Application Of Ensemble Learning In Emitter Identification

In this thesis,ensemble learning algorithm is applied to the problem of emitter identification.There are three typical problems in classifier design,including feature selection,SNR change and how to deal with new samples.These problems are discussed chapter by chapter,and the corresponding solutions are given,of which the effectiveness is proved by simulation experiments.In order to solve the more complex problem of emitter identification,more emitter features must be introduced into the classifier,which will lead to the overburden of signal processing flow.Moreover,the increase of the dimension of the feature vector is easy to cause a large increase in the computational effort during the training of the classifier.Since ensemble learning algorithm consists of many base classifiers,the influence of computational effort on the algorithm will be more obvious.According to the feature selection mechanism of CART algorithm,this paper proposes an encapsulated feature selection algorithm.Simulation results show that the proposed algorithm can achieve a relatively ideal effect of feature selection.Compared with the classical algorithm Relief,the proposed algorithm requires fewer features to maintain the original classification accuracy on the data set of emitter sources,and the dimensionality reduction effect is better.In addition,due to the strong physical significance of emitter characteristics,three feature evaluation indexes are proposed to further enhance the interpretability of emitter classification process.The simulation results also show that these three feature evaluation indexes can effectively quantify the role of a certain feature in the classifier.The change of SNR has a significant effect on the numerical distribution of emitter features,which directly affects the accuracy of the classifier model.To solve the above problems,this paper proposes solutions from two directions: diversity of ensemble and variable weighted ensemble.Firstly,the base classifiers were constructed by means of sample autonomous sampling and random feature combination.And a diversity enhancement method based on different SNR is proposed.Then,a variable weighted ensemble algorithm based on the SNR is proposed,which dynamically adjusts the weight of the base classifier by the SNR estimation results of the samples to be tested,so as to adapt to the fluctuation of the SNR.The simulation results show that both diversity of ensemble and variable weighted ensemble can effectively improve the classification performance of the classifier under different SNR,especially the random forest algorithm can achieve the highest global accuracy when the number of base classifiers is large,but the performance is poor when the number of base classifiers is small.The variable weighted ensemble algorithm can achieve good classification accuracy even when the number of base classifiers is small,but the final classification effect will be affected by the SNR estimation error.Conventional classifiers are not only unable to detect new emitter samples,but also can only acquire the recognition of new samples through retraining,which is inefficient.To solve this problem,this paper proposes a lightweight incremental ensemble learning algorithm.The improvement of open-set recognition ensures that the CART base classifiers will only make effective decisions on the known samples,otherwise "abstentions" will be cast.So we can add base classifier to expand the category which can be recognized by ensemble learning algorithm.The simulation results show that the algorithm can achieve the linear computational complexity of the number of categories,greatly reduce the training cost of adding new categories,and maintain relatively high accuracy within a certain SNR range.