A method for identifying a radiation source based on an integrated neural network and a signal graph

Abstract

The application discloses a radiation source identification method based on an integrated neural network and a signal graph, and comprises the following steps: extracting a signal sequence of radiation source data; constructing a signal graph of the signal sequence; inputting the signal sequence and the signal graph into the integrated neural network, extracting signal sequence features and graph features, and performing feature splicing to form integrated features; respectively placing the signal sequence features, the graph features and the integrated features into a sequence classifier, a graph classifier and an integrated classifier, and calculating a classification result of the radiation source signal; and performing loss value back propagation and updating module parameters. The application can effectively improve the radiation source identification performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radiation source signal identification, and in particular to a radiation source identification method based on an integrated neural network and a signal graph. BACKGROUND

[0002] In today's highly digital and information society, electromagnetic signal radiation has become an important technical means for information transmission, ranging, navigation, imaging and other applications. Today, national defense technology, industrial production and daily life cannot be separated from electromagnetic signal radiation technology. The increasingly developed electromagnetic signal radiation technology also brings more requirements and challenges to radiation source identification. Radiation source identification includes radiation source type identification, radiation source individual identification, electromagnetic signal modulation type identification, electromagnetic signal channel coding identification, etc. Radiation source identification is an important prerequisite for subsequent radiation source identity verification, signal demodulation and decoding, and plays a key role in electromagnetic spectrum application fields such as signal monitoring, spectrum management, reconnaissance and electronic countermeasures.

[0003] In the civil field, radiation source identification technology is mainly used for wireless device identity security verification and spectrum resource detection. Radiation source identification technology realizes the identity identification and authentication of the device by extracting the physical layer fingerprint (usually formed by the unintentional modulation of the radiation source) of the radio equipment, without signal interpretation and key comparison. Therefore, the radiation source identification technology can provide effective protection for wireless network security at the physical layer. As can be seen, the radiation source identification technology has broad application prospects and great research significance in radio station identification, radar equipment and civil communication device identity authentication, and wireless spectrum management.

[0004] The traditional radiation source identification method is limited by the effectiveness of the manual feature extraction method, and has poor applicability. Although the existing deep learning method can alleviate the shortcomings of the traditional method to a certain extent, the network structure and convolution operator are not specially designed, and the recognition accuracy and robustness of the model often cannot meet the application requirements. SUMMARY

[0005] In view of this, the present application provides a radiation source identification method based on an integrated neural network and a signal graph to solve the above technical problems.

[0006] The present application discloses a radiation source identification method based on an integrated neural network and a signal graph, comprising the following steps:

[0007] Step 1: extracting the signal sequence of the radiation source data;

[0008] Step 2: constructing a signal graph of the signal sequence;

[0009] Step 3: inputting the signal sequence and the signal graph into the integrated neural network, extracting the signal sequence features and the graph features, and performing feature splicing to form integrated features;

[0010] Step 4: Put the signal sequence feature, the graph feature and the integrated feature into the sequence classifier, the graph classifier and the integrated classifier respectively, and calculate the classification result of the radiation source signal;

[0011] Step 5: Reverse propagate the loss value, update the module parameters, and then re-execute steps 1 to 4.

[0012] Further, the step 2 comprises:

[0013] By adding adjacent edges between different sampling points of the signal sequence, a signal graph is constructed.

[0014] Further, the expression of the weight coefficient of the adjacent edge is:

[0015]

[0016] Wherein, k represents the weight coefficient of the adjacent edge, d represents the distance between the adjacent node and the center node in the signal sequence, and σ is the shape parameter; The smaller the distance d between the adjacent node and the center node, the greater the influence of the adjacent node value on the center node value, and vice versa.

[0017] Further, the set of weight coefficients is denoted as a weight matrix K, which has a dimension of length x length, and the relationship between the weight matrix K and its element k ij is expressed as:

[0018]

[0019] Wherein, the diag function is a diagonal matrix generator, the first parameter length controls the size of the generated diagonal matrix, the second element d controls the position of the diagonal line relative to the main diagonal line, and the third parameter k |d| controls the value of the diagonal element;

[0020] The weight matrix K and the adjacent matrix A are used to do Hadamard product ° operation, so as to give each adjacent edge a corresponding weight, so as to control the influence degree of different adjacent nodes in the updating process of the center node; The signal graph is represented by the weight matrix K.

