CN114417938B - Electromagnetic target classification method embedded by using knowledge vector - Google Patents
Electromagnetic target classification method embedded by using knowledge vector Download PDFInfo
- Publication number
- CN114417938B CN114417938B CN202210101982.8A CN202210101982A CN114417938B CN 114417938 B CN114417938 B CN 114417938B CN 202210101982 A CN202210101982 A CN 202210101982A CN 114417938 B CN114417938 B CN 114417938B
- Authority
- CN
- China
- Prior art keywords
- neural network
- graph
- electromagnetic target
- vector
- electromagnetic
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Signal Processing (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
- Complex Calculations (AREA)
Abstract
The invention discloses an electromagnetic target classification method utilizing knowledge vector embedding, which utilizes the data of known electromagnetic target information to establish a graph structure of an electromagnetic target, and carries out embedded vector representation on graph nodes corresponding to each electromagnetic target class based on a graph neural network; collecting an electromagnetic target signal, performing short-time Fourier transform on the electromagnetic target data to obtain time-frequency data of the electromagnetic target signal, and preprocessing the electromagnetic target data to serve as a sample for training a convolutional neural network; constructing a convolutional neural network, training the convolutional neural network based on the result of the embedded vector representation of the graph node corresponding to the electromagnetic target class, and finally obtaining a reference vector for classifying and identifying the acquired electromagnetic target signal; and classifying and identifying the acquired electromagnetic target signals by using the obtained reference vectors. The method has strong applicability, combines the category relation knowledge into the network training, and solves the defect that the traditional classification network is only suitable for identifying the categories which appear in the training set.
Description
Technical Field
The invention relates to the field of communication and artificial intelligence, in particular to an electromagnetic target classification method embedded by using a knowledge vector.
Background
In the knowledge graph, the knowledge of the graph structure mainly includes node knowledge and relationship knowledge. The nodes can represent a certain known category, the relation is the relation degree between each category, the relation degree between the nodes can be intuitively reflected, and after the graph structure is embedded into a vector form, the vector characterization of each node can reflect the relation knowledge between the nodes. In recent years, due to the development of the graph neural network, the attribute and the interrelationship of the nodes in the graph structure can be effectively utilized, modeling is carried out according to the interrelationship, the nodes are embedded into vectors, and a computable data format is provided for the processing of subsequent computers. At present, when a graph neural network is used for sample embedding and then electromagnetic targets are classified, the following two problems mainly exist: firstly, the graph neural network architecture is mainly used for classifying nodes, and has lower execution efficiency for electromagnetic target classification tasks embedded by multiple samples; secondly, the graph neural network needs prior relation knowledge of each node and other nodes, and in a practical situation, the relation information is difficult to reflect on a sample level. Convolutional neural networks can classify different inputs according to labels, and conventional convolutional neural networks have the following disadvantages when used for classifying graph structures: (1) The label in the traditional method can not reflect the tightness degree of the relation among samples; (2) The output of a conventional convolutional neural network cannot be an embedded representation of knowledge with a graph structure.
Disclosure of Invention
Aiming at the problem of how to embed a priori knowledge sample with a graph structure and classify electromagnetic targets, the method combines the characteristics of two neural networks with different structures, carries out vector embedding on the sample with the priori knowledge structure, and then uses the sample with the priori knowledge structure in a technology of classifying the neural network training to realize fusion of the prior knowledge such as category relation of the electromagnetic targets and electromagnetic data characteristics in a vector embedding space, thereby constructing an electromagnetic target classifier driven by the data and the knowledge together. According to the degree of closeness of the relations among the classes of the samples, an embedding result which can reflect the relations among the samples can be obtained. The distance of the embedded vector can be used to reflect the degree of correlation of features between samples. The invention converts the traditional hard decision of the classification neural network represented by the convolutional neural network into the soft decision for calculating the similarity between vectors, thereby solving the problem that the current sample embedding result can not reflect the relationship between the classes to which the sample belongs and the hard decision when the classification decision is performed.
The method is generally divided into two steps, namely embedding the graph nodes corresponding to the categories containing the correlations based on the graph neural network, and training the convolutional neural network based on the embedding results of the graph neural network, so that the embedding results of the samples under the categories can reflect the relationships of the categories to which the samples belong.
