CN-122020319-A - Signal modulation recognition method based on fusion of graph neural network and time-frequency network

CN122020319ACN 122020319 ACN122020319 ACN 122020319ACN-122020319-A

Abstract

The invention discloses a signal modulation identification method based on fusion of a graph neural network and a time-frequency network, and belongs to the technical field of signal processing and artificial intelligence. The method comprises the steps of obtaining a signal sample data set, preprocessing the signal sample data set, finally generating a training set, a verification set and a test set, constructing a signal modulation recognition model, training the signal modulation recognition model based on the training set and the verification set, and testing the trained signal modulation recognition model based on the test set. According to the invention, the topology structure information among signal samples and the time sequence dynamic information of the signals are simultaneously utilized through the double-branch architecture, the automatic extraction and depth self-adaptive fusion of multi-mode features are realized through advanced modules such as a map structure, a time sequence feature extraction structure, bidirectional cross attention and the like which are defined by formulas, and the excellent performance, strong robustness and good generalization capability are shown in the signal modulation recognition task through the end-to-end combined training and the precise optimization strategy.

Inventors

PENG YAN
CHEN YUNPENG
PENG YAXIN
WEI HONGYU
KONG HAO
ZHOU YANG
QU DONG

Assignees

上海大学

Dates

Publication Date: 20260512
Application Date: 20260409

Claims (8)

1. A signal modulation recognition method based on fusion of a graph neural network and a time-frequency network is characterized by comprising the following steps: step 1, acquiring a signal sample data set, preprocessing the signal sample data set, and finally generating a training set, a verification set and a test set; Step 2, constructing a signal modulation recognition model, wherein the signal modulation recognition model comprises a graph neural network branch, a time-frequency network branch and a feature fusion module; training the signal modulation recognition model based on the training set and the verification set; and 4, testing the trained signal modulation recognition model based on the test set.
2. The method for identifying signal modulation based on fusion of a graph neural network and a time-frequency network according to claim 1, wherein in step 1, a signal sample data set is obtained and preprocessed, and a training set, a verification set and a test set are finally generated, specifically: Step 101, acquiring a signal sample data set, and labeling each signal sample to obtain a modulation type label corresponding to each signal sample; Step 102, performing multi-mode processing on the signal sample data set, respectively constructing graph structure data and original IQ signal time sequence data, and dividing the graph structure data, the original IQ signal time sequence data and corresponding modulation type labels according to preset proportions to obtain a training set, a verification set and a test set.
3. The method for identifying signal modulation based on fusion of a graph neural network and a time-frequency network according to claim 2, wherein in step 102, multi-mode processing is performed on a signal sample data set to construct graph structure data, specifically: and converting each signal sample in the signal sample data set into a graph structure representation, wherein the node characteristic matrix comprises statistical characteristics of five dimensions, the edge index matrix is constructed by adopting a K neighbor algorithm, the spatial connection relation between nodes is defined, and the batch index information is used for distinguishing node sets of different samples, so that the graph structure data is finally obtained.
4. The method for identifying signal modulation based on fusion of a graph neural network and a time-frequency network according to claim 3, wherein in step 102, a signal sample data set is subjected to multi-mode processing to construct original IQ signal time sequence data, specifically: Each signal sample in the signal sample data set adopts a three-dimensional tensor representation form, wherein the first dimension represents the batch size and supports batch processing, the second dimension represents the number of signal channels and is fixed into 2 channels corresponding to an in-phase component I and a quadrature component Q respectively, the third dimension represents the time sequence length and is fixed into 128 time steps through resampling, and the uniformity of input dimensions is ensured.
5. The method for identifying signal modulation based on fusion of a graph neural network and a time-frequency network according to claim 4, wherein the graph neural network branches are composed of three continuous graph convolution layers.
6. The signal modulation recognition method based on fusion of a graph neural network and a time-frequency network according to claim 5, wherein the time-frequency network branch adopts a double-branch composite architecture, time domain signal characteristics and frequency domain signal characteristics are extracted respectively, and a cross attention module for characteristic fusion is adopted, the time domain signal characteristic extraction module is used for extracting time domain characteristics by adopting a multi-head self attention mechanism after embedding and position coding of an IQ signal, the frequency domain signal characteristic extraction module is used for detecting a main period of the IQ signal through fast Fourier transform, and remolding frequency domain data according to the detected period, and the cross attention module is used for fusing the output of the time domain signal characteristic extraction module and the frequency domain signal characteristic extraction module.
7. The signal modulation recognition method based on fusion of a graph neural network and a time-frequency network according to claim 6, wherein the feature fusion module adopts a bidirectional cross attention mechanism and comprises a first cross attention module and a second cross attention module, the first cross attention module is used for carrying out feature fusion by taking a graph feature as a query and a key and a value, and the second cross attention module is used for carrying out feature fusion by taking the graph feature as a query and a time-frequency domain fusion feature as a key and a value.
8. The method for identifying signal modulation based on fusion of a graphic neural network and a time-frequency network according to claim 7, wherein in step 3, a cross entropy loss function, an Adam optimizer and an adaptive learning rate scheduling strategy comprising an early-stop mechanism and based on performance of a verification set are adopted in the training process.

