Radiation Source Network Role Recognition System Based on Dynamic Graphs and Behavioral Fingerprints
By constructing a radiation source network role recognition system based on dynamic graphs and behavioral fingerprints, the problem of radiation source role recognition in complex electromagnetic environments has been solved, and high-precision radiation source role recognition and individual recognition have been achieved in noisy environments.
Patent Information
- Application Number
- CN202611502513.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing radiation source identification technologies struggle to accurately identify the role and type of radiation sources in complex electromagnetic environments. Traditional methods rely on radio frequency fingerprint features, which are susceptible to noise, and do not fully utilize the cooperative relationship information between radiation sources.
A radiation source network role recognition system based on dynamic graphs and behavioral fingerprints is constructed. Through pulse stream reception, transient fingerprint extraction, spatiotemporal correlation graph construction, dynamic graph feature learning, and behavioral fingerprint generation, the role type of radiation sources in network collaboration is identified.
It improves the accuracy and interpretability of radiation source identification, maintains high identification accuracy even when radio frequency fingerprint features are overwhelmed by noise, reduces the search space for large-scale fingerprint database comparisons through role discrimination and individual comparison, and captures the instantaneous collaborative relationships and long-term behavioral patterns of radiation sources.
Smart Images

Figure CN123116172A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of radar signal processing and radiation source identification technology, specifically a radiation source network role recognition system based on dynamic graphs and behavioral fingerprints. Background Technology
[0002] Individual radiation source identification refers to the technique of identifying the individual radiation source emitting the signal by utilizing the hardware differences contained in the received electromagnetic signal. Traditional individual radiation source identification technology mainly relies on the radio frequency fingerprint characteristics of the radiation source (such as the envelope shape of the signal, phase noise, frequency stability, etc., subtle differences introduced by hardware non-ideals) and the operating parameter characteristics of the radiation source (such as carrier frequency, pulse width, pulse repetition interval, etc., which can be directly extracted through pulse descriptors).
[0003] With the rapid development of phased array radar, multi-functional radar, and radar networking technology, the number of radiation sources in the electromagnetic environment in which radar operates has increased dramatically, and the cooperative relationship between radiation sources has become increasingly complex. Traditional radiation source identification methods are also facing serious technical challenges. Traditional radio frequency fingerprinting methods rely on subtle differences in the hardware of the radiation source itself. However, in non-cooperative scenarios, the signal transmission distance is long and the signal-to-noise ratio is low, and the radio frequency fingerprint features are easily submerged by noise, making it difficult to guarantee identification accuracy. Furthermore, existing radiation source identification methods based on pulse parameters (such as carrier frequency, pulse width, and time of arrival) only utilize the statistical features of the parameters of a single radiation source for identification, failing to fully utilize the cooperative relationship information between multiple radiation sources. At the same time, although there are schemes in the existing technology that use graph neural networks for radar signal sorting, the construction of their graph structures is mostly based on feature similarity or learnable attention coefficients, lacking explicit modeling of the radar network cooperative working mechanism. For example, CN120470380A constructs a dynamic graph based on the radar signal parameter matrix and performs signal sorting through a graph attention network, but its graph edges are only constructed based on feature similarity and do not consider the cooperative relationship between radiation sources in the three dimensions of pulse interval modulation, frequency hopping, and beam pointing.
[0004] Existing technologies include schemes for behavioral identification of communication radiation sources, but the behaviors they refer to are usually physical layer behaviors such as signaling signal type and time slot pattern, rather than the interactive behaviors of radiation sources in network coordination. To date, no technical solution can place radiation sources in the context of network coordination and identify the type of role they play in the cooperative network. Therefore, existing technologies have significant shortcomings. Summary of the Invention
[0005] The purpose of this invention is to provide a radiation source network role recognition system based on dynamic graphs and behavioral fingerprints to solve the problems mentioned in the background art.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a radiation source network role recognition system based on dynamic graphs and behavioral fingerprints, including a pulse stream receiving module, a transient fingerprint extraction module, a spatiotemporal correlation graph construction module, a dynamic graph feature learning module, a behavioral fingerprint generation module, and a joint recognition module; The pulse stream receiving module is used to receive the intercepted radar pulse stream, parse the pulse descriptor sequence from the obtained radar pulse stream pulse by pulse, and extract a baseband sampling segment containing the corresponding leading edge of each pulse. The baseband sampling segment corresponds one-to-one with each pulse descriptor in the pulse descriptor sequence. The pulse descriptor includes at least the arrival time, carrier frequency, pulse width, amplitude, angle of arrival, and intra-pulse modulation type. In this invention, during the process of extracting a baseband sampling segment containing the corresponding leading edge portion for each pulse, for each detected pulse, several sampling points are extracted forward and several sampling points are extracted backward based on the pulse leading edge detection point to form a baseband sampling segment containing the complete pulse leading edge transition process; this baseband sampling segment corresponds one-to-one with each pulse descriptor in the pulse descriptor sequence, and is used by the subsequent transient fingerprint extraction module.
[0007] The transient fingerprint extraction module is used to extract the leading-edge transient features of each baseband sampling segment, generate a transient fingerprint vector corresponding to the baseband sampling segment, and assign the transient fingerprint vector to the graph node of the corresponding pulse as the initial node feature of the graph node. The transient fingerprint extraction module described in this invention extracts the leading-edge transient features for each baseband sampling segment from one or more combinations of pulse rising edge envelope shape parameters, instantaneous phase noise statistics, and pulse top amplitude fluctuation features; The specific extraction method of the leading-edge transient features is as follows: The method for extracting the pulse rising edge envelope shape parameter is as follows: extract the envelope of the pulse leading edge in the baseband sampling segment to obtain the leading edge envelope curve; perform polynomial fitting or exponential fitting on the envelope curve, and extract the fitting coefficient as the envelope shape parameter.
