Variable predictive model based class discriminate (VPMCD) method was introduced to roller bearing fault diagnosis, and a roller bearing fault diagnosis approach based on empirical mode decomposition (EMD) and VPMCD was put forward.Firstly, different feature vectors were extracted with EMD.Then, different working conditions and failures of roller bearing were distinguished by using VPMCD.Analysis results of vibration signals from roller bearing's normal condition, outer ring fault and inner ring fault show the effectiveness of the proposed approach in roller bearing fault diagnosis.What's more, comparative analysis results demonstrate that VPMCD algorithm gains more stable classification performance and better computational efficiency than artificial neural network (ANN) algorithm.