Modulation identification method based on signal amplitude-phase diagram representation and multi-channel graph convolutional network

Abstract

The invention belongs to the technical field of radar radiation source signal intra-pulse analysis, and particularly relates to a modulation identification method based on signal amplitude-phase diagram representation and a multichannel graph convolutional network, which comprises the following steps: firstly, building a multichannel graph convolutional network model, and then carrying out data processing; comprising the following steps: receiving signal data, constructing signal amplitude and phase diagram data, inputting the amplitude and phase diagram data into a multi-channel diagram convolutional network model to calculate features, then performing classification identification, calculating classification loss, training and storing network parameters, and finally performing modulation type identification on new label-free data. 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 radar radiation source signal pulse analysis technology, specifically relating to a modulation recognition method based on signal amplitude-phase diagram representation and multi-channel graph convolutional network. Background Technology

[0002] Intra-Pulse Modulation Recognition (IPMR) automatically determines the type of intra-pulse modulation used by intercepted radar pulse signals in unknown or complex electromagnetic environments by analyzing the signals. Traditional IPMR recognition methods can be divided into two categories: likelihood estimation-based and feature extraction-based (OA Dobre, A. Abdi, Y. Bar-Ness, and W. Su, “Survey of automatic modulation classification techniques: Classical approaches and new trends,” IET Commun., vol. 1, no. 2, pp. 137–156, Apr. 2007.). Likelihood estimation-based methods model the identification problem as multiple hypothesis testing. While theoretically, they can obtain Bayesian optimal solutions, their computational complexity increases dramatically with modulation order, limiting their practical applications (JL Xu, W. Su, and M. Zhou, “Likelihood-ratio approaches to automatic modulation classification,” IEEE Trans. Syst., Man, Cybern. C, Appl. Rev., vol. 41, no. 4, pp. 455–469, Jul. 2011.). Feature extraction-based methods extract features from the signal's time domain, frequency domain, higher-order cumulants, or cyclostationarity, and then combine them with classifiers such as decision trees and support vector machines to achieve identification, striking a balance between complexity and performance. However, this process relies on expert experience to manually design features, and its discriminative ability decreases under complex channels such as low signal-to-noise ratio and multipath fading, and its identification effect on highly similar modulation types is poor.

[0003] In recent years, the rise of deep learning has brought revolutionary changes to intra-pulse modulation recognition. Its core is to use the powerful automatic feature extraction and pattern recognition capabilities of deep neural networks to learn the optimal discriminative representation directly from the original signal data (I / Q data, time-frequency graphs, etc.), which is currently the mainstream method.

[0004] A typical deep learning modulation recognition process consists of two core stages: first, the received signal is preprocessed to convert it into a representation suitable for neural network processing; then, the deep learning neural network is used to automatically extract features and classify the modulation mode.

[0005] Based on the different forms of input signal representation, existing deep learning-based IPMR algorithms can be divided into four categories: feature representation, image representation, sequence representation, and multimodal combination: (1) Feature representation methods use artificially designed features such as high-order statistics and spectral features as input. These methods rely on prior signal knowledge, but can be combined with deep learning to enhance discrimination ability; (2) Image representation methods convert signals into constellation diagrams, time-frequency diagrams, cyclic spectra, etc., which are then recognized by convolutional neural networks. Visualization can intuitively reflect the modulation structure, but there are problems of information redundancy and information loss during conversion; (3) Sequence representation methods directly use the original I / Q sequence or amplitude-phase (AP) sequence as input. I / Q sequences retain complete signal information and are easy to obtain, and are widely used; AP sequences perform better at high signal-to-noise ratios, but their performance is usually not as good as I / Q sequences at low signal-to-noise ratios; (4) Multimodal combination representation: integrates different representation forms (such as sequence + feature, sequence + image, multi-image combination), and improves adaptability in complex channels through multimodal complementarity. Such methods can comprehensively utilize information from multiple domains, such as the time domain and frequency domain, to enhance the robustness of the model. However, these representations are more about adapting to existing classic network models by directly using the original data or making some simple transformations on the original data, without deeply exploring the characteristics of the signal or taking advantage of the multi-domain representation of the signal. Even the multimodal combination method tends to simply combine and directly use features.

