Multifunction radar signal sorting method and system based on hypergraph

Abstract

The application discloses a multifunctional radar signal sorting method and system based on a hypergraph, and the method steps are as follows: S1, for RIPS obtained through radar reconnaissance receiver interception, interference pulses in the RIPS are removed; S2, the remaining pulse sequence is subjected to screening rules and FCM clustering, and clustering labels of partial pulse sequences are obtained; S3, feature parameters and data potential energy values of the remaining pulse sequence are used to construct a hypergraph; and S4, the clustering labels of the partial pulse sequences are used to learn the constructed hypergraph, and a final radar signal sorting result is obtained. The application can sort multifunctional radar signals without a marked sample, can effectively alleviate the 'batch increase' problem, and has good sorting performance.

Description

Technical Field

[0001] The present invention belongs to the field of radar reconnaissance technology, and in particular relates to a hypergraph-based multi-function radar (MFRs, Multi-Function Radars) signal sorting method and system. The present invention can be used for electronic reconnaissance. Background Art

[0002] Radar signal sorting is a key component of radar reconnaissance technology, and its performance is crucial for subsequent processing steps such as interference, positioning, and tracking. Typically, the input data for signal sorting is the pulse description words (PDWs) of radar interleaved pulse sequences (RIPS). These data contain signal parameters such as direction of arrival (DOA), time of arrival (TOA), pulse width (PW), and radio frequency (RF). However, in complex electromagnetic environments, RIPS intercepted by a reconnaissance receiver are subject to varying degrees of noise and interference. Furthermore, the parameters in the PDWs obtained through signal detection and parameter estimation inevitably have certain estimation errors. Large estimation errors can easily lead to the inclusion of some interfering pulses in the RIPS, severely impacting radar signal sorting performance. Furthermore, with the continuous development of technologies such as new phased array radars, MFRs automatically change operating modes based on battlefield conditions. These pulse parameters vary widely, and the correlation between pulse sequences generated under different operating modes is low. Conventional sorting methods can easily misinterpret pulse sequences from the same MFR in different operating modes as pulse sequences from multiple radars, leading to the "overlapping" problem. Therefore, research is urgently needed to investigate effective sorting methods to mitigate this "overlapping" problem in complex electromagnetic environments containing interfering pulses and measurement errors.

[0003] The existing radar signal sorting methods can be divided into two categories: pulse repetition interval (PRI) based sorting method and machine learning (ML) based sorting method. The PRI based sorting method mainly includes cumulative difference histogram, sequence difference histogram and PRI transformation method, etc. This kind of method obtains the secondary parameter PRI by calculating TOA, and sorts by taking PRI as the main characteristic parameter. This kind of method has a good sorting effect on radar signals with conventional pulse interval modulation mode. However, when facing radar signals with various pulse interval modulation types generated by new MFRs, it exposes the shortcomings of large calculation amount and obvious decline in sorting performance. With the development of ML technology, researchers began to study radar signal sorting based on machine learning. At present, the intelligent sorting algorithm of radar radiation source based on machine learning mainly includes supervised model and unsupervised clustering model. The radar signal sorting based on supervised model has good sorting performance, but this kind of method faces great challenges in training sample acquisition and algorithm real-time, especially in the battlefield, it is difficult to obtain a large number of labeled samples to train the model. Unlike this, the radar signal sorting based on unsupervised clustering model does not need labeled data, and does not need to spend a lot of time to train the model, and the specified hyperparameters are less than the supervised sorting method, which is more in line with the actual demand. Although the clustering method is from the perspective of unsupervised learning to sort radar signals, but they only have good sorting performance in simple scenes such as different types of conventional radar and non-overlapping signal parameters. For MFRs signal sorting under multiple working modes, especially under non-ideal conditions such as interference pulses and measurement errors, the sorting performance of this kind of method declines rapidly.

[0004] In recent years, researchers have begun to study the "increased batching" caused by MFRs signal sorting. Some researchers use the K-nearest neighbor method to construct a graph structure and study the deinterlacing of MFRs signals from the perspective of residual graph convolutional networks. Although this approach can achieve good sorting performance, it still requires a small amount of labeled sample data. In addition, some scholars have introduced complex networks into the field of radar signal sorting, providing a new approach to studying how to alleviate the "increased batching" of MFRs signal sorting. Some scholars have proposed a MFRs signal sorting method based on complex networks. They use a sliding window limited penetration visibility graph to construct a complex network from RIPS, and use a label propagation algorithm and a density peak clustering algorithm to detect the community structure, thereby realizing MFRs signal sorting. However, this method first constructs a complex network for each single feature parameter, and then simply merges each complex network to obtain the overall complex network. This network construction method does not consider the correlation between feature parameters when constructing the network. Based on the shortcomings of existing research, the present invention proposes to use a hypergraph method to construct a network. While traditional graph construction only considers the correlation of individual features in the data, hypergraphs exhibit natural high-order correlations in their structure, and the resulting hypergraphs encompass the correlations between multiple features in the data. Therefore, compared to graph structures derived using complex network methods, constructing a hypergraph from a PDW sequence considers the correlations between various parameters in the PDW, making it more advantageous for radar signal sorting.

