A radar signal intelligent sorting method based on a GraphGated graph convolutional neural network

Abstract

The application relates to a radar signal intelligent sorting method based on a GraphGated graph convolutional neural network and belongs to the field of information and communication engineering.The application aims to solve the problem of signal sorting difficulty of unknown signal sensing in the face of high pulse density environment, complex inter-pulse signal modulation and transient condition of current passive radar, electronic reconnaissance, cognitive radar and spectrum situation awareness, and the process is as follows: one, assuming that V pulse description words are received within time T;two, constructing an adjacency matrix A based on the V pulse description words after pretreatment, and constructing a radar pulse description word graph with labeled node category attributes based on the adjacency matrix A;three, obtaining a trained graph convolution model;four, obtaining B to-be-detected pulse description words, outputting feature vectors of the pulse description words in the B pulse description words based on the trained model; and obtaining a clustering result through a clustering algorithm.

Description

Technical Field

[0001] This invention relates to a radar signal intelligent sorting method based on a GraphGated graph convolutional neural network, belonging to the field of information and communication engineering. Background Technology

[0002] Radar signal sorting research involves separating pulse sequences belonging to different radar radiation sources from densely interleaved and superimposed pulse streams in complex electromagnetic environments. This information is then used for subsequent radiation source identification to obtain information such as the radar's model, purpose, and threat level, thereby enabling radar signal perception and subsequent tracking or jamming.

[0003] For current electronic countermeasures systems, radar signal sorting technology is generally based on five parameters of the radar radiation source's pulse description word (PDW): time of arrival (TOA), carrier frequency (CF), pulse width (PW), direction of arrival (DOA), and pulse amplitude (PA). With technological advancements, intra-pulse modulation (PM) and polarization characteristics can be further utilized. However, considering the real-time processing requirements of broadband reconnaissance systems, DOA and polarization parameters require multi-channel sampling systems, making real-time extraction of these parameters difficult. Furthermore, PA parameters have relatively weak stability. Therefore, in engineering applications, especially in the seekers of passive radars, TOA, CF, and PW parameters are more commonly used for sorting, separating pulse sequences belonging to different radiation sources from the complex and interwoven PDW stream.

[0004] The traditional method combining multi-parameter pre-sorting and main sorting is currently the main application method in electronic reconnaissance equipment, but it is gradually becoming inadequate for the modern electronic environment. Current signal sorting relies too heavily on databases storing countermeasure plans, but this mechanism cannot quickly respond to newly emerging or dynamic advanced threats. With the rapid development of electronic countermeasures technology, new challenges have been presented to radar signal sorting, mainly including: ① A large number, high density, wide range, and severe signal overlap of radiation sources. The average number of pulses occurring per unit time can reach millions per second, placing a huge computational burden on traditional sorting algorithms, leading to increasingly serious issues of missed or added batches in the sorting results. ② Complex and variable radar signal parameter modulation. Many new radars can not only change the carrier frequency, repetition frequency, and waveform of the transmitted pulses, but their parameter changes can also be completed within milliseconds. Especially with the application of multi-functional, phased array radars, rapid changes can be achieved in time, frequency, and space. Single, fixed sorting methods relying on a single or several dimensions of combined features are difficult to adapt to the signals of this new radar system. ③ The emergence of unknown radar signals.

[0005] With the continuous application of deep learning technology in the field of radar electronic reconnaissance, radar signal sorting research has transitioned from traditional algorithms based on PRI sequence differences and time series to intelligent algorithms based on neural networks. Deep learning-based methods have attracted scholarly attention, but the main problems are: ① Fundamental framework issues such as network models have not been effectively resolved, lacking global analytical capabilities for multi-dimensional features. Most methods are based on PDW time streams using one-dimensional networks, ignoring the uncertainty of inter-pulse correlation caused by the interleaving of multiple radar signals; and most focus on PRI modulation analysis. Image mapping methods have been preliminarily introduced, but are mainly used for pre-clustering and primary sorting. ② Because signal sorting systems are passive blind receivers and non-matched receivers, fixed network input parameters are difficult to adapt to various application scenarios. Furthermore, in applications such as missile-borne systems, which are often embedded, research on lightweight design is lacking.

[0006] In recent years, Graph Convolutional Networks (GCNs) have demonstrated superior performance in node classification tasks due to their variable node features, simple network structure, and small number of model parameters. These characteristics correspond to the current challenges in radar signal sorting within deep learning. However, constructing the adjacency matrix based on PDWs (PDW Graph Adjacency Matrix) is often difficult. If the affiliation relationships between PDWs are accurately known, the radar signal sorting problem can be solved. Therefore, this paper proposes a unique PDW graph adjacency matrix method and uses GraphGated graph convolutional networks to infer the affiliation of each PDW, thereby achieving intelligent radar signal sorting during the electronic reconnaissance phase. Summary of the Invention

[0007] The purpose of this invention is to solve the problem of signal sorting difficulties in current passive radar, electronic reconnaissance, cognitive radar, and spectrum situational awareness when facing high pulse density environments, complex inter-pulse signal modulation, and transient situations involving unknown signals. Therefore, this invention proposes a radar signal intelligent sorting method based on a GraphGated graph convolutional neural network.

[0008] The specific process of a radar signal intelligent sorting method based on a GraphGated graph convolutional neural network is as follows:

[0009] Step 1: Assume that there are two radar radiation sources within time T, and the PRI values ​​of the two radiation sources are PRI1 and PRI2 respectively.

[0010] Assuming that it was received within time T One pulse description word (PDW);

[0011] Each pulse descriptor word (PDW) has ε features, where ε = 6. The ε features are arrival time, carrier frequency, pulse width, pulse amplitude, angle of arrival, and intra-pulse modulation information, respectively.

[0012] For the received Each pulse description word (PDW) is preprocessed to obtain the preprocessed result. One pulse description word (PDW);

[0013] Step 2: Based on the preprocessed An adjacency matrix is ​​constructed using pulse descriptors. Based on adjacency matrix Construct a radar pulse descriptor word (PDW) graph labeled with node category attributes;

[0014] Step 3: Construct the graph convolutional model GraphGated;

[0015] Set the loss function for the graph convolution model GraphGated;

[0016] The GraphGated convolutional model is trained based on the radar pulse descriptor (PDW) graph with labeled node category attributes constructed in step two until the loss function converges, thus obtaining the trained GraphGated convolutional model.

