A method and device for identifying a repetition frequency type based on graph domain mapping and feature enhancement

Abstract

The application provides a pulse repetition frequency type recognition method and device based on graph domain mapping and feature enhancement, and relates to the technical field of radar signal processing. The method comprises the following steps: determining a pulse repetition frequency sequence corresponding to a screened pulse repetition frequency; taking the reciprocal of the pulse repetition frequency sequence to obtain a pulse repetition interval sequence; dividing the pulse repetition interval sequence into multiple subsequences according to a preset length; transforming the multiple subsequences into a Gram difference angle field image according to a graph domain mapping algorithm, and dividing the image into training samples and test samples; constructing a pulse repetition frequency type recognition network, inputting the training samples into the pulse repetition frequency type recognition network, training the pulse repetition frequency type recognition network, and obtaining a trained pulse repetition frequency type recognition network; and inputting the test samples into the trained pulse repetition frequency type recognition network to recognize multiple pulse repetition frequency types. In this way, the recognition accuracy and robustness of the pulse repetition frequency type are improved in a complex non-ideal scene with a high proportion of missed pulses and false pulses.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing technology, and in particular to a method and apparatus for identifying repetition frequency types based on graph domain mapping and feature enhancement. Background Technology

[0002] Analyzing the Pulse Repetition Frequency (PRF) is a core part of radar source analysis. PRF is defined as the number of pulses emitted per second, and is the reciprocal of the Pulse Repetition Interval (PRI). It is one of the most complex parameters in radar sources, typically including six PRF types: fixed, jittery, uneven, slip, grouped, and sinusoidal. The diversity of PRF types reflects various radar operating modes, applications, and technical characteristics, comprehensively characterizing the temporal structure of radar signals. Accurate analysis of PRF types is crucial for in-depth exploration of radar signal characteristics and inference of radar operating modes. Furthermore, PRF type analysis helps invert radar signal models, providing interference basis and feedback information for radar signal prediction and the formulation of radar countermeasure strategies. Therefore, research on PRF type identification of radar sources is of great significance. Combining PRF type identification with operating mode labeling can be used to identify radar operating modes. PRF type identification focuses on the identification of signal repetition characteristics and can be applied to various types of radar.

[0003] Currently, the identification of repetition frequency types usually adopts a method based on supervised neural networks. However, this method requires starting with sequence processing. In complex and non-ideal scenarios with a high proportion of missing pulses and false pulses, the differences between sequences are not significant. Directly inputting sequences with little difference into supervised neural networks does not yield good results. As a result, the accuracy and robustness of repetition frequency type identification are low in complex and non-ideal scenarios with a high proportion of missing pulses and false pulses. Summary of the Invention

[0004] The purpose of this invention is to provide a method and apparatus for identifying repetition type based on graph domain mapping and feature enhancement, which solves the problem of low accuracy and poor robustness in identifying repetition type in complex and non-ideal scenarios with a high proportion of missing pulses and false pulses.

[0005] To address the aforementioned technical problems, the embodiments of the present invention provide the following technical solutions:

[0006] The first aspect of this invention provides a method for repetition frequency type recognition based on graph domain mapping and feature enhancement, the method comprising:

[0007] Multiple pulse repetition frequencies are filtered according to a preset threshold range to determine the pulse repetition frequency sequence corresponding to the filtered pulse repetition frequencies.

[0008] The pulse repetition frequency sequence is reciprocated to obtain the pulse repetition interval sequence.

[0009] The pulse repetition interval sequence is divided into multiple sub-sequences according to a preset length;

[0010] According to the graph domain mapping algorithm, multiple subsequences are transformed into corresponding Gram difference angle field images, and multiple Gram difference angle field images are divided into training samples and test samples.

[0011] Construct a repetition type recognition network, which includes a convolutional neural subnetwork and an attention mechanism layer;

[0012] The training samples are input into the frequency repetition type recognition network to train the frequency repetition type recognition network and obtain the trained frequency repetition type recognition network.

[0013] The test samples are input into the trained frequency repetition type recognition network, which identifies multiple frequency repetition types.

[0014] A second aspect of this application provides a repetition frequency type recognition device based on graph domain mapping and feature enhancement, the device comprising:

[0015] The filtering module is used to filter multiple pulse repetition frequencies according to a preset threshold range and determine the pulse repetition frequency sequence corresponding to the filtered pulse repetition frequencies.

