A method for identifying individual radiators in frequency hopping communication based on feature space processing

Abstract

The present invention discloses a method for identifying individual frequency hopping communication radiation sources based on feature space processing, the method comprising: preprocessing, normalizing and bispectral estimating the acquired intermediate frequency signal data, mapping to the bispectral feature space, obtaining a signal bispectral feature set; dividing the signal bispectral feature set into a test feature set and a training feature set after moving cumulative averaging, and smoothing the same; extracting a training sample set and a test sample set based on bispectral feature selection and dimension reduction of the smoothed training feature set; inputting the training sample set into a support vector machine for classification training to obtain a classification model; inputting the test sample set into the classification model to test the individual recognition classification performance; and adjusting the feature space processing parameters according to the individual recognition classification performance evaluation result until the classification model with the best classification performance is obtained, and outputting the radiation source individual recognition result. The algorithm complexity of the present invention is low and easy to implement in engineering.

Description

Technical Field

[0001] The invention relates to the technical field of signal recognition, in particular to a method for identifying individual frequency hopping communication radiation sources based on feature space processing. Background Art

[0002] At present, the individual identification of communication radiation sources mainly achieves the purpose of individual identification and identity confirmation by extracting the subtle features of the radiation source radio equipment, namely the signal fingerprint. In recent years, with the help of machine learning technology, this technology research has gradually become a research hotspot in the fields of military electronic reconnaissance and wireless network security, and its application has become more and more extensive.

[0003] Judging from the current research status and achievements at home and abroad, there are roughly two mainstream technical routes. The first is the statistical learning classification method, which is a machine learning process based on signal features. It constructs a classification model by extracting the characteristic parameters of the radiation source. The research mainly focuses on the three key links of individual identification, such as radiation source feature extraction, feature selection and classifier design. The second is the neural network classification method, which mainly uses the deep self-learning ability of the neural network to automatically extract and classify data or feature samples. The research is mainly in the aspects of data sample processing and visualization, network design and model optimization. The former relies on the expert knowledge of signal analysis. Although the latter reduces the dependence on professional knowledge, it requires high-quality and large training data samples, and the model has poor interpretability. The neural network classification method has relatively strict requirements for signal data samples, which brings great challenges to both signal interception and signal processing. In practice, the individual identification of radiation sources based on the statistical learning classification method still has stronger engineering practical value.

[0004] Since the end of the last century, research on individual identification of radiation sources based on statistical learning classification methods has been relatively mature. Common classifiers include decision trees, Bayesian analyzers, nearest neighbor classifiers, support vector machines, etc., especially support vector machines as a classic model of supervised learning. Kernel function design can solve nonlinear classification problems and theoretically obtain the globally optimal partitioning plane. At the same time, it has unique advantages in dealing with small sample data problems and has strong scalability. In the field of research on individual identification of radiation sources, it is selected as one of the classifiers. In recent years, the research on individual identification of radiation sources has basically focused on "feature engineering", including feature extraction, feature selection and other processing. Feature extraction is essentially the process of transforming the original or preprocessed signal data of the radiation source into the feature space for extraction. The features extracted and involved in classification and identification need to meet the characteristics of time-shift invariance, scale variability and phase retention as much as possible. The current differential constellation trajectory, bispectrum and Hilbert-Huang transform all meet the above characteristics to a certain extent and are theoretically feasible; feature selection is mainly to select the features that are most beneficial to classification, on the one hand to achieve dimensionality reduction of feature space, and on the other hand to avoid the influence of trivial features (features that contribute less to classification) on classification performance. The current methods include principal component analysis (PCA), linear discriminant analysis (LDA), independent component analysis (ICA), and information separability analysis.

