Radar pulse cooperative sorting method based on depth correlation and inter-pulse feature fusion

Abstract

The application discloses a radar pulse cooperative sorting method based on depth correlation and inter-pulse feature fusion, and comprises the following steps: performing data coding operation on double-platform PDW data according to a data coding module, so as to ensure complete coding of pulses; performing feature extraction on double-platform features according to a depth time difference feature extraction branch, performing correlation analysis on the extracted features, performing dimension splicing on the analysis results after mean calculation, obtaining a four-dimensional feature body, matching sample points with close time differences according to the four-dimensional feature body, and determining final depth TDOA features; capturing long-range dependence in a pulse sequence by adopting a double-layer bidirectional LSTM network on main platform data according to an inter-pulse sequence feature extraction branch, and extracting inter-pulse features; fusing the depth TDOA features and the inter-pulse features through a classification module, mapping to radar number labels by using a linear layer, and finally mapping the labels corresponding to the coding sequence to original PDW labels to obtain final sorting results.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar electronic reconnaissance, and particularly relates to a radar pulse cooperative sorting method based on deep correlation and inter-pulse feature fusion. BACKGROUND

[0002] The existing radar pulse sorting methods mainly include two categories: one category is a sorting method based on a single reconnaissance platform, such as a cumulative difference histogram (CDIF) algorithm, a sequence difference histogram (SDIF) algorithm, a PRI (Pulse Repetition Interval) transformation algorithm and other PRI-based sorting algorithms; or a multi-parameter clustering sorting method such as K-means clustering, density-based clustering, Bayesian clustering, fuzzy clustering, incremental clustering, deep clustering and the like; and a recent deep learning sorting method based on recurrent neural networks (RNNs), image segmentation networks and the like. However, in the modern complex electromagnetic environment, the received radar pulse sequence often has complex inter-pulse modulation types and non-ideal conditions such as measurement errors, missing pulses and false pulses. The method based on a single reconnaissance platform has less available features and cannot achieve good sorting results.

[0003] The other category is a sorting method based on multiple reconnaissance platforms, which matches the PDWs (Pulse Description Words) intercepted by different platforms through specific constraint rules to obtain the time difference of arrival (TDOA) of the matched pulse pairs. For example, a time difference window (TDW) is combined with the PDW parameter distance to match the pulses, and then recursive histogram statistics, grid clustering or evaluation of the similarity between the TDOA vectors of the pulse pairs are performed to realize sorting. When there are at least three reconnaissance platforms and the TDOA vectors of different radars are different, these methods usually perform well. However, when there are fewer available platforms, time difference ambiguity is prone to occur, resulting in sorting missing batches. In addition, the clustering distance threshold has a significant impact on the deinterleaving performance. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a radar pulse cooperative sorting method based on deep correlation and inter-pulse feature fusion, which solves the above technical problems.

[0005] To achieve the above purpose, the present application provides the following technical solutions:

[0006] A radar pulse cooperative sorting method based on deep correlation and inter-pulse feature fusion, comprising:

[0007] Build a dual-platform radar pulse deinterleaving network to complete the collaborative sorting of dual-platform data;

[0008] The dual-platform radar pulse deinterleaving network includes a data encoding module, an inter-pulse sequence feature extraction branch, a deep time difference feature extraction branch, and a classification module; the deep time difference feature extraction branch includes a single-platform feature extraction module, a bidirectional group correlation module, and a matching module.

[0009] The dual-platform radar pulse deinterleaving network is used to complete the collaborative sorting of dual-platform data, including:

[0010] Step 1: Perform data encoding on the dual-platform PDW data according to the data encoding module to ensure that the time intervals between input data samples are equal;

[0011] Step 2: Extract branches based on the deep time difference feature, perform feature extraction on the dual platform features, perform feature correlation analysis, calculate the mean of the analysis results, and then perform dimension splicing to obtain a four-dimensional feature volume. Match sample points with close time differences based on the four-dimensional feature volume to determine the final deep TDOA feature.