[0016] The processing module is used to take the reciprocal of the pulse repetition frequency sequence to obtain the pulse repetition interval sequence;

[0017] The segmentation module is used to segment the pulse repetition interval sequence according to a preset length to obtain multiple sub-sequences;

[0018] The transformation module is used to transform multiple subsequences into corresponding Gram difference angle field images according to the graph domain mapping algorithm, and to divide the multiple Gram difference angle field images into training samples and test samples.

[0019] The building module is used to construct a repetition type recognition network, which includes a convolutional neural subnetwork and an attention mechanism layer;

[0020] The training module is used to input training samples into the frequency repetition type recognition network, train the frequency repetition type recognition network, and obtain the trained frequency repetition type recognition network.

[0021] The recognition module is used to input test samples into the trained repetition frequency type recognition network and identify multiple repetition frequency types.

[0022] A third aspect of this application provides a computer-readable storage medium including a stored program, wherein, when the program is running, it controls the device where the storage medium is located to execute the repetition frequency type recognition method based on graph domain mapping and feature enhancement of the first aspect or any optional embodiment of the first aspect.

[0023] Compared to existing technologies, the repetition type recognition method and apparatus based on graph domain mapping and feature enhancement provided by this invention transforms PRF sequences with indistinct differences in complex, non-ideal scenarios with a high proportion of missing and false pulses into Gram difference angle field maps with more significant feature differences. The repetition type recognition network constructed through convolutional neural sub-networks and attention mechanism layers helps to highlight important image features and strengthens the role of important features in classification. In complex, non-ideal scenarios with a high proportion of missing and false pulses, the Gram difference angle field map with more significant feature differences can be transformed into a repetition type recognition network that strengthens important features, thereby improving the accuracy and robustness of repetition type recognition. Attached Figure Description

[0024] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein:

[0025] Figure 1 The flowchart of the repetition frequency type recognition method based on graph domain mapping and feature enhancement is illustrated schematically. Figure 1 ;

[0026] Figure 2 The flowchart of the repetition frequency type recognition method based on graph domain mapping and feature enhancement is illustrated schematically. Figure 2 ;

[0027] Figure 3 The diagram schematically illustrates the PRI sequence distribution map and the corresponding Gram difference angle field map corresponding to the repetition frequency type;

[0028] Figure 4 The confusion matrix of the test for all non-ideal cases is illustrated schematically;

[0029] Figure 5 A schematic diagram of a repetition type recognition device based on graph domain mapping and feature enhancement is shown. Detailed Implementation

[0030] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0031] It should be noted that, unless otherwise stated, the technical or scientific terms used in this invention should have the ordinary meaning as understood by one of ordinary skill in the art.

[0032] The methods described in the embodiments of the present invention will be explained in detail below.

[0033] Figure 1 The flowchart of the repetition frequency type recognition method based on graph domain mapping and feature enhancement in an embodiment of the present invention is illustrated schematically. See [link to flowchart illustration]. Figure 1 As shown, the method may include:

[0034] S101. Filter multiple pulse repetition frequencies according to a preset threshold range to determine the pulse repetition frequency sequence corresponding to the filtered pulse repetition frequencies.

[0035] Different radars emit PRFs with different threshold ranges. The preset threshold range can be determined based on the radar model. Different radar models have different preset threshold ranges, and no specific limit is made here for the preset threshold range.

[0036] Pulse repetition frequencies exceeding a preset threshold range are discarded, while those within the preset threshold range are retained to determine the selected PRFs, and a PRF sequence is constructed from the selected PRFs.

[0037] S102. Take the reciprocal of the pulse repetition frequency sequence to obtain the pulse repetition interval sequence.

[0038] Taking the reciprocal of the PRF sequence can convert it into a PRI sequence, which facilitates subsequent processing.

[0039] S103. Divide the pulse repetition interval sequence according to the preset length to obtain multiple subsequences.

[0040] Using a non-overlapping sliding window, the PRI sequence is divided into multiple subsequences according to a preset length, which can divide a long PRI sequence into multiple PRI subsequences of a preset length.

[0041] The preset length can be 64 or 128; there is no specific limitation on the preset length here.

[0042] S104. According to the graph domain mapping algorithm, multiple subsequences are transformed into corresponding Gram difference angle field images, and multiple Gram difference angle field images are divided into training samples and test samples.

[0043] The graph domain mapping algorithm can be used to transform multiple one-dimensional PRI subsequences into a two-dimensional image, namely the Gram difference angular field image.

[0044] One subsequence corresponds to one Gram difference angle field image, and the number of Gram difference angle field images is the same as the number of subsequences.

