Radar signal sorting method and system based on rough set theory and adaptive weighting

Abstract

The invention discloses a radar signal sorting method and system based on a rough set theory and adaptive weighting, and relates to the technical field of baseband communication. The radar signal sorting method and system based on the rough set theory and adaptive weighting comprises the following steps: S1, performing whole-process dynamic acquisition on aliasing full-pulse signals to obtain a pulse description word sequence; s2, dividing the full-pulse data stream to obtain a granularity combination entropy, and constructing a global weighted distance matrix; s3, performing local density estimation, constructing a decision diagram and screening pulse points; and S4, constructing a basic probability distribution and uncertainty function value, and outputting a signal sorting result. According to the method, the stability, accuracy and interpretability of radar signal sorting under parameter agility and strong noise conditions are effectively improved, and the problem that existing unsupervised sorting cannot adaptively describe pulse parameter stability differences under fixed distance measurement and fixed weight is solved.

Description

Technical Field

[0001] This invention relates to the field of baseband communication technology, specifically to a radar signal sorting method and system based on rough set theory and adaptive weighting. Background Technology

[0002] With the rapid evolution of radar systems and the increasing superposition of multi-source signals in complex electromagnetic environments, existing unsupervised sorting methods typically normalize the multi-dimensional parameters of pulse descriptors and then use a fixed distance metric and a single clustering criterion for correlation. This makes it difficult to characterize the stability differences of different parameter dimensions under different scenarios. When carrier frequency and amplitude parameters fluctuate strongly or are significantly affected by obstruction and measurement noise, the fixed metric is prone to misinterpreting fluctuations as differences in radiation sources, leading to the splitting of pulses from the same source and the merging of pulses from different sources in the overlapping area. This further causes problems such as center point drift, disruption of core pulse continuity, and the inability to quantify the uncertainty of overlapping pulse attribution.

[0003] For example, the invention patent with announcement number CN118444275B discloses an intelligent sorting method for parameter-agile radar signals, including: a feature extraction module: constructing a network based on multi-branch dilated convolution to extract pulse parameter features; a feature fusion module: constructing a network based on an attention mechanism to fuse the extracted features; a radiation source mapping module: realizing the mapping between features and radiation sources based on transposed convolution; and an intelligent sorting network construction and training module: cascading the above modules to construct a deep sorting network, and performing supervised training using a large number of labeled radar signal samples, and finally using the trained network to sort parameter-agile radar signals.

[0004] For example, the invention patent with announcement number CN119596244B discloses a radar signal spectrum clustering and sorting method based on SOM anchor point extraction and graph fusion, including: a data preprocessing module: normalizing radar pulse parameters and constructing a KNN graph; an anchor point extraction module: using a self-organizing map network (SOM) for iterative training to extract topological anchor points of the data; a graph fusion module: calculating the similarity between anchor points and nodes to construct an adaptive anchor graph, and performing weighted fusion with the KNN graph; and a spectrum clustering module: constructing a Laplacian matrix based on the fused graph, and achieving final signal sorting by calculating feature vectors and combining them with the K-means algorithm.

[0005] In existing technologies, under scenarios where parameters change rapidly over time and with varying systems, some fields in existing systems exhibit strong fluctuations or are significantly affected by channels and obstructions, leading to increased dispersion of homologous pulses in these dimensions. Simultaneously, multiple sources naturally approximate each other in several dimensions, resulting in inter-class overlap. Fixed metrics may misinterpret fluctuations and measurement noise as differences in radiation sources, causing homologous pulses to be broken up and overlapping regions to be mixed.

[0006] Therefore, in order to address the above problems, there is an urgent need for radar signal sorting methods and systems based on rough set theory and adaptive weighting. Summary of the Invention

[0007] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a radar signal sorting method and system based on rough set theory and adaptive weighting, which solves the problem that existing unsupervised sorting cannot adaptively characterize the stability differences of pulse parameters under fixed distance metrics and fixed weights.

[0008] Technical solution To achieve the above objectives, this invention employs the following technical solution: a radar signal sorting method based on rough set theory and adaptive weighting, comprising: S1, dynamically acquiring the entire process of aliased full-pulse signals, obtaining pulse descriptor sequences, and normalizing, anomaly marking, and isolating the pulse descriptor sequences to obtain a full-pulse data stream; S2, dividing the full-pulse data stream based on the pulse descriptor sequences to obtain granular combined entropy, performing equivalence class partitioning, and constructing a global weighted distance matrix; S3, performing local density estimation based on the global weighted distance matrix, constructing a decision graph, filtering pulse points, and obtaining the shared nearest neighbor number; S4, constructing basic probability assignment and uncertainty function values, generating global confidence, marking noise pulses, and outputting the signal sorting results.

[0009] Furthermore, the entire process of dynamically acquiring the aliased full-pulse signal to obtain the pulse descriptor sequence, and then normalizing, anomaly marking, and isolating the pulse descriptor sequence to obtain the full-pulse data stream, is as follows: The entire process of dynamically acquiring the aliased full-pulse signal intercepted by the radar reconnaissance receiver is performed to obtain the pulse descriptor sequence. The pulse descriptor sequence contains carrier frequency, pulse width, angle of arrival, and amplitude parameters, and also records the arrival timestamp for each pulse. The pulse descriptor sequence is written into the full-pulse data stream through structured encapsulation, unifying the field caliber and recording granularity. The full-pulse data stream is aligned using a maximum-minimum normalization method. Missing pulses, abnormal pulses, and outliers in the full-pulse data stream are marked and isolated to avoid amplifying errors in the sorting link due to dirty data.

[0010] Furthermore, the specific process of partitioning the full pulse data stream based on the pulse descriptor sequence to obtain the granular combined entropy is as follows: Based on the pulse descriptor sequence, the threshold approximation discrimination method is used to calculate the indistinguishability relationship of pulses on the pulse descriptor sequence in the full pulse data stream: For any two pulses, values ​​are taken in the dimensions of carrier frequency, pulse width, angle of arrival, and amplitude parameters, respectively; if the absolute value of the difference between the two pulses does not exceed the discrimination threshold corresponding to the dimension in each parameter dimension, then the two pulses are determined to be indistinguishable on the pulse descriptor sequence and are recorded as satisfying the indistinguishability relationship; the full pulse data stream is partitioned based on the indistinguishability relationship to obtain the pulse equivalence class partitioning result; each equivalence class represents a set of pulses that are mutually indistinguishable in the dimensions of carrier frequency, pulse width, angle of arrival, and amplitude parameters. The process involves several steps: First, a sample is generated. Then, the number of pulses in the full pulse data stream is counted. After dividing the full pulse data stream into non-overlapping, full-coverage segments using indistinguishable relationships, the number of pulses in each equivalence class is counted to obtain the total number of pulses in each equivalence class. The total number of radar pulse equivalence classes is then calculated by counting the number of equivalence classes obtained from the division. These equivalence classes are numbered sequentially. The total number of equivalence classes is obtained by squaring the number of pulses in each equivalence class and summing the results. This sum is then divided by the square of the number of pulses in the full pulse data stream to obtain the granularity term. Finally, the proportion of equivalence classes is calculated by dividing the number of pulses in each equivalence class by the total number of pulses in the full pulse data stream. The logarithm of this proportion is multiplied by the total proportion of equivalence classes, and the sum is obtained by taking the negative of the logarithm. This sum is then multiplied by the granularity term to obtain the granularity combination entropy.

