Multi-station radar signal sorting method based on graph learning

Abstract

The invention discloses a multi-station radar signal sorting method based on graph learning, and the method comprises the steps: carrying out the pulse data collection of a known type of target signals through two reconnaissance stations, and constructing a corresponding pulse sorting data set based on a pulse description word; calculating a time difference window based on the position information of the two investigation stations; specifying a master station and a slave station, and determining arrival time adjacent matrixes corresponding to the master station and the slave station; on the basis of the pulse sorting data sets corresponding to the master station and the slave station, selecting pulse description words again, constructing network training data sets of the master station and the slave station, preprocessing the network training data sets, determining a corresponding pulse adjacency matrix, and normalizing the pulse adjacency matrix; constructing a multi-station intelligent sorting model; taking a corresponding group of pulses in the network training data set as samples, and training the multi-station intelligent sorting model; and utilizing the trained multi-station intelligent sorting model to sort the pulse data of the unknown category acquired by the two investigation stations.

Description

Technical Field

[0001] The present invention relates to fields such as electronic reconnaissance and situational awareness, and in particular to a multi-station radar signal sorting method based on graph learning. Background Art

[0002] Radar signal sorting is one of the key technologies for electronic reconnaissance and battlefield situational awareness, and its accuracy directly affects subsequent decision-making and planning. In a complex electromagnetic environment, after intercepting the interwoven radar pulse streams composed of different radar radiation sources, the ability to accurately sort out different radar signals is a prerequisite for subsequent radiation source identification, situational awareness analysis, and other processing.

[0003] Electronic reconnaissance systems can be categorized as single-station or multi-station systems, depending on the number of reconnaissance stations. Single-station systems are subject to external factors such as line-of-sight limitations, target obstruction, and multipath effects, making them unable to fully cover the monitoring range and effectively address issues such as pulse signal distortion and interference in complex electromagnetic environments. Multi-station systems, however, leverage the spatial diversity of their equipment to enhance sorting efficiency through collaborative observation and data sharing.

[0004] Existing multi-station sorting methods primarily utilize the time difference of arrival (TDOA) parameters introduced by the multi-station system, introducing the time difference information of the same pulse arriving at multiple stations, and implementing the multi-station pulse sorting process through pulse matching. Typical methods include time difference histograms, extended time difference histograms, and TDOA clustering algorithms. The most widely studied sorting method based on time difference histograms is effective for low-repetition-rate pulses, but is limited by the fuzzy characteristics of time difference and has difficulty in effectively sorting high-repetition-rate pulses. Furthermore, this type of method is limited by the hyperparameters of histogram construction, making it difficult to achieve effective pulse sorting in complex electromagnetic environments. The research approach for TDOA clustering is similar to that of single-station clustering. Most clustering algorithms require iteration or search, resulting in high time complexity and difficulty in practical sorting processes. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-station radar signal sorting method based on graph learning, which breaks through the limitations of traditional methods and solves the problem of multi-station sorting in complex electromagnetic environments.

[0006] In order to achieve the above tasks, the present invention adopts the following technical solutions:

[0007] A multi-station radar signal sorting method based on graph learning, comprising:

[0008] Two reconnaissance stations are used to collect pulse data of target signals of known categories, and corresponding pulse sorting datasets are constructed based on pulse description words.

[0009] The time difference window is calculated based on the location information of the two reconnaissance stations; the master station and the slave station are designated, and the arrival time adjacency matrix corresponding to the master station and the slave station is obtained through the time difference window matching process using the time difference window;

[0010] Based on the pulse sorting data sets corresponding to the master station and the slave station, the pulse description words are reselected to construct the network training data sets of the master station and the slave station, the network training data sets are preprocessed, and the corresponding pulse adjacency matrix is ​​determined and normalized.

[0011] Construct a multi-station intelligent sorting model; use a corresponding set of pulses in the network training data set as samples to train the multi-station intelligent sorting model;

[0012] The pulse data of unknown categories collected by the two reconnaissance stations are sorted using the trained multi-station intelligent sorting model to obtain the pulse sorting results.

[0013] Furthermore, the pulse description word includes the arrival time of each pulse, as well as the corresponding carrier frequency, pulse width and pulse amplitude; the target signal is a pulse signal generated by the target radiation source.

