A new non-stationary signal processing technique called Hilbert vibration decomposition (HVD) is introduced to fault diagnosis of roller bearings. The HVD and empirical mode decomposition (EMD) are both based on Hilbert transform, and both methods can decompose multi-component signals adaptively. However, compared with EMD, the HVD method does not involve spline fitting and empirical algorithms and has a better frequency resolution. Moreover, the HVD method can decompose more effectively the multi-component signals which can cause mode mixing while decomposed by the EMD method. Based on this consideration, the HVD method is applied to the experimental data of roller bearing with induced faults. The envelope analysis is performed to the component including dominant fault information, and then the characteristic defect frequency of roller bearing can be identified by means of the envelope spectrum. The experimental results validate the effectiveness of the proposed method for roller bearing fault diagnosis.