Semi-supervised GCN modulation identification method based on Transform mask

Abstract

The invention relates to a semi-supervised GCN modulation identification method based on a Transform mask, and the method comprises the steps: building a semi-supervised GCN model of the Transform mask, and calculating the weight distribution between sampling points in a signal sequence through employing a multi-head attention mechanism of the Transform, and constructing an I / Q signal diagram; updating node features in the I / Q signal graph structure by using a graph convolutional network based on p-Laplacian; training is carried out through semi-supervised supervised loss and non-supervised loss; the full connection layer combines the local features into global features for classification decision, and the softmax layer converts the features output by the full connection layer into a final classification result. According to the method, the model can preferentially sample and capture the most significant feature point of the original signal through the weight distribution strategy of the Transform, so that the capability of the model to accurately identify the signal feature is enhanced; according to the semi-supervised graph convolutional network based on the Transform mask, a semi-supervised loss function is introduced through feature reconstruction of the Transform network, and the recognition accuracy and stability of the model are improved under the condition that the annotation data is insufficient.

Description

Technical Field

[0001] This application relates to the field of computer data communication technology, and in particular, to a semi-supervised GCN modulation recognition method based on Transformer masking. Background Art

[0002] Although deep learning (DL) technology has significantly improved the recognition accuracy, especially when using high-quality, well-annotated datasets, it still faces challenges when dealing with datasets with inaccurate annotations or incomplete signal reception. Semi-supervised learning (SSL) combines a small number of annotated samples and a large amount of unannotated data to train the model, thus reducing the need for a large number of labels. The successful application of SSL in other fields has also promoted the development of AMC (Automatic Modulation Classification) to solve the problem of obtaining a large number of specific annotated signals from non-cooperating parties. Semi-supervised graph convolutional networks (GCNs) perform well in processing graph-structured data, especially in node classification tasks. In recent years, semi-supervised GCNs have received extensive attention due to their ability to automatically learn the feature and structure information of graph data. This has led to a rapid growth of GCNs in the field of non-graph data. Generally, GCNs involve the problem of relationship construction in non-graph data, including applications in the field of AMC.

[0003] Therefore, it is necessary to propose a solution to improve one or more problems existing in the above-mentioned related technical solutions.

[0004] It should be noted that the information disclosed in the above background art section is only used to enhance the understanding of the background of this application, and thus may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0005] The purpose of the present disclosure is to provide a semi-supervised GCN modulation recognition method based on Transformer masking, thereby at least to some extent overcoming one or more problems caused by the limitations and deficiencies of the related art.

[0006] A semi-supervised GCN modulation recognition method based on Transformer masking provided according to an embodiment of the present disclosure includes the following steps:

[0007] Establish a semi-supervised GCN model with Transformer masking, and use the multi-head attention mechanism of Transformer to calculate the weight distribution between sampling points in the signal sequence to construct an I / Q signal graph;

[0008] Use a graph convolutional network based on p-Laplacian to update the node features in the structure of the I / Q signal graph;

[0009] The semi-supervised GCN model with Transformer mask is trained on labeled signals and unlabeled signals through semi-supervised supervised loss and unsupervised loss;

[0010] The fully connected layer combines local features into global features for classification decision-making, and the softmax layer converts the features output by the fully connected layer into the final classification result, classifying the signals into different modulation types.

[0011] In an exemplary embodiment of the present application, in the step of using the multi-head attention mechanism of Transformer to calculate the weight distribution between sampling points in the signal sequence to construct the I / Q signal graph, it further includes:

[0012] Preprocess the original discrete signals of the labeled signals and unlabeled signals into I / Q vectors, and combine the I / Q sampling points over time to obtain a signal sequence;

[0013] Use the multi-head attention mechanism of Transformer to calculate the weight distribution between sampling points in the signal sequence to construct an I / Q fully connected undirected graph, and convert the I / Q vectors into node and edge information in the graph structure.

[0014] In an exemplary embodiment of the present application, in the step of using the multi-head attention mechanism of Transformer to calculate the weight distribution between sampling points in the signal sequence to construct the I / Q signal graph, it further includes:

[0015] By selecting the largest weight and its corresponding node in the I / Q fully connected undirected graph, represent the original signal as a non-fully connected graph.

