Electromagnetic signal clustering method and system

Abstract

The application relates to an electromagnetic signal clustering method and system, and belongs to the technical field of electromagnetic signal processing, and solves the problem of difficult electromagnetic signal sorting and pairing in the prior art. The method comprises the following steps: receiving a plurality of electromagnetic signals, extracting a plurality of multilevel semantic features based on a residual network model; performing multiple iterations according to the following steps to obtain a clustering result of each iteration: randomly dividing the plurality of multilevel semantic features into a plurality of sample sets; determining a second clustering result corresponding to each sample set; taking the plurality of second clustering results corresponding to the plurality of sample sets as third clustering results of the current iteration; and determining a fourth clustering result of the plurality of multilevel semantic features by using a voting mechanism according to a plurality of third clustering results obtained through multiple iterations. The electromagnetic signal is accurately clustered.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electromagnetic signal processing, and in particular to an electromagnetic signal clustering method and system. BACKGROUND

[0002] In the process of platform passive detection of electromagnetic signals, when multiple detection platforms simultaneously detect electromagnetic signals emitted by multiple same type transmitters, and there is no coded transmitter identity information in the electromagnetic signals, it is very difficult to sort and pair the electromagnetic signals, which directly affects the electromagnetic signal identification effect. Therefore, how to realize the clustering of detected electromagnetic signals under low signal-to-noise ratio conditions has become a popular research topic at home and abroad.

[0003] For the clustering problem of detected electromagnetic signals, scholars have proposed k-means clustering algorithm, FCM mean clustering algorithm, k-medoids center point clustering algorithm, etc. The k-means clustering algorithm has a simple framework, is easy to implement, and has fast convergence speed, but is sensitive to noise and isolated points, and the clustering effect depends on the selection of the initial clustering center. The FCM mean clustering algorithm has a good clustering result, but also needs to determine the number of clustering categories, the membership function that meets the condition, and the iteration stop threshold. In order to eliminate the influence of isolated points, scholars have also proposed the k-medoids center point clustering algorithm, but the calculation amount is relatively large.

[0004] Traditional clustering methods have poor clustering performance for low signal-to-noise ratio electromagnetic signals, and have the following shortcomings:

[0005] 1. Difficulty in feature extraction of low signal-to-noise ratio electromagnetic signals;

[0006] 2. The selection of the initial clustering center directly affects the clustering performance and convergence speed, which is set by artificial setting and cannot be determined in real time according to the electromagnetic signals;

[0007] 3. For low signal-to-noise ratio electromagnetic signals, even if ensemble learning is used, counterintuitive results may occur due to the performance of a weak clusterer. SUMMARY

[0008] In view of the above analysis, the embodiments of the present application aim to provide an electromagnetic signal clustering method and system to solve the problem of difficult electromagnetic signal sorting and pairing in the prior art.

[0009] In one aspect, the embodiments of the present application provide an electromagnetic signal clustering method, which comprises:

[0010] Receiving a plurality of electromagnetic signals, extracting a plurality of multi-level semantic features corresponding to each electromagnetic signal based on a residual network model, and obtaining a plurality of multi-level semantic features;

[0011] The following steps are performed to iterate multiple times and obtain the clustering results for each iteration: the plurality of multi-level semantic features are randomly divided into multiple sample sets; for each sample set, multiple clustering methods are used to cluster them, and the first clustering result corresponding to each clustering method is obtained; based on the multiple first clustering results corresponding to the multiple clustering methods, a voting mechanism is used to determine the second clustering result corresponding to each sample set; the multiple second clustering results corresponding to the multiple sample sets are used as the third clustering result of this iteration;

[0012] Based on the multiple third clustering results obtained through multiple iterations, a voting mechanism is used to determine the fourth clustering result of the several multi-level semantic features, and the fourth clustering result is used as the clustering result of the several electromagnetic signals corresponding to the several multi-level semantic features.

[0013] A further improvement to the above method includes: determining the initial cluster centers of the plurality of multi-level semantic features based on granular computation;

[0014] For each sample set, multiple clustering methods are used for clustering, and the first clustering result corresponding to each clustering method includes:

[0015] For each sample set, based on the initial cluster centers, multiple clustering methods are used to perform clustering, resulting in the first clustering result corresponding to each clustering method.

[0016] Based on further improvements to the above method, the residual network model includes a preprocessing layer, a residual block layer, and a multi-scale feature fusion layer;

[0017] The extraction of multi-level semantic features corresponding to each electromagnetic signal based on the residual network model yields several multi-level semantic features, including:

[0018] The preprocessing layer processes each electromagnetic signal into a feature map of a preset format; the residual block layer generates multiple predicted features based on the feature map of the preset format; and the multi-scale feature fusion layer fuses the multiple predicted features to form a multi-level semantic feature corresponding to each electromagnetic signal.

[0019] Based on a further improvement of the above method, the residual block layer includes multiple residual blocks connected sequentially from high to low dimension level, and a number of residual blocks with lower dimension level among the multiple residual blocks are used as preset residual blocks;

[0020] The residual block layer generates multiple predicted features based on the feature map in the preset format, including:

[0021] The feature map in the preset format is input into the residual block layer and passes through the multiple residual blocks in sequence;

[0022] Upsample the output of each preset residual block to generate a first feature map;

[0023] Obtain a second feature map of an output of a neighboring high-dimension hierarchical residual block corresponding to the preset residual block;

[0024] Fuse the first feature map and the second feature map to form a prediction feature corresponding to each preset residual block.

[0025] Based on the further improvement of the above method, the determining of the clustering initial center of the plurality of multi-level semantic features based on the granular computing comprises:

[0026] Calculating the similarity between the plurality of multi-level semantic features;

[0027] According to the similarity and the similarity threshold, a fuzzy similarity matrix between the plurality of multi-level semantic features is established, and similar features of each multi-level semantic feature are obtained based on the fuzzy similarity matrix. Each multi-level semantic feature and the corresponding similar feature set are taken as a particle. Based on the plurality of multi-level semantic features, a plurality of particles are obtained.

