Research On Open Set Recognition Of Radar High Resolution Range Profile Target

High resolution range profile(HRRP)characterizes the one-dimensional distribution of the target scattering center along the line of sight of the radar.It contains structural information such as the radial size of the target and the distribution of the scattering points,which is extremely valuable for target recognition and classification.HRRP has the advantages of small footprint and easy access,and is an important way to realize radar automatic target recognition(RATR).In radar HRRP target recognition,HRRP databases of various types of targets need to be established in advance.Most traditional methods assume that the database covers all types of targets to be identified,and use closed-set identification methods to complete the identification tasks.In practice,since some of the targets of interest are non-cooperative targets,it is relatively difficult to obtain HRRP data,resulting in incomplete target models in the database.In this case,the recognition task needs to be treated as an open set recognition problem.This thesis focuses on the open set recognition technology of radar HRRP targets,and conducts research on the rejection of out-of-library targets,the splitting of data within the class,and the open set recognition.The main research contents and innovations are as follows:1.First,explain the development status of RATR and HRRP,analyze and introduce the sensitivity issues and common features of radar HRRP target recognition technology;then,summarize the HRRP open set recognition problem and the history and research status of the open set recognition problem;Next,the basic theories such as the discriminant mode and generation mode of open set recognition and two classic open set recognition methods are described to lay the foundation for follow-up research.2.Aiming at the problems of traditional rejection methods based on one-class classifiers,such as the difficulty in optimizing nuclear parameters and the limited rejection performance,a method for rejecting targets outside the radar HRRP library based on intra-class splitting is proposed.First,use the autoencoder to reconstruct all training samples to obtain more main components representing the original information while retaining more nonlinear features;secondly,calculate the similarity score between the original sample and the reconstructed sample,using segmentation The ratio criterion splits the training samples into typical and atypical sample subsets,and uses the latter to model targets out-of-library.This approach uses only part of the training sample data to model the samples out-of-library,without any prior knowledge;finally,the distance constraint between the two sample subsets is introduced as a regular term into the cost function,forming a closed and dense Decide the boundary to improve the rejection performance.The simulation experiment results show that compared with the commonly used rejection methods,the proposed method has a significant improvement in rejection performance indicators such as AUC value and ROC curve.3.Aiming at the problem of poor performance of existing open set recognition methods directly used in HRRP recognition tasks,a radar HRRP target open set recognition method based on dynamic intra-class splitting is proposed.The innovation of this method lies in:(a)During the training process,the division of typical and atypical sample subsets is dynamically adjusted to increase the randomness of atypical samples to improve the robustness of the open set recognition performance;(b)through The single-stage strategy of intra-class splitting and deep neural network alternate training realizes the integrated operation of data splitting and open set recognition,which significantly reduces the computational overhead of the training stage;(c)Adopting closed-set regularization subnetworks to achieve pairing While efficiently rejecting the target outside the library,it achieves high accuracy of closed set recognition.The simulation experiment results show that compared with the classic open set recognition method,the proposed method has a significant improvement in the open set recognition performance indicators such as Accuracy value,AUC value,F1-Macro value,F1-Weighted value.