[0016] In an exemplary embodiment of the present application, the expression of the signal sequence is:

[0017] x = {[I0, Q0], [I1, Q1],..., [I N-1 , Q N-1}.

[0018] In an exemplary embodiment of the present application, the expression of the multi-head attention mechanism is:

[0019]

[0020] where w i,j is the sum of K normalized multi-head dot product attentions between nodes i and j, and attention k (x i , x j ) is the k-th attention.

[0021] In an exemplary embodiment of the present application, in the step of updating the node features in the I / Q signal graph structure by using the p-Laplacian-based graph convolutional network, the following steps are further included:

[0022] Calculate the degree matrix according to the adjacency matrix of the I / Q signal graph;

[0023] Calculate the p-Laplacian matrix according to the degree matrix;

[0024] Calculate the first-order approximation of the p-Laplacian matrix to update the node features in the I / Q signal graph structure.

[0025] In an exemplary embodiment of the present application, in the calculation of the first-order approximation of the p-Laplacian matrix, the expression of the definition of the first-order convolution of the p-Laplacian is:

[0026]

[0027] Where g θ (L p ) is the graph convolution operation of the p-Laplacian matrix, θ is the parameter in the graph convolution operation, L p is the p-Laplacian matrix, X represents the node feature matrix in the graph, I N represents the identity matrix of size N×N, N is the number of nodes in the graph, λ max is the maximum eigenvalue of the p-Laplacian, σ is the activation function, W σ is the weight matrix of each convolutional layer, and the structural information obtained by the graph p-Laplacian is a positive semi-definite matrix.

[0028] In an exemplary embodiment of the present application, the total loss function for training the labeled signals and unlabeled signals with the semi-supervised supervised loss and unsupervised loss is the weighted sum of the supervised loss and the unsupervised loss. The expression of the total loss function is:

[0029]

[0030] Where D la represents the set of labeled data, D un represents the set of unlabeled data, x is the input data, y is the label corresponding to x, the supervised loss is L s , the unsupervised loss is L u , α and β represent the weights of the supervised loss and the unsupervised loss, and ψ and φ are the parameters of the encoder and the classifier respectively.

[0031] In an exemplary embodiment of the present application, the expression of the supervised loss is as follows:

[0032]

[0033] where C is the number of types of label types, y c is the true label, is the predicted label using the Softmax classifier.

[0034] In an exemplary embodiment of the present application, the expression of the unsupervised loss is as follows:

[0035]

[0036] where x i and represent the original feature and the reconstructed feature of the signal.

[0037] A semi-supervised GCN modulation recognition method based on Transformer masking proposed in the present application. On the one hand, through the weight distribution strategy of the Transformer, the model can preferentially sample and capture the points with the most significant features of the original signal, thereby enhancing the model's ability to accurately identify signal features; on the other hand, based on the semi-supervised graph convolutional network with Transformer masking, a semi-supervised loss function is introduced through the feature reconstruction of the Transformer network, which improves the recognition accuracy and stability of the model in the case of insufficient labeled data. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The drawings herein are incorporated into the specification and constitute a part of this specification, showing embodiments consistent with the present application, and are used together with the specification to explain the principles of the present application. Obviously, the drawings in the following description are only some embodiments of the present application, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0039] Figure 1 A schematic diagram showing the steps of the semi-supervised GCN modulation recognition method based on Transformer masking in an exemplary embodiment of the present application;

[0040] Figure 2 A schematic diagram showing the construction process of the I / Q signal diagram in an exemplary embodiment of the present application;

[0041] Figure 3 A schematic diagram showing the construction process of the simplified I / Q signal diagram in an exemplary embodiment of the present application;

[0042] Figure 4Schematic diagram of the semi-supervised GpCN modulation recognition model based on Transformer mask in an exemplary embodiment of the present application;

[0043] Figure 5 Schematic diagram of the training loss curve of the semi-TMGpCN model in the simulation experiment of the present disclosure;

[0044] Figure 6 Schematic diagram of the recognition accuracy of different embedding networks at different signal-to-noise ratios in the simulation experiment of the present disclosure;