[0017] Step Four:

[0018] Get A pulse description word (PDW) to be tested, for Each pulse description word (PDW) to be tested is preprocessed to obtain the preprocessed pulse description word (PDW). One pulse description word (PDW);

[0019] Based on preprocessing An adjacency matrix is ​​constructed using pulse descriptors. Based on adjacency matrix Constructing a radar pulse descriptor word (PDW) diagram;

[0020] Input the radar pulse descriptor (PDW) image into the graph convolutional model GraphGated trained in step three. The graph convolutional model GraphGated trained in step three outputs... The feature vector of each pulse descriptor in a pulse descriptor;

[0021] Clustering algorithm Clustering is performed on the feature vectors of each pulse descriptor in the pulse descriptor to obtain the clustering result. In the clustering result, the pulse descriptor PDW corresponding to the feature vectors in the same cluster is the pulse sequence emitted by the same radar radiation source.

[0022] The beneficial effects of this invention are as follows:

[0023] This invention is a method for intelligent sorting of radar radiation sources based on graph convolutional neural networks. It converts pulse data over a period of time into graph structure data through pulse features, and then completes the sorting through graph convolutional inference.

[0024] (1) A novel graph convolutional radar signal sorting architecture is proposed. The PDW (Processing Node) appears as a node in the graph structure, contributing an advanced deep learning strategy to the radar signal sorting task. This graph convolutional network architecture, by using the PDW as a processing node, opens up a new path for radar signal identification and sorting. Importantly, this graph convolutional model effectively simplifies the parameter scale of deep learning models, significantly reducing model complexity. This characteristic makes the deployment and application of the model on hardware devices much simpler, providing great convenience for the practical application of radar signal processing technology.

[0025] (2) An adjacency matrix construction method is proposed to connect the graph and the PDW. To ensure that radar pulses can establish accurate and effective associations with nodes in the graph structure, this invention proposes an innovative adjacency matrix construction method. This method cleverly utilizes the characteristic differences between radar pulses and their harmonic components. In traditional radar signal processing algorithms, the presence of harmonic components often interferes with the setting of parameter thresholds and affects the accurate extraction of the true pulse repetition interval (PRI). However, the adjacency matrix construction method in this invention effectively transforms this challenge by using harmonic components as bridges to establish connections between non-adjacent pulses.

[0026] This method not only improves the accuracy of radar signal sorting but also provides a comprehensive and detailed analysis of potential pulse sequences from a global perspective. In this way, the invention can identify non-adjacent pulses that are discontinuous in time but have inherent connections in complex radar signal environments, thus providing a more comprehensive and detailed solution for radar signal sorting. This global processing approach not only optimizes the radar signal processing flow but also significantly improves the performance and efficiency of radar systems when dealing with variable signal environments.

[0027] (3) A deep clustering magnetic loss function was constructed. To better adapt to the ever-changing new radar signal systems and address the challenges posed by unknown radar signals, this invention reconceptualizes the traditional offline radar signal sorting task as a deep clustering problem. This transformation makes radar signal processing more flexible and efficient, capable of handling more complex environments. Within this framework, this invention constructs a novel magnetic loss function from a multi-dimensional global perspective and an end-to-end processing flow. This loss function possesses unique inter-class repulsion and intra-class attraction characteristics, effectively promoting the separation of pulses from different radiation sources during the clustering process while ensuring that pulses from the same radiation source are tightly clustered. Attached Figure Description

[0028] Figure 1 This is a flowchart of the intelligent radar signal sorting process based on GraphGated;

[0029] Figure 2 It is a sequence of two interleaved pulses from different radiation sources. In the diagram, Radar1 represents radar 1, and Radar2 represents radar 2.

[0030] Figure 3 It is a graph showing the pulse sequences of two radar radiation sources and their corresponding graphical structure data.

[0031] Figure 4 This is the flowchart of the improved pulse feature model conversion;

[0032] Figure 5 It is a GraphGated graph convolutional model built on gated recurrent units;

[0033] Figure 6This is a schematic diagram illustrating the thresholds for magnetic clustering loss. r1 represents the center (mean, also known as expectation) of the feature vector output from the GraphGated convolutional model of the N1 pulse descriptors (PDWs) corresponding to the first radar source; r2 represents the center (mean, also known as expectation) of the feature vector output from the GraphGated convolutional model of the N2 pulse descriptors (PDWs) corresponding to the second radar source. c N represents the radiation source of the c-th radar. c The center of the feature vector (which is the mean, also known as the expectation) of the radar pulse descriptor PDW graph input to the input graph convolutional model GraphGated;

[0034] Figure 7 This is a structural diagram of the PRI continuous and staggered radar conversion diagram;

[0035] Figure 8 This is a structural diagram of the PRI continuous and jitter radar conversion diagram;

[0036] Figure 9 The diagram is constructed from the superimposed data of multiple radar radiation sources, (a) three radar radiation sources, (b) four radar radiation sources;

[0037] Figure 10 These are examples of sorting multiple radiation sources: (a) TOA-CF plane, (b) TOA-PA plane, (c) TOA-PW plane, and (d) the feature space output by the graph convolution model.

[0038] Figure 11 This is a graph showing the effect of different lost / spurious pulse rates on the model regression purity;

[0039] Figure 12 The graph shows the impact of different loss / spurious pulse rates on the regression purity of different models. (a) Performance comparison under different pulse spurious rates, (b) Performance comparison under different pulse loss rates. Detailed Implementation

[0040] Specific Implementation Method 1: The specific process of this implementation method for intelligent radar signal sorting based on a GraphGated graph convolutional neural network is as follows:

[0041] The implementation process of this invention includes pulse data preprocessing, pulse feature map model transformation, offline training of graph convolution model, and online sorting of pulse samples. The technical solution process is as follows: Figure 1As shown, the scheme first transforms the PDW data into a graph structure suitable for graph convolutional models, then uses the labeled graph data to effectively train the graph convolutional model, enabling the model to effectively distinguish and separate pulses radiated by different radiation sources in the feature space. Finally, a clustering algorithm is used to filter the pulse sequences that have been separated in the feature space.