[0028] Calculating the particle density and the average particle density of the plurality of particles, and determining an effective particle set according to the particle density and the average particle density;

[0029] Determining the clustering initial center of the plurality of multi-level semantic features according to the effective particle set.

[0030] Based on the further improvement of the above method, the calculating of the similarity between the plurality of multi-level semantic features comprises:

[0031] The similarity is defined as:

[0032]

[0033] wherein, x i represents the i-th multi-level semantic feature; x j represents the j-th multi-level semantic feature; S(x i ,x j ) represents the similarity between x i and x j ; ω l represents the attribute resolution ability of the l-th attribute of the multi-level semantic feature; B represents the attribute set of the multi-level semantic feature, and |B| represents the number of elements included in B; x il represents the attribute value of the l-th attribute of x i ; x jl represents the attribute value of the l-th attribute of x j ;

[0034] ωl is defined as:

[0035]

[0036] wherein the attribute I is divided into L interval blocks {y1, y2, …, y r ,…,y L}, y r represents the rth interval block in the L interval blocks; N represents the number of the plurality of multi-level semantic features.

[0037] Based on the further improvement of the above method, the similarity threshold is determined by the following way:

[0038] The average similarity of the plurality of multi-level semantic features is calculated.

[0039] wherein the average similarity is defined as:

[0040]

[0041] The similarity threshold is determined according to the average similarity, and the range of the similarity threshold is

[0042] Based on the further improvement of the above method, the calculation of the particle density of the plurality of particles and the average particle density, and the determination of the effective particle set according to the particle density and the average particle density include:

[0043] The particle density of each particle is calculated, and the particle density is defined as:

[0044] gd(X n ) = |X n | / N;

[0045] wherein the total number of particles is the same as the number of multi-level semantic features, which is N, and the plurality of particles is represented as {X1, X2, …, X n ,…,X N}, X n represents the nth particle in the N particles, |X n | represents the number of multi-level semantic features included in the nth particle; and gd(X n ) represents the particle density of the nth particle.

[0046] The average particle density of the plurality of particles is calculated, and the average particle density is defined as:

[0047]

[0048] wherein, represents the average particle density of the plurality of particles.

[0049] The particles in the effective particle set are composed of the particles in the plurality of particles with a particle density greater than or equal to an average particle density.

[0050] Based on the further improvement of the above method, the determining the initial cluster centers of the plurality of multi-level semantic features according to the effective particle set comprises:

[0051] The particle center of each particle in the effective particle set is calculated, and the particle center is defined as:

[0052]

[0053] wherein C q represents the particle center of the qth particle in the effective particle set, the qth particle including Q objects, each object being a multi-level semantic feature; x qk1 represents the multi-level semantic feature of the k1th object in the Q objects, x qk2 represents the multi-level semantic feature of the k2th object in the Q objects.

[0054] The particles are sorted according to the particle density of each particle; the particle center of the particle with the maximum particle density in the effective particle set is taken as the first cluster center; and the particle with the maximum particle density and the farthest Euclidean distance from the particle center of the particle with the maximum particle density is taken as the second cluster center.

[0055] The Euclidean distance between each remaining particle center in the remaining particles in the effective particle set and the determined cluster center is calculated, and the minimum value of the above Euclidean distance is taken as the cluster distance of the remaining particle; the cluster distances of the remaining particles are compared, and the particle center of the particle corresponding to the maximum cluster distance is taken as the next cluster center, and the step is repeated until the maximum cluster distance does not satisfy the set cluster distance threshold condition.

[0056] The obtained cluster centers are taken as the initial cluster centers of the plurality of multi-level semantic features.

[0057] In another aspect, an embodiment of the present application provides an electromagnetic signal clustering system, characterized in that the system comprises:

[0058] A signal feature extraction module is configured to receive a plurality of electromagnetic signals, extract the multi-level semantic features corresponding to each electromagnetic signal based on a residual network model, and obtain a plurality of multi-level semantic features.

[0059] The multiple iteration execution module is configured to perform multiple iterations according to the following steps: randomly dividing the multiple hierarchical semantic features into multiple sample sets; for each sample set, performing clustering by using multiple clustering methods to obtain a first clustering result corresponding to each clustering method; determining a second clustering result corresponding to each sample set by using a voting mechanism according to the multiple first clustering results corresponding to the multiple clustering methods; and taking the multiple second clustering results corresponding to the multiple sample sets as a third clustering result of the current iteration;

[0060] The clustering result determination module is configured to determine a fourth clustering result of the multiple hierarchical semantic features by using a voting mechanism according to the multiple third clustering results obtained through multiple iterations, and take the fourth clustering result as the clustering result of the multiple electromagnetic signals corresponding to the multiple hierarchical semantic features.

[0061] Compared with the prior art, the present application can achieve at least one of the following beneficial effects:

[0062] 1. The residual network model is used to extract the multiple hierarchical semantic features corresponding to each electromagnetic signal, thereby improving the feature extraction of the electromagnetic signal with low signal-to-noise ratio;

[0063] 2. The granular computing is used to determine the initial clustering center of the multiple hierarchical semantic features, thereby improving the accuracy and applicability of the artificial selection;

[0064] 3. The dual integration is used to improve the counter-intuitive problem that may occur in the traditional integration, thereby further improving the clustering accuracy of the electromagnetic signal.

[0065] The above technical solutions can be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present application will be described in the subsequent specification, and some advantages will become apparent from the specification, or will be understood by implementing the present application. The purpose and other advantages of the present application can be achieved and obtained from the contents specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0066] The accompanying drawings are included to provide a further understanding of the present application, and are incorporated herein and constitute a part of the specification. The drawings illustrate embodiments of the present application and, together with the description, serve to explain the principles of the present application. In the drawings:

[0067] Figure 1 A flowchart of an electromagnetic signal clustering method provided by an embodiment of the present application;

[0068] Figure 2 A structural diagram of an electromagnetic signal clustering method provided by an embodiment of the present application;

[0069] Figure 3A structural schematic diagram of an electromagnetic signal clustering method provided by an embodiment of the present application is shown in FIG. 2.

