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

Abstract

The application relates to a full pulse sampling radar sequence sorting method based on a point cloud self-attention network, relates to the field of electronic reconnaissance of information and communication engineering, and specifically relates to a full pulse sampling radar sequence sorting method.The application aims to solve the problems of low signal sorting precision and poor robustness of existing methods in a complex scene, and proposes a full pulse sampling radar sequence sorting method based on a point cloud self-attention network.The process is as follows: 1, obtaining pulse point cloud data;2, constructing a Point Transformer sorting model;3, obtaining a trained sorting model;4, inputting the pulse point cloud data into the trained sorting model to output the classification result of the pulse point cloud data;5, restoring the classified point cloud data to the original time domain signal;6, obtaining the center frequency, bandwidth, pulse width, pulse repetition interval, pulse amplitude and modulation type based on the time domain signal obtained in step 5 to form a radiation source pulse description word.

Description

Technical Field

[0001] This invention relates to the field of electronic reconnaissance in information and communication engineering, specifically to a method for sorting sequences in a full-pulse sampling radar. Background Technology

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

[0003] Full-pulse sampling sequences are the raw signals acquired by passive radar receivers, containing various 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, making them widely applicable to different mission requirements. Current mainstream radar signal sorting research still focuses on sorting methods based on pulse descriptor words (PDWs), the sorting process as follows... Figure 2 As shown, existing PDW-based sorting methods typically ignore the fine-grained intra-pulse modulation characteristics in the original sampled signal, making 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 signal warp drive (PDW), such as arrival time, carrier frequency, pulse width, and pulse amplitude. These parameters are then clustered to sort the pulse signals. However, these methods depend on the completeness and accuracy of the pulse parameters. If noise interference or missing parameters occur during the measurement phase, sufficient criteria for judgment become difficult to provide, significantly weakening the effectiveness and stability of the sorting. This reliance on the quality of initial parameters creates a "front-end constraint," becoming a significant factor limiting the performance improvement of traditional sorting methods. Furthermore, these methods often neglect the rich information in the signal's intrapulse structure, failing to utilize the time-frequency dynamics embodied in complex modulation modes, thus limiting their adaptability to modern radar signal environments. Therefore, a novel approach is urgently needed, starting from the raw full-pulse data to perform more comprehensive feature analysis and sorting of radar signals. 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: Obtain the original full-pulse sampled signal x(n), and truncate the original full-pulse sampled signal x(n) to obtain the time-domain signal segment x. T (n); for time-domain signal segment x T (n) Perform a short-time Fourier transform to obtain three-dimensional pulse point cloud data; normalize the three-dimensional pulse point cloud data to map it to the [0,1] interval; based on the amplitude in the three-dimensional pulse point cloud data mapped to the [0,1] interval, crop the pulse point cloud 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 and frequency points;

[0008] Step 2: Construct the 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 effective pulse point cloud data obtained in Step 1 into the trained PointTransformer sorting model. The trained PointTransformer sorting model outputs the classification results of the N×3 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: Based on the time-domain signal obtained in Step 5, obtain the center frequency, bandwidth, pulse width, pulse repetition interval, pulse amplitude, and modulation type;

[0013] The radiation source pulse descriptor is constructed 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 truncated to obtain a time-domain signal segment x. T (n); for time-domain signal segment x T(n) Perform a short-time Fourier transform to obtain three-dimensional pulse point cloud data; normalize the three-dimensional pulse point cloud data to map it to the [0,1] interval; based on the amplitude in the three-dimensional pulse point cloud data mapped to the [0,1] interval, crop the pulse point cloud 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 and frequency points;

[0015] The specific process is as follows:

[0016] Step 11: Obtain the original full-pulse sampling signal x(n), and truncate the original full-pulse sampling signal x(n) to obtain the time-domain signal segment x. T (n); The specific process is as follows:

[0017] A time window of length T is extracted from the original full-pulse radar signal sequence. The pulse radar signal sequence within the window of length T is taken as the time-domain signal segment x. T (n);

[0018] Step 12: Use a sliding window function to divide the time-domain signal segment x T (n) Divide the data into M windows, and perform a Fourier transform on the time-domain signal within each window to obtain the 3D pulse point cloud data within each window; the specific process is as follows:

[0019] Step 121: Let the length of each window be L and the sliding step size be Δt. Then, divide the time-domain signal segment x... T (n) Divide the signal into M windows, and perform a Fourier transform on the time-domain signal within each window to obtain the time, frequency, amplitude, and phase information after the Fourier transform; the time, frequency, amplitude, and phase information after the Fourier transform are expressed as: (Time i ,Freq i Amp i ,φ i ), i = 1, 2, ..., N;

[0020] Among them, Time i Freq represents the i-th time in the pulse point cloud data. i Amp represents the i-th frequency in the pulse point cloud data. i φ represents the i-th amplitude in the pulse point cloud data. i This 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, retain only the time, frequency, and amplitude information to construct three-dimensional pulse point cloud data:

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

[0023] Where Points represents three-dimensional pulse point cloud data;

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

[0025] Step 14: Based on the amplitude in the 3D pulse point cloud data mapped to the [0,1] interval, the pulse point cloud is cropped 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 and 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 [0,1] interval; the specific process is as follows:

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

[0028]

[0029] Where, x norm The normalized time is represented by x, which represents the time in the 3D pulse point cloud data. min x represents the minimum time value in 3D pulse point cloud data. max This represents the maximum time value in the 3D pulse point cloud data;

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

[0031]

[0032] Where, x′ norm x′ represents the frequency after normalization, and x′ represents the frequency in the 3D pulse point cloud data. min x′ represents the minimum frequency in three-dimensional pulse point cloud data. max This represents the maximum frequency value 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] Where, x″ norm x″ represents the amplitude after normalization, and x″ represents the amplitude in the 3D pulse point cloud data.min x″ represents the minimum amplitude in three-dimensional pulse point cloud data. max This represents the maximum amplitude in the three-dimensional pulse point cloud data.

[0036] Preferably, in step 14, 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 and frequency points; the specific process is as follows:

[0037] Introducing an amplitude threshold (Amp) th ;

[0038] Retain all pulse point clouds whose amplitude is greater than the amplitude threshold, and discard pulse point clouds whose amplitude is less than or equal to the amplitude threshold to construct a valid pulse point cloud:

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

[0040] Points′ represents valid pulse point cloud data.

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

[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 as follows:

[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 sequentially input into the fourth downsampling layer and the fifth self-attention layer, 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] After element-wise summation of features G and E, they are sequentially input into the first upsampling layer and the seventh self-attention layer. The seventh self-attention layer outputs feature H.

[0052] After element-wise summation of features H and D, they are sequentially input into the second upsampling layer and the eighth self-attention layer. The eighth self-attention layer outputs feature I.

[0053] After element-wise summation of features I and C, they are sequentially input into the third upsampling layer and the ninth self-attention layer. The ninth self-attention layer outputs feature J.

[0054] After element-wise summation of features J and B, they are sequentially input into the fourth upsampling layer and the tenth self-attention layer. The tenth self-attention layer outputs feature K.

[0055] The feature K is sequentially input into the third fully connected layer and the Softmax layer. The Softmax layer outputs the semantic label for each pulse point cloud.

[0056] Preferably, in step 3, the constructed Point Transformer sorting model is trained until the loss function converges, thus obtaining the trained Point Transformer sorting model; the specific process is as follows:

[0057] Step 31: In the effective pulse point cloud data obtained in Step 1, each point corresponds to a pulse signal category. Add pulse signal category labels to M points to obtain M pulse point cloud datasets 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 then processed by BatchNorm and ReLU in sequence. ReLU outputs features.

[0059] Features are input into the Point Transformer sorting model, and the Point Transformer sorting model outputs a semantic label for 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 For adaptive weighted cross-entropy loss;

[0064] L reg This is an L2 regularization constraint.

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

[0066]

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

[0068] y i,k p represents the true category of the pulse point cloud. i,k The predicted class probabilities for the PointTransformer sorting model;

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

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

[0071]

[0072] Where λ is the hyperparameter for regularization;

[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 as follows:

[0075] Step 51: Perform inverse normalization on the three dimensions of time, frequency, and amplitude in the point cloud data classified in Step 4 to obtain the inverse normalized point cloud data.