[0005] The above statistical analysis methods can achieve better individual recognition performance in theory and in the laboratory. For example, bispectral analysis can obtain relatively stable features, and because it uses multiple correlations, it has a better suppression effect on Gaussian white noise. However, in practical applications, facing anti-interference communication signals such as frequency hopping and burst, the signal duration is short, the actual channel is complex, and the Gaussian white noise characteristics are not obvious. The obtained classification features are no longer stable due to the limitation of continuous data sample length, unstable noise distribution, measurement differences, etc., and the anti-noise and anti-interference performance is significantly reduced, and the ideal performance cannot be achieved. Summary of the invention

[0006] In view of this, the present invention provides a method for individual identification of frequency hopping communication radiation sources based on feature space processing, which improves and enhances the performance of the current individual identification of radiation sources based on statistical learning classification methods, so as to solve the problems that the performance of the current individual identification of radiation sources is degraded in an actual complex channel environment, and the identification performance of the communication radiation sources using anti-interference communication systems such as frequency hopping and burst is poor.

[0007] The present invention discloses a method for identifying individual frequency hopping communication radiation sources based on feature space processing, comprising the following steps:

[0008] Step 1: pre-processing the intermediate frequency original signal data obtained from multiple frequency hopping communication radiation sources of the same model to obtain multiple zero intermediate frequency signal data; each of the intermediate frequency original signal data includes the intermediate frequency data of the frequency hopping pulse and the corresponding individual labeling information of the frequency hopping communication radiation source;

[0009] Step 2: normalizing and bispectral estimating the multiple zero intermediate frequency signal data, mapping the data to the bispectral feature space, and obtaining a signal bispectral feature set;

[0010] Step 3: performing moving cumulative averaging on the signal bispectrum feature set to obtain an updated signal bispectrum feature set, and dividing the updated signal bispectrum feature set into a test feature set and a training feature set;

[0011] Step 4: Smoothing the test feature set and the training feature set respectively to obtain a smoothed training feature set and a test feature set;

[0012] Step 5: using a multi-classification feature selection method to perform bispectral feature selection and dimension reduction on the smoothed training feature set to obtain a selected bispectrum; extracting a training sample set and a test sample set from the smoothed training feature set and the test feature set respectively according to the effective frequency corresponding to the selected bispectrum;

[0013] Step 6: Input the training sample set into the support vector machine for classification training to obtain a classification model; input the test sample set into the classification model to test the individual recognition classification performance; and adjust the feature space processing parameters according to the individual recognition classification performance evaluation results, repeatedly iterate steps 3 to 6 until the classification model with the best classification performance is obtained, and output the radiation source individual recognition result.

[0014] Furthermore, in step 1, the preprocessing includes filtering, down-conversion and sampling rate conversion; the plurality of zero intermediate frequency signal data are in, is single frequency hopping pulse data, n=1, 2, ..., N, N is the number of frequency hopping pulses; F m is the radiation individual labeling information corresponding to the frequency hopping pulse data, m=1, 2, ..., M, where M is the number of frequency hopping communication radiation sources.

[0015] Furthermore, the step 2 comprises:

[0016] Will Each frequency hopping pulse data As a set of observation data, after power normalization processing is performed on each set of observation data, the signal bispectral feature set is obtained by bispectral estimation. L is the Fourier transform length used in bispectrum estimation.

[0017] Furthermore, the step 3 comprises:

[0018] Step 31: Input the bispectral feature set of the i-th type radiation source signal Set the cumulative average length to L1, the moving step to L2, and perform moving average calculation. The number of calculations N0 is:

[0019]

[0020] Where N is The number of eigenvectors of , fix(·) is the floor function;

[0021] Step 32: Right Take the first L1 eigenvectors and perform cumulative average calculation to obtain the following result:

[0022]

[0023] Step 33: Right Move L2 feature vectors backward, take L1 feature vectors, and repeat N0 times according to step 32 to obtain the bispectral feature set of the i-th category after cumulative averaging. Where n0=1, 2, ..., N0;

[0024] Step 34: Repeat steps 32 and 33 to complete the cumulative average processing of the bispectral feature sets of all types of communication radiation source signals to obtain an updated signal bispectral feature set.

