Radar signal sorting method based on improved affinity propagation clustering

Abstract

The invention discloses a radar signal sorting method based on improved affinity propagation clustering. The method comprises the following steps: S1, detecting pulse description word parameter data of an original signal; s2, segmenting the pulse description word parameter data into a plurality of frames; s3, pre-sorting each frame of pulse data by using an improved affinity propagation algorithm; s4, extracting candidate batch centroids by using an adaptive density peak clustering algorithm; s5, completing the blending detection of the candidate blending cluster corresponding to each candidate blending mass center; and S6, completing blending and outputting a final sorting result. The method comprises the following steps: pre-sorting pulse data by using an IAP algorithm introducing an attenuation factor, extracting candidate batch centroids based on density, and finishing batch detection in combination with statistical characteristics of a pulse cluster arrival direction; geometric characteristics, distribution characteristics and statistical characteristics of radar pulse data are comprehensively utilized, potential relevance and mode information between signals are fully mined, and high sorting accuracy is guaranteed while batch increasing phenomena are effectively reduced.

Description

Technical Field

[0001] The invention belongs to the field of electronic reconnaissance, and in particular relates to a radar signal sorting technology. Background Art

[0002] With the rapid development of information technology and the widespread application of new radar systems, the electromagnetic environment has become increasingly complex in recent years, and electronic reconnaissance has become the core mission scenario of electronic warfare. Radar signal sorting technology, as the basic technology of electronic reconnaissance, has the main task of effectively separating pulse sequences belonging to different radiation sources from the received radar pulse stream, providing key data for subsequent radar signal analysis and target identification, and is the premise and foundation for efficient electronic support measures and accurate electronic countermeasure response.

[0003] Traditional PRI-based signal sorting techniques such as cumulative difference histogram (CDIF), sequence difference histogram (SDIF), and PRI transform are very sensitive to the arrival time of the signal, and only use the time information of the pulse data while ignoring the characteristics of other dimensions. In order to mine the internal structure and pattern of radar signal data and reveal the potential distribution and relationship of radar pulses, signal sorting techniques based on unsupervised clustering have been proposed, and classic clustering methods such as K-means and DBSCAN are widely used in radar signal sorting. Although these clustering methods are effective in specific scenarios, they still have certain limitations. K-means requires the number of clusters to be specified in advance and relies on the initial center selection, while DBSCAN is sensitive to neighborhood parameters. Compared with the above traditional clustering methods, affinity propagation (AP) clustering does not require the number of cluster centers and initialization positions to be specified in advance, and is a stable and effective unsupervised clustering algorithm.

[0004] However, with the widespread application of various new radar systems, the battlefield electromagnetic environment is becoming increasingly complex and changeable, the radar pulse density is surging, and the modulation types of radiation source parameters are changing. When using the AP algorithm to cluster and sort radar pulse signals, there is a phenomenon of dividing a radar radiation source into two or more clusters. Specifically, the radar signal parameters have modulation types such as agile change and sliding change, which causes the parameters to change within a large range, and then causes the long strip distribution of pulse data to be divided into multiple clusters, resulting in batch increase; when processing the pulse signal of the multi-function radar, the pulses from different working modes are classified as different radiation source signals, which leads to the phenomenon of batch increase. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a radar signal sorting method based on improved affinity propagation clustering. The AP algorithm is improved by introducing an attenuation factor, and the density is used to extract the candidate batch centroids. The batch detection is completed in combination with the statistical characteristics of the pulse cluster arrival direction, which effectively reduces the batch increase phenomenon while ensuring a high sorting accuracy.

[0006] The object of the present invention is achieved by the following technical solutions: A radar signal sorting method based on improved affinity propagation clustering, comprising the following steps:

[0007] S1. Detect the pulse description word parameter data of the original signal;

[0008] S2. Divide the detected pulse description word parameter data into several frames according to time frames to form a pulse data set;

[0009] S3. Pre-sort each frame of pulse data in the data set using the improved affinity propagation algorithm, comprising the following steps:

[0010] S31. For the pulse data of the k-th frame Calculate the similarity S(x i and the j-th pulse data point x j ) of the i-th pulse data point: i ,x j ):

[0011]

[0012] In the formula, Δp, Δc, and Δd respectively represent the differences in pulse width, carrier frequency, and arrival angle between x i and x j ;

