Full pulse sampling radar sequence sorting method based on point cloud self-attention network

Abstract

The invention discloses a full-pulse sampling radar sequence sorting method based on a point cloud self-attention network, and relates to the field of electronic reconnaissance of information and communication engineering, in particular to the full-pulse sampling radar sequence sorting method. The invention provides a full-pulse sampling radar sequence sorting method based on a point cloud self-attention network in order to solve the problems that an existing method is low in signal sorting precision and poor in robustness in a complex scene. The method comprises the following steps: 1, acquiring pulse point cloud data; 2, a Point Transform sorting model is constructed, and the Point Transform sorting model is constructed; 3, obtaining a trained sorting model; 4, inputting the pulse point cloud data into the trained sorting model, and outputting a classification result of the pulse point cloud data; 5, recovering the classified point cloud data to an original time domain signal; and 6, obtaining a center frequency, a bandwidth, a pulse width, a pulse repetition interval, a pulse amplitude and a modulation type based on the time domain signal obtained in the step 5, and forming a radiation source pulse description word.

Description

Technical Field

[0001] The present invention relates to the field of electronic reconnaissance in information and communication engineering, and in particular to a full-pulse sampling radar sequence sorting method. Background Art

[0002] In modern electronic warfare and radar signal processing, with the increasing complexity of battlefield environments, highly mixed signals have become the norm in multi-source scenarios. Radar signal sorting, a core technology in electronic warfare, aims to separate pulse sequences from different emitters from a densely mixed pulse stream. This process is crucial for radar signal source identification, target tracking, and decision support. It also directly impacts the accuracy of electronic warfare systems' situational awareness and the timeliness of their response decisions, making it a key component of electronic support measures (ESM) systems.

[0003] The full pulse sampling sequence is the original signal collected by the passive radar receiver, which contains a variety of modulation types, such as linear frequency modulation (LFM), nonlinear frequency modulation (NLFM), binary phase shift keying (BPSK), Costas coding, and Frank coding. These signals have complex and diverse time-frequency structures and are widely applicable to different mission requirements. The current mainstream radar signal sorting method research is still focused on the sorting method based on pulse descriptor words (PDW). The sorting process is as follows: Figure 2 Existing PDW-based sorting methods usually ignore the fine-grained intra-pulse modulation characteristics in the original sampled signal, which makes it difficult to establish structural correlations between pulses, thereby reducing the sorting effect.

[0004] Traditional radar signal sorting methods primarily rely on key parameters contained in the pulsed pulse wave (PDW), such as arrival time, carrier frequency, pulse width, and pulse amplitude. These methods perform cluster analysis on these characteristic parameters to achieve pulse signal sorting. However, these methods rely on the integrity and accuracy of pulse parameters. If noise or missing parameters arise during the measurement phase, they struggle to provide sufficient evidence for discrimination, significantly weakening the effectiveness and stability of sorting. This reliance on the quality of these parameters creates a technical bottleneck known as "forward-backward constraints," a significant factor hindering the performance improvement of traditional sorting methods. Furthermore, these methods often overlook the rich information contained in the signal's intrapulse structure and struggle to exploit the time-frequency dynamics embodied by complex modulation patterns, limiting their adaptability to modern radar signal environments. Therefore, a new approach is urgently needed to enable more comprehensive feature analysis and sorting of radar signals, starting from the raw full-pulse data. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of low signal sorting accuracy and poor robustness of existing methods in complex scenarios, and to propose a full pulse sampling radar sequence sorting method based on point cloud self-attention network.

[0006] The specific process of the full pulse sampling radar sequence sorting method based on point cloud self-attention network is as follows:

[0007] Step 1: Get the original full-pulse sampling signal x(n), intercept the original full-pulse sampling signal x(n), and obtain the time domain signal segment x T (n); for the time domain signal segment x T (n) performing a short-time Fourier transform to obtain three-dimensional pulse point cloud data; normalizing the three-dimensional pulse point cloud data and mapping the three-dimensional pulse point cloud data to the interval [0, 1]; cropping the pulse point cloud based on the amplitude of the three-dimensional pulse point cloud data mapped to the interval [0, 1] to obtain N×3-dimensional effective pulse point cloud data, where the three dimensions are time, frequency, and amplitude, and N is the total number of time-frequency points;

[0008] Step 2: Build a PointTransformer sorting model;

[0009] Step 3: Train the constructed PointTransformer sorting model until the loss function converges to obtain the trained PointTransformer sorting model;

[0010] Step 4: Input the N×3-dimensional effective pulse point cloud data obtained in step 1 into the trained PointTransformer sorting model, and the trained PointTransformer sorting model outputs the classification result of the N×3-dimensional effective pulse point cloud data obtained in step 1;

[0011] Step 5: Restore the point cloud data classified in step 4 to the original time domain signal;

[0012] Step 6: Obtain the center frequency, bandwidth, pulse width, pulse repetition interval, pulse amplitude, and modulation type based on the time domain signal obtained in step 5;

[0013] The radiation source pulse description word is composed based on the center frequency, bandwidth, pulse width, pulse repetition interval, pulse amplitude and modulation type.

[0014] Preferably, in step 1, the original full-pulse sampling signal x(n) is obtained, and the original full-pulse sampling signal x(n) is intercepted to obtain the time domain signal segment x T (n); for the time domain signal segment x T(n) performing a short-time Fourier transform to obtain three-dimensional pulse point cloud data; normalizing the three-dimensional pulse point cloud data and mapping the three-dimensional pulse point cloud data to the interval [0, 1]; cropping the pulse point cloud based on the amplitude of the three-dimensional pulse point cloud data mapped to the interval [0, 1] to obtain N×3-dimensional effective pulse point cloud data, where the three dimensions are time, frequency, and amplitude, and N is the total number of time-frequency points;

[0015] The specific process is:

[0016] Step 11: Get the original full-pulse sampling signal x(n), intercept the original full-pulse sampling signal x(n), and obtain the time domain signal segment x T (n); the specific process is:

[0017] A window of time length T is cut from the original full pulse radar signal sequence, and the pulse radar signal sequence in the window of length T is used as the time domain signal segment x T (n);

[0018] Step 12: Use the sliding window function to convert the time domain signal segment x into T (n) is divided into M windows, and the time domain signal is Fourier transformed in each window to obtain the three-dimensional pulse point cloud data in each window; the specific process is:

[0019] Step 121: Assume that the length of each window is L and the sliding step is Δt, and convert the time domain signal segment x T (n) is divided into M windows, and the time domain signal in each window is Fourier transformed to obtain the time, frequency, amplitude and phase information after Fourier transformation; the time, frequency, amplitude and phase information after Fourier transformation is expressed as: (Time i ,Freq i ,Amp i ,φ i ),i=1,2,…,N;

[0020] Among them, Time i Indicates the i-th time in the pulse point cloud data, Freq i Indicates the i-th frequency in the pulse point cloud data, Amp i represents the i-th amplitude in the pulse point cloud data; φ i Represents the i-th phase in the pulse point cloud data; N is the total number of time-frequency points;

[0021] Step 122: For each window, only the time, frequency, and amplitude information are retained to form three-dimensional pulse point cloud data:

[0022] Points={(Time i ,Freq i ,Ampi )},i=1,2,...,N

[0023] Among them, Points represents three-dimensional pulse point cloud data;

[0024] Step 13: Normalize the 3D pulse point cloud data and map it to the interval [0, 1].