[0025] Step 35: Divided into training feature set and the test feature set Among them, n0=n1+n2.

[0026] Furthermore, the value ranges of L1 and L2 are respectively: 1≤L1≤N / 1000, L1 / 10≤L2≤L1 / 2; the specific values ​​of L1 and L2 can be adjusted according to the classification performance.

[0027] For frequency hopping or burst communication signals, bispectrum estimation is performed according to the resident signals of each frequency hopping or burst period. Drawing on the image processing background denoising method, the bispectrum value obtained by each estimation is used as a collected image, and multiple cumulative averaging is performed to form a new average bispectrum value. This can overcome the image detail blurring caused by the linear filter, further reduce the interference of random noise, and improve the signal-to-noise ratio of the signal. At the same time, it can reduce the differences between multiple frequency hopping or burst signal pulses to a certain extent, and ensure the stability and accuracy of the characteristic measurement values.

[0028] Furthermore, the step 4 comprises:

[0029] Step 41: extracting the smoothed feature set of the i-th category from the test feature set and the training feature set respectively, and the feature space dimension of the feature set is L×L;

[0030] Step 42: For the feature set of the ith classification, the Savitsky-Golay smoothing algorithm is used to perform data smoothing on the feature sequence one by one according to each feature dimension, that is, the least squares fitting in the sliding window is realized by a polynomial, and the weighted average sum of the center point in the sliding window and its surrounding points is calculated; after calculating and traversing all feature dimensions, the smoothed training feature set of the ith classification is obtained. and the test feature set

[0031] Step 43: According to step 41 and step 42, the smoothing of the training feature set and the test feature set of all classes is completed to obtain the smoothed training feature set. and the test feature set

[0032] Furthermore, the step 5 includes:

[0033] Step 51: Based on the training feature set The total number of individual categories and labeling information are statistically arranged into binary classification pairing combinations, and the individual category labeling sequence of the pairing combination is obtained as P = {i, j};

[0034] Step 52: Select a set of individual category labeling information (i, j) in P, and extract the corresponding training feature set of the i-th category respectively and the training feature set for the jth category Calculate the inter-class classifiability measure of these two classes, denoted as m (i，j) (ω);

[0035] Step 53: Place m (i，j) (ω) Sort by size and select the larger Q m (i，j) (ω) corresponds to the effective frequency set ω (i，j) (q), as the selection bispectrum between the feature sets of the i-th category and the j-th category; where q = 1, 2, ..., Q, Q is the number of effective frequencies, and its value is determined comprehensively according to the dimensionality reduction requirements and classification performance;

[0036] Step 54: By analogy with steps 52 and 53, the selected bi-spectra of all binary paired combination feature sets are calculated, and the selected bi-spectra of different binary paired combination feature sets are obtained by merging and removing duplicates. (i，j) (q), based on the same effective frequency, only take it once to get the selected bispectrum ω of all classes (i，j) (q′);

[0037] Step 55: Select bispectrum ω based on (i，j) The effective frequency of (q′) is extracted from the smoothed training feature set to obtain a training sample set, and the corresponding bi-spectrum features are extracted from the smoothed test feature set to obtain a test sample set, and the feature dimension is reduced from the feature set dimension L×L to the sample set dimension q′.

[0038] This method uses Fisher information separability and Bhattacharyya distance with good performance for weighted calculation as inter-class classifiability measures, extracts feature vectors with the strongest inter-class separability as signal classification recognition features, achieves feature dimensionality reduction, and reduces the impact of trivial features (features with small classification contribution) on classification performance.

