Variational mode decomposition(VMD) method is a new time-frequency analysis method. It is an entirely non-recursive model, where the modes are extracted concurrently. However, VMD is sensitive to noise. By making use of this property, a new denoising method based on VMD is proposed. The permutation entropy(PE) is used to determine the amount of noise contained in each variational intrinsic mode function (VIMF), which contains high noise abandoned and low noise smoothed by Savitzky-Golay method. Then, the de-noised signal is decomposed by VMD and the fault features are extracted. The simulation and test results show that the denoising method based on VMD is better than the wavelet transform denoising method and the fault features can be extracted by VMD effectively.