A Method for Constructing a Dynamic Edge Weight Map of Radar Multidimensional Data

Abstract

The present application discloses a method for constructing a dynamic edge weight graph of radar multi-dimensional data, which relates to the technical fields of radar signal processing and graph data modeling. The method includes: obtaining a radar observation sequence; the radar observation sequence includes a plurality of nodes; the nodes correspond to time units in the radar observation sequence; performing feature extraction and feature fusion on the node feature vectors of each node to obtain a plurality of fused feature vectors; constructing dynamic edges and corresponding dynamic edge weights based on the plurality of nodes; the dynamic edges include: temporal adjacency edges, feature similarity edges, and self-loop edges; fusing and normalizing the dynamic edge weights to obtain a fused and normalized dynamic edge weight matrix; constructing a graph structure based on the plurality of fused feature vectors, the dynamic edges, and the dynamic edge weight matrix. The present application improves the modeling ability of radar target multi-dimensional features through a temporal attenuation, a dynamic threshold of feature similarity, and an edge weight fusion mechanism.

Description

Technical Field

[0001] This application relates to the technical fields of radar signal processing and graph data modeling, and particularly to a method for constructing a dynamic edge weight graph of radar multi-dimensional data. Background Art

[0002] Radar target echo data has multi-dimensionality (including Doppler spectrum, azimuth angle, elevation angle, distance, velocity, etc.) and non-linear time-varying characteristics. Traditional graph construction methods usually adopt fixed edge weights or single connection methods, and it is difficult to effectively capture the complex correlations between target dynamic features. Specific problems include: First, insufficient temporal correlation modeling: Traditional temporal graph construction methods adopt fixed neighborhoods or full connection strategies, which cannot adapt to the non-stationary dynamic characteristics of targets, resulting in noise interference and lack of feature continuity; Second, neglect of feature similarity: Existing methods do not fully utilize the intra-class similarity between multi-dimensional features, resulting in insufficient association of highly correlated nodes across time steps; Third, rigid edge weight assignment: The fixed edge weight mechanism cannot dynamically adjust the importance between nodes, weakening the sensitivity of the model to micro-motion features.

[0003] Therefore, how to solve the problems of fixed edge weights and insufficient dynamic association modeling in the processing of multi-dimensional radar data by traditional graph construction methods has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] The purpose of this application is to provide a method for constructing a dynamic edge weight graph of radar multi-dimensional data, which can solve the problems of fixed edge weights and insufficient dynamic association modeling in the processing of multi-dimensional radar data by traditional graph construction methods. Through the temporal attenuation, dynamic threshold of feature similarity, and edge weight fusion mechanism, the modeling ability of the multi-dimensional features of radar targets is improved.

[0005] To achieve the above purpose, the following solutions are provided in this application.

[0006] This application provides a method for constructing a dynamic edge weight graph of radar multi-dimensional data, and the method for constructing a dynamic edge weight graph of radar multi-dimensional data includes the following steps.

[0007] S1: Obtain a radar observation sequence; the radar observation sequence includes a number of nodes; the nodes correspond to time units in the radar observation sequence; the node feature vector of the node is jointly composed of a Doppler spectrum and physical motion parameters.

[0008] S2: Perform feature extraction and feature fusion on the node feature vectors of each node to obtain a number of fused feature vectors.

[0009] S3: Based on a number of nodes, construct dynamic edges and corresponding dynamic edge weights; the dynamic edges include: temporal adjacency edges, feature similarity edges, and self-loop edges.

[0010] S4: fusing and normalizing the dynamic edge weights to obtain a fused and normalized dynamic edge weight matrix.

[0011] S5: constructing a graph structure based on the plurality of fused feature vectors, the dynamic edges and the dynamic edge weight matrix.

[0012] According to the specific embodiments provided in this application, this application discloses the following technical effects.

