A dual-mode radar signal sorting method based on adaptive density peak clustering

Abstract

The application discloses a dual-mode radar signal sorting method based on adaptive density peak clustering, which comprises the following steps: S1, data initialization: using the three-dimensional information of the carrier frequency, pulse width and angle of arrival of the radar signal as a characteristic vector, performing normalization processing and cutting into a series of equal-length time frames according to the time length; S2, establishing an adaptive density peak clustering sorting mode; S3, establishing a Gaussian mixture model clustering sorting mode; S4, establishing a dual-mode switching strategy: the first frame of data adopts the ADPC mode, the second frame of data adopts the GMM mode; for the third frame and the subsequent data, it is determined which sorting mode is used for the input data of the current time frame based on the mode selection strategy. The dual-mode sorting of the application takes into account the slow change and mutation of the radiation source parameters in the electromagnetic environment, can effectively improve the sorting timeliness in the slow change scene of the radiation source parameters, and can still realize high-accuracy sorting in the face of the radiation source parameter mutation of the sorting scene.

Description

Technical Field

[0001] This invention belongs to the field of radar signal sorting, and specifically relates to a dual-mode radar signal sorting method based on adaptive density peak clustering. Background Technology

[0002] Radar signal sorting refers to the process by which a receiver, in electronic countermeasures, selects appropriate sorting methods to differentiate between different radar pulse sequences from interleaved pulse streams intercepted during reconnaissance. Developing radar signal sorting technology enhances the analysis and identification of radiation source signal types, radar parameters, and radar types in the environment by electronic reconnaissance measures. This provides intelligence support for targeted self-defense jamming in electronic countermeasures (ECM), playing a crucial role in modern electronic warfare.

[0003] Research on radar signal sorting methods mainly focuses on three categories: traditional sorting methods based on PRI (Primary Repetition Spectrum), machine learning sorting methods based on clustering, and deep learning methods based on neural networks. The paper "Improved algorithm for estimating pulse repetition intervals" (IEEE transactions on Aerospace and Electronic Systems, vol. 36, no. 2, p. 407-421, 2000) proposes an improved algorithm for estimating pulse repetition intervals of staggered pulse sequences using overlapping PRI bins with an offset time origin. This algorithm can effectively obtain the PRI spectrum of jittered pulse trains and further realize threshold-based average PRI detection, improving upon the problem of existing methods being easily affected by timing jitter (PRI jitter). The paper "Density-based fuzzy c-means multicenter re-clustering radar signal sorting algorithm" (in 2018, 17th IEEE International Conference on Machine Learning and Applications (ICMLA). IEEE, 2018, pp. 891-896) proposes a density-based fuzzy c-means multicenter re-clustering algorithm. It combines the advantages of the DBSCAN density clustering algorithm and the fuzzy c-means FCM clustering algorithm, and uses the CFSFDP algorithm to generate random initial centroids. This achieves superior sorting results and membership descriptions compared to existing algorithms while reducing time costs. The paper "A Novel Radar Signals Sorting Method via Residual Graph Convolutional Network" (IEEE Signal Processing Letters (2023)) proposes an RSS method based on a semi-supervised learning framework using residual graph convolutional networks to improve the generalization ability of signal sorting models in small data scenarios. The K-nearest neighbor algorithm is used to construct a graph structure from the intercepted radar signals. A three-layer ResGCN is designed to adaptively improve feature learning. Sorting is achieved through an end-to-end ResGCN with small labeled graph data containing interleaved radar signals. Compared with some existing methods, this method can achieve better average accuracy with almost no increase in computational cost when the labeled data is small. However, most existing sorting methods, including the above study, depend to some extent on the prior knowledge of the radiation source information in the dataset to be sorted, such as the number of radiation sources and their parameter distribution.

