Research On Modulation Recognition Technology For Frequency Hopping Signals Based On Deep Learning

Frequency hopping communication is a major means of spread-spectrum technology,which controls the frequency hopping of its carrier wave through a pseudo-random changing code sequence,Therefore,it has strong anti-interference and anti-interception capabilities and is widely used in military and civilian communication fields.Modulation recognition of frequency hopping signals,as a key technology in frequency hopping communication reconnaissance,has become one of the research hotspots in frequency hopping communication technology,and its recognition accuracy can directly affect the subsequent demodulation results.In the current information battlefield,the electromagnetic environment is increasingly complex,and the types of signal modulation are also increasing.This paper conducts research on its modulation recognition technology to address issues such as low recognition accuracy and poor robustness due to the highly non-stationary characteristics of frequency hopping signals.The main content is as follows:Firstly,this paper proposes a frequency hopping signal modulation recognition algorithm based on convolutional neural network,which converts the time-domain frequency hopping signal into the time-frequency domain and graph domain for feature extraction and modulation recognition.In the time-frequency domain,SPWVD is used to transform the signal to obtain a time-frequency image.After the convolutional denoising auto-encoder filters out some background noise to improve the noise resistance of the data,traditional convolutional neural networks are used to extract its time-frequency features for modulation recognition.The simulation experiment shows that the average recognition accuracy can reach 97.33% at 0d B;In the graph domain,the signal of each frequency hopping point is taken as the graph node by using the frequency hopping characteristics of the frequency hopping signal.At the same time,the bispectrum features and amplitude phase features of each node are extracted to build the node feature matrix,and the feature correlation is calculated to build the adjacency matrix,so as to realize the conversion from the frequency hopping signal to the topology signal,and the design graph convolution neural network extracts the signal graph features for modulation recognition.Simulation experiments have shown that better recognition performance can be achieved under low signal-to-noise ratio,with-8d B reaching 82.5%,which is 5.17% higher than time-frequency domain recognition methods.Secondly,this paper proposes a modulation recognition algorithm based on multimodal feature fusion,based on the different advantages of recognition algorithms based on timefrequency image features and graph features of frequency hopping signals.Introducing deep separable convolutional layers in traditional convolutional neural network models to improve system real-time performance,and introducing multi-head attention mechanism in graph convolutional neural network models to improve model robustness.Two models are used to extract time-frequency image features and graph features for fusion,and support vector machines are used for modulation recognition of frequency hopping signals.The simulation result shows that a recognition accuracy of 93.17% can be achieved when the signal-to-noise ratio is-8d B.Finally,this paper builds a frequency hopping communication system based on the GNURadio software development platform and USRP software radio equipment,collects 10 kinds of modulation frequency hopping signals to test the recognition algorithm in this paper.At the same time,it was compared with the time-domain recognition algorithm based on deep learning models.The recognition algorithm based on multimodal feature fusion in this paper is superior to other algorithms,with an average recognition accuracy of 98.30%.