[0021] Further, the step 3 comprises:

[0022] The signal sequence X S is input into the sequence feature extractor S to extract the sequence feature z S , and the expression is:

[0023] z S =S(X S )

[0024] The sequence feature extractor S is composed of three convolutional layers.

[0025] The signal graph X G is input into the graph feature extractor G to extract the graph feature z G , and the expression is as follows:

[0026] z G =G(X G )

[0027] The graph feature extractor G is composed of one graph convolutional layer as an input layer, followed by two convolutional layers.

[0028] The sequence feature and the graph feature are fused by using the feature fusion module in the form of feature splicing to obtain the integrated feature z E , and the expression is as follows:

[0029] z E =concat[z S ,z G ]

[0030] The overall structure of the integrated neural network integrates the sequence feature extractor S and the graph feature extractor G.

[0031] Further, the adjacency weight matrix of the signal graph X G is represented by K, and the Hadamard product ° operation is performed on the weight matrix K and the adjacency matrix A to give each adjacent edge a corresponding weight to control the influence degree of different adjacent nodes on the updating process of the center node.

[0032]

[0033] In the formula, H l+1 represents the output of the l+1-layer graph convolution, H l represents the output of the l-layer graph convolution, W l represents the parameter matrix of the l-layer graph convolution, and Gonv represents the graph convolution operation.

[0034] Further, the step 4 includes:

[0035] The overall structure of the integrated neural network includes the sequence classifier C S , the graph classifier C S , and the integrated classifier C E .

[0036] The sequence classifier C S only uses the sequence feature to realize the radiation source identification, and the radiation source identification result is p S , and the expression is as follows:

[0037] p S =C S (z S )=C S (S(X S ))

[0038] Figure classifier C S uses only figure features to implement radiation source identification, and the identification result is p G , and its expression is:

[0039] p G =C G (z G )=C G (S(X G ))

[0040] Integrated classifier C E uses integrated features z E to implement radiation source identification, and the identification result is p E , and its expression is:

[0041] p E =C E (z E )=C E (concat[z S ,z G ])

[0042] The integrated neural network has three outputs, respectively from the sequence classifier, the figure classifier and the integrated classifier.

[0043] Further, three sub-loss functions L S , L G and L E are set for training the sequence classifier, the figure classifier and the integrated classifier respectively, and the three sub-loss functions are cross-entropy loss functions, and their expressions are:

[0044]

[0045]

[0046] Wherein, N is the number of samples, k is the number of categories, y is the sample label, p is the model prediction result, θ S , θ G respectively represent the parameters of the sequence feature extractor and the figure feature extractor, respectively represent the parameters of the sequence classifier, the figure classifier and the integrated classifier.

[0047] Further, the training process of the integrated neural network is divided into two stages of local classifier training and global classifier training;

[0048] The first phase of training is local classifier training, at which time the loss function L step1 Simultaneously training the sequence classifier and the graph classifier;

[0049] The second phase of training is global classifier training, which begins after 90% of the training epochs have been completed, and the second phase uses the overall loss function L step2 Training the sequence classifier, the graph classifier, and the ensemble classifier, which is expressed as:

[0050]

[0051] Further, the loss value is back-propagated to update the module parameters, including:

[0052] The ensemble neural network first uses the sequence feature extractor and the graph feature extractor to extract sequence features and graph features, respectively, and fuses the sequence features and the graph features to obtain ensemble features;

[0053] In the training phase, the first step is to train the sequence classifier and the graph classifier, using the loss function L step1 to update the module parameters θ S , θ G , The second step is to add the output of the ensemble classifier to implement the overall training of the three classifiers of the overall network, using the loss function L step2 to update all the module parameters θ s , θ G , Adjust the network on the basis of the first step of training to obtain the optimal radiation source recognition capability; wherein the loss value is the value of L step1 and L step2 .

[0054] Due to the adoption of the above technical solutions, the present application has the following advantages: by designing a signal graph to convert signal sequences in the Euclidean space into a non-Euclidean space, the connection relationship between signal sequence sampling points is enriched, providing sample reserves for subsequent graph feature extraction, and then an ensemble neural network ENN is further designed to extract features from signal sequences and signal graphs. Subsequently, the ensemble features richer than single feature information are obtained by fusing sequence features and graph features. Finally, the ensemble neural network is trained using a two-step training method. Experiments on the ESP20 dataset show that, compared with traditional radiation source recognition methods and existing deep learning radiation source recognition methods, the present method can effectively improve the radiation source recognition performance. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art based on these drawings.

