Aiming at the defects of parameter estimation in VPMCD, BP neural network nonlinear regression method was used instead of the least squares method to solve the ill-conditioned problem that exists in the least square method. Therefore, a fault diagnosis method for rolling bearing based on improved Variable Predictive Mode on the basis of Class Discriminate (VPMCD) was proposed. Firstly, Ensemble Empirical Mode Decomposition (EEMD) approach was used to decompose the rolling bearing vibration signal into a number of single components; and then, the singular values were abstracted from the component matrix and formed feature vector which will act as an input in the improved VPMCD; finally, the work states and faults pattern of the rolling bearing can be identified. The analysis results from the experimental rolling bearing vibration signals have demonstrated that the proposed method can be effectively applied to the rolling bearing fault diagnosis.