Abstract: Applying lots of primal features to identify fault condition leads to reduce classification correctness. Feature selection can remove redundant features in the primal features to enhance the effect of diagnosis. Filter? method which was widely applied before isn’t satisfied with the further demand for diagnosis correctness. A feature selection method based on wrapper model was proposed. The approaches to the subject include： genetic algorithm for optimal feature subset selection; k-fold cross-validation method for error rate evaluation ana BP neural network for fault classification. As an example, the results of bearing fault diagnosis prove that the method possesses excellent optimization feature subset property, and obtains high correctness rate.

Key words: information processing technique , feature selection , wrapper model , genetic algorithm (GA) , neu?ral network