[0045] Figure 7 Schematic diagram of the recognition rate of the TMGpCN model at different mask ratios in the simulation experiment of the present disclosure;

[0046] Figure 8 Schematic diagram of 10% mask sampling of 8PSK and PAM4 at 10 dB in the simulation experiment of the present disclosure;

[0047] Figure 9 Schematic diagram of the comparison of the recognition rates of the TMGCN and TMGpCN models at different label rates in the simulation experiment of the present disclosure;

[0048] Figure 10 Schematic diagram of the confusion matrix of the recognition rates of the TMGCN and TMGpCN models at 100% label rate in the simulation experiment of the present disclosure;

[0049] Figure 11 Schematic diagram of the comparison of the recognition rates of different semi-supervised AMC models at different label rates in the simulation experiment of the present disclosure. Detailed implementation manners

[0050] Now, example embodiments will be described more fully with reference to the accompanying drawings. However, the example embodiments can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this application will be more thorough and complete, and will fully convey the concept of the example embodiments to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0051] In addition, the accompanying drawings are only schematic illustrations of the present application and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and thus their repeated description will be omitted. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0052] This exemplary embodiment provides a semi-supervised GCN modulation recognition method based on a Transformer mask. As Figure 1 shown, it may include the following steps:

[0053] Step S101: Establish a semi-supervised GCN model with a Transformer mask, and use the multi-head attention mechanism of the Transformer to calculate the weight distribution between sampling points in the signal sequence to construct an I / Q signal graph.

[0054] Step S102: Use a graph convolutional network based on p-Laplacian to update the node features in the I / Q signal graph structure.

[0055] Step S103: The semi-supervised GCN model with a Transformer mask trains the labeled signals and unlabeled signals through semi-supervised supervised loss and unsupervised loss.

[0056] Step S104: The fully connected layer combines local features into global features for classification decision-making, and the softmax layer converts the features output by the fully connected layer into the final classification result, classifying the signals into different modulation types.

[0057] A semi-supervised GCN modulation recognition method based on a Transformer mask proposed in the embodiments of this application, on the one hand, enables the model to preferentially sample and capture the points with the most significant features of the original signal through the weight distribution strategy of the Transformer, thereby enhancing the model's ability to accurately identify signal features; on the other hand, based on the semi-supervised graph convolutional network with a Transformer mask, a semi-supervised loss function is introduced through the feature reconstruction of the Transformer network, improving the recognition accuracy and stability of the model in the case of insufficient labeled data.

[0058] Next, as Figures 1 - 11 shown, a semi-supervised GCN modulation recognition method based on a Transformer mask proposed in this exemplary embodiment will be described in more detail.

[0059] Step S101: Establish a semi-supervised GCN model with a Transformer mask, and use the multi-head attention mechanism of the Transformer to calculate the weight distribution between sampling points in the signal sequence to construct an I / Q signal graph.

[0060] Step S1011: Preprocess the original discrete signals of the labeled signals and unlabeled signals into I / Q vectors, and combine the I / Q sampling points over time to obtain a signal sequence.

[0061] Specifically, the I / Q sequence is time-series. The original discrete signal is usually preprocessed into I / Q vectors, which are two time-related sequences with a length of N, i.e., sampling points:

[0062]

[0063] in

[0064] Combining the I / Q samples over time gives the signal sequence x:

[0065] x={[I0,Q0],[I1,Q1],...,[I N-1 ,Q N-1 ]} (2)

[0066] Step S1012: Use the Transformer's multi-head attention mechanism to calculate the weight distribution between the sampling points in the signal sequence to construct an I / Q fully connected undirected graph, and convert the I / Q vectors into node and edge information in the graph structure.

