A combined method was proposed which employs ensemble empirical mode decomposition (EEMD), singular value decomposition (SVD) and locality preserving projection (LPP) to extract useful fault feature from the complex vibration signals of rolling element bearings with problems of extraction in difficulty. Firstly, the vibration signals were decomposed with EEMD into a set of intrinsic mode functions(IMFs), which then were utilized to construct the timedomain and frequencydomain spatial condition matrix, as well as the timefrequency domain spatial condition matrix. Secondly, SVD was used to extract the fault information of multipledomain spatial condition matrix and among which selected the effective SVs which cumulative percentage were greater than 90% constituted the multipledomain SV feature matrix. Thirdly, LPP was used to extract the lowdimension and highseparability fault features from multipledomain SV feature matrix. Finally, support vector machine (SVM) was used to evaluate the fault feature extraction method proposed. The experimental result illustrate that the fault feature extracted according to this method can effectively reflect the fault patterns of rolling element bearings.