Multifunctional radar signal sorting method based on GCN and ResNet

Abstract

The present invention belongs to the technical field of radar signal classification, and in particular, relates to a multifunctional radar signal sorting method based on GCN and ResNet. It includes: constructing a multifunctional radar PDW data set, normalizing, converting the multifunctional radar PDW data into graph structure data using the KNN algorithm, and adding labels to the graph structure data; constructing a three-layer GCN network and adding a residual network; using labeled graph structure data to train a "semi-supervised" GCN-ResNet network; and then inputting the graph structure data into the trained GCN-ResNet network to output the classification results of the unlabeled graph structure data. The method uses the KNN algorithm to convert the multifunctional radar PDW data into graph structure data, and relies on the "semi-supervised" learning network to improve the generalization ability of GCN under small sample conditions; effectively solves the problem of "increased batches" in multifunctional radar signal sorting, and improves the sorting accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal classification, and in particular to a multifunctional radar signal sorting method based on a graph convolutional network (GCN) and a residual network (ResNet). Background Art

[0002] With the increasingly harsh electromagnetic environment and increasing combat requirements in modern electronic warfare, the radars used by both sides of the war are no longer limited to conventional single-function radars. Advanced integrated multi-function radars are constantly being put into use and playing a vital role. In order to improve the adaptability to the battlefield, multi-function radars usually adopt different working modes, such as conventional fixed parameter mode, frequency domain agility mode, time domain agility mode, and multi-dimensional multi-parameter joint agility mode. For multiple multi-function radar radiation source signals from the same direction, the pre-sorting algorithm based on the arrival direction is difficult to effectively dilute, resulting in the conventional main sorting algorithm easily misclassifying the working mode of a single multi-function radar into multiple radars, that is, the "increase batch" phenomenon, which brings huge challenges to radar signal sorting.

[0003] Compared with traditional convolutional neural networks that can only be used for grid-shaped data, GCN can process irregular graphs or network data. The network treats data samples as graph nodes and searches for potential associations between data by analyzing the connection relationship between graph nodes and the attribute characteristics of nodes. GCN is a "semi-supervised" learning network. The so-called "semi-supervised" means that the network can adapt to environments with smaller training sets. Therefore, GCN has the advantages of being good at analyzing complex related data and coping with ordinary small sample constraints. However, in complex electromagnetic environments, the types of intercepted radar signals are often more than the types of signals in our database, that is, only a very small number of pulse signals have labels, and most radar signals do not have labels, that is, there are extremely small sample constraints in the actual radar signal sorting process. The original GCN can only aggregate second-order neighbor nodes, ignoring a wider range of node information, and the sorting accuracy under the constraints of extremely small samples is difficult to guarantee.

[0004] The purpose of this invention is to solve the above-mentioned "batch increase" problem of multifunctional radar signal sorting and the problem that GCN is difficult to cope with extremely small sample constraints, and proposes a multifunctional radar signal sorting method based on GCN and ResNet. Summary of the invention

[0005] The purpose of the present invention is to solve the "increased batch" problem in multifunctional radar signal sorting and the problem of decreased sorting accuracy caused by fewer labeled samples when GCN is used for signal sorting. A multifunctional radar signal sorting method based on GCN and ResNet is proposed. The network constructed based on GCN and ResNet is hereinafter referred to as GCN-ResNet.

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

[0007] The radar signal sorting system on which the multifunctional radar signal sorting method relies includes: a graph structure data construction module, a GCN-ResNet network construction module, and a GCN-ResNet network training and testing module;

[0008] Among them, the graph structure data construction module uses the K-Nearest Neighbor (KNN) algorithm to process the radar pulse description word (PDW) to obtain the adjacency matrix, and combines the feature parameters to construct the graph structure data; the GCN-ResNet network construction module consists of three convolutional layers, a residual network and an activation function; the GCN-ResNet network training and testing module uses small sample data to train the network, and then tests large-scale unlabeled data.

[0009] The connection relationship between the modules in the multifunctional radar signal sorting method is as follows:

[0010] The graph structure data construction module is connected to the GCN-ResNet network construction module, and the GCN-ResNet network construction module is connected to the GCN-ResNet network training and testing module.

[0011] The purpose of the present invention is achieved through the following technical solutions.