[0045] Multiple Gram difference angle field images are divided into training and testing samples. A portion of all Gram difference angle field images can be randomly selected as training samples, and the other portion as testing samples. The number of Gram difference angle field images included in the training and testing samples can be set manually.

[0046] S105. Construct a network for identifying repetition frequency types.

[0047] The repetition type recognition network includes a convolutional neural subnetwork and an attention mechanism layer.

[0048] By introducing an attention mechanism layer into the convolutional neural subnetwork, it is helpful to highlight the important features of the Gram difference angle field image, suppress the influence of secondary features, and improve the classification performance.

[0049] S106. Input the training samples into the frequency repetition type recognition network, train the frequency repetition type recognition network, and obtain the trained frequency repetition type recognition network.

[0050] The Gram difference angle field image from the training samples is input into the frequency repetition type recognition network to train the frequency repetition type recognition network. The network is then iteratively optimized under the loss function until convergence, resulting in a well-trained frequency repetition type recognition network.

[0051] S107. Input the test samples into the trained frequency repetition type recognition network to identify multiple frequency repetition types.

[0052] Based on the above Figure 1As can be seen from the implementation method, the embodiments of the present invention filter multiple pulse repetition frequencies according to a preset threshold range to determine the pulse repetition frequency sequence corresponding to the filtered pulse repetition frequencies; the pulse repetition frequency sequence is processed by taking the reciprocal to obtain a pulse repetition interval sequence; the pulse repetition interval sequence is divided according to a preset length to obtain multiple sub-sequences; according to the graph domain mapping algorithm, the multiple sub-sequences are transformed into corresponding Gram difference angle field images, and the multiple Gram difference angle field images are divided into training samples and test samples; a frequency repetition type recognition network is constructed, the training samples are input into the frequency repetition type recognition network, the frequency repetition type recognition network is trained, and a trained frequency repetition type recognition network is obtained; the test samples are input into the trained frequency repetition type recognition network to identify multiple frequency repetition types. In this way, PRF sequences with low differences in complex non-ideal scenarios with a high proportion of missing and false pulses can be transformed into Gram difference angle field maps with more significant feature differences. The repetition type recognition network constructed by the convolutional neural sub-network and attention mechanism layer helps to highlight the important features of the image and strengthens the role of important features in classification. In complex non-ideal scenarios with a high proportion of missing and false pulses, the Gram difference angle field map with more significant feature differences can be transformed into a repetition type recognition network that strengthens the important features, thereby improving the accuracy and robustness of repetition type recognition.

[0053] As a refinement and extension of the above embodiments, Figure 2 The following is a flowchart of a repetition type recognition method based on graph domain mapping and feature enhancement in an embodiment of the present invention. Figure 2 See Figure 2 As shown, an embodiment of the present invention provides a method for identifying repetition type based on graph domain mapping and feature enhancement, which may include:

[0054] S201. Filter multiple pulse repetition frequencies according to a preset threshold range to determine the pulse repetition frequency sequence corresponding to the filtered pulse repetition frequencies.

[0055] Different radars emit PRFs with different threshold ranges. The preset threshold range can be determined based on the radar model. Different radar models have different preset threshold ranges, and no specific limit is made here for the preset threshold range.

[0056] The presence of outliers in the input PRFs can distort the feature space after subsequent graph mapping, causing dissimilar data points to become closer together in the transformed feature space, reducing their differences and leading to decreased model performance and stability, ultimately affecting recognition accuracy. To improve the final recognition accuracy, a preset threshold range is set based on the radar model during the preprocessing stage, and outliers exceeding this range are removed. After outlier removal, PRFs within the preset threshold range are retained to identify the selected PRFs, and a PRF sequence is constructed from these selected PRFs.

[0057] S202. Take the reciprocal of the pulse repetition frequency sequence to obtain the pulse repetition interval sequence.

[0058] Taking the reciprocal of the PRF sequence can convert it into a PRI sequence, which facilitates subsequent processing.

[0059] S203. Divide the pulse repetition interval sequence according to the preset length to obtain multiple subsequences.

[0060] The preset length can be 64 or 128; there is no specific limitation on the preset length here.

[0061] Using a non-overlapping sliding window, the PRI sequence is divided into multiple subsequences according to a preset length, which can divide a long PRI sequence into multiple PRI subsequences of a preset length.

[0062] Steps S204-S207 below are specific operations for transforming multiple subsequences into corresponding Gram difference angular field images according to the graph domain mapping algorithm, and dividing the multiple Gram difference angular field images into training samples and test samples.

