Intra-pulse modulation recognition method based on time-frequency fusion map and radar map convolutional network

Abstract

The invention belongs to the technical field of radar radiation source signal intra-pulse analysis, and particularly relates to an intra-pulse modulation recognition method based on a time-frequency fusion graph and a radar graph convolutional network, and the method comprises the steps: constructing a new physically interpretable graph structured representation mode of a radar signal, and comprehensively utilizing the time-frequency information of the signal to form a time-frequency fusion graph; according to the method, the deep learning-based IPMR algorithm is adopted, and a special hierarchical residual image convolution network-radar map convolution network is specifically designed, so that the method can be directly generalized to a complex multi-path time-varying channel scene, and additional retraining is not needed, and therefore, a fundamental limitation of many existing deep learning-based IPMR methods can be broken through; according to the method, training can be carried out only under the additive white Gaussian noise channel, then time-varying multipath channels with different characteristics are directly adapted, priori knowledge of the channel and retraining on target domain data are not needed, and the method has the outstanding generalization advantage.

Description

Technical Field

[0001] This invention belongs to the field of intra-pulse analysis technology of radar radiation source signals, specifically relating to an intra-pulse modulation identification method based on time-frequency fusion graph and radar graph convolutional network. Background Technology

[0002] Intra-Pulse Modulation (IPM) in radar refers to the modulation of the carrier frequency or phase within the duration of a transmitted pulse to alter the signal's spectral characteristics, thereby achieving specific radar functions (primarily pulse compression). This is a core technology of modern high-performance radars (especially pulse compression radars). Intra-Pulse Modulation Recognition (IPMR) automatically determines the type of intra-pulse modulation used in unknown or complex electromagnetic environments by analyzing intercepted radar pulse signals.

[0003] With the continuous development of radar technology, new radar systems and applications are constantly being deployed, and radar signal waveform modulation systems are also continuously improving and perfecting. This has led to an increasing number of radar modulation types, more complex signal forms, and increasingly agile and diverse modulation parameters. Under these circumstances, signal identification is becoming increasingly difficult, especially the identification of intra-pulse modulation types in radar signals, which faces even greater challenges. Traditional modulation identification algorithms can only identify a limited number of signal types, making them unable to adapt to the increasingly diverse signal types and complex signal environments. Furthermore, they cannot meet the demand for robust and accurate identification under complex conditions such as low signal-to-noise ratios, interference signals, changes in transmission conditions, and variations in signal parameters caused by non-ideal factors in the transmission and reception stages.

[0004] Intra-pulse modulation (IPM) recognition technology for radar has evolved from traditional methods to deep learning paradigms. Traditional methods rely on expert knowledge for manual feature design (such as time-frequency analysis and higher-order statistics), which, while highly interpretable, suffers from high feature extraction complexity and limited generalization ability, especially exhibiting significant performance degradation in low signal-to-noise ratio (SNR) environments. Deep learning technology overcomes the limitations of manual design by automatically extracting signal features through end-to-end learning. However, most deep learning-based IPM recognition methods directly apply mature models from other fields to the problem of IPM type recognition. Because they cannot achieve high adaptability to different scenarios or data, they suffer from significant performance bottlenecks. Furthermore, current technologies have limited recognition performance in practical applications, particularly in terms of insufficient generalization ability and robustness.

[0005] This is because deep learning-based automatic modulation recognition technology faces the problem of distribution offset between training conditions and actual applications (such as changes in modulation parameters / channel environment), which limits its generalization ability and leads to significant performance degradation. Most existing research only considers radar signal recognition under the influence of Gaussian white noise, with relatively little research on the multipath fading effect that causes signal distortion. This results in a sharp deterioration in performance under low signal-to-noise ratio and time-varying multipath fading scenarios—specifically manifested as amplitude fluctuations, phase distortion, and channel-specific noise-fading characteristics. These factors collectively weaken the system's robustness (Xuan Zhu, Jiangjun Ruan, Hao Wu, Jiahao Zhang, and Jin Meng, “Radar intra-pulse modulation fast perception via adaptive random magnitude channelpruning and knowledge distillation,” IEEE Signal Processing Letters, vol. 32, pp. 2898–2902, 2025.). However, while some existing studies have acknowledged the impact of channel characteristics, their experiments have only considered the same Rician channel with fixed multipath and fading conditions for training and identification. They lack specific robustness designs tailored to channel effects and fail to consider adaptability to complex real-world scenarios, such as channel variations, especially the encounter with entirely new, untrained channels. These channel characteristics cannot be known in advance and cannot be specifically pre-trained. Therefore, algorithms need to be able to adapt to random channel conditions with unknown statistical properties. Summary of the Invention

