Radiation source sorting identification method based on classification network decoupling and dynamic label library

Abstract

The invention provides a radiation source sorting identification method based on classification network decoupling and a dynamic label library, which can sort radar individuals from a low-dimensional motion information level by adopting carrier frequency splitting and trajectory feature clustering, and avoids the technical problem of high calculation complexity of a traditional method and a deep learning method. Meanwhile, individual labels are determined by adopting branch classification networks in multiple dimensions, so that the method has relatively high interpretability and can adapt to a scene with a high pulse loss rate, and the over-fitting problem and the'black box 'characteristic of an existing deep learning algorithm are avoided; furthermore, interference of a new system radar on other complex modulation information can be avoided through a trajectory feature clustering method, and meanwhile, due to the fact that each branch classification network can independently extract and recognize the modulation type, the carrier frequency type, the pulse repetition interval and other features, the expansion optimization and compatibility are improved, and the method is suitable for being popularized and applied. And finally, the overall generalization ability and identification ability of the identification method are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic countermeasure reconnaissance, and in particular to a radiation source sorting and identification method based on classification network decoupling and a dynamic label library. Background Art

[0002] In the field of electronic countermeasure reconnaissance, radar emitter identification is the core technology for achieving effective perception and threat assessment of complex electromagnetic environments. This technology needs to extract multi-dimensional features such as carrier frequency, pulse repetition interval (PRI), pulse width (PW), angle of arrival (DOA) and intra-pulse modulation from densely overlapping radar pulse streams to accurately determine the type of emitter and its behavior pattern, thereby providing key basis for tasks such as electromagnetic spectrum situation analysis and interference decision-making. However, as radar systems evolve towards multi-band and full-spectrum directions, the electronic reconnaissance environment presents complex characteristics of high density, strong dynamics, and multi-system interweaving, resulting in serious aliasing of the received pulse stream in multiple dimensions such as carrier frequency, pulse width, and arrival time. The signal characteristics of different types of emitters are slightly different, and pulse loss and modulation complexity further increase the difficulty of individual sorting, making traditional emitter identification methods face severe challenges.

[0003] At present, for the problem of radar radiation source sorting and identification, the existing technology mainly adopts the following three methods: First, the traditional algorithms based on parameter clustering and sequence analysis (such as SDIF, hidden Markov model HMM), which realize sorting by counting regular parameters such as pulse repetition interval, have certain effectiveness in low aliasing scenarios; Second, the end-to-end model based on deep learning (such as convolutional neural network CNN, long short-term memory network LSTM), which classifies by automatically learning pulse feature mapping, can process complex modulated signals; Third, the hybrid framework that integrates traditional algorithms and deep learning attempts to combine the advantages of rule analysis and data-driven to improve sorting accuracy.

[0004] However, the above existing methods still have significant limitations in practical applications: traditional sorting algorithms rely on preset rules, and the computational complexity grows exponentially in high-density aliasing scenarios, and cannot dynamically analyze changes in modulation parameters; although deep learning models can mine deep features, they are prone to overfitting due to data sparsity in scenarios with high pulse loss rates, and the "black box" feature makes it difficult to explain sorting logic or adapt to new radar systems; hybrid frameworks have weak cross-scenario generalization capabilities due to feature coupling between modules. In addition, existing technologies have not effectively solved common problems such as strong correlation between multi-dimensional features, inability to incrementally update static label libraries, and failure of aggregation of incomplete pulse trajectories, resulting in low sorting and recognition rates in complex electromagnetic environments, making it difficult to meet real-time and robustness requirements. Summary of the invention

[0005] To solve the above problems existing in the prior art, the present invention provides a method for sorting and identifying radiation sources based on classification network decoupling and a dynamic label library.

[0006] The technical problems to be solved by the present invention are realized through the following technical solutions:

[0007] In a first aspect, the present invention provides a method for sorting and identifying radiation sources based on classification network decoupling and a dynamic label library, including:

[0008] S101. Obtain the pulse data of the radar; the pulse data includes: pulse descriptor data and intra-pulse data;

[0009] S102. Successively perform individual sorting on the pulse data by using carrier frequency splitting and trajectory feature clustering to obtain multiple individual sorting results;

[0010] S103. Input the multiple individual sorting results into a pre-trained multi-branch classification network according to the corresponding dimensions for classification processing to generate individual labels corresponding to the multiple individual sorting results; the pre-trained multi-branch classification network includes: a pre-trained modulation type classification network, a pre-trained carrier frequency type classification network, and a pre-trained pulse repetition interval classification network;

[0011] S104. Use the pulse data and individual label corresponding to each individual sorting result to determine whether the current individual belongs to a pre-constructed radar model label library;

[0012] S105. When S104 is not established, based on a weighted scoring method, use the pulse data, individual label corresponding to each individual sorting result, and the pre-constructed radar model label library to construct a radar model update label library;

[0013] S106. Use the pulse data and individual label corresponding to each individual sorting result, and use the radar model update label library to perform radar model determination to obtain an individual model label.

