Research On Radar Signal Recognition Based On Ensemble Learning

Electronic Countermeasures are the most important manifestation of modern information warfare.And radar signal recognition is the key to electronic defense and electronic interference.Traditional radar signal recognition technology that manually extracts physical characteristics from signal has been unable to work effectively in an increasingly complex electromagnetic environment.Therefore,this paper studies data-based machine learning models with a large amount of radar data,and focuses on the use of ensemble learning to establish relevant models for intentionally and unintentionally modulated radar signals.The proposed models can automatically extract signal characteristics,and identify radar signals effectively and quickly.The main work of this paper is:(1)Aiming at the problem of radar mode recognition,a Variational Relevance Vector Machine(VRVM)model is first established,and then an improved Chaotic Gravitational Search Algorithm(CGSA)is introduced to optimize the hyperparameters of the VRVM.The threshold optimized ensemble strategy(TOVO)forms the final ensemble model CGSA-VRVM-TOVO.Experiments show that the established radar operating mode recognition model can effectively distinguish the four operating modes.(2)For the identification of radar specific emitters,the ensemble model of decision tree Light Gradient Boosting Machine(LightGBM)is used as the basic model,and then the ensemble strategy of Stacking is introduced to form the cascade model LightGMB-LR.Finally,the minimum time scale component extracted by the empirical mode decomposition method form the final model IMF-LightGMB-LR.Experiments show that the established radar specific emitter recognition model can effectively improve the recognition accuracy and reduce a lot of training time.(3)Aiming at the problem of identification of radar specific emitters with imbalanced data,a method called Loss_weighted is proposed to mitigate imbalance,which assigns weights to each class according to class loss.Experiments show that Loss_weighted can balance the recognition effect of each class,so that the variance of the recall rate of each class is smaller than other methods.