[0076] Step 52: Fill the missing amplitude values ​​at the time-frequency positions in the point cloud data obtained after inverse normalization in step 51 with 0;

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

[0078] Step 54: Perform 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 denormalized to obtain denormalized point cloud data; the specific process is as follows:

[0080] 1) Perform inverse normalization on the time data in the point cloud data after classification in step 4 to obtain the inverse normalized time; represented as:

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

[0082] Among them, y rec x represents the time after inverse normalization. norm The time in the point cloud data after classification in step 4, x max and x min These are the minimum and maximum times for step 12, respectively;

[0083] 2) Perform inverse normalization on the frequencies in the point cloud data after classification in step 4 to obtain the inverse normalized frequencies; represented as:

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

[0085] Where, y′ rec The frequency after inverse normalization; x′ norm It is the frequency in the point cloud data after classification in step 4, x′ max and x′ min These are the minimum and maximum values ​​of the frequency in step 12, respectively;

[0086] 3) Perform inverse normalization on the amplitudes of the point cloud data after classification in step 4 to obtain the inverse normalized amplitudes; represented as:

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

[0088] Among them, y″ recThe amplitude after inverse normalization; x″ norm It refers to the amplitude, x″, in the point cloud data after classification in step 4. max and x″ min These are the minimum and maximum values ​​of the amplitude in step 12, respectively.

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

[0090] For each time frequency point, based on all amplitudes (Amp) rec Phase φ of 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 this invention are as follows:

[0097] This invention proposes an end-to-end sorting method for radar full-pulse signals based on a point cloud self-attention segmentation network. This method breaks through the limitations of the traditional PDW framework, directly processing the full-pulse sampling sequence. First, it uses short-time Fourier transform to map the one-dimensional time-domain signal into three-dimensional time-frequency-amplitude point cloud data, preserving the signal's modulation information and time-frequency evolution characteristics. Then, it introduces a point cloud segmentation network based on a self-attention mechanism, achieving automatic modeling and sorting of intrapulse structures without relying on manual feature engineering, thereby significantly improving the signal sorting accuracy and robustness in complex scenarios.

[0098] This invention proposes an end-to-end full-pulse sampling radar signal sorting method based on a self-attention mechanism. This method abandons the reliance on intermediate feature extraction and manual priors, and directly uses the full-pulse sampling time-domain signal as input. By constructing a deep learning framework that integrates a point cloud segmentation network and a self-attention mechanism, it achieves end-to-end automated sorting of radar full-pulse sampling sequences. This solves the technical bottleneck problem of "pre- and post-constraints" caused by the strong dependence on prior parameter measurement and feature representation capability, robustness, and traditional radar signal sorting methods that rely on PDW.

[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 full-pulse sampling sequence with a fixed time interval is extracted, and time-frequency data is extracted using short-time Fourier transform. The converted data is then converted into point clouds according to time, frequency, and amplitude, and the pulse point cloud data is normalized and pruned with an amplitude threshold to ensure that the number of points required by the network is met and to reduce the computational burden on the network.

[0100] PointTransformer learns the relationships between pulse data points through an attention mechanism, achieving efficient feature extraction. The model input is 3D point cloud data (time, frequency, amplitude). The input point cloud data first undergoes preliminary feature mapping through an MLP layer to improve data separability, and then enters PointTransformer for deep feature extraction. The network as a whole adopts an encoder-decoder structure, where the encoder is used to extract multi-scale features, and the decoder is used to restore the original point cloud resolution and perform segmentation prediction.

[0101] To improve the performance of PointTransformer in radar full-pulse signal sorting tasks, an optimization objective function incorporating an attention mechanism is introduced. During network training, the cross-entropy loss function is used to optimize model parameters, and regularization is introduced to avoid overfitting. Furthermore, to further improve sorting accuracy, the optimization objective function employs an adaptive weighting strategy to enhance attention to severely aliased pulse point clouds.

[0102] After model training, it can sort real-time radar full-pulse signals and reconstruct the original pulse sequence based on the classification results. First, the sorted point cloud data is denormalized to restore it to its initial 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 amplitude and phase information. Finally, the time-domain signal is restored through inverse short-time Fourier transform. Attached Figure Description

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

[0104] Figure 2 This is a schematic diagram of the sorting process based on Pulse Description Word (PDW);

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

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

[0107] Figure 5 This is a schematic diagram of time-frequency information extraction from a full-pulse signal;

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

[0109] Figure 7 This is a diagram showing the reconstruction results of the full pulse sequence;

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

[0111] Figure 9 The images show the sorting results at different signal-to-noise ratios. Detailed Implementation