[0012] The multifunctional radar signal sorting method based on the GCN-ResNet network includes a graph structure data construction stage, a GCN-ResNet network construction stage, a training stage, and a classification stage;

[0013] The graph structure data construction phase includes constructing a multi-function radar PDW data set, normalizing it, converting the multi-function radar PDW data into graph structure data using the KNN algorithm, and adding labels to the graph structure data; the GCN-ResNet network construction phase includes constructing a three-layer GCN network and adding a residual network;

[0014] The multifunctional radar signal sorting method comprises the following steps:

[0015] S1, graph structure data construction phase, specifically:

[0016] S11, constructing a multifunctional radar PDW data set X including PW, RF, DOA and PRI parameters;

[0017] The PW, RF, and DOA characteristic parameters are multi-function radar data to be sorted, and each parameter is set to change agilely within a certain range, and the same parameters are interlaced and overlapped;

[0018] The PRI is used to count the number of pulses within the interception time, including conventional, slip, staggered, and jitter modulation modes. The PRI in different modes of the multi-function radar has different degrees of aliasing;

[0019] Among them, PW, namely pulse width, is called Pulse Width in English; RF, namely carrier frequency, is called Radio Frequency in English; DOA, namely direction of arrival, is called Direction of Arrival in English; PRI, namely pulse repetition interval, is called Pulse Repetition Interval in English;

[0020] S12, normalize the multi-function radar PDW data set constructed in S11 to obtain a normalized PDW data set

[0021] The normalization adopts the maximum and minimum value method;

[0022] S13. Use the KNN algorithm to convert the normalized PDW data into graph structure data and establish the association between the data, including constructing directed graph structure data and undirected graph structure data, specifically:

[0023] S131, constructing directed graph structure data, specifically: taking each normalized PDW data sample as a graph node, traversing the remaining N-1 nodes, using the KNN algorithm to find the K most adjacent nodes, defining that there is a relationship between the two, and setting the value of the corresponding position of the directed adjacency matrix A' to 1, thereby obtaining directed graph structure data;

[0024] Among them, N is the number of samples of PDW data, and K is the number of limited neighboring samples around it;

[0025] S132, constructing undirected graph structure data, specifically: symmetrizing the directed graph to obtain a symmetric adjacency matrix A, and constructing an undirected graph based on the data characteristics

[0026] S14, adding labels to the graph structure data, specifically: dividing the graph structure data into a training set and a test set according to a ratio of U:10-U, adding labels to the training set, that is, the labeling rate is U / 10, to meet the small sample constraint condition;

[0027] The U is less than or equal to 1;

[0028] So far, from S11 to S14, the graph structure data construction phase has been completed;

[0029] S2, the "semi-supervised" learning network GCN-ResNet construction phase, including the construction of a three-layer GCN network and the addition of a residual network, specifically:

[0030] S21. Build a three-layer GCN network, including building three convolutional layers, setting activation functions, and setting loss functions. Specifically:

[0031] S211, constructing three convolutional layers including a first convolutional layer, a second convolutional layer and a third convolutional layer;

[0032] The input of the first convolutional layer is undirected graph structure data, and its output is the input of the second convolutional layer; the output of the second convolutional layer is used as the input of the third convolutional layer;

[0033] The propagation functions of the first, second and third convolutional layers are:

[0034]

[0035] Among them, H (l+1) and H (l) They represent the output feature matrix and input feature matrix of the lth convolutional layer respectively. l takes the values ​​of 1, 2 and 3, corresponding to the 1st, 2nd and 3rd convolutional layers respectively. σ is the activation function. W (l) Represents the weight matrix of the lth convolutional layer;

[0036] is the full node adjacency matrix, expressed as:

[0037]

[0038] Where I represents the identity matrix;

[0039] is the node degree matrix, and the elements in the matrix are expressed as:

[0040]

[0041] S212, setting the activation function of the convolution layer, setting the activation function of the first convolution layer and the second convolution layer to the ReLU nonlinear activation function, and setting the activation function of the third convolution layer to the Softmax function, so as to predict the label of each node;

[0042] S213, setting a loss function, using the cross entropy function of all labeled nodes as the loss function to evaluate the classification performance of the model;

[0043] S22, add a residual network, add an identity mapping between the input and output of the second convolutional layer, that is, directly connect the input and output, and then form a "residual network" structure through the activation function to learn a wider range of node information;