[0063] S204. Scale multiple subsequences to the range [-1,1] to obtain the scaled values ​​of each subsequence.

[0064] Specifically, multiple subsequences are scaled to the range [-1, 1] to obtain the scaled values ​​of each subsequence, including:

[0065] Using the first formula below, multiple subsequences are scaled to the range [-1, 1] to obtain the scaled values ​​of each subsequence:

[0066]

[0067] in, For the corresponding p n The scaled value, i.e., the scaled value of the nth value in each subsequence, p nLet P be the nth value in each subsequence, where P represents multiple subsequences and N is the length of each subsequence.

[0068] S205. Convert each value to a polar coordinate system to obtain the angle and radius corresponding to each value.

[0069] The scaled value of the nth value in each subsequence is converted to a polar coordinate system according to the following formula: the value is treated as an angle in the polar coordinate system, and the timestamp corresponding to each value is treated as a radius in the polar coordinate system.

[0070] Specifically, using the following formula, each value is converted to a polar coordinate system to obtain the corresponding angle and radius:

[0071]

[0072] in, This represents the angle corresponding to the nth value in the polar coordinate system. For the corresponding p n The scaled value, i.e., the scaled value of the nth value in each subsequence, r n Let t be the radius corresponding to the nth value. n Let N be the timestamp corresponding to the nth value, and N be the length of the subsequence.

[0073] S206. Determine the Gram difference angular field image based on the angular difference between every two values.

[0074] By considering the angular differences between points corresponding to different values ​​to identify the correlation of events at different time points, the Gramian Angular Difference Field (GADF) image can be determined using the difference angle formula.

[0075] Specifically, the GADF image is determined based on the angular difference between every two values, including:

[0076] Step A1: Determine the corresponding matrix based on the angle difference between every two values ​​and the second formula below:

[0077]

[0078] Here, GADF is the matrix corresponding to the angle between every two values. This represents the angle corresponding to the nth value in the polar coordinate system. Let be the angle corresponding to the m-th value in polar coordinates; I is the identity matrix of the same size as the subsequence P. The normalized subsequence for The transpose of .

[0079] Ultimately, the calculation of GADF in Cartesian coordinates becomes an operation similar to an inner product.

[0080] The above processing can transform PRF sequences with high proportions of missing pulses and false pulses in complex and non-ideal scenarios with indistinct differences into Gram difference angular field images with significant corresponding features. Figure 3 The diagram schematically illustrates the PRI sequence distribution plot and the corresponding Gram difference angle field plot corresponding to the repetition frequency type. Specifically, Figure 3 (a) is the PRI sequence diagram corresponding to a fixed type PRF. Figure 3 (d) is the Gram difference angle field diagram corresponding to a fixed type PRF. Figure 3 (b) is the PRI sequence diagram corresponding to the jitter type PRF. Figure 3 (e) is the Gram difference angular field plot corresponding to the jitter type PRF. Figure 3 (c) is the PRI sequence diagram corresponding to the staggered type PRF. Figure 3 (f) is the Gram difference angle field plot corresponding to the staggered type PRF. Figure 3 (g) is the PRI sequence diagram corresponding to the slip-type PRF. Figure 3 (j) is the Gram difference angle field plot corresponding to the slip type PRF. Figure 3 (h) is the PRI sequence graph corresponding to the group edge type PRF. Figure 3 (k) is the Gram difference angle field diagram corresponding to the grouped edge type PRF. Figure 3 (i) is the PRI sequence diagram corresponding to the sinusoidal type PRF. Figure 3 (l) is the Gram difference angle field diagram corresponding to the sinusoidal type PRF. Figure 3 (a)- Figure 3 (l) Displays the distribution of the original PRI sequences for six different repetition rate types and their corresponding Gram difference angle field plots. By comparing the PRI sequence plots of the same repetition rate type with the Gram difference angle field plots, that is... Figure 3 (a) and Figure 3 (d) Compare, Figure 3 (b) and Figure 3 (e) Compare, Figure 3 (c) and Figure 3 (f) Compare, Figure 3 (g) and Figure 3 (j) is compared, Figure 3 (h) and Figure 3 (k) is compared, Figure 3 (i) with Figure 3(l) By comparison, it can be seen that converting each PRI sequence into the corresponding Gram difference angle field map effectively reduces the redundancy of data information, enhances the aggregation of differential features of different repetition frequency types, and helps the learning process of deep neural networks.