[0067] Specifically, an I / Q signal graph G containing N sampling points is constructed through the signal sequence x s , and use multi-head attention to calculate the value of the weight distribution, output the adjacency matrix between the sampling points and the sampling point mapping feature vector, through the adjacency matrix and mapping features, a fully connected undirected graph G can be constructed s Its construction process, such as Figure 2 As shown, G s It is an undirected graph, whose weight matrix is a symmetric matrix and can be transformed into an upper triangular matrix. i,j is the sum of K normalized multi-head dot product attentions between nodes i and j, as follows:

[0068]

[0069] k-th attention k (x i ,x j ) is calculated as:

[0070]

[0071]

[0072] in is the weight coefficient of the kth attention, W kis the learnable weight shared by two nodes in the k-th attention. The edge constructed by the attention is a scalar, and softmax ensures that the sum of the weight coefficients is 1. The Transformer learns through independent linear mappings and obtains feature information through concatenation and fusion.

[0073] After the feature embedding by the Transformer, it can form the nodes of graph G s Through the judgment of non-zero values of the matrix W k the edges and weights of the graph can be formed. For the undirected fully connected graph of I / Q signals, graph G s is then constructed.

[0074] Step S1013: Represent the original signal as an incomplete connected graph by selecting the largest weight and its corresponding nodes in the I / Q fully connected undirected graph.

[0075] Specifically, since the weight distribution can be converted into a symmetric upper triangular or lower triangular matrix, G s contains L×(L + 1) / 2 effective weights (including the self-connections of the nodes). By selecting the largest T weights and their corresponding nodes, the original signal can be represented as a simpler incomplete connected graph which can effectively reduce the computational overhead of graph convolution and improve its practicality.

[0076] As Figure 3 shown, in the constructed matrix W, the largest T weight values construct a new weight matrix and carry the associated edge indices From the corresponding edge index set a new node set can be constructed Using the new node set edge indices and the edge weight distribution a simpler I / Q signal graph can be constructed Since the number of nodes M in the new signal graph is not fixed, so M ≤ L.

[0077] Step S102: Use the graph convolutional network based on p-Laplacian to update the node features in the I / Q signal graph structure.

[0078] According to the principle of the graph Laplacian operator, it measures the change in node gradients and represents the divergence of the gradient field of each node in the graph. In an undirected graph G, it is defined as the difference between the out-degree and in-degree of each node in the graph, L = D - W, where W represents the matrix of weighted edges, and D is a diagonal matrix, A ij is the adjacency matrix of the graph.

[0079] Spectral domain GCN performs convolution operations on the graph spectrum, similar to applying convolution operations after the signal X is transformed to the frequency domain through Fourier transform. In the frequency domain, these convolutions g θ are designed to enhance or suppress certain frequencies in the signal, effectively performing frequency domain filtering operations. The Laplacian-based graph convolution can be defined as:

[0080]

[0081] where, is a normalized Laplacian matrix. U and Λ represent the eigenvector and eigenvalue matrices of L respectively. To avoid the computational resources required for the eigen-decomposition of L, the K-order Chebyshev polynomial T k (.) is used to approximate the approximation of L, and the corresponding GCN can be expressed as:

[0082]

[0083] where, λ max represents the maximum eigenvalue of the Laplacian operator L. To further optimize the computational process of graph convolution, the first-order approximation is used, and only the structural information between directly connected nodes is considered to update the graph nodes, that is, the local spectral graph convolution with K = 1:

[0084]

[0085] where represents the adjacency matrix including self-connections, is about the degree matrix of. W is the weight parameter of each layer of the network. The normalized Laplacian operator improves the gradient iteration convergence speed of the GCN model, and due to the effective feature extraction of non-Euclidean data, many variants have evolved currently. From the definition of GCN, the structural information of the data is embedded in the graph nodes to obtain more useful features.

[0086] Step S1021: Calculate the degree matrix according to the adjacency matrix of the I / Q signal graph.

[0087] Specifically, the p-Laplacian matrix is a generalization of the graph Laplacian matrix in graph theory, which helps to explore the topological features of graph data and discover the internal structure of the data. The p-Laplacian matrix L p is defined as:

[0088] L p = D p - A (9)

[0089] Among them, A is the adjacency matrix of graph G, and D p is the degree matrix, and its diagonal element D p (i, i) is the p-th power of the degree of node i. The p-Laplacian value between nodes i and j is defined as:

[0090] L p (i, j) = |a ij | p-2 a ij , i ≠ j (10)

[0091] where a ij is an element in the adjacency matrix A.

