In view of the problem that the useful classification information in principal component analysis （PCA）may be lost in the process of network fault feature extraction,a new method named center distance ration weighted principal component analysis（CDRWPCA）. According to sample category information, the center distance ratio of the difference between characteristics is measured by using this algorithm. By doing so, the weight is designed based on the feature discrimination. Then the weighted datasets are used for PCA feature extraction. Finally, the extracted datasets are sent to support vector machines （SVM）so as to verify the effectiveness of the algorithm. Experiments on network fault diagnosis demonstrate that the the proposed algorithm can improve the compression ratio and the final fault recognition rate.