The invention discloses an electromagnetic target classification method embedded by using a knowledge vector, which comprises the following specific steps: establishing a graph structure of the electromagnetic targets by using the data of the known electromagnetic target information, wherein the graph structure comprises graph nodes and relations, the graph nodes are used for representing the known electromagnetic target categories, and the relations are used for representing the association degree between each electromagnetic target category;
s1, based on the graph neural network, embedding vector representation is carried out on graph nodes corresponding to each electromagnetic target class. The basic description process of the electromagnetic target class corresponding graph node comprises the following steps: the relation between the classes of the electromagnetic targets is represented by an adjacent matrix D, and the element D of the ith row and the jth column of the adjacent matrix D ij And representing the relation between the ith electromagnetic target class and the jth electromagnetic target class, wherein the characteristics of all the electromagnetic target classes are represented by a matrix F, the ith row of the matrix F represents the characteristics of the ith electromagnetic target class, D and F are input into a graph neural network, and the graph neural network is trained to obtain the embedded vector representation of all the electromagnetic target classes corresponding to the graph nodes.
The step S1 specifically comprises the following steps:
in the graph nerve, reLU (·) is a linear rectification function; II indicates the l2 norm; softmax (·) represents a logistic regression function; e is an identity matrix, N represents the dimension of the identity matrix E and is also the layer number of the neural network of the graph; w (W) i Weight, W, of the ith hidden layer of the graph neural network 0 ∈R M×N ,W 1 ∈R N×H M is the dimension of the feature vector of the graph node, H is the dimension of the embedded vector output by the graph neural network, and k is the iteration number when training the graph neural network;
wherein (1)>
Representing a normalized Laplace matrix,/->
Representation matrix->
Representation matrix->
Elements of row i, column i,/->
Representation matrix->
Elements of row i, column j,/->
Representing an undirected graph adjacency matrix; l (L) i A tag that is the ith electromagnetic target class; q is the embedded vector representation of the electromagnetic target class correspondence map node.
Training the graph neural network to obtain embedded vector representations of all electromagnetic target class corresponding graph nodes, wherein the specific steps comprise:
s11, initializing a weight matrix W of the graph neural network i I=0, 1,..n, N represents the number of layers of the graph neural network.
S12, calculating the output Y of the graph neural network, wherein the calculation formula is as follows:
wherein f () represents the computational function of the neural network of the graph;
s13, under the output condition, calculating a Loss function Loss of the graph neural network:
wherein L is l A label representing the first electromagnetic target class, F representing a row in the matrix F, i.e. a feature vector of a certain electromagnetic target class, Y f The output obtained after the characteristic vector f is input into the graph neural network; l represents a set of tags of the electromagnetic target class.
S14, updating the weight of the graph neural network by using a batch gradient descent method (BGD) according to the loss function; the iteration number k is added with 1;
s15, repeating the processes from the step S12 to the step S14 until the iteration number k reaches a preset value;
s16, calculating an embedded vector representation Q of the electromagnetic target class corresponding graph node, wherein the calculation formula is as follows:
thus obtaining the embedded vector representation of the electromagnetic target class corresponding graph node.
S2, preprocessing an electromagnetic target signal;
and acquiring an electromagnetic target signal, storing the electromagnetic target signal according to acquisition time, performing short-time Fourier transform on the electromagnetic target data to acquire time-frequency data of the electromagnetic target signal, preprocessing the time-frequency data, and transforming the time-frequency data into a data format which can be processed by the convolutional neural network to serve as a sample for training the convolutional neural network.
S3, constructing a convolutional neural network, training the convolutional neural network based on the result of the embedded vector representation of the graph node corresponding to the electromagnetic target class obtained by the graph neural network, and finally obtaining a reference vector for classifying and identifying the acquired electromagnetic target signal; and utilizing a convolutional neural network to realize the mapping from the embedded representation of the signal characteristics of the electromagnetic target class to the embedded vectors of the electromagnetic target class corresponding graph nodes.
