Radar signal spectrum clustering sorting method based on SOM anchor point extraction and graph fusion

Abstract

The application relates to the technical field of signal processing, data representation and classification, and discloses a radar signal spectrum clustering sorting method based on SOM anchor point extraction and graph fusion, which comprises the following steps: configuring a radar signal sorting cluster number and a radar pulse parameter, obtaining a normalized radar pulse data set, taking each normalized radar pulse in the radar pulse data set as a node, and constructing a KNN graph of the radar pulse; extracting an anchor point of the normalized radar pulse data set based on SOM, calculating the similarity between the extracted SOM anchor point and all nodes, obtaining a similarity matrix, and constructing an anchor graph adjacency matrix based on the similarity matrix, namely an adaptive anchor graph; weightedly fusing the KNN graph and the adaptive anchor graph to obtain a fusion graph; and performing spectrum clustering sorting based on the radar signal sorting cluster number and the fusion graph to obtain a sorting result. The application breaks through the limitation that a classical radar sorting clustering algorithm can only utilize distance information and density information, and improves the sorting performance under complex radar pulse distribution conditions.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of signal processing, data representation and classification, and particularly relates to a radar signal spectrum clustering sorting method based on SOM anchor point extraction and graph fusion. BACKGROUND

[0002] Radar signal sorting (RSS) is one of the key technologies of reconnaissance systems, and its purpose is to separate the pulse sequence transmitted by each radar from the received random mixed pulse stream. The specific process of RSS is: deinterleaving and deinterleaving the received radar pulse sequence, separating the pulse sequence from the same radiation source, and obtaining radiation source information. The obtained radiation source information can be used to assist in evaluating the threat level of the radiation source, and the radiation source parameter information can also be sent to the jamming equipment to realize electronic jamming of the radiation source. However, with the emergence of multifunctional radars with multiple working modes, the complex distribution of radar signals has brought challenges to radar sorting in complex electromagnetic environments.

[0003] Radar signal sorting algorithm is implemented based on radar pulse description word (PDW), including direction of arrival (DOA), time of arrival (TOA), pulse width (PW), radio frequency (RF), and pulse amplitude (PA). Existing radar signal sorting algorithms can be mainly divided into two categories: PRI-based sorting method and multi-parameter-based clustering sorting method. Early single-parameter radar sorting algorithms are mainly based on pulse repetition interval (PRI) and its transformed parameters, including cumulative difference integrated function (CDIF), sequence difference histogram (SDIF), and pulse repetition interval (PRI) transformation method. The algorithm highly depends on the measurement accuracy of TOA parameter, and the sorting accuracy is affected by the pulse loss and false pulses caused by environmental noise and hardware devices, and the robustness is poor. Therefore, unsupervised clustering algorithm based on multiple reliable parameters has gradually become the mainstream, and the clustering methods commonly used for radar signal sorting include hierarchical clustering, fuzzy clustering, support vector clustering, spatial density clustering, and K-Means clustering. However, such methods only rely on the distance information and density information of the original features, and the feature information mining ability of the pulses is limited, and the clustering effect is poor under the parameter overlap condition, and a breakthrough needs to be sought from a new angle.

[0004] In recent years, with the rapid development of graph representation learning, the graph-based unsupervised clustering algorithm spectral clustering has excellent performance in face recognition, data analysis, and image clustering. The algorithm aims to explore the local graph structure data, and can optimally divide the data samples of any shape to obtain good clustering accuracy. The graph-based clustering method first needs to construct a graph model, and the most commonly used similarity graphs in the literature are k-nearest neighbors, ε-neighborhood, and complete connected graph. The main idea of the existing method is to use parameters such as k, ε, and σ to represent the local features of the data set.

[0005] A recent study shows that the clustering result depends on the selection of the similarity graph and its parameters. However, for data sets with clusters of arbitrary shape, varying density, and unbalanced clusters, how to set appropriate parameters and improve clustering performance has become a challenging task. SUMMARY

[0006] The purpose of the present application is to solve the technical problems of the prior art, i.e., the clustering sorting of the existing radar pulse sequence has limited pulse feature information mining capability, and the clustering effect is poor in the case of parameter overlap. A spectrum clustering sorting method based on SOM anchor point extraction and graph fusion is proposed. First, a self-adaptive topological graph of the radar pulse sequence is constructed by using a K nearest neighbor algorithm (KNN) and a self-organizing map (SOM), and then a spectrum clustering algorithm is used to improve the spectrum clustering performance.

[0007] To achieve the above object, the present application provides the following technical scheme.