[0012] Step 3: Extract branches based on the inter-pulse sequence features. A two-layer bidirectional LSTM network is used on the main platform data to capture the long-range dependencies in the pulse sequence and extract the inter-pulse features.

[0013] Step 4: The deep TDOA features and inter-pulse features are fused through the classification module and mapped to the radar number label using a linear layer. Finally, the label corresponding to the coding sequence is mapped to the original PDW label to obtain the final sorting result.

[0014] Furthermore, a radar pulse collaborative sorting method based on deep correlation and inter-pulse feature fusion also includes: the dual platform includes a main platform and a sub-platform; in the process of completing the collaborative sorting of the dual-platform data in the radar pulse deinterleaving network based on the dual platform, the PDW data intercepted by the sub-platform is sent to the main platform for unified processing.

[0015] Furthermore, step 1 includes:

[0016] Step 11: Normalize the radio frequency value of the PDW data to be between 0.1 and 1;

[0017] Step 12: For each pulse x i , calculate the unit time points of its start and end, including [toa j / Δt] and [(toa j +pw j ) / Δt];

[0018] Step 13: Between the start and end time units, normalized radio frequency values ​​and labels are assigned in sequence to complete the encoding of the pulse data.

[0019] Furthermore, step 2 includes:

[0020] Step 21: Perform feature extraction on the data of the two platforms using the single-platform feature extraction module to obtain features of platform A and platform B, respectively; platform A is the primary platform and platform B is the secondary platform;

[0021] Step 22: Based on the bidirectional group correlation module, perform correlation analysis on the features of platform A and the features of platform B, and merge the analysis results to obtain a four-dimensional feature volume;

[0022] Step 23: Based on the matching module, sample points with close time differences are matched according to the four-dimensional feature body to obtain the final deep TDOA feature.

[0023] Further, step 21 includes:

[0024] Step 211: Use a pair of ResNet-like networks with shared parameters to map the data from the two platforms into a high-dimensional feature space:

[0025]

[0026] Among them, A represents the main platform in the dual platform, Conv1 represents the convolution operation of the first layer, including multiple groups of convolution layers with batch normalization and ReLU activation functions, represents the l-layer convolutional network of platform A, t A Represents the encoded input data;

[0027] Step 212: Connect the outputs of different layers of the convolutional network in step 212 to extract the features of platform A:

[0028]

[0029] Among them, f A represents the extracted features of platform A;

[0030] Step 213: According to the method of steps 211 and 212, the feature f of platform B is extracted. B .

[0031] Further, step 22 includes:

[0032] Step 221: Divide the feature f along the channel dimension A and f B N g groups, so that each feature group has N c / Ng channels;

[0033] Step 222: Based on the principle that the TDOA between the dual-platform data can be positive or negative, the bidirectional group correlation feature of the g-th feature group is calculated:

[0034]

[0035] Among them, <·,·> represents the inner product, D max Indicates the maximum time difference level;

[0036] Merge the related features of all groups into a bidirectional intra-group related feature body

[0037] Step 223: Take the average along the time difference level dimension and copy the mean back to the original dimension to get Finally and Splicing along the feature group dimension to obtain a four-dimensional feature body

[0038] Further, step 23 includes:

[0039] The improved U-net architecture is used to analyze the four-dimensional feature volume Process it and upsample its output features to restore the original time resolution to obtain the final deep TDOA feature f tdoa .

[0040] The beneficial effects of the present invention are:

[0041] The present invention proposes a radar pulse collaborative sorting method based on deep correlation and inter-pulse feature fusion, introduces deep neural networks into multi-platform collaborative sorting tasks, and the proposed DPCD-Net can simultaneously extract deep TDOA features and inter-pulse sequence features. The sequence feature extraction branch captures the changing relationship between pulses from the data of the main platform. At the same time, the deep TDOA feature extraction branch adopts a bidirectional group correlation (Bi-GWC) module to perform bidirectional displacement correlation calculation along the time dimension to explore the relationship between the data of the two platforms; compared with the existing single-platform sorting method, the present invention combines multi-platform information and can adapt to more complex radar signal sorting scenarios. Compared with the traditional sorting method based on TDOA vectors, the present invention can extract more effective deep TDOA features and has better sorting performance. At the same time, since the present invention combines inter-pulse features and deep TDOA features at the same time, the present invention can also show better effects for radiation sources with time difference ambiguity.

