Radar operating mode recognition method and system based on graph neural network

Abstract

The application discloses a radar working mode recognition method and system based on a graph neural network, relates to the technical field of radar signal processing, and comprises the following steps: performing pulse arrival time detection on received radar signals to construct a pulse parameter set; extracting a multi-dimensional pulse feature vector; constructing a pulse parameter similarity graph; inputting the pulse parameter similarity graph into a multi-relation graph neural network to perform graph feature aggregation and output high-dimensional embedding representation of each pulse node; dividing the high-dimensional embedding representation into N time subblocks based on a pulse time sequence and performing weighted aggregation of graph embedding pooling results; and performing confidence recognition based on global pulse sequence embedding and outputting a working mode recognition result. The application solves the technical problems of the prior art, such as difficulty in effectively modeling pulse interrelation, insufficient anti-interference capability, insufficient multi-mode radar signal classification precision, and poor working mode recognition accuracy, and achieves the technical effect of improving multi-mode radar signal classification precision and working mode recognition accuracy.

Description

Technical Field

[0001] The present application relates to the technical field related to radar signal processing, and specifically to a radar operating mode recognition method and system based on graph neural network. Background Art

[0002] As the electromagnetic environment becomes more complex, the operating modes of radar systems are becoming increasingly complex, and the radar signal waveforms are becoming more and more complex. Radar signals have the characteristics of high density, high complexity and strong dynamic changes. However, traditional feature recognition is difficult to meet actual needs. Accurate and efficient identification of radar operating modes is crucial for situational awareness decision-making. Existing radar operating mode recognition relies on statistical analysis of pulse descriptors, such as pulse repetition interval (PRI), carrier frequency (RF), pulse width (PW) and other parameters for classification, ignoring the correlation between pulses and making it difficult to effectively characterize the timing dependence and complex modulation characteristics of radar signals. In addition, in high-density signal environments, pulse loss, noise interference and parameter overlap may further reduce recognition performance, resulting in poor recognition accuracy and adaptability of radar operating modes in complex electromagnetic environments.

[0003] Therefore, in the current related technologies, there are technical problems such as difficulty in effectively modeling the correlation between pulses and insufficient anti-interference capabilities, which lead to insufficient accuracy in multi-mode radar signal classification and poor accuracy in working mode recognition. Summary of the Invention

[0004] This application provides a radar working mode recognition method and system based on graph neural network, which solves the technical problems in the existing technology such as difficulty in effectively modeling the correlation between pulses and insufficient anti-interference ability, resulting in insufficient multi-mode radar signal classification accuracy and poor working mode recognition accuracy, and achieves the technical effect of improving the multi-mode radar signal classification accuracy and working mode recognition accuracy.

[0005] The present application provides a radar working mode recognition method based on a graph neural network, which includes: after receiving a radar signal, performing pulse arrival time detection on the radar signal to construct a pulse parameter set; after normalizing the pulse parameter set, extracting a multidimensional pulse feature vector; using the multidimensional pulse feature vector to construct a pulse parameter similarity graph, wherein the evaluation dimensions of the pulse parameter similarity graph include basic parameters, statistical features, and intra-pulse modulation; inputting the pulse parameter similarity graph into a multi-relational graph neural network to perform graph feature aggregation and output a high-dimensional embedding representation of each pulse node; dividing the high-dimensional embedding representation into N time sub-blocks based on the pulse time series, where N is an integer greater than 1, performing graph embedding pooling on the N time sub-blocks respectively, and using an attention mechanism to perform weighted aggregation of the graph embedding pooling results to establish a global pulse sequence embedding; performing confidence recognition based on the global pulse sequence embedding, and outputting a working mode recognition result.

[0006] In a possible implementation, the radar operating mode recognition method based on graph neural network also performs the following processing: configuring the length and step parameters of the sliding window according to the radar signal; using the sliding window to perform sliding traversal of the pulse time series, and calculating the local variability index in the traversal time window of each sliding traversal; using the calculation results of the local variability index to mark the detection mutation points; and dividing the high-dimensional embedding representation into N time sub-blocks according to the detection mutation points.

[0007] In a possible implementation, the radar working mode recognition method based on graph neural network also performs the following processing: establishing multidimensional feature indicators, the multidimensional feature indicators include mean square error indicator, coefficient of variation indicator, quantile difference indicator, and extreme ratio indicator; calculating the local variability indicator within the traversal time window based on the multidimensional feature indicator to generate the local variability indicator calculation result; if the local variability indicator calculation result meets the preset threshold, executing the responsible pulse positioning within the traversal time window; using the responsible pulse positioning result to mark the detection mutation point.

[0008] In a possible implementation, the radar operating mode recognition method based on graph neural network further performs the following processing: configuring a calibrated number of time sub-blocks according to the radar signal; locating basic segmentation points using the detected mutation points; generating a basic number of segments after basic segmentation of the high-dimensional embedding representation using the basic segmentation points; performing deviation identification on the basic number of segments and the calibrated number, and performing re-segmentation of the basic segmentation using the deviation identification result to divide the high-dimensional embedding representation into N time sub-blocks.

[0009] In a possible implementation, the radar working mode recognition method based on graph neural network also performs the following processing: extracting the local structural indicators of the pulse nodes within each time sub-block and establishing the extraction results; using the extraction results to configure the graph structure role of the pulse node, and calling the role weight corresponding to the graph structure role; and performing graph embedding pooling management of the corresponding time sub-block according to the role weight.

