A graph neural network automatic modulation classification method based on feature embedding

Abstract

The present invention discloses a method for automatic modulation classification based on a graph neural network with feature embedding, and belongs to the field of signal processing technology. The method generates multiple enhanced samples by interchanging the positions of sequence points in a received signal, thereby introducing position information into the signal, and effectively expanding the expression of a single modulated signal, providing strong support for node feature embedding. The enhanced samples are then input into multiple feature extraction networks with the same structure, and the obtained feature vectors are used as nodes, and the similarities between feature vectors are used as edges to generate a feature graph structure. Finally, the constructed feature graph structure is input into a graph neural network, and the network is trained to classify the modulation type of the received signal. The method combines the feature extraction capability of a traditional neural network with the topological structure analysis capability of a graph neural network, effectively improving the classification performance of the modulated signal.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing, and relates to the classification of modulation modes of radio signals, and specifically to an automatic modulation classification method of a graph neural network based on feature embedding. Background Art

[0002] The rapid growth of IoT devices has intensified the competition for spectrum resources. In dynamic spectrum management, automatic modulation classification is used to detect the modulation mode of unauthorized users. Dynamic spectrum access (DSA) technology based on cognitive radio (CR) can effectively reduce interference and improve spectrum utilization, and wireless signal modulation classification and identification is a key component of DSA.

[0003] The modulation pattern classification and recognition methods of wireless signals can be divided into two parts: feature extraction and classifier design. Traditional modulation classification mainly designs artificial features based on expert knowledge, and then inputs them into machine learning-based classifiers such as support vector machines to achieve modulation classification. These artificial features mainly include time-frequency domain transformation of signals, high-order statistics, etc., with high computational complexity and difficulty in adapting to different modulation types. Due to the powerful feature extraction capabilities of deep learning, automatic modulation recognition methods based on deep learning are being widely used. Deep neural networks are used to automatically extract features directly from the original signal, and the end-to-end training method simplifies the construction of the classifier while achieving good classification performance.

[0004] Commonly used neural network models include convolutional neural network (CNN), recurrent neural network (RNN) and graph neural network (GNN). Among them, the CNN model can effectively capture local features and spatial structures, so it is widely used in the modulation style classification of mapping IQ signals to images. The RNN model has memory ability and can capture the temporal dependencies in the signal. It can effectively process the timing characteristics of the modulated signal, so it is used in the modulation style classification method based on sequence input. The GNN model is suitable for processing graph structure data and can capture the complex relationship and topological structure between nodes. Compared with directly extracting features from the original IQ signal, the introduction of the graph structure can effectively process the topological information of the modulated signal. However, the conversion speed of the current graph mapping algorithm based on fixed rules is slow, and hyperparameters need to be adjusted to control the graph structure. The graph mapping algorithm based on neural network has strict requirements on the performance of the mapping network, and signals of the same modulation type may produce different graph structures. Therefore, the existing modulation recognition method based on graph neural network is limited by the mapping rules, making it difficult to obtain the graph structure of the modulated signal. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes a graph neural network automatic modulation classification method based on feature embedding. By using multiple different neural networks to extract feature representations from different perspectives, a fixed graph structure is constructed as graph nodes, and then the graph convolutional network is used to fully learn the modulation signal feature information and topological structure to improve the performance of the automatic modulation classification and recognition algorithm.

[0006] A graph neural network automatic modulation classification method based on feature embedding specifically includes the following steps:

[0007] Step 1: Preprocessing

[0008] According to the received modulation signal The signal length randomly selects the switching point , the sequence before the exchange point , The sequence after the exchange point , The positions of the two signals are swapped to obtain a new signal sequence. :

[0009]

[0010] Repeat Selection exchange points, get The modulated signal after data enhancement , .

[0011] Step 2: Feature extraction

[0012] The enhanced signals obtained in step 1 are input into The kth feature embedding network is used to embed the kth feature into the same feature embedding network. Get enhanced samples The feature representation ,in Represents the feature dimension.

[0013] The feature embedding network includes an input layer, four residual blocks and an adaptive pooling layer cascaded in sequence. The input layer includes a one-dimensional convolution of size 7x1 and a maximum pooling of size 3x1. Each residual block includes two one-dimensional convolutions of size 3x1, and the input and output of the residual block are residually connected and then input to the next module. The input of the third residual block first passes through a one-dimensional convolution of size 1x1 and then residually connected with the output. The activation function uses the rectified linear unit ReLU.

