Research On Narrow Band Radar True-false Target Discrimination Based On Deep Learning

Radars are important sensors for space surveillance systems and early warning systems.The functionality of the systems greatly depends on the radars' ability of detecting,tracking the interested targets and recognizing them.During the processes of detecting,tracking and recognizing,false targets severely deteriorates the performance of radars,including active jamming targets and passive decoys.This thesis researches several critical techniques of radar target recognition based on deep learning,for discriminating the interested targets from false targets in radar system formed by interrupted-sampling repeater jamming(ISRJ)and decoys.The techniques are listed as follows:(1)Before radar target detection process,a stacked bidirectional gated recurrent unit(GRU)neural network was designed to discriminate real target echo signals from ISRJ signals in time domain,and the jamming-free signal segments were extracted to generate a particular band pass filter,which suppresses the ISRJ signal components and retains the real target echo simultaneously,leading to a result that the false target formed by the ISRJ signal was suppressed and the detection rate of the real target was greatly improved under ISRJ environment.(2)After target detection but before tracking,a special convolutional neural network(CNN)was designed to enhance the time-frequency distribution(TFD)of the dechirped signal,and the parameters of signal components was extracted from the enhanced TFD.Furtherly,among all the detected targets,the real target was discriminated from ISRJ target according to the corresponding signal component's time-frequency parameters.The proposed method kept ISRJ targets from entering the tracking procedure,and improved the probability of radar tracking the real target.(3)After tracking,radar classifies targets by the acquired data sequence.Two method based on deep learning techniques and radar cross section(RCS)sequences were proposed to extract features and recognize targets.Firstly,a one-dimensional CNN was designed to extract the period of a target's micro-motion from RCS sequence and coarsely separate true targets from decoys.The method achieved better period estimation and classification performance than traditional methods.Secondly,another one-dimensional CNN was designed to discriminate true targets from decoys by directly taking RCS sequence as input,and achieved higher recognition rate comparing with traditional recognition methods.