[0005] Based on the above research status, the present invention proposes a multifunctional radar signal sorting method and system based on hypergraph. Summary of the Invention

[0006] Aiming at the problem of "increased batches" that easily occurs during MFRs signal sorting, and the inevitable presence of interference pulses and parameter estimation errors in RIPS intercepted in complex electromagnetic environments, which lead to a decrease in sorting performance, the present invention proposes a multi-function radar signal sorting method and system based on hypergraph.

[0007] In order to achieve the above objectives, the present invention adopts the following technical solutions:

[0008] A multifunctional radar signal sorting method based on a hypergraph comprises the following steps:

[0009] S1. For the RIPS intercepted by the radar reconnaissance receiver, remove the interference pulses in the RIPS;

[0010] S2. Combine the screening rule and Fuzzy C-means (FCM) clustering to obtain the cluster labels of the remaining pulse sequences.

[0011] S3, constructing a hypergraph based on the characteristic parameters of the remaining pulse sequences and their data potential energy values;

[0012] S4. Use the cluster labels of the remaining pulse sequences to learn the constructed hypergraph to obtain the final radar signal sorting result.

[0013] Furthermore, in step S1, the interference pulse is removed using the data field method, and the specific steps are as follows:

[0014] Step S11: perform maximum and minimum normalization on the three characteristic parameters of RF, PW and DOA in the PDW data:

[0015]

[0016] Among them, x d represents the d-dimensional feature parameter vector, represents the dth feature of the i-th pulse;

[0017] Step S12: Calculate the data potential energy value of each PDW data:

[0018]

[0019] Among them, pdw m represents the mth PDW, d(pdw i ,pdw j ) indicates pdw i with pdw j The Euclidean distance between them, σ is the radiation factor, N is the number of pulses, F(pdw m ) indicates pdw m Data potential energy value;

[0020] Step S13: Calculate the detection threshold of the data potential energy value, and consider the data points whose data potential energy value is much smaller than the average data potential energy value as interference pulses and remove them. Specifically, it is expressed as:

[0021]

[0022] in, is the detection threshold, F(pdw jl ) is the interference pulse pdw jl The data potential energy value. ξ is the detection coefficient set according to the electromagnetic environment, mean represents the average value, represents the sum of the data potential energy values ​​of all pulses, N is the number of pulses, F(pdw i ) is the data potential energy value of the i-th PDW.

[0023] Furthermore, in step S2, the proposed filtering rules are applied to the remaining PDW data to obtain local core point data. Furthermore, FCM clustering is performed on the remaining PDW data to obtain the membership value of each pulse. The local core points are then traversed to obtain cluster labels of local core points with a membership value greater than the set value. These cluster labels are used as known information for subsequent learning stages, as follows:

[0024] Step S21, filtering the remaining PDW data according to the proposed filtering rules to obtain local core points; the filtering rules are expressed as:

[0025]

[0026] Among them, μ(δ) represents the relative distance mean of all data, μ(ρ) represents the local density mean of all data, σ(ρ) represents the local density standard deviation of all data, μ(RNN) represents the number of reverse nearest neighbors of all data, μ and σ represent the mean and standard deviation respectively, local_core_pdw i represents i local core points, δ i represents the relative distance of the i-th data, δ represents the relative distance of all data, ρ i represents the local density of the i-th data, ρ represents the local density of all data, |RNN i | represents the number of reverse nearest neighbors of the i-th data, and |RNN| represents the number of reverse nearest neighbors of all data;

[0027] Step S22: Estimate the number of MFRs P. P is determined based on the inflection point of the cluster sum of squares (CSS) and the maximum value of the silhouette coefficient (SC). CSS and SC are expressed as:

[0028]

[0029] Among them, c represents the number of cluster centers, n represents the number of remaining pulses, and d(C i ,pdw j ) represents the i-th cluster center C i With the j-th PDW data pdw j the distance between them;

[0030]

[0031] Among them, a(pdw i ) is pdw i The average distance to other PDW data points in the same cluster, b(pdw i ) is pdwi To its nearest neighbor cluster (i.e., not including pdw i The average distance of all PDW data points in the cluster where the data point is located;

[0032] Step S23: Perform FCM clustering on the remaining PDW data according to the MFR number P obtained in step S22 to obtain the membership value u of each PDW data. ij , traverse the membership values ​​of the local core points obtained in step S21, and when a certain membership value of a local core point is greater than a set value, extract the cluster label of the local core point; calculate the membership value u ij Expressed as:

[0033]

[0034] Among them, u ij The jth PDW data belongs to the i-th cluster center C i The membership value of is , m is the fuzzy index, which is usually an integer greater than 1.

