Radar mode recognition method based on graph kernel method of graph transfer learning

Abstract

The application discloses a radar working mode recognition method based on a graph kernel method, and belongs to the technical field of radar signal classification, and comprises the following steps: S1, using a graph kernel method to extract features of radar working mode pulse data in a radar working mode recognition task of a source domain and a target domain, and forming source domain features and target domain features; S2, based on the source domain features and the target domain features, calculating a graph kernel similarity index of the source domain and the target domain; S3, based on the size of the graph kernel similarity index, selecting a direct transfer recognition algorithm or an adaptive multi-domain transfer recognition algorithm to perform transfer learning on a graph neural network model, and using the trained graph neural network to perform radar working mode recognition on a target domain radar working mode recognition task. The method disclosed by the application can quickly and accurately recognize the working mode of an unknown radar.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar signal classification, and particularly relates to a radar working mode recognition method based on graph kernel method and graph transfer learning. BACKGROUND

[0002] Radar working mode recognition technology is crucial, however, in the face of unknown radar systems, traditional radar working mode recognition methods often rely on a large amount of labeled data and rich domain knowledge, which is difficult to meet in practical applications, especially in complex and changing environments. In addition, with the advancement of radar technology, new radar systems are constantly emerging, and their working modes are more diverse and covert, which brings greater challenges to recognition.

[0003] Under the condition of mastering a small amount of prior knowledge of unknown radar, that is, a small amount of sample data with target domain labels, how to effectively use the network trained with a large amount of known source domain radar working mode pulse data to recognize the unknown target domain radar working mode is a problem to be solved.

[0004] Graph neural networks and graph kernel methods are widely used in two-dimensional data image fields, but there are still several difficulties in adapting graph neural networks to one-dimensional data radar working mode recognition: first, pixels in two-dimensional images naturally have spatial adjacency relationships, and it is intuitive to construct subgraphs; how to effectively map one-dimensional data to graph structures to reflect pattern characteristics is a difficult point that needs to be broken through; second, the topological structure of image grid graphs is very regular, while the structure of one-dimensional signal graphs is usually highly irregular and data-driven, so how to effectively discriminate one-dimensional signals based on the constructed graph structure is also a difficult point; finally, it is a challenge to adapt the graph neural network structure, different radar working modes have different statistical characteristics and frequency domain characteristics, which requires the constructed graph neural network to have strong dynamic adaptation ability. SUMMARY

[0005] The purpose of the present application is to provide a graph neural network transfer learning radar working mode recognition method based on a graph kernel method to solve the problem that existing radar working mode recognition networks cannot quickly and accurately recognize the working mode of unknown radar with little domain prior knowledge.

[0006] In order to achieve the above purpose, the technical scheme of the present application is as follows:

[0007] The present application relates to a radar working mode recognition method based on a graph kernel method and graph transfer learning, which comprises the following steps:

[0008] S1. Using a graph kernel method to extract features from radar working mode pulse data in source domain and target domain radar working mode recognition tasks, forming source domain features and target domain features;

[0009] S2. Calculate the graph kernel similarity index of the source domain and the target domain based on the source domain features and the target domain features;

[0010] S3. Based on the size of the graph kernel similarity index, select the direct migration recognition algorithm or the adaptive multi-domain migration recognition algorithm for graph neural network model migration learning, and use the trained graph neural network to perform radar working mode recognition on the target domain radar working mode recognition task.

[0011] Preferably, the step of using the graph kernel method in S1 to extract features of radar working mode pulse data in the source domain and target domain radar working mode recognition task is:

[0012] S1.1. The radar working mode pulse sequence is divided into n units in chronological order;

[0013] S1.2. Find the source domain feature vector and the target domain feature vector using the graph kernel method;

[0014] The source domain feature vector is represented as: S i , i =1,2,…, P , i is the number of source domain feature vectors, P is the number of source domain feature vectors; the target domain feature vector is represented as: T j , j =1,2,…, Q , j is the number of source domain feature vectors, Q is the number of source domain feature vectors;

[0015] S1.3. Integrate the source domain feature vector and the target domain feature vector respectively to obtain the source domain feature vector matrix and the target domain feature vector matrix.