[0014] Optionally, S102 includes:

[0015] Perform preprocessing on the pulse data to obtain preprocessed pulse data;

[0016] Successively perform dynamic window division, in-window density clustering, and cross-window label inheritance processing on the preprocessed pulse data to obtain multiple split pulse streams;

[0017] Extract trajectory features from the multiple split pulse streams to obtain multi-dimensional trajectory features; the multi-dimensional trajectory features include: arrival angle statistical features, arrival angle dynamic features, arrival angle extreme value features, and arrival angle segmented features;

[0018] Based on the K-means clustering method combined with the silhouette coefficient, multi-dimensional trajectory features are used for individual sorting to obtain multiple individual sorting results.

[0019] Optionally, preprocess the pulse data to obtain preprocessed pulse data, including:

[0020] Perform time-axis synchronization processing on the pulse descriptor data and in-pulse data to obtain synchronized pulse descriptor data and synchronized in-pulse data;

[0021] Perform deletion processing on the synchronized pulse descriptor data and synchronized in-pulse data to remove abnormal points to obtain preprocessed pulse data; the abnormal points include: abnormal arrival angle, abnormal carrier frequency, abnormal pulse width, abnormal pulse width, and abnormal amplitude.

[0022] Optionally, before S104, it also includes:

[0023] Extract the pulse width and frequency points from the pulse data corresponding to multiple individual sorting results to obtain corresponding pulse width values and frequency point information.

[0024] Optionally, S104 includes:

[0025] Use the pulse width value, frequency point information, and individual label to determine whether the current individual belongs to a pre-constructed radar model label library.

[0026] Optionally, S105 includes:

[0027] When S104 is not established, perform the same-type scoring determination on the frequency point information in the pulse data to obtain the same-type scoring result;

[0028] When the same-type scoring result belongs to the same type, merge the corresponding pulse data and individual labels to obtain updated pulse data and updated individual labels, and calculate the corresponding frequency point weights to obtain frequency point weight information;

[0029] Based on the weighted scoring method and frequency point weight information, perform cross-scenario cross-validation on the updated pulse data and updated individual labels to obtain supplementary radar model label information;

[0030] Add the supplementary radar model label information to the pre-constructed radar model label library to obtain a radar model updated label library.

[0031] Optionally, S105 also includes:

[0032] When S104 is established, use the pulse width value, frequency point information, and individual label to perform radar model determination based on the pre-constructed radar model label library to obtain an individual model label.

[0033] In a second aspect, the present invention provides a radiation source sorting and identification device based on classification network decoupling and a dynamic label library. The radiation source sorting and identification device based on classification network decoupling and a dynamic label library includes: an acquisition unit, an individual sorting unit, a classification processing unit, a judgment unit, a construction unit, and a model determination unit;

[0034] The acquisition unit is configured to: acquire the pulse data of a radar; the pulse data includes: pulse descriptor data and intra-pulse data;

[0035] The individual sorting unit is configured to: sequentially perform individual sorting on the pulse data by using carrier frequency splitting and trajectory feature clustering to obtain multiple individual sorting results;

[0036] The classification processing unit is configured to: input the multiple individual sorting results into a pre-trained multi-branch classification network according to the corresponding dimensions for classification processing to generate individual labels corresponding to the multiple individual sorting results; the pre-trained multi-branch classification network includes: a pre-trained modulation type classification network, a pre-trained carrier frequency type classification network, and a pre-trained pulse repetition interval classification network;

[0037] The judgment unit is configured to: use the pulse data and the individual label corresponding to each individual sorting result to judge whether the current individual belongs to a pre-constructed radar model label library;

[0038] The construction unit is configured to: when the condition of the judgment unit is not satisfied, use the pulse data, the individual label corresponding to each individual sorting result, and the pre-constructed radar model label library to construct a radar model update label library;

[0039] The model determination unit is configured to: use the pulse data and the individual label corresponding to each individual sorting result, and use the radar model update label library to perform radar model determination to obtain an individual model label.

[0040] In a third aspect, the present invention provides a radiation source sorting and identification device based on classification network decoupling and a dynamic label library, including: a processor, a storage medium, and a bus. The storage medium stores machine-readable instructions executable by the processor. When the radiation source sorting and identification device based on classification network decoupling and a dynamic label library runs, the processor communicates with the storage medium through the bus, and the processor executes the machine-readable instructions to perform the steps of the radiation source sorting and identification method based on classification network decoupling and a dynamic label library as described in the first aspect above.