[0042] Other advantages, objectives, and features of the present invention will be described in the following description and will be apparent to those skilled in the art to some extent, or may be taught by those skilled in the art from the practice of the present invention. The purposes and other advantages of the present invention may be realized and obtained through the structures particularly pointed out in the written description and the accompanying drawings.

[0043] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0045] Figure 1 Flowchart of a radar pulse collaborative sorting method based on depth correlation and inter-pulse feature fusion in an embodiment of the present invention;

[0046] Figure 2 Schematic diagram of the structure of a radar pulse deinterleaving network based on a dual-platform in a radar pulse collaborative sorting method based on depth correlation and inter-pulse feature fusion in an embodiment of the present invention;

[0047] Figure 3 This is a module diagram of a bidirectional group correlation module in a radar pulse collaborative sorting method based on depth correlation and inter-pulse feature fusion in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0049] like Figure 1 As shown, the present invention proposes a radar pulse collaborative sorting method based on depth correlation and inter-pulse feature fusion, comprising:

[0050] Build a dual-platform radar pulse deinterleaving network to complete the collaborative sorting of dual-platform data;

[0051] Among them, the radar pulse deinterleaving network based on dual platforms includes a data encoding module, an inter-pulse sequence feature extraction branch, a deep time difference feature extraction branch and a classification module; the deep time difference feature extraction branch includes a single platform feature extraction module, a bidirectional group correlation module and a matching module.

[0052] The dual-platform radar pulse deinterleaving network is used to complete the collaborative sorting of dual-platform data, including:

[0053] Step 1: Perform data encoding on the dual-platform PDW data according to the data encoding module to ensure that the time intervals between input data samples are equal;

[0054] Step 2: Extract branches based on the deep time difference feature, perform feature extraction on the dual platform features, perform feature correlation analysis, calculate the mean of the analysis results, and then perform dimension splicing to obtain a four-dimensional feature volume. Match sample points with close time differences based on the four-dimensional feature volume to determine the final deep TDOA feature.

[0055] Step 3: Extract branches based on the inter-pulse sequence features. A two-layer bidirectional LSTM network is used on the main platform data to capture the long-range dependencies in the pulse sequence and extract the inter-pulse features.

[0056] Step 4: The deep TDOA features and inter-pulse features are fused through the classification module and mapped to radar number labels using a linear layer. Finally, the labels corresponding to the coding sequence are mapped to the original PDW labels to obtain the final sorting results.

[0057] In actual operation, step 2 and step 3 are performed in parallel.

[0058] Among them, the dual platform includes a main platform and a sub-platform; in the process of collaborative sorting of dual-platform data based on the radar pulse deinterleaving network of the dual platforms, the PDW data intercepted by the sub-platform is sent to the main platform for unified processing.

[0059] Step 1 includes:

[0060] Step 11: Normalize the radio frequency value of the PDW data to be between 0.1 and 1;

[0061] Step 12: For each pulse x i , calculate the unit time points of its start and end, including [toa j / Δt] and [(toa j +pw j ) / Δt];

[0062] Step 13: Between the start and end time units, normalized radio frequency values ​​and labels are assigned in sequence to complete the encoding of the pulse data.

[0063] Step 2 includes:

[0064] Step 21: Perform feature extraction on the data of the two platforms using the single-platform feature extraction module to obtain features of platform A and platform B, respectively; platform A is the primary platform and platform B is the secondary platform;

[0065] Step 22: Based on the bidirectional group correlation module, perform correlation analysis on the features of platform A and the features of platform B, and merge the analysis results to obtain a four-dimensional feature volume;

[0066] Step 23: Based on the matching module, sample points with close time differences are matched according to the four-dimensional feature body to obtain the final deep TDOA feature.

