As a kind of complicated mechanical component, rolling element bearing plays a significant role in rotating machines, and bearing fault detection benefits decision-making of maintenance and avoids undesired downtime cost. However, extraction of fault signatures from a collected signal in a practical working environment is always a great challenge. This paper proposes an improved combination of the Hilbert and wavelet transforms to identify early bearing fault signatures. Real rail vehicle bearing and motor bearing data were used to validate the proposed method. A traditional combination of Hilbert and wavelet transforms was employed for comparison purpose. An indicator to evaluate fault detection capability of methods was developed in this research. Analysis results showed that the extraction capability of bearing fault signatures is greatly enhanced by the proposed method.