A radar signal sorting method based on a point cloud segmentation network

Abstract

The application discloses a radar signal sorting method based on a point cloud segmentation network, and belongs to the field of electronic reconnaissance of information and communication engineering.The application solves the problems of the traditional radar signal sorting method, such as the dependence on artificial experience for parameter setting, a complex sorting process, poor sorting ability for unknown signals and difficulty in meeting the radar signal sorting requirements in the current complex electromagnetic environment.The application combines a deep learning model with radar signal sorting, realizes an end-to-end sorting scheme without artificial intervention from original PDW to EDW, and simplifies the sorting implementation process.Because the neural network can deeply mine the feature correlation between PDWs of different types of radars, the application still has good sorting ability for unknown radar signals in a non-cooperative scene.The application method can be applied to radar signal sorting.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of electronic reconnaissance of information and communication engineering, and particularly relates to a radar signal sorting method based on a point cloud segmentation network. BACKGROUND

[0002] The main function of the radar signal sorting technology is to deinterleave and demix the mixed pulse signals received by the receiver, and to distinguish different emitter types according to the differences in radar characteristic parameters, so as to provide decision information for subsequent electronic reconnaissance and electronic countermeasures. Therefore, the radar signal sorting technology is one of the key technologies of electronic countermeasures, and many scholars have conducted in-depth research on this technology and proposed many methods. At present, most of the radar signal sorting algorithms are based on pulse description word (PDW) for analysis and implementation. The PDW contains radar characteristic parameter information such as direction of arrival (DOA), carrier frequency (CF), pulse width (PW), pulse amplitude (PA), and time of arrival (TOA). The pulse grouping is completed by analyzing the differences in PDW of different pulses, and then the radar signal sorting is realized, and the sorting result is described by emitter description word (EDW).

[0003] The traditional radar signal sorting method combining the pre-sorting process and the main sorting process is widely used in practical engineering and is also the traditional radar signal sorting technology. The main function of the pre-sorting process is to dilute the original radar pulse stream density, and the clustering algorithm is generally used, such as the k-means clustering algorithm and its improved algorithm, the density-based clustering algorithm DBSCAN, etc. The main sorting process mainly counts the pulse repetition interval (PRI) of each pulse group after clustering, and completes the main sorting in cooperation with the pulse search. The main methods of this process are PRI transform method, cumulative difference histogram (CDIF), and sequential difference histogram (SDIF). The above traditional radar signal sorting methods have the following limitations:

[0004] (1) The clustering center, the number of clusters, the parameter threshold, and the tolerance, etc. need to be set in advance, and in the non-cooperative scene, the above parameters are unknown, the setting of the parameters depends on artificial experience, and at the same time, it is also unable to adapt to the complex and changeable sorting environment.

[0005] (2) The sorting process is relatively complex, and the mutual restriction between the structures leads to poor flexibility.

[0006] The rapid development of modern electronic countermeasure technology and radar makes the current electromagnetic environment more complex in time domain, frequency domain and space domain than in the past. Firstly, the radar signal pulse stream density has reached the level of millions, and the signals overlap seriously. Secondly, the widespread application of various new complex radar systems makes it more difficult to sort the overlapping radar pulse sequences. Therefore, the traditional radar signal sorting algorithm cannot meet the radar signal sorting requirements in the current complex electromagnetic environment, and the sorting ability of the traditional radar signal sorting algorithm for unknown signals is poor. SUMMARY

[0007] The purpose of the present application is to solve the problems of the traditional radar signal sorting method, such as the need for parameter setting relying on artificial experience, complex sorting process, poor sorting ability for unknown signals, and difficulty in meeting the radar signal sorting requirements in the current complex electromagnetic environment, and a radar signal sorting method based on a point cloud segmentation network is proposed.