[0070] Figure 4 A residual network model schematic diagram provided by an embodiment of the present application is shown in FIG. 3.

[0071] Figure 5 A multi-level semantic feature schematic diagram provided by an embodiment of the present application is shown in FIG. 4.

[0072] Figure 6 A structural schematic diagram of an electromagnetic signal clustering system provided by an embodiment of the present application is shown in FIG. 5. DETAILED DESCRIPTION

[0073] The preferred embodiments of the present application will be described in detail below with reference to the accompanying drawings, which form a part of this application. The accompanying drawings and the detailed description below are used to illustrate the preferred embodiments of the present application, and are not intended to limit the scope of the present application.

[0074] One specific embodiment of the present application discloses an electromagnetic signal clustering method, as shown in FIG. 1, which comprises the following steps. Figure 1

[0075] S101: receiving a plurality of electromagnetic signals, extracting a multi-level semantic feature corresponding to each electromagnetic signal based on a residual network model, and obtaining a plurality of multi-level semantic features;

[0076] S102: performing multiple iterations according to the following steps to obtain a clustering result of each iteration: randomly dividing the plurality of multi-level semantic features into a plurality of sample sets; for each sample set, performing clustering using a plurality of clustering methods to obtain a first clustering result corresponding to each clustering method; determining a second clustering result corresponding to each sample set using a voting mechanism according to a plurality of first clustering results corresponding to the plurality of clustering methods; and taking a plurality of second clustering results corresponding to the plurality of sample sets as a third clustering result of this iteration;

[0077] S103: determining a fourth clustering result of the plurality of multi-level semantic features using a voting mechanism according to a plurality of third clustering results obtained through multiple iterations, and taking the fourth clustering result as a clustering result of the plurality of electromagnetic signals corresponding to the plurality of multi-level semantic features.

[0078] ​Specifically, according to the difference between the radar receiving target echo mode after transmitting an electric wave, it can be divided into primary radar and secondary radar. Primary radar: the radar receiving target echo mode after transmitting an electric wave to obtain target distance and direction information is called primary radar; secondary radar: the radar receiving target echo mode after transmitting an electric wave to obtain target distance and direction information is called secondary radar. Electromagnetic signals can be divided into pulse modulation signals transmitted by primary radar, secondary radar electromagnetic signals, target radiation source signals received by electronic warfare, and target communication and radar radiation source signals received by reconnaissance equipment.

[0079] In step S101, a plurality of electromagnetic signals can be received by an antenna, and the received plurality of electromagnetic signals are respectively input to a residual network model. The features of each electromagnetic signal are extracted, the corresponding multi-level semantic features of each electromagnetic signal are obtained, and the corresponding multi-level semantic features of the plurality of electromagnetic signals are obtained. It should be noted that the multi-level semantic features can include time domain, frequency domain, time-frequency domain features and graph signal spectrum distribution features.

[0080] In step S102, the plurality of multi-level semantic features extracted in step S101 are divided into a plurality of sample sets, such as sample set 1, sample set 2, …, sample set…

[0081] Preferably, the number of sample sets is divided into 4-6 for better effect.

[0082] For any one of the plurality of sample sets, as shown in Figure 2 For sample set 1, a plurality of clustering methods are used to cluster sample set 1 respectively, and the clustering results of each clustering method are obtained, denoted as the first clustering result. The plurality of first clustering results obtained by the plurality of clustering methods are determined as the clustering result of sample set 1 by using a voting mechanism, denoted as the second clustering result. It can be understood that the voting mechanism is that the minority submits to the majority, and when the minority submits to the majority is difficult to achieve, it can be determined by random selection.

[0083] The principle of the voting mechanism is illustrated by the following example: assuming that sample set 1 contains 10 electromagnetic signals, three clustering methods are used for clustering, and for any one of the electromagnetic signals,

[0084] If: the clustering result of the multi-level semantic features corresponding to the electromagnetic signal by using the first clustering method is A (the first clustering result), the clustering result of the multi-level semantic features corresponding to the electromagnetic signal by using the second clustering method is A (the first clustering result), and the clustering result of the multi-level semantic features corresponding to the electromagnetic signal by using the third clustering method is B (the first clustering result), then according to the principle of minority submitting to majority, the clustering result of the multi-level semantic features corresponding to the electromagnetic signal is A (the second clustering result);

[0085] If the clustering result of the multi-level semantic features corresponding to the electromagnetic signal by using the first clustering method is A (the first clustering result), the clustering result of the multi-level semantic features corresponding to the electromagnetic signal by using the second clustering method is B (the first clustering result), and the clustering result of the multi-level semantic features corresponding to the electromagnetic signal by using the third clustering method is C (the first clustering result), then the principle of majority over minority cannot be applied at this time, and a clustering result is randomly selected as the clustering result (the second clustering result) of the multi-level semantic features corresponding to the electromagnetic signal.

[0086] Preferably, the plurality of clustering methods can select more than three clustering methods.

[0087] It can be understood that a plurality of sample sets, i.e., sample set 1, sample set 2, …, sample set …, are included in the plurality of multi-level semantic features, and the plurality of sample sets include a plurality of second clustering results, as shown in the following table. Figure 3 The clustering result formed by combining the plurality of second clustering results is recorded as a third clustering result, and the third clustering result is the clustering result of one iteration of the plurality of multi-level semantic features. Multiple iterations of the plurality of multi-level semantic features result in a plurality of third clustering results.