[0035] Furthermore, in step S3, the K-nearest neighbor (KNN) method is used to construct the hyperedge of each node for the remaining PDW data, thereby obtaining a hypergraph. The hypergraph structure is represented by the association matrix H, which is expressed as:

[0036]

[0037] Among them, v H Represents a node in the hypergraph, e H Represents a hyperedge in the hypergraph.

[0038] Furthermore, in step S4, the steps for learning the constructed hypergraph are as follows:

[0039] Step S41, constructing a node classification function F; taking two-class label prediction as an example, the node classification function F(v H ) is expressed as:

[0040]

[0041] Among them, S and Indicates that the entire hypergraph is divided into two parts, one part is represented by S and the remaining part is represented by express;

[0042] Step S42, construct a regularizer Ω(F) on the hypergraph, through F(v H ) represents the regularizer Ω(F) as:

[0043]

[0044] Where c(S) represents the hypergraph partition, Δ is the normalized Laplace matrix, and E H represents the hyperedge set, e H represents a hyperedge, w(e H ) represents the hyperedge e H The weight of δ(e H ) means e H The hyperedge degree, u H and v H Represents two nodes on the hypergraph, F(u H ) represents u H The label, d(u H ) represents u H The node degree, F(v H ) means v H The label, d(v H ) means v H Node degree;

[0045] Step S43: The objective function of the hypergraph learning model can be divided into two parts. The first part is the empirical loss R of the learning label. emp (F), such as the least squares loss. The second part is the regularizer Ω(F) on the hypergraph, which represents the smoothness of the label distribution on the constructed hypergraph structure; through linear combination, the minimization objective function of hypergraph learning is expressed as:

[0046]

[0047] in, represents the minimization objective function of hypergraph learning, and λ is the trade-off coefficient;

[0048] S44. The hypergraph learning minimization objective function in step S43 is further derived and expressed as:

[0049]

[0050] Among them, Y represents the initial label vector, for the labeled node v H , if the node belongs to the jth class, then Y(v H ) is 1, and the other positions are 0; for the unmarked node u H ,Y(u H ) are all set to 0.5, indicating that there is no prior category information for these unlabeled nodes;

[0051] When the hypergraph learning objective function reaches the minimum and converges, the node classification function is expressed as:

[0052]

[0053] Among them, Dv Denotes the node degree matrix, D e Represents the hyperedge degree matrix, μ generally takes a value between 0 and 1;

[0054] In step S45, when the hypergraph learning objective function reaches the minimum and converges, the label message transmission on the hypergraph reaches a balanced state, each node has corresponding label information, and the radar signal sorting can be completed by classifying the nodes with different labels.

[0055] The present invention also discloses a multifunctional radar signal sorting system based on a hypergraph, which is used to execute the above method and specifically includes the following modules:

[0056] Interference pulse removal module: removes interference pulses in the intercepted radar interleaved pulse sequence RIPS;

[0057] The partial local core point label extraction module uses the screening rule and fuzzy C-means FCM clustering to cluster the remaining pulse sequence and obtain the cluster label of the remaining pulse sequence;

[0058] Hypergraph construction module: constructs a hypergraph based on the characteristic parameters of the remaining pulse sequences and their data potential energy values;

[0059] Hypergraph learning module: Use the cluster labels of the remaining pulse sequences to learn the constructed hypergraph;

[0060] Sorting module: When the hypergraph learning in the hypergraph learning module reaches convergence, the label information of each node can be obtained, and the nodes can be classified to obtain the radar signal sorting results.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] 1. Taking into account the correlation between feature parameters, the present invention proposes to use a hypergraph method to construct a network for the multidimensional feature parameters of a pulse sequence. The resulting hypergraph structure reflects the correlation between the feature parameters of data samples, ensuring the accuracy of hypergraph learning and thus improving the sorting performance.

[0063] 2. This invention uses a data field method to remove interfering pulses from RIPS and uses the resulting data potential energy value of each pulse as a characteristic parameter, combined with existing PDW parameters for subsequent use. The data potential energy value can characterize the local density of each pulse data and, to a certain extent, reflect the spatial concentration of PDW data. Therefore, incorporating the data potential energy value into the PDW parameters can effectively assist in MFRs signal sorting.

