Aiming at the difficult problem of fault feature extraction in locomotive bearing fault diagnosis, the empirical wavelet transform (EWT) was introduced into locomotive bearing vibration signal analysis. Empirical wavelet transform decomposed the signal into multiple intrinsic modal components by constructing compactly supported adaptive filters, which could effectively suppress modal aliasing. In this paper, the empirical wavelet transform (EWT) was improved for bearing vibration characteristics. Firstly, the EWT was used to decompose the vibration signals of locomotive bearings, then the sensitive components were screened by kurtosis, and then the Hilbert envelope demodulation was used to extract the fault features of bearings. The locomotive operation experiment showed that the proposed method is reasonable in dividing the frequency band of vibration signal of locomotive bearing, and can effectively extract the characteristic frequency of bearing fault and accurately diagnose various types of bearing fault.