[0004] The paper "Identifying cluster centroids from decision graph automatically using a statistical outlier detection method" (Neurocomputing, vol. 329, pp. 348-358, 2019) proposed Adaptive Density Peak Clustering (ADPC), which solves the problem of manually selecting cluster centroids based on the idea of ​​outlier detection in a statistical sense. It has been proven to achieve high classification accuracy without knowing the data feature information. However, the high time cost of calculating the distance between any two cluster sample points makes this method unsuitable for the sorting of continuous PDW data. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a dual-mode radar signal sorting method based on adaptive density peak clustering. This method evaluates the real-time parameter changes of radiation sources in the electromagnetic environment and designs a dual-mode signal sorting method based on this. The dual-mode sorting can take into account both the gradual and abrupt changes in radiation source parameters in the electromagnetic environment, effectively improve the sorting timeliness in scenarios with gradual changes in radiation source parameters, and still achieve high-accuracy sorting in the face of abrupt changes in radiation source parameters in the sorting scenario.

[0006] The objective of this invention is achieved through the following technical solution: a dual-mode radar signal sorting method based on adaptive density peak clustering, comprising the following steps:

[0007] S1. Data Initialization: Using the three-dimensional information of the radar signal—carrier frequency, pulse width, and angle of arrival—as the feature vector for clustering and sorting, the feature vector is normalized and then divided into a series of equal-length time frames, denoted as X. k ;

[0008] S2. Establish an adaptive density peak clustering sorting mode, including the following steps:

[0009] S2-1. Determine the cluster centroids according to the following formula:

[0010]

[0011]

[0012] Where is p i Local density, d i It is the cutoff distance, r ij It is the eigenvector x i to the eigenvector x j The distance, rc It is the cutoff distance;

[0013] Combine the feature vector set p i d i The distribution of the density p is calculated. i Under the given conditions, the probability density p of distance y (p i ,y):

[0014]

[0015] Where y is a random variable, N is the number of sample points, and σ p It is set p i The standard deviation of all elements, σ d It is set d i The standard deviation of all elements, with α and β taken as empirical values ​​of 0.5;

[0016] Calculate the conditional expectation μ using the probability density from the previous step. y (p i ),variance Threshold TH d (p i ):

[0017]

[0018]

[0019] TH d (p i )=μ y (p i )+3×σ y (p i (6)

[0020] Outlier detection is performed using the threshold shown in formula (6), with density p i Under the condition, d i Sample points whose values ​​exceed the threshold shown in the formula are recorded as outliers, and outliers are used as cluster centroids.

[0021] S2-2. Classify the remaining sample points into the nearest category to complete the sorting of the current data frame;

[0022] S2-3. Based on the class centers obtained from the detection, generate prior information for the current time frame. Specifically, extract the number of classes C. k Mean μ of each category m Covariance Category percentage n m Where m = 1 to C k The prior knowledge S that constitutes the current time frame;

[0023] S3. Establish a Gaussian mixture model for clustering and sorting. The specific method is as follows: Initialize the Gaussian mixture model using prior knowledge S. The initialized Gaussian mixture model approximates X using a finite number of Gaussian distributions. k Distribution:

[0024]

[0025]

[0026] Among them, C k =C k-1 , P(m) is the prior probability of each class in the data and satisfies Initialize P(m) = n m ;

[0027] The Gaussian mixture model is based on maximum likelihood estimation and uses the EM iterative algorithm. It stops iterating when the parameters or the log-likelihood function converge, thus completing the iteration of the parameters P(m) and μ for the current frame. m , Estimate; use GMM for the current data frame X k Cluster and calculate n m Finally, update the prior knowledge S;

[0028] S4. Establish a dual-mode switching strategy: The first frame of data uses ADPC mode by default, and the second frame of data uses GMM mode by default; for the third frame and subsequent frames of data, the mode selection strategy determines which sorting mode to use for the input data of this time frame. Specifically, if the difference in the amount of sorted feature vector data or the difference in the evaluation index, Calinski-Harabasz-Index, is greater than the preset value, the next frame will be switched to the adaptive density peak clustering sorting mode; otherwise, the Gaussian mixture model clustering sorting mode will continue to be used.