[0056] Figure 1 A flowchart of a radiation source identification method based on an integrated neural network and a signal graph according to an embodiment of the present application. DETAILED DESCRIPTION

[0057] The present application will be further described in conjunction with the drawings and embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. All other embodiments obtained by those skilled in the art should belong to the scope of protection of the embodiments of the present application.

[0058] The method provided by the present application is based on a graph convolutional neural network. By constructing a signal graph, a sequence format signal in a Euclidean space can be converted into a non-Euclidean space. Then, an improved graph convolutional operator is used to extract graph features in the non-Euclidean space, so that signal features are extracted from different spaces to fully extract feature information of a radiation source signal. The present application further designs an integrated neural network with a multi-feature extractor to extract sequence features in the Euclidean space and graph features in the non-Euclidean space, respectively.

[0059] The present application provides an embodiment of a radiation source identification method based on an integrated neural network and a signal graph, which comprises the following steps:

[0060] As shown in Figure 1 , the radiation source identification method provided by the present application comprises:

[0061] S101: extracting a signal sequence of radiation source data;

[0062] S102: constructing a signal graph of the signal sequence;

[0063] S103: inputting the signal sequence and the signal graph into an integrated neural network;

[0064] S104: extracting sequence features and graph features, and performing feature splicing to form integrated features;

[0065] S105: placing the sequence features, the graph features and the integrated features into a sequence classifier, a graph classifier and an integrated classifier, respectively, to calculate a classification result of the radiation source signal;

[0066] S106: loss value back propagation, and updating module parameters.

[0067] In S101 provided by the embodiment of the present application, the signal sequence of the radiation source data is extracted, and the specific process is as follows:

[0068] Different radiation source data signal sequences are extracted according to different radiation source identification purposes, and the radiation source type identification, the modulation mode identification and the coding mode identification can extract any part or all of the signal sequence for radiation source identification. Since the radiation source can use different identity codes to forge different radiation source individuals, the radiation source individual identification needs to extract the part of the signal sequence that does not contain the identity code. The radiation source signal can be directly input into the neural network or can be converted into other forms before being input into the neural network. The original signal without any processing is referred to as the signal sequence in the method. The signal sequence is a sequence with a length of N and a channel number of 2, and the two channels respectively represent the I channel (same direction branch) and Q channel (orthogonal branch) components of the radiation source signal. The signal sequence is data in the Euclidean space, and each element is arranged according to the relationship in the time sequence, and each element has a clear geometric position relationship.

[0069] In S102 provided by the embodiment of the present application, the signal graph of the signal sequence is constructed, and the specific process is as follows:

[0070] In order to fully represent the relationship between the non-continuous sampling points of the signal sequence by using a data structure, the method adds adjacent edges between different sampling points of the signal sequence to construct a signal graph, and then converts the signal sequence in the Euclidean space into a non-Euclidean space. The non-Euclidean features of the signal can be extracted from the signal graph by using the graph convolution. The non-Euclidean features can fully represent the relationship between the non-continuous sampling points of the signal sequence, and the convolutional neural network cannot extract these features in the Euclidean space, so the non-Euclidean features can provide incremental information for the radiation source identification.

[0071] In the signal graph, the influence of different adjacent nodes on the center node is different. The influence of the adjacent nodes on the center node is related to the position distance of the connection point in the signal sequence. Therefore, the standard graph convolution formula is improved in the algorithm. Each adjacent edge of the adjacent matrix is assigned a weight coefficient, and the purpose is to adjust the influence degree of different adjacent nodes in the summation updating process of the center node by the weight coefficient. The expression of the weight coefficient of the adjacent edge is as follows:

[0072]

[0073] wherein k represents the weight coefficient of the adjacent edge, d represents the distance between the adjacent node and the center node in the signal sequence, and σ is a shape parameter. The smaller the distance d between the adjacent node and the center node, the greater the influence of the adjacent node value on the center node value, and vice versa. The set of weight coefficients is written as a weight matrix K, and the shape of the weight matrix K is length x length. The weight matrix K and its element kij The expression of the relationship between them is as follows:

[0074]

[0075] The diag function of which is a diagonal matrix generator. The first parameter length controls the size of the generated diagonal matrix, the second element d controls the position of the diagonal line relative to the main diagonal, and the third parameter k |d| controls the value of the diagonal element. Finally, the Hadamard product ° operation is performed using the weight matrix K and the adjacency matrix A to assign a corresponding weight to each adjacent edge to control the influence degree of different adjacent nodes on the updating process of the center node. The signal graph is represented by the weight matrix K.