The step S3 specifically comprises the following steps:
the embedded vector representation of the graph node corresponding to the class to which the sample belongs is obtained through the step S1, the embedded vector of the graph node corresponding to the class to which the sample belongs is used as a label of the convolutional neural network, the time-frequency data of the electromagnetic target preprocessed in the step S2 is used as a training sample, and the distance between the embedded vector of the graph node corresponding to the class to which the sample belongs and the embedded vector of the training sample is used as a loss function for training the convolutional neural network. And finally obtaining a reference vector for classifying and identifying the acquired electromagnetic target signal after performing repeated iterative optimization training on the convolutional neural network. The multi-iteration optimization training process for the convolutional neural network specifically comprises the following steps:
s31, constructing a convolutional neural network with a two-layer convolutional layer, two-layer pooling layer and two-layer full-connection layer structure, and initializing graph node weights of the network; the convolution layer, the pooling layer and the full-connection layer are sequentially connected;
s32, taking the time frequency data P of the electromagnetic target preprocessed in the step S2 as a training sample to be input into the convolutional neural network, and obtaining a feature vector R of the training sample at the last layer of the convolutional neural network;
s33, calculating a loss function of the network: l= - |r-q| where Q is the embedded vector representation of the graph node corresponding to the class to which the input training sample belongs, II represents the norms;
s34, optimizing parameters of the convolutional neural network according to a random gradient descent method, and updating weights of graph nodes in the network;
s35, repeating the steps S32 to S34 until the iteration times reach a preset value, and completing training of the convolutional neural network;
s36, after training of the convolutional neural network is completed, the obtained output R is the final embedded vector representation result of the sample;
s37, calculating an average value Rc of embedded vector representation results of sample data of the same category, and taking Rc as a reference vector for classifying and identifying the collected electromagnetic target signals subsequently;
s4, classifying and identifying the collected electromagnetic target signals by using the reference vector obtained in the step S3, wherein the specific steps comprise:
s41, performing short-time Fourier transform and preprocessing on the acquired electromagnetic target signal to obtain a sample P 'to be identified, and inputting the sample P' to be identified into the convolutional neural network trained in the step S3.
S42, calculating to obtain an output vector R 'corresponding to the sample P' to be identified through a convolutional neural network.
S43, calculating the vector similarity of the average value of the embedded vector representation results of the output vector R 'and the sample data of each class, wherein the vector similarity of the output vector R' and the ith class is represented as epsilon i I=1,..n, n is the total number of classes, the vector similarity is obtained by calculating pearson correlation coefficients of the two vectors.
S44, comparing the obtained vector similarity with a threshold value eta, if epsilon exists j More than or equal to eta, j epsilon n, judging that the sample P' to be identified belongs to the j-th electromagnetic target class; if to arbitrary epsilon j Has epsilon j And < eta, j epsilon n, judging that the sample P' to be identified belongs to a new electromagnetic target class.
The beneficial effects of the invention are as follows:
the method has strong applicability and wide application range, combines the category relation knowledge into the network training, and solves the defect that the traditional classification network is only suitable for identifying the categories appearing in the training set. And the embedding result of the sample is more effective by utilizing the relation information of the category to which the sample belongs. The invention adopts the embedding result based on the graph neural network to train the convolutional neural network, and effectively reflects the relation information of the category of the sample on the embedding result. Facilitating subsequent analysis of the sample and downstream applications.
Drawings
FIG. 1 is a flow chart of an implementation of the method of the present invention;
fig. 2 is a relationship structure diagram of a category correspondence node in the present embodiment;
FIG. 3 is a diagram showing the result of the dimension reduction of the class-embedded vector in the present embodiment;
fig. 4 is a diagram showing the result of the dimension reduction of the sample embedding vector in the present embodiment.
Detailed Description
For a better understanding of the present disclosure, an embodiment is presented herein.
FIG. 1 is a flow chart of an implementation of the method of the present invention; fig. 2 is a relationship structure diagram of a category correspondence node in the present embodiment; FIG. 3 is a diagram showing the result of the dimension reduction of the class-embedded vector in the present embodiment; fig. 4 is a diagram showing the result of the dimension reduction of the sample embedding vector in the present embodiment.