[0008] In a first aspect, the present application provides a radar signal spectrum clustering sorting method based on SOM anchor point extraction and graph fusion, comprising the following steps:

[0009] S1. Configure the number of radar signal sorting clusters and radar pulse parameters, and construct a radar pulse dataset based on the radar pulse parameters;

[0010] S2. Normalize the radar pulse parameters in the radar pulse dataset to obtain a normalized radar pulse dataset;

[0011] S3. Use each normalized radar pulse parameter in the radar pulse dataset as a node to construct a KNN graph of the radar pulse;

[0012] S4. Extract the anchor points of the normalized radar pulse dataset based on SOM, and construct an adaptive anchor graph based on multiple anchor points; specifically comprising the following sub-steps:

[0013] S40. Assign a random value between 0 and 1 to the weight vector of each neuron in the SOM output layer, and configure the number of iterations;

[0014] S41. Take the normalized radar pulse dataset as input, and iteratively train the SOM until the number of iterations is completed. The final neuron weight vector obtained is the extracted SOM anchor point;

[0015] S42. Calculate the similarity between the SOM anchor points and all nodes to obtain a similarity matrix;

[0016] S43. Construct an anchor graph adjacency matrix based on the similarity matrix, which is the adaptive anchor graph;

[0017] S5. Weighted fusion of the KNN graph and the adaptive anchor graph to obtain a fusion graph;

[0018] S6. Perform spectrum clustering sorting based on the number of radar signal sorting clusters and the fusion graph to obtain a sorting result.

[0019] As a possible implementation manner, the radar pulse parameters are: direction of arrival, carrier frequency and pulse width.

[0020] As a possible implementation manner, S3 specifically comprises the following sub-steps:

[0021] S30. Set the number of radar pulse parameters as N, and the normalized radar pulse parameters as nodes;

[0022] S31. Take each node as a center node respectively, traverse the remaining N-1 nodes, measure the similarity between nodes based on Euclidean distance, find out the K nodes most adjacent to the center node, and construct a KNN graph adjacency matrix;

[0023] S32. Set the matrix values corresponding to each node and its K adjacent nodes in the KNN graph adjacency matrix to 1, and obtain a directed graph adjacency matrix;

[0024] S33. Symmetrize the directed graph adjacency matrix, and construct an undirected graph, which is the KNN graph of the radar pulse.

[0025] As a possible implementation manner, S41 specifically comprises the following sub-steps:

[0026] S410. Configure the learning rate and neighborhood radius, input the normalized radar pulse dataset to SOM for the t-th iteration, and calculate the Euclidean distance between each radar pulse parameter in the radar pulse dataset and each weight vector;

[0027] S411. Find the neuron with the smallest Euclidean distance, which is the winning neuron. Based on the learning rate, correct the weight vectors of the neurons within the center neighborhood radius.

[0028] S412. If t is less than the number of iterations, return to S410 for the next iteration, until t is equal to the number of iterations. If t is equal to the number of iterations, the SOM neuron weight vector obtained at this time is the extracted anchor point.

[0029] As a possible implementation manner, the initial learning rate is greater than 0 and less than 1, and the learning rate decreases with the number of iterations each time.

[0030] As a possible implementation manner, the stability of the anchor graph adjacency matrix is improved by using the KNN graph adjacency matrix, the KNN graph adjacency matrix is assigned a weight, and the fusion graph is the sum of the KNN graph adjacency matrix multiplied by the weight and the anchor graph adjacency matrix.

[0031] As a possible implementation manner, S42 specifically comprises:

[0032] The similarity between the SOM anchor point and all nodes is calculated using a Gaussian kernel function as follows:

[0033]

[0034] The similarity matrix is constructed as follows:

[0035]

[0036] wherein x i represents the i-th pulse node, c j represents the j-th anchor point, M represents the number of anchor points, w ij represents the similarity between the i-th graph node and the j-th anchor point, and σ represents a parameter for controlling the neighborhood width.

[0037] As a possible implementation, S6 specifically comprises the following sub-steps:

[0038] S60. Constructing a fusion graph adjacency matrix based on the KNN graph adjacency matrix and the anchor graph adjacency matrix;

[0039] S61. Constructing a degree matrix and a Laplacian matrix based on the fusion graph adjacency matrix, and calculating a normalized Laplacian matrix;

[0040] S62. Calculating the eigenvectors corresponding to the c smallest eigenvalues of the normalized Laplacian matrix, and constructing an eigenvector matrix with the eigenvectors as elements, wherein c is the number of sorting clusters;

[0041] S63. Each row of the eigenvector matrix represents a sample, and K-means clustering is performed on all samples to obtain final clustering clusters, which are the sorting results.

[0042] As a possible implementation, the weight vector of the neuron within the central neighborhood radius is corrected based on the learning rate with the winning neuron as the center, and the correction is specifically as follows:

[0043]

[0044] wherein j * is the winning neuron, j is the neuron, j≠j * , X is the signal parameter set, W is the weight vector, t is the current iteration number, η(t) is the learning rate of the t-th iteration, N g (t) is the neighborhood radius of the t-th iteration.

[0045] In a second aspect, the present application provides a radar signal spectrum clustering and sorting system based on SOM anchor point extraction and graph fusion, comprising:

[0046] The KNN graph construction unit configures a radar signal sorting cluster number and a radar pulse parameter, constructs a radar pulse dataset based on the radar pulse parameter, normalizes radar pulse parameters in the radar pulse dataset, obtains the normalized radar pulse dataset, takes each normalized radar pulse parameter in the radar pulse dataset as a node, and constructs a KNN graph of the radar pulse.