[0010] In a possible implementation, the radar working mode recognition method based on graph neural network also performs the following processing: using the detection mutation point to perform aggregation partitioning; using the attention mechanism to reconstruct the attention weight within each aggregated partition; performing weighted aggregation of the graph embedding pooling results within the aggregated partition according to the attention weight reconstruction result, and outputting a global pulse sequence embedding with the aggregated partition identifier.

[0011] In a possible implementation, the radar working mode recognition method based on graph neural network also performs the following processing: obtaining training samples of the multi-relationship graph neural network; after configuring the adversarial generative network, using the training samples as original data, performing adversarial sample generation and establishing incremental samples; and using the incremental samples to perform network incremental optimization of the multi-relationship graph neural network.

[0012] In a possible implementation, the radar operating mode recognition method based on graph neural network also performs the following processing: the extracted features of the pulse parameter set include pulse repetition interval, pulse width, carrier frequency, frequency change rate, pulse arrival angle, pulse amplitude, and pulse modulation type.

[0013] In a possible implementation, the radar working mode recognition method based on graph neural network also performs the following processing: embedding the global pulse sequence into the input classification network to obtain the category score of the working mode; converting the category score into confidence probability based on the Softmax function, and outputting the working mode recognition result.

[0014] The present application also provides a radar working mode recognition system based on a graph neural network, which includes: a pulse parameter set construction module, which is used to perform pulse arrival time detection on the radar signal after receiving the radar signal and construct a pulse parameter set; a pulse feature vector extraction module, which is used to extract a multidimensional pulse feature vector after normalizing the pulse parameter set; a parameter similarity graph construction module, which is used to use the multidimensional pulse feature vector to construct a pulse parameter similarity graph, and the evaluation dimensions of the pulse parameter similarity graph include basic parameters, statistical features, and intra-pulse modulation; a graph feature aggregation module, which is used to input the pulse parameter similarity graph into a multi-relational graph neural network to perform graph feature aggregation and output a high-dimensional embedding representation of each pulse node; a global pulse sequence embedding establishment module, which is used to divide the high-dimensional embedding representation into N time sub-blocks based on the pulse time series, where N is an integer greater than 1, perform graph embedding pooling of the N time sub-blocks respectively, and use the attention mechanism to perform weighted aggregation of the graph embedding pooling results to establish a global pulse sequence embedding; and a confidence recognition module, which is used to perform confidence recognition based on the global pulse sequence embedding and output a working mode recognition result.

[0015] The radar working mode recognition method and system based on graph neural network proposed in this application is intended to perform pulse arrival time detection on the received radar signal and construct a pulse parameter set; extract multidimensional pulse feature vectors after normalization processing; use the multidimensional pulse feature vectors to construct a pulse parameter similarity graph; input to a multi-relational graph neural network to perform graph feature aggregation and output a high-dimensional embedding representation of each pulse node; divide the high-dimensional embedding representation into N time sub-blocks based on the pulse time series and perform weighted aggregation of the graph embedding pooling results; perform confidence recognition based on the global pulse sequence embedding and output the working mode recognition result. This solves the technical problems existing in the prior art, such as the difficulty in effectively modeling the correlation between pulses and the lack of anti-interference ability, which lead to insufficient multi-mode radar signal classification accuracy and poor working mode recognition accuracy, and achieves the technical effect of improving the multi-mode radar signal classification accuracy and working mode recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] To more clearly illustrate the technical solutions of the embodiments of the present disclosure, the accompanying drawings of the embodiments of the present disclosure are briefly introduced below. Flowcharts are used in this application to illustrate the operations performed by the systems according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in precise order. Instead, various steps may be processed in reverse order or simultaneously as needed. Furthermore, other operations may be added to these processes, or one or more operations may be removed from these processes.

[0017] Figure 1 A schematic flow chart of a radar operating mode recognition method based on a graph neural network provided in an embodiment of the present application.

[0018] Figure 2 Schematic diagram of the structure of the radar working mode recognition system based on graph neural network provided in an embodiment of the present application.

[0019] Explanation of the reference numerals: pulse parameter set construction module 10 , pulse feature vector extraction module 20 , parameter similarity graph construction module 30 , graph feature aggregation module 40 , global pulse sequence embedding establishment module 50 , confidence recognition module 60 . DETAILED DESCRIPTION

[0020] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below.

[0021] In order to make the purposes, technical solutions and advantages of the present application clearer, the following will make further detailed description to the present application in conjunction with the drawings, and the described embodiments should not be regarded as limitation to the present application, all other embodiments obtained by those skilled in the art without creative labor shall belong to the protection scope of the present application.

[0022] In the following description, "some embodiments" are involved, which describe a subset of all possible embodiments, but it can be understood that "some embodiments" can be the same subset or different subset of all possible embodiments, and can be combined with each other without conflict, and the term "first\second" involved only distinguishes similar objects, and does not represent a specific order for the objects. The terms "include" and "have" and any variations, are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or modules not clearly listed or inherent to these processes, methods, products or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by those skilled in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application.

[0023] The embodiments of the present application provide a radar working mode recognition method based on a graph neural network, as shown in Figure 1 The method comprises the following steps:

[0024] Step S100, after receiving a radar signal, performing pulse arrival time detection on the radar signal to construct a pulse parameter set.

