Radar Signal Intra-pulse Modulation Recognition Methods Based On Time-frequency Domain Features And Deep Learning

Radar signal intra-pulse modulation type recognition is the premise and basis of radar anti-jamming,which plays an important role in electronic reconnaissance,electronic countermeasures,spectrum management and other aspects.With the emergence of new radars and electronic devices,the intra-pulse modulation types of radar signals become more and more diverse,and traditional methods of detecting radar signals have become much less effective.Recently,the application of deep learning in radar signal processing has been a popular topic.Deep learning allows the diversity of internal information features to be explored in depth,and radar signal modulation recognition methods based on deep learning can learn the data more completely and are less affected by external noise,which significantly improve the recognition rates over traditional methods.Although some deep learning-based intra-pulse modulation recognition methods for radar signals have been proposed,however,these methods still suffer from low recognition rates and high computational complexity in low signal-to-noise ratio(SNR)environment.Based on the time-domain,frequency-domain and combined time-frequency-domain characteristics,this paper presents an in-depth study of the problem of identifying the intra-pulse modulation type of radar signals using deep learning.The points of this work are the following:Firstly,the generation and feature extraction of typical radar intra-pulse modulation signals are investigated.A total of seven types of typical intra-pulse modulation radar signals are generated via simulation,including continuous wave signal,binary frequency shift keying,sinusoidal frequency modulation signal,even quadratic frequency modulation signal,linear frequency modulation signal,binary phase-shift keying signal and quaternary phase-shift keying signal,and the one-dimensional features in time domain and frequency domain as well as the two-dimensional time-frequency image(TFI)features are analyzed,which can lay a foundation for subsequent dataset generation.Then,to make up for the shortcoming of low identification rate for traditional methods at low SNR,a method based on time-frequency domain features and deep neural network(DNN)is developed for intra-pulse modulation identification of radar signals.The data set is created using the extracted one-dimensional feature information in time domain and frequency domain,such as frequency domain moment kurtosis coefficient,frequency domain envelope fluctuation,single frequency energy concentration degree,frequency domain moment skewness coefficient,frequency average flatness coefficient and carrier factor with certain differentiation degree.The DNN structure is created and the algorithm flow is given.The final result is verified,and the recognition effect under different network layers,different parameter updating algorithms,different training sets and different activation functions are analyzed.The results show that the method works well and achieves an overall correct recognition rate of over 90% even at SNR of-8d B.Finally,to further improve the recognition effect of signals with low SNR,a method using second-order short-time Fourier transform-based synchrosqueezing transform(FSST2)and convolutional neural network(CNN)is proposed for intra-pulse modulation identification of radar signals.At first,the TFIs of the seven radar modulation signals generated based on FSST2 are pre-processed with greyscale and bilinear size transformation,so that the obtained TFIs have better noise immunity and higher time-frequency aggregation,and the greyscale transformation size is used to generate the data set.Then,the CNN used in this paper is designed based on the residual shrinkage module,and the TFI data set is fed into the CNN for classification.Simulations demonstrate that the recognition method based on the CNN can still achieve over 90%correct recognition rate at SNR of-12 d B,with higher recognition rate and better general adaptation.In comparison with other TFI generation methods and other neural network structures,this method is more accurate in identifying the seven signals in low SNR environment.