Deep Learning-based F Lying Target Recognition From Wireless Signals

Flight target recognition(FTR)is very important in civil aviation security,sensitive area protection and many other fields.Radar cross section(RCS),which contains information to reflect the electromagnetic scattering and motion characteristics of flying target,is one of mostly used signals in FTR.Classical statistical method of RCS-based FTR methods usually calculate the high-order cumulant and then classify the targets regarding some specific known features or by typical classfiers such as SVM and MLP.These cumulant-based methods have strict requirements on the stability and signal-to-interference-and-noise-ratio of the RCS signals,and a prior knowledge of flight targets as well.In practice,these requirements on the RCS signals are hardly met,especially in the case of antagonistic FTR,which greatly degrades the cumulant-based methods.Recently,deep learning(DL)methods have been testified to hold excellent abstract presentation ability and end-to-end learning capability,and then have been applied in target recognition,computer vision and natural language processing.In this thesis,DL-based FTR method is studied,and the main contributions are listed below.1.A hybrid deep learning model is proposed,which combines convolution neural network and bi-directional LSTM network to distill both the spatial and temporal features from RCS signals.The performance of the proposed model is testified,and the results show its advantages over the reference schemes.2.A transfer learning-based FTR method is proposed to greatly decrease the cost of practical RCS signals without loss of performance.Since data-driven deep learning model heavily depends on the labeled samples,it is necessary to obtain sufficient samples to train the model.However,it is very expensive to obtain sufficient RCS signals.\The hybrid deep learning model is sufficiently trained on the synthetic RCS dataset,and then transferred to the small dataset consisting of real RCS data.