Abstract: In allusion to deal with data which contains a lot of irrelevant and redundant features.A breeding binary particle swarm optimization-support vector machines (BPSO-SVM) algorithm was proposed for feature selection.In order to improve diagnosis accuracy and save computing resources of the network fault diagnosis system.The algorithm adopts w rapper mode,the classification accuracy of SVM and feature compression ratio as fitness function guide the breeding BPSO algorithm to search the feature space.Finally the best fitness subset was selected out.Experimental result on KDD’99 shows that the advanced algorithm improve the accuracy of diagnosis and reduce the feature dimension,and can further enhance network fault diagnosis effect.