The invention discloses an electromagnetic target classification method embedded by using a knowledge vector, which comprises the following specific steps: establishing a graph structure of the electromagnetic targets by using the data of the known electromagnetic target information, wherein the graph structure comprises graph nodes and relations, the graph nodes are used for representing the known electromagnetic target categories, and the relations are used for representing the association degree between each electromagnetic target category;
s1, based on the graph neural network, embedding vector representation is carried out on graph nodes corresponding to each electromagnetic target class. The basic description process of the electromagnetic target class corresponding graph node comprises the following steps: the relation between the classes of the electromagnetic targets is represented by an adjacent matrix D, and the element D of the ith row and the jth column of the adjacent matrix D ij And representing the relation between the ith electromagnetic target class and the jth electromagnetic target class, wherein the characteristics of all the electromagnetic target classes are represented by a matrix F, the ith row of the matrix F represents the characteristics of the ith electromagnetic target class, D and F are input into a graph neural network, and the graph neural network is trained to obtain the embedded vector representation of all the electromagnetic target classes corresponding to the graph nodes.
In the graph nerve, reLU (·) is a linear rectification function; II indicates 12 norms; softmax (·) represents a logistic regression function; e is an identity matrix, N represents the dimension of the identity matrix E and is also the layer number of the neural network of the graph; w (W) i Weight, W, of the ith hidden layer of the graph neural network 0 ∈R M×N ,W 1 ∈R N×H M is the dimension of the feature vector of the graph node, H is the dimension of the embedded vector output by the graph neural network, and k is the iteration number when training the graph neural network;
wherein (1)>
Representing a normalized Laplace matrix,/->
Representation matrix->
Representation matrix->
Elements of row i, column i,/->
Representation matrix->
Elements of row i, column j,/->
Representing an undirected graph adjacency matrix; l (L) i A tag that is the ith electromagnetic target class; q is the embedded vector representation of the electromagnetic target class correspondence map node.
Training the graph neural network to obtain embedded vector representations of all electromagnetic target class corresponding graph nodes, wherein the specific steps comprise:
s11, initializing a weight matrix W of the graph neural network i I=0, 1,..n, N represents the number of layers of the graph neural network.
S12, calculating the output Y of the graph neural network, wherein the calculation formula is as follows:
wherein f () represents the computational function of the neural network of the graph;
s13, under the output condition, calculating a Loss function Loss of the graph neural network:
wherein L is l A label representing the first electromagnetic target class, F representing a row in the matrix F, i.e. a feature vector of a certain electromagnetic target class, Y f The output obtained after the characteristic vector f is input into the graph neural network; l represents a set of tags of electromagnetic target class;
the feature vectors representing the electromagnetic target classes are input into the graph neural network, and the sum of all the outputs obtained by the graph neural network is calculated. The characteristic vector of a certain electromagnetic target class, i.e. the vector formed by a plurality of specific parameters of a certain electromagnetic target class.
S14, updating the weight of the graph neural network by using a batch gradient descent method (BGD) according to the loss function; the iteration number k is added with 1;
s15, repeating the processes from the step S12 to the step S14 until the iteration number k reaches a preset value;
s16, calculating an embedded vector representation Q of the electromagnetic target class corresponding graph node, wherein the calculation formula is as follows:
s2, preprocessing an electromagnetic target signal;
the method comprises the steps of collecting electromagnetic target signals, storing the electromagnetic target signals according to the collection time, carrying out short-time Fourier transform on the electromagnetic target data to obtain time-frequency data of the electromagnetic target signals, preprocessing the electromagnetic target data, converting the time-frequency data into a data format which can be processed by the convolutional neural network, and taking the data format which can be processed by the convolutional neural network as a sample for training the convolutional neural network, wherein the data format which can be processed by the convolutional neural network comprises pictures or matrix data, so that the convolutional neural network can carry out convolutional operation conveniently.
S3, constructing a convolutional neural network, training the convolutional neural network based on the result of the embedded vector representation of the graph node corresponding to the electromagnetic target class obtained by the graph neural network, and finally obtaining a reference vector for classifying and identifying the acquired electromagnetic target signal; and utilizing a convolutional neural network to realize the mapping from the embedded representation of the signal characteristics of the electromagnetic target class to the embedded vectors of the electromagnetic target class corresponding graph nodes.