[0041] The present invention provides a method for sorting and identifying radiation sources based on classification network decoupling and a dynamic label library, including: S101, obtaining pulse data of a radar; the pulse data includes: pulse descriptor data and intra-pulse data; S102, sequentially performing individual sorting on the pulse data by using carrier frequency splitting and trajectory feature clustering to obtain multiple individual sorting results; S103, inputting the multiple individual sorting results into a pre-trained multi-branch classification network according to corresponding dimensions for classification processing to generate individual labels corresponding to the multiple individual sorting results; the pre-trained multi-branch classification network includes: a pre-trained modulation type classification network, a pre-trained carrier frequency type classification network, and a pre-trained pulse repetition interval classification network; S104, using the pulse data and individual labels corresponding to each individual sorting result to determine whether the current individual belongs to a pre-constructed radar model label library; S105, when S104 does not hold, based on a weighted scoring method, using the pulse data, individual labels corresponding to each individual sorting result, and the pre-constructed radar model label library to construct a radar model update label library; S106, using the pulse data and individual labels corresponding to each individual sorting result, and using the radar model update label library to perform radar model determination to obtain individual model labels. In the present invention, by using carrier frequency splitting and trajectory feature clustering, the radar individuals can be sorted from the low-dimensional motion information level, avoiding the technical problems of high computational complexity of traditional methods and deep learning methods; at the same time, by using the classification networks of each branch in multiple dimensions to determine individual labels, it not only has strong interpretability but also can adapt to the scenario with a high pulse loss rate, avoiding the overfitting problem and "black box" characteristics of existing deep learning algorithms; further, the method of trajectory feature clustering can avoid the interference of new radar systems on other complex modulation information, and at the same time, since the classification networks of each branch can independently extract and identify features such as modulation type, carrier frequency type, and pulse repetition interval, the expandability and compatibility are improved, and the construction strategy of the dynamic label library supports the update of new radar model information, improving the adaptability and dynamic compatibility with new radar systems; and the independent multi-branch classification network completely decouples features such as carrier frequency, modulation type, and pulse repetition interval at the physical level, eliminating cascaded error propagation and avoiding the problem of weak cross-scene generalization ability caused by feature coupling between modules in a hybrid framework, improving the overall generalization ability and identification ability of the identification method.

[0042] The following will further elaborate on the present invention in detail with reference to the drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic flowchart of a method for sorting and identifying radiation sources based on classification network decoupling and a dynamic label library provided by an embodiment of the present invention;

[0044] Figure 2A schematic diagram showing the radar models and quantities in a certain scenario is exemplarily shown;

[0045] Figure 3 A schematic diagram showing the original data and sorting results in a certain scenario is exemplarily shown;

[0046] Figure 4 A schematic diagram showing the original data and sorting results in a certain scenario is exemplarily shown;

[0047] Figure 5 A schematic diagram showing the sorting results in a certain scenario is exemplarily shown;

[0048] Figure 6 A schematic structural diagram of a radiation source sorting and recognition device based on classification network decoupling and a dynamic label library provided by an embodiment of the present invention;

[0049] Figure 7 A schematic structural diagram of a radiation source sorting and recognition device based on classification network decoupling and a dynamic label library provided by an embodiment of the present invention. Detailed implementation manners

[0050] In view of the problems of strong mixing of data in various dimensions, many radar radiation source models, serious data loss, and complex modulation in the existing radar radiation source sorting and recognition technology, the present invention proposes a radar radiation source sorting and recognition framework based on classification network decoupling and a multi-label model library. This framework combines classical algorithms with deep learning models, flexibly utilizes detection information such as carrier frequency, arrival angle, pulse width, pulse amplitude, modulation type, etc., decouples the recognition network, generates a multi-label radiation source model library, and realizes the sorting and recognition of radar radiation source signals in a complex electromagnetic environment.

[0051] The following further describes the present invention in detail with specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0052] In order to improve the overall generalization ability and recognition ability of the radiation source sorting and recognition method, an embodiment of the present invention provides a radiation source sorting and recognition method based on classification network decoupling and a dynamic label library. Figure 1 A schematic flowchart of a radiation source sorting and recognition method based on classification network decoupling and a dynamic label library provided by an embodiment of the present invention. As Figure 1 shown, it includes:

[0053] S101. Obtain the pulse data of the radar.

[0054] Among them, the pulse data includes: pulse descriptor data and in-pulse data. The pulse data can include the pulse data of one scenario or the pulse data of multiple scenarios. For a scenario, the model types and numbers of radars included in different scenarios are different.

[0055] S102. Sequentially perform carrier frequency splitting and trajectory feature clustering on the pulse data to perform individual sorting, obtaining multiple individual sorting results.