[0006] Due to the rapid changes in radar waveforms, the emergence of complex composite modulation techniques, and the increasingly harsh electromagnetic environment, existing deep learning-based intra-pulse modulation (IPMR) identification technologies still face significant challenges in practical applications, typically manifesting as strong noise and multipath fading. These factors collectively limit the reliability of identification, especially under low signal-to-noise ratio (SNR) and multipath fading channel conditions, where the robustness of existing methods is clearly insufficient. To address this, this paper proposes an IPMR identification method based on a time-frequency fusion graph and a radar graph convolutional network. This method constructs a novel, physically interpretable graph-structured representation of radar signals, comprehensively utilizes the time-frequency information of the signal to form a time-frequency fusion graph, and specifically designs a hierarchical residual graph convolutional network—the radar graph convolutional network—which can directly generalize to complex multipath time-varying channel scenarios without additional retraining, thus overcoming a fundamental limitation of many existing deep learning-based IPMR methods.

[0007] The specific design concept is as follows:

[0008] (1) For the first time, a time-frequency fusion graph structure representation method with clear physical meaning was constructed for radar signals by utilizing the inherent structural features of different modulation types and their time-frequency coupling characteristics. Using signal segments as nodes, dynamically and adaptively setting time-domain characteristics as node attributes, and using spectral frequency domain relationships as edge weights, it helps to maintain the fusion of time-domain and frequency-domain information of the signal and enhances robustness to noise and channel distortion.

[0009] (2) Design a dedicated radar image convolutional network. The main body of the network is a deep graph convolutional neural network, which adopts a dual-path parallel processing structure to achieve collaborative learning of in-phase and orthogonal components of the signal. At the same time, by introducing residual connections and multi-layer differentiable pooling mechanism, the network achieves deep feature extraction and structural compression while maintaining representation ability, thus taking into account computational efficiency.

[0010] The technical solution adopted in this invention is as follows: an intra-pulse modulation recognition method based on time-frequency fusion map and radar map convolutional network. First, the radar map convolutional network model is built, and then data processing is performed, including receiving signal data, constructing time-frequency fusion map data, inputting time-frequency fusion data into the radar map convolutional network model to calculate features, followed by classification recognition, calculating classification loss, training and saving network parameters, and finally, modulation type recognition on new unlabeled data.

[0011] The specific steps of this method are as follows:

[0012] S1 Establish a radar chart convolutional network model

[0013] The radar image convolutional network has a dual-channel structure, consisting of an I-channel and a Q-channel. The two channels are identical in structure, independent, and share parameters, processing the real and imaginary parts of complex signal data respectively. Each channel is composed of three ResGCN units (ResGCN1, ResGCN2, and ResGCN3) stacked in series. Each ResGCN unit includes a ResBlock residual block and a pooling layer. The ResBlock residual block consists of six uniGNN units, a merging operation layer, and a linear layer. The six uniGNN units are connected end-to-end, except for the last uniGNN unit. Each uniGNN unit has two outputs: one directly serves as the input to the next uniGNN, and the other outputs to the merging layer. The last uniGNN unit has only one output to the merging layer. The outputs of the six uniGNN units are merged by the merging layer and then passed through a linear layer, ensuring that the dimension of the output features is the same as the dimension of the original input node attributes. All uniGNN units have the same structure, consisting of a residual structure DenseSAGEConv graph convolutional layer, a ReLU activation function, and a LayerNorm normalization layer connected in series. The size of the DenseSAGEConv graph convolutional layer is 64*64. The pooling layer is DiffPool, with a compression ratio of 0.25. Using a hierarchical pooling DiffPool structure helps to preserve the structural information of the data while compressing the representation, thus facilitating feature mining and subsequent modulation type identification.

