-branch parallel structure. To maximize the model's richness and learning ability, the network structure designed by the system increases both the model's depth and width. That is, the model designs multiple branch convolutional networks, with each branch having different parameters such as Conv unit kernel size, stride, and pooling. The correlation of the features extracted by each branch is also different. Through parallel learning and feature extraction of multiple branches, the model can autonomously learn different features in a wider range and greater dimension. The features extracted by multiple branches are superimposed and integrated into the subsequent network structure. The model can compare and analyze the effectiveness of the features extracted by different branches, and by combining the common and different features of each branch, it can finally obtain more comprehensive and stable signal features.

[0054] S34 sets a differentiated target loss function. To improve the network recognition performance under open set conditions, this section optimizes the target loss function by introducing target feature constraints based on cross-entropy. The optimized target loss function is as follows:

[0055]

[0056] st|c m -c n |2 ≥δ,m,n=1,2,..,M,m≠n

[0057] Among them, c m c n Let f be the N-dimensional feature statistical center vector of the m-th and n-th targets. m,k Let c be the N-dimensional feature vector extracted by the network from the k-th baseband data sample of the m-th target. m f m,k The network uses variables for iterative optimization. δ represents the feature constraint value, and α and β are weighting factors satisfying α + β = 1. α, β, and δ are global variables that do not participate in training but gradually increase or decrease with the number of iterations. Initial values ​​need to be set appropriately to ensure model convergence. H(p, q) is the cross-entropy function of the convolutional network, where p represents the true distribution, q represents the predicted distribution, M represents the number of target N-dimensional feature statistical center vectors, and K represents the total number of baseband data samples. The optimized target loss function ensures that the network model can classify and recognize targets while maximizing the convergence of features for the same target signal, although different target signal features have certain differences.

[0058] The S35 incorporates an input layer, a fully connected layer, and an output layer.

[0059] S36 utilizes a constructed deep residual convolutional network with multi-branch parallel learning to train and validate the training and validation sample sets.

[0060] S37 sends the features extracted by the trained network for each radio station to the target statistical center feature library for storage.

[0061] S4: Construct an individual identification framework based on statistical features and decision thresholds, and test it on a test sample set. The constructed individual identification framework is as follows: Figure 4 As shown, the specific steps for individual identification are as follows:

[0062] S41 feeds the test sample set into the network that has been trained and solidified by S3 for feature extraction.

[0063] S42 inputs the proposed feature vector into the optimized variational autoencoder network to calculate the feature distribution probability p and the reconstruction error e;

[0064] S43 performs a differential comparison between the proposed feature vector and the target statistical center feature library.

[0065] d c =min||xx c || 2 c∈C, where C is the target set.

[0066] S44 will use the feature distribution probability p, reconstruction error e, and feature difference value d. cThe target identification result is output if the result is less than the decision threshold vector τ; otherwise, the target identity is output as unknown.

[0067] To evaluate and validate the feasibility and effectiveness of the designed multi-branch parallel learning method for frequency hopping radio individual identification, such as... Figure 5 As shown, a single acquisition device was used to collect the intermediate frequency (IF) signals of six identical frequency-hopping radios. The collected IF dataset was sent to a computer, where it underwent preprocessing before being sent to a server for training. The server software environment consisted of Windows Server R2 2012 operating system, Tensorflow version 1.10, CUDA 9.0, and cuDNN v7.3; the hardware environment included a 16GB GPU and 64GB of RAM. The hardware connection diagram is shown below. Figure 5 As shown.

[0068] Each frequency-hopping radio has a hopping rate of 1000 hops / s, a BPSK modulation scheme, 128 hopping frequencies, a frequency interval of 50kHz, a total bandwidth of 6.4MHz, a sampling frequency of 30MHz, a signal-to-noise ratio of 20dB, and 26430 baseband sampling data points for each radio. The dataset is divided into training, validation, and test datasets in an 8:1:1 ratio.

