An end-to-end radar signal sorting method and system based on point cloud segmentation

Abstract

The application provides an end-to-end radar signal sorting method and system based on point cloud segmentation, and relates to the technical field of signal processing, and comprises the following steps: S1, acquiring a PDW pulse sequence arranged in the order of pulse arrival time; S2, normalizing each pulse arrival time based on the pulse arrival time of the first pulse in the PDW pulse sequence to generate a time sequence code; S3, establishing a four-dimensional space-time point cloud space, taking the carrier frequency, pulse width and arrival angle corresponding to each pulse as the spatial coordinates of the point cloud data, and taking the time sequence code corresponding to each pulse as the time coordinates of the point cloud data; S4, inputting the point cloud data into a time sequence enhanced instance segmentation network for segmentation, and outputting the signal source instance label corresponding to each pulse according to the space-time joint feature; and S5, aggregating the pulses with the same signal source instance label to obtain the sorting result of each signal source, so that the high-precision end-to-end sorting of dense overlapping radar signal sources is realized.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, specifically to an end-to-end radar signal sorting method and system based on point cloud segmentation. Background Technology

[0002] With the increasing complexity of the modern electromagnetic environment, radar signal pulse flow density has reached the million level, signal patterns are becoming increasingly diverse, and signal overlap is severe. Radar signal sorting technology has become a key technical link in electronic reconnaissance and electronic countermeasures. The purpose of radar signal sorting is to separate the pulse sequences emitted by each signal source from the dense and intertwined pulse flow, thereby realizing the identification and tracking of different signal sources.

[0003] Traditional radar signal sorting methods mainly include histogram methods based on pulse repetition interval (PRI), multi-parameter clustering methods based on angle of arrival (DOA) and carrier frequency (RF), and library search methods based on template matching. However, these methods suffer from the following problems when processing high-density, highly complex modern radar signals: when there are many signal sources and signal parameters overlap significantly, traditional clustering methods struggle to accurately distinguish pulses from different sources; traditional methods are highly dependent on prior knowledge, making them unsuitable for sorting signals from unknown sources or new radar systems; furthermore, traditional methods lack robustness in handling pulse loss and spurious pulses, failing to meet the real-time requirements of radar signal sorting in complex electromagnetic environments.

[0004] In recent years, the development of deep learning technology has provided new ideas for radar signal sorting. Among them, the U-Net-based image segmentation method converts a pulse stream (PDW) into a two-dimensional image, achieving separation of pulses from different signal sources through image segmentation. However, two-dimensional image segmentation methods have the following shortcomings: pulse pixel overlap exists during the mapping of PDW data to a two-dimensional image; the two-dimensional representation loses the high-dimensional spatial structure information of the PDW; and networks such as U-Net have a large number of parameters, making it difficult to meet real-time processing requirements. Deep learning methods based on three-dimensional point clouds offer a new technical direction for this purpose. Each pulse in the PDW pulse stream naturally possesses multi-dimensional parameters, which can be directly mapped to points in the point cloud space.

[0005] Existing technologies include solutions that introduce point cloud segmentation networks into radar signal sorting. These methods segment PDW point cloud data using point cloud segmentation networks and then use clustering methods to distinguish different signal sources. However, they still have the following key technical shortcomings: First, this type of method only achieves semantic segmentation, meaning it can only classify pulses into signal types such as fixed PRI, jittery PRI, and uneven PRI, but cannot directly distinguish individual signal sources of the same signal type. Post-processing clustering is necessary to complete the sorting. In this two-stage process, semantic segmentation errors propagate to the clustering stage, leading to accumulated sorting errors. Second, this type of method normalizes TOA as a normal coordinate, failing to fully utilize the timing information contained in the signal pulse stream, especially the PRI modulation mode, which is a core feature dimension for distinguishing different signal sources. Third, this type of method uses standard farthest point sampling and assumes a uniform distribution of point cloud density. However, in actual radar signal environments, the pulse density of different signal sources varies greatly, with high-power signal sources having a much higher pulse density than low-power signal sources. Standard sampling strategies cannot adequately account for the sampling sufficiency of various signal sources. Fourth, the standard classification loss function used by this type of method only focuses on the classification accuracy of a single pulse, without utilizing the physical prior that signal pulses are naturally continuous in time. When there is severe parameter overlap, it is easy to misclassify continuous pulses from the same signal source into different types. Summary of the Invention

[0006] The purpose of this invention is to provide an end-to-end radar signal sorting method and system based on point cloud segmentation. It mainly maps PDW pulse sequences into a four-dimensional spatiotemporal point cloud space containing spatial coordinates and temporal codes, and uses a temporally enhanced instance segmentation network to directly output the signal source instance labels corresponding to each pulse, without the need for post-processing clustering, thereby achieving high-precision end-to-end sorting of densely overlapping radar signal sources.