[0014] The outputs of the two channels are merged by a merging operation layer and then input into the classifier Cls.

[0015] The classifier Cls consists of a fully connected layer FC1, a ReLU activation function, a fully connected layer FC2, and a softmax layer connected in series. The fully connected layer FC1 has a dimension of 128*64, and the fully connected layer FC2 has a dimension of 64*M, where M represents the total number of modulation type categories.

[0016] Network model weight parameters This indicates that the network model weight parameters are initialized using a random setting method. ;

[0017] S2 Data Reception

[0018] The received signal undergoes front-end processing such as filtering, down-conversion, and analog-to-digital conversion to obtain the discrete signal to be processed. Where n represents the nth sampling point, , This represents the total number of sampling points. The discrete signal to be processed. It is a complex signal with real and imaginary parts. If the total number of signal samples is... Then the first A signal sample can be found Based on this, add subscripts to distinguish it, and represent it as , The corresponding modulation type is tagged with express, That is, shared Types of modulation.

[0019] S3 Time-Frequency Fusion Diagram Construction

[0020] A graph is a data format composed of nodes and edges, primarily described by node attributes and an adjacency matrix. Therefore, the construction method of the time-frequency fusion graph in this invention can be divided into the calculation of node attributes and the calculation of the adjacency matrix, and then the two parts are combined to form the time-frequency fusion graph. The basic idea of ​​constructing the time-frequency fusion graph in this invention is to draw on the Hankel matrix construction form in system identification theory to construct a node attribute matrix, with each row of the matrix representing a node. The values ​​of the matrix serve as node attributes, and the complex cosine similarity of the spectra between different nodes is used as edge weights to form an adjacency matrix. The node attribute matrix and the adjacency matrix are combined to form a complete graph. Because node attributes are time-domain information and the adjacency matrix contains frequency-domain information, this graph data format is called a time-frequency fusion graph. The specific steps are as follows:

[0021] S3-1 Node Attribute Calculation

[0022] Based on the construction form of the Hankel matrix in system identification theory (John Wright and Yi Ma, High-Dimensional Data Analysis with Low-Dimensional Models: Principles, Computation, and Applications, Cambridge University Press.), the discrete signal to be processed is... Constructing a node feature matrix :

[0023] ,

[0024] in The dimension of a matrix can be understood as the length of the subsequence corresponding to each row. express yes A complex matrix of dimension . The number of such subsequences is equal to the number of rows in the matrix, satisfying... Normally take The feature matrix Each row is used as a node. For the first row... Each node has a feature vector. Depend on It consists of a series of complex-valued signal samples:

[0025] ,

[0026] Observation Matrix It can be seen that each row vector It can be viewed as a signal slice of equal length (m segments) with different starting points, serving as nodes in the time-frequency fusion graph, and a continuous row vector. and Offset by one sampling point between them;

[0027] S3-2 Adjacency Matrix Calculation

[0028] The basic idea behind constructing an adjacency matrix is ​​to use the complex cosine similarity of the spectral frequencies between different nodes as edge weights to form the adjacency matrix. The specific calculation steps are as follows:

[0029] S3-2-1 Along the Node Feature Matrix Perform a Fourier transform on each row to obtain the corresponding spectrum matrix. :

[0030] ,

[0031] in, Represents a matrix Perform Fourier transform along the row direction. Indicates to The result of performing a discrete Fourier transform along the time dimension. , Representation matrix The Middle Line 1 Column elements;

[0032] The number of points in the Discrete Fourier Transform is equal to the length of the signal window. , that is, the vector length;

[0033] S3-2-2 Construct the Gram matrix according to the following formula :

[0034] ,

[0035] in, Represents the conjugate transpose, matrix The ) elements It is given by the following formula:

[0036] ,

[0037] express Conjugate;

[0038] S3-2-3 Calculation of Norm (modulus):

[0039] ,

[0040] in, and These represent taking the real part and the imaginary part, respectively, and the same applies below.