[0064] 3. Taking into account that the actually intercepted pulse sequence contains less known information, the present invention proposes a new sample screening rule, which screens the remaining pulse sequence after removing the interference pulses to obtain local core points, and combines FCM clustering to obtain clustering labels of some local core points, thereby automatically obtaining the label data required for subsequent hypergraph learning.

[0065] 4. This invention can separate pulse sequences with different operating modes within the same MFR into the same radar radiation source, effectively alleviating the problem of "increased batches" in sorting. The present invention's sorting performance is superior to existing complex network-based and unsupervised clustering-based methods, and it exhibits greater robustness under non-ideal conditions.

[0066] In summary, in view of the fact that the "increased batch" problem is prone to occur during MFRs signal sorting, and the inevitable interference pulses and parameter estimation errors in the RIPS intercepted in a complex electromagnetic environment, which lead to a decrease in sorting performance, the present invention proposes a multi-function radar signal sorting method and system based on hypergraph, introduces the hypergraph method into the field of radar signal sorting, uses the data field method to eliminate interference pulses in RIPS, and proposes a screening rule to screen local core points, and combines FCM clustering to extract clustering labels of some local core points, so that MFRs signals can be sorted without prior information in the actual battlefield. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a flowchart of a multifunctional radar signal sorting method based on a hypergraph according to a preferred embodiment of the present invention;

[0068] Figure 2 Flowchart of removing interference pulses in RIPS according to a preferred embodiment of the present invention;

[0069] Figure 3 This is a flow chart of extracting cluster labels of some local core points in a preferred embodiment of the present invention;

[0070] Figure 4 A schematic diagram of hyperedge segmentation in a preferred embodiment of the present invention;

[0071] Figure 5 It is a two-dimensional visualization distribution diagram of the original data containing the interference pulse and the two-dimensional visualization distribution diagram of the remaining data after the interference pulse is removed using the present invention;

[0072] Figure 6 The two-dimensional visualization distribution map of the filtered local core points and the two-dimensional visualization distribution map of some local core points with cluster labels;

[0073] Figure 7 It is a two-dimensional visualization distribution diagram after the ablation experiment of the present invention and a two-dimensional sorting result diagram obtained using the present invention;

[0074] Figure 8 A two-dimensional visualization of the distribution of real results using spectral clustering, First Integer Neighbor Clustering Hierarchy (FINCH) and complex network sorting methods;

[0075] Figure 9 It is the sorting accuracy of the present invention under different parameter estimation error ratios and different missed pulse rates;

[0076] Figure 10 This is a block diagram of a multifunctional radar signal sorting system based on a hypergraph according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0077] The following describes the implementation of the present invention using preferred embodiments. Those skilled in the art will readily understand the other advantages and benefits of the present invention from the disclosure herein. The present invention may also be implemented or applied through various other specific embodiments, and the details in this specification may be modified or altered based on different perspectives and applications without departing from the spirit of the present invention.

[0078] See also Figure 1 This embodiment provides a multi-function radar signal sorting method based on a hypergraph. Taking four radars as an example, there are two multi-function radars, Radar1 and Radar3 are conventional radars, and Radar2 and Radar4 are multi-function radars, each with two working modes. Where X represents the intercepted PDW data, J represents the PDW data of the interference pulse, and C X (J) represents the PDW data of the remaining pulse, C XL (J) is the PDW data of the local core point, U is the membership matrix of the remaining PDW data, C X1 (J) represents the sorting result of Radar1, and so on, specifically including the following steps:

[0079] S1. For the RIPS intercepted by the radar reconnaissance receiver, first perform maximum and minimum normalization on the characteristic parameters of RIPS, then calculate the data field potential energy value of all pulse sequences, and judge whether the data potential energy value of all pulse sequences is greater than the detection threshold. The pulses less than the detection threshold are interference pulses, and finally the interference pulses are eliminated. For specific steps, refer to Figure 2 .

[0080] Step S11: perform maximum and minimum normalization on the three characteristic parameters of RF, PW and DOA in the PDW data:

[0081]

[0082] Among them, x d represents the d-dimensional feature parameter vector, represents the dth feature of the i-th pulse.

[0083] Step S12: Calculate the data potential energy value of each PDW data. The calculated data potential energy value can be expressed as:

[0084]

[0085] Among them, pdw m represents the mth PDW, d(pdw i ,pdw j ) indicates pdw i with pdw j The Euclidean distance between them, σ is the radiation factor, N is the number of pulses, F(pdw m ) indicates pdw m The data potential value.