[0006] With the continuous development of radar technology and the constant deployment of new radar systems and applications, the types of radar modulation have increased significantly, signal forms have become increasingly complex, and modulation parameters have shown a trend of rapid and agile changes. Against this backdrop, the identification of intra-pulse modulation types in radar signals faces greater challenges. Traditional modulation identification algorithms can only handle a limited number of signal types, making it difficult to adapt to the current highly diverse signal types and increasingly complex signal environments. Furthermore, deep learning-based methods also struggle to achieve robust and accurate identification under complex factors such as low signal-to-noise ratios, the presence of interference signals, changes in transmission conditions, or dynamic changes in signal parameters. Summary of the Invention

[0007] Currently, radar intra-pulse modulation (IPMR) identification techniques based on deep learning models suffer from poor robustness and low generalization ability in practical applications due to complex factors such as low signal-to-noise ratio and dynamic channel changes. To address this, this paper proposes a modulation identification method based on signal amplitude-phase diagram representation and a multi-channel graph convolutional network. This method constructs a novel physically interpretable graph-structured representation of radar signals—the signal amplitude-phase diagram—and designs a dedicated multi-channel graph convolutional network that can directly generalize to complex multipath time-varying channel scenarios without additional retraining—solving a fundamental limitation of many existing deep learning-based IPMR methods.

[0008] The specific design concept of this invention is as follows:

[0009] (1) By utilizing the inherent structural characteristics of different modulation types and their time-frequency coupling properties, a time-frequency fusion diagram structure representation method with clear physical meaning is constructed for radar signals—signal amplitude-phase diagram representation. Nonlinear dynamic analysis is performed on the signal, and the amplitude and phase of the signal are calculated using signal phase point segments as nodes. The amplitude and phase characteristics in the time domain are dynamically and adaptively set as node attributes, while the complex relationship of the spectrum of different segments is used as the edge weight. The real and imaginary parts of the two node attributes and edge weight relationships are combined to obtain four different specific representation forms, thereby realizing the fusion of time-domain and frequency-domain information of the signal.

[0010] (2) Design a dedicated multi-channel graph convolutional network. The main body of the network is a deep graph convolutional neural network, which adopts a four-way parallel processing structure to process four different representation forms of the signal amplitude-phase diagram structure, and introduces a channel attention mechanism to realize the collaborative learning and fusion of different channels.

[0011] The technical solution adopted in this invention is a modulation recognition method based on signal amplitude-phase diagram representation and multi-channel graph convolutional network. The method first builds a multi-channel graph convolutional network model, then performs data processing, including receiving signal data, constructing signal amplitude-phase diagram data, inputting amplitude-phase diagram data into the multi-channel graph convolutional network model to calculate features, then classifying and recognizing, calculating classification loss, training and saving network parameters, and finally recognizing modulation type on new unlabeled data.

[0012] The specific steps of this method are as follows:

[0013] S1 Establish a multi-channel graph convolutional network model

[0014] The multi-channel graph convolutional network has a four-channel structure (channels 1 to 4). All channels have identical structures, are independent of each other, and share parameters, each used to process graph data with different representations. Each channel consists of three GCNBlock units (GCNBlock1, GCNBlock2, and GCNBlock3) connected in series. Each GCNBlock unit consists of a GNN convolutional block and a pooling layer. The GNN convolutional block consists of six uniGNN units, a merging operation layer, and a linear layer. The outputs of the six uniGNN units are merged by the merging operation layer and then passed through the linear layer. The six uniGNN units have the same structure, consisting of a DenseSAGEConv graph convolutional layer, a ReLU activation function, and a LayerNorm normalization layer stacked in series. The size of the graph convolutional layer is 64*64. The pooling layer is DenseSAGPool, used for graph data compression, with a compression ratio of 0.25.