The embedded vector representation of the graph node corresponding to the class to which the sample belongs is obtained through the step S1, the graph node representation corresponding to the class can reflect the degree of closeness of the relations between the classes, the embedded vector of the graph node corresponding to the class to which the sample belongs is used as a label of the convolutional neural network, the time-frequency data of the electromagnetic target preprocessed in the step S2 is used as a training sample, and the distance between the embedded vector of the graph node corresponding to the class to which the sample belongs and the embedded vector of the training sample is used as a loss function for training the convolutional neural network. And finally obtaining a reference vector for classifying and identifying the acquired electromagnetic target signal after performing repeated iterative optimization training on the convolutional neural network. The multi-iteration optimization training process for the convolutional neural network specifically comprises the following steps:
s31, constructing a convolutional neural network with a two-layer convolutional layer, two-layer pooling layer and two-layer full-connection layer structure, and initializing graph node weights of the network; the convolution layer, the pooling layer and the full-connection layer are sequentially connected;
s32, taking the time frequency data P of the electromagnetic target preprocessed in the step S2 as a training sample to be input into the convolutional neural network, and obtaining a feature vector R of the training sample at the last layer of the convolutional neural network;
s33, calculating a loss function of the network: l= - |r-q| where Q is the embedded vector representation of the graph node corresponding to the class to which the input training sample belongs, II represents the norms;
s34, optimizing parameters of the convolutional neural network according to a random gradient descent method, and updating weights of graph nodes in the network;
s35, repeating the steps S32 to S34 until the iteration times reach a preset value, and completing training of the convolutional neural network;
s36, after training of the convolutional neural network is completed, the obtained output R is the final embedded vector representation result of the sample;
s37, calculating an average value Rc of embedded vector representation results of sample data of the same category, and taking Rc as a reference vector for classifying and identifying the collected electromagnetic target signals subsequently;
s4, classifying and identifying the collected electromagnetic target signals by using the reference vector obtained in the step S3, wherein the specific steps comprise:
s41, performing short-time Fourier transform and preprocessing on the acquired electromagnetic target signal to obtain a sample P 'to be identified, and inputting the sample P' to be identified into the convolutional neural network trained in the step S3.
S42, calculating to obtain an output vector R 'corresponding to the sample P' to be identified through a convolutional neural network.
S43, calculating the vector similarity of the average value of the embedded vector representation results of the output vector R 'and the sample data of each class, wherein the vector similarity of the output vector R' and the ith class is represented as epsilon i I=1,..n, n is the total number of classes, the vector similarity is obtained by calculating pearson correlation coefficients of the two vectors.
S44, comparing the obtained vector similarity with a threshold value eta, if epsilon exists j More than or equal to eta, j epsilon n, judging that the sample P' to be identified belongs to the j-th electromagnetic target class; if to arbitrary epsilon j Has epsilon j And < eta, j epsilon n, judging that the sample P' to be identified belongs to a new electromagnetic target class.
By training the convolutional neural network based on the embedding results of the graph neural network, an embedded representation of the samples of each class may be ultimately obtained. Classification recognition based on the similarity of the embedded vectors can be performed using the embedded representation. The sample embedding result can reflect the similarity between samples and the closeness of the category relationship of the samples. When the similarity is smaller than the set threshold, the new class is determined.
And embedding the graph nodes corresponding to the categories containing the mutual relation information based on the graph neural network, and obtaining the embedding result of the nodes corresponding to each category through training the graph neural network. The distance between the embedded vectors can better reflect the degree of interconnection of the nodes. Fig. 2 is a verification result of embedding a node using the graph neural network.
For class nodes, there are 15 nodes corresponding to the class, the size of the adjacent matrix a is 15×15, and the size of the node characteristic matrix X is 15×15. The convolution layer of the selected graph neural network is two layers, the size of the embedded result Q is 15 multiplied by 128, and the dimension of the embedded vector of each node is 128 dimensions.
The result of the relation structure of the class corresponding node and the embedded vector after the dimension reduction is shown in fig. 3.
