Abstract: This paper proposed a novel method based on Singular Value Decomposition(SVD) and Variable Predictive Model based Class Discriminate(VPMCD) for bearing fault diagnosis attemping to solve the difficulty of detecting fault signals,which are easily interference by noise and smooth signals at a early stage.Firstly,SVD method was used for vibration signal analysis that an effective separation between transient signal and background noise,smooth signal was conducted according to curvature spectrum and maximum variance ratio threshold of inter-class and intra-class.Secondly,transient features were extracted in both time domain and frequency domain,they were then used to construct mixed domain feature vectors which reflect the operation state of the bearing.The fault diagnosis model based on VPMCD method was then built using these vectors.The experiment results proved the effectiveness of this method of rolling bearing fault diagnosis.