[0008] The instantaneous phase noise statistics are extracted as follows: perform phase unwrapping operation on the baseband sampling segment to obtain the instantaneous phase curve; remove the linear phase component (corresponding to the carrier frequency component) from the instantaneous phase curve to obtain the residual phase noise sequence; and calculate the variance, skewness, kurtosis and power spectral density of the residual phase noise sequence, and use these parameters together as instantaneous phase noise features.
[0009] The method for extracting the amplitude fluctuation features at the top of the pulse is as follows: perform amplitude statistics on the sampling points corresponding to the top of the pulse in the baseband sampling segment, calculate the mean, variance, ratio of the maximum to the minimum value of the top amplitude, and the spectral features of the amplitude fluctuation, and combine these features into a top amplitude fluctuation feature vector.
[0010] The spatiotemporal correlation graph construction module uses each pulse as a graph node, calculates the arrival time difference (i.e., pulse repetition interval) between adjacent pulses based on the arrival time sequence, and identifies the pulse repetition interval skeleton period and the corresponding jitter range to obtain pulse interval modulation tracking results. The spatiotemporal correlation graph construction module also utilizes the obtained pulse interval modulation tracking results to establish edge relationships between graph nodes by combining carrier frequency, pulse width, and angle of arrival, and determines the edge attributes of each edge. Furthermore, it divides the continuous pulse stream into multiple time slices according to a preset time window (the width of the time window is adaptively adjusted according to the operating mode of the radiation source; the operating modes of the candidate radiation source include search mode, ...). Tracking and guidance modes; for search mode, the time window can be set to be longer to cover the complete beam scanning cycle; for tracking or guidance mode, the time window can be set to be shorter to capture rapidly changing cooperative relationships. Within each time slice, a dynamic pulse flow graph is formed, which is composed of nodes, initial node features, edges, and edge attributes. (Graphs between different time slices are temporally associated through shared node identifiers, where nodes belonging to the same radiation source maintain consistent identifiers in different time slices.) Based on the edge attribute distribution in the dynamic pulse flow graph, pulse nodes are clustered, and pulse nodes belonging to the same radiation source are associated with the corresponding candidate radiation sources. The specific process by which the spatiotemporal correlation graph construction module in this invention identifies the pulse repetition interval skeleton period and the corresponding jitter range includes: processing the arrival time series using a sequence difference histogram algorithm; constructing a difference histogram by calculating the time difference of all pairwise combinations in the pulse arrival time series; and extracting the most frequently occurring difference from the histogram as a candidate pulse repetition interval. For uneven pulse repetition interval signals, multiple sub-periods and their skeleton periods are further identified, and after identifying the skeleton period, the deviation of each pulse relative to the skeleton period is statistically analyzed to obtain the jitter range.
[0011] The dynamic graph feature learning module receives the obtained dynamic pulse flow graph. Within each time slice, it performs weighted aggregation on the obtained initial node features based on the neighborhood relationship and edge attributes of the nodes to obtain the local behavioral features of each node. Between adjacent time slices, nodes belonging to the same candidate radiation source or the same cooperative link are aggregated in time sequence based on the local behavioral features to generate a sequence of node dynamic behavioral features for each candidate radiation source. The behavior fingerprint generation module performs statistical pooling and pattern encoding operations on the node dynamic behavior feature sequences of each candidate radiation source, and extracts the interaction habit features of each candidate radiation source in terms of network timing patterns, beam pointing switching habits, and parameter coordination methods. Based on the obtained interaction habit features, the behavior fingerprint vector of each candidate radiation source is generated. The joint identification module matches the behavioral fingerprint vectors of each candidate radiation source with the pre-built radiation source behavioral fingerprint database, and combines the collaborative working mode exhibited by each candidate radiation source and the role constraints in the network to output the individual role identification results of each candidate radiation source. The individual role identification results include the network role type and individual identity identifier. In this invention, the transient fingerprint vector generated by the transient fingerprint extraction module is sent as the initial node feature to the spatiotemporal correlation graph construction module; the dynamic pulse flow graph generated by the spatiotemporal correlation graph construction module is input to the dynamic graph feature learning module; the node dynamic behavior feature sequence output by the dynamic graph feature learning module is passed to the behavior fingerprint generation module; and the behavior fingerprint vector output by the behavior fingerprint generation module is passed to the joint recognition module.
[0012] Furthermore, the spatiotemporal correlation graph construction module also utilizes the obtained pulse interval modulation tracking results to establish edge relationships between graph nodes and determine the edge attributes of each edge by combining carrier frequency, pulse width, and angle of arrival. Specifically, this process includes: For any two graph nodes, when the arrival time difference between the two nodes falls within the tolerance window of the jitter range of the corresponding pulse repetition interval skeleton period superposition jitter range of the obtained pulse interval modulation tracking result, a timing edge is established. When the carrier frequency difference between the two nodes is consistent with the previous carrier frequency hopping step size obtained by statistical analysis of adjacent values in the carrier frequency sequence (in this invention, the consistency criterion is that the two are the same or the error between them is within a preset range), a frequency hopping edge is established. When the beam pointing change estimated based on the angle of arrival sequence is the same as the preset sector switching sequence, a beam coordination edge is established. Specifically, the angle of arrival is mapped to the corresponding beam sector (e.g., dividing 360 degrees into several sectors). When the angle of arrival changes of two pulses conform to the preset sector switching sequence (e.g., switching from sector A to the adjacent sector B, or switching according to a specific scanning sequence), a beam coordination edge is established.
[0013] The spatiotemporal correlation graph construction module merges the simultaneously established timing edges, frequency hopping edges, and beam coordination edges into a composite edge, and generates edge attributes for this composite edge that include the aforementioned parameters. The edge attributes include at least the pulse interval modulation type, carrier frequency hopping step size, beam pointing change amount, and coordination period identifier. The coordination period identifier is jointly determined based on the pulse repetition interval skeleton period and the beam synchronization period. When the pulse repetition interval skeleton period and the beam synchronization period satisfy an integer multiple relationship, the period identifier corresponding to this integer multiple is used as the coordination period identifier.
