Wavelet packet decomposition (WPD) can effectively reflect the potential signal characteristics by decomposing the non-stationary signal into the low and high frequencies. Locality preserving projection (LPP) can retain the local features of the analyzed signal in dimensionality reduction. Combining these two benefits, this paper selects the spectral energy of all nodes with WPD as a characterization of the analyzed signal and uses LPP feature extraction to reduce dimensions for pattern recognition of machine faults. Experimental results show the effectiveness of the proposed method by using multiple multi-class dataset of bearing faults with different fault types and defect severities.