[0088] In step S103, the clustering result determined by the voting mechanism according to the plurality of third clustering results obtained in step S102 is recorded as a fourth clustering result, and the fourth clustering result is the final clustering result corresponding to the plurality of multi-level semantic features. It can be understood that the plurality of multi-level semantic features correspond to the plurality of electromagnetic signals, and the clustering result of the plurality of multi-level semantic features is the clustering result of the plurality of electromagnetic signals.

[0089] In implementation, according to the plurality of received electromagnetic signals, a plurality of multi-level semantic features can be extracted by using a residual network model, and the plurality of multi-level semantic features are executed multiple times for iteration to obtain the third clustering result of each iteration execution; in each iteration, the plurality of multi-level semantic features are divided into a plurality of sample sets, and each sample set is clustered by using a plurality of clustering methods to obtain a plurality of first clustering results, and a second clustering result of each sample set is determined by using a voting mechanism; according to the plurality of third clustering results obtained by multiple iteration executions, a final fourth clustering result is determined by using a voting mechanism, and the fourth clustering result is the final clustering result of the plurality of electromagnetic signals.

[0090] Compared with the prior art, the electromagnetic signal clustering method provided by the embodiment of the application extracts the corresponding multi-level semantic features of each electromagnetic signal through a residual network model, improves the feature extraction of low signal-to-noise ratio electromagnetic signals, and iterates the several multi-level semantic features for multiple times, respectively uses multiple clustering methods and a voting mechanism in each iteration to improve the counterintuitive problems that may occur in the traditional integration, and further improves the clustering accuracy of the electromagnetic signals.

[0091] Further, the method comprises: determining clustering initial centers of the several multi-level semantic features based on granular computing;

[0092] The clustering of each sample set by using the multiple clustering methods to obtain the first clustering result corresponding to each clustering method comprises:

[0093] For each sample set, the clustering initial centers are applied to the multiple clustering methods in the sample set to obtain the first clustering result corresponding to each clustering method.

[0094] Specifically, the electromagnetic signal clustering method provided by the embodiment of the application can further comprise determining clustering initial centers of the several multi-level semantic features based on granular computing, and applying the clustering initial centers to the multiple clustering methods in each sample set.

[0095] The multiple clustering methods can comprise a k-means clustering algorithm, a k-medoids center point clustering algorithm, an FCM mean clustering algorithm, and the like.

[0096] It can be understood that, for each sample set, as shown in the sample set 1, the clustering initial centers are applied to the multiple clustering methods, specifically, the k-means clustering algorithm, the k-medoids center point clustering algorithm, and the FCM mean clustering algorithm. For each clustering algorithm, the first clustering result is obtained. Figure 2

[0097] It can be understood that, when processing a large amount of complex information, due to the limited cognitive ability of human beings, the human beings often divide a large amount of complex information into several relatively simple blocks according to the characteristics and performances of the complex information. Each divided block is regarded as a particle. In fact, a particle refers to a block formed by individuals through unclear relationships, similar relationships, adjacent relationships, or functional relationships. Such a process of processing information is called information granulation. For example, if the goods in a shopping mall are not arranged in a certain way, it is difficult to effectively manage the goods. Therefore, people divide the shopping mall into several blocks according to the types, volumes, grades, and the like of the goods arranged on the shelves, and arrange the shelves according to the blocks. Each block will arrange goods of the same type, similar volume, or same grade. The block is the concept of a particle, and the process of dividing the particles is called information granulation. ​

[0098] In implementation, the initial clustering centers of a plurality of multi-level semantic features are determined based on granular computing, and the initial clustering centers are respectively applied to a plurality of clustering methods in a plurality of sample sets to obtain a first clustering result.

[0099] Compared with the prior art, the electromagnetic signal clustering method provided by the embodiment of the application solves the problem of manually selecting the clustering center in the clustering method in the prior art by determining the initial clustering centers of a plurality of multi-level semantic features based on granular computing, and further improves the applicability of the electromagnetic signal clustering method provided by the application.

[0100] Further, the residual network model comprises a preprocessing layer, a residual block layer, and a multi-scale feature fusion layer.

[0101] The residual network model extracts the multi-level semantic features corresponding to each electromagnetic signal based on the residual network model to obtain a plurality of multi-level semantic features.

[0102] The preprocessing layer processes each electromagnetic signal into a feature map in a preset format; the residual block layer generates a plurality of predicted features according to the feature map in the preset format; and the multi-scale feature fusion layer fuses the plurality of predicted features to form the multi-level semantic features corresponding to each electromagnetic signal.

[0103] Specifically, as shown in Figure 4 The residual network model comprises a preprocessing layer, a residual block layer, and a multi-scale feature fusion layer. Each electromagnetic signal is input into the residual network model, and the electromagnetic signal is processed into a feature map in a preset format through the preprocessing layer. The feature map in the preset format is processed through the residual block layer to generate a plurality of predicted features, such as predicted feature 1, predicted feature 2, and predicted feature 3. The plurality of predicted features are fused through the multi-scale feature fusion layer to form multi-level semantic features, which are the features extracted for the electromagnetic signal.

[0104] It is worth noting that the preprocessing layer can process the electromagnetic signal into a feature map in a preset format of 416x416x3.

[0105] Compared with the prior art, the electromagnetic signal clustering method provided by the embodiment of the application extracts features from the electromagnetic signal through the residual network model to obtain the multi-level semantic features corresponding to the electromagnetic signal, thereby improving the accuracy of feature extraction for the electromagnetic signal.

[0106] Further, the residual block layer comprises a plurality of residual blocks connected in order from high to low in dimension level, and a plurality of residual blocks with lower dimension level in the plurality of residual blocks are used as preset residual blocks.

[0107] The residual block layer generates a plurality of predicted features according to the feature map in the preset format.