[0007] Specifically, the present invention adopts the following solution: An end-to-end radar signal sorting method based on point cloud segmentation includes the following: S1. Obtain the PDW pulse sequence arranged in order of pulse arrival time; S2. Extract parameters from each pulse in the PDW pulse sequence to obtain the carrier frequency, pulse width, pulse arrival time, and angle of arrival. Normalize the arrival time of each pulse based on the pulse arrival time of the first pulse in the PDW pulse sequence to generate a timing code and maintain the relative time interval relationship between each pulse. S3. Establish a four-dimensional spatiotemporal point cloud space. Use the carrier frequency, pulse width and angle of arrival corresponding to each pulse as the spatial coordinates of the point cloud data, and use the timing code corresponding to each pulse as the temporal coordinates of the point cloud data. Based on the spatial coordinates and temporal coordinates, map each pulse to point cloud data in the four-dimensional spatiotemporal point cloud space. S4. Input the point cloud data into the temporal enhanced instance segmentation network for segmentation. Extract the spatiotemporal joint features based on the spatial and temporal coordinates in the local temporal neighborhood of the point cloud data, and output the signal source instance label corresponding to each pulse based on the spatiotemporal joint features. S5. Aggregate the pulses of the same signal source instance tags to obtain the sorting results of each signal source.

[0008] Furthermore, the PDW pulse sequence arranged in chronological order of pulse arrival time is obtained, including: Acquire PDW pulse stream data, preprocess the PDW pulse stream data, and sort the preprocessed PDW pulse stream data in ascending order according to the pulse arrival time to form a PDW pulse sequence; Preprocessing includes deduplication and outlier removal.

[0009] Furthermore, the arrival times of each pulse are normalized based on the arrival time of the first pulse in the PDW pulse sequence, specifically as follows: Obtain the minimum pulse arrival time in the PDW pulse sequence. and maximum value Calculate the timing code of the i-th pulse sequentially. : ; in, This represents the arrival time of the i-th pulse.

[0010] Furthermore, each pulse is mapped to point cloud data in a four-dimensional spatiotemporal point cloud space based on spatial and temporal coordinates. ,in, This represents the carrier frequency of the i-th pulse. This represents the pulse width of the i-th pulse. Indicates the angle of arrival of the i-th pulse; also includes: The point cloud data is normalized using the maximum and minimum values ​​to obtain normalized point cloud data. : ; ; ; ; in, and These are the minimum and maximum values ​​of all pulse carrier frequencies, respectively; and These are the minimum and maximum values ​​of all pulse widths, respectively; and These are the minimum and maximum angles of arrival for all pulses, respectively. and These are the minimum and maximum values ​​of all pulse timing codes, respectively.

[0011] Furthermore, S4 includes the following: S41. Adjust the sampling frequency of the point cloud data according to the local density of the point cloud data; S42. Determine the local temporal neighborhood of each point cloud data based on its spatial and temporal coordinates. S43. Extract temporal features in the local temporal neighborhood of each point cloud data, and fuse the temporal features with the spatial features of the point cloud data to obtain spatiotemporal joint features; S44. Predict the signal source instance label corresponding to each pulse based on the spatiotemporal joint features.

[0012] Furthermore, the process of adjusting the sampling frequency of point cloud data based on the local density of the point cloud data is as follows: Using each point cloud data as the center and a preset radius to determine the neighborhood of a sphere, the number of neighborhood points within each sphere's neighborhood is counted as the local density of each point cloud data; based on the local density of each point cloud data, the corresponding probability is determined, wherein the probability is positively correlated with the local density; Sampling points are selected sequentially according to probability, and the sampling results are filtered based on the constraint of the farthest Euclidean distance from the selected sampling points until the preset number of samples is reached.

[0013] Furthermore, the process of determining the local temporal neighborhood of each point cloud data based on its spatial and temporal coordinates is as follows: determine the parameter spatial neighborhood based on the spatial coordinates of each point cloud data, determine the temporal neighborhood based on the temporal coordinates of each point cloud data, and take the union of the parameter spatial neighborhood and the temporal neighborhood as the local temporal neighborhood.

[0014] Furthermore, temporal features are extracted within the local temporal neighborhood of each point cloud data point, and these temporal features are fused with the spatial features of the point cloud data to obtain spatiotemporal joint features, specifically: For each point cloud data, adjacent pulses located within a preset time window are extracted along the time axis within a local temporal neighborhood to form the temporal neighborhood of the point cloud data; the differential arrival time sequence is calculated for the pulse sequence within the temporal neighborhood, and the differential arrival time sequence is used as the local pulse repetition interval sequence; The local pulse repetition interval sequence is encoded by a one-dimensional convolutional network, and the pulse repetition interval mean feature, pulse repetition interval change rate feature, and pulse repetition interval modulation mode feature are extracted as temporal features. The temporal features are then concatenated and fused with the spatial features to obtain the spatiotemporal joint features.

