Signal sorting method and device based on tree diagram and computer equipment

Abstract

The invention relates to a signal sorting method and device based on a tree diagram, and computer equipment, and the method comprises the steps: determining a father-child node relation between pulse data points based on the local density of a pulse signal and the distance measurement of a nearest pulse with a larger local density, thereby constructing a tree diagram, obtaining a clustering center in a self-adaptive manner, and carrying out the self-adaptive clustering of the clustering center; and then distribution of residual pulse data points is carried out, so that automatic clustering and sorting of pulse signals are realized. According to the method, the clustering center can be adaptively estimated, the method is suitable for any cluster shape, non-uniform-density pulse signal sorting is effectively achieved, and the sorting performance is superior to that of an existing method.

Description

Technical Field

[0001] The present invention relates to the technical field of electromagnetic pulse signal sorting, and in particular to a signal sorting method, device and computer equipment based on a dendrogram. Background Art

[0002] Electromagnetic pulse signal sorting is a key link in electronic reconnaissance signal processing. Only by correctly separating multiple randomly interleaved electromagnetic pulse signals into single radiation source pulses can the electronic reconnaissance data be accurately analyzed. Therefore, the accuracy of electromagnetic pulse signal sorting will directly affect the reconnaissance performance of reconnaissance equipment. Electromagnetic pulse signal sorting usually includes two parts: signal pre-sorting and main sorting. Pre-sorting mainly uses parameters such as carrier frequency (CF), direction of arrival (DOA) and pulse width (PW) to cluster, dilute the dense pulse flow, and initially realize signal de-interleaving, laying the foundation for the main sorting of electromagnetic pulse signals. The main sorting is based on the pre-sorting, based on the pulse repetition interval (PRI) feature to de-interleave the pulse sequence. Therefore, good clustering performance is the prerequisite for accurate sorting of electromagnetic pulse signals.

[0003] At present, there are two main categories of pre-sorting algorithms: non-density-based clustering and density-based clustering. Non-density-based clustering algorithms are mainly mainstream algorithms such as mean-class (K-means) clustering algorithm, neural network and deep learning (DL) and their improved applications. However, the performance of the first two methods depends largely on prior knowledge, such as the number of cluster centers, the maximum number of particle swarms, clustering tolerance, the number of output neurons, etc. In addition, the mean-class clustering method only has good clustering performance for spherical clusters, and has poor clustering performance for clusters of arbitrary shapes. For sorting methods based on deep learning, a large amount of data is usually required to train the deep learning model to obtain good performance. Due to the complex and unknown changes in pulse signal parameters, the algorithm is limited in pulse signal sorting. Density-based clustering methods can easily detect clusters of arbitrary shapes. However, density-based clustering algorithms are sensitive to parameters, and it is not easy to select a suitable density threshold or distance threshold. If the threshold is not appropriate, it may lead to reduced clustering efficiency or even clustering failure.

[0004] In response to the problems existing in the above-mentioned pulse signal sorting methods, some researchers have proposed a density-based fuzzy C-means multi-center re-clustering pulse signal sorting algorithm (DFCMRC). This algorithm combines the advantages of the DBSCAN density clustering algorithm and the fuzzy C-means clustering algorithm, overcomes the defect that the traditional algorithm is not suitable for clustering of non-spherical sample sets, and realizes the sorting of pulse signals with unknown target numbers. However, due to the overly simple definitions of local density and distance measurement, the algorithm has poor clustering effect when processing multi-scale pulse radiation source data with non-uniform distribution density. Inspired by DPC and subspace decomposition theory, some researchers have proposed a new adaptive density peak clustering method based on subspace decomposition, namely (SD-ADPC), which is used for pulse signal sorting without any prior knowledge. However, the decomposition of complex pulse signal data based on t-SNE has high computational complexity and is prone to fall into local optimal solutions. In addition, the algorithm has poor adaptability to signals with non-uniform density. Some researchers have proposed a residual graph convolutional network pulse signal sorting method (ResGCN-RSS) based on a semi-supervised learning framework. The algorithm constructs the graph structure of the intercepted pulse signal through the k-nearest neighbor algorithm, designs a three-layer ResGCN to adaptively improve feature learning, and uses the small label graph data of the interlaced signal to effectively and efficiently implement pulse signal sorting through end-to-end ResGCN, effectively improving the generalization ability of the signal sorting model in small data scenarios. However, the algorithm's k-nearest neighbor number takes a fixed value, which makes it difficult to correctly describe the radiation source data with non-uniform distribution density, and the ResGCN model requires known data training, which limits the algorithm in the sorting of pulse signals with complex and unknown parameter changes. Summary of the invention