[0013] S32. Introduce an attenuation factor α, and correct the difference in carrier frequency between x i and x j as:

[0014]

[0015] In the formula, and are respectively the carrier frequency values of x i and x j ; Substitute the corrected difference in carrier frequency into (1) to update the similarity matrix;

[0016] S33. Continuously update the clustering center corresponding to each pulse data point by iteratively updating the attraction information r(i,j) and the belonging information a(i,j):

[0017] r(i,k) = s(i,k) - max{a(i,k′) + s(i,k′)} (3)

[0018] a(i,k) = min{0, r(k,k) + ∑max{0, r(i′,k)} (4)

[0019] Among them, the attraction information r(i,j) describes the clustering center xk Suitability as a pulse data point x i The degree of class representativeness, and the attribution information a(i,j) represents the pulse data point x i Select the pulse data point x k The credibility of using it as its clustering center;

[0020] S34. To ensure the legality of the clustering result, define the constraint condition: c i If it is j, then there is That is If it is x i As the clustering center, then It must be its own clustering center;

[0021] S35. Through the iteration of the attraction information and the attribution information, continuously update the clustering center corresponding to the pulse data point x i To minimize the energy function E[c]: Therein, the preference parameter σ represents the cost of each point as its own representative point, which is a free parameter; δ(c

[0022]

[0023] ,i) is the Kronecker function, that is, δ(c i ,i) = 1 if and only if c i = i, otherwise δ(c i ,i) = 0; i

[0024] By minimizing the energy function, obtain the final corresponding clustering center for each pulse data point x i As its clustering center; for the pulse data X of the k-th frame k A total of n clustering centers are obtained;

[0025] S4. Use the adaptive density peak clustering algorithm to extract candidate batch combination centroids for the pre-sorting result; specifically include the following steps:

[0026] S41. Denote the n pre-sorting clustering centers obtained from the pulse data X k through the improved affinity propagation algorithm as e i , i = 1, 2,..., n; calculate the local density p i of e i and the minimum distance d i ;

[0027] S42. Repeat step S41 until all pre-sorting clustering center points are traversed, and draw a two-dimensional batch combination centroid decision diagram of local density - minimum distance according to (p i , d i ); ​

[0028] S43. Implement adaptive selection of the combined batch centroid by using a combined batch centroid selection method based on statistical theory: Identify the combined batch centroid by using an outlier detection method, which includes three steps: Estimate the specific local density value p in the decision graph i and any minimum distance value p i corresponding to the probability density function P y (p i , y); Use P y (p i , y) to calculate the mean and variance of the y distribution at the specific local density value p i ; Use the mean and variance for outlier detection to obtain candidate combined batch centroids;

[0029] S5. Complete combined batch detection for each candidate combined batch cluster corresponding to the candidate combined batch centroid;

[0030] S6. Complete combined batching and output the final sorting result.

[0031] The beneficial effects of the present invention are as follows: A radar signal sorting method based on improved affinity propagation clustering of the present invention improves the AP clustering algorithm by introducing an attenuation factor, suppresses the influence of large-scale changes in radiation source parameters on the calculation of the similarity matrix, and effectively alleviates the problem of batch increase of the traditional AP algorithm in a complex electromagnetic environment; combines adaptive density peak clustering to extract candidate combined centroids, and fuses the statistical features of the pulse cluster DOA to achieve joint judgment of multi-dimensional features and accurately combine multi-mode pulse sequences of the same radiation source. The present invention constrains parameter drift through geometric and distribution features, and enhances cluster correlation by using statistical characteristics, showing strong robustness and high generalization ability in a complex electromagnetic environment with high pulse density, overlapping pulse parameters, and multi-mode agile modulation of radiation sources. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flowchart of a radar signal sorting scheme provided by an embodiment of the present invention;

[0033] Figure 2 is a flowchart of an IAP pre-sorting algorithm provided by an embodiment of the present invention;

[0034] Figure 3 is a flowchart of extracting candidate combined batch centroids provided by an embodiment of the present invention;

[0035] Figure 4 is a comparison diagram of radar signal sorting test effects provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] The technical solution of the present invention will be further described below with reference to the drawings.

[0037] As Figure 1As shown in the figure, a radar signal sorting method based on improved affinity propagation clustering according to the present invention includes the following steps:

[0038] S1. Detect the pulse description word (PDW) parameter data of the original signal; generally, the PDW parameters specifically include time of arrival (TOA), carrier frequency (CF), direction of arrival (DOA), pulse width (PW), pulse amplitude (PA), etc.

