In order to improve the prediction accuracy of the residual lifetime of rolling bearings, a new model for predicting the residual lifetime of rolling bearings based on the fusion algorithm of variation modal decomposition (VMD) and gated recurrent neural networks (GRU) was proposed. Firstly, the original vibration signal was decomposed by the VMD algorithm which is optimized by Sparrow Search algorithm (SSA). Secondly, the effective modal components were selected by kurtosis criterion to extract the degenerate features. And then, the importance of the features of the stochastic forest algorithm was analyzed, and the decision table of the degradation features was constructed. In order to ensure the accuracy of the model, the super parameters in GRU were optimized by arithmetic optimization algorithm (AOA), and the residual lifetime prediction models of GRU were established according to different fault types. The XJTU-SY standard data set was used to validate the residual lifetime prediction, compared with the traditional method of extracting the degradation characteristics and establishing the prediction model without combining the fault types, the experimental results show that the root mean square error and mean absolute error of VMD-GRU model are reduced by 21.26% and 40.15%, respectively.