[0042] Step 1: Assume that there are two radar radiation sources within time T, and the PRI values ​​of the two radiation sources are PRI1 and PRI2 respectively.

[0043] Assuming that it was received within time T One pulse description word (PDW);

[0044] Each pulse descriptor word (PDW) has ε features, where ε = 6. The ε features are arrival time, carrier frequency, pulse width, pulse amplitude, angle of arrival, and intra-pulse modulation information, respectively.

[0045] For the received Each pulse description word (PDW) is preprocessed to obtain the preprocessed result. One pulse description word (PDW);

[0046] Step 2: Based on the preprocessed An adjacency matrix is ​​constructed using pulse descriptors. Based on adjacency matrix Construct a radar pulse descriptor (PDW) graph labeled with node category attributes (which radar radiation source the node belongs to).

[0047] Step 3:

[0048] Constructing a graph convolutional model GraphGated suitable for radar signal sorting, such as... Figure 5 As shown;

[0049] Set the loss function for the graph convolution model GraphGated;

[0050] The PDW graph of radar pulse descriptors (PDWs) constructed in step two, which is labeled with the node category attributes (which radar radiation source the node belongs to), is used to train the graph convolutional model GraphGated until the loss function converges, and the trained graph convolutional model GraphGated is obtained.

[0051] In this way, the model can effectively separate the radar pulses emitted by each radiation source in the feature space, and realize intelligent clustering of radiation source pulses;

[0052] Step Four:

[0053] Get A pulse description word (PDW) to be tested, for Each pulse description word (PDW) to be tested is preprocessed to obtain the preprocessed pulse description word (PDW). One pulse description word PDW (process as in step one);

[0054] Based on preprocessing An adjacency matrix is ​​constructed using pulse descriptors. Based on adjacency matrix Construct the radar pulse descriptor word (PDW) diagram (the process is the same as step two);

[0055] Input the radar pulse descriptor (PDW) image into the graph convolutional model GraphGated trained in step three. The graph convolutional model GraphGated trained in step three outputs... The feature vector of each pulse descriptor in a pulse descriptor;

[0056] Clustering algorithm The feature vectors of each pulse descriptor in the pulse descriptor are clustered to obtain the clustering result. The pulse descriptor PDWs corresponding to the feature vectors in the same cluster are pulse sequences emitted by the same radar radiation source, thus completing the accurate sorting of radar signals.

[0057] Clustering algorithms such as DBSCAN.

[0058] Online sample sorting. A well-trained graph convolutional model is used to perform inference analysis on unlabeled PDW graphs. The graph convolutional model maps each node to the feature space, achieving effective clustering of emitted pulses from different radiation sources. Subsequently, basic clustering algorithms such as DBSCAN are used to filter pulse sequences from the clustering results, thereby completing the accurate sorting of radar signals.

[0059] In the sample set to be tested, the category of the pulse signals remains unknown. Through inference using a graph convolutional network, we successfully constructed a feature space mapping. In this mapping, each pulse signal is represented as a feature vector. Although effective separation of pulse signals is achieved in the feature space, it must be noted that the network does not provide a direct output of the pulse sequence; that is, the radar radiation source sorting task only completes "separation" and not "screening." In the feature space, the screening process can be implemented using clustering algorithms, such as DBSCAN. In the DBSCAN algorithm, the critical neighborhood radius is the atomic radius ε determined by the magnetic clustering loss during training. v It is important to emphasize that DBSCAN's role here is merely to extract the pulse sequence from the already separated feature space; that is, the sorting of radar signals is essentially accomplished by graph convolutional networks.

[0060] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that: in step one, it is assumed that there are two radar radiation sources within time T, and the PRI of the two radiation sources are PRI1 and PRI2 respectively.

[0061] Assuming that it was received within time T One pulse description word (PDW);

[0062] Each pulse descriptor word (PDW) has ε features, where ε = 6. The ε features are arrival time, carrier frequency, pulse width, pulse amplitude, angle of arrival, and intra-pulse modulation information, respectively.

[0063] For the received Each pulse description word (PDW) is preprocessed to obtain the preprocessed result. One pulse description word (PDW);

[0064] The specific process is as follows:

[0065] Assume that there are two radar radiation sources within time T, and the PRI values ​​of the two radiation sources are PRI1 and PRI2 respectively.

[0066] Assuming that it was received within time T Each pulse description word (PDW) The pulse descriptor for the first-stage radiation source is {1,3,5,6,8,10}, and the pulse descriptor for the second-stage radiation source is {2,4,7,9}, as follows. Figure 2 As shown;

[0067] PRI indicates the pulse repetition period;

[0068] Each pulse descriptor word (PDW) has ε features, where ε = 6. The ε features are arrival time, carrier frequency, pulse width, pulse amplitude, angle of arrival, and intra-pulse modulation information, respectively.

[0069] The arrival time, carrier frequency, pulse width, pulse amplitude, angle of arrival, and intra-pulse modulation information of each pulse description word (PDW) are normalized to obtain the normalized arrival time, carrier frequency, pulse width, pulse amplitude, angle of arrival, and intra-pulse modulation information.

[0070] The normalized arrival time, carrier frequency, pulse width, pulse amplitude, angle of arrival, and intra-pulse modulation information data range is [-1, 1].

[0071] A radar pulse descriptor word (PDW) typically consists of time of arrival (TOA), carrier frequency (CF), pulse width (PW), pulse amplitude (PA), angle of arrival (DOA), and intra-pulse modulation information (IntraPM). Generally, the numerical values ​​of the various pulse characteristics in a PDW differ significantly; for example, TOA may be measured in nanoseconds (ns), and its order of magnitude may be 10⁻⁶. 9PA is usually expressed in dBm, and its order of magnitude is in the range of 10. 2 To balance the differences between the dimensions of each feature and reduce the network model's excessive focus on a particular feature, all features are normalized, and the normalized data range is [-1, 1].

