Study Of Radar Target Recognition Algorithm Based On High Resolution Range Profile

Radar automatic target recognition technology is an important technique component of modern radar system,which indicates one of the main developing direction of radar technology.With the improvement of equipment informatization and intelligence,RATR will affect battlefield situation in a deeper level and accelerate the renewal of the concept of war.Target recognition technology based on high resolution range profile is a mainstream and promising radar automatic target recognition technology.This technique uses wideband signals to obtain high resolution projection of target scattering centers at radial distances,which is a direct reflection of scattering centers,and is related to target size,shape,and surface material characteristics.Therefore,once the high resolution range profile can be reasonably analyzed,it means that the classification and discrimination can be realized to a certain extent.By considering both the theory and engineering application background of radar high resolution range profile recognition,this dissertation,which is supported by National Science Foundation of China,carries out the corresponding research from the database construction,data preprocessing,feature extraction,classification,statistical modeling and robust identification.The main work and progress of this thesis are summarized as follows:(1)The correspondence between the multi-scattering point model and the physical characteristics of HRRP is studied,and a more complete template database of HRRP which covers full azimuth angle is established,mainly to solve the problem that the performance of radar automatic target recognition algorithms is limited by the sample data base.Utilizing this database,an adaptive wavelet threshold function de-noising algorithm is designed to obtain high SNR HRRP in the preprocessing step.The algorithm can adaptively adjust the threshold according to the input HRRP signal,and get over the weakness of rescaling the estimated wavelet coefficients through soft threshold function;it does not produce a new mutation point like the hard threshold function,making the signal distorted,and the application is more flexible.(2)Several nonparametric data compression algorithms which can be applied to HRRP feature extraction are proposed to address four inherent defects of classical linear discriminant analysis algorithm extended to HRRP feature extraction.Based on the local characteristics of the sample distribution,this nonparametric method utilize nearest neighbor means or nearest neighbor to discretize the parametric scatter matrix to make feature extraction problem transform into the eigenvalue decomposition problem.Simulation results show the nonparametric method extends the application area of HRRP feature extraction,increase the separability of heterogeneous data,and has the advantages of high recognition rate and good stability.(3)Under the foundation that four classical feature extraction algorithms based on wavelet transform are applied to the feature extraction of HRRP,an adaptive wavelet neural network feature extraction and recognition algorithm based on evolutionary algorithm is proposed,mainly to settle the problem that the gradient descent algorithm depends on the initial value,converges easily to the local extremum and has low computational efficiency when calculating the parameters.This algorithm not only can be used just for feature extraction of HRRP,but also output the classification results.Moreover,this method utilizes the multi-resolution analysis property of adaptive wavelet network,which can increase the difference of extracted features and improve the recognition rate.Simulation results show that the recognition rate is close to 90% and the processing speed can be improved.(4)A HRRP statistical identification method based on Mixtures of Probabilistic Principal Component Analysis model is proposed to solve the problem that direct calculation of HRRP statistical parameters is inaccurate and the low computation efficiency when the sample number is small.In this method,the classical PPCA model is extended based on the probabilistic combination property of Gaussian mixture probability model.So the model parameters can be estimated by the expectation maximization algorithm,therefore the HRRP can be re-divided according to the clustering property of the mixture probability model(which can be regarded as the criterion of the maximum a posteriori probability)to weaken the HRRP azimuth sensitivity.Moreover,the MPPCA model introduces the latent variable parameter space,the free parameters that need to be estimated are greatly reduced thus the computational efficiency is improved.Simulation results show that the MPPCA model has better recognition performance than PPCA model.(5)The model parameter compensation method and t distribution based PPCA model are both put forward,these two methods are mainly aimed at the noise sensitive defects of PPCA model.In parameter compensation method,by estimating the noise energy in advance,the vector relation between the noise and the signal are utilized to compensate the classical PPCA model to achieve the noise robust purpose.The simulation results show this method can weaken the recognition rate decline caused by different environmental factors through reducing the difference between training sample's PPCA model and test sample's PPCA model.t distribution based PPCA model can utilize the noise robustness properties of t distribution,and can be further extended to mixture t distribution based PPCA model.In the numerical experiments,the validity of the mixture t distribution PPCA model is verified by accurately describing the covariance matrix of the artificial data.In the simulation experiments,by analyzing the clustering effect of HRRP in noisy environment,it is further proved that the mixture t distribution model is robust to noise and weakens the azimuth sensitivity of HRRP at the same time.