[0014] Furthermore, the spatiotemporal correlation graph construction module has a built-in node pregrouping unit. Before establishing edge relationships, the node pregrouping unit pregroups each pulse node according to the carrier frequency jump pattern, pulse width jitter amplitude, and the continuity of the angle of arrival change (specifically, pulses with the same carrier frequency within the same jump step size range, the same pulse width within the same jitter range, and the same angle of arrival within the same sector range are grouped into the same pregroup). Subsequently, candidate edges are only established between node pairs that are within the same pregroup or between adjacent pregroups.
[0015] Furthermore, the edge attributes also include a cooperative link identifier and a synchronization error range; the cooperative link identifier is determined by the recurrence pattern of node pairs that simultaneously satisfy the timing edge condition and the beam cooperative edge condition on the time axis; the synchronization error range is jointly determined based on the jitter statistics of the pulse repetition interval skeleton period and the time deviation statistics of the beam pointing change moment in the obtained pulse interval modulation tracking results.
[0016] Furthermore, the dynamic graph feature learning module includes intra-time-slice aggregation units and inter-time-slice aggregation units; The aggregation unit within a time slice uses the carrier frequency hopping step size and beam pointing change in the edge attributes as gating reference values. It normalizes the obtained carrier frequency hopping step size and beam pointing change, and determines the weighting coefficients using an exponential decay method. The weighting coefficients are then used to weight and sum the initial node features of each neighboring node within the same time slice to obtain the corresponding local behavioral features. Specifically, for a target node within a certain time slice, its neighboring nodes include all nodes connected to that node through composite edges. During aggregation, the carrier frequency hopping step size and beam pointing change in the edge attributes are used as gating reference values. These two quantities are first normalized (mapped to between 0 and 1), and then the weighting coefficients are determined using an exponential decay method. The determined weights are... ,in This represents the normalized distance metric. This represents the attenuation coefficient. The obtained weights are used to weight and sum the initial node features of each neighboring node within the same time slice to obtain the local behavioral features of the target node.
[0017] The time-slice aggregation unit determines whether nodes in adjacent time slices belong to the same candidate radiation source or the same cooperative link based on the cooperative link identifier in the edge attribute. For nodes determined to belong to the same source, the corresponding local behavioral features are recursively weighted and summed in chronological order. The weight coefficient of historical features further away from the current time slice is smaller, and finally, the dynamic behavioral feature sequence of nodes of each candidate radiation source is obtained.
[0018] Specifically, the calculation formulas involved in obtaining the node dynamic behavior feature sequence of each candidate radiation source are as follows:
[0019] in, This represents the local behavioral characteristics of the current time slice. This represents the aggregation characteristics of the previous time slice. Represents the weight coefficient of the current feature and The value range is [0, 1].
[0020] Furthermore, the behavioral fingerprint generation module includes: The timing habit coding submodule is used to statistically analyze the transition probability, pulse repetition interval skeleton period, and jitter range of candidate radiation sources in multiple time slices to obtain the corresponding network timing habit characteristics; the transition probability is obtained by normalizing the pulse interval modulation type transition frequency of each node pair on the same cooperative link between adjacent time slices. The beam habit coding submodule is used to statistically analyze the beam pointing switching order, histogram distribution of dwell time of each beam, sector coverage order, and statistical distribution characteristics of beam pointing change of candidate radiation sources to obtain beam pointing habit features. Specifically, the method for obtaining the beam pointing switching order is to record the beam pointing (i.e., the sector corresponding to the angle of arrival) sequence of the candidate radiation source in each time slice and extract the switching order pattern therein; The method for obtaining the histogram distribution characteristics of dwell time for each beam is to accumulate and count the number of pulses that continuously fall into the same angle of arrival interval, and then obtain the dwell time histogram for each beam direction. The method for obtaining the sector coverage order is to encode the beam pointing switching order according to the sector number to form a sector coverage order sequence; The method to obtain the statistical distribution characteristics of beam pointing variation is to calculate the statistical distribution of beam pointing variation between adjacent pulses (such as mean, variance, kurtosis, etc.).
[0021] The parameter coordination encoding submodule is used to statistically analyze the joint change trends of the carrier frequency difference, pulse width difference, and time of arrival difference between the candidate radiation source and its cooperating nodes over time, thereby obtaining parameter coordination characteristics. The joint change trend is obtained by calculating the first-order difference of the carrier frequency difference, pulse width difference, and time of arrival difference of corresponding node pairs in adjacent time slices within the same cooperative link, and performing distribution statistics (including mean, variance, skewness, and kurtosis) on the obtained difference results. In this invention, a cooperative node represents another radiation source associated with the current candidate radiation source through a cooperative link.
[0022] The fusion submodule is used to splice and fuse or weighted fuse the obtained network timing habitual features, beam pointing habitual features and parameter matching features to obtain the behavioral fingerprint vector of the corresponding candidate radiation source.
[0023] Furthermore, when the fusion submodule performs weighted fusion of the obtained network timing habitual characteristics, beam pointing habitual characteristics, and parameter coordination characteristics, the weighting coefficients used are dynamically adjusted according to the current working mode of the candidate radiation source. The current working mode of the candidate radiation source includes search mode, tracking mode, and guidance mode. When the candidate radiation source is currently in the search mode, increase the weight corresponding to the beam pointing habit feature; When the candidate radiation source is currently in tracking or guidance mode, the weights corresponding to the parameters and features are increased. The weight increase scheme in this invention is preset.
[0024] Furthermore, the joint identification module includes a role identification submodule and an individual comparison submodule; The role recognition submodule identifies candidate radiation sources as search roles, tracking roles, guidance roles, or blind spot filling roles based on the beam pointing habit features and parameter matching features in the obtained behavioral fingerprint vector. In the process of distinguishing different roles in this invention, the characteristics corresponding to each role are as follows: The characteristics of the search character are that the beam direction scans regularly over a large range, the dwell time of each beam is short and uniform, and the parameter coordination characteristics are weak (i.e., there is little parameter coordination with other radiation sources). The characteristics of the tracking role are that the beam is focused in a specific direction, the beam dwell time is long, and the parameter matching characteristics are obvious (parameter adjustment is consistent with the cooperative radiation source). The guidance role is characterized by a highly concentrated beam pointing and strong parameter matching characteristics (requiring high-precision frequency and time synchronization). The characteristics of a gap-filling role are that the beam is pointed to scan a specific area, the parameter matching characteristics are moderate, and it usually works in areas where the main radar has insufficient coverage.