[0041] S3-2-4 Calculate the sum of all vectors Norm, to obtain the norm vector :

[0042] ,

[0043] Represents a norm vector Elements in;

[0044] S3-2-5 Calculate the norm vector outer product matrix :

[0045] ,

[0046] S3-2-6 Calculate the adjacency matrix :

[0047] ,

[0048] Where ε is the numerical stability constant, which can usually be taken as ε = 10. −8 .

[0049] Adjacency Matrix elements for:

[0050] ;

[0051] S3-3 Time-Frequency Fusion Diagram Construction

[0052] S3-3-1 Adjacency Matrix With node feature matrix Together they form a complete time-frequency fusion diagram : ;

[0053] S3-3-2 Take The real part: ;

[0054] S3-3-3 Take The imaginary part: ;

[0055] S4 Feature Calculation

[0056] S4-1 Channel I Feature Calculation

[0057] The result of step S3-3-2 The input I channel is passed sequentially through three ResGCN units in the I channel, and the final output is the embedded feature of the I channel. :

[0058] S4-1-1 In the ResGCN1 unit, First, the output node attribute matrix intermediate values ​​are processed through the ResBlock residual block. ;

[0059] S4-1-2 will and As a new graph structure data The input is fed into the DiffPool pooling layer, which compresses the number of nodes and outputs the updated graph. ;

[0060] S4-1-3 Result of S4-1-2 The input is fed into the ResGCN2 cell, first passing through the ResBlock residual block, and then outputting the intermediate values ​​of the node attribute matrix. ;

[0061] S4.1.4 will and As a new graph structure data The input is fed into the DiffPool pooling layer, which compresses the number of nodes and outputs the updated graph. ;

[0062] S4-1-5 Result of S4.1.4 The input is fed into the ResGCN3 cell, first passing through the ResBlock residual block, and then outputting the intermediate values ​​of the node attribute matrix. ;

[0063] S4.1.6 will and As a new graph structure data The input is fed into the DiffPool pooling layer, which compresses the number of nodes and outputs the updated graph. , where node attributes As an embedding feature of the I channel, using express.

[0064] S4-2 Q Channel Feature Calculation

[0065] The result of step S3.3.3 The input Q-channel is computed, and the data is sequentially passed through three ResGCN units in the Q-channel to finally output the Q-channel embedded features. For specific processing ideas, please refer to the calculation process of the I channel features in S4.1.

[0066] S4-3 Dual-channel feature calculation

[0067] Embedded features of the I channel Q-channel embedding features The combined features are obtained by merging the two channels. :

[0068] ;

[0069] S5 Classification and Recognition

[0070] Dual-channel features The input is fed into the classifier Cls, which outputs the predicted label for radar signal modulation recognition. ;

[0071] S6 Calculate classification loss

[0072] Predictive tags identified by modulating radar signals output from the S5 With the actual modulation type label in S2 Calculate the classification loss corresponding to the training process. : ;

[0073] S7 Network Training and Parameter Saving

[0074] S7-1 Network Training

[0075] S7-1-1 The Adam algorithm (Kingma DP, Ba J. Adam: A method for stochastic optimization[J]. arXiv preprint arXiv:1412.6980, 2014.) is used to perform gradient descent on the radar chart convolutional network model, and the network model weight parameters are updated in real time. This leads to training loss. Continuing to decline;

[0076] S7-1-2 sets the initial learning rate to 1e-4. After each epoch, the learning rate is updated using a cosine annealing algorithm with restart (Loshchilov I, Hutter F. SGDR: Stochastic gradient descent with warmrestarts[J]. arXiv preprint arXiv:1608.03983, 2016.). The restart period is set to 2.

