Transform and recurrence plot-based signal modulation identification method and device

Abstract

The invention discloses a signal modulation identification method and device based on Transform and a recurrence plot, and the method comprises the steps: S1, carrying out the preprocessing of a frequency component of an inputted to-be-identified IQ signal through a frequency selection module, and filtering out the redundant frequency component through the joint analysis of the frequency spectrum information of an IQ channel; s2, converting the IQ signal after frequency selection into a trajectory vector of a phase space through a time delay embedding method; the component of each dimension of the trajectory vector is the node feature of the graph. And S3, transmitting the node features of the graph into graph construction of a Transform architecture with a contraction depth convolution bypass and a graph convolution layer for training, and finally obtaining a classification result of the modulation signal. According to the method, the IQ signal is mapped to a graph domain by using a phase space recurrence graph, the calculation complexity of graph convolution is reduced by using a self-attention mechanism, meanwhile, the performance of the model is further improved by designing a frequency selection module and ShrinkageDWC, finally, relatively good performance is obtained, parameters are kept efficient, and the method has good re-characterization capability on signals with different lengths.

Description

Technical Field

[0001] The present invention relates to the fields of phase space, recursive graph, neural network architecture (graph neural network, Transformer) and modulated signal recognition, and in particular to a method and device for signal modulation recognition using the phase space recursive graph of the modulated signal and the Transformer architecture. Background Art

[0002] Automatic modulation recognition is a key technology for non-cooperative communication and intelligent communication. How to quickly and accurately determine the modulation type has become an important issue in intelligent signal processing. In communication systems, signals are usually transmitted in a time series manner, which makes it difficult to capture the dynamic characteristics of the system. Most researchers focus on automatic modulation recognition methods based on deep learning (reference [1] F. Zhang, C. Luo, J. Xu, Y. Luo, F.-C. Zheng, Deep learning based automatic modulation recognition: Models, datasets, and challenges, Digital Signal Processing 129 (2022) 103650, i.e. F. Zhang, C. Luo, J. Xu, Y. Luo, F.-C. Zheng, “Automatic modulation recognition based on deep learning: models, datasets and challenges” Digital Signal Processing 129 (2022) 103650.). By constructing a multi-layer neural network, it is possible to adaptively learn signal features and perform classification, avoiding parameter estimation and complex feature engineering (reference [2] W. Xiao, Z. Luo, Q. Hu, A Review of Research on Signal Modulation Recognition Based on Deep Learning, Electronics 11 (2022) 2764). Most of the above methods use time series as the starting point for feature mining. For data-based deep learning methods, the original time domain representation limits the diversity of features. Therefore, the re-representation method is used to enrich the representation of the signal from the perspective of complex networks or graph theory, thereby improving the signal feature expression ability and enriching the mining of features, which is of great significance to improving recognition performance.

[0003] There has been a lot of progress in the method of re-representing signals as graphs. First, the signals are mapped to graphs, and then they are classified using powerful graph neural networks (GNNs). A signal recognition method based on visibility graph (VG) is proposed (reference [3] Q. Xuan, J. Zhou, K. Qiu, Z. Chen, D. Xu, S. Zheng, X. Yang, AvgNet: Adaptive Visibility Graph Neural Network and Its Application in Modulation Classification, Ieee Trans. Netw. Sci. Eng. 9 (2022) 1516–1526, i.e. Q. Xuan, J. Zhou, K. Qiu, Z. Chen, D. Xu, S. Zheng, X. Yang, “AvgNet: Adaptive Visibility Graph Neural Network and Its Application in Modulation Classification”, Ieee Trans. Netw. Sci. Eng. 9 (2022) 1516–1526.). By combining VG with DiffPool, an end-to-end graph learning framework is realized. Then, other workers deeply analyzed the mechanism of the visible graph and simplified the mapping process of AvgNet by limiting the visible range, which significantly improved the mapping efficiency with little loss of accuracy (reference [4] K. Qiu, S. Zheng, L. Zhang, C. Lou, X. Yang, DeepSIG: A Hybrid Heterogeneous Deep Learning Framework for Radio Signal Classification, IEEE Transactions on Wireless Communications 23 (2024) 775–788). Although the above work achieved high modulation recognition accuracy, its model architecture serialized signal re-representation and data mining, resulting in problems such as inefficient parameters, larger model size and longer iterative training.Mapping time series to phase space by embedding time delays to construct recursive graphs is another way to implement graph representation algorithms (reference [5] W. Ren, N. Jin, L. Ou Yang, Phase Space Graph Convolutional Network for Chaotic Time Series Learning, IEEE Transactions on Industrial Informatics 20 (2024) 7576–7584). Recursive graphs can achieve parallel re-representation of time series and data mining. However, there is no relevant research on signal recognition based on recursive graphs in the task of signal modulation recognition.