[0013] The present application provides a method for constructing a dynamic edge weight graph of radar multi-dimensional data. By acquiring a radar observation sequence, the radar observation sequence includes a plurality of nodes, performing feature extraction and feature fusion on the node feature vector of each node, and obtaining a plurality of fused feature vectors, more representative and critical information can be mined, redundant or relatively minor parts can be removed, and features of different sources and different properties but interrelated can be integrated together to form a more compact and more expressive feature representation; by constructing dynamic edges and corresponding dynamic edge weights based on a plurality of nodes, the dynamic edges include: time-series adjacent edges, feature-similar edges and self-loop edges, the sequential logic between nodes can be clarified from a time-series perspective, the feature-similar edges can be used to mine the implicit connection between nodes close to each other at the feature level, and the self-loop edges also take into account factors such as the state continuation of the node itself, and provide diversified connection information and weight basis for constructing a more expressive graph structure; by fusing and normalizing the dynamic edge weights, a fused and normalized dynamic edge weight matrix can be obtained, the relative importance of different edges can be reflected more objectively and accurately, and the problem of some edge information being ignored due to excessive differences in edge weight values can be avoided. In this way, when the graph structure is subsequently constructed based on the edge weight matrix, the relationship information represented by each dynamic edge can be better integrated, so that the graph structure can more reasonably reflect the real and comprehensive relationship between nodes, and improve the expression ability and accuracy of the entire graph structure for the dynamic characteristics of the target; finally, based on a number of the fused feature vectors, the dynamic edges and the dynamic edge weight matrix, the graph structure is constructed. This application solves the problems of fixed edge weights and insufficient dynamic association modeling in traditional graph construction methods in multi-dimensional radar data processing. Through multi-stream feature extraction, the fusion design of temporal adjacent attenuation edges, feature similarity edges and self-loop edges, combined with dynamic edge weight allocation and normalization mechanism, adaptive modeling and efficient fusion of multi-dimensional features of radar targets are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 This is an application environment diagram of a method for constructing a dynamic edge weight graph of radar multi-dimensional data in an embodiment of the present application.

[0016] Figure 2 This is a schematic flowchart of a method for constructing a dynamic edge weight graph of radar multi-dimensional data provided in an embodiment of the present application.

[0017] Figure 3 This is a schematic diagram of the overall process of the method for constructing a dynamic edge weight graph of radar multi-dimensional data provided in an embodiment of the present application.

[0018] Figure 4 This is a schematic diagram of the curve of the time series attenuation weight changing with distance provided in an embodiment of the present application.

[0019] Figure 5 This is a schematic diagram of the dynamic threshold of feature similar edges provided in an embodiment of the present application.

[0020] Figure 6 This is a schematic diagram of the dynamic edge weight fusion normalization matrix provided in an embodiment of the present application.

[0021] Figure 7 This is a schematic diagram of the complete graph structure provided in an embodiment of the present application. Detailed implementation manners

[0022] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0023] To make the above objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0024] The method for constructing a dynamic edge weight graph of radar multi-dimensional data provided in the embodiments of the present application can be applied to, for example Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set separately, integrated on the server 104, placed on the cloud or other servers. The terminal 102 can send the radar observation sequence to the server 104. The radar observation sequence includes a number of nodes. The nodes correspond to the time units in the radar observation sequence. The node feature vector of the nodes is jointly composed of the Doppler spectrum and the physical motion parameters. After receiving the radar observation sequence, for the radar observation sequence, the server 104 extracts and fuses the feature vectors of each node to obtain a number of fused feature vectors. Based on a number of nodes, dynamic edges and corresponding dynamic edge weights are constructed. The dynamic edges include: temporal adjacency edges, feature similarity edges, and self-loops. The dynamic edge weights are fused and normalized to obtain a fused and normalized dynamic edge weight matrix. Based on a number of the fused feature vectors, the dynamic edges, and the dynamic edge weight matrix, a graph structure is constructed. The server 104 can feedback the obtained graph structure to the terminal 102. In addition, in some embodiments, the method for constructing a multi-dimensional data dynamic edge weight graph can also be implemented independently by the server 104 or the terminal 102. For example, the terminal 102 can directly construct a radar multi-dimensional data dynamic edge weight graph for the radar observation sequence, or the server 104 can obtain the radar observation sequence from the data storage system and construct a radar multi-dimensional data dynamic edge weight graph for the radar observation sequence.