[0081] Step A2: Perform grayscale conversion on the matrix to obtain the Gram difference angular field image.

[0082] One subsequence corresponds to one Gram difference angle field image, and the number of Gram difference angle field images is the same as the number of subsequences.

[0083] S207. Divide multiple Gram difference angular field images into training samples and test samples.

[0084] Multiple Gram difference angle field images are divided into training and testing samples. A portion of all Gram difference angle field images can be randomly selected as training samples, and the other portion as testing samples. The number of Gram difference angle field images included in the training and testing samples can be set manually.

[0085] S208. Construct a network for identifying repetition frequency types.

[0086] The repetition type recognition network consists of a convolutional neural subnetwork and an attention mechanism layer. By introducing an attention mechanism layer into the convolutional neural subnetwork, it helps to highlight the important features of the Gram difference angular field image, suppress the influence of secondary features, and improve the classification performance.

[0087] The convolutional neural network consists of two convolutional layers, two normalization layers, two activation layers, two pooling layers, and two fully connected layers. The convolutional and pooling layers are responsible for feature extraction and dimensionality reduction; the activation layers increase the network's non-linearity and enhance the model's feature representation capabilities; the normalization layers are typically placed after the activation layers to prevent overfitting; and the fully connected layers convert two-dimensional features into one-dimensional features.

[0088] The attention mechanism layer can be called the channel feature enhancement module. The channel feature enhancement module adopts the attention mechanism and is an end-to-end network structure. This module mainly includes two operations: compression and activation, which are used to process input features.

[0089] The frequency repetition type recognition network consists of one convolutional layer, one normalization layer, one activation layer, one pooling layer, one attention mechanism layer, one convolutional layer, one normalization layer, one activation layer, one pooling layer, and two fully connected layers.

[0090] To address the low repetition rate recognition rate in complex, non-ideal scenarios with a high proportion of missing and spurious pulses, the constructed repetition type recognition network adds an attention mechanism layer between two convolutional layers of a convolutional neural subnetwork. Global average pooling is used to compress the output feature map of the first convolutional layer into a single feature vector, generating a channel weight vector. This weighting allows the neural network to focus on feature channels with high differences. The adaptively weighted feature map is then input into the second convolutional layer, which facilitates feature filtering and enhancement before convolution, thereby improving the overall expressive power of the network.

[0091] S209. Input the training samples into the frequency repetition type recognition network, train the frequency repetition type recognition network, and obtain the trained frequency repetition type recognition network.

[0092] The Gram difference angle field image from the training samples is input into the frequency repetition type recognition network to train the frequency repetition type recognition network. The network is then iteratively optimized under the loss function until convergence, resulting in a well-trained frequency repetition type recognition network.

[0093] S210. Input the test samples into the trained frequency repetition type recognition network to identify multiple frequency repetition types.

[0094] The trained repetition type recognition network consists of one convolutional layer, one normalization layer, one activation layer, one pooling layer, one attention mechanism layer, one convolutional layer, one normalization layer, one activation layer, one pooling layer, and two fully connected layers.

[0095] Specifically, the test samples are input into the trained repetition frequency type recognition network to identify multiple repetition frequency types, including:

[0096] Step B1: Input the test sample into one convolutional layer, one normalization layer, one activation layer, and one pooling layer in sequence, and output the input features corresponding to the test sample.

[0097] Step B2: Input the input features into one attention mechanism layer and output the final feature representation.

[0098] Specifically, the input features are fed into an attention mechanism layer, and the final feature representation is output, including:

[0099] Step B21: Compress the input features to obtain channel-level global features.

[0100] Step B22: Activate the global features at the channel level to obtain the weights of each channel.

[0101] Step B23: Multiply the channel weights by the input features to obtain the final feature representation.

[0102] Step B3: Input the final feature representation into one convolutional layer, one normalization layer, one activation layer, one pooling layer, and two fully connected layers in sequence, and output multiple repetition frequency types.

[0103] To investigate the impact of different measurement errors, missing pulses, and faulty pulse levels on frequency repetition type identification, the effectiveness of the method of this invention is verified through simulation experiments.

[0104] The simulation conditions were as follows: The experiment combined different levels of measurement error, missing pulses, and spurious pulses to simulate data for six repetition rate (PRF) types: fixed, jittery, uneven, slip, dwell and switching, and sinusoidal. The PRF value ranges for different modulation types are shown in Table 1. Each modulation type generated 38,400 pulses; the ratio of missing pulses and spurious pulses was set between 10% and 60%; the measurement error was achieved by adding Gaussian white noise with a mean of 0 and a variance of 2 μs to each pulse.