[0015] Furthermore, the temporal-enhanced instance segmentation network is trained using a temporal consistency loss function, which includes a classification loss, an instance triplet loss, and a temporal consistency loss term. The temporal consistency loss term is used to enhance the consistency of pulse instance labels that are temporally adjacent and belong to the same signal source.

[0016] An end-to-end radar signal sorting system based on point cloud segmentation applies an end-to-end radar signal sorting method based on point cloud segmentation, including: Data acquisition module: Acquires PDW pulse sequences arranged in chronological order of pulse arrival time; Parameter extraction module: Extracts parameters from each pulse in the PDW pulse sequence to obtain carrier frequency, pulse width, pulse arrival time, and angle of arrival. Normalizes the arrival time of each pulse based on the pulse arrival time of the first pulse in the PDW pulse sequence to generate timing codes and maintain the relative time interval relationship between each pulse. Four-dimensional spatiotemporal point cloud mapping module: Establish a four-dimensional spatiotemporal point cloud space, using the carrier frequency, pulse width, and angle of arrival corresponding to each pulse as the spatial coordinates of the point cloud data, and using the timing code corresponding to each pulse as the temporal coordinates of the point cloud data. Based on the spatial coordinates and temporal coordinates, each pulse is mapped to point cloud data in the four-dimensional spatiotemporal point cloud space. Temporally Enhanced Instance Segmentation Module: Inputs point cloud data into the temporally enhanced instance segmentation network for segmentation. Extracts spatiotemporal joint features based on spatial and temporal coordinates within the local temporal neighborhood of the point cloud data, and outputs the signal source instance label corresponding to each pulse based on the spatiotemporal joint features. Signal source sorting module: Aggregates pulses from the same signal source instance tags to obtain the sorting results for each signal source.

[0017] The beneficial effects of this invention are as follows: (1) Based on the traditional three-dimensional spatial coordinates (carrier frequency RF, pulse width PW, angle of arrival DOA), timing coding is introduced as the fourth dimension coordinate, which explicitly preserves the relative time interval relationship between pulses, so that pulses that are close in time and space will naturally gather in the point cloud space. The timing structure information of the pulses can be used to assist the signal source sorting, solve the problem that it is difficult to distinguish overlapping pulses by relying solely on the parameter spatial coordinates, and improve the sorting accuracy.

[0018] (2) In view of the non-uniform distribution characteristics of signal pulse density, a local density estimation and probability sampling mechanism is introduced in the traditional farthest point sampling. The sampling points are adaptively increased in high-density areas to capture fine structure, and the sampling points are reduced in low-density areas to avoid oversampling. This can take into account the pulse density differences of different signal sources and improve the detection and sorting capabilities of weak signal sources.

[0019] (3) The local pulse repetition interval (PRI) sequence is encoded by a one-dimensional convolutional network, and key temporal features such as the pulse repetition interval mean feature, pulse repetition interval change rate feature and pulse repetition interval modulation mode feature are automatically extracted. In the feature extraction stage, the key temporal features are fused with the parameter space features, which can enhance the network's ability to distinguish different signal sources and realize the automatic identification of PRI modulation type.

[0020] (4) Add a time regularization term to the instance segmentation loss. By utilizing the prior continuity of signal pulses in time, the temporal consistency is constrained, so that the instance embedding vectors of temporally adjacent pulses tend to be consistent, reducing the segmentation fragmentation caused by instantaneous parameter overlap, and improving the integrity and continuity of sorting.

[0021] (5) By using a learnable MLP to map the original parameters of the PDW to a high-dimensional feature space, the distinguishability of pulses that are adjacent but belong to different signal sources in the parameter space is enhanced, and the sorting accuracy is improved in scenarios where signal parameters overlap significantly. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the end-to-end radar signal sorting method in Embodiment 1 of the present invention.

[0023] Figure 2 This is a comparative diagram showing the effect of timing consistency constraints on instance segmentation in Embodiment 1 of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0026] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0027] Furthermore, for clarity and brevity, descriptions of well-known structures, functions, and configurations may have been omitted. Those skilled in the art will recognize that various changes and modifications can be made to the examples described herein without departing from the spirit and scope of this disclosure.