[0067] Step 21 includes:

[0068] Step 211: Use a pair of ResNet-like networks with shared parameters to map the data from the two platforms into a high-dimensional feature space:

[0069]

[0070] Among them, A represents the main platform in the dual platform, Conv1 represents the convolution operation of the first layer, including multiple groups of convolution layers with batch normalization and ReLU activation functions, represents the l-layer convolutional network of platform A, t A Represents the encoded input data;

[0071] Step 212: Connect the outputs of different layers of the convolutional network in step 212 to extract the features of platform A:

[0072]

[0073] Among them, f A represents the extracted features of platform A;

[0074] Step 213: According to the method of steps 211 and 212, the feature f of platform B is extracted. B .

[0075] Step 22 includes:

[0076] Step 221: Divide the feature f along the channel dimension A and f B N g groups, so that each feature group has N c / N g channels;

[0077] Step 222: Based on the principle that the TDOA between the dual-platform data can be positive or negative, the bidirectional group correlation feature of the g-th feature group is calculated:

[0078]

[0079] Among them, <·,·> represents the inner product, D max Indicates the maximum time difference level;

[0080] Merge the related features of all groups into a bidirectional intra-group related feature body

[0081] Step 223: Take the average along the time difference level dimension and copy the mean back to the original dimension to get Finally and Splicing along the feature group dimension to obtain a four-dimensional feature body

[0082] Step 23 includes:

[0083] The improved U-net architecture is used to analyze the four-dimensional feature volume Process it and upsample its output features to restore the original time resolution to obtain the final deep TDOA feature f tdoa .

[0084] The working principle and beneficial effects of the above technical solution are as follows: The purpose of the present invention is to provide a radar pulse cooperative sorting method based on deep correlation and inter-pulse feature fusion, which sends the PDW data intercepted by the secondary platform to the main platform for unified processing, and explores stable TDOA features and complex inter-pulse modulation features; based on this, the present invention designs a dual-platform cooperative deinterleaving network (Dual-Platform Cooperative Deinterleaving Network, DPCD-Net), and the overall network framework is detailed in [1]. Figure 2 , which mainly includes two feature extraction branches: the deep TDOA feature extraction branch based on dual-platform data and the inter-pulse sequence feature extraction branch based on main platform data, which can simultaneously extract and fuse the deep features of the two branches.

[0085] First, in order to mine the correlation between the data of the two platforms, the data needs to be uniformly time-encoded. The specific steps adopted by the present invention are as follows:

[0086] 1. Normalize the RF value of the PDW data to be between 0.1 and 1.

[0087] 2. For each pulse x i , calculate the unit time points of its start and end [toa j / Δt] and [(toa j +pw j ) / Δt];

[0088] 3. Between the start and end time units, normalized RF values ​​and labels are assigned in sequence to complete the encoding of the pulse data.

[0089] The specific encoding process in the above steps is relatively mature in the prior art and will not be described in detail here. For the specific encoding process, it is preferred to refer to the relevant encoding process in the prior art "Z. Kang, Y. Zhong, Y. Wu, and Y. Cai, "Signal deinterleaving based on u-net networks," in 2023 8th International Conference on Computer and Communication Systems (ICCCS), 2023, pp. 62–67."

[0090] The specific steps of deep TDOA feature extraction branch are as follows:

[0091] 1. Single-platform feature extraction module. This paper uses a pair of ResNet-like networks with shared parameters to map data from two platforms into a high-dimensional feature space (taking platform A as an example):

[0092]

[0093] Conv1 represents the convolution operation of the first layer, including multiple groups of convolution layers with batch normalization and ReLU activation functions. A represents the encoded input data. At the same time, in order to better extract the relevant features between the data of the two platforms, it is necessary to capture as much useful information as possible from the data of each platform separately. Therefore, we connect the outputs of different layers of the network to make full use of information at different levels:

[0094]

[0095] Among them, f A is the extracted feature of platform A.