[0069] Based on a specific embodiment of the present invention:

[0070] To verify the performance of the deep learning-based frequency-hopping network station individual identification system under different numbers of stations, experiments were conducted to verify the individual identification performance under different numbers of stations. Using the provided data samples, Experiment 1 randomly selected three stations for individual identification, while Experiment 2 identified all six stations. Furthermore, to verify that the deep learning-based system for frequency-hopping network station individual identification can automatically import the existing fixed network model and improve the model training speed, Experiment 2 further optimized the model based on Experiment 1. Specifically, the model parameters trained and fixed in Experiment 1 were used as the initial values ​​for the model parameters in Experiment 2, and further training and optimization were performed based on these initial values. The data used in Experiments 1 and 2 were randomly divided into training, test, and validation sets in an 8:1:1 ratio.

[0071] The results of Experiment 1 are as follows Figure 6As shown in the figure, the model converges relatively quickly. Using the converged model, feature extraction is performed on the test data of three radio stations. The extracted features, after dimensionality reduction, demonstrate that the deep learning-based radiation source identification system, after training with a sufficient number of samples, can effectively and automatically extract subtle features of each radio station. Furthermore, the features extracted for each radio station exhibit good convergence and central consistency. In addition, during model training, the network model automatically calculates and saves the statistical centers of each radio station's features. To characterize the differences between the features of different radio stations, the distance matrix of the statistical centers of the features of three different radio stations is calculated, as shown in the figure. This proves that the features extracted by the network model for each radio station have good differences, thus demonstrating high recognition accuracy on the test data under closed-set conditions.

[0072] Experiment 2, based on the training results of Experiment 1, used the model parameters solidified from Experiment 1 as the initial values ​​for training the model in Experiment 2. Compared to Experiment 1, the model training in Experiment 2 converged faster, such as... Figure 7 As shown in the figure. In Experiment 2, after the model was trained and converged, it was tested using a validation dataset. The recognition accuracy reached 98%, and the validation dataset, after feature extraction by the model, showed good feature convergence. The statistical centers of features from different radio stations showed significant differences. Experiment 2 verifies that the designed deep learning-based frequency hopping radio individual identification system can effectively extract and accurately identify features of different numbers of radio stations under sufficient sample data. Moreover, by utilizing the transferability of model features, the identifiable target set can be flexibly and effectively expanded.

[0073] Based on a specific embodiment of the present invention:

[0074] To verify the impact of different signal-to-noise ratios (SNRs) on the model's individual recognition performance, this embodiment uses six radio station signal samples, divided into training, testing, and validation sets in an 8:1:1 ratio. The training and testing sets are left untouched by noise; the model acquisition files are preprocessed before training the model. During training, the test set is used to test the model's generalization ability and recognition performance. After model convergence, the model parameters are fixed. To verify the impact of different SNR conditions on the model's feature extraction and individual recognition performance after training, Gaussian white noise with varying SNR intensities is added to the validation dataset. The trained network model is then used to perform feature extraction and recognition tests on the validation sets with added noise at different SNR ratios. This experiment, based on the model trained in Experiment 2, conducts SNR impact tests. A total of 15,858 validation data points from the six radio stations are used in this experiment.

[0075] When SNR = 30dB, the average identification accuracy of the six radio station verification data is still as high as 99%, and the system automatically calculates the results as follows: Figure 8 As shown. The features extracted from the verification data of six radio stations, after dimensionality reduction, are as follows. Figure 9As shown, under this signal-to-noise ratio condition, the proposed features of each radio station still exhibit good clustering convergence and distinguishability. Taking radio station 1 as an example, a total of 2643 signal samples from all validation data of this station were statistically analyzed. The feature vectors proposed by the model were compared with the statistical center vectors of the station's features stored in the model. The Euclidean distance and the angle between the vectors are shown below. Figure 10 As shown, the characteristics of each verification data point of radio station 1 are highly consistent with the statistical characteristics of the radio station center.