[0028] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0029] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0030] The present invention will now be described in detail with reference to the accompanying drawings and embodiments: Example 1 In this embodiment, an end-to-end radar signal sorting method based on point cloud segmentation is proposed. This end-to-end radar signal sorting method, as follows: Figure 1 As shown, it includes the following steps: S1. Obtain the PDW pulse sequence arranged in order of pulse arrival time; S2. Extract parameters from each pulse in the PDW pulse sequence to obtain the carrier frequency, pulse width, pulse arrival time, and angle of arrival. Normalize the arrival time of each pulse based on the pulse arrival time of the first pulse in the PDW pulse sequence to generate a timing code and maintain the relative time interval relationship between each pulse. S3. Establish a four-dimensional spatiotemporal point cloud space. Use the carrier frequency, pulse width and angle of arrival corresponding to each pulse as the spatial coordinates of the point cloud data, and use the timing code corresponding to each pulse as the temporal coordinates of the point cloud data. Based on the spatial coordinates and temporal coordinates, map each pulse to point cloud data in the four-dimensional spatiotemporal point cloud space. S4. Input the point cloud data into the temporal enhanced instance segmentation network for segmentation. Extract the spatiotemporal joint features based on the spatial and temporal coordinates in the local temporal neighborhood of the point cloud data, and output the signal source instance label corresponding to each pulse based on the spatiotemporal joint features. S5. Aggregate the pulses of the same signal source instance tags to obtain the sorting results of each signal source.

[0031] Based on the above principles, steps S1-S5 are explained as follows.

[0032] In step S1, raw PDW pulse stream data is acquired from the signal receiver. Each PDW record contains at least four parameters: carrier frequency (RF), pulse width (PW), pulse arrival time (TOA), and angle of arrival (DOA). The acquired PDW pulse stream data is preprocessed, including deduplication and outlier removal.

[0033] Deduplication is used to remove duplicate pulse records caused by multipath effects or repeated acquisition from multiple channels of the receiver. Specifically, a time window is set. During the time window If there are multiple PDW records with the same pulse arrival time (TOA) and both the carrier frequency (RF) and pulse width (PW) deviations within the set threshold, only one of them will be retained, and the rest will be discarded as duplicate records.

[0034] Outlier removal is used to eliminate pulse data whose parameters significantly deviate from the normal range. Specifically, the statistical mean and standard deviation of each parameter are calculated, and pulse data whose parameter values ​​exceed the mean ± 3 times the standard deviation are removed.

[0035] The preprocessed PDW pulse stream data is sorted in ascending order according to the pulse arrival time to form a PDW pulse sequence arranged in order of pulse arrival time.

[0036] In step S2, four parameters are extracted from the PDW pulse sequence: carrier frequency (RF), pulse width (PW), pulse arrival time (TOA), and angle of arrival (DOA). The pulse arrival time (TOA) is then relative-time encoded to generate a timing code. The purpose is to preserve the timing information in the PDW pulse stream, that is, to maintain the relative time interval relationship between each pulse.

[0037] Specifically, in the process of relative time coding, time-series coding Normalization is performed based on the arrival time of the first pulse in the PDW pulse sequence. The process is as follows: Obtain the minimum pulse arrival time in the PDW pulse sequence. and maximum value ; Calculate the timing code of the i-th pulse sequentially. : ; in, This represents the arrival time of the i-th pulse.

[0038] It can be seen that this timing code Normalization is performed based on the arrival time of the first pulse in the PDW pulse sequence, which maintains the relative time interval between each pulse, allowing temporally adjacent pulses to naturally cluster in the subsequently constructed point cloud space.

[0039] Based on this, step S3 establishes a four-dimensional spatiotemporal point cloud space, using the carrier frequency RF, pulse width PW, and angle of arrival (DOA) corresponding to each pulse as the spatial coordinates of the point cloud data, thus forming a traditional three-dimensional spatial coordinate system; simultaneously, the timing code corresponding to each pulse is used... Using the time coordinates of point cloud data, a four-dimensional spatiotemporal joint point cloud representation is constructed. Unlike existing schemes that treat pulse arrival time as ordinary coordinates, this invention chooses to introduce temporal coding. As the fourth dimension of coordinates, while preserving the relative time interval relationship between each pulse, it enables temporally adjacent pulses to form a natural cluster in the four-dimensional spatiotemporal point cloud space.

[0040] Based on spatial and temporal coordinates, each pulse is mapped to point cloud data in a four-dimensional spatiotemporal point cloud space, meaning one pulse corresponds to one point, and the coordinates of that point are represented as follows: ,in, Indicates the carrier frequency of the i-th pulse. , This represents the pulse width of the i-th pulse. , Indicates the angle of arrival of the i-th pulse. ; This represents the timing code of the i-th pulse.

[0041] Perform maximum and minimum value normalization on point cloud data in four-dimensional spatiotemporal point cloud space: ; ; ; ; in, and These are the minimum and maximum values ​​of all pulse carrier frequencies, respectively; and These are the minimum and maximum values ​​of all pulse widths, respectively; and These are the minimum and maximum angles of arrival for all pulses, respectively. and These are the minimum and maximum values ​​of all pulse timing codes, respectively.

[0042] Specifically, when At that time, take ;when At that time, take ;when At that time, take ;when At that time, take .

