The rolling bearing vibration signal can be decomposed into a number of intrinsic mode functions (IMF) adaptively according to its own scale by the empirical mode decomposition (EMD) method. The rolling bearing failure will vary the distribution of IMF energy, and the bearing fault diagnosis can be realized by establishing the relationship between the IMF energy distributions and bearing conditions based on the gray relational grade theory. In order to make up for the defects of the traditional gray analysis in pattern recognition, a gray similar relational grade model which can reflect the curve’s shape feature was proposed based on the similarity of the slope. Then combined with the traditional approaching relativity model, the gray comprehensive relativity diagnosis model which can reflect both the curve’s position and shape features was constructed. The simulation results show that the new model can recognize rolling bearing faults more effectively and accurately．