Radar high-resolution range profile recognition method based on graph neural network

Abstract

The present invention discloses a radar high-resolution range image recognition method based on graph neural network, comprising the following steps: S1, data preprocessing, S2, feature extraction, S3, output classification results, S4, repeating steps S1-S3 to complete the test, first classifying the HRRP test data through step S1-1, and then sending it to the model that has completed the training of steps S1-S3 for testing. Using the above technical solution, a method of sequence construction graph (seq2graph) is proposed, which extracts the relationship weights between nodes from the transformer, and combines them with the node features into a graph structure data, so as to facilitate processing using the graph neural network method, thereby laying a foundation for improving the recognition effect and reducing the amount of calculation.

Description

Technical Field

[0001] The present invention relates to the field of radar target recognition technology, and in particular to a radar high-resolution range image recognition method based on a graph neural network. Background Art

[0002] High-resolution range profile (HRRP) is the sum of the scattered point echoes generated by the high-resolution radar when it scans the target. Since the range resolution of the high-resolution bandwidth radar is much smaller than the target size, the fluctuations and peaks in the high-resolution radar echo reflect the relative geometric relationship of the target structure. In addition, except for the signal bandwidth, there are no other stringent requirements for the radar. The HRRP data of the target can be easily obtained, and the HRRP data is one-dimensional, so it is very convenient to store the data, and the amount of calculation is relatively reduced. Due to the above advantages, the radar automatic target recognition method based on HRRP has become a very popular direction in the field of radar target recognition.

[0003] Since the HRRP original sample data of the target to be identified is one-dimensional, its essential attributes are difficult to be directly expressed by its HRRP original data. Therefore, feature extraction and utilization of HRRP original sample data has become an important process for most current HRRP target recognition systems. Traditional HRRP feature extraction methods are mainly divided into two parts: (1) Transformer-based feature extraction methods, such as spectrograms. These methods project the HRRP signal into the frequency domain and then model and identify its frequency domain features. (2) Feature extraction methods based on reduced-dimensional data. Although traditional feature extraction methods have good recognition performance, most of these methods are unsupervised and lossy, and the selection of feature extraction methods is highly dependent on the researchers' understanding and experience of the data. Therefore, in many cases, it is difficult to achieve good results.

[0004] In recent years, due to the rise of deep learning, the above-mentioned shortcomings of traditional methods, and the pursuit of higher accuracy, people have tried to use deep learning methods to classify HRRP target samples. Common methods include convolutional neural networks (CNN), recurrent neural networks (RNN), and the application of natural language processing methods to HRRP classification. These methods are based on sequence or grid data. For CNN, the information of scattering points that are close to each other in HRRP is fully utilized. However, when the network is shallow, it is difficult to extract the features of two scattering points that are far apart in space. Therefore, it is generally necessary to stack a deeper network to achieve better results, which invisibly increases the amount of calculation. RNN has a natural advantage in processing one-dimensional sequences. It can use previous features to predict the features of the next node, that is, the new features of the current point are obtained from its left features, so the node features on its right are not extracted. Even for bidirectional RNN, the node features around a certain point can be extracted as the new features of the point, but when the sequence is very long, the features of the previous nodes will gradually weaken. BERT, a newcomer in NLP, is mainly based on the transformer network structure, which is composed of multiple transformers stacked in multiple dimensions. The transformer can effectively extract the features between two scattering points that are far apart in the HRRP sample, thereby producing better results. However, due to the stacking of a large number of transformers, the amount of calculation and memory consumption cannot be ignored. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes a radar high-resolution range profile recognition method based on graph neural network, which can not only improve the recognition effect, but also reduce the amount of calculation and improve the recognition efficiency.