[0044] So far, from S21 to S22, the GCN-ResNet network construction phase has been completed;

[0045] S3, training phase, using labeled graph structure data to train the "semi-supervised" GCN-ResNet network, specifically:

[0046] S31, initialization parameters of the GCN-ResNet network constructed for S2, including learning rate, number of random initialization seeds, maximum number of iterations, and weight decay value;

[0047] S32, using the graph structure data in S1 to train the network, and stopping the training after the maximum number of iterations is met;

[0048] The main steps of the training phase are:

[0049] S321, aggregate neighbor nodes, and represent the characteristics of each node by the weighted sum of the first-order neighbor node characteristics and the node's own characteristics;

[0050] S322, using the activation function to update each node information;

[0051] S323, the network iterates and learns layer by layer, repeating the two operations of aggregation and updating;

[0052] So far, from S31 to S32, the GCN-ResNet network training phase has been completed;

[0053] S4, classification stage, specifically: input the graph structure data in S1 into the trained GCN-ResNet network, and output the classification results of the unlabeled graph structure data.

[0054] Beneficial Effects

[0055] Compared with the existing signal sorting methods, the multifunctional radar signal sorting method based on GCN and ResNet has the following beneficial effects:

[0056] 1. The sorting method uses the KNN algorithm to convert the multi-function radar PDW data into graph structure data, establishes the association between the data, and provides data support for the application of GCN in the field of radar signal sorting;

[0057] 2. The sorting method is based on a "semi-supervised" learning network, which is not only conducive to classifying complex related data, but also improves the generalization ability of GCN under small sample conditions;

[0058] 3. The sorting method uses GCN-ResNet to learn graph structure data and realizes the prediction of unknown nodes through aggregation and update operations, which effectively solves the problem of "increased batches" in multi-function radar signal sorting and improves the sorting accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a flow chart of a multifunctional radar signal sorting method based on GCN and ResNet of the present invention;

[0060] Figure 2 This is a schematic diagram of PDW parameters of the radar part in Example 1 of the multifunctional radar signal sorting method based on GCN and ResNet of the present invention;

[0061] Figure 3 It is a sample distribution diagram of radar simulation parameters in feature space in Example 1 of the multifunctional radar signal sorting method based on GCN and ResNet of the present invention;

[0062] Figure 4 It is a schematic diagram of the radar chart structure data construction process in Example 1 of the multifunctional radar signal sorting method based on GCN and ResNet of the present invention;

[0063] Figure 5 This is a diagram of the sorting results in Example 1 of the multifunctional radar signal sorting method based on GCN and ResNet of the present invention;

[0064] Figure 6 This is a curve showing the change of sorting accuracy with marking rate in Example 1 of the multifunctional radar signal sorting method based on GCN and ResNet of the present invention. DETAILED DESCRIPTION

[0065] In order to better illustrate the purpose and advantages of the present method, the specific implementation content of the multifunctional radar signal sorting method based on GCN and ResNet of the present invention is further described in detail in combination with the accompanying drawings and specific embodiments.

[0066] Example 1

[0067] Electronic warfare includes electronic support, electronic protection and electronic attack. Among them, the main task of electronic support is to search, intercept, identify and locate the electromagnetic energy radiated intentionally or unintentionally by the enemy, so as to provide decision makers with richer and more accurate tactical intelligence support. As a key technology in electronic support, radar signal sorting is the process of deinterleaving the signals of different radar radiation sources intercepted by the radar reconnaissance system.

[0068] This embodiment describes the specific implementation of the multifunctional radar signal sorting method based on GCN and ResNet described in this application in sorting different radar signals. The implementation flow chart of this application is as follows: Figure 1 shown.

[0069] The simulation environment includes software and hardware configuration, as follows:

[0070] 1. Software configuration:

[0071] Operating system: Windows 11; Simulation software: Python 3.9;

[0072] 2. Hardware configuration:

[0073] CPU: AMD R7 4800U (1.8GHz); Memory: 16GB.

[0074] Most radar signal sorting algorithms rely on the PDW parameters of the pulse signal, which is a digital representation of the pulse signal parameters by the radar reconnaissance system. Taking the LFM signal as an example, some parameters such as Figure 2 shown.

