Research On Non-cooperative Target Micro-motion Feature Recognition Method Based On Convolutional Neural Networ

The unmanned and stealth technology of modern non-cooperative targets is becoming more and more advanced.How to effectively detect and warn non-cooperative targets such as ballistic targets and unmanned aerial vehicles is a challenge in the field of radar target recognition.Micro-motion features and deep learning have been the research hotspots in the field of radar target recognition in recent years.The convolution neural network and other deep learning methods are used to extract the micro motion features of the target has important application prospects.On this basis,this paper studies the inertial feature extraction method of the space cone target with the tails and the classification method of the unmanned aerial vehicles.The main research contents include the following aspects:The third chapter studies the inertial feature extraction method of space cone target based on micro-motion features.The acquisition methods of radar echo data in this chapter include theoretical derivation,numerical simulation and experimental testing.In terms of theoretical derivation,starting from the point scattering model,the micro-Doppler frequency was obtained by micro-motion modeling of the blunt-head flat-bottomed cone target,and the correctness of the subsequent numerical simulation results was verified.In terms of numerical simulation,the obtained echo data was used to construct a time-frequency map dataset,a convolutional neural network was built to extract inertial features,and the boundary values of signal-to-noise ratio,pulse repetition frequency and radar carrier frequency were discussed in combination with measured data.In order to reduce the influence of noise on the quality of the time-frequency map,the feed-forward denoising convolutional neural networks and singular value decomposition were introduced to optimize the time-frequency map dataset,reducing the network training time and memory to 48% and 40% of the original,respectively.And compared with classical convolutional neural networks such as Alex Net,Res Net50 and Goog Le Net at0 d B,the inertial feature parameter estimation error is reduced by more than 4.5%.The fourth chapter studies the unmanned aerial vehicles target classification method based on micro-motion features.The acquisition methods of radar echo data in this chapter include theoretical derivation and numerical simulation.In terms of theoretical derivation,starting from the point scattering model,the unmanned aerial vehicles’ rotor blades were micro-motion modeling to obtain the micro-Doppler frequency,and the correctness of the subsequent numerical simulation results was verified.In terms of numerical simulation,an unmanned aerial vehicle target simulation method based on multilevel fast multipole algorithm was proposed,which significantly reduces the simulation time and memory under the premise of ensuring the accuracy of the echo.On this basis,the obtained echo data was used to construct a timefrequency map dataset,a convolutional neural network was built for unmanned aerial vehicle target classification.Transfer learning was used to improve the classification accuracy of the network,the classification accuracy can reach 98.5% when the signal-to-noise ratio is 10 d B,and the classification accuracy below 10 d B is improved by up to 3%.Finally,the influence of signal-to-noise ratio,spin frequency and radar elevation angle on classification accuracy was discussed.Finally,the methods proposed in this paper have universal applicability and can be further extended to other target recognition fields.