[0016] Preferably, the graph kernel method used in S1.2 adopts the PMK pyramid matching kernel method, and the specific steps of finding the source domain feature vector and the target domain feature vector using this method are:

[0017] S1.2.1. The n units divided in chronological order are respectively corresponding to each node in the graph neural network, forming a subgraph, the edges are constructed based on the time connection or proximity of the nodes, and the weight of the edge is calculated, the calculation formula is:

[0018] ,

[0019] wherein, ω ab indicates the nodea and nodes b between the nodes, n a , n b representing the nodes of the input graph neural network a and nodes b of the specific feature values, is a scale parameter;

[0020] S1.2.2. Constructing a layer pyramid for each subgraph, the feature space of the first K layer pyramid is divided into k grid cells, is the number of the layer, k is the feature dimension of the subgraph; d

[0021] S1.2.3. Counting the number of nodes of the subgraph within the grid, generating a histogram vector , which is the corresponding source domain feature vector or target domain feature vector,

[0022] The calculation formula of the histogram vector is as follows:

[0023] ,

[0024] wherein, c is the subgraph number corresponding to the source domain or the target domain; represents the feature value in the first grid of the k layer of the c th subgraph, and there are grid cells in total.

[0025] Preferably, the calculation formula of the S2 graph kernel similarity index is:

[0026] ,

[0027] wherein, k △ represents the graph kernel similarity index, G S is the source domain radar working mode subgraph, G T is the target domain radar working mode subgraph, and respectively represent the histogram of the source domain and the target domain radar working mode subgraph in the layer, I is a histogram intersection function, K is the number of layers of the source domain and the target domain radar working mode subgraph pyramid kernel method; the expression of the histogram intersection function is: ​

[0028] ,

[0029] wherein, n is the total number of grids of the graph neural network, i is the grid node number.

[0030] Preferably, the S3 selects a direct transfer recognition algorithm or an adaptive multi-domain transfer recognition algorithm to perform transfer learning on the graph neural network model based on the size of the graph kernel similarity index, which is: setting a similarity threshold α When the graph kernel similarity index is not less than the similarity threshold, the direct transfer recognition algorithm is selected to perform transfer learning on the graph neural network model; when the graph kernel similarity index is less than the similarity threshold, the adaptive multi-domain transfer recognition algorithm is selected to perform transfer learning on the graph neural network model.

[0031] Preferably, when the direct transfer recognition algorithm is selected to perform transfer learning on the graph neural network model in S3, the specific method is: fixing the model structure of the selected graph neural network model of the source domain radar working mode recognition task and the network parameters from the input layer to the feature extraction layer, and adaptively training the graph neural network model with the labeled data in the target domain radar working mode recognition task to construct a network suitable for the target domain radar working mode recognition task.

[0032] Preferably, when the adaptive multi-domain transfer recognition algorithm is selected to perform transfer learning on the graph neural network model in S3, the specific steps are:

[0033] S3.1. Select at least 3 source domain radar working mode recognition tasks in descending order of the graph kernel similarity index;

[0034] S3.2. Select a source domain radar working mode recognition task with the largest graph kernel similarity index to construct a graph neural network model, and the graph neural network model has a total of M layers;

[0035] S3.3. Select an unselected source domain radar working mode recognition task in descending order of the graph kernel similarity index;

[0036] S3.4. Fix the parameters of the front l layers of the graph neural network model, initialize the parameters of the l +1 to M layers, train the graph neural network model with the source domain radar working mode recognition task selected in S3.3, and update the parameters of the l +1 to M layers of the graph neural network model;

[0037] S3.5. The source domain radar operating mode recognition task selected in S3.3 is respectively brought into the graph neural network model constructed in S3.2 and the graph neural network model updated in S3.4, and the recognition rates are compared, if the recognition rate of the updated graph neural network model is better, the first layer of the graph neural network model is marked as general l , if the recognition rate of the constructed graph neural network model is better, the first layer of the graph neural network model is marked as general l , and the first layer of the graph neural network model is marked as special.