Training a convolutional neural network based on the graph neural network embedding result; after the embedding of the class to which the sample belongs is completed, the sample needs to be embedded by using the prior knowledge of the class embedding result. There are 100 samples in each class, there are 1500 samples in total in 15 classes, and the samples are subjected to short-time Fourier transform to obtain a time-frequency diagram. The embedded vector obtained by a convolutional neural network with a two-layer convolutional layer, a two-layer pooling layer and a two-layer fully-connected layer structure is 128-dimensional. The distance between the embedded vectors of the samples may reflect the degree of closeness of the relationship between the samples. After the embedded vector of samples is reduced in dimension, the result shown in fig. 4 can be obtained.
After the sample embedding result is obtained, the subsequent analysis of sample data and the downstream application based on the embedding vector are facilitated.
The foregoing is merely exemplary of the present application and is not intended to limit the present application. Various modifications and changes may be made to the present application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. which are within the spirit and principles of the present application are intended to be included within the scope of the claims of the present application.
Claims (2)
1. The electromagnetic target classification method embedded by using the knowledge vector is characterized in that a graph structure of an electromagnetic target is established by using data of known electromagnetic target information, wherein the graph structure comprises graph nodes and relations, the graph nodes are used for representing known electromagnetic target categories, and the relations are used for representing the association degree between each electromagnetic target category; the method comprises the following specific steps:
s1, carrying out embedded vector representation on graph nodes corresponding to each electromagnetic target class based on a graph neural network; the description process of the electromagnetic target class corresponding graph node comprises the following steps: the relation between the classes of the electromagnetic targets is represented by an adjacent matrix D, and the element D of the ith row and the jth column of the adjacent matrix D ij Representing the relation between the ith electromagnetic target class and the jth electromagnetic target class, wherein the characteristics of all the electromagnetic target classes are represented by a matrix F, the ith row of the matrix F represents the characteristics of the ith electromagnetic target class, inputting D and F into a graph neural network, and training the graph neural network to obtain embedded vector representations of all the electromagnetic target classes corresponding to graph nodes;
s2, preprocessing an electromagnetic target signal;
collecting an electromagnetic target signal, storing the electromagnetic target signal according to the collection time, performing short-time Fourier transform on the electromagnetic target data to obtain time-frequency data of the electromagnetic target signal, preprocessing the time-frequency data, and transforming the time-frequency data into a data format which can be processed by a convolutional neural network to be used as a sample for training the convolutional neural network;
s3, constructing a convolutional neural network, training the convolutional neural network based on the result of the embedded vector representation of the graph node corresponding to the electromagnetic target class obtained by the graph neural network, and finally obtaining a reference vector for classifying and identifying the acquired electromagnetic target signal; the convolution neural network is utilized to realize the mapping from the embedded representation of the signal characteristics of the electromagnetic target class to the embedded vector of the electromagnetic target class corresponding graph node;
s4, classifying and identifying the collected electromagnetic target signals by using the reference vector obtained in the step S3;
the step S3 specifically comprises the following steps:
the embedded vector representation of the graph node corresponding to the class to which the sample belongs is obtained through the step S1, the embedded vector of the graph node corresponding to the class to which the sample belongs is used as a label of the convolutional neural network, the time-frequency data of the electromagnetic target preprocessed in the step S2 is used as a training sample, and the distance between the embedded vector of the graph node corresponding to the class to which the sample belongs and the embedded vector of the training sample is used as a loss function for training the convolutional neural network; after performing repeated iterative optimization training on the convolutional neural network, finally obtaining a reference vector for classifying and identifying the acquired electromagnetic target signals; the multi-iteration optimization training process for the convolutional neural network specifically comprises the following steps:
s31, constructing a convolutional neural network with a two-layer convolutional layer, two-layer pooling layer and two-layer full-connection layer structure, and initializing graph node weights of the network; the convolution layer, the pooling layer and the full-connection layer are sequentially connected;
