Study Of Radar High Range Resolution Profiles Target Recognition Based On Auto-Encoder

Radar high-resolution range profile(HRRP)denotes the amplitude of coherent summation of complex time returns from target scatterers in each range cell.HRRP contains the intensity distribution of target's scatterers and the radial structural information of the target.Moreover,compared with other high-resolution radar signals,HRRP is easier to obtain and store.Therefore,it has become an important method of radar real-time target recognition.However,for radar HRRP target recognition,how to obtain the robust and effective recognition features of the target is the question that we always want to solve.The thesis focuses on the study of HRRP target recognition,mainly discussing and analyzing the inherent characteristics of HRRP and the feature extraction method for HRRP target recognition.The main contents of the thesis are summarized as follows:1.The basic principles of HRRP target recognition are studied.The three key problems in HRRP target recognition: the sensitivity of HRRP,the extraction and selection of target features,and the design of classifier are described in detail,which provide a theoretical basis for the following research.2.The radar target recognition method based on deep neural network is studied.Most of the traditional feature extraction methods are shallow linear structures,so it's difficult to obtain the hierarchical effective recognition features of the target.In order to solve this problem,we introduce the radar HRRP target recognition method based on deep network: Stacked Denoising Autoencoder(SDAE)and Deep Belief Network(DBN).SDAE optimizes the model parameters by minimizing the reconstruction error,while DBN studies the model by maximizing the logarithmic likelihood function.Comparing these two models,we find that although they have different principles,the training process is similar.Both of them use the unsupervised greedy layer-wise training method to pre-train the network,and then use the back propagation algorithm to fine-tune the parameters.In this way,they can avoid problems existing in traditional training methods,such as long training time and vanishing gradient.Finally,the experiment based on the measured HRRP data verifies that the deep network structure can improve the feature extraction capability of the model.3.The radar target recognition method based on Robust Variational Autoencoder(RVAE)is studied.Because of the limitation of the model structure and ignoring the inherent characteristics of HRRP,the performance in the practical application of traditional neural network is limited.To solve this problem,we propose a target recognition algorithm based on RVAE.First,we introduce the meaning and the algorithm principle of Variational Autoencoder(VAE)in detail,and demonstrate its superiority by compared with the traditional Autoencoder(AE).Moreover,considering the characteristics of HRRP,we use the average profile to alleviate its orientation sensitivity.And then based on the original VAE we add HRRP frame and average processing to construct RVAE.It utilizes the average profile as a constraint to learn the robust features of the single HRRP sample,which involve detailed information of the sample and reflect the stable property of the corresponding average profile.In this way,we can obtain the robust and effective recognition features of the target.Finally,the experiment based on the measured HRRP data verifies the effectiveness of the proposed method.