Aiming at the problem that the difference among BP neural networks as the base classifier in Bagging ensemble learning is small， a feature perturbation method is Keywords ensemble learning； neural network； rotor； fault diagnosis introduced to improve the classification performance of the model of ensemble learning. Firstly， the Relief-F feature evaluation algorithm is integrated with the improved roulette wheel selection algorithm and the number of base classifiers is set to thirty. Next， thirty feature subsets， where the feature dimension are thirty， are selected from the rotor fault feature set. Then， the training set and the test set are respectively projected on the corresponding thirty fault feature subsets to obtain a series of training and test subsets corresponding to the thirty base classifiers， which realize the feature perturbation. Afterwards， each training subset is processed using the self-service sampling method （bootstrap sampling） included in the Bagging ensemble learning machine. Thus， they has certain differences in the feature space and sample set when they are finally input to each base classifier， which indirectly makes the trained base classifiers show higher differences， so as to achieve the purpose of making the final classification results more credible. Moreover， a low-dimensional double-span rotor fault data set is used to classify in the ensemble learning method. The results show that this method can significantly improve the accuracy of class identification of the BP network. In addition， it also has good performance in terms of anti-interference.