Research On Online Learning Method Of Aircraft Target Classification

In modern warfare,using radar echoes to obtain the category of aircraft target is of great significance for timely and accurate battlefield decision-making.The physical structure and motion parameters of the micro-moving components on helicopters,propellers and jets are different,which can produce different micro-Doppler modulations on radar echoes.Different modulations can be used to classify aircraft targets.Considering the non-cooperation of some targets and other factors,it is usually difficult to obtain a training sample set that meets the needs in the actual situation,which limits the classification performance.The traditional method is to store new samples collected during the normal operation on the radar,periodically merge the old and new sample sets,and retrain the classification model.However,this method has low learning efficiency and wastes resources.A more feasible idea is to combine online learning algorithms with traditional classification methods,which automatically update the original model with new samples collected to continuously improve the classification accuracy of the model.Therefore,we propose three aircraft target classification methods based on online learning.The main work is summarized as follows:1.First,we establish the rotor echo parameter model and analyze the difference in microDoppler modulation produced by the three types of aircraft;then,the aircraft target classification method based on the micro-Doppler effect is introduced;finally,through simulation experiments,the impact of insufficient or incomplete training sample set on the classification task is analyzed.2.Aiming at the problem that traditional classifiers cannot automatically update the model,we introduce the Self-adaptive Incremental Support Vector Machine(SD-ISVM)and Mondrian Forests(MF)algorithm,and propose two online learning methods of aircraft target classification.The SD-ISVM-based classification method selects the samples used to construct and maintain the Support Vector Machine(SVM)classification hyperplane from the original retained set and the newly added sample set as the new retained set,and uses it to update the model.The classifier used in the MF-based classification method is composed of multiple Mondrian trees.Applying the new sample set can directly expand the structure of each tree to complete the update of the original model.The experimental results show that the proposed methods can use newly added samples to effectively improve the classification performance.3.In view of the classification performance of traditional classifiers on an adequate and complete training sample set reaches a bottleneck.We introduce a method of aircraft target classification based on Convolutional Neural Networks(CNN).When a new sample set is collected,it is costly to retrain the network model by the offline learning method,and if the new sample set is used to update the network model directly,it will cause the model to forget the original learning results.Therefore,we propose an aircraft target classification method based on CNN online learning.This method adopts the Elastic Weight Consolidation(EWC)algorithm.When learning a new task,the importance of each node parameter in the model to the original task is evaluated,and the important node parameters are consolidated during model update,so that the updated model has good classification performance on the original task and the new task.Experimental results show that the proposed method can not only solve the problems of the offline learning method,but also obtain higher classification accuracy on large sample sets than the SD-ISVM-based classification method.