[0072] Data preprocessing. Besides numerical features, pulses may also possess string-like features, such as intra-pulse modulation patterns. Therefore, non-numerical samples need to be encoded. Furthermore, given that different features often have their own numerical ranges and units of measurement, features with wider ranges may be overemphasized by the model, thus overshadowing features with narrower ranges. This phenomenon prevents the model from effectively utilizing all feature information. Therefore, to ensure the model can handle all features evenly, we need to normalize the data to eliminate the influence of different units of measurement between features.

[0073] The other steps and parameters are the same as in Specific Implementation Method 1.

[0074] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that: in step two, based on the preprocessed... An adjacency matrix is ​​constructed using pulse descriptors. Based on adjacency matrix Construct a radar pulse descriptor (PDW) graph labeled with node category attributes (which radar radiation source the node belongs to).

[0075] The specific process is as follows:

[0076] Assuming that it was received within time T If each pulse has ε features, then for an image, it also has ε pulses. There are ε independent nodes, each with ε features, and the relationships between nodes are represented by the adjacency matrix. Constructed using the difference in TOA;

[0077] Step Two One:

[0078] The Time of Arrival (TOA) vector, composed of pulse descriptors, is t = [t1, t2, t3, ..., t4]. v ], take the difference between any two elements in t to obtain a difference matrix D;

[0079] The elements in the difference matrix D are represented as follows:

[0080]

[0081] Among them, t i t represents the arrival time of the i-th pulse descriptor word.j D represents the arrival time of the j-th pulse descriptor word. ij This represents the time difference between the arrival time of the i-th pulse descriptor and the arrival time of the j-th pulse descriptor;

[0082] Then the difference matrix D is expressed as equation (2);

[0083] When considering the pulse arrival time differences between different radiation sources, these differences typically do not exhibit a regular pattern. To simplify the problem and facilitate understanding, we will temporarily denote these irregular arrival time differences as parameters. In the actual calculation of the frequency matrix, the pulse differences between these different radiation sources are also counted. Observing formula (2), it can be seen that the difference matrix contains not only the fundamental wave but also higher-order harmonics. The existence of higher-order harmonics establishes the connection between non-adjacent pulses.

[0084]

[0085] in,

[0086] PRI1 represents the arrival time difference between adjacent pulse descriptors in the first radiation source; 2PRI1 represents the arrival time difference between two pulse descriptors separated by 1 pulse descriptor in the first radiation source; 3PRI1 represents the arrival time difference between two pulse descriptors separated by 2 pulse descriptors in the first radiation source; 4PRI1 represents the arrival time difference between two pulse descriptors separated by 3 pulse descriptors in the first radiation source; 5PRI1 represents the arrival time difference between two pulse descriptors separated by 4 pulse descriptors in the first radiation source.

[0087] PRI2 represents the arrival time difference between adjacent pulse descriptors in the second radiation source; 2PRI2 represents the arrival time difference between two pulse descriptors separated by 1 pulse descriptor in the second radiation source; 3PRI2 represents the arrival time difference between two pulse descriptors separated by 2 pulse descriptors in the second radiation source.

[0088] This represents the time difference between the arrival of two pulse descriptors;

[0089] Rows 1-10 in the difference matrix D represent the first to the tenth pulse descriptor words;

[0090] In the difference matrix D, columns 1-10 represent the 1st to the 10th pulse descriptor words;

[0091] Step 22: Calculate the difference between the maximum and minimum values ​​in the difference matrix D;

[0092] Divide the difference between the maximum and minimum values ​​into k consecutive intervals B1, B2, ..., B according to the interval width.l ,(,B k ;

[0093] B1 represents the first interval; B2 represents the second interval; B l B represents the l-th interval; k Indicates the k-th interval;

[0094] The interval width is defined as the TOA measurement error;

[0095] Define the counting function h(B) l ), calculate each element D in matrix D ij Falling in interval B l The number of elements in the array;

[0096]

[0097] in It is an indicator function, when D ij Falling in interval B l Intratime, indicator function The value is 1 if it is not 1, otherwise it is 0.

[0098] For the interval B1, B2, ..., B l ,…,B k This yields a frequency vector H = [h(B1), h(B2), ..., h(B2)]. l ),…,h(B k )], normalizing H yields each element D in the difference matrix D ij The frequency of occurrence H′ is represented as:

[0099]

[0100] Steps 2 and 3: Use the elements in the difference matrix of equation (2) to form the adjacency matrix between nodes in the graph structure according to equation (4).

[0101] Step Two Four: Based on the reception within time T Each pulse descriptor has ε features and an adjacency matrix. Construct a pulse descriptor (PDW) graph labeled with node category attributes (which radar radiation source a node belongs to). See Figure 3 ;

[0102] Step 25: Calculate the Pulse Description Word (PDW) graph constructed in Step 24. The number of nodes with a degree of 0;

[0103] A node degree of 0 indicates that the node is an isolated node, and there are no other nodes related to it.

[0104] If the number of nodes with a degree of 0 is less than or equal to 5, then the Pulse Description Word (PDW) graph constructed in step two and four is... This is the final pulse descriptor (PDW) graph labeled with the node category attribute (which radar source the node belongs to).

[0105] If the number of nodes with a degree of 0 is greater than 5, then proceed as follows:

[0106] 1) Expand the width of each interval based on Scott's rule (in step two, the k intervals B1, B2, ..., B...). l ,…,B k The width of the intermediate interval is fixed and defined as the TOA measurement error; here, based on the intervals in step two, the width of each interval is expanded (each interval has the same width), resulting in new intervals B′1, B′2, ..., B′. l ,…,B′ k ;

[0107] 2) Based on the new intervals B′1, B′2, ..., B′ l ,…,B′ k We obtain a new frequency vector, and normalize the new frequency vector to obtain each element D in the difference matrix D. ij The new frequency of occurrence;

[0108] 3) The elements in the difference matrix of equation (2) are used to form a new adjacency matrix between nodes in the graph structure according to the new occurrence frequency;

[0109] 4) Construct a new Pulse Descriptor Word (PDW) graph based on the new adjacency matrix. The new Pulse Descriptor Word (PDW) graph is the final Pulse Descriptor Word (PDW) graph labeled with node category attributes (which radar radiation source the node belongs to).