[0025] The individual comparison submodule filters out registered radiation source templates that match the roles of candidate radiation sources from the obtained radiation source behavior fingerprint database. Within this filtering range, it calculates the similarity between the behavior fingerprint vector of the candidate radiation source and the preset behavior fingerprint vector corresponding to each registered radiation source template (obtained by cosine similarity calculation). The radiation source individual corresponding to the registered radiation source template with the highest similarity exceeding the preset threshold is output as the identification result. If the similarity of all registered radiation source templates is less than the preset threshold, the candidate radiation source is marked as an unknown radiation source, and an alarm message is generated.
[0026] The construction method of the radiation source behavioral fingerprint database in this invention is as follows: During the testing phase, known radiation source individuals are observed, and behavioral fingerprint vectors of each radiation source are extracted according to the processing flow of the behavioral fingerprint generation module. These vectors are then associated with the radiation source individual identifier and role type and stored to form an initial fingerprint database. During the system operation phase, for new observation samples that are successfully identified and have high confidence, their behavioral fingerprint vectors are updated with the corresponding templates in the fingerprint database using a moving average method to achieve dynamic maintenance of the fingerprint database.
[0027] Furthermore, the transient fingerprint vector serves as an initial node feature to provide physical difference information at the pulse level for the graph structure. This physical difference information, after being aggregated by the dynamic graph feature learning module, indirectly participates in the subsequent individual role discrimination process.
[0028] Compared with the prior art, the beneficial effects achieved by the present invention are: (1) This invention constructs composite edges with three physical meanings: time-series edges, frequency-jump edges, and beam coordination edges, thereby constructing the coordination relationship between radiation sources in the three dimensions of time, frequency, and space. Compared with the graph construction method based on feature similarity, it has stronger interpretability. (2) By introducing the concept of behavioral fingerprint, this invention extracts the interactive behavioral features of radiation sources from three dimensions: networking timing habits, beam pointing switching habits, and parameter matching habits. This elevates radiation source identification from traditional physical layer fingerprint identification to behavioral layer identification, and can maintain high identification accuracy even when radio frequency fingerprint features are submerged by noise. (3) The present invention first performs role identification (search, tracking, guidance and blind spot filling) on the radiation source, and then performs individual comparison in the same role template library. Through the hierarchical architecture of role recognition and individual recognition, the search space of large-scale fingerprint database comparison is reduced and the accuracy of individual recognition is improved. (4) The present invention adopts a two-level feature learning architecture of intra-time slice aggregation and inter-time slice aggregation, which can simultaneously capture the instantaneous collaborative relationship and long-term behavioral evolution pattern of radiation sources. Attached Figure Description
[0029] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the structure of the radiation source network role recognition system based on dynamic graphs and behavioral fingerprints of the present invention. Detailed Implementation
[0030] 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.
[0031] Please see Figure 1 The radiation source network role recognition system based on dynamic graphs and behavioral fingerprints in this embodiment includes a pulse stream receiving module, a transient fingerprint extraction module, a spatiotemporal correlation graph construction module, a dynamic graph feature learning module, a behavioral fingerprint generation module, and a joint recognition module. The pulse stream receiving module is used to receive the intercepted radar pulse stream, parse the pulse descriptor sequence from the obtained radar pulse stream pulse by pulse, and extract a baseband sampling segment containing the corresponding leading edge of each pulse. The baseband sampling segment corresponds one-to-one with each pulse descriptor in the pulse descriptor sequence. The pulse descriptor includes at least the arrival time, carrier frequency, pulse width, amplitude, angle of arrival, and intra-pulse modulation type. In this invention, during the process of extracting a baseband sampling segment containing the corresponding leading edge portion for each pulse, for each detected pulse, several sampling points are extracted forward and several sampling points are extracted backward based on the pulse leading edge detection point to form a baseband sampling segment containing the complete pulse leading edge transition process; this baseband sampling segment corresponds one-to-one with each pulse descriptor in the pulse descriptor sequence, and is used by the subsequent transient fingerprint extraction module.
[0032] The transient fingerprint extraction module is used to extract the leading-edge transient features of each baseband sampling segment, generate a transient fingerprint vector corresponding to the baseband sampling segment, and assign the transient fingerprint vector to the graph node of the corresponding pulse as the initial node feature of the graph node. The transient fingerprint vector, as the initial node feature, is used to provide physical difference information at the pulse level for the graph structure. After being aggregated by the dynamic graph feature learning module, the physical difference information indirectly participates in the subsequent individual role discrimination process.
[0033] In this invention, the transient fingerprint extraction module extracts the leading-edge transient features for each baseband sampling segment from one or more combinations of pulse rising edge envelope shape parameters, instantaneous phase noise statistics, and pulse top amplitude fluctuation features; The specific extraction method of the leading-edge transient features is as follows: The extraction method for the pulse rising edge envelope shape parameters is as follows: The envelope of the pulse leading edge in the baseband sampling segment is extracted to obtain the leading edge envelope curve (in this embodiment, the analytic signal is obtained by performing a Hilbert transform on the sampling points of the leading edge, and the envelope curve is obtained by taking the magnitude of the analytic signal); polynomial or exponential fitting is performed on the envelope curve, and the fitting coefficients are extracted as envelope shape parameters. In this embodiment, the rise time (time from 10% amplitude to 90% amplitude), rise rate (maximum value of the envelope derivative), and statistics of the second derivative of the envelope are calculated, and these parameters are combined into an envelope shape feature vector.