[0015] The outputs of the four channels are input into a channel attention mechanism fusion module for multi-channel fusion, and then input into a classifier CLS consisting of a fully connected layer FC1, a ReLU activation function, a fully connected layer FC2, and a Softmax layer for modulation type identification.

[0016] Network model weight parameters This indicates that the network model weight parameters are initialized using a randomized 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 data to be processed. Where n represents the nth sampling point, , This represents the total number of sampling points; if the total number of signal samples is... Then the first A signal sample can be found To distinguish them, add subscripts, specifically as follows: , The corresponding modulation type is tagged with express, That is, shared Types of modulation;

[0019] S3 signal amplitude-phase diagram construction

[0020] A graph is a data form composed of nodes and edges, and its complete description is mainly composed of node attributes and adjacency matrix. Therefore, the construction method of the signal amplitude phase diagram of the present invention can be divided into the calculation of node attributes and the calculation of adjacency matrix, and then the two parts are combined to form the signal amplitude phase diagram.

[0021] The basic idea behind this invention for constructing amplitude-phase diagrams is to obtain a phase space matrix using phase space reconstruction theory, construct a node attribute matrix, and treat each row of the matrix (i.e., the phase point of the phase space matrix) as a node. The amplitude and phase values ​​of the phase points are used as node attributes, and the complex cosine similarity of the spectra between different nodes is used as edge weights to form an adjacency matrix. By combining the node attribute matrix (amplitude or phase) and the adjacency matrix (real or imaginary part) pairwise, four different forms of graph data are obtained. Because the node attributes are time-domain information (amplitude or phase information), this form of graph data is called a signal amplitude-phase diagram.

[0022] The specific steps are as follows:

[0023] S3-1 Constructing the phase space matrix

[0024] For signals The phase space is reconstructed to obtain the phase space matrix, as follows:

[0025] S3-1-1 Calculation Time Delay and embedding dimension

[0026] Phase space reconstruction requires two important parameters: time delay. and embedding dimension When time delay and embedding dimension After satisfying Takens' theorem (Takens' theorem is also known as the embedding theorem; see Takens F. Detecting strange attractors in turbulence[G] / / Dynamical systems and turbulence, Warwick 1980. Springer, 1981: 366-381.), it can be considered that the differential phase space matrix obtained by phase space reconstruction is topologically equivalent to the nonlinear system characteristics corresponding to the original signal. Therefore, the characteristics of the differential phase space matrix can be calculated to measure the fingerprint information of the radiation source in the signal, thereby realizing the fingerprint identification of the radiation source.

[0027] Calculated using the correlation integral method Corresponding time delay parameters (For details, see Kim H, Eykholt R, Salas J. Nonlinear dynamics, delay times, and embedding windows[J]. PhysicalD: Nonlinear Phenomena, 1999, 127(1-2): 48∼60.);

[0028] Calculated using the Cao method Corresponding embedding dimension (For details, see Cao L, Mees A, Judd K. Dynamics from multivariate time series[J]. Physical D Nonlinear Phenomena, 1998, 121(1-2): 75-88.)

[0029] S3-1-2 Coordinate Delay Reconstruction Method for Constructing Phase Space Matrix

[0030] The phase space matrix is ​​calculated using the coordinate delay reconstruction method. The specific method is as follows:

[0031] ,

[0032] in, Denotes the row number of the difference phase space matrix, satisfying , Represents an N×m dimensional complex space matrix; phase space matrix Each row is also called a phase point;

[0033] S3-2 Node Attribute Calculation

[0034] S3-2-1 Calculate the matrix The amplitude and phase of the middle element

[0035] With matrix The Middle Line number Column corresponding elements For example, calculate the corresponding amplitude. and phase :

[0036] ,

[0037] ,

[0038] in, This indicates taking the real part of the corresponding complex number. This indicates taking the imaginary part of the corresponding complex number, and the same applies below.

[0039] S3-2-2 Organize all amplitudes into an amplitude node attribute matrix :

[0040] ,

[0041] in Represent an N×m dimensional real matrix; Each row serves as a node in the amplitude field.