[0006] A radar high-resolution range profile recognition method based on graph neural network includes the following steps:

[0007] S1. Data preprocessing

[0008] S1-1, collect raw data, HRRP data collected by radar;

[0009] S1-2, perform intensity normalization on the data in turn;

[0010] S1-3, performing gravity center alignment processing on the data after intensity normalization processing;

[0011] S1-4, converting the preprocessed sequence data into linear graph structure data;

[0012] S2. Feature extraction

[0013] S2-1, use ResGatedGCN and GCN to expand the feature dimension of linear graph structure data respectively;

[0014] S2-2, the linear graph structure data expanded by GCN is transformed into nonlinear graph structure data through Transformer, and the edges with small node associations are deleted;

[0015] S2-3, aggregate edge features and node features of nonlinear graph structure data through Graph Transformer, and add the result to the output of ResGatedGCN;

[0016] S2-4, repeat steps S2-1 to S2-3;

[0017] S3. Output classification results

[0018] S3-1, adding the linear graph structure data to the output of step S2-4 through ResGatedGCN;

[0019] S3-2, use the fully connected network and softmax to build a classifier, and the output of step S3-1 is passed through the classifier to obtain the classification result;

[0020] S4. Repeat steps S1-S3 to complete the test

[0021] The HRRP test data is first classified through step S1-1, and then sent to the model that has completed the training of steps S1-S3 for testing.

[0022] Preferably, in step S1-1, each category of HRRP data is sampled, and a training set and a test set are selected respectively, and then merged to ensure that the data form of each category is sampled, and the ratio of the training set to the test set is maintained at 7:3. The training data set is denoted as X, where x i represents the i-th sample, y i Represents the category of the i-th sample. A total of C types of targets were collected, and the total number of samples is recorded as N.

[0023] Preferably, in step S1-2, the intensity is normalized, assuming that the original HRRP data is represented by x raw =[x 1 , x 2 , ..., x M ], where M represents the total number of distance units in the HRRP data, and the normalized HRRP is expressed as:

[0024]

[0025] Preferably, in step S1-3, the center of gravity is aligned. In order to make the training and testing have a unified standard, a center of gravity alignment method is used to eliminate translation sensitivity. The center of gravity alignment method translates the HRRP so that its center of gravity is located near the center point of the HRRP. The calculation method of the center of gravity g is as follows:

[0026]

[0027] Let the sample data after preprocessing be x pre .

[0028] Preferably, in step S1-4, the originally isolated nodes of M are connected together by adding adjacent edges according to the original sequence.

[0029] The graph constructed by the above method The node data and edge data are:

[0030] V=X pre

[0031] E={e ij}

[0032] Among them, |ij|=1 and 0≤i,j≤M. The graph structure generated by this method is called a linear graph structure, denoted as

[0033]

[0034] Preferably, in S2-1,

[0035] The linear graph structure data is expanded by GCN from 1×M to 32×M, and the neighbor point features are initially extracted. The process is as follows:

[0036]

[0037] in is the adjacency matrix with self-connection, A is the adjacency matrix of the graph, I is the identity matrix of the corresponding shape, for The degree matrix of W is a trainable weight parameter, X GCN ∈R 32×M It is the output of a linear graph structure data through GCN.

[0038] The linear graph structure data is expanded by ResGatedGCN from 1×M to 32×M, and the neighbor point features are preliminarily extracted. The process is as follows:

[0039]

[0040] where xj is the feature of a single node in the linear graph structure data, N(i) is the set of all nodes connected to node i, ⊙ represents the dot product of two matrices, Represents X output by ResGatedGCN res The i-th node feature in r i，j represents the threshold function, which is defined as:

[0041] r i，j =σ(W 3 x i +W 4 x j )

[0042] where σ(·) is the sigmoid function; W 1 ~W 4 These are the weight parameters for the science department.

[0043] Preferably, in S2-2,

[0044] Using transformer to calculate the edge weight, the edge weight vector between node i and other nodes is:

[0045]

[0046] Then the weight matrix of all edges of a sample is:

[0047]

[0048] In order to reduce parameters, each node only takes the K nodes with the largest edge weights as connected nodes, and sets the edges with other nodes to 0, that is, there is no connection with the remaining nodes. The specific principle is shown in the following formula:

[0049]

[0050] in, express The largest K values ​​in the vector elements, is the weight matrix W′ e The value of the i-th row and j-th column of ;

[0051] Calculate the new representation of the node after passing through the transformer, multiply the weight obtained above with the value vector of each node, and then sum them up to get the node representation of the point:

[0052]

[0053] This can obtain the nonlinear graph structure data composed of transformer, which is recorded as in, Vt A node feature in which V t is the node feature set of nonlinear graph structure data, E t is the set of edges in the graph structure data, W e ′ is the weight matrix after parameter reduction.