[0086] Step S13: Calculate the detection threshold of the data potential energy value, and consider the data points whose data potential energy value is much smaller than the average data potential energy value as interference pulses and eliminate them. The detection threshold can be expressed as:

[0087]

[0088] in, is the detection threshold, F(pdw jl ) is the interference pulse pdw jl The data potential energy value. ξ is the detection coefficient set according to the electromagnetic environment, mean represents the average value, represents the sum of the data potential energy values ​​of all pulses, N is the number of pulses, F(pdw i ) is the data potential energy value of the i-th PDW.

[0089] S2. After removing the interference pulses in step S1, on the one hand, the proposed filtering rules are used to obtain the local core point data for the remaining PDW data, and on the other hand, the FCM clustering is performed on the remaining PDW data to obtain the membership value of each pulse. Then, the local core points are traversed to obtain the cluster labels of the local core points that are greater than the set membership value. The cluster labels of these local core points are used as known information for subsequent learning. For specific steps, refer to Figure 3 .

[0090] Step S21: Filter the remaining PDW data according to the proposed filtering rules to obtain local core points. The filtering rules can be expressed as:

[0091]

[0092] Among them, μ(δ) represents the relative distance mean of all data, μ(ρ) represents the local density mean of all data, σ(ρ) represents the local density standard deviation of all data, μ(|RNN|) represents the number of reverse nearest neighbors of all data, μ and σ represent the mean and standard deviation respectively, local_core_pdw i represents i local core points, δ i represents the relative distance of the i-th data, δ represents the relative distance of all data, ρ i represents the local density of the i-th data, ρ represents the local density of all data, |RNN i | represents the number of reverse nearest neighbors of the i-th data, and |RNN| represents the number of reverse nearest neighbors of all data;

[0093] Step S22: Estimate the MFR quantity P. P is determined based on when CSS has an inflection point and when SC reaches a maximum value. CSS and SC can be expressed as:

[0094]

[0095] Among them, c represents the number of cluster centers, n represents the number of remaining pulses, and d(C i ,pdw j ) represents the i-th cluster center C i With the j-th PDW data pdw j the distance between them;

[0096]

[0097] Among them, a(pdw i ) is pdw i The average distance to other PDW data points in the same cluster, b(pdw i ) is pdw i To its nearest neighbor cluster (i.e., not including pdw i The average distance of all PDW data points in the cluster.

[0098] Step S23: Perform FCM clustering on the remaining PDW data according to the MFR number P obtained in step S22 to obtain the membership value u of each PDW data. ij , the membership value of the local core point obtained by traversal, when a certain membership value of a local core point is greater than 0.5, the cluster label of the local core point is extracted and the membership value u is calculated ij It can be expressed as:

[0099]

[0100] Among them, u ij The jth PDW data belongs to the i-th cluster center Ci where m is the fuzzy index, usually an integer greater than 1.

[0101] S3, the method of using KNN to the remaining PDW data to construct the hyperedge of each node, and then get the hypergraph, usually using the association matrix H to represent the hypergraph structure, the association matrix H can be represented as:

[0102]

[0103] where v H represents the node in the hypergraph, e H represents the hyperedge in the hypergraph.

[0104] S4, learning the constructed hypergraph with the obtained cluster label pair of partial pulse sequence to get the final radar signal sorting result. The specific steps are as follows:

[0105] Step S41, constructing a node classification function F(v H ), taking two-class label prediction as an example, the node classification function F(v H ) can be represented as:

[0106]

[0107] where S and represent that the entire hypergraph is divided into two parts, one part is represented by S, and the remaining part is represented by , for details, refer to Figure 4 .

[0108] Step S42, constructing a regularizer Ω(F) on the hypergraph, through F(v H ), the regularizer Ω(F) can be represented as:

[0109]

[0110] where c(S) represents the hypergraph division, Δ is the normalized Laplacian matrix, E H represents the hyperedge set, e H represents a hyperedge, w(e H ) represents the weight of the hyperedge e H , δ(e H ) represents the hyperedge degree of e H , u H and v H represent two nodes on the hypergraph, F(u H ) represents the label of u H , d(u H ) represents the node degree of u H , F(v H ) represents the label of v H , d(vH represents the node degree of v H

[0111] Step S43, the objective function of the general hypergraph learning model can be divided into two parts, the first part is the empirical loss R emp (F) such as least square loss. The second part is the regularizer Ω(F) on the hypergraph, which represents the smoothness of the label distribution on the constructed hypergraph structure. Through linear combination, the minimization objective function of hypergraph learning can be represented as:

[0112]

[0113] wherein, represents the minimization objective function of hypergraph learning, and λ is the trade-off coefficient.

