Detailed Description
The following description of the embodiments of the present application will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present application, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.
In order that the above-recited objects, features and advantages of the present application will become more readily apparent, a more particular description of the application will be rendered by reference to the appended drawings and appended detailed description.
The method for constructing the radar multidimensional data dynamic side weight graph provided by the embodiment of the application can be applied to an application environment shown in fig. 1. Wherein the terminal 102 communicates with the server 104 via a network. The data storage system may store data that the server 104 needs to process. The data storage system may be provided separately, may be integrated on the server 104, or may be placed on a cloud or other server. The terminal 102 may send a radar observation sequence to the server 104, where the radar observation sequence includes a plurality of nodes, the nodes correspond to time units in the radar observation sequence, node feature vectors of the nodes are formed by combining doppler spectrum and physical motion parameters, after the server 104 receives the radar observation sequence, the server 104 performs feature extraction and feature fusion on the node feature vectors of each node to obtain a plurality of fused feature vectors, constructs a dynamic edge and a corresponding dynamic edge weight based on the plurality of nodes, where the dynamic edge includes a time sequence adjacent edge, a feature similar edge and a self-loop edge, performs fusion normalization on the dynamic edge weights to obtain a dynamic edge weight matrix after fusion normalization, and constructs a graph structure based on the plurality of fused feature vectors, the dynamic edge and the dynamic edge weight matrix. The server 104 may feed back the resulting graph structure to the terminal 102. In addition, in some embodiments, the method for constructing the multi-dimensional data dynamic edge map may be implemented by the server 104 or the terminal 102 alone, for example, the terminal 102 may directly construct the radar multi-dimensional data dynamic edge map for the radar observation sequence, or the server 104 may acquire the radar observation sequence from the data storage system and construct the radar multi-dimensional data dynamic edge map for the radar observation sequence.
The terminal 102 may be, but is not limited to, a variety of desktop computers, notebook computers, smart phones, and tablet computers. The server 104 may be implemented as a stand-alone server or a server cluster composed of a plurality of servers, or may be a cloud server.
In an exemplary embodiment, as shown in fig. 2, a method for constructing a dynamic edge map of radar multidimensional data is provided, and the method is executed by a computer device, specifically, may be executed by a computer device such as a terminal or a server, or may be executed by the terminal and the server together, and in an embodiment of the present application, the method is applied to the server 104 in fig. 1 and is described as an example, and includes the following steps S1 to S5.
S1, acquiring a radar observation sequence, wherein the radar observation sequence comprises a plurality of nodes, the nodes correspond to time units in the radar observation sequence, and node feature vectors of the nodes are formed by combining Doppler frequency spectrums and physical motion parameters.
And S2, carrying out feature extraction and feature fusion on the node feature vectors of each node to obtain a plurality of fusion feature vectors.
And S3, constructing dynamic edges and corresponding dynamic edge weights based on a plurality of nodes, wherein the dynamic edges comprise time sequence adjacent edges, characteristic similar edges and self-loop edges.
And S4, carrying out fusion normalization on the dynamic side weights to obtain a fusion normalized dynamic side weight matrix.
S5, constructing a graph structure based on the fusion feature vectors, the dynamic edges and the dynamic edge weight matrix.
By implementing the steps S1 to S5, through the fusion design of feature extraction and fusion, the time sequence adjacent attenuation edge, the feature similar edge and the self-annular edge, and combining a dynamic edge weight distribution and normalization mechanism, the self-adaptive modeling and efficient fusion of the radar target multidimensional features are realized. Compared with the traditional method, the graph data constructed by the dynamic side weight mechanism can be used as the input of the graph convolution network, so that the target classification precision and the noise resistance can be obviously improved.
As an optional implementation manner, in step S2, feature extraction and feature fusion are performed on node feature vectors of each node to obtain a plurality of fused feature vectors, which specifically includes the following steps.
And S21, carrying out feature extraction on Doppler frequency spectrums of each node by adopting a one-dimensional convolution layer to obtain local time-frequency features.
