Aimed at the phenomenon of mode mixing of LMD extracting the fault information from the elevator guide shoe, a feature extraction method of the elevator guide shoe vibration signal based on singular value decomposition(SVD) optimizing local mean decomposition(LMD) is proposed. First of all, the method structures the Hankel matrix of original signal with singular value contribution principle. It Uses SVD to decompose the Hankel matrix. Then, in order to obtain the useful singular value which contains the main fault information, the method selects singular values with the principle of curvature spectrum and the principle of singular value contribution. The selected singular values come into being mutation signal through reconstructing. Finally, the fault feature component which represents the original signal vibration characteristics could be extracted from the mutation signal by LMD method. The method was applied to diagnosis the elevator up guide shoe fault, the elevator down guide boots fault. Application examples show that this method is effective to improve the mode mixing of LMD and extract the fault characteristic components from the elevator guide shoe vibration signal more accurately .