[0004] There are many attempts to apply Transformer to the graph field. Among them, a relatively effective method is to replace softmax attention with some key modules in the classic GNN variant. A classic graph attention network is inspired by the attention mechanism (reference [6] P. G. Cucurull, A. Casanova, A. Romero, P. Liò, Y. Bengio, GraphAttention Networks, (2018), that is, P. G.Cucurull, A.Casanova, A.Romero, P.Liò, Y.Bengio, Graph Attention Networks, (2018). Although the above methods have achieved satisfactory results in graph classification tasks, they cannot be used for signals without ready-made graph structures. Summary of the invention

[0005] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provides a signal modulation recognition method and device based on Transformer and recursive graph.

[0006] In order to improve the accuracy of signal modulation recognition, while improving parameter efficiency, reducing model size and providing more diverse features, the present invention constructs a recursive graph in the signal phase space, and utilizes the advantages of Transformer parallel computing to propose a signal modulation recognition method based on Transformer and recursive graph.

[0007] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0008] The first aspect of the present invention relates to a signal modulation identification method based on Transformer and recursive graph, comprising the following steps:

[0009] S1: The input IQ signal to be identified is first pre-processed by the frequency selection module, and the redundant frequency components are filtered out by jointly analyzing the spectrum information of the IQ channel.

[0010] S2: The IQ signal after frequency selection is converted into a trajectory vector in phase space through time delay embedding. The component of each dimension of the trajectory vector is the node feature of the graph.

[0011] S3: The node features of the graph are passed into the graph construction and graph convolution layer of the Transformer architecture with shrinkage depth convolution (ShrinkageDWC) bypass for training, and finally the classification result of the modulated signal is obtained.

[0012] Further, the step S1 comprises the following contents:

[0013] S1.1: The time domain IQ signal is transformed into the frequency domain by fast real Fourier transform (rFFT) to obtain the frequency component of the IQ signal. The IQ input signal is expressed as:

[0014]

[0015] Where n is the number of sampling points of each signal data. The frequency components of the IQ signal are calculated separately, and

[0016]

[0017] S1.2: Input the frequency components to the frequency selection module to automatically filter out redundant frequency components.

[0018] S1.3: The processed frequency components are restored to IQ signals through inverse real Fourier transform (irFFT),

[0019]

[0020] Further, the step S2 comprises the following contents:

[0021] S2.1: Convert the preprocessed IQ signal into a phase space trajectory vector. Use the time delay embedding method to convert the IQ signal into a phase space trajectory vector:

[0022]

[0023] Where m represents the dimension of the time-delay embedding and τ represents the delay time.

[0024] S2.2: Splice the IQ trajectory vectors according to the time points to obtain the complete phase space trajectory vector of the IQ signal, which is recorded as the node vector X of the graph:

[0025] X=[x1,x2,...,x n ] (5)

[0026]

[0027] Further, the step S3 includes the following contents:

[0028] S3.1: Linearly project the node vector and (k=0,1,...,h-1), respectively mapped to query vector, key vector, and value vector, each of which has h heads, and the formula is as follows:

[0029]

[0030] Among them, d t Represents the embedding dimension of the node token, d k represents the embedding dimension of the key vector of each head, d v The value vector representing the embedding dimension of each head.

[0031] S3.2: Through the linearized attention mechanism, the graph convolution output of each head is calculated, denoted as The attention mechanism formula used is as follows:

[0032]

[0033] Where 1 is an N×1 dimensional all-ones vector.

[0034] S3.3: The value vector obtains richer features through the ShrinkageDWC module. Shringkage uses soft thresholding to implement feature denoising. The soft thresholding function is expressed as:

[0035]

[0036] Among them, x is the input feature, y is the output feature, and τ represents the threshold, which is a positive parameter.

[0037] S3.4: The output of the attention mechanism and the output of ShrinkageDWC are added element by element and linearly projected to achieve feature fusion between h heads.

[0038] S3.5: The output obtained from S3.4 passes through a feed-forward network (FFN) and a pooling layer in the time dimension to improve the diversity of features and reduce the number of graph nodes, and serves as the input of the next layer of graph convolution process.

[0039] S3.6: Use global average pooling to read out the graph after graph convolution, and use a two-layer fully connected network to map the number of features to the category of the modulated signal.

[0040] The second aspect of the present invention relates to a signal modulation identification device based on Transformer and recursive graph, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement a signal modulation identification method based on Transformer and recursive graph of the present invention.