[0110] When a radar radiation source with a large PRI fluctuation range exists in the electromagnetic environment, PRI jitter or slip radar exists. In this case, the adjustment range B... l The width of the pulses can be used to correlate the pulses that fluctuate within the PRI range. Figure 1 The process of transforming the mid-pulse characteristic model is further described as follows: Figure 4 .

[0111] Other steps and parameters are the same as in specific implementation method one or two.

[0112] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that, in steps two and three, the elements in the difference matrix of equation (2) are used to form the adjacency matrix between nodes in the graph structure according to equation (4). Represented as:

[0113]

[0114] Adjacency matrix Each element This represents the relationship between node i and node j;

[0115] This indicates that there is no relationship between node i and node j;

[0116] This indicates that there is a relationship between node i and node j, and the weight represents the degree of tightness of the relationship;

[0117] E 10×10 The unit matrix represents a node that is related to other nodes while retaining its own inherent characteristics.

[0118] When pulse number When there are a certain number of pulses, the time-of-arrival (TOA) difference between pulses from different radiation sources occurs with a frequency close to 0, and their relationship can be basically ignored. In practice, as the number of pulses to be sorted increases, the arrival time differences between pulses from different radiation sources may exhibit certain patterns. These high-frequency arrival time differences are also considered as a kind of relationship between pulses and used to construct the node adjacency matrix.

[0119] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0120] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that step three involves constructing a graph convolutional model, GraphGated, suitable for radar signal sorting, such as... Figure 5 As shown;

[0121] Set the loss function for the graph convolution model GraphGated;

[0122] The PDW graph of radar pulse descriptors (PDWs) constructed in step two, which is labeled with the node category attributes (which radar radiation source the node belongs to), is used to train the graph convolutional model GraphGated until the loss function converges, and the trained graph convolutional model GraphGated is obtained.

[0123] In this way, the model can effectively separate the radar pulses emitted by each radiation source in the feature space, and realize intelligent clustering of radiation source pulses;

[0124] The specific process is as follows:

[0125] Step 31: The graph convolutional model GraphGated consists of three time-consuming steps: GatedGraphConv and a linear layer.

[0126] The three-step GatedGraphConv consists of a first gated recurrent unit (GRU), a second gated recurrent unit, and a third gated recurrent unit.

[0127] Step 32: Set the loss function L for the graph convolution model GraphGated;

[0128] Using L with intra-class attraction var and inter-class repulsion L dist Characteristic magnetic loss function constraint z v Simultaneously considering class distribution L reg The constraints complete the joint design of (6); for the explanation of each threshold in the loss, see Figure 6 As shown.

[0129] Step 33: Based on the radar pulse descriptor PDW graph constructed in Step 2, which is labeled with node category attributes (which radar radiation source the node belongs to), input it into a 3-step time-long GatedGraphConv. The 3-step time-long GatedGraphConv outputs a feature vector, which is input into a linear layer. The linear layer outputs... The feature vector of each pulse descriptor in a pulse descriptor;

[0130] Train the graph convolutional model GraphGated until the loss function converges to obtain the trained graph convolutional model GraphGated.

[0131] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0132] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that the loss function L of the graph convolution model GraphGated is:

[0133] L=α×L var +β×L dist +γ×L reg (6)

[0134] Among them, L var Indicates intra-class attraction; L dist Indicates inter-class exclusion; L reg Represents class distribution;

[0135] α, β, and γ represent weighting coefficients, where α = 1, β = 1, and γ = 0.01.

[0136] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0137] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One to Six in that the intraclass attraction L... var Represented as:

[0138]

[0139] in,

[0140] C′ represents the radar radiation source of unit C′;

[0141] N c This indicates that the c-th radar radiation source has N c Each pulse description word (PDW)

[0142] z i The input graph of the radar pulse descriptor PDW corresponding to the i-th pulse descriptor PDW of the c-th radar radiation source is the feature vector output by the graph convolution model GraphGated.

[0143] r c N represents the radiation source of the c-th radar. c The center of the feature vector (which is the mean, also known as the expectation) of the radar pulse descriptor PDW graph input to the input graph convolutional model GraphGated;

[0144] δ v Indicates the maximum distance within the class;

[0145] L var This indicates that each PDW corresponds to The distance from a feature vector to its feature center must be less than or equal to δ. v In other words, under ideal conditions, PDWs from the same radar will be limited to a radius of δ. v inside the sphere;

[0146] [] + This indicates that the lower bound of the elements in the matrix is ​​0.

[0147] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0148] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One to Seven in that the inter-class exclusion L... dist Represented as:

[0149]

[0150] in,

[0151] C′ represents the radar radiation source of unit C′;

[0152] δ d Represents the minimum distance between classes;

[0153] r p N represents the radiation source of the p-th radar. p The center of the feature vector (which is the mean, also known as the expectation) of the radar pulse descriptor PDW graph input to the input graph convolutional model GraphGated;

[0154] r q N represents the radiation source of the q-th radar. q The center of the feature vector (which is the mean, also known as the expectation) of the radar pulse descriptor PDW graph input to the input graph convolutional model GraphGated;

[0155] N p This indicates that the p-th radar radiation source has N p Each pulse description word (PDW)

[0156] N q This indicates that the q-th radar radiation source has N q Each pulse description word (PDW)

[0157] L dist Indicates r p and r q The distance between them must be greater than or equal to v d ;

[0158] Maximum intraclass distance δ v Less than the minimum distance δ between classes d .

[0159] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0160] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that the class distribution L... reg Represented as:

[0161]

[0162] To prevent network over-optimization, L reg It is desirable that the center of each radar signal characteristic be as close as possible to the center of the characteristic space. It can be seen that L... reg It is not a primary optimization term; for an explanation of the thresholds in the magnetic clustering loss, see [link to relevant documentation]. Figure 6 As shown.