[0096] 2. Bidirectional Group-Wise Correlation (Bi-GWC) module. For details on the module diagram, see Figure 3 The process of estimating TDOA is to perform correlation analysis on the data of the two platforms. The present invention uses a deep learning model to construct a correlation feature body to extract deep TDOA features. The features obtained by the single-platform feature extraction module are subjected to bidirectional shift correlation calculation along the time dimension to explore the relationship between the data of the two platforms at different time differences.

[0097] Specifically, the bidirectional group correlation module first divides the feature f along the channel dimension A and f B Ng groups, so each feature group has N c / N g channels. Since the TDOA between the two platform data can be positive or negative, the bidirectional group correlation feature of the g-th feature group is calculated as follows:

[0098]

[0099] Where <·,·> represents the inner product. D max Indicates the maximum time difference level. Combine the related features of all groups into a two-way intra-group related feature body

[0100] Next, in order to capture the relationship between different time difference levels, Take the average along the time difference level dimension and copy the mean back to the original dimension to get Finally and Splice along the feature group dimension to obtain a combined feature body

[0101] 3. Feature matching module. After constructing the four-dimensional feature body After that, it is necessary to further match the sample points with close time differences and refine the TDOA features.

[0102] Although a 3D convolutional network can directly process 4D features, this paper chooses a 2D convolutional network for layer-by-layer processing to improve real-time performance. Specifically, a modified U-net architecture is adopted, in which cross-layer connections help preserve fine-grained temporal details, which are crucial for accurate TDOA estimation. The features output by the network are then upsampled to restore the original temporal resolution, resulting in the final deep TDOA feature f tdoa .

[0103] The specific steps of the inter-pulse sequence feature extraction branch are as follows:

[0104] In addition to the TDOA characteristics between platforms, it is also essential to capture the inter-pulse sequence characteristics of the radar.

[0105] In order to effectively model these complex changing relationships, the present invention uses a two-layer bidirectional LSTM network to capture the long-range dependencies in the pulse sequence in a parallel feature extraction branch for the data of the main platform A, and extracts the inter-pulse feature f s .

[0106] The specific steps of feature fusion and classification are as follows:

[0107] After obtaining the deep TDOA features and inter-pulse sequence features, they are fused in the classification module and mapped to radar number labels using a linear layer.

[0108] Finally, the labels corresponding to the coding sequences are mapped to the original PDW labels to obtain the final sorting results.

[0109] This paper introduces deep neural networks into the multi-platform collaborative sorting task. The proposed DPCD-Net can simultaneously extract deep TDOA features and inter-pulse sequence features. The sequence feature extraction branch captures the changing relationships between pulses from the data of the main platform. Simultaneously, the deep TDOA feature extraction branch uses a bidirectional group correlation (Bi-GWC) module to perform bidirectional displacement correlation calculations along the time dimension, exploring the relationship between the data of the two platforms.

[0110] Compared to existing single-platform sorting methods, this method combines information from multiple platforms, adapting to more complex radar signal sorting scenarios. Compared to traditional TDOA vector-based sorting methods, this method extracts more effective deep TDOA features, resulting in improved sorting performance. Furthermore, because it combines both inter-pulse and deep TDOA features, it can also achieve superior results for emitters with time-of-day ambiguity.