s32, taking the time frequency data P of the electromagnetic target preprocessed in the step S2 as a training sample to be input into the convolutional neural network, and obtaining a feature vector R of the training sample at the last layer of the convolutional neural network;
s33, calculating a loss function of the network: l= - |r-q| where Q is the embedded vector representation of the graph node corresponding to the class to which the input training sample belongs, II represents the norms;
s34, optimizing parameters of the convolutional neural network according to a random gradient descent method, and updating weights of graph nodes in the network;
s35, repeating the steps S32 to S34 until the iteration times reach a preset value, and completing training of the convolutional neural network;
s36, after training of the convolutional neural network is completed, the obtained output R is the final embedded vector representation result of the sample;
s37, calculating an average value Rc of embedded vector representation results of sample data of the same category, and taking Rc as a reference vector for classifying and identifying the collected electromagnetic target signals subsequently;
the step S1 specifically comprises the following steps:
in the graph nerve, reLU (·) is a linear rectification function; II indicates the l2 norm;
softmax (·) represents a logistic regression function; e is an identity matrix, N represents the dimension of the identity matrix E and is also the layer number of the neural network of the graph; w (W) i Weight, W, of the ith hidden layer of the graph neural network 0 ∈R M×N ，W 1 ∈R N×H M is the dimension of the feature vector of the graph node, H is the dimension of the embedded vector output by the graph neural network, and k is the iteration number when training the graph neural network;
wherein (1)>
Representing a normalized Laplace matrix,/->
Representation matrix->
Representation matrix->
Elements of row i, column i,/->
Representation matrix->
Elements of row i, column j,/->
Representing an undirected graph adjacency matrix; l (L) i A tag that is the ith electromagnetic target class; q is the embedded vector representation of the electromagnetic target class corresponding graph node; the size of the adjacent matrix A is 15 multiplied by 15, and the size of the node characteristic matrix X is 15 multiplied by 15; the convolution layer of the graph neural network is two layers, the size of the embedded vector Q is 15 multiplied by 128, and the dimension of the embedded vector of each node is 128 dimensions;
training the graph neural network to obtain embedded vector representations of all electromagnetic target class corresponding graph nodes, wherein the specific steps comprise:
s11, initializing a weight matrix W of the graph neural network i I=0, 1..n, N represents the number of layers of the graph neural network:
s12, calculating the output Y of the graph neural network, wherein the calculation formula is as follows:
wherein f () represents the computational function of the neural network of the graph;
s13, under the output condition, calculating a Loss function Loss of the graph neural network:
wherein L is l A label representing the first electromagnetic target class, F representing a row in the matrix F, i.e. a feature vector of a certain electromagnetic target class, Y f The output obtained after the characteristic vector f is input into the graph neural network; l represents a set of tags of electromagnetic target class;
s14, according to the loss function, updating the weight of the graph neural network by using a batch gradient descent method; the iteration number k is added with 1;
s15, repeating the processes from the step S12 to the step S14 until the iteration number k reaches a preset value;
s16, calculating an embedded vector representation Q of the electromagnetic target class corresponding graph node, wherein the calculation formula is as follows:
thus obtaining the embedded vector representation of the electromagnetic target class corresponding graph node.
2. The method for classifying an electromagnetic object using knowledge vector embedding as recited in claim 1,
the step S4 comprises the following specific steps:
s41, performing short-time Fourier transform and preprocessing on the acquired electromagnetic target signal to obtain a sample P 'to be identified, and inputting the sample P' to be identified into the convolutional neural network trained in the step S3;
s42, calculating to obtain an output vector R 'corresponding to the sample P' to be identified through a convolutional neural network;
s43, calculating the vector similarity of the average value of the embedded vector representation results of the output vector R 'and the sample data of each class, wherein the vector similarity of the output vector R' and the ith class is represented as epsilon i I=1,..n, n is the total number of classes, the vector similarity is obtained by calculating pearson correlation coefficients of two vectors;