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

[0113] Step 1: Obtain the original full-pulse sampled signal x(n), and truncate the original full-pulse sampled signal x(n) to obtain the time-domain signal segment x. T (n); for time-domain signal segment x T (n) Perform a short-time Fourier transform to obtain three-dimensional pulse point cloud data; normalize the three-dimensional pulse point cloud data to map it to the [0,1] interval; based on the amplitude in the three-dimensional pulse point cloud data mapped to the [0,1] interval, crop the pulse point cloud 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 and frequency points; as shown Figure 4 As shown;

[0114] Step 2: Construct the 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 (unlabeled) obtained in Step 1 into the trained PointTransformer sorting model. The trained PointTransformer sorting model outputs the classification results 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: Based on the time-domain signal obtained in Step 5, obtain the center frequency, bandwidth, pulse width, pulse repetition interval, pulse amplitude, and modulation type;

[0119] The emitter description word (EDW) of a radiation source is constructed based on the center frequency, bandwidth, pulse width, pulse repetition interval, pulse amplitude, and modulation type.

[0120] For each selected radiation source, the pulse descriptor word corresponding to that radiation source is calculated. 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 beneficial effects of the present invention are:

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

[0126] (3) The full-pulse sorting system proposed in this invention can adaptively mine key features in the time-frequency space, significantly improving the signal sorting accuracy in complex electromagnetic environments. By converting the time-frequency information of radar signals into point cloud representation, it achieves accurate characterization of fine-grained features between radar pulses, thereby effectively solving the problem of strong parameter dependence in traditional methods.

[0127] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that, in step 1, the original full-pulse sampling signal x(n) is obtained, and the original full-pulse sampling signal x(n) is truncated to obtain a time-domain signal segment x. T(n); for time-domain signal segment x T (n) Perform a short-time Fourier transform to obtain three-dimensional pulse point cloud data; normalize the three-dimensional pulse point cloud data to map it to the [0,1] interval; based on the amplitude in the three-dimensional pulse point cloud data mapped to the [0,1] interval, crop the pulse point cloud 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 and frequency points;

[0128] The specific process is as follows:

[0129] Step 11: Acquire the original full-pulse sampling signal x(n) through the passive radar receiving system, and truncate the original full-pulse sampling signal x(n) to obtain the time-domain signal segment x. T (n); The specific process is as follows:

[0130] A fixed-time window of length T is extracted from the original full-pulse radar signal sequence. The pulse radar signal sequence within the window of length T is taken as the time-domain signal segment x. T (n);

[0131] Step 12, as follows Figure 5 As shown, the time-domain signal segment x is divided by a sliding window function. T (n) Divide the data into M windows, and perform a Fourier transform on the time-domain signal within each window to obtain the 3D pulse point cloud data within each window; the specific process is as follows:

[0132] Step 121: Let the length of each window be L and the sliding step size be Δt. Then, divide the time-domain signal segment x... T (n) Divide the signal into M windows, and perform a Fourier transform on the time-domain signal within each window to obtain the time, frequency, amplitude, and phase information after the Fourier transform; the time, frequency, amplitude, and phase information after the Fourier transform are expressed as: (Time i ,Freq i Amp i ,φ i ), i = 1, 2, ..., N;

[0133] Among them, Time i Freq represents the i-th time in the pulse point cloud data. i Amp represents the i-th frequency in the pulse point cloud data. i φ represents the i-th amplitude in the pulse point cloud data. i This 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 performing a Fourier transform on the time-domain signal within the time window; the amplitude is the magnitude corresponding to the frequency component; and the phase is the phase information corresponding to the frequency component.

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

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

[0137] Where Points represents three-dimensional pulse point cloud data;

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

[0139] Step 14: Based on the amplitude in the 3D pulse point cloud data mapped to the [0,1] interval, the pulse point cloud is cropped 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 and frequency points.

[0140] The other steps and parameters are the same as in Specific Implementation Method 1.

[0141] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that step 13 performs normalization processing on the three-dimensional pulse point cloud data, mapping the three-dimensional pulse point cloud data to the [0,1] interval; the specific process is as follows:

[0142] Since the data in the three dimensions of time, frequency, and amplitude differ significantly in magnitude, if not properly processed, it will lead to problems such as gradient shift 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; represented as:

[0144]

[0145] Where, x norm The normalized time is represented by x, which represents the time in the 3D pulse point cloud data. min x represents the minimum time value in 3D pulse point cloud data. max This represents the maximum time value in the 3D pulse point cloud data;

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

[0147]

[0148] Where, x′ norm x′ represents the frequency after normalization, and x′ represents the frequency in the 3D pulse point cloud data. min x′ represents the minimum frequency in three-dimensional pulse point cloud data. max This represents the maximum frequency value 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 x″ represents the amplitude after normalization, and x″ represents the amplitude in the 3D pulse point cloud data. min x″ represents the minimum amplitude in three-dimensional pulse point cloud data. max This represents the maximum amplitude in the three-dimensional pulse point cloud data.