[0054] Preferably, in S2-3,

[0055] Overfitting is not prevented, and the node feature set V in the linear graph structure data is l The node feature set V in the nonlinear graph structure data constructed by Transformer t Add together to form a new node feature V′ t , that is, V′ t =V t +V l , the nonlinear graph structure data is

[0056] Will Through Graph Transformer, feature expressions are further extracted by aggregating edge features and adjacent node features. So the multi-head attention mechanism for each edge from j to i is as follows:

[0057]

[0058]

[0059]

[0060]

[0061] Its reason represents the query vector of the i-th node in the c-th attention mechanism of the l-th layer, represents the key vector of the jth node in the cth attention mechanism of the lth layer, e c，ij represents the relationship weight between the i-th node and the j-th node in the c-th attention mechanism, is the exponentially scaled dot product function, d is the dimension of the key vector in each attention mechanism, That is, the attention weight between the i-th node and the j-th node in the c-th attention mechanism of the l-th layer. After obtaining the multi-head attention weight of the graph, it is aggregated from j to i:

[0062]

[0063]

[0064] in Represents the value vector of the jth node in the cth attention mechanism of the lth layer, which is composed of the feature vector of the jth node Through a fully connected layer, represents all nodes connected to node i, is the feature obtained by graphtransformer At the output of layer l, assuming there are L layers in total, the output of Graph Transformer is To prevent overfitting, the output of GraphTransformer is combined with the output of ResGatedGCN X res Add, that is

[0065] X′X res +H (L)

[0066] Repeat steps S2-1 to S2-3, and the output result is X″.

[0067] Preferably, in S3-1,

[0068] The single-dimensional feature information of the preprocessed data is obtained through ResGatedGCN, and the output node information is recorded as X′ res ;

[0069] X′ res And X″ to get the node feature expression X″′∈R 1×M .

[0070] Preferably, in S3-2,

[0071] Pass X″′ through the fully connected layer to obtain Y∈R 1×C , C is the number of species;

[0072] Pass Y through the softmax function, HRRP training sample X train The probability corresponding to the i-th target in the target set is expressed as:

[0073]

[0074] Among them, exp(·) represents the exponential operation, y c Represents the training sample x train Through the network, the cth element of the sequence is output at the fully connected layer, and the test HRRP sample x is converted to train Classification to the maximum target probability c 0 middle:

[0075]

[0076] During the training process, cross entropy is used as the loss function, which is defined as:

[0077]

[0078] Where N is the number of training samples in each batch, c i represents the category of the i-th sample, P(c i |x i ) indicates that the i-th sample is c i The probability value of the category.

[0079] The present invention has the following characteristics and beneficial effects:

[0080] 1. A sequence-to-graph method is proposed to extract the relationship weights between nodes from the transformer and combine them with the node features to form a graph structure data, which is convenient for processing using graph neural network methods, thus laying the foundation for improving recognition effects and reducing computational complexity.

[0081] 2. Using the graph network structure, it is easy to extract features between scattering points that are far apart in HRRP samples without the need to stack a very deep network structure like CNN, thereby reducing computational costs and memory consumption and improving recognition efficiency.

[0082] 3. While extracting the features between scattering points that are far apart, the structural features of the sequence itself are retained, and the two features are fused as the input of the classification layer. By recognizing multiple features, the recognition effect is further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 It is a schematic diagram of the principle of an embodiment of the present invention;

[0084] Figure 2 It is a schematic diagram of the principle of converting sequence data into linear graph structure data in an embodiment of the present invention. DETAILED DESCRIPTION

[0085] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0086] The present invention discloses a radar high-resolution range image recognition method based on graph neural network, such as Figure 1 As shown, the following steps are included:

[0087] S1. Data preprocessing

[0088] S1-1. Collect raw data, HRRP data collected by radar,

[0089] Specifically, we sampled each category of HRRP data, selected training sets and test sets respectively, and then merged them to ensure that data from each category was sampled. The ratio of training set to test set was kept at 7:3. The training data set was denoted as X, where x i represents the i-th sample, y i Represents the category of the i-th sample. A total of C types of targets were collected, and the total number of samples is recorded as N.