[0005] Aiming at the problem that the existing pulse signal sorting algorithm is not applicable to the density non-uniform data, a method for solving the problem that the existing pulse signal sorting algorithm is not applicable to the density non-uniform data is provided.

[0006] A signal sorting method based on a dendrogram, the method comprising:

[0007] For each pulse in the signal set to be sorted, the pulse with the largest local density closest to it is taken as its parent node, thereby constructing the signal set to be sorted into a tree structure diagram;

[0008] The cluster center is obtained from the parent node by searching, and each remaining pulse is assigned to the corresponding cluster center according to the tree structure diagram, thereby realizing pulse signal sorting.

[0009] A signal sorting device based on a dendrogram, the device comprising:

[0010] A tree diagram construction module is used for taking the pulse with the largest local density closest to each pulse signal in the signal set to be sorted as its parent node, thereby constructing the signal set to be sorted into a tree structure diagram;

[0011] The signal sorting module is used to obtain the cluster center from the parent node by searching, and assign each remaining pulse to the corresponding cluster center according to the tree structure diagram, so as to realize pulse signal sorting.

[0012] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0013] For each pulse signal in the signal set to be sorted, the pulse with the largest local density closest to it is taken as its parent node, thereby constructing the signal set to be sorted into a tree structure diagram;

[0014] The cluster center is obtained from the parent node by searching, and each remaining pulse is assigned to the corresponding cluster center according to the tree structure diagram, thereby realizing signal sorting.

[0015] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0016] For each pulse signal in the signal set to be sorted, the pulse with the largest local density closest to it is taken as its parent node, thereby constructing the signal set to be sorted into a tree structure diagram;

[0017] The cluster center is obtained from the parent node by searching, and each remaining pulse is assigned to the corresponding cluster center according to the tree structure diagram, thereby realizing pulse signal sorting.

[0018] The above-mentioned signal sorting method, device and computer equipment based on the dendrogram determines the parent-child node relationship between pulse data points based on the local density of the pulse signal and the distance measurement of the nearest pulse with larger local density, thereby constructing a dendrogram, and adaptively obtaining the cluster center, and then allocating the remaining pulse data points, thereby realizing automatic cluster sorting of pulse signals. This method can adaptively estimate the cluster center and is applicable to any cluster shape, effectively realizing the sorting of pulse signals with non-uniform density, and the sorting performance is better than the existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the radiation source data of the unsorted original pulse signal;

[0020] Figure 2Problems in existing pulse signal sorting methods; (a) is the k-means clustering method; (b) is the DBSCAN clustering method;

[0021] Figure 3 Is the clustering result of the density peak clustering algorithm in pulse signal sorting; (a) is the clustering result; (b) is the decision graph;

[0022] Figure 4 Is the flow schematic diagram of a signal sorting method based on dendrogram in an embodiment;

[0023] Figure 5 Is the schematic diagram of the tree structure of the data set; (a) is an example of the data set; (b) is the constructed tree structure;

[0024] Figure 6 Is the schematic diagram for explaining the subordinate relationship between nodes;

[0025] Figure 7 Is the schematic diagram of the algorithm idea of a signal sorting method based on dendrogram;

[0026] Figure 8 Is the schematic diagram of the point distribution of 4 typical data sets;

[0027] Fig. 9 Is the schematic diagram of the processing results of four groups of pulse signal data sets; among them, (a) is the processing result of data set 1, (b) is the processing result of data set 2, (c) is the processing result of data set 3, and (d) is the processing result of data set 4;

[0028] Fig.10 Is the structural block diagram of a signal sorting device based on dendrogram in an embodiment;

[0029] Fig.11 Is the internal structure diagram of a computer device in an embodiment. Detailed implementation manners

[0030] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0031] In pulse signal sorting, the input data is the pulse stream intercepted by an electronic reconnaissance receiver The clustering center is At this time i≠j, and M << N. Where N and M respectively represent the number of input data and pulse classes. Then, the definition of the optimal clustering center is

[0032]

[0033] In the formula, L(·) is the distance measure between the data set and the cluster center. The distance between the PDW of the same reconnaissance receiver and its cluster center is the smallest, and C* represents the optimal cluster center. After obtaining the cluster center, the PDW data in X can be assigned to different classes through distance discrimination, thereby realizing the sorting of pulse signals.