[0014] Step 3: Graph Construction

[0015] Will The feature representation output by the feature embedding network is regarded as a graph node, and the graph node set . Consider the feature similarity between nodes as edges and construct a graph structure ,in is the edge set.

[0016] Step 4: Modulation style classification

[0017] The graph structure constructed in step 3 Input into the graph neural network, the signal sequence corresponding to the graph structure The modulation mode is output, and the cross entropy loss function Loss is used to train the graph neural network. The trained graph neural network is used to classify the modulation style of the signal.

[0018] The present invention has the following beneficial effects:

[0019] (1) The feature representations of the original IQ signal from different perspectives are extracted through a neural network and regarded as graph nodes, thereby constructing a graph structure. The feature extraction capability of traditional neural networks and the topological structure analysis capability of graph neural networks are combined, effectively improving the classification performance of modulated signals.

[0020] (2) Randomly swapping the front and back sequence segments of the IQ signal to generate new samples, introducing position information into the modulated signal and effectively expanding the expression of a single modulated signal, providing strong support for node feature embedding. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of the automatic modulation classification method of graph neural network based on feature embedding;

[0022] Figure 2 A schematic diagram of a feature embedding network structure constructed in an embodiment;

[0023] Figure 3 is the classification accuracy of different methods under different signal-to-noise ratios in the embodiment;

[0024] Figure 4 It is the recognition confusion matrix of CNN1D network on different data sets in the embodiment;

[0025] Figure 5 It is the recognition confusion matrix of the CNN2D network on different data sets in the embodiment;

[0026] Figure 6 It is the recognition confusion matrix of the GRU network on different data sets in the embodiment;

[0027] Figure 7 It is the recognition confusion matrix of the LSTM network on different data sets in the embodiment;

[0028] Figure 8 It is the recognition confusion matrix of the MCLDNN network on different data sets in the embodiment;

[0029] Fig. 9 It is the recognition confusion matrix of the AvgNet network on different data sets in the embodiment;

[0030] Fig.10 : This is the recognition confusion matrix of the method in the embodiment on different data sets. DETAILED DESCRIPTION

[0031] The present invention will be further explained below with reference to the accompanying drawings;

[0032] definition It is the zero intermediate frequency received signal obtained after down-conversion:

[0033]

[0034] in, is the transmitted baseband signal, represents the wireless channel, Indicates the carrier frequency offset, Indicates phase offset. It means the mean is 0 and the variance is Additive Gaussian white noise, , is the signal length. The modulation recognition task is to Analyze and infer the transmitted baseband signal The modulation method used can be defined as a multi-classification problem: for a given received signal , giving its corresponding emission signal Modulation type ,in Indicates the number of modulation types.

[0035] Represent the received signal in complex format , whose IQ components are:

[0036]

[0037] in and Respectively The in-phase and quadrature components of and Respectively represent signals The real and imaginary parts of .

[0038] A graph neural network automatic modulation classification method based on feature embedding, such as Figure 1 As shown, first the received signal The original IQ components are preprocessed, and the single modulated signal is enhanced by sequence interchange to generate The next step is to Enhanced samples are input simultaneously Different residual networks are used to extract feature representations in different feature spaces. Then, The feature vectors output by the residual network are regarded as a graph node , considering the similarity between feature vectors as edge , construct feature map ,in Represents a collection of nodes. Represents the edge set. Finally, the constructed feature graph Input the graph neural network to understand the topological relationship of the signal feature graph and the interaction mode of the nodes, so as to obtain the modulation type of the signal. The specific steps are as follows:

[0039] Step 1: Preprocessing

[0040] Compared with directly mapping the IQ signal sequence into a graph, this method can obtain the received modulated signal Perform preprocessing and randomly select exchange points based on signal length , the sequence before the exchange point , The sequence after the exchange point , The positions of the two signals are swapped to obtain a new signal sequence. :

[0041]

[0042] Repeat Selection exchange points, get The modulated signal after data enhancement , , thus converting a single modulated signal Expand to An enhanced sample.

[0043] Step 2: Feature extraction

[0044] The enhanced signals obtained in step 1 are input into From the feature embedding networks Net1~NetK with the same structure, Get enhanced samples The feature representation ,in Represents the feature dimension. Compared with a feature embedding network, using A feature embedding network can obtain feature representations in different feature spaces from a single signal, expand the feature differences of each graph node, and facilitate graph construction.

