Research For Classification Method Of Complex Radar Signals Based On Neural Network

Radar signal sorting is a very important process in the electronic detection system,which contains the separation of the overlapping pulse stream signals received in the reconnaissance plane,and the recognization of radar signals belongs to the same pulsed radiation source.The technology of radar signal sorting is increasingly innovating with the development of the modulation mode of radiation source signals and investigation equipment.The traditional methods of radar signal sorting are mainly based on the pulse description word(PDW)of the pulse feature.The carrier frequency,pulse arrival angle and other parameters are used for presorting,and then the pulse repetition interval(PRI)is used to achieve deinterleaving of the pulse stream to complete the main sorting.Traditional methods were frequently used in practical engineering.With the diversification of signal modulation methods,the electromagnetic scene is increasingly dense and complex,and the sorting methods based on pulse characteristics cannot meet the demand of modern radar signals sorting.The researchers exploited the inpulse characteristics of signals and further developed noval sorting methods based on it.This article focuses on the research of sorting techniques of complex modulated pulse radar signal.Sorting signals are the following six kinds of complex modulation signals: LFM signal,SLFM signal,FMCW signal,Costas frequency hopping signal,Barker code signal and Frank code phase modulation signal.The research content is as follows:1.The classic sorting methods based on the characteristics of the pulse,cumulative difference histogram(CDIF)algorithm,sequential difference histogram(SDIF)algorithm and PRI transform method were studied.The application scene and the sorting performance were discussed.Simulation results show that the CDIF algorithm and SDIF algorithm are suitable for the fixed PRI signals,and the modified PRI transform algorithm works better on jitter signal and partial pulse loss signal.2.This thesis exploits the extraction methods of two kinds of signal pulse features: the timefrequency feature based on wavelet transform and the bispectrum slice feature based on highorder cumulant.The time-frequency features obtained by the wavelet transform have clear physical meanings,but the deviation in the ridge line of the frequency hopping signal is easy to be created;the bispectrum slice feature has good performance on suppressing the Gaussian noise,and make the subsequent classification algorithms more effectively to achieve the sorting of complex modulated signals.3.Two feature classification algorithms based on full-connection neural network and deep belief network(DBN)are designed.The input of both networks is the intra-pulse feature vector obtained in the previous text.The results show that the two networks can achieve good separation of these kinds of complex signals.The DBN converges faster during training.The correct rate of sorting signals with DBN is more than 80% when the SNR of the signal is not less than-5d B.The correct rate of sorting signals with full-connected neural network is more than 91% when the SNR of the signal is not less than-5d B.