[0047] The SOM anchor point extraction unit assigns a random value between [0, 1] to a weight vector of each neuron of a SOM output layer and configures an iteration number; takes the normalized radar pulse dataset as input, iteratively trains the SOM until the iteration number is completed, and finally obtains a neuron weight vector as an extracted SOM anchor point.

[0048] The adaptive anchor graph construction unit calculates a similarity between the SOM anchor point and all nodes, obtains a similarity matrix, and constructs an anchor graph adjacency matrix based on the similarity matrix, that is, the adaptive anchor graph.

[0049] The fusion graph construction unit improves the stability of the anchor graph adjacency matrix by using the KNN graph adjacency matrix, assigns a weight value to the KNN graph adjacency matrix, and fuses the graph as a sum of the KNN graph adjacency matrix multiplied by the weight value and the anchor graph adjacency matrix.

[0050] The spectral clustering sorting unit constructs a fusion graph adjacency matrix based on the KNN graph adjacency matrix and the anchor graph adjacency matrix, constructs a degree matrix and a Laplacian matrix based on the fusion graph adjacency matrix, calculates a standardized Laplacian matrix, calculates a feature vector corresponding to c smallest eigenvalues of the standardized Laplacian matrix, takes the feature vector as an element to construct a feature matrix, where c is a sorting cluster number, each row of the feature matrix represents a sample, and K-means clustering is performed on all samples to obtain a final clustering cluster, that is, a sorting result.

[0051] Compared with the prior art, the present application has the following beneficial effects:

[0052] 1. The radar signal spectral clustering sorting method based on SOM anchor point extraction and graph fusion provided by the present application fully utilizes the function of SOM in capturing the topological structure of input samples, obtains more uniformly distributed and more representative anchor points, and can better reflect the graph structure characteristics of input samples compared with the random selection of anchor points in the prior art.

[0053] 2. The radar signal spectral clustering sorting method based on SOM anchor point extraction and graph fusion provided by the present application introduces a clustering method based on graph theory for radar signal sorting, takes the topological structure of pulses as one of the features, breaks through the limitation of classical radar sorting clustering algorithms that can only use distance information and density information, and improves the sorting performance under complex radar pulse distribution conditions.

[0054] 3. The radar signal spectrum clustering sorting method based on SOM anchor point extraction and graph fusion provided by the present application fuses an anchor point graph and a KNN graph to construct a fusion graph, and uses the KNN graph to correct the anchor point graph, thereby further improving the clustering performance of spectrum clustering. BRIEF DESCRIPTION OF DRAWINGS

[0055] The accompanying drawings, which are included to provide a further understanding of the present application and are incorporated in and constitute a part of this application, illustrate embodiments of the present application and together with the description serve to explain the present application. In the drawings:

[0056] Figure 1 A flow chart of the radar signal spectrum clustering sorting method based on SOM anchor point extraction and graph fusion in the embodiment of the present application;

[0057] Figure 2 A flow chart of the method for constructing an adaptive anchor graph after extracting anchor points based on SOM in the embodiment of the present application;

[0058] Figure 3 A sorting result graph of a type 4 radar based on the sorting method of the embodiment of the present application;

[0059] Figure 4 A sorting result graph of a type 4 radar based on the sorting method of the embodiment of the present application;

[0060] Figure 5 A sorting result graph of a type 4 radar based on the sorting method of the embodiment of the present application;

[0061] Figure 6 A sorting result graph of a type 4 radar based on the sorting method of the embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to clearly describe the technical solutions of the embodiments of the present application, in the embodiments of the present application, the same items or similar items with basically the same functions and effects are distinguished by using “first”, “second”, etc. For example, the first threshold value and the second threshold value are only used to distinguish different threshold values, and do not limit the order. Those skilled in the art can understand that “first”, “second”, etc. do not limit the quantity and execution order, and “first”, “second”, etc. also do not necessarily mean different.

[0063] It should be noted that in the present application, the words “exemplary” or “for example” are used to mean serving as an example, instance, or illustration. Any embodiment or design presented as “exemplary” or “for example” in the present application should not be interpreted as being more preferred or advantageous than other embodiments or designs. Rather, the use of “exemplary” or “for example” is intended to present concepts in a particular manner.

[0064] In the present application, "at least one" means one or more, and "multiple" means two or more. The association relationship of "and / or" describes the association relationship of the associated objects, which means that there can be three kinds of relationships, for example, A and / or B, which can represent the following three cases: A exists alone, A and B exist together, and B exists alone, where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after it. The following at least one or similar expressions mean any combination of these items, including any combination of single or multiple items. For example, at least one of a, b or c can represent a, b, c, the combination of a and b, the combination of a and c, the combination of b and c, or the combination of a, b and c, where a, b and c can be single or multiple.