[0043] The normalized four-dimensional spatiotemporal point cloud dataset is represented as follows: ; Where N is the total number of pulses.

[0044] In step S4, the enhanced four-dimensional spatiotemporal point cloud dataset is input into the temporally enhanced PointNet++ (TE-PointNet++) instance segmentation network for segmentation, including the following sub-steps S41 to S44.

[0045] S41. Adjust the sampling frequency of the point cloud data according to the local density of the point cloud data. The local density of each point in the point cloud data has non-uniform distribution characteristics. In order to improve the sampling efficiency, this step chooses to use adaptive density-aware sampling (ADA-FPS) instead of traditional farthest point sampling.

[0046] Specifically, a sphere neighborhood is determined with each point cloud data point as the center and a preset radius r. The number of point cloud data points contained within each sphere neighborhood is counted and used as the local density of each point cloud data point. Based on the local density of each point cloud data The calculation uses a weighted rate distribution. Specifically, when alpha = 0.5, and the radius r = 0.05 is used in density estimation, Then, normalize it into a probability. The probability With local density Positive correlation; according to probability Select candidate points, and then choose the farthest point among the candidate points. During the iterative sampling process, each time with probability... Candidate points are selected from the remaining points, while considering the constraint of the farthest Euclidean distance from the selected point set. While maintaining geometric coverage, sampling is increased in dense areas first, so as to achieve adaptive increase of sampling in high-density areas and decrease of sampling in low-density areas.

[0047] S42. Determine the local temporal neighborhood of each point cloud data point based on its spatial and temporal coordinates. Specifically, a grouping strategy based on both parameter space Euclidean distance and temporal proximity is proposed. A parameter space neighborhood Ns and a temporal neighborhood Nt are constructed, and the union of these two neighborhoods is taken as the local temporal neighborhood N = Ns ⊆ Nt. The sphere radius increases progressively with each layer, starting at 0.2 for the first layer and 0.4 for the second layer, and the time window... Adaptive adjustment.

[0048] In this embodiment, a three-layer set abstraction layer is adopted, and the parameter settings of each layer are shown in Table 1 below: Table 1 Parameter Setting Table S43. Extract temporal features within the local temporal neighborhood of each point cloud data set, and fuse these temporal features with the spatial features of the point cloud data to obtain spatiotemporal joint features. Specifically, PointNet is used to extract features for each neighborhood point set. Spatial coordinates, enhanced features, and temporal coordinates are concatenated into a comprehensive feature vector. Local spatial-temporal joint features (spatiotemporal joint features) are extracted using MLP and max pooling, as follows: ; in, For the j-th sampling point, for The neighborhood point set, where concat represents the feature concatenation operation. This represents the coordinates of the i-th point. To enhance features, For sequential coding, MLP stands for Multilayer Perceptron, and MAX stands for Channel-wise Max Pooling. The extracted local spatial-temporal joint features.

[0049] Specifically, for each point cloud data Within a local temporal neighborhood, adjacent pulses located within a preset time window are extracted along the time axis to form the temporal neighborhood of the point cloud data. : ; Calculate the differential arrival time series of pulse sequences within the temporal neighborhood. : ; Differential arrival time series As a local pulse repetition interval sequence, i.e., a local PRI sequence, the local PRI sequence is encoded using a one-dimensional convolutional network. This one-dimensional convolutional network can employ three layers with kernel sizes of 3, 5, and 3, and channels of 16, 32, and 64 respectively. The pulse repetition interval mean, pulse repetition interval change rate, and pulse repetition interval modulation mode features are extracted as temporal features. Specifically, the PRI mean, PRI change rate, and PRI modulation mode features are converged into a temporal feature vector using global average pooling. Then the time series feature vector With parameter space features Fusion: The fused features contain both parameter space and temporal pattern information.

[0050] S44. Predict the signal source instance label corresponding to each pulse based on the spatiotemporal joint features.

[0051] In this step, the extracted spatiotemporal joint features are input into the segmentation head for instance prediction. The segmentation head consists of a classification branch and an instance branch. The classification branch uses cross-entropy loss for classification prediction; the instance branch outputs the instance embedding vector of each impulse point, and uses triplet loss to minimize the embedding vector distance between instances from the same signal source and maximize the embedding vector distance between instances from different signal sources.

[0052] Preferably, the predicted signal source instance labels are optimized using a temporal consistency constraint. The temporal consistency constraint is implemented through a temporal consistency loss function, which includes a classification loss, an instance triplet loss, and a temporal consistency loss term, expressed as: ; in, For classification loss; For instance triplet loss; This is the time-series consistency loss term. , and An instance embedding vector for two temporally adjacent pulse points; To balance the weighting coefficients, the embedding vectors of pulse point instances that are temporally adjacent and belong to the same signal source are made more consistent, reducing instance segmentation fragmentation caused by parameter overlap.

