Radiation source signal classification method based on pulse-level multi-source separation

Abstract

This invention discloses a radiation source signal classification method based on pulse-level multi-source separation, mainly addressing the problem of poor radiation source signal classification performance in complex electromagnetic environments using existing technologies. The implementation steps are as follows: 1) Receiving broadband intermediate frequency or baseband signals and performing pulse detection; 2) Estimating the direction of arrival and constructing pulse descriptors for the detected pulses based on a multi-element antenna structure; 3) Modeling a pulse-level multi-dimensional feature space and performing density peak clustering based on temporal consistency constraints within this space; 4) Adaptively estimating the number of radiation sources and reconstructing the pulse sequence, performing multi-physical domain feature extraction; 5) Fusing the multi-physical domain features to complete the classification and identification of radiation source signals. This invention can effectively distinguish pulses from different radiation sources in complex electromagnetic environments, while effectively reducing the system's storage and computing resource consumption while ensuring high recognition performance.

Description

Technical Field

[0001] This invention belongs to the field of electromagnetic spectrum intelligent sensing technology, and further relates to radiation source signal classification technology. Specifically, it is a radiation source signal classification method based on pulse-level multi-source separation, which can be used to simultaneously extract features and identify individuals from multiple co-frequency / adjacent-frequency radiation sources in complex electromagnetic environments. Background Technology

[0002] With the rapid development of radar detection, communication countermeasures, and electronic support reconnaissance (ESM / ELINT) systems, radiation sources in complex electromagnetic spaces are exhibiting trends such as diverse systems, maneuvering parameters, dense co-channel activity, and increased low probability of intercept (LPI). Receivers, under broadband, long-duration, and strong interference conditions, often do not acquire "clean signals" from a single radiation source, but rather highly overlapping and mixed pulse descriptor word (PDW) sequences of multiple radiation source pulses across multiple dimensions, including time of arrival (TOA), frequency (RF / IF), pulse width (PW), amplitude (PA), and direction of arrival (DOA). Simultaneously, due to propagation fading, noise, co-channel suppression / deception interference, and hardware non-ideal factors (clock drift, quantization error, frequency offset, measurement jitter, etc.), PDW parameters and their statistical characteristics undergo significant distortion, transforming radiation source identification from an "ideal classification problem" into a system-level challenge of "multi-source aliasing separation + fingerprint / system identification + online incremental updates."

[0003] In the engineering process, radiation source identification typically includes: pulse detection and parameter measurement → pulse sorting / deinterleaving → feature construction (statistical / time-frequency / fingerprint) → classification and identification (system / individual). Among these, sorting / deinterleaving determines whether subsequent features "come from the same radiation source," which is a typical front-end bottleneck: once cross-source misassignment occurs, even if the subsequent classification network is powerful, its recognition performance will significantly decrease due to the contamination of the input sequence. In multi-source aliasing scenarios, the key technical challenges faced by existing methods mainly include: 1. Difficulty in effectively separating co-frequency or adjacent-frequency radiation sources: Traditional methods based on frame-level statistics or coarse-grained features struggle to accurately characterize individual radiation source features at the pulse granularity, leading to blurred boundaries between sources; 2. Prone to cross-source mis-clustering / splitting of homogeneous sources under multi-source aliasing: Existing clustering and sorting methods are prone to cross-source mis-merging or homogeneous source splitting under conditions of strong aliasing, noise, and parameter maneuvering, affecting subsequent classification accuracy; 3. Difficulty in adaptive estimation when the number of radiation sources is unknown and dynamically changing: Existing methods often rely on preset source numbers or manual thresholds to determine the number of clusters, making it difficult to adapt to changes in the number of sources caused by power-on / off cycles, intermittent emission, etc.; 4. Deep learning methods often rely on fixed-class training and suffer from catastrophic forgetting: When new radiation sources are added online, the lack of a reasonable continuous / incremental update mechanism can easily lead to a decline in the performance of old classes and affect long-term operation.