[0034] However, the number of pulse classes is usually unknown, and due to the influence of factors such as the diversity of signal systems and the agility of waveform parameters, the distribution density of radiation source data is uneven and the cluster shape is variable, such as Figure 1 As shown. Figure 1 In the figure, due to the different types of carrier frequency changes, the distribution shapes and densities of radiation sources 1, 2, 3, and 5 all show different characteristics.

[0035] Therefore, the following three problems must be solved when using clustering methods to sort pulse signals: 1) the unknown number of clusters, 2) the uneven data distribution density, and 3) the arbitrary cluster shape. For the above three problems, neither non-density-based clustering methods nor density-based clustering methods can effectively solve them. The processing results of typical algorithms such as k-means and DBSCAN are as follows: Figure 2 shown.

[0036] exist Figure 2 In (a), in addition to the k-means clustering result errors caused by the unknown number of cluster centers, there are also errors caused by uneven distribution density, such as Figure 2 (a) is shown as region 1 and region 2. Figure 2 In (b), although the distribution density of regions 2 and 3 is different, the distribution density within the cluster is relatively constant. The clustering effect of the DBSCAN method is good, but there is an error in the clustering of the data shown in region 1. Moreover, the DBSCAN algorithm is sensitive to parameters such as distance and density threshold. The clustering results of different threshold settings are quite different and cannot be automatically clustered.

[0037] The clustering method based on density peak can adapt to data with different distribution shapes and does not require prior information such as the number of cluster centers. It has great potential in improving the signal sorting of transmitter radiation sources. The basic idea of ​​this method is that the local density of the cluster center is higher than that of its neighboring centers, and the distance to other centers is relatively large. Then the density peak node can be distinguished from other non-density peak nodes based on the above two characteristics. To this end, the algorithm defines two quantities, namely the local density f(i) of the sample point and the distance δ(i) to the higher density point, as shown in formula (2)

[0038]

[0039] Where, d ij Indicates the distance between the i-th sample point and the j-th sample point, d c represents the cutoff distance. When t<0, χ(t)=1, and when t is other values, χ(t)=0. First, a decision diagram is established by calculating f(i) and δ(i), as shown in Figure 3 As shown in (b), the ideal cluster center is selected according to the maximum criterion of f(i) and δ(i); then, based on the information obtained from the calculated value of δ(i), each remaining point will be assigned to the cluster to which its denser neighbor belongs, and finally clustering is completed.

[0040] Although the experimental results of DPC show that it can play a good role in many cases, it also has obvious disadvantages when the data distribution density is uneven, such as Figure 3 shown.

[0041] Figure 3 (a) shows the clustering result of the DPC algorithm in pulse signal sorting, where the density is calculated using the Gaussian kernel distance formula. Figure 3 As shown in (a), there are obviously four clusters with different densities and shapes in the data set. The density calculation formula based on distance cannot be applied to all clusters, resulting in classification errors.

[0042] Figure 3 (b) shows the decision graph composed of the f(i) and δ(i) values ​​of each data point. The data points with relatively large f(i) and δ(i) values ​​are the cluster centers. It can be seen from the figure that the f(i) values ​​of most data points in the high-density clusters represented by points A and B are larger than the data points in the low-density clusters, including the cluster centers of the low-density clusters, such as data points D and E. At the same time, it can be observed that the δ(i) value of the cluster represented by data point E is smaller than that of most data points, so the cluster centers of the clusters represented by data points D and E cannot be found.