[0076] When SNR = 15dB, the average recognition accuracy of the six radio station verification datasets is 94%. The system recognition statistics are as follows: Figure 11 As shown. At this point, the dimensionality reduction of the features extracted from the verification data of the six radio stations is as follows: Figure 12 As shown, comparing the feature dimensionality reduction clustering diagram when SNR=30dB, it is clear that the convergence and distinguishability of the six radio station features decrease significantly at this point. Furthermore, taking radio station 1 as an example, comparing the features of all validation data for this station with the feature centers saved in the model, their Euclidean distance and vector angle are shown below. Figure 13 As shown, the difference between the features proposed by the model and the central features saved during the training of the radio station increases after strong noise is added to the validation data, and the convergence of the radio station validation data features decreases.

[0077] Embodiment 3 proposed according to the method of the present invention:

[0078] In practical applications, it is impossible to label and accumulate data samples from all radio station radiation sources. That is, the model cannot be trained on all radio station signal samples. When the test set includes targets outside the training set, it is considered open-set recognition. In practical applications of individual radiation source recognition, open-set recognition is almost always used. Therefore, the model, through feature extraction, can effectively identify targets within the training set and effectively distinguish targets outside the set, which has significant practical value. This embodiment tests individual recognition under open-set conditions. Radio stations 1, 2, 3, and 5 are randomly selected for model training. Radio stations 4 and 6 are used as targets outside the training set, i.e., open-set targets. After convergence and solidification, the model's feature extraction and individual recognition performance are tested. Figure 14 The diagram shows the dimensionality reduction and clustering of the verification data from six radio stations after feature extraction by the model. As can be seen from the diagram, the features extracted by the designed network model for target signals outside the set are significantly different from the features of the signals from the four radio stations within the set. Therefore, the model can automatically detect and identify most signal bursts of targets outside the set as targets outside the set, while targets within the set can still be accurately identified. Thus, the model has a certain ability to identify open sets.

[0079] The invention also provides a multi-branch parallel learning frequency hopping radio individual identification system to implement the aforementioned multi-branch parallel learning frequency hopping radio individual identification method.

[0080] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is used to execute the program in the memory to implement the aforementioned multi-branch parallel learning frequency hopping radio individual identification method.

[0081] The present invention also provides a computer-readable storage medium storing a computer program that implements the aforementioned multi-branch parallel learning method for individual identification of frequency-hopping radio stations.

[0082] In summary, the technical solution proposed in this invention has the following technical effects: First, regarding the applicability to frequency-hopping radios, this invention extracts the signal of each hop of the frequency-hopping radio through preprocessing methods such as detection, framing, down-conversion, and normalization, thus providing a prerequisite for individual identification. Second, addressing the issues of inaccurate and inefficient feature extraction and insufficient robustness of differential features, this invention, under the premise of comprehensively considering and measuring hardware computing power and model training speed, differs from existing publicly available algorithms that only use network models as feature classifiers or use simple, shallow network structures as feature extractors. This invention designs a deep convolutional network with multiple branches, where the Conv unit convolution kernel size, stride, pooling, and other parameters are different in each branch, and the correlation degree of the features extracted by each branch is also different. Through parallel learning and feature extraction of multiple branches, the model can autonomously learn different features over a wider range and greater dimensions, striving to obtain stable, rich, and distinguishable subtle features of the radiation source carried by each hop of the radio's burst signal. Third, regarding the open-set identification problem, this invention constructs an individual identification framework based on statistical features and decision thresholds, making it adaptable to open-set problems as well. Based on the method provided by this invention, the accuracy and speed of identifying individual radio stations are improved, communication security is ensured, and a stable and effective foundation is provided for subsequent communication interactions.