[0014] Furthermore, the arrival time adjacency matrix corresponding to the master station and the slave station is obtained by using the time difference window through the time difference window matching process, including:

[0015] For the pulse sorting data set D0 of the master station, the time difference pairing range of the arrival time TOA0 of the first pulse is [TOA0-T diff ,TOA0+T diff ]; where the time difference window is [-T diff ,T diff ], where T diff represents the time difference calculated from the positions of two reconnaissance stations;

[0016] Determine the arrival time TOA of all pulses in the pulse sorting data set of the slave station; if the TOA of a pulse of the slave station falls within the time difference pairing range, a connection is established between the arrival time TOA0 of the first pulse of the master station and the arrival time TOA of the pulse of the slave station, and the adjacency matrix A constructed at the master station is g0 The value of the corresponding element in is 1, otherwise the value of the corresponding element is 0; the same method is used to calculate each pulse in the pulse sorting data set D0 of the master station in turn, so as to obtain the arrival time adjacency matrix A of the master station g0 ;

[0017] Swap the roles of the master station and the slave station, and determine the arrival time adjacency matrix A corresponding to the slave station according to the method of constructing the arrival time adjacency matrix of the master station. g1 .

[0018] Furthermore, the number of pulses in the pulse sorting data sets D0 and D1 of the master and slave stations is recorded as M and N, where M>N;

[0019] The arrival time adjacency matrix A of the limited master station g0 The size is M×M, and the arrival time adjacency matrix A g0 The rows in correspond to the number of pulses M in D0, and the columns correspond to the number of pulses N in D1; zeros are added to the part with less than M columns;

[0020] The adjacency matrix A of the arrival time of the slave station g1 The rows correspond to the number of pulses N in D1, and the columns correspond to the number of pulses M in D0. Since M>N, the redundant part is cut off to obtain the TOA adjacency matrix A of size N×N. g1 .

[0021] Furthermore, based on the pulse sorting data sets corresponding to the master station and the slave station, the pulse description words are reselected to construct the network training data sets of the master station and the slave station, and the network training data sets are preprocessed; including:

[0022] Based on the pulse sorting data sets D0 and D1 of the master station and the slave station, the carrier frequency CF, pulse width PW and pulse amplitude PA of each pulse in D0 and D1 are extracted to form the network training data sets D0' and D1' of the master station and the slave station; the carrier frequency CF, pulse width PW and pulse amplitude PA of all pulses in the network training data sets D0' and D1' are normalized respectively to obtain the normalized network training data sets and The normalization process is maximum and minimum normalization.

[0023] Furthermore, the corresponding impulse adjacency matrix is ​​determined, including:

[0024] Training the network with a dataset and Use KNN algorithm to perform clustering operations respectively to obtain network training data sets and The corresponding impulse adjacency matrices A0 and A1;

[0025] For the network training data set For a pulse in , among its K nearest neighbor pulses, the pulse description word is used to calculate the Euclidean distance between the pulse and each neighbor pulse, and the neighbor pulse with the smallest Euclidean distance is selected. It is considered that there is a connection relationship between the pulse and the neighbor pulse, and the element in the corresponding pulse adjacency matrix A0 is 1; the same operation is performed on each pulse to obtain the complete pulse adjacency matrix A0; the same method can be used to obtain the network training data set of the slave station The corresponding impulse adjacency matrix A1.

[0026] Furthermore, the value of K in the KNN algorithm is based on the network training data set. It is determined by the maximum and minimum repetition frequencies of all pulses in the system: if the maximum repetition frequency is n times the minimum repetition frequency, then the value of K is 2 to 3 times n.

[0027] Furthermore, the formula for normalizing the pulse adjacency matrix A0 of the master station is:

[0028] A0'=D0 -1 / 2 (A0+I)D0 -1 / 2

[0029] Where A0' is the normalized impulse adjacency matrix, D0 is the degree matrix of the impulse adjacency matrix A0, and I is the identity matrix.

[0030] Furthermore, the multi-station intelligent sorting model includes two parallel branches, each branch includes a fully connected layer, an activation function layer, a normalization layer, a graph attention network and a cross-attention mechanism layer connected in sequence; finally, the results of the two branches are output to the fusion layer, and a fully connected layer and a classification function are set after the fusion layer.