[0004] To address the issues of radar pulse deinterleaving and radiation source identification, various schemes have been proposed, but they still have shortcomings under conditions of dense co-frequency pulses and multi-source aliasing. Kenan Gençol; Ali Kara; Nuray At. Improvementson deinterleaving of radar pulses in dynamically varying signal environments. Digital Signal Processing, 2017, 69: 86–93. DOI: 10.1016 / j.dsp.2017.06.010. This paper proposes a multi-parameter fusion approach for pulse deinterleaving, which extracts PRI based on TOA and combines it with PDW parameters such as DOA, PW, and RF for sorting to improve multi-source separation capability. While this scheme can improve the separation effect to some extent, it is sensitive to measurement errors, missing pulses, and parameter jitter; it is still prone to cross-source misallocation when co-frequency pulses are dense and PRIs are maneuvering / interleaving; and it usually relies on manual threshold and rule settings, resulting in insufficient generalization. Tao Chen; Xiaoqi Guo; Jinxin Li. RadarSignal Sorting Method with Mimetic Image Mapping Based on Antenna ScanPattern via SOLOv2 Network. Remote Sensing, 2024, 16(24): 4639 (Articlenumber). DOI: 10.3390 / rs16244639. This paper proposes a sorting / online clustering approach based on density clustering (such as DBSCAN combined with heuristic optimization) to address noise points and aliasing issues. While this approach can alleviate some noise interference, density clustering is sensitive to hyperparameters such as parameter scale and neighborhood radius; it is still prone to "homogeneous multi-clustering / heterogeneous clustering" in high aliasing scenarios; and it often implicitly assumes fixed clustering, lacking adaptability to dynamic changes in the number of radiation sources.Zhijie Qu; Jinquan Zhang; Yuewei Zhou; Lina Ni. The Intelligent Evolution of Radar Signal Deinterleaving: A Systematic Review from Foundational Algorithms to Cognitive AI Frontiers. Sensors, 2026, 26(1): 248 (Article number). DOI: 10.3390 / s26010248. This paper summarizes the evolution of radar signal deinterleaving from traditional algorithms to intelligent methods, and points out representative works of various density clustering, template matching and two-step sorting. These methods mostly treat "sorting" as an independent module and lack integrated optimization with the back-end classification target; once cross-source misassignment occurs in the front-end sorting, the back-end classification error will be amplified by cascade, resulting in insufficient overall robustness.

[0005] In summary, existing technologies are insufficient to simultaneously meet the core requirements of "accurate pulse-level separation, robust clustering, and dynamic adaptive identification" in radiation source identification tasks under complex electromagnetic environments. Therefore, there is an urgent need in this field for a radiation source signal classification scheme capable of adaptive estimation of the number of radiation sources, reliable pulse-level separation of multiple sources, and further robust classification and online expansion under complex electromagnetic environments, in order to address the severe challenges posed by the current electromagnetic environment. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a radiation source signal classification method based on pulse-level multi-source separation. This method aims to solve the problems of insufficient accuracy, stability, and engineering applicability in radiation source identification under complex electromagnetic environments. This invention achieves accurate separation of multiple radiation sources at the same frequency without requiring a pre-defined number of radiation sources by introducing an improved density peak clustering mechanism with temporal consistency constraints. Furthermore, by combining multi-physics domain feature fusion and incremental learning mechanisms, it effectively improves the accuracy and stability of radiation source identification, maintains high robustness in classification and identification under strong aliasing / interference conditions, and supports online expansion.

[0007] To achieve the above objectives, the technical solution of the present invention includes the following steps:

[0008] (1) Receive broadband signals, detect and extract pulse arrival time TOA, pulse width PW and signal amplitude parameters;

[0009] (2) Based on the multi-element antenna structure, the direction of arrival (DOA) of the detected pulse is estimated; and a pulse descriptor word (PDW) is constructed for each pulse, including at least the TOA, PW and DOA parameters;

[0010] (3) Using a single pulse as the basic processing unit, a pulse-level multidimensional feature vector is constructed based on the pulse descriptor word (PDW), and the feature vector is normalized to form a pulse-level feature space.

