Bearing fault diagnosis method based on similarity measure and ensemble learning

Abstract

Ensemble learning aggregates the predictions of diverse predictors. Most existing ensemble learning methods gather the base classifiers based on the weight of accuracy. However, if the weight is adjusted only according to the correct rate of the base classifiers, it ignores the different types of errors caused by different working principles of the base classifiers. Blindly taking the correct rate of the basis classifier as the weight will increase the probability of the classifier making the same mistake and thus reduce the correct rate. To address this problem, this paper proposes a multi-round voting method based on similarity measurement. In the proposed method, the characteristic parameters of the data are extracted and PCA dimensionality reduction is used to train several base classifiers, which achieved high accuracy of about 90% at first. Then, we extract the confusion matrix of these base classifiers to facilitate the calculation of the similarity measurement matrix and the obtainment of the similarity matrix between base classifiers. Later, on the basis of the number of base classifiers and the values of the similarity matrix, we set a certain number of similar thresholds. Finally, according to the comparison of similarity value and similarity thresholds, a multi-round voting classifier is trained. By comparing the hard voting classifier, multi-round voting classifier and their base classifiers, the experimental results verify that the multi-round voting classifier has higher accuracy than other methods.