[0076] In S103 provided by the embodiment of the application, the signal sequence and the signal graph are input to the integrated neural network.

[0077] In S104 provided by the embodiment of the application, the sequence features and the graph features are extracted, and the integrated features are formed by feature splicing, and the specific process is as follows:

[0078] In order to fully use the sequence features and the graph features in two different spaces, the method designs an integrated neural network (ENN). The overall structure of the integrated neural network integrates two feature extractors, i.e., a sequence feature extractor and a graph feature extractor, and has three classifiers, i.e., a sequence classifier, a graph classifier and an integrated classifier. The sequence feature extractor S is composed of three convolutional layers, and the input thereof is the sequence signal X S , and the extracted features are sequence features z S , and the expression thereof is as follows:

[0079] z S =S(X S )

[0080] The graph feature extractor G is composed of one graph convolutional layer as an input layer and two convolutional layers connected after the input layer. The input thereof is the signal graph X G , and the extracted features are graph features z G , and the expression thereof is as follows:

[0081] z G =G(X G )

[0082] The adjacency weight matrix of the signal graph X G is represented by K, and the Hadamard product ° operation is performed using the weight matrix K and the adjacency matrix A to assign a corresponding weight to each adjacent edge to control the influence degree of different adjacent nodes on the updating process of the center node. Therefore, the expression of the improved graph convolution used for the signal graph features in the patent is as follows:

[0083]

[0084] In the formula, H l+1 represents the output of l+1 layer graph convolution, H l represents the output of l layer graph convolution, W l represents the parameter matrix of l layer graph convolution, and Gonv is a graph convolution operation; a plurality of layer graph convolution operators are stacked to form a graph convolution network (GCN), and a signal graph is subjected to the graph convolution network to obtain a graph feature.

[0085] A plurality of layer graph convolution operators are stacked to form a graph convolution network, and a signal graph is subjected to the graph convolution network to obtain a graph feature.

[0086] By integrating the sequence feature extractor and the graph feature extractor, the neural network extracts sequence features from the signal sequence in the Euclidean space and extracts graph features from the signal graph in the non-Euclidean space. Then, the sequence features and the graph features are fused by using a feature fusion module in a feature splicing manner to obtain integrated features z E , and the expression is as follows:

[0087] z E =concat[z S ,z G ]

[0088] The integrated features have both the feature information in the Euclidean space and the feature information in the non-Euclidean space, so the integrated features have more sufficient feature information than the sequence features and the graph features for the radiation source identification.

[0089] In the S105 provided by the embodiment of the application, the sequence features, the graph features and the integrated features are respectively put into a sequence classifier, a graph classifier and an integrated classifier, and the classification results of the radiation source signals are calculated, and the specific process is as follows:

[0090] The sequence classifier C S only uses the sequence features to realize the radiation source identification, and the radiation source identification result is p S , and the expression is as follows:

[0091] p S =C S (z S )=C s (S(X S ))

[0092] The graph classifier C G only uses the graph features to realize the radiation source identification, and the identification result is p G , and the expression is as follows:

[0093] p G =C G (zG ) = C G (S(X G ))

[0094] Integrated classifier C E Using integrated features z fused from sequence features and graph features E Radiation source identification is implemented, and the identification result is p E , the expression of which is as follows:

[0095] p E = C E (z E ) = C E (concat[z S ,z G ])

[0096] The integrated neural network has three outputs, respectively from the sequence classifier, the graph classifier and the integrated classifier. The method sets three sub-loss functions L S , L G and L E are respectively used for training the sequence classifier, the graph classifier and the integrated classifier. The three sub-loss functions are all cross-entropy loss functions, the expressions of which are as follows:

[0097]

[0098]

[0099] Where N is the number of samples, k is the number of categories, y is the sample label, p is the model prediction result, and θ S , θ G respectively represent the parameters of the sequence feature extractor and the graph feature extractor, respectively represent the parameters of the sequence classifier, the graph classifier and the integrated classifier.