[0042] S3-2-3 Form a phase node attribute matrix from all phases. :

[0043] ,

[0044] Will Each row serves as a node in the phase field.

[0045] S3-3 Adjacency Matrix Calculation

[0046] S3-3-1 Phase Space Matrix Perform a Fourier transform on each row to obtain the corresponding spectrum matrix. :

[0047] ,

[0048] in, Indicates along the matrix Perform Fourier transform in the row direction. Indicates the first The result of the discrete Fourier transform of each row vector , Representation matrix The Middle Line number Column elements;

[0049] S3-3-2 Construct the Gram matrix according to the following formula. :

[0050] ,

[0051] in, Represents the conjugate transpose, matrix elements It is given by the following formula:

[0052] ,

[0053] express Conjugate;

[0054] S3-3-3 Calculate row vectors of Norm (modulus):

[0055]

[0056] S3-3-4 Combining the norms of all vectors, we obtain the norm vector. :

[0057]

[0058] in Represents a norm vector The i-th element;

[0059] S3-3-5 Calculate the norm vector outer product matrix :

[0060] ,

[0061] S3-3-6 Calculate the adjacency matrix :

[0062] ,

[0063] Where ε is the numerical stability constant, which is usually taken as ε=10. −8 ;

[0064] matrix elements for:

[0065] ,

[0066] S3-3-7 Take The real part is used as the adjacency matrix of the real part. :

[0067] ,

[0068] S3-3-8 Take The imaginary part is used as the imaginary adjacency matrix. :

[0069] ,

[0070] S3-4 Graph Structure Data Construction

[0071] Amplitude node attribute matrix Phase node attribute matrix adjacency matrix of real part Imaginary adjacency matrix By combining the data, four different forms of signal amplitude-phase diagram data are obtained, as follows:

[0072] S3-4-1 The real part adjacency matrix With amplitude node attribute matrix Common Composition Diagram ;

[0073] S3.4.2 Connect the imaginary adjacency matrix With amplitude node attribute matrix Common Composition Diagram ;

[0074] S3.4.3 Connect the real part adjacency matrix Phase node attribute matrix Common Composition Diagram ;

[0075] S3.4.4 Connect the imaginary adjacency matrix Phase node attribute matrix Common Composition Diagram ;

[0076] S4 Feature Calculation

[0077] S4-1 Channel 1 Feature Calculation

[0078] Will Input channel 1, which passes through three GCNBlock units of channel 1 in sequence, and finally outputs the embedded features of channel 1. The specific steps are as follows:

[0079] S4-1-1 will Input the first GCNBlock unit GCNBlock1, output the updated graph. The specific calculation process is as follows:

[0080] In unit S4-1-1-1, the nodes are sequentially passed through six GNN convolutional blocks. The node attributes input to each convolutional block are the node attributes output by the previous convolutional block, while the input adjacency matrix remains consistent. The embedding features output by the six convolutional blocks are represented as follows: ,…, The dimensions of the embedded features are all ;

[0081] S4-1-1-2 Embedded Features arrive After the merging operation layer, the combined features of all GNN convolutional blocks in channel 1 are formed. Dimension is :

[0082] ,

[0083] S4-1-1-3 will The median value of the output node attribute matrix in the linear layer of the input GNN convolutional block. Dimension is ;

[0084] S4-1-1-4 The median value of the node attribute matrix adjacency matrix of real part Input a pooling layer DenseSAGPool, update the node attribute matrix and adjacency matrix, and output the updated graph. ;picture The number of nodes is lower than that of the input graph. The compression ratio is 0.25, which is the updated version. The number of graph nodes is reduced to the original input. One-quarter;

[0085] S4-1-2 will Input the second GCNBlock unit GCNBlock2, and output the updated graph. For details, please refer to step S4.1.1.

[0086] S4-1-3 will Input the third GCNBlock unit GCNBlock3, output the updated graph. For details, please refer to step S4.1.1.

