Multi-radar cooperative target identification method based on graph neural network
By employing a multi-radar cooperative target recognition method based on graph neural networks, feature extraction is performed using graph adjacency matrices and cascaded networks. This solves the problems of insufficient feature extraction capability and inadequate utilization of physical spatial information, achieving higher recognition accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2026-01-21
- Publication Date
- 2026-05-19
AI Technical Summary
Existing multi-radar cooperative target recognition methods suffer from insufficient feature extraction capabilities and inadequate utilization of multi-radar physical space information, resulting in low recognition accuracy.
A multi-radar collaborative target recognition method based on graph neural networks is adopted. By constructing a graph adjacency matrix and a cascaded feature extraction subnetwork and graph neural recognition subnetwork, multiple graph convolutions and nonlinear transformations are performed to obtain more discriminative features.
It effectively improves the accuracy of target recognition in complex scenarios and makes full use of the physical space information of multiple radars for dynamic collaborative enhancement.
Smart Images
Figure FT_1 
Figure FT_2 
Figure QLYQS_17
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar signal processing technology and relates to a multi-radar cooperative target recognition method, specifically a multi-radar cooperative target recognition method based on graph neural networks, which can be applied to fields such as air traffic control. Background Technology
[0002] Multi-radar cooperative target recognition involves deploying multiple radars in space to simultaneously acquire target information from different perspectives. By leveraging the complementary characteristics of this information, a more comprehensive and complete view of the target is constructed, overcoming the limitations of a single perspective and effectively mitigating attitude sensitivity while improving recognition accuracy. Currently, multi-radar cooperative target recognition mainly employs two methods: decision-level fusion and feature-level fusion. The feature-level fusion method, in particular, demonstrates relatively better utilization of target information.
[0003] Improving feature extraction capabilities and fully utilizing the physical space information of multiple radars are key factors in enhancing recognition accuracy. For example, patent document CN120122071A discloses a multi-radar cooperative target recognition method based on adaptive manifold discriminant regression. This invention constructs an objective function containing parameters such as viewpoint weights, projection matrices, and nearest neighbor weight matrices, and iteratively optimizes an initial projection matrix using training data. In the recognition stage, this method constructs a data matrix from the multi-view HRRP data to be identified and projects features onto it using the optimized target projection matrix, mapping high-dimensional data to a low-dimensional feature space. Finally, the target category is determined based on the obtained low-dimensional feature vectors. This invention improves the accuracy of target recognition. However, it is essentially a shallow dimensionality reduction method based on linear projection, lacking deep nonlinear feature extraction of HRRP data. Furthermore, it only models the multi-radar cooperative relationship through global viewpoint weights, failing to utilize the physical space information of multiple radars, making it difficult to achieve complementary enhancement between different viewpoint information, thus limiting further improvement in accuracy. Summary of the Invention
[0004] The purpose of this invention is to overcome the defects of the prior art and propose a multi-radar cooperative target recognition method based on graph neural networks to solve the technical problem of low recognition accuracy caused by insufficient feature extraction capability and insufficient utilization of physical space information of multiple radars in the prior art.
[0005] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:
[0006] (1) Obtain the training sample set and the test sample set:
[0007] Get Each radar receives data from different angles. The HRRP data of each target is collected, and the target category is labeled after preprocessing the HRRP data of each target received by each radar. At the same time, a graph adjacency matrix is constructed for each target. Then, more than half of the preprocessed HRRP data and their labels and the corresponding target graph adjacency matrix are used to form a training sample set, and the remaining preprocessed HRRP data and the corresponding target graph adjacency matrix are used to form a test sample set.
[0008] (2) Construct a multi-radar cooperative target recognition network model:
[0009] Construct a multi-radar cooperative target recognition network model comprising a cascaded feature extraction subnetwork and a graph neural network recognition subnetwork; wherein the graph neural network recognition subnetwork comprises a cascaded feature extraction subnetwork and a graph neural network recognition subnetwork. The graph convolution module, consisting of a graph convolutional layer and a ReLU activation function layer loaded between adjacent graph convolutional layers, is used to perform multiple graph convolutions and nonlinear transformations on the node features based on the node features of each target, the features of the node's neighboring nodes, and the weights of the edges between the node and its neighboring nodes in the graph adjacency matrix, in order to obtain more discriminative features.
[0010] (3) Iteratively train the multi-radar cooperative target recognition network model:
[0011] The multi-radar cooperative target recognition network model is iteratively trained using a training sample set to obtain a well-trained target recognition network model.
[0012] (4) Obtain target recognition results from multi-radar collaborative methods:
[0013] By using the test sample set as input to the trained target recognition network model, the multi-radar cooperative target recognition results corresponding to the test sample set are obtained.
