Application Of Deep Learning In Cognitive Radar Behavior Identification

The cognitive radar can adjust the radar transmitter parameters autonomously according to the electromagnetic environment,which enables the cognitive radar to flexibly switch the state of the interference on the battlefield to ensure that the cognitive radar is always in an optimal working state.The emergence of cognitive radar makes traditional reconnaissance methods based on ELINT library matching no longer effective.In this paper,cognitive radar is studied from the perspective of radar behavior,and the deep learning is used to make the reconnaissance system have the ability to extract and identity cognitive radar behavior characteristics.1.The basic model of cognitive radar is studied.This paper studies the signal characteristics of cognitive radar from the radar signal level,and studies the radar signal modulation method,cognitive radar adaptive waveform selection mechanism and radar working mode.2.Cognitive radar is characterized from the perspective of radar behavior.The concept and division of radar behavior are studied,and the behavioral model and description of cognitive radar are established.3.The deep learning network structure was studied.The basic network model is studied for one-dimensional and two-dimensional convolutional neural networks,and the training methods of deep neural networks are studied.The fast r-cnn network model in the field of image detection is further studied.4.The deep learning method is used to realize the modulation recognition of the radar signal.Firstly,the transform methods of radar signals in the frequency domain and time-frequency domain are studied and simulated,including Fourier transform,square spectral transform and time-frequency transform.In the case of low SNR,combined spectrum,square spectrum and time-frequency diagram,multiple parallel CNNs are used to realize the modulation recognition of 16 complex modulated signals.After testing,the signal-to-noise ratio is above-2dB,16 kinds of signals The recognition accuracy rate is above 90%.In the case of pulse overlap,the detection,identification and parameter extraction of overlapping signals are implemented by fast r-cnn according to the time-frequency diagram.The simulation test shows that the signal noise is mixed when the signals are mixed.When the ratio is above 0dB,the missed detection rate and false alarm rate of the signal are kept below 5%,the recognition accuracy of the signal is above 92%,and the error rate of parameter estimation is below 4%.5.Design a multi-channel one-dimensional CNN network to identify four radar modes.After testing,when the pulse loss rate is below 4%,the recognition accuracy of working mode is above 90%;when the pulse loss rate is below 10%,the recognition accuracy remains above 70%.The technical method of cognitive radar signal waveform prediction is studied.A cognitive radar waveform acquisition and database construction scheme is proposed,and a deep network model for waveform prediction is designed.6.The deep network implementation platform constructed in this paper is designed.The platform structure contains all algorithms for cognitive radar behavior feature extraction and behavior recognition.According to the size of the deep network,two schemes including acceleration chip and no acceleration chip are designed.The key technology and deployment process of deep network hardware acceleration are studied by using the FGPA deep learning acceleration platform of Xilinx.