[0038] S3.6. S3.4-S3.5 is repeated until the source domain radar operating mode recognition task selected in S3.2 is migrated to the source domain radar operating mode recognition task selected in S3.3, the M layer of the graph neural network model is determined as general or special, the layer marked as special is frozen, the layer marked as general is adaptively trained, and the updated migrated graph neural network model is updated.

[0039] S3.7. It is judged whether the source domain radar operating mode recognition task selected in S3.1 is selected, if not, return to S3.3, if yes, output the final graph neural network model.

[0040] Compared with the prior art, the technical scheme provided by the present application has the following beneficial effects:

[0041] The radar operating mode recognition method based on the graph kernel method provided by the present application is used for the unique properties of unknown radar few-sample signal data, and the transfer learning using the graph neural network is introduced for the radar operating mode recognition task; the graph neural network containing the attention mechanism allows the importance weight of the adjacent node to be dynamically calculated according to the feature content when the adjacent node information is aggregated; and the graph kernel similarity feature optimized model is calculated according to the graph kernel method, and the adjacency matrix is automatically updated and optimized; the long radar operating mode signal is divided into short semantic signals, and the radar operating mode features in the long sequence are effectively extracted by using the pyramid matching kernel method, so that the calculation amount is saved, and the feature extraction of the one-dimensional radar signal is realized; the graph kernel similarity between the recognized radar signal sequence of the source domain and the radar signal sequence to be recognized of the target domain is calculated, and the distance relationship between the source domain and the target domain is established; for the radar operating mode recognition of the source domain and the target domain with large graph kernel similarity, the deep network adaptive training is used by using the transfer learning method, so that the source domain radar operating mode recognition classification knowledge can be migrated to the recognition work of the target domain with few labels and few prior knowledge; for the source domain and the target domain with small graph kernel similarity, the step-by-step migration network of the source domain and the target domain is built by using the multi-source domain, the knowledge that can be migrated in the multiple source domains is step-by-step migrated to the target domain, the negative migration problem caused by direct migration is avoided, the network model suitable for the radar operating mode recognition task of the target domain is constructed, and then the unknown radar operating mode is quickly and accurately recognized. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 A principle block diagram of a radar working mode recognition method based on a graph kernel method and graph transfer learning of the present application;

[0043] Figure 2 A network structure diagram of a graph neural network based on source domain radar working mode recognition of the present application;

[0044] Figure 3 A schematic diagram of a direct transfer recognition algorithm of the present application;

[0045] Figure 4 A schematic diagram of an adaptive multi-domain transfer recognition algorithm of the present application. DETAILED DESCRIPTION

[0046] In order to further understand the content of the present application, the present application will be described in detail in conjunction with the embodiments, and the following embodiments are used to illustrate the present application but not to limit the scope of the present application.

[0047] Referring to the accompanying drawings Figure 1 The present application relates to a radar working mode recognition method based on a graph kernel method and graph transfer learning, which comprises the following steps:

[0048] S1. Using a graph kernel method to extract features of radar working mode pulse data in a source domain and a target domain radar working mode recognition task, forming source domain features and target domain features, wherein the source domain and the target domain are both radar working mode pulse data sets, the source domain is a radar working mode data set with classification labels, and the target domain is a radar working mode data set containing rare labels, wherein the rare label means that each classification contains only a few data with labels, and the specific steps are as follows:

[0049] S1.1. Dividing the radar working mode pulse sequence into n units in chronological order;

[0050] S1.2. Using a graph kernel method to find source domain feature vectors and target domain feature vectors, wherein the graph kernel method adopts a PMK pyramid matching kernel method, and the specific steps are as follows:

[0051] S1.2.1. The n units divided in chronological order are respectively corresponding to each node in the graph neural network, that is, the graph neural network has n nodes, forming a subgraph, constructing edges based on the time connection or proximity of the nodes, and calculating the weight of the edges, and the calculation formula is as follows:

[0052] ,

[0053] wherein, ω ab indicates the nodea and nodes b The weight of the edge between n a , n b Nodes representing the input graph neural network a and nodes b The specific characteristic value of is the scale parameter;