[0108] inputting the feature map in the preset format into the residual block layer and sequentially passing through the plurality of residual blocks;

[0109] performing upsampling operation on the output of each preset residual block to generate a first feature map;

[0110] obtaining a second feature map of the output of the adjacent high-dimension level residual block corresponding to the preset residual block;

[0111] performing feature fusion on the first feature map and the second feature map to form a prediction feature corresponding to each preset residual block.

[0112] Specifically, as shown in Figure 4 the residual block layer includes a plurality of residual blocks sequentially connected from high to low in dimension level, and several residual blocks with low dimension level in the plurality of residual blocks are taken as preset residual blocks, for example, the residual block layer includes residual block m, residual block 4, residual block 3, residual block 2, residual block 1 from high to low in dimension level, and residual block 1, residual block 2 and residual block 3 with low dimension level therein are taken as preset residual block 1, preset residual block 2 and preset residual block 3 respectively. The upsampling method can be methods such as bilinear interpolation, transpose convolution and unpooling, which are used to convert the feature map of low dimension level into the feature map of high dimension level.

[0113] It can be understood that the more preset residual blocks are preset in the residual block layer, the more prediction features will be generated. For example, if the preset number of preset residual blocks is 4-6, 4-6 prediction features will be generated accordingly.

[0114] In implementation, the residual block layer receives the feature map in the preset format, and then sequentially passes through residual block m, residual block 4, preset residual block 3, preset residual block 2 and preset residual block 1. In the residual block layer, upsampling operation is performed on the preset residual block to generate a new feature map, which is recorded as a first feature map; the output of the adjacent high-dimension level residual block corresponding to the preset residual block is recorded as a second feature map, and the first feature map and the second feature map are fused to form a prediction feature corresponding to each preset residual block.

[0115] It is worth noting that the residual network model provided by the embodiment of the present application can be the residual network model used in YOLO V3.

[0116] As shown in Figure 5 , for the ADS-B signal, the residual network model provided by the embodiment of the present application can generate Figure 5 64-dimensional high-level semantic features.

[0117] Compared with the prior art, the electromagnetic signal clustering method provided in this embodiment of the invention fuses the features of different levels of electromagnetic signals to form predicted features through preset residual blocks in the residual block layer, and uses multiple predicted features to form multi-level semantic features, which further improves the accuracy of electromagnetic signal feature extraction.

[0118] Furthermore, the determination of the initial cluster centers for the plurality of multi-level semantic features based on granular computation includes:

[0119] Calculate the similarity between the aforementioned multi-level semantic features;

[0120] A fuzzy similarity matrix is ​​established among the several multi-level semantic features based on the similarity and similarity threshold. Based on the fuzzy similarity matrix, the similarity features of each multi-level semantic feature are obtained. Each multi-level semantic feature and its corresponding set of similar features are taken as a particle. Several particles are obtained based on the several multi-level semantic features.

[0121] Calculate the particle density and average particle density of the plurality of particles, and determine the effective particle set based on the particle density and the average particle density;

[0122] The initial cluster centers of the plurality of multi-level semantic features are determined based on the effective particle set.

[0123] Specifically, suppose there are N multi-level semantic features among several multi-level semantic features, denoted as x1, x2, x3, ..., x... i , ..., x j , ..., x N x i and x j Let each of the N multi-level semantic features be any one of them. Calculating the similarity between several multi-level semantic features is equivalent to calculating the similarity between any one of the N multi-level semantic features and all the N multi-level semantic features, S(x). i ,x j ) is denoted as x i and x j The similarity can be calculated using methods such as Jaccard correlation coefficient, cosine similarity, or Pearson correlation coefficient.

[0124] Understandably, the similarity threshold can be preset manually, denoted as .

[0125] It is worth noting that the fuzzy similarity matrix between several multi-level semantic features can be used to represent the similarity value of two multi-level semantic features, where [M(i,j)] is denoted as x. i and x jThe fuzzy similarity matrix. When When, it can be written as [M(i,j)]=1, x i and x j They have similar features; when When, it can be written as [M(i,j)]=0, x i and x j Not similar features. Therefore, based on the fuzzy similarity matrix, the similarity features of any one of the several multi-level semantic features can be determined, and each multi-level semantic feature and its corresponding set of similar features can be treated as a particle.

[0126] It is understandable that if there are N multi-level semantic features, then there are corresponding N particles. The particle density and average particle density of the N particles are calculated. Based on the particle density and average particle density, multiple particles that meet the conditions are selected. These selected particles are taken as the effective particle set. The initial cluster centers of the N multi-level semantic features are determined based on the effective particle set.

[0127] Compared with the prior art, the electromagnetic signal clustering method provided in this embodiment of the invention determines several particles by using the similarity and similarity threshold between several multi-level semantic features, and then determines the initial clustering center of several multi-level semantic features by using several particles, which greatly improves the accuracy and applicability of the selection of the initial clustering center.

[0128] Preferably, calculating the similarity between the plurality of multi-level semantic features includes:

[0129] The similarity is defined as:

[0130]

[0131] Where, x i x represents the i-th multi-level semantic feature; j S(x) represents the j-th multi-level semantic feature; i ,x j ) represents x i and x j Similarity between them; ω l Let represent the attribute discrimination capability of the l-th attribute of the multi-level semantic features; B represents the attribute set of the multi-level semantic features, and |B| represents the number of elements included in B; x il x represents i The attribute value of the lth attribute; x jl x represents j The attribute value of the lth attribute;

[0132] ω l Defined as:

[0133]

[0134] wherein the attribute l is divided into L interval blocks {y1, y2, …, y r ,…,y L}, y r represents the rth interval block in the L interval blocks; N represents the number of the plurality of multi-level semantic features.

[0135] Preferably, the similarity threshold is determined by:

[0136] calculating the average similarity of the plurality of multi-level semantic features;

[0137] wherein the average similarity is defined as:

[0138]

[0139] determining the similarity threshold according to the average similarity, the similarity threshold ranges from 0 to 1.