[0025] Step 14: Crop the pulse point cloud based on the amplitude in the three-dimensional pulse point cloud data mapped to the [0,1] interval to obtain N×3-dimensional effective pulse point cloud data, where the three dimensions are time, frequency, and amplitude, and N is the total number of time-frequency points.

[0026] Preferably, in step 13, the three-dimensional pulse point cloud data is normalized to map the three-dimensional pulse point cloud data to the interval [0, 1]. The specific process is:

[0027] 1) Normalize the time in the 3D pulse point cloud data, that is, map the time to the [0,1] interval; expressed as:

[0028]

[0029] Among them, x norm represents the time after normalization, x represents the time in the three-dimensional pulse point cloud data, x min Indicates the minimum time value in the three-dimensional pulse point cloud data, x max Indicates the maximum value of time in 3D pulse point cloud data;

[0030] 2) Normalize the frequency in the 3D pulse point cloud data, that is, map the frequency to the [0,1] interval; expressed as:

[0031]

[0032] Where x′ norm represents the frequency after normalization, x′ represents the frequency in the three-dimensional pulse point cloud data, and x′ min Represents the minimum value of the frequency in the three-dimensional pulse point cloud data, x′ max Indicates the maximum value of the frequency in the three-dimensional pulse point cloud data;

[0033] 3) Normalize the amplitude in the 3D pulse point cloud data, that is, map the amplitude to the [0,1] interval; expressed as:

[0034]

[0035] Among them, x″ norm represents the amplitude after normalization, x″ represents the amplitude in the three-dimensional pulse point cloud data, and x″min Indicates the minimum amplitude in the 3D pulse point cloud data, x″ max Indicates the maximum amplitude in the 3D pulse point cloud data.

[0036] Preferably, in step 14, the pulse point cloud is clipped based on the amplitude in the three-dimensional pulse point cloud data mapped to the interval [0, 1] to obtain N×3-dimensional effective pulse point cloud data, where the three dimensions are time, frequency and amplitude, and N is the total number of time-frequency points; the specific process is:

[0037] Introducing amplitude threshold Amp th ;

[0038] All pulse point clouds with amplitudes greater than the amplitude threshold are retained, and pulse point clouds with amplitudes less than or equal to the amplitude threshold are discarded to construct a valid pulse point cloud:

[0039] Points′={(Time i ,Freq i ,Amp i )|Amp i >Amp th},i=1,2,...,N

[0040] Among them, Points′ represents the valid pulse point cloud data.

[0041] Preferably, in step 2, a Point Transformer sorting model is constructed; the specific process is:

[0042] The Point Transformer sorting model includes:

[0043] First fully connected layer, first self-attention layer, first downsampling layer, second self-attention layer, second downsampling layer, third self-attention layer, third downsampling layer, fourth self-attention layer, fourth downsampling layer, fifth self-attention layer, second fully connected layer, sixth self-attention layer, first upsampling layer, seventh self-attention layer, second upsampling layer, eighth self-attention layer, third upsampling layer, ninth self-attention layer, fourth upsampling layer, tenth self-attention layer, third fully connected layer, Softmax layer;

[0044] The working process of the PointTransformer sorting model is:

[0045] Feature A is sequentially input into the first fully connected layer and the first self-attention layer, and the first self-attention layer outputs feature B;

[0046] Feature B is sequentially input into the first downsampling layer and the second self-attention layer, and the second self-attention layer outputs feature C;

[0047] Feature C is sequentially input into the second downsampling layer and the third self-attention layer, and the third self-attention layer outputs feature D;

[0048] Feature D is sequentially input into the third downsampling layer and the fourth self-attention layer, and the fourth self-attention layer outputs feature E;

[0049] Feature E is input into the fourth downsampling layer and the fifth self-attention layer in sequence, and the fifth self-attention layer outputs feature F;

[0050] Feature F is sequentially input into the second fully connected layer and the sixth self-attention layer, and the sixth self-attention layer outputs feature G;

[0051] Feature G and feature E are added element by element and then input into the first upsampling layer and the seventh self-attention layer in sequence. The seventh self-attention layer outputs feature H.

[0052] Feature H and feature D are added element by element and then input into the second upsampling layer and the eighth self-attention layer in sequence. The eighth self-attention layer outputs feature I.

[0053] Feature I and feature C are added element by element and then input into the third upsampling layer and the ninth self-attention layer in sequence. The ninth self-attention layer outputs feature J.

[0054] Feature J and feature B are added element by element and then input into the fourth upsampling layer and the tenth self-attention layer in sequence. The tenth self-attention layer outputs feature K.

[0055] Feature K is sequentially input into the third fully connected layer and the Softmax layer, and the Softmax layer outputs the semantic label of each pulse point cloud.

[0056] Preferably, in step 3, the constructed Point Transformer sorting model is trained until the loss function converges to obtain a trained Point Transformer sorting model; the specific process is:

[0057] Step 31: Each point in the valid pulse point cloud data obtained in step 1 corresponds to a pulse signal category, and pulse signal category labels are added to M points to obtain M pulse point cloud data sets with pulse signal category labels;

[0058] Step 32: The M pulse point cloud data with pulse signal category labels obtained in step 31 are upgraded to 32 dimensions through linear transformation. The pulse point cloud data after linear transformation are sequentially subjected to BatchNorm and ReLU, and ReLU outputs features.

[0059] The features are input into the Point Transformer sorting model, which then outputs the semantic label of each pulse point cloud.

[0060] Step 33: Repeat step 32 until the loss function L converges to obtain the trained Point Transformer sorting model;

[0061] The loss function L is:

[0062] L=L WCE +L reg

[0063] L WCE is the adaptive weighted cross entropy loss;

[0064] L reg is the L2 regularization constraint.

[0065] Preferably, the adaptive weighted cross entropy loss L WCE Expressed as:

[0066]

[0067] Among them, M is the number of points in the point cloud, and C is the number of categories;

[0068] y i,k is the true category of the pulse point cloud, p i,k The category probability predicted by the PointTransformer classification model;

[0069] ω i is the weight; α is the hyperparameter;

[0070] The L2 regularization constraint L reg Expressed as:

[0071]

[0072] Among them, λ is the regularization hyperparameter;

[0073] θ is a parameter in the PointTransformer sorting model.

[0074] Preferably, in step 5, the point cloud data classified in step 4 is restored to the original time domain signal; the specific process is:

[0075] Step 51: Denormalize the time, frequency, and amplitude dimensions of the point cloud data classified in step 4 to obtain denormalized point cloud data.