[0065] With the emergence of multifunctional radars with multiple working modes, the complex distribution of radar signals poses a challenge to radar signal sorting in complex electromagnetic environments. In view of the technical status that the existing radar signal spectrum clustering sorting algorithm has limited pulse feature information mining capability and poor clustering effect in the case of parameter overlap, a radar signal spectrum clustering sorting method based on SOM anchor point extraction and graph fusion is proposed. First, K nearest neighbor algorithm (K Nearest Neighbors, KNN) and self-organizing map (Self-Organizing Maps, SOM) are used to construct an adaptive topological graph of radar pulse sequences, and then a spectral clustering algorithm is used to improve the spectral clustering performance of pulse sequence deinterleaving.

[0066] In a first aspect, the present application provides a radar signal spectrum clustering sorting method based on SOM anchor point extraction and graph fusion, as shown in Figure 1 , comprising the following steps:

[0067] S1. Configure the number of radar signal sorting clusters and radar pulse parameters, and construct a radar pulse dataset based on the radar pulse parameters;

[0068] As a possible implementation manner, the radar pulse parameters are: direction of arrival, carrier frequency and pulse width.

[0069] S2. Normalize the radar pulse parameters in the radar pulse dataset to obtain a normalized radar pulse dataset;

[0070] As an example, the normalization of the radar pulse parameters in the radar pulse dataset is specifically: normalizing each radar pulse parameter to [0, 1];

[0071] S3. Using each normalized radar pulse parameter in the radar pulse dataset as a node, a KNN graph of the radar pulse is constructed;

[0072] As a possible implementation manner, S3 specifically comprises the following sub-steps:

[0073] S30. Set the number of radar pulse parameters as N, and take the normalized radar pulse parameters as nodes;

[0074] S31. Take each node as a center node respectively, traverse the remaining N-1 nodes, measure the similarity between nodes based on Euclidean distance, find out the K nodes most adjacent to the center node, and construct a KNN graph adjacency matrix;

[0075] S32. Set the matrix values corresponding to each node and its K adjacent nodes in the KNN graph adjacency matrix to 1, and obtain a directed graph adjacency matrix;

[0076] S33. Symmetrize the directed graph adjacency matrix, construct an undirected graph, and obtain the KNN graph of the radar pulse.

[0077] S4. Extract the anchor points of the normalized radar pulse data set based on SOM, and construct an adaptive anchor graph based on the extracted multiple anchor points; see Figure 2 , specifically comprising the following sub-steps:

[0078] S40. Assign a random value between 0 and 1 to the weight vector of each neuron in the SOM output layer, and configure the number of iterations;

[0079] S41. Take the normalized radar pulse data set as input, iteratively train the SOM until the number of iterations is completed, and the final neuron weight vector obtained is the extracted SOM anchor point;

[0080] As a possible implementation manner, S41 specifically comprises the following sub-steps:

[0081] S410. For the tth iteration, input the normalized radar pulse data set to the SOM, and calculate the Euclidean distance between each radar pulse parameter in the radar pulse data set and each weight vector;

[0082] S411. Find the neuron with the smallest Euclidean distance, which is the winning neuron, and modify the weight vector of the neuron within the center neighborhood radius based on the learning rate with the winning neuron as the center;

[0083] As a possible implementation manner, the weight vector of the neuron within the center neighborhood radius is modified with the winning neuron as the center, specifically as follows:

[0084]

[0085] Wherein, j * is the winning neuron, j is the neuron, and j≠j* , X is a signal parameter set, W is a weight vector, t is the current iteration number, η(t) is the learning rate of the tth iteration, N g (t) is the neighborhood radius of the tth iteration.

[0086] S412. Update the size N g (t) and the learning rate η(t), the field radius and the learning rate function select the exponential function, specifically:

[0087] N g (t) = INT[N g (0)·exp(-t / (5·T))]

[0088] η(t) = η(0)·exp(-t / (5·T))

[0089] Where, INT[·] is the integer function, T is the total iteration number.

[0090] Let t = t + 1, return S410, when t = T, the SOM network stops running. At this time, the SOM neuron weight vector obtained is the extracted anchor point.

[0091] It should be noted that the first iteration, the weight vector of the neurons within the initial radius of the center neighborhood is corrected based on the initial learning rate, and the neighborhood radius and learning rate are updated after each correction, and the next iteration is entered until the configured iteration number is completed.

[0092] As a possible implementation, the initial learning rate is greater than 0 and less than 1, and the learning rate of each iteration decreases with the iteration number.

[0093] S42. Calculate the similarity between the SOM anchor points and all nodes to obtain a similarity matrix;

[0094] As a possible implementation, S42 is specifically:

[0095] The similarity between the SOM anchor points and all nodes is calculated using the Gaussian kernel function as follows:

[0096]

[0097] The similarity matrix is constructed as follows:

[0098]

[0099] In the formula, x i represents the i-th graph node, c j represents the j-th anchor point, M represents the number of anchor points, w ijrepresents the similarity between the ith graph node and the jth anchor point, and σ represents a parameter for controlling the neighborhood width.