[0090] S1-2, in order to reduce the interference of inconsistent intensity, the data are subjected to intensity normalization processing in turn;

[0091] Specifically, intensity normalization, assuming that the original HRRP data is represented by x raw =[x 1 , x 2 , ..., x M ], where M represents the total number of distance units in the HRRP data, and the normalized HRRP is expressed as:

[0092] S1-3, in order to reduce the interference of translation sensitivity, the data after intensity normalization are subjected to center of gravity alignment processing in turn;

[0093] Specifically, in order to make training and testing have a unified standard, the center of gravity alignment method is used to eliminate translation sensitivity. The center of gravity alignment method translates the HRRP so that its center of gravity is near the center point of the HRRP. The calculation method of the center of gravity g is as follows:

[0094]

[0095] Let the sample data after preprocessing be x pre .

[0096] It can be understood that in the above technical solution, in order to reduce the influence of intensity sensitivity and translation sensitivity on prediction accuracy, it is necessary to perform intensity normalization and center of gravity alignment processing on the original data in sequence;

[0097] S1-4, converting the preprocessed sequence data into linear graph structure data;

[0098] Specifically, Figure 2 As shown, the originally isolated nodes of M are connected together by adding adjacent edges in the original sequence. The graph constructed by the above method is The node data and edge data are:

[0099] V=X pre

[0100] E={e ij}

[0101] Among them, |ij|=1 and 0≤i,j≤M. The graph structure generated by this method is called a linear graph structure, denoted as

[0102]

[0103] In the above technical solution, by converting the preprocessed sequence data into linear graph structure data, it is convenient to process it using the graph neural network method.

[0104] S2. Feature extraction

[0105] S2-1, use ResGatedGCN and GCN to expand the feature dimension of linear graph structure data respectively;

[0106] Specifically, the linear graph structure data is expanded from 1×M to 32×M through GCN, and the neighbor point features are preliminarily extracted. The process is as follows:

[0107]

[0108] in is the adjacency matrix with self-connection, A is the adjacency matrix of the graph, I is the identity matrix of the corresponding shape, for The degree matrix of W is a trainable weight parameter, X GCN ∈R 32×M It is the output of a linear graph structure data through GCN.

[0109] The linear graph structure data is expanded by ResGatedGCN from 1×M to 32×M, and the neighbor point features are preliminarily extracted. The process is as follows:

[0110]

[0111] where x j is the feature of a single node in the linear graph structure data, N(i) is the set of all nodes connected to node i, ⊙ represents the dot product of two matrices, Represents X output by ResGatedGCN res The i-th node feature in r i，j represents the threshold function, which is defined as:

[0112] η i，j =σ(W 3 x i+W 4 x j )

[0113] where σ(·) is the sigmoid function; W 1 ~W 4 These are the weight parameters for the science department.

[0114] S2-2, the linear graph structure data expanded by GCN is transformed into nonlinear graph structure data through Transformer, and the edges with small node associations are deleted;

[0115] Specifically, the transformer is used to calculate the edge weights, and the edge weight vector between node i and other nodes is:

[0116]

[0117] Then the weight matrix of all edges of a sample is:

[0118]

[0119] In order to reduce parameters, each node only takes the K nodes with the largest edge weights as connected nodes, and sets the edges with other nodes to 0, that is, there is no connection with the remaining nodes. The specific principle is shown in the following formula:

[0120]

[0121] in, express The largest K values ​​in the vector elements, is the weight matrix W′ e The value of the i-th row and j-th column of ;

[0122] Calculate the new representation of the node after passing through the transformer, multiply the weight obtained above with the value vector of each node, and then sum them up to get the node representation of the point:

[0123]

[0124] This can obtain the nonlinear graph structure data composed of transformer, which is recorded as in, V t A node feature in which V t is the node feature set of nonlinear graph structure data, E t is the set of edges in the graph structure data, W e ′ is the weight matrix after parameter reduction.

