This paper presents a fault feature extraction method based on multifractal and singularity value decomposition for multi-sensor, according to the interference and coupling of fault information and complex non-linear, non-stationary characteristics of the vibration signal in reciprocating compressor. The generalized fractal dimension can characterize local scale behavior of signal more appropriately, so an initial feature matrix was built by calculating the generalized fractal dimension of multi-sensor signal. The matrix was compressed by singular value decomposition method, and the eigenvalues of matrix were taken as feature vectors. Taken reciprocating compressor transmission mechanism as research object, feature vectors of bearing clearance faults on different position were extracted from vibration signal. Support vector machine was established as pattern classifier to identify faults. Compared with results of single sensor multifractal method and multi-sensor fractal method, the validity of this method is proved.