[0152] Other steps and parameters are the same as in specific implementation method one or two.

[0153] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that, in step 14, 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 and frequency points; the specific process is as follows:

[0154] Because STFT transforms the signal across the entire time domain, it generates spectral points even in time intervals where no pulse signal exists. This results in some point clouds consisting only of background noise, reducing the accuracy and efficiency of the processing. To reduce interference from redundant data, the point cloud data needs to be pruned. This step introduces an amplitude threshold, Amp. th This process filters out low-amplitude point cloud data; it retains all pulse point clouds with amplitudes greater than the amplitude threshold and discards pulse point clouds with amplitudes less than or equal to the amplitude threshold, thereby constructing a high-confidence, well-structured, and effective pulse point cloud.

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

[0156] Points′ represents valid pulse point cloud data.

[0157] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0158] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that step 2 involves constructing a Point Transformer sorting model; 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 as follows:

[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 sequentially input into the fourth downsampling layer and the fifth self-attention layer, 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] After element-wise summation of features G and E, they are sequentially input into the first upsampling layer and the seventh self-attention layer. The seventh self-attention layer outputs feature H.

[0169] After element-wise summation of features H and D, they are sequentially input into the second upsampling layer and the eighth self-attention layer. The eighth self-attention layer outputs feature I.

[0170] After element-wise summation of features I and C, they are sequentially input into the third upsampling layer and the ninth self-attention layer. The ninth self-attention layer outputs feature J.

[0171] After element-wise summation of features J and B, they are sequentially input into the fourth upsampling layer and the tenth self-attention layer. The tenth self-attention layer outputs feature K.

[0172] The feature K is sequentially input into the third fully connected layer and the Softmax layer. The Softmax layer outputs the semantic label for each pulse point cloud.

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

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

[0175] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0176] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that, in step 3, the constructed Point Transformer sorting model is trained until the loss function converges, thus obtaining the trained Point Transformer sorting model; the specific process is as follows:

[0177] Step 31: In the effective pulse point cloud data obtained in Step 1, each point corresponds to a pulse signal category. Add pulse signal category labels to M points to obtain M pulse point cloud datasets 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 pulse point cloud data after linear transformation are then processed by BatchNorm and ReLU in sequence. ReLU outputs features.

[0179] Features are input into the PointTransformer sorting model, and the PointTransformer sorting model outputs a semantic label for 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 For adaptive weighted cross-entropy loss;

[0184] L reg This is an L2 regularization constraint.

[0185] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0186] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that the adaptive weighted cross-entropy loss L... WCE Represented as:

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

[0188]

[0189] Where M is the number of points in the point cloud, and C is the number of categories; y i,k It is the true category of the pulse point cloud, p i,k It predicts the class probability;

[0190] To improve the focus on severely aliased pulse point clouds, an adaptive weighted cross-entropy loss Li is used. WCE :

[0191]

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

[0193] y i,k p represents the true category of the pulse point cloud. i,k The predicted class probabilities for the PointTransformer sorting model;

[0194] ω i The weights are dynamically adjusted based on the classification difficulty of the point cloud. When the prediction entropy of the pulse point cloud is high, it means that the classification uncertainty is greater, and the weights will be increased accordingly, making the hard-to-classify pulse point cloud contribute more to the loss function;

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

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

[0197] To prevent the model from overfitting, an L2 regularization constraint is added.

[0198]

[0199] Here, λ is a hyperparameter for regularization, used to control the strength of the regularization term. If λ is large, the model parameters will be smaller; if λ is small, the effect of regularization 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, and a category label needs to be added to it. 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 a training set, a validation set, and a test set 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 were set with a batch size of 16 and a total of 200 training epochs. The optimizer used was Adam, with an initial learning rate of 0.001 and a regularization hyperparameter λ of 0.0001. Weight decay was introduced to effectively suppress overfitting and improve the model's generalization ability. 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 large learning rates. Therefore, the learning rate was updated every 20 epochs with a decay factor of 0.5. The trained model was saved for online sorting when the loss function converged.