[0077] S7-1-3 until total training loss When the number of iterations stops decreasing, the network is in a convergent state, i.e., the network has reached a steady state, and training should be stopped.

[0078] S7-2 Network Parameter Saving

[0079] Save the optimal network model weight parameters when the network reaches steady state. ;

[0080] S8 identifies modulation type label

[0081] The optimal network model weight parameters are applied to the trained radar chart convolutional network model to identify the modulation of signals within a pulse, thereby enabling the identification of new signal modulation types.

[0082] The present invention has the following technical effects:

[0083] 1. A dynamic mapping paradigm from signal to time-frequency fusion graph structure is proposed to realize the adaptive structured transformation from the original complex signal to the topological graph structure data. The time domain representation is used as the node attribute and the frequency domain relationship is used as the adjacency matrix of the edge. The fusion of time and frequency domain information at the graph level is realized. It has inherent robustness to noise and multipath fading, and provides a new structured representation dimension for radar signal analysis.

[0084] 2. Design a dedicated graph convolutional network for the complex domain, adopt a dual-path architecture to support the collaborative learning of real and imaginary parts, and combine residual connections and differentiable pooling mechanisms to achieve efficient model compression while maintaining training stability;

[0085] 3. This invention has good recognition performance and good robustness to low signal-to-noise ratio and multipath channels.

[0086] It is worth emphasizing that this method can be trained only in an additive white Gaussian noise channel and then directly adapted to time-varying multipath channels with different characteristics. It does not require prior knowledge of the channel or retraining on the target domain data, and has outstanding generalization advantages. Attached Figure Description

[0087] Figure 1 is a flowchart of the implementation of the present invention;

[0088] Figure 2 This is the time-frequency fusion graph data construction process (S3);

[0089] Figure 3 It is the feature calculation and classification recognition process (S4-S5);

[0090] Figure 4 The structure of the ResGCN unit is shown in: (a) schematic diagram and (b) detailed diagram.

[0091] Figure 5 It is the structural composition of ResBlock residual blocks;

[0092] Figure 6 It is the structural composition of uniGNN small units;

[0093] Figure 7 It is the structural composition of the classifier Cls;

[0094] Figure 8 This is the performance curve of the correct recognition rate of the present invention as a function of SNR under AWGN channel;

[0095] Figure 9 This is the performance curve of the correct recognition rate as a function of SNR under a random multipath fading channel with unknown characteristics.

[0096] Figure 10 This is the embedded feature of the present invention. Distribution map after feature reduction;

[0097] Figure 11 This is the confusion matrix of the identification results of the method of the present invention. Detailed Implementation

[0098] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0099] Figure 1 This is a flowchart of the implementation of the present invention. The present invention proposes an intra-pulse modulation recognition method based on a time-frequency fusion graph and a radar graph convolutional network, which consists of the following steps:

[0100] S1. Establish a radar chart convolutional network model;

[0101] S2 data reception;

[0102] S3 time-frequency fusion graph data construction;

[0103] S4 feature calculation;

[0104] S5 classification and recognition;

[0105] S6 Calculates the classification loss;

[0106] S7 network training and parameter saving;

[0107] S8 identifies the modulation type label.

[0108] Figure 2 This is the time-frequency fusion graph data construction process (S3). The basic process involves constructing a Hankel matrix from the received signal, using the row vectors of each row as nodes, and the complex values ​​of the row vectors as the attribute matrix of the nodes. Then, a Fourier transform is performed on the corresponding nodes in each row to obtain the corresponding spectral features. The complex cosine similarity between all pairs of nodes is calculated as the edge attribute relationship, forming an adjacency matrix. Finally, the attribute matrix of the nodes and the adjacency matrix constitute the complete time-frequency coupled graph data. This is the new paradigm for converting signals into graph structure data proposed in this invention.