And S22, extracting the characteristics of the physical motion parameters of each node by adopting a one-dimensional convolution layer to obtain the physical motion characteristics.
And S23, splicing the local time-frequency characteristics and the physical motion characteristics corresponding to each node to obtain a plurality of fusion characteristic vectors.
Specifically, the expression of the fusion feature vector is as follows.
;
;
;
Wherein, theExtracting local time-frequency characteristics of Doppler frequency spectrum through a one-dimensional convolution layer; is a one-dimensional convolution layer with a kernel size of 3 and an output channel of 64; Is the firstDoppler spectrum characteristics of individual time units; mapping physical motion parameters through a one-dimensional convolution layer; Is a physical movement feature; is a fused feature vector.
As an alternative embodiment, in step S3, the time sequence adjacent edge and edge weight construction includes the following steps.
A1, establishing local connection among nodes based on a time sliding window to obtain a time sequence adjacent edge, wherein the time sliding window is adaptively adjusted according to the rate of change of the template acceleration.
A2, determining the edge weight of the time sequence adjacent edge by adopting a mixed attenuation function based on the time sequence adjacent edge, wherein the mixed attenuation function is a function obtained by combining an exponential attenuation function and a Gaussian kernel function.
Specifically, the expression of the local connection between the nodes is as follows.
;
;
Wherein, theIs the firstA node; Is the firstA node; Is a local connection between nodes; the total number of time steps is equal to the total number of nodes; Is the firstA window radius; Is the reference window radius; Is the acceleration change rate coefficient; Is the minimum value of the radius of the window; is the maximum value of the radius of the window; Is the firstAcceleration at time.
The calculation formula of the adjacent edge weight of the time sequence is shown as follows.
;
Wherein, theThe adjacent edge weight is the time sequence; Is an exponential decay weight coefficient; Is the exponential decay rate; is Gaussian kernel width; AndIs constant.
As an alternative embodiment, in step S3, the feature similarity edge and edge weight construction includes the following steps.
And B1, calculating cosine similarity among the nodes based on the fusion feature vector.
And B2, calculating cosine similarity of all similar node pairs.
And B3, taking the quantiles of cosine similarity of all similar node pairs as dynamic thresholds.
And B4, taking the edge with cosine similarity larger than the dynamic threshold value between the nodes as a characteristic similarity edge.
And B5, based on the feature similarity sides, adopting Z-score standardization to obtain feature similarity side weights.
Specifically, the calculation formula of the feature similarity side weight is shown as follows.
;
Wherein, theSimilar edge weights are the features; is the average value of the similarity of the similar nodes; The standard deviation of the similarity of the similar nodes; distributing coefficients for the weights; Is a nodeAndCosine similarity of (c).
As an alternative embodiment, in step S3, the self-lookaside and side weight construction includes the following steps.
And C1, performing self-connection on a plurality of nodes to obtain a self-ring edge.
And C2, adding fixed weight to each node based on the self-loop edge to obtain the self-loop edge weight.
As an alternative embodiment, the fusion normalized expression is as follows.
;
;
Wherein, theThe side weight matrix is the side weight matrix after fusion; the adjacent edge weight is the time sequence; similar edge weights are the features; is a secondary weighting coefficient; Fusing the normalized dynamic side weight matrix; to sum and index the variables, represent the nodesTraversing all neighbor edge weights of (a).
In an exemplary embodiment, a method for constructing a radar multidimensional data dynamic side-weight graph is provided, and the overall flowchart is shown in fig. 3, and includes the following steps.
And 1, initializing a graph structure.
Modeling radar observation sequences as graph structuresWhereinFor a set of nodes,For the collection of edges,Is an edge weight matrix. Each nodeThe characteristic vector is formed by combining Doppler frequency spectrum and motion parameters (including azimuth angle, pitch angle, distance, speed and the like) corresponding to a time unit in the radar observation sequence.