[0025] Preferably, the radar signal is usually composed of multiple discrete pulses, and the arrival time of each pulse is a key parameter for signal processing. Specifically, after receiving the radar signal, the received radio frequency signal is sampled, and the start and end time of the pulse is detected by using energy detection, matched filtering or time domain envelope analysis, and in a high-density signal environment, the pulses of multiple radars can be superimposed, the pulses of different radiation sources are preliminarily separated by the parameters such as pulse arrival time, carrier frequency, pulse width, and the accurate arrival time is marked for each detected pulse to form time sequence data; and then a group of key parameters describing the characteristics of the radar pulse, i.e. the pulse parameter set, is obtained, which can include pulse arrival time, carrier frequency, pulse width, amplitude, and can also include pulse repetition interval and modulation types such as single frequency, linear frequency modulation and phase coding.

[0026] Step S200, after normalizing the pulse parameter set, extracting a multi-dimensional pulse feature vector.

[0027] Step S200 further includes that the extracted features of the pulse parameter set include pulse repetition interval, pulse width, carrier frequency, frequency change rate, pulse arrival angle, pulse amplitude, and pulse modulation type.

[0028] Preferably, the original pulse parameters are standardized and key features are extracted from them so that the radar working mode can be learned and classified more effectively. Specifically, since the numerical ranges and units of different pulse parameters vary greatly, directly importing them into the model as input data will lead to unstable training or deviation. The pulse parameters in the pulse parameter set are normalized, that is, each parameter is mapped to a unified scale. For example, the maximum and minimum values ​​of each pulse parameter are obtained by Min-Max normalization, and the normalized value of the non-zero element is calculated; after the normalization is completed, multiple key features are extracted from the pulse parameter set, including pulse repetition interval, pulse width, carrier frequency, frequency change rate, pulse arrival angle, pulse amplitude, pulse Pulse modulation type: The pulse repetition interval is the difference in arrival time between adjacent pulses, reflecting the radar's scanning pattern and used to distinguish radar types. Pulse width is the duration of a single pulse; narrow pulses may correspond to high-resolution radars, while wide pulses may be used for long-range detection. Carrier frequency refers to the center frequency of the pulse and is used to distinguish radar frequency bands. Frequency variation rate is the pattern of carrier frequency variation between pulses, such as frequency hopping or linear frequency modulation slope. Pulse arrival angle refers to the incident direction of the pulse, determined by multi-antenna direction finding and used to assist radar sorting and emitter location. Pulse amplitude refers to the received power of the pulse, reflecting radar range or antenna scanning pattern. Pulse modulation type refers to the intra-pulse modulation method, such as single frequency, linear frequency modulation, and phase coding, used to identify radar functions. Finally, the extracted features corresponding to each pulse are used as elements of a vector to generate a multidimensional pulse feature vector for each pulse. By combining basic parameters with advanced features (such as modulation type), the accuracy of pattern recognition is improved.

[0029] Step S300: constructing a pulse parameter similarity graph using the multi-dimensional pulse feature vector, wherein the evaluation dimensions of the pulse parameter similarity graph include basic parameters, statistical features, and intra-pulse modulation.

[0030] Preferably, the radar pulse sequence is modeled as a graph structure by evaluating the similarity of pulse characteristics in multiple dimensions. That is, the multi-dimensional pulse feature vector is constructed as a pulse parameter similarity graph, in which each pulse is a node in the graph and the similarity relationship between pulses is an edge. Specifically, the evaluation is performed from three key evaluation dimensions: basic parameters, statistical characteristics, and intra-pulse modulation. The basic parameter dimension is used to evaluate the basic electromagnetic characteristics of the pulse, including directly measured parameters such as pulse repetition interval, pulse width, and carrier frequency, which reflect the basic operating mode of the radar. For example, a fixed pulse repetition interval corresponds to a conventional scanning radar, while a variable pulse repetition interval may correspond to an agile or staggered mode. The statistical feature dimension is used to analyze the statistical laws of the pulse sequence, such as the jitter characteristics of the pulse repetition interval, the carrier frequency hopping law, the distribution of pulse amplitude, etc., which can capture the dynamic behavior of the radar signal, such as the statistical characteristics of the carrier frequency variation of the frequency hopping radar or the periodic law of the pulse group. The intra-pulse modulation dimension evaluates the modulation characteristics within the pulse, including complex modulation methods such as linear frequency modulation and phase coding. Modulation parameters such as the frequency modulation slope and coding sequence are extracted through time-frequency analysis or correlation detection. These parameters are used to distinguish between radar's fine-grained operating modes, such as synthetic aperture radar and conventional search radar. When constructing a similarity graph, the characteristics of the three evaluation dimensions are integrated, and the strength of the association between pulse nodes is measured using Euclidean distance or cosine similarity. For example, two pulses with similar pulse repetition intervals, similar frequency hopping patterns, and the same modulation are assigned a stronger connection weight. This ultimately yields a pulse parameter similarity graph that comprehensively characterizes the spatiotemporal correlations and modulation characteristics of radar signals.

[0031] Step S400: Input the pulse parameter similarity graph into a multi-relational graph neural network to perform graph feature aggregation and output a high-dimensional embedding representation of each pulse node.