[0039] Furthermore, m (i，j) The calculation formula of (ω) is:

[0040] m (i，j) (ω)=a*f (i，j) (ω)+b*d (i，j) (ω)

[0041] Among them, f (i，j) (ω) is the Fisher information separability, which reflects the contribution of the difference in the mean distribution of the two categories of the bispectrum to the classification, d (i，j) (ω) is the Bhattacharyya distance metric, which reflects the contribution of the difference in the variance of the two-category distribution to the classification. a and b are weighting coefficients, which are determined by the random variable The prior probability is determined, and the value range is a, b∈[0,1], and the specific value can be adjusted and determined according to the classification performance;

[0042] f (i，j) The calculation formula of (ω) is:

[0043]

[0044] d (i，j) The calculation formula of (ω) is:

[0045]

[0046] in, and Respectively represent the mean of all sample bispectra of the i-th signal and the j-th signal at frequency ω, Represents the mean of all sample bispectra of the lth type signal at frequency ω.

[0047] Furthermore, in step 6, the feature space processing parameters include cumulative average length, moving step, number of effective frequencies, and inter-class classifiability measure weighting coefficient.

[0048] Furthermore, the classification performance evaluation indicators include Precision, Recall and F1.

[0049] Precision=TP / (TP+FP)

[0050] Recall=TP / (TP+FN)

[0051] F1=(2*Precision*Recall) / (Precision+Recall)

[0052] Among them, Precision, Recall and F1 represent precision, recall and the combination of the two respectively. TP is a positive sample whose prediction result is correct; FP is a positive sample whose prediction result is wrong; FN is a negative sample whose prediction result is wrong.

[0053] Due to the adoption of the above technical solution, the present invention has the following advantages:

[0054] (1) The present invention adopts two types of feature processing methods in the classification feature vector space: First, the moving cumulative average method is used for the feature vectors of multiple measurements, which effectively reduces the unstable influence of the communication signal on the extracted features due to time-frequency hopping and discontinuity, and further suppresses the background noise in the feature space; second, the smoothing method is used for the multi-dimensional feature sequence, which further removes the influence of outlier noise (outliers) caused by multiple measurements of the communication signal, and at the same time ensures the stability and continuity of the features. By processing the features, the measurement of the features is improved, and the effect of suppressing various noises is further achieved, especially effectively avoiding the problems of difficult feature extraction and easy distortion in the individual identification of communication radiation sources such as frequency hopping or burst. At the same time, the algorithm complexity is low and easy to implement in engineering.

[0055] (2) The present invention provides a method for adjusting feature space processing parameters for performance optimization, and a classification inter-class classifiable measurement model based on a weighted combination of multiple distance metrics and information metrics, and provides multiple adjustable feature space processing parameters such as cumulative average length, moving step, number of effective frequencies, and inter-class classifiable measurement weighting coefficients. The provided feature space processing parameters can be dynamically adjusted and optimized in multiple dimensions based on the feedback of classification performance evaluation results, and the established classifiable measurement model can flexibly integrate and expand multiple similarity measurement methods based on classification problems, providing an effective means for optimizing the performance of communication radiation source individual identification in actual application environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0057] Figure 1 A schematic flow chart of a method for identifying individual frequency hopping communication radiation sources based on feature space processing according to an embodiment of the present invention;

[0058] FIG2( a ) is a schematic diagram of a conventional bispectral selection method without feature space processing;

[0059] FIG2( b ) is a schematic diagram of a method for selecting a bispectrum using feature moving average processing according to an embodiment of the present invention;

[0060] FIG2( c ) is a schematic diagram of a method for selecting a bispectrum using feature moving average and smoothing according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The present invention is further described in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field should fall within the scope of protection of the embodiments of the present invention.

[0062] The present invention aims to solve the problem of poor recognition performance of communication radiation sources in actual complex channel environments for anti-interference communication systems such as frequency hopping and burst. On the basis of the mainstream individual radiation source recognition based on statistical learning classification methods, by adding and improving feature processing methods, the recognition of individual communication radiation sources in actual complex noise backgrounds is realized, and the individual recognition performance is improved.