[0092] The GCN (GpCN) based on p-Laplacian is used to fully obtain the internal structural relationship information in the signal. According to the definition of formula (4.7), the K-th order approximation of GpCN can be expressed as:

[0093]

[0094] Among them, the parameter θ is the weight coefficient of the K-th order operator.

[0095] Step S1022: Calculate the p-Laplacian matrix according to the degree matrix.

[0096] Specifically, the graph p-Laplacian operator L p is a generalization of the Laplacian operator considering the p-norm of node differences. It is normalized by the maximum eigenvalue λ max of p-Laplacian to obtain the normalized p-Laplacian matrix Expression:

[0097]

[0098] Among them, I N represents the identity matrix of size N×N, where N is the number of nodes in the graph. To simplify the calculation process,

[0099] Step S1023: Calculate the first-order approximation of the p-Laplacian matrix and update the node features in the I / Q signal graph structure.

[0100] Referring to formula (8), the first-order approximation of the p-Laplacian matrix can be calculated. This approximation simplifies the hierarchical convolution operation and transforms it into a linear function in the p-Laplacian spectral domain. This method not only reduces the computational complexity but also retains the basic characteristics of graph signal processing. Its simplification process is:

[0101]

[0102] In practical applications, the shared θ0 and θ1 in each convolutional layer increase the computational complexity. To distinguish self-connections and avoid negative numbers in the structural information, the definition of the first-order spectral graph p-Laplacian convolution is further optimized:

[0103]

[0104] where g θ (L p ) is the graph convolution operation of the p-Laplacian matrix, θ is the parameter in the graph convolution operation, L p is the p-Laplacian matrix, X represents the node feature matrix in the graph, I N represents the identity matrix of size N×N, N is the number of nodes in the graph, λ max is the largest eigenvalue of the p-Laplacian, σ is the activation function, W σ is the weight matrix of each convolutional layer, and the structural information obtained by the graph p-Laplacian is a positive semi-definite matrix. It will be preprocessed before model training, so it will not increase the computational complexity of the GCN model.

[0105] Step S103: The semi-supervised GCN model with Transformer mask trains the labeled signals and unlabeled signals through semi-supervised supervised loss and unsupervised loss.

[0106] Specifically, in semi-supervised AMC, the signal dataset X is divided into a training set X tr ={X la ,X un} and a test set X te ={X te}. Among them, X la ={(x1,y1),(x2,y2),...,(x La ,y La )} is the labeled data, X un =x1,x2,...,x Un is the unlabeled data, and La ≤ Un. Semi-supervised learning is to use the labeled data X la and the unlabeled data X un to learn a prediction model to predict the modulation method for the signals in X te , that is, g φ (f ψ (x)):X→Y, where ψ and φ are the parameters of the encoder and classifier respectively. The total loss function L tDenoted as the supervised loss L s and the unsupervised loss L u as a weighted sum, i.e.:

[0107]

[0108] where D la represents the set of labeled data, D un represents the set of unlabeled data, x is the input data, y is the label corresponding to x, the supervised loss is L s and the unsupervised loss is L u , α and β represent the weights of the supervised loss and the unsupervised loss, and ψ and φ are the parameters of the encoder and the classifier respectively. During the training process, the cross-entropy loss function minimizes the error and continuously optimizes the parameters, i.e.:

[0109]

[0110] where C is the number of types of label classes, y c is the true label, and y c is the predicted label using the Softmax classifier. The unsupervised loss L u is represented by the feature encoding reconstruction loss, i.e., the mean squared error (MSE) is used to measure the difference in the feature reconstruction process, which is expressed as:

[0111]

[0112] where x i and represent the original feature and the reconstructed feature of the signal respectively.

[0113] Step S104: The fully connected layer combines the local features into global features for classification decision-making, and the softmax layer converts the features output by the fully connected layer into the final classification result, classifying the signal into different modulation types.

[0114] It should be understood that the GCN based on p-Laplacian is GpCN, and TMGpCN is the semi-supervised GpCN based on the Transformer mask.