[0032] Step S400 further includes step S401, obtaining training samples of the multi-relationship graph neural network; step S402, after configuring the adversarial generative network, using the training samples as original data, performing adversarial sample generation and establishing incremental samples; step S403, using the incremental samples to perform network incremental optimization of the multi-relationship graph neural network.

[0033] Preferably, pulse sequences are collected through electronic support measures or radar reconnaissance equipment to obtain real radar signals and mark their operating modes, such as search, tracking, imaging, etc.; a radar signal simulator is used to generate pulse streams of different operating modes, and simulation data is determined to ensure coverage of diverse parameter combinations such as pulse repetition intervals, carrier frequencies, and modulation types; and then an initial training sample of a multi-relational graph neural network is constructed, wherein each sample contains a pulse similarity graph, in which nodes represent single pulses, edges represent similarities between pulses, and classification labels corresponding to radar modes. Then, an adversarial generative network is configured. The generator is used to input random noise and output a synthesized pulse feature vector, such as a forged pulse repetition interval, carrier frequency, modulation parameters, etc. The discriminator is used to distinguish between real similarity pulse graphs and similarity pulse graphs forged by the generator. Specifically, the generator attempts to generate data that is close to the real pulse feature distribution but has slight perturbations, such as slightly adjusting the statistical laws of the pulse repetition interval and modifying the modulation parameters to deceive the discriminator so that it cannot distinguish between real and generated samples; the discriminator learns to identify the differences between real pulse graphs and generated pulse graphs to improve robustness; the trained generator is used to generate incremental samples, including adversarial samples and enhanced samples, where adversarial samples are perturbed data generated to target the weaknesses of the existing model, and enhanced samples are used to supplement the scarce patterns in the real data.

[0034] Preferably, the incremental samples are finally used to perform network incremental optimization of the multi-relationship graph neural network to improve its ability to recognize unknown signals. Specifically, the original training samples are used to train the multi-relationship graph neural network so that it can initially master the radar working mode classification ability. The adversarial samples and enhanced samples output by the adversarial generative network are then added to the training set to form an incremental training data set. The graph neural network is subjected to adversarial training and continuous learning, so that when aggregating node features, the multi-relationship graph neural network not only considers the original pulse similarity, but also needs to adapt to the hidden correlation in the adversarial samples. For example, the weights of different relationships are dynamically adjusted through the attention mechanism to improve the generalization of the multi-relationship graph neural network, and it has higher fault tolerance to pulse loss, noise interference and intentional deception, while improving adaptability to complex electromagnetic environments.

[0035] Preferably, the pulse parameter similarity graph is input into a multi-relationship graph neural network to perform graph feature aggregation, which is to convert each pulse node into a corresponding high-dimensional embedding representation through graph structure data aggregation, that is, a dense vector, such as 128 dimensions or 256 dimensions, which comprehensively characterizes the global characteristics of the pulse and its contextual relationship in the radar signal, and not only considers the similarity between pulses, but also models the correlation of different dimensions, such as modulation similarity, frequency hopping pattern similarity, etc. Specifically, the pulse parameter similarity graph is used as input data, and feature aggregation is realized through a multi-relationship graph neural network, including aggregating the neighbor information of each pulse node under different relationships, for example, aggregating pulse nodes with similar basic parameters, aggregating pulse nodes with the same modulation type, and using the relational attention mechanism to dynamically assign weights of different relationships, and fusing the results of different relationship aggregation to generate an intermediate representation of the node; then, by stacking multiple graph neural network layers, higher-order neighbor information is gradually fused, and finally a high-dimensional embedding representation of the pulse node is output for high-precision classification, which can intuitively reflect the difference in radar modes.

[0036] Step S500: Divide the high-dimensional embedding representation into N time sub-blocks based on the pulse time series, where N is an integer greater than 1, perform graph embedding pooling on the N time sub-blocks respectively, and use the attention mechanism to perform weighted aggregation of the graph embedding pooling results to establish a global pulse series embedding.

[0037] Step S500 further includes step S510, configuring the length and step parameters of the sliding window according to the radar signal; step S520, using the sliding window to perform a sliding traversal of the pulse time series, and calculating the local variability index in the traversal time window of each sliding traversal; step S530, using the calculation result of the local variability index to mark the detection mutation point; step S540, dividing the high-dimensional embedding representation into N time sub-blocks according to the detection mutation point.

[0038] Preferably, the high-dimensional embedding representation is divided into N time sub-blocks based on the pulse time series, that is, the high-dimensional embedding representation is cut into multiple continuous sub-blocks according to the time dimension using the radar pulse sequence, and each sub-block contains a set of pulse embeddings within a time window, wherein N is an integer greater than 1, such as N=5. The pulse time series refers to discrete time series data formed by the electromagnetic pulse signals emitted or received by the radar arranged in order of arrival time. Each pulse is described by a set of pulse parameters. Specifically, the length and step parameters of the sliding window are configured according to the radar signal. The sliding window is used to capture the time-varying characteristics of the pulse sequence. The window length needs to cover the minimum working mode period of the radar. If the window length is too short, the mode change will be missed, and if it is too long, the time resolution will be reduced. The step parameter is usually set to 1 / 5~1 / 2 of the window length. A smaller step size can improve the accuracy of mutation point detection, but increase the amount of calculation.

[0039] Preferably, a sliding window is used to perform a sliding traversal of the pulse time series. That is, the window slides along the time axis, each time moving a step size to intercept a subsequence. Local variability indices are calculated within each sliding traversal time window. These indices reflect signal mutations, including the variance of the pulse repetition interval, carrier frequency variance, pulse amplitude entropy, short-time Fourier transform energy mutation detection within the sliding window, and the cosine similarity matrix of the high-dimensional embedding of pulses within the window. The results of the local variability indices are then used to mark mutation points. For example, if the local variability index exceeds k standard deviations of the historical mean (e.g., k = 3), it is marked as a mutation point. Alternatively, local extreme points in the variability index sequence are identified as mutation points. A list of timestamps for the detected mutation points is then output, indicating the moments when the radar operating mode may change. Finally, the high-dimensional embedding representation of the pulse nodes is segmented by the detected mutation points, ensuring consistent patterns within the same sub-block. This includes a set of high-dimensional embeddings containing continuous pulses. For example, if the detected mutation point is t1 = 50 ms, it is divided into sub-block 1 (0-50 ms) and sub-block 2 (50 ms to the end).

[0040] Furthermore, step S530 also includes step S531, establishing multidimensional feature indicators, wherein the multidimensional feature indicators include a mean square error indicator, a coefficient of variation indicator, a quantile difference indicator, and an extreme ratio indicator; step S532, calculating the local variability indicator within the traversal time window based on the multidimensional feature indicators, and generating a local variability indicator calculation result; step S533, if the local variability indicator calculation result meets the preset threshold, executing the responsible pulse positioning within the traversal time window; step S534, using the responsible pulse positioning result to mark the detection mutation point.

[0041] Preferably, a multidimensional characteristic index is established to quantify the local variability of the pulse parameters, including the mean square error index, the coefficient of variation index, the quantile difference index, and the extreme ratio index. Among them, the mean square error index reflects the absolute fluctuation intensity of the pulse parameter within the window; the coefficient of variation index is used to standardize the volatility and compare the degree of variation of pulse parameters of different magnitudes; the quantile difference index uses the interquartile range to measure the data dispersion; and the extreme ratio index is used to detect the proportion of extreme values, such as sudden PRI jumps. Then, the local variability index within the traversal time window is calculated based on the multidimensional characteristic index. That is, the four characteristic indicators are calculated for each pulse parameter to form a characteristic vector, and weighted aggregation is performed according to the importance of different pulse parameters to generate the local variability index calculation result.

[0042] Preferably, the preset threshold is determined through historical data statistics or simulation data. If the calculated result of the local variability index meets the preset threshold, indicating that there is a significant mode change in the current window, such as when the radar switches from search mode to tracking mode, the responsible pulse location within the traversal time window is performed. That is, the key responsible pulse that causes the sudden increase in variability is located within the window. This includes combining the maximum deviation pulse in the analysis window and comparing the mean values ​​at both ends of the window (comparing the mean values ​​before and after the window) to determine the responsible pulse location result within the traversal time window. Finally, the responsible pulse location result is used to mark the detection mutation point, that is, the arrival time of the responsible pulse is recorded as the detection mutation point. This provides a reliable time series segmentation basis for radar operating mode recognition, facilitating agile radar signal analysis and recognition in complex electromagnetic environments.

[0043] Furthermore, step S540 also includes step S541, configuring a calibrated number of time sub-blocks according to the radar signal; step S542, locating a basic segmentation point using the detected mutation point; step S543, generating a basic segmentation number after basic segmentation of the high-dimensional embedding representation using the basic segmentation point; step S544, performing deviation identification on the basic segmentation number and the calibrated number, and performing re-segmentation of the basic segmentation using the deviation identification result to divide the high-dimensional embedding representation into N time sub-blocks.

[0044] Preferably, a calibration number of time sub-blocks is configured according to the periodic characteristics of the radar signal, wherein the calibration number is not necessarily equal to N. The basic segmentation points are then located using the detection mutation points, and the high-dimensional embedded representation is subjected to basic segmentation with the detection mutation points as boundaries to obtain M basic segments, and then the number of basic segments is determined, wherein M is the number of detection mutation points + 1. Deviation identification is performed on the basic segment number and the calibration number. For example, if the basic segment number is less than the calibration number, it means that short-term interference is not identified and some mode switching points may be missed. If the basic segment number is greater than the calibration number, it means that instantaneous pulse loss is misjudged as mode switching, and false mutation points may be detected due to noise. The relative deviation ratio of the basic segment number to the calibration number is calculated to obtain a deviation identification result. The deviation identification result is then used to perform re-segmentation of the basic segmentation. If the basic segment number is less than the calibration number, secondary mutation points are detected in the long basic segment and the long basic segment is further segmented. If the basic segment number is greater than the calibration number, the cosine similarity of the sub-block embeddings is calculated and continuous sub-blocks with high similarity are merged. Finally, the high-dimensional embedded representation is divided into N time sub-blocks.

[0045] Furthermore, step S500 also includes step S550, extracting the local structural indicators of the pulse nodes within each time sub-block and establishing the extraction results; step S560, configuring the graph structure role of the pulse node using the extraction results, and calling the role weight corresponding to the graph structure role; step S570, performing graph embedding pooling management of the corresponding time sub-block according to the role weight.