[0100] S43. Constructing an anchor graph adjacency matrix based on the similarity matrix, that is, an adaptive anchor graph;

[0101] S5. Weighted fusion of the KNN graph and the adaptive anchor graph to obtain a fusion graph;

[0102] As a possible implementation, the stability of the anchor graph adjacency matrix is improved by using the KNN graph adjacency matrix, the KNN graph adjacency matrix is assigned a weight value, and the fusion graph is the sum of the KNN graph adjacency matrix multiplied by the weight value and the anchor graph adjacency matrix.

[0103] S6. Performing spectral clustering sorting based on the radar signal sorting cluster number and the fusion graph to obtain a sorting result.

[0104] As a possible implementation, S6 specifically includes the following sub-steps:

[0105] S60. Constructing a fusion graph adjacency matrix based on the KNN graph adjacency matrix and the anchor graph adjacency matrix;

[0106] S61. Constructing a degree matrix and a Laplacian matrix based on the fusion graph adjacency matrix, and calculating a normalized Laplacian matrix;

[0107] S62. Calculating the eigenvectors corresponding to the c smallest eigenvalues of the normalized Laplacian matrix, and using them as elements to construct an eigenvector matrix, where c is the sorting cluster number;

[0108] S63. Each row of the eigenvector matrix represents a sample, and K-means clustering is performed on all samples to obtain the final clustering cluster, that is, the sorting result.

[0109] In a second aspect, the present application provides a radar signal spectral clustering sorting system based on SOM anchor point extraction and graph fusion, comprising:

[0110] A KNN graph construction unit configured with a radar signal sorting cluster number and radar pulse parameters, constructs a radar pulse dataset based on the radar pulse parameters, normalizes the radar pulse parameters in the radar pulse dataset, obtains a normalized radar pulse dataset, and constructs a KNN graph of the radar pulse by taking each normalized radar pulse parameter in the radar pulse dataset as a node.

[0111] A SOM anchor point extraction unit assigns a random value between [0, 1] to the weight vector of each neuron in the SOM output layer, and configures the number of iterations; takes the normalized radar pulse dataset as input, iteratively trains the SOM until the number of iterations is completed, and the final neuron weight vector obtained is the extracted SOM anchor point.

[0112] An adaptive anchor graph construction unit calculates the similarity between the SOM anchor points and all nodes to obtain a similarity matrix; and constructs an anchor graph adjacency matrix based on the similarity matrix, i.e., the adaptive anchor graph;

[0113] A fusion graph construction unit improves the stability of the anchor graph adjacency matrix by using the KNN graph adjacency matrix, assigns a weight value to the KNN graph adjacency matrix, and the fusion graph is the sum of the KNN graph adjacency matrix multiplied by the weight value and the anchor graph adjacency matrix;

[0114] A spectral clustering sorting unit constructs a fusion graph adjacency matrix based on the KNN graph adjacency matrix and the anchor graph adjacency matrix, constructs a degree matrix and a Laplacian matrix based on the fusion graph adjacency matrix, calculates a normalized Laplacian matrix, calculates the eigenvectors corresponding to the c smallest eigenvalues of the normalized Laplacian matrix, and uses the eigenvectors as elements to construct a feature matrix, where c is the number of sorted clusters, and each row of the feature matrix represents a sample. K-means clustering is performed on all samples to obtain the final clustering clusters, i.e., the sorting result.

[0115] To facilitate understanding of the technical solutions of the present application, further description will be made below in combination with specific embodiments.

[0116] Taking a four-type multifunctional radar as an example, each type of radar has one to three working modes, and a simulation signal set of the multifunctional radar is configured, which contains three PDW parameters: direction of arrival (DOA), radio frequency (RF), and pulse width (PW). Each type of radar has multiple modulation types for corresponding DOA, RF, and PW, such as conventional, sliding, staggered, and jittering. The same parameter types between different radars or working modes overlap to different degrees. Referring to Table 1, for example, for the three working modes of radar 2, the DOA variation range is consistent, and the PW and RF parameter overlap rates are 40% and 70%, respectively. For radar 2 and radar 3, the RF and DOA parameter overlap rates are 64.29% and 33.33%, respectively.

[0117] Table 1: Multifunctional radar simulation data set

[0118]

[0119] First, a KNN graph of the radar pulse sequence is constructed, and the specific steps are as follows:

[0120] S100. Normalize each radar pulse parameter in the radar pulse data set to [0, 1], and the radar pulse data set can be represented as N and d represent the number of pulses and the feature dimension, respectively. The normalization process is defined by the following formula:

[0121]

[0122] wherein, x represents the i-th pulsei the jth dimension feature of the ith radar pulse.