[0008] The technical solution adopted by the present application to solve the above technical problems is: a radar signal sorting method based on a point cloud segmentation network, which specifically comprises the following steps:

[0009] Step one, PDW data of N consecutive pulses are intercepted from the original radar pulse sequence received by the receiver in time sequence to form PDW point cloud data;

[0010] The PDW data of each pulse includes TOA, PW, CF and PA data, i.e. the dimension of the PDW point cloud data is N x 4;

[0011] Step two, the PDW point cloud data is normalized;

[0012] Step three, the normalized PDW point cloud data is input into the trained point cloud segmentation network, the PDW data of each pulse is predicted to belong to the radar signal type through the trained point cloud segmentation network, and the PDW point cloud data with label information with a dimension of N x 5 is output, i.e. the radar signal type to which the normalized PDW data of each pulse belongs is labeled;

[0013] Step four, the PDW point cloud data with label information output in step three is de-normalized to obtain the original PDW point cloud data with labels;

[0014] Step five, in the original PDW point cloud data with labels obtained in step four, the PDW data with the same label is classified as the PDW data of the same type of radar signal pulse;

[0015] For any type of radar signal pulse PDW data, a clustering method is used to process the PDW data to distinguish the different radar radiation sources corresponding to this type of radar signal pulse and extract the EDW information of each radar radiation source.

[0016] Similarly, the PDW data of each type of radar signal pulse is processed.

[0017] The beneficial effects of this invention are:

[0018] This invention proposes a radar signal sorting method based on point cloud segmentation networks. It demonstrates excellent sorting performance for point-cloud signal generators (PDWs) with severe overlap in the time, frequency, and spatial domains and high pulse flow density. Furthermore, it maintains a high sorting accuracy even with a certain pulse loss rate, making it adaptable to current complex electromagnetic environments. By combining deep learning models with radar signal sorting, it achieves an end-to-end sorting scheme directly from the original PDW to the initial signal generator (EDW) without human intervention, simplifying the sorting process. Moreover, because neural networks can deeply mine the feature correlations between PDWs of different radar types, it still exhibits good sorting capabilities for unknown radar signals in non-cooperative scenarios.

[0019] Experimental results show that the method of the present invention can achieve a sorting accuracy of 97.46% when the pulse loss rate is 20%. At the same time, the method of the present invention also has the advantage of high sorting efficiency. Attached Figure Description

[0020] Figure 1 This is a flowchart of a radar signal sorting method based on a point cloud segmentation network according to the present invention;

[0021] Figure 2 This is a schematic diagram of the PointNet++ network structure;

[0022] Figure 3(a) is a schematic diagram of the original PDW pulse sequence;

[0023] Figure 3(b) is a side view of the PDW point cloud prediction results;

[0024] Figure 3(c) is a top view of the PDW point cloud prediction results;

[0025] Figure 3(d) is a front view of the PDW point cloud prediction results;

[0026] Figure 4 This is a comparison chart of the sorting effects of different sorting methods. Detailed Implementation

[0027] The application will be described in further detail below with specific embodiments in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the application.

[0028] Specific embodiment one, combination Figure 1 This embodiment is described. The radar signal sorting method based on the point cloud segmentation network in this embodiment specifically comprises the following steps:

[0029] Step one, PDW data of N continuous pulses is intercepted from the original radar pulse sequence received by the receiver in time sequence, to form PDW point cloud data;

[0030] The PDW data of each pulse comprises TOA, PW, CF and PA data, that is, the dimension of the PDW point cloud data is N x 4;

[0031] The method of the application can realize radar signal sorting only by using TOA, PW, CF and PA data in the PDW data, which reduces the amount of data to be processed compared with the prior art, and significantly improves the accuracy of radar signal sorting by combining the selected data with the PointNet++ model;

[0032] Step two, the PDW point cloud data is normalized;

[0033] Step three, the normalized PDW point cloud data is input into the trained point cloud segmentation network, the radar signal type to which each pulse of PDW data belongs is predicted through the trained point cloud segmentation network, and PDW point cloud data with label information with a dimension of N x 5 is output, that is, the radar signal type to which each pulse of PDW data after normalization belongs is labeled; the last dimension stores the label information of the radar signal type to which each pulse of PDW data belongs;

[0034] Step four, the PDW point cloud data with label information output in step three is de-normalized to obtain original PDW point cloud data with labels;

[0035] Step five, in the original PDW point cloud data with labels obtained in step four, PDW data with the same label is classified as PDW data of the same class of radar signal pulses;

[0036] For any class of radar signal pulse PDW data, a clustering method is further used to process the PDW data to distinguish different radar radiation sources corresponding to the class of radar signal pulses, and extract EDW information of each radar radiation source;

[0037] Similarly, the PDW data of each type of radar signal pulse is processed.