[0046] Preferably, local structural indicators of the impulse nodes within each time sub-block are extracted to quantify the local topological importance of each impulse node within its time sub-block. This includes extracting degree centrality, which counts the number of edges connecting the node and reflects the local correlation strength of the impulses; extracting the clustering coefficient, which measures the closeness of connections between node neighbors and reflects local signal consistency; and extracting betweenness centrality, which identifies key transition nodes connecting different patterns. The extracted results are then used to configure the graph structural roles of the impulse nodes. This involves classifying nodes into predefined roles based on a combination of indicators. Specifically, nodes with high centrality combined with low betweenness are considered core nodes, corresponding to a role weight of 0.1; high betweenness centrality is considered a bridge node, corresponding to a role weight of 0.7; and low degree centrality is considered an edge node, corresponding to a role weight of 0.2. Finally, graph embedding pooling management is performed for the corresponding time sub-block based on the role weights. This involves weighted summing the embedding representations of all impulse nodes in each sub-block according to the role weights to generate a sub-block-level global feature. The output graph embedding pooling result is primarily used for radar behavior analysis and to ensure classification and recognition accuracy.

[0047] Furthermore, step S500 also includes step S570, performing aggregation partitioning by detecting mutation points; step S580, reconstructing the attention weight within each aggregated partition by using the attention mechanism; step S590, performing weighted aggregation of the graph embedding pooling results within the aggregated partition according to the attention weight reconstruction results, and outputting a global pulse sequence embedding with the aggregated partition identifier.

[0048] Preferably, the mutation point is detected as a natural boundary for aggregated partitioning, the pulse sequence is divided into a plurality of semantically continuous aggregated partitions, and it is ensured that the radar operating mode is consistent within each partition, for example, the pulse sequence is divided into search mode pulses with a fixed pulse repetition interval, tracking mode pulses with a jittered pulse repetition interval, and frequency hopping jamming mode pulses; a dynamic weight is assigned to each pulse node within each aggregated partition using an attention mechanism, highlighting the key features of the mode. Specifically, the high-dimensional embedding of the pulse node is taken as the input, the correlation score between nodes is calculated through the attention mechanism, and the weight is normalized to ensure that the sum of the weights is 1, and then the attention weight reconstruction within the partition is performed, so that the weight of the core pulse is increased and the weight of the noise pulse is reduced to close to 0. Finally, the weighted aggregation of the graph embedding pooling result within the aggregated partition is performed according to the attention weight reconstruction result, that is, the node embedding within the aggregated partition is aggregated by weighted summation according to the attention weight, to generate a partition feature vector, and all partition feature vectors are spliced in time sequence and appended with partition identifiers to construct a global pulse sequence embedding, and the global pulse sequence embedding is also segmented and can be directly used for radar operating mode classifier to realize end-to-end radar operating mode recognition.

[0049] Step S600, performing confidence recognition according to the global pulse sequence embedding, and outputting an operating mode recognition result.

[0050] Step S600 further includes step S610, inputting the global pulse sequence embedding into a classification network to obtain a class score of the operating mode; and step S620, converting the class score into a confidence probability based on a Softmax function, and outputting an operating mode recognition result.

[0051] Preferably, the global pulse sequence embedding with an aggregated partition identifier is input into a classification network, wherein the classification network is composed of a multi-layer perceptron (MLP), which is a kind of feedforward artificial neural network containing a plurality of trainable neurons, and is used to map the input global pulse sequence embedding to a class space of radar operating modes. Specifically, the classification network maps the high-dimensional global pulse sequence embedding to the class space, outputs a multi-dimensional vector through a fully connected layer, corresponds to a plurality of radar operating modes, simultaneously calculates a class score of each output dimension, that is, a class score of the operating mode, and the higher the value is, the more likely it belongs to the mode; then the class score is converted into a confidence probability according to the Softmax function, that is, the score is taken as an exponential, the negative value is eliminated and the difference is amplified, and then the Softmax function is used to calculate the confidence probability distribution of the output class score, which intuitively reflects the prediction confidence of each operating mode, and the class with the highest probability is selected as the final operating mode recognition result. If the highest probability is lower than a preset probability threshold, it is determined as an unknown mode, thereby ensuring the classification accuracy of the multi-mode radar signal and the accuracy of the operating mode recognition.

[0052] In the foregoing, reference is made to Figure 1 The radar working mode recognition method based on graph neural network according to an embodiment of the present invention is described in detail. Figure 2 A radar operating mode recognition system based on a graph neural network according to an embodiment of the present invention is described.

[0053] The radar working mode recognition system based on graph neural network according to the embodiment of the present invention is used to solve the technical problems existing in the prior art, such as the difficulty in effectively modeling the correlation between pulses and the insufficient anti-interference ability, which leads to insufficient multi-mode radar signal classification accuracy and poor working mode recognition accuracy. It achieves the technical effect of improving the multi-mode radar signal classification accuracy and working mode recognition accuracy. Figure 2 As shown, the radar working mode recognition system based on graph neural network includes: a pulse parameter set construction module 10, a pulse feature vector extraction module 20, a parameter similarity graph construction module 30, a graph feature aggregation module 40, a global pulse sequence embedding establishment module 50, and a confidence recognition module 60.