[0205] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0206] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods 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 results are denormalized to restore the original time, frequency, and amplitude ranges. Then, the point cloud data is associated with the original pulse signal according to the classification results, and the frequency-domain signal is converted to a time-domain signal using an inverse short-time Fourier transform. Finally, the complete pulse sequence is restored, as shown in the figure. Figure 7 As shown. Finally, statistical analysis is performed on the separated pulse signals. The specific implementation steps are as follows:

[0208] Step 51: Perform inverse normalization on the three dimensions of time, frequency, and amplitude in the point cloud data classified in Step 4 to obtain the inverse normalized point cloud data.

[0209] Step 52: Fill the missing amplitude values ​​at the time-frequency positions in the point cloud data obtained after inverse normalization in step 51 with 0; and retain the corresponding time and frequency indices;

[0210] Since the point cloud was pruned based on an amplitude threshold in step 14, some points in low-amplitude or background noise regions were removed. While this improved data quality, it also introduced sparsity into the time-frequency distribution of the point cloud, thus affecting the integrity of the time-frequency matrix. To ensure consistent input matrix dimensions during the Inverse Short-Time Fourier Transform (ISTFT) operation, interpolation padding is required for the missing points. This solution employs a simple and effective zero-padding strategy: points with an amplitude value of 0 are added to the missing time-frequency positions after pruning, while retaining the corresponding time and frequency indices.

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

[0212] Step 54: Perform 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 reconstructs the complete time series by performing an inverse Fourier transform on the spectrum of each time window and using a windowing and overlapping summation strategy.

[0214] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0215] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One to Eight in that, in step 51, the three dimensions of time, frequency, and amplitude in the point cloud data classified in step 4 are respectively denormalized to obtain point cloud data after denormalization.

[0216] The specific process is as follows:

[0217] 1) Perform inverse normalization on the time data in the point cloud data after classification in step 4 to obtain the inverse normalized time; represented as:

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

[0219] Among them, y rec x represents the time after inverse normalization. norm The time in the point cloud data after classification in step 4, x max and x min These are the minimum and maximum times for step 12, respectively;

[0220] 2) Perform inverse normalization on the frequencies in the point cloud data after classification in step 4 to obtain the inverse normalized frequencies; represented as:

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

[0222] Where, y′ rec The frequency after inverse normalization; x′ norm It is the frequency in the point cloud data after classification in step 4, x′ max and x′ min These are the minimum and maximum values ​​of the frequency in step 12, respectively;

[0223] 3) Perform inverse normalization on the amplitudes of the point cloud data after classification in step 4 to obtain the inverse normalized amplitudes; represented as:

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

[0225] Among them, y″ rec The amplitude after inverse normalization; x″ norm It refers to the amplitude, x″, in the point cloud data after classification in step 4. max and x″ min These are the minimum and maximum values ​​of the amplitude in step 12, respectively.

[0226] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0227] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One to Nine in that step 53 involves time-frequency matrix reconstruction based on step 52; the specific process is as follows:

[0228] The short-time Fourier transform essentially outputs complex spectral coefficients, containing both amplitude and phase information of the signal. During pulse sampling point cloud conversion, 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 rebuilt.

[0229] For each time frequency point, based on all 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 (including the zero-padded amplitude) corresponding to time point n and frequency f;

[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 any of the specific implementation methods one to nine.

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

[0237] Figure 8 The image shows the visualization results of the Point Transformer network output when using this method for signal sorting. Four radiation sources were used in the experiment: two with LFM intrapulse modulation, and the other two with NLFM and Barker code, respectively. Specific parameter information is shown in Table 3. The visualization results show that the network output labels accurately sort signals from different radiation sources, verifying the feasibility of Point Transformer in signal sorting tasks.

[0238] Based on the PointTransformer network, a simulated sorting experiment was conducted on multiple full-pulse radar signals (2, 4, 6, and 8 signals). The sorting performance was evaluated under different signal-to-noise ratios. The experimental results are as follows: Figure 9 As shown in the figure, the analysis results indicate that with the increase in the number of radiation sources, the distribution of radar signal characteristic parameters in the characteristic space gradually becomes intertwined and overlapping, significantly increasing the degree of parameter aliasing. This leads to increased complexity in the sorting task and a decrease in sorting accuracy. Nevertheless, PointTransformer exhibits strong feature learning ability and sorting robustness under different signal-to-noise ratio (SNR) conditions: the sorting accuracy reaches 99.3% at an SNR of 20dB; even under severe aliasing and an SNR reduced to -10dB, the sorting accuracy remains above 85%. This result fully validates the effectiveness and robustness of PointTransformer in full-pulse radar signal sorting tasks.

