Integrated clustering radar signal sorting method based on data enhancement and quality screening

Abstract

The application provides an integrated clustering radar signal sorting method based on data enhancement and quality screening, comprising: preprocessing a received pulse description word to obtain a normalized parameter matrix; based on the normalized parameter matrix, obtaining multiple groups of weighted feature data by randomly weighting the pulse description word, and processing the multiple groups of weighted feature data by using multiple clustering algorithms to generate multiple base clustering results; screening an optimal base clustering set from the multiple base clustering results according to signal physical characteristics and differences; and performing consensus fusion on the optimal base clustering set by using spectral clustering to obtain a radar signal sorting result. The application generates a diversified base clustering pool by mixing multiple view feature weighting and multiple clustering algorithms, and combines a quality screening mechanism based on a radar signal physical constraint to effectively eliminate low-quality division results, thereby significantly improving the accuracy and robustness of radar signal sorting in a complex electromagnetic environment.

Description

Technical Field

[0001] This application belongs to the field of radar radiation source signal technology, specifically relating to an integrated clustering radar signal sorting method based on data augmentation and quality screening. Background Technology

[0002] Radar source signal sorting is a crucial component of electronic reconnaissance systems. Its core task is to separate the independent pulse sequences of each radiation source from aliased and interleaved pulse streams, providing fundamental intelligence support for subsequent situation assessment and electronic countermeasures decision-making. With the increasing complexity of modern electronic warfare environments, the density of radar signals in the battlefield electromagnetic space has risen dramatically. New radars widely employ advanced technologies such as frequency agility, staggered repetition rates, and intra-pulse modulation, resulting in severe overlap and aliasing of different radar radiation sources in the pulse descriptor space. Against this backdrop, traditional sorting methods relying on single feature dimensions, fixed sorting rules, or simple clustering criteria are no longer sufficient to accurately characterize the multidimensional structural characteristics of complex radar signals, leading to a significant decrease in sorting accuracy. Therefore, achieving accurate deinterleaving of radar radiation source pulse sequences in high-density, strongly overlapping signal environments has become a critical technical challenge that urgently needs to be overcome in the field of electronic reconnaissance.

[0003] The radar signal sorting method based on cluster analysis measures the similarity between characteristic parameters of pulse stream signals using statistical methods. Its core objective is to divide aliased radar pulse signals into several "clusters" subsets, and to satisfy the principles of "maximizing intra-cluster similarity" and "maximizing inter-cluster separation" as much as possible during the division process, thereby achieving effective separation of different radiation sources (Sui Jinping, Liu Zhen, Liu Li, et al. Research progress on radar radiation source signal sorting [J]. Journal of Radar, 2022(11)). Under this technical framework, the performance of the clustering algorithm directly determines the final sorting effect. According to the differences in algorithm mechanism, existing research mainly covers multiple directions such as partitioning clustering, hierarchical clustering, density clustering, and fuzzy clustering. Among them, partitioning clustering, represented by K-means, is the most widely used basic method. It achieves signal sorting by pre-dividing data into K clusters and iteratively optimizing them. Hierarchical clustering organizes data by constructing nested phylogenetic trees, without making any assumptions about the shape of the clusters, and shows strong adaptability when dealing with non-convex radar signals. Density clustering assumes that the inter-pulse parameters of the same radiation source signal are continuously distributed in the density domain. Theoretically, it can discover clusters of arbitrary shapes and has excellent noise resistance. Fuzzy clustering introduces fuzzy set theory, allowing samples to belong to multiple categories with a certain membership degree, and can more objectively reflect the inherent fuzzy boundary characteristics of radar signal parameters.

[0004] To further enhance the robustness of clustering, ensemble clustering has been proposed as an effective improvement approach, which consists of two core parts: "base cluster generation" and "consensus ensemble". Regarding base cluster generation, existing methods mainly include: using different clustering algorithms, setting different parameter values ​​in the same algorithm, using subsets of different features, using subsets of different data, and projecting the dataset into subspaces, among other generation strategies. Consensus ensemble is responsible for fusing different base clustering results to generate the final cluster partition. Currently, the main approaches to implementing consensus functions include methods based on co-linked matrices (FRED AL, JAIN A K. Combining multiple clusterings using evidence accumulation[J].IEEE Transactions on Pattern Analysis and Machine Intelligence, 2005, 27(6):835-850), voting-based methods, hypergraph partitioning-based methods (e.g., bipartite graph partitioning, FERN XZ, BRODLEY CE. Solving cluster ensemble problems by bipartite graph partitioning[C]. Proceedings of the Twenty-First International Conference on Machine Learning(ICML), Banff, Alberta, Canada, July 4-July 8, 2004: 36), information theory-based methods, and hybrid model-based methods.