[0087] Pick The node attribute matrix in the graph data is used as the embedding feature of channel 1, and... express.

[0088] S4-2 Channel 2 Feature Calculation

[0089] Will The input channel 2 is passed through three GCNBlock units in sequence, and the final output is the embedding feature of channel 2. For details, please refer to step S4-2.

[0090] S4-2 Channel 3 Feature Calculation

[0091] Will The input channel 3 is passed through three GCNBlock units in sequence, and the final output is the embedding feature of channel 3. For details, please refer to step S4-2.

[0092] S4-2 Channel 4 Feature Calculation

[0093] Will The input channel 4 is passed through three GCNBlock units in sequence, and the final output is the embedding feature of channel 4. For details, please refer to step S4.2.

[0094] S4-3 Multi-channel Feature Fusion

[0095] Embedded features obtained from four channels , , and The input channel attention mechanism fusion module obtains multi-channel fused features. ;

[0096] S5 Classification and Recognition

[0097] Multi-channel fusion features Input classifier CLS, output predicted label for radar signal modulation recognition ;

[0098] S6 Calculate classification loss

[0099] Predictive tags identified by modulating radar signals output from the S5 With the actual modulation type label in S2 Calculate the cross-entropy loss and use it as the classification loss during training. :

[0100] ;

[0101] S7 Network Training and Parameter Saving

[0102] S7-1 Network Training

[0103] 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 a multi-channel graph convolutional network model, and the network model weight parameters are updated in real time. This causes the classification loss during training to be higher. Continuing to decline;

[0104] 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.

[0105] S7-1-3 Until Classification 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.

[0106] S7-2 Network Parameter Saving

[0107] Save the optimal network model weight parameters when the network reaches steady state. ;

[0108] S8 identifies modulation type labels

[0109] Optimal network model weight parameters The trained multi-channel graph convolutional network model is applied to identify the modulation of intra-pulse signals.

[0110] The present invention has the following technical effects:

[0111] (1) A dynamic mapping paradigm for amplitude-phase diagram structure fusion from signal sequence to time-frequency information is proposed to realize the adaptive structured transformation from original complex signal to topological graph structure data. It has inherent robustness against noise and multipath fading, and provides a new structured representation dimension for radar signal analysis.

[0112] (2) Design a dedicated multi-channel graph convolutional network for the complex domain. Adopt a multi-path architecture to support the collaborative learning of multi-dimensional information such as amplitude and phase, real part and imaginary part, and combine it with the channel attention fusion mechanism to achieve efficient fusion of multi-channel information.

[0113] (3) The present invention has good recognition performance and good robustness to low signal-to-noise ratio and multipath channels.

[0114] 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

[0115] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0116] Figure 2 This is the signal amplitude-phase diagram data construction process (S3).

[0117] Figure 3 It is the feature calculation and classification recognition process (S4-S5);

[0118] Figure 4 Here is the structure of GCNBlock: (a) outline, (b) detailed diagram;

[0119] Figure 5 It is the structural component of uniGNN;

[0120] Figure 6This is the performance curve of the correct recognition rate of the present invention as a function of SNR under AWGN channel;

[0121] Figure 7 This is the performance curve of the correct identification rate as a function of SNR under a random multipath fading channel with unknown characteristics. Detailed Implementation

[0122] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0123] Figure 1 This is a flowchart illustrating the implementation of the present invention. The present invention proposes a modulation recognition method based on signal amplitude-phase diagram representation and multi-channel graph convolutional network, which consists of the following steps:

[0124] S1. Establish a multi-channel graph convolutional network model;

[0125] S2 data reception;

[0126] S3 signal amplitude-phase diagram data construction;

[0127] S4 feature calculation;

[0128] S5 classification and recognition;

[0129] S6 Calculates the classification loss;

[0130] S7 network training and parameter saving;

[0131] S8 identifies the modulation type label.