[0239] Table 3 Radar radiation source parameter settings

[0240]

[0241] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A method for sorting full-pulse sampling radar sequences based on point cloud self-attention networks, characterized in that: The specific process of the method is as follows: Step 1: Obtain the original full-pulse sampled signal x(n), and truncate the original full-pulse sampled signal x(n) to obtain the time-domain signal segment x. T (n); for time-domain signal segment x T (n) Perform a short-time Fourier transform to obtain three-dimensional pulse point cloud data; The three-dimensional pulse point cloud data is normalized and mapped 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 effective pulse point cloud data, where the three dimensions are time, frequency and amplitude, and N is the total number of time and frequency points. Step 2: Construct the 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 effective pulse point cloud data obtained in Step 1 into the trained Point Transformer sorting model. The trained Point Transformer sorting model outputs the classification results of the N×3 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: Based on the time-domain signal obtained in Step 5, obtain the center frequency, bandwidth, pulse width, pulse repetition interval, pulse amplitude, and modulation type; The radiation source pulse descriptor is constructed based on the center frequency, bandwidth, pulse width, pulse repetition interval, pulse amplitude, and modulation type.

2. The method for sorting full-pulse sampling radar sequences based on point cloud self-attention networks according to claim 1, characterized in that: In step 1, the original full-pulse sampling signal x(n) is obtained, and the original full-pulse sampling signal x(n) is truncated to obtain the time-domain signal segment x. T (n); for time-domain signal segment x T (n) Perform a short-time Fourier transform to obtain three-dimensional pulse point cloud data; The three-dimensional pulse point cloud data is normalized and mapped 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 effective pulse point cloud data, where the three dimensions are time, frequency and amplitude, and N is the total number of time and frequency points. The specific process is as follows: Step 11: Obtain the original full-pulse sampling signal x(n), and truncate the original full-pulse sampling signal x(n) to obtain the time-domain signal segment x. T (n); The specific process is as follows: A time window of length T is extracted from the original full-pulse radar signal sequence. The pulse radar signal sequence within the window of length T is taken as the time-domain signal segment x. T (n); Step 12: Use a sliding window function to divide the time-domain signal segment x 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 as follows: Step 121: Let the length of each window be L and the sliding step size be Δt. Then, divide the time-domain signal segment x... T (n) Divide the signal into M windows, and perform a Fourier transform on the time-domain signal within each window to obtain the time, frequency, amplitude, and phase information after the Fourier transform; the time, frequency, amplitude, and phase information after the Fourier transform are expressed as: (Time i ,Freq i Amp i ,φ i ), i = 1, 2, ..., N; Among them, Time i Freq represents the i-th time in the pulse point cloud data. i Amp represents the i-th frequency in the pulse point cloud data. i φ represents the i-th amplitude in the pulse point cloud data. i This 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, retain only the time, frequency, and amplitude information to construct three-dimensional pulse point cloud data: Points={(Time i ,Freq i ,Amp i )},i=1,2,…,N Where Points represents three-dimensional pulse point cloud data; Step 13: Normalize the 3D pulse point cloud data and map it to the [0,1] interval; Step 14: Based on the amplitude in the 3D pulse point cloud data mapped to the [0,1] interval, the pulse point cloud is cropped 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 and frequency points.

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

4. The method for sorting full-pulse sampling radar sequences based on point cloud self-attention networks according to claim 3, characterized in that: In step 14, 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 and frequency points; the specific process is as follows: Introducing an amplitude threshold (Amp) th ; Retain all pulse point clouds whose amplitude is greater than the amplitude threshold, and discard pulse point clouds whose amplitude is less than or equal to the amplitude threshold to construct a valid pulse point cloud: Points′={(Time i ,Freq i ,Amp i )|Amp i >Amp th },i=1,2,...,N Points′ represents valid pulse point cloud data.