[0076] Step 52: Supplement the missing amplitude values at the time-frequency position in the point cloud data obtained after the denormalization processing in step 51 with 0;

[0077] Step 53: reconstruct the time-frequency matrix based on step 52;

[0078] Step 54: Perform an inverse short-time Fourier transform (ISTFT) on the time-frequency matrix reconstructed in step 53 to restore the reconstructed time-frequency matrix to a time-domain signal.

[0079] Preferably, in step 51, the time, frequency and amplitude dimensions of the point cloud data classified in step 4 are respectively denormalized to obtain denormalized point cloud data; the specific process is:

[0080] 1) Denormalize the time in the point cloud data classified in step 4 to obtain the denormalized time; it is expressed as:

[0081] y rec =x norm (x max -x min )+x min

[0082] Among them, y rec is the time after denormalization; x norm is the time in the point cloud data after classification in step 4, x max and x min are the minimum and maximum values of the time in step 12 respectively;

[0083] 2) Denormalize the frequency of the point cloud data classified in step 4 to obtain the denormalized frequency; it is expressed as:

[0084] y′ rec =x′ norm (x′ max -x′ min )+x′ min

[0085] Among them, y′ rec is the frequency after denormalization; x′ norm is the frequency in the point cloud data after classification in step 4, x′ max and x′ min are the minimum and maximum values of the frequency of step 12 respectively;

[0086] 3) Denormalize the amplitude of the point cloud data classified in step 4 to obtain the denormalized amplitude; it is expressed as:

[0087] y″ rec =x″ norm (x″ max -x″ min )+x″ min

[0088] Among them, y recis the amplitude after denormalization; x″ norm is the amplitude of the point cloud data after classification in step 4, x″ max and x″ min are the minimum and maximum values of the step 12 amplitude respectively.

[0089] Preferably, in step 53, the time-frequency matrix is reconstructed based on step 52; the specific process is:

[0090] For each time-frequency point, according to all the amplitudes Amp rec Phase φ with step 12 i , synthesize the corresponding time-frequency matrix:

[0091] S rec (n,f)=Amp rec (n,f)e j×φ(n,f)

[0092] Among them, S rec (n,f) represents the complex time-frequency matrix value at time point n and frequency f, where n represents the time point and f represents the frequency;

[0093] Amp rec (n,f) represents the amplitude corresponding to time point n and frequency f;

[0094] φ(n,f) represents the phase information obtained in step 12;

[0095] j represents the imaginary unit, j 2 =-1.

[0096] The beneficial effects of the present invention are:

[0097] This paper proposes an end-to-end radar full-pulse signal sorting method based on a point cloud self-attention segmentation network. This method breaks through the limitations of the traditional PDW framework and directly processes the full-pulse sampling sequence. First, the short-time Fourier transform (SFT) is used to map the one-dimensional time-domain signal into a three-dimensional time-frequency-amplitude point cloud, preserving the signal's modulation information and time-frequency evolution characteristics. Subsequently, a point cloud segmentation network based on a self-attention mechanism is introduced to automatically model and sort the intra-pulse structure without relying on manual feature engineering, significantly improving signal sorting accuracy and robustness in complex scenarios.

[0098] The present invention proposes an end-to-end full-pulse sampling radar signal sorting method based on the self-attention mechanism, which abandons the dependence on intermediate feature extraction and artificial priors, and directly uses the full-pulse sampling time domain signal as input. By constructing a deep learning framework that integrates the point cloud segmentation network and the self-attention mechanism, it realizes end-to-end automatic sorting of the radar full-pulse sampling sequence, and solves the technical bottleneck problem of "forward and backward constraints" brought about by the traditional radar signal sorting method that relies on PDW in feature representation ability, robustness and strong dependence on early parameter measurement.

[0099] This method first converts the radar full-pulse signal to construct a time-frequency-amplitude point cloud dataset. After acquiring the original radar signal, a fixed-time segment of the full-pulse sampling sequence is captured, and the time-frequency data is extracted using a short-time Fourier transform. The converted data is then converted into a point cloud based on time, frequency, and amplitude. The pulse point cloud data is then normalized and clipped using an amplitude threshold to ensure that the network's point count requirements are met and reduce the computational burden.

[0100] PointTransformer uses an attention mechanism to learn the relationships between spike data points, enabling efficient feature extraction. The model input is 3D point cloud data (time, frequency, and amplitude). The input point cloud data first undergoes preliminary feature mapping through an MLP layer to improve data separability, before entering PointTransformer for deep feature extraction. The network employs an encoder-decoder architecture, where the encoder extracts multi-scale features and the decoder restores the original point cloud resolution and performs segmentation predictions.

[0101] To improve the performance of the PointTransformer in radar full-pulse signal sorting, an optimization objective function combined with an attention mechanism was introduced. During network training, a cross-entropy loss function was used to optimize model parameters, and regularization was introduced to prevent overfitting. Furthermore, to further enhance sorting accuracy, an adaptive weighting strategy was employed in the optimization objective function to increase attention to pulse point clouds with severe aliasing.

[0102] After model training is complete, the real-time radar full pulse signal can be sorted and the original pulse sequence reconstructed based on the classification results. First, the sorted point cloud data is denormalized to restore it to its original time, frequency, and amplitude range. Then, based on the sorting results, the pulse point cloud data is separated, and the time-frequency matrix is reconstructed by combining the amplitude and phase information. The time domain signal is then restored using an inverse short-time Fourier transform. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] Figure 1 This is a block diagram of the online sorting method for radar full pulse signals based on point cloud self-attention segmentation network;

[0104] Figure 2 It is a schematic diagram of the sorting process based on pulse description words (PDW);

[0105] Figure 3 This is a schematic diagram of the sequence sorting process of the full pulse sampling radar of the present invention;

[0106] Figure 4 It is a visualization diagram of radar full pulse signal point cloud data;

[0107] Figure 5 This is a schematic diagram of extracting time-frequency information of full pulse signals;

[0108] Figure 6 This is a schematic diagram of the PointTransformer depth segmentation process;

[0109] Figure 7 is the result of full pulse sequence reconstruction;

[0110] Figure 8 This is a visualization of the full-pulse radar signal point cloud prediction results;

[0111] Figure 9 These are the sorting results at different signal-to-noise ratios. DETAILED DESCRIPTION

[0112] Specific implementation method 1: Combination Figure 1 and Figure 3 This embodiment describes the specific process of the full pulse sampling radar sequence sorting method based on the point cloud self-attention network:

[0113] Step 1: Get the original full-pulse sampling signal x(n), intercept the original full-pulse sampling signal x(n), and obtain the time domain signal segment x T (n); for the time domain signal segment x T (n) performing short-time Fourier transform to obtain three-dimensional pulse point cloud data; performing normalization processing on the three-dimensional pulse point cloud data and mapping the three-dimensional pulse point cloud data to the interval [0,1]; cropping the pulse point cloud based on the amplitude in the three-dimensional pulse point cloud data mapped to the interval [0,1] to obtain N×3-dimensional effective pulse point cloud data, where the three dimensions are time, frequency, and amplitude, and N is the total number of time-frequency points; Figure 4 As shown;