[0014] Compared with the prior art, the present invention has the following advantages:
[0015] The graph convolution module in this invention performs multiple graph convolutions and nonlinear transformations on the node features based on the node features of each target, the features of the node's neighboring nodes, and the weights of the edges containing the node and its neighboring nodes in the graph adjacency matrix. This not only enables the data received by the radar from different angles to be dynamically and complementaryly enhanced based on the correlation of the physical spatial information of multiple radars, but also allows for deep nonlinear feature extraction to obtain more discriminative features. Compared with existing technologies, this effectively improves the accuracy of target recognition in complex scenarios. Attached Figure Description
[0016] Figure 1 This is a flowchart illustrating the implementation of the present invention.
[0017] Figure 2 This is a schematic diagram of the structure of the multi-radar cooperative target recognition network of the present invention. Detailed Implementation
[0018] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0019] Reference Figure 1 The present invention includes the following steps:
[0020] Step 1) Obtain the training sample set and the test sample set:
[0021] Get Each radar receives data from different angles. The HRRP data of each target is collected, and the HRRP data of each target received by each radar is preprocessed and labeled with the target category. At the same time, a graph adjacency matrix of each target is constructed. Then, more than half of the preprocessed HRRP data and their labels and the corresponding target graph adjacency matrix are used to form a training sample set, and the remaining preprocessed HRRP data and the corresponding target graph adjacency matrix are used to form a test sample set. In this embodiment, , .
[0022] HRRP data, or High Resolution Range Profile, is an important data form in radar target identification. It represents the distribution of scattering intensity of a target in the radar line of sight, and is essentially a one-dimensional image of the target in the range dimension.
[0023] The steps for preprocessing the HRRP data received by each radar for each target are as follows:
[0024] The HRRP data of each target received by each radar is normalized, and the segment with the highest amplitude of the normalized HRRP data is extracted to obtain the preprocessed HRRP data of each target.
[0025] Construct the adjacency matrix for each target graph The implementation steps are as follows:
[0026] (1a) through the first Position of the radar in the global coordinate system and its distance relative to the first radar Calculate the position of each target in the global coordinate system and according to and The calculated first Range vector of each radar relative to each target Calculate the first The radar observes the azimuth angle of each target. With pitch angle :
[0027] ;
[0028] ;
[0029] ;
[0030] ;
[0031] ;
[0032] ;
[0033] ;
[0034] in, This indicates the operation of taking the mold length. This indicates the remainder operation. This is the position of the first radar in the global coordinate system. , , They represent In the global coordinate system , , Components on the axis, , , express In the global coordinate system , , Components on the axis, , , express In the global coordinate system , , Components on the axis, ;
[0035] (1b) By azimuth angle With pitch angle Calculate the weight of the edge between every two radars And construct a target with the serial number of each radar as the node, and the edge weights as the nodes. The graph adjacency matrix with value ,in:
[0036] ;
[0037] in, , .
[0038] This invention utilizes multi-radar physical spatial information to construct a graph adjacency matrix, which integrates multi-radar physical spatial information as core prior knowledge into the network. This enables data received by radars from different angles to be dynamically and collaboratively enhanced based on the correlation of multi-radar physical spatial information, thereby improving the recognition accuracy in complex scenarios.
[0039] Step 2) Construct a multi-radar cooperative target recognition network model, the structure of which is as follows: Figure 2 As shown:
[0040] Construct a multi-radar cooperative target recognition network model comprising a cascaded feature extraction subnetwork and a graph neural network recognition subnetwork; wherein the graph neural network recognition subnetwork comprises a cascaded feature extraction subnetwork and a graph neural network recognition subnetwork. A graph convolution module, consisting of a graph convolutional layer and a ReLU activation function layer loaded between adjacent graph convolutional layers, is used to perform multiple graph convolutions and nonlinear transformations on the node features of each target, the features of the node's neighboring nodes, and the weights of the edges between the node and its neighboring nodes in the graph adjacency matrix, to obtain more discriminative features. In this embodiment... ,in:
[0041] The feature extraction subnetwork includes a cascaded convolutional module consisting of three stacked convolutional layers and a global average pooling layer. A max-pooling layer is loaded between adjacent convolutional layers in the convolutional module. The kernel size of the first-level convolutional layer is 5×1, and the kernel size of the second and third-level convolutional layers is 3×1. The pooling window size of the max-pooling layer between adjacent convolutional layers is 2×1, with a stride of 2.
[0042] The graph neural recognition subnetwork consists of a graph convolutional module composed of three stacked graph convolutional layers and a ReLU activation function layer loaded between adjacent graph convolutional layers. The output of the graph convolutional module is cascaded with a global average pooling layer and a fully connected classification head. The fully connected classification head includes stacked fully connected layers and a Softmax function layer.