[0054] S1.2.2. Construct for each subgraph K Layer pyramid, for the k The feature space of the layer pyramid is divided into A grid, k For which layer, d is the feature dimension of the subgraph;

[0055] S1.2.3. Count the number of nodes in the subgraph in the grid and generate a histogram vector , the histogram vector is the corresponding source domain feature vector or target domain feature vector,

[0056] The calculation formula of the histogram vector is as follows:

[0057] ,

[0058] in, c Number the subgraph corresponding to the source domain or target domain; Indicates the k Layer c The eigenvalues ​​in the first grid of the subgraph are grids;

[0059] The source domain feature vector is expressed as: S i , i =1,2,…, P , i is the number of the source domain feature vector, P is the number of source domain feature vectors; the target domain feature vector is expressed as: T j , j =1,2,…, Q , j is the number of the source domain feature vector, Q is the number of source domain feature vectors;

[0060] S1.3. Integrate the source domain eigenvectors and the target domain eigenvectors to obtain the source domain eigenvector matrix and the target domain eigenvector matrix.

[0061] The source domain eigenvector matrix Sis expressed as:

[0062] ,

[0063] The target domain feature vector is expressed as: T The target domain feature vector is expressed as:

[0064] .

[0065] S2. Based on the source domain feature and the target domain feature, the graph kernel similarity index of the source domain and the target domain is calculated, and the calculation formula of the graph kernel similarity index is:

[0066] ,

[0067] Wherein, k △ The graph kernel similarity index is expressed as: G S The source domain radar working mode recognition task is: G T The target domain radar working mode recognition task is: And The source domain and the target domain radar working mode subgraph in the layer are expressed as: The histogram intersection function is: I The number of layers of the source domain and the target domain radar working mode recognition task is: L

[0068] The expression of the histogram intersection function is:

[0069] ,

[0070] Wherein, n The total number of grid nodes of the graph neural network is: i The grid node number is:

[0071] S3. Based on the size of the graph kernel similarity index, the direct migration recognition algorithm or the adaptive multi-domain migration recognition algorithm is selected to perform transfer learning on the graph neural network model. The graph neural network model adopts the radar working mode recognition network combined with the GNN and the attention mechanism as shown in Figure 2 .

[0072] ​The input is the radar operation mode data sliced in time sequence as described above, the pulse sequence in each time unit is input into a node of the graph neural network to form a subgraph, the subgraphs are connected in the form of edges, a graph attention mechanism is added in each layer of feature extraction layer to optimize the weight calculation of the edges between each subgraph, after the feature extraction of the graph neural network with several layers of graph attention mechanism, two layers of full connection layer are added to accept the features extracted by the above graph neural network, finally the extracted features are classified and recognized by the classifier containing Softmax, the activation function is RELU, and the above steps realize the task of classifying the original radar pulse data into different radar operation modes.

[0073] The way of transferring learning of the graph neural network model based on the size of the graph kernel similarity index is to set a similarity critical value α When the graph kernel similarity index is not less than the similarity critical value, that is, It means that the similarity between the two domains reaches a certain standard, at this time, the radar operation mode classification knowledge in the source domain can be directly transferred to the target domain, therefore, the direct transfer recognition algorithm is selected to transfer the learning of the graph neural network model. The specific way of direct transfer learning is shown in Figure 3 , the model structure and network parameters of the selected graph neural network model of the source domain radar operation mode recognition task are fixed, and the graph neural network model is adaptively trained with the rare labeled data in the target domain radar operation mode recognition task, and a network suitable for the target domain radar operation mode recognition task is further constructed, wherein the node feature vector of the 0th layer in the source domain and the target domain structure is represented as h v0 ; C v0 Then the feature vector after extracting the surrounding associated node information by the attention mechanism is y i and y j represent the output source domain node classification output and target domain node classification output respectively.