[0029] The beneficial effects of this invention are as follows: First, it decomposes the signal sorting process into a series of sub-sorting problems divided into equal time frames. Simultaneously, it introduces prior information about the radiation source, establishing a connection between the two sorting modes based on this prior information. Finally, based on the prior information, a mode switching strategy is designed: adaptive density peak clustering is used for sorting when radiation source parameters change abruptly in the electromagnetic environment, while Gaussian mixture model clustering is used when parameters change slowly. Compared with existing technologies, the dual-mode radar signal sorting method proposed in this invention can effectively utilize the parameter characteristics of radiation sources in the electromagnetic environment for mode switching, exhibiting advantages in sorting accuracy and timeliness, and possessing strong practical value. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0031] Figure 2 This is a schematic diagram of the radar signal clustering and sorting model in this embodiment;

[0032] Figure 3 This is a schematic diagram of the radar dataset distribution in this implementation method.

[0033] Figure 4 The flowchart is shown below for the ADPC-GMM algorithm.

[0034] Figure 5 This is a comparison chart showing the effects of different methods in this embodiment; where, Figure 5 (a) is a comparison chart of weighted accuracy curves. Figure 5 (b) is a comparison chart of running times. Detailed Implementation

[0035] This invention is primarily verified using simulation experiments. All steps and conclusions have been verified correctly on the Windows 11 operating system platform using Matlab 2022b. To facilitate understanding of the technical content of this invention by those skilled in the art, the following description, in conjunction with accompanying drawings and tables, further elaborates on the invention.

[0036] like Figure 1 As shown, the present invention provides a dual-mode radar signal sorting method based on adaptive density peak clustering, comprising the following steps:

[0037] S1. Data Initialization: The three-dimensional information of the radar signal—carrier frequency, pulse width, and angle of arrival—is used as the feature vector for clustering and sorting. The radar data source is the Pulse Description Word (PDW) data stream continuously received by the reconnaissance receiver. This data stream consists of a series of PDW feature vectors sorted by Time of Arrival (TOA). A typical PDW includes five dimensions of feature information: Time of Arrival (TOA), Angle of Arrival (DOA), Carrier Frequency (CF), Pulse Width (PW), and Pulse Amplitude (PA). Clustering and sorting uses one or more dimensions of the PDW as the feature vector for clustering. For three-dimensional clustering and sorting, this invention uses CF, PW, and DOA parameters to form the feature vector. The feature vector is then subjected to min-max normalization and divided into a series of time frames of equal length according to the duration. In this embodiment, the length of the divided time frame is 100ms, and the resulting time frame is denoted as X. k The i-th eigenvector is represented as x. i The specific process is as follows: Figure 2 As shown, to illustrate the spatial distribution of the three-dimensional clustering parameters, feature mapping is used to map the set of sample points to a three-dimensional space, such as... Figure 2 As shown on the right. In this embodiment, the radar dataset distribution is shown in Appendix 1 and Appendix 2. Figure 3 As shown, where, Figure 3 (a) is a schematic diagram of the three-dimensional CF-PW-DOA distribution. Figure 3 (b) is a schematic diagram of the two-dimensional DOA-CF distribution. Figure 3 (c) is a schematic diagram of the two-dimensional DOA-PW distribution. Figure 3 (d) is a schematic diagram of the two-dimensional CF-PW distribution.

[0038] Table 1. Radar signal simulation parameters

[0039] Radar DOA / ° PW / μs RF / MHz RF modulation type Duration / s 1 52(±3) 1.28(±0.05) 6100～6300(±30) Jitter 0～5 2 49(±3) 1.32(±0.05) 6200～6400(±30) Jitter 0～5 3 52(±3) 1.37(±0.05) 6400(±30) Fixed 0～5 4 55(±3) 1.42(±0.05) 6300～6500(±30) Jitter 0～5 5 54(±3) 1.5(±0.05) 6500(±30) Fixed 0.4～5 6 58(±3) 1.55(±0.05) 6400～6600(±30) Jitter 0.4～4.5