[0031] Furthermore, the process of training the multi-station intelligent sorting model using the model training dataset is as follows:

[0032] Network training data sets for master and slave stations and The corresponding set of pulses in the two branches are input respectively; in the two branches, the network training data sets of the master station and the slave station and The pulses in the network are firstly embedded through the fully connected layer to mine high-dimensional features, and the feature embedding representation is obtained and then processed by the activation function layer and the normalization layer. The feature embedding representation of the master station and the corresponding pulse adjacency matrix A'0, and the feature embedding representation of the slave station and the corresponding pulse adjacency matrix A1' are respectively input into the graph attention network of their respective branches to obtain the graph feature matrix of their respective branches.

[0033] After the features of the two branches are embedded and spliced, the TOA adjacency matrix A corresponding to the master station and the slave station is aggregated at the graph aggregation layer. g0 、A g1 , get the global adjacency matrix A g ;

[0034] The graph feature matrix of each branch and the global connection matrix A gThey enter the cross-attention mechanism layer of the branch together. The output features of the cross-attention mechanism layers of the two branches are spliced ​​and fused through the gated fusion mechanism in the fusion layer. After the features after the fusion layer enter the fully connected layer for processing, the corresponding sorting results are obtained through the classification function.

[0035] Select the optimizer and loss function to train the multi-station intelligent sorting model. After the loss function converges or reaches the preset number of training rounds, save the trained multi-station intelligent sorting model.

[0036] A terminal device includes a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, the multi-station radar signal sorting method based on graph learning is implemented.

[0037] A computer-readable storage medium stores a computer program; when the computer program is executed by a processor, the multi-station radar signal sorting method based on graph learning is implemented.

[0038] Compared with the prior art, the present invention has the following technical features:

[0039] In order to solve the problems of insufficient fusion of multi-station features, difficulty in sorting complex modulated pulses such as high repetition rate, and low overall sorting accuracy in existing histogram-based methods and algorithms in multi-station scenarios, the present invention proposes a multi-station intelligent sorting method based on graph learning and a multi-station sorting framework based on graph processing. It utilizes the multi-head attention mechanism and aggregation mechanism to effectively integrate multi-station pulse data, give full play to the information gain advantage brought by multiple stations, and improve the sorting performance of complex modulated pulses. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Flowchart of the method of the present invention;

[0041] Figure 2 Schematic diagram of the model framework of the method of the present invention;

[0042] Figure 3 Schematic diagram of the distribution of simulation data carrier frequency parameter sequence;

[0043] Figure 4 Schematic diagram of the distribution of pulse width parameter sequence of simulation data;

[0044] Figure 5 Schematic diagram of the distribution of pulse amplitude parameter sequence of simulation data;

[0045] Figure 6 Schematic diagram of loss function iteration;

[0046] Figure 7 Schematic diagram of various accuracy iterations of training data;

[0047] Figure 8Schematic diagram of various accuracy iterations of test data;

[0048] Figure 9 Schematic diagram of training and testing average accuracy iterations. DETAILED DESCRIPTION

[0049] Based on graph neural networks, this paper deeply analyzes the characteristics of various parameters of pulse description words (PDWs) and constructs a framework for multi-station intelligent sorting. It utilizes graph attention and aggregation mechanisms to fuse multi-station data, breaking through the limitations of single-station processing and improving the sorting performance of complex modulated pulses in multi-station scenarios.

[0050] The basic idea behind this invention is to use the electronic reconnaissance system of dual reconnaissance stations to collect pulse data of target signals and output PDW information; use the KNN method to construct the pulse connection relationship of single-station pulses, namely the pulse adjacency matrix, calculate the time difference window between stations based on the location of the reconnaissance stations, and further use the time difference matching method to obtain the connection relationship between stations, namely the TOA adjacency matrix; standardize the pulse data set and input it into the multi-station intelligent sorting model for learning iteration to realize the pulse classification process; the specific instructions are as follows:

[0051] Step 1: Use two reconnaissance stations to collect pulse data of target signals of known categories. Extract pulse descriptors from the pulse data collected by each reconnaissance station to construct pulse sorting datasets D0 and D1. The extracted pulse descriptors include the arrival time (TOA) of each pulse, as well as the corresponding carrier frequency (CF), pulse width (PW), and pulse amplitude (PA). The pulse descriptors of all pulses constitute the pulse sorting dataset.

[0052] The target signal is a pulse signal generated by a target radiation source; one of the two reconnaissance stations is designated as a master station and the other as a slave station.

[0053] Step 2: Calculate the time difference window based on the location information of the two reconnaissance stations; specify the master station and the slave station, and obtain the arrival time adjacency matrix A corresponding to the master station and the slave station through the time difference window matching process. g0 and A g1 .