[0043] From the above analysis, it can be seen that the density calculation formula based on distance cannot select the correct cluster center for a data set with variable density. This is because the distance and density measurement results are unreasonable without considering the influence of the neighborhood of each point, and low-density clusters usually have very small f(i) values. Even if the δ(i) value is large, it cannot effectively improve their status in the decision graph. Therefore, DPC can easily select the wrong cluster center, resulting in incorrect clustering results. In addition, points in high-density clusters often have higher f(i) values ​​because the branches are generally far apart, and these points are more likely to be selected as cluster centers than points in low-density clusters. Therefore, the algorithm can easily select the wrong point as the cluster center, resulting in incorrect clustering results. It can be seen that improving the status of low-density cluster centers in the decision graph and adaptively selecting cluster centers are the key to the success of clustering.

[0044] In one embodiment, to solve the above problem, Figure 4 As shown, a signal sorting method based on a dendrogram is provided, comprising the following steps:

[0045] Step 402: For each pulse in the signal set to be sorted, the pulse with the closest local density and the largest local density is used as its parent node, thereby constructing the signal set to be sorted into a tree structure diagram.

[0046] Step 404, obtain the cluster center from the parent node by searching, and assign each remaining pulse to the corresponding cluster center according to the tree structure diagram, so as to realize pulse signal sorting.

[0047] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0048] In one embodiment, the local density of the pulse signal is calculated as follows:

[0049] 1) The common k nearest neighbors of the first pulse and the second pulse are taken as the shared nearest neighbors of the two. Specifically, for any point pdw in the data set X i and pdw j , click pdwi The k nearest neighbor set is NN k (pdw i ), click pdw j The k nearest neighbor set is NN k (pdw j ), click pdw i and pdw j The shared neighbors of are their common neighbors, expressed as

[0050] SNN(pdw i ,pdw j )=NN k (pdw i )∩NN k (pdw j ) (3)

[0051] 2) When the first pulse and the second pulse are in their shared neighbors, the shared neighbor similarity between them is calculated. Specifically, for any point pdw in the data set X, i and pdw j , their SNN similarity is defined as

[0052]

[0053] Where d is the point pdw i and pdw j The distance between the two points is calculated only when the two points are in the SNN set. Otherwise, the point pdw i and pdw j The SNN similarity between them is 0.

[0054] The non-zero part of (4) can be written as

[0055]

[0056] The meaning of SNN similarity can be clearly seen from formula (5).

[0057] The left side of the formula represents pdw i and pdw j The number of shared neighbors, the right part is pdw i and pdw j The inverse of the mean distance from a point to all shared neighbors represents the density around the two points to a certain extent. By simultaneously examining the shared neighborhood and density of two points, SNN similarity can better adapt to various environments.

[0058] 3) Calculate the sum of the shared neighbor similarities of the k pulses with the highest shared neighbor similarity to the first pulse as the local density of the first pulse.

[0059] Assume point pdw i is any point in the data set X, Γ(i)={pdw i1 ,pdw i2 ,...,pdw ik} is with point pdw i The set of k points with the highest similarity. Then, point pdw i The local density is defined as the point pdw i The sum of the similarities of the k points with the highest similarity. The local density formula is as follows:

[0060]

[0061] From equations (3) to (6), we can see that the local density f(i) at the i-th point has the following properties: 1) When the k value is large, the point pdw i and one of its shared neighbors pdw j When the number of shared neighbors is large, the local density is large, and when the k value is small, the local density is small; therefore, k can reflect the local density of the i-th point. 2) Number of shared neighbors|SNN(pdw i ,pdw j )| is a constant, if point pdw i and pdw j The smaller the distance to each of their shared neighbor points, the larger the local density of the i-th point; thus, the closer the spatial distance is, the larger the value of f(i). 3) Point pdw i and pdw j When they belong to the same class, they share many neighbors and the local density is large. That is, if pdw i Most of the surrounding points belong to the same cluster, then pdw i Obviously, points belonging to the same cluster contribute more to f(i).

[0062] It can be seen that the local density definition shown in formula (6) has good adaptability to density non-uniform data. It not only utilizes the distance information of data points, but also can describe the structural information of clustering. Compared with the definition shown in formula (2), it can better reveal the intrinsic relationship between points.