[0056] Optionally, S102 includes:

[0057] Preprocess the pulse data to obtain preprocessed pulse data;

[0058] Perform dynamic window partitioning, in-window density clustering, and cross-window label inheritance processing on the preprocessed pulse data in sequence to obtain multiple split pulse streams;

[0059] Extract trajectory features from the multiple split pulse streams to obtain multi-dimensional trajectory features; the multi-dimensional trajectory features include: arrival angle statistical features, arrival angle dynamic features, arrival angle extreme value features, and arrival angle segmented features;

[0060] Based on the K-means mean clustering method combined with the silhouette coefficient, perform individual sorting using the multi-dimensional trajectory features to obtain multiple individual sorting results.

[0061] The dynamic window partitioning mechanism includes: the preprocessed pulse data can be cut into several windows of different sizes, and the signal rhythm change is detected through the pulse arrival time difference during this period. Specifically, calculate the time interval between adjacent pulses. When the interval exceeds the preset threshold (usually taking 95% of the historical maximum interval), it is determined as the radar scan cycle switching point, and this is used as the window segmentation basis. This adaptive partitioning maintains the natural structure of the pulse group, and the window length dynamically changes between 50 - 200 pulses, which can not only capture the signal features but also avoid feature breaks caused by forced segmentation.

[0062] In-window density clustering means that within each time window, use the DBSCAN algorithm to perform clustering analysis on the carrier frequency dimension. Set a neighborhood radius of 200 kHz, allowing individual pulses to form clusters independently to adapt to the sparse signal environment. The DBSCAN algorithm can automatically identify pulse groups with density connection, generate temporary cluster labels, use the geometric center of each cluster as the characteristic frequency, and mark isolated points as noise at the same time.

[0063] Cross-window label inheritance is to implement cross-window label inheritance using the dynamic label library mechanism. By maintaining the global alphabet label library (such as 'A' and 'B') and its latest center frequency value, perform frequency similarity determination on the cluster centers generated by the new window clustering: when the frequency difference between the new cluster center and the historical label center is less than 400 kHz (normalized threshold 0.0002), inherit the original label and update the center value, otherwise assign a new label in alphabetical order. At the same time, append and store the pulse data of each cluster according to the label classification to the corresponding data file, and ensure data traceability through the association of the original line numbers. Noise points are processed independently and do not participate in label inheritance, and finally multiple split pulse streams are obtained.

[0064] Further, a modulation type classification network (2DCNN) is used to extract trajectory features from multiple split pulse streams, obtaining multi-dimensional trajectory features. Further, by comprehensively using the K-means mean clustering method and arrival angle statistical features (mean, standard deviation, minimum, maximum), arrival angle dynamic features (angular velocity and angular acceleration), arrival angle extreme value features (peaks and valleys processed by Gaussian filtering and their position information), and arrival angle segmented features (segmented mean and standard deviation of the arrival angle), the pulse streams with similar trajectories are efficiently aggregated to form a set of "individual pulse streams", that is, the final multiple individual sorting results.

[0065] Specifically, the K-means mean clustering method can be expressed as:

[0066]

[0067] is the individual sorting result, representing the clustering set of each radar radiation source individual (which clusters multiple split pulse streams into multiple radar radiation source individuals). C k : represents the k-th clustering cluster, corresponding to the pulse set of a radar radiation source individual. μ k : the cluster center of the k-th clustering cluster, characterizing the trajectory feature center of this radar individual, K: represents the optimal number of clusters, which is jointly determined by silhouette coefficient analysis and manual adjustment.

[0068]

[0069] F(·): represents the feature extraction function, which converts the original pulse into a trajectory feature vector available for clustering, X i : represents the i-th pulse stream, F(x i ): represents the basic trajectory feature parameters extracted from the i-th pulse stream (pulse original data), specifically including the following four parts of features:

[0070] Among them, the arrival angle statistical features μ θ , σ θ , min(θ), max(θ) respectively represent the mean of the arrival angle (DOA), the standard deviation of the arrival angle, the minimum of the arrival angle, and the maximum of the arrival angle.

[0071] Arrival angle dynamic features respectively represent the angular velocity (first-order difference) of the arrival angle, the angular acceleration (second-order difference) of the arrival angle, the standard deviation of the angular velocity of the arrival angle, and the standard deviation of the angular acceleration of the arrival angle.

[0072] Arrival angle extreme value features N p , N v , p p , p vrespectively represent the number of peaks of the angle of arrival after Gaussian filtering, the number of valleys of the angle of arrival after Gaussian filtering, the normalized position vectors of each peak of the angle of arrival after Gaussian filtering, and the normalized position vectors of each valley of the angle of arrival after Gaussian filtering.

[0073] Angle of arrival segmentation feature respectively represent the mean and standard deviation of the angle of arrival corresponding to each segment after the trajectory is divided into 10 segments (or other segmentation values) by the segmentation feature, j represents the j-th segmentation feature, and T represents the transpose process.

