Research On HRRP Generative Data Augmentation Method Based On Feature Decoupling

With the rapid development of deep learning,radar High Resolution Range Profile(HRRP)recognition algorithm based on deep neural network has achieved good results in Radar Automatic Target Recognition(RATR)task.Due to the high attitude sensitivity of HRRP signal,this kind of radar target recognition model based on deep learning often needs sufficient full attitude HRRP data as training samples.However,in practical application,there is often a phenomenon that the uncooperative target data is missing and cannot capture the full attitude HRRP data.As a result,the performance of the data-driven deep network in such tasks drops sharply.In order to alleviate the attitude sensitivity of HRRP data and the degradation of model performance caused by missing data,the existing schemes alleviate the problem of small samples from the data level,that is,using the Generative Adversarial Network(GAN)to expand the data.However,the feature source of the existing generation algorithms is single,resulting in less diversity of generated data.Based on the existing research,this paper increases the data source of the generation algorithms,introduces the simulation data,and proposes the HRRP expansion algorithms based on feature decoupling,so as to expand the small sample target feature space and improve the recognition accuracy.The main research contents and innovations of this paper are as follows:firstly,on the basis of the existing augmentation scheme for small sample HRRP,the simulation data is introduced for auxiliary training for the first time,and the real domain samples are generated from the simulation data by means of feature transformation and other ways.Secondly,this paper uses the current commonly used image translation models for reference,and applies them to the data augmentation task of HRRP,also researches the adjustment schemes of these models.According to the characteristics of HRRP data and application requirements,the generation scheme is improved,which alleviates the difficulty of small sample recognition to a certain extent.Thirdly,according to the experimental results of the existing models,this paper proposes a radar data generation algorithm based on feature decoupling,which mixes the complete attitude information of the simulation data with the real data,and introduces more attitude diversity for the generated data.A large number of experiments on real and simulated data sets show that the proposed scheme has the best performance in the small sample HRRP expansion task,and further improves the recognition performance of the existing deep recognition model in the case of missing data.