[0114] Step 2: Build a Point Transformer sorting model;

[0115] Step 3: Train the constructed Point Transformer sorting model until the loss function converges to obtain the trained Point Transformer sorting model;

[0116] Step 4: Input the N×3-dimensional effective pulse point cloud data (without labels) obtained in step 1 into the trained Point Transformer sorting model, and the trained Point Transformer sorting model outputs the classification result of the N×3-dimensional effective pulse point cloud data obtained in step 1;

[0117] Step 5: Restore the point cloud data classified in step 4 to the original time domain signal;

[0118] Step 6: Obtain the center frequency, bandwidth, pulse width, pulse repetition interval, pulse amplitude, and modulation type based on the time domain signal obtained in step 5;

[0119] The emitter description word (EDW) is composed of the center frequency, bandwidth, pulse width, pulse repetition interval, pulse amplitude and modulation type.

[0120] For each selected radiation source, the pulse descriptor corresponding to the radiation source is counted. The EDW consists of the following key parameters, as shown in Table 2:

[0121] Table 2 Radar radiation source parameter information

[0122]

[0123] Compared with the prior art, the present invention has the following beneficial effects:

[0124] (1) The present invention adopts an end-to-end deep learning framework, from the original full-pulse sampling sequence input to the final sorting result output, that is, from the broadband intermediate frequency signal input to the EDW information output, without the need for manually designed feature extraction steps, which significantly improves the sorting efficiency.

[0125] (2) The present invention introduces a self-attention mechanism, which can fully capture the local and global features of the full pulse sampling sequence data. The self-attention mechanism can dynamically adjust the feature weights, significantly enhancing the feature extraction capability, thereby improving the sorting accuracy.

[0126] (3) The proposed full pulse sorting system can adaptively mine key features in the time-frequency space, significantly improving signal sorting accuracy in complex electromagnetic environments. By converting the radar signal's time-frequency information into a point cloud representation, it accurately characterizes the fine-grained features between radar pulses, effectively addressing the strong parameter dependence of traditional methods.

[0127] Specific embodiment 2: This embodiment differs from the specific embodiment 1 in that in step 1, the original full pulse sampling signal x(n) is obtained, the original full pulse sampling signal x(n) is intercepted, and the time domain signal segment x is obtained. T(n); for the time domain signal segment x T (n) performing a short-time Fourier transform to obtain three-dimensional pulse point cloud data; normalizing the three-dimensional pulse point cloud data and mapping the three-dimensional pulse point cloud data to the interval [0, 1]; cropping the pulse point cloud based on the amplitude of the three-dimensional pulse point cloud data mapped to the interval [0, 1] to obtain N×3-dimensional effective pulse point cloud data, where the three dimensions are time, frequency, and amplitude, and N is the total number of time-frequency points;

[0128] The specific process is:

[0129] Step 11: Obtain the original full-pulse sampling signal x(n) through the passive radar receiving system, intercept the original full-pulse sampling signal x(n), and obtain the time domain signal segment x T (n); the specific process is:

[0130] A window with a fixed time length of T is cut from the original full pulse radar signal sequence, and the pulse radar signal sequence in the window with a length of T is used as the time domain signal segment x T (n);

[0131] Step 12: Figure 5 As shown, the time domain signal segment x is transformed into T (n) is divided into M windows, and the time domain signal is Fourier transformed in each window to obtain the three-dimensional pulse point cloud data in each window; the specific process is:

[0132] Step 121: Assume that the length of each window is L and the sliding step is Δt, and convert the time domain signal segment x T (n) is divided into M windows, and the time domain signal in each window is Fourier transformed to obtain the time, frequency, amplitude and phase information after Fourier transformation; the time, frequency, amplitude and phase information after Fourier transformation is expressed as: (Time i ,Freq i ,Amp i ,φ i ),i=1,2,...,N;

[0133] Among them, Time i Indicates the i-th time in the pulse point cloud data, Freq i Indicates the i-th frequency in the pulse point cloud data, Amp i represents the i-th amplitude in the pulse point cloud data; φ i Represents the i-th phase in the pulse point cloud data; N is the total number of time-frequency points;

[0134] The time of each window is the center of the time window; the frequency is composed of multiple frequency components obtained by Fourier transforming the time domain signal in the time window; the amplitude is the amplitude corresponding to the frequency component; and the phase is the phase information corresponding to the frequency component.

[0135] Step 122: For each window, only the time, frequency, and amplitude information are retained to form three-dimensional pulse point cloud data:

[0136] Points={(Time i ,Freq i ,Amp i )},i=1,2,...,N

[0137] Among them, Points represents three-dimensional pulse point cloud data;

[0138] Step 13: Normalize the 3D pulse point cloud data and map it to the interval [0, 1].

[0139] Step 14: Crop the pulse point cloud based on the amplitude in the three-dimensional pulse point cloud data mapped to the [0,1] interval to obtain N×3-dimensional effective pulse point cloud data, where the three dimensions are time, frequency, and amplitude, and N is the total number of time-frequency points.

[0140] Other steps and parameters are the same as those in the first embodiment.

[0141] Specific embodiment three: This embodiment differs from specific embodiment one or two in that in step 13, the three-dimensional pulse point cloud data is normalized and mapped to the interval [0, 1]. The specific process is as follows:

[0142] Since the data in the three dimensions of time, frequency and amplitude differ significantly in order of magnitude, if they are not processed, it will lead to problems such as gradient offset and feature imbalance during model training, affecting the network's learning ability and convergence efficiency.

[0143] 1) Normalize the time in the 3D pulse point cloud data, that is, map the time to the [0,1] interval; expressed as:

[0144]

[0145] Among them, x norm represents the time after normalization, x represents the time in the three-dimensional pulse point cloud data, x min Indicates the minimum time value in the three-dimensional pulse point cloud data, x max Indicates the maximum value of time in 3D pulse point cloud data;

[0146] 2) Normalize the frequency in the 3D pulse point cloud data, that is, map the frequency to the [0,1] interval; expressed as:

[0147]

[0148] Where x′ norm represents the frequency after normalization, x′ represents the frequency in the three-dimensional pulse point cloud data, and x′ min Represents the minimum value of the frequency in the three-dimensional pulse point cloud data, x′ max Indicates the maximum value of the frequency in the three-dimensional pulse point cloud data;

[0149] 3) Normalize the amplitude in the 3D pulse point cloud data, that is, map the amplitude to the [0,1] interval; expressed as:

[0150]

[0151] Where x″ norm represents the amplitude after normalization, x″ represents the amplitude in the three-dimensional pulse point cloud data, and x″ min Indicates the minimum amplitude in the 3D pulse point cloud data, x″ max Indicates the maximum amplitude in the 3D pulse point cloud data.

