A fault feature extraction method for hydraulic pump vibration signal based on singular value decomposition and wavelet packets analysis is presented. In the method, the noise signal is decomposed into noise distributes uniform vectors, and then through wavelet packets threshold de-noising on every vector, finally, reconstructing de-noised vectors. De-noised signal is decomposed by wavelet packets, and extracting every power of frequency bands. Take gear pump as an example, the feature of gear pump air pocket fault, gear abrasion, side plate abrasion, and normal state vibration signal is extracted by using the method. The result indicates that the method can extract the feature of gear pump effectively.