In this paper, the application of nonlinear feature extraction based on wavelet kernel KPCA for faults diagnosis is presented. Mexican hat wavelet kernel is intruded to enhance Kernel-PCA nonlinear mapping capability. The experimental data sets of rotor working under four conditions: normal, oil whirling, rub and unbalance are used to test the WKPCA method. The feature reduction results of WKPCA are compared with that of PCA method and KPCA method. The results indicate that WKPCA can classify the rotor fault type efficiently. The WKPCA is more suitable for nonlinear feature reduction in fault diagnosis area.