Radar signal intelligent sorting method based on GraphGated graph convolutional neural network

Abstract

The invention discloses a radar signal intelligent sorting method based on a GraphGated graph convolutional neural network, relates to the radar signal intelligent sorting method based on the GraphGated graph convolutional neural network, and belongs to the field of information and communication engineering. The objective of the invention is to solve the problem of difficulty in signal sorting of unknown signal sensing in a high-pulse-density environment, complex inter-pulse signal modulation and transient conditions of a current passive radar, an electronic reconnaissance radar, a cognitive radar, frequency spectrum situation sensing and the like. The method comprises the following steps of: 1, assuming that V pulse description words are received within time T; 2, constructing an adjacent matrix A based on the V preprocessed pulse description words, and constructing a radar pulse description word graph marked with node category attributes based on the adjacent matrix A; 3, obtaining a trained graph convolution model; 4, acquiring B to-be-tested pulse description words, and outputting feature vectors of the B pulse description words based on the trained model; and obtaining a clustering result through a clustering algorithm.

Description

Technical Field

[0001] The present invention relates to an intelligent sorting method for radar signals based on a Graph Gated graph convolutional neural network, belonging to the field of information and communication engineering. Background Technique

[0002] The research on radar signal sorting is to sort out the pulse sequences belonging to different radar radiation sources from the dense, interleaved and overlapping pulse streams in a complex electromagnetic environment, for subsequent radiation source identification to obtain information such as the model, purpose and threat level of the radar, and then complete the perception of radar signals and subsequent tracking or interference, etc.

[0003] For current electronic countermeasure systems, the implementation of radar signal sorting technology is generally based on the five major parameters of the Pulse Description Word (PDW) of radar radiation sources, including the time of arrival (TOA), carrier frequency (CF), pulse width (PW), direction of arrival (DOA), and pulse amplitude (PA) and other radar waveform characteristic parameters. With the development of technology, intra-pulse modulation information (PM), polarization characteristics, etc. can be further adopted. However, generally considering the real-time processing of broadband reconnaissance systems, since the DOA and polarization parameters require the support of a multi-channel sampling system, it is not easy to extract the DOA, PM, polarization and other parameters in real time; and the stability of the PA parameter is relatively weak, so in engineering, especially in the seeker of a passive radar, more parameters such as TOA, CF, and PW are used to achieve sorting, and the pulse sequences belonging to different radiation sources are separated from the complex and interleaved PDW stream.

[0004] The traditional method combining multi-parameter pre-sorting and main sorting is the main application method in current electronic reconnaissance equipment, but it has gradually become unable to adapt to the modern electronic environment. Current signal sorting relies too much on the database storing countermeasure plans, but this mechanism cannot quickly respond to newly emerging or dynamic advanced threats. With the rapid development of electronic countermeasure technology, new challenges have been posed to radar signal sorting, mainly including: ① There are a large number of radiation sources, high density, wide range, and serious signal overlap. The average number of pulses appearing per unit time can reach several million per second, bringing a huge computational load to traditional sorting algorithms, and the phenomena of missed batches and additional batches in sorting results are becoming more and more serious. ② The parameter modulation of radar signals is complex and variable. Many new types of radars can not only change the carrier frequency, repetition frequency, and waveform of transmitted pulses, but also complete the parameter changes within a few milliseconds. Especially with the application of multi-functional and phased array radars, they can achieve rapid changes in time, frequency, and space, and it is very difficult for single or fixed joint sorting relying on certain dimensions or several dimensions of features to adapt to this new type of radar signal. ③ The emergence of unknown radar signals.

[0005] With the continuous application of deep learning technology in the field of radar electronic reconnaissance, the research on radar signal sorting has shifted from traditional algorithms based on PRI sequence differences and time series to intelligent algorithms based on neural networks. Relevant methods based on deep learning have attracted the attention of scholars, but the main problems are as follows: ① Basic framework problems such as network models have not been effectively solved, and they do not have the global analysis ability of multi-dimensional features. Most are implemented using one-dimensional networks based on the PDW time stream, ignoring the problem of uncertain pulse correlation caused by the interweaving of multiple radar signals; and most focus on PRI modulation analysis. The method of image mapping has been initially introduced and is mainly used for pre-clustering and main sorting. ② Since the signal sorting system belongs to passive blind reception and non-matching reception, it is difficult for fixed network input parameters to adapt to various application scenarios. On the other hand, in the case of on-board applications, etc., which are all embedded applications, the research on lightweight problems is also blank.

[0006] In recent years, Graph Convolutions Networks (GCN) have shown excellent performance in node classification tasks. Its node features are variable, the network structure is simple, and the number of model parameters is small. These characteristics correspond to the problems in current radar signal sorting in deep learning. However, it is often difficult to construct the adjacency matrix when building the graph structure based on PDW. If the attribution relationship between PDWs is accurately known, the problem of radar signal sorting has been solved. Therefore, a unique method for constructing the PDW graph adjacency matrix is proposed, and the attribution of each PDW is inferred based on the GraphGated graph convolution network, so as to realize intelligent sorting of radar signals in the electronic reconnaissance stage. Summary of the Invention

[0007] The object of the present invention is to solve the problem of difficult signal sorting in current passive radars, electronic reconnaissance, cognitive radars, spectrum situation awareness, etc., when facing high pulse density environments, complex inter-pulse signal modulation, and unknown signal perception in transient situations, and to propose an intelligent radar signal sorting method based on a GraphGated graph convolutional neural network.