[0152] Other steps and parameters are the same as those in the first or second embodiment.

[0153] Specific embodiment 4: This embodiment differs from any one of specific embodiments 1 to 3 in that, in step 14, the pulse point cloud is clipped based on the amplitude in the three-dimensional pulse point cloud data mapped to the [0, 1] interval to obtain N×3-dimensional effective pulse point cloud data, where the 3 dimensions are time, frequency, and amplitude, and N is the total number of time-frequency points. The specific process is as follows:

[0154] Since STFT transforms the full time domain signal, spectral points will be generated even in the time interval when there is no pulse signal, resulting in part of the point cloud consisting only of background noise, which reduces the accuracy and efficiency of the processing. In order to reduce the interference caused by redundant data, the point cloud data needs to be cropped. This step introduces the amplitude threshold Amp th , to filter low-amplitude point cloud data; retain all pulse point clouds with amplitudes greater than the amplitude threshold, and discard pulse point clouds with amplitudes less than or equal to the amplitude threshold, thereby constructing a valid pulse point cloud with high confidence and clear structure:

[0155] Points′={(Time i ,Freq i ,Amp i )|Amp i >Amp th},i=1,2,...,N

[0156] Among them, Points′ represents the valid pulse point cloud data.

[0157] The other steps and parameters are the same as those in the first to third embodiments.

[0158] Specific embodiment 5: This embodiment differs from specific embodiments 1 to 4 in that a Point Transformer sorting model is constructed in step 2; the specific process is as follows:

[0159] The Point Transformer sorting model includes:

[0160] First fully connected layer, first self-attention layer, first downsampling layer, second self-attention layer, second downsampling layer, third self-attention layer, third downsampling layer, fourth self-attention layer, fourth downsampling layer, fifth self-attention layer, second fully connected layer, sixth self-attention layer, first upsampling layer, seventh self-attention layer, second upsampling layer, eighth self-attention layer, third upsampling layer, ninth self-attention layer, fourth upsampling layer, tenth self-attention layer, third fully connected layer, Softmax layer;

[0161] The working process of the Point Transformer sorting model is:

[0162] Feature A is sequentially input into the first fully connected layer and the first self-attention layer, and the first self-attention layer outputs feature B;

[0163] Feature B is sequentially input into the first downsampling layer and the second self-attention layer, and the second self-attention layer outputs feature C;

[0164] Feature C is sequentially input into the second downsampling layer and the third self-attention layer, and the third self-attention layer outputs feature D;

[0165] Feature D is sequentially input into the third downsampling layer and the fourth self-attention layer, and the fourth self-attention layer outputs feature E;

[0166] Feature E is input into the fourth downsampling layer and the fifth self-attention layer in sequence, and the fifth self-attention layer outputs feature F;

[0167] Feature F is sequentially input into the second fully connected layer and the sixth self-attention layer, and the sixth self-attention layer outputs feature G;

[0168] Feature G and feature E are added element by element and then input into the first upsampling layer and the seventh self-attention layer in sequence. The seventh self-attention layer outputs feature H.

[0169] Feature H and feature D are added element by element and then input into the second upsampling layer and the eighth self-attention layer in sequence. The eighth self-attention layer outputs feature I.

[0170] Feature I and feature C are added element by element and then input into the third upsampling layer and the ninth self-attention layer in sequence. The ninth self-attention layer outputs feature J.

[0171] Feature J and feature B are added element by element and then input into the fourth upsampling layer and the tenth self-attention layer in sequence. The tenth self-attention layer outputs feature K.

[0172] Feature K is sequentially input into the third fully connected layer and the Softmax layer, and the Softmax layer outputs the semantic label of each pulse point cloud.

[0173] The Point Transformer deep sorting model uses a self-attention mechanism to perform deep segmentation on the valid spike point cloud data input in step 1, enabling the Point Transformer to effectively capture both local and global features of the spike point cloud data. This module calculates attention weights within a local neighborhood through the self-attention mechanism and combines it with relative position encoding to enhance the perception of the structure of the spike point cloud data.

[0174] The specific depth sorting process based on Point Transformer is as follows Figure 6 As shown. The Point Transformer sorting model uses an encoding-decoding structure to segment the full pulse point cloud data. The encoder uses four-level downsampling, gradually reducing the number of points (N / 4, N / 16, N / 64, N / 256) through the farthest point sampling (FPS), and uses KNN (K=16) to construct a local neighborhood, combined with a shared MLP (64→128→256→512) to extract features. After each level of downsampling, a self-attention layer is introduced to calculate the attention weights of the neighborhood points and dynamically adjust the feature aggregation to enhance the feature discrimination ability. The decoder gradually restores the point cloud resolution through four-level upsampling, expands the number of points using trilinear interpolation, and performs jump connections with the corresponding layers of the encoder to fuse fine-grained information. After upsampling, the features are optimized through the MLP and self-attention layers, and finally classified through the fully connected layer and Softmax to output the semantic label of each pulse point cloud.

[0175] The other steps and parameters are the same as those in the first to fourth embodiments.

[0176] Specific embodiment 6: This embodiment differs from any one of specific embodiments 1 to 5 in that in step 3, the constructed Point Transformer sorting model is trained until the loss function converges to obtain a trained Point Transformer sorting model; the specific process is as follows:

[0177] Step 31: Each point in the valid pulse point cloud data obtained in step 1 corresponds to a pulse signal category, and pulse signal category labels are added to M points to obtain M pulse point cloud data sets with pulse signal category labels;

[0178] Step 32: The M pulse point cloud data (time, frequency, amplitude) with pulse signal category labels obtained in step 31 are upgraded to 32 dimensions through linear transformation. The linearly transformed pulse point cloud data are sequentially subjected to BatchNorm and ReLU, and ReLU outputs features.

[0179] The features are input into the PointTransformer sorting model, which then outputs the semantic label of each pulse point cloud.

[0180] Step 33: Repeat step 32 until the loss function L converges to obtain the trained PointTransformer sorting model;

[0181] The loss function L is:

[0182] L=L WCE +L reg

[0183] L WCE is the adaptive weighted cross entropy loss;

[0184] L reg is the L2 regularization constraint.

[0185] The other steps and parameters are the same as those in the first to fifth embodiments.