[0100] The training process of the integrated neural network can be divided into two steps of local classifier training and global classifier training. The first stage of training is local classifier training, at this time the loss function L step1 The sequence classifier and the graph classifier are trained at the same time. The second stage of training is global classifier training, which starts after 90% of the total epoch training. The second stage uses the total loss function L step2 to train the sequence classifier, the graph classifier and the integrated classifier. The expression is as follows:

[0101]

[0102] In the S106 provided by the embodiment of the application, the loss value is back propagated to update the module parameters; the specific implementation process is as follows:

[0103] The integrated neural network first extracts sequence features and graph features using a sequence feature extractor and a graph feature extractor respectively, and fuses the above two features to obtain integrated features. In the training phase, in the first step, the sequence classifier and the graph classifier are trained first, and the loss function is L step1 , and the module parameter θ S is updated θ G , In the second step, the output of the integrated classifier is added to realize the whole training of the three classifiers of the network, and the loss function is L step2 , and the whole module parameter θ S is updated θ G , On the basis of the first step of training, the network is adjusted to obtain the optimal radiation source identification capability; wherein the loss value is the value of L step1 and L step2 .

[0104] In order to verify the beneficial effect of the radiation source identification algorithm based on the integrated neural network and the signal graph, the patent inventors carried out the experiment of the radiation source signal identification task on the ESP20 data set collected independently. In order to evaluate the algorithm performance, the accuracy evaluation index is used. The ESP20 data set is a wireless local area network WiFi radiation source signal data set provided by the patent inventors, and the data set contains signal data of 20 WiFi radiation sources with the same mode (EPS8266) and batch. The legacy long training field (LLTF) of the WiFi signal is extracted as the signal sequence for radiation source identification, and the dimension of each signal sequence is 160x2, 160 represents the number of sampling points of each branch, and 2 represents the in-phase branch and the quadrature branch.

[0105] In the experiment, the integrated neural network is trained using the real-time domain signal, and the above evaluation index is used to comprehensively evaluate the basic performance of the integrated neural network. In order to prove the superior performance of the algorithm in this paper, the proposed integrated neural network will be compared with two deep learning methods and two traditional machine learning methods. In addition, in the training phase of the network, the experiment of the algorithm is based on the PyTorch deep learning framework, the running environment is Windows 10, and the training is carried out on the NVIDIA GeForceGTX 950M GPU. The whole network is optimized using the ADAM optimizer, the initial learning rate is 0.001, and the iteration number is 50 epochs. The specific experimental results can be seen in Table 1.

[0106] Table 1 Comparison results of several algorithms

[0107]

[0108] As can be seen from Table 1, the integrated neural network ENN achieves the best result compared with its comparative methods. The comparative algorithms include convolutional neural network (CNN), graph convolution network (GCN), K-nearest neighbor algorithm (KNN), and decision tree algorithm (TREE). The graph convolution network GCN ranks second in each index and has better performance than the ordinary convolutional neural network CNN. Through the comparison between GCN and CNN, it can be concluded that the signal graph signal data representation method proposed in the method is effective. Although the KNN and TREE algorithms do not need to train the model with a large amount of data, the recognition rates are 79.91% and 69.9% respectively, which are much lower than the deep learning method. The radiation source recognition performance of ENN is better than that of CNN and GCN, which shows that the integrated neural network designed in this chapter can effectively extract signal fingerprint features and identify the identity of the radiation source.

[0109] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, but not to limit it, although the present application has been described in detail with reference to the above examples, those skilled in the art should understand that: the specific embodiments of the present application can still be modified or replaced by the equivalent, without departing from the spirit and scope of the present application. Any modification or equivalent replacement, which should be covered within the protection scope of the claims of the present application.

Claims

1. A method for emitter identification based on integrated neural networks and signal diagrams, characterized in that, The method comprises the following steps: Step 1: extracting a signal sequence of the radiation source data; Step 2: constructing a signal graph of the signal sequence; Step 3: inputting the signal sequence and the signal graph into an integrated neural network, extracting signal sequence features and graph features, and performing feature splicing to form integrated features; Step 4: placing the signal sequence features, the graph features and the integrated features into a sequence classifier, a graph classifier and an integrated classifier respectively, and calculating a classification result of the radiation source signal; Step 5: performing back propagation of a loss value, updating module parameters, and then re-executing steps 1 to 4; The step 2 comprises: A signal graph is constructed by adding adjacent edges between different sampling points of the signal sequence; An expression of a weight coefficient of the adjacent edge is: Wherein, k represents the weight coefficient of the adjacent edge, d represents a distance between an adjacent node and a center node in the signal sequence, and sigma is a shape parameter; the smaller the distance d between the adjacent node and the center node, the greater the influence of the adjacent node value on the center node value, and vice versa; The set of weight coefficients is denoted as a weight matrix K, which has dimensions length x length, and the weight matrix K has an element k ij The expression for the relationship between its elements k ij is: where the diag function is a diagonal matrix generator, the first parameter length controls the size of the generated diagonal matrix, the second element d controls the position of the diagonal line relative to the main diagonal, and the third parameter k |d| controls the value of the diagonal line elements. A Hadamard product ° operation is performed on the weight matrix K and the adjacent matrix A to give each adjacent edge a corresponding weight, so as to control the influence degree of different adjacent nodes on the updating process of the center node; the signal graph is represented by the weight matrix K.