[0034] The instantaneous phase noise statistics are extracted as follows: perform phase unwrapping operation on the baseband sampling segment to obtain the instantaneous phase curve; remove the linear phase component (corresponding to the carrier frequency component) from the instantaneous phase curve to obtain the residual phase noise sequence; and calculate the variance, skewness, kurtosis and power spectral density of the residual phase noise sequence, and use these parameters together as instantaneous phase noise features.
[0035] The method for extracting the amplitude fluctuation features at the top of the pulse is as follows: perform amplitude statistics on the sampling points corresponding to the top of the pulse in the baseband sampling segment, calculate the mean, variance, ratio of the maximum to the minimum value of the top amplitude, and the spectral features of the amplitude fluctuation (specifically, the amplitude spectrum features of the low-frequency components extracted after performing a fast Fourier transform on the amplitude sequence), and combine these features into a top amplitude fluctuation feature vector.
[0036] The spatiotemporal correlation graph construction module uses each pulse as a graph node, calculates the arrival time difference (i.e., pulse repetition interval) between adjacent pulses based on the arrival time sequence, and identifies the pulse repetition interval skeleton period and the corresponding jitter range to obtain the pulse interval modulation tracking result. The spatiotemporal correlation graph construction module also uses the obtained pulse interval modulation tracking result to establish edge relationships between graph nodes in conjunction with carrier frequency, pulse width, and angle of arrival, and determines the edge attributes of each edge. Furthermore, it divides the continuous pulse stream into multiple time slices according to a preset time window (the width of the time window is adaptively adjusted according to the operating mode of the radiation source; the operating modes of the candidate radiation source include search mode, tracking mode, and guidance mode; for search mode, the time window can be set longer to cover the complete beam scanning cycle; for tracking mode, the time window can be set longer to cover the entire beam scanning cycle; for other modes, the time window can be set longer to cover the entire beam scanning cycle). In the tracking or guidance mode, the time window can be set to be shorter to capture rapidly changing cooperative relationships. Within each time slice, a dynamic pulse flow graph is formed, which is composed of nodes, initial node features, edges, and edge attributes. (Graphs in different time slices are temporally associated through shared node identifiers, where nodes belonging to the same radiation source maintain consistent identifiers in different time slices.) The pulse nodes are clustered according to the edge attribute distribution in the dynamic pulse flow graph. (In this embodiment, a density-based clustering algorithm or a graph-based Louvain algorithm is used to group pulse nodes with dense edge connections into the same cluster. Each cluster corresponds to a candidate radiation source, and all pulse nodes within the cluster belong to the same radiation source.) Pulse nodes belonging to the same radiation source are associated with the corresponding candidate radiation source. The specific process by which the spatiotemporal correlation graph construction module in this invention identifies the pulse repetition interval skeleton period and the corresponding jitter range includes: processing the arrival time series using a sequence difference histogram algorithm; constructing a difference histogram by calculating the time difference of all pairwise combinations in the pulse arrival time series; and extracting the most frequently occurring difference from the histogram as a candidate pulse repetition interval. For uneven pulse repetition interval signals, multiple sub-periods and their skeleton periods are further identified, and after identifying the skeleton period, the deviation of each pulse relative to the skeleton period is statistically analyzed to obtain the jitter range.
[0037] The spatiotemporal correlation graph construction module also utilizes the obtained pulse interval modulation tracking results to establish edge relationships between graph nodes and determine the edge attributes of each edge by combining carrier frequency, pulse width, and angle of arrival. Specifically, this process includes: For any two graph nodes, when the arrival time difference between the two nodes falls within the tolerance window of the jitter range of the corresponding pulse repetition interval skeleton period superposition jitter range of the obtained pulse interval modulation tracking result, a timing edge is established. When the carrier frequency difference between the two nodes is consistent with the previous carrier frequency hopping step size obtained by statistical analysis of adjacent values in the carrier frequency sequence (in this invention, the consistency criterion is that the two are the same or the error between them is within a preset range), a frequency hopping edge is established. When the beam pointing change estimated based on the angle of arrival sequence is the same as the preset sector switching sequence, a beam coordination edge is established. Specifically, the angle of arrival is mapped to the corresponding beam sector (e.g., dividing 360 degrees into several sectors). When the angle of arrival changes of two pulses conform to the preset sector switching sequence (e.g., switching from sector A to the adjacent sector B, or switching according to a specific scanning sequence), a beam coordination edge is established.
[0038] The spatiotemporal correlation graph construction module merges the simultaneously established timing edges, frequency hopping edges, and beam coordination edges into a composite edge, and generates edge attributes for this composite edge that include the aforementioned parameters. The edge attributes include at least the pulse interval modulation type, carrier frequency hopping step size, beam pointing change amount, and coordination period identifier. The coordination period identifier is jointly determined based on the pulse repetition interval skeleton period and the beam synchronization period. When the pulse repetition interval skeleton period and the beam synchronization period satisfy an integer multiple relationship, the period identifier corresponding to this integer multiple is used as the coordination period identifier.
[0039] The spatiotemporal correlation graph construction module has a built-in node pregrouping unit. Before establishing edge relationships, the node pregrouping unit pregroups each pulse node according to the carrier frequency jump pattern, pulse width jitter amplitude, and the continuity of the angle of arrival change (specifically, pulses with the same carrier frequency within the same jump step size range, the same pulse width within the same jitter range, and the same angle of arrival within the same sector range are grouped into the same pregroup). Subsequently, candidate edges are only established between node pairs that are within the same pregroup or between adjacent pregroups.
[0040] The edge attributes also include a cooperative link identifier and a synchronization error range; the cooperative link identifier is determined by the recurrence pattern of node pairs that simultaneously satisfy the timing edge condition and the beam cooperative edge condition on the time axis; the synchronization error range is jointly determined by the jitter statistics of the pulse repetition interval skeleton period and the time deviation statistics of the beam pointing change moment in the obtained pulse interval modulation tracking results.