[0040] A prior information model is established to characterize the radiation source parameter information of the previous time frame in the sorting scenario. This model serves as the initialization parameter for the Gaussian Mixture Model (GMM) clustering algorithm and the basis for mode switching decisions. Specific information includes the number of radiation sources, the mean values ​​of the three-dimensional parameters (radiation source carrier frequency, pulse width, and angle of arrival). When establishing the information model, the temporal continuity of PDW data is considered; radar parameters typically do not change frequently within a short period, meaning that the PDW data sources for multiple consecutive time frames are all fixed radars in the sorting scenario. Prior knowledge is defined to describe the radars in the sorting scenario, including four parts: the number of radar categories, the mean of the categories, the covariance, and the category proportion. Prior knowledge is updated frame by frame. A decision criterion is used to determine the validity of the prior knowledge from the previous time frame. Two adjacent frames are considered valid prior information, having the same number of radar categories, while minor differences are allowed in the other three parts of the information.

[0041] The ADPC-GMM algorithm flow is as follows: Figure 4 As shown, the process includes two steps: establishing an adaptive density peak clustering (ADPC) sorting pattern and establishing a Gaussian mixture model clustering sorting pattern.

[0042] S2. Establish an adaptive density peak clustering (ADPC) sorting pattern, including the following steps:

[0043] S2-1. Determine the cluster centroids according to the following formula:

[0044]

[0045]

[0046] Where is p i Local density, d i It is the cutoff distance, r ij It is the eigenvector x i to the eigenvector x j The distance, r c It is the cutoff distance; because p i and di The calculation involves all feature vectors, therefore the time cost of ADPC mode is O(N). 2 ).

[0047] Combine the feature vector set p i d i The distribution of the density p is calculated. i Under the given conditions, the probability density p of distance y (p i ,y):

[0048]

[0049] Where y is a random variable, N is the number of sample points, and σ p It is set p i The standard deviation of all elements, σ d It is set d i The standard deviation of all elements, with α and β taken as empirical values ​​of 0.5;

[0050] Calculate the conditional expectation μ using the probability density from the previous step. y (p i ),variance Threshold TH d (p i ):

[0051]

[0052]

[0053] TH d (p i )=μ y (p i )+3×σ y (p i (6)

[0054] Outlier detection is performed using the threshold shown in formula (6), with density p i Under the condition, d i Sample points whose values ​​exceed the threshold shown in the formula are recorded as outliers, and outliers are used as cluster centroids (class centers).

[0055] S2-2. Classify the remaining sample points into the nearest category with higher density to complete the sorting of the current data frame;

[0056] S2-3. Based on the class centers obtained from the detection, generate prior information for the current time frame. Specifically, extract the number of classes C. k Mean μ of each category m Covariance Category percentage n mWhere m = 1 to C k The prior knowledge S that constitutes the current time frame;

[0057] S3. Establish a Gaussian mixture model for clustering and sorting. The specific method is as follows: Initialize the Gaussian mixture model using prior knowledge S. The initialized Gaussian mixture model approximates X using a finite number of Gaussian distributions. k Distribution:

[0058]

[0059]

[0060] Among them, C k =C k-1 , P(m) is the prior probability of each class in the data and satisfies Initialize P(m) = n m ;

[0061] The Gaussian mixture model is based on maximum likelihood estimation and uses the EM iterative algorithm. It stops iterating when the parameters or the log-likelihood function converge, thus completing the iteration of the parameters P(m) and μ for the current frame. m , Estimate; use GMM for the current data frame X k Cluster and calculate n m Finally, update the prior knowledge S;

[0062] S4. Establish a dual-mode switching strategy: The first frame of data uses ADPC mode by default, and the second frame of data uses GMM mode by default; for the third frame and subsequent frames of data, the mode selection strategy determines which sorting mode to use for the input data of this time frame. Specifically, if the difference in the amount of sorted feature vector data or the difference in the evaluation index, Calinski-Harabasz-Index, is greater than the preset value, the next frame will be switched to the adaptive density peak clustering sorting mode; otherwise, the Gaussian mixture model clustering sorting mode will continue to be used.