[0025] Among them, the terminal 102 can be, but is not limited to, various desktop computers, laptop computers, smart phones, and tablet computers. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers, and can also be a cloud server.

[0026] In an exemplary embodiment, as Figure 2 shown, a method for constructing a radar multi-dimensional data dynamic edge weight graph is provided. This method is executed by a computer device, and can be specifically executed independently by a computer device such as a terminal or a server, or jointly executed by a terminal and a server. In the embodiments of the present application, taking this method applied to Figure 1 the server 104 in

[0027] S1: Obtain a radar observation sequence. The radar observation sequence includes a number of nodes. The nodes correspond to the time units in the radar observation sequence. The node feature vector of the nodes is jointly composed of the Doppler spectrum and the physical motion parameters.

[0028] S2: Extract and fuse the feature vectors of each node's node feature vector to obtain a number of fused feature vectors.

[0029] S3: Based on a number of nodes, construct dynamic edges and corresponding dynamic edge weights; the dynamic edges include: temporal adjacency edges, feature similarity edges, and self-loop edges.

[0030] S4: Fuse and normalize the dynamic edge weights to obtain a fused and normalized dynamic edge weight matrix.

[0031] S5: Based on a number of the fused feature vectors, the dynamic edges, and the dynamic edge weight matrix, construct a graph structure.

[0032] Implementing the above steps S1 to S5, through feature extraction and fusion, the fusion design of temporal adjacency decay edges, feature similarity edges, and self-loop edges, combined with the dynamic edge weight allocation and normalization mechanism, realizes the adaptive modeling and efficient fusion of multi-dimensional features of radar targets. Compared with traditional methods, the graph data constructed by the dynamic edge weight mechanism of this application, as the input of the graph convolutional network, can significantly improve the target classification accuracy and anti-noise performance.

[0033] As an alternative implementation, in step S2, perform feature extraction and feature fusion on the node feature vectors of each node to obtain a number of fused feature vectors, specifically including the following steps.

[0034] S21: Use a one-dimensional convolutional layer to extract features from the Doppler spectrum of each node to obtain local time-frequency features.

[0035] S22: Use a one-dimensional convolutional layer to extract features from the physical motion parameters of each node to obtain physical motion features.

[0036] S23: Concatenate the local time-frequency features and physical motion features corresponding to each node to obtain a number of fused feature vectors.

[0037] Specifically, the expression of the fused feature vector is as follows.

[0038] ;

[0039] ;

[0040] ;

[0041] Among them, is the local time-frequency feature of the Doppler spectrum extracted by the one-dimensional convolutional layer; is a one-dimensional convolutional layer with a kernel size of 3 and 64 output channels; is the th time unit's Doppler spectrum feature; is the mapping of physical motion parameters through the one-dimensional convolutional layer; is the physical motion feature; is the fused feature vector.

[0042] As an alternative implementation, in step S3, the construction of the temporal adjacent edges and edge weights includes the following steps.

[0043] A1: Establish local connections between nodes based on a time sliding window to obtain temporal adjacent edges; the time sliding window is adaptively adjusted according to the template acceleration change rate.

[0044] A2: Based on the temporal adjacent edges, use a hybrid decay function to determine the edge weights of the temporal adjacent edges; the hybrid decay function is a function obtained by combining an exponential decay function and a Gaussian kernel function.

[0045] Specifically, the expression for the local connection between nodes is as follows.