[0074] Optionally, preprocess the pulse data to obtain preprocessed pulse data, including:

[0075] Perform time-axis synchronization processing on the pulse descriptor data and the in-pulse data to obtain synchronized pulse descriptor data and synchronized in-pulse data;

[0076] Perform abnormal point deletion processing on the synchronized pulse descriptor data and the synchronized in-pulse data to obtain preprocessed pulse data; abnormal points include: abnormal angle of arrival, abnormal carrier frequency, abnormal pulse width, abnormal pulse width, and abnormal amplitude.

[0077] S103. Input the multiple individual sorting results into a pre-trained multi-branch classification network according to the corresponding dimensions for classification processing to generate individual labels corresponding to the multiple individual sorting results.

[0078] The pre-trained multi-branch classification network includes: a pre-trained modulation type classification network, a pre-trained carrier frequency type classification network, and a pre-trained pulse repetition interval classification network.

[0079] In the embodiment of the present invention, first split the multiple individual sorting results to obtain a modulation type signal, a carrier frequency type signal, and a pulse repetition interval signal, and then sequentially input them into the pre-trained modulation type classification network, the pre-trained carrier frequency type classification network, and the pre-trained pulse repetition interval classification network according to the corresponding signal dimensions to obtain individual labels in multiple dimensions. The individual labels may specifically include: modulation type labels, carrier frequency type labels, and pulse repetition interval labels.

[0080] It can be understood that the decoupled network architecture supports independent optimization and expansion, can significantly improve the classification efficiency and algorithm adaptability, and can meet complex sorting requirements.

[0081] Optionally, before S104, it further includes:

[0082] Extract the pulse width and frequency points from the pulse data corresponding to the multiple individual sorting results to obtain the corresponding pulse width values and frequency point information.

[0083] S104. Use the pulse data and individual label corresponding to the sorting result of each individual to determine whether the current individual belongs to a pre-constructed radar model label library.

[0084] Optionally, S104 includes:

[0085] Use the pulse width value, frequency point information, and individual label to determine whether the current individual belongs to a pre-constructed radar model label library.

[0086] S105. When S104 is not established, based on the weighted scoring method, use the pulse data, individual label, and pre-constructed radar model label library corresponding to the sorting result of each individual to construct a radar model update label library.

[0087] Optionally, S105 includes:

[0088] When S104 is not established, perform the same-type scoring determination on the frequency point information in the pulse data to obtain the same-type scoring result;

[0089] When the same-type scoring results belong to the same type, merge the corresponding pulse data and individual label to obtain updated pulse data and updated individual label, and calculate the corresponding frequency point weight to obtain the frequency point weight information;

[0090] Based on the weighted scoring method and frequency point weight information, perform cross-scenario cross-validation on the updated pulse data and updated individual label to obtain supplementary radar model label information;

[0091] Add the supplementary radar model label information to the pre-constructed radar model label library to obtain a radar model update label library.

[0092] In the embodiments of the present invention, based on the individual label, combined with the distribution of different models of radars in each scenario, a model label library is constructed through cross-processing (the specific function is to determine which model of radar each radar individual is, while assigning a model label to the radar individual, and constructing a radar model update label library). The above processing method not only integrates rich domain knowledge and historical data, but also can be dynamically updated in a timely manner according to new radar models, ensuring that the coverage range and accuracy of the label library are continuously optimized. This dual strategy improves the adaptability and reliability of the signal library, providing a solid guarantee for radar emitter recognition in complex environments.

[0093] In addition, by combining expert prior model information and the dynamic supplement strategy of the model library, the radar model library can be effectively constructed and improved. The expert prior model information provides rich domain knowledge and historical data, while the dynamic supplement of the model library can dynamically update the model library in a timely manner according to actual application requirements and new radar models, ensuring continuous improvement of its coverage and accuracy. This dual strategy not only optimizes the construction process of the radar signal library but also improves its adaptability and reliability in complex environments.

[0094] The specific process is as follows:

[0095] (1) Frequency point matching score in the same scene

[0096] Objective: To determine whether two sets of frequency points belong to the same radar model.

[0097] Frequency point matching is adopted: that is, for each frequency point f, if there exists a frequency point g in another set such that |f - g| ≤ 0.01, it is determined to be a match.

[0098] (2) Merging data of the same model in the same scene

[0099] Objective: To merge the attributes of two radar individuals in the same scene and generate a unified model description.

[0100] Merging rules:

[0101] Ordinary fields (such as carrier frequency cf, pulse repetition interval pri, modulation type mod): If the values are the same, they are retained; otherwise, they are saved as a list.

[0102] Frequency point ratio: Take the union of the two sets of frequency points, and for each frequency point f, calculate the combined frequency and weight.

[0103] (3) Cross-validation

[0104] Objective: To cross-validate the radar model consistency across scenes and generate a model database.