[0011] (4) In the pulse-level feature space, the pulse samples are clustered using the density peak clustering method to obtain the local density and relative distance of each pulse; and based on the local density calculation results, a penalty mechanism based on pulse timing consistency is introduced to correct the local density.

[0012] (5) Obtain the density peak distribution based on the corrected local density and relative distance, adaptively determine the number of radiation sources based on the distribution, and output the pulse sequence corresponding to each radiation source;

[0013] (6) Extract multi-physical domain features for the pulse sequence corresponding to each radiation source, and characterize the radiation source characteristics through collaborative modeling to improve the ability to distinguish radiation sources with the same modulation scheme.

[0014] (7) Input the multi-physical domain features into a multi-branch CNN+ feature fusion classification network for fusion processing to achieve classification and identification of radiation source signals;

[0015] (8) When a new radiation source category is detected, the ability to identify the new category is expanded by incremental learning while keeping the parameters of the existing feature extraction layer unchanged.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] First, because this invention uses a single pulse as the smallest processing unit, rather than a traditional pulse sequence or signal segment, it constructs a pulse-level multidimensional feature vector containing time of arrival (TOA), direction of arrival (DOA), and pulse width (PW), and completes the separation processing of signals from multiple radiation sources at the same frequency within this feature space; by modeling the emission behavior characteristics of radiation sources at the pulse-level granularity, it is still possible to effectively distinguish pulses from different radiation sources in complex electromagnetic environments where multiple radiation sources are working simultaneously and signals are highly overlapping.

[0018] Secondly, this invention introduces a temporal consistency penalty mechanism in the density calculation stage of Density Peak Clustering (DPC): by dynamically weakening the local density of temporally discontinuous samples through the angular change of the direction of arrival (DOA) of adjacent pulses. When the DOA of adjacent pulses changes abruptly, the local density of that pulse is automatically weighted down, making it more difficult for it to be incorrectly merged into the same cluster. The direct effect of this design is to improve the purity of clusters, enhance the stability of multi-source separation, significantly reduce the probability of cross-source mis-clustering, and ensure that the output pulse sequences are more consistent in the same-frequency multi-source scenario, closer to the ideal state of "one source, one sequence". This provides more reliable basic data for subsequent radiation source classification and identification, directly improving the accuracy and stability of classification.

[0019] Third, based on the density peak distribution after introducing temporal consistency constraints, this invention achieves adaptive estimation of the number of radiation sources, and can automatically complete the source number determination and pulse sequence reconstruction without pre-setting the number of radiation sources. This mechanism avoids the limitation of traditional algorithms that rely on the preset number of sources, enhances the engineering adaptability to scenarios with dynamic changes in the number of radiation sources, and is especially suitable for practical scenarios with complex electromagnetic environments and real-time fluctuations in the number of radiation sources.

[0020] Fourth, this invention proposes an incremental learning mechanism based on small sample buffering and periodic triggering to update the classification model online, so as to adapt to the actual application scenario where the operating status of the radiation source changes dynamically over time. By limiting the number of samples required for a single update and setting a reasonable update cycle, the system's storage and computing resource consumption can be effectively controlled while continuously improving the recognition performance. This can meet the real-time requirements in engineering applications and is especially suitable for electronic reconnaissance systems that require long-term deployment and dynamic evolution. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall implementation of the method of the present invention;

[0022] Figure 2 This is a schematic diagram of the implementation architecture for multi-physical domain feature extraction and feature fusion recognition in this invention;

[0023] Figure 3 This is a flowchart of incremental learning / online access update in this invention. Detailed Implementation

[0024] The present invention will now be further described with reference to the accompanying drawings.