[0105] Table 1. PRF value range for different modulation types

[0106] Repetition frequency type Value range (kHz) Constant 0.5～10 Jitter 0.5～10(10％) Sine (Sin) 0.5～3.3 Dwell & Switch 2.5～3.3 Stagger 0.5～10 Sliding 5～10

[0107] The experimental setup is as follows: The parameters of the frequency repetition type recognition network used in this invention are shown in Table 2. In Table 2, Batch size is the number of samples used to update the model weights in each iteration, epoch is the iteration batch, Learningrate is the learning rate, optimizer is the optimizer, Weight decay is the weight decay coefficient (a parameter to prevent overfitting), Dropout is the dropout coefficient (a parameter used to prevent model overfitting), Conv1 and Conv2 are the convolutional layer sizes, and Pool1 and Pool 2 are the pooling layer sizes. 80% of all samples are randomly selected for training, and the remaining 20% ​​are used for validation and testing.

[0108] Table 2 Parameters of the Frequency Repetition Type Identification Network

[0109]

[0110]

[0111] The results analysis specifically involves testing the performance of the proposed method under different proportions of missing and spurious pulses. The results show that the algorithm's performance gradually decreases as the proportion of spurious pulses increases. This is because spurious pulses have high randomness, causing all repetition frequency types to tend towards characteristics similar to jitter repetition frequencies. Conversely, the algorithm's accuracy remains relatively stable as the proportion of missing pulses gradually increases. This indicates that the proposed algorithm has good robustness under the missing pulse condition. It is worth noting that the accuracy under the missing pulse condition is about 10% lower than that under the spurious pulse condition. This is because as the number of lost pulses increases, the number of pulse samples gradually decreases, leading to insufficient training.

[0112] Figure 4 The confusion matrix of the test for all non-ideal cases is schematically shown, according to Figure 4 The confusion matrix results show the sample test results under all non-ideal conditions. Figure 4 The horizontal axis represents the predicted label, and the vertical axis represents the actual label. The graph uses red, orange, and yellow boxes to indicate accuracy; darker colors indicate higher accuracy. The accuracy of the method in this invention is the average of the values ​​in the red boxes along the matrix diagonal, which is 87.72%. This means the overall recognition accuracy of the method reaches 87.72%, and the boundaries between different repetition frequency features are clear, exhibiting high discriminability. This result further demonstrates that the method of this invention has good adaptability under non-ideal conditions.

[0113] This invention utilizes a graph mapping algorithm to convert PRF sequences with indistinct differences in complex, non-ideal scenes containing a high proportion of missing and spurious pulses into feature-rich Gram difference angular field images, reducing information redundancy and enhancing feature aggregation. Building upon this, to address the low recognition rate of repetition types in complex, non-ideal scenes with a high proportion of missing and spurious pulses, a repetition type recognition network is constructed. Based on the channel feature enhancement module (attention mechanism layer) in this network, the network assigns greater weight to significantly different channel features, increasing the dominant role of important features in the classification process. This strengthens the role of important features in classification, improving robustness and accuracy of repetition type recognition in complex, non-ideal scenes.

[0114] Based on the same inventive concept, as an implementation of the above-mentioned method for identifying repetition types based on graph domain mapping and feature enhancement, this embodiment of the invention also provides a device for identifying repetition types based on graph domain mapping and feature enhancement. Figure 5 This is a structural diagram of the device in an embodiment of the present invention. See also: Figure 5 As shown, the device may include:

[0115] The filtering module 501 is used to filter multiple pulse repetition frequencies according to a preset threshold range and determine the pulse repetition frequency sequence corresponding to the filtered pulse repetition frequencies.

[0116] Processing module 502 is used to perform reciprocal processing on the pulse repetition frequency sequence determined by filtering module 501 to obtain pulse repetition interval sequence;

[0117] The segmentation module 503 is used to segment the pulse repetition interval sequence obtained by the processing module 502 according to a preset length to obtain multiple sub-sequences;

[0118] The transformation module 504 is used to transform the multiple subsequences obtained by the segmentation module 503 into corresponding Gram difference angular field images according to the graph domain mapping algorithm, and to divide the multiple Gram difference angular field images into training samples and test samples.

[0119] Module 505 is used to construct a repetition type recognition network, which includes a convolutional neural subnetwork and an attention mechanism layer.

