Radio signal modulation identification method based on multi-scale graph neural network

Abstract

The invention discloses a radio signal modulation identification method based on a multi-scale map neural network. The radio signal modulation identification method specifically comprises the following steps: acquiring time sequence data of a radio signal and splitting the time sequence data into two groups of data of an I channel and a Q channel with the same scale; converting the signal into channel data under a plurality of different scales by using a convolutional network; processing channel data of different scales by using a long-short-term memory network to construct a weighted adjacent matrix, and obtaining views of signals under different scales; and constructing a multi-scale graph neural network model by using graph convolution and a pooling layer to realize automatic modulation identification of radio signals. According to the technical scheme, the problem that in the prior art, time sequence feature extraction is insufficient in the graph structure mapping process is solved, and meanwhile the accuracy of signal modulation recognition is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of signal recognition, and in particular to a radio signal modulation recognition method based on a multi-scale graph neural network. Background Art

[0002] With the rapid development of wireless communication technology and the large-scale commercial use of 5G, the communication environment is becoming increasingly dense, and signal modulation styles are emerging in an endless stream. Therefore, automatically and accurately identifying the modulation mode of wireless signals is of great significance to the application field of radio communication. Automatic modulation identification is to identify the modulation type of the received signal under unknown or lacking prior information. This is the key for the receiver to achieve correct signal demodulation in the wireless communication system. At present, signal modulation identification plays an important role in cognitive radio and wireless communication scenarios. In the military field, signal modulation identification is widely used in scenarios such as electronic support, electronic reconnaissance and electronic attack; in the civilian field, it is mainly used in radio scenarios such as spectrum sensing and management.

[0003] Traditional automatic modulation recognition methods are mainly divided into two categories: the likelihood ratio method based on decision theory and the feature extraction method based on statistics. The likelihood ratio method based on decision theory uses rigorous mathematical derivation and calculation to obtain the optimal solution in the Bayesian sense using the likelihood function. However, this type of method is highly dependent on prior knowledge and has the problem of high complexity, so it is difficult to be widely applied to practical problems. The feature extraction method based on statistics requires extracting representative features from manually calculated training samples and then identifying them through a classifier. Since this algorithm is overly dependent on manually calculated features and is greatly affected by the environment and has poor generalization ability, it is often difficult to achieve an ideal recognition rate in practical applications.

[0004] In recent years, deep learning technology has made breakthroughs in many application fields, especially in natural language processing and image processing. Therefore, automatic modulation recognition methods based on deep learning technology have been proposed one after another and have achieved good results. The deep learning-based method mainly completes the signal modulation recognition task by converting the time domain signal into a time-frequency image and then using a deep neural network model to implement classification. However, this type of method mainly focuses on the extraction of local features and cannot fully capture the global characteristics of the signal. At the same time, the deep learning model cannot capture the potential spatiotemporal dependencies between time series signals, which is an important limitation for modulation recognition tasks that require understanding the changes of signals over time.

[0005] Recently, graph neural network technology has shown significant advantages in time series data processing. It can effectively capture the spatiotemporal dependencies of time series signals and provide a new perspective for signal modulation recognition. Therefore, based on this feature, some methods propose to map radio signals into graph structures and use graph neural networks to classify the obtained visual graphs to achieve automatic modulation recognition tasks. This type of method has promoted the development of signal processing technology to a certain extent. However, the existing graph mapping methods have great limitations, that is, they rely on fixed rules to construct graph structures and ignore the rich channel information of the signal itself. This processing method limits the potential of graph neural networks in the field of signal processing. At the same time, existing technical means cannot fully utilize the global and local information of the signal at the same time, especially under low signal-to-noise ratio conditions. Summary of the invention

[0006] The purpose of the present invention is to provide a radio signal modulation recognition method based on a multi-scale graph neural network to solve the limitations of existing signal modulation recognition methods based on graph neural networks in terms of graph structure mapping and insufficient information utilization.