[0025] Example 1: Refer to Appendix Figure 1-3 This invention proposes a radiation source signal classification method based on pulse-level multi-source separation, which specifically includes the following steps:

[0026] Step 1) Signal Acquisition and Pulse Detection:

[0027] The system receives a broadband signal and detects and extracts parameters such as pulse arrival time (TOA), pulse width (PW), and signal amplitude. In this embodiment, the intermediate frequency (IF) or baseband IQ signal of the broadband signal is digitally channelized and divided into multiple sub-channels. Pulse detection is performed in parallel in each sub-channel to extract parameters such as pulse arrival time (TOA), pulse width (PW), and signal amplitude.

[0028] Step 2) Arrival direction estimation and pulse descriptor construction:

[0029] Based on a multi-element antenna array structure, the direction of arrival (DOA) of the detected radio frequency pulse signal is estimated, and high-precision algorithms such as MUSIC and ESPRIT can be used to achieve angle resolution. Subsequently, a pulse descriptor word (PDW) is constructed for each pulse, which includes at least the following core parameters:

[0030] TOA (Time of Arrival): Pulse arrival time, reflecting the timing information of the pulse;

[0031] PW (Pulse Width): Pulse width, reflecting the temporal duration characteristics of a pulse;

[0032] DOA (Direction of Arrival): The direction of pulse arrival, reflecting the spatial orientation information of the radiation source;

[0033] RF (Radio Frequency): Pulse carrier frequency, reflecting the operating frequency of the radiation source.

[0034] Step 3) Using a single pulse as the basic processing unit, construct a pulse-level multidimensional feature vector based on the pulse descriptor word (PDW), and normalize the feature vector to form a pulse-level feature space. This embodiment specifically obtains this feature space according to the following steps:

[0035] (3.1) Using a single pulse as the basic processing unit, a pulse-level multidimensional feature vector composed of TOA, DOA, and PW is constructed based on the pulse descriptor word PDW:

[0036] ,

[0037] in, , and These represent the arrival time, direction of arrival, and pulse width of the i-th pulse, respectively. Indicates the transpose operation;

[0038] (3.2) Normalize each component of the pulse-level multidimensional feature vector to map its numerical range to the [0,1] interval, so as to eliminate the influence of different dimensions on the clustering results and obtain the pulse-level feature space, which provides a unified computational basis for subsequent clustering algorithms.

[0039] Step 4) In the pulse-level feature space, the pulse samples are clustered using the density peak clustering method to obtain the local density and relative distance of each pulse; and based on the local density calculation results, a penalty mechanism based on pulse timing consistency is introduced to correct the local density.

[0040] The local density of the pulse mentioned above in this embodiment is calculated according to the following formula:

[0041] ,

[0042] in, This represents the local density of the i-th pulse. This represents the distance between the i-th pulse and the j-th pulse in the multidimensional feature space, where i is not equal to j; This is the cutoff distance parameter; This is a step function; it takes the value 1 when the value inside the parentheses is less than zero, and 0 otherwise.

[0043] The relative distance between the above pulses is obtained according to the following expression:

[0044] ,

[0045] ;

[0046] in, This represents the relative distance to the i-th pulse; This represents the relative distance between pulse samples with the highest local density.

[0047] Based on the local density calculation results, a penalty mechanism based on pulse timing consistency is introduced to correct the local density, as follows:

[0048] ,

[0049] in, The improved local density value is based on the original density. The result after performing timing consistency correction; It represents the change in the direction of arrival between the i-th pulse and its predecessor (the (i-1)-th pulse), reflecting the temporal continuity of the pulse sequence; This is a direction change control parameter (penalty factor) used to adjust the strength of timing consistency constraints; The smaller the value, the greater the penalty for changes in the DOA of adjacent pulses; the larger the value, the weaker the penalty. That is, when the arrival direction of adjacent pulses changes significantly, the local density of the corresponding pulses is significantly weakened, thereby reducing the probability of them being incorrectly clustered into the same radiation source and improving the accuracy of clustering.

