Research On Online Learning Method Of Radar High Resolution Range Profile Target Recognition

Radar can realize the target monitoring,tracking and type judgment,and plays an important role in the improvement of air defense system and battlefield enemy situation judgment.In order to meet the needs of information warfare,the radar automatic target recognition(RATR)has been developed rapidly.High resolution range profile(HRRP)contains rich structural information of the target,which has the advantages of easy access and easy processing.It is an important way to realize RATR.In recent years,the deep learning theory has developed rapidly,and the neural network model has been widely used in HRRP target recognition,and has achieved remarkable results.The prerequisite of effective recognition by neural network model is to establish a complete radar target HRRP training database.In practical engineering application,it is usually difficult to establish a complete database in advance due to the factors of non-cooperation.For this reason,the model is required to be able to continuously learn with new samples during use.Existing methods are prone to catastrophic forgetting problems when using new samples to update model parameters,which seriously affects recognition performance.In response to this problem,the online learning method is introduced into the HRRP recognition task in this thesis,which enables the model to learn new knowledge while still retaining the old knowledge it has learned.The main research contents of this thesis are as follows:(1)First,starting from the scattering point model,the characteristics,sensitivity issues and preprocessing methods of HRRP are introduced.Then,several traditional HRRP recognition methods are introduced,and the traditional methods are used to carry out HRRP recognition experiments under the condition of a complete training database.Finally,several classic online learning methods are introduced,and their recognition performance is compared and analyzed.(2)Aiming at the application scenario of target sample expansion in the incomplete database,an online learning method for radar HRRP target recognition is proposed,which can be refer as adaptive comacting and growing(ACG).This method introduces a model retraining strategy.It can retrain parameters for the compressed model,and improves the model's ability to extract important features of the original sample.At this time,the feature discrimination of different targets is high,and the model recognition ability is strong.At the same time,this method introduces a group sparse regularization strategy.It can eliminate redundant nodes when the model is expanded,and overcomes the problem of blind growth of the model.The simulation experiment results show that ACG can extract more separable features,and it can improve the recognition rate and greatly reduce the model scale.(3)Aiming at the application scenario of target model expansion in the incomplete database,an online learning method for radar HRRP target recognition is proposed,which can be refer as balanced progressive segmented training(BPST).This method introduces an automatic sub-frame sampling strategy,which can obtain a more representative HRRP sample set.It makes the features extracted by the model closer to the true distribution of the sample in the feature space,which can help improve model recognition capabilities.At the same time,this method introduces an incremental parameter freezing rate strategy to allocate more sufficient model space for recognition tasks with a large total number of models.This strategy enhances the learning ability of the model.The simulation experiment results show that compared with similar online learning methods,the BPST method can better learn new knowledge and retain old knowledge,which improves the recognition rate.