[0123] S101. Take the normalized radar pulse parameters as nodes, take each node as a graph node, traverse the remaining N-1 nodes, measure the similarity between nodes based on Euclidean distance, find the K nearest neighbors, and construct a KNN graph adjacency matrix A k , wherein the Euclidean distance is defined by the following formula:

[0124]

[0125] wherein x i1 and x i2 represent nodes i1 and i2, respectively;

[0126] S102. Define that the graph node is associated with its K nearest neighbors; that is, set the elements in the corresponding position of the KNN graph adjacency matrix A k to 1, for example, node i1 is associated with node i2, then let a i1i2 = 1, thereby obtaining a directed graph, symmetrizing the KNN graph adjacency matrix A k , let the elements in the symmetric position a i1 = a i2 , construct an undirected graph, and after the above construction process, the radar pulse dataset is converted into an undirected graph structure data, that is, G = (X, A).

[0127] The specific implementation steps of extracting anchor points based on SOM and constructing an adaptive anchor graph based on anchor points are as follows:

[0128] S103. Assuming that the output layer space of SOM has M neurons randomly distributed, the weight vector of each neuron in the output layer is assigned a random value between [0, 1], the weight matrix W = [w1, w2,..., w N ] T , i = 1, 2,..., N, the initial radius N(0) of the neighborhood radius N(t) is determined, the initial learning rate η(0) is determined, 0 < η(0) < 1, and the number of iterations T, wherein t represents the tth iteration;

[0129] S104. Input the normalized radar pulse dataset X, calculate the Euclidean distance d = ||w i - X || 2 , i = 1, 2,..., N between each pulse parameter in the radar pulse dataset and each weight vector;

[0130] S105. Find the neuron with the smallest Euclidean distance d, determine the neuron as the winning neuron j * , The weight vector of the neuron within the center neighborhood radius is corrected with the winning neuron as the center:

[0131]

[0132] wherein j * is the winning neuron, j is the neuron, j≠j * , X is the signal parameter set, W is the weight vector, t is the current iteration number, η(t) is the learning rate of the tth iteration, N g (t) is the neighborhood radius of the tth iteration;

[0133] S106. Update the size N g (t) and the learning rate η(t) of the winning neighborhood radius, the field radius and the learning rate function are selected as exponential functions, specifically:

[0134] N g (t) = INT[N g (0)·exp(-t / (5·T))]

[0135] η(t) = η(0)·exp(-t / (5·T))

[0136] Let t = t + 1, return to step S104 to continue training until t = T, and the training is completed.

[0137] S107. After the training is completed, the weight vectors of the M neurons are finally obtained and used as anchor points, that is, the extraction of the M anchor points is completed.

[0138] The method fully utilizes the function of SOM to capture the topological structure of the input sample space, and can obtain more uniform and representative anchor points.

[0139] Based on the M anchor points, an adaptive anchor graph is constructed, and the specific implementation steps are as follows:

[0140] S108. The similarity between the SOM anchor points and all nodes is calculated using a Gaussian kernel function, and the formula is as follows:

[0141]

[0142] The similarity matrix is constructed as follows:

[0143]

[0144] In the formula, x i represents the ith impulse node, c j represents the jth anchor point, M represents the number of anchor points, w ij represents the similarity between the ith graph node and the jth anchor point, and σ represents a parameter controlling the neighborhood width.

[0145] S109. Constructing an anchor graph adjacency matrix based on a similarity matrix: A a =WΛ -z W T The diagonal matrix Λ is defined as follows:

[0146] S110. Using the KNN graph adjacency matrix A k Improved anchor graph adjacency matrix A a The stability, given A k The weight is α. The KNN graph and the adaptive anchor graph are weighted and fused to obtain the following fused graph:

[0147] A=αA k +A a

[0148] In practice, based on experience, weights are preset in the weight interval [0, 1]. Experiments are conducted to obtain sorting results. Then, the preset weights are increased in increments of 0.1. Experiments are performed on each obtained weight, and the sorting results are compared. The weight value with the best result is selected as given A. k The weight α.

[0149] S111.A is the adjacency matrix of the fused graph. The degree matrix D is constructed as follows:

[0150]

[0151] Degree matrix D∈R N×N Let d be a diagonal matrix. ii Represents the i-th element of the matrix;

[0152] S112. Construct the Laplacian matrix as follows:

[0153] L=DA

[0154] S113. Calculate the standardized Laplacian matrix, defined as follows:

[0155]

[0156] S114. Calculation The eigenvectors corresponding to the c smallest eigenvalues ​​are denoted as the eigenma matrix Q = [q1, q2, ..., qc]. c ]∈R N×k Where c is the number of sorting clusters;

[0157] S115. Using each row of the feature matrix Q to represent a sample, perform K-means clustering on the N samples to obtain clusters.

[0158] The final result of radar signal sorting is given by the result of spectral clustering, containing the cluster label corresponding to each pulse, i.e. the radar category information of each pulse.