[0063] See also Figure 1 The present invention provides an embodiment of a method for identifying individual frequency hopping communication radiation sources based on feature space processing, which comprises the following steps:

[0064] Step 1: pre-processing the intermediate frequency raw signal data obtained from multiple frequency hopping communication radiation sources of the same model to obtain multiple zero intermediate frequency signal data; each intermediate frequency raw signal data includes the intermediate frequency data of the frequency hopping pulse and the corresponding individual labeling information of the frequency hopping communication radiation source;

[0065] Step 2: normalize and estimate the bispectrum of multiple zero-IF signal data, map the data to the bispectrum feature space, and obtain the signal bispectrum feature set;

[0066] Step 3: Perform moving cumulative averaging on the signal bispectrum feature set to obtain an updated signal bispectrum feature set, and divide the updated signal bispectrum feature set into a test feature set and a training feature set;

[0067] Step 4: Smooth the test feature set and the training feature set respectively to obtain the smoothed training feature set and the test feature set;

[0068] Step 5: Using a multi-classification feature selection method, perform bispectral feature selection and dimension reduction on the smoothed training feature set to obtain a selected bispectrum; according to the effective frequency corresponding to the selected bispectrum, extract the training sample set and the test sample set from the smoothed training feature set and the test feature set respectively;

[0069] Step 6: Input the training sample set into the support vector machine for classification training to obtain a classification model; input the test sample set into the classification model to test the individual recognition classification performance; and adjust the feature space processing parameters according to the individual recognition classification performance evaluation results, repeatedly iterate steps 3 to 6 until the classification model with the best classification performance is obtained, and output the radiation source individual recognition results.

[0070] This method adds a feature space processing link between the feature extraction and feature selection links. By moving average and smoothing the feature vector, it removes the influence of channel changes and pulse measurement on the feature vector, removes signal measurement glitches and outlier noise, enhances the stability of the feature process, avoids the influence of feature outliers and measurement errors on classification performance, ensures the quality of classification features, and improves the fault tolerance and performance of the individual recognition method.

[0071] See also Figure 2(a) to Figure 2(c) , which counts the classifier training performance under different numbers of features and gives the performance comparison results between the traditional bispectral selection method and this method.

[0072] In this embodiment, in step 1, preprocessing includes filtering, down-conversion and sampling rate conversion; multiple zero intermediate frequency signal data are in, is single frequency hopping pulse data, n=1, 2, ..., N, N is the number of frequency hopping pulses; F m is the radiation individual labeling information corresponding to the frequency hopping pulse data, m=1, 2, ..., M, where M is the number of frequency hopping communication radiation sources.

[0073] In this embodiment, step 2 includes:

[0074] Will Each frequency hopping pulse data As a set of observation data, after power normalization processing is performed on each set of observation data, the signal bispectral feature set is obtained by bispectral estimation. L is the Fourier transform length used in bispectrum estimation.

[0075] In this embodiment, step 3 includes:

[0076] Step 31: Input the bispectral feature set of the i-th type radiation source signal Set the cumulative average length to L1, the moving step to L2, and perform moving average calculation. The number of calculations N0 is:

[0077]

[0078] Where N is The number of eigenvectors of , fix(·) is the floor function;

[0079] Step 32: Right Take the first L1 eigenvectors and perform cumulative average calculation to obtain the following result:

[0080]

[0081] Step 33: Right Move L2 feature vectors backward, take L1 feature vectors, and repeat N0 times according to step 32 to obtain the bispectral feature set of the i-th category after cumulative averaging. Where n0=1, 2, ..., N0;

[0082] Step 34: Repeat steps 32 and 33 to complete the cumulative average processing of the bispectral feature sets of all types of communication radiation source signals to obtain an updated signal bispectral feature set.

[0083] Step 35: Divided into training feature set and the test feature set Among them, n0=n1+n2.

[0084] In this embodiment, the value ranges of L1 and L2 are: 1≤L1≤N / 1000, L1 / 10≤L2≤L1 / 2; the specific values ​​of L1 and L2 can be adjusted according to the classification performance.