[0054] The positions of the two reconnaissance stations are (x i ,y i ,z i ), where i = 0, 1 represents the master station and the slave station in the two reconnaissance stations respectively; according to the position difference between the two reconnaissance stations, the time difference range of the pulse data arriving at the two reconnaissance stations can be obtained, that is, the time difference window [-Tdiff ,T diff ], where T diff represents the time difference, and c is the speed of light:

[0055]

[0056] Taking the arrival time TOA of the pulse as the node, for the pulse sorting data set D0 of the master station, the time difference matching range of the arrival time TOA0 of the first pulse is [TOA0-T diff ,TOA0+T diff ]; determine the arrival time TOA of all pulses in the pulse sorting data set of the slave station; if the TOA of a pulse of the slave station falls within the time difference pairing range, a connection relationship is established between the arrival time TOA0 of the first pulse of the master station and the arrival time TOA of the pulse of the slave station, that is, there is an edge between the two nodes, then the adjacency matrix A constructed in the master station is g0 The value of the corresponding element in is 1, otherwise the value of the corresponding element is 0; using the same method, the time difference pairing range of the arrival time of each pulse in the pulse sorting data set D0 of the master station is calculated in turn, and whether the arrival time of all pulses of the slave station falls into the time difference pairing range is determined to determine whether there is a connection relationship between the corresponding pulses of the master station and the slave station, thereby constructing the TOA adjacency matrix A of the master station g0 .

[0057] The number of pulses in the pulse sorting data sets D0 and D1 of the master station and the slave station is recorded as M and N, where M>N; the TOA adjacency matrix A of the master station is limited g0 The size is M×M, that is, the TOA adjacency matrix A g0 The rows in correspond to the number of pulses M in D0, and the columns correspond to the number of pulses N in D1; and in order to meet the formal requirements of the adjacency matrix, the part with less than M columns is filled with zeros, thus forming an M×M TOA adjacency matrix A g0 .

[0058] Construct the TOA adjacency matrix A of the slave station g1 Specifically, the roles of the master station and the slave station are exchanged and the same method as above is used to obtain the TOA adjacency matrix A constructed when the slave station is used as the master station. g1 , size is N×N; TOA adjacency matrix A g1 The rows correspond to the number of pulses N in D1, and the columns correspond to the number of pulses M in D0. However, since M>N, the redundant part needs to be cut off to obtain the TOA adjacency matrix A of size N×N. g1 .

[0059] Step 3: Based on the pulse sorting data sets D0 and D1 of the master station and the slave station, the carrier frequency CF, pulse width PW and pulse amplitude PA of each pulse in D0 and D1 are extracted to form the network training data sets D0' and D1' of the master station and the slave station; the carrier frequency CF, pulse width PW and pulse amplitude PA of all pulses in the network training data sets D0' and D1' are normalized respectively to obtain the normalized network training data sets. and The normalization process is maximum and minimum normalization; that is, each type of parameter in each network training data set is normalized separately; standardization eliminates the impact of different dimensions and is a key step in intelligent algorithm processing.

[0060] Training the network with a dataset and Use KNN algorithm to perform clustering operations respectively to obtain network training data sets and The corresponding impulse adjacency matrices A0 and A1.

[0061] Among them, when the KNN algorithm is clustered, for example, for the network training data set For a pulse in the network, among its K nearest neighbor pulses, the Euclidean distance between the pulse and each neighbor pulse is calculated using the carrier frequency CF, pulse width PW and pulse amplitude PA. The neighbor pulse with the smallest Euclidean distance is selected and it is considered that there is a connection between the pulse and the neighbor pulse. The corresponding element in the pulse adjacency matrix A0 is 1. The same operation is performed on each pulse to obtain the complete pulse adjacency matrix A0. The rows and columns of the pulse adjacency matrix A0 correspond to the number of pulses and the size is N×N. The key to the KNN algorithm is the selection of the K value and the distance metric. According to the network training data set It is determined by the maximum and minimum repetition frequencies of all pulses in the system; for example, if the maximum repetition frequency is n times the minimum repetition frequency, then the value of K is 2 to 3 times n.

[0062] Using the same method, the network training data set of the slave station can be obtained The corresponding impulse adjacency matrix A1.