[0114] Step S44, the minimization objective function of hypergraph learning in step S43 can be further derived as:

[0115]

[0116] wherein, Y represents the initial label vector, for the labeled node v H , if the node belongs to the j-th class, the j-th element of Y(v H ) is 1, and the other positions are 0. For the unlabeled node u H , all elements of Y(u H ) are set to 0.5, which represents that this part of the unlabeled node has no prior class information.

[0117] When the hypergraph learning objective function reaches the minimum and converges, the node classification function can be represented as:

[0118]

[0119] wherein, D v represents the node degree matrix, D e represents the hyperedge degree matrix, and μ generally takes a value between 0 and 1.

[0120] Step S45, when the hypergraph learning reaches convergence, at this time the label message passing on the hypergraph reaches a balanced state, each node has corresponding label information, and the classification of nodes with different labels can complete the radar signal sorting.

[0121] ​Based on the above examples, experiments were conducted using datasets from two multifunction radars and two conventional radars to verify the effectiveness of sorting. In addition to real-world scenarios where MFRs switch between different operating modes in real time, Radiators 1 and 3 operated in only one mode, while Radiators 2 and 4 each operated in three modes. PW, RF, and PRI modulation types included jitter, stagger, and agility. Furthermore, in addition to testing the sorting effectiveness with varying numbers of interference pulses, the proposed algorithm also tested sorting performance with a uniform number of 50 interference pulses, resulting in a total of 2450 RIPS.

[0122] The present invention is used to remove the interference pulse in RIPS. The visualization result is shown in Figure 5 . Figure 5 In (a), the black five-pointed star represents the interference pulse, the small circle represents the normal pulse, PW represents the pulse width normalization, and RF represents the carrier frequency normalization. Similarly, in other figures, it can be seen that the distribution of interference data points is relatively sparse, while most other pulse sequences have data clustering and are distributed relatively densely. (b) shows the distribution of the remaining data points after removing the interference data points. It can be seen that after removing the data points with data potential energy values ​​less than the detection threshold value, there are very few interference data points in the remaining data. The interference pulses in RIPS are removed, and then the present invention is used to obtain the clustering labels of part of the PDW data for the remaining pulse data. The visualization results of the local core points and the clustering labels of some local core points obtained by screening can be found in Figure 6 shown. Figure 6 In (a), the inverted triangles represent the local core points obtained through screening. The distribution of local core points obtained through screening is characterized by relatively dense local core points in each part and relatively large distances between local core points in adjacent parts. (b) shows the distribution of local core points with cluster labels. The four gray diamonds represent the four cluster centers, and the four different shapes represent the local core point cluster labels for the four types of data. Figure 7 The ablation experiment of the present invention for each step is shown. (a), (b) and (c) are respectively without data field interference pulse removal, without screening rules and without hypergraph method. (d) is the sorting result obtained by using the present invention. Figure 7 In this example, predict-radar1 indicates the distribution of sorting results predicted for radar 1, and so on. The same applies to other figures. Figure 7 This shows the effectiveness and importance of each step in the method proposed in this invention. Figure 8 The comparison results of the proposed method with spectral clustering, FINCH clustering and complex network are shown in Figure 2. To ensure fairness, no interference pulses were added to the three comparison methods. Specifically, Figure 8In the figure, (a) is the sorting result obtained using the spectral clustering method, (b) is the sorting result obtained using the FINCH clustering, (c) is the sorting result obtained using the complex network method, and (d) is the true result distribution after interference pulses are removed using the data field method, wherein radar1 represents the data distribution of the real radar 1, and the like, and interferencepulse represents the remaining interference pulses. It can be seen that there are very few interference pulses in (d). Figure 8 It is shown that the method has better sorting effect than the other three methods when facing the sorting of multi-functional radar signals. Figure 9 It is shown that the performance of the sorting under different parameter estimation error ratios and different missing pulse rates, wherein Proportion represents the proportion of different error ratios or different missing pulse rates, Accuracy represents the sorting accuracy, Radar1 represents the sorting accuracy of radar 1 under different parameter estimation error ratios and different missing pulse rates, and the like. It can be seen from (a) that when the estimation error is less than or equal to 2%, the sorting accuracy of each radar is more than 86%, and the sorting accuracy of three radars is more than 96%. It can be seen from (b) that when the missing pulse rate is less than or equal to 25%, the sorting accuracy of each radar is more than 86%, which shows that the method has robustness. Therefore, the multi-functional radar sorting method based on the hypergraph has good sorting performance for sorting the MFRs signals, and lays a foundation for further processing such as recognition of the radar signals.