[0132] Figure 2 This is the signal amplitude-phase diagram construction process (S3). The basic process involves constructing a phase space matrix from the received signal. Each row of the phase space matrix corresponds to a row vector (phase point in the phase space), which is then used as a node. The amplitude and phase, represented in polar coordinates by the complex values ​​of the row vectors, are used as the node's attribute matrix, resulting in amplitude node attribute matrices and phase node attributes, respectively. Then, a Fourier transform is performed on each row's corresponding node to obtain the corresponding spectral features. The complex cosine similarity between all pairs of nodes is calculated as the edge attribute relationship. The real and imaginary parts are used to construct the real part adjacency matrix and the imaginary part adjacency matrix, respectively. Finally, the node attribute matrices and adjacency matrices are combined pairwise to construct signal amplitude-phase diagram data in four different representation forms. This is the new paradigm for converting signals into graph-structured data proposed in this invention.

[0133] Figure 3This is the feature calculation and classification process (S4-S5). As you can see, the four complete image data are input into channels 1 through 4 for independent processing. Then, the features output from all channels are fused using a channel attention mechanism to obtain the final embedded features. Finally, the classifier outputs the modulated recognition category label. It's important to note that all channels have the same structure, operate independently during processing, and share parameters.

[0134] Figure 4 (a) is a schematic diagram of the GCNBlock structure: GCNBlock consists of a GNN convolutional block and a pooling layer. The pooling layer uses DenseSAGPool pooling, and the GNN convolutional block consists of six uniGNN units, a merging operation layer, and a linear layer. The six uniGNN units are concatenated end-to-end and merged together, receiving the same adjacency matrix during computation to progressively update node attributes. The merging operation layer combines the outputs of all uniGNN units. In other words, these six uniGNN units are concatenated 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 is input to the merging operation layer. The last uniGNN unit, however, only has one output to the merging operation layer. The merging operation layer merges the node attributes output by the six uniGNN units, then performs a dimensionality transformation through the linear layer, outputting the updated node attributes of the GNN convolutional block. Finally, the original adjacency matrix and the updated node attributes are input into the DenseSAGPool pooling layer to further compress and update the node attributes and adjacency matrix. The number of nodes corresponding to the updated node attributes and adjacency matrix is ​​less than the original number. Figure 4 (b) is a detailed diagram of the structure of GCNBlock, which shows that... Figure 4 and Figure 5 Details of the composition of a single GCNBlock unit after combination.

[0135] Figure 5 This is the structure of uniGNN. As shown in the figure, uniGNN consists of DenseSAGConv convolutional layers, ReLU activation layers, and normalization layers (LayerNorm). It takes the adjacency matrix and node attributes as input and outputs the updated node attributes.

[0136] Figure 6This 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 6 As shown, the method of this invention achieves over 90% accuracy at 0dB and 95.19% correct recognition rate for 18 modulation types at 2dB, fully demonstrating its effectiveness. Furthermore, it still achieves approximately 80% accuracy at -6dB, proving its robustness to noise.

[0137] Figure 7 This is the performance curve of the correct recognition rate as a function of SNR under a random multipath fading channel with unknown characteristics, which is derived from... Figure 6 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 maintains a 90% correct recognition rate at 0dB on unknown multipath channels without training, and still maintains 71.73% at -6dB, demonstrating its robustness and generalization ability to unknown signals.

Claims

1. A modulation recognition method based on signal amplitude-phase diagram representation and multi-channel graph convolutional network, characterized in that, The specific steps of this method are as follows: S1 Establish a multi-channel graph convolutional network model Multichannel graph convolutional networks have a four-channel structure. All channels are identical in structure, independent of each other, and share parameters. They are used to process graph data with different representations. Each channel is composed of three GCNBlock units, GCNBlock1, GCNBlock2, and GCNBlock3, connected in series. Each GCNBlock unit consists of a GNN convolutional block and a pooling layer. The GNN convolutional block consists of six uniGNN units, a merging operation layer, and a linear layer. The outputs of the six uniGNN units are merged by the merging operation layer and then passed through the linear layer. The six uniGNN units have the same structure, each consisting of a DenseSAGEConv graph convolutional layer, a ReLU activation function, and a LayerNorm normalization layer stacked in series. The graph convolutional layer is 64*64 in size; the pooling layer is DenseSAGPool, used for graph data compression, with a compression ratio of 0.