[0053] In this embodiment, the Adam optimizer is used to train the TE-PointNet++ instance segmentation network. The initial learning rate is set to 0.001, the number of training epochs is 200, and the batch size is set to 16. During training, the learning rate is decayed using a cosine annealing strategy, with a minimum learning rate of 1e-5.

[0054] In addition, in this embodiment, the downsampled global features are progressively upsampled and propagated back to the original point cloud size through feature propagation, using a distance-weighted trilinear interpolation method: ; in, , For Euclidean distance, Let be the feature vector of the i-th sampling point. For each point to be interpolated, take the k nearest sampling points for weighted interpolation, where k is typically 3. The interpolated features are then concatenated with the encoded features of the corresponding layer via skip connections, and feature fusion is performed using an MLP.

[0055] In summary, the TE-PointNet++ instance segmentation network proposed in this invention has made the following innovative improvements that are different from the existing technology based on the standard PointNet++: (1) The standard FPS sampling is replaced with adaptive density-aware sampling ADA-FPS to solve the problem of not being adapted to non-uniform pulse density caused by the use of standard FPS; (2) A new temporal coding module is added to extract local PRI mode features to overcome the defect of not using PRI modulation mode as feature dimension; (3) Multimodal grouping is based on both spatial distance and temporal proximity, breaking through the limitation of standard PointNet++ grouping based only on Euclidean distance; (4) Temporal consistency constraint loss function is used to replace standard classification loss, and fragmentation is reduced by using temporal continuity prior.

[0056] In step S5, the signal source descriptor generation process includes: (1) Pulse aggregation: Aggregate the PDW pulses corresponding to points belonging to the same instance label to form pulse clusters; (2) Feature statistics: Perform statistical analysis on each pulse cluster to extract statistical features such as the average carrier frequency and carrier frequency variation range, the average pulse width and pulse width variation range, the average pulse repetition interval PRI and PRI variation mode (fixed, uneven, jitter, slip, etc.), and the average angle of arrival; (3) Descriptor generation: Generate the signal source descriptor based on the statistical features, with the format: {Signal source ID, carrier frequency variation range, average pulse width, pulse width type, average pulse repetition interval PRI, PRI variation mode, average angle of arrival}.

[0057] To verify the technical effectiveness of this invention, a comparative experiment was conducted with traditional clustering methods, U-Net image segmentation methods, standard PointNet++ clustering methods, and end-to-end radar signal sorting methods. Specifically, the experimental conditions were: 8 signal sources of different systems (including four pulse repetition interval modulation types: fixed, staggered, jittery, and slip-mode), a signal density of 500,000 pulses / second, a signal-to-noise ratio of 10dB, and a pulse loss rate of 20%. PDW parameters included measurement errors (carrier frequency ±1MHz, pulse width ±0.1μs, angle of arrival ±1 degree).

[0058] The experimental results are shown in Table 2: Table 2. Comparison of sorting performance of different methods As can be seen, the end-to-end radar signal sorting method proposed in this invention outperforms other methods in evaluation metrics such as sorting accuracy, pulse overlap scene accuracy, and PRI modulation recognition accuracy. Specifically, in terms of PRI modulation recognition accuracy, traditional clustering methods achieve only 74.6%, U-Net image segmentation methods achieve 82.3%, and the standard PointNet++ clustering method achieves 87.9%, while the method of this invention reaches 93.5%. Furthermore, compared with the standard PointNet++ clustering method, the sorting accuracy is improved by 7.1 percentage points, the PRI modulation recognition accuracy reaches 93.5%, and the pulse overlap scene accuracy is improved by 12.1 percentage points.

[0059] In one embodiment, such as Figure 2 As shown, signal sorting is performed on the same radiation source A pulse and radiation source B pulse using both the method without timing consistency constraints and the method with timing consistency constraints proposed in this embodiment. It can be seen that signal sorting without timing consistency constraints results in instance fragmentation, that is, the same radiation source is divided into multiple instances; while using timing consistency constraints can reduce fragmentation based on timing continuity constraints, so that the same radiation source pulse remains continuous.