[0120] The training module 506 is used to input the training samples divided by the transformation module 504 into the repetition type recognition network constructed by the construction module 505, train the repetition type recognition network, and obtain the trained repetition type recognition network.

[0121] The recognition module 507 is used to input the test samples divided by the transformation module 504 into the repetition frequency type recognition network trained by the training module 506, and to identify multiple repetition frequency types.

[0122] The transformation module 504 is specifically used to transform multiple subsequences into corresponding Gram difference angular field images according to the graph domain mapping algorithm, including: scaling multiple subsequences to the range of [-1,1] to obtain the scaled values ​​of each subsequence; converting each value to a polar coordinate system to obtain the angle and radius corresponding to each value; and determining the Gram difference angular field image based on the angle difference between every two values.

[0123] Module 505 is specifically used for the convolutional neural subnetwork, which includes 2 convolutional layers, 2 normalization layers, 2 activation layers, 2 pooling layers, and 2 fully connected layers.

[0124] The construction module 505 is specifically used for the frequency repetition type recognition network, which includes one convolutional layer, one normalization layer, one activation layer, one pooling layer, one attention mechanism layer, one convolutional layer, one normalization layer, one activation layer, one pooling layer, and two fully connected layers.

[0125] The recognition module 507 is specifically used to input test samples sequentially into one convolutional layer, one normalization layer, one activation layer, and one pooling layer, and output the input features corresponding to the test samples; input the input features into one attention mechanism layer and output the final feature representation; and input the final feature representation sequentially into one convolutional layer, one normalization layer, one activation layer, one pooling layer, and two fully connected layers to output multiple repetition frequency types.

[0126] The recognition module 507 is specifically used to input the input features into an attention mechanism layer and output the final feature representation, including: compressing the input features to obtain channel-level global features; activating the channel-level global features to obtain the channel weights; and multiplying the channel weights with the input features to obtain the final feature representation.

[0127] The transformation module 504 is specifically used to scale multiple subsequences to the range [-1, 1] to obtain the scaled values ​​of each subsequence, including: using the following first formula to scale multiple subsequences to the range [-1, 1] to obtain the scaled values ​​of each subsequence:

[0128]

[0129] in, For the corresponding p n The scaled value, i.e., the scaled value of the nth value in each subsequence, p n Let P be the nth value in each subsequence, where P represents multiple subsequences and N is the length of each subsequence.

[0130] The transformation module 504 is specifically used to determine the Gram difference angular field image based on the angular difference between every two values, including: determining the corresponding matrix based on the angular difference between every two values ​​and the following second formula:

[0131]

[0132] Here, GADF is the matrix corresponding to the angle between every two values. This represents the angle corresponding to the nth value in the polar coordinate system. This represents the angle corresponding to the m-th value in the polar coordinate system.

[0133] The matrix is ​​converted to grayscale to obtain the Gram difference angular field image.

[0134] It should be noted that the above description of the repetition type recognition device based on graph domain mapping and feature enhancement is similar to the description of the method embodiment above, and has similar beneficial effects. For technical details not disclosed in the embodiments of the repetition type recognition device based on graph domain mapping and feature enhancement of this invention, please refer to the description of the method embodiment of this invention for understanding.

[0135] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the computer-readable storage medium including a stored program, wherein, when the program is running, it controls the device where the storage medium is located to execute the methods in one or more of the above embodiments.

[0136] It should be noted that the descriptions of the above computer-readable storage medium embodiments are similar to those of the above method embodiments, and have similar beneficial effects. For technical details not disclosed in the embodiments of the computer-readable storage medium of this invention, please refer to the descriptions of the method embodiments of this invention for understanding.