25. The outputs of the four channels are input into a channel attention mechanism fusion module for multi-channel fusion, and then input into a classifier CLS consisting of a fully connected layer FC1, a ReLU activation function, a fully connected layer FC2, and a Softmax layer for modulation type identification. Network model weight parameters This indicates that the network model weight parameters are initialized using a randomized method. ; S2 Data Reception The received signal is processed through filtering, down-conversion, and analog-to-digital conversion to obtain the discrete signal data to be processed. Where n represents the nth sampling point, , This represents the total number of sampling points; If the total number of signal samples is Then the first A signal sample can be found To distinguish them, add subscripts, specifically as follows: , The corresponding modulation type is tagged with express, That is, shared Types of modulation; S3 signal amplitude-phase diagram construction The specific steps are as follows: S3-1 Constructing the phase space matrix For signals The phase space is reconstructed to obtain the phase space matrix, as follows: S3-1-1 Calculation Time Delay and embedding dimension ; S3-1-2 Coordinate Delay Reconstruction Method for Constructing Phase Space Matrix The phase space matrix is ​​calculated using the coordinate delay reconstruction method. The specific method is as follows: ， in, Denotes the row number of the difference phase space matrix, satisfying , Represents an N×m dimensional complex space matrix; phase space matrix Each row is also called a phase point; S3-2 Node Attribute Calculation S3-2-1 Calculate the matrix The amplitude and phase of the middle element With matrix The Middle Line number Column corresponding elements For example, calculate the corresponding amplitude. and phase : ， ， in, This indicates taking the real part of the corresponding complex number. This indicates taking the imaginary part of the corresponding complex number; S3-2-2 Organize all amplitudes into an amplitude node attribute matrix : ， in Represent an N×m dimensional real matrix; Each row serves as a node in the amplitude field; S3-2-3 Form a phase node attribute matrix from all phases. : ， Will Each row serves as a node in the phase field; S3-3 Adjacency Matrix Calculation S3-3-1 Phase Space Matrix Perform a Fourier transform on each row to obtain the corresponding spectrum matrix. : ， in, Indicates along the matrix Perform Fourier transform in the row direction. Indicates the first The result of the discrete Fourier transform of each row vector , Representation matrix The Middle Line number Column elements; S3-3-2 Construct the Gram matrix according to the following formula. : ， in, Represents the conjugate transpose, matrix elements It is given by the following formula: ， express Conjugate; S3-3-3 Calculate row vectors of Norm: ； S3-3-4 Combining the norms of all vectors, we obtain the norm vector. : ， in Represents a norm vector The i-th element; S3-3-5 Calculate the norm vector outer product matrix : ， S3-3-6 Calculate the adjacency matrix : ， Where ε is the numerical stability constant; matrix elements for: ， S3-3-7 Take The real part is used as the adjacency matrix of the real part. : ， S3-3-8 Take The imaginary part is used as the imaginary adjacency matrix. : ， S3-4 Graph Structure Data Construction Amplitude node attribute matrix Phase node attribute matrix adjacency matrix of real part Imaginary adjacency matrix By combining the data, four different forms of signal amplitude-phase diagram data are obtained, as follows: S3-4-1 The real part adjacency matrix With amplitude node attribute matrix Common Composition Diagram ; S3.4.2 Connect the imaginary adjacency matrix With amplitude node attribute matrix Common Composition Diagram ; S3.4.3 Connect the real part adjacency matrix Phase node attribute matrix Common Composition Diagram ; S3.4.4 Connect the imaginary adjacency matrix Phase node attribute matrix Common Composition Diagram ; S4 Feature Calculation S4-1 Channel 1 Feature Calculation Will Input channel 1, which passes through three GCNBlock units of channel 1 in sequence, and finally outputs the embedded features of channel 1. The specific steps are as follows: S4-1-1 will Input the first GCNBlock unit GCNBlock1, output the updated graph. The specific calculation process is as follows: In unit S4-1-1-1, the nodes are sequentially passed through six GNN convolutional blocks. The node attributes input to each convolutional block