[0011] Further, the specific process of performing equivalence class partitioning and constructing a global weighted distance matrix is ​​as follows: For any parameter in the pulse descriptor sequence, construct a pulse descriptor sequence after removing the parameter; recalculate the indistinguishable relationship and partition equivalence classes based on the pulse descriptor sequence after removing the parameter to obtain the granular combination entropy after removing the parameter; obtain the parameter discrimination contribution by comparing the granular combination entropy of all parameters with the granular combination entropy after removing the parameter through a step-by-step elimination comparison method and granular combination entropy difference evaluation; map the parameter discrimination contribution to sorting weights through a segmented mapping method; construct a weighted Euclidean distance based on the sorting weights of each parameter to measure the difference between two radar pulses: For any two radar pulses, extract the values ​​of their carrier frequency, pulse width, angle of arrival, and amplitude parameters; calculate the differences in each dimension and sum them up according to the sorting weights of the dimensions to obtain the weighted distance between the two pulses; perform weighted distance calculations on each pair of all pulses in the full pulse data stream to form a global weighted distance matrix.

[0012] Furthermore, the specific process of local density estimation based on the global weighted distance matrix is ​​as follows: Based on the global weighted distance matrix, the k nearest neighbor pulses of each radar pulse by weighted distance are taken as the nearest neighbor set; a variable bandwidth Cauchy kernel is used for local density estimation, with the weighted Euclidean distance between the current pulse and neighbor pulses as the distance term, and the distance from the neighbor pulse to its own nearest neighbor set as the adaptive bandwidth; by accumulating the Cauchy kernel responses of all neighbor pulses in the nearest neighbor set, the Cauchy kernel density value of the current pulse is obtained.

[0013] Furthermore, the specific process of constructing the decision graph and selecting pulse points is as follows: For each neighbor pulse in the nearest neighbor set, the distance term is divided by the adaptive bandwidth to obtain the scaled distance. The scaled distance is squared and then one is added, and the reciprocal is taken to obtain the contribution value of the neighbor pulse. The contribution values ​​of all neighbor pulses in the nearest neighbor set are accumulated sequentially to obtain the accumulated Cauchy kernel density value of the current pulse. Within the same nearest neighbor set, the maximum value of the variable bandwidth Cauchy kernel density value corresponding to each neighbor pulse is found as the normalization benchmark. The accumulated Cauchy kernel density value of the current pulse is divided by the normalization benchmark to obtain the relative Cauchy kernel density value. A decision graph is constructed based on the relative Cauchy kernel density value and the pulse samples, and pulse points with relative Cauchy kernel density values ​​not less than the density threshold are determined as radiation source centers.

[0014] Furthermore, the specific process of obtaining the shared nearest neighbor number is as follows: obtain the nearest neighbor sets of two radar pulses respectively, calculate the number of pulses contained in the intersection of the two nearest neighbor sets, and define the number of pulses as the shared nearest neighbor number of the two pulses; determine and associate the core pulse based on the shared nearest neighbor number: for any pulse and the radiation source center, calculate the shared nearest neighbor number of the pulse and the radiation source center; when the shared nearest neighbor number is greater than the shared number threshold, the pulse is determined to be the core pulse of the radiation source and directly associated with the corresponding radiation source center.

[0015] Furthermore, the specific process of constructing the basic probability assignment and uncertainty function value to generate the global confidence score is as follows: Pulses that failed to be identified and associated with the core pulse are considered overlapping pulses. Neighbor pulses in the nearest neighbor set are searched. Evidence theory is used to construct the basic probability assignment and uncertainty function value for each neighbor pulse relative to each radiation source center: For any neighbor pulse and any radiation source center, the weighted Euclidean distance between the overlapping pulse and the neighbor pulse is used as the distance attenuation term. Simultaneously, the number of samples falling into the radiation source center in the nearest neighbor set of the neighbor pulse is normalized according to the number of nearest neighbors as the neighborhood support term. The distance attenuation term and the neighborhood support term are multiplied to obtain the basic probability assignment of the neighbor pulse relative to the radiation source center. For the same neighbor pulse, the probability assignment is calculated as follows: After summing the support probabilities of all radiation source centers, subtract the sum from one to obtain the uncertainty function value of the neighbor pulses assigned to the uncertainty set. Using the evidence theory combination rule, all basic probability assignments are fused. First, the basic probability assignments and uncertainty function values ​​provided by each neighbor pulse of the overlapping pulses are added together to obtain the support composite quantity. Then, all support composite quantities are multiplied sequentially to obtain the comprehensive support term. Finally, the uncertainty function values ​​are multiplied sequentially to obtain the comprehensive uncertainty term. Subtracting the comprehensive uncertainty term from the comprehensive support term yields the net support quantity. Multiplying the comprehensive uncertainty term by one and subtracting the number of radiation source centers gives the correction term. Adding the correction term to the comprehensive support term yields the normalization term. Dividing the net support quantity by the normalization term gives the global confidence value of the overlapping pulses belonging to the radiation source centers.

[0016] Furthermore, the specific process of marking noise pulses and outputting signal sorting results is as follows: The global confidence values ​​of overlapping pulses to the centers of each radiation source are compared, and the radiation source with the highest global confidence value is selected as the candidate for assignment. Simultaneously, the maximum global confidence value is compared with the confidence decision threshold. When the maximum global confidence value is not less than the confidence decision threshold, the overlapping pulses are classified to the corresponding radiation source. When the maximum global confidence value is less than the confidence decision threshold, the overlapping pulses are determined as uncertain pulses and marked as noise pulses. The core pulse correlation results and overlapping pulse decision results are summarized, and a radiation source number label is assigned to each pulse in the full pulse data stream. Noise pulses are identified and marked, and the signal sorting results and the global confidence value corresponding to each pulse are output.

[0017] Furthermore, the second aspect of the present invention provides a radar signal sorting system based on rough set theory and adaptive weighting, applied to a radar signal sorting method based on rough set theory and adaptive weighting, comprising: a full-pulse data acquisition module, used to dynamically acquire aliased full-pulse signals throughout the entire process, obtain pulse descriptor sequences, and normalize, anomaly-marking, and isolate the pulse descriptor sequences to obtain a full-pulse data stream; a rough set granular entropy weighting module, used to partition the full-pulse data stream based on the pulse descriptor sequences to obtain granular combined entropy, perform equivalence class partitioning, and construct a global weighted distance matrix; a core pulse fast association module, used to perform local density estimation based on the global weighted distance matrix, construct a decision graph, filter pulse points, and obtain the number of shared nearest neighbors; and an overlapping pulse evidence fusion module, used to construct basic probability assignment and uncertainty function values, generate global confidence, mark noise pulses, and output signal sorting results.

[0018] Beneficial effects The present invention has the following beneficial effects: (1) In this invention, the parameter differentiation contribution is obtained by eliminating and comparing one by one and the segmented mapping is used to generate sorting weights to construct a unified weighted distance space, so that the pulse similarity measurement can adaptively suppress the influence of agile fluctuations and measurement noise, highlight the contribution of stable parameters, thereby reducing the risk of homogeneous pulses being split and heterogeneous pulses being mixed. (2) In this invention, the long-tailed kernel function is used to suppress the interference of clutter and outlier pulses on the density peak by using variable bandwidth Cauchy kernel density estimation based on nearest neighbor set, and the connectivity of agile signal is maintained by adaptively adjusting the bandwidth with local sparsity, thereby improving the robustness of radiation source center extraction. (3) This invention achieves rapid association of core pulses by sharing the nearest neighbor consistency criterion, prioritizes locking the high-confidence core set as a reliable support for the decision, reduces the accumulation of errors in association expansion, and suppresses the propagation of false associations from a mechanism perspective. It solves the problem of center omission caused by the difference in the number of pulses from different radiation sources, and ensures that center screening still has a consistent judgment caliber in multi-cluster imbalance and overlapping scenarios; (4) This invention calculates the global confidence of overlapping pulses by using basic probability allocation and combination fusion based on evidence theory. It then expresses the support and uncertainty in a structured way and uses it for threshold decision-making, avoiding forced assignment of overlapping pulses when evidence is insufficient, which would lead to misclassification. This facilitates tracking, verification, and quality assessment, and improves reproducibility and engineering adaptability under complex electromagnetic environments and parameter agility conditions.