[0186] Specific embodiment 7: This embodiment is different from any one of the specific embodiments 1 to 6 in that the adaptive weighted cross entropy loss L WCE Expressed as:

[0187] The model uses the cross entropy loss function as the basic optimization objective function:

[0188]

[0189] Among them, M is the number of points in the point cloud, C is the number of categories; y i,k is the true category of the pulse point cloud, p i,k is the predicted class probability;

[0190] In order to increase the attention to the pulse point cloud with severe aliasing, the adaptive weighted cross entropy loss L WCE :

[0191]

[0192] Among them, M is the number of points in the point cloud, and C is the number of categories;

[0193] y i,k is the true category of the pulse point cloud, p i,k The category probability predicted by the PointTransformer classification model;

[0194] ω i is the weight, which is dynamically adjusted according to the classification difficulty of the point cloud. When the predicted entropy of the pulse point cloud is high, it means that the classification uncertainty is large, and the weight will be increased accordingly, making the difficult-to-classify pulse point cloud contribute more to the loss function;

[0195] α is a hyperparameter used to balance the influence of easy-to-classify and difficult-to-classify point clouds;

[0196] The L2 regularization constraint L reg Expressed as:

[0197] To prevent the model from overfitting, add L2 regularization constraints;

[0198]

[0199] Where λ is a regularization hyperparameter that controls the strength of the regularization term. If λ is large, the model parameters will be smaller; if λ is small, the regularization effect will be weakened.

[0200] θ is a parameter in the Point Transformer sorting model.

[0201] Table 1 Model hyperparameter settings

[0202]

[0203] Each point in the pulse point cloud data corresponds to a pulse signal category, requiring a category label. The label column is appended to the point cloud data matrix to obtain the final pulse point cloud dataset. The pulse point cloud dataset is divided into training, validation, and test sets in a ratio of 8:1:1. The hyperparameter settings for the model during training are shown in Table 1.

[0204] The model training hyperparameters set a batch size of 16 and a total number of training epochs of 200. Adam was used as the optimizer, with an initial learning rate of 0.001 and a regularization hyperparameter of λ of 0.0001. Weight decay was introduced to effectively prevent overfitting and improve the model's generalization. As training progressed, the learning rate was periodically reduced to achieve more refined parameter updates in the later stages of training, thereby improving the model's final performance while avoiding oscillations or jumps caused by a large learning rate. Therefore, the learning rate was updated every 20 epochs with a decay factor of 0.5. When the loss function converged, the trained model was saved for online sorting.

[0205] The other steps and parameters are the same as those in the first to sixth embodiments.

[0206] Specific embodiment eight: This embodiment differs from any one of specific embodiments one to seven in that, in step 5, the point cloud data classified in step 4 is restored to the original time domain signal; the specific process is as follows:

[0207] The main task of the full pulse sequence reconstruction module is to restore the classified point cloud data to the original time domain signal. First, the point cloud coordinates in the classification result are denormalized to restore them to the original time, frequency, and amplitude value ranges. According to the classification results, the point cloud data is associated with the original pulse signal, and the frequency domain signal is converted into a time domain signal through the inverse short-time Fourier transform. Finally, the complete pulse sequence is restored. The result is as follows Figure 7 Finally, the separated pulse signals are statistically analyzed. The specific implementation steps are as follows:

[0208] Step 51: Denormalize the time, frequency, and amplitude dimensions of the point cloud data classified in step 4 to obtain denormalized point cloud data.

[0209] Step 52: Supplement the missing amplitude values at the time-frequency positions in the point cloud data obtained after the denormalization processing in step 51 with 0; and retain the corresponding time and frequency indexes;

[0210] Because the point cloud is cropped based on the amplitude threshold in step 14, some points with low amplitude or background noise areas are removed. Although this improves data quality, it also leads to sparsity in the time-frequency distribution of the point cloud, which in turn affects the integrity of the time-frequency matrix. To ensure the consistency of the input matrix dimensions in the inverse short-time Fourier transform (ISTFT) operation, missing points need to be interpolated and filled. This solution adopts a simple and effective zero-filling strategy, which supplements the missing time-frequency positions after cropping with points with an amplitude value of 0 and retains the corresponding time and frequency indexes.

[0211] Step 53: reconstruct the time-frequency matrix based on step 52;

[0212] Step 54: Perform an inverse short-time Fourier transform (ISTFT) on the time-frequency matrix reconstructed in step 53 to restore the reconstructed time-frequency matrix to a time-domain signal.

[0213] ISTFT performs an inverse Fourier transform on the spectrum of each time window and restores the complete time series through a windowing and overlap-summing strategy.

[0214] The other steps and parameters are the same as those in the first to seventh embodiments.

[0215] Specific embodiment 9: This embodiment differs from any one of specific embodiments 1 to 8 in that, in step 51, the time, frequency, and amplitude dimensions of the point cloud data classified in step 4 are respectively denormalized to obtain denormalized point cloud data;

[0216] The specific process is:

[0217] 1) Denormalize the time in the point cloud data classified in step 4 to obtain the denormalized time; it is expressed as:

[0218] y rec =x norm (x max -x min )+x min

[0219] Among them, y rec is the time after denormalization; x norm is the time in the point cloud data after classification in step 4, x max and x min are the minimum and maximum values of the time in step 12 respectively;

[0220] 2) Denormalize the frequency of the point cloud data classified in step 4 to obtain the denormalized frequency; it is expressed as:

[0221] y′ rec =x′ norm (x′ max -x′ min )+x′ min

[0222] Among them, y′ rec is the frequency after denormalization; x′ norm is the frequency in the point cloud data after classification in step 4, x′ max and x′ min are the minimum and maximum values of the frequency of step 12 respectively;

[0223] 3) Denormalize the amplitude of the point cloud data classified in step 4 to obtain the denormalized amplitude; it is expressed as:

[0224] y″ rec =x″ norm (x″ max -x″ min )+x″ min

[0225] Among them, y rec is the amplitude after denormalization; x″ norm is the amplitude of the point cloud data after classification in step 4, x″ max and x″ min are the minimum and maximum values of the step 12 amplitude respectively.

[0226] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0227] Specific embodiment ten: This embodiment differs from any one of specific embodiments one to nine in that the time-frequency matrix is reconstructed in step 53 based on step 52; the specific process is as follows:

[0228] The short-time Fourier transform essentially outputs complex spectral coefficients, which contain both the amplitude and phase information of the signal. During the pulse sampling point cloud conversion process, only the amplitude information is retained for the point cloud input, while the phase information is temporarily stored. To reconstruct the time-domain signal, the complete complex time-frequency matrix must be reconstructed.

[0229] For each time-frequency point, according to all the amplitudes Amp rec (including all amplitudes after zero padding) and the phase φ of step 12 i , synthesize the corresponding time-frequency matrix:

[0230] S rec (n,f)=Amp rec (n,f)e j·φ(n,f)

[0231] Among them, S rec (n,f) represents the complex time-frequency matrix value at time point n and frequency f, where n represents the time point and f represents the frequency;

[0232] Amp rec (n,f) represents the amplitude corresponding to time point n and frequency f (including the amplitude after zero padding);

[0233] φ(n,f) represents the phase information obtained in step 12;

[0234] j represents the imaginary unit, j 2 =-1.

[0235] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0236] The effectiveness of the present invention is verified as follows:

[0237] Figure 8 The following table shows the visualization of the Point Transformer network output when using this method for signal sorting. Four radiation sources were selected in the experiment, two of which used LFM intra-pulse modulation, and the other two used NLFM and Barker coding, respectively. Specific parameters are shown in Table 3. The visualization shows that the labels output by the network accurately distinguish between signals from different radiation sources, validating the feasibility of the Point Transformer for signal sorting tasks.