[0075] In order to verify the effectiveness of the method and its generalization under the constraint of small samples, this embodiment simulates the PDW parameters of four radars, as shown in Table 1. Specifically, the interception time of the simulated electronic reconnaissance aircraft is 0.2 seconds, and the transmission signals of two conventional radars (Radar No. 1 and Radar No. 3) and two multi-function radars (Radar No. 2 and Radar No. 4) are intercepted, and their four characteristic parameters are extracted: PW, RF, DOA and PRI. PRI has multiple modulation modes such as conventional, slip, staggered, and jitter. The same parameters of the four radars have different degrees of overlap. In addition, PRI is used to count the number of pulses within the interception time, so the three-dimensional characteristic parameters of PW, RF, and DOA are selected as the data to be sorted, and the distribution is as follows: Figure 3 As shown, Figure 3 (3a) is a three-dimensional feature space distribution diagram, and (3b) is a two-dimensional feature space distribution diagram.

[0076] Table 1 Radar signal simulation parameter settings

[0077]

[0078] The multi-function radar signal sorting method proposed in this application is divided into four stages, mainly including: multi-function radar image structure data construction stage, GCN-ResNet network construction stage, training stage and classification stage. The following is the specific implementation process.

[0079] S1, graph structure data construction stage, mainly using KNN algorithm to convert PDW data into graph structure data with embedded correlation, to provide data support for subsequent signal sorting, specifically:

[0080] S11. Construct a multi-function radar PDW data set X, simulate two conventional radars and two multi-function radars, each radar includes PW, RF, DOA and PRI parameters, and the parameters overlap with each other, as shown in Table 1. The distribution in the feature space is as follows: Figure 3 As shown, the three parameters of PW, RF and DOA are selected as the multi-function radar data to be sorted;

[0081] S12. Process the PDW data set according to the maximum and minimum value normalization method to obtain a normalized PDW data set It is expressed as follows:

[0082]

[0083] Among them, the radar PDW data is expressed as n is the data sample number, j is the feature number, n = 1, 2, ..., N, j = 1, 2, ..., J, N is the total number of samples in the data set, and J is the feature dimension of the sample;

[0084] S13. Use the KNN algorithm to convert the normalized PDW data into graph structure data and establish the association between the data, including constructing directed graph structure data and undirected graph structure data. The construction process is as follows: Figure 4 As shown, specifically:

[0085] S131, construct directed graph structure data, treat each normalized data sample as a graph node, and traverse the remaining N-1 nodes. At the same time, use the KNN algorithm to find the K most adjacent nodes, define the relationship between the two, and set the value of the corresponding position of the adjacency matrix A' to 1. For example, if node n1 is associated with node n2, then A' n1n2 =1, thus we get a directed graph;

[0086] Where K is the number of neighboring samples, ranging from [1, 20], and n1 and n2 are the subscript indices of the elements in the adjacency matrix;

[0087] S132, construct undirected graph structure data. In a directed graph, the association relationship does not appear in pairs. In order to adapt to the learning process of GCN, the directed graph needs to be converted into an undirected graph, that is, the unidirectional association is converted into a bidirectional association. Specifically, the directed graph is symmetric, and the element A' in the directed adjacency matrix is n1n2 =A' n2n1 , we get the symmetric adjacency matrix A, and construct an undirected graph based on the data features.

[0088] S14, adding labels to the graph structure data, specifically: dividing the graph structure data into a training set and a test set according to a ratio of U:10-U, adding labels to the training set, that is, the labeling rate is U / 10, to meet the small sample constraint condition;

[0089] Wherein, U is set to 1 in this embodiment;

[0090] So far, from S11 to S14, the graph structure data construction phase has been completed;

[0091] S2, the "semi-supervised" network GCN-ResNet construction phase, including building a three-layer GCN network, adding a residual network, and setting the loss function, specifically:

[0092] S21, set the learning process of the first convolutional layer, expressed as:

[0093]

[0094] Among them, H (2) ∈R N×C Output feature matrix for the first convolutional layer; X∈R N×J is the characteristic matrix of normalized PDW data, W (1) ∈R J×C is the weight matrix of the first convolutional layer, R represents a real number set, J = 3 is the dimension of the input data, and C = 16 is the output feature dimension of the first convolutional layer;

[0095] S22, setting the learning process of the second convolutional layer, specifically: based on the residual network, adding an identity mapping between the input and output of the second convolutional layer, that is, directly connecting the input and the output, and then forming a "residual network" structure through the activation function, which is conducive to increasing the depth of the network so as to learn more node information. The learning process of the second convolutional layer is expressed as:

[0096]

[0097] Among them, H (3) ∈R N×C is the output feature matrix of the second convolutional layer, and its feature dimension is the same as the input feature matrix H (2) Same, W(2) ∈R C×C is the weight matrix of the second convolutional layer;

[0098] S23, setting the learning process of the third convolutional layer, expressed as:

[0099]

[0100] Among them, the input of the third convolutional layer is the output H of the second convolutional layer (3) , the activation function is the Softmax function, which can predict the label Z∈R of each node N×F , F is the total number of categories in the data set, W (3) ∈R C×F is the weight matrix of the third convolutional layer;

[0101] S24. Set the loss function. For the GCN-based semi-supervised learning model, in order to evaluate the classification performance of the model, the cross entropy function of all labeled nodes is used as the loss function:

[0102]

[0103] Among them, y L is the set of all labeled nodes, L is the size of the set, and Y represents the true label of the node;

[0104] So far, from S21 to S24, the GCN-ResNet network construction phase has been completed;

[0105] S3, training phase, using graph structure data to train the GCN-ResNet network, specifically:

[0106] S31. Initialize the hyperparameters of the GCN-ResNet network constructed in S2, including the parameters of KNN and GCN. For KNN, the K value is set to 12, the distance parameter is set to sqeuclidean, and for GCN, the number of random initialization seeds is set to 42, the maximum number of iterations is set to 100, the initial learning rate is set to 0.01, and the weight decay value is 0.0005; the input dimension of the first convolutional layer is J=3, and the output dimension is C=16; the input and output dimensions of the second convolutional layer are the same as the output dimension of the first convolutional layer, that is, C=16; similarly, the input dimension of the third convolutional layer is the same as the output dimension of the second convolutional layer, that is, C=16, and the output dimension of the third convolutional layer is F=4;

[0107] S32, use the graph structure data with a labeling rate of 0.1 in S1 to train the network, and stop training after the maximum number of iterations is met;

[0108] So far, from S31 to S32, the GCN-ResNet network training phase has been completed;

[0109] S4, classification stage, input the graph structure data into the trained GCN-ResNet network and output the signal classification result;

[0110] At this point, the GCN-ResNet classification stage ends.

[0111] In the specific implementation, in order to test the sorting performance and generalization performance of GCN-ResNet, based on the radar PDW dataset shown in Table 1, the following two experiments are designed:

[0112] Experiment 1: Three commonly used signal sorting algorithms are selected as comparison algorithms, including K-means clustering algorithm, density based spatial clustering of applications with noise (DBSCAN) algorithm, and support vector machine (SVM) algorithm. The first two algorithms are unsupervised clustering algorithms, and the last algorithm is a supervised learning algorithm. After parameter adjustment, the hyperparameters are set as follows:

[0113] 1. For the K-means algorithm, the initial number of clusters n_clusters = 4, the maximum number of iterations max_iter = 100, and the distance parameter Distance is set to Euclidean distance;

[0114] 2. For DBSCAN, set the neighborhood radius epsilon = 0.08 and the minimum number of neighborhood sample points MinPts = 4;

[0115] 3. For SVM, set the penalty coefficient of the error term c = 0.8, select the linear kernel function as the kernel function type, and the maximum number of iterations max_iter = 100.

[0116] The sorting results of the GCN-ResNet-based sorting algorithm and three comparison algorithms on the radar PDW dataset are shown in the figure below. Figure 5 As shown. Among them, (5a) is the sorting result of the K-means algorithm, (5b) is the sorting result of the DBSCAN algorithm, and (5c) is the sorting result of the SVM algorithm. Figure 5 (5d) is the sorting result of the GCN-ResNet algorithm.

[0117] By analyzing Figure 5 , we can find:

[0118] 1. Since the K-means algorithm is based on the Euclidean distance between samples and cannot process clusters of arbitrary shapes, it cannot accurately sort multi-function radar data with overlapping parameters, such as Figure 5 As shown in (5a);

[0119] 2. The DBSCAN algorithm uses the high-density connectivity of the data for classification and divides the data within the density range into one category. However, in this experiment, the four radars overlap seriously in the parameter feature space. The DBSCAN algorithm mistakenly divides the four radars into two radars, that is, not only the No. 1 and No. 3 conventional radars are classified into multi-function radars, but also the working mode III of the No. 2 multi-function radar is selected as part of the No. 4 multi-function radar. Figure 5 As shown in (5b);

[0120] 3. Due to the small size of the training set in this experiment, the SVM algorithm did not fully learn all the data information, and mistakenly classified part of the No. 3 radar data into the No. 2 radar data. In addition, the sorting effect of the overlapping No. 1 and No. 4 radar data was also poor. Figure 5 As shown in (5c);

[0121] 4. From Figure 5 As can be seen from (5d), GCN-ResNet can correctly sort out the four radars, which solves the problem of "increased batches" in sorting. Although a small part of the data is misclassified, the overall sorting effect is better than the other three algorithms.

[0122] This experiment uses sorting accuracy and running time to evaluate the sorting performance of the GCN-ResNet-based sorting algorithm and three comparison algorithms on the simulation data set, as shown in Table 2. The average sorting accuracy of the four algorithms is 66.38%, 69.85%, 89.17% and 94.57% respectively. It can be seen that the sorting accuracy of GCN-ResNet is higher than that of other algorithms. This is because GCN-ResNet relies more on the data graph structure rather than parameter features and the number of labels. Unlike the other three comparison algorithms, even under small sample constraints (labeling rate is 0.1), it can also better analyze the association and distinction between multi-function radar signals. In addition, although the running time of the DBSCAN algorithm and the SVM algorithm is lower than that of the algorithm of the present invention, the DBSCAN algorithm is greatly affected by the two parameters of the neighborhood radius and the number of domain samples, and is prone to the problem of "increase in batches", and cannot effectively sort multi-function radar signals; for the SVM algorithm, there are fewer labeled training samples, resulting in unsatisfactory sorting effect. In short, the algorithm of the present invention can achieve excellent results in both sorting accuracy and computational cost.

[0123] Table 2 Sorting performance of four algorithms

[0124]

[0125] Experiment 2: In order to test the generalization performance of GCN-ResNet under the constraint of small samples, this experiment selected SVM and GCN as comparison algorithms. First, the radar simulation data set in Table 1 is divided into a small number of labeled training sets and a large number of unlabeled test sets, and then the three algorithm models are trained on the training set. Finally, the unlabeled test set is used to test the sorting accuracy of the three algorithms. The curve of sorting accuracy changing with labeling rate is shown in the figure below. Figure 6 shown.

[0126] By analyzing Figure 6 , we can find:

[0127] 1. As the labeling rate increases, the classification accuracy of GCN-ResNet, GCN and SVM algorithms generally tends to increase. However, compared with the GCN and SVM algorithms, the classification performance of the GCN-ResNet algorithm is more excellent and stable. Specifically, when the labeling rate is 0.025, the classification accuracy of the GCN-ResNet algorithm is as high as 94.42%; and when the labeling rate is 0.1, the classification accuracy is 94.57%, indicating that the classification accuracy of the algorithm is relatively high;

[0128] 2. When the marking rate gradually increases in the range of [0.025, 0.1], the sorting accuracy does not fluctuate significantly, indicating that the algorithm is insensitive to changes in the marking rate;

[0129] 3. When the marking rate is very low, such as 0.005, the algorithm also has a good sorting effect, with an accuracy rate of 87.64%.

[0130] It can be concluded that the GCN-ResNet described in the present invention effectively solves the "batch increase" problem when deinterleaving multi-function radar pulse signals, and performs superiorly in terms of sorting accuracy and computational overhead. Even under small sample constraints, the algorithm still has good generalization performance.

[0131] The above is only a preferred embodiment of the present invention, and the present invention should not be limited to the contents disclosed in the embodiment and the drawings. Any equivalent or modification completed without departing from the spirit disclosed in the present invention shall fall within the scope of protection of the present invention.

Claims

1. A multifunctional radar signal sorting method based on GCN and ResNet, and a radar signal sorting system based on it, comprising: a graph structure data construction module, a GCN-ResNet network construction module, and a GCN-ResNet network training and testing module; the graph structure data construction module uses the K nearest neighbor algorithm to process the radar pulse description word to obtain the adjacency matrix, and combines the characteristic parameters to construct the graph structure data; the GCN-ResNet network construction module consists of three convolutional layers, a residual network, and an activation function; the GCN-ResNet network training and testing module uses small sample data to train the network, and then tests large-scale unlabeled data; the graph structure data construction module is connected to the GCN-ResNet network construction module, and the GCN-ResNet network construction module is connected to the GCN-ResNet network training and testing module, characterized in that, The steps include: S1, graph structure data construction stage, including construction of multi-function radar PDW data set, normalization processing, and conversion of multi-function radar PDW data into graph structure data using KNN algorithm; dividing the graph structure data into training set and test set according to the ratio of U:10-U, adding labels to the graph structure data included in the training set, and defining it as a training set with labeled graph structure data; S2, the "semi-supervised" learning network GCN-ResNet construction phase, including building a three-layer GCN network and adding a residual network; The GCN-ResNet construction phase is used to construct a "semi-supervised" learning network GCN-ResNet; S2, specifically: S21. Build a three-layer GCN network, including building three convolutional layers, setting activation functions, and setting loss functions. Specifically: S211, constructing a first convolutional layer, a second convolutional layer, and a third convolutional layer; S212, setting the activation function of the convolution layer, setting the activation function of the first convolution layer and the second convolution layer to the ReLU nonlinear activation function, and setting the activation function of the third convolution layer to the Softmax function, so as to predict the label of each node; S213, setting a loss function, using the cross entropy function of all labeled nodes as the loss function to evaluate the classification performance of the model; S22, add a residual network, add an identity mapping between the input and output of the second convolutional layer, that is, directly connect the input and output, and then form a "residual network" structure through the activation function to learn a wider range of node information; So far, from S21 to S22, the GCN-ResNet network construction phase has been completed; S3, training phase, using labeled graph structure data to train the "semi-supervised" GCN-ResNet network, specifically: S31, initialization parameters of the GCN-ResNet network constructed for S2, including learning rate, number of random initialization seeds, maximum number of iterations, and weight decay value; S32, training the network using the training set of labeled graph structure data in S1, and stopping the training after the maximum number of iterations is met; S4, classification stage, specifically: input the graph structure data in the test set in S1 into the trained GCN-ResNet network, and output the classification results of the unlabeled graph structure data.

2. The multifunctional radar signal sorting method according to claim 1, characterized in that: S1. Specifically: S11, constructing a multi-function radar PDW data set X including PW, RF, DOA and PRI parameters; S12, normalize the multi-function radar PDW data set constructed in S11 to obtain a normalized PDW data set S13. Use the KNN algorithm to convert the normalized PDW data into graph structure data and establish the association between the data, including constructing directed graph structure data and undirected graph structure data, specifically: S131, constructing directed graph structure data, specifically: taking each normalized PDW data sample as a graph node, traversing the remaining N-1 nodes, using the KNN algorithm to find the K most adjacent nodes, and when there is a correlation between two most adjacent nodes, setting the value of the corresponding position of the directed adjacency matrix A' to 1, thereby obtaining the directed graph structure data; Among them, N is the number of samples of PDW data, and K is the number of limited neighboring samples; S132, constructing undirected graph structure data, specifically: symmetrizing the directed graph to obtain a symmetric adjacency matrix A, and constructing an undirected graph based on data features.

3. The multifunctional radar signal sorting method according to claim 2, characterized in that: The PW, RF, and DOA characteristic parameters described in S11 are the multifunctional radar data to be sorted. Each parameter is set to change agilely within a certain range, and the same parameters are interlaced and overlapped. The PRI, i.e., pulse repetition interval, is used to count the number of pulses within the interception time, including conventional, sliding, staggered, and jitter modulation methods. The PRI in different modes of the multifunctional radar has different degrees of aliasing. Among them, PW, i.e., pulse width; RF, i.e., carrier frequency; DOA, i.e., direction of arrival.

4. The multifunctional radar signal sorting method according to claim 2, characterized in that: The normalization described in S12 adopts the maximum and minimum value method.

5. The multifunctional radar signal sorting method according to claim 1, characterized in that: S32 trains the network, specifically: S321, aggregate neighbor nodes, and the characteristics of each node are represented by the weighted sum of the first-order neighbor node characteristics and the node's own characteristics; S322, using the activation function to update each node information; S323, the network iteratively learns layer by layer, repeating the two operations of aggregation and updating.

Images