The training process of Variable predictive model based class discriminate (VPMCD) approach is to establish the predictive model with Polynomial Response Surface (PRS) method. However, the model fitting precision of PRS can't significantly improve with the increase of the capacity of training sample. According to this defect, the PRS method is improved in this paper and the method of VPMCD based on Improved Polynomial Response Surface (IPRS) is proposed, and then applied to rolling bearing fault diagnosis. Through the experiment, the pattern classification accuracy of the original method and the improved method are compared under the condition of different training sample sizes. The results show that compared with the original VPMCD method, improved VPMCD method not only has better effect of pattern classification, but also its classification precision enhances more obviously with the increase of the training sample size.