[0105] Process: Group by scene and model: Extract the occurrence records of each model in different scenes from the preset scenes; Multi-scene matching: Call the frequency point matching algorithm to perform frequency point matching on the same model individuals across scenes; Merge valid data: If the matching is successful, write the data into the multi-dimensional signal tag library through data merging; Output: A database file containing model ID, scene ID, and merged parameters.

[0106] Optionally, S105 further includes:

[0107] When S104 holds, based on the pre-constructed radar model tag library, use the pulse width value, frequency point information, and individual label to determine the radar model and obtain the individual model label.

[0108] Specifically, the determination rule can be to calculate the distance between the pulse width value, frequency point information, and individual label and the corresponding pulse width comparison value, frequency point information comparison value, and individual label comparison value in the pre-constructed radar model label library to determine the final individual model label.

[0109] S106. Use the pulse data and individual label corresponding to each individual sorting result, and adopt the radar model update label library to determine the radar model and obtain the individual model label.

[0110] The present invention provides a radiation source sorting and recognition method based on classification network decoupling and dynamic label library, including: S101. Obtain the pulse data of the radar; the pulse data includes pulse descriptor data and in-pulse data; S102. Successively perform individual sorting on the pulse data by using carrier frequency splitting and trajectory feature clustering to obtain multiple individual sorting results; S103. Input the multiple individual sorting results into a pre-trained multi-branch classification network according to the corresponding dimensions for classification processing to generate individual labels corresponding to the multiple individual sorting results; the pre-trained multi-branch classification network includes a pre-trained modulation type classification network, a pre-trained carrier frequency type classification network, and a pre-trained pulse repetition interval classification network; S104. Use the pulse data and individual label corresponding to each individual sorting result to determine whether the current individual belongs to the pre-constructed radar model label library; S105. When S104 does not hold, based on the weighted scoring method, use the pulse data, individual label corresponding to each individual sorting result, and the pre-constructed radar model label library to construct a radar model update label library; S106. Use the pulse data and individual label corresponding to each individual sorting result, and adopt the radar model update label library to determine the radar model and obtain the individual model label. In the present invention, by using carrier frequency splitting and trajectory feature clustering, the radar individuals can be sorted from the low-dimensional motion information level, avoiding the technical problems of high computational complexity of traditional methods and deep learning methods; at the same time, by using the classification network of each branch in multiple dimensions to determine the individual label, it not only has strong interpretability but also can adapt to the scenario with a high pulse loss rate, avoiding the overfitting problem and "black box" characteristics of the existing deep learning algorithms; furthermore, the method of trajectory feature clustering can avoid the interference of new radar systems in other complex modulation information, and at the same time, since the classification network of each branch can independently extract and identify features such as modulation type, carrier frequency type, and pulse repetition interval, the expandability and compatibility are improved. The construction strategy of the dynamic label library supports the update of new radar model information, improving the adaptability and dynamic compatibility with new radar systems; and the independent multi-branch classification network completely decouples features such as carrier frequency, modulation type, and pulse repetition interval at the physical level, eliminating the cascade error propagation and avoiding the problem of weak cross-scene generalization ability caused by feature coupling between modules in the hybrid framework, improving the overall generalization ability and recognition ability of the recognition method.

[0111] To verify the effectiveness of the method of the present invention, simulation experiments were also carried out as follows:

[0112] Figure 2 Exemplarily, a schematic diagram showing the radar models and quantities in a certain scenario is presented. In Figure 2 the shown scenario, there are 3 radars of model 49, 2 radars of model 31, 2 radars of model 8, and 3 radars of model 21. As Figure 2 shown, a three-dimensional visualization result diagram of the original pulse data is given from three dimensions of arrival time, arrival angle, and carrier frequency. It can be seen that the original data overlaps in multiple dimensions.

[0113] Figure 3 Exemplarily, a schematic diagram of the original data and sorting results in a certain scenario is presented, where Figure 3 Figure (a) shows the schematic diagram of the original data, Figure 3 and Figure (b) shows the schematic diagram of the sorting results.

[0114] Figure 4 Exemplarily, a schematic diagram of the original data and sorting results in a certain scenario is presented, where Figure 4 Figure (a) shows the schematic diagram of the original data, Figure 4 and Figure (b) shows the schematic diagram of the sorting results.

[0115] Figure 5 Exemplarily, a schematic diagram of the sorting results in a certain scenario is presented. It can be seen that the algorithm can still obtain accurate sorting results even in the case of serious pulse loss.