[0046] ;

[0047] ;

[0048] where is the node; is the node; is the local connection between nodes; is the total number of time steps, equal to the total number of nodes; is the th window radius; is the reference window radius; is the acceleration change rate coefficient; is the minimum value of the window radius; is the maximum value of the window radius; is the acceleration at the

[0049] The calculation formula for the edge weights of the temporal adjacent edges is as follows.

[0050] ;

[0051] where is the edge weight of the temporal adjacent edge; is the exponential decay weight coefficient; is the exponential decay rate; is the Gaussian kernel width; and are constants.

[0052] As an alternative implementation, in step S3, the construction of the feature similarity edges and edge weights includes the following steps.

[0053] B1: Calculate the cosine similarity between nodes based on the fused feature vectors.

[0054] B2: Calculate the cosine similarity of all pairs of nodes of the same type.

[0055] B3: Use the quantile of the cosine similarity of all pairs of nodes of the same type as the dynamic threshold.

[0056] B4: Consider the edges with the cosine similarity between the nodes greater than the dynamic threshold as feature-similar edges.

[0057] B5: Based on the feature-similar edges, use Z-score normalization to obtain the edge weights of the feature-similar edges.

[0058] Specifically, the calculation formula for the edge weights of the feature-similar edges is as follows.

[0059] ;

[0060] Where, is the edge weight of the feature-similar edge; is the mean of the similarities of nodes of the same type; is the standard deviation of the similarities of nodes of the same type; is the weight distribution coefficient; is node and 's cosine similarity.

[0061] As an alternative implementation, in step S3, the construction of self-loop edges and edge weights includes the following steps.

[0062] C1: Self-connect several nodes to obtain self-loop edges.

[0063] C2: Based on the self-loop edges, add fixed weights to each node to obtain the edge weights of the self-loop edges.

[0064] As an alternative implementation, the expression for the fused normalization is as follows.

[0065] ;

[0066] ;

[0067] Where, is the fused edge weight matrix; is the edge weight of the temporal adjacent edge; is the edge weight of the feature-similar edge; is the secondary weight coefficient; is the fused and normalized dynamic edge weight matrix; is the summation index variable, indicating traversing all neighbor edge weights of node ​

[0068] In an exemplary embodiment, a method for constructing a dynamic edge weight graph of radar multi-dimensional data is provided, and the overall flowchart is as Figure 3 shown, including the following steps.

[0069] Step 1: Initialization of the graph structure.

[0070] Model the radar observation sequence as a graph structure , where is the node set, is the edge set, is the edge weight matrix. Each node corresponds to a time unit in the radar observation sequence, and its feature vector is jointly composed of the Doppler spectrum and motion parameters (including azimuth, elevation, distance, speed, etc.).

[0071] ;

[0072] Among them, is the Doppler spectrum feature of the th time unit ; contains 5D physical motion features such as azimuth , elevation , distance , speed , and acceleration .

[0073] Step 2: Feature extraction and fusion.

[0074] Extract the Doppler flow and physical flow features through independent convolutional layers respectively. Use a one-dimensional convolutional layer to process the Doppler spectrum and output local time-frequency features, and use a convolutional layer with the same structure to process the physical motion parameters and output physical motion features. Concatenate the two-stream features into a fused feature to retain the complementarity of multi-modal information.

[0075] Extract the local time-frequency features of the Doppler spectrum through a one-dimensional convolutional layer.

[0076] ;

[0077] Among them, is the local time-frequency feature of the Doppler spectrum extracted through a one-dimensional convolutional layer; is a one-dimensional convolutional layer with a kernel size of 3 and an output channel of 64; is the th time unit of the Doppler spectrum feature .

[0078] Map the physical motion parameters through a one-dimensional convolutional layer.

[0079] ;

[0080] Among them, is to map physical motion parameters through a one-dimensional convolutional layer; is the physical motion feature.

[0081] Concatenate the two-stream features into a fused feature vector.

[0082] ;

[0083] Among them, is the fused feature vector.

