Study On Radar Target High-resolution Range Profile Simulation And Recognition

Radar automatic target recognition (RATR) plays an important role in military and civilian field.With the development of wideband radar techniques, more and more useful target information can beobtained, and it provides strong supports for RATR. As a form of wideband radar target returned ech-oes, high-resolution range profile (HRRP) is the amplitude distribution of returned echoes along theradar line-of-sight, and it contains target structure information. Moreover, HRRP has the advantagesof easy acquisition and processing, thereby radar target HRRP recognition is of great value in theRATR community. This thesis focuses on radar target HRRP recognition, and mainly consists of twoparts. In the first part, high-frequency electromagnetic scattering calculation and HRRP simulation forcomplex radar targets are studied and corresponding valuable software is also developed for engi-neering, which aims at providing relatively high quality data support for radar target HRRP recogni-tion. In the second part, based on the simulated HRRP dataset, radar target HRRP recognition is re-searched from tree aspects for solving HRRP target aspect sensitivity problem, including feature scal-ing coefficients optimization, semi-parametric statistical modeling, and subspace feature extraction.Altogether, the main purpose of this thesis intends to make a useful exploration and technology ac-cumulation for RATR engineering application in the future.In summary, the main contents and contribution of this thesis are listed as follows:1. Research and development of radar target HRRP recognition can not do without the support ofHRRP data. In this thesis, HRRP dataset is established by electromagnetic scattering simulation. Ac-cording to the feature of electromagnetic scattering calculation for radar targets and for the purpose ofRATR application, high-frequency and asymptotic method for radar target high-frequency electro-magnetic scattering calculation and HRRP simulation is adopted. In addition, near field high-fre-quency electromagnetic scattering and HRRP simulation for extremely electrically large radar target,like ship target, is also studied. Meanwhile, KdTree based ray tracing algorithm and high performancecomputing (HPC) based task-level parallelism are both used for calculating acceleration. Finally, onthe basis of theoretical work, the radar target high-frequency electromagnetic scattering characteristicsanalysis software containing geometric modeling, high-frequency electromagnetic scattering calcula-tion, and data analyzing is developed.2. Feature scaling coefficients optimization based HRRP recognition is studied. For the reasonthat feature scaling HRRP template can be seem as a template which is independent with statistic model of HRRP returned echoes, it changes the geometric distance metric compared with originalHRRP feature space by coordinate-scale transformation, and avoids statistic model establishing forHRRP returned echoes. In this thesis, by defining the separability criterion of HRRP patterns in thetransformed feature scaling space, a HRRP feature scaling coefficients optimization objective functionis designed, and a HRRP recognition approach based on feature scaling optimization is proposed. Be-sides, feature scaling coefficients optimization based on kernel method is also studied. This thesisprovides a HRRP feature scaling coefficients optimization approach in view of the support vectormachine (SVM) classifier. The given approach works well owing to the consistency between kerneloptimization and SVM solution.3. Semi-parametric statistical modeling based HRRP recognition is studied. According to theclass-conditional probability density model selection problem caused by the complex statistic distri-bution of returned echo in each HRRP range cells, a radar HRRP recognition approach based on thesemi-parametric probability density is proposed. The proposed approach unifies the statistic distribu-tion model, and both advantages of parametric method and nonparametric method are merged in thesemi-parametric density estimation. However, when the Parzen window based semi-parametric isused and large data quantity is appeared, the execution efficiency reduces. In order to solve this prob-lem, a semi-parametric probability density approach based on kernel method is proposed. The pro-posed approach reduces the needs of samples for probability density function representation, and im-proves the computational efficiency.4. Subspace feature extraction based HRRP recognition is studied. In this thesis, two subspacefeature extraction methods including kernel fisher discriminant analysis (KFDA) and kernel principalcomponent analysis (KPCA) is applied to HRRP recognition. For KFDA, considering the numericalinstability problem when optimizing kernel parameters using Fisher’s discriminant criterion, the lowerbound of Fisher’s discriminant criterion is used as the kernel parameter optimization objective func-tion for KFDA, and applied it to HRRP feature extraction and recognition. For KPCA, an approachbased on kernel principle component analysis reconstruction is proposed. To this approach, the type oftest sample is determined by the minimum reconstruction error instead of distance measure, whichsolves the HRRP non-Guassian distribution problems and relaxes the angle division rules.