[0074] When the graph kernel similarity index is less than the similarity critical value, that is, It can be considered that the source domain has less relevance to the target domain, and the radar working mode classification knowledge of the source domain cannot be directly migrated to the target domain. The neural network extracts features that can be divided into general features and special features. The general features have less relevance to the task and can be directly used for other tasks. The special features have strong task relevance and are only applicable to specific tasks. The features in the early layers of the network show stronger generality, while the features in the later layers have stronger specificity. In order to avoid the negative transfer problem caused by directly migrating the source domain radar working mode recognition model to the target domain radar, an adaptive multi-domain transfer recognition algorithm is selected to perform transfer learning on the graph neural network model. The specific steps are as shown in Figure 4

[0075] S3.1. Select at least 3 source domain radar working mode recognition tasks in descending order of graph kernel similarity index. In this embodiment, 3 source domain radar working mode recognition tasks are selected S A , S B , S C ,

[0076] S3.2. Select the source domain radar working mode recognition task with the maximum graph kernel similarity index S A . N S A , the graph neural network model has a total of M layers;

[0077] S3.3. Select an unselected source domain radar working mode recognition task in descending order of graph kernel similarity index S B .

[0078] S3.4. Fix the parameters of the first N S A layers of the graph neural network model l l +1 to M layers are initialized, and the source domain radar working mode recognition task selected in S3.3 is used S B to train the graph neural network model and update the parameters of the l +1 to M layers of the graph neural network model;

[0079] ​​​​S3.5. Substitute the source domain radar working mode recognition task selected in S3.3 into the graph neural network model constructed in S3.2 and the graph neural network model updated in S3.4, and compare the recognition rates. If the recognition rate of the updated graph neural network model is better, it proves that the source domain radar working mode recognition task selected in S3.2 is better. S A Migrate to the source domain radar working mode recognition task selected in S3.3 S B In the process, the first l The layer is general. If the recognition rate of the constructed graph neural network model is better, it proves that the source domain radar working mode recognition task selected in S3.2 S A Migrate to the source domain radar working mode recognition task selected in S3.3 S B In the process, the first l Layers have specificity;

[0080] S3.6. Repeat S3.4-S3.5 until the source domain radar working mode recognition task selected in S3.2 is transferred to the source domain radar working mode recognition task selected in S3.3. M Each layer is judged as general or special, and then the S A Migrate to S B In the process N ( S A ) The generality and specificity of each layer, freeze the layers marked as specific, adaptively train the layers marked as general, and update the migrated graph neural network model, denoted as N ( S A* S B );

[0081] S3.7. Determine whether all the source domain radar working mode recognition tasks selected in S3.1 have been selected. If not, return to S3.3. If so, output the final graph neural network model and finally build a model suitable for the target task. T D Network N ( S A* S B* S C* T D ).

[0082] Finally, the trained graph neural network is used to perform radar mode recognition on a target domain radar mode recognition task.

[0083] The application has been described in detail with reference to the embodiments above, but the above description is only the preferred embodiments of the application and should not be considered as limiting the scope of the application. Any equivalent changes and improvements made in the scope of the application should still fall within the scope of the patent.

Claims

1. A radar operating mode recognition method based on graph kernel method-based graph transfer learning, characterized in that: It comprises the following steps: S1. Feature extraction is performed on the radar working mode pulse data in the source domain and target domain radar working mode recognition tasks using a graph kernel method, forming source domain features and target domain features, and the steps are as follows: S1.

1. divide the sequence of radar operating mode pulses into n units in chronological order; S1.

2. The source domain feature vector and the target domain feature vector are found using the graph kernel method, and the specific steps are as follows: S1.2.

1. dividing the time sequence into n Each unit corresponds to each node in the graph neural network, forming a subgraph. The edges are constructed based on the time connection or proximity of the nodes, and the weight of the edges is calculated. The calculation formula is: ， wherein, ω ab representing a weight of an edge between a node a and a node b , n a , n b representing a specific feature value of a node a and a node b input to a graph neural network, is a scale parameter; S1.2.

2. Constructing for each sub-graph K layer pyramid, for the k feature space of the layer pyramid is divided into grids, k for the layer, d is the feature dimension of the sub-graph; S1.2.