[0041] The dynamic graph feature learning module receives the obtained dynamic pulse flow graph. Within each time slice, it performs weighted aggregation on the obtained initial node features based on the neighborhood relationship and edge attributes of the nodes to obtain the local behavioral features of each node. Between adjacent time slices, nodes belonging to the same candidate radiation source or the same cooperative link are aggregated in time sequence based on the local behavioral features to generate a sequence of node dynamic behavioral features for each candidate radiation source. The dynamic graph feature learning module includes intra-time-slice aggregation units and inter-time-slice aggregation units; The aggregation unit within a time slice uses the carrier frequency hopping step size and beam pointing change in the edge attributes as gating reference values. It normalizes the obtained carrier frequency hopping step size and beam pointing change, and determines the weighting coefficients using an exponential decay method. The weighting coefficients are then used to weight and sum the initial node features of each neighboring node within the same time slice to obtain the corresponding local behavioral features. Specifically, for a target node within a certain time slice, its neighboring nodes include all nodes connected to that node through composite edges. During aggregation, the carrier frequency hopping step size and beam pointing change in the edge attributes are used as gating reference values. These two quantities are first normalized (mapped to between 0 and 1), and then the weighting coefficients are determined using an exponential decay method. The determined weights are... ,in This represents the normalized distance metric. This represents the attenuation coefficient. The obtained weights are used to weight and sum the initial node features of each neighboring node within the same time slice to obtain the local behavioral features of the target node.
[0042] The time-slice aggregation unit determines whether nodes in adjacent time slices belong to the same candidate radiation source or the same cooperative link based on the cooperative link identifier in the edge attribute. For nodes determined to belong to the same source, the corresponding local behavioral features are recursively weighted and summed in chronological order. The weight coefficient of historical features further away from the current time slice is smaller, and finally, the dynamic behavioral feature sequence of nodes of each candidate radiation source is obtained.
[0043] Specifically, the calculation formulas involved in obtaining the node dynamic behavior feature sequence of each candidate radiation source are as follows:
[0044] in, This represents the local behavioral characteristics of the current time slice. This represents the aggregation characteristics of the previous time slice. Represents the weight coefficient of the current feature and The value range is [0, 1].
[0045] The behavior fingerprint generation module performs statistical pooling and pattern encoding operations on the node dynamic behavior feature sequences of each candidate radiation source, and extracts the interaction habit features of each candidate radiation source in terms of network timing patterns, beam pointing switching habits, and parameter coordination methods. Based on the obtained interaction habit features, the behavior fingerprint vector of each candidate radiation source is generated. The behavioral fingerprint generation module includes: The timing habit coding submodule is used to statistically analyze the transition probability, pulse repetition interval skeleton period, and jitter range of candidate radiation sources in multiple time slices to obtain the corresponding network timing habit characteristics; the transition probability is obtained by normalizing the pulse interval modulation type transition frequency of each node pair on the same cooperative link between adjacent time slices. The beam habit coding submodule is used to statistically analyze the beam pointing switching order, histogram distribution of dwell time of each beam, sector coverage order, and statistical distribution characteristics of beam pointing change of candidate radiation sources to obtain beam pointing habit features. Specifically, the method for obtaining the beam pointing switching order is to record the beam pointing (i.e., the sector corresponding to the angle of arrival) sequence of the candidate radiation source in each time slice and extract the switching order pattern therein; The method for obtaining the histogram distribution characteristics of dwell time for each beam is to accumulate and count the number of pulses that continuously fall into the same angle of arrival interval, and then obtain the dwell time histogram for each beam direction. The method for obtaining the sector coverage order is to encode the beam pointing switching order according to the sector number to form a sector coverage order sequence; The method to obtain the statistical distribution characteristics of beam pointing variation is to calculate the statistical distribution of beam pointing variation between adjacent pulses (such as mean, variance, kurtosis, etc.).
[0046] The parameter coordination encoding submodule is used to statistically analyze the joint change trends of the carrier frequency difference, pulse width difference, and time of arrival difference between the candidate radiation source and its cooperating nodes over time, thereby obtaining parameter coordination characteristics. The joint change trend is obtained by calculating the first-order difference of the carrier frequency difference, pulse width difference, and time of arrival difference of corresponding node pairs in adjacent time slices within the same cooperative link, and performing distribution statistics (including mean, variance, skewness, and kurtosis) on the obtained difference results. In this invention, a cooperative node represents another radiation source associated with the current candidate radiation source through a cooperative link.
[0047] The fusion submodule is used to splice and fuse or weighted fuse the obtained network timing habitual features, beam pointing habitual features and parameter matching features to obtain the behavioral fingerprint vector of the corresponding candidate radiation source.
[0048] When the fusion submodule performs weighted fusion of the obtained network timing habitual characteristics, beam pointing habitual characteristics and parameter coordination characteristics, the weighting coefficients used are dynamically adjusted according to the current working mode of the candidate radiation source. The current working mode of the candidate radiation source includes search mode, tracking mode and guidance mode. When the candidate radiation source is currently in the search mode, increase the weight corresponding to the beam pointing habit feature; When the candidate radiation source is currently in tracking or guidance mode, the weights corresponding to the parameters and features are increased. The weight increase scheme in this invention is preset.
[0049] The joint identification module matches the behavioral fingerprint vectors of each candidate radiation source with the pre-built radiation source behavioral fingerprint database, and combines the collaborative working mode exhibited by each candidate radiation source and the role constraints in the network to output the individual role identification results of each candidate radiation source. The individual role identification results include the network role type and individual identity identifier. The joint identification module includes a role identification submodule and an individual comparison submodule; The role recognition submodule identifies candidate radiation sources as search roles, tracking roles, guidance roles, or blind spot filling roles based on the beam pointing habit features and parameter matching features in the obtained behavioral fingerprint vector. In the process of distinguishing different roles in this invention, the characteristics corresponding to each role are as follows: The characteristics of the search character are that the beam direction scans regularly over a large range, the dwell time of each beam is short and uniform, and the parameter coordination characteristics are weak (i.e., there is little parameter coordination with other radiation sources). The characteristics of the tracking role are that the beam is focused in a specific direction, the beam dwell time is long, and the parameter matching characteristics are obvious (parameter adjustment is consistent with the cooperative radiation source). The guidance role is characterized by a highly concentrated beam pointing and strong parameter matching characteristics (requiring high-precision frequency and time synchronization). The characteristics of a gap-filling role are that the beam is pointed to scan a specific area, the parameter matching characteristics are moderate, and it usually works in areas where the main radar has insufficient coverage.