[0041] The third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements a signal modulation identification method based on a Transformer and a recursive graph of the present invention.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] This paper proposes a graph mapping algorithm from signal to recursive graph, which can analyze the complex modulation type of the signal from the perspective of graph theory. It uses the self-attention mechanism of Transformer to optimize the distance calculation of phase space trajectory vectors and the threshold selection process.

[0044] The present invention realizes parallel graph structure learning and graph convolution operation process with linear computational complexity. At the same time, the maximum pooling commonly used in convolutional neural networks is used to realize layered graph pooling, thereby realizing graph readout operation, reducing the number of nodes and achieving further acceleration. The above optimization of computational complexity enables the graph-based signal recognition model to be used on modulated signals of longer lengths, greatly improving the recognition performance on long signal data samples.

[0045] The present invention introduces a shrinkage deep convolution bypass to achieve denoising and enrich feature diversity. In addition, the present invention designs a learning-based frequency selection module for signal preprocessing, which is very helpful for the graph mapping process. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Flow chart of the method of the present invention.

[0047] Figure 2 It is a framework diagram of the method of the present invention.

[0048] Figure 3This is a framework diagram of the frequency selection module proposed in the present invention.

[0049] Figure 4 This is a schematic diagram of the Shrinkage module threshold learning used in the present invention.

[0050] Figure 5 It is a schematic diagram of the device of the present invention. DETAILED DESCRIPTION

[0051] The following is attached to the instruction manual Figure 1 , 2 The present invention is further described in detail by taking the modulation signal dataset RML2016.10a as an example.

[0052] Example 1

[0053] This embodiment provides a signal modulation recognition method based on Transformer and recursive graph, which is specifically divided into the following steps:

[0054] S1: Divide the data set, and then the input IQ signal to be identified is first pre-processed by the frequency selection module to filter out redundant frequency components by jointly analyzing the spectrum information of the IQ channel.

[0055] S2: The IQ signal after frequency selection is converted into a trajectory vector in phase space through time delay embedding. The component of each dimension of the trajectory vector is the node feature of the graph.

[0056] S3: The node features of the graph are passed into the graph construction and graph convolution layer of the Transformer architecture with shrinkage depth convolution (ShrinkageDWC) bypass for training, and finally the classification result of the modulated signal is obtained.

[0057] In step S1, the specific operation process is as follows:

[0058] The RML2016.10a dataset is divided into a training set and a test set in a ratio of 4:1. First, the IQ signal is input into the frequency selection module. Figure 3 . The frequency selection process can be described as: the input IQ signal with a shape of 128*2 is subjected to real Fourier transform to obtain the frequency components of the IQ signal with a shape of 65*2. Then, the local frequency correlation of the IQ channel is learned through two layers of large-scale 1D convolution, and each frequency component is mapped to two eigenvalues. The length of the convolution kernel is generally set to 8. Then, it is normalized by softmax as the probability of selecting the frequency component. Next, random sampling is implemented using the Gumbel-trick, which allows the gradient to backpropagate during discrete sampling. Finally, the selected frequency component is transformed into an IQ signal by the inverse real Fourier transform (irFFT).

[0059] In step S2, the specific operation process is as follows:

[0060] The frequency-selected IQ signal is obtained from S1, with a shape of 128*2. In the time delay embedding method, we set the embedding dimension to 16 and the delay time to 1. First, the phase space trajectory vectors of the IQ signal are obtained separately, where the trajectory vector shape of the I signal is 113*16, and the trajectory vector shape of the Q signal is also 113*16. Then, the IQ trajectory vectors are spliced ​​to obtain the overall phase space trajectory vector, with a shape of 113*32.

[0061] In step S3, the specific operation process is as follows:

[0062] For the phase space trajectory vector obtained in S2, i.e. the node vector, we first input a linear layer and a Gelu activation function to obtain the node embedding token of the attention mechanism, which has a shape of 113*64. The formula of this process is as follows:

[0063] X emb =Gelu(Linear(X)) (10)

[0064] Next, refer to Figure 2 , the node embedding is mapped to query vector, key vector, and value vector through different linear projections as the input of the linear attention mechanism. The shape of the node embedding after linear attention aggregation is h*113*(64 / h).

[0065] At the same time, the value vector is enriched by two layers of large-scale deep convolutional networks (DWC) to enrich the feature representation of the linear attention mechanism. Then the features output by DWC are denoised using the shrinkage mechanism. The learning process of the denoising threshold is as follows: Figure 4 As shown: The features of each channel take the absolute value and use global average pooling, interact with the information between channels through two fully connected layers, and then apply sigmoid at the end to make the scaling factor within the range of (0,1).

[0066] The output of the attention mechanism and the output of ShrinkageDWC are added element by element and linearly projected to achieve feature fusion between h heads to obtain the output of the graph convolution. Then, after the FFN and pooling layers in the time dimension, the diversity of features is improved and the number of graph nodes is reduced, which serves as the input of the next layer of graph convolution process.

[0067] Finally, the graph after graph convolution is read out using global average pooling, whose shape is 64, and a two-layer fully connected network is used to map the number of features to the category of the modulated signal, that is, the shape is 11.

[0068] As described above, the present invention introduces an embodiment of signal modulation recognition using the modulated signal dataset RML2016.10a. The present invention uses a phase space recursive graph to map the IQ signal to the graph domain, and uses a self-attention mechanism to reduce the computational complexity of graph convolution. At the same time, a frequency selection module and ShrinkageDWC are designed to further improve the performance of the model, and finally obtain better performance while maintaining efficient parameters, and have good re-characterization capabilities on signals of different lengths.

[0069] Example 2

[0070] This embodiment relates to a signal modulation identification device based on Transformer and recursive graph, such as Figure 5 , including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement a signal modulation identification method based on Transformer and recursive graph of Example 1.

[0071] Example 3

[0072] This embodiment provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, a signal modulation identification method based on a Transformer and a recursive graph of Embodiment 1 is implemented.

[0073] The above embodiments are merely illustrative and not restrictive of the present invention. Those skilled in the art will appreciate that many changes, modifications, and even equivalences may be made to the embodiments within the spirit and scope of the claims, but all of these will fall within the scope of protection of the present invention.

Claims

1. A signal modulation recognition method based on Transformer and recursive graph, comprising the following steps: S1: The input IQ signal to be identified is first pre-processed by the frequency selection module, and the redundant frequency components are filtered out by jointly analyzing the spectrum information of the IQ channel; S2: The IQ signal after frequency selection is converted into a trajectory vector in phase space through time delay embedding. The component of each dimension of the trajectory vector is the node feature of the graph. S3: The node features of the graph are passed into the graph construction and graph convolution layer of the Transformer architecture with shrinkage depth convolution (ShrinkageDWC) bypass for training, and finally the classification result of the modulated signal is obtained.

2. A signal modulation recognition method based on Transformer and recursive graph, wherein step S1 specifically comprises: S1.1: The time domain IQ signal is transformed into the frequency domain by fast real Fourier transform rFFT to obtain the frequency component of the IQ signal; the IQ input signal is expressed as: in, n is the number of sampling points of each signal data; the frequency components of the IQ signal are calculated respectively, by S1.2: Input the frequency components to the frequency selection module to automatically filter out redundant frequency components; S1.3: The processed frequency components are restored to IQ signals through inverse real Fourier transform irFFT, 3. A signal modulation recognition method based on Transformer and recursive graph, step S2 specifically comprises: S2.1: Convert the preprocessed IQ signal into a phase space trajectory vector; Use the time delay embedding method to convert the IQ signal into a phase space trajectory vector respectively: in, m represents the dimension of time delay embedding, and τ represents the delay time; S2.2: Splice the IQ trajectory vectors according to the time points to obtain the complete phase space trajectory vector of the IQ signal, which is recorded as the node vector X of the graph: X=[x1,x2,...,x n ] (5) 4. A signal modulation recognition method based on Transformer and recursive graph, step S3 specifically comprises: S3.1: Linearly project the node vector and They are mapped to query vector, key vector, and value vector respectively, each of which has h heads, and the formula is as follows: in, d t Represents the embedding dimension of the node token, d k represents the embedding dimension of the key vector of each head, d v The value vector embedding dimension representing each head; S3.2: Through the linearized attention mechanism, the graph convolution output of each head is calculated, denoted as The attention mechanism formula used is as follows: Among them, 1 is an N×1-dimensional all-one vector; S3.3: The value vector obtains richer features through the ShrinkageDWC module; Shrinkage uses soft thresholding to achieve feature denoising, and the soft thresholding function is expressed as: Among them, x is the input feature, y is the output feature, τ represents the threshold, which is a positive parameter; S3.4: The output of the attention mechanism and the output of ShrinkageDWC are added element by element and linearly projected to achieve feature fusion between the h heads; S3.5: The output obtained by S3.4 passes through the feed-forward network (FFN) and the pooling layer in the time dimension to improve the diversity of features and reduce the number of graph nodes, and serves as the input of the next layer of graph convolution process; S3.6: Use global average pooling to read out the graph after graph convolution, and use a two-layer fully connected network to map the number of features to the category of the modulated signal.

5. A signal modulation recognition device based on Transformer and recursive graph, characterized in that: It comprises a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement a signal modulation recognition method based on Transformer and recursive graph according to any one of claims 1 to 4.

6. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, a signal modulation identification method based on a Transformer and a recursive graph as described in any one of claims 1 to 4 is implemented.

Images