[0050] In this embodiment, by introducing a dynamic penalty for the change in DOA, the pulse timing pattern of the real radiation source signal is effectively matched, making it particularly suitable for radiation source sorting scenarios in complex electromagnetic environments; the penalty factor The algorithm can be flexibly adjusted according to the complexity of the actual electromagnetic environment, balancing its versatility and scenario adaptability. Furthermore, since the correction process is embedded in the pre-clustering calculation, it does not add extra computational complexity to post-processing, ensuring the algorithm's real-time performance. This approach effectively suppresses cross-source mis-clustering problems caused by pulses from different radiation sources during the clustering process.

[0051] Step 5) Obtain the density peak distribution based on the corrected local density and relative distance, adaptively determine the number of radiation sources based on the distribution, and output the pulse sequence corresponding to each radiation source;

[0052] Step 6) Refer to the appendix Figure 2 For each radiation source, multi-physics domain features are extracted from the pulse sequence. Collaborative modeling is used to characterize the radiation source's properties, thereby improving the distinguishability between radiation sources with the same modulation scheme. In this embodiment, the aforementioned multi-physics domain features include at least one of the following: transient features based on the transient changes of the IQ signal, statistical features based on the pulse repetition interval (PRI) and its transition relationship, and RF fingerprint features based on the non-ideal characteristics of RF hardware. The transient features are extracted based on the transient variation patterns of the amplitude, phase, and frequency of the IQ signal, used to characterize the instantaneous modulation behavior of the radiation source signal. The statistical features are extracted based on the distribution characteristics of the pulse repetition interval (PRI) and the transition probability matrix parameters, used to characterize the operating rhythm of the radiation source and the temporal pattern of the pulse sequence. The RF fingerprint features are extracted based on the non-ideal characteristics of RF hardware, such as nonlinear distortion, phase noise, and frequency offset, used to characterize the inherent hardware properties of the radiation source. The aforementioned collaborative modeling characterization of radiation source properties involves correlating and fusing multi-physics domain features from three dimensions: transient behavior, operating rhythm, and hardware properties, to achieve a multi-dimensional and differentiated characterization of the radiation source's properties.

[0053] Step 7) Input the multi-physical domain features into a multi-branch CNN+ feature fusion classification network for fusion processing to achieve classification and identification of radiation source signals;

[0054] Step 8) Refer to Appendix Figure 3When a new radiation source category is detected, the new category recognition capability is expanded through incremental learning while keeping the existing feature extraction layer parameters unchanged. Specifically, in this embodiment, only the classification output layer is expanded, and knowledge distillation is used to update the multi-branch CNN+ feature fusion classification network to reduce the impact on the existing radiation source recognition performance.

[0055] The effects of the present invention will be further explained below with reference to simulation experiments.

[0056] 1. Simulation conditions:

[0057] The simulation experiments of this invention were conducted in a hardware environment of Intel Core i9-14900K and NVIDIA RTX 4090 (24GB) and a software environment of Ubuntu 22.04.

[0058] 2. Simulation content:

[0059] Self-built Synthetic PDW Dataset: Generates multi-source aliased PDW sequences according to the radiation source pulse emission model, including TOA (Time of Arrival), DOA (Direction of Arrival / Angle), and PW (Pulse Width) for each pulse. Experimental settings: Number of radiation sources: K = 3 (main experiment), and K = 5 (to verify the adaptive capability of the number of sources). The operating frequency bands of each source overlap (not dependent on frequency distinction). Radiation sources adopt different PRI modes: fixed PRI / jittery PRI / staggered PRI. The pulse loss rate is 10% (miss=0.1). The aliasing method is to sort the pulses of each source by TOA and then alias them into a sequence (simulating actual reception). SNR is set to 10dB, 5dB, and 0dB. When K=3: DOA center is in [-0.8°, 0°, 0.8°], and when K=5: DOA center is in [-1.2°, -0.6°, 0°, 0.6°, 1.2°].