2. The method of claim 1, wherein, The step 3 comprises: The signal sequence X S The input sequence feature extractor S extracts sequence features z S The expression is: z S = S(X S ) Wherein, the sequence feature extractor S is composed of 3 convolutional layers; The signal pattern X G The input pattern feature extractor G extracts the pattern feature z G The expression is z G = G(X G ) Wherein, the graph feature extractor G is composed of one graph convolutional layer as an input layer, and then two convolutional layers are connected; The sequence feature and the graph feature are fused by using a feature fusion module in a feature splicing manner to obtain integrated features z E The expression is: z E = concat[z S ,z G ] The overall structure of the integrated neural network integrates the sequence feature extractor S and the graph feature extractor G.

3. The method of claim 2, wherein, The signal graph X G The adjacency weight matrix of the signal graph is denoted as K, and the Hadamard product ° operation is performed using the weight matrix K and the adjacency matrix A to assign a corresponding weight to each adjacent edge to control the influence degree of different adjacent nodes on the updating process of the center node. The expression of the improved graph convolution for the signal graph features is: In the formula, H l+1 represents the output of l+1 layer graph convolution, H l represents the output of l layer graph convolution, W l represents the parameter matrix of l layer graph convolution, Gonv is the graph convolution operation; a plurality of layer graph convolution operators are stacked to form a graph convolution network, and a signal graph passes through the graph convolution network to obtain a graph feature.

4. The method of claim 2, wherein, The step 4 comprises: The overall structure of the ensemble neural network comprises a sequence classifier C S , a graph classifier C S , and an ensemble classifier C E ; Sequence classifier C S Radiation source identification is achieved using only sequence features, with the radiation source identification result being p S The expression is: p S =C S (z S )=C S (S(X S )) Figure classifier C S Radiation source identification is achieved using only figure features, with the result being p G The expression is: p G =C G (z G )=C G (S(X G )) The integrated classifier C E Using the integrated feature z E A radiation source identification is implemented, the result of which is p E The expression of which is: p E =C E (z E )=C E (concat[z S ,z G ]) The integrated neural network has three outputs, respectively from the sequence classifier, the graph classifier and the integrated classifier.

5. The method of claim 4, wherein, Three sub-loss functions L S , L G , and L E are set for training the sequence classifier, the graph classifier, and the ensemble classifier, respectively, and all of the three sub-loss functions are cross-entropy loss functions, and the expression thereof is: where N is the number of samples, k is the number of classes, y is the sample label, p is the model prediction result, and θ S G denote the parameters of the sequence feature extractor and the graph feature extractor, respectively, denote the parameters of the sequence classifier, the graph classifier, and the ensemble classifier, respectively.​ 6. The method of claim 5, wherein, The training process of the integrated neural network is divided into two stages of local classifier training and global classifier training; The first phase of training is local classifier training, at which point the loss function L step1 Simultaneously train the sequence classifier and the graph classifier. The second phase of training is global classifier training, which begins after 90% of the training epochs have been completed, and the second phase uses the overall loss function L step2 The three parts of training the sequence classifier, the graph classifier, and the ensemble classifier are expressed as:

7. The method of claim 6, wherein, The loss value back propagation and the module parameter updating comprise: The integrated neural network firstly extracts the sequence features and the graph features using the sequence feature extractor and the graph feature extractor respectively, and fuses the sequence features and the graph features to obtain the integrated features; In the training phase, the first step, train the sequence classifier and the graph classifier, with L step1 as the loss function, update the module parameters θ S , θ G , The second step, add the output of the integrated classifier to achieve the overall network three classifiers, with L step2 as the loss function, update all module parameters θ s , θ G , On the basis of the first step of training, adjust the network to obtain the optimal radiation source identification ability; wherein, the loss value is the value of L dtep1 and L step2 .

Images