[0063] In one embodiment, for each pulse signal in the to-be-sorted signal set, the pulse with the closest local density and the largest distance thereto is taken as its parent node, including:

[0064] For the first pulse, find the second pulse with larger local density; minimize the distance between the first pulse and the second pulse multiplied by the distances of the first pulse and the second pulse to their nearest neighbors respectively, obtain the pulse with larger local density closest to the first pulse, and use it as the parent node of the first pulse.

[0065] Assume point pdw i is any point in the data set X, find a local density greater than pdw i Point pdw j Then, minimize pdw i and pdw j Multiply the distance between by pdw i and pdw j The distance to their nearest neighbor; take this as the point pdw i The delta value is:

[0066]

[0067] As shown in formula (7), the δ value of the point with the highest local density is the maximum value among other points, that is,

[0068]

[0069] From (7) to (8), we can see that the δ value has the following characteristics: for low-density clusters, the k value has a significant impact on the value of δ, and due to pdw i and pdw j The neighbors of the points are far away from them, and the value of δ is large; while for high-density clusters, the k value has a slight effect on the value of δ, and the value of δ is small. The distance δ from the points with higher density not only utilizes the distance information, but also considers the neighbor information of each point, which compensates for the δ value of the points in the low-density clusters and enhances the fairness of the δ value.

[0070] Assume point pdw i is any point in the data set X, data point pdw i is recorded as a node; as shown in formula (7), if the point pdw j Is the density greater than pdw i The closest point to the j For pdw i The parent node of pdw i For pdw j It can be seen that δ can not only describe the distance between data points, but also the density relationship between data points. Then, according to the calculation results of (6) to (7), the points in the data set X to be sorted can be constructed into a tree structure, such as Figure 5 shown.

[0071] Each node of the tree represents a data point. Except for the root node, each node has a unique parent node. Each parent node is the closest point with higher density among all its child nodes. The root node is the point with the highest density, which is the result of formula (8). Each node is the closest point with higher density among all its child nodes. For example, point 13 is the root node because it has the highest density among all points; points 5, 7, and 8 are the child nodes of point 6 because point 6 is the closest point with higher density among these three points.

[0072] like Figure 5 As shown in the figure, each cluster of data corresponds to a subgraph. The tree graph is composed of multiple subgraphs, and each subgraph has its own structure and unique root node. Therefore, the following conclusions can be drawn: (1) The root node of each subgraph can be used as the cluster center of the cluster data, and other data in the cluster are connected to each other according to the subgraph structure; (2) The cluster center set C is a subset of the set P composed of all parent nodes. For example, in the tree structure, the density of points 3, 6, 9, 13, 14, and 16 is higher than that of other data points. They are the parent nodes of other points and form a tree graph subgraph with other points. Among them, each subgraph represented by points 3, 6, 9, and 13 represents a cluster, and the number of data points with the highest density in the cluster is equal to the number of subgraphs. Therefore, the data point with the highest density can be used as the representative of the cluster, which is called the local density peak.

[0073] From this, we can see that the cluster center can be obtained from the parent node of the data point through a certain search strategy. After finding the cluster center, we assign each remaining point to the corresponding cluster center according to the tree diagram, and then we can complete the clustering of the pulse radiation source data and realize signal sorting.

[0074] Because of the collection To obtain the cluster center, the data points in the set P must be distinguished according to certain decision criteria.

[0075] In one embodiment, obtaining the cluster center from the parent node by searching includes:

[0076] 1) For a parent node, the union of all its child nodes’ k nearest neighbors is taken as its corresponding parent node’s nearest neighbors. Specifically, suppose point pdw i is any parent node in the data set X, G is the set of child nodes of the parent node, and the parent node neighbors are defined as the union of all the child nodes’ k neighbors, that is

[0077]

[0078] 2) If the first parent node is the parent node of the second parent node, then the second parent node is a new cluster center if and only if the number of shared neighbors between the second parent node and the parent node neighbors of the first parent node does not exceed a preset threshold, otherwise, the second parent node is a subordinate parent node of the first parent node.

[0079] Assume that parent nodes A, B∈P, and A is the parent node of B; then, B is a subordinate parent node of A if and only if B satisfies the following conditions, that is,

[0080] |PNN(A)∩PNN(B)|≥th (10)

[0081] th represents a threshold related to the number of data points. That is, if point A and point B are considered to belong to the same cluster, then their respective parent node neighborhoods have at least a certain number of neighbors shared with each other.