[0125] S2-3, aggregate edge features and node features of nonlinear graph structure data through Graph Transformer, and add the result to the output of ResGatedGCN;

[0126] Specifically, without preventing overfitting, the node feature set V in the linear graph structure data l The node feature set V in the nonlinear graph structure data constructed by Transformer t Add together to form a new node feature V′ t , that is, V′ t =V t +V l , the nonlinear graph structure data is

[0127] Will Through Graph Transformer, feature expressions are further extracted by aggregating edge features and adjacent node features. So the multi-head attention mechanism for each edge from j to i is as follows:

[0128]

[0129]

[0130]

[0131]

[0132] in represents the query vector of the i-th node in the c-th attention mechanism of the l-th layer, represents the key vector of the jth node in the cth attention mechanism of the lth layer, e c，ij represents the relationship weight between the i-th node and the j-th node in the c-th attention mechanism, is the exponentially scaled dot product function, d is the dimension of the key vector in each attention mechanism, That is, the attention weight between the i-th node and the j-th node in the c-th attention mechanism of the l-th layer. After obtaining the multi-head attention weight of the graph, it is aggregated from j to i:

[0133]

[0134]

[0135] in Represents the value vector of the jth node in the cth attention mechanism of the lth layer, which is composed of the feature vector of the jth node Through a fully connected layer, represents all nodes connected to node i, is the feature obtained by graphtransformer At the output of layer l, assuming there are L layers in total, the output of Graph Transformer is To prevent overfitting, the output of GraphTransformer is combined with the output of ResGatedGCN X res Add, that is

[0136] X′=X res +H (L)

[0137] S2-4, repeat steps S2-1 to S2-3, and output the result as X″.

[0138] S3. Output classification results

[0139] S3-1, adding the linear graph structure data to the output of step S2-4 through ResGatedGCN;

[0140] Specifically, the single-dimensional feature information of the preprocessed data is obtained through ResGatedGCN, and the output node information is recorded as X′ res ;

[0141] X′ res And X″ to get the node feature expression X″′∈R 1×M .

[0142] S3-2, use the fully connected network and softmax to build a classifier, and the output of step S3-1 is passed through the classifier to obtain the classification result;

[0143] Specifically, pass X″′ through the fully connected layer to obtain Y∈R 1×C , C is the number of species;

[0144] Pass Y through the softmax function, HRRP training sample X train The probability corresponding to the i-th target in the target set is expressed as:

[0145]

[0146] Among them, exp(·) represents the exponential operation, y c Represents the training sample x train Through the network, the cth element of the sequence is output at the fully connected layer, and the test HRRP sample x is converted to train Classification to the maximum target probability c 0 middle:

[0147]

[0148] During the training process, cross entropy is used as the loss function, which is defined as:

[0149]

[0150] Where N is the number of training samples in each batch, c i represents the category of the i-th sample, P(c i |x i ) indicates that the i-th sample is c i The probability value of the category.

[0151] S4. Repeat steps S1-S3 to complete the test

[0152] The HRRP test data is first classified through step S1-1, and then sent to the model that has completed the training of steps S1-S3 for testing.