[0163] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0164] This invention, from the perspective of variable input feature dimensions and saving computational resources, constructs a graph convolutional model GraphGated suitable for radar signal sorting, such as... Figure 5 As shown, the model consists of a 3-step GatedGraphConv and a linear layer. It utilizes the characteristics of intra-class attraction and inter-class repulsion to construct a magnetic loss function to perform PDW graph classification tasks, achieving radar signal sorting. The core of GatedGraphConv is to control the node feature update process through a gated recurrent unit (GRU). Specifically, the features of node v at time step t are... The update process is as follows:

[0165]

[0166] in This represents the aggregation of neighbor node information for node v. Specifically, each node with a time step of t updates its features over multiple time steps, gradually fusing multi-hop neighbor information. The number of time steps (i.e., the number of iterations) controls the receptive field size. For example, 3 time steps can aggregate 3-hop neighbors. All time steps share the same set of GRU parameters and weight matrix W, reducing the number of parameters.

[0167] After inference processing by the PDW graph using the GraphGated model, feature vectors for each pulse in the feature space are obtained. Based on this, the feature vectors in the feature space are precisely adjusted using a constraint loss function to ensure that each pulse node can be correctly classified into its corresponding category. Iterative training of the GraphGated model using a certain number of different PDW samples can lead to model convergence. The hyperparameter settings used in the training are shown in Table 1.

[0168] Table 1 Model Hyperparameter Settings

[0169]

[0170] The evaluation index for radiation source sorting is determined by the pulse sorting success rate, namely the regression purity (RP), which represents the best matching result of the sorted pulse sequence relative to the label, and is given by equation (11).

[0171]

[0172] in Represents the total number of pulses, Ω={ω1,ω2,...,ω K} represents the division of the pulse sequence after sorting, C={c1,c2,...,c J} represents the division of the label cluster. Equation (10) shows that the regression purity assigns a label cluster to each sorted pulse sequence, and the sample in this label cluster appears most frequently in the sorted pulse sequence. RP∈[0,1], the closer to 1, the better the clustering result.

[0173] Validation of the effectiveness of the method of the present invention

[0174] (1) Feasibility verification

[0175] Feasibility verification of PDW graph structure data construction

[0176] To verify the feasibility of the adjacency matrix calculation method in this invention, adjacency matrix calculations were performed in the following two scenarios: constructing graph structure data corresponding to radar radiation sources with different PRI modulation types; and constructing graph structure data corresponding to different numbers of radar radiation sources with different PRI modulation types. Common PRI modulation types for radar radiation sources include continuous, staggered, and jittery.

[0177] a) PRI continuous

[0178] Taking two PRI continuous radar radiation sources as an example, an adjacency matrix of 10 pulses is constructed, see... Figure 3 .

[0179] b) PRI variation

[0180] Taking a continuous PRI radar source and a staggered PRI radar source as an example, construct a... A graph structure is shown. The waveform parameters of the two radar radiation sources are shown in Table 2. Using the adjacency matrix calculation method of this invention, the transformed graph structure is shown below. Figure 7 .

[0181] Table 2 Hybrid PRI Continuous and PRI Staggered Radar

[0182]

[0183] c) PRI jitter

[0184] Taking a continuous PRI radar source and a staggered PRI radar source as an example, construct a... A graph structure is shown. The waveform parameters of the two radar radiation sources are shown in Table 3. The transformed graph structure, calculated using the adjacency matrix method of this invention, is shown below. Figure 8 .

[0185] Table 3. Hybrid PRI Continuous and PRI Jitter Radar

[0186]

[0187] For the electromagnetic environment of multiple radar radiation sources (more than two), graph data is constructed according to the adjacency matrix method proposed in this invention, see [link to invention]. Figure 9 .

[0188] Feasibility verification of radar signal sorting based on graph convolution model:

[0189] According to the PDW graph transformation method proposed in this invention, a training set for the graph convolution model GrapgGated is constructed. The training set simulates radar radiation sources with five modulation types, and their waveform parameter settings are shown in Table 4. Each sample contains between 100 and 200 pulses, and the number of radiation sources in each sample ranges from 1 to 8. The PDW category attributes in the training samples are all known.

[0190] Table 4. Training Set Radar Radar Source Waveform Parameter Settings

[0191]

[0192]

[0193] A test set of 100 samples was constructed based on Table 4. For the model, the pulse class attributes of all samples in the test set are unknown; that is, the graph convolutional model learns on a graph where all node class attributes are known, and then infers from a completely unknown graph. A sample containing 5 radiation sources was sorted. The results are shown below. Figure 10 The results show that the sample contained a total of 9 radiation sources, and the graph convolution model accurately identified all 9 radar radiation sources. The regression purity of this sample reached 99.3%. The feature planes also show that the edge pulses of two radar radiation sources were not correctly identified. This can be attributed to the relatively small number of connected pulses when constructing the adjacency matrix. However, this does not affect the extraction of radiation source descriptors.

[0194] (2) Reliability verification

[0195] In real electromagnetic environments, due to the performance limitations of various communication radiation sources and reconnaissance equipment, the received pulse sequences often contain not only spurious pulses, but also, within a certain timeframe, the continuous reception of radar radiation source pulses can be affected, leading to discontinuities in the reception process. Furthermore, factors such as thermal noise can cause insufficient measurement accuracy when reconnaissance equipment measures PDW characteristics. In reliability verification, the effects of different lost pulse rates and spurious pulse rates on regression purity were simulated.

[0196] The experimental results are presented in Figure 11The results show that, under reasonable measurement error, lost pulse rate, and stray pulse rate conditions, the intelligent sorting algorithm based on the graph convolution model exhibits high concentration of regression purity for different sampled samples. Although there are some discrete points, this is sufficient to prove that the algorithm of this invention still has strong sorting performance and robustness in harsh electromagnetic environments.

[0197] (3) Comparison and performance analysis

[0198] To further verify the effectiveness and advancement of this invention, we employed various deep learning frameworks, including RCNN, LSTM, and autoencoders, as well as traditional sorting algorithms such as SDIF, and conducted comprehensive comparative tests on both lost pulse rate and spurious pulse rate.