[0115] Specifically, the signal is classified by performing the fully connected layer and the softmax operation. The fully connected layer can integrate the features extracted by the previous layers, while the softmax layer is responsible for converting these features into the final classification result, making the TMGpCN algorithm perform excellently in the modulation recognition task, especially in the case of limited labeled samples, and can effectively improve the recognition accuracy.

[0116] Such as Figure 4As shown in Table 4.1, the process of the TMGpCN modulation recognition algorithm is as follows:

[0117] First, the algorithm converts the original I / Q vector into node and edge information in the graph structure through the attention mechanism. This step can capture the key features of the signal and lay the foundation for subsequent graph convolution operations.

[0118] Next, the algorithm simplifies the graph structure by selecting nodes with larger weights. This can reduce the computational complexity while retaining the core information of the signal.

[0119] Then, it uses the p-Laplacian based graph convolutional network (GpCN) to update the node features. This process can further extract and strengthen the feature representation of the signal. The introduction of the p-Laplacian operator can capture the topological information of graph-structured data more effectively than traditional GCN, enhancing the model's ability to extract signal features.

[0120] Finally, the signal is classified by performing a fully connected layer and a softmax operation. The fully connected layer can integrate the features extracted by the previous layers, while the softmax layer is responsible for converting these features into the final classification results. The careful design of this process makes the TMGpCN algorithm perform well in the modulation recognition task, especially in the case of limited labeled samples, which can effectively improve the recognition accuracy.

[0121] Table 4.1. Training process of the TMGpCN modulation recognition algorithm

[0122]

[0123] To evaluate the effectiveness of the method we proposed, experiments were conducted using the publicly available dataset RML2016.10a, which is a standard benchmark for evaluating the performance of algorithms in the field of modulation recognition. Other information supplementing the RML2016.10a dataset is shown in Table 4.2. The RML2016.10a dataset contains a large number of samples with rich styles and is still a commonly used dataset for current deep learning AMC.

[0124] Table 4.2. RML2016.10a dataset

[0125]

[0126]

[0127] This experiment was evenly selected under various modulation types and various signal-to-noise ratios, with 70% used for training, 15% for validation, and the remaining 15% for testing. In the semi-supervised experiment, 20%, 40%, 60%, and 80% of the training set were respectively selected as labeled data, and the rest were used as unlabeled data. For the fairness of the experiment and the reliability of the results, the same number of samples were randomly selected from each signal type at each signal-to-noise ratio level. This experiment was conducted under the Pytorch framework and used an NVIDIA Quadro T2000 GPU. The experimental model adopted an end-to-end training mode and an Adam optimizer, with the initial learning rate set to 0.001.

[0128] (1) Training loss

[0129] As shown in Figure 5 , the loss value tracking during the training process of the GCN model is shown, including supervised and unsupervised loss components. The hyperparameters of these losses are respectively set as 1 for the supervised loss and 0.1 for the unsupervised loss. Through the experiment, it was observed that the supervised loss reached convergence after approximately 135 iterations, and its performance on the validation set presented a relatively smooth curve, indicating that the model learned effective feature representations during the training process and had good generalization ability.

[0130] At the 185th iteration, the unsupervised model also achieved convergence. Its performance on the validation set was excellent, indicating that the model had strong fitting ability. The improvement of the fitting ability means that the model can not only learn the patterns in the labeled data but also capture the implicit structure in the data through the unsupervised learning component, which is crucial for improving the recognition accuracy of the model under limited labeled samples. As the training progresses, the accumulation of local features leads to a significant superposition of local features, resulting in a difference between the reconstructed features and the original features. Therefore, a small fluctuation in the unsupervised loss of the validation set occurred. This variability enhanced the diversity of the extracted features, thus enhancing the feature extraction ability of the network.

[0131] (2) Ablation experiment of TMGCN

[0132] To test the feature extraction ability of the Transformer-based encoder, ablation experiments were conducted, involving the selection of the embedding network and the choice of attention weights.