[0063] The validity of prior information from the previous time frame is determined based on the following two criteria:

[0064] 1) Current time frame PDW data volume N k Compared to the previous frame N k-1 Significant changes can be described using the following formula:

[0065] |N k -N k-1 |>TH n ×N k-1 (9)

[0066] Among them THn It is a threshold, set to TH based on experience. n =0.1. The most direct possibility for this situation is that a radar suddenly appears or disappears in the sorting scene of the current time frame, causing a change in data density. Alternatively, it may be due to a change in the radar's working mode, causing a change in its PRI.

[0067] 2) There is a significant difference in the clustering results between the first two frames. To evaluate the clustering results, the Calinski-Harabasz Index (CHI), an evaluation metric that does not require prior knowledge of the actual sorting labels, is introduced and calculated according to the following formula:

[0068]

[0069] Where N k Q is the number of samples in the k-th frame. m It is the feature set of the m-th class, q m The center is the number of samples in the m-th class. m It is the class center of the m-th category, g c It refers to the global cluster centers. CHI measures intra-cluster density by calculating intra-cluster covariance and inter-cluster density by calculating inter-cluster covariance. Higher values ​​indicate denser clusters and more dispersed clusters, resulting in better clustering. Based on CHI, the following formula describes the significant difference in clustering results between the first two frames:

[0070] |CHI k-2 -CHI k-1 |>TH c ×CHI k-2 (11)

[0071] Among them TH c It is a threshold, set to TH based on experience. c =0.1. Even if there is no significant change in the amount of PDW data in the sorting scenario of adjacent time frames, the sudden change in CF and PW parameters of the current frame caused by the switching of radar working mode will lead to the failure of prior knowledge and a decrease in the clustering index CHI of the current frame. Therefore, formula (10) is used to capture the change of CHI.

[0072] When one of the above two situations occurs, prior knowledge is no longer valid: the current frame uses the adaptive density peak clustering sorting mode, which means jumping to the processing of step S2; otherwise, the Gaussian mixture model sorting mode is maintained, which means jumping to the processing of step S3.

[0073] To evaluate the sorting methods, comparison methods were set up in the same dataset scenario, including the following four: Gaussian Mixture Model Clustering with Random Initialization Strategy (RandSample-GMM), Gaussian Mixture Model Clustering with Kmeans Clustering Initialization Strategy (Kmeans++-GMM), Adaptive Density Peak Clustering (ADPC), and the dual-mode clustering based on Adaptive Density Peak Clustering proposed in this invention (ADPC-GMM).

[0074] Since both algorithms require the number of clusters as input, the elbow rule is used to infer the degree of distortion according to the formula:

[0075]

[0076] Here, k is the number of categories. The degree of distortion decreases with iterative increases in the number of categories k. The goal is to find the k value that causes a sharp decrease in distortion as the number of clusters. A weighted accuracy is introduced to evaluate the sorting accuracy of the above algorithm on an average basis for all radiation source categories.

[0077]

[0078] The following is a comparison of the sorting performance of several different algorithms:

[0079] 1) Sorting accuracy of several algorithms for different radars: Table 2 shows the sorting accuracy of the four methods for each type of radiation source, based on the average value across all detection time frames. Except for ADPC-GMM, the algorithms show differences in their identification accuracy for different radiation sources. This is because there is a certain degree of aliasing between the parameters of different radiation sources, making source sorting difficult. Comparing ADPC-GMM with RandSample-GMM and Kmeans++-GMM, it can be seen that the prior knowledge S makes the initial model value close to the convergent global optimum, ensuring that the iteration process does not get trapped in local optima, thus improving the sorting accuracy for each category. Comparing ADPC-GMM with ADPC, it can be seen that GMM-mode has a sorting accuracy comparable to ADPC in most categories, and maintains comparable sorting capability to other radars even in cases like Radar3 where parameters overlap with multiple radars.