[0199] In the experiment investigating the effect of the lost pulse rate, the stray pulse rate was fixed at 5%. Similarly, in investigating the effect of the stray pulse rate, the lost pulse rate was fixed at 5%. The experimental results are as follows: Figure 12 As shown, the data demonstrates a significant performance improvement over traditional algorithms. Even compared to advanced deep learning frameworks such as R-CNN, LSTM, or intelligent sorting algorithms based on Mask R-CNN image segmentation, the algorithm of this invention still exhibits superior performance. Specifically, when the lost pulse rate and spurious pulse rate reach 30%, the average regression purity of the proposed method can still reach over 98%. While intelligent sorting algorithms based on Mask R-CNN image segmentation are based on pure visual analysis and closely linked to current popular artificial intelligence technologies, they are susceptible to interference from spurious pulses, leading to variations in the shape of the radar source's radiated pulses. Simultaneously, lost pulses also cause changes in shape features. Therefore, under harsh conditions, the robustness of image segmentation-based intelligent sorting algorithms faces challenges. In contrast, the intelligent algorithm proposed in this invention possesses high robustness thanks to its unique adjacency matrix construction method. In this invention, the construction of the adjacency matrix not only considers the relationship between adjacent pulses of the radar source but also calculates the correlation coefficient between non-adjacent pulses. Although this method introduces some redundancy, it significantly improves the robustness of the algorithm. Comparative experiments with other excellent algorithms further verify the effectiveness and advancement of the algorithm proposed in this invention.

[0200] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A method for intelligent sorting of radar signals based on a GraphGated graph convolutional neural network, characterized in that: The specific process of the method is as follows: Step 1: Assume that there are two radar radiation sources within time T, and the PRI values ​​of the two radiation sources are PRI1 and PRI2 respectively. Assuming that it was received within time T One pulse description word (PDW); Each pulse descriptor word (PDW) has ε features, where ε = 6. The ε features are arrival time, carrier frequency, pulse width, pulse amplitude, angle of arrival, and intra-pulse modulation information, respectively. For the received Each pulse description word (PDW) is preprocessed to obtain the preprocessed result. One pulse description word (PDW); Step 2: Based on the preprocessed An adjacency matrix is ​​constructed using pulse descriptors. Based on adjacency matrix Construct a radar pulse descriptor word (PDW) graph labeled with node category attributes; Step 3: Construct a graph convolutional model GraphGated; Set the loss function for the graph convolution model GraphGated; The GraphGated convolutional model is trained based on the radar pulse descriptor (PDW) graph with labeled node category attributes constructed in step two until the loss function converges, thus obtaining the trained GraphGated convolutional model. Step Four: Get A pulse description word (PDW) to be tested, for Each pulse description word (PDW) to be tested is preprocessed to obtain the preprocessed pulse description word (PDW). One pulse description word (PDW); Based on preprocessing An adjacency matrix is ​​constructed using pulse descriptors. Based on adjacency matrix Constructing a radar pulse descriptor word (PDW) diagram; Input the radar pulse descriptor (PDW) image into the graph convolutional model GraphGated trained in step three. The graph convolutional model GraphGated trained in step three outputs... The feature vector of each pulse descriptor in a pulse descriptor; Clustering algorithm Clustering is performed on the feature vectors of each pulse descriptor in the pulse descriptor to obtain the clustering result. In the clustering result, the pulse descriptor PDW corresponding to the feature vectors in the same cluster is the pulse sequence emitted by the same radar radiation source.

2. The intelligent radar signal sorting method based on GraphGated graph convolutional neural network according to claim 1, characterized in that: In step one, it is assumed that there are two radar radiation sources within time T, and the PRI of the two radiation sources are PRI1 and PRI2 respectively. Assuming that it was received within time T One pulse description word (PDW); Each pulse descriptor word (PDW) has ε features, where ε = 6. The ε features are arrival time, carrier frequency, pulse width, pulse amplitude, angle of arrival, and intra-pulse modulation information, respectively. For the received Each pulse description word (PDW) is preprocessed to obtain the preprocessed result. One pulse description word (PDW); The specific process is as follows: Assume that there are two radar radiation sources within time T, and the PRI values ​​of the two radiation sources are PRI1 and PRI2 respectively. Assuming that it was received within time T Each pulse description word (PDW) The pulse descriptor for the first radiation source is {1,3,5,6,8,10}, and the pulse descriptor for the second radiation source is {2,4,7,9}. PRI indicates the pulse repetition period; Each pulse descriptor word (PDW) has ε features, where ε = 6. The ε features are arrival time, carrier frequency, pulse width, pulse amplitude, angle of arrival, and intra-pulse modulation information, respectively. The arrival time, carrier frequency, pulse width, pulse amplitude, angle of arrival, and intra-pulse modulation information of each pulse description word (PDW) are normalized to obtain the normalized arrival time, carrier frequency, pulse width, pulse amplitude, angle of arrival, and intra-pulse modulation information. The normalized arrival time, carrier frequency, pulse width, pulse amplitude, angle of arrival, and intra-pulse modulation information data range is [-1, 1].