[0085] For frequency hopping or burst communication signals, bispectrum estimation is performed according to the resident signals of each frequency hopping or burst period. Drawing on the image processing background denoising method, the bispectrum value obtained by each estimation is used as a collected image, and multiple cumulative averaging is performed to form a new average bispectrum value. This can overcome the image detail blurring caused by the linear filter, further reduce the interference of random noise, and improve the signal-to-noise ratio of the signal. At the same time, it can reduce the differences between multiple frequency hopping or burst signal pulses to a certain extent, and ensure the stability and accuracy of the characteristic measurement values.

[0086] In this embodiment, step 4 includes:

[0087] Step 41: extract the feature set of the i-th category from the smoothed test feature set and the training feature set respectively, and the feature space dimension of the feature set is L×L;

[0088] Step 42: For the feature set of the ith classification, the Savitsky-Golay smoothing algorithm is used to perform data smoothing on the feature sequence one by one according to each feature dimension, that is, the least squares fitting in the sliding window is realized by a polynomial, and the weighted average sum of the center point in the sliding window and its surrounding points is calculated; after calculating and traversing all feature dimensions, the smoothed training feature set of the ith classification is obtained. and the test feature set

[0089] Step 43: According to step 41 and step 42, the smoothing of the training feature set and the test feature set of all classes is completed to obtain the smoothed training feature set. and the test feature set

[0090] In this embodiment, step 5 includes:

[0091] Step 51: Based on the training feature set The total number of individual categories and labeling information are statistically arranged into binary classification pairing combinations, and the individual category labeling sequence of the pairing combination is obtained as P = {i, j};

[0092] Step 52: Select a set of individual category labeling information (i, j) in P, and extract the corresponding training feature set of the i-th category respectively and the training feature set for the jth category Calculate the inter-class classifiability measure of these two classes, denoted as m (i，j) (ω);

[0093] Step 53: Place m (i，j) (ω) Sort by size and select the larger Q m (i，j) (ω) corresponds to the effective frequency set ω (i，j) (q), as the selection bispectrum between the feature sets of the i-th category and the j-th category; where q = 1, 2, ..., Q, Q is the number of effective frequencies, and its value is determined comprehensively according to the dimensionality reduction requirements and classification performance;

[0094] Step 54: By analogy with steps 52 and 53, the selected bi-spectra of all binary paired combination feature sets are calculated, and the selected bi-spectra of different binary paired combination feature sets are obtained by merging and removing duplicates. (i，j) (q), based on the same effective frequency, only take it once to get the selected bispectrum ω of all classes (i，j)(q′);

[0095] Step 55: Select bispectrum ω based on (i，j) The effective frequency of (q′) is extracted from the smoothed training feature set to obtain the training sample set, and the corresponding bispectral features are extracted from the smoothed test feature set to obtain the test sample set. The feature dimension is reduced from the feature set dimension L×L to the sample set dimension q′.

[0096] This method uses Fisher information separability and Bhattacharyya distance with good performance for weighted calculation as inter-class classifiability measures, extracts feature vectors with the strongest inter-class separability as signal classification recognition features, achieves feature dimensionality reduction, and reduces the impact of trivial features (features with small classification contribution) on classification performance.

[0097] In this embodiment, m (i，j) The calculation formula of (ω) is:

[0098] m (i，j) (ω)=a*f (i，j) (ω)+b*d (i，j) (ω)

[0099] Among them, f (i，j) (ω) is the Fisher information separability, which reflects the contribution of the difference in the mean distribution of the two categories of the bispectrum to the classification, d (i，j) (ω) is the Bhattacharyya distance metric, which reflects the contribution of the difference in the variance of the two-category distribution to the classification. a and b are weighting coefficients, which are determined by the random variable The prior probability is determined by the range of a, b∈[0,1], and its specific value can be adjusted and determined according to the classification performance;

[0100] f (i，j) The calculation formula of (ω) is:

[0101]

[0102] d (i，j) The calculation formula of (ω) is:

[0103]

[0104] in, and Respectively represent the mean of all sample bispectra of the i-th signal and the j-th signal at frequency ω, Represents the mean of all sample bispectra of the lth type signal at frequency ω.

[0105] In this embodiment, in step 6, the feature space processing parameters include the cumulative average length, the moving step, the number of effective frequencies, and the weighted coefficient of the inter-class classifiability measure.

[0106] In this embodiment, the classification performance evaluation indicators include Precision, Recall and F1.

[0107] Precision=TP / (TP+FP)

[0108] Recall=TP / (TP+FN)

[0109] F1=(2*Precision*Recall) / (Precision+Recall)

[0110] Among them, Precision, Recall and F1 represent precision, recall and the combination of the two respectively. TP is a positive sample whose prediction result is correct; FP is a positive sample whose prediction result is wrong; FN is a negative sample whose prediction result is wrong.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for identifying individual frequency hopping communication radiation sources based on feature space processing, characterized in that: The following steps are involved: Step 1: pre-processing the intermediate frequency original signal data obtained from multiple frequency hopping communication radiation sources of the same model to obtain multiple zero intermediate frequency signal data; each of the intermediate frequency original signal data includes the intermediate frequency data of the frequency hopping pulse and the corresponding individual labeling information of the frequency hopping communication radiation source; Step 2: normalizing and bispectral estimating the multiple zero intermediate frequency signal data, mapping the data to the bispectral feature space, and obtaining a signal bispectral feature set; Step 3: performing moving cumulative averaging on the signal bispectrum feature set to obtain an updated signal bispectrum feature set, and dividing the updated signal bispectrum feature set into a test feature set and a training feature set; Step 4: Smoothing the test feature set and the training feature set respectively to obtain a smoothed training feature set and a test feature set; Step 5: using a multi-classification feature selection method to perform bispectral feature selection and dimension reduction on the smoothed training feature set to obtain a selected bispectrum; extracting a training sample set and a test sample set from the smoothed training feature set and the test feature set respectively according to the effective frequency corresponding to the selected bispectrum; Step 6: Input the training sample set into the support vector machine for classification training to obtain a classification model; input the test sample set into the classification model to test the individual identification classification performance; and adjust the feature space processing parameters according to the individual identification classification performance evaluation results, repeatedly iterate steps 3 to 6 until the classification model with the best classification performance is obtained, and output the radiation source individual identification result.

2. The method according to claim 1, characterized in that In step 1, the preprocessing includes filtering, down-conversion and sampling rate conversion; the plurality of zero intermediate frequency signal data are in, is single frequency hopping pulse data, n=1, 2, ..., N, N is the number of frequency hopping pulses; F m is the radiation individual labeling information corresponding to the frequency hopping pulse data, m=1, 2, ..., M, where M is the number of frequency hopping communication radiation sources.

3. The method according to claim 2, characterized in that The step 2 comprises: Will Each frequency hopping pulse data As a set of observation data, after power normalization processing is performed on each set of observation data, the signal bispectral feature set is obtained by bispectral estimation. L is the Fourier transform length used in bispectrum estimation.

4. The method according to claim 3, characterized in that: The step 3 comprises: Step 31: Input the bispectral feature set of the i-th radiation source signal Set the cumulative average length to L1, the moving step to L2, and perform moving average calculation. The number of calculations N0 is: Where N is The number of eigenvectors of , fix(·) is the floor function; Step 32: Right Take the first L1 eigenvectors and perform cumulative average calculation to obtain the following result: Step 33: Right Move L2 feature vectors backward, take L1 feature vectors, and repeat N0 times according to step 32 to obtain the bispectral feature set of the i-th category after cumulative averaging. Where n0=1, 2, ..., N0; Step 34: Repeat steps 32 and 33 to complete the cumulative average processing of the bispectral feature sets of all types of communication radiation source signals to obtain an updated signal bispectral feature set. Step 35: Divided into training feature set and the test feature set Among them, n0=n1+n2.

5. The method according to claim 4, characterized in that The value ranges of L1 and L2 are: 1≤L1≤N / 1000, L1 / 10≤L2≤L1 / 2; the specific values ​​of L1 and L2 can be adjusted according to the classification performance.