[0063] Step 4: Perform matrix normalization on the pulse adjacency matrices A0 and A1 corresponding to the master station and the slave station respectively using the corresponding degree matrices to obtain normalized pulse adjacency matrices A'0 and A1'; the calculation formula of A'0 is as follows:

[0064] A0'=D0 -1 / 2 (A0+I)D0 -1 / 2

[0065] Where D0 is the degree matrix of the impulse adjacency matrix A0; I is the identity matrix, which enables the model to learn its own features. The calculation formula for A1' is the same as above.

[0066] Step 5: Construct a model training dataset; a sample in the model training dataset is generated by the network training dataset. and It is composed of a group of pulses corresponding to the , and sets a label for the sample, that is, which radiation source the pulse comes from.

[0067] Step 6: Build a multi-station intelligent sorting model, which includes two parallel branches. Each branch includes a fully connected layer, an activation function layer, a normalization layer, a graph attention network, and a cross-attention mechanism layer connected in sequence. Finally, the results of the two branches are output to the fusion layer, and a fully connected layer and a classification function are set after the fusion layer.

[0068] The process of training the multi-station intelligent sorting model using the model training dataset is as follows:

[0069] Network training data sets for master and slave stations and The corresponding set of pulses in the two branches are input into the two branches respectively; in the two branches, the network training data sets of the master station and the slave station and The pulses in the network are firstly subjected to feature embedding through the fully connected layer to mine high-dimensional features, and the feature embedding representation is obtained and then processed by the activation function layer and the normalization layer. The feature embedding representation of the master station and the corresponding pulse adjacency matrix A'0, and the feature embedding representation of the slave station and the corresponding pulse adjacency matrix A1' are respectively input into the graph attention network of their respective branches to obtain the graph feature matrix of their respective branches. In this step, the attention weights between pulses are calculated by combining the pulse adjacency matrix through the attention network to capture the importance differences of different pulse relationships.

[0070] After the features of the two branches are embedded and spliced, the TOA adjacency matrix A corresponding to the master station and the slave station is aggregated at the graph aggregation layer. g0 、A g1 , and obtain a global adjacency matrix A of size (M+N)×(M+N) g .

[0071] The graph feature matrix of each branch and the global connection matrix A g They enter the cross-attention mechanism layer of this branch together, and the cross-attention mechanism is introduced in this layer for matching, effectively integrating the outputs of the graph attention network and the graph aggregation layer, thereby establishing lateral connections between multi-layer architectures and enhancing the classification ability of the model.

[0072] The output features of the last two branches across the attention mechanism layers are concatenated and fused in the fusion layer, and a gated fusion mechanism is further introduced to automatically determine the fusion ratio through learnable gating weights to achieve more intelligent feature combination. After the features after the fusion layer enter the fully connected layer for processing, the corresponding sorting results are obtained through the classification function, that is, the probability that the pulse comes from different radiation sources.

[0073] Select the Adam optimizer and cross entropy loss function to train the multi-station intelligent sorting model. After the loss function converges or reaches the preset number of training rounds, save the trained multi-station intelligent sorting model.

[0074] Step 7: In actual application, two reconnaissance stations are used to collect pulse data of target signals of unknown categories, and the TOA adjacency matrix and pulse adjacency matrix of the two reconnaissance stations are constructed respectively. The corresponding pulses in the two reconnaissance stations are extracted, and the pulse carrier frequency CF, pulse width PW and pulse amplitude PA parameters as well as the TOA adjacency matrix and pulse adjacency matrix are input into the trained multi-station intelligent sorting model to obtain the sorting results of the pulses.

[0075] Example:

[0076] In one embodiment of the present invention, the simulation process is close to the complex pulse modulation mode in the actual environment. A total of five types of radar radiation sources are simulated, and 10% of stray pulses are added. The simulation data is shown in Table 1:

[0077] Table 1 Simulation data parameter table

[0078]

[0079] The distribution diagram of the pulse CF, PW and PA parameters received by station 0 is as shown above Figure 3 、 Figure 4 and Figure 5 As shown, it can be seen that each pulse parameter has complex modulation and the parameters overlap seriously.

[0080] The method of the present invention is used to perform the following treatments:

[0081] The four parameter sequences of arrival time (TOA), carrier frequency (CF), pulse width (PW) and pulse amplitude (PA) of the received pulse sequences of the two stations (station 0 and station 1) within the simulated 0.2s range are extracted to form pulse sorting data sets D0 and D1. The number of pulses at the two stations are 11746 and 12385 respectively.