[0007] To achieve the above object, the present invention adopts the following technical solutions: A radio signal modulation recognition method based on a multi-scale graph neural network is characterized by comprising the following steps: Step S1: Acquire radio signal data and split it into co-directional channel data I and orthogonal channel data Q of the same scale; Step S2: The same-direction channel data I and the orthogonal channel data Q are respectively subjected to two layers of one-dimensional convolution operations to obtain channel data of two different scales. , and , ,in , is the output of the first convolution of the same-direction channel data I and the orthogonal channel data Q. , They are the outputs of the second convolution layer for the same-direction channel data I and the orthogonal channel data Q respectively; Step S3: using the long short-term memory network to map the channel data of different scales into two-dimensional adjacency matrices, thereby constructing visualization graphs of different scales; Step S4: Construct a multi-scale graph neural network model by combining graph convolution and pooling layers, taking visibility graphs of different scales as input to obtain feature vectors of different scales; Step S5: concatenate the characteristic vectors of different scales and output the signal modulation type through the classifier, and use the signal data set to train the model; Step S6: Input the signal to be identified into the trained model to predict the modulation type of the signal.

[0008] On the basis of adopting the above technical solutions, the present invention may also adopt the following further technical solutions, or use these further technical solutions in combination: In the step S1, the radio signal data is obtained and split into co-directional channel data I and orthogonal channel data Q of the same scale, including the following steps: Step S101: For each radio signal sample, it is expressed as ,in n is the scale of the channel data, and the shape of the signal sample is ; Step S102: transform each shape into The signal samples are split into The same-direction channel data I and the orthogonal channel data Q are respectively expressed as and .

[0009] In step S2, the same-direction channel data I and the orthogonal channel data Q are respectively subjected to two layers of one-dimensional convolution operations to obtain channel data of two different scales, including the following steps: Step S201: Send the same-direction channel data I and the orthogonal channel data Q to the first one-dimensional convolution layer to obtain the same-direction channel data of the intermediate scale Hezhengtong road data , respectively expressed as and ; Step S202: The same-direction channel data of the intermediate scale Hezhengtong road data Input the second-dimensional convolutional layer to obtain smaller-scale same-direction channel data Hezhengtong road data , respectively expressed as and .

[0010] In step S3, the channel data of different scales are mapped to the adjacency matrix using the long short-term memory network, so as to construct the visualization graphs of different scales, including the following steps: Step 301: For channel data of different scales, the output features of different scales are calculated using the long short-term memory network, which is specifically expressed as:

[0011] in, Represents the input channel data, including , and , , represents the output of the long short-term memory network, and its scale is , s is an adjustable parameter used to control the number of eigenvalues, is the maximum preset value, is the output feature of the long short-term memory network; Step 302: Use the ReLU activation function to transform the different scale output features of the long short-term memory network to create a The feature matrix of the transformed image is transformed, and the transformation results are arranged along the diagonal to obtain the weighted adjacency matrix corresponding to the output features of different scales; Step 303: The weighted adjacency matrix corresponding to the output features of different scales is input into the graph structure mapper in sequence to obtain the weighted adjacency matrix at each scale, and finally obtain the visible graph of different scales.

[0012] In step S4, a multi-scale graph neural network model is constructed by combining graph convolution and pooling layers, and visual graphs of different scales are used as input to obtain feature vectors of different scales, including the following steps: Step 401: construct a multi-scale graph neural network model consisting of alternately stacked graph convolution and pooling layers; Step 402: Input the visibility graphs of different scales into the multi-scale graph neural network model to obtain corresponding feature vectors.

[0013] In step S5, the characteristic vectors of different scales are concatenated and the signal modulation type is output through a classifier, and the signal data set is used for model training, which includes the following steps: Step 501: concatenate feature vectors of different scales to obtain a final feature vector; Step 502: The concatenated feature vectors are sequentially reduced in dimension through two layers of fully connected networks, and then output the signal modulation type through a SoftMax classifier; Step 503: Use the signal data set to train the multi-scale graph neural network model.

[0014] Compared with the prior art, the present invention has the following characteristics: Scalability: The graph structure method provided by the present invention fully combines and utilizes the time information and spatial information in the information through the long short-term memory network (LSTM), thereby extracting a more discriminative graph structure representation. At the same time, the method can control the number of eigenvalues ​​(edges) by adjusting parameters during the mapping process, effectively solving the problem that the existing mapping methods are inflexible based on fixed rules. Based on this, a recognition model based on a graph neural network is constructed, and a multi-layer graph neural network is used to perform deep calculations on channel features. This method can more effectively capture the structure and information of graph data. The method of the present invention can change the mapping method of the graph structure as needed, which increases flexibility and adaptability.