[0060] The comparison methods are as follows: Baseline-1 (frame-level statistical clustering): The aliased sequences are subjected to statistical features (mean DOA, mean PW, mean Δ TOA) within a fixed window (200 pulses / frame), and then clustered → typical "frame-level / statistical" clustering; Baseline-2 (DBSCAN): Density clustering is performed using DOA and PW → typical engineering clustering; Baseline-3 (original DPC): DPC is performed using pulse-level features, but without temporal consistency penalty. This invention adds temporal adjacency information (such as Δ DOA, Δ TOA) to the pulse-level feature vectors and introduces a temporal consistency penalty factor in the local density stage of DPC, while simultaneously performing adaptive source number estimation.

[0061] 3. Simulation Result Analysis:

[0062] Table 1 shows the separation effect and source number estimation results under different SNRs with K=3 and a missing rate of 10%; Table 2 shows the validation results with K=5 (validating the "adaptive source number estimation" capability), SNR=5, and a missing rate of 10%.

[0063] Table 1

[0064]

[0065] Table 2

[0066]

[0067] Experiments show that, under multi-source aliasing simulation conditions with 3 and 5 radiation source types, a missing pulse rate of 10%, and SNR of 10 dB / 5 dB / 0 dB, the results compare frame-level statistical clustering, DBSCAN density clustering, and the original density peak clustering method. The experimental results are shown in Tables 1 and 2: This invention maintains a higher clustering consistency index (ARI / NMI) even with decreasing SNR and increasing source number, and achieves more accurate adaptive source number estimation in K=3 and K=5 scenarios. (closer to the true K), thereby verifying that the present invention, by introducing pulse-level temporal consistency constraints and source number adaptive estimation mechanism, can effectively reduce cross-source misclustering and improve the reliability of multi-source separation and identification in complex electromagnetic environments.

[0068] The classification model constructed in this invention demonstrates superior performance under different signal-to-noise ratios and interference scenarios, with its advantages being particularly prominent in low signal-to-noise ratio environments. Currently, the model has completed simulation verification, and the relevant code has been packaged into a directly executable program, possessing the foundation for engineering applications. Further improvements can be made through model lightweighting and adaptation to edge computing devices. Balancing classification accuracy and training efficiency, it can be widely applied in practical scenarios such as electromagnetic monitoring, spectrum sensing, and electronic countermeasures, showing broad application prospects.

[0069] The parts of this invention not described in detail are common knowledge to those skilled in the art.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Obviously, those skilled in the art, after understanding the content and principle of the present invention, may make various modifications and changes in form and detail without departing from the principle and structure of the present invention. However, these modifications and changes based on the concept of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A radiation source signal classification method based on pulse-level multi-source separation, characterized in that, Includes the following steps: (1) Receive broadband signals, detect and extract pulse arrival time TOA, pulse width PW and signal amplitude parameters; (2) Based on the multi-element antenna structure, the direction of arrival (DOA) of the detected pulse is estimated; and a pulse descriptor word (PDW) is constructed for each pulse, including at least the TOA, PW and DOA parameters; (3) Using a single pulse as the basic processing unit, a pulse-level multidimensional feature vector is constructed based on the pulse descriptor word (PDW), and the feature vector is normalized to form a pulse-level feature space. (4) In the pulse-level feature space, the pulse samples are clustered using the density peak clustering method to obtain the local density and relative distance of each pulse; and based on the local density calculation results, a penalty mechanism based on pulse timing consistency is introduced to correct the local density. (5) Obtain the density peak distribution based on the corrected local density and relative distance, adaptively determine the number of radiation sources based on the distribution, and output the pulse sequence corresponding to each radiation source; (6) Extract multi-physical domain features for the pulse sequence corresponding to each radiation source, and characterize the radiation source characteristics through collaborative modeling to improve the ability to distinguish radiation sources with the same modulation scheme. (7) Input the multi-physical domain features into a multi-branch CNN+ feature fusion classification network for fusion processing to achieve classification and identification of radiation source signals; (8) When a new radiation source category is detected, the ability to identify the new category is expanded by incremental learning while keeping the parameters of the existing feature extraction layer unchanged.