s44, comparing the obtained vector similarity with a threshold value eta, if epsilon exists j More than or equal to eta, j epsilon n, judging that the sample P' to be identified belongs to the j-th electromagnetic target class; if to arbitrary epsilon j Has epsilon j And < eta, j epsilon n, judging that the sample P' to be identified belongs to a new electromagnetic target class.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210101982.8A CN114417938B (en) | 2022-01-27 | 2022-01-27 | Electromagnetic target classification method embedded by using knowledge vector |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210101982.8A CN114417938B (en) | 2022-01-27 | 2022-01-27 | Electromagnetic target classification method embedded by using knowledge vector |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114417938A CN114417938A (en) | 2022-04-29 |
| CN114417938B true CN114417938B (en) | 2023-06-27 |
Family
ID=81279571
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210101982.8A Active CN114417938B (en) | 2022-01-27 | 2022-01-27 | Electromagnetic target classification method embedded by using knowledge vector |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114417938B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116304885B (en) * | 2023-05-11 | 2023-08-22 | 之江实验室 | Event identification method, device and equipment based on graph node embedding |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109977232A (en) * | 2019-03-06 | 2019-07-05 | 中南大学 | A kind of figure neural network visual analysis method for leading figure based on power |
| CN110086737A (en) * | 2019-03-13 | 2019-08-02 | 西安电子科技大学 | A kind of recognition methods of the modulation mode of communication signal based on figure neural network |
| CN110210330A (en) * | 2019-05-13 | 2019-09-06 | 清华大学 | Electromagnetic signal recognition methods and device based on Tacit Knowledge structure figures convolutional network |
| CN110427877A (en) * | 2019-08-01 | 2019-11-08 | 大连海事大学 | A method of the human body three-dimensional posture estimation based on structural information |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11693924B2 (en) * | 2019-06-06 | 2023-07-04 | Hitachi, Ltd. | System and method for maintenance recommendation in industrial networks |
-
2022
- 2022-01-27 CN CN202210101982.8A patent/CN114417938B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109977232A (en) * | 2019-03-06 | 2019-07-05 | 中南大学 | A kind of figure neural network visual analysis method for leading figure based on power |
| CN110086737A (en) * | 2019-03-13 | 2019-08-02 | 西安电子科技大学 | A kind of recognition methods of the modulation mode of communication signal based on figure neural network |
| CN110210330A (en) * | 2019-05-13 | 2019-09-06 | 清华大学 | Electromagnetic signal recognition methods and device based on Tacit Knowledge structure figures convolutional network |
| CN110427877A (en) * | 2019-08-01 | 2019-11-08 | 大连海事大学 | A method of the human body three-dimensional posture estimation based on structural information |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114417938A (en) | 2022-04-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112906720B (en) | Multi-label image identification method based on graph attention network | |
| CN111723675B (en) | Remote sensing image scene classification method based on multiple similarity measurement deep learning | |
| CN111079847B (en) | Remote sensing image automatic labeling method based on deep learning | |
| CN109993100B (en) | Method for realizing facial expression recognition based on deep feature clustering | |
| CN110866530A (en) | Character image recognition method and device and electronic equipment | |
| CN113392931B (en) | Hyperspectral open set classification method based on self-supervision learning and multitask learning | |
| CN112685504B (en) | Production process-oriented distributed migration chart learning method | |
| CN112347970B (en) | Remote sensing image ground object identification method based on graph convolution neural network | |
| CN107169117B (en) | Hand-drawn human motion retrieval method based on automatic encoder and DTW | |
| CN111428511A (en) | Event detection method and device | |
| CN114417938B (en) | Electromagnetic target classification method embedded by using knowledge vector | |
| Defriani et al. | Recognition of regional traditional house in Indonesia using Convolutional Neural Network (CNN) method | |
| CN113962262B (en) | Continuous learning-based intelligent radar signal sorting method | |
| CN114492581A (en) | Method for classifying small sample pictures based on transfer learning and attention mechanism element learning application | |
| CN108960005B (en) | Method and system for establishing and displaying object visual label in intelligent visual Internet of things | |
| CN114037866B (en) | Generalized zero sample image classification method based on distinguishable pseudo-feature synthesis | |
| CN115510948A (en) | Block chain fishing detection method based on robust graph classification | |
| CN116012878A (en) | Pedestrian re-identification method and system | |
| CN112307914B (en) | Open domain image content identification method based on text information guidance | |
| CN113283519A (en) | Deep neural network approximate model analysis method based on discrete coefficients | |
| Tuya | Graph convolutional enhanced discriminative broad learning system for hyperspectral image classification | |
| Watanabe et al. | Understanding community structure in layered neural networks | |
| CN116310463B (en) | Remote sensing target classification method for unsupervised learning | |
| Miyamot | A rotation and scale robust texture recognition using moment features and neural network | |
| CN116186513A (en) | Vibration signal identification method based on one-dimensional convolutional neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |