Radiation source identification method based on deep Bayesian clustering and category association

Abstract

The invention discloses a radiation source identification method based on deep Bayesian clustering and category association, and the method comprises the steps: inputting radar data into a simulation program, constructing a radiation source data set, carrying out the simulation of the data set, obtaining the data of a training set and a test set, constructing a category judgment model, and carrying out the training to obtain a feature extractor. After training is completed, fixing network parameters for subsequent feature extraction; a classification model based on a time sequence is constructed, the extracted pulse repetition interval is used for training the network, and after network convergence, fixed parameters are used for subsequent category association; constructing a Chinese restaurant clustering algorithm based on a Gaussian model for clustering without a specified category number, obtaining clustered clusters, and calculating the distance between the clusters in a time period to perform micro-cluster merging; and inputting the clustered clusters into a classification model constructed based on pulse repetition intervals to realize category association. The method not only adapts to a complex and changeable signal environment in a real environment, but also improves the recognition efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of radiation source sorting and identification, and in particular to a radiation source identification method based on deep Bayesian clustering and category association. Background Art

[0002] Radar emitter sorting and identification is one of the key technologies in electronic countermeasures, and is of great significance for battlefield situation awareness and target detection, tracking and classification in electronic countermeasures. The purpose of radar emitter sorting is to distinguish and identify different radar emitters in a complex electromagnetic environment through the characteristics of the received radar signal, thereby providing support for electronic countermeasures, target identification and threat analysis. The pulse descriptor word (PDW) is an important parameter used to describe the characteristics of radar pulse signals, including pulse arrival time (TOA), pulse width (PW), pulse arrival angle (OA), carrier frequency (CF) and pulse amplitude (Amplitude), which are key information in the sorting and identification of radar emitter signals.

[0003] Traditional signal sorting research mostly adopts a combination of pre-sorting and main sorting, using k-means clustering algorithm and density-based clustering method DBSCAN for preliminary clustering, and then performing main sorting through statistical algorithms. However, these clustering methods require pre-setting of cluster centers or thresholds, etc., which makes it difficult to flexibly adapt to different data distributions. In addition, methods based on statistics such as cumulative difference histograms, sequence difference histograms, and PRI transformation methods have high computational complexity and are time-consuming. In addition, as electronic countermeasures are becoming more diverse and modulation methods are more varied, traditional algorithms are difficult to handle complex situations. Summary of the invention

[0004] The purpose of the present invention is to propose a radiation source identification method based on deep Bayesian clustering and category association, aiming to solve the problems of large sorting errors due to unknown number of categories and insufficient flexibility caused by feature changes in existing radiation source signal sorting and identification technologies.

[0005] To achieve the above object, the present invention provides the following technical solution: a radiation source identification method based on deep Bayesian clustering and category association, comprising the following steps:

[0006] S1. Input radar data into the simulation program to construct a radiation source data set, which includes the pulse description words of the radiation source signal and the corresponding label data.

[0007] S2, according to the data set obtained in step S1, respectively simulate to obtain training set and test set data, build a category determination model, train to obtain a feature extractor, and after the training is completed, fix the network parameters for subsequent feature extraction;

[0008] S3, build a time series-based classification model, use the extracted pulse repetition interval to train the network, and fix the parameters after the network converges for subsequent category association;

[0009] S4. Construct a deep Bayesian clustering model based on the Gaussian model and the Chinese restaurant model for clustering without specifying the number of categories, use the Chinese restaurant model to obtain clusters after clustering, and calculate the distance between clusters within a time period to merge micro-clusters;

[0010] S5. Input the clusters clustered in step S4 into a classification model constructed based on the pulse repetition interval to achieve category association.

[0011] Preferably, in step S1, the radar data includes power-on time, power-off time, pulse repetition interval, carrier frequency, pulse width, amplitude, azimuth and parameter distribution, and the radar radiation source signal parameter distribution includes the start time and end time of the radar radiation source signal generation.

[0012] Preferably, in step S1, the steps of constructing the radiation source data set are as follows:

[0013] 1) According to the designed radiation source signal generation time, carrier frequency, pulse width, amplitude, azimuth parameter range, construct the corresponding radiation source signal data and send it to the radiation source signal simulation program. The signal parameters are described as PDW i ={TOA i ,CF i ,PW i ,PA i ,AOA i};

[0014] 2) The obtained PDW parameter features retain the carrier frequency, pulse width and azimuth angle as the training input parameters of the feature extraction network and the input parameters of the deep clustering.

[0015] Preferably, in step S1, the number of categories is set according to the required number of radiation sources, and for each radiation source pulse description word, five parameters including generation time, carrier frequency, pulse width, amplitude, and azimuth are selected for simulation, and the power-on time and power-off time of each radiation source are set. At the same time, according to the existing radar characteristic modulation modes, different modulation types are simulated for each radar PRI, CF, and PW, including fixed, staggered, agile, and jitter, and a certain amount of noise interference is added to simulate real environmental noise and measurement errors.

[0016] Preferably, in step S2, the feature extractor is trained using the training set, and the model parameters are optimized to minimize the loss function. After the training is completed, the fixed network parameters are used for subsequent feature extraction; wherein the input of the category determination model is a 1*3 feature vector corresponding to the three features of carrier frequency, pulse width, and azimuth in the pulse description word, and each input corresponds to a label of a radar radiation source category;

[0017] The feature extraction network is composed of linear layers, the optimizer is set to Adam, the learning rate is 1e-3, the decay factor is 0.1, and the loss is the cross entropy loss. C is the number of categories, y i is the i-th element in the true label vector, is the i-th element in the probability distribution predicted by the model.

[0018] Preferably, in step S3, the specific steps include the following:

[0019] The PRI sequence corresponding to each category of radiation source signal is extracted, the corresponding category label is generated, the PRI dataset is constructed and split into a training set and a validation set at a ratio of 80% and 20%, and a classification network model is constructed. The feature extractor is trained using the training set, and the model parameters are optimized to minimize the loss function. After the training is completed, the network parameters are fixed for subsequent category association.

[0020] Preferably, in step S3, the input of the time series-based classification model is a PRI vector of 1*200, and each type of PRI data set is a time series slice obtained by extracting PRI data of each type of radar in a continuous time with a length of 200 and a step length of 25, and the output is the radar category corresponding to each PRI sequence;

[0021] The class association network is composed of convolutional layers, linear layers, and batch normalization layers. The optimizer is set to Adam, the learning rate is set to a variable learning rate to prevent overfitting and accelerate network convergence, the decay factor is 0.1, and the loss is the cross entropy loss.

[0022] Preferably, step S4 specifically includes the following:

[0023] The features of the test data are input into the feature extractor to obtain the data feature vector, which is then input into the Bayesian clustering framework to calculate the likelihood function between each new data point and the existing cluster. The probability that the new data point belongs to the existing cluster and the probability that it belongs to the new cluster are obtained according to the Chinese restaurant process.

[0024] P={p1,p2,p3,…,p C ,p C+1},

[0025] Where C is the number of existing clusters, C+1 represents a new cluster, and the initial deinterleaved radiation source pulse signal is obtained; when the clustering time reaches T th When , the distance between each cluster is calculated, and the distance between the i-th cluster and the k-th cluster is expressed as:

[0026]

[0027] in, and Represents cluster C i and cluster C j , S is the covariance matrix of the samples. When the distance is less than the set association threshold ∈, the i-th cluster and the j-th cluster are merged and the clustering result is updated, which effectively reduces the batch increase and improves the accuracy of subsequent association operations.

[0028] Preferably, in step S5, the arrival time information corresponding to each cluster of the clustering results obtained in step 4 is differentiated, and the PRI sequence of each cluster obtained by clustering is extracted with a length of 200 and a variable step length set according to the sequence length, and input into the trained classification network. The decision is made by PRI sequence category prediction to obtain the category information of each micro-cluster, thereby realizing the association between the clustering results and the actual category.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. The present invention introduces a deep Bayesian clustering model and an end-to-end category association algorithm. Under the Bayesian framework, clustering is performed in combination with the Chinese restaurant process model and Gaussian fitting. New clusters can be created dynamically without specifying the number of clusters, and the clustering results can be gradually adjusted according to the existing signal characteristics, thereby improving the flexibility and accuracy of the sorting and identification of radiation source signals and adapting to the complex and changeable signal environment in the real environment.

[0031] 2. In the present invention, a classification network based on PRI sequences is introduced to identify the categories to which clusters belong, and to achieve category association between micro-clusters. The PRI features of the clustered clusters are input into the neural network for identification. Compared with traditional methods, the present invention overcomes the problems that the radar radiation source carrier frequency and azimuth angle undergo large migration changes over time and are difficult to identify associations, and that the same category of radiation sources contains multiple different individuals. In addition, end-to-end identification also greatly improves the recognition efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0033] In the attached picture:

[0034] Figure 1 It is a schematic diagram of the flow of the radiation source identification method of the present invention;

[0035] Figure 2 It is a deep Bayesian clustering and category association model diagram of the present invention. DETAILED DESCRIPTION

[0036] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0037] Embodiment: The present invention proposes a radiation source identification method based on deep Bayesian clustering and category association, in order to improve the flexibility and accuracy of sorting and identifying radiation source signals. It combines the feature extraction capability of deep learning and the flexibility of the Chinese restaurant process without specifying the number of categories, and introduces a classification network based on PRI sequence to identify the category to which the cluster belongs, and realize the category association between micro-clusters; Figure 1 and Figure 2 As shown, the specific steps include:

[0038] Step 1: Input the radar's appearance time, disappearance time, pulse repetition interval, carrier frequency, pulse width, amplitude, and azimuth data into the simulation program to construct a radiation source data set, which includes the pulse description word of the radiation source signal and the corresponding label data;

[0039] The distribution of the radar radiation source signal parameters designed for this project is shown in the figure below: Figure 2 As shown, it includes the start time and end time of the radar radiation source signal.

[0040] In step 1, the steps of constructing the radiation source data set and the test set are as follows:

[0041] 1) According to the designed radiation source signal generation time, carrier frequency, pulse width, amplitude, azimuth parameter range, construct the corresponding radiation source signal data and send it to the radiation source signal simulation program. The signal parameters can be described as PDW i ={TOA i ,CF i ,PW i ,PA i ,AOA i};

[0042] 2) The obtained PDW parameter features retain carrier frequency, pulse width and azimuth as the training input parameters of the feature extraction network and the input parameters of deep clustering, because these three features can reflect the main characteristics of the radar signal;

[0043] Step 2: Split the dataset obtained in step 1 into a training set and a test set, and build a deep neural network as a feature extractor to extract high-level features from PDW. Use the training set to train the feature extractor and optimize the model parameters to minimize the loss function. After training, fix the network parameters for subsequent feature extraction. The input of the training dataset is a 1*3 feature vector, and the corresponding label is a 1-dimensional radar emitter category label;

[0044] The feature extraction network is composed of linear layers, the optimizer is set to Adam, the learning rate is 1e-3, the decay factor is 0.1, and the loss is the cross entropy loss. C is the number of categories, y i is the i-th element in the true label vector, is the i-th element in the probability distribution predicted by the model.

[0045] Step 3: Extract the PRI sequence corresponding to each category of radiation source signal, generate the corresponding category label, construct the PRI dataset and split it into a training set and a validation set in a ratio of 80% and 20%.

[0046] Constructing Figure 2 The classification network model shown uses the training set to train the feature extractor, and uses the training set to train the feature extractor, and optimizes the model parameters to minimize the loss function. After the training is completed, the fixed network parameters are used for subsequent category association.

[0047] The class association network is composed of convolutional layers, linear layers, and batch normalization layers. The optimizer is set to Adam, the learning rate is set to a variable learning rate to prevent overfitting and accelerate network convergence, the decay factor is 0.1, and the loss is the cross entropy loss. Where y is the true value, is the predicted value.

[0048] Step 4: Construct a deep Bayesian clustering model based on the Chinese Restaurant Process (CRP) and Gaussian model to cluster the radiation source signals. The Chinese restaurant model is a probabilistic process used to describe infinite mixture models. The CRP model can automatically determine the number of clusters and simulate the process of customers entering the restaurant and choosing a table to sit. Assume that there are currently n customers entering the restaurant, and there are n customers on the kth and nth tables. k customers, there are K tables with customers seated, and the probability of a customer joining an existing table is The probability of opening a new table is where α is a positive hyperparameter called the concentration parameter.

[0049] For a new radiation source signal data point x, the high-dimensional feature f(x) of the data is obtained through the feature extraction network, and its belonging to each existing cluster C is calculated i The likelihood function of :

[0050]

[0051] in, represents multi-source Gaussian distribution, μ i and∑ i They are clusters C i The mean vector and covariance matrix of the prior probability P(C i ) is defined by CRP:

[0052]

[0053] Among them, |C i | represents cluster C i The probability of creating a new cluster P(new) is:

[0054]

[0055] Finally, the new data point x belongs to cluster C i The probability P(C i |x):

[0056] P(C i |x)=L(f(x)|C i )·P(C i )

[0057] For the data points that need to create a new cluster, the mean of the cluster needs to be initialized, where the mean can be directly taken as the value of the new data point, and the covariance matrix is ​​based on the initial parameter α init And the value of the new point is constructed. Whenever a new point is added to an existing cluster, the mean and covariance matrix of the cluster are updated:

[0058]

[0059] After each time period T th , calculate the distance between each existing cluster:

[0060]

[0061] in, and Represents cluster C i and cluster C j, S is the mean vector of the samples, and S is the covariance matrix of the samples. If the distance between the i-th cluster and the j-th cluster is less than the set association threshold ∈, the i-th cluster and the j-th cluster are merged and the clustering result is updated.

[0062] Step 5: Differencing the arrival time information corresponding to each cluster of the clustering results obtained in step 4, setting a variable step size according to the size of the cluster, extracting a PRI group with a length of 200, PRI i =TOA i -TOA i-1 ,The PRI group of each cluster can be expressed as {PRI1, PRI2, PRI3, ..., PRI N}, N is determined according to the size of the cluster, and the obtained PRI group is input into the trained classification network. According to the prediction results of this group of data, the final category label is determined.

[0063] Finally, it should be noted that the above description is only a preferred example of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A radiation source identification method based on deep Bayesian clustering and category association, characterized in that: The following steps are involved: S1. Input radar data into the simulation program to construct a radiation source data set, which includes the pulse description words of the radiation source signal and the corresponding label data. S2, according to the data set obtained in step S1, respectively simulate to obtain training set and test set data, build a category determination model, train to obtain a feature extractor, and after the training is completed, fix the network parameters for subsequent feature extraction; S3, build a time series-based classification model, use the extracted pulse repetition interval to train the network, and fix the parameters after the network converges for subsequent category association; S4. Construct a deep Bayesian clustering model based on the Gaussian model and the Chinese restaurant model for clustering without specifying the number of categories, use the Chinese restaurant model to obtain clusters after clustering, and calculate the distance between clusters within a time period to merge micro-clusters; S5. Input the clusters clustered in step S4 into a classification model constructed based on the pulse repetition interval to achieve category association.

2. The radiation source identification method based on deep Bayesian clustering and category association according to claim 1, characterized in that: In step S1, the radar data includes power-on time, power-off time, pulse repetition interval, carrier frequency, pulse width, amplitude, azimuth angle and parameter distribution, and the radar radiation source signal parameter distribution includes the start time and end time of the radar radiation source signal.

3. The radiation source identification method based on deep Bayesian clustering and category association according to claim 2 is characterized in that: In step S1, the steps of constructing the radiation source dataset are as follows: 1) According to the designed radiation source signal generation time, carrier frequency, pulse width, amplitude, azimuth parameter range, construct the corresponding radiation source signal data and send it to the radiation source signal simulation program. The signal parameters are described as PDW i ={TOA i ,CF i ,PW i ,PA i ,AOA i }; 2) The obtained PDW parameter features retain the carrier frequency, pulse width and azimuth angle as the training input parameters of the feature extraction network and the input parameters of the deep clustering.