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

[0009] Step 1: Assume that within time T, there are two radar emitters, and the PRIs of the two emitters are PRI1 and PRI2 respectively;

[0010] Assume that within time T, pulse description words PDW are received;

[0011] Each pulse description word PDW has ε features, ε = 6, and the ε features are arrival time, carrier frequency, pulse width, pulse amplitude, arrival angle, and intra-pulse modulation information respectively;

[0012] Preprocess the received pulse description words PDW to obtain the preprocessed pulse description words PDW;

[0013] Step 2: Based on the preprocessed pulse description words, construct an adjacency matrix Based on the adjacency matrix construct a radar pulse description word PDW graph labeled with node category attributes;

[0014] Step 3: Construct a graph convolution model GraphGated;

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

[0016] Based on the radar pulse description word PDW graph labeled with node category attributes constructed in Step 2, train the graph convolution model GraphGated until the loss function converges to obtain a trained graph convolution model GraphGated;

[0017] Step 4:

[0018] Obtain test pulse description words PDW, preprocess the test pulse description words PDW to obtain the preprocessed pulse description words PDW;

[0019] Based on the preprocessed pulse description words, construct an adjacency matrix Based on the adjacency matrix Construct a radar Pulse Descriptor Word (PDW) graph;

[0020] Input the radar PDW graph into the graph convolutional model GraphGated trained in Step 3, and the graph convolutional model GraphGated trained in Step 3 outputs the feature vectors of each pulse descriptor word among the pulse descriptor words;

[0021] Through a clustering algorithm, cluster the feature vectors of each pulse descriptor word among the pulse descriptor words to obtain a clustering result. The PDW of the measured pulses corresponding to the feature vectors in the same cluster in the clustering result is a pulse sequence emitted by the same radar emitter.

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

[0023] The present invention is a method for intelligent sorting of radar emitters based on a graph convolutional neural network, which converts pulse data within a period of time into graph-structured data through pulse features and then completes the inference of graph convolution.

[0024] (1) A novel graph convolutional radar signal sorting architecture is proposed. PDW appears in the form of nodes in the graph structure, contributing an advanced deep learning strategy to the sorting task of radar signals. This graph convolutional-based network architecture, by taking PDW as the processing node, opens up a new path for radar signal identification and sorting. Importantly, the graph convolutional model effectively streamlines the parameter scale of the deep learning model and significantly reduces the complexity of the model. This feature makes the deployment and application of the model on hardware devices more convenient and provides great convenience for the practical application of radar signal processing technology.

[0025] (2) A method for constructing an adjacency matrix is proposed to connect the graph and PDW. In order to ensure that radar pulses can establish accurate and effective associations with the nodes in the graph structure, the present invention proposes an innovative method for constructing an adjacency matrix. This method cleverly utilizes the feature differences between radar pulses and their harmonic components of each order. 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 method for constructing the adjacency matrix in the present invention effectively transforms this challenge and uses the harmonic components of each order as a bridge to establish connections between non-adjacent pulses.

[0026] This method not only improves the accuracy of radar signal sorting, but also, from a global perspective, conducts in-depth scrutiny and analysis of potential pulse sequences. In this way, the present invention can identify non-adjacent pulses that are discontinuous in time but have internal connections in a complex radar signal environment, thus providing a more comprehensive and detailed solution for radar signal sorting. This global processing means not only optimizes the radar signal processing flow, but also significantly enhances the performance and efficiency of the radar system in dealing with changing signal environments.

[0027] (3) A deep clustering magnetic loss function is constructed. To better adapt to the continuously changing new system signals and address the challenges posed by unknown radar signals, the present 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. Under this framework, the present invention constructs a new type of magnetic loss function from a multi-dimensional global perspective and an end-to-end processing flow. This loss function has unique inter-class repulsion and intra-class attraction characteristics, which can effectively promote the separation of pulses from different radiation sources during the clustering process while ensuring the tight aggregation of pulses from the same radiation source. Description of the Drawings

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

[0029] Figure 2 is the interleaved pulse sequence of two radiation sources Figure, Radar1 is radar 1, and Radar2 is radar 2;

[0030] Figure 3 is the pulse sequence of two radar radiation sources and its corresponding graph structure data diagram;

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

[0032] Figure 5 is the GraphGated graph convolutional model diagram built based on the gated recurrent unit;

[0033] Figure 6It is a schematic diagram showing the thresholds of magnetic clustering loss. r1 represents the center (which is the mean value and can also be called the expectation) of the feature vectors output after the radar pulse descriptor word (PDW) map corresponding to N1 PDWs of the first radar radiation source is input into the graph convolutional model GraphGated; r2 represents the center (which is the mean value and can also be called the expectation) of the feature vectors output after the radar pulse descriptor word (PDW) map corresponding to N2 PDWs of the second radar radiation source is input into the graph convolutional model GraphGated; r c represents the center (which is the mean value and can also be called the expectation) of the feature vectors output after the radar pulse descriptor word (PDW) map corresponding to N c PDWs of the c-th radar radiation source is input into the graph convolutional model GraphGated;

[0034] Figure 7 It is the structure diagram of the PRI continuous and staggered radar conversion diagram;

[0035] Figure 8 It is the structure diagram of the PRI continuous and jitter radar conversion diagram;

[0036] Figure 9 It is the diagram for constructing graph data when multiple radar radiation sources are aliased, (a) three radar radiation sources, (b) four radar radiation sources;

[0037] Figure 10 It is the diagram of the sorting example of multiple radiation sources, (a) TOA-CF plane, (b) TOA-PA plane, (c) TOA-PW plane, (d) the feature space output by the graph convolutional model;

[0038] Figure 11 It is the diagram showing the influence of different loss / spurious pulse rates on the regression purity of the model;

[0039] Figure 12 It is the diagram showing the influence 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 manners

[0040] Detailed implementation manner 1: The specific process of a radar signal intelligent sorting method based on the GraphGated graph convolutional neural network in this implementation manner is as follows:

[0041] The implementation process of the present invention includes pulse data preprocessing, pulse feature map model conversion, offline training of the graph convolutional model, and online sorting of pulse samples. The technical solution process is as Figure 1As shown in the figure. The solution first converts the PDW data into a graph structure suitable for the graph convolutional model, and uses the labeled graph data to effectively train the graph convolutional model, so that the model can effectively distinguish and separate the pulses radiated by different radiation sources in the feature space. Finally, a clustering algorithm is used to screen 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 PRIs of the two radiation sources are PRI1 and PRI2 respectively;

[0043] Assume that pulse description words PDW are received within time T;

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

[0045] Preprocess the received pulse description words PDW to obtain the preprocessed pulse description words PDW;

[0046] Step 2: Based on the preprocessed pulse description words, construct an adjacency matrix Based on the adjacency matrix construct a radar pulse description word PDW graph labeled with node category attributes (which radar radiation source's pulse description word the node belongs to);

[0047] Step 3:

[0048] Construct a graph convolutional model GraphGated suitable for radar signal sorting, as Figure 5 shown;

[0049] Set the loss function of the graph convolutional model GraphGated;

[0050] Based on the radar pulse description word PDW graph labeled with node category attributes (which radar radiation source's pulse description word the node belongs to) constructed in Step 2, train the graph convolutional model GraphGated until the loss function converges to obtain the trained graph convolutional model GraphGated;

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

[0052] Step 4:

[0053] Obtain test pulse description words PDW, and for Preprocess the pulse description words (PDWs) to be measured to obtain the preprocessed PDWs (the process is the same as in Step 1);

[0054] Based on the preprocessed PDWs, construct an adjacency matrix Based on the adjacency matrix Construct a radar pulse description word (PDW) graph (the process is the same as in Step 2);

[0055] Input the radar PDW graph into the graph convolutional model GraphGated trained in Step 3, and the graph convolutional model GraphGated trained in Step 3 outputs the feature vectors of each PDW among the PDWs;

[0056] Use a clustering algorithm to cluster the feature vectors of each PDW among the PDWs to obtain a clustering result. In the clustering result, the PDWs corresponding to the feature vectors in the same cluster are the pulse sequences emitted by the same radar radiation source, thus completing the accurate sorting of radar signals;

[0057] The clustering algorithm is, for example, DBSCAN.

[0058] Online sorting of samples. Use a trained graph convolutional model to perform inference and analysis on the unlabeled PDW graph. Map each node to the feature space through the graph convolutional model to achieve effective clustering of pulses emitted by different radiation sources. Subsequently, use a basic clustering algorithm such as DBSCAN to screen out the pulse sequences from the clustering result, thus completing the accurate sorting of radar signals.

[0059] In the set of samples to be measured, the class attribution of the pulse signals remains unknown. Through the inference of the graph convolutional network, we have 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 pointed out 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 has not completed "screening". In the feature space, the screening process can be achieved through a clustering algorithm, such as DBSCAN. In the DBSCAN algorithm, the key neighborhood radius is the atomic radius ε determined by the magnetic clustering loss during the training process v . It should be emphasized that the role of DBSCAN here is only to extract the pulse sequence from the already separated feature space, that is, the sorting of radar signals is essentially completed by the graph convolutional network.

[0060] Embodiment 2: The difference between this embodiment and Embodiment 1 is that in step 1, it is assumed that there are two radar radiation sources within time T, and the PRIs of the two radiation sources are PRI1 and PRI2 respectively;

[0061] It is assumed that pulse description words (PDWs) are received within time T;

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

[0063] The received pulse description words (PDWs) are preprocessed to obtain the preprocessed pulse description words (PDWs);

[0064] The specific process is as follows:

[0065] It is assumed that there are two radar radiation sources within time T, and the PRIs of the two radiation sources are PRI1 and PRI2 respectively;

[0066] It is assumed that pulse description words (PDWs) are received within time T, and the pulse description words belonging to the first radiation source are {1, 3, 5, 6, 8, 10}, and the pulse description words belonging to the second radiation source are {2, 4, 7, 9}, as Figure 2 shown;

[0067] PRI represents the pulse repetition interval;

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

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

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

[0071] A radar pulse description word (PDW) generally consists of arrival time (TOA), carrier frequency (CF), pulse width (PW), pulse amplitude (PA), arrival angle (DOA), and intra-pulse modulation information (IntraPM). Generally, the numerical differences between the pulse features of PDWs are relatively large. For example, TOA may be in nanoseconds, and its order of magnitude may be on the order of 10 9, PA is usually expressed in dBm, and its order of magnitude is in the range of 10 2 . In order to balance the differences between the dimensions of each feature quantity and reduce the excessive focus of the network model on a specific feature, all features are normalized, and the range of the normalized data is [-1, 1].

[0072] Data preprocessing. In addition to numerical features, pulses may also have string features, such as the in-pulse modulation method. Therefore, non-numerical type samples need to be encoded. In addition, considering that different features often have their own different numerical ranges and measurement units, those features with a wider range of values may cause the model to pay too much attention, thus drowning out the features with a smaller range of values. This phenomenon will cause the model to be unable to effectively utilize all feature information. Therefore, in order to ensure that the model can process each feature evenly, we need to normalize the data to eliminate the influence of the dimension between features.

[0073] Other steps and parameters are the same as those in the first specific implementation manner.