[0238] Based on the PointTransformer network, a simulation sorting experiment was conducted on multiple full-pulse radar signals (2, 4, 6, and 8), and its sorting performance was evaluated under different signal-to-noise ratio conditions. The experimental results are as follows: Figure 9 As shown. The analysis results show that as the number of radiation sources increases, the distribution of radar signal characteristic parameters in the feature space gradually becomes intertwined and overlapping, and the degree of parameter aliasing increases significantly, resulting in an increase in the complexity of the sorting task and a decrease in sorting accuracy. Despite this, PointTransformer exhibits strong feature learning ability and sorting robustness under different signal-to-noise ratio conditions: when the signal-to-noise ratio is 20dB, the sorting accuracy can reach 99.3%; even when the aliasing is severe and the signal-to-noise ratio is reduced to -10dB, the sorting accuracy can still be maintained at above 85%. This result fully verifies the effectiveness and robustness of PointTransformer in the full-pulse radar signal sorting task.

[0239] Table 3 Radar radiation source parameter settings

[0240]

[0241] The present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A full-pulse sampling radar sequence sorting method based on point cloud self-attention network, characterized by: The specific process of the method is: Step 1: Get the original full-pulse sampling signal x(n), intercept the original full-pulse sampling signal x(n), and obtain the time domain signal segment x T (n); for the time domain signal segment x T (n) performing short-time Fourier transform to obtain three-dimensional pulse point cloud data; Normalize the 3D pulse point cloud data and map it to the [0,1] interval; The pulse point cloud is cropped based on the amplitude in the three-dimensional pulse point cloud data mapped to the [0,1] interval to obtain N×3-dimensional effective pulse point cloud data, where the three dimensions are time, frequency, and amplitude, and N is the total number of time-frequency points; Step 2: Build a PointTransformer sorting model; Step 3: Train the constructed PointTransformer sorting model until the loss function converges to obtain the trained PointTransformer sorting model; Step 4: Input the N×3-dimensional effective pulse point cloud data obtained in step 1 into the trained Point Transformer sorting model, and the trained Point Transformer sorting model outputs the classification result of the N×3-dimensional effective pulse point cloud data obtained in step 1; Step 5: Restore the point cloud data classified in step 4 to the original time domain signal; Step 6: Obtain the center frequency, bandwidth, pulse width, pulse repetition interval, pulse amplitude, and modulation type based on the time domain signal obtained in step 5; The radiation source pulse description word is composed based on the center frequency, bandwidth, pulse width, pulse repetition interval, pulse amplitude and modulation type.

2. The full pulse sampling radar sequence sorting method based on point cloud self-attention network according to claim 1 is characterized by: In step 1, the original full-pulse sampling signal x(n) is obtained, and the original full-pulse sampling signal x(n) is intercepted to obtain the time domain signal segment x T (n); for the time domain signal segment x T (n) performing short-time Fourier transform to obtain three-dimensional pulse point cloud data; Normalize the 3D pulse point cloud data and map it to the [0,1] interval; The pulse point cloud is cropped based on the amplitude in the three-dimensional pulse point cloud data mapped to the [0,1] interval to obtain N×3-dimensional effective pulse point cloud data, where the three dimensions are time, frequency, and amplitude, and N is the total number of time-frequency points; The specific process is: Step 11: Get the original full-pulse sampling signal x(n), intercept the original full-pulse sampling signal x(n), and obtain the time domain signal segment x T (n); the specific process is: A window of time length T is cut from the original full pulse radar signal sequence, and the pulse radar signal sequence in the window of length T is used as the time domain signal segment x T (n); Step 12: Use the sliding window function to convert the time domain signal segment x into T (n) is divided into M windows, and the time domain signal is Fourier transformed in each window to obtain the three-dimensional pulse point cloud data in each window; The specific process is: Step 121: Assume that the length of each window is L and the sliding step is Δt, and convert the time domain signal segment x T (n) is divided into M windows, and the time domain signal in each window is Fourier transformed to obtain the time, frequency, amplitude and phase information after Fourier transformation; the time, frequency, amplitude and phase information after Fourier transformation is expressed as: (Time i ,Freq i ,Amp i ,φ i ),i=1,2,...,N; Among them, Time i Indicates the i-th time in the pulse point cloud data, Freq i Indicates the i-th frequency in the pulse point cloud data, Amp i represents the i-th amplitude in the pulse point cloud data; φ i Represents the i-th phase in the pulse point cloud data; N is the total number of time-frequency points; Step 122: For each window, only the time, frequency, and amplitude information are retained to form three-dimensional pulse point cloud data: Points={(Time i ,Freq i ,Amp i )},i=1,2,…,N Among them, Points represents three-dimensional pulse point cloud data; Step 13: Normalize the 3D pulse point cloud data and map it to the interval [0, 1]. Step 14: Crop the pulse point cloud based on the amplitude in the three-dimensional pulse point cloud data mapped to the [0,1] interval to obtain N×3-dimensional effective pulse point cloud data, where the three dimensions are time, frequency, and amplitude, and N is the total number of time-frequency points.

3. The full pulse sampling radar sequence sorting method based on point cloud self-attention network according to claim 2 is characterized by: In step 13, the three-dimensional pulse point cloud data is normalized to map the three-dimensional pulse point cloud data to the interval [0, 1]. The specific process is: 1) Normalize the time in the 3D pulse point cloud data, that is, map the time to the [0,1] interval; expressed as: Among them, x norm represents the time after normalization, x represents the time in the three-dimensional pulse point cloud data, x min Indicates the minimum time value in the three-dimensional pulse point cloud data, x max Indicates the maximum value of time in 3D pulse point cloud data; 2) Normalize the frequency in the 3D pulse point cloud data, that is, map the frequency to the [0,1] interval; expressed as: Where x′ norm represents the frequency after normalization, x′ represents the frequency in the three-dimensional pulse point cloud data, and x′ min Represents the minimum value of the frequency in the three-dimensional pulse point cloud data, x′ max Indicates the maximum value of the frequency in the three-dimensional pulse point cloud data; 3) Normalize the amplitude in the 3D pulse point cloud data, that is, map the amplitude to the [0,1] interval; expressed as: Among them, x″ norm represents the amplitude after normalization, x″ represents the amplitude in the three-dimensional pulse point cloud data, and x″ min Indicates the minimum amplitude in the 3D pulse point cloud data, x″ max Indicates the maximum amplitude in the 3D pulse point cloud data.