[0050] The individual comparison submodule filters out registered radiation source templates that match the roles of candidate radiation sources from the obtained radiation source behavior fingerprint database. Within this filtering range, it calculates the similarity between the behavior fingerprint vector of the candidate radiation source and the preset behavior fingerprint vector corresponding to each registered radiation source template (obtained by cosine similarity calculation). The radiation source individual corresponding to the registered radiation source template with the highest similarity exceeding the preset threshold is output as the identification result. If the similarity of all registered radiation source templates is less than the preset threshold, the candidate radiation source is marked as an unknown radiation source, and an alarm message is generated.
[0051] The construction method of the radiation source behavioral fingerprint database in this invention is as follows: During the testing phase, known radiation source individuals are observed, and behavioral fingerprint vectors of each radiation source are extracted according to the processing flow of the behavioral fingerprint generation module. These vectors are then associated with the radiation source individual identifier and role type and stored to form an initial fingerprint database. During the system operation phase, for new observation samples that are successfully identified and have high confidence, their behavioral fingerprint vectors are updated with the corresponding templates in the fingerprint database using a moving average method to achieve dynamic maintenance of the fingerprint database.
[0052] In this invention, the transient fingerprint vector generated by the transient fingerprint extraction module is sent as the initial node feature to the spatiotemporal correlation graph construction module; the dynamic pulse flow graph generated by the spatiotemporal correlation graph construction module is input to the dynamic graph feature learning module; the node dynamic behavior feature sequence output by the dynamic graph feature learning module is passed to the behavior fingerprint generation module; and the behavior fingerprint vector output by the behavior fingerprint generation module is passed to the joint recognition module.
[0053] 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.
[0054] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A radiation source network role recognition system based on dynamic graphs and behavioral fingerprints, characterized in that, It includes a pulse stream receiving module, a transient fingerprint extraction module, a spatiotemporal correlation graph construction module, a dynamic graph feature learning module, a behavioral fingerprint generation module, and a joint recognition module; The pulse stream receiving module is used to receive the intercepted radar pulse stream, parse the pulse descriptor sequence from the obtained radar pulse stream pulse by pulse, and extract a baseband sampling segment containing the corresponding leading edge of each pulse. The baseband sampling segment corresponds one-to-one with each pulse descriptor in the pulse descriptor sequence. The pulse descriptor includes at least the arrival time, carrier frequency, pulse width, amplitude, angle of arrival, and intra-pulse modulation type. The transient fingerprint extraction module is used to extract the leading-edge transient features of each baseband sampling segment, generate a transient fingerprint vector corresponding to the baseband sampling segment, and assign the transient fingerprint vector to the graph node of the corresponding pulse as the initial node feature of the graph node. The spatiotemporal correlation graph construction module uses each pulse as a graph node, calculates the arrival time difference between adjacent pulses based on the arrival time sequence, and identifies the pulse repetition interval skeleton period and the corresponding jitter range to obtain pulse interval modulation tracking results. The spatiotemporal correlation graph construction module also uses the obtained pulse interval modulation tracking results to establish edge relationships between each graph node in conjunction with the carrier frequency, pulse width, and angle of arrival, and determines the edge attributes of each edge. It also divides the continuous pulse stream into multiple time slices according to a preset time window, forming a dynamic pulse flow graph composed of nodes, initial node features, edges, and edge attributes within each time slice. Based on the edge attribute distribution in the dynamic pulse flow graph, the pulse nodes are clustered, and pulse nodes belonging to the same radiation source are associated with the corresponding candidate radiation sources. The dynamic graph feature learning module receives the obtained dynamic pulse flow graph. Within each time slice, it performs weighted aggregation on the obtained initial node features based on the neighborhood relationship and edge attributes of the nodes to obtain the local behavioral features of each node. Between adjacent time slices, nodes belonging to the same candidate radiation source or the same cooperative link are aggregated in time sequence based on the local behavioral features to generate a sequence of node dynamic behavioral features for each candidate radiation source. The behavior fingerprint generation module performs statistical pooling and pattern encoding operations on the node dynamic behavior feature sequences of each candidate radiation source, and extracts the interaction habit features of each candidate radiation source in terms of network timing patterns, beam pointing switching habits, and parameter coordination methods. Based on the obtained interaction habit features, the behavior fingerprint vector of each candidate radiation source is generated. The joint identification module matches the behavioral fingerprint vectors of each candidate radiation source with the pre-built radiation source behavioral fingerprint database, and combines the collaborative working mode exhibited by each candidate radiation source with the role constraints in the network to output the individual role identification results of each candidate radiation source. The individual role identification results include the network role type and individual identity identifier.
2. The radiation source network role recognition system based on dynamic graphs and behavioral fingerprints according to claim 1, characterized in that, The spatiotemporal correlation graph construction module also utilizes the obtained pulse interval modulation tracking results to establish edge relationships between graph nodes and determine the edge attributes of each edge by combining carrier frequency, pulse width, and angle of arrival. Specifically, this process includes: For any two graph nodes, when the arrival time difference between the two nodes falls within the tolerance window of the jitter range of the corresponding pulse repetition interval skeleton period superposition jitter range of the obtained pulse interval modulation tracking result, a timing edge is established. When the carrier frequency difference between the two nodes is consistent with the previous hop carrier frequency hopping step size obtained by statistical analysis of adjacent values in the carrier frequency sequence, a frequency hopping edge is established. When the beam pointing change estimated based on the angle of arrival sequence is the same as the preset sector switching sequence, a beam coordination edge is established. The spatiotemporal correlation graph construction module merges the simultaneously established timing edges, frequency hopping edges, and beam coordination edges into a composite edge, and generates edge attributes for this composite edge that include the aforementioned parameters. The edge attributes include at least the pulse interval modulation type, carrier frequency hopping step size, beam pointing change amount, and coordination period identifier. The coordination period identifier is jointly determined based on the pulse repetition interval skeleton period and the beam synchronization period. When the pulse repetition interval skeleton period and the beam synchronization period satisfy an integer multiple relationship, the period identifier corresponding to this integer multiple is used as the coordination period identifier.