[0015] Accuracy: In the real-world automatic modulation recognition of radio signals, the acquired radio signals inevitably contain some irrelevant information or noise, which affects the accuracy of the recognition results. The present invention can conduct targeted learning through multi-scale structural features to minimize noise interference, thereby effectively improving the accuracy of recognition. Compared with the existing signal modulation recognition technology, the recognition accuracy rate is improved by 1.6% under low signal-to-noise ratio.

[0016] In summary, the present invention uses a visual graph mapping model based on LSTM to complete the effective mapping from channel signals to graph structures on the basis of multi-scale channel data transformation of radio signals, and more flexibly obtains graph structure representations of different scales, and then realizes modulation signal recognition through a graph neural network classifier. The method of the present invention can flexibly adjust the graph structure mapping process according to different tasks and environments, while improving the accuracy of recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a radio signal modulation recognition method based on a multi-scale graph neural network according to an embodiment of the present invention.

[0018] Figure 2 Schematic diagram of multi-scale feature extraction according to an embodiment of the present invention.

[0019] Figure 3 The figure is a schematic diagram of a visual graph mapping process based on LSTM according to an embodiment of the present invention.

[0020] Figure 4 This is a schematic diagram of the structure of a multi-scale graph neural network model based on graph convolution and pooling layers according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The technical solution provided by the present invention will be further described below in conjunction with the accompanying drawings. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0022] The present invention is further described in detail below in conjunction with the accompanying drawings: The present invention provides a radio signal modulation recognition method based on a multi-scale graph neural network. The method mines the relationship between the global and local signals through multi-scale transformation to obtain a richer feature representation; then, a graph structure mapping method based on LSTM is used to better and more flexibly convert the multi-scale channel information into a graph structure form, and a multi-scale graph neural network is designed to extract the features of the multi-scale visible graph, thereby constructing a signal modulation recognition model based on the graph neural network.

[0023] Reference Figure 1 As shown, the radio signal modulation recognition method based on multi-scale graph neural network of the present invention comprises the following steps: Step S1: Obtain radio signal time series data and split it into The two channel data of the same direction channel data I and the orthogonal channel data Q are specifically comprised of the following steps: Step S101: For each radio signal sample, it is expressed as ,in, n is the scale of the channel data, and the shape of the signal sample is ; Step S102: transform each shape into The signal samples are split into The same-direction channel data I and the orthogonal channel data Q are respectively expressed as and , as the first layer channel data (scale 1), can be found in Figure 2 .

[0024] Step S2: The same-direction channel data I and the orthogonal channel data Q are respectively subjected to two layers of one-dimensional convolution operations to obtain channel data of different scales. , and , ,in , is the output after the first layer of convolution, , They are the outputs after the second layer of convolution respectively.

[0025] Step S3: Use LSTM to map different channel data to adjacency matrices respectively, so as to construct visualization graphs of different scales.

[0026] Step S4: Combine graph convolution and pooling layers to build a multi-scale graph neural network model, take visibility graphs of different scales as input, and obtain feature vectors of different scales.

[0027] Step S5: concatenate the feature vectors of different scales and output the signal modulation type through the classifier, and use the signal data set to train the model. Step S6: input the signal to be identified into the trained model to predict the modulation type of the signal.

[0028] Reference Figure 2 , in step S2, the scale is The same-direction channel data I and the orthogonal channel data Q are respectively subjected to two layers of one-dimensional convolution operations to obtain channel data of different scales, further comprising: S201: Send the same-direction channel data I and the orthogonal channel data Q to the first one-dimensional convolution layer to obtain the channel data of the second layer and (Scale 2), respectively expressed as and .

[0029] S202: Layer 2 channel data and Send it to another one-dimensional convolutional layer to get the channel data of the third layer and channel data (Scale 3), respectively expressed as and ; Reference Figure 3 , in the step S3, using LSTM to map different channel data to an adjacency matrix, thereby constructing visualization graphs of different scales, further comprising; S301: For channel data of different scales, the output features of different scales are calculated using the long short-term memory network, which is specifically expressed as:

[0030] in, Represents the input channel data, including , and , , represents the output of the long short-term memory network, and its scale is , s is an adjustable parameter used to control the number of eigenvalues, is the maximum preset value, is the output feature of the long short-term memory network.

[0031] S302: Use the ReLU activation function to transform the output features of different scales of the long short-term memory network. The transformation results are expressed as , ,like Figure 3 As shown. Create a The characteristic matrix of , arrange the transformation results along the diagonal to obtain the corresponding weighted adjacency matrix, such as Figure 3 As shown in the grayscale matrix, different grayscales represent different weights.