[0005] However, the aforementioned existing technical solutions still have significant technical shortcomings in practical applications. First, traditional single-cluster sorting algorithms are highly dependent on data distribution assumptions, exhibiting poor adaptability and limited sorting accuracy when faced with severe parameter overlap and non-ideal sampling data in complex and agile signals. Furthermore, existing ensemble clustering algorithms are rarely used in radar signal sorting, and their base clustering generation strategies are mostly based on a single perspective, making it difficult to fully mine the multidimensional structural information of the data from different feature subspaces and different algorithmic assumptions. This results in insufficient diversity in the base cluster pool, reducing the upper limit of the ensemble fusion gain. Second, existing ensemble clustering methods generally lack a scientific evaluation and screening mechanism for the quality of base cluster members. This leads to the potential inclusion of a large number of low-quality partitioning results in the ensemble pool, such as incorrectly dividing the same radiation source into multiple fragments or confusing different radiation sources into a single cluster. These inferior base members introduce a significant error accumulation effect during consensus fusion, directly limiting the improvement of the final sorting accuracy. Furthermore, when faced with non-ideal electromagnetic environments such as high pulse loss rates and strong false interference, existing clustering methods have low tolerance for incomplete data and anomalous samples, resulting in a sharp deterioration in sorting performance and making it difficult to meet the application requirements of practical electronic reconnaissance systems for high reliability and strong robustness. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this application provides an integrated clustering radar signal sorting method based on data augmentation and quality screening. The technical problem to be solved by this application is achieved through the following technical solution: An integrated clustering radar signal sorting method based on data augmentation and quality screening includes: S1, preprocess the received pulse descriptor to obtain the normalized parameter matrix; S2, based on the normalized parameter matrix, multiple sets of weighted feature data are obtained by randomly weighting the pulse descriptor, and multiple clustering algorithms are used to process the multiple sets of weighted feature data to generate multiple base clustering results; S3, Based on the physical characteristics and differences of the signals, select a preferred set of base clusters from the multiple base clustering results; S4. Consensus fusion of the preferred base cluster set is performed using spectral clustering to obtain the radar signal sorting result.

[0007] Compared with the prior art, this application has the following beneficial effects: 1. This application generates multi-view data representation by random feature weighting and constructs a base cluster pool by combining three clustering algorithms with different mechanisms: K-means, Gaussian mixture model and hierarchical clustering. It fully mines the multi-dimensional structural information of radar signals from multiple feature subspaces and multiple data distribution assumptions, effectively enhancing the differences between base cluster members. It overcomes the shortcomings of single-view clustering in capturing complex data structures and the lack of differentiation of existing integrated cluster base members, providing a high-quality input foundation for subsequent consensus fusion.

[0008] 2. This application innovatively introduces a quality assessment system based on the physical constraints of radar signals. By calculating the standard deviation of each cluster in the angle of arrival dimension, it determines whether its spatial aggregation degree conforms to the physical characteristics of radar signal point source emission. It also introduces signal agility reward and stability reward by combining the standard deviation of pulse width and radio frequency dimensions. This can effectively identify and eliminate low-quality classification results that incorrectly segment the same radiation source or confuse different radiation sources. It avoids the error accumulation effect introduced by inferior base members in the consensus fusion process from the source, and significantly improves the physical consistency and reliability of the sorting results.

[0009] 3. This application adopts the maximum marginal relevance strategy to iteratively screen the base clustering results. While ensuring that the selected base members have high-quality scores, it maximizes their differences from the selected set, taking into account both the quality and diversity of base clustering, and effectively solves the problems of high redundancy and insufficient diversity in ensemble learning.

[0010] 4. This application constructs a weighted consensus matrix and uses the quality score of each base clustering result as the basis for weight allocation, so that high-quality base clusters can make a greater contribution to consensus fusion. It combines spectral theory to perform feature decomposition and automatically estimates the optimal number of clusters through feature gap statistics, thereby achieving adaptive and accurate determination of the number of unknown radar radiation sources and avoiding the limitation of traditional methods that require pre-specifying the number of clusters.