[0060] Example 2 In this embodiment, an end-to-end radar signal sorting system based on point cloud segmentation is proposed, which applies an end-to-end radar signal sorting method based on point cloud segmentation, including: Data acquisition module: Acquires PDW pulse sequences arranged in chronological order of pulse arrival time; Parameter extraction module: Extracts parameters from each pulse in the PDW pulse sequence to obtain carrier frequency, pulse width, pulse arrival time, and angle of arrival. Normalizes the arrival time of each pulse based on the pulse arrival time of the first pulse in the PDW pulse sequence to generate timing codes and maintain the relative time interval relationship between each pulse. Four-dimensional spatiotemporal point cloud mapping module: Establish a four-dimensional spatiotemporal point cloud space, using the carrier frequency, pulse width, and angle of arrival corresponding to each pulse as the spatial coordinates of the point cloud data, and using the timing code corresponding to each pulse as the temporal coordinates of the point cloud data. Based on the spatial coordinates and temporal coordinates, each pulse is mapped to point cloud data in the four-dimensional spatiotemporal point cloud space. Temporally Enhanced Instance Segmentation Module: Inputs point cloud data into the temporally enhanced instance segmentation network for segmentation. Extracts spatiotemporal joint features based on spatial and temporal coordinates within the local temporal neighborhood of the point cloud data, and outputs the signal source instance label corresponding to each pulse based on the spatiotemporal joint features. Signal source sorting module: Aggregates pulses from the same signal source instance tags to obtain the sorting results for each signal source.

[0061] In this embodiment, the time-enhanced instance segmentation module includes the following modules: The Adaptive Density Sensing Sampling Module (ADA-FPS) is used to adaptively adjust the sampling density according to the local density of the point cloud. It increases the number of sampling points in high-density areas to capture fine structures and reduces the number of sampling points in low-density areas to avoid oversampling. This overcomes the shortcomings of the standard FPS algorithm, which assumes a uniform distribution of point cloud density and is not adapted to the non-uniform density characteristics of signal pulses. The multi-scale grouping and feature extraction module groups points based on both Euclidean distance in parameter space and temporal proximity. It extracts parameter features and temporal features from the point set within each group, overcoming the shortcomings of the standard PointNet++ clustering method, which only groups points based on Euclidean distance and cannot associate temporally adjacent but parameter-discrepant pulses. The temporal coding module encodes the local temporal neighborhood of each point and extracts the pulse repetition interval (PRI) variation pattern features as additional features of the point, overcoming the deficiency of existing methods that cannot utilize the PRI modulation pattern as a key distinguishing dimension. The feature propagation layer is used to progressively upsample the downsampled global features and propagate them back to the original point cloud size; The segmentation head is used to perform instance prediction on the features of each point and output the signal source instance label for each point. The instance prediction is trained using a temporal consistency constraint loss function, which utilizes the physical prior of the continuous temporal distribution of signal pulses to reduce instance segmentation fragmentation.

[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Based on the technical essence of the present invention, any simple modifications, equivalent substitutions, and improvements made to the above embodiments within the spirit and principles of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An end-to-end radar signal sorting method based on point cloud segmentation, characterized in that, Including the following: S1. Obtain the PDW pulse sequence arranged in order of pulse arrival time; S2. Extract parameters from each pulse in the PDW pulse sequence to obtain the carrier frequency, pulse width, pulse arrival time, and angle of arrival. Normalize the arrival time of each pulse based on the pulse arrival time of the first pulse in the PDW pulse sequence to generate a timing code and maintain the relative time interval relationship between each pulse. S3. Establish a four-dimensional spatiotemporal point cloud space. Use the carrier frequency, pulse width and angle of arrival corresponding to each pulse as the spatial coordinates of the point cloud data, and use the timing code corresponding to each pulse as the temporal coordinates of the point cloud data. Based on the spatial coordinates and temporal coordinates, map each pulse to point cloud data in the four-dimensional spatiotemporal point cloud space. S4. Input the point cloud data into the temporal enhanced instance segmentation network for segmentation. Extract the spatiotemporal joint features based on the spatial and temporal coordinates in the local temporal neighborhood of the point cloud data, and output the signal source instance label corresponding to each pulse based on the spatiotemporal joint features. S5. Aggregate the pulses of the same signal source instance tags to obtain the sorting results of each signal source.

2. The end-to-end radar signal sorting method based on point cloud segmentation according to claim 1, characterized in that, Obtain the PDW pulse sequence arranged in chronological order of pulse arrival time, including: Acquire PDW pulse stream data, preprocess the PDW pulse stream data, and sort the preprocessed PDW pulse stream data in ascending order according to the pulse arrival time to form a PDW pulse sequence; Preprocessing includes deduplication and outlier removal.

3. The end-to-end radar signal sorting method based on point cloud segmentation according to claim 1, characterized in that, The arrival times of each pulse are normalized based on the arrival time of the first pulse in the PDW pulse sequence, specifically as follows: Obtain the minimum pulse arrival time in the PDW pulse sequence. and maximum value Calculate the number in sequence Timing encoding of each pulse : ; in, Indicates the first The pulse arrival time of each pulse.

4. The end-to-end radar signal sorting method based on point cloud segmentation according to claim 3, characterized in that, Each pulse is mapped to point cloud data in a four-dimensional spatiotemporal point cloud space based on spatial and temporal coordinates. ,in, Indicates the first The carrier frequency of each pulse This represents the pulse width of the i-th pulse. Indicates the first The angle of arrival of each pulse; also includes: The point cloud data is normalized using the maximum and minimum values ​​to obtain normalized point cloud data. : ； ； ； ； in, and These are the minimum and maximum values ​​of all pulse carrier frequencies, respectively; and These are the minimum and maximum values ​​of all pulse widths, respectively; and These are the minimum and maximum angles of arrival for all pulses, respectively. and These are the minimum and maximum values ​​of all pulse timing codes, respectively.