[0045] The feature embedding network is modified based on ResNet18, such as Figure 2 As shown, it includes a cascaded input layer, 4 residual blocks and an adaptive pooling layer. The input layer includes a one-dimensional convolution of size 7x1 and a maximum pooling of size 3x1. Each residual block includes two one-dimensional convolutions of size 3x1, and the input and output of the residual block are residually connected and then input to the next module. The input of the third residual block first passes through a one-dimensional convolution of size 1x1 and then residually connected with the output. The activation function uses the rectified linear unit ReLU.

[0046] Step 3: Graph Construction

[0047] Will The feature representation output by the feature embedding network is regarded as a graph node, and the graph node set . Calculate the similarity matrix B between graph nodes:

[0048]

[0049] in Represented as a node and nodes The feature similarity of , i, j = 1, 2, ... Since the similarity of the graph nodes themselves is 1, for each row vector B of the matrix B i Divide by the diagonal elements of the row , and use the ReLU function Processing, finally get the similarity matrix :

[0050]

[0051] Similarity matrix Elements in As a node and nodes The similarity matrix is the edge set, thus obtaining the graph structure .

[0052] Direct signal sequence Compared with constructing a graph structure by treating the sampling points as graph nodes, the graph structure constructed by the present application using the feature representation of the signal sequence effectively reduces the number of graph nodes and the complexity of the constructed graph structure, and the fixed fully connected structure reduces the difficulty of the subsequent model to perform modulation classification tasks.

[0053] Step 4: Modulation classification

[0054] The graph structure constructed in step 3 Input into the graph neural network, the signal sequence corresponding to the graph structure The modulation mode is output, and the cross entropy loss function Loss is used to train the graph neural network:

[0055]

[0056] in Indicates signal The modulation method belongs to the category The true probability of Represents the output signal of the graph neural network The modulation method belongs to the category The predicted probability of , M is the batch size of training.

[0057] The graph neural network includes a GraphSAGE layer and a DiffPool layer. The GraphSAGE layer learns the node embedding of the graph structure through neighborhood sampling and feature aggregation. For each node in the layer , according to the connection information between nodes, from its neighbor node set Randomly sample a certain number of neighbor nodes as the sampling point set Then its own characteristics With sampling point set Aggregate the node features in to obtain a new node representation :

[0058]

[0059] in Representation Node The feature set of neighbor nodes, represents the weight matrix, is the activation function. Multiple layers are stacked to improve the current node feature representation through local and neighborhood information.

[0060] The DiffPool layer aggregates the nodes in the graph into smaller subgraphs by learning the hierarchical structure of the graph, and generates a pooled representation of each subgraph, thereby achieving graph dimensionality reduction. The DiffPool layer enables the network to learn how to effectively cluster the nodes in the graph, thereby capturing the hierarchical features of different structures of the graph in each layer.

[0061] In order to verify the effectiveness of this method, two public modulation signal datasets RML2016.10a and RML2016.10b are used for comparative experiments. The RML2016.10a dataset contains 220,000 samples and 11 modulation types, including BPSK, QPSK, 8PSK, 16QAM, 64QAM, BFSK, CPFSK and PAM4, and WBFM, AM-SSB and AM-DSB, respectively. The signal-to-noise ratio of each modulation signal ranges from -20dB to 18dB, with an interval of 2dB, and each signal sample is in the form of IQ dual channels with a length of 128. 80% of the samples are randomly selected from each signal-to-noise ratio of each modulation type in the RML2016.10a dataset as the training set, and the remaining 20% ​​are used as the test set. The dataset RML2016.10b contains 1,200,000 samples and 10 modulation types. The digital modulation types include BPSK, QPSK, 8PSK, 16QAM, 64QAM, BFSK, CPFSK and PAM4, and the analog modulation types include WBFM and AM-DSB. The signal-to-noise ratio range and signal length are the same as those of the RML2016.10a dataset. Samples are randomly selected from them in a 1:1 ratio to form the training set and test set.