3. Count the number of nodes of the subgraph within the grid, generate a histogram vector The histogram vector is the corresponding source domain feature vector or target domain feature vector. The calculation formula of the histogram vector is as follows: ， wherein, c is the subgraph number corresponding to the source domain or the target domain; represents the first k layer in the c first grid of the total of grid. The source domain feature vector is represented as: S i , i =1, 2, …, P , i is the number of source domain feature vectors, P is the number of source domain feature vectors; the target domain feature vector is represented as: T j , j =1, 2, …, Q , j is the number of source domain feature vectors, Q is the number of source domain feature vectors; S1.

3. The source domain feature vector and the target domain feature vector are integrated respectively to obtain the source domain feature vector matrix and the target domain feature vector matrix; S2. Based on the source domain features and the target domain features, the graph kernel similarity index of the source domain and the target domain is calculated, and the calculation formula is as follows: ， wherein, k △ denotes a graph kernel similarity index, G S is a source domain radar operating mode subgraph, G T is a target domain radar operating mode subgraph, and denotes a histogram of the source domain and target domain radar operating mode subgraph in the layer, I is a histogram intersection function, K is a number of layers of the source domain and target domain radar operating mode subgraph pyramid kernel method; and an expression of the histogram intersection function is: ， wherein, n is the total number of grids for the graph neural network, i is the grid node number; S3. Based on the size of the graph kernel similarity index, select the direct migration identification algorithm or the adaptive multi-domain migration identification algorithm for graph neural network model transfer learning, the specific way is: set the similarity threshold α When the graph kernel similarity index is not less than the similarity threshold, the direct migration identification algorithm is selected for graph neural network model transfer learning; when the graph kernel similarity index is less than the similarity threshold, the adaptive multi-domain migration identification algorithm is selected for graph neural network model transfer learning; and the trained graph neural network is used for radar working mode recognition of target domain radar working mode recognition task.

2. The graph kernel method based graph transfer learning radar operating mode recognition method according to claim 1, characterized in that: When the direct transfer recognition algorithm is selected to perform transfer learning on the graph neural network model in S3, the specific method is: the model structure and the network parameters from the input layer to the feature extraction layer of the selected graph neural network model of the source domain radar working mode recognition task are fixed, the graph neural network model is adaptively trained using the labeled data in the target domain radar working mode recognition task, and a network suitable for the target domain radar working mode recognition task is constructed.

3. The graph kernel method based graph transfer learning radar operating mode recognition method according to claim 1, characterized in that: When the adaptive multi-domain transfer recognition algorithm is selected to perform transfer learning on the graph neural network model in S3, the specific steps are as follows: S3.

1. At least three source domain radar working mode recognition tasks are selected in descending order of the graph kernel similarity index; S3.

2. Select the source domain radar working mode recognition task to build a graph neural network model with the maximum graph kernel similarity index, and the graph neural network model is composed of M layers; S3.

3. An unselected source domain radar working mode recognition task is selected in descending order of the graph kernel similarity index; S3.

4. fixing parameters of the front l layers of the graph neural network model, initializing parameters of the l +1 to M layers of the graph neural network model, training the graph neural network model on the source domain radar operating mode recognition task selected in S3.3, and updating the parameters of the l +1 to M layers of the graph neural network model; S3.

5. Bring the source domain radar working mode recognition task selected in S3.3 into the graph neural network model constructed in S3.2 and the graph neural network model updated in S3.4, respectively, and compare the recognition rates. If the recognition rate of the updated graph neural network model is better, mark the graph neural network model as the first l The layer is general. If the recognition rate of the constructed graph neural network model is better, the first layer of the labeled graph neural network model l Layers have specificity; S3.

6. Repeat S3.4-S3.5 until the source domain radar working mode recognition task selected in S3.2 is migrated to the source domain radar working mode recognition task selected in S3.3, and the parameters of the graph neural network model in the process are updated M all layers are determined as general or special, the layers marked as special are frozen, the layers marked as general are adaptively trained, and the migrated graph neural network model is updated; S3.

7. It is judged whether the source domain radar working mode recognition tasks selected in S3.1 have all been selected, if not, return to S3.3, if yes, output the final graph neural network model.

Images