Abstract: Gear crack level identification is of great significance for gear box fault diagnosis．Regarding instability and performance limitation in current identification,we generated a set of neighborhood separable subspaces based on randomized attribute reduction,on which a set of base classifiers were obtained．The weight vector of the base classifiers was learned by optimizing loss of ensemble margin and regularization learning to change margin distribution．Base classifiers were ranked according to the weight value and a set of classifiers that make the ensemble classification accuracy highest on the training set was got．The experiment analysis shows that the proposed method is much better than other state of the art methods in crack level identification.

Key words: Neighborhood rough set,Randomized reducts,Ensemble learning,Margin distribution