Owing to the fact that the characteristic data of networks are huge and have redundancy and conflict each other

a method of data pretreatment based on fuzzy clustering is proposed. The center of each kind is used to be a core

with adapting the radius to sample from training samples. During learning of the characteristic data

Levenberg-Marquardt algorithm is imported in order to resolve the weaknesses of BP

such as large training times

slow convergence and incidental surge. Theoretical analysis and experimental results show that this method is efficient for network fault diagnosis.