[0054] A pulse parameter set construction module 10 is used to perform pulse arrival time detection on the radar signal after receiving the radar signal and construct a pulse parameter set; a pulse feature vector extraction module 20 is used to extract a multidimensional pulse feature vector after normalizing the pulse parameter set; a parameter similarity graph construction module 30 is used to construct a pulse parameter similarity graph using the multidimensional pulse feature vector, wherein the evaluation dimensions of the pulse parameter similarity graph include basic parameters, statistical features, and intra-pulse modulation; a graph feature aggregation module 40 is used to input the pulse parameter similarity graph into a multi-relational graph neural network to perform graph feature aggregation and output a high-dimensional embedding representation of each pulse node; a global pulse sequence embedding establishment module 50 is used to divide the high-dimensional embedding representation into N time sub-blocks based on the pulse time series, where N is an integer greater than 1, perform graph embedding pooling on the N time sub-blocks respectively, and use the attention mechanism to perform weighted aggregation of the graph embedding pooling results to establish a global pulse sequence embedding; a confidence recognition module 60 is used to perform confidence recognition based on the global pulse sequence embedding and output a working mode recognition result.

[0055] The specific configuration of the global pulse sequence embedding establishment module 50 will be described in detail below. The global pulse sequence embedding establishment module 50 further includes: configuring the length and step size parameters of the sliding window based on the radar signal; performing a sliding traversal of the pulse time series using the sliding window, and calculating a local variability index in the traversal time window of each sliding traversal; using the local variability index calculation result to mark the detected mutation points; and dividing the high-dimensional embedding representation into N time sub-blocks based on the detected mutation points.

[0056] The specific configuration of the global pulse sequence embedding establishment module 50 will be described in detail below. The global pulse sequence embedding establishment module 50 further includes: establishing multidimensional feature indicators, including a mean square error indicator, a coefficient of variation indicator, a quantile difference indicator, and a limit ratio indicator; calculating a local variability indicator within the traversal time window based on the multidimensional feature indicators to generate a local variability indicator calculation result; locating the responsible pulse within the traversal time window if the local variability indicator calculation result meets a preset threshold; and marking the detection mutation point using the responsible pulse location result.

[0057] The specific configuration of the global pulse sequence embedding establishment module 50 will be described in detail below. The global pulse sequence embedding establishment module 50 further includes: configuring a calibrated number of time sub-blocks based on the radar signal; locating basic segmentation points using the detected mutation points; performing basic segmentation of the high-dimensional embedded representation using the basic segmentation points to generate a basic number of segments; identifying deviations between the basic number of segments and the calibrated number, and re-segmenting the basic segmentation using the deviation identification result to divide the high-dimensional embedded representation into N time sub-blocks.

[0058] The specific configuration of the global pulse sequence embedding establishment module 50 will be described in detail below. The global pulse sequence embedding establishment module 50 further includes: extracting local structural indicators of the pulse nodes within each time sub-block and generating an extraction result; configuring the graph structure role of the pulse node using the extraction result and invoking the role weight corresponding to the graph structure role; and performing graph embedding pooling management for the corresponding time sub-block based on the role weight.

[0059] The specific configuration of the global pulse sequence embedding establishment module 50 will be described in detail below. The global pulse sequence embedding establishment module 50 further includes: using mutation point detection to perform clustering and partitioning; using an attention mechanism to reconstruct the attention weight within each clustered partition; performing weighted aggregation of the graph embedding pooling results within the clustered partition based on the attention weight reconstruction results, and outputting a global pulse sequence embedding with a clustered partition identifier.

[0060] The specific configuration of graph feature aggregation module 40 will be described in detail below. Graph feature aggregation module 40 further includes: obtaining training samples for a multi-relational graph neural network; after configuring an adversarial generative network, using the training samples as raw data, performing adversarial sample generation to create incremental samples; and utilizing the incremental samples to perform incremental network optimization of the multi-relational graph neural network.

[0061] The following describes in detail the specific configuration of the pulse feature vector extraction module 20. The pulse feature vector extraction module 20 further includes: the extraction features of the pulse parameter set include pulse repetition interval, pulse width, carrier frequency, frequency change rate, pulse arrival angle, pulse amplitude, and pulse modulation type.

[0062] The following describes the specific configuration of the confidence recognition module 60 in detail. The confidence recognition module 60 further includes: embedding the global pulse sequence into an input classification network to obtain the category score of the working mode; converting the category score into a confidence probability based on the Softmax function, and outputting the working mode recognition result.

[0063] The radar working mode recognition system based on graph neural network provided by an embodiment of the present invention can execute the radar working mode recognition method based on graph neural network provided by any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0064] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules may be used and run on the user terminal and / or server, and the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other and are not used to limit the scope of protection of the present invention.

[0065] The above specific embodiments do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application shall be included within the scope of protection of this application.