[0082] The two stations are located at (10e3, 0, 0)m and (0, 10e3, 0)m respectively, and the time difference window range between the stations is calculated to be [-47us, 47us]. Station 0 is selected as the master station, and the master station pulses are traversed for time difference matching to obtain the arrival time adjacency matrix A. g0 Similarly, station 1 is selected as the master station, and the time difference matching is performed on the master station pulses to obtain the arrival time adjacency matrix A g1 , sizes are 11746×11746 and 12385×12385 respectively.

[0083] After KNN algorithm processing, the impulse adjacency matrices A0 and A1 of station 0 and station 1 are obtained, with sizes of 11746×11746 and 12385×12385 respectively; after normalization, they are impulse adjacency matrices A'0 and A1'.

[0084] The dataset was divided into training and test sets in a 6:4 ratio and labeled. The labels of radars 0 to 4 were set to 0 to 4, and the label of stray pulses was 5. The multi-station intelligent sorting model was trained with 1500 training rounds, Adma optimizer, 0.005 learning rate, and cross entropy loss as the loss function. The loss was calculated using the training set labels.

[0085] Table 2 Division statistics

[0086]

[0087] The test set is used to evaluate the model, and the sorting effect is as follows Figure 6 、 Figure 7 、 Figure 8 and Figure 9 As shown in Figure 2, the model has excellent sorting performance, with an average accuracy of over 97%. Figure 6 It can be seen that the model reaches a convergence state around 1200 rounds, and there is no large fluctuation in the whole process, and no abnormal decline or rise process. Figure 7 、 Figure 8 and Figure 9 The model demonstrates excellent sorting performance for high-repetition-rate radiation sources, which are difficult to handle with traditional histogram methods. This effectively addresses the problem of large amounts of fuzzy time differences that traditional histogram methods struggle to handle. Even with complex parameter modulations such as agility, jitter, jitter, and slip, the model's sorting accuracy remains around 95%.

[0088] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A multi-station radar signal sorting method based on graph learning, characterized in that: include: Two reconnaissance stations are used to collect pulse data of target signals of known categories, and corresponding pulse sorting datasets are constructed based on pulse description words. Calculate the time difference window based on the position information of the two reconnaissance stations; Specify the master station and the slave station, and use the time difference window to obtain the arrival time adjacency matrix corresponding to the master station and the slave station through the time difference window matching process; Based on the pulse sorting data sets corresponding to the master station and the slave station, the pulse description words are reselected to construct the network training data sets of the master station and the slave station, the network training data sets are preprocessed, and the corresponding pulse adjacency matrix is ​​determined and normalized. Build a multi-station intelligent sorting model; The corresponding set of pulses in the network training data set is used as samples to train the multi-station intelligent sorting model; The pulse data of unknown categories collected by the two reconnaissance stations are sorted using the trained multi-station intelligent sorting model to obtain the pulse sorting results.

2. The multi-station radar signal sorting method based on graph learning according to claim 1 is characterized in that: The pulse description word includes the arrival time of each pulse, as well as the corresponding carrier frequency, pulse width and pulse amplitude; the target signal is the pulse signal generated by the target radiation source.

3. The multi-station radar signal sorting method based on graph learning according to claim 1 is characterized in that: The arrival time adjacency matrix corresponding to the master station and the slave station is obtained by using the time difference window matching process, including: For the pulse sorting data set D0 of the master station, the time difference pairing range of the arrival time TOA0 of the first pulse is [TOA0-T diff ,TOA0+T diff ]; where the time difference window is [-T diff ,T diff ], where T diff represents the time difference calculated from the positions of two reconnaissance stations; Determine the arrival time TOA of all pulses in the pulse sorting data set of the slave station; if the TOA of a pulse of the slave station falls within the time difference pairing range, a connection is established between the arrival time TOA0 of the first pulse of the master station and the arrival time TOA of the pulse of the slave station, and the adjacency matrix A constructed at the master station is g0 The value of the corresponding element in is 1, otherwise the value of the corresponding element is 0; the same method is used to calculate each pulse in the pulse sorting data set D0 of the master station in turn, so as to obtain the arrival time adjacency matrix A of the master station g0 ; Swap the roles of the master station and the slave station, and determine the arrival time adjacency matrix A corresponding to the slave station according to the method of constructing the arrival time adjacency matrix of the master station. g1 .