[0039] In the actual detection process, due to the wide coverage range and numerous targets, the collected PDW data is often a continuous data stream of a large number of pulse signals, with a huge amount of data and high time variability. Direct processing is difficult to meet the requirements of computational efficiency. Therefore, in order to facilitate subsequent effective sorting of data according to different features and time-domain changes, it is necessary to divide the PDW data detected in step S1 into several frames according to time or other features to achieve batch processing.

[0040] S2. Divide the detected pulse description word parameter data into several frames according to time frames to form a pulse data set.

[0041] The PDW data received within the working time T of the receiver in step S1 is divided into L f frames according to the time frame length τ, and the pulse data of the k-th frame is expressed as:

[0042]

[0043] where N k is the number of pulses in the k-th frame. Since TOA and PA are easily affected by environmental interference, the three relatively stable parameters of CF, PW, and DOA are often selected for the signal sorting algorithm based on clustering. The pulse data x i in this frame is:

[0044] x i ={CF i , PW i , DOA i}, i = 1, 2,..., N k .

[0045] S3. Use the improved affinity propagation (IAP) algorithm to perform preliminary sorting on the pulse data of each frame in the data set. Specifically: for a given pulse data sample set, construct a similarity matrix to measure the similarity between any two pulse data points. By iteratively updating the attraction information r(i,j) and the membership information a(i,j), each pulse data point can select the most suitable clustering center, so as to achieve the goal of minimizing the overall energy function. Finally, each pulse data point x i is assigned to the cluster belonging to the best clustering center to achieve the purpose of preliminary sorting. In this embodiment, the IAP algorithm process is as shown in Figure 2As shown in the figure; it includes the following steps:

[0046] S31. For the pulse data of the k-th frame Calculate the similarity S(x i between the i-th pulse data point x j and the j-th pulse data point x i : j )

[0047]

[0048] In the formula, Δp, Δc, and Δd respectively represent the differences in pulse width, carrier frequency, and arrival angle between x i and x j ;

[0049] S32. Introduce the attenuation factor α and correct the difference in carrier frequency between x i and x j to:

[0050]

[0051] In the formula, and are respectively the carrier frequency values of x i and x j ; Substitute the corrected difference in carrier frequency into (1) to update the similarity matrix;

[0052] S33. By iteratively updating the attraction information r(i,k) and the membership information a(i,k), continuously update the cluster center corresponding to each pulse data point:

[0053] r(i,k) = S(i,k) - max{a(i,k′) + S(i,k′)} (3)

[0054] a(i,k) = min{0, r(k,k) + ∑max{0, r(i′,k)}} (4)

[0055] Among them, the attraction information r(i,k) describes the degree to which the cluster center x k is suitable as the class representative of the pulse data point x i , and the membership information a(i,k) indicates the credibility of the pulse data point x i selecting the pulse data point x k as its cluster center; a(i,k′) represents the membership information of the remaining pulse data points to x k except for the pulse data point x i , and S(i,k′) represents the similarity between the remaining pulse data points x k except for the pulse data point x k′ and x iThe similarity, r(i′, k) represents the attraction information of the remaining pulse data points except the pulse data point x i to x k . max{0, r(i′, k)} represents selecting the larger value from 0 and r(i′, k).

[0056] S34. To ensure the legality of the clustering result, a constraint condition is defined: if c i = j, then there is that is If it is the clustering center of x i , then must be its own clustering center;

[0057] S35. Through the iteration of the attraction information and the attribution information in step S33, continuously update the clustering center corresponding to the pulse data point x i so as to minimize the energy function E[c]: where the preference parameter σ represents the cost of each point as its own representative point and is a free parameter; δ(c

[0058]

[0059] , i) is the Kronecker function, that is, δ(c i , i) = 1 if and only if c i = i, otherwise δ(c i , i) = 0; i

[0060] By minimizing the energy function, obtain the final corresponding clustering center of each pulse data point x i as its clustering center; for the pulse data X k of the k-th frame, a total of n clustering centers are obtained.