[0074] The third specific implementation manner: The difference between this implementation manner and the first or second specific implementation manner is that in step 2, an adjacency matrix is constructed based on the preprocessed pulse description words Based on the adjacency matrix Construct a radar pulse description word PDW graph annotated with the node category attribute (which radar radiation source's pulse description word the node belongs to);

[0075] The specific process is as follows:

[0076] Assume that pulses are received within time T, and each PDW has ε features. Then, for a graph, there are also independent nodes, and each node also has ε features. The connection between nodes, that is, the adjacency matrix is constructed through the TOA difference;

[0077] Step 2-1:

[0078] The arrival time (TOA) vector composed of v pulse description words is t = [t1, t2, t3,..., t

[0079] Take the difference between any two elements in t to obtain a difference matrix D;

[0080]

[0081] Among them, t i represents the arrival time of the i-th pulse description word, and tj denotes the arrival time of the j-th pulse descriptor, D ij denotes 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 time difference of pulse arrival between different radiation sources, these differences usually do not show a regular pattern. To simplify the problem and facilitate understanding, we uniformly mark these irregular arrival time differences as parameter In the actual process of calculating the frequency matrix, the pulse differences between these different radiation sources will also be counted for frequency. Observing Equation (2), it can be seen that there are not only fundamental waves but also higher harmonics in the difference matrix. The existence of higher harmonics establishes the connection between non-adjacent pulses.

[0084]

[0085] Among them,

[0086] PRI1 denotes the arrival time difference between adjacent pulse descriptors in the first radiation source; 2PRI1 denotes the arrival time difference between two pulse descriptors with one pulse descriptor interval in the first radiation source; 3PRI1 denotes the arrival time difference between two pulse descriptors with two pulse descriptor intervals in the first radiation source; 4PRI1 denotes the arrival time difference between two pulse descriptors with three pulse descriptor intervals in the first radiation source; 5PRI1 denotes the arrival time difference between two pulse descriptors with four pulse descriptor intervals in the first radiation source;

[0087] PRI2 denotes the arrival time difference between adjacent pulse descriptors in the second radiation source; 2PRI2 denotes the arrival time difference between two pulse descriptors with one pulse descriptor interval in the second radiation source; 3PRI2 denotes the arrival time difference between two pulse descriptors with two pulse descriptor intervals in the second radiation source;

[0088] denotes the arrival time difference between two pulse descriptors;

[0089] Rows 1 - 10 in the difference matrix D represent the 1st to 10th pulse descriptors;

[0090] Columns 1 - 10 in the difference matrix D represent the 1st to 10th pulse descriptors;

[0091] Step two: Calculate the difference between the maximum value and the minimum value in the difference matrix D;

[0092] Divide the difference between the maximum value and the minimum value into k consecutive intervals B1, B2, …, B according to the interval widthl ,(,B k ;

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

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

[0095] Define the counting function h(B l ), and calculate the number of elements in matrix D that each element D ij falls within the interval B l ;

[0096]

[0097] where is the indicator function. When D ij falls within the interval B l , the value of the indicator function is 1, otherwise it is 0;

[0098] For the intervals B1, B2, …, B l , …, B k , a frequency vector H = [h(B1), h(B2), …, h(B l ), …, h(B k )] is obtained. Normalize H to get the occurrence frequency H′ of each element D ij in the difference matrix D, which is expressed as:

[0099]

[0100] Step Two and Three: Form the adjacency matrix between the nodes in the graph structure according to Equation (4) for the elements in the difference matrix of Equation (2)

[0101] Step Two and Four: Based on the pulse description words received within time T, each pulse description word having ε features, and the adjacency matrix construct a pulse description word (PDW) graph labeled with the node category attributes (which radar radiation source's pulse description word the node belongs to) See Figure 3 ;

[0102] Step Two and Five: Calculate the number of nodes with degree 0 in the pulse description word (PDW) graph constructed in Step Two and Four ;

[0103] A node with degree 0 means the node is an isolated point and there is no relationship between this node and other nodes;

[0104] If the number of nodes with degree 0 is less than or equal to 5, the Pulse Description Word (PDW) graph constructed in Steps 2 and 4 is the Pulse Description Word (PDW) graph finally labeled with the node class attribute (the pulse description word of which radar radiation source the node belongs to).

[0105] If the number of nodes with degree 0 is greater than 5, the following process is performed:

[0106] 1), Expand the width of each interval based on the Scott rule (in Step 22, the widths of the k intervals B1, B2, …, B l , …, B k are fixed and defined as the TOA measurement error; here, expand the width of each interval on the basis of the intervals in Step 22, and the widths of each interval are the same), to obtain 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 obtain a new frequency vector, and normalize the new frequency vector to obtain the new occurrence frequency of each element D ij in the difference matrix D;

[0108] 3), Form a new adjacency matrix between the nodes in the graph structure according to the new occurrence frequency of the elements in the difference matrix of Equation (2);

[0109] 4), Construct a new Pulse Description Word (PDW) graph based on the new adjacency matrix. The new Pulse Description Word (PDW) graph is the Pulse Description Word (PDW) graph finally labeled with the node class attribute (the pulse description word of which radar radiation source the node belongs to).

[0110] When there are radar radiation sources with a large PRI fluctuation range in the electromagnetic environment, that is, there are PRI jitter and slip radar. At this time, adjusting the width of the interval B l can correlate the pulses with a fluctuating PRI range. Figure 1 The conversion process of the pulse feature model in Figure 4 is further described as

[0111] Other steps and parameters are the same as those in Embodiment 1 or 2.