[0062] First, based on the above test set and training set, the number of exchange points in this method is discussed. The influence of different training methods on the final classification results was studied, and the accuracy, recall and F1 score were selected as evaluation indicators. During the training process, the Adam optimizer was used, and the learning rate was decayed with a fixed step size. The learning rate was reduced to 50% of the current value every 10 training batches. The initial learning rate and batch size were 0.001 and 256, respectively. All simulation experiments were run on the PyTorch deep learning framework, and the GPU model used was NVIDIA Geforce GTX 1080.

[0063] different The classification results of the two modulation signal data sets are shown in Table 1.

[0064] Table 1

[0065]

[0066] From Table 1, we can see that with With the increase of , the classification performance of this method in both datasets first increases and then decreases, and when When it is equal to 5, the classification performance reaches the best.

[0067] Secondly, using the same data and training method, the classification effect difference between the present application and the commonly used method in the prior art is compared. In step 1, the present method selects 5 switching points for each received signal to generate enhanced samples. The classification results of different methods on the data sets RML2016.10a and RML2016.10b are shown in Tables 2 and 3. The classification accuracy under different signal-to-noise ratios is shown in Tables 2 and 3. Figure 3 shown.

[0068] Table 2

[0069] method Accuracy F1 score Recall CNN1D 60.07% 0.6193 0.6007 CNN2D 54.31% 0.5534 0.5430 GRU 61.32% 0.6280 0.6095 LSTM 59.86% 0.6145 0.5986 MCLDNN 61.79% 0.6390 0.6179 AvgNet 61.98% 0.6353 0.6198 This method 62.43% 0.6465 0.6242

[0070] Table 3

[0071] method Accuracy F1 score Recall CNN1D 62.93% 0. 6303 0.6293 CNN2D 56.88% 0. 5653 0.5688 GRU 63.25% 0. 6324 0.6325 LSTM 54.86% 0. 5426 0.5487 MCLDNN 64.11% 0. 6435 0.6411 AvgNet 63.84% 0. 6429 0.6384 This method 65.06% 0. 6520 0.6506

[0072] On the RML2016.10a dataset, from Table 2 and Figure 3As shown in (a), firstly, the accuracy, F1 score and recall rate of this method are better than those of CNN1D and CNN2D methods, and the classification accuracy at each signal-to-noise ratio is also higher than that of CNN1D and CNN2D. Secondly, compared with GRU, LSTM and fusion method MCLDNN, this method also has obvious advantages. In particular, compared with the fusion method MCLDNN, the three indicators are improved by 0.63%, 0.0075 and 0.0063 respectively. It is worth noting that modeling the modulated signal in a time series manner is more advantageous than the CNN model, but the data enhancement method of sequence exchange proposed in this method relies on randomly selected exchange points to add position information to the sample to make up for the lack of time information of the CNN model. Finally, comparing this method with the GNN model-based method AvgNet, it can be seen that this method is significantly better than the AvgNet method, and its three evaluation indicators are improved by 0.45%, 0.0112 and 0.0045 respectively. Compared with directly using sampling points as graph nodes to construct a graph structure, the proposed method of constructing a graph structure based on feature embedding can more effectively characterize the modulation signal and obtain better classification performance. In general, compared with the existing modulation classification methods of various network types, this method has better classification performance. For the RML2016.10b dataset with a larger sample size, from Table 3 and Figure 3 The same conclusion can be drawn from (b) in Figure 1, and the performance is more significantly improved compared to RML2016.10a. In particular, compared with the AvgNet method, the three evaluation indicators of this method are improved by 1.22%, 0.0091, and 0.0122, respectively. This shows that this method has a strong ability to generalize data sets.

[0073] In order to further compare the classification performance of different methods, the confusion matrices of the six methods in three categories for signal modulation recognition of the datasets RML2016.10a and RML2016.10b are given respectively. Figure 4~Figure 9 As shown in Figure 2, the confusion matrix of this method for signal modulation recognition of the datasets RML2016.10a and RML2016.10b is as follows: Fig.10 As shown. Figure 4~Figure 10 As can be seen from (a) in the figure, on the RML2016.10a dataset, all methods tend to misclassify other modulation types as AM-SSB, and CNN1D and CNN2D tend to confuse QAM16 and QAM64 with each other; while this method effectively improves the classification accuracy of AM-SSB. Figure 4~Figure 10As can be seen from (b) in the figure, on the RML2016.10b dataset, all methods easily misclassify WBFM as the modulation type AM-DSB, GRU, MCLDNN, AvgNet and FENET are more likely to distinguish QAM16 and QAM64, and the recognition accuracy of this method for ten modulation signals is higher than that of GRU and LSTM, the recognition accuracy of nine modulation signals is higher than that of AvgNet, and the recognition accuracy of six modulation signals is higher than that of MCLDNN; it effectively improves the correct classification performance of AM-SSB, BPSK, PAM4 and QPSK.