[0140] Compared with the prior art, the electromagnetic signal clustering method provided by the embodiment of the present application determines the similarity threshold through the similarity and the average similarity of the plurality of multi-level semantic features, can adjust the number of multi-level semantic features included in the particle by adjusting the similarity threshold, and can further improve the accuracy and applicability of the selection of the initial clustering center.

[0141] Preferably, the calculating the particle density and the average particle density of the plurality of particles and determining the effective particle set according to the particle density and the average particle density comprises:

[0142] calculating the particle density of each particle, and the particle density is defined as:

[0143] gd(X n ) = |X n | / N;

[0144] wherein the total number of particles is the same as the number of multi-level semantic features, which is N, and the plurality of particles is represented as {X1, X2, …, X n ,…,X N}, X n represents the nth particle in the N particles, |X n | represents the number of multi-level semantic features included in the nth particle; and gd(X n ) represents the particle density of the nth particle.

[0145] calculating the average particle density of the plurality of particles, and the average particle density is defined as:

[0146]

[0147] wherein, represents the average particle density of the plurality of particles;

[0148] The particles in the plurality of particles with a particle density greater than or equal to the average particle density are grouped into an effective particle set.

[0149] Preferably, the determining the initial cluster centers of the plurality of multi-level semantic features according to the effective particle set comprises:

[0150] The particle center of each particle in the effective particle set is calculated, and the particle center is defined as:

[0151]

[0152] wherein, C q represents the particle center of the qth particle in the effective particle set, the qth particle including Q objects, each object being a multi-level semantic feature; x qk1 represents the multi-level semantic feature of the k1th object in the Q objects, x qk2 represents the multi-level semantic feature of the k2th object in the Q objects, x qk2 as the particle center of the qth particle;

[0153] The particles are sorted according to the particle density of each particle, the particle center of the particle with the maximum particle density in the effective particle set is taken as the first cluster center, and the particle with the maximum particle density and the farthest Euclidean distance from the particle center of the particle with the maximum particle density is taken as the second cluster center.

[0154] The Euclidean distance between each remaining particle center in the remaining particles in the effective particle set and the determined cluster center is calculated, the minimum value of the above Euclidean distance is taken as the cluster distance of the remaining particle, the cluster distances of the remaining particles are compared, the particle center of the particle corresponding to the maximum cluster distance is taken as the next cluster center, and the step is repeated until the maximum cluster distance does not satisfy the set cluster distance threshold condition.

[0155] The obtained cluster centers are taken as the initial cluster centers of the plurality of multi-level semantic features.

[0156] Specifically, the particles are sorted according to the particle density of each particle, the particle center of the particle with the maximum particle density in the effective particle set is taken as the first cluster center o1, and the particle with the maximum particle density and the farthest Euclidean distance from the particle center of the particle with the maximum particle density is taken as the second cluster center o2.

[0157] It should be noted that the Euclidean distance between two particles is the Euclidean distance between the particle centers of the two particles.

[0158] After the first and second cluster centers are determined, the third and subsequent cluster centers need to be determined in the following manner:

[0159] The Euclidean distance between each of the remaining particle centers in the effective particle set and the determined cluster center is calculated, and the minimum value of the Euclidean distance is taken as the cluster distance of the remaining particle; the cluster distances of the remaining particles are compared, and the particle center of the particle corresponding to the maximum cluster distance is taken as the next cluster center, and the step is repeatedly executed until the maximum cluster distance does not satisfy the set cluster distance threshold condition;

[0160] For example, it is assumed that the effective particle set includes D particles, and k cluster centers have been determined, then the number of remaining particles is D-k, and the Euclidean distance between the particle center of the pth particle in the remaining particles and the k determined cluster centers can be defined as: d p1 ,d p2 ,…,d pk , where o1, o2,..., o k represent the determined cluster centers.

[0161] The cluster distance d p of the pth particle in the remaining particles can be defined as:

[0162] d p =min(d p1 ,d p2 ,…,d pk )。

[0163] The cluster distances of each of the D-k remaining particles are compared, and the particle center of the particle corresponding to the maximum cluster distance max(d1, d2,..., d p ,..., d D-k ) is selected as the next cluster center, denoted as o k+1 .

[0164] The step is repeatedly executed until the maximum cluster distance does not satisfy the set cluster distance threshold condition, and the obtained cluster centers are taken as the cluster initial centers of the multiple multi-level semantic features.

[0165] In implementation, the effective particle set is determined according to the particle density and the average particle density, and the cluster initial centers of the multiple multi-level semantic features are determined according to the effective particle set screening, which can further improve the accuracy and applicability of the selection of the cluster initial centers.

[0166] The technical effects of the electromagnetic signal clustering method provided by the embodiments of the application are illustrated by two specific examples below.

[0167] Example one:

[0168] For Iris and Wine public data sets in UCI, as shown in Table 1:

[0169] Table 1 Iris, Wine standard data sets

[0170] Dataset Number of objects Attribute dimension Number of classes Number of objects per class Iris 150 4 3 50,50,50 Wine 178 13 3 59,71,48

[0171] Through 1000 times of Monte Carlo simulation, the initial clustering center of the multi-level semantic feature is determined based on granular computing, and then the data set clustering is determined according to the k-means clustering algorithm, the k-medoids center point clustering algorithm and the FCM mean clustering algorithm, and the results are shown in Table 2:

[0172] Table 2 Clustering accuracy comparison (%)

[0173]

[0174] Among them, the three rows of k-means clustering algorithm, k-medoids center point clustering algorithm and FCM mean clustering algorithm represent the clustering accuracy reached by using the clustering method alone; the k-means clustering algorithm based on granular computing, the k-medoids center point clustering algorithm based on granular computing and the FCM mean clustering algorithm based on granular computing represent the clustering accuracy reached by the improved clustering method after determining the initial clustering center based on granular computing.