4. The multi-station radar signal sorting method based on graph learning according to claim 1 is characterized in that: The number of pulses in the pulse sorting data sets D0 and D1 of the master station and the slave station is recorded as M and N, where M>N; The arrival time adjacency matrix A of the limited master station g0 The size is M×M, and the arrival time adjacency matrix A g0 The rows in correspond to the number of pulses M in D0, and the columns correspond to the number of pulses N in D1; zeros are added to the part with less than M columns; The adjacency matrix A of the arrival time of the slave station g1 The rows correspond to the number of pulses N in D1, and the columns correspond to the number of pulses M in D0. Since M>N, the redundant part is cut off to obtain the TOA adjacency matrix A of size N×N. g1 .

5. The multi-station radar signal sorting method based on graph learning according to claim 1 is characterized in that: Based on the pulse sorting data sets corresponding to the master station and the slave station, the pulse description words are reselected to construct the network training data sets of the master station and the slave station, and the network training data sets are preprocessed; include: Based on the pulse sorting data sets D0 and D1 of the master station and the slave station, the carrier frequency CF, pulse width PW and pulse amplitude PA of each pulse in D0 and D1 are extracted to form the network training data sets D0' and D1' of the master station and the slave station; the carrier frequency CF, pulse width PW and pulse amplitude PA of all pulses in the network training data sets D0' and D1' are normalized respectively to obtain the normalized network training data sets and The normalization process is maximum and minimum normalization.

6. The multi-station radar signal sorting method based on graph learning according to claim 1 is characterized in that: Determine the corresponding impulse adjacency matrix, including: Training the network with a dataset and Use KNN algorithm to perform clustering operations respectively to obtain network training data sets and The corresponding impulse adjacency matrices A0 and A1; For the network training data set For a pulse in , among its K nearest neighbor pulses, the pulse description word is used to calculate the Euclidean distance between the pulse and each neighbor pulse, and the neighbor pulse with the smallest Euclidean distance is selected. It is considered that there is a connection relationship between the pulse and the neighbor pulse, and the element in the corresponding pulse adjacency matrix A0 is 1; the same operation is performed on each pulse to obtain the complete pulse adjacency matrix A0; the same method can be used to obtain the network training data set of the slave station The corresponding impulse adjacency matrix A1.

7. The multi-station radar signal sorting method based on graph learning according to claim 1 is characterized in that: The multi-station intelligent sorting model includes two parallel branches, each branch includes a fully connected layer, an activation function layer, a normalization layer, a graph attention network and a cross-attention mechanism layer connected in sequence; finally, the results of the two branches are output to the fusion layer, and a fully connected layer and a classification function are set after the fusion layer.

8. The multi-station radar signal sorting method based on graph learning according to claim 7 is characterized in that: The process of training the multi-station intelligent sorting model using the model training dataset is as follows: Network training data sets for master and slave stations and The corresponding set of pulses are input into the two branches respectively; In both branches, the network training datasets of the master and slave stations and The pulses in the network are firstly embedded through the fully connected layer to mine high-dimensional features, and the feature embedding representation is obtained and then processed by the activation function layer and the normalization layer. The feature embedding representation of the master station and the corresponding pulse adjacency matrix A'0, and the feature embedding representation of the slave station and the corresponding pulse adjacency matrix A1' are respectively input into the graph attention network of their respective branches to obtain the graph feature matrix of their respective branches. After the features of the two branches are embedded and spliced, the TOA adjacency matrix A corresponding to the master station and the slave station is aggregated at the graph aggregation layer. g0 、A g1 , get the global adjacency matrix A g ; The graph feature matrix of each branch and the global connection matrix A g They enter the cross-attention mechanism layer of the branch together, and the output features of the cross-attention mechanism layers of the two branches are spliced ​​and fused through the gated fusion mechanism in the fusion layer; After the features after the fusion layer enter the fully connected layer for processing, the corresponding sorting results are obtained through the classification function; Select the optimizer and loss function to train the multi-station intelligent sorting model. After the loss function converges or reaches the preset number of training rounds, save the trained multi-station intelligent sorting model.

9. A terminal device comprising a processor, a memory, and a computer program stored in the memory; characterized in that: When the processor executes the computer program, it implements the multi-station radar signal sorting method based on graph learning according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program; wherein: When the computer program is executed by a processor, the multi-station radar signal sorting method based on graph learning according to any one of claims 1 to 8 is implemented.

Images