Description

Signal modulation recognition method based on fusion of graph neural network and time-frequency network Technical Field The invention relates to the technical field of signal processing and artificial intelligence, in particular to a signal modulation identification method based on fusion of a graph neural network and a time-frequency network. Background With the rapid development of wireless communication technology, signal modulation recognition plays an important role in the fields of spectrum monitoring, electronic reconnaissance, cognitive radio and the like. The traditional signal modulation recognition method mainly depends on manually extracting features such as instantaneous features, statistical features, transform domain features and the like, and the method is seriously dependent on priori knowledge of domain experts and has poor robustness under different signal-to-noise ratio conditions. Modulation identification techniques have generally undergone three stages of development. The method is initially based on decision theory, is identified through hypothesis testing and likelihood ratio calculation, is complete in theory, is complex in calculation, requires known channel parameters, and has limited practicability. Subsequently, a pattern recognition method based on feature extraction becomes the mainstream, and the method classifies by combining a manually designed feature extractor (such as high-order accumulation of calculated signals, spectrum features and the like) with a traditional classifier (such as a support vector machine, a decision tree and the like), so that dependence on priori knowledge is reduced, but the robustness and generalization capability of the features are still bottlenecks. In recent years, deep learning techniques have been widely used for signal modulation recognition tasks. The Convolutional Neural Network (CNN) can automatically extract local characteristics of signals, but is difficult to capture time sequence dependency when processing long sequence signals, and the cyclic neural network (RNN) and variants thereof can process sequence data, but the serial calculation characteristics of the cyclic neural network (RNN) lead to low training efficiency and have gradient disappearance and explosion problems. More importantly, the existing recognition method based on deep learning still has a plurality of inherent defects when facing signals in a complex electromagnetic environment, and severely restricts the performance upper limit of the method, namely, firstly, the existing model (comprising CNN, RNN and transducer) generally treats each signal sample as an independent individual, lacks modeling capability of potential topological association among the signal samples, ignores structural information of a sample domain, secondly, on the processing of the time domain, the transducer model based on a self-attention mechanism can capture long-range dependence, but the calculation complexity of the transducer model is increased in a quadratic way along with the sequence length, and calculation cost is huge when processing long-sequence high-sampling signals, and RNN is limited by the difficulty of parallelization of serial calculation and low efficiency, thirdly, the existing method is single processing, even if research attempts are made to fuse various features, the method cannot realize deep interaction and self-adaptation complementation between structural features of the sample domain and dynamic features of the time domain by adopting a simple splicing or weighting summation and other shallow fusion strategies. Furthermore, existing architectures also lack specialized, efficient modeling mechanisms for complex periodic and non-stationary characteristics that are prevalent in communication signals. Therefore, the existing single-mode deep learning method often cannot simultaneously utilize the structural information and the time sequence information of the signals, has obvious defects in the aspects of calculation efficiency, feature depth fusion and periodic mining, and limits the further improvement of the recognition performance. In view of the above, there is a need for a signal modulation recognition method based on fusion of a graph neural network and a time-frequency network, which solves the above problems of the conventional method. Disclosure of Invention The invention aims to provide a signal modulation recognition method based on fusion of a graph neural network and a time-frequency network, which utilizes topological structure information among signal samples and time sequence dynamic information of the signals by a double-branch architecture, realizes automatic extraction and depth self-adaptive fusion of multi-modal characteristics by advanced modules such as a graph structure, a time-frequency characteristic extraction structure, bidirectional cross attention and the like defined by formulas, and shows excellent performance, strong robustness and goo