;
Wherein, theIs the firstDoppler spectrum characteristics of individual time units;Including azimuth anglePitch angleDistance and distanceSpeed and velocity ofAcceleration ofAnd 5-dimensional physical motion characteristics.
And 2, feature extraction and fusion.
Doppler flow and physical flow characteristics are extracted through independent convolution layers respectively. And processing Doppler frequency spectrum by using a one-dimensional convolution layer, outputting local time-frequency characteristics, processing physical motion parameters by adopting the convolution layer with the same structure, and outputting physical motion characteristics. And splicing the double-flow characteristics into fusion characteristics, and reserving complementarity of the multi-mode information.
Local time-frequency characteristics of the Doppler frequency spectrum are extracted through the one-dimensional convolution layer.
;
Wherein, theExtracting local time-frequency characteristics of Doppler frequency spectrum through a one-dimensional convolution layer; is a one-dimensional convolution layer with a kernel size of 3 and an output channel of 64; Is the firstDoppler spectrum characteristics of individual time units。
The physical motion parameters are mapped by a one-dimensional convolution layer.
;
Wherein, theMapping physical motion parameters through a one-dimensional convolution layer; Is a physical movement characteristic.
And splicing the double-flow features into a fusion feature vector.
;
Wherein, theIs a fused feature vector.
And 3, constructing dynamic side rights.
The method comprises the steps of generating three types of edges based on target motion characteristics and feature similarity, namely (1) dynamically adjusting a time window according to acceleration change rate, distributing weights by adjacent time nodes through a mixed attenuation function, balancing local association and noise immunity, (2) calculating node similarity based on fusion features, screening high-correlation nodes by dynamic thresholds, weighting and distributing edge weights after standardization, and (3) adding fixed weight self-connection from each node to ensure that self features are reserved in information transmission.
(1) And (5) constructing time sequence adjacent edges and edge weights.
The local connection is established based on a time sliding window.
;
Wherein, theIs the firstA node; Is the firstA node; Is a local connection between nodes; the total number of time steps is equal to the total number of nodes; Is the firstThe window radius is adaptively adjusted according to the target acceleration change rate, and the formula is shown as follows.
;
Wherein, theIs the reference window radius; Is the acceleration change rate coefficient; Is the firstAcceleration at time; The minimum value of the radius of the window can be set according to the actual setting; is the maximum value of the radius of the window, and can be set according to practice.
To balance local sensitivity with noise immunity, the timing edge weights decay with separation distance. The mixed decay function combines exponential decay with a gaussian kernel function.
;
Wherein, theThe adjacent edge weight is the time sequence; Is an exponential decay weight coefficient; Is the exponential decay rate; The width of the Gaussian kernel can be set according to practice; AndIs constant. The time-series decay weight is plotted against distance as shown in fig. 4.
(2) And (5) constructing similar sides and side weights of the features.
In order to overcome the defect of insufficient sensitivity of the pure time sequence connection to the target fine feature difference, feature similarity edges based on dynamic threshold values of similarity distribution are introduced, and effective connection is determined by adopting a dynamic threshold value strategy. Calculating cosine similarity of all similar node pairsTaking out25% Quantile of (2) as dynamic thresholdReserved, reserveGreater thanIs a side of (c).
First based on the fusion feature vectorComputing nodeAndCosine similarity of (c).
;
Calculating cosine similarity of all similar node pairs。
;
Taking outNumber of digits of (a)As a dynamic thresholdThe feature similarity edge dynamic threshold is schematically shown in fig. 5.
;
Active connections are filtered by dynamic threshold, i.e. reserved by dynamic thresholdGreater thanIs a side of (c).
;
Similarity is mapped to weights by Z-score normalization.
;
Wherein, theIs the average value of the similarity of the similar nodes; The standard deviation of the similarity of the similar nodes; distributing coefficients for the weights according to actual settings; Representing nodesAndNodes belonging to the same class of radar targets.
(3) And constructing from the ring edge and the edge weight.
