The Research Of Two-stage Feature Selection Ensemble Classifier Based On Bagging

The ensemble classification is a multi-classifier ensemble method that utilizes the complementary relationship of the classifiers to improve the generalization ability effectively.Ensemble classifier becomes a new field in machine learning and has been widely used in many fields.The performance of the ensemble classifier is mainly restricted by two factors.One is that the accuracy of the individual classifier.The other is the diversity of the base classifiers.At present,in order to design a large difference between the members of the classifier,most of the algorithms are designed based on perturbation training set,such as Bagging and Boosting algorithm.How to improve the accuracy of the member classifier at the same time to obtain the bigger difference between the classifiers has become the hotspot of the current ensemble learning research.In this thesis,two-stage feature selection ensemble classifier based on Bagging is proposed,which can improve the effect of integration from two aspects,namely,the differential membership classifier and the improvement of the precision of the member classifier.The ensemble classifier can improve the accuracy and diversity of the member classifiers by perturbation training set with the two-stage feature selection.Then,we selectively integrate the classifiers with larger difference in composition.Finally using weighted voting ideas to produce classification results.This ensemble classifier has been used in the practical application of partial discharge fault diagnosis.It can effectively identify the type of partial discharge and evaluate the insulation state of transformer in a timely and accurate manner.The main duties include following aspects:(1)An improved ensemble classifier is proposed,which combines the Bagging algorithm and the two stage feature selection to perturbation the training set and construct different input spaces,so as to improve the diversity and correct rate of the individual classifier.(2)Use IAS algorithm to select the member classifiers with larger difference,and then use the genetic algorithm to select the optimal classifier.Finally,the final result is output by weighted voting method.Experimental results show that the ensemble classifier can obtain better classifier accuracy.(3)The improved ensemble classifier is applied in the practical problem of partial discharge fault diagnosis.Through the experiment,it chooses entropy method,mutual information method,feature increasing and decreasing method as the elements of the two stage feature selection,also SVM algorithm is selected as the construction method of member classifier.Experimental results show that the ensemble classifier can identify the type of discharge accurately.And it has been applications as a commercial transformer partial discharge fault detection products embedded blocks.