3. The intelligent radar signal sorting method based on a GraphGated graph convolutional neural network according to claim 2, characterized in that: In step two, based on the preprocessed... An adjacency matrix is ​​constructed using pulse descriptors. Based on adjacency matrix Construct a radar pulse descriptor word (PDW) graph labeled with node category attributes; the specific process is as follows: Step Two One: The arrival time vector composed of pulse descriptors is Taking the difference between any two elements in t yields a difference matrix D; The elements in the difference matrix D are represented as follows: Among them, t i t represents the arrival time of the i-th pulse descriptor word. j D represents the arrival time of the j-th pulse descriptor word. ij This represents the time difference between the arrival time of the i-th pulse descriptor and the arrival time of the j-th pulse descriptor; Then the difference matrix D is expressed as equation (2); in, PRI1 represents the arrival time difference between adjacent pulse descriptors in the first radiation source; 2PRI1 represents the arrival time difference between two pulse descriptors separated by 1 pulse descriptor in the first radiation source; 3PRI1 represents the arrival time difference between two pulse descriptors separated by 2 pulse descriptors in the first radiation source; 4PRI1 represents the arrival time difference between two pulse descriptors separated by 3 pulse descriptors in the first radiation source; 5PRI1 represents the arrival time difference between two pulse descriptors separated by 4 pulse descriptors in the first radiation source. PRI2 represents the arrival time difference between adjacent pulse descriptors in the second radiation source; 2PRI2 represents the arrival time difference between two pulse descriptors separated by 1 pulse descriptor in the second radiation source; 3PRI2 represents the arrival time difference between two pulse descriptors separated by 2 pulse descriptors in the second radiation source. This represents the time difference between the arrival of two pulse descriptors; Step 22: Calculate the difference between the maximum and minimum values ​​in the difference matrix D; Divide the difference between the maximum and minimum values ​​into k consecutive intervals B1, B2, ..., B according to the interval width. l ,…,B k ; B1 represents the first interval; B2 represents the second interval; B l B represents the l-th interval; k Indicates the k-th interval; The interval width is defined as the TOA measurement error; Define the counting function h(B) l ), calculate each element D in matrix D ij Falling in interval B l The number of elements in the array; in It is an indicator function, when D ij Falling in interval B l Intratime, indicator function The value is 1 if it is not 1, otherwise it is 0. For the interval B1, B2, ..., B l ,…,B k This yields a frequency vector H = [h(B1), h(B2), ..., h(B... l ),…,h(B k )], normalizing H yields each element D in the difference matrix D ij The frequency of occurrence H′ is represented as: Steps 2 and 3: Use the elements in the difference matrix of equation (2) to form the adjacency matrix between nodes in the graph structure according to equation (4). Step Two Four: Based on the reception within time T Each pulse descriptor has ε features and an adjacency matrix. Construct a pulse descriptor (PDW) graph labeled with node category attributes. Step 25: Calculate the Pulse Description Word (PDW) graph constructed in Step 24. The number of nodes with a degree of 0; If the number of nodes with a degree of 0 is less than or equal to 5, then the Pulse Description Word (PDW) graph constructed in step two and four is... This is the final pulse descriptor word (PDW) graph labeled with node category attributes. If the number of nodes with a degree of 0 is greater than 5, then proceed as follows: 1) Expand the width of each interval based on Scott's rule to obtain new intervals B1′, B′2, ..., B l ′,…,B′ k ; 2) Based on the new intervals B1′, B′2, ..., B l ′,…,B′ k We obtain a new frequency vector, and normalize the new frequency vector to obtain each element D in the difference matrix D. ij The new frequency of occurrence; 3) The elements in the difference matrix of equation (2) are used to form a new adjacency matrix between nodes in the graph structure according to the new occurrence frequency; 4) Construct a new pulse descriptor word (PDW) graph based on the new adjacency matrix. The new pulse descriptor word graph is the final pulse descriptor word (PDW) graph labeled with node category attributes.

4. The intelligent radar signal sorting method based on a GraphGated graph convolutional neural network according to claim 3, characterized in that: In steps two and three, the elements in the difference matrix of equation (2) are used to form the adjacency matrix between nodes in the graph structure according to equation (4). Represented as: Adjacency matrix Each element This represents the relationship between node i and node j; This indicates that there is no relationship between node i and node j; This indicates that there is a relationship between node i and node j; E 10×10 Represents a unit array.

5. The intelligent radar signal sorting method based on a GraphGated graph convolutional neural network according to claim 4, characterized in that: In step three, a graph convolutional model GraphGated is constructed. Set the loss function for the graph convolution model GraphGated; The GraphGated convolutional model is trained based on the radar pulse descriptor (PDW) graph with labeled node category attributes constructed in step two until the loss function converges, thus obtaining the trained GraphGated convolutional model. The specific process is as follows: Step 31: The graph convolutional model GraphGated consists of three time-consuming steps: GatedGraphConv and a linear layer. The three-step gated graphconv consists of a first gated loop unit, a second gated loop unit, and a third gated loop unit. Step 32: Set the loss function L for the graph convolution model GraphGated; Step 33: Based on the radar pulse descriptor PDW graph constructed in Step 2 with node category attributes labeled, input it into a 3-step time-long GatedGraphConv. The 3-step time-long GatedGraphConv outputs a feature vector, which is then input into a linear layer. The linear layer outputs... The feature vector of each pulse descriptor in a pulse descriptor; Train the graph convolutional model GraphGated until the loss function converges to obtain the trained graph convolutional model GraphGated.

6. The intelligent radar signal sorting method based on a GraphGated graph convolutional neural network according to claim 5, characterized in that: The loss function L of the graph convolutional model GraphGated is: L=α×L var +β×L dist +γ×L reg (6) Among them, L var Indicates intra-class attraction; L dist Indicates inter-class exclusion; L reg Represents class distribution; α, β, and γ represent weighting coefficients.

7. The intelligent radar signal sorting method based on a GraphGated graph convolutional neural network according to claim 6, characterized in that: The intraclass attraction L var Represented as: in, C′ represents the radar radiation source of unit C′; N c This indicates that the c-th radar radiation source has N c Each pulse description word (PDW) z i The input graph of the radar pulse descriptor PDW corresponding to the i-th pulse descriptor PDW of the c-th radar radiation source is the feature vector output by the graph convolution model GraphGated. r c N represents the radiation source of the c-th radar. c The center of the feature vector output by the input graph convolution model GraphGated for each radar pulse descriptor (PDW) graph; δ v Indicates the maximum distance within the class; L var This indicates that each PDW corresponds to The distance from a feature vector to its feature center must be less than or equal to δ. v ; [ ] + This indicates that the lower bound of the elements in the matrix is ​​0.

8. The intelligent radar signal sorting method based on a GraphGated graph convolutional neural network according to claim 7, characterized in that: The interclass exclusion L dist Represented as: in, C′ represents the radar radiation source of unit C′; δ d Represents the minimum distance between classes; r p N represents the radiation source of the p-th radar. p The center of the feature vector output by the input graph convolution model GraphGated for each radar pulse descriptor (PDW) graph; r q N represents the radiation source of the q-th radar. q The center of the feature vector output by the input graph convolution model GraphGated for each radar pulse descriptor (PDW) graph; L dist Indicates r p and r q The distance between them must be greater than or equal to δ d ; Maximum intraclass distance δ v Less than the minimum distance δ between classes d .

9. The intelligent radar signal sorting method based on a GraphGated graph convolutional neural network according to claim 8, characterized in that: The class distribution L reg Represented as:

Images