[0133] To evaluate the feature mapping effect of the Transformer, a comparative analysis was carried out with several traditional feature embedding networks, including linear embedding, single-layer CNN embedding, single-layer LSTM embedding, and hybrid CNN-LSTM embedding methods. Among these methods, linear embedding showed slightly lower performance. At the same time, it was found that combining CNN or LSTM and small convolutional kernels was beneficial to enhancing the feature extraction ability of the network. As shown inFigure 6 As shown, for the recognition accuracy of different embedding networks at different signal-to-noise ratios, the average accuracy of the TMGpCN model reached 58.2%, which is 1.08% higher than that of the CNN-LSTM embedding network.

[0134] According to the attention weight distribution, relatively important edges are selected in corresponding proportions, and the remaining edges and nodes are used to construct a more simplified graph and matrix to achieve the masking function of signal sampling. As Figure 7 shown, the recognition accuracy at various masking rates under different signal-to-noise ratio levels. The TMGpCN model shows almost the best recognition performance at a masking rate of 10%. Even when 20% of the edges are masked, the TMGpCN still maintains a recognition rate of 52.2%. Although higher masking rates lead to a more significant loss of accuracy. The experiment confirms that selecting the largest distribution point can reduce the computational overhead while achieving almost optimal recognition performance.

[0135] After 10% masking of 8PSK and PAM4 signals at 10 dB, their remaining sampling points are as Figure 8 shown.

[0136] To verify the effect of the p-Laplacian matrix on optimizing performance, a comparative experiment was conducted before and after optimizing the p-Laplacian formula. The experimental results are recorded in Table 4.3, showing the recognition rates of the two models under different p parameter values and different proportions of labeled samples. From the data in the table, it can be seen that when the p parameter value is 2, the graph p-Laplacian performs the same as the traditional Laplacian operator. When the p parameter value is adjusted to 1.5, the recognition accuracy of the TMGpCN model exceeds that of the TMGCN model in most cases of labeled sample proportions, except when the labeled samples account for 60% of the data. By adjusting the p parameter value, the feature extraction ability of the model is optimized, thereby improving the recognition accuracy of the model.

[0137] In addition, as Figure 9 shown, the comparison of the recognition rates of the two algorithms at different labeling rates, and at a 100% labeling rate, the confusion matrices of the two methods are as Figure 10 shown. This further confirms the performance of the TMGpCN model under different proportions of labeled samples. Especially in the case of a high proportion of labeled samples, the TMGpCN model demonstrates stronger classification ability.

[0138] Through experiments, it is verified that optimizing the model with p-Laplacian-based GCN is an effective method. By enhancing the feature extraction ability of GCN, the overall recognition accuracy of the model is improved. This method is particularly suitable for dealing with small-sample modulation recognition problems and can achieve efficient modulation recognition with limited labeled samples, providing an effective solution to the problem of insufficient labeled data in practical applications.

[0139] Table 4.3 Comparison of recognition rates of TMGCN and TMGpCN models with different parameters

[0140]

[0141]

[0142] (3) Comparison with other semi-supervised models

[0143] To evaluate the comparative recognition performance of various semi-supervised algorithms, experiments with different sizes of labeled samples were conducted. The recognition accuracy was averaged at labeling rates of 10%, 20%, 30%, 40%, 50%, and 100%. The proposed TMGpCN was compared with CGDNet, MCNet, CLDNN, and TcssAMR, as Figure 11 shown. The results show that TMGpCN is significantly superior to other semi-supervised methods in terms of recognition accuracy. Even at a labeling rate of 50%, TMGpCN achieved a recognition accuracy of 0.536, exceeding the peak recognition rates of MCNet and TcssAMR. When the weights of signal sampling points were masked, the model maintained robust performance in semi-supervised learning. At a masking rate of 10%, TMGpCN outperformed the other four algorithms in all cases where the labeled amount was greater than 50%, with a performance loss of only 0.8%.

[0144] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present application, "a plurality" means two or more, unless otherwise specifically defined.

[0145] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0146] Other embodiments of the present application will be readily envisioned by those skilled in the art after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any variations, uses, or adaptations of the present application, which follow the general principles of the present application and include known common general knowledge or conventional technical means in the technical field not disclosed in the present application.

