In the early stages of rolling bearing fault diagnosis, the characteristic frequencies of different parts are weak and difficult to identify. Thus, this paper presents a method to enhance the interested frequency components of rolling bearing. This method has better results and higher resistance to interference frequencies than do traditional methods using phase-space reconstruction to expand the time-domain signal into a high dimension space and independent component analysis with reference (ICA-R) to find the closest component to the reference signal. The simulation results and actual measurements of signals prove that this method is effective and feasible in early stage fault diagnosis, and also has certain values for engineering applications.