[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0020] Figure 1A flowchart illustrating a parameter-agile radar signal sorting method based on coarse evidence theory and adaptive weighting, provided for embodiments of this application; Figure 2 A diagram illustrating the architecture of a parameter-agile radar signal sorting system based on coarse evidence theory and adaptive weighting, provided for embodiments of this application; Figure 3 This is a schematic diagram showing the distribution of the full-pulse data stream in the carrier frequency and pulse width dimensions, provided for embodiments of this application. Figure 4 A decision diagram of the radar radiation source center provided in the embodiments of this application; Figure 5 A schematic diagram showing the performance comparison of the method provided in this application embodiment with the prior art in terms of sorting accuracy; Figure 6 This is a schematic diagram showing the sorting results of the DBSCAN algorithm in the prior art for the data in the embodiments of this application; Figure 7 This is a schematic diagram showing the sorting results of the anchor-based spectral clustering algorithm (USPEC) in the prior art on the data of the embodiments of this application; Figure 8 This is a schematic diagram illustrating the sorting results of data from the embodiments of this application using a deep convolutional network (CNN) in the prior art. Figure 9 This is a schematic diagram illustrating the sorting results of the same data using the method provided in this embodiment of the application. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figures 1-9 This invention provides a technical solution: a radar signal sorting method based on rough set theory and adaptive weighting, comprising: S1, dynamically acquiring the entire process of aliased full-pulse signals, obtaining pulse descriptor sequences, and normalizing, anomaly marking, and isolating the pulse descriptor sequences to obtain a full-pulse data stream; S2, dividing the full-pulse data stream based on the pulse descriptor sequences to obtain granular combined entropy, performing equivalence class partitioning, and constructing a global weighted distance matrix; S3, performing local density estimation based on the global weighted distance matrix, constructing a decision graph, filtering pulse points, and obtaining the number of shared nearest neighbors; S4, constructing basic probability assignment and uncertainty function values, generating global confidence, marking noise pulses, and outputting the signal sorting results.

[0023] Specifically, the process of dynamically acquiring the aliased full-pulse signal throughout the entire process, obtaining the pulse descriptor sequence, and normalizing, anomaly marking, and isolating the pulse descriptor sequence to obtain the full-pulse data stream is as follows: dynamically acquiring the aliased full-pulse signal intercepted by the radar reconnaissance receiver throughout the entire process, obtaining the pulse descriptor sequence, which includes carrier frequency, pulse width, angle of arrival, and amplitude parameters, and also records the arrival timestamp for each pulse.

[0024] The pulse descriptor sequence is written into the full pulse data stream through structured encapsulation, unifying field caliber and record granularity, binding unique sequence numbers and acquisition channel identifiers, and using a maximum-minimum value normalization method to align the dimensions of the full pulse data stream. Missing pulses, abnormal pulses, and outliers in the full pulse data stream are marked and isolated. The determination of missing pulses adopts the arrival timestamp continuity check rule: the arrival timestamps of pulses in the same acquisition channel are sorted in chronological order, the arrival time interval of adjacent pulses is calculated, and when the arrival time interval has a break exceeding the interval threshold, or a time window with a significant increase in continuous intervals, the pulse sequence corresponding to the break interval is marked as having a missing pulse, and the missing position index and missing window range are recorded. The determination of abnormal pulses adopts a joint rule of field completeness and value validity: when any pulse record has missing parameter fields, fields that cannot be parsed, fields that exceed the device's allowable range, or combinations that do not conform to physical constraints, the pulse is marked as an abnormal pulse, and the abnormality type flag and trigger field are written. Outlier pulses are identified using a local statistical deviation rule: within a sliding time window or a nearest neighbor sample set, local statistical benchmarks are calculated for carrier frequency, pulse width, angle of arrival, and amplitude, respectively. When a pulse deviates from the local statistical benchmark in any parameter dimension beyond the deviation threshold, or when the multi-dimensional joint deviation score exceeds the joint threshold, the pulse is marked as an outlier pulse. The marked missing pulse intervals, abnormal pulses, and outlier pulses are isolated and stored and are not included in the sorting calculation to avoid dirty data amplifying the error in the sorting link.

[0025] The full-pulse simulation dataset shown in Table 1 contains the full-pulse simulation characteristic parameters of three radiation sources and environmental noise. Radiation source 1: operating mode is extreme frequency agility, carrier frequency is 2000-4000 Hz, pulse width is 10±0.2 Hz, angle of arrival is 45±1 Hz, amplitude follows a normal distribution with mean -60 Hz and variance 1 (N(-60,1)), and pulse number is 300. Radiation source 2: operating mode is conventional fixed, carrier frequency is 3000±20 Hz, pulse width is 20±0.8 Hz, angle of arrival is 60±1 Hz, amplitude follows a normal distribution with mean -50 Hz and variance 1. (N(-50,1)), pulse number is 250; Radiation source 3: Operating mode is conventional fixed, carrier frequency is 3000±20, pulse width is 24±0.8, angle of arrival is 60±1, amplitude follows a normal distribution with mean -55 and variance 1 (N(-55,1)), pulse number is 250; Ambient noise: Operating mode is random clutter, carrier frequency is 1500-4500, pulse width is 5-30, angle of arrival is 0-360, amplitude follows a uniform distribution with lower limit -90 and upper limit -40 (U(-90,-40)), pulse number is 100.

[0026] Table 1 Full-pulse simulation dataset

[0027] like Figure 3 The diagram illustrates the distribution of the full-pulse data stream in the carrier frequency and pulse width dimensions provided in this embodiment. The horizontal axis represents the carrier frequency, and the vertical axis represents the pulse width. Each point in the diagram represents a pulse record in the full-pulse dataset. Different markers and colors are used to distinguish pulse samples from environmental noise and different radar radiation sources. In the legend, "Radiation Source 1 (Agile)" refers to the radiation source in the diagram that exhibits a banded distribution and reflects the agile characteristics of parameters. It can be seen that one type of radiation source pulses exhibits a banded distribution in the pulse width dimension and spans a relatively wide carrier frequency range, reflecting that under the condition of parameter agility, pulses from the same source fluctuate greatly in the carrier frequency dimension but are relatively concentrated in the pulse width dimension. The other two types of radiation source pulses exhibit a relatively concentrated clustered distribution near the carrier frequency, and there are adjacent and locally overlapping areas with the banded distribution. At the same time, environmental noise points are discretely distributed throughout the entire domain and mixed with valid pulse samples. This diagram reflects the typical sorting difficulties of the data in this embodiment, which simultaneously possess the characteristics of discreteness within the same source class, proximity between different source classes, noise dispersion, and local overlap. It provides a data background for adopting adaptive weighted measurement, density center screening, and overlapping pulse confidence judgment.