[0084] Step 3: Dynamic edge weight construction.

[0085] Generate three types of edges based on the target motion characteristics and feature similarity: (1) Temporal adjacency edge: Dynamically adjust the time window according to the acceleration change rate, and assign weights to neighboring time nodes through a mixed decay function to balance local correlation and anti-noise ability; (2) Feature similarity edge: Calculate the node similarity based on the fused features, filter out highly correlated nodes through a dynamic threshold, and assign edge weights after normalization; (3) Self-loop edge: Add a fixed-weight self-connection to each node to ensure the retention of its own features in information transmission.

[0086] (1) Construction of temporal adjacency edge and edge weight.

[0087] Establish local connections based on a time sliding window.

[0088] ;

[0089] Among them, is the node; is the node; is the local connection between nodes; is the total number of time steps, equal to the total number of nodes; is the th window radius, adaptively adjusted according to the target acceleration change rate, and the formula is as follows.

[0090] ;

[0091] Among them, is the reference window radius; is the acceleration change rate coefficient; is the acceleration at the th moment; is the minimum value of the window radius, which can be set according to the actual situation;

[0092] To balance local sensitivity and anti-noise ability, the temporal edge weight decays with the interval distance. The hybrid decay function combines exponential decay and Gaussian kernel function.

[0093] ;

[0094] Among them, is the edge weight of the temporal adjacent edge; is the exponential decay weight coefficient; is the exponential decay rate; is the Gaussian kernel width, which can be set according to the actual situation; and are constants. The curve of the temporal decay weight changing with distance is as shown in Figure 4 shown.

[0095] (2)Construction of feature-similar edges and edge weights.

[0096] To overcome the insufficient sensitivity of pure temporal connections to the subtle feature differences of the target, feature-similar edges based on the dynamic threshold of the intra-class similarity distribution are introduced, and the dynamic threshold strategy is adopted to determine the effective connections. Calculate the cosine similarity of all pairs of nodes of the same class, take the 25% quantile as the dynamic threshold , and retain the edges where is greater than .

[0097] First, calculate the cosine similarity of nodes and and based on the fused feature vector.

[0098] ;

[0099] Calculate the cosine similarity of all pairs of nodes of the same class.

[0100] ;

[0101] Take the quantile as the dynamic threshold , and the schematic diagram of the dynamic threshold of the feature-similar edge is as shown in Figure 5 shown.

[0102] ;

[0103] Screen the effective connections through the dynamic threshold, that is, retain the edges where is greater than through the dynamic threshold.

[0104] ;

[0105] Map the similarity to weights through Z-score normalization.

[0106] ;

[0107] Among them, is the mean of the similarities of nodes of the same type; is the standard deviation of the similarities of nodes of the same type; is the weight distribution coefficient, set according to the actual situation; represents node and are nodes belonging to the same category of radar targets.

[0108] (3) Construction of self-loop edges and edge weights.

[0109] Add self-loop edges with a fixed weight of 1.0 to each node to retain the characteristics of the node itself.

[0110] ;

[0111] Step 4: Dynamic edge weight fusion and normalization.

[0112] Perform weighted fusion on the temporal edges and feature edges, giving priority to retaining the information of the high-weight edges, and the secondary edge weights are superimposed proportionally. The fused edge weight matrix is processed by normalization to avoid numerical instability problems and ensure the reliability of the input data for subsequent tasks.

[0113] If there are both temporal edges and feature edges between nodes, adopt the weighted summation strategy.

[0114] ;

[0115] Among them, is the fused edge weight matrix; is the secondary weight coefficient, set according to the actual situation. The self-loop edges are retained independently and do not conflict with the other edges.

[0116] Perform L2 normalization on the fused edge weight matrix by row to ensure numerical stability, and the weight matrix distribution is shown in Figure 6.