Claims

1. A radar working mode recognition method based on graph neural network, characterized in that: The method comprises: After receiving the radar signal, performing pulse arrival time detection on the radar signal to construct a pulse parameter set; After normalizing the pulse parameter set, extracting a multidimensional pulse feature vector; Constructing a pulse parameter similarity graph using the multidimensional pulse feature vector, wherein evaluation dimensions of the pulse parameter similarity graph include basic parameters, statistical characteristics, and intra-pulse modulation; Inputting the pulse parameter similarity graph into a multi-relational graph neural network to perform graph feature aggregation and output a high-dimensional embedding representation of each pulse node; Divide the high-dimensional embedding representation into N time sub-blocks based on the spike time series, where N is an integer greater than 1, perform graph embedding pooling on each of the N time sub-blocks, and use an attention mechanism to perform weighted aggregation of the graph embedding pooling results to establish a global spike sequence embedding; Perform confidence recognition based on the global pulse sequence embedding and output a working mode recognition result; The step of dividing the high-dimensional embedding representation into N time sub-blocks based on the pulse time series includes: Configuring the length and step size parameters of the sliding window according to the radar signal; Utilizing the sliding window to perform a sliding traversal of the pulse time series, and calculating a local variability index in the traversal time window of each sliding traversal; The results of local variability index calculation are used to mark and detect mutation points; The high-dimensional embedding representation is divided into N time sub-blocks according to the detected mutation points.

2. The radar operating mode recognition method based on graph neural network according to claim 1, characterized in that: The method of using the calculation result of the local variability index to mark and detect the mutation point includes: Establishing multidimensional characteristic indicators, wherein the multidimensional characteristic indicators include a mean square error indicator, a coefficient of variation indicator, a quantile difference indicator, and a limit ratio indicator; Calculating a local variability index within the traversal time window according to the multidimensional feature index to generate a local variability index calculation result; If the calculation result of the local variability index meets the preset threshold, the responsible pulse positioning within the traversal time window is performed; The detection mutation point is marked using the responsible pulse positioning results.

3. The radar operating mode recognition method based on graph neural network according to claim 1, characterized in that: The step of dividing the high-dimensional embedding representation into N time sub-blocks according to the detected mutation points includes: configuring a calibrated number of time sub-blocks according to the radar signal; Utilizing the detected mutation point to locate the basic segmentation point; After basic segmentation of the high-dimensional embedding representation using the basic segmentation points, a basic number of segments is generated; Deviations are identified between the basic segment quantity and the calibration quantity, and a re-segmentation of the basic segmentation is performed using the deviation identification result to divide the high-dimensional embedding representation into N time sub-blocks.

4. The radar operating mode recognition method based on graph neural network according to claim 1, characterized in that: The graph embedding pooling of the N time sub-blocks is performed separately, including: Extract the local structural indicators of the pulse nodes in each time sub-block and establish the extraction results; Using the extraction results, the graph structure role of the impulse node is configured, and the role weight corresponding to the graph structure role is called; Graph embedding pooling management of the corresponding time sub-block is performed according to the role weight.

5. The radar operating mode recognition method based on graph neural network according to claim 4, characterized in that: The method uses the attention mechanism to perform weighted aggregation of the graph embedding pooling results to establish a global spike train embedding, including: Use detection mutation points for aggregation partitioning; Use the attention mechanism to reconstruct the attention weight within each aggregation partition; After performing weighted aggregation of the graph embedding pooling results within the aggregation partition according to the attention weight reconstruction results, the global spike train embedding with the aggregation partition identifier is output.

6. The radar operating mode recognition method based on graph neural network according to claim 1, characterized in that: Before inputting the pulse parameter similarity graph into a multi-relational graph neural network to perform graph feature aggregation, the method includes: Obtain training samples for multi-relational graph neural networks; After configuring the adversarial generative network, using the training samples as original data, performing adversarial sample generation to establish incremental samples; The incremental samples are used to perform network incremental optimization of the multi-relational graph neural network.

7. The radar operating mode recognition method based on graph neural network according to claim 1, characterized in that: The extracted features of the pulse parameter set include pulse repetition interval, pulse width, carrier frequency, frequency change rate, pulse arrival angle, pulse amplitude, and pulse modulation type.

8. The radar operating mode recognition method based on graph neural network according to claim 1, characterized in that: The performing confidence recognition according to the global pulse sequence embedding and outputting the working mode recognition result includes: Embedding the global pulse sequence into an input classification network to obtain a category score of the working mode; After converting the category score into confidence probability based on the Softmax function, the working mode recognition result is output.

9. A radar working mode recognition system based on graph neural network, characterized by: The system is used to implement the radar operating mode recognition method based on graph neural network according to any one of claims 1 to 8, and the system includes: a pulse parameter set construction module, configured to, after receiving a radar signal, perform pulse arrival time detection on the radar signal and construct a pulse parameter set; A pulse feature vector extraction module, configured to extract a multi-dimensional pulse feature vector after normalizing the pulse parameter set; A parameter similarity graph construction module is used to construct a pulse parameter similarity graph using the multidimensional pulse feature vector, wherein the evaluation dimensions of the pulse parameter similarity graph include basic parameters, statistical characteristics, and intra-pulse modulation; A graph feature aggregation module is used to input the pulse parameter similarity graph into a multi-relational graph neural network to perform graph feature aggregation and output a high-dimensional embedding representation of each pulse node; A global spike train embedding establishment module is used to divide the high-dimensional embedding representation into N time sub-blocks based on the spike time series, where N is an integer greater than 1, perform graph embedding pooling on each of the N time sub-blocks, and use an attention mechanism to perform weighted aggregation of the graph embedding pooling results to establish a global spike train embedding; The confidence recognition module is used to perform confidence recognition according to the global pulse sequence embedding and output a working mode recognition result.

Images