2. The method according to claim 1, characterized in that: The detection and extraction of pulse arrival time (TOA), pulse width (PW), and signal amplitude parameters in step (1) are specifically achieved by digitally channelizing the intermediate frequency or baseband IQ signal of the broadband signal into multiple sub-channels; and by performing pulse detection in parallel in each sub-channel to extract parameters such as pulse arrival time (TOA), pulse width (PW), and signal amplitude.

3. The method according to claim 1, characterized in that: The pulse-level feature space mentioned in step (3) is obtained specifically according to the following steps: (3.1) Using a single pulse as the basic processing unit, a pulse-level multidimensional feature vector composed of TOA, DOA, and PW is constructed based on the pulse descriptor word PDW: ， in, , and These represent the arrival time, direction of arrival, and pulse width of the i-th pulse, respectively. Indicates the transpose operation; (3.2) Normalize each component of the pulse-level multidimensional feature vector to map its numerical range to the [0,1] interval, so as to eliminate the influence of different dimensions on the clustering results and obtain the pulse-level feature space.

4. The method according to claim 1, characterized in that: The local density of the pulse mentioned in step (4) is calculated according to the following formula: ， in, This represents the local density of the i-th pulse. This represents the distance between the i-th pulse and the j-th pulse in the multidimensional feature space, where i is not equal to j; This is the cutoff distance parameter; This is a step function; it takes the value 1 when the value inside the parentheses is less than zero, and 0 otherwise.

5. The method according to claim 4, characterized in that: The relative distance of the pulses mentioned in step (4) is obtained according to the following expression: ， ； in, This represents the relative distance to the i-th pulse; This represents the relative distance between pulse samples with the highest local density.

6. The method according to claim 5, characterized in that: In step (4), based on the local density calculation results, a penalty mechanism based on pulse timing consistency is introduced to correct the local density, as follows: ， in, For the improved local density, This represents the change in the direction of arrival between the i-th pulse and its predecessor. These are the directional change control parameters used to adjust the strength of timing consistency constraints.

7. The method according to claim 1, characterized in that: The multi-physical domain features mentioned in step (6) include at least one of the following: transient features based on IQ signal transient changes, statistical features based on pulse repetition intervals and their transfer relationships, and radio frequency fingerprint features based on non-ideal characteristics of radio frequency hardware.

8. The method according to claim 7, characterized in that: The transient features mentioned in step (6) are extracted based on the transient change law of the amplitude, phase and frequency of the IQ signal, and are used to characterize the instantaneous modulation behavior of the radiation source signal; the statistical features are extracted based on the distribution characteristics of the pulse repetition interval PRI and the parameters of the transition probability matrix, and are used to characterize the working rhythm of the radiation source and the temporal law of the pulse sequence; the radio frequency fingerprint features are extracted based on the non-ideal characteristics of the radio frequency hardware such as nonlinear distortion, phase noise and frequency offset, and are used to characterize the inherent hardware properties of the radiation source.

9. The method according to claim 8, characterized in that: The step (6) describes the characteristics of the radiation source through collaborative modeling, which involves linking and fusing multi-physical domain features from three dimensions: transient behavior, operating rhythm, and hardware attributes, to achieve a multi-dimensional and differentiated characterization of the radiation source characteristics.

10. The method according to claim 1, characterized in that: The expansion of the new category recognition capability through incremental learning in step (8) is specifically achieved by expanding only the classification output layer and updating the multi-branch CNN+ feature fusion classification network using knowledge distillation, so as to reduce the impact on the existing radiation source recognition performance.

Images