[0123] As shown in Figure 10 The embodiment discloses a multi-functional radar signal sorting system based on a hypergraph, which is used for executing the above method embodiment, and specifically includes the following modules:

[0124] The interference pulse removal module removes the interference pulses in the intercepted radar staggered pulse sequence RIPS. Specifically, the characteristic parameters of the RIPS are first normalized by maximum and minimum, then the data field potential values of all pulse sequences are calculated, it is judged whether the data field potential values of all pulse sequences are greater than a detection threshold, the pulses less than the detection threshold are interference pulses, and finally the interference pulses are removed.

[0125] The partial local core point label extraction module obtains the clustering labels of the remaining pulse sequences by using a screening rule and a fuzzy C-means (FCM) clustering. Specifically, the screening rule is used to screen the characteristic parameters of the remaining pulses to obtain local core points, and the clustering labels of the partial local core points are extracted as known labels in the subsequent sorting stage by combining the FCM clustering algorithm.

[0126] Supergraph construction module: the feature parameters of the remaining pulse sequence and the data potential energy value thereof are constructed into a supergraph; specifically, the KNN method is used to construct a supergraph for the feature parameters and the data potential energy value of the remaining pulse;

[0127] Supergraph learning module: the obtained clustering label pair of the remaining pulse sequence is used to learn the constructed supergraph; specifically, the clustering label of the partial local core point obtained in the partial local core point label extraction module is used as known information, and then the remaining unmarked nodes are learned by first learning the partially marked node data on the supergraph;

[0128] Sorting module: when the objective function in the supergraph learning module reaches the minimum and converges, the label information of each node can be obtained, the nodes are classified, and the radar signal sorting result can be obtained.

[0129] Other contents of the embodiment can refer to the above method embodiments.

[0130] The above is only the preferred embodiment of the present application and the applied technical principle. Those skilled in the art will understand that the present application is not limited to the specific embodiments described herein, and various obvious changes, re-adjustments and substitutions can be made without departing from the scope of the present application. Therefore, although the present application has been described in detail through the above embodiments, the present application is not limited to the above embodiments, and more other equivalent embodiments can be included without departing from the concept of the present application, and the scope of the present application is determined by the scope of the appended claims.

Claims

1. A multifunctional radar signal sorting method based on hypergraph, characterized by: The specific steps include: S1. Eliminate the interference pulses in the intercepted radar interleaved pulse sequence RIPS; S2, clustering the remaining pulse sequences using screening rules and fuzzy C-means FCM to obtain cluster labels of the remaining pulse sequences; S3, constructing a hypergraph based on the characteristic parameters of the remaining pulse sequences and their data potential energy values; S4. The cluster labels of the remaining pulse sequences are used to learn the constructed hypergraph to obtain the final radar signal sorting result.

2. The multifunctional radar signal sorting method based on hypergraph according to claim 1, characterized in that: In step S1, the interference pulses in the RIPS are eliminated using the data field method, as follows: S11. Perform maximum and minimum normalization on the three characteristic parameters of carrier frequency RF, pulse width PW and direction of arrival DOA in the pulse description word PDW data of RIPS: Among them, x d represents the d-dimensional feature parameter vector, represents the dth feature of the i-th pulse; S12. Calculate the data potential energy value of each PDW data. The calculated data potential energy value is expressed as: Among them, pdw m Indicates the mth PDW, d(pdw i ,pdw j ) indicates pdw i with pdw j The Euclidean distance between them, σ is the radiation factor, N is the number of pulses, F(pdw m ) indicates pdw m Data potential energy value; S13, calculating the detection threshold of the data potential energy value, identifying the data points whose data potential energy value is less than the set average data potential energy value as interference pulses, and eliminating them, which is specifically expressed as: in, is the detection threshold, Interference pulse The data potential energy value; ξ is the detection coefficient set according to the electromagnetic environment, mean represents the average value, represents the sum of the data potential energy values ​​of all pulses, N is the number of pulses, F(pdw i ) is the data potential energy value of the i-th PDW.