Claims

1. A semi-supervised GCN modulation recognition method based on Transformer masks, characterized in that, It includes the following steps: Establish a semi-supervised GCN model with a Transformer mask, and use the multi-head attention mechanism of the Transformer to calculate the weight distribution between sampling points in the signal sequence to construct an I / Q signal graph; Use a graph convolutional network based on p-Laplacian to update the node features in the I / Q signal graph structure; The semi-supervised GCN model with a Transformer mask trains the labeled signals and unlabeled signals through semi-supervised supervised loss and unsupervised loss; The fully connected layer combines local features into global features for classification decision-making, and the softmax layer converts the features output by the fully connected layer into the final classification result, classifying the signals into different modulation types.

2. The semi-supervised GCN modulation recognition method based on Transformer masking according to claim 1, wherein In the step of using the multi-head attention mechanism of the Transformer to calculate the weight distribution between sampling points in the signal sequence to construct an I / Q signal graph, it further includes: Preprocess the original discrete signals of the labeled signals and unlabeled signals into I / Q vectors, and combine the I / Q sampling points over time to obtain a signal sequence; Use the multi-head attention mechanism of the Transformer to calculate the weight distribution between sampling points in the signal sequence to construct a fully connected undirected I / Q graph, and convert the I / Q vectors into node and edge information in the graph structure.

3. The semi-supervised GCN modulation recognition method based on Transformer masking according to claim 2, wherein In the step of using the multi-head attention mechanism of the Transformer to calculate the weight distribution between sampling points in the signal sequence to construct an I / Q signal graph, it further includes: By selecting the largest weight and its corresponding node in the fully connected undirected I / Q graph, represent the original signal as a non-fully connected graph.

4. The semi-supervised GCN modulation recognition method based on Transformer masking according to claim 2, wherein The expression of the signal sequence is: x = {[I0, Q0], [I1, Q1],..., [I N-1 , Q N-1}.

5. The semi-supervised GCN modulation recognition method based on Transformer masking according to claim 1, characterized in that, The expression of the multi-head attention mechanism is: where, w i,j is the sum of K normalized multi-head dot-product attentions between nodes i and j, attention k (x i , x j ) is the k-th attention.

6. The semi-supervised GCN modulation recognition method based on Transformer masking according to claim 1, characterized in that In the step of using a graph convolutional network based on p-Laplacian to update the node features in the I / Q signal graph structure, it further includes: Calculate the degree matrix according to the adjacency matrix of the I / Q signal graph; Calculate the p-Laplacian matrix according to the degree matrix; Calculate the first-order approximation of the p-Laplacian matrix to update the node features in the I / Q signal graph structure.

7. The semi-supervised GCN modulation recognition method based on Transformer masking according to claim 6, wherein, In the calculation of the first-order approximation of the p-Laplacian matrix, the expression of the definition of the first-order convolution of p-Laplacian is: Among them, g θ (L p ) is the graph convolution operation of the p-Laplacian matrix, θ is the parameter in the graph convolution operation, L p is the p-Laplacian matrix, X represents the node feature matrix in the graph, I N represents the identity matrix of size N×N, N is the number of nodes in the graph, λ max is the largest eigenvalue of the p-Laplacian, σ is the activation function, W σ is the weight matrix of each convolutional layer, and the structural information obtained by the graph p-Laplacian is a positive semi-definite matrix.

8. The semi-supervised GCN modulation recognition method based on Transformer masking according to claim 1, wherein The total loss function for training the labeled signals and unlabeled signals with the semi-supervised supervised loss and unsupervised loss is the weighted sum of the supervised loss and the unsupervised loss, and the expression of the total loss function is: Among them, D la represents the set of labeled data, and D un represents the set of unlabeled data. x is the input data, y is the label corresponding to x, and the supervised loss is L s , and the unsupervised loss is L u . α and β represent the weights of the supervised loss and the unsupervised loss, and ψ and φ are the parameters of the encoder and the classifier respectively.

9. The semi-supervised GCN modulation recognition method based on Transformer masking according to claim 8, characterized in that The expression of the supervised loss is: Among them, C is the number of categories of label types, and y c is the true label, and is the predicted label using the Softmax classifier.

10. The semi-supervised GCN modulation recognition method based on Transformer masking according to claim 9, characterized in that, The expression of the unsupervised loss is: where x i and represent the original features and reconstructed features of the signal.

Images