[0032] S303: Input the weighted adjacency matrices corresponding to the output features of different scales into the graph structure mapper in sequence to obtain the weighted adjacency matrices at each scale, and finally obtain the visible graphs of different scales; Figure 3 As shown, circles represent graph nodes and straight lines represent edges between nodes.

[0033] Reference Figure 4 In step S4, a multi-scale graph neural network model is constructed by combining graph convolution and pooling layers, and visual graphs of different scales are used as input to obtain feature vectors of different scales, further comprising: S401: construct a multi-scale graph neural network classification model consisting of three graph convolution layers and two pooling layers stacked alternately; Figure 4 The figure shows the process of channel data after graph mapping, and then processing through a single graph convolution and pooling layer, and finally obtaining the feature representation corresponding to the channel data.

[0034] S402: Inputting visibility graphs of different scales into the multi-scale graph neural network model to obtain corresponding feature representations.

[0035] In the step S5, the characteristic vectors of different scales are concatenated and the signal modulation type is output through a classifier, and the model training is performed using the signal data set, further comprising: S501: concatenate feature vectors of different scales to obtain a final feature vector; S502: The concatenated feature vectors are sequentially reduced in dimension through two layers of fully connected networks, and then output the signal modulation type through a SoftMax classifier; S503: Train the multi-scale graph neural network model using the signal dataset.

[0036] Specifically, in the multi-scale graph neural network model, the graph convolution layer and the pooling layer are implemented using GraphSAGE and Diffpool respectively. The specific process is: first, the visible graph is input into the GraphSAGE layer to calculate the embedded features of each node in the visible graph, and then the Diffpool layer performs a pooling operation to obtain a new embedded feature. It should be pointed out that this process can be implemented using a similar graph structure learning method.

[0037] The effectiveness of the technology proposed in the present invention is verified by a specific example below, and is compared with the existing similar methods. The specific steps are as follows: (1) Radio signal dataset RML2016.10b is used to realize radio signal modulation recognition. The RML2016.10B dataset is generated by GNURadio and contains 10 modulation signals, including 8 digital modulation signals BPSK, QPSK, 8PSK, 16QAM, GFSK, CPFSK, PAM4, QAM64, and 2 analog modulation signals WBFM and AM-DSB. The signal-to-noise ratio of each modulation signal ranges from -20dB to 18dB. The dataset contains 1.2 million signal samples, and each signal sample consists of 2×128 I / Q dual-channel signals. For each signal-to-noise ratio, there are 6000 samples for each modulation type. In order to simulate the real communication environment, additive white Gaussian noise, which may cause channel damage, is added to the dataset to form a signal dataset; (2) The data set is randomly divided into a training set (50%) and a test set (50%), that is, 3,000 samples of each modulation signal are randomly selected for training and 3,000 for testing under a single signal-to-noise ratio. Signal labels are uniformly processed using one-hot encoding; (3) Use the Adam optimizer to optimize the training process of the multi-scale graph neural network model until the iteration stop condition is reached and the converged optimal graph classification model is obtained; (4) Input the test data into the trained multi-scale graph neural network model to predict the label, and use the recognition accuracy, recall rate and F1 score as evaluation indicators to evaluate the results. Compared with the existing technology, including 3 automatic modulation recognition models based on deep neural networks (CNN, LSTM and MCLDNN) and 2 models based on graph neural networks (VG, LPVG), the comparison results are shown in Table 1.

[0038] Table 1 Performance comparison of the proposed method and other methods on signal modulation recognition tasks method Accuracy F1 score Recall CNN 0.551 0.556 0.553 LSTM 0.631 0.631 0.631 MCLDNN 0.639 0.641 0.640 VG 0.598 0.601 0.596 LPVG 0.616 0.614 0.616 Method of the present invention 0.649 0.651 0.649 By comparison, it is found that the performance of the method provided by the present invention is significantly better than other methods, especially under low signal-to-noise ratio. The method of the present invention achieves better performance mainly because: on the one hand, the method of the present invention adds multi-scale changes to deeply integrate channel information, better explore the relationship between the global and the local, and obtain richer feature representations; on the other hand, the LSTM-based graph structure mapping method is used to better capture the spatial and temporal information of channel data. At the same time, the number of edges in the graph structure can be controlled by hyperparameters, and different types of data can be flexibly responded to to obtain better results.