[0109] Figure 3 This is the feature calculation and classification process (S4-S5). As you can see, the real and imaginary parts of the obtained complete image data are extracted and input into the I and Q channels respectively, processed independently. Then, the features output from the two channels are merged to obtain the final embedded features, which are finally modulated and labeled by the classifier. It is important to note that the processing of the two channels is independent, but parameters are shared.

[0110] Figure 4 It is the composition structure of ResGCN units, in which Figure 4 (a) is a schematic diagram. Each ResGCN unit has the same structure, consisting of a ResBlock block and a pooling layer. Figure 4 (b) is a detailed view, which shows that... Figure 4 and Figure 5 The complete composition details of a single ResGNN unit after combination.

[0111] The structure of the ResBlock block is as follows: Figure 5As shown, the pooling layer is DiffPool. On the left, there are six identical repeating uniGNN units, distinguished by shading. Each uniGNN unit consists of a DenseSAGEConv layer, a ReLU layer, and a LayerNorm layer. These six uniGNN units are connected end-to-end, and the adjacency matrix remains unchanged during computation, while the node attributes are continuously updated. Except for the last uniGNN unit, each other uniGNN unit has two outputs: one directly serves as the input to the next uniGNN, and the other is input to the merging operation layer. The last uniGNN unit, however, has only one output to the merging operation layer. The outputs of the six different uniGNN units are merged and passed through a linear layer, making the dimension of the output features the same as the dimension of the original input node attributes. That is, the rightmost linear layer performs a dimensionality transformation, reducing it to a matrix of the same size as the output node attributes of a single uniGNN unit.

[0112] Figure 6 This is the structure of uniGNN. uniGNN uses a residual structure, consisting of a convolutional layer (DenseSAGEConv), an activation layer (ReLU), and a normalization layer (LayerNorm). Specifically, the adjacency matrix and node attributes are input into the convolutional layer (DenseSAGEConv), which outputs new node attributes. These attributes then pass through the activation layer (ReLU) and are added to the original input node attributes. Finally, a normalization layer outputs the new node attributes.

[0113] Figure 7 This is the structure of the classifier Cls. Cls consists of a fully connected layer FC1, a ReLU activation function, a fully connected layer FC2, and a softmax layer. The first fully connected layer FC1 has a dimension of 128*64, and the second fully connected layer FC2 has a dimension of 64*M, where M represents the total number of modulation type categories.

[0114] Figure 8 This is the performance curve of the correct recognition rate of this invention under an AWGN channel as a function of SNR. Eighteen LPI modulation types were selected for experiments, including CP, LFM, NLFM, QPSK, BPSK, Costas, Frank, P1, P2, P3, P4, T1, T2, T3, T4, and composite modulation types LFM_BPSK, FSK_BPSK, and FSK_LFM. The simulated signal parameters were all random values, with random additive white Gaussian noise (AWGN) added, and a signal-to-noise ratio ranging from -12dB to 18dB. Model training was performed according to steps S1-S7 to obtain the parameters. Then, the new signal passing through the AWGN channel is identified, and the identification result is as follows: Figure 9As shown, the method of this invention achieves a correct recognition rate of 96.3% for 18 modulation types at 0dB and 98.89% at 6dB, which fully demonstrates its effectiveness. Furthermore, it still achieves approximately 90% at -6dB, proving its robustness to noise.

[0115] Figure 9 This is the performance curve of the correct identification rate as a function of SNR under a random multipath fading channel with unknown characteristics. It is derived from... Figure 8 The model trained on AWGN The proposed method was directly tested on a multipath channel to examine its generalization ability on novel and unknown complex channels. To simulate complex time-varying channel scenarios, all signals passed through randomly generated channels with a multipath number ranging from 1 to 5, and the channel attenuation value was also randomly generated, meaning that the channel conditions for each signal were different. As can be seen, the proposed method, without training, maintains an 88.4% correct recognition rate at 0dB on unknown multipath channels and 90.6% at 6dB, demonstrating its robustness and generalization ability to unknown signals.

