Radar Signal Detection And Parameter Estimation Based On Deep Learning

Electronic warfare(EW),also known as electronic countermeasures(ECM),is a struggle between two opposing sides to compete for the electromagnetic spectrum of the battlefield.ECM technology is one of the key combat forces on the` contemporary information battlefield.The detected and estimated technology of electronic reconnaissance(ER)is the primary task and prerequisite for the successful implementation of electronic attacks and protection in EW.However,with the rapid development of military electronic information technology,the electromagnetic spectrum of the battlefield is dynamic,changeable,dense and complex.It brings unprecedented challenges to traditional detection and estimation methods.At the same time,with the rapid development of deep learning(DL)technology,the application of related technologies in EW has also become a current research focus.Based on the traditional technology of detection and estimation,the intelligent detection and intelligent parameter estimation methods of radar signals are proposed respectively with the self-learning framework of DL.The experiment results show that the proposed algorithm is effective.The specific work in this thesis is as follows:When the radar signals of different modulation types have aliasing in the time and frequency domains at the same time,the conventional detection methods may miss or unable to detect.Aiming at this problem,this thesis proposes a new radar signal detection method which is based on DeepLabv3+.First,perform short-time Fourier transform(STFT)of the received mixed signal to obtain time-frequency data and store it as a time-frequency map.Secondly,While the model is being trained,the pixels of background noise,signal components,and the area where the overlap is located are labeled differently.And then,the model is trained with the input time-frequency maps and labels to be able to detect and identify aliased signals.After the signals can be accurately detected and identified,this thesis investigates how to separate mixed signals,and proposes an technique which is based on mask filtering and generative adversarial network(GAN)to separate and reconstruct the mixed signals.After DeepLabv3+ detects the time-frequency map of aliased signals,it can obtain the pixel areas where different signals are located.By mapping the pixel areas of different signals on the time-frequency map to the time-frequency matrix,the target signal can be masked to obtain the time-frequency matrix of a single signal.The time-frequency matrix is transformed into the signal waveform of the incomplete overlapping part after inverse timefrequency transformation.The trained GAN network can reconstruct the incomplete waveform of the signal and then separate the mixed signal.This thesis also studies the method of radar signal parameter estimation which is based on deep learning.Aiming at the uncertainty principle in time-frequency transformation,this thesis proposes a time-frequency analysis method based on multi-scale fusion,and combines the target detection model of DL to estimate the parameters of frequency-modulated signals.The method integrates the time-frequency analysis results of different window scales,which can effectively alleviate the defects in the uncertainty principle,and improve the time-frequency resolution.When the YOLOv5 m network obtains the fused time-frequency map and the real position of the signal,it can estimate the initial frequency,bandwidth(BW),pulse time of arrival(TOA)and pulse width(PW)of the signal.However,this method can only estimate the linear time-frequency relationship,and is not suitable for estimating the frequency-time relationship in nonlinear frequency modulation(NLFM)signals.Therefore,based on the accurate estimation of the time-frequency range of the signal,the method intercepts the time-frequency region where the NLFM signal is located,and extracts the extreme value from the intercepted data as the time-frequency curve of the signal.Then,the time-frequency curve is fitted by the least square method to obtain a more accurate timefrequency curve,and the frequency modulation coefficient of the fitting curve is solved to obtain the parameters of the nonlinear frequency modulation signal.