[0011] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating an integrated clustering radar signal sorting method based on data augmentation and quality screening provided in this application. Figure 2 This is a schematic diagram showing the accuracy comparison curves between the single clustering algorithm and different missing / false pulse ratios provided in this application; Figure 3 This is a schematic diagram showing the accuracy comparison curves between the algorithm and the classical ensemble clustering algorithm under different leak / false pulse ratios provided in this application. Detailed Implementation

[0013] The present application will be described in further detail below with reference to specific embodiments, but the implementation of the present application is not limited thereto.

[0014] like Figure 1 As shown, this application provides an integrated clustering radar signal sorting method based on data augmentation and quality screening, including: S1, preprocess the received pulse descriptor to obtain the normalized parameter matrix; S2, based on the normalized parameter matrix, multiple sets of weighted feature data are obtained by randomly weighting the pulse descriptor, and multiple clustering algorithms are used to process the multiple sets of weighted feature data to generate multiple base clustering results; This step introduces random feature weighting in the data space to construct a multi-view description. Then, in the algorithm space, three clustering models with different principles are mixed: K-Means, Gaussian mixture model, and HC clustering, to generate a highly differentiated base cluster pool.

[0015] S3, Based on the physical characteristics and differences of the signals, select a preferred set of base clusters from the multiple base clustering results; S4. Consensus fusion of the preferred base cluster set is performed using spectral clustering to obtain the radar signal sorting result.

[0016] This application provides an integrated clustering radar signal sorting method based on data augmentation and quality screening, comprising: preprocessing received pulse descriptors to obtain a normalized parameter matrix; based on the normalized parameter matrix, obtaining multiple sets of weighted feature data by randomly weighting the pulse descriptors, and processing the multiple sets of weighted feature data using various clustering algorithms to generate multiple base clustering results; selecting a preferred base clustering set from the multiple base clustering results according to the signal physical characteristics and differences; and performing consensus fusion on the preferred base clustering set using spectral clustering to obtain the radar signal sorting result. This application generates diverse base clustering pools by combining multi-view feature weighting with multiple clustering algorithms, and combined with a quality screening mechanism based on radar signal physical constraints, effectively eliminating low-quality partitioning results and significantly improving the accuracy and robustness of radar signal sorting in complex electromagnetic environments.

[0017] In one specific embodiment of this application, S1 includes: S11, combine the arrival angle, radio frequency, arrival time, pulse width and pulse amplitude parameters of each received pulse to form a pulse description word; The radar reconnaissance system receives individual pulse signals and processes them through a time-frequency receiver, a time-phase receiver, and a parameter measurement module to obtain the pulse's DOA, RF, TOA, PW, and PA. These parameters are combined to form a pulse descriptor word (PDW). The PWD of each pulse is then combined to form the PDW, which can be represented as: (1) in This indicates the total number of radiation sources. Indicates the total number of pulses. This indicates the radiation source number to which the pulse belongs. Indicates the pulse sequence number.

[0018] S12, combine the pulse descriptors of multiple pulses to form a parameter matrix; parameter matrix It can be represented as: (2) S13, standardize the parameter matrix to obtain a normalized parameter matrix.

[0019] The parameter matrix is ​​standardized using the following formula: (3) In the formula, This represents the mean vector of the original parameter matrix X across each feature dimension. This represents the standard deviation vector of the parameter matrix X along each feature dimension.

[0020] In one specific embodiment of this application, S2 includes: S21, a variety of different weight allocation templates are preset and they are combined into a weight template set; wherein, the weight allocation template includes the weights corresponding to the radio frequency parameter, pulse width parameter and arrival time parameter respectively; By enhancing the parameters and combining them with different weight ratios, unique feature perspectives are generated. The algorithm pre-sets... A set of different weight allocation templates is formed, the first... Each template is represented as: (4) in, These are the corresponding weights for the RF parameters. These are the corresponding weights for the PW parameters. These are the corresponding weights for the TOA parameters.

[0021] S22, introduce a random perturbation factor into each weight template in the weight template set to generate multiple sets of perturbed feature weights; wherein, the random perturbation factor is a multidimensional uniform random vector with a value range of [0,1], used to jitter the weights of each parameter in the weight allocation template, and the perturbed feature weights are obtained by the weight allocation template and the random perturbation factor.

[0022] To construct a unique feature view for each base clustering generator, a perturbation factor is introduced. It is a 3-dimensional uniform random vector with values ​​in the range [0,1], which implements weights. Jitter. The specific weights used for each set of parameters. for: (5) S23, the normalized parameter matrix is ​​weighted using the perturbed feature weights of each group to generate multiple sets of weighted feature data; The weighted feature data of this application are : (6) in, This represents the Hadamard product. The method creates a difference in perspective, allowing subsequent clustering algorithms to uncover structural information under different feature combinations.

[0023] S24. For each group of weighted feature data, K-means clustering algorithm, Gaussian mixture model clustering algorithm and hierarchical clustering algorithm are mixed according to a preset ratio for processing, and the target cluster number of each algorithm is randomly sampled within a preset interval to generate multiple base clustering results.

[0024] The algorithm combines three classic algorithms—K-Means, GMM, and HC—with completely different assumptions about data distribution in a specific ratio, and the number of target clusters is not a fixed value but falls within a preset range. Uniform random sampling is performed within the matrix. M base clusters are obtained by processing the feature matrix using various base clusterers.

[0025] In one specific embodiment of this application, S3 includes: S31, for each candidate partition in the multiple base clustering results, extract all clusters contained in the candidate partition, and calculate the standard deviation of each cluster in the angle of arrival dimension, pulse width dimension and radio frequency dimension respectively; This application performs a sequential traversal and physical quality evaluation of each cluster partition in the candidate pool. For the first... Candidate partitions Extract all the clusters it contains, and for each cluster Calculate the standard deviations of its three key dimensions: DOA, PW, and RF, and denote them as follows: , , .

[0026] S32, based on the relationship between the standard deviation of each cluster in the angle of arrival dimension and the preset standard deviation threshold of the angle of arrival, determine whether the spatial aggregation degree of each cluster meets the physical characteristics of radar signal point source transmission, and combine the standard deviation of each cluster in the pulse width dimension and the radio frequency dimension to introduce signal agility reward and stability reward, calculate the physical quality score of each cluster, and then obtain the overall quality score of each candidate division. S33, the candidate partition with the highest overall quality score among the multiple base clustering results is selected into the preferred set. In each iteration, the comprehensive marginal gain of each remaining candidate partition is calculated based on the overall quality score of the remaining candidate partitions and their normalized mutual information with the selected partitions in the preferred set. The candidate partition with the largest comprehensive marginal gain is then moved into the preferred set until the number of candidate partitions in the preferred set reaches the preset preferred number threshold. The iteration is then stopped, and the preferred base clustering set is obtained.

[0027] In one specific embodiment of this application, S32 includes: S321. For each cluster, determine whether the sample size of the cluster satisfies statistical significance. If it does, and the standard deviation of the cluster in the angle of arrival dimension is greater than or equal to the preset standard deviation threshold of the angle of arrival, then determine that the spatial aggregation degree of the cluster does not meet the physical characteristics of radar signal point source emission, and set the quality score of the cluster to zero. This step calculates the validity indicator function and space compactness. Validity indicator function The calculation method is as follows: A DOA standard deviation threshold is set. Once the sample size of a cluster meets the statistical significance and This means that the cluster is too spatially discrete, which does not conform to the physical characteristics of radar signal point source emission. Therefore, the mass score of the cluster is directly set to zero. .on the contrary, .

[0028] S322, For a cluster whose spatial aggregation satisfies the physical characteristics, the spatial compactness score of the cluster is calculated using a Gaussian kernel function based on the standard deviation of the cluster in the angle of arrival dimension. The Gaussian kernel function uses the standard deviation of the angle of arrival as an input variable and controls the rate at which the score decays as the standard deviation increases through a scale parameter. If the clustering degree of DOA meets the requirements, a basic score is calculated based on the standard deviation of DOA using a Gaussian kernel function to reflect the spatial compactness of the cluster. Spatial compactness is defined as: (7) in, The scale parameter used to control the decay rate.

[0029] S323, based on the standard deviation of the cluster in the pulse width dimension or radio frequency dimension, a signal agility reward is introduced. The signal agility reward is given when the standard deviation of the pulse width dimension or radio frequency dimension exceeds the corresponding preset threshold, provided that the angle of arrival dimension remains stable. This step introduces a signal agility reward mechanism: if a cluster maintains a stable DOA but has a large standard deviation in either its PW or RF, this usually means that the algorithm has successfully correlated frequency-agile or staggered signals from the same radar. Therefore, additional connectivity and jump rewards are given. .

[0030] S324, Based on the relationship between the standard deviation of the cluster in the radio frequency dimension and the preset radio frequency stability threshold, a stability reward is introduced, which is granted when the standard deviation in the radio frequency dimension is less than the radio frequency stability threshold; This step sets the RF stability threshold. For clusters with extremely stable RF, i.e. Provide stability rewards .

[0031] S325, the spatial compactness score, the signal agility reward, and the stability reward are added together to obtain the physical quality score of the cluster; Any single cluster The formula for calculating the physical quality score is: (8) In the formula, It's a jump reward. It's a stability reward.

[0032] S326, average the physical quality scores of all clusters within the candidate partition to obtain the global base score of the candidate partition; After completing the cluster-level scoring, the algorithm will divide... The average score of all clusters within the cluster is used as the global base score for this clustering partition. .

[0033] S327, determine whether the number of clusters in the candidate partition is less than a preset threshold. If so, apply a decay coefficient penalty to the global base score of the candidate partition to obtain the overall quality score of the candidate partition. To prevent the algorithm from degenerating by generating too few clusters, a penalty term is included in the logic. When the number of clusters generated is less than a preset threshold, the total score is penalized by a decay coefficient. The final score is: (9) In the formula, It is the attenuation coefficient, and its value ranges from (0,1).

[0034] S328, Traverse all candidate partitions in the multiple base clustering results, obtain the overall quality score of each candidate partition, and perform normalization processing.

[0035] After all candidate partitions have been traversed, the resulting quality score vector will be normalized to the [0, 1] interval to prepare for subsequent selection. Finally, the... Overall quality score of candidate partitions The calculation formula is: (10) In the formula, This represents the clustering result of the m-th candidate cluster; This indicates that in a certain candidate partition, the first... Clusters.

[0036] In one specific embodiment of this application, S33 includes: S331, In each iteration, for the remaining candidate partitions that are not selected into the preferred set, calculate the normalized mutual information between each remaining candidate partition and the selected partitions in the preferred set. A maximum marginal relevance (MMR) strategy is employed for iterative selection to address the diversity issue in ensemble learning. First, the partitions with the highest quality scores are directly selected into the "optimal set". Then, an iterative loop begins. In each iteration, for the remaining unselected candidate partitions, their relationship with the current "preferred set" is calculated. The maximum normalized mutual information (NMI) of the partition already exists. The NMI is calculated as follows: (11) in, The definition is as follows: (12) The definition is as follows: (13) in, and Representing candidate partitions With the selected division The number of clusters contained in it. Indicates division In this process, a randomly selected radar pulse sample is assigned to the first... The marginal probabilities of each cluster Similarly, Indicates division In this case, the sample was classified as the first... The marginal probabilities of each cluster . This represents the joint probability, that is, the probability that the pulse sample is divided into... China was classified as the first Clusters, and in the division China was simultaneously classified as the first The probability of each cluster.

[0037] S332, Calculate the comprehensive marginal gain of each remaining candidate partition based on the overall quality score of each remaining candidate partition and its maximum normalized mutual information with the selected partitions in the preferred set; The algorithm calculates the comprehensive marginal gain for each candidate based on the MMR formula, which weights the difference between its physical quality score and the selected set through a balancing factor. The formula for calculating the MMR score is: (14) In the formula, It is a balancing factor, and its value range is [0,1].

[0038] Select Gain The largest partition, move it into The termination condition for the iteration loop is: when the 'preferred set' is reached... The number of candidate partitions in the data reaches the preset preferred number threshold. The loop terminates when the specified time is reached.

[0039] S333, move the candidate partition with the largest comprehensive marginal gain into the preferred set; S334, Repeat steps S331 to S333 until the number of candidate partitions in the preferred set reaches the preset preferred number threshold, and obtain the preferred base cluster set.

[0040] In one specific embodiment of this application, S4 includes: S41, Calculate the weight of each base clustering result based on the overall quality score of each base clustering result in the preferred base clustering set; the weight is positively correlated with the overall quality score; After receiving base cluster label sets generated from different parameter configurations or algorithms, the system assigns weights to each result based on its quality score. The resulting optimized base cluster set is then selected. Include The result of the first base clustering is then the first base clustering result. The weights of each basis clustering result are defined as follows: (15) S42, construct a weighted consensus matrix based on the weights of each base clustering result, and perform spectral decomposition on the weighted consensus matrix to obtain an eigenvalue sequence; each element value in the weighted consensus matrix represents the cumulative weighted probability that two corresponding samples are classified into the same cluster in all base clustering results; S43, estimate the optimal number of clusters based on the feature value sequence, and complete the sample division in the normalized feature space based on the optimal number of clusters to obtain the radar signal sorting result.

[0041] In one specific embodiment of this application, S42 includes: S421. Traverse all base clustering results. For any two samples, if they are classified into the same cluster in any base clustering result, then accumulate the weight corresponding to the base clustering result to the matrix element position corresponding to the two samples. After traversal, use the accumulated value of each element position as the cumulative weighted probability of the two samples to construct the weighted consensus matrix. A sparse weighted consensus matrix was constructed. Each element in this matrix represents the cumulative probability that the corresponding two samples are grouped into the same cluster across all weighted base clustering results. Matrix elements represent samples With sample The weighted probability of being assigned to the same cluster in all base clustering results. This transforms discrete label information into a continuous measure of sample similarity. The formula for calculating the sparse matrix is: (16) S422, calculate the degree matrix of the weighted consensus matrix, perform a normalized Laplace transform on the weighted consensus matrix based on the degree matrix to obtain a normalized affinity matrix, and perform eigenvalue decomposition on the normalized affinity matrix to obtain an eigenvalue sequence arranged in descending order.

[0042] Subsequently, the algorithm utilizes Spectral Graph Theory (SGT) to transform the consensus matrix into a partitioning problem in a low-dimensional feature space. By calculating the degree matrix of the consensus matrix and performing a normalized Laplace transform, the algorithm constructs a normalized affinity matrix to eliminate the influence of uneven sample density distribution. Affinity Matrix Defined as: (17) Wherein, degree matrix The diagonal elements are represented as: (18) Next, the affinity matrix is ​​decomposed into eigenvalues ​​to obtain a set of eigenvalues ​​arranged in descending or ascending order. .

[0043] In one specific embodiment of this application, S43 includes: S431, calculate the difference between adjacent feature values ​​in the feature value sequence, and take the feature value index corresponding to the maximum difference as the optimal number of clusters; This step automatically determines the optimal number of clusters K. The method introduces a feature gap statistic, the mathematical formula of which is: (19) S432, Select a corresponding number of feature vectors according to the optimal number of clusters, and construct a feature matrix by row from the selected feature vectors; By analyzing the point of maximum drop in the descent rate of the eigenvalue sequence, a judgment criterion is adopted. The system automatically estimates the true number of potential radar radiation sources, using this as a basis for determining the internal cluster structure.

[0044] S433, perform row normalization on the feature matrix to project it onto the unit sphere; S434, on the projected unit sphere, each row vector of the feature matrix is ​​used as the feature representation of the sample point, and the K-means algorithm based on cosine distance is used to cluster each row vector to obtain the final clustering label of each pulse sample as the radar signal sorting result.

[0045] The selected feature vectors are assembled into a feature matrix and projected onto a unit sphere. Finally, the K-Means algorithm based on cosine distance is used to complete the sample partitioning in the normalized feature space, realizing the mapping from local similarity to global structure partitioning.

[0046] The effectiveness of this application can be further illustrated by the following simulation experiments.

[0047] The experiment used a hybrid pulse stream dataset containing radiation sources from six different radar systems, including fixed-frequency, frequency-agile, frequency-diversity, and jitter signals. During the basal cluster generation stage, a basal cluster pool of size M=150 was generated. After physical quality screening, the 10 highest-quality and most differentiated basal clusters were retained to participate in weighted spectral clustering ensemble.

[0048] The experiment was conducted under extreme conditions with a leak pulse ratio of 0% to 50% and a spurious pulse ratio of 0% to 50%. Figure 2 Accuracy comparison curves of the proposed method and a single clustering algorithm under different missing / false pulse ratios are presented. Figure 2It can be seen that the F1 score of the algorithm in this application is always stable above 90%, while the sorting accuracy of each single clustering algorithm decreases significantly with the increase of pulse loss rate and false pulse ratio, indicating that this application effectively overcomes the defect of poor robustness of traditional single clustering algorithms in complex electromagnetic environment.

[0049] Figure 3 Accuracy comparison curves of the proposed method, classical ensemble clustering algorithms (i.e., evidence-based ensemble clustering and standard bipartite graph partitioning algorithms), are presented under different missing / false pulse ratios. Figure 3 As can be seen, when faced with up to 50% missing pulses, this application can effectively piece together global consensus using local residual information, maintaining stable model performance. Even in environments with 40% to 50% spurious pulses, it can still accurately distinguish noise from real targets, with all evaluation indicators remaining stable and undiminished. Especially in environments with 50% severe missing pulses or strong clutter, compared to the two classic ensemble clustering algorithms mentioned above, this method improves sorting accuracy by approximately 10% to 20%.

[0050] In summary, this method demonstrates extremely high clustering accuracy and excellent stability in both signal loss and high-noise scenarios, fundamentally improving the problem of difficult separation of overlapping pulses in complex electromagnetic environments by traditional clustering methods.

[0051] It is worth noting that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0052] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. An integrated clustering radar signal sorting method based on data augmentation and quality screening, characterized in that, include: S1, preprocess the received pulse descriptor to obtain the normalized parameter matrix; S2, based on the normalized parameter matrix, multiple sets of weighted feature data are obtained by randomly weighting the pulse descriptor, and multiple clustering algorithms are used to process the multiple sets of weighted feature data to generate multiple base clustering results; S3, Based on the physical characteristics and differences of the signals, select a preferred set of base clusters from the multiple base clustering results; S4. Consensus fusion of the preferred base cluster set is performed using spectral clustering to obtain the radar signal sorting result.

2. The integrated clustering radar signal sorting method based on data augmentation and quality screening according to claim 1, characterized in that, S1 includes: S11, combine the arrival angle, radio frequency, arrival time, pulse width and pulse amplitude parameters of each received pulse to form a pulse description word; S12, combine the pulse descriptors of multiple pulses to form a parameter matrix; S13, standardize the parameter matrix to obtain a normalized parameter matrix.

3. The integrated clustering radar signal sorting method based on data augmentation and quality screening according to claim 1, characterized in that, S2 include: S21, preset multiple different weight allocation templates and form a weight template set; S22, introduce a random perturbation factor into each weight allocation template in the weight template set to generate multiple sets of perturbed feature weights; S23, the normalized parameter matrix is ​​weighted using the perturbed feature weights of each group to generate multiple sets of weighted feature data; S24. For each group of weighted feature data, K-means clustering algorithm, Gaussian mixture model clustering algorithm and hierarchical clustering algorithm are mixed according to a preset ratio for processing, and the target cluster number of each algorithm is randomly sampled within a preset interval to generate multiple base clustering results.

4. The integrated clustering radar signal sorting method based on data augmentation and quality screening according to claim 3, characterized in that, The weight allocation template described in S21 includes the weights corresponding to the radio frequency parameters, pulse width parameters, and arrival time parameters, respectively. The random perturbation factor mentioned in S22 is a multidimensional uniform random vector with a value range of [0,1]. It is used to jitter the weights of each parameter in the weight allocation template. The perturbed feature weights are obtained by the weight allocation template and the random perturbation factor.

5. The integrated clustering radar signal sorting method based on data augmentation and quality screening according to claim 1, characterized in that, S3 include: S31, for each candidate partition in the multiple base clustering results, extract all clusters contained in the candidate partition, and calculate the standard deviation of each cluster in the angle of arrival dimension, pulse width dimension and radio frequency dimension respectively; S32, based on the relationship between the standard deviation of each cluster in the angle of arrival dimension and the preset standard deviation threshold of the angle of arrival, determine whether the spatial aggregation degree of each cluster meets the physical characteristics of radar signal point source transmission, and combine the standard deviation of each cluster in the pulse width dimension and the radio frequency dimension to introduce signal agility reward and stability reward, calculate the physical quality score of each cluster, and then obtain the overall quality score of each candidate division. S33, the candidate partition with the highest overall quality score among the multiple base clustering results is selected into the preferred set. In each iteration, the comprehensive marginal gain of each remaining candidate partition is calculated based on the overall quality score of the remaining candidate partitions and their normalized mutual information with the selected partitions in the preferred set. The candidate partition with the largest comprehensive marginal gain is then moved into the preferred set until the number of candidate partitions in the preferred set reaches the preset preferred number threshold. The iteration is then stopped, and the preferred base clustering set is obtained.

6. The integrated clustering radar signal sorting method based on data augmentation and quality screening according to claim 5, characterized in that, S32 includes: S321. For each cluster, determine whether the sample size of the cluster satisfies statistical significance. If it does, and the standard deviation of the cluster in the angle of arrival dimension is greater than or equal to the preset standard deviation threshold of the angle of arrival, then determine that the spatial aggregation degree of the cluster does not meet the physical characteristics of radar signal point source emission, and set the quality score of the cluster to zero. S322, For a cluster whose spatial aggregation satisfies the physical characteristics, the spatial compactness score of the cluster is calculated using a Gaussian kernel function based on the standard deviation of the cluster in the angle of arrival dimension. The Gaussian kernel function uses the standard deviation of the angle of arrival as an input variable and controls the rate at which the score decays as the standard deviation increases through a scale parameter. S323, based on the standard deviation of the cluster in the pulse width dimension or radio frequency dimension, a signal agility reward is introduced. The signal agility reward is given when the standard deviation of the pulse width dimension or radio frequency dimension exceeds the corresponding preset threshold, provided that the angle of arrival dimension remains stable. S324, Based on the relationship between the standard deviation of the cluster in the radio frequency dimension and the preset radio frequency stability threshold, a stability reward is introduced, which is granted when the standard deviation in the radio frequency dimension is less than the radio frequency stability threshold; S325, the spatial compactness score, the signal agility reward, and the stability reward are added together to obtain the physical quality score of the cluster; S326, average the physical quality scores of all clusters within the candidate partition to obtain the global base score of the candidate partition; S327, determine whether the number of clusters in the candidate partition is less than a preset threshold. If so, apply a decay coefficient penalty to the global base score of the candidate partition to obtain the overall quality score of the candidate partition. S328, Traverse all candidate partitions in the multiple base clustering results, obtain the overall quality score of each candidate partition, and perform normalization processing.

7. The integrated clustering radar signal sorting method based on data augmentation and quality screening according to claim 5, characterized in that, S33 includes: S331, In each iteration, for the remaining candidate partitions that are not selected into the preferred set, calculate the normalized mutual information between each remaining candidate partition and the selected partitions in the preferred set. S332, Calculate the comprehensive marginal gain of each remaining candidate partition based on the overall quality score of each remaining candidate partition and its maximum normalized mutual information with the selected partitions in the preferred set; S333, move the candidate partition with the largest comprehensive marginal gain into the preferred set; S334, Repeat steps S331 to S333 until the number of candidate partitions in the preferred set reaches the preset preferred number threshold, and obtain the preferred base cluster set.

8. The integrated clustering radar signal sorting method based on data augmentation and quality screening according to claim 1, characterized in that, S4 include: S41, Calculate the weight of each base clustering result based on the overall quality score of each base clustering result in the preferred base clustering set; the weight is positively correlated with the overall quality score; S42, construct a weighted consensus matrix based on the weights of each base clustering result, and perform spectral decomposition on the weighted consensus matrix to obtain an eigenvalue sequence; each element value in the weighted consensus matrix represents the cumulative weighted probability that two corresponding samples are classified into the same cluster in all base clustering results; S43, estimate the optimal number of clusters based on the feature value sequence, and complete the sample division in the normalized feature space based on the optimal number of clusters to obtain the radar signal sorting result.

9. The integrated clustering radar signal sorting method based on data augmentation and quality screening according to claim 8, characterized in that, S42 includes: S421, Traverse all base clustering results. For any two samples, if they are classified into the same cluster in any base clustering result, then accumulate the weight corresponding to the base clustering result to the matrix element position corresponding to the two samples. After traversal, use the accumulated value of each element position as the cumulative weighted probability of the two samples to construct a weighted consensus matrix; if they are not classified into the same cluster in a certain base clustering result, then do not accumulate the weight of the base clustering result to the matrix element. S422, calculate the degree matrix of the weighted consensus matrix, perform a normalized Laplace transform on the weighted consensus matrix based on the degree matrix to obtain a normalized affinity matrix, and perform eigenvalue decomposition on the normalized affinity matrix to obtain an eigenvalue sequence arranged in descending order.

10. The integrated clustering radar signal sorting method based on data augmentation and quality screening according to claim 8, characterized in that, S43 includes: S431, calculate the difference between adjacent feature values ​​in the feature value sequence, and take the feature value index corresponding to the maximum difference as the optimal number of clusters; S432, Select a corresponding number of feature vectors according to the optimal number of clusters, and construct a feature matrix by row from the selected feature vectors; S433, perform row normalization on the feature matrix to project it onto the unit sphere; S434, on the projected unit sphere, each row vector of the feature matrix is ​​used as the feature representation of the sample point, and the K-means algorithm based on cosine distance is used to cluster each row vector to obtain the final clustering label of each pulse sample as the radar signal sorting result.

Images