[0116] As Figures 3 - 5 shown in the verification results of the sorting and recognition performance of pulse data under multi-scenario conditions, it can be seen that through the comprehensive analysis of features such as pulse arrival time (TOA), carrier frequency (CF), and arrival angle (DOA), points of different colors represent different correctly sorted radar individuals. The verification results show that in the multi-dimensional feature space, the method of the present invention can effectively separate different radar individuals more accurately, achieving clear clustering and recognition of individuals. This indicates that the adopted sorting and recognition method has high adaptability and accuracy in complex scenarios, providing reliable support for the accurate classification and model recognition of radar signals. Through the above result display, it is verified that the method of the present invention can achieve efficient and accurate model recognition in a complex and changeable radar signal environment, and has good practical application value.

[0117] The method provided by the embodiments of the present invention can be applied to electronic devices. Specifically, the electronic device can be: a desktop computer, a portable computer, a smart mobile terminal, a server, etc., which is not limited in the embodiments of the present invention.

[0118] Based on the same inventive concept, an embodiment of the present invention further provides a radiation source sorting and identification device based on classification network decoupling and a dynamic label library. Figure 6 FIG. is a schematic structural diagram of a radiation source sorting and identification device based on classification network decoupling and a dynamic label library provided by an embodiment of the present invention. As Figure 6 shown, it includes: an acquisition unit 601, an individual sorting unit 602, a classification processing unit 603, a judgment unit 604, a construction unit 605, and a model determination unit 606;

[0119] The acquisition unit 601 is configured to: acquire pulse data of a radar; the pulse data includes: pulse descriptor data and in-pulse data;

[0120] The individual sorting unit 602 is configured to: sequentially perform individual sorting on the pulse data by using carrier frequency splitting and trajectory feature clustering to obtain a plurality of individual sorting results;

[0121] The classification processing unit 603 is configured to: input the plurality of individual sorting results into a pre-trained multi-branch classification network according to corresponding dimensions for classification processing to generate individual labels corresponding to the plurality of individual sorting results; the pre-trained multi-branch classification network includes: a pre-trained modulation type classification network, a pre-trained carrier frequency type classification network, and a pre-trained pulse repetition interval classification network;

[0122] The judgment unit 604 is configured to: use the pulse data and individual labels corresponding to each individual sorting result to judge whether the current individual belongs to a pre-constructed radar model label library;

[0123] The construction unit 605 is configured to: when the condition of the judgment unit is not satisfied, use the pulse data, individual labels corresponding to each individual sorting result, and the pre-constructed radar model label library to construct a radar model update label library;

[0124] The model determination unit 606 is configured to: use the pulse data and individual labels corresponding to each individual sorting result, and perform radar model determination by using the radar model update label library to obtain an individual model label.

[0125] Figure 7 FIG. is a schematic structural diagram of a radiation source sorting and identification device based on classification network decoupling and a dynamic label library provided by an embodiment of the present invention, including: a processor 710, a storage medium 720, and a bus 730. The storage medium 720 stores machine-readable instructions executable by the processor 710. When the radiation source sorting and identification device based on classification network decoupling and a dynamic label library runs, the processor 710 communicates with the storage medium 720 through the bus 730, and the processor 710 executes the machine-readable instructions to execute the steps of the above method embodiment. The specific implementation manners and technical effects are similar and will not be elaborated here.

[0126] The storage medium may include a Random Access Memory (RAM), or may also include a Non-Volatile Memory (NVM), such as at least one disk memory. Optionally, the storage medium may also be at least one storage device located away from the aforementioned processor.

[0127] The aforementioned processor may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0128] It should be noted that the terms "first", "second", etc. are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present invention.

[0129] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0130] Although the present invention has been described in connection with various embodiments, those skilled in the art can understand and implement other variations of the above-described disclosed embodiments by referring to the accompanying drawings and the disclosure during the implementation of the claimed invention. In the description of the present invention, the term "comprising" does not exclude other components or steps, the indefinite article "a" or "an" does not exclude a plurality, and the meaning of "plurality" is two or more, unless otherwise specifically defined. In addition, certain measures are described in different embodiments, but this does not mean that these measures cannot be combined to produce good results.

[0131] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for sorting and identifying radiation sources based on the decoupling of a classification network and a dynamic label library, characterized in that, Including: S101. Obtain the pulse data of the radar; The pulse data includes: pulse descriptor data and in-pulse data; S102. Successively perform carrier frequency splitting and trajectory feature clustering on the pulse data for individual sorting to obtain multiple individual sorting results; S103. Input the multiple individual sorting results into a pre-trained multi-branch classification network according to the corresponding dimensions for classification processing to generate individual labels corresponding to the multiple individual sorting results; the pre-trained multi-branch classification network includes: a pre-trained modulation type classification network, a pre-trained carrier frequency type classification network, and a pre-trained pulse repetition interval classification network; S104. Use the pulse data and the individual label corresponding to each individual sorting result to determine whether the current individual belongs to a pre-constructed radar model label library; S105. When S104 is not established, based on a weighted scoring method, use the pulse data, the individual label, and the pre-constructed radar model label library corresponding to each individual sorting result to construct a radar model update label library; S106. Use the pulse data and the individual label corresponding to each individual sorting result, and use the radar model update label library to perform radar model determination to obtain an individual model label.