4. The full pulse sampling radar sequence sorting method based on point cloud self-attention network according to claim 3 is characterized by: In step 14, the pulse point cloud is clipped based on the amplitude in the three-dimensional pulse point cloud data mapped to the [0, 1] interval to obtain N×3-dimensional effective pulse point cloud data, where the three dimensions are time, frequency, and amplitude, and N is the total number of time-frequency points. The specific process is as follows: Introducing amplitude threshold Amp th ; All pulse point clouds with amplitudes greater than the amplitude threshold are retained, and pulse point clouds with amplitudes less than or equal to the amplitude threshold are discarded to construct a valid pulse point cloud: Points′={(Time i ,Freq i ,Amp i )|Amp i >Amp th },i=1,2,...,N Among them, Points′ represents the valid pulse point cloud data.

5. The full pulse sampling radar sequence sorting method based on point cloud self-attention network according to claim 4 is characterized in that: In step 2, a PointTransformer sorting model is constructed; the specific process is as follows: The PointTransformer sorting model includes: First fully connected layer, first self-attention layer, first downsampling layer, second self-attention layer, second downsampling layer, third self-attention layer, third downsampling layer, fourth self-attention layer, fourth downsampling layer, fifth self-attention layer, second fully connected layer, sixth self-attention layer, first upsampling layer, seventh self-attention layer, second upsampling layer, eighth self-attention layer, third upsampling layer, ninth self-attention layer, fourth upsampling layer, tenth self-attention layer, third fully connected layer, Softmax layer; The working process of the PointTransformer sorting model is: Feature A is sequentially input into the first fully connected layer and the first self-attention layer, and the first self-attention layer outputs feature B; Feature B is sequentially input into the first downsampling layer and the second self-attention layer, and the second self-attention layer outputs feature C; Feature C is sequentially input into the second downsampling layer and the third self-attention layer, and the third self-attention layer outputs feature D; Feature D is sequentially input into the third downsampling layer and the fourth self-attention layer, and the fourth self-attention layer outputs feature E; Feature E is input into the fourth downsampling layer and the fifth self-attention layer in sequence, and the fifth self-attention layer outputs feature F; Feature F is sequentially input into the second fully connected layer and the sixth self-attention layer, and the sixth self-attention layer outputs feature G; Feature G and feature E are added element by element and then input into the first upsampling layer and the seventh self-attention layer in sequence. The seventh self-attention layer outputs feature H. Feature H and feature D are added element by element and then input into the second upsampling layer and the eighth self-attention layer in sequence. The eighth self-attention layer outputs feature I. Feature I and feature C are added element by element and then input into the third upsampling layer and the ninth self-attention layer in sequence. The ninth self-attention layer outputs feature J. Feature J and feature B are added element by element and then input into the fourth upsampling layer and the tenth self-attention layer in sequence. The tenth self-attention layer outputs feature K. Feature K is sequentially input into the third fully connected layer and the Softmax layer, and the Softmax layer outputs the semantic label of each pulse point cloud.

6. The full pulse sampling radar sequence sorting method based on point cloud self-attention network according to claim 5 is characterized by: In step 3, the constructed Point Transformer sorting model is trained until the loss function converges to obtain a trained Point Transformer sorting model; the specific process is as follows: Step 31: Each point in the valid pulse point cloud data obtained in step 1 corresponds to a pulse signal category, and pulse signal category labels are added to M points to obtain M pulse point cloud data sets with pulse signal category labels; Step 32: The M pulse point cloud data with pulse signal category labels obtained in step 31 are upgraded to 32 dimensions through linear transformation. The pulse point cloud data after linear transformation are sequentially subjected to BatchNorm and ReLU, and ReLU outputs features. The features are input into the Point Transformer sorting model, which then outputs the semantic label of each pulse point cloud. Step 33: Repeat step 32 until the loss function L converges to obtain the trained Point Transformer sorting model; The loss function L is: L=L WCE +L reg L WCE is the adaptive weighted cross entropy loss; L reg is the L2 regularization constraint.

7. The full pulse sampling radar sequence sorting method based on point cloud self-attention network according to claim 6 is characterized by: The adaptive weighted cross entropy loss L WCE Expressed as: Among them, M is the number of points in the point cloud, and C is the number of categories; y i,k is the true category of the pulse point cloud, p i,k The category probability predicted by the PointTransformer classification model; ω i is the weight; α is the hyperparameter; The L2 regularization constraint L reg Expressed as: Among them, λ is the regularization hyperparameter; θ is a parameter in the PointTransformer sorting model.

8. The full pulse sampling radar sequence sorting method based on point cloud self-attention network according to claim 7 is characterized in that: In step 5, the point cloud data classified in step 4 is restored to the original time domain signal; the specific process is: Step 51: Denormalize the time, frequency, and amplitude dimensions of the point cloud data classified in step 4 to obtain denormalized point cloud data. Step 52: Supplement the missing amplitude values at the time-frequency position in the point cloud data obtained after the denormalization processing in step 51 with 0; Step 53: reconstruct the time-frequency matrix based on step 52; Step 54: Perform an inverse short-time Fourier transform (ISTFT) on the time-frequency matrix reconstructed in step 53 to restore the reconstructed time-frequency matrix to a time-domain signal.

9. The full pulse sampling radar sequence sorting method based on point cloud self-attention network according to claim 8 is characterized by: In step 51, the time, frequency and amplitude dimensions of the point cloud data classified in step 4 are respectively denormalized to obtain the denormalized point cloud data. The specific process is as follows: 1) Denormalize the time in the point cloud data classified in step 4 to obtain the denormalized time; it is expressed as: y rec =x norm (x max -x min )+x min Among them, y rec is the time after denormalization; x norm is the time in the point cloud data after classification in step 4, x max and x min are the minimum and maximum values of the time in step 12 respectively; 2) Denormalize the frequency of the point cloud data classified in step 4 to obtain the denormalized frequency; it is expressed as: y′ rec =x′ norm (x′ max -x′ min )+x′ min Among them, y′ rec is the frequency after denormalization; x′ norm is the frequency in the point cloud data after classification in step 4, x′ max and x′ min are the minimum and maximum values of the frequency of step 12 respectively; 3) Denormalize the amplitude of the point cloud data classified in step 4 to obtain the denormalized amplitude; it is expressed as: y″ rec =x″ norm (x″ max -x″ min )+x″ min Among them, y rec is the amplitude after denormalization; x″ norm is the amplitude of the point cloud data after classification in step 4, x″ max and x″ min are the minimum and maximum values of the step 12 amplitude respectively.

10. The full pulse sampling radar sequence sorting method based on point cloud self-attention network according to claim 9 is characterized in that: In step 53, the time-frequency matrix is reconstructed based on step 52; the specific process is: For each time-frequency point, according to all the amplitudes Amp rec Phase φ with step 12 i , synthesize the corresponding time-frequency matrix: S rec (n,f)=Amp rec (n,f)e j×φ(n,f) Among them, S rec (n,f) represents the complex time-frequency matrix value at time point n and frequency f, where n represents the time point and f represents the frequency; Amp rec (n,f) represents the amplitude corresponding to time point n and frequency f; φ(n,f) represents the phase information obtained in step 12; j represents the imaginary unit, j 2 =-1.

Images