A new deep learning method is proposed to automatically learn the useful fault features from the raw vibration signals. A new deep auto-encoder model is constructed for the enhancement of feature learning ability. Locality preserving projection is adopted to fuse the deep features to extract the most representative information.