[0159] The sorting result obtained by the present application is compared with some existing classical and other advanced clustering methods, including K-Means, DBSCAN, PSO-based K-means (i.e. ABC-K, PSO-K), SOFM and KNN graph-based spectral clustering (KNN-SC), and the comparison result is shown in Table 2:

[0160] Table 2 Experimental results of different methods

[0161]

[0162] As can be seen from the table, the present application achieves the best performance and has almost the highest sorting accuracy for four types of radar signals, with an average sorting accuracy of 93.01%, and the sorting rates for radar 1, radar 2, radar 3 and radar 4 are 88.80%, 87.09%, 100% and 99.93% respectively. Obviously, K-Means, DBSCAN and SOFM perform very poorly in terms of accuracy, with only 66.01%, 69.85% and 66.10%. These three types of methods only rely on the distance and density information of the original data, and are suitable for cases where the density and distance distribution between different data classes are significantly different. However, our data simulates four multi-function radars under the condition of parameter overlap, making the simulation closer to the actual scenario of multi-function radar signal sorting in a complex electromagnetic environment. Compared with KNN-SC, the average sorting accuracy of the present application is improved by 21.08%, and the sorting accuracy for radar 2 and radar 4 is significantly improved, reaching 60.56% and 28.93%.

[0163] Referring to Figure 3 , the sorting method based on K-Means classifies radar 2 with two working modes into two radiation sources, resulting in "batching"; and in the case of overlap between two radar parameters, it completely sorts it into one radar. Referring to Figure 4 , the density clustering DNSCAN finally sorts the four types of radars into only two types of radars; referring to Figure 5 , for the SVM algorithm, there is still a problem that radar 1 is sorted into radar 4 and radar 2 is sorted into radar 3; referring to Figure 6 , the present application sorts the four types of radars almost completely correctly.

[0164] Although the present application has been described in connection with various embodiments thereof, it will be understood that the application is capable of further modifications and that this application is intended to cover any and all such variations, using the scope of the application, which is defined by the appended claims. In the description of the application, the term "comprising" does not exclude other components or steps, and the indefinite article "a" or "an" does not exclude a plurality. A single processor or other unit can fulfill the functions of several items recited in the description. Certain measures can be implemented in different embodiments of the application, and these measures are not necessarily mutually exclusive.

[0165] Although the present application has been described in connection with specific features thereof, it will be evident to an artisan of ordinary skill that various modifications and changes can be made to the application without departing from the spirit and scope thereof. Accordingly, it is intended that the description and drawings be regarded as illustrative rather than restrictive. The scope of the application is indicated by the appended claims, rather than the foregoing description, and all changes that come within the meaning and range of equivalents thereof are intended to be embraced therein.

Claims

1. A radar signal spectrum clustering sorting method based on SOM anchor point extraction and graph fusion, characterized in that, The method comprises the following steps: S1. configuring a radar signal sorting cluster number and a radar pulse parameter, and constructing a radar pulse data set based on the radar pulse parameter; S2. normalizing the radar pulse parameter in the radar pulse data set to obtain a normalized radar pulse data set; S3. taking each normalized radar pulse parameter in the radar pulse data set as a node to construct a KNN graph of the radar pulse; S4. extracting an anchor point of the normalized radar pulse data set based on SOM, and constructing an adaptive anchor graph based on a plurality of anchor points; specifically comprising the following sub-steps: S40. assigning a random value between 0 and 1 to the weight vector of each neuron of the SOM output layer, and configuring an iteration number; S41. taking the normalized radar pulse data set as input to iteratively train the SOM until the iteration number is completed, and the finally obtained neuron weight vector is the extracted SOM anchor point; S42. calculating the similarity between the SOM anchor point and all nodes to obtain a similarity matrix; S43. constructing an anchor graph adjacency matrix based on the similarity matrix, which is the adaptive anchor graph; S5. weighting and fusing the KNN graph and the adaptive anchor graph to obtain a fusion graph; improving the stability of the anchor graph adjacency matrix by using the KNN graph adjacency matrix, assigning a weight value to the KNN graph adjacency matrix, and the fusion graph is the sum of the KNN graph adjacency matrix multiplied by the weight value and the anchor graph adjacency matrix; S6. performing spectral clustering sorting based on the radar signal sorting cluster number and the fusion graph to obtain a sorting result.

2. The radar signal spectrum clustering sorting method based on SOM anchor point extraction and graph fusion according to claim 1, characterized in that, The radar pulse parameter is a direction of arrival, a carrier frequency and a pulse width.

3. The radar signal spectrum clustering and sorting method based on SOM anchor point extraction and graph fusion according to claim 1, characterized in that, The S3 specifically comprises the following sub-steps: S30. setting the number of radar pulse parameters as N, and taking the normalized radar pulse parameter as a node; S31. taking each node as a center node, traversing the remaining N-1 nodes, measuring the similarity between nodes based on Euclidean distance, finding out K nodes most adjacent to the center node, and constructing a KNN graph adjacency matrix; S32. setting the matrix values of each node and its K adjacent nodes in the KNN graph adjacency matrix to 1 to obtain a directed graph adjacency matrix; S33. symmetrizing the directed graph adjacency matrix to construct an undirected graph, which is the KNN graph of the radar pulse.