[0112] Embodiment 4: The difference between this embodiment and one of Embodiments 1 to 3 is that in Step 23, the elements in the difference matrix of Equation (2) form the adjacency matrix between the nodes in the graph structure according to Equation (4) Expressed as:

[0113]

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

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

[0116] represents that there is a relationship between node i and node j, and the weight represents the degree of closeness of a certain connection;

[0117] E 10×10 represents the identity matrix, meaning that the current node is associated with other nodes and will not lose its own inherent characteristics;

[0118] When the number of pulses reaches a certain quantity, the frequency of the TOA difference between pulses from different radiation sources approaches 0, and their connection can be basically ignored. In practice, as the number of pulses to be sorted continuously increases, the time difference of arrival between pulses from different radiation sources may show certain patterns, and then these frequently occurring time differences of arrival are also regarded as a certain connection between pulses for constructing the node adjacency matrix.

[0119] Other steps and parameters are the same as those in any one of the specific embodiments 1 to 3.

[0120] Specific embodiment 5: The difference between this embodiment and any one of the specific embodiments 1 to 4 is that in step 3, a graph convolutional model GraphGated suitable for radar signal sorting is constructed, as Figure 5 shown;

[0121] Set the loss function of the graph convolutional model GraphGated;

[0122] Based on the radar pulse descriptor word PDW graph labeled with the node category attribute (the pulse descriptor word of which radar radiation source the node belongs to) constructed in step 2, train the graph convolutional model GraphGated until the loss function converges to obtain the trained graph convolutional model GraphGated;

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

[0124] The specific process is as follows:

[0125] Step 3-1: The graph convolutional model GraphGated sequentially includes 3-time-step GatedGraphConv and a linear layer;

[0126] The 3-time-step GatedGraphConv sequentially includes a first gated recurrent unit (GRU), a second gated recurrent unit, and a third gated recurrent unit;

[0127] Step 3-2: Set the loss function L of the graph convolutional model GraphGated;

[0128] Adopt a magnetic loss function with intra-class attraction L var and inter-class repulsion L dist to constrain z v , and at the same time consider the constraint of class distribution L reg to complete the joint design of (6); for the description of each threshold in the loss, see Figure 6 as shown.

[0129] Step 3-3: Input the radar pulse description word PDW graph labeled with node category attributes (the pulse description word to which the node belongs to which radar radiation source) into the 3-time-step GatedGraphConv. The 3-time-step GatedGraphConv outputs a feature vector, and the feature vector is input into the linear layer. The linear layer outputs the feature vectors of each pulse description word among the pulse description words;

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

[0131] Other steps and parameters are the same as those in any one of the specific embodiments 1 to 4.

[0132] Specific Embodiment 6: Different from any one of the specific embodiments 1 to 5, the loss function L of the graph convolutional model GraphGated is:

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

[0134] where L var represents intra-class attraction; L dist represents inter-class repulsion; L reg represents class distribution;

[0135] α, β, and γ represent weight coefficients, α = 1, β = 1, γ = 0.01;

[0136] Other steps and parameters are the same as those in any one of the first to fifth specific embodiments.

[0137] Specific Embodiment Seven: The difference between this embodiment and any one of the first to sixth specific embodiments is that the within-class attraction L var is expressed as:

[0138]

[0139] Among them,

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

[0141] N c represents that the c-th radar radiation source has N c pulse description words PDW,

[0142] z i represents the eigenvector output after the radar pulse description word PDW graph corresponding to the i-th pulse description word PDW of the c-th radar radiation source is input into the graph convolutional model GraphGated;

[0143] r c represents the center (which is the mean value and can also be called the expectation) of the eigenvectors output after the radar pulse description word PDW graph corresponding to the N c pulse description words PDW of the c-th radar radiation source are input into the graph convolutional model GraphGated;

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

[0145] L var represents that the distance from the eigenvector corresponding to each PDW to the eigenvector center should satisfy being less than or equal to δ ; that is, in the ideal state, the PDWs from the same radar will be restricted within a sphere with a radius of δ v ; v ;

[0146] [][] + represents that the lower limit of the elements in the matrix is 0.

[0147] Other steps and parameters are the same as those in any one of the first to sixth specific embodiments.

[0148] Specific Embodiment Eight: The difference between this embodiment and any one of the first to seventh specific embodiments is that the between-class repulsion L dist is expressed as:

[0149]

[0150] Among them,

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

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

[0153] r p represents the center (which is the mean value and can also be called the expectation) of the feature vector output after the radar pulse description word PDW graph corresponding to the Nth pulse description word PDW of the pth radar radiation source is input into the graph convolutional model GraphGated; p

[0154] r q represents the center (which is the mean value and can also be called the expectation) of the feature vector output after the radar pulse description word PDW graph corresponding to the Nth pulse description word PDW of the qth radar radiation source is input into the graph convolutional model GraphGated; q

[0155] N p represents that the pth radar radiation source has N p pulse description words PDW,

[0156] N q represents that the qth radar radiation source has N q pulse description words PDW,

[0157] L dist represents that the distance between r p and r q should be greater than or equal to v d ;

[0158] The maximum distance δ within the class v is less than the minimum distance δ between classes d .

[0159] Other steps and parameters are the same as those in any one of the specific embodiments one to seven.

[0160] Specific embodiment nine: The difference between this embodiment and any one of the specific embodiments one to eight is that the class distribution L reg is expressed as:

[0161]

[0162] To prevent over-optimization of the network, L reg hopes that the center of each radar signal feature should be as close as possible to the center of the feature space. It can be seen that L reg is not the main optimization item; See Figure 6 for the description of each threshold in the magnetic clustering loss.