[0061] S4. Use the Adaptive Density Peak Clustering (ADPC) algorithm to extract candidate batch combination centroids for the pre-sorting result; the process of extracting candidate batch combination centroids in this embodiment is as Figure 3 shown; specifically includes the following steps:

[0062] S41. Denote the n pre-sorting clustering centers obtained by the improved affinity propagation algorithm for the pulse data X k as e i , i = 1, 2,..., n; calculate the local density p i of e i and the minimum distance d i :

[0063]

[0064] where is the pre-sorting clustering center e​i With e j The distance between th is the cutoff distance threshold; usually, if point e i p i is the local or global maximum density value, then its d i The value of is much larger than the minimum distance of its neighboring points. Therefore, the pre-sorted cluster center points with larger minimum distance are more likely to be selected as candidate batch centroid points.

[0065] S42, repeat step S41 until all pre-sorted cluster centers are traversed, according to (p i ,d i ) Draw a local density-minimum distance two-dimensional batch centroid decision diagram, and the batch centroid can be screened out in the decision diagram;

[0066] S43, using the batch centroid selection method based on statistical theory to realize adaptive selection of batch centroid: if the pre-sorted cluster center e i is the local density peak point, then its minimum distance value d i is much larger than the average point in its vicinity. Therefore, an important feature of detecting the centroid of a batch is its d i The value is abnormally large. Therefore, the outlier detection method is used to identify the batch centroid. This method mainly includes three steps: estimating the specific local density value p in the decision graph i and any minimum distance value p i The corresponding probability density function P y (p i ,y); using P y (p i ,y) at a specific local density value p i Calculate the mean and variance of the y distribution; use the mean and variance to detect outliers and obtain candidate batch centroids.

[0067] Use a two-dimensional Gaussian function to estimate the local density value p i The probability density function P y (p i ,y) is:

[0068]

[0069] Where n is the total number of pre-sorted cluster centers, a and b are the widths of the two-dimensional core;

[0070] Calculate a specific p i The minimum distance d i The expected value and variance of are as follows:

[0071]

[0072] So far, the minimum distance d i at a specific p i and its variance are obtained; the 3σ criterion in statistics is used for outlier detection, and the detection threshold is given by the following formula:

[0073] TH d (p i ) = μ y (p i ) + 3 × σ y (p i ) (11)

[0074] For any pre-sorted clustering center e i , if its minimum distance d i > TH d (p i ), where TH d (p i ) is a preset threshold; it will be recognized as an outlier and thus marked as a candidate batch centroid e can ;

[0075] S5. Perform batch detection on each candidate batch cluster corresponding to each candidate batch centroid; extract the DOA statistical feature quantities of each candidate batch cluster corresponding to each candidate batch centroid e can obtained in step S4, and perform threshold judgment; update the DOA statistical feature quantities of the new cluster after batching. This embodiment specifically includes the following steps:

[0076] S51. For the same radar radiation source, within one scanning time, the DOA does not change suddenly and has good stability. Therefore, extract the statistical feature quantities of the DOA of the candidate batch pulse cluster corresponding to e can and perform batch detection. For the i-th candidate batch cluster, the statistical feature quantity of its DOA is DOA i (μ i , Σ i , N i ). Among them, μ i , Σ i are the mean value and covariance matrix of the DOA of this cluster respectively, and N i is the number of pulses in this cluster:

[0077]

[0078] where d j is the DOA value of the j-th pulse in this cluster, j = 1, 2,..., N i .

[0079] S52. Extract the DOA statistical feature quantities of all candidate batch clusters through step S51. Subsequently, perform threshold judgment. If:

[0080]

[0081] It is regarded that the batch determination of the p-th cluster and the q-th cluster is successful, where ε and β are both preset thresholds; the two clusters are batched, and the DOA statistical feature quantity of the new cluster after batching is updated.

[0082] S6. Complete batching and output the final sorting result.

[0083] For the sorting performance test of the proposed radar signal sorting method based on improved affinity propagation clustering, specifically: using the PDW data of the test samples for IAP clustering to realize the pre-sorting of radar signals, extracting candidate batching centroids using the ADPC algorithm according to the pre-sorting results, and combining the statistical features of the DOA of pulse clusters for batching detection to complete the sorting of radar signals. The sorting accuracy rate is defined as the ratio of the number of correctly sorted pulses to the total number of pulses, the additional batching rate of a single radiation source is defined as the average number of clusters that the pulses of this radiation source are multi-sorted into per frame, and the purity is defined as the weighted average of the maximum proportion of the true categories to which the samples in the cluster belong for all clusters.