[0038] The PDW data with the same label is classified as a group in step five, and the radar radiation source is further divided according to the PDW information of the group, and the EDW information is formed.

[0039] The EDW information extraction method of different types of radar radiation sources is as follows:

[0040] 1. PRI conventional radar. For the PRI conventional radar, the PRI is almost fixed or the jitter is not more than 1%, and the CF and PW of the same PRI conventional radar remain unchanged or the jitter is not more than 1%. Therefore, for the PRI conventional radar, the dynamic k-means clustering algorithm can be used to perform two-level clustering on the CF and PW in turn to obtain the preliminary grouping result, and then the possible PRI value of each group result is counted, and the PRI conventional radar individuals are further divided according to the difference of the PRI value.

[0041] 2. PRI jitter radar. The PRI jitter radar is a radar whose PRI jitters around a center PRI value and the jitter rate is generally not more than 30%, and the CF and PW of the same PRI jitter radar remain unchanged or the jitter is not more than 1%. Therefore, for the PRI jitter radar, the dynamic k-means clustering algorithm can also be used to perform two-level clustering on the CF and PW in turn to obtain the preliminary grouping result, and then the center PRI value and the jitter rate of each group result are counted and calculated, and the PRI jitter radar individuals are further divided according to the difference.

[0042] 3. PRI staggered radar. The PRI staggered radar is composed of multiple sub-periodic PRIs to form a frame period, and repeats in turn within the frame period PRI. The CF and PW of the same PRI staggered radar remain unchanged or the jitter is not more than 1%. For the PRI staggered radar, the dynamic k-means clustering algorithm can also be used to perform two-level clustering on the CF and PW in turn to obtain the preliminary grouping result, and then the sub-periodic PRI value, frame period PRI value and staggered number of each group result are counted and calculated, and the different PRI staggered radar individuals are further distinguished according to the difference of the values.

[0043] 4, Frequency agile radar. The PRI value and PW of the frequency agile radar are generally fixed values, and the CF changes constantly on different pulses. The CF of the frequency agile radar randomly jumps within a certain agile range, and the agile range is 15% to 20% of the working center frequency of the agile frequency radar. For such a radiation source, the dynamic k-means clustering algorithm can be used to cluster the PW to complete the preliminary grouping, then the PRI of each group of pulse sequences is calculated and the radar radiation source individuals are further distinguished according to the difference, and finally the center frequency and agile range of each radiation source individual are calculated to complete the final grouping and form the EDW information of each radiation source individual.

[0044] 5, Pulse group frequency agile radar. The CF of the pulse group frequency agile radar changes in a frame period, and the frame period contains several subperiods, which is a special form of the agile frequency radar. For such a radiation source, the dynamic k-means clustering algorithm is also used to cluster the PW to form a preliminary grouping result, then the PRI of each group of pulse sequences is calculated and further divided according to the PRI difference, and finally the agile frequency and agile range of each radiation source individual are counted to form the EDW information of each radiation source individual.

[0045] The existing sorting method using an image segmentation network needs to map the original PDW data to an image, which has a large amount of data processing, and the limitation of picture resolution leads to the problem of point overlap in the mapping, and the processing efficiency and performance are limited. The present application first divides the original PDW data into multiple sub-clusters according to different radar types by using the PointNet++ point cloud segmentation network, realizes the grouping of high-density and severely overlapping PDW pulse streams according to the radar type, and also dilutes the original PDW pulse stream density, and then uses a clustering method to divide different radar radiation sources. In the sorting process, there is no manual intervention, and it does not depend on manual parameter setting, and there is no restriction between structures, so it can be applied to the current complex electromagnetic environment. And this method does not need to make complex transformation to the original PDW data, so the amount of data processing is small, and the lightweight structure of the model also makes the sorting efficiency significantly improved compared with the existing machine learning method, while avoiding the loss or distortion of PDW data, which can completely preserve the original PDW information. It is an end-to-end intelligent sorting scheme.