[0175] As can be seen from Table 2, the clustering method based on granular computing to determine the initial clustering center has different degrees of improvement in clustering accuracy compared with the original clustering method, which verifies that the improved clustering method based on granular computing has more advantages.

[0176] Example two:

[0177] Based on the ADS-B signals of 10 targets actually collected, based on the 64-dimensional features extracted by the residual network model provided in the embodiment of the application, for 5 targets, 8 targets and 10 targets, 1000 times of Monte Carlo simulation are carried out by using the k-means clustering algorithm, the k-medoids center point clustering algorithm, the FCM mean clustering algorithm and the electromagnetic signal clustering method provided by the application, and the results are shown in Table 3:

[0178] Table 3 Clustering accuracy comparison (%)

[0179]

[0180] As can be seen from Table 3, the clustering accuracy of the electromagnetic signal clustering method provided by the embodiment of the application reaches 100%, which has a qualitative leap compared with the clustering accuracy of the k-means clustering algorithm, the k-medoids center point clustering algorithm or the FCM mean clustering algorithm alone in the prior art.

[0181] Furthermore, the embodiment of the present application also provides an electromagnetic signal clustering system, which comprises:

[0182] The signal feature extraction module 601 is configured to receive a plurality of electromagnetic signals, extract a multi-level semantic feature corresponding to each electromagnetic signal based on a residual network model, and obtain a plurality of multi-level semantic features.

[0183] The multiple iteration execution module 602 is configured to perform multiple iterations according to the following steps to obtain a clustering result of each iteration: randomly divide the plurality of multi-level semantic features into a plurality of sample sets; for each sample set, perform clustering by using a plurality of clustering methods to obtain a first clustering result corresponding to each clustering method; determine a second clustering result corresponding to each sample set by using a voting mechanism according to the plurality of first clustering results corresponding to the plurality of clustering methods; and take the plurality of second clustering results corresponding to the plurality of sample sets as a third clustering result of the current iteration.

[0184] The clustering result determination module 603 is configured to determine a fourth clustering result of the plurality of multi-level semantic features by using a voting mechanism according to the plurality of third clustering results obtained through multiple iterations, and take the fourth clustering result as a clustering result of the plurality of electromagnetic signals corresponding to the plurality of multi-level semantic features.

[0185] Those skilled in the art can understand that all or part of the processes of the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium. The computer readable storage medium includes a magnetic disk, an optical disk, a read-only memory, a random access memory, etc.

[0186] The above is only a preferred embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An electromagnetic signal clustering method, characterized in that, The method includes: Several electromagnetic signals are received, and multi-level semantic features corresponding to each electromagnetic signal are extracted based on the residual network model to obtain several multi-level semantic features. The following steps are performed to iterate multiple times and obtain the clustering results for each iteration: the plurality of multi-level semantic features are randomly divided into multiple sample sets; for each sample set, multiple clustering methods are used to cluster them, and the first clustering result corresponding to each clustering method is obtained; based on the multiple first clustering results corresponding to the multiple clustering methods, a voting mechanism is used to determine the second clustering result corresponding to each sample set; the multiple second clustering results corresponding to the multiple sample sets are used as the third clustering result of this iteration; Based on the multiple third clustering results obtained through multiple iterations, a voting mechanism is used to determine the fourth clustering result of the several multi-level semantic features, and the fourth clustering result is used as the clustering result of the several electromagnetic signals corresponding to the several multi-level semantic features. The method includes: determining the initial cluster centers of the plurality of multi-level semantic features based on granular computation; For each sample set, multiple clustering methods are used for clustering, and the first clustering result corresponding to each clustering method includes: For each sample set, based on the initial cluster centers, multiple clustering methods are used to perform clustering, and the first clustering result corresponding to each clustering method is obtained; The determination of the initial cluster centers for the plurality of multi-level semantic features based on granular computation includes: Calculate the similarity between the aforementioned multi-level semantic features; A fuzzy similarity matrix is ​​established among the several multi-level semantic features based on the similarity and similarity threshold. Based on the fuzzy similarity matrix, the similarity features of each multi-level semantic feature are obtained. Each multi-level semantic feature and its corresponding set of similar features are taken as a particle. Several particles are obtained based on the several multi-level semantic features. Calculate the particle density and average particle density of the plurality of particles, and determine the effective particle set based on the particle density and the average particle density; The initial cluster centers of the plurality of multi-level semantic features are determined based on the effective particle set; The step of determining the initial cluster centers of the plurality of multi-level semantic features based on the effective particle set includes: Calculate the particle center of each particle in the effective particle set, whereby the particle center is defined as: in, C q This represents the particle center of the q-th particle in the effective particle set. The q-th particle contains Q objects, each of which is a multi-level semantic feature; x qk1 x represents the multi-level semantic features of the k1-th object among Q objects. qk2 This represents the multi-level semantic features of the k2-th object among Q objects; The particles are sorted according to their particle density; the particle center of the particle with the highest particle density in the effective particle set is taken as the first cluster center; the particle with the highest particle density and the greatest Euclidean distance from the particle center of the particle with the highest particle density is taken as the second cluster center. Calculate the Euclidean distance between the center of each remaining particle in the effective particle set and the determined cluster center. Take the minimum value from the above Euclidean distances as the cluster distance of the remaining particle. Compare the cluster distances of each remaining particle. Take the particle center corresponding to the particle with the maximum cluster distance as the next cluster center. Repeat this step until the maximum cluster distance does not meet the set cluster distance threshold condition. The obtained cluster centers are used as the initial cluster centers for the multiple multi-level semantic features.

2. The electromagnetic signal clustering method according to claim 1, characterized in that, The residual network model includes a preprocessing layer, a residual block layer, and a multi-scale feature fusion layer. The extraction of multi-level semantic features corresponding to each electromagnetic signal based on the residual network model yields several multi-level semantic features, including: The preprocessing layer processes each electromagnetic signal into a feature map of a preset format; the residual block layer generates multiple predicted features based on the feature map of the preset format; and the multi-scale feature fusion layer fuses the multiple predicted features to form a multi-level semantic feature corresponding to each electromagnetic signal.