[0084] Table 1 shows the data set used in the test of the radar signal sorting method based on improved affinity propagation clustering in the embodiment. Among them, the first and second radars are multi-functional radars, and each radar has 3 working modes. The other 3 radars have only one working mode. The working duration of the receiver is 10 s. During this period, the pulses generated by each radar are subjected to simulation processing such as white noise interference, pulse aliasing, and pulse missing, and an aliased PDW sequence is formed. The comparison of the sorting accuracy rate obtained from the test with the AP clustering, ADPC, K-means, DBSCAN, and Dual-Mode methods is as Figure 4 shown. The overall recognition rate of the present invention reaches 98.93%; the test results of the additional batching rate and purity of sorting are shown in Table 2. The additional batching rate of sorting of the present invention is 1.66%, and the purity reaches 98.86%.

[0085] Table 1 Data set for the test of the radar signal sorting method based on improved affinity propagation clustering

[0086]

[0087] Table 2 Test results of sorting accuracy rate, additional batching rate, and purity

[0088]

[0089]

[0090] Those of ordinary skill in the art will realize that the embodiments described herein are provided to assist the reader in understanding the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present invention based on these technical revelations disclosed in the present invention, and these deformations and combinations are still within the scope of protection of the present invention.

Claims

1. A radar signal sorting method based on improved affinity propagation clustering, characterized in that: The following steps are involved: S1, detect the pulse description word parameter data of the original signal; S2, dividing the detected pulse description word parameter data into several frames according to the time frame to form a pulse data set; S3, using the improved affinity propagation algorithm to pre-sort each frame of pulse data in the data set, including the following steps: S31, for the pulse data of the kth frame Calculate the i-th pulse data point x i With the jth pulse data point x j The similarity S(x i ,x j ): Where Δp, Δc and Δd represent x i and x j The difference in pulse width, carrier frequency and angle of arrival; S32, introduce the attenuation factor α, and change x i and x j The carrier frequency difference is corrected to: In the formula, and They are x i and x j The carrier frequency value of the modified carrier frequency is substituted into (1) to update the similarity matrix; S33, iteratively update the attraction information r(i, k) and the belonging information a(i, k): r(i,k)=S(i,k)-max{a(i,k ′ )+S(i,k ′ )} (3) a(i,k)=min{0,r(k,k)+∑max {0,r(i′,k)}} (4) Among them, a(i,k ′ ) means except for the pulse data point x k The remaining pulse data points x i The attribution information, S(i,k ′ ) means except for the pulse data point x k The remaining pulse data points are i The similarity of r(i′,k) is the similarity of the pulse data point x. i The attraction information of other pulse data points to xk; S34. In order to ensure the legitimacy of the clustering results, define the constraints: c i =j then we have Right now If x i The cluster center of Must be its own cluster center; S35, continuously update the pulse data point x through the iteration of attraction information and belonging information i The corresponding cluster center Thus minimizing the energy function E[c]: Among them, the preference parameter σ represents the cost of each point as its own representative point and is a free parameter; δ(c i ,i) is a Kronecker function, that is, if and only if c i = i when δ(c i ,i)=1, otherwise δ(c i ,i)=0; Each pulse data point x is obtained by minimizing the energy function i The final corresponding cluster center is used as its cluster center; the pulse data X of the kth frame k A total of n cluster centers are obtained; S4, using the adaptive density peak clustering algorithm to extract the candidate batch centroids from the pre-sorting results; specifically comprising the following steps: S41, the pulse data X k The n pre-sorted cluster centers obtained by the improved affinity propagation algorithm are denoted as e i ,i=1,2,…,n; calculate e i The local density p i and the minimum distance d i ; S42, repeat step S41 until all pre-sorted cluster centers are traversed, according to (p i ,d i ) Draw a local density-minimum distance two-dimensional batch centroid decision diagram; S43, using the batch centroid selection method based on statistical theory to achieve adaptive selection of batch centroids: using the outlier detection method to identify the batch centroid, the method includes three steps: estimating the specific local density value p in the decision graph i and any minimum distance value p i The corresponding probability density function P y (p i ,y); using P y (p i ,y) at a specific local density value p i Calculate the mean and variance of the y distribution; use the mean and variance to detect outliers and obtain candidate batch centroids; S5, completing batch detection for each candidate batch cluster corresponding to the candidate batch centroid; S6. Complete batching and output the final sorting results.

Images