And adding a self-loop edge with fixed weight of 1.0 for each node, and reserving the self-characteristics of the node.
;
And 4, dynamic side weight fusion normalization.
And carrying out weighted fusion on the time sequence side and the characteristic side, preferentially reserving high-weight side information, and overlapping secondary side weights in proportion. The fused side weight matrix is normalized, so that the problem of unstable numerical value is avoided, and the reliability of the input data of the subsequent task is ensured.
And if the time sequence edges and the characteristic edges exist between the nodes at the same time, adopting a weighted summation strategy.
;
Wherein, theThe side weight matrix is the side weight matrix after fusion; And the secondary weight coefficient is set according to the actual setting. And the self-ring edge is reserved independently and does not collide with the rest edges.
L2 normalization is carried out on the fused side weight matrix according to the rows, so that the numerical stability is ensured, and the weight matrix distribution is shown in a figure 6.
;
Wherein, theFusing the normalized dynamic side weight matrix; to sum and index the variables, represent the nodesTraversing all neighbor edge weights (including time sequence edges, feature similarity edges and self-loop edges); not parameters to be set, but traversal and nodesAll nodes connected(I.e. satisfyIs a node of (c). For example, if the nodeThere are 10 neighbors (with self-loop edge), thenTake values of 1 to 10. By normalizing the edge weights by row (node dimension), the sum of the edge weights of each node is ensured to be 1, and the problem of unstable numerical values (such as gradient explosion) is avoided.
And 5, constructing and outputting graph data.
Generating containing node featuresAnd dynamic side weight matrixIs of the figure structureAs shown in fig. 7, as input data for downstream tasks (e.g., classification, tracking). The graph structure is fused with multiple modes through self-adaptive modeling time-space association, and provides a high robustness foundation for radar target analysis.
The method for constructing the radar target multidimensional data dynamic side weight graph has the following beneficial effects.
(1) And a dynamic side weight mechanism, namely, self-adaptively adjusting time sequence and characteristic side weight by mixing an attenuation function and a dynamic threshold strategy, and enhancing modeling capability of a target motion state.
(2) And (3) multi-edge type fusion, namely capturing local continuity by a time sequence edge, enhancing cross-time step association by a characteristic edge, and reserving the self characteristics of the node from a ring edge.
(3) Feature extraction and fusion, namely extracting features through a one-dimensional convolution layer and then splicing the features, and fully mining the complementarity of Doppler frequency spectrum and physical parameters.
(4) Noise immunity optimization, namely, noise interference is restrained by combining dynamic threshold screening and normalization strategies, and robustness of graph data is improved.
The application also provides an application scene, which applies the method for constructing the radar multidimensional data dynamic side weight graph. The method for constructing the radar multidimensional data dynamic side weight graph can be applied to radar signal processing and graph data modeling scenes. The radar signal processing and graph data modeling scene comprises 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, wherein a radar observation sequence is firstly acquired, the radar observation sequence comprises a plurality of nodes, the nodes correspond to time units in the radar observation sequence, node feature vectors of the nodes are formed by combining Doppler frequency spectrums and physical motion parameters, feature extraction and feature fusion are carried out on the node feature vectors of the nodes to obtain a plurality of fusion feature vectors, then a dynamic edge and corresponding dynamic edge weight are constructed based on the plurality of nodes, the dynamic edge comprises a time sequence adjacent edge, a feature similar edge and a self-loop edge, then fusion normalization is carried out on the dynamic edge weight to obtain a dynamic edge weight matrix after fusion normalization, and finally, the graph structure can be constructed based on the fusion feature vectors, the dynamic edge and the dynamic edge weight matrix.
The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.
The principles and embodiments of the present application have been described herein with reference to specific examples, which are intended to facilitate an understanding of the principles and concepts of the application and are to be varied in scope and detail by persons of ordinary skill in the art based on the teachings herein. In view of the foregoing, this description should not be construed as limiting the application.