Aiming at the strong background noise involved in the signals of roller bearing and the difficulty to obtain fault frequencies in practice, a new fault diagnosis method is proposed based on dual-tree complex wavelet transform and singular value difference spectrum. Firstly, original fault signals are decomposed into several different frequency band components through DT-CWT; Secondly, Hankel matrix is constructed by the component which contains the fault information, and the singular value difference spectrum can be obtained after SVD. Then the maximum catastrophe point is used to identify the number of singular value reconstruction component. Finally, the fault frequency can be identified accurately by Hilbert envelope spectrum. The results of the experiments and engineering application show that the fault feature of roller bearing can be separated effectively and the fault feature were extracted, the feasibility and effectiveness of the method were verified.