[0046] Specific implementation method two: the difference between this implementation method and specific implementation method one is that the value of the parameter N is not less than 1024.

[0047] In the present embodiment, the value of the parameter N is set to be not less than 1024, which can reduce the amount of data processing as much as possible while ensuring the sorting accuracy.

[0048] The other steps and parameters are the same as those in specific implementation method one.

[0049] Specific implementation three: the difference between this implementation and specific implementation one or two is that the specific process of step two is:

[0050] For TOA data in PDW data, the process of normalizing TOA data is:

[0051]

[0052]

[0053]

[0054] Where t i is TOA data in PDW data of the i-th pulse, t i,nor is normalized TOA data in PDW data of the i-th pulse, t means is the mean of TOA data in PDW data of N pulses, and σ is the variance of TOA data in PDW data of N pulses.

[0055] For CF, PW and PA data in PDW data, the normalization method is the same as that of TOA data.

[0056] By normalizing the size of the parameter to [-1, 1], the dimensional influence between different characteristic parameters can be eliminated, and the statistical distribution of the unified sample can be summarized, avoiding the problem of subsequent model prediction accuracy decline caused by singular point cloud data. While normalizing, the normalized parameters t means and σ of the batch of PDW point cloud data need to be retained in order to extract the original PDW point cloud data in subsequent reverse normalization.

[0057] The other steps and parameters are the same as those in specific implementation one or two.

[0058] Specific implementation four: in combination with Figure 2 This implementation is described. The difference between this implementation and one of specific implementations one to three is that the point cloud segmentation network is a PointNet++ model, and the PointNet++ model specifically includes: a first SA unit (Set Abstraction Layer, SA), a second SA unit, a third SA unit, a first FP unit (Feature Propagation, FP), a second FP unit, a third FP unit, and a fully connected layer.

[0059] As Figure 2As shown, the left half is the encoding part of the PointNet++ network, which gradually down-samples the point cloud through the encoding part, mainly realizing the local feature space coding of the point cloud; the right half is the decoding part of the PointNet++ network, which up-samples each point in the point cloud, gradually restores the structure of the original point cloud, and realizes the aggregation of the feature information of each point in the point cloud; finally, the semantic discrimination of each point is performed through the full connection layer.

[0060] In the encoding part, the purpose of the SA unit structure is to learn the structure and geometric information between points, encode it to partial points through down-sampling, and then make each point aggregate its neighborhood information through subsequent decoding operations. The normalized PDW point cloud data is first passed through the first SA unit, 256 points that are farthest from each other are randomly sampled from the normalized PDW point cloud data points, then local spaces are constructed by finding neighborhood points, each space contains 16 points, and finally the feature information of each space is obtained through the PointNet layer. The PDW point cloud data obtains 256 sampling points and the global features of the local space of each sampling point through the first SA unit. After obtaining the overall feature of the local space, it will be used in the subsequent decoding layer. The 256 sampling points continue to pass through the second SA unit and the third SA unit to construct deeper features for the decoding part.

[0061] The main purpose of the decoding part is to map the local space information aggregated by the encoding to all point clouds in the point cloud set to obtain the point features of each original PDW point cloud in the PDW point cloud. The specific implementation process is to use the FP unit to propagate the features of the points from the previous unit to the next unit until the original PDW point cloud data is restored.

[0062] The other steps and parameters are the same as one of the first to third embodiments.

[0063] Embodiment five: different from one of the first to fourth embodiments, the structure of the first SA unit, the second SA unit and the third SA unit is the same, and each SA unit includes a sampling layer, a grouping layer and a PointNet layer.

[0064] Each SA unit is composed of three associated network structures of a sampling layer, a grouping layer and a PointNet layer, wherein the sampling layer (Sampling Layer) is used for point cloud sampling point selection, the grouping layer (Grouping Layer) is used for constructing a local space, and the PointNet layer is used for feature aggregation.

[0065] The other steps and parameters are the same as one of the first to fourth embodiments.