[0163] ​​The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0164] The present invention constructs a graph convolution model GraphGated suitable for radar signal sorting from the perspective of variable input feature dimension and saving computing resources. Figure 5 As shown in the figure. The model consists of a 3-step long GatedGraphConv and a linear layer. It combines the characteristics of intra-class attraction and inter-class repulsion to construct a magnetic loss function to complete the PDW graph classification task and realize radar signal sorting. The core of GatedGraphConv is to control the update process of node features through the gated recurrent unit (GatedRecurrentUnit, GRU). Specifically, the feature of node v at time step t is The update process is:

[0165]

[0166] in Represents the aggregation of neighbor node information of node v. That is, each node of t steps updates node features through multiple time steps, gradually fusing multi-hop neighbor information. The number of time steps (i.e., the number of iterations) controls the size of the receptive field. 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 the PDW graph is processed by the GraphGated model, the feature vectors of each pulse in the feature space are obtained. On this basis, the feature vectors in the feature space are precisely adjusted through the constraint loss function to ensure that each pulse node can be correctly classified into its category. Iterative training of the GraphGated model using a certain number of different PDW samples can make the model converge. The hyperparameter settings used in the training are shown in Table 1:

[0168] Table 1 Model hyperparameter settings

[0169]

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

[0171]

[0172] in Represents the total number of pulses, Ω={ω1,ω2,...,ω K} represents the pulse sequence division after sorting, C={c1,c2,...,c J}\(\{\}\) represents the division of label clusters. It can be seen from Equation (10) that the regression purity assigns a label cluster to each sorted pulse sequence, and the samples in this label cluster appear the most times in the sorted pulse sequence. \(RP\in[0,1]\), and the closer it is to 1, the better the clustering result.

[0173] Effectiveness Verification of the Method of the Present Invention

[0174] (1) Feasibility Verification

[0175] Feasibility Verification of PDW Graph Structure Data Construction

[0176] The adjacency matrix is calculated for the following two scenarios to verify the feasibility of the adjacency matrix calculation method in the present invention: constructing graph structure data corresponding to different PRI modulation type radar radiation sources; constructing graph structure data corresponding to different numbers of different PRI modulation type radar radiation sources. Common PRI modulation types of radar radiation sources include continuous, staggered, and jittered.

[0177] a) PRI Continuous

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

[0179] b) PRI Staggered

[0180] Taking one PRI continuous radar radiation source and one PRI staggered radar radiation source as an example, a graph structure is constructed. The waveform parameters of the two radar radiation sources are shown in Table 2. Through the adjacency matrix calculation method in the present invention, the converted graph structure is shown in Figure 7 .

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

[0182]

[0183] c) PRI Jittered

[0184] Taking one PRI continuous radar radiation source and one PRI staggered radar radiation source as an example, a graph structure is constructed. The waveform parameters of the two radar radiation sources are shown in Table 3. Through the adjacency matrix calculation method in the present invention, the converted graph structure is shown in Figure 8 .

[0185] Table 3 PRI Continuous and PRI Jittered Radar Mixing

[0186]

[0187] Construct graph data for the electromagnetic environment of multiple radar radiation sources exceeding two according to the adjacency matrix method proposed in the present invention, as shown in Figure 9 .

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

[0189] According to the PDW graph conversion method proposed in the present invention, construct a training set for the graph convolutional model GrapgGated, where the training set simulates radar radiation sources of 5 modulation types, and the waveform parameters are set as shown in Table 4. Each sample contains 100 to 200 pulses, and the number of radiation sources in the sample ranges from 1 to 8. The category attributes of PDW in the training samples are all known.

[0190] Table 4 Waveform parameter settings of radar radiation sources in the training set

[0191]

[0192]

[0193] Construct a test set containing 100 samples according to Table 4. For the model, the pulse category attributes of the test set samples are all unknown, that is, the graph convolutional model learns on a graph with all known node category attributes and then infers a completely unknown graph. Sort a sampling sample containing 5 radiation sources. The results are shown in Figure 10 . It can be seen from the results that a total of 9 radiation sources are included in this sample, and the graph convolutional model accurately sorts out 9 radar radiation sources. The regression purity of this sample reaches 99.3%. It can also be seen from each feature plane that the edge pulses of two radar radiation sources are not correctly sorted, and the reason can be attributed to the small number of connected pulses when constructing the adjacency matrix. However, it does not affect the extraction of radiation source descriptors.

[0194] (2) Reliability verification

[0195] In the actual electromagnetic environment, due to the performance influence of various communication radiation sources and reconnaissance equipment, the received pulse sequence often contains not only stray pulses, but also the continuous reception of radar radiation source pulses will be affected within a certain period of time, resulting in discontinuity in the reception process. In addition, affected by factors such as thermal noise, the reconnaissance equipment may have insufficient measurement accuracy when measuring PDW characteristics. In the reliability verification, the influence of different pulse loss rates and stray pulse rates on the regression purity is simulated.

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

[0197] (3) Comparison and performance analysis

[0198] In order to further verify the effectiveness and advancement of the present invention, we adopted a variety of deep learning frameworks, including RCNN, LSTM and autoencoder, as well as traditional sorting algorithms such as SDIF, and conducted comprehensive comparative tests in terms of lost pulse rate and spurious pulse rate.