[0083] Please note that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for individual identification of frequency-hopping radio stations using multi-branch parallel learning, characterized in that, include: Step S1: Collect frequency-hopping intermediate frequency sampling data transmitted by multiple frequency-hopping radio stations to generate the original data sample set; Step S2: Perform detection, framing, down-conversion, and normalization preprocessing on each intermediate frequency sample data in the original data sample set obtained in step S1 to form baseband data samples; Step S3: Construct a deep residual convolutional network with multi-branch parallel learning, train it on the baseband data samples to obtain a solidified convolutional network; use the solidified deep residual convolutional network after training to extract features from each frequency hopping radio station to form a feature vector, and store the feature vector in the target statistical center feature library; Step S4: Input the frequency hopping radio data to be identified into the solidified convolutional network, process the data based on statistical features and decision thresholds, and obtain the final individual identification result of the frequency hopping radio. Specifically, step S4 includes: Step S41: Input the frequency hopping radio data to be identified into the solidified convolutional network for feature extraction to obtain a feature vector; Step S42: Input the feature vector extracted in step S41 into the variational autoencoder network and calculate the feature distribution probability. p and reconstruction error e ; Step S43: Compare the feature vector extracted in step S41 with the target statistical center feature library for differentiation: in, d c For feature differentiation values; C For the target set; The extracted feature vector; The feature vector is the corresponding feature vector in the target statistical center feature library; Step S44: Divide the feature distribution probability p Reconstruction error e and feature difference value With decision threshold vector τ Perform a comparison; if they are all less than the decision threshold vector... τ If the corresponding threshold value is set, the target recognition result will be output; otherwise, the target identity will be output as unknown.

2. The multi-branch parallel learning method for frequency hopping radio station individual identification according to claim 1, characterized in that, Step S1 involves collecting frequency-hopping intermediate frequency sampling data transmitted by multiple frequency-hopping radios, including: during the data collection process, the sampling frequency of the acquisition device is a fixed value, and the center frequency of the frequency-hopping radio is changed so that the collected data sample set covers the operating frequency of the frequency-hopping radio.

3. The multi-branch parallel learning method for individual frequency hopping radio station identification according to claim 1, characterized in that, The construction of a deep residual convolutional network with multi-branch parallel learning in step S3 includes: setting an input layer; setting a multi-branch parallel structure, wherein the parallel structure includes multiple parallel single-layer learning networks, each single-layer learning network consisting of multiple residual network units and an attention unit, and each residual network unit consisting of three convolutional units, one pooling unit and one shortcut connection unit, wherein the kernel size, stride and pooling parameters of the convolutional units on each branch are different, and the feature correlation degree of each branch is also different; setting a fully connected layer and an output layer.

4. The multi-branch parallel learning method for frequency hopping radio station individual identification according to claim 3, characterized in that, The differentiated objective loss function in the deep residual convolutional network is: by introducing objective feature constraints based on cross-entropy, the optimized objective loss function is: in, c m , c n For the first m , n One goal N 3D feature statistical center vector, f m,k For the network to the first m The first goal k Extracted from the baseband data samples N 3D feature vectors c m , f m,k Optimize through iteration as a variable. For feature constraint values, are weighting factors and satisfy , It is a global variable that does not participate in training but gradually increases or decreases with the number of iterations. H ( p , q ) represents the cross-entropy function of the convolutional network. p Represents the true distribution. q Indicates the predicted distribution. M Indicate target N The number of central vectors in the feature statistics. K This represents the total number of baseband data samples.

5. The multi-branch parallel learning method for individual frequency hopping radio station identification according to claim 3, characterized in that, The construction of a multi-branch parallel learning deep residual convolutional network in step S3 further includes: in the multiple branches of the deep residual convolutional network, the features extracted by each branch are superimposed and integrated, and the effectiveness of the features extracted by different branches is compared and analyzed. The common and different features of each branch are combined to obtain the final signal features.

6. A multi-branch parallel learning frequency hopping radio individual identification system, characterized in that, The system is used to implement the multi-branch parallel learning frequency hopping radio individual identification method according to any one of claims 1-5.

7. An electronic device, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program in the memory to implement the multi-branch parallel learning frequency hopping radio individual identification method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that implements the frequency hopping radio individual identification method of multi-branch parallel learning as described in any one of claims 1-5.

Images