[0028] In this implementation scheme, the aliased full-pulse signal is encapsulated in a pulse-granular structure and standardized in terms of caliber to form an indexable and traceable full-pulse data stream. Based on normalized dimensional alignment, a regularized marking and isolation mechanism for missing pulses, abnormal pulses and outliers is introduced, which significantly reduces the amplification effect of dirty data and measurement noise on the sorting link and improves the robustness, reproducibility and consistency of the sorting process.

[0029] Specifically, the process of dividing the full pulse data stream based on the pulse descriptor sequence to obtain the granular combined entropy is as follows: Based on the pulse descriptor sequence, the threshold approximation discrimination method is used to calculate the indistinguishability relationship of pulses on the pulse descriptor sequence in the full pulse data stream: For any two pulses, the values ​​of carrier frequency, pulse width, angle of arrival, and amplitude parameters are taken respectively; when there is an abnormal marker or missing marker in a certain dimension, it is directly determined that they are not comparable and the current comparison is terminated. If the absolute value of the difference between the two pulses does not exceed the discrimination threshold corresponding to the dimension in each parameter dimension, the two pulses are determined to be indistinguishable on the pulse descriptor sequence and are recorded as satisfying the indistinguishability relationship; when the absolute difference in all four dimensions does not exceed the threshold, the two pulses are determined to be indistinguishable under the current parameter set, and the pulse pair is written into the indistinguishability relationship set. The full pulse data stream is divided based on the indistinguishability relationship to obtain the pulse equivalence class partitioning result, and an equivalence class identifier is generated for each group; each equivalence class represents a group of pulse samples that are mutually indistinguishable in the dimensions of carrier frequency, pulse width, angle of arrival, and amplitude parameters.

[0030] The number of pulses in the full pulse data stream is counted. After dividing the full pulse data stream into non-overlapping full-coverage segments using indistinguishable relationships, the number of pulses in each equivalence class is counted to obtain the total number of pulses in each equivalence class. The total number of radar pulse equivalence classes is obtained by counting the number of equivalence classes obtained from the division. The equivalence classes are numbered sequentially. The number of pulses contained in each equivalence class is squared and summed sequentially to obtain the total number of equivalence classes. This total number is divided by the square of the number of pulses in the full pulse data stream to obtain the granularity term. The number of pulses contained in each equivalence class is divided by the number of pulses in the full pulse data stream to obtain the proportion of equivalence classes. The logarithm of the proportion of equivalence classes is multiplied by the proportion of equivalence classes and summed. The negative of the sum is obtained to obtain the entropy term. The entropy term is subtracted from the entropy term and multiplied by the granularity term to obtain the granular combination entropy.

[0031] The specific formula for calculating the granularity combination entropy is as follows: ; In the formula, It represents the granularity combination entropy, used to evaluate the ability of pulse descriptor sequences to sort and distinguish radar pulses; This represents the number of pulses in the full pulse data stream, which serves as the basic quantization benchmark for calculating granular combined entropy. This represents the number of pulses in each equivalence class, used to calculate the distribution concentration term and information entropy term in the granular combinatorial entropy. This represents the total number of radar pulse equivalence classes. It is obtained by statistically dividing the full pulse data stream through indistinguishable relationships and is used to determine the summation range in the granular combination entropy calculation. The index number represents the equivalence class, which is obtained by sequentially numbering the equivalence classes obtained through partitioning by unidentifiable relations. It is used for the iterative calculation of the summation term in the granular combination entropy formula. This represents a pulse descriptor sequence used to define the parameter range of unidentifiable relationships.

[0032] In this implementation scheme, by constructing indistinguishable relationships and forming equivalence class partitions on the full pulse data stream based on the threshold approximation discrimination method, and using the number of equivalence classes to obtain the granular combination entropy, the contribution of the pulse descriptor sequence to the sorting and distinguishing ability can be quantified on a unified scale. This provides a stable and interpretable evaluation benchmark for parameter distinguishing contribution assessment, sorting weight mapping and weighted distance construction, and improves the adaptability and consistency of the sorting link to parameter agility, overlapping distribution and noise interference scenarios.

[0033] Specifically, the process of performing equivalence class partitioning and constructing a global weighted distance matrix is ​​as follows: For any parameter in the pulse descriptor sequence, construct a pulse descriptor sequence after removing the parameter; recalculate the indistinguishable relation and partition equivalence classes based on the pulse descriptor sequence after removing the parameter to obtain the granular combination entropy after removing the parameter; obtain the parameter discrimination contribution by comparing the granular combination entropy of the full parameter with the granular combination entropy after removing the parameter through a step-by-step elimination comparison method and granular combination entropy difference evaluation; add the granular combination entropy after removing the parameter to the granular combination entropy of the full parameter as the numerator, subtract the granular combination entropy of the full parameter from the granular combination entropy after removing the parameter as the denominator, and divide the numerator by the denominator to obtain the parameter discrimination contribution.

[0034] The difference assessment employs a normalized relative difference measure: using the full-parameter granular combination entropy as a reference, the changes in the full-parameter granular combination entropy and the granular combination entropy after parameter removal are scaled and normalized to avoid the impact of entropy magnitude differences caused by different data scales or equivalence class distributions on the comparability of contributions. To reduce the impact of correlation between parameters on the single-parameter contribution assessment, a conditional removal consistency check is used for limitation: when calculating the contribution of a single parameter, the changes in the number of equivalence classes before and after parameter removal, the rearrangement ratio of equivalence class members, and the changes in the concentration of equivalence class distribution are recorded simultaneously. When there is a situation where the change in the number of equivalence classes is very small but the rearrangement ratio of members is significant, the parameter is judged to have redundant correlation with other parameters, and the contribution is reduced according to the redundancy coefficient, making the contribution more biased towards the parameter dimension that can provide independent distinguishing information.

[0035] The parameter distinguishing contribution is mapped to a sorting weight using a segmented mapping method. When the parameter distinguishing contribution is zero, the sorting weight of the parameter is equal to the reciprocal of the number of equivalence class partitions after removing the parameter; when the parameter distinguishing contribution is greater than zero, the sorting weight of the parameter is equal to one plus the parameter distinguishing contribution. The number of equivalence class partitions after removing the parameter represents the number of equivalence classes obtained by partitioning the entire pulse data stream according to the indistinguishable relationship after removing the parameter from the pulse parameter set. It is obtained by statistically analyzing the equivalence class partitioning results after removing the parameter and is used to reflect the change in distinguishability after removing the parameter, thus providing a usable weighting scale even when the importance is zero.

[0036] Based on the sorting weights of each parameter, a weighted Euclidean distance is constructed to measure the difference between two radar pulses: For any two radar pulses, the values ​​of their carrier frequency, pulse width, angle of arrival, and amplitude parameters are extracted. The differences in each dimension are calculated and weighted according to the sorting weights of the dimensions to obtain the weighted distance between the two pulses. Weighted distance calculations are performed pairwise on all pulses in the full pulse data stream to form a global weighted distance matrix. The weighted Euclidean distance represents the weighted difference value between two radar pulses; the number of pulse parameter dimensions represents the number of parameters contained in the pulse descriptor; and the sorting weight of a certain parameter dimension represents the contribution weight of the parameter in the sorting process.

[0037] In this implementation scheme, by eliminating pulse descriptor parameters one by one and recalculating unidentifiable relationships and equivalence class partitions, the parameter differentiation contribution is stably converted into sorting weights through segmented mapping. Weighted Euclidean distance and global weighted distance matrices are constructed for agile and noisy scenarios, which improves the ability of the sorting link to suppress unstable parameter disturbances and the efficiency of utilizing stable differentiation information, reduces the risk of homogeneous splitting and heterogeneous mixing, and improves the consistency of results.