4. The radiation source identification method based on deep Bayesian clustering and category association according to claim 3 is characterized by: In step S1, the number of categories is set according to the required number of radiation sources. For each radiation source pulse description word, five parameters including generation time, carrier frequency, pulse width, amplitude and azimuth are selected for simulation. The power-on time and power-off time of each radiation source are set. At the same time, according to the existing radar characteristic modulation modes, different modulation types are simulated for each radar PRI, CF, and PW, including fixed, staggered, agile and jitter, and a certain amount of noise interference is added to simulate the actual environmental noise and measurement errors.

5. The radiation source identification method based on deep Bayesian clustering and category association according to claim 1, characterized in that: In step S2, the feature extractor is trained using the training set to optimize the model parameters to minimize the loss function. After the training is completed, the fixed network parameters are used for subsequent feature extraction; wherein the input of the category determination model is a 1*3 feature vector corresponding to the three features of carrier frequency, pulse width, and azimuth in the pulse description word, and each input corresponds to a label of a radar radiation source category; The feature extraction network is composed of linear layers, the optimizer is set to Adam, the learning rate is 1e-3, the decay factor is 0.1, and the loss is the cross entropy loss. C is the number of categories, y i is the i-th element in the true label vector, is the i-th element in the probability distribution predicted by the model.

6. The radiation source identification method based on deep Bayesian clustering and category association according to claim 1, characterized in that: In step S3, the specific steps include the following: The PRI sequence corresponding to each category of radiation source signal is extracted, the corresponding category label is generated, the PRI dataset is constructed and split into a training set and a validation set at a ratio of 80% and 20%, and a classification network model is constructed. The feature extractor is trained using the training set, and the model parameters are optimized to minimize the loss function. After the training is completed, the network parameters are fixed for subsequent category association.

7. The radiation source identification method based on deep Bayesian clustering and category association according to claim 1, characterized in that: In step S3, the input of the time series-based classification model is a 1*200 PRI vector, and each type of PRI data set is a time series slice obtained by extracting PRI data of each type of radar in a continuous time with a length of 200 and a step length of 25, and the output is the radar category corresponding to each PRI sequence; The class association network is composed of convolutional layers, linear layers, and batch normalization layers. The optimizer is set to Adam, the learning rate is set to a variable learning rate to prevent overfitting and accelerate network convergence, the decay factor is 0.1, and the loss is the cross entropy loss.

8. The radiation source identification method based on deep Bayesian clustering and category association according to claim 1, characterized in that: Step S4 specifically includes the following: The features of the test data are input into the feature extractor to obtain the data feature vector, which is then input into the Bayesian clustering framework to calculate the likelihood function between each new data point and the existing cluster. The probability that the new data point belongs to the existing cluster and the probability that it belongs to the new cluster are obtained according to the Chinese restaurant process. P={p1,p2,p3,…,p C ,p C+1 }, Where C is the number of existing clusters, C+1 represents a new cluster, and the initial deinterleaved radiation source pulse signal is obtained; when the clustering time reaches T th When , the distance between each cluster is calculated, and the distance between the i-th cluster and the k-th cluster is expressed as: in, and Represents cluster C i and cluster C i The mean vector of , S is the covariance matrix of the sample. When the distance is less than the set association threshold ∈, the i-th cluster and the j-th cluster are merged and the clustering result is updated.

9. The radiation source identification method based on deep Bayesian clustering and category association according to claim 1, characterized in that: In step S5, the arrival time information corresponding to each cluster of the clustering results obtained in step 4 is differentiated, and the PRI sequence of each cluster obtained by clustering is extracted with a length of 200 and a variable step length set according to the sequence length, and input into the trained classification network. The decision is made by the PRI sequence category prediction to obtain the category information of each micro-cluster, thereby realizing the association between the clustering results and the actual category.

Images