5. The method for sorting full-pulse sampling radar sequences based on point cloud self-attention networks according to claim 4, characterized in that: Step 2 involves constructing the PointTransformer sorting model; 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 as follows: 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 sequentially input into the fourth downsampling layer and the fifth self-attention layer, 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. After element-wise summation of features G and E, they are sequentially input into the first upsampling layer and the seventh self-attention layer. The seventh self-attention layer outputs feature H. After element-wise summation of features H and D, they are sequentially input into the second upsampling layer and the eighth self-attention layer. The eighth self-attention layer outputs feature I. After element-wise summation of features I and C, they are sequentially input into the third upsampling layer and the ninth self-attention layer. The ninth self-attention layer outputs feature J. After element-wise summation of features J and B, they are sequentially input into the fourth upsampling layer and the tenth self-attention layer. The tenth self-attention layer outputs feature K. The feature K is sequentially input into the third fully connected layer and the Softmax layer. The Softmax layer outputs the semantic label for each pulse point cloud.

6. The method for sorting full-pulse sampling radar sequences based on point cloud self-attention networks according to claim 5, characterized in that: In step 3, the constructed Point Transformer sorting model is trained until the loss function converges, thus obtaining the trained Point Transformer sorting model; the specific process is as follows: Step 31: In the effective pulse point cloud data obtained in Step 1, each point corresponds to a pulse signal category. Add pulse signal category labels to M points to obtain M pulse point cloud datasets 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 then processed by BatchNorm and ReLU in sequence. ReLU outputs features. Features are input into the Point Transformer sorting model, and the Point Transformer sorting model outputs a semantic label for 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 For adaptive weighted cross-entropy loss; L reg This is an L2 regularization constraint.

7. The method for sorting full-pulse sampling radar sequences based on point cloud self-attention networks according to claim 6, characterized in that: The adaptive weighted cross-entropy loss L WCE Represented as: Where M is the number of points in the point cloud, and C is the number of categories; y i,k p represents the true category of the pulse point cloud. i,k The predicted class probabilities for the PointTransformer sorting model; ω i α is the weight; α is the hyperparameter. The L2 regularization constraint L reg Represented as: Where λ is the hyperparameter for regularization; θ is a parameter in the PointTransformer sorting model.

8. The method for sorting full-pulse sampling radar sequences based on point cloud self-attention networks according to claim 7, 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 as follows: Step 51: Perform inverse normalization on the three dimensions of time, frequency, and amplitude in the point cloud data classified in Step 4 to obtain the inverse normalized point cloud data. Step 52: Fill the missing amplitude values ​​at the time-frequency positions in the point cloud data obtained after inverse normalization in step 51 with 0; Step 53: Reconstruct the time-frequency matrix based on step 52; Step 54: Perform 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 method for sorting full-pulse sampling radar sequences based on point cloud self-attention networks according to claim 8, characterized in that: In step 51, the time, frequency, and amplitude dimensions of the point cloud data classified in step 4 are denormalized to obtain denormalized point cloud data; the specific process is as follows: 1) Perform inverse normalization on the time data in the point cloud data after classification in step 4 to obtain the inverse normalized time; represented as: y rec =x norm (x max -x min )+x min Among them, y rec x represents the time after inverse normalization. norm The time in the point cloud data after classification in step 4, x max and x min These are the minimum and maximum times for step 12, respectively; 2) Perform inverse normalization on the frequencies in the point cloud data after classification in step 4 to obtain the inverse normalized frequencies; represented as: y′ rec =x′ norm (x′ max -x′ min )+x′ min Where, y′ rec The frequency after inverse normalization; x′ norm It is the frequency in the point cloud data after classification in step 4, x′ max and x′ min These are the minimum and maximum values ​​of the frequency in step 12, respectively; 3) Perform inverse normalization on the amplitudes of the point cloud data after classification in step 4 to obtain the inverse normalized amplitudes; represented as: y″ rec =x″ norm (x″ max -x″ min )+x″ min Among them, y″ rec The amplitude after inverse normalization; x″ norm It refers to the amplitude, x″, in the point cloud data after classification in step 4. max and x″ min These are the minimum and maximum values ​​of the amplitude in step 12, respectively.

10. The method for sorting full-pulse sampling radar sequences based on point cloud self-attention networks according to claim 9, characterized in that: In step 53, time-frequency matrix reconstruction is performed based on step 52; the specific process is as follows: For each time frequency point, based on all amplitudes (Amp) rec Phase φ of 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