[0039] The following is a system embodiment of the present invention, which can be used to implement the method embodiment of the present invention. For details not described in the system embodiment, please refer to the method embodiment of the present invention.

[0040] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A radio signal modulation recognition method based on a multi-scale graph neural network, characterized in that: The following steps are involved: Step S1: Acquire radio signal data and split it into co-directional channel data I and orthogonal channel data Q of the same scale; Step S2: The same-direction channel data I and the orthogonal channel data Q are respectively subjected to two layers of one-dimensional convolution operations to obtain channel data of two different scales. , and , ,in , is the output of the first convolution of the same-direction channel data I and the orthogonal channel data Q. , They are the outputs of the second convolution layer for the same-direction channel data I and the orthogonal channel data Q respectively; Step S3: using the long short-term memory network to map the channel data of different scales into two-dimensional adjacency matrices, thereby constructing visualization graphs of different scales; Step S4: Construct a multi-scale graph neural network model by combining graph convolution and pooling layers, taking visibility graphs of different scales as input to obtain feature vectors of different scales; Step S5: concatenate the characteristic vectors of different scales and output the signal modulation type through the classifier, and use the signal data set to train the model; Step S6: Input the signal to be identified into the trained model to predict the modulation type of the signal.

2. The radio signal modulation recognition method based on multi-scale graph neural network according to claim 1 is characterized in that: In the step S1, the radio signal data is obtained and split into co-directional channel data I and orthogonal channel data Q of the same scale, including the following steps: Step S101: For each radio signal sample, it is expressed as ,in, n is the scale of the channel data, and the shape of the signal sample is ; Step S102: transform each shape into The signal samples are split into The same-direction channel data I and the orthogonal channel data Q are respectively expressed as and .

3. The radio signal modulation recognition method based on multi-scale graph neural network according to claim 1 is characterized in that: In step S2, the same-direction channel data I and the orthogonal channel data Q are respectively subjected to two layers of one-dimensional convolution operations to obtain channel data of two different scales, including the following steps: Step S201: Send the same-direction channel data I and the orthogonal channel data Q to the first one-dimensional convolution layer to obtain the same-direction channel data of the intermediate scale Hezhengtong road data , respectively expressed as and ; Step S202: The same-direction channel data of the intermediate scale Hezhengtong road data Input the second-dimensional convolutional layer to obtain smaller-scale co-directional channel data Hezhengtong road data , respectively expressed as and .

4. The radio signal modulation recognition method based on multi-scale graph neural network according to claim 1 is characterized in that: In step S3, the channel data of different scales are mapped to the adjacency matrix using the long short-term memory network, so as to construct the visualization graphs of different scales, including the following steps: Step 301: For channel data of different scales, the output features of different scales are calculated using the long short-term memory network, which is specifically expressed as: ， in, Represents the input channel data, including , and , , represents the output of the long short-term memory network, and its scale is , s is an adjustable parameter used to control the number of eigenvalues, is the maximum preset value, is the output feature of the long short-term memory network; Step 302: Use the ReLU activation function to transform the different scale output features of the long short-term memory network to create a The feature matrix of is arranged along the diagonal to obtain the weighted adjacency matrix corresponding to the output features of different scales; Step 303: The weighted adjacency matrices corresponding to the output features of different scales are sequentially input into the graph structure mapper to obtain the weighted adjacency matrices at each scale, and finally obtain the visibility graphs of different scales.

5. The radio signal modulation recognition method based on multi-scale graph neural network according to claim 4 is characterized in that: In step S4, a multi-scale graph neural network model is constructed by combining graph convolution and pooling layers, and visual graphs of different scales are used as input to obtain feature vectors of different scales, including the following steps: Step 401: construct a multi-scale graph neural network model consisting of alternately stacked graph convolution and pooling layers; Step 402: Input the visibility graphs of different scales into the multi-scale graph neural network model to obtain corresponding feature vectors.

6. The radio signal modulation recognition method based on multi-scale graph neural network according to claim 5 is characterized in that: In step S5, the characteristic vectors of different scales are concatenated and the signal modulation type is output through a classifier, and the signal data set is used for model training, which includes the following steps: Step 501: concatenate feature vectors of different scales to obtain a final feature vector; Step 502: The concatenated feature vectors are sequentially reduced in dimension through two layers of fully connected networks, and then output the signal modulation type through a SoftMax classifier; Step 503: Use the signal data set to train the multi-scale graph neural network model.

Images