[0116] Figure 10 This is the embedded feature of the present invention. Distribution plot after feature dimensionality reduction. The embedded features corresponding to the 18 modulation types are displayed after feature dimensionality reduction, with different colors representing different modulation types. It can be seen that signals of different types can be well distinguished, while signals of the same type can be well clustered together, indicating that the proposed method can extract effective features and accurately characterize the differences between different modulation schemes.

[0117] Figure 11 This is the confusion matrix of the identification results of the method of this invention. The figure shows the confusion matrix of all test signal samples under the AWGN channel for 18 modulation schemes from -12dB to 10dB. It can be seen that, except for LFM_BPSK, the correct identification rate of all modulation types is greater than 83%, which proves the effectiveness of the method and also shows the complexity of composite modulation type identification.

Claims

1. A method for intra-pulse modulation recognition based on time-frequency fusion graph and radar chart convolutional network, characterized in that, The specific steps of the method are as follows: S1, a radar chart convolutional network model is established The radar chart convolutional network is a double-channel structure, which is I channel and Q channel respectively, the two channel structures are completely same, independent of each other, parameter sharing, and are respectively used for processing the real part and the imaginary part of the complex signal data: each channel is composed of three ResGCN units ResGCN1, ResGCN2 and ResGCN3 which are stacked in series, and each ResGCN unit includes a ResBlock residual block and a pooling layer; the ResBlock residual block is composed of six uniGNN small units, a merging operation layer and a linear layer, the six uniGNN small units are connected in a head-to-tail manner, except for the last uniGNN unit, the other five uniGNN units have two outputs: one is directly used as the input of the next uniGNN, and the other is output to the merging operation layer, and the last uniGNN unit only has one output to the merging operation layer; The output of the six uniGNN units is merged through the merging operation layer and then passed through a linear layer, so that the dimension of the output feature is the same as that of the original input node attribute; All uniGNN small units have the same structure, which is composed of a DenseSAGEConv graph convolution layer with a residual structure, a ReLU activation function and a LayerNorm normalization layer in series, wherein the size of the DenseSAGEConv graph convolution layer is 64*64; the pooling layer is DiffPool, and the compression ratio is 0.25; The outputs of the two channels are merged through the merging operation layer and then input into the classifier Cls; The classifier Cls is composed of a fully connected layer FC1, a ReLu activation function, a fully connected layer FC2 and a softmax layer in series, the dimension of the fully connected layer FC1 is 128*64, and the dimension of the fully connected layer FC2 is 64*M, M represents the total number of modulation type categories; The network model weight parameters are initialized in a random setting manner represented, the network model weight parameters are initialized in a random setting manner ; S2, data receiving The received signal undergoes front-end processing such as filtering, down-conversion, and analog-to-digital conversion to obtain the discrete signal to be processed. Where n represents the nth sampling point, , Total number of sampling points; discrete signal to be processed It is a complex signal with real and imaginary parts; if the total number of signal samples is Then the first A signal sample can be found Based on this, add subscripts to distinguish it, and represent it as , The corresponding modulation type is tagged with express, That is, shared Types of modulation; S3, time-frequency fusion graph construction The specific steps are as follows: S3-1, node attribute calculation According to the construction form of Hankel matrix in system identification theory, the discrete signal to be processed is constructed into a node feature matrix : ， wherein denotes the dimension of the matrix, which can be understood as the length of each row corresponding to the length of the subsequence, denotes is a complex matrix of dimension ; the number of such sub-sequences is the number of rows of the matrix, satisfying , usually taken as ; each row of the feature matrix is taken as a node; for the th node, eigenvectors comprises of consecutive complex-valued signal samples: ， Observation matrix It can be seen that each row vector It can be seen that each row vector and is offset by one sample point from the previous one. S3-2, adjacency matrix calculation The basic idea of constructing the adjacency matrix is to take the spectral complex cosine similarity between different nodes as the