Radar High Resolution Range Profiles Target Recognition With Continual Learning

High-resolution range image(HRRP)data obtained through wideband radar has been widely used for radar target recognition.Because HRRP is of high resolution,contains fine features and sufficient information of targets,and it is easy to obtain and process.As a research hotspot in recent years,deep learning can discover the inherent laws of data in complex environments.And the HRRP target features extracted by deep neural networks can improve recognition performance.However,in the actual target recognition scenario,the radar echo data require updating in real time,and tasks may change with the environment.In order to ensure the real-time update efficiency and feature memory ability and avoid the catastrophic forgetting phenomenon under continuous tasks,it is necessary to explore continual learning algorithms to achieve optimal performance.Therefore,this paper focuses on continual learning algorithms and studies radar HRRP target recognition with continual learning.The specific content is as follows:1.The basic concepts of HRRP data are introduced,then three kinds of data sensitivity and corresponding preprocessing methods of HRRP are analyzed.The development of neural networks for extracting HRRP features is introduced,and continual learning methods based on deep neural networks are discussed.2.The HRRP target recognition methods with continual learning based on parameter memory are studied in this paper.Two continual learning methods that memorize the characteristics of previous tasks through parameters are introduced,which are Elastic Weight Consolidation(EWC)and Variational Continual Learning(VCL)respectively.EWC is based on biological synaptic mechanisms and flexibly retains information from old tasks according to the importance of parameters.However,the network weights of EWC based model are constant,leading to limited characterization ability.And VCL further introduces weight uncertainty through variational distribution,and combines online variational inference and Monte Carlo variational inference to enhance the model performance with complex tasks,making it capable of continual learning.Based on the characteristics of radar HRRP data,a continual learning HRRP target recognition method based on VCL framework is studied in the paper.Experiments on MNIST dataset have confirmed the continual learning ability of EWC and VCL,and those on HRRP measured data have verified that the HRRP target recognition methods with continual learning based on parametric memory are effective.3.The HRRP target recognition methods with continual learning based on dynamic networks are studied in this paper.This paper first introduces two kinds of network structures for continual learning,Progressive Neural Network(PNN)that adds sub-networks to retain pre-trained models,and Dynamically Expandable Network(DEN)that performs selective node sharing and network expanding.And they have been applied to radar HRRP target recognition.In the existing dynamic networks,there are so many parameters to learn that comes to structural redundancy.To address the problem,this paper proposes a self-learning algorithm for hyperparameters without manual finetune,which can automatically select the appropriate hyperparameter configuration and the optimal network structure to efficiently control the parameter size.Further,the Dynamic Network Self-learning(DNSL)method is proposed for continual HRRP target recognition.Two kinds of continual learning methods respectively based on parametric memory and dynamic network are compared and analyzed through MNIST dataset and HRRP measured data.Experimental results prove that the method proposed in this paper can reduce the amount of network parameters while improve the continual learning property of deep networks,and thus achieve high performance on continual HRRP target recognition.