Variable predictive model based class discriminate (VPMCD) establishes the feature learning model by using the intrinsic correlation of the sample sets. The model can solve the problems with nonlinearity and high orders, but leads to inaccurate model prediction when the number of the samples is small. The thesis puts forward a fault recognition method based on the recurrence quantification analysis (RQA) and voted variable predictive model based class discriminate (V-VPMCD). This method applies strong robustness of RQA for the non-linear and non-stationary signals as well as the advantages for processing samples with inferior quality. Then, the prediction model based VPMCD is optimized by the method based on V-VPMCD. It has higher stability and recognition rate. The experiment-based fault recognition of different types and different degrees of rolling bearing demonstrates that the optimization algorithm has higher recognition rate and stability.