[0111] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A radar pulse collaborative sorting method based on depth correlation and inter-pulse feature fusion, characterized in that: include: Build a dual-platform radar pulse deinterleaving network to complete the collaborative sorting of radar signals; The dual-platform radar pulse deinterleaving network includes a data encoding module, an inter-pulse sequence feature extraction branch, a deep time difference feature extraction branch, and a classification module; the deep time difference feature extraction branch includes a single-platform feature extraction module, a bidirectional group correlation module, and a matching module. The radar signal is collaboratively sorted based on a dual-platform radar pulse deinterleaving network, including: Step 1: Perform data encoding on the dual-platform PDW data according to the data encoding module to ensure that the time intervals between input data samples are equal; Step 2: Extract branches based on the deep time difference feature, perform feature extraction on the dual platform features, perform feature correlation analysis, calculate the mean of the analysis results, and then perform dimension splicing to obtain a four-dimensional feature volume. Match sample points with close time differences based on the four-dimensional feature volume to determine the final deep TDOA feature. Step 3: Extract branches based on the inter-pulse sequence features. A two-layer bidirectional LSTM network is used on the main platform data to capture the long-range dependencies in the pulse sequence and extract the inter-pulse features. Steps 2 and 3 are performed in parallel. Step 4: The deep TDOA features and inter-pulse features are fused through the classification module and mapped to the radar number label using a linear layer. Finally, the label corresponding to the coding sequence is mapped to the original PDW label to obtain the final sorting result.

2. The radar pulse collaborative sorting method based on depth correlation and inter-pulse feature fusion according to claim 1 is characterized in that: Also includes: The dual platform includes a main platform and a secondary platform; In the process of collaborative sorting of dual-platform data based on the dual-platform radar pulse deinterleaving network, the PDW data intercepted by the secondary platform is sent to the main platform for unified processing.

3. The radar pulse collaborative sorting method based on depth correlation and inter-pulse feature fusion according to claim 1 is characterized in that: Step 1 includes: Step 11: Normalize the radio frequency value of the PDW data to be between 0.1 and 1; Step 12: For each pulse , calculate the unit time points of its start and end, including and ; Step 13: Between the start and end time units, normalized radio frequency values ​​and labels are assigned in sequence to complete the encoding of the pulse data.

4. The radar pulse collaborative sorting method based on depth correlation and inter-pulse feature fusion according to claim 1 is characterized in that: Step 2 includes: Step 21: Perform feature extraction on the data of the two platforms using the single-platform feature extraction module to obtain features of platform A and platform B, respectively; platform A is the primary platform and platform B is the secondary platform; Step 22: Based on the bidirectional group correlation module, perform correlation analysis on the features of platform A and the features of platform B, and merge the analysis results to obtain a four-dimensional feature volume; Step 23: Based on the matching module, sample points with close time differences are matched according to the four-dimensional feature body to obtain the final deep TDOA feature.

5. The radar pulse collaborative sorting method based on depth correlation and inter-pulse feature fusion according to claim 4 is characterized in that: Step 21 includes: Step 211: Use a pair of ResNet-like networks with shared parameters to map the data from the two platforms into a high-dimensional feature space: Among them, A represents the main platform in the dual platform. Represents the convolution operation of layer 1, including multiple groups of convolution layers with batch normalization and ReLU activation functions. Indicates platform A l layer convolutional networks, Represents the encoded input data; Step 212: Connect the outputs of different layers of the convolutional network in step 212 to extract the features of platform A: in, represents the extracted features of platform A; Step 213: Extract the features of platform B according to the methods of steps 211 and 212. .

6. The radar pulse collaborative sorting method based on depth correlation and inter-pulse feature fusion according to claim 4 is characterized in that: Step 22 includes: Step 221: Divide features along the channel dimension and for groups, so that each feature group has channels; Step 222: Based on the principle that the TDOA between the two platform data can be positive or negative, calculate the Two-way group correlation features for feature groups: in, represents the inner product, Indicates the maximum time difference level; Merge the related features of all groups into a bidirectional intra-group related feature body ; Step 223: Take the average along the time difference level dimension and copy the mean back to the original dimension to get , and finally and Splicing along the feature group dimension to obtain a four-dimensional feature body .

7. The radar pulse collaborative sorting method based on depth correlation and inter-pulse feature fusion according to claim 4 is characterized in that: Step 23 includes: The improved U-net architecture is used to analyze the four-dimensional feature volume Process it and upsample its output features to restore the original time resolution to obtain the final deep TDOA features .

Images