are the node attributes output by the previous convolutional block, while the input adjacency matrix remains consistent. The embedding features output by the six convolutional blocks are represented as follows: ,…, The dimensions of the embedded features are all ; S4-1-1-2 Embedded Features arrive After the merging operation layer, the combined features of all GNN convolutional blocks in channel 1 are formed. Dimension is : ， S4-1-1-3 will The median value of the output node attribute matrix in the linear layer of the input GNN convolutional block. Dimension is ; S4-1-1-4 The median value of the node attribute matrix adjacency matrix of real part Input a pooling layer DenseSAGPool, update the node attribute matrix and adjacency matrix, and output the updated graph. ;picture The number of nodes is lower than that of the input graph. The compression ratio is 0.25, which is the updated version. The number of graph nodes is reduced to the original input. One-quarter; S4-1-2 will Input the second GCNBlock unit GCNBlock2, and output the updated graph. ; S4-1-3 will Input the third GCNBlock unit GCNBlock3, output the updated graph. ; Pick The node attribute matrix in the graph data is used as the embedding feature of channel 1, and... express; S4-2 Channel 2 Feature Calculation Will The input channel 2 is passed through three GCNBlock units in sequence, and the final output is the embedding feature of channel 2. ; S4-2 Channel 3 Feature Calculation Will The input channel 3 is passed through three GCNBlock units in sequence, and the final output is the embedding feature of channel 3. ; S4-2 Channel 4 Feature Calculation Will The input channel 4 is passed through three GCNBlock units in sequence, and the final output is the embedding feature of channel 4. ; S4-3 Multi-channel Feature Fusion Embedded features obtained from four channels , , and The input channel attention mechanism fusion module obtains multi-channel fused features. ; S5 Classification and Recognition Multi-channel fusion features Input classifier CLS, output predicted label for radar signal modulation recognition ; S6 Calculate classification loss Predictive tags identified by modulating radar signals output from the S5 With the actual modulation type label in S2 Calculate the cross-entropy loss and use it as the classification loss during training. : ； S7 Network Training and Parameter Saving S7-1 Network Training S7-1-1 applies gradient descent to a multi-channel graph convolutional network model and updates the network model weight parameters in real time. This causes the classification loss during training to be higher. Continuing to decline; S7-1-2 updates the learning rate; S7-1-3 Until Classification 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. S7-2 Network Parameter Saving Save the optimal network model weight parameters when the network reaches steady state. ; S8 identifies modulation type labels Optimal network model weight parameters The trained multi-channel graph convolutional network model is applied to identify the modulation of intra-pulse signals.

2. The modulation recognition method based on signal amplitude-phase diagram representation and multi-channel graph convolutional network according to claim 1, characterized in that: In S3-1-1, the correlation integral method is used to calculate... Corresponding time delay parameters Calculation using the Cao method Corresponding embedding dimension .

3. The modulation recognition method based on signal amplitude-phase diagram representation and multi-channel graph convolutional network according to claim 1, characterized in that: In S3-3-6, the numerical stability constant is taken as ε=10. −8 .

4. The modulation recognition method based on signal amplitude-phase diagram representation and multi-channel graph convolutional network according to claim 1, characterized in that: In S7-1-1, the Adam algorithm is used to perform gradient descent on a multi-channel graph convolutional network model.

5. The modulation recognition method based on signal amplitude-phase diagram representation and multi-channel graph convolutional network according to claim 1, characterized in that: In S7-1-2, the initial learning rate is set to 1e-4.

6. The modulation recognition method based on signal amplitude-phase diagram representation and multi-channel graph convolutional network according to claim 1, characterized in that: In S7-1-2, a cosine annealing algorithm with restart is used to update the learning rate, and the restart period is set to 2.

Images