4. The radar signal spectrum clustering and sorting method based on SOM anchor extraction and graph fusion according to claim 1, characterized in that, The S41 specifically comprises the following sub-steps: S410. configuring a learning rate and a neighborhood radius, inputting the normalized radar pulse dataset to the SOM, and calculating the Euclidean distance between each radar pulse parameter in the radar pulse dataset and each weight vector. t S410. configuring a learning rate and a neighborhood radius, inputting the normalized radar pulse dataset to the SOM, and calculating the Euclidean distance between each radar pulse parameter in the radar pulse dataset and each weight vector. S411. finding out the neuron with the smallest Euclidean distance, which is the winning neuron, and correcting the weight vector of the neuron within the center neighborhood radius based on the learning rate with the winning neuron as the center; S412. if t is less than the iteration number, returning to S410 for the next iteration until t is equal to the iteration number; If t is equal to the iteration number, the SOM neuron weight vector obtained at this time is the extracted anchor point.

5. The radar signal spectrum clustering and sorting method based on SOM anchor point extraction and graph fusion according to claim 4, characterized in that, The initial learning rate is greater than 0 and less than 1, and the learning rate decreases with the iteration number each time.

6. The radar signal spectrum clustering and sorting method based on SOM anchor extraction and graph fusion according to claim 1, characterized in that, The S42 is specifically as follows: The similarity between the SOM anchor point and all nodes is calculated by using a Gaussian kernel function as follows: The similarity matrix is constructed as follows: wherein, denotes the i-th pulse node, denotes the i-th anchor node, denotes the number of anchor nodes, denotes the i-th anchor node, denotes the number of anchor nodes, denotes the similarity between the i-th graph node and the j-th anchor node, denotes the similarity between the i-th graph node and the j-th anchor node, denotes the similarity between the i-th graph node and the j-th anchor node, denotes the similarity between the i-th graph node and the j-th anchor node.

7. The radar signal spectrum clustering and sorting method based on SOM anchor extraction and graph fusion according to claim 3, characterized in that, The S6 specifically comprises the following sub-steps: S60. Constructing a fusion graph adjacency matrix based on the KNN graph adjacency matrix and the anchor graph adjacency matrix; S61. Constructing a degree matrix and a Laplacian matrix based on the fusion graph adjacency matrix, and calculating a normalized Laplacian matrix; S62. Calculating the eigenvectors corresponding to the c smallest eigenvalues of the normalized Laplacian matrix, and using the eigenvectors as elements to construct a feature matrix, wherein c is the number of sorting clusters; S63. Each row of the feature matrix represents a sample, and K-means clustering is performed on all samples to obtain final clustering clusters, which are the sorting results.

8. The radar signal spectrum clustering and sorting method based on SOM anchor extraction and graph fusion according to claim 4, characterized in that, The weight vector of the neuron within the central neighborhood radius is corrected, and the correction is specifically as follows: ； in, For the winning neuron, For neurons, , For signal parameter set, For weight vectors, This represents the current iteration number. For the first t The learning rate for the next iteration. For the first The neighborhood radius of the next iteration.

9. A radar signal spectrum clustering and sorting system based on SOM anchor extraction and graph fusion, characterized in that, The method comprises the following steps: The KNN graph construction unit configures the number of radar signal sorting clusters and radar pulse parameters, constructs a radar pulse dataset based on the radar pulse parameters, normalizes the radar pulse parameters in the radar pulse dataset, obtains the normalized radar pulse dataset, and constructs a KNN graph of the radar pulse by taking each normalized radar pulse parameter in the radar pulse dataset as a node; The SOM anchor extraction unit assigns a random value between 0 and 1 to the weight vector of each neuron in the SOM output layer, and configures the number of iterations; The SOM is iteratively trained with the normalized radar pulse dataset as input until the number of iterations is completed, and the final neuron weight vector obtained is the extracted SOM anchor; The adaptive anchor graph construction unit calculates the similarity between the SOM anchor and all nodes to obtain a similarity matrix, and constructs an anchor graph adjacency matrix based on the similarity matrix, which is an adaptive anchor graph; The fusion graph construction unit improves the stability of the anchor graph adjacency matrix by using the KNN graph adjacency matrix, and assigns weights to the KNN graph adjacency matrix. The fusion graph is the sum of the KNN graph adjacency matrix multiplied by the weights and the anchor graph adjacency matrix; The spectral clustering sorting unit constructs a fusion graph adjacency matrix based on the KNN graph adjacency matrix and the anchor graph adjacency matrix, constructs a degree matrix and a Laplacian matrix based on the fusion graph adjacency matrix, calculates a normalized Laplacian matrix, calculates the eigenvectors corresponding to the c smallest eigenvalues of the normalized Laplacian matrix, and uses the eigenvectors as elements to construct a feature matrix, wherein c is the number of sorting clusters. Each row of the feature matrix represents a sample, and K-means clustering is performed on all samples to obtain final clustering clusters, which are the sorting results.

Images