[0117] ;

[0118] Among them, is the dynamically fused and normalized edge weight matrix; is the summation index variable, indicating traversing all the neighbor edge weights of node (including temporal edges, feature similarity edges, and self-loop edges); The value of is not a parameter that needs to be set, but traverses all the nodes connected to node (i.e., nodes satisfying ). For example, if node has 10 neighbors (including self-loop edges), then takes values from 1 to 10. By normalizing the edge weights row by row (node dimension), it is ensured that the sum of the out-edge weights of each node is 1, avoiding numerical instability problems (such as gradient explosion).

[0119] Step 5: Construct and output graph data.

[0120] Generate a graph structure containing node features and the dynamic edge weight matrix (as shown in Figure 7 ), as the input data for downstream tasks (such as classification, tracking). This graph structure provides a highly robust basis for radar target analysis through adaptive modeling of spatio-temporal associations and multi-modal fusion.

[0121] The beneficial effects of a method for constructing a dynamic edge weight graph of radar target multi-dimensional data proposed in this application are as follows.

[0122] (1) Dynamic edge weight mechanism: By mixing decay functions and dynamic threshold strategies, adaptively adjust the temporal and feature edge weights to enhance the ability to model the target motion state.

[0123] (2) Multi-edge type fusion: Temporal edges capture local continuity, feature edges enhance cross-time step associations, and self-loop edges retain the node's own features.

[0124] (3) Feature extraction and fusion: Through the strategy of first extracting features and then splicing using a one-dimensional convolutional layer, fully exploit the complementarity of Doppler spectra and physical parameters.

[0125] (4) Anti-noise performance optimization: The combination of dynamic threshold screening and normalization strategies suppresses noise interference and improves the robustness of graph data.

[0126] The present application also provides an application scenario, which applies the above-mentioned method for constructing a radar multi-dimensional data dynamic edge weight graph. Specifically: The method for constructing a radar multi-dimensional data dynamic edge weight graph provided in this embodiment can be applied in the scenarios of radar signal processing and graph data modeling. The scenarios of radar signal processing and graph data modeling include: a data acquisition link, a feature extraction and feature fusion link, a dynamic edge and edge weight construction link, a fusion normalization link, and a graph structure construction link; First, obtain a radar observation sequence; the radar observation sequence includes a number of nodes; the nodes correspond to time units in the radar observation sequence; the node feature vector of each node is jointly composed of a Doppler spectrum and physical motion parameters; perform feature extraction and feature fusion on the node feature vectors of each node to obtain a number of fused feature vectors; Secondly, based on a number of nodes, construct dynamic edges and corresponding dynamic edge weights; the dynamic edges include: time-series adjacency edges, feature similarity edges, and self-loop edges; Then, fuse and normalize the dynamic edge weights to obtain a fused and normalized dynamic edge weight matrix; Finally, based on a number of the fused feature vectors, the dynamic edges, and the dynamic edge weight matrix, a graph structure can be constructed.

[0127] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0128] Specific examples are used in this article to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A method for constructing a dynamic edge weight graph of radar multi-dimensional data, characterized in that, The method for constructing a dynamic edge weight graph of radar multi-dimensional data includes: Obtaining a radar observation sequence; the radar observation sequence includes a number of nodes; the nodes correspond to time units in the radar observation sequence; the node feature vector of each node is jointly composed of a Doppler spectrum and physical motion parameters; Performing feature extraction and feature fusion on the node feature vectors of each node to obtain a number of fused feature vectors; Based on a number of nodes, constructing dynamic edges and corresponding dynamic edge weights; the dynamic edges include: temporal adjacency edges, feature similarity edges, and self-loop edges; Fusing and normalizing the dynamic edge weights to obtain a fused and normalized dynamic edge weight matrix; Based on a number of the fused feature vectors, the dynamic edges, and the dynamic edge weight matrix, constructing a graph structure.