5. The end-to-end radar signal sorting method based on point cloud segmentation according to claim 1, characterized in that, S4 includes the following: S41. Adjust the sampling frequency of the point cloud data according to the local density of the point cloud data; S42. Determine the local temporal neighborhood of each point cloud data based on its spatial and temporal coordinates. S43. Extract temporal features in the local temporal neighborhood of each point cloud data, and fuse the temporal features with the spatial features of the point cloud data to obtain spatiotemporal joint features; S44. Predict the signal source instance label corresponding to each pulse based on the spatiotemporal joint features.

6. The end-to-end radar signal sorting method based on point cloud segmentation according to claim 5, characterized in that, The process of adjusting the sampling frequency of point cloud data based on the local density of the point cloud data is as follows: Using each point cloud data as the center and a preset radius to determine the neighborhood of a sphere, the number of neighborhood points within each sphere's neighborhood is counted as the local density of each point cloud data; based on the local density of each point cloud data, the corresponding probability is determined, wherein the probability is positively correlated with the local density; Sampling points are selected sequentially according to probability, and the sampling results are filtered based on the constraint of the farthest Euclidean distance from the selected sampling points until the preset number of samples is reached.

7. The end-to-end radar signal sorting method based on point cloud segmentation according to claim 5, characterized in that, The process of determining the local temporal neighborhood of each point cloud data based on its spatial and temporal coordinates is as follows: determine the parameter spatial neighborhood based on the spatial coordinates of each point cloud data, determine the temporal neighborhood based on the temporal coordinates of each point cloud data, and take the union of the parameter spatial neighborhood and the temporal neighborhood as the local temporal neighborhood.

8. The end-to-end radar signal sorting method based on point cloud segmentation according to claim 5, characterized in that, Temporal features are extracted within the local temporal neighborhood of each point cloud data set. These temporal features are then fused with the spatial features of the point cloud data to obtain spatiotemporal joint features, specifically: For each point cloud data, adjacent pulses located within a preset time window are extracted along the time axis within a local temporal neighborhood to form the temporal neighborhood of the point cloud data; the differential arrival time sequence is calculated for the pulse sequence within the temporal neighborhood, and the differential arrival time sequence is used as the local pulse repetition interval sequence; The local pulse repetition interval sequence is encoded by a one-dimensional convolutional network, and the pulse repetition interval mean feature, pulse repetition interval change rate feature, and pulse repetition interval modulation mode feature are extracted as temporal features. The temporal features are then concatenated and fused with the spatial features to obtain the spatiotemporal joint features.

9. The end-to-end radar signal sorting method based on point cloud segmentation according to claim 5, characterized in that, The temporal-enhanced instance segmentation network is trained using a temporal consistency loss function, which includes a classification loss, an instance triplet loss, and a temporal consistency loss term. The temporal consistency loss term is used to enhance the consistency of pulse instance labels that are temporally adjacent and belong to the same signal source.

10. An end-to-end radar signal sorting system based on point cloud segmentation, characterized in that, An end-to-end radar signal sorting method based on point cloud segmentation as described in any one of claims 1-9 is provided, comprising: Data acquisition module: Acquires PDW pulse sequences arranged in chronological order of pulse arrival time; Parameter extraction module: Extracts parameters from each pulse in the PDW pulse sequence to obtain carrier frequency, pulse width, pulse arrival time, and angle of arrival. Normalizes the arrival time of each pulse based on the pulse arrival time of the first pulse in the PDW pulse sequence to generate timing codes and maintain the relative time interval relationship between each pulse. Four-dimensional spatiotemporal point cloud mapping module: Establish a four-dimensional spatiotemporal point cloud space, using the carrier frequency, pulse width, and angle of arrival corresponding to each pulse as the spatial coordinates of the point cloud data, and using the timing code corresponding to each pulse as the temporal coordinates of the point cloud data. Based on the spatial coordinates and temporal coordinates, each pulse is mapped to point cloud data in the four-dimensional spatiotemporal point cloud space. Temporally Enhanced Instance Segmentation Module: Inputs point cloud data into the temporally enhanced instance segmentation network for segmentation. Extracts spatiotemporal joint features based on spatial and temporal coordinates within the local temporal neighborhood of the point cloud data, and outputs the signal source instance label corresponding to each pulse based on the spatiotemporal joint features. Signal source sorting module: Aggregates pulses from the same signal source instance tags to obtain the sorting results for each signal source.

Images