[0199] When exploring the influence of lost pulse rate in the experiment, the spurious pulse rate is fixed at 5%. Similarly, when exploring the influence of spurious pulse rate, the lost pulse rate is fixed at 5%. The experimental results are as follows Figure 12 As shown, the data shows that the performance of the present invention has been significantly improved compared with the traditional algorithm. Even compared with advanced deep learning frameworks such as RCNN, LSTM, or intelligent sorting algorithms based on MaskRCNN image segmentation, the algorithm of the present invention still shows excellent performance. Specifically, when the missing pulse rate and the stray pulse rate reach 30%, the average regression purity of the method proposed in the present invention can still reach more than 98%. The intelligent sorting algorithm based on MaskRCNN image segmentation, although it is based on pure visual analysis and is closely linked to the current popular artificial intelligence technology, is susceptible to interference from stray pulses, resulting in the shape formed by the radar radiation source radiating pulses. At the same time, the phenomenon of missing pulses will also cause changes in shape features. Therefore, under harsh conditions, the robustness of the intelligent sorting algorithm based on image segmentation faces challenges. In contrast, the reason why the intelligent algorithm proposed in the present invention has higher robustness is attributed to its unique adjacency matrix construction method. In the present invention, the construction of the adjacency matrix not only considers the relationship between adjacent pulses of the radar radiation source, but also calculates the correlation coefficient between non-adjacent pulses. Although this method introduces a certain 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] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. An intelligent sorting method for radar signals based on a Graph Gated 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 PRIs of the two radiation sources are PRI1 and PRI2 respectively; Suppose that pulse description words (PDWs) are received within time T; Each Pulse Descriptor Word (PDW) has ε features, where ε = 6, and the ε features are arrival time, carrier frequency, pulse width, pulse amplitude, arrival angle, and intra-pulse modulation information respectively; Preprocess the received pulse description words PDW to obtain the preprocessed pulse description words PDW; Step 2: Based on the preprocessed pulse descriptor words, construct an adjacency matrix Based on the adjacency matrix construct a radar pulse descriptor word PDW graph annotated with node category attributes; Step 3: Construct a graph convolutional model GraphGated; Set the loss function of the graph convolutional model GraphGated; Train the graph convolutional model GraphGated based on the radar Pulse Descriptor Word (PDW) graph labeled with node category attributes constructed in Step 2 until the loss function converges to obtain the trained graph convolutional model GraphGated; Step 4: Obtain pulse description words PDW to be measured, and perform preprocessing on pulse description words PDW to be measured to obtain the preprocessed pulse description words PDW; Based on the preprocessed pulse description words to construct an adjacency matrix Based on the adjacency matrix construct a radar pulse description word PDW graph; Input the radar pulse description word PDW graph into the graph convolutional model GraphGated trained in step three, and the graph convolutional model GraphGated trained in step three outputs the feature vectors of each pulse description word among pulse description words; Cluster the feature vectors of each pulse descriptor word among the pulse descriptor words through a clustering algorithm to obtain a clustering result. In the clustering result, the pulse descriptor words (PDWs) to be measured corresponding to the feature vectors in the same cluster are the pulse sequences emitted by the same radar emitter.

2. The intelligent sorting method for radar signals based on the Graph Gated graph convolutional neural network according to claim 1, characterized in that: In Step 1, it is assumed that there are two radar radiation sources within time T, and the PRIs of the two radiation sources are PRI1 and PRI2 respectively; Suppose that pulse description words (PDWs) are received within time T; Each Pulse Descriptor Word (PDW) has ε features, where ε = 6, and the ε features are arrival time, carrier frequency, pulse width, pulse amplitude, arrival angle, and intra-pulse modulation information respectively; Preprocess the received pulse description words (PDWs) to obtain the preprocessed pulse description words (PDWs); The specific process is as follows: Assume that there are two radar radiation sources within time T, and the PRIs of the two radiation sources are PRI1 and PRI2 respectively; Assume that pulse description words PDW are received within time T, and the pulse description words belonging to the first radiation source are {1, 3, 5, 6, 8, 10}, and those belonging to the second radiation source are {2, 4, 7, 9}; PRI represents the Pulse Repetition Interval; Each Pulse Descriptor Word (PDW) has ε features, where ε = 6, and the ε features are arrival time, carrier frequency, pulse width, pulse amplitude, arrival angle, and intra-pulse modulation information respectively; Normalize the arrival time, carrier frequency, pulse width, pulse amplitude, arrival angle, and intra-pulse modulation information of each Pulse Descriptor Word (PDW) respectively to obtain the normalized arrival time, carrier frequency, pulse width, pulse amplitude, arrival angle, and intra-pulse modulation information; The data range of the normalized arrival time, carrier frequency, pulse width, pulse amplitude, arrival angle, and intra-pulse modulation information is [-1, 1].