3. The electromagnetic signal clustering method according to claim 2, characterized in that, The residual block layer includes multiple residual blocks connected sequentially from high to low dimension level, and a number of residual blocks with lower dimension level among the multiple residual blocks are used as preset residual blocks; The residual block layer generates multiple predicted features based on the feature map of the preset format, including: The feature map in the preset format is input into the residual block layer and passes through the multiple residual blocks in sequence; The output of each preset residual block is upsampled to generate the first feature map; Obtain the second feature map of the output of the adjacent high-dimensional residual blocks corresponding to the preset residual block; The first feature map and the second feature map are fused to form the prediction features corresponding to each preset residual block.

4. The electromagnetic signal clustering method according to claim 1, characterized in that, The calculation of the similarity between the plurality of multi-level semantic features includes: The similarity is defined as: Where, x i x represents the i-th multi-level semantic feature; j S(x) represents the j-th multi-level semantic feature; i ,x j ) represents x i and x j Similarity between them; ω l Let represent the attribute discrimination capability of the l-th attribute of the multi-level semantic features; B represents the attribute set of the multi-level semantic features, and |B| represents the number of elements included in B; x il x represents i The attribute value of the lth attribute; x jl x represents j The attribute value of the lth attribute; ω l Defined as: Among them, attribute l is divided into L interval blocks {y1, y2, ..., y r ,…,y L }, y r This represents the r-th interval block among L interval blocks; N represents the number of the plurality of multi-level semantic features.

5. The electromagnetic signal clustering method according to claim 1, characterized in that, The similarity threshold is determined as follows: Calculate the average similarity of the aforementioned multi-level semantic features; The average similarity is defined as follows: The similarity threshold is determined based on the average similarity, and the range of the similarity threshold is [missing information].

6. The electromagnetic signal clustering method according to claim 1, characterized in that, The calculation of the particle density and average particle density of the plurality of particles, and the determination of the effective particle set based on the particle density and the average particle density, includes: Calculate the particle density for each particle, which is defined as: gd(X n )=|X n | / N; The total number of particles is the same as the number of multi-level semantic features, which is N. These particles are represented as {X1, X2, ..., X...} n ,…,X N }, X n Let |X| represent the nth particle out of N particles. n | represents the number of multi-level semantic features included in the nth particle; gd(X n ) represents the particle density of the nth particle; Calculate the average particle density of the given particles. The average particle density is defined as: in, This represents the average particle density of the aforementioned particles; The effective particle set is composed of particles whose particle density is greater than or equal to the average particle density.

7. An electromagnetic signal clustering system, characterized in that, The system includes: The signal feature extraction module is used to receive several electromagnetic signals and extract multi-level semantic features corresponding to each electromagnetic signal based on the residual network model to obtain several multi-level semantic features. The multiple iteration execution module is used to perform multiple iterations according to the following steps to obtain the clustering result of each iteration: randomly dividing the several multi-level semantic features into multiple sample sets; for each sample set, using multiple clustering methods to perform clustering, obtaining the first clustering result corresponding to each clustering method; based on the multiple first clustering results corresponding to the multiple clustering methods, using a voting mechanism to determine the second clustering result corresponding to each sample set; and using the multiple second clustering results corresponding to the multiple sample sets as the third clustering result of this iteration. The clustering result determination module is used to determine the fourth clustering result of the plurality of multi-level semantic features based on the plurality of third clustering results obtained through multiple iterations, using a voting mechanism, and to use the fourth clustering result as the clustering result of the plurality of electromagnetic signals corresponding to the plurality of multi-level semantic features. The system includes: determining the initial cluster centers of the several multi-level semantic features based on granular computing; For each sample set, multiple clustering methods are used for clustering, and the first clustering result corresponding to each clustering method includes: For each sample set, based on the initial cluster centers, multiple clustering methods are used to perform clustering, and the first clustering result corresponding to each clustering method is obtained; The determination of the initial cluster centers for the plurality of multi-level semantic features based on granular computation includes: Calculate the similarity between the aforementioned multi-level semantic features; A fuzzy similarity matrix is ​​established among the several multi-level semantic features based on the similarity and similarity threshold. Based on the fuzzy similarity matrix, the similarity features of each multi-level semantic feature are obtained. Each multi-level semantic feature and its corresponding set of similar features are taken as a particle. Several particles are obtained based on the several multi-level semantic features. Calculate the particle density and average particle density of the plurality of particles, and determine the effective particle set based on the particle density and the average particle density; The initial cluster centers of the plurality of multi-level semantic features are determined based on the effective particle set; The step of determining the initial cluster centers of the plurality of multi-level semantic features based on the effective particle set includes: Calculate the particle center of each particle in the effective particle set, whereby the particle center is defined as: in, C q This represents the particle center of the q-th particle in the effective particle set. The q-th particle contains Q objects, each of which is a multi-level semantic feature; x qk1 x represents the multi-level semantic features of the k1-th object among Q objects. qk2 This represents the multi-level semantic features of the k2-th object among Q objects; The particles are sorted according to their particle density; the particle center of the particle with the highest particle density in the effective particle set is taken as the first cluster center; the particle with the highest particle density and the greatest Euclidean distance from the particle center of the particle with the highest particle density is taken as the second cluster center. Calculate the Euclidean distance between the center of each remaining particle in the effective particle set and the determined cluster center. Take the minimum value from the above Euclidean distances as the cluster distance of the remaining particle. Compare the cluster distances of each remaining particle. Take the particle center corresponding to the particle with the maximum cluster distance as the next cluster center. Repeat this step until the maximum cluster distance does not meet the set cluster distance threshold condition. The obtained cluster centers are used as the initial cluster centers for the multiple multi-level semantic features.

Images