[0137] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying repetition frequency types based on graph domain mapping and feature enhancement, characterized in that, The method includes: Multiple pulse repetition frequencies are filtered according to a preset threshold range to determine the pulse repetition frequency sequence corresponding to the filtered pulse repetition frequencies. The pulse repetition frequency sequence is reciprocated to obtain the pulse repetition interval sequence; The pulse repetition interval sequence is divided into multiple sub-sequences according to a preset length; According to the graph domain mapping algorithm, the multiple subsequences are transformed into corresponding Gram difference angle field images, and the multiple Gram difference angle field images are divided into training samples and test samples; Construct a repetition type recognition network, which includes a convolutional neural subnetwork and an attention mechanism layer; The training samples are input into the frequency repetition type recognition network to train the frequency repetition type recognition network and obtain the trained frequency repetition type recognition network. The test samples are input into the trained frequency repetition type recognition network to identify multiple frequency repetition types; The step of transforming the multiple subsequences into corresponding Gram difference angular field images according to the graph domain mapping algorithm includes: The multiple subsequences are scaled to the range [-1, 1] to obtain the scaled values ​​of each subsequence. Transform each value into a polar coordinate system to obtain the corresponding angle and radius; The Gram difference angular field image is determined based on the angular difference between every two values; The convolutional neural subnetwork includes two convolutional layers, two normalization layers, two activation layers, two pooling layers, and two fully connected layers. The frequency repetition type recognition network sequentially includes one convolutional layer, one normalization layer, one activation layer, one pooling layer, one attention mechanism layer, one convolutional layer, one normalization layer, one activation layer, one pooling layer, and two fully connected layers; the attention mechanism layer includes two operations, compression and activation, for processing input features.

2. The method according to claim 1, characterized in that, The test samples are input into the trained frequency repetition type recognition network to identify multiple frequency repetition types, including: The test sample is sequentially input into one convolutional layer, one normalization layer, one activation layer, and one pooling layer, and the input features corresponding to the test sample are output. The input features are fed into one of the attention mechanism layers, and the final feature representation is output. The final feature representation is sequentially input into one convolutional layer, one normalization layer, one activation layer, one pooling layer, and two fully connected layers to output the multiple repetition frequency types.

3. The method according to claim 2, characterized in that, The step of inputting the input features into one attention mechanism layer and outputting the final feature representation includes: The input features are compressed to obtain channel-level global features; The channel-level global features are stimulated to obtain the weights of each channel; The final feature representation is obtained by multiplying the weights of each channel by the input features.

4. The method according to claim 1, characterized in that, The step of scaling the multiple subsequences to the range [-1, 1] to obtain the scaled values ​​of each subsequence includes: Using the following first formula, the multiple subsequences are scaled to the range [-1, 1] to obtain the scaled values ​​of each subsequence: ； in, The value after scaling the nth value in each subsequence. For the first subsequence in each of the subsequences A number, For the multiple subsequences, is the length of the subsequence.

5. The method according to claim 1, characterized in that, Determining the Gram difference angular field image based on the angular difference between every two values ​​includes: Based on the angular difference between every two values ​​and the following second formula, determine the corresponding matrix: ； in, This is the matrix corresponding to the angle between each pair of values. For the first polar coordinate system The angles corresponding to each value. For the first polar coordinate system The angle corresponding to each value; The matrix is ​​converted to grayscale to obtain the Gram difference angular field image.

6. A repetition frequency type recognition device based on graph domain mapping and feature enhancement, characterized in that, The device includes: The filtering module is used to filter multiple pulse repetition frequencies according to a preset threshold range and determine the pulse repetition frequency sequence corresponding to the filtered pulse repetition frequencies. The processing module is used to take the reciprocal of the pulse repetition frequency sequence to obtain the pulse repetition interval sequence; The segmentation module is used to segment the pulse repetition interval sequence according to a preset length to obtain multiple sub-sequences; The transformation module is used to transform the multiple subsequences into corresponding Gram difference angular field images according to the graph domain mapping algorithm, and to divide the multiple Gram difference angular field images into training samples and test samples. The construction module is used to construct a repetition type recognition network, which includes a convolutional neural subnetwork and an attention mechanism layer; The training module is used to input the training samples into the repetition frequency type recognition network, train the repetition frequency type recognition network, and obtain a trained repetition frequency type recognition network. The recognition module is used to input the test sample into the trained repetition frequency type recognition network and identify multiple repetition frequency types; The transformation module is specifically used to scale the multiple subsequences to the range of [-1, 1] to obtain the scaled values ​​of each subsequence; to convert each value to a polar coordinate system to obtain the angle and radius corresponding to each value; and to determine the Gram difference angular field image based on the angle difference between every two values. The convolutional neural subnetwork includes two convolutional layers, two normalization layers, two activation layers, two pooling layers, and two fully connected layers. The frequency repetition type recognition network sequentially includes one convolutional layer, one normalization layer, one activation layer, one pooling layer, one attention mechanism layer, one convolutional layer, one normalization layer, one activation layer, one pooling layer, and two fully connected layers; the attention mechanism layer includes two operations, compression and activation, for processing input features.

7. A computer-readable storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, it controls the device where the storage medium is located to perform the repetition type identification method based on graph domain mapping and feature enhancement as described in any one of claims 1 to 5.

Images