[0080] Table 2. Sorting accuracy of different radars

[0081]

[0082]

[0083] 2) Weighted Precision: The frame-by-frame weighted average precision of the four methods is as follows: Figure 5As shown in (a), RandSample-GMM initializes the class centers of the GMM using random sampling, while Kmeans++-GMM, which uses Kmeans++ for initialization, lacks the ability to identify parameter distributions with non-spherical shapes in the dataset. These inherent limitations mean that the class centers determined by the initialization algorithms may differ significantly from the true class centers, potentially causing the model to converge to a local optimum. This is reflected in the large fluctuations in the accuracy of the two algorithms shown in the figure. In contrast, ADPC-GMM utilizes prior knowledge S to provide GMM initial values ​​close to the global optimum for each time frame, resulting in more stable and higher frame-by-frame sorting accuracy. Furthermore, examining the accuracy in frames 5 and 46 confirms that the algorithm can recognize changes in the number of radars and switch to ADPC-mode. Compared to ADPC, GMM-mode exhibits similar sorting accuracy performance.

[0084] 3) Time cost: The frame-by-frame execution time of the four methods is as follows: Figure 5 As shown in (b), a comparison of ADPC-GMM with RandSample-GMM and Kmeans++-GMM reveals that, thanks to prior knowledge optimizing the number of GMM iterations, the ADPC-GMM algorithm has a lower time cost in most time frames, with the exception of when the algorithm uses ADPC mode for sorting. A comparison of ADPC-GMM and ADPC shows that using GMM-mode significantly reduces the time cost required for operation compared to ADPC, and indeed accelerates sorting in multi-frame PDW sorting scenarios.

[0085] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A dual-mode radar signal sorting method based on adaptive density peak clustering, characterized in that, Comprising the following steps: S1, data initialization: use the carrier frequency, pulse width, and angle of arrival of the radar signal as the feature vector for clustering and sorting. Normalize the feature vector, and then divide it into a series of equal-length time frames according to the time length, denoted as X k ; S2, establishing an adaptive density peak clustering sorting mode, comprising the following steps: S2-1, determining the clustering centroid according to the following formula: where p i local density, d i is the cutoff distance, r ij is the feature vector x i to the feature vector x j is the distance, r c is the cutoff distance; The set of feature vectors p i , d i is calculated from the distribution of the distances p i The probability density p y (p i , y) is calculated under the condition that the distance is where y is a random variable, N is the number of sample points, σ p is the set p i is the standard deviation of all elements, σ d is the set d i is the standard deviation of all elements, α, β take the empirical value 0.5; Calculate the conditional expectation value μ using the probability density from the previous step y (p i ) and variance Threshold value TH d (p i ): TH d (p i ) = μ y (p i )+3×σ y (p i ) (6) The threshold value shown in formula (6) is used to carry out outlier detection, and the density p i Under the condition, d i The sample point whose value exceeds the threshold value shown in formula (6) is recorded as an outlier, and the outlier is taken as a cluster center. S2-2, classifying the remaining sample points to the nearest category, and completing the sorting of the current data frame; S2-3, based on the obtained class center, generating prior information of the current time frame, and the specific method is to extract the number of categories C k , the mean value μ of each category m , the covariance The proportion of each category n m , where m = 1 ~ C k , the prior knowledge S of the current time frame S3, a Gaussian mixture model clustering sorting mode is established, and the specific method is as follows: initializing a Gaussian mixture model by using prior knowledge S, and the initialized Gaussian mixture model is approximated to the distribution of X by a limited number of Gaussian distributions k ​ where C k = C k-1 , P(m) is the prior probability of each class in the data and satisfies Initialize P(m) = n m ; The Gaussian mixture model is based on maximum likelihood estimation, adopts an EM iterative algorithm, stops iteration when parameters or a log likelihood function converges, and completes estimation of parameters P(m), μ m , of the current frame; the current data frame X k is clustered and n m is calculated using the GMM, and prior knowledge S is finally updated; S4, establishing a switching strategy of the double modes: the first frame of data adopts the ADPC mode by default, the second frame of data adopts the GMM mode by default; for the third frame and the data thereafter, it is determined which sorting mode is used for the input data of the current time frame based on the mode selection strategy, and the specific method is: when the difference between the adjacent time frame sorting feature vector data quantity or the difference between the evaluation index Calinski-Harabasz-Index is greater than a preset value, the next frame is switched to the adaptive density peak clustering sorting mode, otherwise the Gaussian mixed model clustering sorting mode is continued to be used.

Images