[0082] Assume that the parent node A, B∈P, and A is the parent node of B; then, if and only if B satisfies the following conditions, it is a new cluster center point, that is,

[0083] |PNN(A)∩PNN(B)|≤th (11)

[0084] That is to say, if point A and point B are considered to be the root nodes of different clusters, then there are very few or even no certain number of neighbors shared with each other in the neighborhood of their respective parent nodes. Therefore, in an ideal situation, the clusters represented by A and B are completely disjoint, and the value of the threshold th is zero; considering the proximity of the clusters represented by A and B and the presence of noise points, this embodiment selects the threshold th as 1 / 4 of the number of neighbors of the parent node of point B.

[0085] The decision formulas shown in equations (10) and (11) can be obtained by Figure 6 Explain intuitively.

[0086] The two circles in the upper right corner of the figure represent the parent node neighborhoods of points 13 and 14, and the other three circles represent the parent node neighborhoods of points 3, 6, and 9. The intersection of the neighborhoods of 13 and 14 is not an empty set, which indicates that 13 and 14 are very close and can be considered as the same cluster; while the intersection of the neighborhoods of points 3, 6, and 9, as well as the intersection of the neighborhoods of points 3, 6, 9 and points 13 and 14 are all empty sets, which indicates that they are far apart and are not in the same cluster.

[0087] In one embodiment, the cluster center is obtained from the parent node by searching, and each remaining pulse is assigned to the corresponding cluster center according to the tree structure diagram, including:

[0088] Search cluster centers from all parent nodes, and then decompose the constructed tree structure graph into several subgraphs;

[0089] According to the connection relationship of the pulses in the subgraph, the remaining pulses are assigned to each cluster center to form a cluster.

[0090] In summary, the algorithm ideas of this application are as follows Figure 7 As shown in the figure, the algorithm is roughly divided into four stages: In the first stage, the density value f(i) of the data point and the distance δ(i) to the nearest larger density point are adaptively obtained based on the natural nearest neighbor and the shared nearest neighbor. In the second stage, the points in the data set X to be sorted are constructed into a tree according to the values ​​of f(i) and δ(i). In the third stage, the threshold th is determined, and the cluster center point is searched from all parent nodes according to the decision formula shown in equations (12) to (13), and the constructed tree diagram is decomposed into several subgraphs. In the fourth stage, data points are assigned to each cluster center point according to the connection relationship of the subgraph to form a cluster cluster. The specific implementation of the key algorithm is shown in Algorithms 1, 2, 3, and 4.

[0091] The first stage: Based on the natural nearest neighbor and shared nearest neighbor, the density value f(i) of the data point and the distance δ(i) to the nearest larger density point are adaptively obtained. The specific process of obtaining the natural eigenvalue is detailed in Table 1.

[0092] Table 1 Specific process of Algorithm 1

[0093]

[0094]

[0095] Among them, NN r (pdw j ) is pdw i For a data point pdw in the dataset X i , if pdw i In pdw j Within the r-neighborhood of , and for any other point pdw in the data set m , there is at least one data point whose r-neighborhood contains pdw m , then the data point pdw j is the data point pdw i The natural nearest neighbor of .

[0096] Data point pdw i The reverse nearest neighbor of the dataset X is the set of all nearest neighbors that contains the data point pdw i A set of data points, namely

[0097] RNN r (pdw i )={pdw j ∈X|pdw i ∈NN r (pdwj )} (12)

[0098] The natural eigenvalue of a data set is defined as making any data point pdw i are at least one other data point pdw j (pdw j ≠pdw i )’s r-neighborhood contains the minimum r value, which is mathematically defined as follows:

[0099]

[0100] In the formula, NN r (pdw j ) represents the data point pdw j The r-neighborhood of , supk is the natural eigenvalue. That is, in the process of forming natural adjacencies, when the number of points without reverse adjacencies remains unchanged, the adjacency search range r is the natural eigenvalue, represented by k.

[0101] The second stage: According to the values ​​of f(i) and δ(i), the specific process of constructing the data set tree diagram is shown in Table 2.