[0066] Specific implementation six: different from one of the specific implementations one to five, the structures of the first FP unit, the second FP unit and the third FP unit are the same, each FP unit includes an IDW layer, a Concate layer and a PointNet layer;

[0067] The normalized PDW point cloud data is taken as the input of the first SA unit, and the output of the first SA unit is obtained;

[0068] The output of the first SA unit is taken as the input of the second SA unit, and the output of the second SA unit is obtained;

[0069] The output of the second SA unit is taken as the input of the third SA unit, and the output of the third SA unit is obtained;

[0070] The output of the third SA unit is taken as the input of the IDW layer of the first FP unit, the output of the IDW layer of the first FP unit is spliced with the output of the third SA unit, and the spliced result is taken as the input of the Concate layer of the first FP unit;

[0071] The output of the PointNet layer of the first FP unit is taken as the input of the IDW layer of the second FP unit, the output of the IDW layer of the second FP unit is spliced with the output of the second SA unit, and the spliced result is taken as the input of the Concate layer of the second FP unit;

[0072] The output of the PointNet layer of the second FP unit is taken as the input of the IDW layer of the third FP unit, the output of the IDW layer of the third FP unit is spliced with the output of the first SA unit, and the spliced result is taken as the input of the Concate layer of the third FP unit;

[0073] The output of the PointNet layer of the third FP unit is further taken as the input of the full connection layer, and the radar signal type prediction result of the PDW data of each pulse in the normalized PDW point cloud data is output through the full connection layer.

[0074] The output of the IDW layer is taken as the input of the PointNet layer in the same FP unit, the IDW layer is used to calculate the interpolation features of each point of the shallow network, the Concate layer is used for feature splicing, and the PointNet layer is used for feature aggregation.

[0075] Each hierarchical structure includes a PointNet layer, and the component structures of the PointNet layers in different hierarchical structures are as shown in Table 1:

[0076] Table 1: PointNet structure of each level

[0077]

[0078] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0079] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that the interpolation method for the IDW layer is as follows:

[0080]

[0081]

[0082] In the formula, f (j) (x) is the j-th eigenvalue of the unknown point, f l (j) Let C be the eigenvalue of the l-th known point, where C is the number of features, l = 1, 2, ..., L, and L represents the number of known points selected from the set of known points. l (x) is the weight between the l-th known point and the unknown point. It is inversely proportional to the distance between the unknown point and the known point; the closer the distance, the greater the influence. d(x,x) l ) is the distance between the unknown point and the l-th known point, and p represents the influence factor of distance on weight.

[0083] This invention employs trilinear interpolation, where p=2 and k=3. The three points closest to the unknown point are selected from the known point set for interpolation calculation, resulting in interpolated features of 64 points. These interpolated features are then concatenated with the interpolated features obtained in the second SA unit to generate new features, which are then aggregated using a PointNet layer. Subsequently, the point features obtained from the first FP unit are passed sequentially through the second and third FP units to obtain new point features. Finally, the features obtained from the third FP unit are processed by a fully connected layer to obtain the prediction result for the PDW data of each pulse. The network output is PDW point cloud data with prediction labels.

[0084] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0085] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One through Seven in that the training process of the PointNet++ model is as follows:

[0086] PDW data of pulse sequences from a single radar source and PDW data of pulse sequences from multiple radar sources (2, 3, or more radars can be collected to meet the diversity of training data) are collected separately. Then, PDW data of each pulse sequence is extracted to form PDW point cloud data. The PDW point cloud data is then normalized and labeled to obtain the dataset.

[0087] The obtained data set is randomly divided into a training set, a validation set and a test set in a ratio of 8:1:1, the PointNet++ model is trained by using the training set, the effect of training the PointNet++ model is evaluated and adjusted by using the validation set, and the training is stopped and the PointNet++ model is saved when the prediction accuracy of the PointNet++ model on the test set reaches a threshold value, and the saved PointNet++ model is the trained PointNet++ model.

[0088] The hyperparameter setting of the PointNet++ model during training is shown in Table 2:

[0089] Table 2 Model Hyperparameter Setting

[0090]

[0091] The model parameter optimization function is selected as Adam to adaptively adjust the learning rate of each parameter. The initial learning rate is 0.001, the total number of training set samples is 1800, the batch processing sample number is 16, and the training iteration number is 120 to ensure that the training samples are fully covered, and the training is stopped and the successfully trained model is saved when the prediction accuracy meets the demand.