[0043] Step 3) Iteratively train the multi-radar cooperative target recognition network model:
[0044] The multi-radar cooperative target recognition network model is iteratively trained using a training sample set to obtain a well-trained target recognition network model. The steps are as follows:
[0045] (3a) Initialize the number of iterations to be The maximum number of iterations is , , No. The weights in the next iteration of the multi-radar cooperative target recognition network model are: and order In this embodiment ;
[0046] (3b) The training sample set is used as the input to the multi-radar cooperative target recognition network model, and the feature extraction sub-network extracts the first... The global features representing the scattering characteristics of each target category and possessing translation invariance in the HRRP data of each target are used to form a node feature matrix. The graph neural recognition subnetwork passes through the first Node feature matrix of each target And graph adjacency matrix For the The target is classified to obtain the first target. The probability that an object is predicted as the true object category. , , This represents the total number of targets corresponding to the HRRP data in the training sample set, in this embodiment. ;
[0047] (3b1) Each radar receives HRRP data of each target from different angles, with a dimension of 1 × 256. The global features in the HRRP data of each target are extracted through the feature extraction subnetwork to form a node feature matrix. ;
[0048] Among them, the node feature matrix The node is the sequence number of each radar, and the node feature is the first feature received from the radar corresponding to that node by the feature extraction network. Global features extracted from the HRRP data of each target.
[0049] (3b2) The first graph convolutional layer in the graph convolution module utilizes the node feature matrix And graph adjacency matrix Convolution is performed on each node, and the features of each node are updated by a non-linear transformation using the ReLU activation function, resulting in the output of the first graph convolutional layer. The second graph convolutional layer will convert the output of the first graph convolutional layer into a single layer. As input, a graph convolution operation is performed again, and a non-linear transformation is applied using the ReLU activation function to obtain the output of the second graph convolutional layer. The output of the third graph convolutional layer to the second graph convolutional layer. Perform graph convolution operations to obtain the final node feature matrix. ;
[0050] The formula for calculating the output feature matrix from the input of each layer is as follows:
[0051] ;
[0052] in, , The ReLU activation function is used. Representing the graph adjacency matrix The normalized matrix, .
[0053] (3b3) Global average pooling layer on the final node feature matrix Perform feature fusion to obtain the first Global graph features of each target;
[0054] (3b4) The fully connected classification head identifies global features of the graph and obtains the first... The probability that an object is predicted as the true object category. ;
[0055] (3c) The cross-entropy loss function is used, and the loss value of the target recognition network model is calculated by the probability that each target is predicted as the true target category and its corresponding true target category label. Then through weight The target recognition network model for this iteration is then updated.
[0056] The loss value The calculation formula is:
[0057] ;
[0058] in, This represents the logarithmic operation. For the first The true category label of each target.
[0059] calculate right partial derivatives and according to right Update:
[0060] ;
[0061] in, Represents weight The update results Indicates the learning rate. express right Take the partial derivative.
[0062] (3d) Judgment If the condition is met, a well-trained multi-radar cooperative target recognition network model is obtained; otherwise, let... Then proceed with step (3b).
[0063] Step 4) Obtain the target recognition results of multi-radar collaborative operation:
[0064] By using the test sample set as input to the trained target recognition network model, the multi-radar cooperative target recognition results corresponding to the test sample set are obtained.
Claims
1. A multi-radar cooperative target recognition method based on graph neural networks, characterized in that, Includes the following steps: (1) Obtain the training sample set and the test sample set: Get Each radar receives data from different angles. The HRRP data of each target is collected, and the target category is labeled after preprocessing the HRRP data of each target received by each radar. At the same time, a graph adjacency matrix is constructed for each target. Then, more than half of the preprocessed HRRP data and their labels and the corresponding target graph adjacency matrix are used to form a training sample set, and the remaining preprocessed HRRP data and the corresponding target graph adjacency matrix are used to form a test sample set. (2) Construct a multi-radar cooperative target recognition network model: Construct a multi-radar cooperative target recognition network model comprising a cascaded feature extraction subnetwork and a graph neural network recognition subnetwork; wherein the graph neural network recognition subnetwork comprises a cascaded feature extraction subnetwork and a graph neural network recognition subnetwork. The graph convolution module, consisting of a graph convolutional layer and a ReLU activation function layer loaded between adjacent graph convolutional layers, is used to perform multiple graph convolutions and nonlinear transformations on the node features based on the node features of each target, the features of the node's neighboring nodes, and the weights of the edges between the node and its neighboring nodes in the graph adjacency matrix, in order to obtain more discriminative features. (3) Iteratively train the multi-radar cooperative target recognition network model: The multi-radar cooperative target recognition network model is iteratively trained using a training sample set to obtain a well-trained target recognition network model. (4) Obtain target recognition results from multi-radar collaborative methods: By using the test sample set as input to the trained target recognition network model, the multi-radar cooperative target recognition results corresponding to the test sample set are obtained.