3. The intelligent sorting method for radar signals based on the Graph Gated graph convolutional neural network according to claim 2, wherein: In the second step, based on the preprocessed pulse descriptor words, an adjacency matrix is constructed Based on the adjacency matrix a radar pulse descriptor word PDW graph marked with node category attributes is constructed; the specific process is as follows: Step 2-1: The arrival time vector composed of Take the difference between any two elements in t to obtain a difference matrix D; The elements in the difference matrix D are expressed as: where, t i represents the arrival time of the i-th pulse descriptor word, t j represents the arrival time of the j-th pulse descriptor word, D ij represents the time difference between the arrival time of the i-th pulse descriptor word and the arrival time of the j-th pulse descriptor word; Then the difference matrix D is expressed as Equation (2); Among them, PRI1 represents the arrival time difference between adjacent Pulse Descriptor Words in the first radiation source; 2PRI1 represents the arrival time difference between two Pulse Descriptor Words with a one-Pulse Descriptor Word interval in the first radiation source; 3PRI1 represents the arrival time difference between two Pulse Descriptor Words with a two-Pulse Descriptor Word interval in the first radiation source; 4PRI1 represents the arrival time difference between two Pulse Descriptor Words with a three-Pulse Descriptor Word interval in the first radiation source; 5PRI1 represents the arrival time difference between two Pulse Descriptor Words with a four-Pulse Descriptor Word interval in the first radiation source; PRI2 represents the arrival time difference between adjacent Pulse Descriptor Words in the second radiation source; 2PRI2 represents the arrival time difference between two Pulse Descriptor Words with a one-Pulse Descriptor Word interval in the second radiation source; 3PRI2 represents the arrival time difference between two Pulse Descriptor Words with a two-Pulse Descriptor Word interval in the second radiation source; Indicates the time difference of arrival of two pulse descriptor words; Step 2-2: Calculate the difference between the maximum value and the minimum value in the difference matrix D; Divide the difference between the maximum value and the minimum value into k consecutive intervals B1, B2, …, B l , …, B k ; B1 represents the first interval; B2 represents the second interval; B l represents the l-th interval; B k represents the k-th interval; The interval width is defined as the TOA measurement error; Define the counting function h(B l ), which calculates the number of elements in matrix D that each element D ij falls within the interval B l ; wherein is an indicator function, when D ij falls within the interval B l the value of the indicator function is 1, otherwise 0; For intervals B1, B2, …, B l , …, B k , a frequency vector H = [h(B1), h(B2), …, h(B l ), …, h(B k )] is obtained. Normalize H to get the occurrence frequency H′ of each element D ij in the difference matrix D, expressed as: Step 2-3: Form the adjacency matrix between the nodes in the graph structure based on the elements in the difference matrix of Equation (2) according to Equation (4). Step 24: Based on the pulse descriptor words received within time T, each pulse descriptor word having ε features, and the adjacency matrix construct a pulse descriptor word (PDW) graph annotated with node class attributes Step 25: Calculate the number of nodes with degree 0 in the Pulse Description Word (PDW) graph constructed in Step 24 among them; If the number of nodes with degree 0 is less than or equal to 5, the Pulse Description Word (PDW) graph constructed in Step 2.4 is the final Pulse Description Word (PDW) graph with node class attributes labeled If the number of nodes with degree 0 is greater than 5, then perform the following process: 1), Expand the width of each interval based on the Scott 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 to obtain a new frequency vector, and normalize the new frequency vector to obtain the new occurrence frequency of each element D in the difference matrix D ij ; 3) Form a new adjacency matrix between the nodes in the graph structure according to the new occurrence frequencies of the elements in the difference matrix of Equation (2); 4) Construct a new Pulse Description Word (PDW) graph based on the new adjacency matrix. The new PDW graph is the PDW graph with the node class attributes finally labeled.

4. An intelligent sorting method for radar signals based on a Graph Gated Graph Convolutional Neural Network according to claim 3, characterized in that: In the second step 23, the elements in the difference matrix of equation (2) are used to form the adjacency matrix between the nodes in the graph structure according to equation (4). It is expressed as: Adjacency matrix Each element in represents the relationship between node i and node j; Indicates that there is no relationship between node i and node j; Indicates that there is a relationship between node i and node j; E 10×10 Denotes the identity matrix.

5. The intelligent sorting method for radar signals based on the Graph Gated graph convolutional neural network according to claim 4, characterized in that: In step 3, construct the graph convolutional model GraphGated; Set the loss function of the graph convolutional model GraphGated; Train the graph convolutional model GraphGated based on the radar pulse descriptor PDW graph labeled with node category attributes constructed in step 2 until the loss function converges to obtain the trained graph convolutional model GraphGated; The specific process is as follows: Step 3-1: The graph convolutional model GraphGated sequentially includes 3-time-step GatedGraphConv and a linear layer; The 3-time-step GatedGraphConv sequentially includes a first gated recurrent unit, a second gated recurrent unit, and a third gated recurrent unit; Step 3-2: Set the loss function L of the graph convolutional model GraphGated; Step 3: Input the PDW graph with node category attributes constructed in Step 2 into the GatedGraphConv with a 3-step time length. The GatedGraphConv with a 3-step time length outputs a feature vector, which is input into a linear layer, and the linear layer outputs the feature vectors of each pulse descriptor among the pulse descriptors; Train the graph convolutional model GraphGated until the loss function converges to obtain the trained graph convolutional model GraphGated.

6. The intelligent sorting method for radar signals based on the Graph Gated 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 represents intra-class attraction; L dist represents inter-class repulsion; L reg represents class distribution; α, β, and γ represent weight coefficients.

7. The intelligent sorting method for radar signals based on the Graph Gated graph convolutional neural network according to claim 6, wherein: The intra-class attraction L var is expressed as: Where, C′ represents the C′-part radar emitter; N c It means that the c-th radar radiation source has N c pulse description words (PDWs), z i It represents the feature vector output after the input radar pulse description word (PDW) diagram corresponding to the i-th PDW of the c-th radar radiation source is convolved by the input diagram convolution model GraphGated; r c represents the center of the feature vector output after the input radar pulse description word (PDW) map corresponding to the N pulse description words (PDWs) of the c-th radar radiation source is convolved by the graph gated convolution model GraphGated; c ​ δ v represents the maximum distance within the class; L var Indicates that the distance from the dimensional feature vector corresponding to each PDW to the feature center should satisfy being less than or equal to δ v ; [ ] + Indicates that the lower limit of the elements in the matrix is 0.

8. The intelligent sorting method for radar signals based on the Graph Gated graph convolutional neural network according to claim 7, wherein: The inter-class repulsion L dist is expressed as: Where, C′ represents the C′-part radar emitter; δ d represents the minimum distance between classes; r p represents the center of the eigenvector output after the input graph convolution model GraphGated of the radar pulse description word PDW diagram corresponding to the N p pulse description words PDW of the p-th radar radiation source; r q represents the center of the feature vector output after the input radar pulse descriptor word (PDW) map corresponding to the N pulse descriptor words (PDWs) of the q-th radar radiation source passes through the graph gated convolution model GraphGated; q ​ L dist represents r p and r q the distance between them should be greater than or equal to δ d ; The maximum distance δ within the class v is less than the minimum distance δ between classes d .

9. The intelligent sorting method for radar signals based on the Graph Gated Graph Convolutional Neural Network according to claim 8, characterized in that: The class distribution L reg is represented as:

Images