[0092] The other steps and parameters are the same as one of the first to seventh embodiments.

[0093] Experimental Part

[0094] (1) Feasibility verification. A group of PDW pulse sequences are randomly generated, the pulse number is 1024, which contains PRI conventional radar, PRI jitter radar, PRI staggered radar, agile frequency radar and pulse group agile radar, each type of radar has two, a total of ten radars. After visualizing TOA, PW and CF as coordinates, as shown in Figure 3(a).

[0095] Considering that each characteristic parameter of PDW exists a certain jitter in the actual sorting environment, therefore, PW, CF and PA are set to have a certain jitter in the simulation process. After being predicted by the PointNet++ model, the total number of successful and correct sorting is 1015, and the overall sorting accuracy is 99.12%, the sorting accuracy of different types of radars is shown in Table 3, and the visualization results are shown in Figures 3(b), 3(c) and 3(d). Through the experiment, the feasibility of the method of the present application for multi-radar signal sorting is verified.

[0096] Table 3 Sorting Accuracy of Each Type of Radar

[0097]

[0098] (2) Reliability verification. In order to further verify the reliability of the method of the application in sorting under complex environment, a group of PDW pulse sequences are randomly generated, the number of pulses is 1024, including the radar types and the number contained in the feasibility verification process, and each radar parameter also has a certain jitter. The difference is that the pulse has a 20% loss rate, which simulates the PDW received under complex sorting environment. After the prediction of the PointNet++ model, the total number of successful and correct sorting pulses is 998, and the sorting accuracy is 97.46%, and the sorting accuracy of different types of radars is shown in Table 4. It can be seen that under the condition of high pulse loss rate, the method of the application can still maintain a high sorting accuracy. For PRI conventional radar, PRI staggered radar and pulse group frequency agile radar, this type of radar feature parameter presents a certain regularity, and under the condition of high pulse loss rate, the regularity is partially destroyed, and the traditional method is difficult to correctly sort. And because the method of the application deeply excavates and combines the local features and global features between each PDW, the prediction pays attention to both local features and global features, and the two can complement each other. Therefore, under the condition of high pulse loss rate, the method of the application can still correctly sort.

[0099] Table 4: Sorting accuracy of each type of radar

[0100]

[0101] (3) Comparison of sorting methods and performance analysis. In order to further illustrate the sorting performance of the method of the application, and compare the sorting effect of the method of the application with other sorting methods, under different pulse loss rates, the same PDW pulse sequence is sorted by different methods, and the relationship curve of the sorting accuracy of different methods with the change of pulse loss rate is obtained, as shown in Figure 4 From Figure 4 , it can be seen that the traditional method has low sorting accuracy under the condition of high pulse loss rate, and is sensitive to pulse loss. The sorting effect of the method of the application under the same condition is better than that of other sorting methods, and is not sensitive to pulse loss, and still maintains a high sorting accuracy under complex environment. It shows that the sorting performance of the method of the application is better than that of other methods, can adapt to the current complex electromagnetic environment, and meets the sorting needs of the current electronic warfare.

[0102] The above examples of the application are only used to illustrate the calculation model and calculation process of the application, and are not limited to the embodiments of the application. Based on the above description, those skilled in the art can make other different forms of changes or changes, and it is impossible to enumerate all the embodiments here. Any obvious changes or changes derived from the technical solutions of the application still fall within the protection scope of the application.