Claims

1. A graph neural network automatic modulation classification method based on feature embedding, characterized by: The original IQ component of the received signal r(n) is preprocessed, a data point is randomly selected, the sequence before and after the point is swapped, and data enhancement is performed on a single modulated signal. This is repeated K times to generate K enhanced samples. Then, the K enhanced samples are simultaneously input into K different feature extraction networks. Then, the feature vector output by each feature extraction network is regarded as a graph node v k , calculate the similarity matrix B between graph nodes: in Represented as node v i and node v j The feature similarity of ; each row vector of the similarity matrix B is divided by the diagonal element of the row, and processed with the ReLU function ρ(·), and finally the similarity matrix E is obtained: The element e in the similarity matrix E ij As node v i and node v j The edges of the feature graph G = (V, E), where V = {v1, ..., v K } represents a node set, E = {e ij ,1≤i,j≤K} represents the edge set; finally, the constructed feature graph G=(V,E) is input into the graph neural network, and the modulation mode of the signal sequence r(n) corresponding to the graph structure is output; The graph neural network includes a GraphSAGE layer and a DiffPool layer; the GraphSAGE layer learns the node embedding of the graph structure through neighborhood sampling and feature aggregation; in the first layer, for each node vi, according to the connection information between nodes, the neighbor node set Neighbor (v i ) randomly samples a certain number of neighbor nodes as the sampling point set G(v i ), and then its own characteristics With the sampling point set G(v i ) to obtain a new node representation in Represents node v i The feature set of neighbor nodes, W represents the weight matrix, σ(·) is the activation function; multiple layers are stacked to improve the current node feature representation through local and neighborhood information; The DiffPool layer aggregates the nodes in the graph into smaller subgraphs by learning the hierarchical structure of the graph and generates a pooled representation of each subgraph, thereby achieving graph dimensionality reduction. The DiffPool layer enables the network to learn how to effectively cluster the nodes in the graph, thereby capturing the hierarchical features of different structures of the graph in each layer.

2. The method for automatic modulation classification based on graph neural network with feature embedding according to claim 1, characterized in that: The received modulated signal According to the signal length L, randomly select the exchange point n = i k ∈[5,L-5], the sequence before the exchange point I(0),…,I(i k -1)、Q(0),…,Q(i k -1) and the sequence I(i k ),…,I(L-1),Q(i k ),…,Q(L-1) are swapped to obtain a new signal sequence R' k : K switching points are repeatedly selected, thereby expanding a single modulated signal r(n) into K enhanced samples.

3. The method for automatic modulation classification based on graph neural network with feature embedding according to claim 1, characterized in that: The feature extraction network is a residual network.

4. The method for automatic modulation classification based on graph neural network with feature embedding according to claim 3, characterized in that: The residual network includes an input layer, four residual blocks and an adaptive pooling layer cascaded in sequence; the input layer includes a one-dimensional convolution layer of size 7x1 and a maximum pooling layer of size 3x1; each residual block includes two one-dimensional convolution layers of size 3x1, and the input and output of the residual block are residually connected and then input into the next module.

5. The method for automatic modulation classification based on graph neural network with feature embedding according to claim 4, characterized in that: The input of the third residual block first undergoes a one-dimensional convolution of size 1x1 and then performs a residual connection with the output.

6. The method for automatic modulation classification based on graph neural network with feature embedding according to claim 1, characterized in that: The cross entropy loss function Loss is used to train the graph neural network: where y i represents the true probability that the modulation mode of the signal r(n) belongs to category i, p i (R) represents the predicted probability that the modulation mode of the graph neural network output signal r(n) belongs to category i, and M is the batch size of training.

7. The method for automatic modulation classification based on graph neural network with feature embedding according to claim 1, characterized in that: For the training of graph neural networks, the Adam optimizer is used, and the learning rate is decayed with a fixed step size. The learning rate is reduced to 50% of the current value every 10 training batches; the initial learning rate and batch size are 0.001 and 256, respectively.

8. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1 to 7.

Images