3. The radiation source network role recognition system based on dynamic graphs and behavioral fingerprints according to claim 2, characterized in that, The spatiotemporal correlation graph construction module has a built-in node pregrouping unit. Before establishing edge relationships, the node pregrouping unit pregroups each pulse node according to the carrier frequency jump law, pulse width jitter amplitude and the continuity of the angle of arrival change. Subsequently, candidate edges are only established between node pairs that are within the same pregroup or between adjacent pregroups.
4. The radiation source network role recognition system based on dynamic graphs and behavioral fingerprints according to claim 3, characterized in that, The edge attributes also include a cooperative link identifier and a synchronization error range; the cooperative link identifier is determined by the recurrence pattern of node pairs that simultaneously satisfy the timing edge condition and the beam cooperative edge condition on the time axis; the synchronization error range is jointly determined by the jitter statistics of the pulse repetition interval skeleton period and the time deviation statistics of the beam pointing change moment in the obtained pulse interval modulation tracking results.
5. The radiation source network role recognition system based on dynamic graphs and behavioral fingerprints according to claim 1, characterized in that, The dynamic graph feature learning module includes intra-time-slice aggregation units and inter-time-slice aggregation units; The aggregation unit within the time slice uses the carrier frequency hopping step size and beam pointing change in the edge attributes as gating reference values, normalizes the obtained carrier frequency hopping step size and beam pointing change respectively, and determines the weight coefficient in the form of exponential decay. The obtained weight coefficient is used to perform weighted summation on the initial node features of each neighboring node in the same time slice to obtain the corresponding local behavioral features. The time-slice aggregation unit determines whether nodes in adjacent time slices belong to the same candidate radiation source or the same cooperative link based on the cooperative link identifier in the edge attribute. For nodes determined to belong to the same source, the corresponding local behavioral features are recursively weighted and summed in chronological order. The weight coefficient of historical features further away from the current time slice is smaller, and finally, the dynamic behavioral feature sequence of nodes of each candidate radiation source is obtained.
6. The radiation source network role recognition system based on dynamic graphs and behavioral fingerprints according to claim 1, characterized in that, The behavioral fingerprint generation module includes: The timing habit coding submodule is used to statistically analyze the transition probability of pulse interval modulation type, pulse repetition interval skeleton period and jitter range of candidate radiation sources in multiple time slices to obtain the corresponding network timing habit characteristics; the transition probability is obtained by normalizing the pulse interval modulation type transition frequency of each node pair on the same cooperative link between adjacent time slices. The beam habit coding submodule is used to statistically analyze the beam pointing switching order, the histogram distribution of each beam dwell time, the sector coverage order, and the statistical distribution characteristics of beam pointing changes of candidate radiation sources to obtain beam pointing habit features; the histogram of the beam dwell time is obtained by accumulating and statistically analyzing the number of pulses that continuously fall into the same angle of arrival interval. The parameter coordination coding submodule is used to statistically analyze the joint change trend of the carrier frequency difference, pulse width difference, and time of arrival difference between the candidate radiation source and its cooperating node over time, thereby obtaining the parameter coordination characteristics. The joint change trend is obtained by calculating the first-order difference of the carrier frequency difference, pulse width difference, and time of arrival difference of the corresponding node pairs in adjacent time slices within the same cooperative link and performing distribution statistics on the obtained difference results. The fusion submodule is used to splice and fuse or weighted fuse the obtained network timing habitual features, beam pointing habitual features and parameter matching features to obtain the behavioral fingerprint vector of the corresponding candidate radiation source.
7. The radiation source network role recognition system based on dynamic graphs and behavioral fingerprints according to claim 6, characterized in that, When the fusion submodule performs weighted fusion of the obtained network timing habitual characteristics, beam pointing habitual characteristics and parameter coordination characteristics, the weighting coefficients used are dynamically adjusted according to the current working mode of the candidate radiation source. The current working mode of the candidate radiation source includes search mode, tracking mode and guidance mode. When the candidate radiation source is currently in the search mode, increase the weight corresponding to the beam pointing habit feature; When the candidate radiation source is currently in tracking or guidance mode, increase the weight of the parameter matching feature.
8. The radiation source network role recognition system based on dynamic graphs and behavioral fingerprints according to claim 1, characterized in that, The joint identification module includes a role identification submodule and an individual comparison submodule; The role recognition submodule identifies candidate radiation sources as search roles, tracking roles, guidance roles, or blind spot filling roles based on the beam pointing habit features and parameter matching features in the obtained behavioral fingerprint vector. The individual comparison submodule filters out registered radiation source templates that match the roles of candidate radiation sources from the obtained radiation source behavior fingerprint database. Within this filtering range, it calculates the similarity between the behavior fingerprint vector of the candidate radiation source and the preset behavior fingerprint vectors corresponding to each registered radiation source template. The radiation source individual corresponding to the registered radiation source template with the highest similarity exceeding the preset threshold is output as the identification result. If the similarity of all registered radiation source templates is less than the preset threshold, the candidate radiation source is marked as an unknown radiation source, and an alarm message is generated.
9. The radiation source network role recognition system based on dynamic graphs and behavioral fingerprints according to claim 1, characterized in that, The transient fingerprint extraction module extracts the leading-edge transient features for each baseband sampling segment from one or more combinations of pulse rising edge envelope shape parameters, instantaneous phase noise statistics, and pulse top amplitude fluctuation features. The transient fingerprint vector serves as the initial node feature to provide physical difference information at the pulse level for the graph structure. After being aggregated by the dynamic graph feature learning module, this physical difference information indirectly participates in the subsequent individual role discrimination process.
