Study Of Radar HRRP Target Recognition Based On Detailed Statistical Modeling

Radar high-resolution range profile(HRRP)is the amplitude of the coherent summations of the complex time returns from target scatterers in each range resolution cell,which represents the projection of the complex returned echoes from the Radar-Cross Section of target sacttering centers onto the radar line-of-sight.It contains the abundant physical structure information of targets.Compared with synthetic aperture radar image or Inverse SAR image,the HRRPs have been extensively studied and successfully applied to the radar target recognition community with the property of easy acquisition and processing.The main contents of this dissertation are summarized as follows:The first part makes a brief analysis on the physical characteristics of radar HRRP echoes.Three sensitivity problems of HRRP samples and how to deal with the above problems are described.Next,starting from the basic concepts of Bayesian theory,we introduce the common inference methods of model parameters based on Bayesian theory.Then basic steps of statistical modeling for HRRP samples are briefly illustrated.In the second part,the detailed statistical modeling methods of the HRRPs is studied from the perspective of the description of data distribution.After briefly introducing the HMM-Gaussian model,we point out that the Gaussian distribution of the observation probability is assumed by the HMM-Gaussian model.However,considering that the non-Gaussian property of HRRP echoes,the Gaussian distribution assumption may not be able to accurately describe the data.Therefore,a hybrid HMM-SVM model combining Hidden Markov Model(HMM)and Support Vector Machine(SVM)is proposed.The HMM-SVM hybrid model no longer assumes data distribution of the observation,but utilizes nonlinear classification ability of the kernel SVM to provide the observation probability of the HMM with the probability outputs of SVM,so that the HMM-SVM hybrid model is able to better fit the non-Gaussian observation data.Finally,experiments based on synthetic data and measured data show that the proposed model is an effective target recognition method.The third part studies the detailed statistical modeling methods of the HRRPs from the perspective of introducing prior information.Focusing on the Factor Analysis(FA)model,the traditional FA model and multi-task learning FA model for radar HRRPs are briefly introduced.Considering that above two models only use the generative information of HRRP samples,this dissertation proposes a similarity preserving FA(SP-MTL)model based on the multi-task learning FA model.The SP-MTL model introduces similarity constraints on the basis of the multi-task learning FA model.With the introduction of discriminative prior knowledge,the prior knowledge can be utilized to assist in model learning when multi-task learning FA model is insufficient to accurately describe the data,so that the estimated accuracy of model parameters are improved.In addition,this dissertation also applies the SP-MTL model to model prediction based on the transfer learning theory.Using the idea of transfer learning,the shared model parameters in the original modeling phase were transferred to the statistical modeling process of new observed data,which plays an auxiliary role in model prediction.