[0153] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments including components are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A radar high-resolution range profile recognition method based on graph neural network, characterized in that: The following steps are involved: S1. Data preprocessing S1-1, collect raw data, HRRP data collected by radar; S1-2. Perform intensity normalization on the data in turn: Assume that the original HRRP data is represented by x raw =[x1,x2,…,x M ], where M represents the total number of distance units in the HRRP data, and the normalized HRRP is expressed as: S1-3. Perform gravity center alignment on the data after intensity normalization: The center of gravity alignment method is used to eliminate translation sensitivity. The center of gravity alignment method translates the HRRP so that its center of gravity is near the center point of the HRRP. The calculation method of the center of gravity g is as follows: Let the sample data after preprocessing be x pre ; S1-4. Convert the preprocessed sequence data into linear graph structure data: Connect the originally isolated nodes of M together by adding adjacent edges in the original sequence. The graph constructed by the above method The node data and edge data are: V=X pre And={and ij } Among them, |ij|=1 and 0≤i,j≤M. The graph structure generated by this method is called a linear graph structure, denoted as S2. Feature extraction S2-1. Use ResGatedGCN and GCN to expand the feature dimension of linear graph structure data respectively: The linear graph structure data is expanded by GCN from 1×M to 32×M, and the neighbor point features are initially extracted. The process is as follows: in is the adjacency matrix with self-connection, A is the adjacency matrix of the graph, I is the identity matrix of the corresponding shape, for The degree matrix of W is a trainable weight parameter, X GCN ∈R 32×m It is the output of a linear graph structure data through GCN; The linear graph structure data is expanded by ResGatedGCN from 1×M to 32×M, and the neighbor point features are preliminarily extracted. The process is as follows: where x j is the feature of a single node in the linear graph structure data, N(i) is the set of all nodes connected to node i, ⊙ represents the dot product of two matrices, Represents X output by ResGatedGCN res The i-th node feature in i,j represents the threshold function, which is defined as: or i,j =σ(W3x i +W4x j ) Where σ(·) is the sigmoid function; W1~W4 are the weight parameters of the science department; S2-2. The linear graph structure data expanded by GCN is transformed into nonlinear graph structure data through Transformer, and the edges with small node associations are deleted: Using transformer to calculate the edge weight, the edge weight vector between node i and other nodes is: Then the weight matrix of all edges of a sample is: In order to reduce parameters, each node only takes the K nodes with the largest edge weights as connected nodes, and sets the edges with other nodes to 0, that is, there is no connection with the remaining nodes. The specific principle is shown in the following formula: in, express The largest K values ​​in the vector elements, The weight matrix W′ behind the parameter reduction e The value of the i-th row and j-th column of ; Calculate the new representation of the node after passing through the transformer, multiply the weight obtained above with the value vector of each node, and then sum them up to get the node representation of the point: This can obtain the nonlinear graph structure data composed of transformer, which is recorded as in, V t A node feature in which V t is the node feature set of nonlinear graph structure data, E t is the set of edges in the graph structure data, W e ′ is the weight matrix after parameter reduction; S2-3. Aggregate edge features and node features of nonlinear graph structure data through Graph Transformer, and add the result to the output of ResGatedGCN: To prevent overfitting, the node feature set V in the linear graph structure data is l The node feature set V in the nonlinear graph structure data constructed by Transformer t Add together to form a new node feature V′ t , that is, V′ t =V t +V l , the nonlinear graph structure data is Will Through Graph Transformer, feature expressions are further extracted by aggregating edge features and adjacent node features. So the multi-head attention mechanism for each edge from j to i is as follows: in represents the query vector of the i-th node in the c-th attention mechanism of the l-th layer, represents the key vector of the jth node in the cth attention mechanism of the lth layer, e c,ij represents the relationship weight between the i-th node and the j-th node in the c-th attention mechanism, is the exponentially scaled dot product function, d is the dimension of the key vector in each attention mechanism, That is, the attention weight between the i-th node and the j-th node in the c-th attention mechanism of the l-th layer. After obtaining the multi-head attention weight of the graph, it is aggregated from j to i: in Represents the value vector of the jth node in the cth attention mechanism of the lth layer, which is composed of the feature vector of the jth node Through a fully connected layer, represents all nodes connected to node i, is the feature obtained by graphtransformer At the output of layer l, assuming there are L layers in total, the output of Graph Transformer is To prevent overfitting, the output of GraphTransformer is combined with the output of ResGatedGCN X res Add, that is X′=X res +H (L) Repeat steps S2-1 to S2-3, and the output result is X″; S2-4, repeat steps S2-1 to S2-3; S3. Output classification results S3-1, adding the linear graph structure data to the output of step S2-4 through ResGatedGCN; S3-2, use the fully connected network and softmax to build a classifier, and the output of step S3-1 is passed through the classifier to obtain the classification result: Pass X″′ through the fully connected layer to obtain Y∈R 1×C , C is the number of species; Pass Y through the softmax function, HRRP training sample X train The probability corresponding to the i-th target in the target set is expressed as: Among them, exp(·) represents the exponential operation, y c Represents the training sample x train Through the network, the cth element of the sequence is output at the fully connected layer, and the test HRRP sample x is converted to train Classify into the maximum target probability c0: During the training process, cross entropy is used as the loss function, which is defined as: Where N is the number of training samples in each batch, c i represents the category of the i-th sample, P(c i |x i ) indicates that the i-th sample is c i The probability value of the category; S4. Repeat steps S1-S3 to complete the test The HRRP test data is first classified through step S1-1, and then sent to the model that has completed the training of steps S1-S3 for testing.

2. The radar high-resolution range profile recognition method based on graph neural network according to claim 1 is characterized in that: In step S1-1, each category of HRRP data is sampled, and a training set and a test set are selected respectively, and then merged to ensure that the data form of each category is sampled, and the ratio of the training set to the test set is maintained at 7:

3. The training data set is denoted as X, where x i represents the i-th sample, y i Represents the category of the i-th sample. A total of C types of targets were collected, and the total number of samples is recorded as N.

Images