Research On Generation And Recognition Of Radar HRRP Signal Under Sample Missing Condition

Radar High Resolution Range Profile(HRRP),as the vector sum of the target scattering point echo on the radar line of sight,contains the geometric size and physical characteristics of the target,and has the advantages of easy acquisition and simple calculation,so it has become an important part of radar target automatic recognition.However,HRRP has translation sensitivity,amplitude sensitivity and attitude sensitivity,which make it difficult to obtain high quality HRRP data with complete target.In the actual scenario,the non-cooperative nature of the target and the complexity of the environment will also lead to the problem of limited sample quantity and low sample quality.Due to the small amount of data or uneven data distribution,the data is incomplete,which restricts the application of data-driven recognition model and algorithm in HRRP target recognition to a certain extent,resulting in the performance of recognition algorithm degradation and poor generalization ability.At this point,the depth generation model is used to expand the data,meet the training requirements of the recognition model,and provide an effective way to improve the performance of target recognition.In order to improve the recognition performance,this thesis discusses and studies the generation method of radar range image under the condition of incomplete data.The main work contents and innovation points are summarized as follows:(1)Aiming at the problem of incomplete data with a small number of samples,the HRRP generation method of conditional one-dimensional convolution generation adduction network based on attention mechanism is proposed.Based on the deep convolution generation adversarial network,this method reduces the training parameters and replaces the pooled form of fixed sampling by means of step convolution,and adds category condition information to guide the generation process.At the same time,a feature refining module composed of self-attention mechanism is introduced to work cooperatively with the convolutional network which extracts target features layer by layer,so that the network adaptively strengthens the remote dependence between HRRP scattering points and optimizes the global geometric features of samples.The experimental results show that this method can stably generate samples with higher quality and better diversity,and can effectively improve the classification ability of the recognition model after adding the generated samples to expand the original data set.(2)Aiming at the problem of incomplete data with fewer or missing attitude angles in the sample part,an HRRP generation method based on mutual information constraints of attitude angles was proposed.Aiming at the difficulty of introducing continuous attitude corner tags into the generation task,an attitude corner tag embedding mechanism is designed to introduce it into the model non-invasively.At the same time,the supervised auxiliary classification generating adversarial network and the unsupervised mutual information generating adversarial network are combined to maximize the mutual information between the attitude Angle and the generated samples.The least square loss function is introduced to improve the original loss function and optimize the training steps of the model.The experimental results show that this method can generate the data of the specified azimuth and complete the missing azimuth to a certain extent.Moreover,it can effectively improve the classification ability of the recognition model after adding the generated samples to expand the original data set.