3. The multifunctional radar signal sorting method based on hypergraph according to claim 2, characterized in that: Step S2 is specifically as follows: S21, filtering the remaining pulse sequences according to a filtering rule to obtain a local core point, wherein the filtering rule is expressed as: Among them, μ(δ) represents the relative distance mean of all data, μ(ρ) represents the local density mean of all data, σ(ρ) represents the local density standard deviation of all data, μ(|RNN|) represents the number of reverse nearest neighbors of all data, μ and σ represent the mean and standard deviation respectively, local_core_pdw i represents i local core points, δ i represents the relative distance of the i-th data, δ represents the relative distance of all data, ρ i represents the local density of the i-th data, ρ represents the local density of all data, |RNN i | represents the number of reverse nearest neighbors of the i-th data, and |RNN| represents the number of reverse nearest neighbors of all data; S22. Estimate the number of multi-function radar MFRs P. P is determined based on the inflection point of the intra-cluster sum of squares CSS and the maximum value of the silhouette coefficient SC. CSS and SC are expressed as: Among them, c represents the number of cluster centers, n represents the number of remaining pulses, and d(C i ,pdw j ) represents the i-th cluster center C i With the j-th PDW data pdw j the distance between them; Among them, a(pdw i ) is pdw i The average distance to other PDW data points in the same cluster, b(pdw i ) is pdw i The average distance to all PDW data points in its nearest neighbor cluster; S23, perform FCM clustering on the remaining pulse sequences according to the MFR number P obtained in step S22, and obtain the membership value u of each pulse sequence ij , the membership value of the local core point obtained by traversal, when a certain membership value of a local core point is greater than the set value, the cluster label of the local core point is extracted and the membership value u is calculated ij Expressed as: Among them, u ij The jth PDW data belongs to the i-th cluster center C i The membership value of is , and m is the fuzzy index.

4. The multifunctional radar signal sorting method based on hypergraph according to claim 3 is characterized in that: In step S3, the K-nearest neighbor (KNN) method is used to construct the hyperedge of each node for the remaining pulse sequences to obtain a hypergraph.

5. The multifunctional radar signal sorting method based on hypergraph according to claim 4, characterized in that: In step S3, the hypergraph structure is represented by the association matrix H, which is expressed as: Among them, v H Represents a node in the hypergraph, e H Represents a hyperedge in the hypergraph.

6. The multifunctional radar signal sorting method based on hypergraph according to claim 5, characterized in that: Step S4 is specifically as follows: S41, construct node classification function F(v H ), taking two types of label prediction as an example, the node classification function F(v H ) is expressed as: Among them, S and Indicates that the entire hypergraph is divided into two parts, one part is represented by S and the remaining part is represented by express; S42, construct a regularizer Ω(F) on the hypergraph, through F(v H ) represents the regularizer Ω(F) as: Where c(S) represents the hypergraph partition, Δ is the normalized Laplace matrix, and E H represents the hyperedge set, e H represents a hyperedge, w(e H ) represents the hyperedge e H The weight of δ(e H ) means e H The hyperedge degree, u H and v H Represents two nodes on the hypergraph, F(u H ) represents u H The label, d(u H ) represents u H The node degree, F(v H ) indicates v H The label, d(v H ) indicates v H Node degree; S43. The objective function in the hypergraph learning model is divided into two parts. The first part is the empirical loss R of the learning label. emp (F), the second part is the regularizer Ω(F) on the hypergraph, which represents the smoothness of the label distribution on the constructed hypergraph structure; through linear combination, the minimization objective function of hypergraph learning is expressed as: in, represents the minimization objective function of hypergraph learning, and λ is the trade-off coefficient; S44. The hypergraph learning minimization objective function in step S43 is further derived and expressed as: Among them, Y represents the initial label vector, for the labeled node v H , if the node belongs to the jth class, then Y(v H ) is 1, and the other positions are 0; for the unmarked node u H ,Y(u H ) are all set to 0.5, indicating that there is no prior category information for these unlabeled nodes; When the hypergraph learning objective function reaches the minimum and converges, the node classification function is expressed as: Among them, D v Denotes the node degree matrix, D e represents the hyperedge degree matrix, μ takes values ​​between 0 and 1; S45. When hypergraph learning reaches convergence, the label message transmission on the hypergraph reaches a balanced state, each node has corresponding label information, and the radar signal sorting is completed by classifying the nodes with different labels.

7. A multifunctional radar signal sorting system based on a hypergraph, for executing the method according to any one of claims 1 to 6, characterized in that Includes the following modules: Interference pulse removal module: removes interference pulses in the intercepted radar interleaved pulse sequence RIPS; Partial local core point label extraction module: The remaining pulse sequence is clustered using screening rules and fuzzy C-means FCM to obtain the cluster label of the remaining pulse sequence; Hypergraph construction module: constructs a hypergraph based on the characteristic parameters of the remaining pulse sequences and their data potential energy values; Hypergraph learning module: Use the cluster labels of the remaining pulse sequences to learn the constructed hypergraph; Sorting module: When the hypergraph learning in the hypergraph learning module reaches convergence, the label information of each node is obtained, and each node is classified to obtain the radar signal sorting results.

Images