edge weight to form the adjacency matrix; the specific calculation steps are as follows: S3-2-1 Fourier transform is performed along each row of the node feature matrix to obtain a corresponding frequency spectrum matrix : ， wherein, denotes the Fourier transform of the matrix a Fourier transform is performed in the row direction, denotes the Fourier transform of the result of a discrete Fourier transform along the time dimension, , denotes the matrix the element in the i-th row and the j-th column of the matrix the element in the i-th row and the j-th column of the matrix the element in the i-th row and the j-th column of the matrix The number of points of the discrete Fourier transform is the signal window length i.e. the vector length; S3-2-2 The Gram matrix is constructed according to the following formula : ， wherein denotes the conjugate transpose, the matrix has its ) element is given by ， representing conjugation of S3-2-3 Calculation of Norm: ， wherein and denote taking the real and imaginary parts, respectively; S3-2-4 Compute norm of all vectors , resulting in norm vector : ， denotes the norm vector of the elements in ; S3-2-5 Compute norm vector of outer product matrix : ， S3-2-6 Computing an adjacency matrix : ， Wherein ε is a numerical stability constant; The element of the adjacency matrix is: is: ； S3-3, time-frequency fusion graph construction S3-3-1 adjacency matrix with node feature matrix together constitute a complete time-frequency fusion graph : ; S3-3-2 Take the real part of: ; S3-3-3 Take the imaginary part of: ; S4, feature calculation S4-1, I channel feature calculation The results of step S3-3-2 are input into the I channel, sequentially pass through the three ResGCN units of the I channel, and finally output the I channel embedding features The input I channel, sequentially pass through the three ResGCN units of the I channel, and finally output the I channel embedding features : S4-1-1 In the ResGCN1 unit, the following is performed First, the output node attribute matrix intermediate value is passed through a ResBlock residual block ; S4-1-2 will and as new graph structure data into the DiffPool pooling layer, compress the number of nodes, output the updated graph ; S4-1-3 results of S4-1-2 Input into the ResGCN2 unit, first through the ResBlock residual block, output the node attribute matrix intermediate value ; S4.1.4 will be and as new graph structure data into the DiffPool pooling layer, compressing the number of nodes, outputting the updated graph ; S4-1-5 Results from S4.1.4 Input into the ResGCN3 unit, first through the ResBlock residual block, output the node attribute matrix intermediate value ; S4.1.6 will be and as new graph structure data into the DiffPool pooling layer, compressing the number of nodes, outputting the updated graph where the node attribute is represented as the embedding feature of the I channel, denoted as ; S4-2, Q channel feature calculation The results of step S3.3.3 are input into the Q channel calculation, sequentially pass through the three ResGCN units of the Q channel, and finally output the Q channel embedding features ; and ; S4-3, double-channel feature calculation Embedding features for channel I Embedding features for channel Q Merging to get dual channel features : ； S5, classification and identification The two-channel features are input into a classifier Cls, which outputs a predicted label of radar signal modulation recognition ; S6, classification loss calculation Predicted labels of radar signal modulation recognition using s5 output True modulation type labels in s2 Calculate classification loss corresponding to training process : ; S7, network training and parameter saving S7-1, network training S7-1-1 performs gradient descent on the radar chart convolutional network model and updates the network model weight parameters in real time , so that the training loss continuously decreases; S7-1-2, learning rate updating is performed; S7-1-3 until total training loss When the decrease is no longer, at this time the network is in a state of convergence, that is, the network reaches a steady state, stop training; S7-2, network parameter saving Saving the optimal network model weight parameters of the network at steady state ; S8, identification of modulation type label The optimal network model weight parameter is applied to the trained radar chart convolutional network model to identify the modulation of the intrapulse signal, so as to realize the identification of the modulation type of the new signal.

2. The method of claim 1, wherein the method further comprises: In S3-2-6, take ε = 10 −8 .

3. The method of claim 1, wherein the method further comprises: In S7-1-1, the Adam algorithm is used to perform gradient descent on the radar chart convolutional network model.

4. The method of claim 1, wherein the method further comprises: In S7-1-2, the initial learning rate is set to 1e-4.

5. The method of claim 1, wherein: The cosine annealing algorithm with restart is used to update the learning rate, and the restart period is set to 2.

Images