[0038] Specifically, the process of local density estimation based on the global weighted distance matrix is ​​as follows: Based on the global weighted distance matrix, the k nearest neighbor pulses of each radar pulse by weighted distance are taken as the nearest neighbor set; k represents the number of nearest neighbors, used to determine the local range within the weighted distance space for each pulse to participate in nearest neighbor retrieval, density estimation, shared nearest neighbor calculation, and evidence fusion; k is adaptively determined by the dataset size and local density: based on the number of pulses and the density of radiation source distribution within the sliding time window of the full pulse data stream, candidate value ranges for k are given; then, within the candidate range, local connectivity and stability are evaluated for different k, and the k that allows the nearest neighbor set to cover the local structure of pulses from the same source without introducing too many pulses from different sources is selected as the number of nearest neighbors. Taking the current pulse as the query object, based on the global weighted distance matrix constructed by the sorting weights, the weighted distance between each of the remaining pulses in the full pulse data stream and the current pulse is calculated, and sorted in ascending order of distance; the neighbor pulses are the group of pulse samples with the highest similarity and the smallest weighted distance to the current pulse in the weighted distance space.

[0039] A variable-bandwidth Cauchy kernel is used for local density estimation. The weighted Euclidean distance between the current pulse and its neighboring pulses is used as the distance term, and the distance from a neighboring pulse to its nearest neighbor set is used as the adaptive bandwidth. The Cauchy kernel density value of the current pulse is obtained by summing the Cauchy kernel responses of all neighboring pulses within the nearest neighbor set. The method utilizes the long-tailed characteristic of the Cauchy distribution to suppress the interference of clutter and anomalous pulses on the density peak, and maintains the connectivity of the parameter-agile signal by adaptively changing the bandwidth with the local sparsity.

[0040] In this implementation, by constructing a nearest neighbor set for each pulse on the global weighted distance matrix, the density estimation is limited to the range of local samples most similar to the current pulse, avoiding interference from global noise and distant heterogeneous pulses on the local structure, and suppressing density distortion caused by introducing too many heterogeneous samples in the overlapping area; it can still stably characterize the density peak and connectivity structure even in scenarios where parameter agility causes intraclass stretching, noise point dispersion and irregular cluster morphology.

[0041] Specifically, the process of constructing the decision graph and selecting pulse points is as follows: For each neighbor pulse in the nearest neighbor set, divide the distance term by the adaptive bandwidth to obtain the scaled distance, square the scaled distance and add one, and take the reciprocal to obtain the contribution value of the neighbor pulse; accumulate the contribution values ​​of all neighbor pulses in the nearest neighbor set to obtain the accumulated Cauchy kernel density value of the current pulse; within the same nearest neighbor set, find the maximum value of the variable bandwidth Cauchy kernel density value corresponding to each neighbor pulse as the normalization benchmark; divide the accumulated Cauchy kernel density value of the current pulse by the normalization benchmark to obtain the relative Cauchy kernel density value. The basis for using the maximum density within the nearest neighbor set as the normalization benchmark is that the variable bandwidth Cauchy kernel density is a local neighborhood accumulation. Different radiation sources have different pulse numbers, local density, and bandwidth scales, which leads to the absolute density values ​​not being directly comparable. Normalizing by using the maximum density within the same nearest neighbor set as a local upper bound can convert the absolute density of the current pulse into a proportional score relative to the local peak, so that the relative density values ​​fall into a uniform comparable range and maintain local ranking consistency. This avoids center omissions in small sample clusters or sparse clusters due to low absolute density and facilitates center screening on the decision graph using a uniform threshold.

[0042] The specific formula for calculating the relative Cauchy kernel density value is as follows: ; In the formula, Indicates pulse sample The relative Cauchy kernel density value is used to normalize the pulse density of different radiation sources, avoiding the problem of missing the center of the radiation source due to the difference in the number of pulses from different radiation sources; k represents the number of nearest neighbors, which is obtained by searching for pulse samples. The k closest pulses, weighted by distance and sorted from nearest to farthest, are used to determine... The local estimation range; Indicates pulse sample k nearest neighbor set A single neighboring pulse sample in the data is retrieved by comparing it with the pulse sample. The k closest pulses by weighted distance, each retrieved pulse is... Used as a calculation pulse sample The basic neighbor unit of Cauchy kernel density; Represents the nearest neighbor set, used to determine neighbor impulses. The local distribution range; Indicates pulse sample Pulse with neighbor The weighted Euclidean distance between them, obtained by weighted Euclidean distance, is used to quantize pulse samples. Pulse with neighbor Similarity; Indicates neighbor pulse The variable bandwidth Cauchy kernel density value is obtained by local density estimation using the variable bandwidth Cauchy kernel, and is used to provide pulse samples. Density peak reference within the nearest neighbor range.

[0043] A decision graph is constructed based on the relative Cauchy kernel density value and pulse samples. When the relative Cauchy kernel density of a pulse in the nearest neighbor set is the maximum value, the maximum weighted Euclidean distance between the pulse and each of the nearest neighbor pulses in the nearest neighbor set is taken as the relative distance of the pulse. When the relative Cauchy kernel density of a pulse in the nearest neighbor set is not the maximum value, all pulses with a relative Cauchy kernel density greater than the pulse are selected, the weighted Euclidean distance between the pulse and these pulses is calculated, and the minimum value is taken as the relative distance of the pulse. Pulse points with a relative Cauchy kernel density value not less than the density threshold are determined as the radiation source center.

[0044] like Figure 4 The diagram shows a decision map of radar radiation source centers provided in this application embodiment. Each point in the diagram corresponds to a pulse sample in the full-pulse data stream: first, the local density of the sample is calculated based on the weighted distance and the nearest neighbor set, and then the relative distance of the sample is calculated; when a sample has the maximum local density in the nearest neighbor set, its relative distance is determined according to the maximum weighted distance from the sample to each sample in the nearest neighbor set. The resulting two-dimensional distribution is used to screen radiation source centers: sample points that simultaneously satisfy the conditions of local density not less than a density threshold and relative distance not less than a distance threshold are selected as radiation source centers. In the diagram, the red pentagrams represent the screened radiation source centers (center 1, center 2, center 3), and the blue dots represent non-center pulses; the selected center points will serve as reference centers for core pulse association and overlapping pulse attribution decisions, realizing the initialization and stable numbering of radiation source clusters.

[0045] In this implementation scheme, by using a variable bandwidth Cauchy kernel to accumulate the nearest neighbor contribution in the global weighted distance space and introducing neighborhood maximum density normalization, the absolute density of different radiation sources under the differences in sample size, local density and bandwidth scale is unified to a comparable relative density scale, making the local density evaluation more stable for data size and cluster imbalance.

[0046] Specifically, the process of obtaining the shared nearest neighbor number is as follows: Obtain the nearest neighbor sets of two radar pulses respectively, calculate the number of pulses contained in the intersection of their nearest neighbor sets, and define this number as the shared nearest neighbor number of the two pulses. A larger shared nearest neighbor number indicates stronger consistency in the neighborhood structure of the two pulses, and a higher probability of them originating from the same source. Pulses from the same source typically have similar local neighborhood distributions in the weighted distance space, meaning that the sources and structures of their surrounding neighbor samples are similar. When the nearest neighbor sets of two pulses highly overlap, it indicates that they are located within the same local manifold or the same radiation source cluster. When the nearest neighbor sets of two pulses overlap less, it indicates that they are located in different clusters or at overlapping boundaries, requiring a reduction in the confidence level of direct attribution and subsequent evidence fusion processing.

[0047] Core pulse determination and association based on shared nearest neighbor count: For any pulse and radiation source center, the shared nearest neighbor count between the pulse and the radiation source center is calculated. The radiation source center represents the central pulse point selected by the decision graph of relative density and relative distance, serving as the representative point of the radiation source cluster. The shared nearest neighbor count is calculated by intersecting the nearest neighbor set of the central pulse point with the nearest neighbor set of the pulse to be determined. When the shared nearest neighbor count is greater than the shared count threshold, the pulse is determined to be a core pulse of the radiation source and directly associated with the corresponding radiation source center. When the shared nearest neighbor count is not greater than the shared count threshold, the pulse is determined to be a non-core pulse and is not forcibly assigned. Instead, the process proceeds to the global confidence fusion decision process for overlapping pulses to avoid false associations and subsequent propagation in the overlapping area or under noise interference.

[0048] In this implementation scheme, the core determination does not depend on the instantaneous fluctuation of a certain dimension parameter, but focuses more on the local sample composition and connectivity structure, effectively suppressing the propagation of false associations caused by overlapping boundaries and noisy neighbors, and improving the reliability of radiation source number initialization and the consistency, robustness and interpretability of the overall sorting results.

[0049] Specifically, the process of constructing basic probability assignments and uncertainty function values ​​to generate global confidence is as follows: Pulses that failed to be identified and associated with the core pulse are designated as overlapping pulses. These overlapping pulses are used to represent inter-class overlap areas, cluster boundary areas, and pulse samples whose attribution is uncertain due to noise. Neighbor pulses in the nearest neighbor set are searched, and evidence theory is used to construct basic probability assignments and uncertainty function values ​​for each neighbor pulse relative to each radiation source center: For any neighbor pulse and any radiation source center, the weighted Euclidean distance between the overlapping pulse and the neighbor pulse is used as a distance attenuation term. Simultaneously, the number of samples falling into the radiation source center in the nearest neighbor set of the neighbor pulse is normalized according to the number of nearest neighbors to serve as a neighborhood support term. The distance attenuation term and the neighborhood support term are multiplied to obtain the basic probability assignment of the neighbor pulse to the radiation source center. For the same neighbor pulse, the support probabilities for all radiation source centers are summed, and the sum is subtracted to obtain the uncertainty function value assigned to the uncertainty set by the neighbor pulse. For each neighbor pulse, a set of support probabilities for each radiation source center and a residual probability assigned to the uncertainty set are generated, thus explicitly distinguishing between local support provided by a single neighbor and insufficient evidence.

[0050] By combining all basic probability assignments using evidence theory combination rules, the basic probability assignments and uncertainty function values ​​provided by each neighboring pulse of the overlapping pulse are first added together to obtain the support composition quantity; all support composition quantities are then multiplied sequentially to obtain the comprehensive support term; the uncertainty function values ​​are also multiplied sequentially to obtain the comprehensive uncertainty term; the comprehensive support term is subtracted from the comprehensive uncertainty term to obtain the net support quantity; the comprehensive uncertainty term is multiplied by one and the number of radiation source centers is subtracted to obtain the correction term; the comprehensive support term is added to the correction term to obtain the normalization term; the net support quantity is divided by the normalization term to obtain the global confidence value that the overlapping pulse belongs to the radiation source center.

[0051] The specific formula for calculating the global confidence value is as follows: ; In the formula, This represents the global confidence value that a pulse sample belongs to the center of a radiation source, and is used to quantify the reliability of overlapping pulses belonging to a certain radiation source; The nearest neighbor count of a pulse sample is obtained by considering the size of the pulse dataset and the distribution characteristics of the radiation source. It is used to define the local range for density calculation, nearest neighbor retrieval, and evidence fusion. The index number represents the center of the radiation source. It is obtained by sequentially numbering all the radiation source centers obtained through filtering, and is used to distinguish different radar radiation source centers. This represents the basic probability assignment, used to quantify the degree of local support for the assignment of overlapping pulses by a single neighbor; This represents the uncertainty function value, used to quantify the degree of uncertainty of a single neighbor regarding the attribution of overlapping pulses.

[0052] In this implementation scheme, excessive interference from distant or low-similarity neighbors is avoided, thereby improving the reliability of overlapping pulse assignment in scenarios with inter-class overlap, cluster boundaries, and noise contamination, and providing interpretable quantitative basis for confidence threshold determination, noise labeling, and stable number output.

[0053] Specifically, the process of marking noise pulses and outputting signal sorting results is as follows: The global confidence values ​​of overlapping pulses to each radiation source center are compared, and the radiation source with the highest global confidence value is selected as the candidate. Simultaneously, the maximum global confidence value is compared with the confidence decision threshold. When the maximum global confidence value is not less than the confidence decision threshold, the overlapping pulses are classified to the corresponding radiation source, and the maximum confidence value, candidate center number, and effective threshold version are recorded as traceable decision annotations. When the maximum global confidence value is less than the confidence decision threshold, the overlapping pulses are determined as uncertain pulses and marked as noise pulses, and the trigger reason code is recorded for verification or re-judgment after threshold update. The core pulse association results and overlapping pulse decision results are summarized, a radiation source number label is assigned to each pulse in the full pulse data stream, and noise pulses are identified and marked. The signal sorting results and the global confidence value corresponding to each pulse are output. The output results are written back to the radiation source number label at the granularity of pulse recording unit, and the original fields of carrier frequency, pulse width, angle of arrival, amplitude and arrival timestamp are retained for indexing and playback. The core pulse directly inherits the center association number, the overlapping pulse adopts the confidence decision number, and the noise pulse is written with a noise mark and assigned an empty number or a dedicated noise number. At the same time, the corresponding radiation source center number is output for each pulse to form a structured output.

[0054] like Figure 5 The diagram illustrates a performance comparison of the sorting accuracy of the method provided in this application and existing technologies. It shows the sorting accuracy comparison results of four methods on the same dataset. The horizontal axis represents the comparison methods (DBSCAN, USPEC, CNN, and the method of this invention), and the vertical axis represents the sorting accuracy. As can be seen from the bars in the diagram, the method of this invention has the highest sorting accuracy, exceeding that of the three comparison methods: DBSCAN, USPEC, and CNN. This demonstrates that the present invention can achieve more stable sorting results and a higher overall accuracy under conditions of parameter agility, noise interference, and inter-class overlap.

[0055] like Figure 6The diagram illustrates the sorting results of the DBSCAN algorithm on the data of this application embodiment, demonstrating the sorting results of DBSCAN in a two-dimensional plane of carrier frequency and pulse width. Different colors represent different radiation source clusters, and crosses represent environmental noise points. In the legend, "Radiation Source 1 (Agile)" refers to the radiation source in the diagram that exhibits a banded distribution and reflects parameter agile characteristics. Under the density discrimination of DBSCAN, the banded pulse samples are split into multiple clusters, generating additional incorrectly sorted clusters; simultaneously, noise points and cluster boundaries still interfere with each other near the overlapping areas. This indicates that when homologous pulses exhibit strong fluctuations or local sparsity in some parameter dimensions, clustering based on the fixed density connectivity criterion is prone to homologous splitting and mis-clustering.

[0056] like Figure 7 The diagram illustrates the sorting results of the anchor-based spectral clustering algorithm (USPEC) on the data of this application embodiment. In the legend, "Radiation Source 1 (Agile)" refers to the radiation source in the image that exhibits a banded distribution and reflects the agile parameter characteristics. USPEC can divide the main pulse samples into multiple radiation source clusters, forming a grouping selection that matches the data distribution as a whole. However, in areas with dense noise points and near cluster boundaries, some samples are still close to noise points, which can easily lead to unstable assignment of boundary samples. This reflects that when the data has scattered noise points and clusters that are banded or overlapping, sorting based solely on the approximate global graph structure may still result in insufficient local boundary discrimination.

[0057] like Figure 8 The diagram illustrates the sorting results of the data from the embodiments of this application using a deep convolutional network (CNN) in the prior art. In the legend, "Radiation Source 1 (Agile)" refers to the radiation source in the diagram that exhibits a banded distribution and reflects the agile parameter characteristics. In the banded pulse region (the region where homologous pulses are continuously distributed in a banded pattern on the carrier frequency and pulse width planes), multiple categories of colors are interleaved, meaning that the same distribution pattern is assigned to different radiation source categories; simultaneously, category mixing also exists near the overlapping regions. This indicates that when parameter agility causes intra-class distribution stretching and inter-class distance reduction, if the model does not adequately characterize the differences in parameter stability and local neighborhood structure, homologous pulse misclassification and mixed classification in overlapping areas are likely to occur.

[0058] like Figure 9The diagram illustrates the sorting results of the same data using the method provided in this application. In the legend, "Radiation Source 1 (Agile)" refers to the radiation source in the diagram that exhibits a banded distribution and reflects parameter agile characteristics. Banded pulses from the same source can be maintained as the same radiation source cluster, without being split into additional clusters; the two concentrated radiation source clusters and the banded cluster remain clearly separated under two-dimensional projection; environmental noise points are still identified as noise and distinguished from valid clusters. This result demonstrates that the present invention, through adaptive weighted measurement and subsequent center screening, core pulse association, and overlapping pulse confidence judgment mechanism, achieves stable assignment and consistent numbering output under noise and overlap conditions.

[0059] In this implementation scheme, by establishing the assignment decision of overlapping pulses on the dual constraints of the global confidence maximum value and the confidence decision threshold, the assignment of samples in the overlapping area no longer depends on a single distance or a single clustering result, but uses the confidence formed by the fusion of multiple neighbor evidence as a unified decision basis, thereby effectively reducing the spread of misjudgments caused by forced assignment in scenarios of inter-class overlap and noise mixture.

[0060] Specifically, the second aspect of this invention provides a radar signal sorting system based on rough set theory and adaptive weighting, applied to a radar signal sorting method based on rough set theory and adaptive weighting, comprising: a full-pulse data acquisition module, used for dynamic acquisition of aliased full-pulse signals throughout the entire process, obtaining pulse descriptor sequences; the dynamic acquisition process includes pulse stream output by the reconnaissance receiver being analyzed and continuously received at the pulse granularity, structurally encapsulating the carrier frequency, pulse width, angle of arrival, amplitude, and arrival timestamp of each pulse, and normalizing, anomaly marking, and isolating the pulse descriptor sequences to obtain a full-pulse data stream; and a rough set granularity entropy weighting module, used for partitioning the full-pulse data stream based on the pulse descriptor sequences to obtain granularity combination entropy, and performing equivalence class partitioning. The system employs a threshold approximation discrimination method to calculate indistinguishable relationships across carrier frequency, pulse width, angle of arrival, and amplitude parameters. Based on this, it generates non-overlapping, fully covered equivalence class partitioning results for the entire pulse data stream and constructs a global weighted distance matrix. The core pulse fast association module performs local density estimation based on the global weighted distance matrix, constructs a decision map, and filters pulse points. It filters radiation source center pulse points according to density and distance thresholds to obtain the number of shared nearest neighbors. The overlapping pulse evidence fusion module constructs basic probability assignments and uncertainty function values. It constructs basic probability assignments for each radiation source center in the nearest neighbor set for each neighbor, and obtains the uncertainty function value by subtracting the support probability, generating global confidence, marking noise pulses, and outputting signal sorting results.

[0061] This implementation scheme significantly reduces the risks of homogeneous separation and heterogeneous merging; at the same time, it implements thresholded noise labeling and confidence level annotation output for uncertain samples, improving the consistency, traceability and verifiability of sorting results, and enhancing the robustness and reliability of signal sorting in complex electromagnetic environments.

[0062] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0063] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A radar signal sorting method based on rough set theory and adaptive weighting, characterized in that, Includes the following steps: S1, dynamically acquire the aliased full-pulse signal throughout the entire process, obtain the pulse descriptor sequence, and normalize, anomaly mark and isolate the pulse descriptor sequence to obtain the full-pulse data stream; S2, the full pulse data stream is divided based on the pulse descriptor sequence to obtain the granular combined entropy, and equivalence class partitioning is performed to construct a global weighted distance matrix; S3. Based on the global weighted distance matrix, local density estimation is performed, a decision graph is constructed, and pulse points are selected to obtain the number of shared nearest neighbors; S4 constructs the basic probability assignment and uncertainty function value, generates global confidence, marks noise impulses, and outputs the signal sorting results.

2. The radar signal sorting method based on rough set theory and adaptive weighting according to claim 1, characterized in that: The specific process of dynamically acquiring the aliased full-pulse signal throughout the entire process, obtaining the pulse descriptor sequence, and normalizing, anomaly marking, and isolating the pulse descriptor sequence to obtain the full-pulse data stream is as follows: The entire process of dynamic acquisition of the aliased full-pulse signal intercepted by the radar reconnaissance receiver is performed to obtain the pulse descriptor sequence. The pulse descriptor sequence contains carrier frequency, pulse width, angle of arrival and amplitude parameters, and at the same time, the arrival timestamp of each pulse is recorded. The pulse descriptor sequence is written into the full pulse data stream through structured encapsulation, unifying the field caliber and record granularity. The full pulse data stream is aligned in terms of dimensions using the maximum and minimum value normalization method. Missing pulses, abnormal pulses, and outliers in the full pulse data stream are marked and isolated to avoid amplifying the error of the sorting link by dirty data.

3. The radar signal sorting method based on rough set theory and adaptive weighting according to claim 1, characterized in that: The specific process of dividing the full pulse data stream based on the pulse descriptor sequence to obtain the granular combined entropy is as follows: Based on the pulse descriptor sequence, a threshold approximation discrimination method is used to calculate the indistinguishability relationship of pulses on the pulse descriptor sequence in the full pulse data stream: for any two pulses, the values ​​of the carrier frequency, pulse width, angle of arrival and amplitude parameters are taken respectively; if the absolute value of the difference between the two pulses does not exceed the discrimination threshold corresponding to the dimension in each parameter dimension, then the two pulses are determined to be indistinguishable on the pulse descriptor sequence and are recorded as satisfying the indistinguishability relationship. The full pulse data stream is partitioned based on the unidentifiable relationship to obtain the pulse equivalence class partitioning result; Each equivalence class represents a set of pulse samples that are mutually indistinguishable in the dimensions of carrier frequency, pulse width, angle of arrival, and amplitude parameters; The number of pulses in the full pulse data stream is counted. After dividing the full pulse data stream into non-overlapping full coverage segments through indistinguishable relationships, the number of pulses in each equivalence class is counted to obtain the number of pulses in each equivalence class. The total number of radar pulse equivalence classes is obtained by counting the number of equivalence classes obtained from the division. The equivalence classes are numbered sequentially. The total number of equivalence classes is obtained by squaring the number of pulses in each equivalence class and summing the results. This sum is then divided by the square of the number of pulses in the full pulse data stream to obtain the granularity term. The proportion of equivalence classes is obtained by dividing the number of pulses in each equivalence class by the number of pulses in the full pulse data stream. The logarithm of the proportion of equivalence classes is multiplied by the proportion of equivalence classes and summed. The negative of this sum is then used to obtain the entropy term. This entropy term is subtracted from the entropy term and multiplied by the granularity term to obtain the granular combination entropy.

4. The radar signal sorting method based on rough set theory and adaptive weighting according to claim 1, characterized in that: The specific process of performing equivalence class partitioning and constructing the global weighted distance matrix is ​​as follows: For any parameter in the pulse descriptor sequence, construct a pulse descriptor sequence after removing the parameter. Based on the pulse descriptor sequence after removing the parameter, recalculate the indistinguishable relation and divide the equivalence class to obtain the granular combination entropy after removing the parameter. The parameter differentiation contribution is obtained by comparing the granular combination entropy of all parameters and the granular combination entropy after parameter removal, and by evaluating the difference between the granular combination entropy and the method of elimination one by one. The contribution of parameters is mapped to sorting weights using a segmented mapping method. Based on the sorting weights of each parameter, a weighted Euclidean distance is constructed to measure the difference between two radar pulses: for any two radar pulses, the values ​​of their carrier frequency, pulse width, angle of arrival, and amplitude parameters are extracted, the differences are calculated in each dimension, and the weighted summation is performed according to the sorting weights of the dimensions to obtain the weighted distance between the two pulses; weighted distance calculation is performed pairwise on all pulses in the full pulse data stream to form a global weighted distance matrix.

5. The radar signal sorting method based on rough set theory and adaptive weighting according to claim 1, characterized in that: The specific process of local density estimation based on the global weighted distance matrix is ​​as follows: Based on the global weighted distance matrix, the k nearest neighbor pulses of each radar pulse by weighted distance are taken as the nearest neighbor set. A variable bandwidth Cauchy kernel is used for local density estimation, with the weighted Euclidean distance between the current pulse and neighbor pulses as the distance term and the distance from the neighbor pulse to its own nearest neighbor set as the adaptive bandwidth. The Cauchy kernel density value of the current pulse is obtained by accumulating the Cauchy kernel responses of all neighbor pulses in the nearest neighbor set.

6. The radar signal sorting method based on rough set theory and adaptive weighting according to claim 1, characterized in that: The specific process of constructing the decision graph and selecting pulse points is as follows: For each neighbor pulse in the nearest neighbor set, divide the distance term by the adaptive bandwidth to obtain the scaled distance. Square the scaled distance, add one, and take the reciprocal to obtain the contribution value of the neighbor pulse. Accumulate the contribution values ​​of all neighbor pulses in the nearest neighbor set to obtain the accumulated Cauchy kernel density value of the current pulse. Within the same nearest neighbor set, find the maximum value of the variable bandwidth Cauchy kernel density value corresponding to each neighbor pulse as the normalization benchmark. Divide the accumulated Cauchy kernel density value of the current pulse by the normalization benchmark to obtain the relative Cauchy kernel density value. A decision map is constructed based on the relative Cauchy nuclear density value and pulse samples, and pulse points with a relative Cauchy nuclear density value not less than the density threshold are identified as radiation source centers.

7. The radar signal sorting method based on rough set theory and adaptive weighting according to claim 1, characterized in that: The specific process for obtaining the shared nearest neighbor number is as follows: Obtain the nearest neighbor sets of two radar pulses respectively, calculate the number of pulses contained in the intersection of the two nearest neighbor sets, and define the number of pulses as the shared nearest neighbor number of the two pulses; determine and associate the core pulse based on the shared nearest neighbor number: for any pulse and radiation source center, calculate the shared nearest neighbor number of the pulse and radiation source center; when the shared nearest neighbor number is greater than the shared number threshold, the pulse is determined to be the core pulse of the radiation source and directly associated with the corresponding radiation source center.

8. The radar signal sorting method based on rough set theory and adaptive weighting according to claim 1, characterized in that: The specific process of constructing the basic probability assignment and uncertainty function value to generate the global confidence score is as follows: Pulses that failed to be identified and associated with the core pulse are considered overlapping pulses. Neighbor pulses in the nearest neighbor set are searched, and evidence theory is used to construct the basic probability assignment and uncertainty function value of each neighbor pulse to each radiation source center: For any neighbor pulse and any radiation source center, the weighted Euclidean distance between the overlapping pulse and the neighbor pulse is used as the distance attenuation term, while the number of samples falling into the radiation source center in the nearest neighbor set of the neighbor pulse is normalized according to the number of nearest neighbors as the neighborhood support term; the distance attenuation term and the neighborhood support term are multiplied to obtain the basic probability assignment of the neighbor pulse to the radiation source center. For the same neighbor pulse, the support probabilities for all radiation source centers are summed, and the sum is subtracted from the sum to obtain the uncertainty function value of the neighbor pulse assigned to the uncertainty set. By combining all basic probability assignments using evidence theory combination rules, the basic probability assignments and uncertainty function values ​​provided by each neighboring pulse of the overlapping pulse are first added together to obtain the support composition quantity; all support composition quantities are then multiplied sequentially to obtain the comprehensive support term; the uncertainty function values ​​are also multiplied sequentially to obtain the comprehensive uncertainty term; the comprehensive support term is subtracted from the comprehensive uncertainty term to obtain the net support quantity; the comprehensive uncertainty term is multiplied by one and the number of radiation source centers is subtracted to obtain the correction term; the comprehensive support term is added to the correction term to obtain the normalization term; the net support quantity is divided by the normalization term to obtain the global confidence value that the overlapping pulse belongs to the radiation source center.

9. The radar signal sorting method based on rough set theory and adaptive weighting according to claim 1, characterized in that: The specific process of marking noise pulses and outputting signal sorting results is as follows: The global confidence values ​​of overlapping pulses at the centers of each radiation source are compared, and the radiation source with the largest global confidence value is selected as the candidate for assignment. At the same time, the maximum global confidence value is compared with the confidence decision threshold. When the maximum global confidence value is not less than the confidence decision threshold, the overlapping pulses are classified into the corresponding radiation source. When the maximum global confidence value is less than the confidence decision threshold, the overlapping pulses are judged as uncertain pulses and marked as noise pulses; The core pulse correlation results and overlapping pulse decision results are summarized, each pulse in the full pulse data stream is assigned a radiation source number label, and noise pulses are identified and marked. The signal sorting results and the global confidence value corresponding to each pulse are output.

10. A radar signal sorting system based on rough set theory and adaptive weighting, employing the radar signal sorting method based on rough set theory and adaptive weighting as described in any one of claims 1-9, characterized in that, include: The full-pulse data acquisition module is used to dynamically acquire the aliased full-pulse signal throughout the entire process, obtain the pulse descriptor sequence, and normalize, anomaly mark and isolate the pulse descriptor sequence to obtain the full-pulse data stream. The rough set granular entropy weighting module is used to divide the full pulse data stream based on the pulse descriptor sequence, obtain the granular combined entropy, perform equivalence class partitioning, and construct a global weighted distance matrix; The core pulse fast association module is used to perform local density estimation based on the global weighted distance matrix, construct a decision graph and filter pulse points to obtain the number of shared nearest neighbors; The overlapping pulse evidence fusion module is used to construct the basic probability assignment and uncertainty function value, generate global confidence, label noise pulses, and output signal sorting results.

Images