Study On Target Recognition Method Of High Resolution Range Profile Based On Deep Attention Networks

High Resolution Range Profile(HRRP)is the vector sum of the target scattering point echo obtained by broadband radar signal projected on the radar line of sight.It can provide the distribution of target scattering points along the range direction under a certain radar angle of view.It contains important identifiable information such as target size,shape,orientation and distance.The target recognition technology based on HRRP has become one of the research hotspots in the radar target recognition.The traditional HRRP target recognition algorithm has complex steps,and needs to manually extract HRRP features and prior knowledge,so the degree of intelligence is low.In recent years,deep learning methods are widely used in HRRP target recognition field.This kind of methods do not need to manually design feature extractors and a large amount of professional knowledge,and can automatically extract features effective for classification from data.Aiming at the characteristics of HRRP and the defects of existing HRRP target recognition algorithms,this paper completes the research on HRRP target recognition method based on deep attention network.The contents of the paper are as follows:1.Aiming at the problems that the traditional HRRP target recognition algorithm needs manual feature extraction and classification design,which depends on expert knowledge and complicated steps.The automatic feature extraction and classification of HRRP based on Convolutional Neural Networks(CNN),Long Short Term Memory(LSTM)and Convolutional Long Short Term Memory(CNN-LSTM)are studied.CNN extracts features by modeling the local correlation of sequence data,and LSTM learns the temporal correlation of the target.CNN-LSTM uses LSTM to further extract temporal features based on the feature sequence extracted by CNN.Experiments based on HRRP sequences verify the effectiveness of the three methods.2.Aiming at the problems that the traditional depth network does not consider the separability difference of HRRP data in different regions,and can not effectively extract separable features,and the existing attention network depends on the inherent sequence of sequences and can not operate in parallel,a dual attention converter network Bi Attention Transformer(BiAT)is proposed,which is composed of channel spatial attention network Channel Spatial Attention Network(CSAN),It is composed of transform encoder Transformer Encoder(TE).CSAN takes convolution as the benchmark structure,and uses channel spatial attention to extract the preliminary features of the input HRRP sequence.Te extracts the high separability features and timing related features of different feature blocks in parallel and efficiently through the multi head self attention mechanism.Finally,experiments based on the measured HRRP sequences verify the effectiveness of BiAT.3.Aiming at the problem that the visual converter Vision Transformer(Vi T)and BiAT networks with self attention as the core do not consider the data characteristics of HRRP samples,and the information interaction between multiple attention heads and different attention heads,a horizontal fringe converter network Horizontal Stripe Transformer(HST)is proposed.Compared with the traditional attention and self attention,the horizontal stripe converter performs the calculation of self attention in the horizontal stripe of equal width,which not only reduce computing costs,but also achieve a better accuracy of target recognition.Finally,experiments based on HRRP sequences verify the noise robustness and advanced nature of HST.