2. The method for constructing a radar multi-dimensional data dynamic edge weight graph according to claim 1, wherein Performing feature extraction and feature fusion on the node feature vectors of each node to obtain a number of fused feature vectors, specifically including: Using a one-dimensional convolutional layer to perform feature extraction on the Doppler spectrum of each node to obtain local time-frequency features; Using a one-dimensional convolutional layer to perform feature extraction on the physical motion parameters of each node to obtain physical motion features; Concatenating the local time-frequency features and physical motion features corresponding to each node to obtain a number of fused feature vectors.

3. The method for constructing a radar multi-dimensional data dynamic edge weight graph according to claim 2, wherein The expression of the fused feature vector is: ； ； ； Among them, is to extract the local time-frequency features of the Doppler spectrum through a one-dimensional convolutional layer; is a one-dimensional convolutional layer with a kernel size of 3 and 64 output channels; is the Doppler spectrum feature of the th time unit; is to map the physical motion parameters through a one-dimensional convolutional layer; is the physical motion feature; is the fused feature vector.

4. The method for constructing a radar multi-dimensional data dynamic edge weight graph according to claim 1, wherein Based on a number of nodes, constructing dynamic edges and corresponding dynamic edge weights, specifically including: Establishing local connections between nodes based on a time sliding window to obtain temporal adjacency edges; the time sliding window is adaptively adjusted according to the template acceleration change rate; Based on the temporal adjacency edges, using a mixed decay function to determine the temporal adjacency edge weights; the mixed decay function is a function obtained by combining an exponential decay function and a Gaussian kernel function.

5. The method for constructing a radar multi-dimensional data dynamic edge weight graph according to claim 4, wherein The expression of the local connection between nodes is: ； ； Among them, is the node; is the node; is the local connection between nodes; is the total number of time steps, equal to the total number of nodes; is the th window radius; is the reference window radius; is the acceleration change rate coefficient; is the minimum value of the window radius; is the maximum value of the window radius; is the acceleration at the moment.

6. The method for constructing a radar multi-dimensional data dynamic edge weight graph according to claim 4, wherein The calculation formula of the temporal adjacency edge weight is: ； Among them, is the edge weight of the temporal adjacent edge; is the exponential decay weight coefficient; is the exponential decay rate; is the Gaussian kernel width; and are constants.

7. The method for constructing a radar multi-dimensional data dynamic edge weight graph according to claim 1, wherein Based on a number of nodes, constructing dynamic edges and corresponding dynamic edge weights, specifically including: Calculating the cosine similarity between nodes based on the fused feature vectors; Calculating the cosine similarity of all pairs of similar nodes; Taking the quantile of the cosine similarity of all pairs of similar nodes as the dynamic threshold; Taking the edges with the cosine similarity between nodes greater than the dynamic threshold as feature similarity edges; Based on the feature similarity edges, using Z-score normalization to obtain the feature similarity edge weights.

8. The method for constructing a radar multi-dimensional data dynamic edge weight graph according to claim 7, wherein, The calculation formula of the feature similarity edge weight is: ； Among them, is the edge weight of feature-similar edges; is the mean of the similarities of nodes of the same type; is the standard deviation of the similarities of nodes of the same type; is the weight distribution coefficient; is the node and cosine similarity.

9. The method for constructing a radar multi-dimensional data dynamic edge weight graph according to claim 1, wherein Based on a number of nodes, constructing dynamic edges and corresponding dynamic edge weights, specifically including: Performing self-connection on a number of nodes to obtain self-loop edges; Based on the self-loop edges, adding a fixed weight to each node to obtain the self-loop edge weights.

10. The method for constructing a radar multi-dimensional data dynamic edge weight graph according to claim 1, wherein The expression of the fusion normalization is: ； ； Among them, is the fused edge weight matrix; is the edge weight of the temporal adjacent edge; is the edge weight of the feature similar edge; is the secondary weight coefficient; is the fused and normalized dynamic edge weight matrix; is the summation index variable, indicating traversing all neighbor edge weights of node ​

Images