2. The method according to claim 1, characterized in that, The preprocessing of HRRP data for each target received by each radar, as described in step (1), is implemented as follows: The HRRP data of each target received by each radar is normalized, and the segment with the highest amplitude of the normalized HRRP data is extracted to obtain the preprocessed HRRP data of each target.
3. The method according to claim 1, characterized in that, The construction of the graph adjacency matrix for each target as described in step (1) The implementation steps are as follows: (1a) through the first Position of the radar in the global coordinate system and its distance relative to the first radar Calculate the position of each target in the global coordinate system and according to and The calculated first Range vector of each radar relative to each target Calculate the first The radar observes the azimuth angle of each target. With pitch angle : ； ； ； ； ； ； ； in, This indicates the operation of taking the mold length. This indicates the remainder operation. This is the position of the first radar in the global coordinate system. , , They represent In the global coordinate system , , Components on the axis, , , express In the global coordinate system , , Components on the axis, , , express In the global coordinate system , , Components on the axis, ; (1b) By azimuth angle With pitch angle Calculate the weight of the edge between every two radars And construct a target with the serial number of each radar as the node, and the edge weights as the nodes. The graph adjacency matrix with value ,in: ； in, , .
4. The method according to claim 1, characterized in that, The multi-radar cooperative target recognition network model described in step (2) includes: The feature extraction subnetwork includes cascaded convolutional modules consisting of multiple stacked convolutional layers and global average pooling layers; max pooling layers are loaded between adjacent convolutional layers in the convolutional modules. The graph neural recognition subnetwork consists of a global average pooling layer and a fully connected classification head cascaded at the output of the graph convolution module. The fully connected classification head includes stacked fully connected layers and a Softmax function layer.
5. The method according to claim 4, characterized in that, The iterative training of the multi-radar cooperative target recognition network model described in step (3) is implemented as follows: (3a) Initialize the number of iterations to be The maximum number of iterations is , , No. The weights in the next iteration of the multi-radar cooperative target recognition network model are: and order ; (3b) The training sample set is used as the input to the multi-radar cooperative target recognition network model, and the feature extraction sub-network extracts the first... The global features representing the scattering characteristics of each target category and possessing translation invariance in the HRRP data of each target are used to form a node feature matrix. The graph neural recognition subnetwork passes through the first Node feature matrix of each target And graph adjacency matrix For the The target row category is identified, and the first... The probability that an object is predicted as the true object category. , , This represents the total number of targets corresponding to the HRRP data in the training sample set; (3c) The cross-entropy loss function is used, and the loss value of the target recognition network model is calculated by the probability that each target is predicted as the true target category and its corresponding true target category label. Then through weight The target recognition network model for this iteration is then updated. (3d) Judgment If the condition is met, a well-trained multi-radar cooperative target recognition network model is obtained; otherwise, let... Then proceed with step (3b).
6. The method according to claim 5, characterized in that, The node feature matrix described in step (3b) This refers to using each radar as a node, and extracting the first signal received by each node through a feature extraction network. A matrix composed of global features from the HRRP data of each target.
7. The method according to claim 6, characterized in that, The step (3b) described above for the first To perform category identification on each target, the steps are as follows: In the graph convolution module The convolutional layer is based on the first graph convolutional layer Node feature matrix of each target Features of each node, features of its neighboring nodes, and the graph adjacency matrix. The weights of the edges between the node and its neighbors. Perform multiple graph convolutions on the node's features, and in each graph convolution, perform a convolution with the previous one. The ReLU activation function connected to each graph convolutional layer performs a non-linear transformation on the features after graph convolution, and the last graph convolutional layer outputs the final node feature matrix. Global average pooling layer Feature fusion is performed; the fully connected classification head identifies the global features of the fused graph to obtain the first feature. The probability that an object is predicted as the true object category. .
8. The method according to claim 7, characterized in that, The final node feature matrix described in step (3b) The method to obtain it is as follows: ； in, The ReLU activation function is used. Representing the graph adjacency matrix The normalized matrix.
9. The method according to claim 5, characterized in that, The loss value described in step (3c) The calculation formula is: ； in, This represents the logarithmic operation. For the first The true category label of each target.
10. The method according to claim 5, characterized in that, The weighting described in step (3c) The update is performed using the following formula: ； in, Represents weight The update results Indicates the learning rate. express right Take the partial derivative.