6. The method according to claim 1, characterized in that The step 4 comprises: Step 41: extracting the smoothed feature set of the i-th category from the test feature set and the training feature set respectively, and the feature space dimension of the feature set is L×L; Step 42: For the feature set of the ith classification, the Savitsky-Golay smoothing algorithm is used to perform data smoothing on the feature sequence one by one according to each feature dimension, and the least squares fitting in the sliding window is realized by polynomials, and the weighted average sum of the center point in the sliding window and its surrounding points is calculated; after calculating and traversing all feature dimensions, the smoothed training feature set of the ith classification is obtained. and the test feature set Step 43: According to step 41 and step 42, the smoothing of the training feature set and the test feature set of all classes is completed to obtain the smoothed training feature set. and the test feature set 7. The method according to claim 1, characterized in that The step 5 includes: Step 51: Based on the training feature set The total number of individual categories and labeling information are statistically arranged into binary classification pairing combinations, and the individual category labeling sequence of the pairing combination is obtained as P = {i, j}; Step 52: Select a set of individual category labeling information (i, j) in P, and extract the corresponding training feature set of the i-th category respectively and the training feature set for the jth category Calculate the inter-class classifiability measure of these two classes, denoted as m (i，j) (ω); Step 53: Place m (i，j) (ω) Sort by size and select the larger Q m (i，j) (ω) corresponds to the effective frequency set ω (i，j) (q), as the selection bispectrum between the feature sets of the i-th category and the j-th category; where q = 1, 2, ..., Q, Q is the number of effective frequencies, and its value is determined comprehensively according to the dimensionality reduction requirements and classification performance; Step 54: By analogy with steps 52 and 53, the selected bi-spectra of all binary paired combination feature sets are calculated, and the selected bi-spectra of different binary paired combination feature sets are obtained by merging and removing duplicates. (i，j) (q), based on the same effective frequency, only take it once to get the selected bispectrum ω of all classes (i，j) (q′); Step 55: Select bispectrum ω based on (i，j) The effective frequency of (q′) is extracted from the smoothed training feature set to obtain a training sample set, and the corresponding bi-spectrum features are extracted from the smoothed test feature set to obtain a test sample set, and the feature dimension is reduced from the feature set dimension L×L to the sample set dimension q′.

8. The method according to claim 7, characterized in that m (i，j) The calculation formula of (ω) is: m (i，j) (ω)=a*f (i，j) (ω)+b*d (i，j) (oh) Among them, f (i，j) (ω) is the Fisher information separability, which reflects the contribution of the difference in the mean distribution of the two categories of the bispectrum to the classification, d (i，j) (ω) is the Bhattacharyya distance metric, which reflects the contribution of the difference in the variance of the two-category distribution to the classification. a and b are weighting coefficients, which are determined by the random variable The prior probability is determined, and the value range is a, b∈[0,1], and the specific value can be adjusted and determined according to the classification performance; f (i，j) The calculation formula of (ω) is: d (i，j) The calculation formula of (ω) is: in, and Respectively represent the mean of all sample bispectra of the i-th signal and the j-th signal at frequency ω, Represents the mean of all sample bispectra of the lth type signal at frequency ω.

9. The method according to claim 1, characterized in that: In step 6, the feature space processing parameters include the cumulative average length, the moving step, the number of effective frequencies, and the weighted coefficient of the inter-class classifiable measure.

10. The method according to claim 1, characterized in that The indicators for evaluating the classification performance include Precision n , Recall and F1, Precision=TP / (TP+FP) Recall=TP / (TP+FN) F1=(2*Precision*Recall) / (Precision+Recall) Among them, Precision n , Recall and F1 represent precision, recall and the combination of the two respectively. TP is a positive sample with a correct prediction result; FP is a positive sample with an incorrect prediction result; and FN is a negative sample with an incorrect prediction result.

Images