[0102] Table 2 Specific process of Algorithm 2

[0103]

[0104]

[0105] The third stage: The specific process of obtaining the root node of the subtree graph is detailed in Table 3 Algorithm 3.

[0106] Table 3 Specific process of Algorithm 3

[0107]

[0108]

[0109] Stage 4: According to the results of the above calculations, the specific process of the pulse sorting algorithm based on adaptive peak density clustering is detailed in Table 4 Algorithm 4.

[0110] Table 4 Specific process of Algorithm 4

[0111]

[0112]

[0113] In summary, this application introduces the concepts of natural neighbors and shared neighbors (Shared Nearest Neighbors, SNN), improves the definitions of distance, local density f(i), and the distance between local density f(i) and the nearest larger density point δ(i) to deal with problems caused by non-uniform density; and based on the new definitions of f(i) and δ(i), an adaptive clustering center selection algorithm based on a dendrogram is proposed to achieve automatic sorting. Since the f(i) and δ(i) defined by the SNNN-ADPC (Adaptive Density PeakClustering Algorithm Based on Shared Natural Nearest Neighbors) algorithm reflect both the local characteristics of the data set and the natural structure of the data, and can adaptively select clustering centers and allocate data points based on the dendrogram subgraph structure. Therefore, the SNNN-ADPC algorithm can improve the clustering results of signal radiation source data with uneven distribution density.

[0114] In order to verify and evaluate the performance of the algorithm in this paper, synthetic datasets of various shapes and non-uniform density clustering are synthesized according to the dataset construction parameter settings shown in Table 1. The point distributions of the four typical datasets are as follows: Figure 8 This section conducts four sets of experiments on a synthetic dataset containing clusters of various shapes and non-uniform density, and compares our method with DFCMRC, SD-ADPC, and ResGNN-RSS algorithms.

[0115] Table-1 Dataset setting parameters

[0116]

[0117]

[0118] Note: [·] indicates a set of numerical points, and ~ indicates a numerical range.

[0119] The algorithm in this paper is used together with DFCMRC, SD-ADPC, and ResGNN-RSS algorithms to process the four sets of synthesized data respectively. The k-nearest neighbor value of the ResGNN-RSS algorithm is set to 10. The results are shown in Fig. 9 shown.

[0120] contrast Figure 8It can be seen that in the four groups of experiments, the DFCMRC algorithm has errors in clustering radar 1, 3, 7, 8, 11, 12, and 13; among them, the frequencies of radar 1, 12, and 13 signals jump at several frequency values, and the density distribution of data points is locally clustered and overall discrete. The DFCMRC algorithm is affected in the process of predicting cluster centers and has poor adaptability to such distributions, making the predicted number of cluster centers more than the actual number, resulting in signal sorting and batch addition; radar 3 and 8, 7 and 8 have different distribution densities, and the distribution of clusters is too close. The DFCMRC algorithm will treat the two as the same cluster in the process of predicting cluster centers, making the predicted number of cluster centers less than the actual number, resulting in signal sorting and batch addition; radar 11 has a large data distribution range and a small distribution density, so the algorithm regards it as multiple data, resulting in signal sorting and batch addition.

[0121] In the four experiments, the SD-ADPC algorithm made mistakes in clustering radars 1, 3, 4, 7, 8, 11, 12, and 13. Among them, although radars 1 and 4 in dataset 1 have different density distributions, they are incorrectly clustered due to batch addition. This is mainly because the SD-ADPC algorithm uses a method based on differential factor outlier detection to estimate the cluster center. This method makes data points with high density more likely to become cluster centers. Therefore, the probability of errors in the algorithm in estimating cluster centers of datasets with various density distributions is greatly increased, resulting in errors in the final clustering results. For example, in dataset 2, radars 6 and 7 are partially clustered into the same cluster, radar2 and radar 1 are partially clustered into the same cluster, radars 11 and 12 in dataset 3 are divided into 4 clusters, and radars 11 and 12 are incorrectly classified in dataset 4. These are all examples of poor robustness of the algorithm under uneven density distribution conditions.