2. The method for sorting and identifying radiation sources based on classification network decoupling and dynamic label library according to claim 1, wherein S102 includes: Preprocess the pulse data to obtain preprocessed pulse data; Successively perform dynamic window division, in-window density clustering, and cross-window label inheritance processing on the preprocessed pulse data to obtain multiple split pulse streams; Extract trajectory features from the multiple split pulse streams to obtain multi-dimensional trajectory features; the multi-dimensional trajectory features include: arrival angle statistical features, arrival angle dynamic features, arrival angle extreme value features, and arrival angle segment features; Based on the K-means mean clustering method combined with the silhouette coefficient, use the multi-dimensional trajectory features for individual sorting to obtain the multiple individual sorting results.

3. The method for sorting and identifying radiation sources based on classification network decoupling and dynamic label library according to claim 2, wherein The preprocessing of the pulse data to obtain preprocessed pulse data includes: Perform time-axis synchronization processing on the pulse descriptor data and the in-pulse data to obtain synchronized pulse descriptor data and synchronized in-pulse data; Perform deletion processing of abnormal points on the synchronized pulse descriptor data and the synchronized in-pulse data to obtain the preprocessed pulse data; the abnormal points include: arrival angle abnormality, carrier frequency abnormality, pulse width abnormality, pulse width abnormality, and amplitude abnormality.

4. The method for sorting and identifying radiation sources based on classification network decoupling and dynamic tag library according to claim 1, characterized in that Before S104, it also includes: Extract the pulse width and frequency points of the pulse data corresponding to the multiple individual sorting results to obtain the corresponding pulse width values and frequency point information.

5. The radiation source sorting and recognition method based on classification network decoupling and dynamic label library according to claim 4, wherein S104 includes: Use the pulse width value, the frequency point information, and the individual label to determine whether the current individual belongs to a pre-constructed radar model label library.

6. The method for sorting and identifying radiation sources based on classification network decoupling and dynamic label library according to claim 4, wherein S105 includes: When S104 is not established, perform same-type scoring determination on the frequency point information in the pulse data to obtain a same-type scoring result; When the same-type scoring result belongs to the same type, merge the corresponding pulse data and individual label to obtain updated pulse data and updated individual label, and calculate the corresponding frequency point weight to obtain frequency point weight information; Based on the weighted scoring method and the frequency point weight information, perform cross-scenario cross-validation on the updated pulse data and the updated individual labels to obtain supplementary radar model label information; Add the supplementary radar model label information to the pre-constructed radar model label library to obtain the updated radar model label library.

7. The radiation source sorting and identification method based on classification network decoupling and dynamic label library according to claim 5, wherein S105 also includes: When S104 holds, use the pulse width value, the frequency point information, and the individual label to perform radar model determination based on the pre-constructed radar model label library to obtain individual model labels.

8. A radiation source sorting and recognition device based on classification network decoupling and dynamic label library, characterized in that, The radiation source sorting and identification device based on classification network decoupling and dynamic label library includes: an acquisition unit, an individual sorting unit, a classification processing unit, a judgment unit, a construction unit, and a model determination unit; The acquisition unit is used to: acquire the pulse data of the radar; the pulse data includes: pulse descriptor data and in-pulse data; The individual sorting unit is used to: sequentially perform individual sorting on the pulse data by using carrier frequency splitting and trajectory feature clustering to obtain multiple individual sorting results; The classification processing unit is used to: input the multiple individual sorting results into a pre-trained multi-branch classification network according to the corresponding dimensions for classification processing to generate individual labels corresponding to the multiple individual sorting results; the pre-trained multi-branch classification network includes: a pre-trained modulation type classification network, a pre-trained carrier frequency type classification network, and a pre-trained pulse repetition interval classification network; The judgment unit is used to: use the pulse data and the individual label corresponding to each individual sorting result to judge whether the current individual belongs to the pre-constructed radar model label library; The construction unit is used to: when the condition of the judgment unit is not satisfied, use the pulse data, the individual label, and the pre-constructed radar model label library corresponding to each individual sorting result to construct an updated radar model label library; The model determination unit is used to: use the pulse data and the individual label corresponding to each individual sorting result, and use the updated radar model label library to perform radar model determination to obtain individual model labels.

9. A radiation source sorting and identification device based on classification network decoupling and dynamic label library, characterized in that, Includes: A processor, a storage medium, and a bus. The storage medium stores machine-readable instructions executable by the processor. When the radiation source sorting and identification device based on classification network decoupling and dynamic label library runs, the processor communicates with the storage medium through the bus, and the processor executes the machine-readable instructions to perform the steps of the radiation source sorting and identification method according to any one of claims 1-7.

Images