Claims

1. A radar signal sorting method based on point cloud segmentation networks, characterized in that, The method specifically includes the following steps: Step 1: Extract N consecutive PDW data from the original radar pulse sequence received by the receiver in chronological order to form PDW point cloud data; Each pulse's PDW data includes TOA, PW, CF, and PA data, meaning the PDW point cloud data has a dimension of N×4. Step 2: Normalize the PDW point cloud data; Step 3: Input the normalized PDW point cloud data into the trained point cloud segmentation network. Predict the radar signal type of each pulse's PDW data through the trained point cloud segmentation network, and output PDW point cloud data with N×5 dimension and label information, that is, label the radar signal type of each pulse's PDW data after normalization. The point cloud segmentation network is a PointNet++ model, which specifically includes: a first SA unit, a second SA unit, a third SA unit, a first FP unit, a second FP unit, a third FP unit, and a fully connected layer; The first SA unit, the second SA unit, and the third SA unit have the same structure, and each SA unit includes a sampling layer, a grouping layer, and a PointNet layer; The first FP unit, the second FP unit, and the third FP unit have the same structure, and each FP unit includes an IDW layer, a Concaten layer, and a PointNet layer; Step 4: Perform inverse normalization on the PDW point cloud data with label information output in Step 3 to obtain the original PDW point cloud data with labels. Step 5: In the original PDW point cloud data with labels obtained in Step 4, PDW data with the same label are classified as PDW data of the same type of radar signal pulse; For any type of radar signal pulse PDW data, a clustering method is used to process the PDW data to distinguish the different radar radiation sources corresponding to this type of radar signal pulse and extract the EDW information of each radar radiation source. Similarly, the PDW data of each type of radar signal pulse is processed.

2. The radar signal sorting method based on point cloud segmentation network according to claim 1, characterized in that, The value of N is not less than 1024.

3. The radar signal sorting method based on point cloud segmentation network according to claim 2, characterized in that, The specific process of step two is as follows: For TOA data in PDW data, the process of normalizing TOA data is as follows: （1） （2） （3） in, For the first cut TOA data in the PDW data of each pulse For the normalized first TOA data in the PDW data of each pulse for The mean of TOA data in the PDW data of each pulse for The variance of TOA data in the PDW data of each pulse; For the CF, PW, and PA data in the PDW data, the normalization method is the same as that for the TOA data.

4. The radar signal sorting method based on point cloud segmentation network according to claim 3, characterized in that, The normalized PDW point cloud data is used as the input to the first SA unit to obtain the output of the first SA unit; The output of the first SA unit is used as the input of the second SA unit to obtain the output of the second SA unit; The output of the second SA unit is used as the input of the third SA unit to obtain the output of the third SA unit; The output of the third SA unit is used as the input of the IDW layer of the first FP unit. The output of the IDW layer of the first FP unit is concatenated with the output of the third SA unit, and the concatenation result is used as the input of the Concate layer of the first FP unit. The output of the PointNet layer of the first FP unit is used as the input of the IDW layer of the second FP unit. The output of the IDW layer of the second FP unit is concatenated with the output of the second SA unit. The concatenation result is used as the input of the Concate layer of the second FP unit. The output of the PointNet layer of the second FP unit is used as the input of the IDW layer of the third FP unit. The output of the IDW layer of the third FP unit is concatenated with the output of the first SA unit, and the concatenation result is used as the input of the Concate layer of the third FP unit. Then, the output of the PointNet layer of the third FP unit is used as the input of the fully connected layer, and the radar signal type prediction result of each pulse of the PDW data in the normalized PDW point cloud data is obtained through the output of the fully connected layer.

5. The radar signal sorting method based on point cloud segmentation network according to claim 4, characterized in that, The interpolation method for the IDW layer is as follows: （4） （5） In the formula, For the unknown point 1 eigenvalue, For the selected first The first known point 1 eigenvalue, It is the characteristic number. , This represents the number of known points selected from the set of known points. It is the first The weights of known and unknown points It is an unknown point and the first Distance between known points This represents the influence factor of distance on weight.

6. The radar signal sorting method based on point cloud segmentation network according to claim 5, characterized in that, The training process of the PointNet++ model is as follows: PDW data of pulse sequences from a single radar source and PDW data of pulse sequences from multiple radar sources were collected separately. Then, PDW data of each pulse sequence was extracted to form PDW point cloud data. The PDW point cloud data was then normalized and labeled to obtain the dataset. The obtained dataset is randomly divided into three parts: training set, validation set, and test set in a ratio of 8:1:

1. The PointNet++ model is trained using the training set, and the training effect of the PointNet++ model is evaluated and adjusted using the validation set. Training is stopped and the PointNet++ model is saved when the prediction accuracy of the PointNet++ model on the test set reaches the threshold. The saved PointNet++ model is the trained PointNet++ model.

Images