[0122] In the four groups of experiments, both the ResGNN-RSS algorithm and the algorithm in this paper can correctly cluster the data sets, and the overall clustering performance of the algorithms is better than that of the DFCMRC algorithm and the SD-ADPC algorithm. However, since the ResGNN-RSS algorithm is a semi-supervised learning algorithm and cannot adaptively determine the k-nearest neighbor value, it is limited when applied to the sorting of unknown pulse signals; while the algorithm in this paper adopts a k-nearest neighbor value search algorithm based on natural neighbors, which is more suitable for the sorting of unknown pulse signals. From the above analysis, it can be seen that the algorithm in this paper has a good effect when sorting pulse signals with uneven density distribution. In one embodiment, Fig.10 As shown, a signal sorting device based on a dendrogram is provided, comprising: a dendrogram building module and a signal sorting module, wherein:

[0123] A tree diagram construction module is used for taking the pulse with the largest local density closest to each pulse signal in the signal set to be sorted as its parent node, thereby constructing the signal set to be sorted into a tree structure diagram;

[0124] The signal sorting module is used to obtain the cluster center from the parent node by searching, and assign each remaining pulse to the corresponding cluster center according to the tree structure diagram, so as to realize pulse signal sorting.

[0125] For the specific definition of a signal sorting device based on a dendrogram, please refer to the definition of a signal sorting method based on a dendrogram above, which will not be repeated here. Each module in the above-mentioned signal sorting device based on a dendrogram can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0126] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Fig.11 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data such as pulse streams. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a signal sorting method based on a dendrogram is implemented.

[0127] Those skilled in the art will understand that Fig.11 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0128] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method in the above embodiment when executing the computer program.

[0129] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method in the above embodiment are implemented.

[0130] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0131] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0132] The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A signal sorting method based on a dendrogram, characterized in that: The method comprises: For each electromagnetic pulse in the signal set to be sorted, the pulse with the largest local density closest to it is taken as its parent node, thereby constructing the signal set to be sorted into a tree structure diagram; The cluster center is obtained from the parent node by searching, and each remaining pulse is assigned to the corresponding cluster center according to the tree structure diagram, thereby realizing signal sorting.

2. The method according to claim 1, characterized in that Obtain the cluster center from the parent node by searching, including: For a parent node, the union of all its child nodes’ k nearest neighbors is taken as its corresponding parent node’s nearest neighbors; If the first parent node is the parent node of the second parent node, then the second parent node is a new cluster center if and only if the number of shared neighbors between the second parent node and the parent node neighbors of the first parent node does not exceed a preset threshold, otherwise, the second parent node is a subordinate parent node of the first parent node.

3. The method according to claim 2, characterized in that The preset threshold is determined according to the number of neighbors of the second parent node.

4. The method according to claim 1, characterized in that: The cluster center is obtained from the parent node by searching, and each remaining pulse is assigned to the corresponding cluster center according to the tree structure diagram, including: Search cluster centers from all parent nodes, and then decompose the constructed tree structure graph into several subgraphs; According to the connection relationship of the pulses in the subgraph, the remaining pulses are assigned to each cluster center to form a cluster.

5. The method according to claim 1, characterized in that: The local density of the pulse is calculated as follows: The common k nearest neighbors of the first pulse and the second pulse are regarded as the shared nearest neighbors of the two; When the first pulse and the second pulse are in their shared neighbors, the shared neighbor similarity between the two is calculated; The sum of the shared neighbor similarities of the k pulses with the highest shared neighbor similarity to the first pulse is calculated as the local density of the first pulse.

6. The method according to claim 1, characterized in that For each pulse in the signal set to be sorted, the pulse with the largest local density closest to it is taken as its parent node, including: For the first pulse, look for the second pulse with a larger local density; Minimize the distance between the first pulse and the second pulse multiplied by the distances of the first pulse and the second pulse to their nearest neighbors, obtain the pulse with the closest local density to the first pulse, and use it as the parent node of the first pulse.

7. A signal sorting device based on a dendrogram, characterized in that: The device comprises: A tree diagram construction module is used for taking, for each pulse in the to-be-sorted signal set, the pulse with the largest local density closest to it as its parent node, thereby constructing the to-be-sorted signal set into a tree structure diagram; The signal sorting module is used to obtain the cluster center from the parent node by searching, and assign each remaining pulse to the corresponding cluster center according to the tree structure diagram, so as to realize signal sorting.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images